xref: /freebsd/sys/cddl/contrib/opensolaris/uts/common/dtrace/dtrace.c (revision 725a9f47324d42037db93c27ceb40d4956872f3e)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
24  * Copyright (c) 2016, Joyent, Inc. All rights reserved.
25  * Copyright (c) 2012, 2014 by Delphix. All rights reserved.
26  */
27 
28 /*
29  * DTrace - Dynamic Tracing for Solaris
30  *
31  * This is the implementation of the Solaris Dynamic Tracing framework
32  * (DTrace).  The user-visible interface to DTrace is described at length in
33  * the "Solaris Dynamic Tracing Guide".  The interfaces between the libdtrace
34  * library, the in-kernel DTrace framework, and the DTrace providers are
35  * described in the block comments in the <sys/dtrace.h> header file.  The
36  * internal architecture of DTrace is described in the block comments in the
37  * <sys/dtrace_impl.h> header file.  The comments contained within the DTrace
38  * implementation very much assume mastery of all of these sources; if one has
39  * an unanswered question about the implementation, one should consult them
40  * first.
41  *
42  * The functions here are ordered roughly as follows:
43  *
44  *   - Probe context functions
45  *   - Probe hashing functions
46  *   - Non-probe context utility functions
47  *   - Matching functions
48  *   - Provider-to-Framework API functions
49  *   - Probe management functions
50  *   - DIF object functions
51  *   - Format functions
52  *   - Predicate functions
53  *   - ECB functions
54  *   - Buffer functions
55  *   - Enabling functions
56  *   - DOF functions
57  *   - Anonymous enabling functions
58  *   - Consumer state functions
59  *   - Helper functions
60  *   - Hook functions
61  *   - Driver cookbook functions
62  *
63  * Each group of functions begins with a block comment labelled the "DTrace
64  * [Group] Functions", allowing one to find each block by searching forward
65  * on capital-f functions.
66  */
67 #include <sys/errno.h>
68 #include <sys/param.h>
69 #include <sys/types.h>
70 #ifndef illumos
71 #include <sys/time.h>
72 #endif
73 #include <sys/stat.h>
74 #include <sys/conf.h>
75 #include <sys/systm.h>
76 #include <sys/endian.h>
77 #ifdef illumos
78 #include <sys/ddi.h>
79 #include <sys/sunddi.h>
80 #endif
81 #include <sys/cpuvar.h>
82 #include <sys/kmem.h>
83 #ifdef illumos
84 #include <sys/strsubr.h>
85 #endif
86 #include <sys/sysmacros.h>
87 #include <sys/dtrace_impl.h>
88 #include <sys/atomic.h>
89 #include <sys/cmn_err.h>
90 #ifdef illumos
91 #include <sys/mutex_impl.h>
92 #include <sys/rwlock_impl.h>
93 #endif
94 #include <sys/ctf_api.h>
95 #ifdef illumos
96 #include <sys/panic.h>
97 #include <sys/priv_impl.h>
98 #endif
99 #ifdef illumos
100 #include <sys/cred_impl.h>
101 #include <sys/procfs_isa.h>
102 #endif
103 #include <sys/taskq.h>
104 #ifdef illumos
105 #include <sys/mkdev.h>
106 #include <sys/kdi.h>
107 #endif
108 #include <sys/zone.h>
109 #include <sys/socket.h>
110 #include <netinet/in.h>
111 #include "strtolctype.h"
112 
113 /* FreeBSD includes: */
114 #ifndef illumos
115 #include <sys/callout.h>
116 #include <sys/ctype.h>
117 #include <sys/eventhandler.h>
118 #include <sys/limits.h>
119 #include <sys/linker.h>
120 #include <sys/kdb.h>
121 #include <sys/jail.h>
122 #include <sys/kernel.h>
123 #include <sys/malloc.h>
124 #include <sys/lock.h>
125 #include <sys/mutex.h>
126 #include <sys/ptrace.h>
127 #include <sys/random.h>
128 #include <sys/rwlock.h>
129 #include <sys/sx.h>
130 #include <sys/sysctl.h>
131 
132 
133 #include <sys/mount.h>
134 #undef AT_UID
135 #undef AT_GID
136 #include <sys/vnode.h>
137 #include <sys/cred.h>
138 
139 #include <sys/dtrace_bsd.h>
140 
141 #include <netinet/in.h>
142 
143 #include "dtrace_cddl.h"
144 #include "dtrace_debug.c"
145 #endif
146 
147 #include "dtrace_xoroshiro128_plus.h"
148 
149 /*
150  * DTrace Tunable Variables
151  *
152  * The following variables may be tuned by adding a line to /etc/system that
153  * includes both the name of the DTrace module ("dtrace") and the name of the
154  * variable.  For example:
155  *
156  *   set dtrace:dtrace_destructive_disallow = 1
157  *
158  * In general, the only variables that one should be tuning this way are those
159  * that affect system-wide DTrace behavior, and for which the default behavior
160  * is undesirable.  Most of these variables are tunable on a per-consumer
161  * basis using DTrace options, and need not be tuned on a system-wide basis.
162  * When tuning these variables, avoid pathological values; while some attempt
163  * is made to verify the integrity of these variables, they are not considered
164  * part of the supported interface to DTrace, and they are therefore not
165  * checked comprehensively.  Further, these variables should not be tuned
166  * dynamically via "mdb -kw" or other means; they should only be tuned via
167  * /etc/system.
168  */
169 int		dtrace_destructive_disallow = 0;
170 #ifndef illumos
171 /* Positive logic version of dtrace_destructive_disallow for loader tunable */
172 int		dtrace_allow_destructive = 1;
173 #endif
174 dtrace_optval_t	dtrace_nonroot_maxsize = (16 * 1024 * 1024);
175 size_t		dtrace_difo_maxsize = (256 * 1024);
176 dtrace_optval_t	dtrace_dof_maxsize = (8 * 1024 * 1024);
177 size_t		dtrace_statvar_maxsize = (16 * 1024);
178 size_t		dtrace_actions_max = (16 * 1024);
179 size_t		dtrace_retain_max = 1024;
180 dtrace_optval_t	dtrace_helper_actions_max = 128;
181 dtrace_optval_t	dtrace_helper_providers_max = 32;
182 dtrace_optval_t	dtrace_dstate_defsize = (1 * 1024 * 1024);
183 size_t		dtrace_strsize_default = 256;
184 dtrace_optval_t	dtrace_cleanrate_default = 9900990;		/* 101 hz */
185 dtrace_optval_t	dtrace_cleanrate_min = 200000;			/* 5000 hz */
186 dtrace_optval_t	dtrace_cleanrate_max = (uint64_t)60 * NANOSEC;	/* 1/minute */
187 dtrace_optval_t	dtrace_aggrate_default = NANOSEC;		/* 1 hz */
188 dtrace_optval_t	dtrace_statusrate_default = NANOSEC;		/* 1 hz */
189 dtrace_optval_t dtrace_statusrate_max = (hrtime_t)10 * NANOSEC;	 /* 6/minute */
190 dtrace_optval_t	dtrace_switchrate_default = NANOSEC;		/* 1 hz */
191 dtrace_optval_t	dtrace_nspec_default = 1;
192 dtrace_optval_t	dtrace_specsize_default = 32 * 1024;
193 dtrace_optval_t dtrace_stackframes_default = 20;
194 dtrace_optval_t dtrace_ustackframes_default = 20;
195 dtrace_optval_t dtrace_jstackframes_default = 50;
196 dtrace_optval_t dtrace_jstackstrsize_default = 512;
197 int		dtrace_msgdsize_max = 128;
198 hrtime_t	dtrace_chill_max = MSEC2NSEC(500);		/* 500 ms */
199 hrtime_t	dtrace_chill_interval = NANOSEC;		/* 1000 ms */
200 int		dtrace_devdepth_max = 32;
201 int		dtrace_err_verbose;
202 hrtime_t	dtrace_deadman_interval = NANOSEC;
203 hrtime_t	dtrace_deadman_timeout = (hrtime_t)10 * NANOSEC;
204 hrtime_t	dtrace_deadman_user = (hrtime_t)30 * NANOSEC;
205 hrtime_t	dtrace_unregister_defunct_reap = (hrtime_t)60 * NANOSEC;
206 #ifndef illumos
207 int		dtrace_memstr_max = 4096;
208 int		dtrace_bufsize_max_frac = 128;
209 #endif
210 
211 /*
212  * DTrace External Variables
213  *
214  * As dtrace(7D) is a kernel module, any DTrace variables are obviously
215  * available to DTrace consumers via the backtick (`) syntax.  One of these,
216  * dtrace_zero, is made deliberately so:  it is provided as a source of
217  * well-known, zero-filled memory.  While this variable is not documented,
218  * it is used by some translators as an implementation detail.
219  */
220 const char	dtrace_zero[256] = { 0 };	/* zero-filled memory */
221 
222 /*
223  * DTrace Internal Variables
224  */
225 #ifdef illumos
226 static dev_info_t	*dtrace_devi;		/* device info */
227 #endif
228 #ifdef illumos
229 static vmem_t		*dtrace_arena;		/* probe ID arena */
230 static vmem_t		*dtrace_minor;		/* minor number arena */
231 #else
232 static taskq_t		*dtrace_taskq;		/* task queue */
233 static struct unrhdr	*dtrace_arena;		/* Probe ID number.     */
234 #endif
235 static dtrace_probe_t	**dtrace_probes;	/* array of all probes */
236 static int		dtrace_nprobes;		/* number of probes */
237 static dtrace_provider_t *dtrace_provider;	/* provider list */
238 static dtrace_meta_t	*dtrace_meta_pid;	/* user-land meta provider */
239 static int		dtrace_opens;		/* number of opens */
240 static int		dtrace_helpers;		/* number of helpers */
241 static int		dtrace_getf;		/* number of unpriv getf()s */
242 #ifdef illumos
243 static void		*dtrace_softstate;	/* softstate pointer */
244 #endif
245 static dtrace_hash_t	*dtrace_bymod;		/* probes hashed by module */
246 static dtrace_hash_t	*dtrace_byfunc;		/* probes hashed by function */
247 static dtrace_hash_t	*dtrace_byname;		/* probes hashed by name */
248 static dtrace_toxrange_t *dtrace_toxrange;	/* toxic range array */
249 static int		dtrace_toxranges;	/* number of toxic ranges */
250 static int		dtrace_toxranges_max;	/* size of toxic range array */
251 static dtrace_anon_t	dtrace_anon;		/* anonymous enabling */
252 static kmem_cache_t	*dtrace_state_cache;	/* cache for dynamic state */
253 static uint64_t		dtrace_vtime_references; /* number of vtimestamp refs */
254 static kthread_t	*dtrace_panicked;	/* panicking thread */
255 static dtrace_ecb_t	*dtrace_ecb_create_cache; /* cached created ECB */
256 static dtrace_genid_t	dtrace_probegen;	/* current probe generation */
257 static dtrace_helpers_t *dtrace_deferred_pid;	/* deferred helper list */
258 static dtrace_enabling_t *dtrace_retained;	/* list of retained enablings */
259 static dtrace_genid_t	dtrace_retained_gen;	/* current retained enab gen */
260 static dtrace_dynvar_t	dtrace_dynhash_sink;	/* end of dynamic hash chains */
261 static int		dtrace_dynvar_failclean; /* dynvars failed to clean */
262 #ifndef illumos
263 static struct mtx	dtrace_unr_mtx;
264 MTX_SYSINIT(dtrace_unr_mtx, &dtrace_unr_mtx, "Unique resource identifier", MTX_DEF);
265 static eventhandler_tag	dtrace_kld_load_tag;
266 static eventhandler_tag	dtrace_kld_unload_try_tag;
267 #endif
268 
269 /*
270  * DTrace Locking
271  * DTrace is protected by three (relatively coarse-grained) locks:
272  *
273  * (1) dtrace_lock is required to manipulate essentially any DTrace state,
274  *     including enabling state, probes, ECBs, consumer state, helper state,
275  *     etc.  Importantly, dtrace_lock is _not_ required when in probe context;
276  *     probe context is lock-free -- synchronization is handled via the
277  *     dtrace_sync() cross call mechanism.
278  *
279  * (2) dtrace_provider_lock is required when manipulating provider state, or
280  *     when provider state must be held constant.
281  *
282  * (3) dtrace_meta_lock is required when manipulating meta provider state, or
283  *     when meta provider state must be held constant.
284  *
285  * The lock ordering between these three locks is dtrace_meta_lock before
286  * dtrace_provider_lock before dtrace_lock.  (In particular, there are
287  * several places where dtrace_provider_lock is held by the framework as it
288  * calls into the providers -- which then call back into the framework,
289  * grabbing dtrace_lock.)
290  *
291  * There are two other locks in the mix:  mod_lock and cpu_lock.  With respect
292  * to dtrace_provider_lock and dtrace_lock, cpu_lock continues its historical
293  * role as a coarse-grained lock; it is acquired before both of these locks.
294  * With respect to dtrace_meta_lock, its behavior is stranger:  cpu_lock must
295  * be acquired _between_ dtrace_meta_lock and any other DTrace locks.
296  * mod_lock is similar with respect to dtrace_provider_lock in that it must be
297  * acquired _between_ dtrace_provider_lock and dtrace_lock.
298  */
299 static kmutex_t		dtrace_lock;		/* probe state lock */
300 static kmutex_t		dtrace_provider_lock;	/* provider state lock */
301 static kmutex_t		dtrace_meta_lock;	/* meta-provider state lock */
302 
303 #ifndef illumos
304 /* XXX FreeBSD hacks. */
305 #define cr_suid		cr_svuid
306 #define cr_sgid		cr_svgid
307 #define	ipaddr_t	in_addr_t
308 #define mod_modname	pathname
309 #define vuprintf	vprintf
310 #ifndef crgetzoneid
311 #define crgetzoneid(_a)        0
312 #endif
313 #define ttoproc(_a)	((_a)->td_proc)
314 #define SNOCD		0
315 #define CPU_ON_INTR(_a)	0
316 
317 #define PRIV_EFFECTIVE		(1 << 0)
318 #define PRIV_DTRACE_KERNEL	(1 << 1)
319 #define PRIV_DTRACE_PROC	(1 << 2)
320 #define PRIV_DTRACE_USER	(1 << 3)
321 #define PRIV_PROC_OWNER		(1 << 4)
322 #define PRIV_PROC_ZONE		(1 << 5)
323 #define PRIV_ALL		~0
324 
325 SYSCTL_DECL(_debug_dtrace);
326 SYSCTL_DECL(_kern_dtrace);
327 #endif
328 
329 #ifdef illumos
330 #define curcpu	CPU->cpu_id
331 #endif
332 
333 
334 /*
335  * DTrace Provider Variables
336  *
337  * These are the variables relating to DTrace as a provider (that is, the
338  * provider of the BEGIN, END, and ERROR probes).
339  */
340 static dtrace_pattr_t	dtrace_provider_attr = {
341 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
342 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
343 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
344 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
345 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
346 };
347 
348 static void
349 dtrace_nullop(void)
350 {}
351 
352 static dtrace_pops_t dtrace_provider_ops = {
353 	.dtps_provide =	(void (*)(void *, dtrace_probedesc_t *))dtrace_nullop,
354 	.dtps_provide_module =	(void (*)(void *, modctl_t *))dtrace_nullop,
355 	.dtps_enable =	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
356 	.dtps_disable =	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
357 	.dtps_suspend =	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
358 	.dtps_resume =	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
359 	.dtps_getargdesc =	NULL,
360 	.dtps_getargval =	NULL,
361 	.dtps_usermode =	NULL,
362 	.dtps_destroy =	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
363 };
364 
365 static dtrace_id_t	dtrace_probeid_begin;	/* special BEGIN probe */
366 static dtrace_id_t	dtrace_probeid_end;	/* special END probe */
367 dtrace_id_t		dtrace_probeid_error;	/* special ERROR probe */
368 
369 /*
370  * DTrace Helper Tracing Variables
371  *
372  * These variables should be set dynamically to enable helper tracing.  The
373  * only variables that should be set are dtrace_helptrace_enable (which should
374  * be set to a non-zero value to allocate helper tracing buffers on the next
375  * open of /dev/dtrace) and dtrace_helptrace_disable (which should be set to a
376  * non-zero value to deallocate helper tracing buffers on the next close of
377  * /dev/dtrace).  When (and only when) helper tracing is disabled, the
378  * buffer size may also be set via dtrace_helptrace_bufsize.
379  */
380 int			dtrace_helptrace_enable = 0;
381 int			dtrace_helptrace_disable = 0;
382 int			dtrace_helptrace_bufsize = 16 * 1024 * 1024;
383 uint32_t		dtrace_helptrace_nlocals;
384 static dtrace_helptrace_t *dtrace_helptrace_buffer;
385 static uint32_t		dtrace_helptrace_next = 0;
386 static int		dtrace_helptrace_wrapped = 0;
387 
388 /*
389  * DTrace Error Hashing
390  *
391  * On DEBUG kernels, DTrace will track the errors that has seen in a hash
392  * table.  This is very useful for checking coverage of tests that are
393  * expected to induce DIF or DOF processing errors, and may be useful for
394  * debugging problems in the DIF code generator or in DOF generation .  The
395  * error hash may be examined with the ::dtrace_errhash MDB dcmd.
396  */
397 #ifdef DEBUG
398 static dtrace_errhash_t	dtrace_errhash[DTRACE_ERRHASHSZ];
399 static const char *dtrace_errlast;
400 static kthread_t *dtrace_errthread;
401 static kmutex_t dtrace_errlock;
402 #endif
403 
404 /*
405  * DTrace Macros and Constants
406  *
407  * These are various macros that are useful in various spots in the
408  * implementation, along with a few random constants that have no meaning
409  * outside of the implementation.  There is no real structure to this cpp
410  * mishmash -- but is there ever?
411  */
412 #define	DTRACE_HASHSTR(hash, probe)	\
413 	dtrace_hash_str(*((char **)((uintptr_t)(probe) + (hash)->dth_stroffs)))
414 
415 #define	DTRACE_HASHNEXT(hash, probe)	\
416 	(dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_nextoffs)
417 
418 #define	DTRACE_HASHPREV(hash, probe)	\
419 	(dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_prevoffs)
420 
421 #define	DTRACE_HASHEQ(hash, lhs, rhs)	\
422 	(strcmp(*((char **)((uintptr_t)(lhs) + (hash)->dth_stroffs)), \
423 	    *((char **)((uintptr_t)(rhs) + (hash)->dth_stroffs))) == 0)
424 
425 #define	DTRACE_AGGHASHSIZE_SLEW		17
426 
427 #define	DTRACE_V4MAPPED_OFFSET		(sizeof (uint32_t) * 3)
428 
429 /*
430  * The key for a thread-local variable consists of the lower 61 bits of the
431  * t_did, plus the 3 bits of the highest active interrupt above LOCK_LEVEL.
432  * We add DIF_VARIABLE_MAX to t_did to assure that the thread key is never
433  * equal to a variable identifier.  This is necessary (but not sufficient) to
434  * assure that global associative arrays never collide with thread-local
435  * variables.  To guarantee that they cannot collide, we must also define the
436  * order for keying dynamic variables.  That order is:
437  *
438  *   [ key0 ] ... [ keyn ] [ variable-key ] [ tls-key ]
439  *
440  * Because the variable-key and the tls-key are in orthogonal spaces, there is
441  * no way for a global variable key signature to match a thread-local key
442  * signature.
443  */
444 #ifdef illumos
445 #define	DTRACE_TLS_THRKEY(where) { \
446 	uint_t intr = 0; \
447 	uint_t actv = CPU->cpu_intr_actv >> (LOCK_LEVEL + 1); \
448 	for (; actv; actv >>= 1) \
449 		intr++; \
450 	ASSERT(intr < (1 << 3)); \
451 	(where) = ((curthread->t_did + DIF_VARIABLE_MAX) & \
452 	    (((uint64_t)1 << 61) - 1)) | ((uint64_t)intr << 61); \
453 }
454 #else
455 #define	DTRACE_TLS_THRKEY(where) { \
456 	solaris_cpu_t *_c = &solaris_cpu[curcpu]; \
457 	uint_t intr = 0; \
458 	uint_t actv = _c->cpu_intr_actv; \
459 	for (; actv; actv >>= 1) \
460 		intr++; \
461 	ASSERT(intr < (1 << 3)); \
462 	(where) = ((curthread->td_tid + DIF_VARIABLE_MAX) & \
463 	    (((uint64_t)1 << 61) - 1)) | ((uint64_t)intr << 61); \
464 }
465 #endif
466 
467 #define	DT_BSWAP_8(x)	((x) & 0xff)
468 #define	DT_BSWAP_16(x)	((DT_BSWAP_8(x) << 8) | DT_BSWAP_8((x) >> 8))
469 #define	DT_BSWAP_32(x)	((DT_BSWAP_16(x) << 16) | DT_BSWAP_16((x) >> 16))
470 #define	DT_BSWAP_64(x)	((DT_BSWAP_32(x) << 32) | DT_BSWAP_32((x) >> 32))
471 
472 #define	DT_MASK_LO 0x00000000FFFFFFFFULL
473 
474 #define	DTRACE_STORE(type, tomax, offset, what) \
475 	*((type *)((uintptr_t)(tomax) + (uintptr_t)offset)) = (type)(what);
476 
477 #if !defined(__x86) && !defined(__aarch64__)
478 #define	DTRACE_ALIGNCHECK(addr, size, flags)				\
479 	if (addr & (size - 1)) {					\
480 		*flags |= CPU_DTRACE_BADALIGN;				\
481 		cpu_core[curcpu].cpuc_dtrace_illval = addr;	\
482 		return (0);						\
483 	}
484 #else
485 #define	DTRACE_ALIGNCHECK(addr, size, flags)
486 #endif
487 
488 /*
489  * Test whether a range of memory starting at testaddr of size testsz falls
490  * within the range of memory described by addr, sz.  We take care to avoid
491  * problems with overflow and underflow of the unsigned quantities, and
492  * disallow all negative sizes.  Ranges of size 0 are allowed.
493  */
494 #define	DTRACE_INRANGE(testaddr, testsz, baseaddr, basesz) \
495 	((testaddr) - (uintptr_t)(baseaddr) < (basesz) && \
496 	(testaddr) + (testsz) - (uintptr_t)(baseaddr) <= (basesz) && \
497 	(testaddr) + (testsz) >= (testaddr))
498 
499 #define	DTRACE_RANGE_REMAIN(remp, addr, baseaddr, basesz)		\
500 do {									\
501 	if ((remp) != NULL) {						\
502 		*(remp) = (uintptr_t)(baseaddr) + (basesz) - (addr);	\
503 	}								\
504 } while (0)
505 
506 
507 /*
508  * Test whether alloc_sz bytes will fit in the scratch region.  We isolate
509  * alloc_sz on the righthand side of the comparison in order to avoid overflow
510  * or underflow in the comparison with it.  This is simpler than the INRANGE
511  * check above, because we know that the dtms_scratch_ptr is valid in the
512  * range.  Allocations of size zero are allowed.
513  */
514 #define	DTRACE_INSCRATCH(mstate, alloc_sz) \
515 	((mstate)->dtms_scratch_base + (mstate)->dtms_scratch_size - \
516 	(mstate)->dtms_scratch_ptr >= (alloc_sz))
517 
518 #define DTRACE_INSCRATCHPTR(mstate, ptr, howmany) \
519 	((ptr) >= (mstate)->dtms_scratch_base && \
520 	(ptr) <= \
521 	((mstate)->dtms_scratch_base + (mstate)->dtms_scratch_size - (howmany)))
522 
523 #define	DTRACE_LOADFUNC(bits)						\
524 /*CSTYLED*/								\
525 uint##bits##_t								\
526 dtrace_load##bits(uintptr_t addr)					\
527 {									\
528 	size_t size = bits / NBBY;					\
529 	/*CSTYLED*/							\
530 	uint##bits##_t rval;						\
531 	int i;								\
532 	volatile uint16_t *flags = (volatile uint16_t *)		\
533 	    &cpu_core[curcpu].cpuc_dtrace_flags;			\
534 									\
535 	DTRACE_ALIGNCHECK(addr, size, flags);				\
536 									\
537 	for (i = 0; i < dtrace_toxranges; i++) {			\
538 		if (addr >= dtrace_toxrange[i].dtt_limit)		\
539 			continue;					\
540 									\
541 		if (addr + size <= dtrace_toxrange[i].dtt_base)		\
542 			continue;					\
543 									\
544 		/*							\
545 		 * This address falls within a toxic region; return 0.	\
546 		 */							\
547 		*flags |= CPU_DTRACE_BADADDR;				\
548 		cpu_core[curcpu].cpuc_dtrace_illval = addr;		\
549 		return (0);						\
550 	}								\
551 									\
552 	*flags |= CPU_DTRACE_NOFAULT;					\
553 	/*CSTYLED*/							\
554 	rval = *((volatile uint##bits##_t *)addr);			\
555 	*flags &= ~CPU_DTRACE_NOFAULT;					\
556 									\
557 	return (!(*flags & CPU_DTRACE_FAULT) ? rval : 0);		\
558 }
559 
560 #ifdef _LP64
561 #define	dtrace_loadptr	dtrace_load64
562 #else
563 #define	dtrace_loadptr	dtrace_load32
564 #endif
565 
566 #define	DTRACE_DYNHASH_FREE	0
567 #define	DTRACE_DYNHASH_SINK	1
568 #define	DTRACE_DYNHASH_VALID	2
569 
570 #define	DTRACE_MATCH_NEXT	0
571 #define	DTRACE_MATCH_DONE	1
572 #define	DTRACE_ANCHORED(probe)	((probe)->dtpr_func[0] != '\0')
573 #define	DTRACE_STATE_ALIGN	64
574 
575 #define	DTRACE_FLAGS2FLT(flags)						\
576 	(((flags) & CPU_DTRACE_BADADDR) ? DTRACEFLT_BADADDR :		\
577 	((flags) & CPU_DTRACE_ILLOP) ? DTRACEFLT_ILLOP :		\
578 	((flags) & CPU_DTRACE_DIVZERO) ? DTRACEFLT_DIVZERO :		\
579 	((flags) & CPU_DTRACE_KPRIV) ? DTRACEFLT_KPRIV :		\
580 	((flags) & CPU_DTRACE_UPRIV) ? DTRACEFLT_UPRIV :		\
581 	((flags) & CPU_DTRACE_TUPOFLOW) ?  DTRACEFLT_TUPOFLOW :		\
582 	((flags) & CPU_DTRACE_BADALIGN) ?  DTRACEFLT_BADALIGN :		\
583 	((flags) & CPU_DTRACE_NOSCRATCH) ?  DTRACEFLT_NOSCRATCH :	\
584 	((flags) & CPU_DTRACE_BADSTACK) ?  DTRACEFLT_BADSTACK :		\
585 	DTRACEFLT_UNKNOWN)
586 
587 #define	DTRACEACT_ISSTRING(act)						\
588 	((act)->dta_kind == DTRACEACT_DIFEXPR &&			\
589 	(act)->dta_difo->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING)
590 
591 /* Function prototype definitions: */
592 static size_t dtrace_strlen(const char *, size_t);
593 static dtrace_probe_t *dtrace_probe_lookup_id(dtrace_id_t id);
594 static void dtrace_enabling_provide(dtrace_provider_t *);
595 static int dtrace_enabling_match(dtrace_enabling_t *, int *);
596 static void dtrace_enabling_matchall(void);
597 static void dtrace_enabling_reap(void);
598 static dtrace_state_t *dtrace_anon_grab(void);
599 static uint64_t dtrace_helper(int, dtrace_mstate_t *,
600     dtrace_state_t *, uint64_t, uint64_t);
601 static dtrace_helpers_t *dtrace_helpers_create(proc_t *);
602 static void dtrace_buffer_drop(dtrace_buffer_t *);
603 static int dtrace_buffer_consumed(dtrace_buffer_t *, hrtime_t when);
604 static intptr_t dtrace_buffer_reserve(dtrace_buffer_t *, size_t, size_t,
605     dtrace_state_t *, dtrace_mstate_t *);
606 static int dtrace_state_option(dtrace_state_t *, dtrace_optid_t,
607     dtrace_optval_t);
608 static int dtrace_ecb_create_enable(dtrace_probe_t *, void *);
609 static void dtrace_helper_provider_destroy(dtrace_helper_provider_t *);
610 uint16_t dtrace_load16(uintptr_t);
611 uint32_t dtrace_load32(uintptr_t);
612 uint64_t dtrace_load64(uintptr_t);
613 uint8_t dtrace_load8(uintptr_t);
614 void dtrace_dynvar_clean(dtrace_dstate_t *);
615 dtrace_dynvar_t *dtrace_dynvar(dtrace_dstate_t *, uint_t, dtrace_key_t *,
616     size_t, dtrace_dynvar_op_t, dtrace_mstate_t *, dtrace_vstate_t *);
617 uintptr_t dtrace_dif_varstr(uintptr_t, dtrace_state_t *, dtrace_mstate_t *);
618 static int dtrace_priv_proc(dtrace_state_t *);
619 static void dtrace_getf_barrier(void);
620 static int dtrace_canload_remains(uint64_t, size_t, size_t *,
621     dtrace_mstate_t *, dtrace_vstate_t *);
622 static int dtrace_canstore_remains(uint64_t, size_t, size_t *,
623     dtrace_mstate_t *, dtrace_vstate_t *);
624 
625 /*
626  * DTrace Probe Context Functions
627  *
628  * These functions are called from probe context.  Because probe context is
629  * any context in which C may be called, arbitrarily locks may be held,
630  * interrupts may be disabled, we may be in arbitrary dispatched state, etc.
631  * As a result, functions called from probe context may only call other DTrace
632  * support functions -- they may not interact at all with the system at large.
633  * (Note that the ASSERT macro is made probe-context safe by redefining it in
634  * terms of dtrace_assfail(), a probe-context safe function.) If arbitrary
635  * loads are to be performed from probe context, they _must_ be in terms of
636  * the safe dtrace_load*() variants.
637  *
638  * Some functions in this block are not actually called from probe context;
639  * for these functions, there will be a comment above the function reading
640  * "Note:  not called from probe context."
641  */
642 void
643 dtrace_panic(const char *format, ...)
644 {
645 	va_list alist;
646 
647 	va_start(alist, format);
648 #ifdef __FreeBSD__
649 	vpanic(format, alist);
650 #else
651 	dtrace_vpanic(format, alist);
652 #endif
653 	va_end(alist);
654 }
655 
656 int
657 dtrace_assfail(const char *a, const char *f, int l)
658 {
659 	dtrace_panic("assertion failed: %s, file: %s, line: %d", a, f, l);
660 
661 	/*
662 	 * We just need something here that even the most clever compiler
663 	 * cannot optimize away.
664 	 */
665 	return (a[(uintptr_t)f]);
666 }
667 
668 /*
669  * Atomically increment a specified error counter from probe context.
670  */
671 static void
672 dtrace_error(uint32_t *counter)
673 {
674 	/*
675 	 * Most counters stored to in probe context are per-CPU counters.
676 	 * However, there are some error conditions that are sufficiently
677 	 * arcane that they don't merit per-CPU storage.  If these counters
678 	 * are incremented concurrently on different CPUs, scalability will be
679 	 * adversely affected -- but we don't expect them to be white-hot in a
680 	 * correctly constructed enabling...
681 	 */
682 	uint32_t oval, nval;
683 
684 	do {
685 		oval = *counter;
686 
687 		if ((nval = oval + 1) == 0) {
688 			/*
689 			 * If the counter would wrap, set it to 1 -- assuring
690 			 * that the counter is never zero when we have seen
691 			 * errors.  (The counter must be 32-bits because we
692 			 * aren't guaranteed a 64-bit compare&swap operation.)
693 			 * To save this code both the infamy of being fingered
694 			 * by a priggish news story and the indignity of being
695 			 * the target of a neo-puritan witch trial, we're
696 			 * carefully avoiding any colorful description of the
697 			 * likelihood of this condition -- but suffice it to
698 			 * say that it is only slightly more likely than the
699 			 * overflow of predicate cache IDs, as discussed in
700 			 * dtrace_predicate_create().
701 			 */
702 			nval = 1;
703 		}
704 	} while (dtrace_cas32(counter, oval, nval) != oval);
705 }
706 
707 /*
708  * Use the DTRACE_LOADFUNC macro to define functions for each of loading a
709  * uint8_t, a uint16_t, a uint32_t and a uint64_t.
710  */
711 /* BEGIN CSTYLED */
712 DTRACE_LOADFUNC(8)
713 DTRACE_LOADFUNC(16)
714 DTRACE_LOADFUNC(32)
715 DTRACE_LOADFUNC(64)
716 /* END CSTYLED */
717 
718 static int
719 dtrace_inscratch(uintptr_t dest, size_t size, dtrace_mstate_t *mstate)
720 {
721 	if (dest < mstate->dtms_scratch_base)
722 		return (0);
723 
724 	if (dest + size < dest)
725 		return (0);
726 
727 	if (dest + size > mstate->dtms_scratch_ptr)
728 		return (0);
729 
730 	return (1);
731 }
732 
733 static int
734 dtrace_canstore_statvar(uint64_t addr, size_t sz, size_t *remain,
735     dtrace_statvar_t **svars, int nsvars)
736 {
737 	int i;
738 	size_t maxglobalsize, maxlocalsize;
739 
740 	if (nsvars == 0)
741 		return (0);
742 
743 	maxglobalsize = dtrace_statvar_maxsize + sizeof (uint64_t);
744 	maxlocalsize = maxglobalsize * NCPU;
745 
746 	for (i = 0; i < nsvars; i++) {
747 		dtrace_statvar_t *svar = svars[i];
748 		uint8_t scope;
749 		size_t size;
750 
751 		if (svar == NULL || (size = svar->dtsv_size) == 0)
752 			continue;
753 
754 		scope = svar->dtsv_var.dtdv_scope;
755 
756 		/*
757 		 * We verify that our size is valid in the spirit of providing
758 		 * defense in depth:  we want to prevent attackers from using
759 		 * DTrace to escalate an orthogonal kernel heap corruption bug
760 		 * into the ability to store to arbitrary locations in memory.
761 		 */
762 		VERIFY((scope == DIFV_SCOPE_GLOBAL && size <= maxglobalsize) ||
763 		    (scope == DIFV_SCOPE_LOCAL && size <= maxlocalsize));
764 
765 		if (DTRACE_INRANGE(addr, sz, svar->dtsv_data,
766 		    svar->dtsv_size)) {
767 			DTRACE_RANGE_REMAIN(remain, addr, svar->dtsv_data,
768 			    svar->dtsv_size);
769 			return (1);
770 		}
771 	}
772 
773 	return (0);
774 }
775 
776 /*
777  * Check to see if the address is within a memory region to which a store may
778  * be issued.  This includes the DTrace scratch areas, and any DTrace variable
779  * region.  The caller of dtrace_canstore() is responsible for performing any
780  * alignment checks that are needed before stores are actually executed.
781  */
782 static int
783 dtrace_canstore(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
784     dtrace_vstate_t *vstate)
785 {
786 	return (dtrace_canstore_remains(addr, sz, NULL, mstate, vstate));
787 }
788 
789 /*
790  * Implementation of dtrace_canstore which communicates the upper bound of the
791  * allowed memory region.
792  */
793 static int
794 dtrace_canstore_remains(uint64_t addr, size_t sz, size_t *remain,
795     dtrace_mstate_t *mstate, dtrace_vstate_t *vstate)
796 {
797 	/*
798 	 * First, check to see if the address is in scratch space...
799 	 */
800 	if (DTRACE_INRANGE(addr, sz, mstate->dtms_scratch_base,
801 	    mstate->dtms_scratch_size)) {
802 		DTRACE_RANGE_REMAIN(remain, addr, mstate->dtms_scratch_base,
803 		    mstate->dtms_scratch_size);
804 		return (1);
805 	}
806 
807 	/*
808 	 * Now check to see if it's a dynamic variable.  This check will pick
809 	 * up both thread-local variables and any global dynamically-allocated
810 	 * variables.
811 	 */
812 	if (DTRACE_INRANGE(addr, sz, vstate->dtvs_dynvars.dtds_base,
813 	    vstate->dtvs_dynvars.dtds_size)) {
814 		dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
815 		uintptr_t base = (uintptr_t)dstate->dtds_base +
816 		    (dstate->dtds_hashsize * sizeof (dtrace_dynhash_t));
817 		uintptr_t chunkoffs;
818 		dtrace_dynvar_t *dvar;
819 
820 		/*
821 		 * Before we assume that we can store here, we need to make
822 		 * sure that it isn't in our metadata -- storing to our
823 		 * dynamic variable metadata would corrupt our state.  For
824 		 * the range to not include any dynamic variable metadata,
825 		 * it must:
826 		 *
827 		 *	(1) Start above the hash table that is at the base of
828 		 *	the dynamic variable space
829 		 *
830 		 *	(2) Have a starting chunk offset that is beyond the
831 		 *	dtrace_dynvar_t that is at the base of every chunk
832 		 *
833 		 *	(3) Not span a chunk boundary
834 		 *
835 		 *	(4) Not be in the tuple space of a dynamic variable
836 		 *
837 		 */
838 		if (addr < base)
839 			return (0);
840 
841 		chunkoffs = (addr - base) % dstate->dtds_chunksize;
842 
843 		if (chunkoffs < sizeof (dtrace_dynvar_t))
844 			return (0);
845 
846 		if (chunkoffs + sz > dstate->dtds_chunksize)
847 			return (0);
848 
849 		dvar = (dtrace_dynvar_t *)((uintptr_t)addr - chunkoffs);
850 
851 		if (dvar->dtdv_hashval == DTRACE_DYNHASH_FREE)
852 			return (0);
853 
854 		if (chunkoffs < sizeof (dtrace_dynvar_t) +
855 		    ((dvar->dtdv_tuple.dtt_nkeys - 1) * sizeof (dtrace_key_t)))
856 			return (0);
857 
858 		DTRACE_RANGE_REMAIN(remain, addr, dvar, dstate->dtds_chunksize);
859 		return (1);
860 	}
861 
862 	/*
863 	 * Finally, check the static local and global variables.  These checks
864 	 * take the longest, so we perform them last.
865 	 */
866 	if (dtrace_canstore_statvar(addr, sz, remain,
867 	    vstate->dtvs_locals, vstate->dtvs_nlocals))
868 		return (1);
869 
870 	if (dtrace_canstore_statvar(addr, sz, remain,
871 	    vstate->dtvs_globals, vstate->dtvs_nglobals))
872 		return (1);
873 
874 	return (0);
875 }
876 
877 
878 /*
879  * Convenience routine to check to see if the address is within a memory
880  * region in which a load may be issued given the user's privilege level;
881  * if not, it sets the appropriate error flags and loads 'addr' into the
882  * illegal value slot.
883  *
884  * DTrace subroutines (DIF_SUBR_*) should use this helper to implement
885  * appropriate memory access protection.
886  */
887 static int
888 dtrace_canload(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
889     dtrace_vstate_t *vstate)
890 {
891 	return (dtrace_canload_remains(addr, sz, NULL, mstate, vstate));
892 }
893 
894 /*
895  * Implementation of dtrace_canload which communicates the uppoer bound of the
896  * allowed memory region.
897  */
898 static int
899 dtrace_canload_remains(uint64_t addr, size_t sz, size_t *remain,
900     dtrace_mstate_t *mstate, dtrace_vstate_t *vstate)
901 {
902 	volatile uintptr_t *illval = &cpu_core[curcpu].cpuc_dtrace_illval;
903 	file_t *fp;
904 
905 	/*
906 	 * If we hold the privilege to read from kernel memory, then
907 	 * everything is readable.
908 	 */
909 	if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0) {
910 		DTRACE_RANGE_REMAIN(remain, addr, addr, sz);
911 		return (1);
912 	}
913 
914 	/*
915 	 * You can obviously read that which you can store.
916 	 */
917 	if (dtrace_canstore_remains(addr, sz, remain, mstate, vstate))
918 		return (1);
919 
920 	/*
921 	 * We're allowed to read from our own string table.
922 	 */
923 	if (DTRACE_INRANGE(addr, sz, mstate->dtms_difo->dtdo_strtab,
924 	    mstate->dtms_difo->dtdo_strlen)) {
925 		DTRACE_RANGE_REMAIN(remain, addr,
926 		    mstate->dtms_difo->dtdo_strtab,
927 		    mstate->dtms_difo->dtdo_strlen);
928 		return (1);
929 	}
930 
931 	if (vstate->dtvs_state != NULL &&
932 	    dtrace_priv_proc(vstate->dtvs_state)) {
933 		proc_t *p;
934 
935 		/*
936 		 * When we have privileges to the current process, there are
937 		 * several context-related kernel structures that are safe to
938 		 * read, even absent the privilege to read from kernel memory.
939 		 * These reads are safe because these structures contain only
940 		 * state that (1) we're permitted to read, (2) is harmless or
941 		 * (3) contains pointers to additional kernel state that we're
942 		 * not permitted to read (and as such, do not present an
943 		 * opportunity for privilege escalation).  Finally (and
944 		 * critically), because of the nature of their relation with
945 		 * the current thread context, the memory associated with these
946 		 * structures cannot change over the duration of probe context,
947 		 * and it is therefore impossible for this memory to be
948 		 * deallocated and reallocated as something else while it's
949 		 * being operated upon.
950 		 */
951 		if (DTRACE_INRANGE(addr, sz, curthread, sizeof (kthread_t))) {
952 			DTRACE_RANGE_REMAIN(remain, addr, curthread,
953 			    sizeof (kthread_t));
954 			return (1);
955 		}
956 
957 		if ((p = curthread->t_procp) != NULL && DTRACE_INRANGE(addr,
958 		    sz, curthread->t_procp, sizeof (proc_t))) {
959 			DTRACE_RANGE_REMAIN(remain, addr, curthread->t_procp,
960 			    sizeof (proc_t));
961 			return (1);
962 		}
963 
964 		if (curthread->t_cred != NULL && DTRACE_INRANGE(addr, sz,
965 		    curthread->t_cred, sizeof (cred_t))) {
966 			DTRACE_RANGE_REMAIN(remain, addr, curthread->t_cred,
967 			    sizeof (cred_t));
968 			return (1);
969 		}
970 
971 #ifdef illumos
972 		if (p != NULL && p->p_pidp != NULL && DTRACE_INRANGE(addr, sz,
973 		    &(p->p_pidp->pid_id), sizeof (pid_t))) {
974 			DTRACE_RANGE_REMAIN(remain, addr, &(p->p_pidp->pid_id),
975 			    sizeof (pid_t));
976 			return (1);
977 		}
978 
979 		if (curthread->t_cpu != NULL && DTRACE_INRANGE(addr, sz,
980 		    curthread->t_cpu, offsetof(cpu_t, cpu_pause_thread))) {
981 			DTRACE_RANGE_REMAIN(remain, addr, curthread->t_cpu,
982 			    offsetof(cpu_t, cpu_pause_thread));
983 			return (1);
984 		}
985 #endif
986 	}
987 
988 	if ((fp = mstate->dtms_getf) != NULL) {
989 		uintptr_t psz = sizeof (void *);
990 		vnode_t *vp;
991 		vnodeops_t *op;
992 
993 		/*
994 		 * When getf() returns a file_t, the enabling is implicitly
995 		 * granted the (transient) right to read the returned file_t
996 		 * as well as the v_path and v_op->vnop_name of the underlying
997 		 * vnode.  These accesses are allowed after a successful
998 		 * getf() because the members that they refer to cannot change
999 		 * once set -- and the barrier logic in the kernel's closef()
1000 		 * path assures that the file_t and its referenced vode_t
1001 		 * cannot themselves be stale (that is, it impossible for
1002 		 * either dtms_getf itself or its f_vnode member to reference
1003 		 * freed memory).
1004 		 */
1005 		if (DTRACE_INRANGE(addr, sz, fp, sizeof (file_t))) {
1006 			DTRACE_RANGE_REMAIN(remain, addr, fp, sizeof (file_t));
1007 			return (1);
1008 		}
1009 
1010 		if ((vp = fp->f_vnode) != NULL) {
1011 			size_t slen;
1012 #ifdef illumos
1013 			if (DTRACE_INRANGE(addr, sz, &vp->v_path, psz)) {
1014 				DTRACE_RANGE_REMAIN(remain, addr, &vp->v_path,
1015 				    psz);
1016 				return (1);
1017 			}
1018 			slen = strlen(vp->v_path) + 1;
1019 			if (DTRACE_INRANGE(addr, sz, vp->v_path, slen)) {
1020 				DTRACE_RANGE_REMAIN(remain, addr, vp->v_path,
1021 				    slen);
1022 				return (1);
1023 			}
1024 #endif
1025 
1026 			if (DTRACE_INRANGE(addr, sz, &vp->v_op, psz)) {
1027 				DTRACE_RANGE_REMAIN(remain, addr, &vp->v_op,
1028 				    psz);
1029 				return (1);
1030 			}
1031 
1032 #ifdef illumos
1033 			if ((op = vp->v_op) != NULL &&
1034 			    DTRACE_INRANGE(addr, sz, &op->vnop_name, psz)) {
1035 				DTRACE_RANGE_REMAIN(remain, addr,
1036 				    &op->vnop_name, psz);
1037 				return (1);
1038 			}
1039 
1040 			if (op != NULL && op->vnop_name != NULL &&
1041 			    DTRACE_INRANGE(addr, sz, op->vnop_name,
1042 			    (slen = strlen(op->vnop_name) + 1))) {
1043 				DTRACE_RANGE_REMAIN(remain, addr,
1044 				    op->vnop_name, slen);
1045 				return (1);
1046 			}
1047 #endif
1048 		}
1049 	}
1050 
1051 	DTRACE_CPUFLAG_SET(CPU_DTRACE_KPRIV);
1052 	*illval = addr;
1053 	return (0);
1054 }
1055 
1056 /*
1057  * Convenience routine to check to see if a given string is within a memory
1058  * region in which a load may be issued given the user's privilege level;
1059  * this exists so that we don't need to issue unnecessary dtrace_strlen()
1060  * calls in the event that the user has all privileges.
1061  */
1062 static int
1063 dtrace_strcanload(uint64_t addr, size_t sz, size_t *remain,
1064     dtrace_mstate_t *mstate, dtrace_vstate_t *vstate)
1065 {
1066 	size_t rsize;
1067 
1068 	/*
1069 	 * If we hold the privilege to read from kernel memory, then
1070 	 * everything is readable.
1071 	 */
1072 	if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0) {
1073 		DTRACE_RANGE_REMAIN(remain, addr, addr, sz);
1074 		return (1);
1075 	}
1076 
1077 	/*
1078 	 * Even if the caller is uninterested in querying the remaining valid
1079 	 * range, it is required to ensure that the access is allowed.
1080 	 */
1081 	if (remain == NULL) {
1082 		remain = &rsize;
1083 	}
1084 	if (dtrace_canload_remains(addr, 0, remain, mstate, vstate)) {
1085 		size_t strsz;
1086 		/*
1087 		 * Perform the strlen after determining the length of the
1088 		 * memory region which is accessible.  This prevents timing
1089 		 * information from being used to find NULs in memory which is
1090 		 * not accessible to the caller.
1091 		 */
1092 		strsz = 1 + dtrace_strlen((char *)(uintptr_t)addr,
1093 		    MIN(sz, *remain));
1094 		if (strsz <= *remain) {
1095 			return (1);
1096 		}
1097 	}
1098 
1099 	return (0);
1100 }
1101 
1102 /*
1103  * Convenience routine to check to see if a given variable is within a memory
1104  * region in which a load may be issued given the user's privilege level.
1105  */
1106 static int
1107 dtrace_vcanload(void *src, dtrace_diftype_t *type, size_t *remain,
1108     dtrace_mstate_t *mstate, dtrace_vstate_t *vstate)
1109 {
1110 	size_t sz;
1111 	ASSERT(type->dtdt_flags & DIF_TF_BYREF);
1112 
1113 	/*
1114 	 * Calculate the max size before performing any checks since even
1115 	 * DTRACE_ACCESS_KERNEL-credentialed callers expect that this function
1116 	 * return the max length via 'remain'.
1117 	 */
1118 	if (type->dtdt_kind == DIF_TYPE_STRING) {
1119 		dtrace_state_t *state = vstate->dtvs_state;
1120 
1121 		if (state != NULL) {
1122 			sz = state->dts_options[DTRACEOPT_STRSIZE];
1123 		} else {
1124 			/*
1125 			 * In helper context, we have a NULL state; fall back
1126 			 * to using the system-wide default for the string size
1127 			 * in this case.
1128 			 */
1129 			sz = dtrace_strsize_default;
1130 		}
1131 	} else {
1132 		sz = type->dtdt_size;
1133 	}
1134 
1135 	/*
1136 	 * If we hold the privilege to read from kernel memory, then
1137 	 * everything is readable.
1138 	 */
1139 	if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0) {
1140 		DTRACE_RANGE_REMAIN(remain, (uintptr_t)src, src, sz);
1141 		return (1);
1142 	}
1143 
1144 	if (type->dtdt_kind == DIF_TYPE_STRING) {
1145 		return (dtrace_strcanload((uintptr_t)src, sz, remain, mstate,
1146 		    vstate));
1147 	}
1148 	return (dtrace_canload_remains((uintptr_t)src, sz, remain, mstate,
1149 	    vstate));
1150 }
1151 
1152 /*
1153  * Convert a string to a signed integer using safe loads.
1154  *
1155  * NOTE: This function uses various macros from strtolctype.h to manipulate
1156  * digit values, etc -- these have all been checked to ensure they make
1157  * no additional function calls.
1158  */
1159 static int64_t
1160 dtrace_strtoll(char *input, int base, size_t limit)
1161 {
1162 	uintptr_t pos = (uintptr_t)input;
1163 	int64_t val = 0;
1164 	int x;
1165 	boolean_t neg = B_FALSE;
1166 	char c, cc, ccc;
1167 	uintptr_t end = pos + limit;
1168 
1169 	/*
1170 	 * Consume any whitespace preceding digits.
1171 	 */
1172 	while ((c = dtrace_load8(pos)) == ' ' || c == '\t')
1173 		pos++;
1174 
1175 	/*
1176 	 * Handle an explicit sign if one is present.
1177 	 */
1178 	if (c == '-' || c == '+') {
1179 		if (c == '-')
1180 			neg = B_TRUE;
1181 		c = dtrace_load8(++pos);
1182 	}
1183 
1184 	/*
1185 	 * Check for an explicit hexadecimal prefix ("0x" or "0X") and skip it
1186 	 * if present.
1187 	 */
1188 	if (base == 16 && c == '0' && ((cc = dtrace_load8(pos + 1)) == 'x' ||
1189 	    cc == 'X') && isxdigit(ccc = dtrace_load8(pos + 2))) {
1190 		pos += 2;
1191 		c = ccc;
1192 	}
1193 
1194 	/*
1195 	 * Read in contiguous digits until the first non-digit character.
1196 	 */
1197 	for (; pos < end && c != '\0' && lisalnum(c) && (x = DIGIT(c)) < base;
1198 	    c = dtrace_load8(++pos))
1199 		val = val * base + x;
1200 
1201 	return (neg ? -val : val);
1202 }
1203 
1204 /*
1205  * Compare two strings using safe loads.
1206  */
1207 static int
1208 dtrace_strncmp(char *s1, char *s2, size_t limit)
1209 {
1210 	uint8_t c1, c2;
1211 	volatile uint16_t *flags;
1212 
1213 	if (s1 == s2 || limit == 0)
1214 		return (0);
1215 
1216 	flags = (volatile uint16_t *)&cpu_core[curcpu].cpuc_dtrace_flags;
1217 
1218 	do {
1219 		if (s1 == NULL) {
1220 			c1 = '\0';
1221 		} else {
1222 			c1 = dtrace_load8((uintptr_t)s1++);
1223 		}
1224 
1225 		if (s2 == NULL) {
1226 			c2 = '\0';
1227 		} else {
1228 			c2 = dtrace_load8((uintptr_t)s2++);
1229 		}
1230 
1231 		if (c1 != c2)
1232 			return (c1 - c2);
1233 	} while (--limit && c1 != '\0' && !(*flags & CPU_DTRACE_FAULT));
1234 
1235 	return (0);
1236 }
1237 
1238 /*
1239  * Compute strlen(s) for a string using safe memory accesses.  The additional
1240  * len parameter is used to specify a maximum length to ensure completion.
1241  */
1242 static size_t
1243 dtrace_strlen(const char *s, size_t lim)
1244 {
1245 	uint_t len;
1246 
1247 	for (len = 0; len != lim; len++) {
1248 		if (dtrace_load8((uintptr_t)s++) == '\0')
1249 			break;
1250 	}
1251 
1252 	return (len);
1253 }
1254 
1255 /*
1256  * Check if an address falls within a toxic region.
1257  */
1258 static int
1259 dtrace_istoxic(uintptr_t kaddr, size_t size)
1260 {
1261 	uintptr_t taddr, tsize;
1262 	int i;
1263 
1264 	for (i = 0; i < dtrace_toxranges; i++) {
1265 		taddr = dtrace_toxrange[i].dtt_base;
1266 		tsize = dtrace_toxrange[i].dtt_limit - taddr;
1267 
1268 		if (kaddr - taddr < tsize) {
1269 			DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
1270 			cpu_core[curcpu].cpuc_dtrace_illval = kaddr;
1271 			return (1);
1272 		}
1273 
1274 		if (taddr - kaddr < size) {
1275 			DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
1276 			cpu_core[curcpu].cpuc_dtrace_illval = taddr;
1277 			return (1);
1278 		}
1279 	}
1280 
1281 	return (0);
1282 }
1283 
1284 /*
1285  * Copy src to dst using safe memory accesses.  The src is assumed to be unsafe
1286  * memory specified by the DIF program.  The dst is assumed to be safe memory
1287  * that we can store to directly because it is managed by DTrace.  As with
1288  * standard bcopy, overlapping copies are handled properly.
1289  */
1290 static void
1291 dtrace_bcopy(const void *src, void *dst, size_t len)
1292 {
1293 	if (len != 0) {
1294 		uint8_t *s1 = dst;
1295 		const uint8_t *s2 = src;
1296 
1297 		if (s1 <= s2) {
1298 			do {
1299 				*s1++ = dtrace_load8((uintptr_t)s2++);
1300 			} while (--len != 0);
1301 		} else {
1302 			s2 += len;
1303 			s1 += len;
1304 
1305 			do {
1306 				*--s1 = dtrace_load8((uintptr_t)--s2);
1307 			} while (--len != 0);
1308 		}
1309 	}
1310 }
1311 
1312 /*
1313  * Copy src to dst using safe memory accesses, up to either the specified
1314  * length, or the point that a nul byte is encountered.  The src is assumed to
1315  * be unsafe memory specified by the DIF program.  The dst is assumed to be
1316  * safe memory that we can store to directly because it is managed by DTrace.
1317  * Unlike dtrace_bcopy(), overlapping regions are not handled.
1318  */
1319 static void
1320 dtrace_strcpy(const void *src, void *dst, size_t len)
1321 {
1322 	if (len != 0) {
1323 		uint8_t *s1 = dst, c;
1324 		const uint8_t *s2 = src;
1325 
1326 		do {
1327 			*s1++ = c = dtrace_load8((uintptr_t)s2++);
1328 		} while (--len != 0 && c != '\0');
1329 	}
1330 }
1331 
1332 /*
1333  * Copy src to dst, deriving the size and type from the specified (BYREF)
1334  * variable type.  The src is assumed to be unsafe memory specified by the DIF
1335  * program.  The dst is assumed to be DTrace variable memory that is of the
1336  * specified type; we assume that we can store to directly.
1337  */
1338 static void
1339 dtrace_vcopy(void *src, void *dst, dtrace_diftype_t *type, size_t limit)
1340 {
1341 	ASSERT(type->dtdt_flags & DIF_TF_BYREF);
1342 
1343 	if (type->dtdt_kind == DIF_TYPE_STRING) {
1344 		dtrace_strcpy(src, dst, MIN(type->dtdt_size, limit));
1345 	} else {
1346 		dtrace_bcopy(src, dst, MIN(type->dtdt_size, limit));
1347 	}
1348 }
1349 
1350 /*
1351  * Compare s1 to s2 using safe memory accesses.  The s1 data is assumed to be
1352  * unsafe memory specified by the DIF program.  The s2 data is assumed to be
1353  * safe memory that we can access directly because it is managed by DTrace.
1354  */
1355 static int
1356 dtrace_bcmp(const void *s1, const void *s2, size_t len)
1357 {
1358 	volatile uint16_t *flags;
1359 
1360 	flags = (volatile uint16_t *)&cpu_core[curcpu].cpuc_dtrace_flags;
1361 
1362 	if (s1 == s2)
1363 		return (0);
1364 
1365 	if (s1 == NULL || s2 == NULL)
1366 		return (1);
1367 
1368 	if (s1 != s2 && len != 0) {
1369 		const uint8_t *ps1 = s1;
1370 		const uint8_t *ps2 = s2;
1371 
1372 		do {
1373 			if (dtrace_load8((uintptr_t)ps1++) != *ps2++)
1374 				return (1);
1375 		} while (--len != 0 && !(*flags & CPU_DTRACE_FAULT));
1376 	}
1377 	return (0);
1378 }
1379 
1380 /*
1381  * Zero the specified region using a simple byte-by-byte loop.  Note that this
1382  * is for safe DTrace-managed memory only.
1383  */
1384 static void
1385 dtrace_bzero(void *dst, size_t len)
1386 {
1387 	uchar_t *cp;
1388 
1389 	for (cp = dst; len != 0; len--)
1390 		*cp++ = 0;
1391 }
1392 
1393 static void
1394 dtrace_add_128(uint64_t *addend1, uint64_t *addend2, uint64_t *sum)
1395 {
1396 	uint64_t result[2];
1397 
1398 	result[0] = addend1[0] + addend2[0];
1399 	result[1] = addend1[1] + addend2[1] +
1400 	    (result[0] < addend1[0] || result[0] < addend2[0] ? 1 : 0);
1401 
1402 	sum[0] = result[0];
1403 	sum[1] = result[1];
1404 }
1405 
1406 /*
1407  * Shift the 128-bit value in a by b. If b is positive, shift left.
1408  * If b is negative, shift right.
1409  */
1410 static void
1411 dtrace_shift_128(uint64_t *a, int b)
1412 {
1413 	uint64_t mask;
1414 
1415 	if (b == 0)
1416 		return;
1417 
1418 	if (b < 0) {
1419 		b = -b;
1420 		if (b >= 64) {
1421 			a[0] = a[1] >> (b - 64);
1422 			a[1] = 0;
1423 		} else {
1424 			a[0] >>= b;
1425 			mask = 1LL << (64 - b);
1426 			mask -= 1;
1427 			a[0] |= ((a[1] & mask) << (64 - b));
1428 			a[1] >>= b;
1429 		}
1430 	} else {
1431 		if (b >= 64) {
1432 			a[1] = a[0] << (b - 64);
1433 			a[0] = 0;
1434 		} else {
1435 			a[1] <<= b;
1436 			mask = a[0] >> (64 - b);
1437 			a[1] |= mask;
1438 			a[0] <<= b;
1439 		}
1440 	}
1441 }
1442 
1443 /*
1444  * The basic idea is to break the 2 64-bit values into 4 32-bit values,
1445  * use native multiplication on those, and then re-combine into the
1446  * resulting 128-bit value.
1447  *
1448  * (hi1 << 32 + lo1) * (hi2 << 32 + lo2) =
1449  *     hi1 * hi2 << 64 +
1450  *     hi1 * lo2 << 32 +
1451  *     hi2 * lo1 << 32 +
1452  *     lo1 * lo2
1453  */
1454 static void
1455 dtrace_multiply_128(uint64_t factor1, uint64_t factor2, uint64_t *product)
1456 {
1457 	uint64_t hi1, hi2, lo1, lo2;
1458 	uint64_t tmp[2];
1459 
1460 	hi1 = factor1 >> 32;
1461 	hi2 = factor2 >> 32;
1462 
1463 	lo1 = factor1 & DT_MASK_LO;
1464 	lo2 = factor2 & DT_MASK_LO;
1465 
1466 	product[0] = lo1 * lo2;
1467 	product[1] = hi1 * hi2;
1468 
1469 	tmp[0] = hi1 * lo2;
1470 	tmp[1] = 0;
1471 	dtrace_shift_128(tmp, 32);
1472 	dtrace_add_128(product, tmp, product);
1473 
1474 	tmp[0] = hi2 * lo1;
1475 	tmp[1] = 0;
1476 	dtrace_shift_128(tmp, 32);
1477 	dtrace_add_128(product, tmp, product);
1478 }
1479 
1480 /*
1481  * This privilege check should be used by actions and subroutines to
1482  * verify that the user credentials of the process that enabled the
1483  * invoking ECB match the target credentials
1484  */
1485 static int
1486 dtrace_priv_proc_common_user(dtrace_state_t *state)
1487 {
1488 	cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1489 
1490 	/*
1491 	 * We should always have a non-NULL state cred here, since if cred
1492 	 * is null (anonymous tracing), we fast-path bypass this routine.
1493 	 */
1494 	ASSERT(s_cr != NULL);
1495 
1496 	if ((cr = CRED()) != NULL &&
1497 	    s_cr->cr_uid == cr->cr_uid &&
1498 	    s_cr->cr_uid == cr->cr_ruid &&
1499 	    s_cr->cr_uid == cr->cr_suid &&
1500 	    s_cr->cr_gid == cr->cr_gid &&
1501 	    s_cr->cr_gid == cr->cr_rgid &&
1502 	    s_cr->cr_gid == cr->cr_sgid)
1503 		return (1);
1504 
1505 	return (0);
1506 }
1507 
1508 /*
1509  * This privilege check should be used by actions and subroutines to
1510  * verify that the zone of the process that enabled the invoking ECB
1511  * matches the target credentials
1512  */
1513 static int
1514 dtrace_priv_proc_common_zone(dtrace_state_t *state)
1515 {
1516 #ifdef illumos
1517 	cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1518 
1519 	/*
1520 	 * We should always have a non-NULL state cred here, since if cred
1521 	 * is null (anonymous tracing), we fast-path bypass this routine.
1522 	 */
1523 	ASSERT(s_cr != NULL);
1524 
1525 	if ((cr = CRED()) != NULL && s_cr->cr_zone == cr->cr_zone)
1526 		return (1);
1527 
1528 	return (0);
1529 #else
1530 	return (1);
1531 #endif
1532 }
1533 
1534 /*
1535  * This privilege check should be used by actions and subroutines to
1536  * verify that the process has not setuid or changed credentials.
1537  */
1538 static int
1539 dtrace_priv_proc_common_nocd(void)
1540 {
1541 	proc_t *proc;
1542 
1543 	if ((proc = ttoproc(curthread)) != NULL &&
1544 	    !(proc->p_flag & SNOCD))
1545 		return (1);
1546 
1547 	return (0);
1548 }
1549 
1550 static int
1551 dtrace_priv_proc_destructive(dtrace_state_t *state)
1552 {
1553 	int action = state->dts_cred.dcr_action;
1554 
1555 	if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE) == 0) &&
1556 	    dtrace_priv_proc_common_zone(state) == 0)
1557 		goto bad;
1558 
1559 	if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER) == 0) &&
1560 	    dtrace_priv_proc_common_user(state) == 0)
1561 		goto bad;
1562 
1563 	if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG) == 0) &&
1564 	    dtrace_priv_proc_common_nocd() == 0)
1565 		goto bad;
1566 
1567 	return (1);
1568 
1569 bad:
1570 	cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1571 
1572 	return (0);
1573 }
1574 
1575 static int
1576 dtrace_priv_proc_control(dtrace_state_t *state)
1577 {
1578 	if (state->dts_cred.dcr_action & DTRACE_CRA_PROC_CONTROL)
1579 		return (1);
1580 
1581 	if (dtrace_priv_proc_common_zone(state) &&
1582 	    dtrace_priv_proc_common_user(state) &&
1583 	    dtrace_priv_proc_common_nocd())
1584 		return (1);
1585 
1586 	cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1587 
1588 	return (0);
1589 }
1590 
1591 static int
1592 dtrace_priv_proc(dtrace_state_t *state)
1593 {
1594 	if (state->dts_cred.dcr_action & DTRACE_CRA_PROC)
1595 		return (1);
1596 
1597 	cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1598 
1599 	return (0);
1600 }
1601 
1602 static int
1603 dtrace_priv_kernel(dtrace_state_t *state)
1604 {
1605 	if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL)
1606 		return (1);
1607 
1608 	cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1609 
1610 	return (0);
1611 }
1612 
1613 static int
1614 dtrace_priv_kernel_destructive(dtrace_state_t *state)
1615 {
1616 	if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL_DESTRUCTIVE)
1617 		return (1);
1618 
1619 	cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1620 
1621 	return (0);
1622 }
1623 
1624 /*
1625  * Determine if the dte_cond of the specified ECB allows for processing of
1626  * the current probe to continue.  Note that this routine may allow continued
1627  * processing, but with access(es) stripped from the mstate's dtms_access
1628  * field.
1629  */
1630 static int
1631 dtrace_priv_probe(dtrace_state_t *state, dtrace_mstate_t *mstate,
1632     dtrace_ecb_t *ecb)
1633 {
1634 	dtrace_probe_t *probe = ecb->dte_probe;
1635 	dtrace_provider_t *prov = probe->dtpr_provider;
1636 	dtrace_pops_t *pops = &prov->dtpv_pops;
1637 	int mode = DTRACE_MODE_NOPRIV_DROP;
1638 
1639 	ASSERT(ecb->dte_cond);
1640 
1641 #ifdef illumos
1642 	if (pops->dtps_mode != NULL) {
1643 		mode = pops->dtps_mode(prov->dtpv_arg,
1644 		    probe->dtpr_id, probe->dtpr_arg);
1645 
1646 		ASSERT((mode & DTRACE_MODE_USER) ||
1647 		    (mode & DTRACE_MODE_KERNEL));
1648 		ASSERT((mode & DTRACE_MODE_NOPRIV_RESTRICT) ||
1649 		    (mode & DTRACE_MODE_NOPRIV_DROP));
1650 	}
1651 
1652 	/*
1653 	 * If the dte_cond bits indicate that this consumer is only allowed to
1654 	 * see user-mode firings of this probe, call the provider's dtps_mode()
1655 	 * entry point to check that the probe was fired while in a user
1656 	 * context.  If that's not the case, use the policy specified by the
1657 	 * provider to determine if we drop the probe or merely restrict
1658 	 * operation.
1659 	 */
1660 	if (ecb->dte_cond & DTRACE_COND_USERMODE) {
1661 		ASSERT(mode != DTRACE_MODE_NOPRIV_DROP);
1662 
1663 		if (!(mode & DTRACE_MODE_USER)) {
1664 			if (mode & DTRACE_MODE_NOPRIV_DROP)
1665 				return (0);
1666 
1667 			mstate->dtms_access &= ~DTRACE_ACCESS_ARGS;
1668 		}
1669 	}
1670 #endif
1671 
1672 	/*
1673 	 * This is more subtle than it looks. We have to be absolutely certain
1674 	 * that CRED() isn't going to change out from under us so it's only
1675 	 * legit to examine that structure if we're in constrained situations.
1676 	 * Currently, the only times we'll this check is if a non-super-user
1677 	 * has enabled the profile or syscall providers -- providers that
1678 	 * allow visibility of all processes. For the profile case, the check
1679 	 * above will ensure that we're examining a user context.
1680 	 */
1681 	if (ecb->dte_cond & DTRACE_COND_OWNER) {
1682 		cred_t *cr;
1683 		cred_t *s_cr = state->dts_cred.dcr_cred;
1684 		proc_t *proc;
1685 
1686 		ASSERT(s_cr != NULL);
1687 
1688 		if ((cr = CRED()) == NULL ||
1689 		    s_cr->cr_uid != cr->cr_uid ||
1690 		    s_cr->cr_uid != cr->cr_ruid ||
1691 		    s_cr->cr_uid != cr->cr_suid ||
1692 		    s_cr->cr_gid != cr->cr_gid ||
1693 		    s_cr->cr_gid != cr->cr_rgid ||
1694 		    s_cr->cr_gid != cr->cr_sgid ||
1695 		    (proc = ttoproc(curthread)) == NULL ||
1696 		    (proc->p_flag & SNOCD)) {
1697 			if (mode & DTRACE_MODE_NOPRIV_DROP)
1698 				return (0);
1699 
1700 #ifdef illumos
1701 			mstate->dtms_access &= ~DTRACE_ACCESS_PROC;
1702 #endif
1703 		}
1704 	}
1705 
1706 #ifdef illumos
1707 	/*
1708 	 * If our dte_cond is set to DTRACE_COND_ZONEOWNER and we are not
1709 	 * in our zone, check to see if our mode policy is to restrict rather
1710 	 * than to drop; if to restrict, strip away both DTRACE_ACCESS_PROC
1711 	 * and DTRACE_ACCESS_ARGS
1712 	 */
1713 	if (ecb->dte_cond & DTRACE_COND_ZONEOWNER) {
1714 		cred_t *cr;
1715 		cred_t *s_cr = state->dts_cred.dcr_cred;
1716 
1717 		ASSERT(s_cr != NULL);
1718 
1719 		if ((cr = CRED()) == NULL ||
1720 		    s_cr->cr_zone->zone_id != cr->cr_zone->zone_id) {
1721 			if (mode & DTRACE_MODE_NOPRIV_DROP)
1722 				return (0);
1723 
1724 			mstate->dtms_access &=
1725 			    ~(DTRACE_ACCESS_PROC | DTRACE_ACCESS_ARGS);
1726 		}
1727 	}
1728 #endif
1729 
1730 	return (1);
1731 }
1732 
1733 /*
1734  * Note:  not called from probe context.  This function is called
1735  * asynchronously (and at a regular interval) from outside of probe context to
1736  * clean the dirty dynamic variable lists on all CPUs.  Dynamic variable
1737  * cleaning is explained in detail in <sys/dtrace_impl.h>.
1738  */
1739 void
1740 dtrace_dynvar_clean(dtrace_dstate_t *dstate)
1741 {
1742 	dtrace_dynvar_t *dirty;
1743 	dtrace_dstate_percpu_t *dcpu;
1744 	dtrace_dynvar_t **rinsep;
1745 	int i, j, work = 0;
1746 
1747 	for (i = 0; i < NCPU; i++) {
1748 		dcpu = &dstate->dtds_percpu[i];
1749 		rinsep = &dcpu->dtdsc_rinsing;
1750 
1751 		/*
1752 		 * If the dirty list is NULL, there is no dirty work to do.
1753 		 */
1754 		if (dcpu->dtdsc_dirty == NULL)
1755 			continue;
1756 
1757 		if (dcpu->dtdsc_rinsing != NULL) {
1758 			/*
1759 			 * If the rinsing list is non-NULL, then it is because
1760 			 * this CPU was selected to accept another CPU's
1761 			 * dirty list -- and since that time, dirty buffers
1762 			 * have accumulated.  This is a highly unlikely
1763 			 * condition, but we choose to ignore the dirty
1764 			 * buffers -- they'll be picked up a future cleanse.
1765 			 */
1766 			continue;
1767 		}
1768 
1769 		if (dcpu->dtdsc_clean != NULL) {
1770 			/*
1771 			 * If the clean list is non-NULL, then we're in a
1772 			 * situation where a CPU has done deallocations (we
1773 			 * have a non-NULL dirty list) but no allocations (we
1774 			 * also have a non-NULL clean list).  We can't simply
1775 			 * move the dirty list into the clean list on this
1776 			 * CPU, yet we also don't want to allow this condition
1777 			 * to persist, lest a short clean list prevent a
1778 			 * massive dirty list from being cleaned (which in
1779 			 * turn could lead to otherwise avoidable dynamic
1780 			 * drops).  To deal with this, we look for some CPU
1781 			 * with a NULL clean list, NULL dirty list, and NULL
1782 			 * rinsing list -- and then we borrow this CPU to
1783 			 * rinse our dirty list.
1784 			 */
1785 			for (j = 0; j < NCPU; j++) {
1786 				dtrace_dstate_percpu_t *rinser;
1787 
1788 				rinser = &dstate->dtds_percpu[j];
1789 
1790 				if (rinser->dtdsc_rinsing != NULL)
1791 					continue;
1792 
1793 				if (rinser->dtdsc_dirty != NULL)
1794 					continue;
1795 
1796 				if (rinser->dtdsc_clean != NULL)
1797 					continue;
1798 
1799 				rinsep = &rinser->dtdsc_rinsing;
1800 				break;
1801 			}
1802 
1803 			if (j == NCPU) {
1804 				/*
1805 				 * We were unable to find another CPU that
1806 				 * could accept this dirty list -- we are
1807 				 * therefore unable to clean it now.
1808 				 */
1809 				dtrace_dynvar_failclean++;
1810 				continue;
1811 			}
1812 		}
1813 
1814 		work = 1;
1815 
1816 		/*
1817 		 * Atomically move the dirty list aside.
1818 		 */
1819 		do {
1820 			dirty = dcpu->dtdsc_dirty;
1821 
1822 			/*
1823 			 * Before we zap the dirty list, set the rinsing list.
1824 			 * (This allows for a potential assertion in
1825 			 * dtrace_dynvar():  if a free dynamic variable appears
1826 			 * on a hash chain, either the dirty list or the
1827 			 * rinsing list for some CPU must be non-NULL.)
1828 			 */
1829 			*rinsep = dirty;
1830 			dtrace_membar_producer();
1831 		} while (dtrace_casptr(&dcpu->dtdsc_dirty,
1832 		    dirty, NULL) != dirty);
1833 	}
1834 
1835 	if (!work) {
1836 		/*
1837 		 * We have no work to do; we can simply return.
1838 		 */
1839 		return;
1840 	}
1841 
1842 	dtrace_sync();
1843 
1844 	for (i = 0; i < NCPU; i++) {
1845 		dcpu = &dstate->dtds_percpu[i];
1846 
1847 		if (dcpu->dtdsc_rinsing == NULL)
1848 			continue;
1849 
1850 		/*
1851 		 * We are now guaranteed that no hash chain contains a pointer
1852 		 * into this dirty list; we can make it clean.
1853 		 */
1854 		ASSERT(dcpu->dtdsc_clean == NULL);
1855 		dcpu->dtdsc_clean = dcpu->dtdsc_rinsing;
1856 		dcpu->dtdsc_rinsing = NULL;
1857 	}
1858 
1859 	/*
1860 	 * Before we actually set the state to be DTRACE_DSTATE_CLEAN, make
1861 	 * sure that all CPUs have seen all of the dtdsc_clean pointers.
1862 	 * This prevents a race whereby a CPU incorrectly decides that
1863 	 * the state should be something other than DTRACE_DSTATE_CLEAN
1864 	 * after dtrace_dynvar_clean() has completed.
1865 	 */
1866 	dtrace_sync();
1867 
1868 	dstate->dtds_state = DTRACE_DSTATE_CLEAN;
1869 }
1870 
1871 /*
1872  * Depending on the value of the op parameter, this function looks-up,
1873  * allocates or deallocates an arbitrarily-keyed dynamic variable.  If an
1874  * allocation is requested, this function will return a pointer to a
1875  * dtrace_dynvar_t corresponding to the allocated variable -- or NULL if no
1876  * variable can be allocated.  If NULL is returned, the appropriate counter
1877  * will be incremented.
1878  */
1879 dtrace_dynvar_t *
1880 dtrace_dynvar(dtrace_dstate_t *dstate, uint_t nkeys,
1881     dtrace_key_t *key, size_t dsize, dtrace_dynvar_op_t op,
1882     dtrace_mstate_t *mstate, dtrace_vstate_t *vstate)
1883 {
1884 	uint64_t hashval = DTRACE_DYNHASH_VALID;
1885 	dtrace_dynhash_t *hash = dstate->dtds_hash;
1886 	dtrace_dynvar_t *free, *new_free, *next, *dvar, *start, *prev = NULL;
1887 	processorid_t me = curcpu, cpu = me;
1888 	dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[me];
1889 	size_t bucket, ksize;
1890 	size_t chunksize = dstate->dtds_chunksize;
1891 	uintptr_t kdata, lock, nstate;
1892 	uint_t i;
1893 
1894 	ASSERT(nkeys != 0);
1895 
1896 	/*
1897 	 * Hash the key.  As with aggregations, we use Jenkins' "One-at-a-time"
1898 	 * algorithm.  For the by-value portions, we perform the algorithm in
1899 	 * 16-bit chunks (as opposed to 8-bit chunks).  This speeds things up a
1900 	 * bit, and seems to have only a minute effect on distribution.  For
1901 	 * the by-reference data, we perform "One-at-a-time" iterating (safely)
1902 	 * over each referenced byte.  It's painful to do this, but it's much
1903 	 * better than pathological hash distribution.  The efficacy of the
1904 	 * hashing algorithm (and a comparison with other algorithms) may be
1905 	 * found by running the ::dtrace_dynstat MDB dcmd.
1906 	 */
1907 	for (i = 0; i < nkeys; i++) {
1908 		if (key[i].dttk_size == 0) {
1909 			uint64_t val = key[i].dttk_value;
1910 
1911 			hashval += (val >> 48) & 0xffff;
1912 			hashval += (hashval << 10);
1913 			hashval ^= (hashval >> 6);
1914 
1915 			hashval += (val >> 32) & 0xffff;
1916 			hashval += (hashval << 10);
1917 			hashval ^= (hashval >> 6);
1918 
1919 			hashval += (val >> 16) & 0xffff;
1920 			hashval += (hashval << 10);
1921 			hashval ^= (hashval >> 6);
1922 
1923 			hashval += val & 0xffff;
1924 			hashval += (hashval << 10);
1925 			hashval ^= (hashval >> 6);
1926 		} else {
1927 			/*
1928 			 * This is incredibly painful, but it beats the hell
1929 			 * out of the alternative.
1930 			 */
1931 			uint64_t j, size = key[i].dttk_size;
1932 			uintptr_t base = (uintptr_t)key[i].dttk_value;
1933 
1934 			if (!dtrace_canload(base, size, mstate, vstate))
1935 				break;
1936 
1937 			for (j = 0; j < size; j++) {
1938 				hashval += dtrace_load8(base + j);
1939 				hashval += (hashval << 10);
1940 				hashval ^= (hashval >> 6);
1941 			}
1942 		}
1943 	}
1944 
1945 	if (DTRACE_CPUFLAG_ISSET(CPU_DTRACE_FAULT))
1946 		return (NULL);
1947 
1948 	hashval += (hashval << 3);
1949 	hashval ^= (hashval >> 11);
1950 	hashval += (hashval << 15);
1951 
1952 	/*
1953 	 * There is a remote chance (ideally, 1 in 2^31) that our hashval
1954 	 * comes out to be one of our two sentinel hash values.  If this
1955 	 * actually happens, we set the hashval to be a value known to be a
1956 	 * non-sentinel value.
1957 	 */
1958 	if (hashval == DTRACE_DYNHASH_FREE || hashval == DTRACE_DYNHASH_SINK)
1959 		hashval = DTRACE_DYNHASH_VALID;
1960 
1961 	/*
1962 	 * Yes, it's painful to do a divide here.  If the cycle count becomes
1963 	 * important here, tricks can be pulled to reduce it.  (However, it's
1964 	 * critical that hash collisions be kept to an absolute minimum;
1965 	 * they're much more painful than a divide.)  It's better to have a
1966 	 * solution that generates few collisions and still keeps things
1967 	 * relatively simple.
1968 	 */
1969 	bucket = hashval % dstate->dtds_hashsize;
1970 
1971 	if (op == DTRACE_DYNVAR_DEALLOC) {
1972 		volatile uintptr_t *lockp = &hash[bucket].dtdh_lock;
1973 
1974 		for (;;) {
1975 			while ((lock = *lockp) & 1)
1976 				continue;
1977 
1978 			if (dtrace_casptr((volatile void *)lockp,
1979 			    (volatile void *)lock, (volatile void *)(lock + 1)) == (void *)lock)
1980 				break;
1981 		}
1982 
1983 		dtrace_membar_producer();
1984 	}
1985 
1986 top:
1987 	prev = NULL;
1988 	lock = hash[bucket].dtdh_lock;
1989 
1990 	dtrace_membar_consumer();
1991 
1992 	start = hash[bucket].dtdh_chain;
1993 	ASSERT(start != NULL && (start->dtdv_hashval == DTRACE_DYNHASH_SINK ||
1994 	    start->dtdv_hashval != DTRACE_DYNHASH_FREE ||
1995 	    op != DTRACE_DYNVAR_DEALLOC));
1996 
1997 	for (dvar = start; dvar != NULL; dvar = dvar->dtdv_next) {
1998 		dtrace_tuple_t *dtuple = &dvar->dtdv_tuple;
1999 		dtrace_key_t *dkey = &dtuple->dtt_key[0];
2000 
2001 		if (dvar->dtdv_hashval != hashval) {
2002 			if (dvar->dtdv_hashval == DTRACE_DYNHASH_SINK) {
2003 				/*
2004 				 * We've reached the sink, and therefore the
2005 				 * end of the hash chain; we can kick out of
2006 				 * the loop knowing that we have seen a valid
2007 				 * snapshot of state.
2008 				 */
2009 				ASSERT(dvar->dtdv_next == NULL);
2010 				ASSERT(dvar == &dtrace_dynhash_sink);
2011 				break;
2012 			}
2013 
2014 			if (dvar->dtdv_hashval == DTRACE_DYNHASH_FREE) {
2015 				/*
2016 				 * We've gone off the rails:  somewhere along
2017 				 * the line, one of the members of this hash
2018 				 * chain was deleted.  Note that we could also
2019 				 * detect this by simply letting this loop run
2020 				 * to completion, as we would eventually hit
2021 				 * the end of the dirty list.  However, we
2022 				 * want to avoid running the length of the
2023 				 * dirty list unnecessarily (it might be quite
2024 				 * long), so we catch this as early as
2025 				 * possible by detecting the hash marker.  In
2026 				 * this case, we simply set dvar to NULL and
2027 				 * break; the conditional after the loop will
2028 				 * send us back to top.
2029 				 */
2030 				dvar = NULL;
2031 				break;
2032 			}
2033 
2034 			goto next;
2035 		}
2036 
2037 		if (dtuple->dtt_nkeys != nkeys)
2038 			goto next;
2039 
2040 		for (i = 0; i < nkeys; i++, dkey++) {
2041 			if (dkey->dttk_size != key[i].dttk_size)
2042 				goto next; /* size or type mismatch */
2043 
2044 			if (dkey->dttk_size != 0) {
2045 				if (dtrace_bcmp(
2046 				    (void *)(uintptr_t)key[i].dttk_value,
2047 				    (void *)(uintptr_t)dkey->dttk_value,
2048 				    dkey->dttk_size))
2049 					goto next;
2050 			} else {
2051 				if (dkey->dttk_value != key[i].dttk_value)
2052 					goto next;
2053 			}
2054 		}
2055 
2056 		if (op != DTRACE_DYNVAR_DEALLOC)
2057 			return (dvar);
2058 
2059 		ASSERT(dvar->dtdv_next == NULL ||
2060 		    dvar->dtdv_next->dtdv_hashval != DTRACE_DYNHASH_FREE);
2061 
2062 		if (prev != NULL) {
2063 			ASSERT(hash[bucket].dtdh_chain != dvar);
2064 			ASSERT(start != dvar);
2065 			ASSERT(prev->dtdv_next == dvar);
2066 			prev->dtdv_next = dvar->dtdv_next;
2067 		} else {
2068 			if (dtrace_casptr(&hash[bucket].dtdh_chain,
2069 			    start, dvar->dtdv_next) != start) {
2070 				/*
2071 				 * We have failed to atomically swing the
2072 				 * hash table head pointer, presumably because
2073 				 * of a conflicting allocation on another CPU.
2074 				 * We need to reread the hash chain and try
2075 				 * again.
2076 				 */
2077 				goto top;
2078 			}
2079 		}
2080 
2081 		dtrace_membar_producer();
2082 
2083 		/*
2084 		 * Now set the hash value to indicate that it's free.
2085 		 */
2086 		ASSERT(hash[bucket].dtdh_chain != dvar);
2087 		dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
2088 
2089 		dtrace_membar_producer();
2090 
2091 		/*
2092 		 * Set the next pointer to point at the dirty list, and
2093 		 * atomically swing the dirty pointer to the newly freed dvar.
2094 		 */
2095 		do {
2096 			next = dcpu->dtdsc_dirty;
2097 			dvar->dtdv_next = next;
2098 		} while (dtrace_casptr(&dcpu->dtdsc_dirty, next, dvar) != next);
2099 
2100 		/*
2101 		 * Finally, unlock this hash bucket.
2102 		 */
2103 		ASSERT(hash[bucket].dtdh_lock == lock);
2104 		ASSERT(lock & 1);
2105 		hash[bucket].dtdh_lock++;
2106 
2107 		return (NULL);
2108 next:
2109 		prev = dvar;
2110 		continue;
2111 	}
2112 
2113 	if (dvar == NULL) {
2114 		/*
2115 		 * If dvar is NULL, it is because we went off the rails:
2116 		 * one of the elements that we traversed in the hash chain
2117 		 * was deleted while we were traversing it.  In this case,
2118 		 * we assert that we aren't doing a dealloc (deallocs lock
2119 		 * the hash bucket to prevent themselves from racing with
2120 		 * one another), and retry the hash chain traversal.
2121 		 */
2122 		ASSERT(op != DTRACE_DYNVAR_DEALLOC);
2123 		goto top;
2124 	}
2125 
2126 	if (op != DTRACE_DYNVAR_ALLOC) {
2127 		/*
2128 		 * If we are not to allocate a new variable, we want to
2129 		 * return NULL now.  Before we return, check that the value
2130 		 * of the lock word hasn't changed.  If it has, we may have
2131 		 * seen an inconsistent snapshot.
2132 		 */
2133 		if (op == DTRACE_DYNVAR_NOALLOC) {
2134 			if (hash[bucket].dtdh_lock != lock)
2135 				goto top;
2136 		} else {
2137 			ASSERT(op == DTRACE_DYNVAR_DEALLOC);
2138 			ASSERT(hash[bucket].dtdh_lock == lock);
2139 			ASSERT(lock & 1);
2140 			hash[bucket].dtdh_lock++;
2141 		}
2142 
2143 		return (NULL);
2144 	}
2145 
2146 	/*
2147 	 * We need to allocate a new dynamic variable.  The size we need is the
2148 	 * size of dtrace_dynvar plus the size of nkeys dtrace_key_t's plus the
2149 	 * size of any auxiliary key data (rounded up to 8-byte alignment) plus
2150 	 * the size of any referred-to data (dsize).  We then round the final
2151 	 * size up to the chunksize for allocation.
2152 	 */
2153 	for (ksize = 0, i = 0; i < nkeys; i++)
2154 		ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
2155 
2156 	/*
2157 	 * This should be pretty much impossible, but could happen if, say,
2158 	 * strange DIF specified the tuple.  Ideally, this should be an
2159 	 * assertion and not an error condition -- but that requires that the
2160 	 * chunksize calculation in dtrace_difo_chunksize() be absolutely
2161 	 * bullet-proof.  (That is, it must not be able to be fooled by
2162 	 * malicious DIF.)  Given the lack of backwards branches in DIF,
2163 	 * solving this would presumably not amount to solving the Halting
2164 	 * Problem -- but it still seems awfully hard.
2165 	 */
2166 	if (sizeof (dtrace_dynvar_t) + sizeof (dtrace_key_t) * (nkeys - 1) +
2167 	    ksize + dsize > chunksize) {
2168 		dcpu->dtdsc_drops++;
2169 		return (NULL);
2170 	}
2171 
2172 	nstate = DTRACE_DSTATE_EMPTY;
2173 
2174 	do {
2175 retry:
2176 		free = dcpu->dtdsc_free;
2177 
2178 		if (free == NULL) {
2179 			dtrace_dynvar_t *clean = dcpu->dtdsc_clean;
2180 			void *rval;
2181 
2182 			if (clean == NULL) {
2183 				/*
2184 				 * We're out of dynamic variable space on
2185 				 * this CPU.  Unless we have tried all CPUs,
2186 				 * we'll try to allocate from a different
2187 				 * CPU.
2188 				 */
2189 				switch (dstate->dtds_state) {
2190 				case DTRACE_DSTATE_CLEAN: {
2191 					void *sp = &dstate->dtds_state;
2192 
2193 					if (++cpu >= NCPU)
2194 						cpu = 0;
2195 
2196 					if (dcpu->dtdsc_dirty != NULL &&
2197 					    nstate == DTRACE_DSTATE_EMPTY)
2198 						nstate = DTRACE_DSTATE_DIRTY;
2199 
2200 					if (dcpu->dtdsc_rinsing != NULL)
2201 						nstate = DTRACE_DSTATE_RINSING;
2202 
2203 					dcpu = &dstate->dtds_percpu[cpu];
2204 
2205 					if (cpu != me)
2206 						goto retry;
2207 
2208 					(void) dtrace_cas32(sp,
2209 					    DTRACE_DSTATE_CLEAN, nstate);
2210 
2211 					/*
2212 					 * To increment the correct bean
2213 					 * counter, take another lap.
2214 					 */
2215 					goto retry;
2216 				}
2217 
2218 				case DTRACE_DSTATE_DIRTY:
2219 					dcpu->dtdsc_dirty_drops++;
2220 					break;
2221 
2222 				case DTRACE_DSTATE_RINSING:
2223 					dcpu->dtdsc_rinsing_drops++;
2224 					break;
2225 
2226 				case DTRACE_DSTATE_EMPTY:
2227 					dcpu->dtdsc_drops++;
2228 					break;
2229 				}
2230 
2231 				DTRACE_CPUFLAG_SET(CPU_DTRACE_DROP);
2232 				return (NULL);
2233 			}
2234 
2235 			/*
2236 			 * The clean list appears to be non-empty.  We want to
2237 			 * move the clean list to the free list; we start by
2238 			 * moving the clean pointer aside.
2239 			 */
2240 			if (dtrace_casptr(&dcpu->dtdsc_clean,
2241 			    clean, NULL) != clean) {
2242 				/*
2243 				 * We are in one of two situations:
2244 				 *
2245 				 *  (a)	The clean list was switched to the
2246 				 *	free list by another CPU.
2247 				 *
2248 				 *  (b)	The clean list was added to by the
2249 				 *	cleansing cyclic.
2250 				 *
2251 				 * In either of these situations, we can
2252 				 * just reattempt the free list allocation.
2253 				 */
2254 				goto retry;
2255 			}
2256 
2257 			ASSERT(clean->dtdv_hashval == DTRACE_DYNHASH_FREE);
2258 
2259 			/*
2260 			 * Now we'll move the clean list to our free list.
2261 			 * It's impossible for this to fail:  the only way
2262 			 * the free list can be updated is through this
2263 			 * code path, and only one CPU can own the clean list.
2264 			 * Thus, it would only be possible for this to fail if
2265 			 * this code were racing with dtrace_dynvar_clean().
2266 			 * (That is, if dtrace_dynvar_clean() updated the clean
2267 			 * list, and we ended up racing to update the free
2268 			 * list.)  This race is prevented by the dtrace_sync()
2269 			 * in dtrace_dynvar_clean() -- which flushes the
2270 			 * owners of the clean lists out before resetting
2271 			 * the clean lists.
2272 			 */
2273 			dcpu = &dstate->dtds_percpu[me];
2274 			rval = dtrace_casptr(&dcpu->dtdsc_free, NULL, clean);
2275 			ASSERT(rval == NULL);
2276 			goto retry;
2277 		}
2278 
2279 		dvar = free;
2280 		new_free = dvar->dtdv_next;
2281 	} while (dtrace_casptr(&dcpu->dtdsc_free, free, new_free) != free);
2282 
2283 	/*
2284 	 * We have now allocated a new chunk.  We copy the tuple keys into the
2285 	 * tuple array and copy any referenced key data into the data space
2286 	 * following the tuple array.  As we do this, we relocate dttk_value
2287 	 * in the final tuple to point to the key data address in the chunk.
2288 	 */
2289 	kdata = (uintptr_t)&dvar->dtdv_tuple.dtt_key[nkeys];
2290 	dvar->dtdv_data = (void *)(kdata + ksize);
2291 	dvar->dtdv_tuple.dtt_nkeys = nkeys;
2292 
2293 	for (i = 0; i < nkeys; i++) {
2294 		dtrace_key_t *dkey = &dvar->dtdv_tuple.dtt_key[i];
2295 		size_t kesize = key[i].dttk_size;
2296 
2297 		if (kesize != 0) {
2298 			dtrace_bcopy(
2299 			    (const void *)(uintptr_t)key[i].dttk_value,
2300 			    (void *)kdata, kesize);
2301 			dkey->dttk_value = kdata;
2302 			kdata += P2ROUNDUP(kesize, sizeof (uint64_t));
2303 		} else {
2304 			dkey->dttk_value = key[i].dttk_value;
2305 		}
2306 
2307 		dkey->dttk_size = kesize;
2308 	}
2309 
2310 	ASSERT(dvar->dtdv_hashval == DTRACE_DYNHASH_FREE);
2311 	dvar->dtdv_hashval = hashval;
2312 	dvar->dtdv_next = start;
2313 
2314 	if (dtrace_casptr(&hash[bucket].dtdh_chain, start, dvar) == start)
2315 		return (dvar);
2316 
2317 	/*
2318 	 * The cas has failed.  Either another CPU is adding an element to
2319 	 * this hash chain, or another CPU is deleting an element from this
2320 	 * hash chain.  The simplest way to deal with both of these cases
2321 	 * (though not necessarily the most efficient) is to free our
2322 	 * allocated block and re-attempt it all.  Note that the free is
2323 	 * to the dirty list and _not_ to the free list.  This is to prevent
2324 	 * races with allocators, above.
2325 	 */
2326 	dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
2327 
2328 	dtrace_membar_producer();
2329 
2330 	do {
2331 		free = dcpu->dtdsc_dirty;
2332 		dvar->dtdv_next = free;
2333 	} while (dtrace_casptr(&dcpu->dtdsc_dirty, free, dvar) != free);
2334 
2335 	goto top;
2336 }
2337 
2338 /*ARGSUSED*/
2339 static void
2340 dtrace_aggregate_min(uint64_t *oval, uint64_t nval, uint64_t arg)
2341 {
2342 	if ((int64_t)nval < (int64_t)*oval)
2343 		*oval = nval;
2344 }
2345 
2346 /*ARGSUSED*/
2347 static void
2348 dtrace_aggregate_max(uint64_t *oval, uint64_t nval, uint64_t arg)
2349 {
2350 	if ((int64_t)nval > (int64_t)*oval)
2351 		*oval = nval;
2352 }
2353 
2354 static void
2355 dtrace_aggregate_quantize(uint64_t *quanta, uint64_t nval, uint64_t incr)
2356 {
2357 	int i, zero = DTRACE_QUANTIZE_ZEROBUCKET;
2358 	int64_t val = (int64_t)nval;
2359 
2360 	if (val < 0) {
2361 		for (i = 0; i < zero; i++) {
2362 			if (val <= DTRACE_QUANTIZE_BUCKETVAL(i)) {
2363 				quanta[i] += incr;
2364 				return;
2365 			}
2366 		}
2367 	} else {
2368 		for (i = zero + 1; i < DTRACE_QUANTIZE_NBUCKETS; i++) {
2369 			if (val < DTRACE_QUANTIZE_BUCKETVAL(i)) {
2370 				quanta[i - 1] += incr;
2371 				return;
2372 			}
2373 		}
2374 
2375 		quanta[DTRACE_QUANTIZE_NBUCKETS - 1] += incr;
2376 		return;
2377 	}
2378 
2379 	ASSERT(0);
2380 }
2381 
2382 static void
2383 dtrace_aggregate_lquantize(uint64_t *lquanta, uint64_t nval, uint64_t incr)
2384 {
2385 	uint64_t arg = *lquanta++;
2386 	int32_t base = DTRACE_LQUANTIZE_BASE(arg);
2387 	uint16_t step = DTRACE_LQUANTIZE_STEP(arg);
2388 	uint16_t levels = DTRACE_LQUANTIZE_LEVELS(arg);
2389 	int32_t val = (int32_t)nval, level;
2390 
2391 	ASSERT(step != 0);
2392 	ASSERT(levels != 0);
2393 
2394 	if (val < base) {
2395 		/*
2396 		 * This is an underflow.
2397 		 */
2398 		lquanta[0] += incr;
2399 		return;
2400 	}
2401 
2402 	level = (val - base) / step;
2403 
2404 	if (level < levels) {
2405 		lquanta[level + 1] += incr;
2406 		return;
2407 	}
2408 
2409 	/*
2410 	 * This is an overflow.
2411 	 */
2412 	lquanta[levels + 1] += incr;
2413 }
2414 
2415 static int
2416 dtrace_aggregate_llquantize_bucket(uint16_t factor, uint16_t low,
2417     uint16_t high, uint16_t nsteps, int64_t value)
2418 {
2419 	int64_t this = 1, last, next;
2420 	int base = 1, order;
2421 
2422 	ASSERT(factor <= nsteps);
2423 	ASSERT(nsteps % factor == 0);
2424 
2425 	for (order = 0; order < low; order++)
2426 		this *= factor;
2427 
2428 	/*
2429 	 * If our value is less than our factor taken to the power of the
2430 	 * low order of magnitude, it goes into the zeroth bucket.
2431 	 */
2432 	if (value < (last = this))
2433 		return (0);
2434 
2435 	for (this *= factor; order <= high; order++) {
2436 		int nbuckets = this > nsteps ? nsteps : this;
2437 
2438 		if ((next = this * factor) < this) {
2439 			/*
2440 			 * We should not generally get log/linear quantizations
2441 			 * with a high magnitude that allows 64-bits to
2442 			 * overflow, but we nonetheless protect against this
2443 			 * by explicitly checking for overflow, and clamping
2444 			 * our value accordingly.
2445 			 */
2446 			value = this - 1;
2447 		}
2448 
2449 		if (value < this) {
2450 			/*
2451 			 * If our value lies within this order of magnitude,
2452 			 * determine its position by taking the offset within
2453 			 * the order of magnitude, dividing by the bucket
2454 			 * width, and adding to our (accumulated) base.
2455 			 */
2456 			return (base + (value - last) / (this / nbuckets));
2457 		}
2458 
2459 		base += nbuckets - (nbuckets / factor);
2460 		last = this;
2461 		this = next;
2462 	}
2463 
2464 	/*
2465 	 * Our value is greater than or equal to our factor taken to the
2466 	 * power of one plus the high magnitude -- return the top bucket.
2467 	 */
2468 	return (base);
2469 }
2470 
2471 static void
2472 dtrace_aggregate_llquantize(uint64_t *llquanta, uint64_t nval, uint64_t incr)
2473 {
2474 	uint64_t arg = *llquanta++;
2475 	uint16_t factor = DTRACE_LLQUANTIZE_FACTOR(arg);
2476 	uint16_t low = DTRACE_LLQUANTIZE_LOW(arg);
2477 	uint16_t high = DTRACE_LLQUANTIZE_HIGH(arg);
2478 	uint16_t nsteps = DTRACE_LLQUANTIZE_NSTEP(arg);
2479 
2480 	llquanta[dtrace_aggregate_llquantize_bucket(factor,
2481 	    low, high, nsteps, nval)] += incr;
2482 }
2483 
2484 /*ARGSUSED*/
2485 static void
2486 dtrace_aggregate_avg(uint64_t *data, uint64_t nval, uint64_t arg)
2487 {
2488 	data[0]++;
2489 	data[1] += nval;
2490 }
2491 
2492 /*ARGSUSED*/
2493 static void
2494 dtrace_aggregate_stddev(uint64_t *data, uint64_t nval, uint64_t arg)
2495 {
2496 	int64_t snval = (int64_t)nval;
2497 	uint64_t tmp[2];
2498 
2499 	data[0]++;
2500 	data[1] += nval;
2501 
2502 	/*
2503 	 * What we want to say here is:
2504 	 *
2505 	 * data[2] += nval * nval;
2506 	 *
2507 	 * But given that nval is 64-bit, we could easily overflow, so
2508 	 * we do this as 128-bit arithmetic.
2509 	 */
2510 	if (snval < 0)
2511 		snval = -snval;
2512 
2513 	dtrace_multiply_128((uint64_t)snval, (uint64_t)snval, tmp);
2514 	dtrace_add_128(data + 2, tmp, data + 2);
2515 }
2516 
2517 /*ARGSUSED*/
2518 static void
2519 dtrace_aggregate_count(uint64_t *oval, uint64_t nval, uint64_t arg)
2520 {
2521 	*oval = *oval + 1;
2522 }
2523 
2524 /*ARGSUSED*/
2525 static void
2526 dtrace_aggregate_sum(uint64_t *oval, uint64_t nval, uint64_t arg)
2527 {
2528 	*oval += nval;
2529 }
2530 
2531 /*
2532  * Aggregate given the tuple in the principal data buffer, and the aggregating
2533  * action denoted by the specified dtrace_aggregation_t.  The aggregation
2534  * buffer is specified as the buf parameter.  This routine does not return
2535  * failure; if there is no space in the aggregation buffer, the data will be
2536  * dropped, and a corresponding counter incremented.
2537  */
2538 static void
2539 dtrace_aggregate(dtrace_aggregation_t *agg, dtrace_buffer_t *dbuf,
2540     intptr_t offset, dtrace_buffer_t *buf, uint64_t expr, uint64_t arg)
2541 {
2542 	dtrace_recdesc_t *rec = &agg->dtag_action.dta_rec;
2543 	uint32_t i, ndx, size, fsize;
2544 	uint32_t align = sizeof (uint64_t) - 1;
2545 	dtrace_aggbuffer_t *agb;
2546 	dtrace_aggkey_t *key;
2547 	uint32_t hashval = 0, limit, isstr;
2548 	caddr_t tomax, data, kdata;
2549 	dtrace_actkind_t action;
2550 	dtrace_action_t *act;
2551 	uintptr_t offs;
2552 
2553 	if (buf == NULL)
2554 		return;
2555 
2556 	if (!agg->dtag_hasarg) {
2557 		/*
2558 		 * Currently, only quantize() and lquantize() take additional
2559 		 * arguments, and they have the same semantics:  an increment
2560 		 * value that defaults to 1 when not present.  If additional
2561 		 * aggregating actions take arguments, the setting of the
2562 		 * default argument value will presumably have to become more
2563 		 * sophisticated...
2564 		 */
2565 		arg = 1;
2566 	}
2567 
2568 	action = agg->dtag_action.dta_kind - DTRACEACT_AGGREGATION;
2569 	size = rec->dtrd_offset - agg->dtag_base;
2570 	fsize = size + rec->dtrd_size;
2571 
2572 	ASSERT(dbuf->dtb_tomax != NULL);
2573 	data = dbuf->dtb_tomax + offset + agg->dtag_base;
2574 
2575 	if ((tomax = buf->dtb_tomax) == NULL) {
2576 		dtrace_buffer_drop(buf);
2577 		return;
2578 	}
2579 
2580 	/*
2581 	 * The metastructure is always at the bottom of the buffer.
2582 	 */
2583 	agb = (dtrace_aggbuffer_t *)(tomax + buf->dtb_size -
2584 	    sizeof (dtrace_aggbuffer_t));
2585 
2586 	if (buf->dtb_offset == 0) {
2587 		/*
2588 		 * We just kludge up approximately 1/8th of the size to be
2589 		 * buckets.  If this guess ends up being routinely
2590 		 * off-the-mark, we may need to dynamically readjust this
2591 		 * based on past performance.
2592 		 */
2593 		uintptr_t hashsize = (buf->dtb_size >> 3) / sizeof (uintptr_t);
2594 
2595 		if ((uintptr_t)agb - hashsize * sizeof (dtrace_aggkey_t *) <
2596 		    (uintptr_t)tomax || hashsize == 0) {
2597 			/*
2598 			 * We've been given a ludicrously small buffer;
2599 			 * increment our drop count and leave.
2600 			 */
2601 			dtrace_buffer_drop(buf);
2602 			return;
2603 		}
2604 
2605 		/*
2606 		 * And now, a pathetic attempt to try to get a an odd (or
2607 		 * perchance, a prime) hash size for better hash distribution.
2608 		 */
2609 		if (hashsize > (DTRACE_AGGHASHSIZE_SLEW << 3))
2610 			hashsize -= DTRACE_AGGHASHSIZE_SLEW;
2611 
2612 		agb->dtagb_hashsize = hashsize;
2613 		agb->dtagb_hash = (dtrace_aggkey_t **)((uintptr_t)agb -
2614 		    agb->dtagb_hashsize * sizeof (dtrace_aggkey_t *));
2615 		agb->dtagb_free = (uintptr_t)agb->dtagb_hash;
2616 
2617 		for (i = 0; i < agb->dtagb_hashsize; i++)
2618 			agb->dtagb_hash[i] = NULL;
2619 	}
2620 
2621 	ASSERT(agg->dtag_first != NULL);
2622 	ASSERT(agg->dtag_first->dta_intuple);
2623 
2624 	/*
2625 	 * Calculate the hash value based on the key.  Note that we _don't_
2626 	 * include the aggid in the hashing (but we will store it as part of
2627 	 * the key).  The hashing algorithm is Bob Jenkins' "One-at-a-time"
2628 	 * algorithm: a simple, quick algorithm that has no known funnels, and
2629 	 * gets good distribution in practice.  The efficacy of the hashing
2630 	 * algorithm (and a comparison with other algorithms) may be found by
2631 	 * running the ::dtrace_aggstat MDB dcmd.
2632 	 */
2633 	for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
2634 		i = act->dta_rec.dtrd_offset - agg->dtag_base;
2635 		limit = i + act->dta_rec.dtrd_size;
2636 		ASSERT(limit <= size);
2637 		isstr = DTRACEACT_ISSTRING(act);
2638 
2639 		for (; i < limit; i++) {
2640 			hashval += data[i];
2641 			hashval += (hashval << 10);
2642 			hashval ^= (hashval >> 6);
2643 
2644 			if (isstr && data[i] == '\0')
2645 				break;
2646 		}
2647 	}
2648 
2649 	hashval += (hashval << 3);
2650 	hashval ^= (hashval >> 11);
2651 	hashval += (hashval << 15);
2652 
2653 	/*
2654 	 * Yes, the divide here is expensive -- but it's generally the least
2655 	 * of the performance issues given the amount of data that we iterate
2656 	 * over to compute hash values, compare data, etc.
2657 	 */
2658 	ndx = hashval % agb->dtagb_hashsize;
2659 
2660 	for (key = agb->dtagb_hash[ndx]; key != NULL; key = key->dtak_next) {
2661 		ASSERT((caddr_t)key >= tomax);
2662 		ASSERT((caddr_t)key < tomax + buf->dtb_size);
2663 
2664 		if (hashval != key->dtak_hashval || key->dtak_size != size)
2665 			continue;
2666 
2667 		kdata = key->dtak_data;
2668 		ASSERT(kdata >= tomax && kdata < tomax + buf->dtb_size);
2669 
2670 		for (act = agg->dtag_first; act->dta_intuple;
2671 		    act = act->dta_next) {
2672 			i = act->dta_rec.dtrd_offset - agg->dtag_base;
2673 			limit = i + act->dta_rec.dtrd_size;
2674 			ASSERT(limit <= size);
2675 			isstr = DTRACEACT_ISSTRING(act);
2676 
2677 			for (; i < limit; i++) {
2678 				if (kdata[i] != data[i])
2679 					goto next;
2680 
2681 				if (isstr && data[i] == '\0')
2682 					break;
2683 			}
2684 		}
2685 
2686 		if (action != key->dtak_action) {
2687 			/*
2688 			 * We are aggregating on the same value in the same
2689 			 * aggregation with two different aggregating actions.
2690 			 * (This should have been picked up in the compiler,
2691 			 * so we may be dealing with errant or devious DIF.)
2692 			 * This is an error condition; we indicate as much,
2693 			 * and return.
2694 			 */
2695 			DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
2696 			return;
2697 		}
2698 
2699 		/*
2700 		 * This is a hit:  we need to apply the aggregator to
2701 		 * the value at this key.
2702 		 */
2703 		agg->dtag_aggregate((uint64_t *)(kdata + size), expr, arg);
2704 		return;
2705 next:
2706 		continue;
2707 	}
2708 
2709 	/*
2710 	 * We didn't find it.  We need to allocate some zero-filled space,
2711 	 * link it into the hash table appropriately, and apply the aggregator
2712 	 * to the (zero-filled) value.
2713 	 */
2714 	offs = buf->dtb_offset;
2715 	while (offs & (align - 1))
2716 		offs += sizeof (uint32_t);
2717 
2718 	/*
2719 	 * If we don't have enough room to both allocate a new key _and_
2720 	 * its associated data, increment the drop count and return.
2721 	 */
2722 	if ((uintptr_t)tomax + offs + fsize >
2723 	    agb->dtagb_free - sizeof (dtrace_aggkey_t)) {
2724 		dtrace_buffer_drop(buf);
2725 		return;
2726 	}
2727 
2728 	/*CONSTCOND*/
2729 	ASSERT(!(sizeof (dtrace_aggkey_t) & (sizeof (uintptr_t) - 1)));
2730 	key = (dtrace_aggkey_t *)(agb->dtagb_free - sizeof (dtrace_aggkey_t));
2731 	agb->dtagb_free -= sizeof (dtrace_aggkey_t);
2732 
2733 	key->dtak_data = kdata = tomax + offs;
2734 	buf->dtb_offset = offs + fsize;
2735 
2736 	/*
2737 	 * Now copy the data across.
2738 	 */
2739 	*((dtrace_aggid_t *)kdata) = agg->dtag_id;
2740 
2741 	for (i = sizeof (dtrace_aggid_t); i < size; i++)
2742 		kdata[i] = data[i];
2743 
2744 	/*
2745 	 * Because strings are not zeroed out by default, we need to iterate
2746 	 * looking for actions that store strings, and we need to explicitly
2747 	 * pad these strings out with zeroes.
2748 	 */
2749 	for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
2750 		int nul;
2751 
2752 		if (!DTRACEACT_ISSTRING(act))
2753 			continue;
2754 
2755 		i = act->dta_rec.dtrd_offset - agg->dtag_base;
2756 		limit = i + act->dta_rec.dtrd_size;
2757 		ASSERT(limit <= size);
2758 
2759 		for (nul = 0; i < limit; i++) {
2760 			if (nul) {
2761 				kdata[i] = '\0';
2762 				continue;
2763 			}
2764 
2765 			if (data[i] != '\0')
2766 				continue;
2767 
2768 			nul = 1;
2769 		}
2770 	}
2771 
2772 	for (i = size; i < fsize; i++)
2773 		kdata[i] = 0;
2774 
2775 	key->dtak_hashval = hashval;
2776 	key->dtak_size = size;
2777 	key->dtak_action = action;
2778 	key->dtak_next = agb->dtagb_hash[ndx];
2779 	agb->dtagb_hash[ndx] = key;
2780 
2781 	/*
2782 	 * Finally, apply the aggregator.
2783 	 */
2784 	*((uint64_t *)(key->dtak_data + size)) = agg->dtag_initial;
2785 	agg->dtag_aggregate((uint64_t *)(key->dtak_data + size), expr, arg);
2786 }
2787 
2788 /*
2789  * Given consumer state, this routine finds a speculation in the INACTIVE
2790  * state and transitions it into the ACTIVE state.  If there is no speculation
2791  * in the INACTIVE state, 0 is returned.  In this case, no error counter is
2792  * incremented -- it is up to the caller to take appropriate action.
2793  */
2794 static int
2795 dtrace_speculation(dtrace_state_t *state)
2796 {
2797 	int i = 0;
2798 	dtrace_speculation_state_t curstate;
2799 	uint32_t *stat = &state->dts_speculations_unavail, count;
2800 
2801 	while (i < state->dts_nspeculations) {
2802 		dtrace_speculation_t *spec = &state->dts_speculations[i];
2803 
2804 		curstate = spec->dtsp_state;
2805 
2806 		if (curstate != DTRACESPEC_INACTIVE) {
2807 			if (curstate == DTRACESPEC_COMMITTINGMANY ||
2808 			    curstate == DTRACESPEC_COMMITTING ||
2809 			    curstate == DTRACESPEC_DISCARDING)
2810 				stat = &state->dts_speculations_busy;
2811 			i++;
2812 			continue;
2813 		}
2814 
2815 		if (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2816 		    curstate, DTRACESPEC_ACTIVE) == curstate)
2817 			return (i + 1);
2818 	}
2819 
2820 	/*
2821 	 * We couldn't find a speculation.  If we found as much as a single
2822 	 * busy speculation buffer, we'll attribute this failure as "busy"
2823 	 * instead of "unavail".
2824 	 */
2825 	do {
2826 		count = *stat;
2827 	} while (dtrace_cas32(stat, count, count + 1) != count);
2828 
2829 	return (0);
2830 }
2831 
2832 /*
2833  * This routine commits an active speculation.  If the specified speculation
2834  * is not in a valid state to perform a commit(), this routine will silently do
2835  * nothing.  The state of the specified speculation is transitioned according
2836  * to the state transition diagram outlined in <sys/dtrace_impl.h>
2837  */
2838 static void
2839 dtrace_speculation_commit(dtrace_state_t *state, processorid_t cpu,
2840     dtrace_specid_t which)
2841 {
2842 	dtrace_speculation_t *spec;
2843 	dtrace_buffer_t *src, *dest;
2844 	uintptr_t daddr, saddr, dlimit, slimit;
2845 	dtrace_speculation_state_t curstate, new = 0;
2846 	intptr_t offs;
2847 	uint64_t timestamp;
2848 
2849 	if (which == 0)
2850 		return;
2851 
2852 	if (which > state->dts_nspeculations) {
2853 		cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2854 		return;
2855 	}
2856 
2857 	spec = &state->dts_speculations[which - 1];
2858 	src = &spec->dtsp_buffer[cpu];
2859 	dest = &state->dts_buffer[cpu];
2860 
2861 	do {
2862 		curstate = spec->dtsp_state;
2863 
2864 		if (curstate == DTRACESPEC_COMMITTINGMANY)
2865 			break;
2866 
2867 		switch (curstate) {
2868 		case DTRACESPEC_INACTIVE:
2869 		case DTRACESPEC_DISCARDING:
2870 			return;
2871 
2872 		case DTRACESPEC_COMMITTING:
2873 			/*
2874 			 * This is only possible if we are (a) commit()'ing
2875 			 * without having done a prior speculate() on this CPU
2876 			 * and (b) racing with another commit() on a different
2877 			 * CPU.  There's nothing to do -- we just assert that
2878 			 * our offset is 0.
2879 			 */
2880 			ASSERT(src->dtb_offset == 0);
2881 			return;
2882 
2883 		case DTRACESPEC_ACTIVE:
2884 			new = DTRACESPEC_COMMITTING;
2885 			break;
2886 
2887 		case DTRACESPEC_ACTIVEONE:
2888 			/*
2889 			 * This speculation is active on one CPU.  If our
2890 			 * buffer offset is non-zero, we know that the one CPU
2891 			 * must be us.  Otherwise, we are committing on a
2892 			 * different CPU from the speculate(), and we must
2893 			 * rely on being asynchronously cleaned.
2894 			 */
2895 			if (src->dtb_offset != 0) {
2896 				new = DTRACESPEC_COMMITTING;
2897 				break;
2898 			}
2899 			/*FALLTHROUGH*/
2900 
2901 		case DTRACESPEC_ACTIVEMANY:
2902 			new = DTRACESPEC_COMMITTINGMANY;
2903 			break;
2904 
2905 		default:
2906 			ASSERT(0);
2907 		}
2908 	} while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2909 	    curstate, new) != curstate);
2910 
2911 	/*
2912 	 * We have set the state to indicate that we are committing this
2913 	 * speculation.  Now reserve the necessary space in the destination
2914 	 * buffer.
2915 	 */
2916 	if ((offs = dtrace_buffer_reserve(dest, src->dtb_offset,
2917 	    sizeof (uint64_t), state, NULL)) < 0) {
2918 		dtrace_buffer_drop(dest);
2919 		goto out;
2920 	}
2921 
2922 	/*
2923 	 * We have sufficient space to copy the speculative buffer into the
2924 	 * primary buffer.  First, modify the speculative buffer, filling
2925 	 * in the timestamp of all entries with the curstate time.  The data
2926 	 * must have the commit() time rather than the time it was traced,
2927 	 * so that all entries in the primary buffer are in timestamp order.
2928 	 */
2929 	timestamp = dtrace_gethrtime();
2930 	saddr = (uintptr_t)src->dtb_tomax;
2931 	slimit = saddr + src->dtb_offset;
2932 	while (saddr < slimit) {
2933 		size_t size;
2934 		dtrace_rechdr_t *dtrh = (dtrace_rechdr_t *)saddr;
2935 
2936 		if (dtrh->dtrh_epid == DTRACE_EPIDNONE) {
2937 			saddr += sizeof (dtrace_epid_t);
2938 			continue;
2939 		}
2940 		ASSERT3U(dtrh->dtrh_epid, <=, state->dts_necbs);
2941 		size = state->dts_ecbs[dtrh->dtrh_epid - 1]->dte_size;
2942 
2943 		ASSERT3U(saddr + size, <=, slimit);
2944 		ASSERT3U(size, >=, sizeof (dtrace_rechdr_t));
2945 		ASSERT3U(DTRACE_RECORD_LOAD_TIMESTAMP(dtrh), ==, UINT64_MAX);
2946 
2947 		DTRACE_RECORD_STORE_TIMESTAMP(dtrh, timestamp);
2948 
2949 		saddr += size;
2950 	}
2951 
2952 	/*
2953 	 * Copy the buffer across.  (Note that this is a
2954 	 * highly subobtimal bcopy(); in the unlikely event that this becomes
2955 	 * a serious performance issue, a high-performance DTrace-specific
2956 	 * bcopy() should obviously be invented.)
2957 	 */
2958 	daddr = (uintptr_t)dest->dtb_tomax + offs;
2959 	dlimit = daddr + src->dtb_offset;
2960 	saddr = (uintptr_t)src->dtb_tomax;
2961 
2962 	/*
2963 	 * First, the aligned portion.
2964 	 */
2965 	while (dlimit - daddr >= sizeof (uint64_t)) {
2966 		*((uint64_t *)daddr) = *((uint64_t *)saddr);
2967 
2968 		daddr += sizeof (uint64_t);
2969 		saddr += sizeof (uint64_t);
2970 	}
2971 
2972 	/*
2973 	 * Now any left-over bit...
2974 	 */
2975 	while (dlimit - daddr)
2976 		*((uint8_t *)daddr++) = *((uint8_t *)saddr++);
2977 
2978 	/*
2979 	 * Finally, commit the reserved space in the destination buffer.
2980 	 */
2981 	dest->dtb_offset = offs + src->dtb_offset;
2982 
2983 out:
2984 	/*
2985 	 * If we're lucky enough to be the only active CPU on this speculation
2986 	 * buffer, we can just set the state back to DTRACESPEC_INACTIVE.
2987 	 */
2988 	if (curstate == DTRACESPEC_ACTIVE ||
2989 	    (curstate == DTRACESPEC_ACTIVEONE && new == DTRACESPEC_COMMITTING)) {
2990 		uint32_t rval = dtrace_cas32((uint32_t *)&spec->dtsp_state,
2991 		    DTRACESPEC_COMMITTING, DTRACESPEC_INACTIVE);
2992 
2993 		ASSERT(rval == DTRACESPEC_COMMITTING);
2994 	}
2995 
2996 	src->dtb_offset = 0;
2997 	src->dtb_xamot_drops += src->dtb_drops;
2998 	src->dtb_drops = 0;
2999 }
3000 
3001 /*
3002  * This routine discards an active speculation.  If the specified speculation
3003  * is not in a valid state to perform a discard(), this routine will silently
3004  * do nothing.  The state of the specified speculation is transitioned
3005  * according to the state transition diagram outlined in <sys/dtrace_impl.h>
3006  */
3007 static void
3008 dtrace_speculation_discard(dtrace_state_t *state, processorid_t cpu,
3009     dtrace_specid_t which)
3010 {
3011 	dtrace_speculation_t *spec;
3012 	dtrace_speculation_state_t curstate, new = 0;
3013 	dtrace_buffer_t *buf;
3014 
3015 	if (which == 0)
3016 		return;
3017 
3018 	if (which > state->dts_nspeculations) {
3019 		cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
3020 		return;
3021 	}
3022 
3023 	spec = &state->dts_speculations[which - 1];
3024 	buf = &spec->dtsp_buffer[cpu];
3025 
3026 	do {
3027 		curstate = spec->dtsp_state;
3028 
3029 		switch (curstate) {
3030 		case DTRACESPEC_INACTIVE:
3031 		case DTRACESPEC_COMMITTINGMANY:
3032 		case DTRACESPEC_COMMITTING:
3033 		case DTRACESPEC_DISCARDING:
3034 			return;
3035 
3036 		case DTRACESPEC_ACTIVE:
3037 		case DTRACESPEC_ACTIVEMANY:
3038 			new = DTRACESPEC_DISCARDING;
3039 			break;
3040 
3041 		case DTRACESPEC_ACTIVEONE:
3042 			if (buf->dtb_offset != 0) {
3043 				new = DTRACESPEC_INACTIVE;
3044 			} else {
3045 				new = DTRACESPEC_DISCARDING;
3046 			}
3047 			break;
3048 
3049 		default:
3050 			ASSERT(0);
3051 		}
3052 	} while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
3053 	    curstate, new) != curstate);
3054 
3055 	buf->dtb_offset = 0;
3056 	buf->dtb_drops = 0;
3057 }
3058 
3059 /*
3060  * Note:  not called from probe context.  This function is called
3061  * asynchronously from cross call context to clean any speculations that are
3062  * in the COMMITTINGMANY or DISCARDING states.  These speculations may not be
3063  * transitioned back to the INACTIVE state until all CPUs have cleaned the
3064  * speculation.
3065  */
3066 static void
3067 dtrace_speculation_clean_here(dtrace_state_t *state)
3068 {
3069 	dtrace_icookie_t cookie;
3070 	processorid_t cpu = curcpu;
3071 	dtrace_buffer_t *dest = &state->dts_buffer[cpu];
3072 	dtrace_specid_t i;
3073 
3074 	cookie = dtrace_interrupt_disable();
3075 
3076 	if (dest->dtb_tomax == NULL) {
3077 		dtrace_interrupt_enable(cookie);
3078 		return;
3079 	}
3080 
3081 	for (i = 0; i < state->dts_nspeculations; i++) {
3082 		dtrace_speculation_t *spec = &state->dts_speculations[i];
3083 		dtrace_buffer_t *src = &spec->dtsp_buffer[cpu];
3084 
3085 		if (src->dtb_tomax == NULL)
3086 			continue;
3087 
3088 		if (spec->dtsp_state == DTRACESPEC_DISCARDING) {
3089 			src->dtb_offset = 0;
3090 			continue;
3091 		}
3092 
3093 		if (spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
3094 			continue;
3095 
3096 		if (src->dtb_offset == 0)
3097 			continue;
3098 
3099 		dtrace_speculation_commit(state, cpu, i + 1);
3100 	}
3101 
3102 	dtrace_interrupt_enable(cookie);
3103 }
3104 
3105 /*
3106  * Note:  not called from probe context.  This function is called
3107  * asynchronously (and at a regular interval) to clean any speculations that
3108  * are in the COMMITTINGMANY or DISCARDING states.  If it discovers that there
3109  * is work to be done, it cross calls all CPUs to perform that work;
3110  * COMMITMANY and DISCARDING speculations may not be transitioned back to the
3111  * INACTIVE state until they have been cleaned by all CPUs.
3112  */
3113 static void
3114 dtrace_speculation_clean(dtrace_state_t *state)
3115 {
3116 	int work = 0, rv;
3117 	dtrace_specid_t i;
3118 
3119 	for (i = 0; i < state->dts_nspeculations; i++) {
3120 		dtrace_speculation_t *spec = &state->dts_speculations[i];
3121 
3122 		ASSERT(!spec->dtsp_cleaning);
3123 
3124 		if (spec->dtsp_state != DTRACESPEC_DISCARDING &&
3125 		    spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
3126 			continue;
3127 
3128 		work++;
3129 		spec->dtsp_cleaning = 1;
3130 	}
3131 
3132 	if (!work)
3133 		return;
3134 
3135 	dtrace_xcall(DTRACE_CPUALL,
3136 	    (dtrace_xcall_t)dtrace_speculation_clean_here, state);
3137 
3138 	/*
3139 	 * We now know that all CPUs have committed or discarded their
3140 	 * speculation buffers, as appropriate.  We can now set the state
3141 	 * to inactive.
3142 	 */
3143 	for (i = 0; i < state->dts_nspeculations; i++) {
3144 		dtrace_speculation_t *spec = &state->dts_speculations[i];
3145 		dtrace_speculation_state_t curstate, new;
3146 
3147 		if (!spec->dtsp_cleaning)
3148 			continue;
3149 
3150 		curstate = spec->dtsp_state;
3151 		ASSERT(curstate == DTRACESPEC_DISCARDING ||
3152 		    curstate == DTRACESPEC_COMMITTINGMANY);
3153 
3154 		new = DTRACESPEC_INACTIVE;
3155 
3156 		rv = dtrace_cas32((uint32_t *)&spec->dtsp_state, curstate, new);
3157 		ASSERT(rv == curstate);
3158 		spec->dtsp_cleaning = 0;
3159 	}
3160 }
3161 
3162 /*
3163  * Called as part of a speculate() to get the speculative buffer associated
3164  * with a given speculation.  Returns NULL if the specified speculation is not
3165  * in an ACTIVE state.  If the speculation is in the ACTIVEONE state -- and
3166  * the active CPU is not the specified CPU -- the speculation will be
3167  * atomically transitioned into the ACTIVEMANY state.
3168  */
3169 static dtrace_buffer_t *
3170 dtrace_speculation_buffer(dtrace_state_t *state, processorid_t cpuid,
3171     dtrace_specid_t which)
3172 {
3173 	dtrace_speculation_t *spec;
3174 	dtrace_speculation_state_t curstate, new = 0;
3175 	dtrace_buffer_t *buf;
3176 
3177 	if (which == 0)
3178 		return (NULL);
3179 
3180 	if (which > state->dts_nspeculations) {
3181 		cpu_core[cpuid].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
3182 		return (NULL);
3183 	}
3184 
3185 	spec = &state->dts_speculations[which - 1];
3186 	buf = &spec->dtsp_buffer[cpuid];
3187 
3188 	do {
3189 		curstate = spec->dtsp_state;
3190 
3191 		switch (curstate) {
3192 		case DTRACESPEC_INACTIVE:
3193 		case DTRACESPEC_COMMITTINGMANY:
3194 		case DTRACESPEC_DISCARDING:
3195 			return (NULL);
3196 
3197 		case DTRACESPEC_COMMITTING:
3198 			ASSERT(buf->dtb_offset == 0);
3199 			return (NULL);
3200 
3201 		case DTRACESPEC_ACTIVEONE:
3202 			/*
3203 			 * This speculation is currently active on one CPU.
3204 			 * Check the offset in the buffer; if it's non-zero,
3205 			 * that CPU must be us (and we leave the state alone).
3206 			 * If it's zero, assume that we're starting on a new
3207 			 * CPU -- and change the state to indicate that the
3208 			 * speculation is active on more than one CPU.
3209 			 */
3210 			if (buf->dtb_offset != 0)
3211 				return (buf);
3212 
3213 			new = DTRACESPEC_ACTIVEMANY;
3214 			break;
3215 
3216 		case DTRACESPEC_ACTIVEMANY:
3217 			return (buf);
3218 
3219 		case DTRACESPEC_ACTIVE:
3220 			new = DTRACESPEC_ACTIVEONE;
3221 			break;
3222 
3223 		default:
3224 			ASSERT(0);
3225 		}
3226 	} while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
3227 	    curstate, new) != curstate);
3228 
3229 	ASSERT(new == DTRACESPEC_ACTIVEONE || new == DTRACESPEC_ACTIVEMANY);
3230 	return (buf);
3231 }
3232 
3233 /*
3234  * Return a string.  In the event that the user lacks the privilege to access
3235  * arbitrary kernel memory, we copy the string out to scratch memory so that we
3236  * don't fail access checking.
3237  *
3238  * dtrace_dif_variable() uses this routine as a helper for various
3239  * builtin values such as 'execname' and 'probefunc.'
3240  */
3241 uintptr_t
3242 dtrace_dif_varstr(uintptr_t addr, dtrace_state_t *state,
3243     dtrace_mstate_t *mstate)
3244 {
3245 	uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3246 	uintptr_t ret;
3247 	size_t strsz;
3248 
3249 	/*
3250 	 * The easy case: this probe is allowed to read all of memory, so
3251 	 * we can just return this as a vanilla pointer.
3252 	 */
3253 	if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
3254 		return (addr);
3255 
3256 	/*
3257 	 * This is the tougher case: we copy the string in question from
3258 	 * kernel memory into scratch memory and return it that way: this
3259 	 * ensures that we won't trip up when access checking tests the
3260 	 * BYREF return value.
3261 	 */
3262 	strsz = dtrace_strlen((char *)addr, size) + 1;
3263 
3264 	if (mstate->dtms_scratch_ptr + strsz >
3265 	    mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
3266 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3267 		return (0);
3268 	}
3269 
3270 	dtrace_strcpy((const void *)addr, (void *)mstate->dtms_scratch_ptr,
3271 	    strsz);
3272 	ret = mstate->dtms_scratch_ptr;
3273 	mstate->dtms_scratch_ptr += strsz;
3274 	return (ret);
3275 }
3276 
3277 /*
3278  * Return a string from a memoy address which is known to have one or
3279  * more concatenated, individually zero terminated, sub-strings.
3280  * In the event that the user lacks the privilege to access
3281  * arbitrary kernel memory, we copy the string out to scratch memory so that we
3282  * don't fail access checking.
3283  *
3284  * dtrace_dif_variable() uses this routine as a helper for various
3285  * builtin values such as 'execargs'.
3286  */
3287 static uintptr_t
3288 dtrace_dif_varstrz(uintptr_t addr, size_t strsz, dtrace_state_t *state,
3289     dtrace_mstate_t *mstate)
3290 {
3291 	char *p;
3292 	size_t i;
3293 	uintptr_t ret;
3294 
3295 	if (mstate->dtms_scratch_ptr + strsz >
3296 	    mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
3297 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3298 		return (0);
3299 	}
3300 
3301 	dtrace_bcopy((const void *)addr, (void *)mstate->dtms_scratch_ptr,
3302 	    strsz);
3303 
3304 	/* Replace sub-string termination characters with a space. */
3305 	for (p = (char *) mstate->dtms_scratch_ptr, i = 0; i < strsz - 1;
3306 	    p++, i++)
3307 		if (*p == '\0')
3308 			*p = ' ';
3309 
3310 	ret = mstate->dtms_scratch_ptr;
3311 	mstate->dtms_scratch_ptr += strsz;
3312 	return (ret);
3313 }
3314 
3315 /*
3316  * This function implements the DIF emulator's variable lookups.  The emulator
3317  * passes a reserved variable identifier and optional built-in array index.
3318  */
3319 static uint64_t
3320 dtrace_dif_variable(dtrace_mstate_t *mstate, dtrace_state_t *state, uint64_t v,
3321     uint64_t ndx)
3322 {
3323 	/*
3324 	 * If we're accessing one of the uncached arguments, we'll turn this
3325 	 * into a reference in the args array.
3326 	 */
3327 	if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9) {
3328 		ndx = v - DIF_VAR_ARG0;
3329 		v = DIF_VAR_ARGS;
3330 	}
3331 
3332 	switch (v) {
3333 	case DIF_VAR_ARGS:
3334 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_ARGS);
3335 		if (ndx >= sizeof (mstate->dtms_arg) /
3336 		    sizeof (mstate->dtms_arg[0])) {
3337 			int aframes = mstate->dtms_probe->dtpr_aframes + 2;
3338 			dtrace_provider_t *pv;
3339 			uint64_t val;
3340 
3341 			pv = mstate->dtms_probe->dtpr_provider;
3342 			if (pv->dtpv_pops.dtps_getargval != NULL)
3343 				val = pv->dtpv_pops.dtps_getargval(pv->dtpv_arg,
3344 				    mstate->dtms_probe->dtpr_id,
3345 				    mstate->dtms_probe->dtpr_arg, ndx, aframes);
3346 			else
3347 				val = dtrace_getarg(ndx, aframes);
3348 
3349 			/*
3350 			 * This is regrettably required to keep the compiler
3351 			 * from tail-optimizing the call to dtrace_getarg().
3352 			 * The condition always evaluates to true, but the
3353 			 * compiler has no way of figuring that out a priori.
3354 			 * (None of this would be necessary if the compiler
3355 			 * could be relied upon to _always_ tail-optimize
3356 			 * the call to dtrace_getarg() -- but it can't.)
3357 			 */
3358 			if (mstate->dtms_probe != NULL)
3359 				return (val);
3360 
3361 			ASSERT(0);
3362 		}
3363 
3364 		return (mstate->dtms_arg[ndx]);
3365 
3366 	case DIF_VAR_REGS:
3367 	case DIF_VAR_UREGS: {
3368 		struct trapframe *tframe;
3369 
3370 		if (!dtrace_priv_proc(state))
3371 			return (0);
3372 
3373 		if (v == DIF_VAR_REGS)
3374 			tframe = curthread->t_dtrace_trapframe;
3375 		else
3376 			tframe = curthread->td_frame;
3377 
3378 		if (tframe == NULL) {
3379 			DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
3380 			cpu_core[curcpu].cpuc_dtrace_illval = 0;
3381 			return (0);
3382 		}
3383 
3384 		return (dtrace_getreg(tframe, ndx));
3385 	}
3386 
3387 	case DIF_VAR_CURTHREAD:
3388 		if (!dtrace_priv_proc(state))
3389 			return (0);
3390 		return ((uint64_t)(uintptr_t)curthread);
3391 
3392 	case DIF_VAR_TIMESTAMP:
3393 		if (!(mstate->dtms_present & DTRACE_MSTATE_TIMESTAMP)) {
3394 			mstate->dtms_timestamp = dtrace_gethrtime();
3395 			mstate->dtms_present |= DTRACE_MSTATE_TIMESTAMP;
3396 		}
3397 		return (mstate->dtms_timestamp);
3398 
3399 	case DIF_VAR_VTIMESTAMP:
3400 		ASSERT(dtrace_vtime_references != 0);
3401 		return (curthread->t_dtrace_vtime);
3402 
3403 	case DIF_VAR_WALLTIMESTAMP:
3404 		if (!(mstate->dtms_present & DTRACE_MSTATE_WALLTIMESTAMP)) {
3405 			mstate->dtms_walltimestamp = dtrace_gethrestime();
3406 			mstate->dtms_present |= DTRACE_MSTATE_WALLTIMESTAMP;
3407 		}
3408 		return (mstate->dtms_walltimestamp);
3409 
3410 #ifdef illumos
3411 	case DIF_VAR_IPL:
3412 		if (!dtrace_priv_kernel(state))
3413 			return (0);
3414 		if (!(mstate->dtms_present & DTRACE_MSTATE_IPL)) {
3415 			mstate->dtms_ipl = dtrace_getipl();
3416 			mstate->dtms_present |= DTRACE_MSTATE_IPL;
3417 		}
3418 		return (mstate->dtms_ipl);
3419 #endif
3420 
3421 	case DIF_VAR_EPID:
3422 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_EPID);
3423 		return (mstate->dtms_epid);
3424 
3425 	case DIF_VAR_ID:
3426 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3427 		return (mstate->dtms_probe->dtpr_id);
3428 
3429 	case DIF_VAR_STACKDEPTH:
3430 		if (!dtrace_priv_kernel(state))
3431 			return (0);
3432 		if (!(mstate->dtms_present & DTRACE_MSTATE_STACKDEPTH)) {
3433 			int aframes = mstate->dtms_probe->dtpr_aframes + 2;
3434 
3435 			mstate->dtms_stackdepth = dtrace_getstackdepth(aframes);
3436 			mstate->dtms_present |= DTRACE_MSTATE_STACKDEPTH;
3437 		}
3438 		return (mstate->dtms_stackdepth);
3439 
3440 	case DIF_VAR_USTACKDEPTH:
3441 		if (!dtrace_priv_proc(state))
3442 			return (0);
3443 		if (!(mstate->dtms_present & DTRACE_MSTATE_USTACKDEPTH)) {
3444 			/*
3445 			 * See comment in DIF_VAR_PID.
3446 			 */
3447 			if (DTRACE_ANCHORED(mstate->dtms_probe) &&
3448 			    CPU_ON_INTR(CPU)) {
3449 				mstate->dtms_ustackdepth = 0;
3450 			} else {
3451 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3452 				mstate->dtms_ustackdepth =
3453 				    dtrace_getustackdepth();
3454 				DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3455 			}
3456 			mstate->dtms_present |= DTRACE_MSTATE_USTACKDEPTH;
3457 		}
3458 		return (mstate->dtms_ustackdepth);
3459 
3460 	case DIF_VAR_CALLER:
3461 		if (!dtrace_priv_kernel(state))
3462 			return (0);
3463 		if (!(mstate->dtms_present & DTRACE_MSTATE_CALLER)) {
3464 			int aframes = mstate->dtms_probe->dtpr_aframes + 2;
3465 
3466 			if (!DTRACE_ANCHORED(mstate->dtms_probe)) {
3467 				/*
3468 				 * If this is an unanchored probe, we are
3469 				 * required to go through the slow path:
3470 				 * dtrace_caller() only guarantees correct
3471 				 * results for anchored probes.
3472 				 */
3473 				pc_t caller[2] = {0, 0};
3474 
3475 				dtrace_getpcstack(caller, 2, aframes,
3476 				    (uint32_t *)(uintptr_t)mstate->dtms_arg[0]);
3477 				mstate->dtms_caller = caller[1];
3478 			} else if ((mstate->dtms_caller =
3479 			    dtrace_caller(aframes)) == -1) {
3480 				/*
3481 				 * We have failed to do this the quick way;
3482 				 * we must resort to the slower approach of
3483 				 * calling dtrace_getpcstack().
3484 				 */
3485 				pc_t caller = 0;
3486 
3487 				dtrace_getpcstack(&caller, 1, aframes, NULL);
3488 				mstate->dtms_caller = caller;
3489 			}
3490 
3491 			mstate->dtms_present |= DTRACE_MSTATE_CALLER;
3492 		}
3493 		return (mstate->dtms_caller);
3494 
3495 	case DIF_VAR_UCALLER:
3496 		if (!dtrace_priv_proc(state))
3497 			return (0);
3498 
3499 		if (!(mstate->dtms_present & DTRACE_MSTATE_UCALLER)) {
3500 			uint64_t ustack[3];
3501 
3502 			/*
3503 			 * dtrace_getupcstack() fills in the first uint64_t
3504 			 * with the current PID.  The second uint64_t will
3505 			 * be the program counter at user-level.  The third
3506 			 * uint64_t will contain the caller, which is what
3507 			 * we're after.
3508 			 */
3509 			ustack[2] = 0;
3510 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3511 			dtrace_getupcstack(ustack, 3);
3512 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3513 			mstate->dtms_ucaller = ustack[2];
3514 			mstate->dtms_present |= DTRACE_MSTATE_UCALLER;
3515 		}
3516 
3517 		return (mstate->dtms_ucaller);
3518 
3519 	case DIF_VAR_PROBEPROV:
3520 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3521 		return (dtrace_dif_varstr(
3522 		    (uintptr_t)mstate->dtms_probe->dtpr_provider->dtpv_name,
3523 		    state, mstate));
3524 
3525 	case DIF_VAR_PROBEMOD:
3526 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3527 		return (dtrace_dif_varstr(
3528 		    (uintptr_t)mstate->dtms_probe->dtpr_mod,
3529 		    state, mstate));
3530 
3531 	case DIF_VAR_PROBEFUNC:
3532 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3533 		return (dtrace_dif_varstr(
3534 		    (uintptr_t)mstate->dtms_probe->dtpr_func,
3535 		    state, mstate));
3536 
3537 	case DIF_VAR_PROBENAME:
3538 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3539 		return (dtrace_dif_varstr(
3540 		    (uintptr_t)mstate->dtms_probe->dtpr_name,
3541 		    state, mstate));
3542 
3543 	case DIF_VAR_PID:
3544 		if (!dtrace_priv_proc(state))
3545 			return (0);
3546 
3547 #ifdef illumos
3548 		/*
3549 		 * Note that we are assuming that an unanchored probe is
3550 		 * always due to a high-level interrupt.  (And we're assuming
3551 		 * that there is only a single high level interrupt.)
3552 		 */
3553 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3554 			return (pid0.pid_id);
3555 
3556 		/*
3557 		 * It is always safe to dereference one's own t_procp pointer:
3558 		 * it always points to a valid, allocated proc structure.
3559 		 * Further, it is always safe to dereference the p_pidp member
3560 		 * of one's own proc structure.  (These are truisms becuase
3561 		 * threads and processes don't clean up their own state --
3562 		 * they leave that task to whomever reaps them.)
3563 		 */
3564 		return ((uint64_t)curthread->t_procp->p_pidp->pid_id);
3565 #else
3566 		return ((uint64_t)curproc->p_pid);
3567 #endif
3568 
3569 	case DIF_VAR_PPID:
3570 		if (!dtrace_priv_proc(state))
3571 			return (0);
3572 
3573 #ifdef illumos
3574 		/*
3575 		 * See comment in DIF_VAR_PID.
3576 		 */
3577 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3578 			return (pid0.pid_id);
3579 
3580 		/*
3581 		 * It is always safe to dereference one's own t_procp pointer:
3582 		 * it always points to a valid, allocated proc structure.
3583 		 * (This is true because threads don't clean up their own
3584 		 * state -- they leave that task to whomever reaps them.)
3585 		 */
3586 		return ((uint64_t)curthread->t_procp->p_ppid);
3587 #else
3588 		if (curproc->p_pid == proc0.p_pid)
3589 			return (curproc->p_pid);
3590 		else
3591 			return (curproc->p_pptr->p_pid);
3592 #endif
3593 
3594 	case DIF_VAR_TID:
3595 #ifdef illumos
3596 		/*
3597 		 * See comment in DIF_VAR_PID.
3598 		 */
3599 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3600 			return (0);
3601 #endif
3602 
3603 		return ((uint64_t)curthread->t_tid);
3604 
3605 	case DIF_VAR_EXECARGS: {
3606 		struct pargs *p_args = curthread->td_proc->p_args;
3607 
3608 		if (p_args == NULL)
3609 			return(0);
3610 
3611 		return (dtrace_dif_varstrz(
3612 		    (uintptr_t) p_args->ar_args, p_args->ar_length, state, mstate));
3613 	}
3614 
3615 	case DIF_VAR_EXECNAME:
3616 #ifdef illumos
3617 		if (!dtrace_priv_proc(state))
3618 			return (0);
3619 
3620 		/*
3621 		 * See comment in DIF_VAR_PID.
3622 		 */
3623 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3624 			return ((uint64_t)(uintptr_t)p0.p_user.u_comm);
3625 
3626 		/*
3627 		 * It is always safe to dereference one's own t_procp pointer:
3628 		 * it always points to a valid, allocated proc structure.
3629 		 * (This is true because threads don't clean up their own
3630 		 * state -- they leave that task to whomever reaps them.)
3631 		 */
3632 		return (dtrace_dif_varstr(
3633 		    (uintptr_t)curthread->t_procp->p_user.u_comm,
3634 		    state, mstate));
3635 #else
3636 		return (dtrace_dif_varstr(
3637 		    (uintptr_t) curthread->td_proc->p_comm, state, mstate));
3638 #endif
3639 
3640 	case DIF_VAR_ZONENAME:
3641 #ifdef illumos
3642 		if (!dtrace_priv_proc(state))
3643 			return (0);
3644 
3645 		/*
3646 		 * See comment in DIF_VAR_PID.
3647 		 */
3648 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3649 			return ((uint64_t)(uintptr_t)p0.p_zone->zone_name);
3650 
3651 		/*
3652 		 * It is always safe to dereference one's own t_procp pointer:
3653 		 * it always points to a valid, allocated proc structure.
3654 		 * (This is true because threads don't clean up their own
3655 		 * state -- they leave that task to whomever reaps them.)
3656 		 */
3657 		return (dtrace_dif_varstr(
3658 		    (uintptr_t)curthread->t_procp->p_zone->zone_name,
3659 		    state, mstate));
3660 #elif defined(__FreeBSD__)
3661 	/*
3662 	 * On FreeBSD, we introduce compatibility to zonename by falling through
3663 	 * into jailname.
3664 	 */
3665 	case DIF_VAR_JAILNAME:
3666 		if (!dtrace_priv_kernel(state))
3667 			return (0);
3668 
3669 		return (dtrace_dif_varstr(
3670 		    (uintptr_t)curthread->td_ucred->cr_prison->pr_name,
3671 		    state, mstate));
3672 
3673 	case DIF_VAR_JID:
3674 		if (!dtrace_priv_kernel(state))
3675 			return (0);
3676 
3677 		return ((uint64_t)curthread->td_ucred->cr_prison->pr_id);
3678 #else
3679 		return (0);
3680 #endif
3681 
3682 	case DIF_VAR_UID:
3683 		if (!dtrace_priv_proc(state))
3684 			return (0);
3685 
3686 #ifdef illumos
3687 		/*
3688 		 * See comment in DIF_VAR_PID.
3689 		 */
3690 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3691 			return ((uint64_t)p0.p_cred->cr_uid);
3692 
3693 		/*
3694 		 * It is always safe to dereference one's own t_procp pointer:
3695 		 * it always points to a valid, allocated proc structure.
3696 		 * (This is true because threads don't clean up their own
3697 		 * state -- they leave that task to whomever reaps them.)
3698 		 *
3699 		 * Additionally, it is safe to dereference one's own process
3700 		 * credential, since this is never NULL after process birth.
3701 		 */
3702 		return ((uint64_t)curthread->t_procp->p_cred->cr_uid);
3703 #else
3704 		return ((uint64_t)curthread->td_ucred->cr_uid);
3705 #endif
3706 
3707 	case DIF_VAR_GID:
3708 		if (!dtrace_priv_proc(state))
3709 			return (0);
3710 
3711 #ifdef illumos
3712 		/*
3713 		 * See comment in DIF_VAR_PID.
3714 		 */
3715 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3716 			return ((uint64_t)p0.p_cred->cr_gid);
3717 
3718 		/*
3719 		 * It is always safe to dereference one's own t_procp pointer:
3720 		 * it always points to a valid, allocated proc structure.
3721 		 * (This is true because threads don't clean up their own
3722 		 * state -- they leave that task to whomever reaps them.)
3723 		 *
3724 		 * Additionally, it is safe to dereference one's own process
3725 		 * credential, since this is never NULL after process birth.
3726 		 */
3727 		return ((uint64_t)curthread->t_procp->p_cred->cr_gid);
3728 #else
3729 		return ((uint64_t)curthread->td_ucred->cr_gid);
3730 #endif
3731 
3732 	case DIF_VAR_ERRNO: {
3733 #ifdef illumos
3734 		klwp_t *lwp;
3735 		if (!dtrace_priv_proc(state))
3736 			return (0);
3737 
3738 		/*
3739 		 * See comment in DIF_VAR_PID.
3740 		 */
3741 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3742 			return (0);
3743 
3744 		/*
3745 		 * It is always safe to dereference one's own t_lwp pointer in
3746 		 * the event that this pointer is non-NULL.  (This is true
3747 		 * because threads and lwps don't clean up their own state --
3748 		 * they leave that task to whomever reaps them.)
3749 		 */
3750 		if ((lwp = curthread->t_lwp) == NULL)
3751 			return (0);
3752 
3753 		return ((uint64_t)lwp->lwp_errno);
3754 #else
3755 		return (curthread->td_errno);
3756 #endif
3757 	}
3758 #ifndef illumos
3759 	case DIF_VAR_CPU: {
3760 		return curcpu;
3761 	}
3762 #endif
3763 	default:
3764 		DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
3765 		return (0);
3766 	}
3767 }
3768 
3769 
3770 typedef enum dtrace_json_state {
3771 	DTRACE_JSON_REST = 1,
3772 	DTRACE_JSON_OBJECT,
3773 	DTRACE_JSON_STRING,
3774 	DTRACE_JSON_STRING_ESCAPE,
3775 	DTRACE_JSON_STRING_ESCAPE_UNICODE,
3776 	DTRACE_JSON_COLON,
3777 	DTRACE_JSON_COMMA,
3778 	DTRACE_JSON_VALUE,
3779 	DTRACE_JSON_IDENTIFIER,
3780 	DTRACE_JSON_NUMBER,
3781 	DTRACE_JSON_NUMBER_FRAC,
3782 	DTRACE_JSON_NUMBER_EXP,
3783 	DTRACE_JSON_COLLECT_OBJECT
3784 } dtrace_json_state_t;
3785 
3786 /*
3787  * This function possesses just enough knowledge about JSON to extract a single
3788  * value from a JSON string and store it in the scratch buffer.  It is able
3789  * to extract nested object values, and members of arrays by index.
3790  *
3791  * elemlist is a list of JSON keys, stored as packed NUL-terminated strings, to
3792  * be looked up as we descend into the object tree.  e.g.
3793  *
3794  *    foo[0].bar.baz[32] --> "foo" NUL "0" NUL "bar" NUL "baz" NUL "32" NUL
3795  *       with nelems = 5.
3796  *
3797  * The run time of this function must be bounded above by strsize to limit the
3798  * amount of work done in probe context.  As such, it is implemented as a
3799  * simple state machine, reading one character at a time using safe loads
3800  * until we find the requested element, hit a parsing error or run off the
3801  * end of the object or string.
3802  *
3803  * As there is no way for a subroutine to return an error without interrupting
3804  * clause execution, we simply return NULL in the event of a missing key or any
3805  * other error condition.  Each NULL return in this function is commented with
3806  * the error condition it represents -- parsing or otherwise.
3807  *
3808  * The set of states for the state machine closely matches the JSON
3809  * specification (http://json.org/).  Briefly:
3810  *
3811  *   DTRACE_JSON_REST:
3812  *     Skip whitespace until we find either a top-level Object, moving
3813  *     to DTRACE_JSON_OBJECT; or an Array, moving to DTRACE_JSON_VALUE.
3814  *
3815  *   DTRACE_JSON_OBJECT:
3816  *     Locate the next key String in an Object.  Sets a flag to denote
3817  *     the next String as a key string and moves to DTRACE_JSON_STRING.
3818  *
3819  *   DTRACE_JSON_COLON:
3820  *     Skip whitespace until we find the colon that separates key Strings
3821  *     from their values.  Once found, move to DTRACE_JSON_VALUE.
3822  *
3823  *   DTRACE_JSON_VALUE:
3824  *     Detects the type of the next value (String, Number, Identifier, Object
3825  *     or Array) and routes to the states that process that type.  Here we also
3826  *     deal with the element selector list if we are requested to traverse down
3827  *     into the object tree.
3828  *
3829  *   DTRACE_JSON_COMMA:
3830  *     Skip whitespace until we find the comma that separates key-value pairs
3831  *     in Objects (returning to DTRACE_JSON_OBJECT) or values in Arrays
3832  *     (similarly DTRACE_JSON_VALUE).  All following literal value processing
3833  *     states return to this state at the end of their value, unless otherwise
3834  *     noted.
3835  *
3836  *   DTRACE_JSON_NUMBER, DTRACE_JSON_NUMBER_FRAC, DTRACE_JSON_NUMBER_EXP:
3837  *     Processes a Number literal from the JSON, including any exponent
3838  *     component that may be present.  Numbers are returned as strings, which
3839  *     may be passed to strtoll() if an integer is required.
3840  *
3841  *   DTRACE_JSON_IDENTIFIER:
3842  *     Processes a "true", "false" or "null" literal in the JSON.
3843  *
3844  *   DTRACE_JSON_STRING, DTRACE_JSON_STRING_ESCAPE,
3845  *   DTRACE_JSON_STRING_ESCAPE_UNICODE:
3846  *     Processes a String literal from the JSON, whether the String denotes
3847  *     a key, a value or part of a larger Object.  Handles all escape sequences
3848  *     present in the specification, including four-digit unicode characters,
3849  *     but merely includes the escape sequence without converting it to the
3850  *     actual escaped character.  If the String is flagged as a key, we
3851  *     move to DTRACE_JSON_COLON rather than DTRACE_JSON_COMMA.
3852  *
3853  *   DTRACE_JSON_COLLECT_OBJECT:
3854  *     This state collects an entire Object (or Array), correctly handling
3855  *     embedded strings.  If the full element selector list matches this nested
3856  *     object, we return the Object in full as a string.  If not, we use this
3857  *     state to skip to the next value at this level and continue processing.
3858  *
3859  * NOTE: This function uses various macros from strtolctype.h to manipulate
3860  * digit values, etc -- these have all been checked to ensure they make
3861  * no additional function calls.
3862  */
3863 static char *
3864 dtrace_json(uint64_t size, uintptr_t json, char *elemlist, int nelems,
3865     char *dest)
3866 {
3867 	dtrace_json_state_t state = DTRACE_JSON_REST;
3868 	int64_t array_elem = INT64_MIN;
3869 	int64_t array_pos = 0;
3870 	uint8_t escape_unicount = 0;
3871 	boolean_t string_is_key = B_FALSE;
3872 	boolean_t collect_object = B_FALSE;
3873 	boolean_t found_key = B_FALSE;
3874 	boolean_t in_array = B_FALSE;
3875 	uint32_t braces = 0, brackets = 0;
3876 	char *elem = elemlist;
3877 	char *dd = dest;
3878 	uintptr_t cur;
3879 
3880 	for (cur = json; cur < json + size; cur++) {
3881 		char cc = dtrace_load8(cur);
3882 		if (cc == '\0')
3883 			return (NULL);
3884 
3885 		switch (state) {
3886 		case DTRACE_JSON_REST:
3887 			if (isspace(cc))
3888 				break;
3889 
3890 			if (cc == '{') {
3891 				state = DTRACE_JSON_OBJECT;
3892 				break;
3893 			}
3894 
3895 			if (cc == '[') {
3896 				in_array = B_TRUE;
3897 				array_pos = 0;
3898 				array_elem = dtrace_strtoll(elem, 10, size);
3899 				found_key = array_elem == 0 ? B_TRUE : B_FALSE;
3900 				state = DTRACE_JSON_VALUE;
3901 				break;
3902 			}
3903 
3904 			/*
3905 			 * ERROR: expected to find a top-level object or array.
3906 			 */
3907 			return (NULL);
3908 		case DTRACE_JSON_OBJECT:
3909 			if (isspace(cc))
3910 				break;
3911 
3912 			if (cc == '"') {
3913 				state = DTRACE_JSON_STRING;
3914 				string_is_key = B_TRUE;
3915 				break;
3916 			}
3917 
3918 			/*
3919 			 * ERROR: either the object did not start with a key
3920 			 * string, or we've run off the end of the object
3921 			 * without finding the requested key.
3922 			 */
3923 			return (NULL);
3924 		case DTRACE_JSON_STRING:
3925 			if (cc == '\\') {
3926 				*dd++ = '\\';
3927 				state = DTRACE_JSON_STRING_ESCAPE;
3928 				break;
3929 			}
3930 
3931 			if (cc == '"') {
3932 				if (collect_object) {
3933 					/*
3934 					 * We don't reset the dest here, as
3935 					 * the string is part of a larger
3936 					 * object being collected.
3937 					 */
3938 					*dd++ = cc;
3939 					collect_object = B_FALSE;
3940 					state = DTRACE_JSON_COLLECT_OBJECT;
3941 					break;
3942 				}
3943 				*dd = '\0';
3944 				dd = dest; /* reset string buffer */
3945 				if (string_is_key) {
3946 					if (dtrace_strncmp(dest, elem,
3947 					    size) == 0)
3948 						found_key = B_TRUE;
3949 				} else if (found_key) {
3950 					if (nelems > 1) {
3951 						/*
3952 						 * We expected an object, not
3953 						 * this string.
3954 						 */
3955 						return (NULL);
3956 					}
3957 					return (dest);
3958 				}
3959 				state = string_is_key ? DTRACE_JSON_COLON :
3960 				    DTRACE_JSON_COMMA;
3961 				string_is_key = B_FALSE;
3962 				break;
3963 			}
3964 
3965 			*dd++ = cc;
3966 			break;
3967 		case DTRACE_JSON_STRING_ESCAPE:
3968 			*dd++ = cc;
3969 			if (cc == 'u') {
3970 				escape_unicount = 0;
3971 				state = DTRACE_JSON_STRING_ESCAPE_UNICODE;
3972 			} else {
3973 				state = DTRACE_JSON_STRING;
3974 			}
3975 			break;
3976 		case DTRACE_JSON_STRING_ESCAPE_UNICODE:
3977 			if (!isxdigit(cc)) {
3978 				/*
3979 				 * ERROR: invalid unicode escape, expected
3980 				 * four valid hexidecimal digits.
3981 				 */
3982 				return (NULL);
3983 			}
3984 
3985 			*dd++ = cc;
3986 			if (++escape_unicount == 4)
3987 				state = DTRACE_JSON_STRING;
3988 			break;
3989 		case DTRACE_JSON_COLON:
3990 			if (isspace(cc))
3991 				break;
3992 
3993 			if (cc == ':') {
3994 				state = DTRACE_JSON_VALUE;
3995 				break;
3996 			}
3997 
3998 			/*
3999 			 * ERROR: expected a colon.
4000 			 */
4001 			return (NULL);
4002 		case DTRACE_JSON_COMMA:
4003 			if (isspace(cc))
4004 				break;
4005 
4006 			if (cc == ',') {
4007 				if (in_array) {
4008 					state = DTRACE_JSON_VALUE;
4009 					if (++array_pos == array_elem)
4010 						found_key = B_TRUE;
4011 				} else {
4012 					state = DTRACE_JSON_OBJECT;
4013 				}
4014 				break;
4015 			}
4016 
4017 			/*
4018 			 * ERROR: either we hit an unexpected character, or
4019 			 * we reached the end of the object or array without
4020 			 * finding the requested key.
4021 			 */
4022 			return (NULL);
4023 		case DTRACE_JSON_IDENTIFIER:
4024 			if (islower(cc)) {
4025 				*dd++ = cc;
4026 				break;
4027 			}
4028 
4029 			*dd = '\0';
4030 			dd = dest; /* reset string buffer */
4031 
4032 			if (dtrace_strncmp(dest, "true", 5) == 0 ||
4033 			    dtrace_strncmp(dest, "false", 6) == 0 ||
4034 			    dtrace_strncmp(dest, "null", 5) == 0) {
4035 				if (found_key) {
4036 					if (nelems > 1) {
4037 						/*
4038 						 * ERROR: We expected an object,
4039 						 * not this identifier.
4040 						 */
4041 						return (NULL);
4042 					}
4043 					return (dest);
4044 				} else {
4045 					cur--;
4046 					state = DTRACE_JSON_COMMA;
4047 					break;
4048 				}
4049 			}
4050 
4051 			/*
4052 			 * ERROR: we did not recognise the identifier as one
4053 			 * of those in the JSON specification.
4054 			 */
4055 			return (NULL);
4056 		case DTRACE_JSON_NUMBER:
4057 			if (cc == '.') {
4058 				*dd++ = cc;
4059 				state = DTRACE_JSON_NUMBER_FRAC;
4060 				break;
4061 			}
4062 
4063 			if (cc == 'x' || cc == 'X') {
4064 				/*
4065 				 * ERROR: specification explicitly excludes
4066 				 * hexidecimal or octal numbers.
4067 				 */
4068 				return (NULL);
4069 			}
4070 
4071 			/* FALLTHRU */
4072 		case DTRACE_JSON_NUMBER_FRAC:
4073 			if (cc == 'e' || cc == 'E') {
4074 				*dd++ = cc;
4075 				state = DTRACE_JSON_NUMBER_EXP;
4076 				break;
4077 			}
4078 
4079 			if (cc == '+' || cc == '-') {
4080 				/*
4081 				 * ERROR: expect sign as part of exponent only.
4082 				 */
4083 				return (NULL);
4084 			}
4085 			/* FALLTHRU */
4086 		case DTRACE_JSON_NUMBER_EXP:
4087 			if (isdigit(cc) || cc == '+' || cc == '-') {
4088 				*dd++ = cc;
4089 				break;
4090 			}
4091 
4092 			*dd = '\0';
4093 			dd = dest; /* reset string buffer */
4094 			if (found_key) {
4095 				if (nelems > 1) {
4096 					/*
4097 					 * ERROR: We expected an object, not
4098 					 * this number.
4099 					 */
4100 					return (NULL);
4101 				}
4102 				return (dest);
4103 			}
4104 
4105 			cur--;
4106 			state = DTRACE_JSON_COMMA;
4107 			break;
4108 		case DTRACE_JSON_VALUE:
4109 			if (isspace(cc))
4110 				break;
4111 
4112 			if (cc == '{' || cc == '[') {
4113 				if (nelems > 1 && found_key) {
4114 					in_array = cc == '[' ? B_TRUE : B_FALSE;
4115 					/*
4116 					 * If our element selector directs us
4117 					 * to descend into this nested object,
4118 					 * then move to the next selector
4119 					 * element in the list and restart the
4120 					 * state machine.
4121 					 */
4122 					while (*elem != '\0')
4123 						elem++;
4124 					elem++; /* skip the inter-element NUL */
4125 					nelems--;
4126 					dd = dest;
4127 					if (in_array) {
4128 						state = DTRACE_JSON_VALUE;
4129 						array_pos = 0;
4130 						array_elem = dtrace_strtoll(
4131 						    elem, 10, size);
4132 						found_key = array_elem == 0 ?
4133 						    B_TRUE : B_FALSE;
4134 					} else {
4135 						found_key = B_FALSE;
4136 						state = DTRACE_JSON_OBJECT;
4137 					}
4138 					break;
4139 				}
4140 
4141 				/*
4142 				 * Otherwise, we wish to either skip this
4143 				 * nested object or return it in full.
4144 				 */
4145 				if (cc == '[')
4146 					brackets = 1;
4147 				else
4148 					braces = 1;
4149 				*dd++ = cc;
4150 				state = DTRACE_JSON_COLLECT_OBJECT;
4151 				break;
4152 			}
4153 
4154 			if (cc == '"') {
4155 				state = DTRACE_JSON_STRING;
4156 				break;
4157 			}
4158 
4159 			if (islower(cc)) {
4160 				/*
4161 				 * Here we deal with true, false and null.
4162 				 */
4163 				*dd++ = cc;
4164 				state = DTRACE_JSON_IDENTIFIER;
4165 				break;
4166 			}
4167 
4168 			if (cc == '-' || isdigit(cc)) {
4169 				*dd++ = cc;
4170 				state = DTRACE_JSON_NUMBER;
4171 				break;
4172 			}
4173 
4174 			/*
4175 			 * ERROR: unexpected character at start of value.
4176 			 */
4177 			return (NULL);
4178 		case DTRACE_JSON_COLLECT_OBJECT:
4179 			if (cc == '\0')
4180 				/*
4181 				 * ERROR: unexpected end of input.
4182 				 */
4183 				return (NULL);
4184 
4185 			*dd++ = cc;
4186 			if (cc == '"') {
4187 				collect_object = B_TRUE;
4188 				state = DTRACE_JSON_STRING;
4189 				break;
4190 			}
4191 
4192 			if (cc == ']') {
4193 				if (brackets-- == 0) {
4194 					/*
4195 					 * ERROR: unbalanced brackets.
4196 					 */
4197 					return (NULL);
4198 				}
4199 			} else if (cc == '}') {
4200 				if (braces-- == 0) {
4201 					/*
4202 					 * ERROR: unbalanced braces.
4203 					 */
4204 					return (NULL);
4205 				}
4206 			} else if (cc == '{') {
4207 				braces++;
4208 			} else if (cc == '[') {
4209 				brackets++;
4210 			}
4211 
4212 			if (brackets == 0 && braces == 0) {
4213 				if (found_key) {
4214 					*dd = '\0';
4215 					return (dest);
4216 				}
4217 				dd = dest; /* reset string buffer */
4218 				state = DTRACE_JSON_COMMA;
4219 			}
4220 			break;
4221 		}
4222 	}
4223 	return (NULL);
4224 }
4225 
4226 /*
4227  * Emulate the execution of DTrace ID subroutines invoked by the call opcode.
4228  * Notice that we don't bother validating the proper number of arguments or
4229  * their types in the tuple stack.  This isn't needed because all argument
4230  * interpretation is safe because of our load safety -- the worst that can
4231  * happen is that a bogus program can obtain bogus results.
4232  */
4233 static void
4234 dtrace_dif_subr(uint_t subr, uint_t rd, uint64_t *regs,
4235     dtrace_key_t *tupregs, int nargs,
4236     dtrace_mstate_t *mstate, dtrace_state_t *state)
4237 {
4238 	volatile uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
4239 	volatile uintptr_t *illval = &cpu_core[curcpu].cpuc_dtrace_illval;
4240 	dtrace_vstate_t *vstate = &state->dts_vstate;
4241 
4242 #ifdef illumos
4243 	union {
4244 		mutex_impl_t mi;
4245 		uint64_t mx;
4246 	} m;
4247 
4248 	union {
4249 		krwlock_t ri;
4250 		uintptr_t rw;
4251 	} r;
4252 #else
4253 	struct thread *lowner;
4254 	union {
4255 		struct lock_object *li;
4256 		uintptr_t lx;
4257 	} l;
4258 #endif
4259 
4260 	switch (subr) {
4261 	case DIF_SUBR_RAND:
4262 		regs[rd] = dtrace_xoroshiro128_plus_next(
4263 		    state->dts_rstate[curcpu]);
4264 		break;
4265 
4266 #ifdef illumos
4267 	case DIF_SUBR_MUTEX_OWNED:
4268 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
4269 		    mstate, vstate)) {
4270 			regs[rd] = 0;
4271 			break;
4272 		}
4273 
4274 		m.mx = dtrace_load64(tupregs[0].dttk_value);
4275 		if (MUTEX_TYPE_ADAPTIVE(&m.mi))
4276 			regs[rd] = MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER;
4277 		else
4278 			regs[rd] = LOCK_HELD(&m.mi.m_spin.m_spinlock);
4279 		break;
4280 
4281 	case DIF_SUBR_MUTEX_OWNER:
4282 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
4283 		    mstate, vstate)) {
4284 			regs[rd] = 0;
4285 			break;
4286 		}
4287 
4288 		m.mx = dtrace_load64(tupregs[0].dttk_value);
4289 		if (MUTEX_TYPE_ADAPTIVE(&m.mi) &&
4290 		    MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER)
4291 			regs[rd] = (uintptr_t)MUTEX_OWNER(&m.mi);
4292 		else
4293 			regs[rd] = 0;
4294 		break;
4295 
4296 	case DIF_SUBR_MUTEX_TYPE_ADAPTIVE:
4297 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
4298 		    mstate, vstate)) {
4299 			regs[rd] = 0;
4300 			break;
4301 		}
4302 
4303 		m.mx = dtrace_load64(tupregs[0].dttk_value);
4304 		regs[rd] = MUTEX_TYPE_ADAPTIVE(&m.mi);
4305 		break;
4306 
4307 	case DIF_SUBR_MUTEX_TYPE_SPIN:
4308 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
4309 		    mstate, vstate)) {
4310 			regs[rd] = 0;
4311 			break;
4312 		}
4313 
4314 		m.mx = dtrace_load64(tupregs[0].dttk_value);
4315 		regs[rd] = MUTEX_TYPE_SPIN(&m.mi);
4316 		break;
4317 
4318 	case DIF_SUBR_RW_READ_HELD: {
4319 		uintptr_t tmp;
4320 
4321 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
4322 		    mstate, vstate)) {
4323 			regs[rd] = 0;
4324 			break;
4325 		}
4326 
4327 		r.rw = dtrace_loadptr(tupregs[0].dttk_value);
4328 		regs[rd] = _RW_READ_HELD(&r.ri, tmp);
4329 		break;
4330 	}
4331 
4332 	case DIF_SUBR_RW_WRITE_HELD:
4333 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
4334 		    mstate, vstate)) {
4335 			regs[rd] = 0;
4336 			break;
4337 		}
4338 
4339 		r.rw = dtrace_loadptr(tupregs[0].dttk_value);
4340 		regs[rd] = _RW_WRITE_HELD(&r.ri);
4341 		break;
4342 
4343 	case DIF_SUBR_RW_ISWRITER:
4344 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
4345 		    mstate, vstate)) {
4346 			regs[rd] = 0;
4347 			break;
4348 		}
4349 
4350 		r.rw = dtrace_loadptr(tupregs[0].dttk_value);
4351 		regs[rd] = _RW_ISWRITER(&r.ri);
4352 		break;
4353 
4354 #else /* !illumos */
4355 	case DIF_SUBR_MUTEX_OWNED:
4356 		if (!dtrace_canload(tupregs[0].dttk_value,
4357 			sizeof (struct lock_object), mstate, vstate)) {
4358 			regs[rd] = 0;
4359 			break;
4360 		}
4361 		l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
4362 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4363 		regs[rd] = LOCK_CLASS(l.li)->lc_owner(l.li, &lowner);
4364 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4365 		break;
4366 
4367 	case DIF_SUBR_MUTEX_OWNER:
4368 		if (!dtrace_canload(tupregs[0].dttk_value,
4369 			sizeof (struct lock_object), mstate, vstate)) {
4370 			regs[rd] = 0;
4371 			break;
4372 		}
4373 		l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
4374 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4375 		LOCK_CLASS(l.li)->lc_owner(l.li, &lowner);
4376 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4377 		regs[rd] = (uintptr_t)lowner;
4378 		break;
4379 
4380 	case DIF_SUBR_MUTEX_TYPE_ADAPTIVE:
4381 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (struct mtx),
4382 		    mstate, vstate)) {
4383 			regs[rd] = 0;
4384 			break;
4385 		}
4386 		l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
4387 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4388 		regs[rd] = (LOCK_CLASS(l.li)->lc_flags & LC_SLEEPLOCK) != 0;
4389 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4390 		break;
4391 
4392 	case DIF_SUBR_MUTEX_TYPE_SPIN:
4393 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (struct mtx),
4394 		    mstate, vstate)) {
4395 			regs[rd] = 0;
4396 			break;
4397 		}
4398 		l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
4399 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4400 		regs[rd] = (LOCK_CLASS(l.li)->lc_flags & LC_SPINLOCK) != 0;
4401 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4402 		break;
4403 
4404 	case DIF_SUBR_RW_READ_HELD:
4405 	case DIF_SUBR_SX_SHARED_HELD:
4406 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
4407 		    mstate, vstate)) {
4408 			regs[rd] = 0;
4409 			break;
4410 		}
4411 		l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
4412 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4413 		regs[rd] = LOCK_CLASS(l.li)->lc_owner(l.li, &lowner) &&
4414 		    lowner == NULL;
4415 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4416 		break;
4417 
4418 	case DIF_SUBR_RW_WRITE_HELD:
4419 	case DIF_SUBR_SX_EXCLUSIVE_HELD:
4420 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
4421 		    mstate, vstate)) {
4422 			regs[rd] = 0;
4423 			break;
4424 		}
4425 		l.lx = dtrace_loadptr(tupregs[0].dttk_value);
4426 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4427 		regs[rd] = LOCK_CLASS(l.li)->lc_owner(l.li, &lowner) &&
4428 		    lowner != NULL;
4429 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4430 		break;
4431 
4432 	case DIF_SUBR_RW_ISWRITER:
4433 	case DIF_SUBR_SX_ISEXCLUSIVE:
4434 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
4435 		    mstate, vstate)) {
4436 			regs[rd] = 0;
4437 			break;
4438 		}
4439 		l.lx = dtrace_loadptr(tupregs[0].dttk_value);
4440 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4441 		LOCK_CLASS(l.li)->lc_owner(l.li, &lowner);
4442 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4443 		regs[rd] = (lowner == curthread);
4444 		break;
4445 #endif /* illumos */
4446 
4447 	case DIF_SUBR_BCOPY: {
4448 		/*
4449 		 * We need to be sure that the destination is in the scratch
4450 		 * region -- no other region is allowed.
4451 		 */
4452 		uintptr_t src = tupregs[0].dttk_value;
4453 		uintptr_t dest = tupregs[1].dttk_value;
4454 		size_t size = tupregs[2].dttk_value;
4455 
4456 		if (!dtrace_inscratch(dest, size, mstate)) {
4457 			*flags |= CPU_DTRACE_BADADDR;
4458 			*illval = regs[rd];
4459 			break;
4460 		}
4461 
4462 		if (!dtrace_canload(src, size, mstate, vstate)) {
4463 			regs[rd] = 0;
4464 			break;
4465 		}
4466 
4467 		dtrace_bcopy((void *)src, (void *)dest, size);
4468 		break;
4469 	}
4470 
4471 	case DIF_SUBR_ALLOCA:
4472 	case DIF_SUBR_COPYIN: {
4473 		uintptr_t dest = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
4474 		uint64_t size =
4475 		    tupregs[subr == DIF_SUBR_ALLOCA ? 0 : 1].dttk_value;
4476 		size_t scratch_size = (dest - mstate->dtms_scratch_ptr) + size;
4477 
4478 		/*
4479 		 * This action doesn't require any credential checks since
4480 		 * probes will not activate in user contexts to which the
4481 		 * enabling user does not have permissions.
4482 		 */
4483 
4484 		/*
4485 		 * Rounding up the user allocation size could have overflowed
4486 		 * a large, bogus allocation (like -1ULL) to 0.
4487 		 */
4488 		if (scratch_size < size ||
4489 		    !DTRACE_INSCRATCH(mstate, scratch_size)) {
4490 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4491 			regs[rd] = 0;
4492 			break;
4493 		}
4494 
4495 		if (subr == DIF_SUBR_COPYIN) {
4496 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4497 			dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
4498 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4499 		}
4500 
4501 		mstate->dtms_scratch_ptr += scratch_size;
4502 		regs[rd] = dest;
4503 		break;
4504 	}
4505 
4506 	case DIF_SUBR_COPYINTO: {
4507 		uint64_t size = tupregs[1].dttk_value;
4508 		uintptr_t dest = tupregs[2].dttk_value;
4509 
4510 		/*
4511 		 * This action doesn't require any credential checks since
4512 		 * probes will not activate in user contexts to which the
4513 		 * enabling user does not have permissions.
4514 		 */
4515 		if (!dtrace_inscratch(dest, size, mstate)) {
4516 			*flags |= CPU_DTRACE_BADADDR;
4517 			*illval = regs[rd];
4518 			break;
4519 		}
4520 
4521 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4522 		dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
4523 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4524 		break;
4525 	}
4526 
4527 	case DIF_SUBR_COPYINSTR: {
4528 		uintptr_t dest = mstate->dtms_scratch_ptr;
4529 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4530 
4531 		if (nargs > 1 && tupregs[1].dttk_value < size)
4532 			size = tupregs[1].dttk_value + 1;
4533 
4534 		/*
4535 		 * This action doesn't require any credential checks since
4536 		 * probes will not activate in user contexts to which the
4537 		 * enabling user does not have permissions.
4538 		 */
4539 		if (!DTRACE_INSCRATCH(mstate, size)) {
4540 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4541 			regs[rd] = 0;
4542 			break;
4543 		}
4544 
4545 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4546 		dtrace_copyinstr(tupregs[0].dttk_value, dest, size, flags);
4547 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4548 
4549 		((char *)dest)[size - 1] = '\0';
4550 		mstate->dtms_scratch_ptr += size;
4551 		regs[rd] = dest;
4552 		break;
4553 	}
4554 
4555 #ifdef illumos
4556 	case DIF_SUBR_MSGSIZE:
4557 	case DIF_SUBR_MSGDSIZE: {
4558 		uintptr_t baddr = tupregs[0].dttk_value, daddr;
4559 		uintptr_t wptr, rptr;
4560 		size_t count = 0;
4561 		int cont = 0;
4562 
4563 		while (baddr != 0 && !(*flags & CPU_DTRACE_FAULT)) {
4564 
4565 			if (!dtrace_canload(baddr, sizeof (mblk_t), mstate,
4566 			    vstate)) {
4567 				regs[rd] = 0;
4568 				break;
4569 			}
4570 
4571 			wptr = dtrace_loadptr(baddr +
4572 			    offsetof(mblk_t, b_wptr));
4573 
4574 			rptr = dtrace_loadptr(baddr +
4575 			    offsetof(mblk_t, b_rptr));
4576 
4577 			if (wptr < rptr) {
4578 				*flags |= CPU_DTRACE_BADADDR;
4579 				*illval = tupregs[0].dttk_value;
4580 				break;
4581 			}
4582 
4583 			daddr = dtrace_loadptr(baddr +
4584 			    offsetof(mblk_t, b_datap));
4585 
4586 			baddr = dtrace_loadptr(baddr +
4587 			    offsetof(mblk_t, b_cont));
4588 
4589 			/*
4590 			 * We want to prevent against denial-of-service here,
4591 			 * so we're only going to search the list for
4592 			 * dtrace_msgdsize_max mblks.
4593 			 */
4594 			if (cont++ > dtrace_msgdsize_max) {
4595 				*flags |= CPU_DTRACE_ILLOP;
4596 				break;
4597 			}
4598 
4599 			if (subr == DIF_SUBR_MSGDSIZE) {
4600 				if (dtrace_load8(daddr +
4601 				    offsetof(dblk_t, db_type)) != M_DATA)
4602 					continue;
4603 			}
4604 
4605 			count += wptr - rptr;
4606 		}
4607 
4608 		if (!(*flags & CPU_DTRACE_FAULT))
4609 			regs[rd] = count;
4610 
4611 		break;
4612 	}
4613 #endif
4614 
4615 	case DIF_SUBR_PROGENYOF: {
4616 		pid_t pid = tupregs[0].dttk_value;
4617 		proc_t *p;
4618 		int rval = 0;
4619 
4620 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4621 
4622 		for (p = curthread->t_procp; p != NULL; p = p->p_parent) {
4623 #ifdef illumos
4624 			if (p->p_pidp->pid_id == pid) {
4625 #else
4626 			if (p->p_pid == pid) {
4627 #endif
4628 				rval = 1;
4629 				break;
4630 			}
4631 		}
4632 
4633 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4634 
4635 		regs[rd] = rval;
4636 		break;
4637 	}
4638 
4639 	case DIF_SUBR_SPECULATION:
4640 		regs[rd] = dtrace_speculation(state);
4641 		break;
4642 
4643 	case DIF_SUBR_COPYOUT: {
4644 		uintptr_t kaddr = tupregs[0].dttk_value;
4645 		uintptr_t uaddr = tupregs[1].dttk_value;
4646 		uint64_t size = tupregs[2].dttk_value;
4647 
4648 		if (!dtrace_destructive_disallow &&
4649 		    dtrace_priv_proc_control(state) &&
4650 		    !dtrace_istoxic(kaddr, size) &&
4651 		    dtrace_canload(kaddr, size, mstate, vstate)) {
4652 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4653 			dtrace_copyout(kaddr, uaddr, size, flags);
4654 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4655 		}
4656 		break;
4657 	}
4658 
4659 	case DIF_SUBR_COPYOUTSTR: {
4660 		uintptr_t kaddr = tupregs[0].dttk_value;
4661 		uintptr_t uaddr = tupregs[1].dttk_value;
4662 		uint64_t size = tupregs[2].dttk_value;
4663 		size_t lim;
4664 
4665 		if (!dtrace_destructive_disallow &&
4666 		    dtrace_priv_proc_control(state) &&
4667 		    !dtrace_istoxic(kaddr, size) &&
4668 		    dtrace_strcanload(kaddr, size, &lim, mstate, vstate)) {
4669 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4670 			dtrace_copyoutstr(kaddr, uaddr, lim, flags);
4671 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4672 		}
4673 		break;
4674 	}
4675 
4676 	case DIF_SUBR_STRLEN: {
4677 		size_t size = state->dts_options[DTRACEOPT_STRSIZE];
4678 		uintptr_t addr = (uintptr_t)tupregs[0].dttk_value;
4679 		size_t lim;
4680 
4681 		if (!dtrace_strcanload(addr, size, &lim, mstate, vstate)) {
4682 			regs[rd] = 0;
4683 			break;
4684 		}
4685 
4686 		regs[rd] = dtrace_strlen((char *)addr, lim);
4687 		break;
4688 	}
4689 
4690 	case DIF_SUBR_STRCHR:
4691 	case DIF_SUBR_STRRCHR: {
4692 		/*
4693 		 * We're going to iterate over the string looking for the
4694 		 * specified character.  We will iterate until we have reached
4695 		 * the string length or we have found the character.  If this
4696 		 * is DIF_SUBR_STRRCHR, we will look for the last occurrence
4697 		 * of the specified character instead of the first.
4698 		 */
4699 		uintptr_t addr = tupregs[0].dttk_value;
4700 		uintptr_t addr_limit;
4701 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4702 		size_t lim;
4703 		char c, target = (char)tupregs[1].dttk_value;
4704 
4705 		if (!dtrace_strcanload(addr, size, &lim, mstate, vstate)) {
4706 			regs[rd] = 0;
4707 			break;
4708 		}
4709 		addr_limit = addr + lim;
4710 
4711 		for (regs[rd] = 0; addr < addr_limit; addr++) {
4712 			if ((c = dtrace_load8(addr)) == target) {
4713 				regs[rd] = addr;
4714 
4715 				if (subr == DIF_SUBR_STRCHR)
4716 					break;
4717 			}
4718 
4719 			if (c == '\0')
4720 				break;
4721 		}
4722 		break;
4723 	}
4724 
4725 	case DIF_SUBR_STRSTR:
4726 	case DIF_SUBR_INDEX:
4727 	case DIF_SUBR_RINDEX: {
4728 		/*
4729 		 * We're going to iterate over the string looking for the
4730 		 * specified string.  We will iterate until we have reached
4731 		 * the string length or we have found the string.  (Yes, this
4732 		 * is done in the most naive way possible -- but considering
4733 		 * that the string we're searching for is likely to be
4734 		 * relatively short, the complexity of Rabin-Karp or similar
4735 		 * hardly seems merited.)
4736 		 */
4737 		char *addr = (char *)(uintptr_t)tupregs[0].dttk_value;
4738 		char *substr = (char *)(uintptr_t)tupregs[1].dttk_value;
4739 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4740 		size_t len = dtrace_strlen(addr, size);
4741 		size_t sublen = dtrace_strlen(substr, size);
4742 		char *limit = addr + len, *orig = addr;
4743 		int notfound = subr == DIF_SUBR_STRSTR ? 0 : -1;
4744 		int inc = 1;
4745 
4746 		regs[rd] = notfound;
4747 
4748 		if (!dtrace_canload((uintptr_t)addr, len + 1, mstate, vstate)) {
4749 			regs[rd] = 0;
4750 			break;
4751 		}
4752 
4753 		if (!dtrace_canload((uintptr_t)substr, sublen + 1, mstate,
4754 		    vstate)) {
4755 			regs[rd] = 0;
4756 			break;
4757 		}
4758 
4759 		/*
4760 		 * strstr() and index()/rindex() have similar semantics if
4761 		 * both strings are the empty string: strstr() returns a
4762 		 * pointer to the (empty) string, and index() and rindex()
4763 		 * both return index 0 (regardless of any position argument).
4764 		 */
4765 		if (sublen == 0 && len == 0) {
4766 			if (subr == DIF_SUBR_STRSTR)
4767 				regs[rd] = (uintptr_t)addr;
4768 			else
4769 				regs[rd] = 0;
4770 			break;
4771 		}
4772 
4773 		if (subr != DIF_SUBR_STRSTR) {
4774 			if (subr == DIF_SUBR_RINDEX) {
4775 				limit = orig - 1;
4776 				addr += len;
4777 				inc = -1;
4778 			}
4779 
4780 			/*
4781 			 * Both index() and rindex() take an optional position
4782 			 * argument that denotes the starting position.
4783 			 */
4784 			if (nargs == 3) {
4785 				int64_t pos = (int64_t)tupregs[2].dttk_value;
4786 
4787 				/*
4788 				 * If the position argument to index() is
4789 				 * negative, Perl implicitly clamps it at
4790 				 * zero.  This semantic is a little surprising
4791 				 * given the special meaning of negative
4792 				 * positions to similar Perl functions like
4793 				 * substr(), but it appears to reflect a
4794 				 * notion that index() can start from a
4795 				 * negative index and increment its way up to
4796 				 * the string.  Given this notion, Perl's
4797 				 * rindex() is at least self-consistent in
4798 				 * that it implicitly clamps positions greater
4799 				 * than the string length to be the string
4800 				 * length.  Where Perl completely loses
4801 				 * coherence, however, is when the specified
4802 				 * substring is the empty string ("").  In
4803 				 * this case, even if the position is
4804 				 * negative, rindex() returns 0 -- and even if
4805 				 * the position is greater than the length,
4806 				 * index() returns the string length.  These
4807 				 * semantics violate the notion that index()
4808 				 * should never return a value less than the
4809 				 * specified position and that rindex() should
4810 				 * never return a value greater than the
4811 				 * specified position.  (One assumes that
4812 				 * these semantics are artifacts of Perl's
4813 				 * implementation and not the results of
4814 				 * deliberate design -- it beggars belief that
4815 				 * even Larry Wall could desire such oddness.)
4816 				 * While in the abstract one would wish for
4817 				 * consistent position semantics across
4818 				 * substr(), index() and rindex() -- or at the
4819 				 * very least self-consistent position
4820 				 * semantics for index() and rindex() -- we
4821 				 * instead opt to keep with the extant Perl
4822 				 * semantics, in all their broken glory.  (Do
4823 				 * we have more desire to maintain Perl's
4824 				 * semantics than Perl does?  Probably.)
4825 				 */
4826 				if (subr == DIF_SUBR_RINDEX) {
4827 					if (pos < 0) {
4828 						if (sublen == 0)
4829 							regs[rd] = 0;
4830 						break;
4831 					}
4832 
4833 					if (pos > len)
4834 						pos = len;
4835 				} else {
4836 					if (pos < 0)
4837 						pos = 0;
4838 
4839 					if (pos >= len) {
4840 						if (sublen == 0)
4841 							regs[rd] = len;
4842 						break;
4843 					}
4844 				}
4845 
4846 				addr = orig + pos;
4847 			}
4848 		}
4849 
4850 		for (regs[rd] = notfound; addr != limit; addr += inc) {
4851 			if (dtrace_strncmp(addr, substr, sublen) == 0) {
4852 				if (subr != DIF_SUBR_STRSTR) {
4853 					/*
4854 					 * As D index() and rindex() are
4855 					 * modeled on Perl (and not on awk),
4856 					 * we return a zero-based (and not a
4857 					 * one-based) index.  (For you Perl
4858 					 * weenies: no, we're not going to add
4859 					 * $[ -- and shouldn't you be at a con
4860 					 * or something?)
4861 					 */
4862 					regs[rd] = (uintptr_t)(addr - orig);
4863 					break;
4864 				}
4865 
4866 				ASSERT(subr == DIF_SUBR_STRSTR);
4867 				regs[rd] = (uintptr_t)addr;
4868 				break;
4869 			}
4870 		}
4871 
4872 		break;
4873 	}
4874 
4875 	case DIF_SUBR_STRTOK: {
4876 		uintptr_t addr = tupregs[0].dttk_value;
4877 		uintptr_t tokaddr = tupregs[1].dttk_value;
4878 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4879 		uintptr_t limit, toklimit;
4880 		size_t clim;
4881 		uint8_t c = 0, tokmap[32];	 /* 256 / 8 */
4882 		char *dest = (char *)mstate->dtms_scratch_ptr;
4883 		int i;
4884 
4885 		/*
4886 		 * Check both the token buffer and (later) the input buffer,
4887 		 * since both could be non-scratch addresses.
4888 		 */
4889 		if (!dtrace_strcanload(tokaddr, size, &clim, mstate, vstate)) {
4890 			regs[rd] = 0;
4891 			break;
4892 		}
4893 		toklimit = tokaddr + clim;
4894 
4895 		if (!DTRACE_INSCRATCH(mstate, size)) {
4896 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4897 			regs[rd] = 0;
4898 			break;
4899 		}
4900 
4901 		if (addr == 0) {
4902 			/*
4903 			 * If the address specified is NULL, we use our saved
4904 			 * strtok pointer from the mstate.  Note that this
4905 			 * means that the saved strtok pointer is _only_
4906 			 * valid within multiple enablings of the same probe --
4907 			 * it behaves like an implicit clause-local variable.
4908 			 */
4909 			addr = mstate->dtms_strtok;
4910 			limit = mstate->dtms_strtok_limit;
4911 		} else {
4912 			/*
4913 			 * If the user-specified address is non-NULL we must
4914 			 * access check it.  This is the only time we have
4915 			 * a chance to do so, since this address may reside
4916 			 * in the string table of this clause-- future calls
4917 			 * (when we fetch addr from mstate->dtms_strtok)
4918 			 * would fail this access check.
4919 			 */
4920 			if (!dtrace_strcanload(addr, size, &clim, mstate,
4921 			    vstate)) {
4922 				regs[rd] = 0;
4923 				break;
4924 			}
4925 			limit = addr + clim;
4926 		}
4927 
4928 		/*
4929 		 * First, zero the token map, and then process the token
4930 		 * string -- setting a bit in the map for every character
4931 		 * found in the token string.
4932 		 */
4933 		for (i = 0; i < sizeof (tokmap); i++)
4934 			tokmap[i] = 0;
4935 
4936 		for (; tokaddr < toklimit; tokaddr++) {
4937 			if ((c = dtrace_load8(tokaddr)) == '\0')
4938 				break;
4939 
4940 			ASSERT((c >> 3) < sizeof (tokmap));
4941 			tokmap[c >> 3] |= (1 << (c & 0x7));
4942 		}
4943 
4944 		for (; addr < limit; addr++) {
4945 			/*
4946 			 * We're looking for a character that is _not_
4947 			 * contained in the token string.
4948 			 */
4949 			if ((c = dtrace_load8(addr)) == '\0')
4950 				break;
4951 
4952 			if (!(tokmap[c >> 3] & (1 << (c & 0x7))))
4953 				break;
4954 		}
4955 
4956 		if (c == '\0') {
4957 			/*
4958 			 * We reached the end of the string without finding
4959 			 * any character that was not in the token string.
4960 			 * We return NULL in this case, and we set the saved
4961 			 * address to NULL as well.
4962 			 */
4963 			regs[rd] = 0;
4964 			mstate->dtms_strtok = 0;
4965 			mstate->dtms_strtok_limit = 0;
4966 			break;
4967 		}
4968 
4969 		/*
4970 		 * From here on, we're copying into the destination string.
4971 		 */
4972 		for (i = 0; addr < limit && i < size - 1; addr++) {
4973 			if ((c = dtrace_load8(addr)) == '\0')
4974 				break;
4975 
4976 			if (tokmap[c >> 3] & (1 << (c & 0x7)))
4977 				break;
4978 
4979 			ASSERT(i < size);
4980 			dest[i++] = c;
4981 		}
4982 
4983 		ASSERT(i < size);
4984 		dest[i] = '\0';
4985 		regs[rd] = (uintptr_t)dest;
4986 		mstate->dtms_scratch_ptr += size;
4987 		mstate->dtms_strtok = addr;
4988 		mstate->dtms_strtok_limit = limit;
4989 		break;
4990 	}
4991 
4992 	case DIF_SUBR_SUBSTR: {
4993 		uintptr_t s = tupregs[0].dttk_value;
4994 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4995 		char *d = (char *)mstate->dtms_scratch_ptr;
4996 		int64_t index = (int64_t)tupregs[1].dttk_value;
4997 		int64_t remaining = (int64_t)tupregs[2].dttk_value;
4998 		size_t len = dtrace_strlen((char *)s, size);
4999 		int64_t i;
5000 
5001 		if (!dtrace_canload(s, len + 1, mstate, vstate)) {
5002 			regs[rd] = 0;
5003 			break;
5004 		}
5005 
5006 		if (!DTRACE_INSCRATCH(mstate, size)) {
5007 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5008 			regs[rd] = 0;
5009 			break;
5010 		}
5011 
5012 		if (nargs <= 2)
5013 			remaining = (int64_t)size;
5014 
5015 		if (index < 0) {
5016 			index += len;
5017 
5018 			if (index < 0 && index + remaining > 0) {
5019 				remaining += index;
5020 				index = 0;
5021 			}
5022 		}
5023 
5024 		if (index >= len || index < 0) {
5025 			remaining = 0;
5026 		} else if (remaining < 0) {
5027 			remaining += len - index;
5028 		} else if (index + remaining > size) {
5029 			remaining = size - index;
5030 		}
5031 
5032 		for (i = 0; i < remaining; i++) {
5033 			if ((d[i] = dtrace_load8(s + index + i)) == '\0')
5034 				break;
5035 		}
5036 
5037 		d[i] = '\0';
5038 
5039 		mstate->dtms_scratch_ptr += size;
5040 		regs[rd] = (uintptr_t)d;
5041 		break;
5042 	}
5043 
5044 	case DIF_SUBR_JSON: {
5045 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
5046 		uintptr_t json = tupregs[0].dttk_value;
5047 		size_t jsonlen = dtrace_strlen((char *)json, size);
5048 		uintptr_t elem = tupregs[1].dttk_value;
5049 		size_t elemlen = dtrace_strlen((char *)elem, size);
5050 
5051 		char *dest = (char *)mstate->dtms_scratch_ptr;
5052 		char *elemlist = (char *)mstate->dtms_scratch_ptr + jsonlen + 1;
5053 		char *ee = elemlist;
5054 		int nelems = 1;
5055 		uintptr_t cur;
5056 
5057 		if (!dtrace_canload(json, jsonlen + 1, mstate, vstate) ||
5058 		    !dtrace_canload(elem, elemlen + 1, mstate, vstate)) {
5059 			regs[rd] = 0;
5060 			break;
5061 		}
5062 
5063 		if (!DTRACE_INSCRATCH(mstate, jsonlen + 1 + elemlen + 1)) {
5064 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5065 			regs[rd] = 0;
5066 			break;
5067 		}
5068 
5069 		/*
5070 		 * Read the element selector and split it up into a packed list
5071 		 * of strings.
5072 		 */
5073 		for (cur = elem; cur < elem + elemlen; cur++) {
5074 			char cc = dtrace_load8(cur);
5075 
5076 			if (cur == elem && cc == '[') {
5077 				/*
5078 				 * If the first element selector key is
5079 				 * actually an array index then ignore the
5080 				 * bracket.
5081 				 */
5082 				continue;
5083 			}
5084 
5085 			if (cc == ']')
5086 				continue;
5087 
5088 			if (cc == '.' || cc == '[') {
5089 				nelems++;
5090 				cc = '\0';
5091 			}
5092 
5093 			*ee++ = cc;
5094 		}
5095 		*ee++ = '\0';
5096 
5097 		if ((regs[rd] = (uintptr_t)dtrace_json(size, json, elemlist,
5098 		    nelems, dest)) != 0)
5099 			mstate->dtms_scratch_ptr += jsonlen + 1;
5100 		break;
5101 	}
5102 
5103 	case DIF_SUBR_TOUPPER:
5104 	case DIF_SUBR_TOLOWER: {
5105 		uintptr_t s = tupregs[0].dttk_value;
5106 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
5107 		char *dest = (char *)mstate->dtms_scratch_ptr, c;
5108 		size_t len = dtrace_strlen((char *)s, size);
5109 		char lower, upper, convert;
5110 		int64_t i;
5111 
5112 		if (subr == DIF_SUBR_TOUPPER) {
5113 			lower = 'a';
5114 			upper = 'z';
5115 			convert = 'A';
5116 		} else {
5117 			lower = 'A';
5118 			upper = 'Z';
5119 			convert = 'a';
5120 		}
5121 
5122 		if (!dtrace_canload(s, len + 1, mstate, vstate)) {
5123 			regs[rd] = 0;
5124 			break;
5125 		}
5126 
5127 		if (!DTRACE_INSCRATCH(mstate, size)) {
5128 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5129 			regs[rd] = 0;
5130 			break;
5131 		}
5132 
5133 		for (i = 0; i < size - 1; i++) {
5134 			if ((c = dtrace_load8(s + i)) == '\0')
5135 				break;
5136 
5137 			if (c >= lower && c <= upper)
5138 				c = convert + (c - lower);
5139 
5140 			dest[i] = c;
5141 		}
5142 
5143 		ASSERT(i < size);
5144 		dest[i] = '\0';
5145 		regs[rd] = (uintptr_t)dest;
5146 		mstate->dtms_scratch_ptr += size;
5147 		break;
5148 	}
5149 
5150 #ifdef illumos
5151 	case DIF_SUBR_GETMAJOR:
5152 #ifdef _LP64
5153 		regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR64) & MAXMAJ64;
5154 #else
5155 		regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR) & MAXMAJ;
5156 #endif
5157 		break;
5158 
5159 	case DIF_SUBR_GETMINOR:
5160 #ifdef _LP64
5161 		regs[rd] = tupregs[0].dttk_value & MAXMIN64;
5162 #else
5163 		regs[rd] = tupregs[0].dttk_value & MAXMIN;
5164 #endif
5165 		break;
5166 
5167 	case DIF_SUBR_DDI_PATHNAME: {
5168 		/*
5169 		 * This one is a galactic mess.  We are going to roughly
5170 		 * emulate ddi_pathname(), but it's made more complicated
5171 		 * by the fact that we (a) want to include the minor name and
5172 		 * (b) must proceed iteratively instead of recursively.
5173 		 */
5174 		uintptr_t dest = mstate->dtms_scratch_ptr;
5175 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
5176 		char *start = (char *)dest, *end = start + size - 1;
5177 		uintptr_t daddr = tupregs[0].dttk_value;
5178 		int64_t minor = (int64_t)tupregs[1].dttk_value;
5179 		char *s;
5180 		int i, len, depth = 0;
5181 
5182 		/*
5183 		 * Due to all the pointer jumping we do and context we must
5184 		 * rely upon, we just mandate that the user must have kernel
5185 		 * read privileges to use this routine.
5186 		 */
5187 		if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) == 0) {
5188 			*flags |= CPU_DTRACE_KPRIV;
5189 			*illval = daddr;
5190 			regs[rd] = 0;
5191 		}
5192 
5193 		if (!DTRACE_INSCRATCH(mstate, size)) {
5194 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5195 			regs[rd] = 0;
5196 			break;
5197 		}
5198 
5199 		*end = '\0';
5200 
5201 		/*
5202 		 * We want to have a name for the minor.  In order to do this,
5203 		 * we need to walk the minor list from the devinfo.  We want
5204 		 * to be sure that we don't infinitely walk a circular list,
5205 		 * so we check for circularity by sending a scout pointer
5206 		 * ahead two elements for every element that we iterate over;
5207 		 * if the list is circular, these will ultimately point to the
5208 		 * same element.  You may recognize this little trick as the
5209 		 * answer to a stupid interview question -- one that always
5210 		 * seems to be asked by those who had to have it laboriously
5211 		 * explained to them, and who can't even concisely describe
5212 		 * the conditions under which one would be forced to resort to
5213 		 * this technique.  Needless to say, those conditions are
5214 		 * found here -- and probably only here.  Is this the only use
5215 		 * of this infamous trick in shipping, production code?  If it
5216 		 * isn't, it probably should be...
5217 		 */
5218 		if (minor != -1) {
5219 			uintptr_t maddr = dtrace_loadptr(daddr +
5220 			    offsetof(struct dev_info, devi_minor));
5221 
5222 			uintptr_t next = offsetof(struct ddi_minor_data, next);
5223 			uintptr_t name = offsetof(struct ddi_minor_data,
5224 			    d_minor) + offsetof(struct ddi_minor, name);
5225 			uintptr_t dev = offsetof(struct ddi_minor_data,
5226 			    d_minor) + offsetof(struct ddi_minor, dev);
5227 			uintptr_t scout;
5228 
5229 			if (maddr != NULL)
5230 				scout = dtrace_loadptr(maddr + next);
5231 
5232 			while (maddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
5233 				uint64_t m;
5234 #ifdef _LP64
5235 				m = dtrace_load64(maddr + dev) & MAXMIN64;
5236 #else
5237 				m = dtrace_load32(maddr + dev) & MAXMIN;
5238 #endif
5239 				if (m != minor) {
5240 					maddr = dtrace_loadptr(maddr + next);
5241 
5242 					if (scout == NULL)
5243 						continue;
5244 
5245 					scout = dtrace_loadptr(scout + next);
5246 
5247 					if (scout == NULL)
5248 						continue;
5249 
5250 					scout = dtrace_loadptr(scout + next);
5251 
5252 					if (scout == NULL)
5253 						continue;
5254 
5255 					if (scout == maddr) {
5256 						*flags |= CPU_DTRACE_ILLOP;
5257 						break;
5258 					}
5259 
5260 					continue;
5261 				}
5262 
5263 				/*
5264 				 * We have the minor data.  Now we need to
5265 				 * copy the minor's name into the end of the
5266 				 * pathname.
5267 				 */
5268 				s = (char *)dtrace_loadptr(maddr + name);
5269 				len = dtrace_strlen(s, size);
5270 
5271 				if (*flags & CPU_DTRACE_FAULT)
5272 					break;
5273 
5274 				if (len != 0) {
5275 					if ((end -= (len + 1)) < start)
5276 						break;
5277 
5278 					*end = ':';
5279 				}
5280 
5281 				for (i = 1; i <= len; i++)
5282 					end[i] = dtrace_load8((uintptr_t)s++);
5283 				break;
5284 			}
5285 		}
5286 
5287 		while (daddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
5288 			ddi_node_state_t devi_state;
5289 
5290 			devi_state = dtrace_load32(daddr +
5291 			    offsetof(struct dev_info, devi_node_state));
5292 
5293 			if (*flags & CPU_DTRACE_FAULT)
5294 				break;
5295 
5296 			if (devi_state >= DS_INITIALIZED) {
5297 				s = (char *)dtrace_loadptr(daddr +
5298 				    offsetof(struct dev_info, devi_addr));
5299 				len = dtrace_strlen(s, size);
5300 
5301 				if (*flags & CPU_DTRACE_FAULT)
5302 					break;
5303 
5304 				if (len != 0) {
5305 					if ((end -= (len + 1)) < start)
5306 						break;
5307 
5308 					*end = '@';
5309 				}
5310 
5311 				for (i = 1; i <= len; i++)
5312 					end[i] = dtrace_load8((uintptr_t)s++);
5313 			}
5314 
5315 			/*
5316 			 * Now for the node name...
5317 			 */
5318 			s = (char *)dtrace_loadptr(daddr +
5319 			    offsetof(struct dev_info, devi_node_name));
5320 
5321 			daddr = dtrace_loadptr(daddr +
5322 			    offsetof(struct dev_info, devi_parent));
5323 
5324 			/*
5325 			 * If our parent is NULL (that is, if we're the root
5326 			 * node), we're going to use the special path
5327 			 * "devices".
5328 			 */
5329 			if (daddr == 0)
5330 				s = "devices";
5331 
5332 			len = dtrace_strlen(s, size);
5333 			if (*flags & CPU_DTRACE_FAULT)
5334 				break;
5335 
5336 			if ((end -= (len + 1)) < start)
5337 				break;
5338 
5339 			for (i = 1; i <= len; i++)
5340 				end[i] = dtrace_load8((uintptr_t)s++);
5341 			*end = '/';
5342 
5343 			if (depth++ > dtrace_devdepth_max) {
5344 				*flags |= CPU_DTRACE_ILLOP;
5345 				break;
5346 			}
5347 		}
5348 
5349 		if (end < start)
5350 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5351 
5352 		if (daddr == 0) {
5353 			regs[rd] = (uintptr_t)end;
5354 			mstate->dtms_scratch_ptr += size;
5355 		}
5356 
5357 		break;
5358 	}
5359 #endif
5360 
5361 	case DIF_SUBR_STRJOIN: {
5362 		char *d = (char *)mstate->dtms_scratch_ptr;
5363 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
5364 		uintptr_t s1 = tupregs[0].dttk_value;
5365 		uintptr_t s2 = tupregs[1].dttk_value;
5366 		int i = 0, j = 0;
5367 		size_t lim1, lim2;
5368 		char c;
5369 
5370 		if (!dtrace_strcanload(s1, size, &lim1, mstate, vstate) ||
5371 		    !dtrace_strcanload(s2, size, &lim2, mstate, vstate)) {
5372 			regs[rd] = 0;
5373 			break;
5374 		}
5375 
5376 		if (!DTRACE_INSCRATCH(mstate, size)) {
5377 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5378 			regs[rd] = 0;
5379 			break;
5380 		}
5381 
5382 		for (;;) {
5383 			if (i >= size) {
5384 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5385 				regs[rd] = 0;
5386 				break;
5387 			}
5388 			c = (i >= lim1) ? '\0' : dtrace_load8(s1++);
5389 			if ((d[i++] = c) == '\0') {
5390 				i--;
5391 				break;
5392 			}
5393 		}
5394 
5395 		for (;;) {
5396 			if (i >= size) {
5397 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5398 				regs[rd] = 0;
5399 				break;
5400 			}
5401 
5402 			c = (j++ >= lim2) ? '\0' : dtrace_load8(s2++);
5403 			if ((d[i++] = c) == '\0')
5404 				break;
5405 		}
5406 
5407 		if (i < size) {
5408 			mstate->dtms_scratch_ptr += i;
5409 			regs[rd] = (uintptr_t)d;
5410 		}
5411 
5412 		break;
5413 	}
5414 
5415 	case DIF_SUBR_STRTOLL: {
5416 		uintptr_t s = tupregs[0].dttk_value;
5417 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
5418 		size_t lim;
5419 		int base = 10;
5420 
5421 		if (nargs > 1) {
5422 			if ((base = tupregs[1].dttk_value) <= 1 ||
5423 			    base > ('z' - 'a' + 1) + ('9' - '0' + 1)) {
5424 				*flags |= CPU_DTRACE_ILLOP;
5425 				break;
5426 			}
5427 		}
5428 
5429 		if (!dtrace_strcanload(s, size, &lim, mstate, vstate)) {
5430 			regs[rd] = INT64_MIN;
5431 			break;
5432 		}
5433 
5434 		regs[rd] = dtrace_strtoll((char *)s, base, lim);
5435 		break;
5436 	}
5437 
5438 	case DIF_SUBR_LLTOSTR: {
5439 		int64_t i = (int64_t)tupregs[0].dttk_value;
5440 		uint64_t val, digit;
5441 		uint64_t size = 65;	/* enough room for 2^64 in binary */
5442 		char *end = (char *)mstate->dtms_scratch_ptr + size - 1;
5443 		int base = 10;
5444 
5445 		if (nargs > 1) {
5446 			if ((base = tupregs[1].dttk_value) <= 1 ||
5447 			    base > ('z' - 'a' + 1) + ('9' - '0' + 1)) {
5448 				*flags |= CPU_DTRACE_ILLOP;
5449 				break;
5450 			}
5451 		}
5452 
5453 		val = (base == 10 && i < 0) ? i * -1 : i;
5454 
5455 		if (!DTRACE_INSCRATCH(mstate, size)) {
5456 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5457 			regs[rd] = 0;
5458 			break;
5459 		}
5460 
5461 		for (*end-- = '\0'; val; val /= base) {
5462 			if ((digit = val % base) <= '9' - '0') {
5463 				*end-- = '0' + digit;
5464 			} else {
5465 				*end-- = 'a' + (digit - ('9' - '0') - 1);
5466 			}
5467 		}
5468 
5469 		if (i == 0 && base == 16)
5470 			*end-- = '0';
5471 
5472 		if (base == 16)
5473 			*end-- = 'x';
5474 
5475 		if (i == 0 || base == 8 || base == 16)
5476 			*end-- = '0';
5477 
5478 		if (i < 0 && base == 10)
5479 			*end-- = '-';
5480 
5481 		regs[rd] = (uintptr_t)end + 1;
5482 		mstate->dtms_scratch_ptr += size;
5483 		break;
5484 	}
5485 
5486 	case DIF_SUBR_HTONS:
5487 	case DIF_SUBR_NTOHS:
5488 #if BYTE_ORDER == BIG_ENDIAN
5489 		regs[rd] = (uint16_t)tupregs[0].dttk_value;
5490 #else
5491 		regs[rd] = DT_BSWAP_16((uint16_t)tupregs[0].dttk_value);
5492 #endif
5493 		break;
5494 
5495 
5496 	case DIF_SUBR_HTONL:
5497 	case DIF_SUBR_NTOHL:
5498 #if BYTE_ORDER == BIG_ENDIAN
5499 		regs[rd] = (uint32_t)tupregs[0].dttk_value;
5500 #else
5501 		regs[rd] = DT_BSWAP_32((uint32_t)tupregs[0].dttk_value);
5502 #endif
5503 		break;
5504 
5505 
5506 	case DIF_SUBR_HTONLL:
5507 	case DIF_SUBR_NTOHLL:
5508 #if BYTE_ORDER == BIG_ENDIAN
5509 		regs[rd] = (uint64_t)tupregs[0].dttk_value;
5510 #else
5511 		regs[rd] = DT_BSWAP_64((uint64_t)tupregs[0].dttk_value);
5512 #endif
5513 		break;
5514 
5515 
5516 	case DIF_SUBR_DIRNAME:
5517 	case DIF_SUBR_BASENAME: {
5518 		char *dest = (char *)mstate->dtms_scratch_ptr;
5519 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
5520 		uintptr_t src = tupregs[0].dttk_value;
5521 		int i, j, len = dtrace_strlen((char *)src, size);
5522 		int lastbase = -1, firstbase = -1, lastdir = -1;
5523 		int start, end;
5524 
5525 		if (!dtrace_canload(src, len + 1, mstate, vstate)) {
5526 			regs[rd] = 0;
5527 			break;
5528 		}
5529 
5530 		if (!DTRACE_INSCRATCH(mstate, size)) {
5531 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5532 			regs[rd] = 0;
5533 			break;
5534 		}
5535 
5536 		/*
5537 		 * The basename and dirname for a zero-length string is
5538 		 * defined to be "."
5539 		 */
5540 		if (len == 0) {
5541 			len = 1;
5542 			src = (uintptr_t)".";
5543 		}
5544 
5545 		/*
5546 		 * Start from the back of the string, moving back toward the
5547 		 * front until we see a character that isn't a slash.  That
5548 		 * character is the last character in the basename.
5549 		 */
5550 		for (i = len - 1; i >= 0; i--) {
5551 			if (dtrace_load8(src + i) != '/')
5552 				break;
5553 		}
5554 
5555 		if (i >= 0)
5556 			lastbase = i;
5557 
5558 		/*
5559 		 * Starting from the last character in the basename, move
5560 		 * towards the front until we find a slash.  The character
5561 		 * that we processed immediately before that is the first
5562 		 * character in the basename.
5563 		 */
5564 		for (; i >= 0; i--) {
5565 			if (dtrace_load8(src + i) == '/')
5566 				break;
5567 		}
5568 
5569 		if (i >= 0)
5570 			firstbase = i + 1;
5571 
5572 		/*
5573 		 * Now keep going until we find a non-slash character.  That
5574 		 * character is the last character in the dirname.
5575 		 */
5576 		for (; i >= 0; i--) {
5577 			if (dtrace_load8(src + i) != '/')
5578 				break;
5579 		}
5580 
5581 		if (i >= 0)
5582 			lastdir = i;
5583 
5584 		ASSERT(!(lastbase == -1 && firstbase != -1));
5585 		ASSERT(!(firstbase == -1 && lastdir != -1));
5586 
5587 		if (lastbase == -1) {
5588 			/*
5589 			 * We didn't find a non-slash character.  We know that
5590 			 * the length is non-zero, so the whole string must be
5591 			 * slashes.  In either the dirname or the basename
5592 			 * case, we return '/'.
5593 			 */
5594 			ASSERT(firstbase == -1);
5595 			firstbase = lastbase = lastdir = 0;
5596 		}
5597 
5598 		if (firstbase == -1) {
5599 			/*
5600 			 * The entire string consists only of a basename
5601 			 * component.  If we're looking for dirname, we need
5602 			 * to change our string to be just "."; if we're
5603 			 * looking for a basename, we'll just set the first
5604 			 * character of the basename to be 0.
5605 			 */
5606 			if (subr == DIF_SUBR_DIRNAME) {
5607 				ASSERT(lastdir == -1);
5608 				src = (uintptr_t)".";
5609 				lastdir = 0;
5610 			} else {
5611 				firstbase = 0;
5612 			}
5613 		}
5614 
5615 		if (subr == DIF_SUBR_DIRNAME) {
5616 			if (lastdir == -1) {
5617 				/*
5618 				 * We know that we have a slash in the name --
5619 				 * or lastdir would be set to 0, above.  And
5620 				 * because lastdir is -1, we know that this
5621 				 * slash must be the first character.  (That
5622 				 * is, the full string must be of the form
5623 				 * "/basename".)  In this case, the last
5624 				 * character of the directory name is 0.
5625 				 */
5626 				lastdir = 0;
5627 			}
5628 
5629 			start = 0;
5630 			end = lastdir;
5631 		} else {
5632 			ASSERT(subr == DIF_SUBR_BASENAME);
5633 			ASSERT(firstbase != -1 && lastbase != -1);
5634 			start = firstbase;
5635 			end = lastbase;
5636 		}
5637 
5638 		for (i = start, j = 0; i <= end && j < size - 1; i++, j++)
5639 			dest[j] = dtrace_load8(src + i);
5640 
5641 		dest[j] = '\0';
5642 		regs[rd] = (uintptr_t)dest;
5643 		mstate->dtms_scratch_ptr += size;
5644 		break;
5645 	}
5646 
5647 	case DIF_SUBR_GETF: {
5648 		uintptr_t fd = tupregs[0].dttk_value;
5649 		struct filedesc *fdp;
5650 		file_t *fp;
5651 
5652 		if (!dtrace_priv_proc(state)) {
5653 			regs[rd] = 0;
5654 			break;
5655 		}
5656 		fdp = curproc->p_fd;
5657 		FILEDESC_SLOCK(fdp);
5658 		/*
5659 		 * XXXMJG this looks broken as no ref is taken.
5660 		 */
5661 		fp = fget_noref(fdp, fd);
5662 		mstate->dtms_getf = fp;
5663 		regs[rd] = (uintptr_t)fp;
5664 		FILEDESC_SUNLOCK(fdp);
5665 		break;
5666 	}
5667 
5668 	case DIF_SUBR_CLEANPATH: {
5669 		char *dest = (char *)mstate->dtms_scratch_ptr, c;
5670 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
5671 		uintptr_t src = tupregs[0].dttk_value;
5672 		size_t lim;
5673 		int i = 0, j = 0;
5674 #ifdef illumos
5675 		zone_t *z;
5676 #endif
5677 
5678 		if (!dtrace_strcanload(src, size, &lim, mstate, vstate)) {
5679 			regs[rd] = 0;
5680 			break;
5681 		}
5682 
5683 		if (!DTRACE_INSCRATCH(mstate, size)) {
5684 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5685 			regs[rd] = 0;
5686 			break;
5687 		}
5688 
5689 		/*
5690 		 * Move forward, loading each character.
5691 		 */
5692 		do {
5693 			c = (i >= lim) ? '\0' : dtrace_load8(src + i++);
5694 next:
5695 			if (j + 5 >= size)	/* 5 = strlen("/..c\0") */
5696 				break;
5697 
5698 			if (c != '/') {
5699 				dest[j++] = c;
5700 				continue;
5701 			}
5702 
5703 			c = (i >= lim) ? '\0' : dtrace_load8(src + i++);
5704 
5705 			if (c == '/') {
5706 				/*
5707 				 * We have two slashes -- we can just advance
5708 				 * to the next character.
5709 				 */
5710 				goto next;
5711 			}
5712 
5713 			if (c != '.') {
5714 				/*
5715 				 * This is not "." and it's not ".." -- we can
5716 				 * just store the "/" and this character and
5717 				 * drive on.
5718 				 */
5719 				dest[j++] = '/';
5720 				dest[j++] = c;
5721 				continue;
5722 			}
5723 
5724 			c = (i >= lim) ? '\0' : dtrace_load8(src + i++);
5725 
5726 			if (c == '/') {
5727 				/*
5728 				 * This is a "/./" component.  We're not going
5729 				 * to store anything in the destination buffer;
5730 				 * we're just going to go to the next component.
5731 				 */
5732 				goto next;
5733 			}
5734 
5735 			if (c != '.') {
5736 				/*
5737 				 * This is not ".." -- we can just store the
5738 				 * "/." and this character and continue
5739 				 * processing.
5740 				 */
5741 				dest[j++] = '/';
5742 				dest[j++] = '.';
5743 				dest[j++] = c;
5744 				continue;
5745 			}
5746 
5747 			c = (i >= lim) ? '\0' : dtrace_load8(src + i++);
5748 
5749 			if (c != '/' && c != '\0') {
5750 				/*
5751 				 * This is not ".." -- it's "..[mumble]".
5752 				 * We'll store the "/.." and this character
5753 				 * and continue processing.
5754 				 */
5755 				dest[j++] = '/';
5756 				dest[j++] = '.';
5757 				dest[j++] = '.';
5758 				dest[j++] = c;
5759 				continue;
5760 			}
5761 
5762 			/*
5763 			 * This is "/../" or "/..\0".  We need to back up
5764 			 * our destination pointer until we find a "/".
5765 			 */
5766 			i--;
5767 			while (j != 0 && dest[--j] != '/')
5768 				continue;
5769 
5770 			if (c == '\0')
5771 				dest[++j] = '/';
5772 		} while (c != '\0');
5773 
5774 		dest[j] = '\0';
5775 
5776 #ifdef illumos
5777 		if (mstate->dtms_getf != NULL &&
5778 		    !(mstate->dtms_access & DTRACE_ACCESS_KERNEL) &&
5779 		    (z = state->dts_cred.dcr_cred->cr_zone) != kcred->cr_zone) {
5780 			/*
5781 			 * If we've done a getf() as a part of this ECB and we
5782 			 * don't have kernel access (and we're not in the global
5783 			 * zone), check if the path we cleaned up begins with
5784 			 * the zone's root path, and trim it off if so.  Note
5785 			 * that this is an output cleanliness issue, not a
5786 			 * security issue: knowing one's zone root path does
5787 			 * not enable privilege escalation.
5788 			 */
5789 			if (strstr(dest, z->zone_rootpath) == dest)
5790 				dest += strlen(z->zone_rootpath) - 1;
5791 		}
5792 #endif
5793 
5794 		regs[rd] = (uintptr_t)dest;
5795 		mstate->dtms_scratch_ptr += size;
5796 		break;
5797 	}
5798 
5799 	case DIF_SUBR_INET_NTOA:
5800 	case DIF_SUBR_INET_NTOA6:
5801 	case DIF_SUBR_INET_NTOP: {
5802 		size_t size;
5803 		int af, argi, i;
5804 		char *base, *end;
5805 
5806 		if (subr == DIF_SUBR_INET_NTOP) {
5807 			af = (int)tupregs[0].dttk_value;
5808 			argi = 1;
5809 		} else {
5810 			af = subr == DIF_SUBR_INET_NTOA ? AF_INET: AF_INET6;
5811 			argi = 0;
5812 		}
5813 
5814 		if (af == AF_INET) {
5815 			ipaddr_t ip4;
5816 			uint8_t *ptr8, val;
5817 
5818 			if (!dtrace_canload(tupregs[argi].dttk_value,
5819 			    sizeof (ipaddr_t), mstate, vstate)) {
5820 				regs[rd] = 0;
5821 				break;
5822 			}
5823 
5824 			/*
5825 			 * Safely load the IPv4 address.
5826 			 */
5827 			ip4 = dtrace_load32(tupregs[argi].dttk_value);
5828 
5829 			/*
5830 			 * Check an IPv4 string will fit in scratch.
5831 			 */
5832 			size = INET_ADDRSTRLEN;
5833 			if (!DTRACE_INSCRATCH(mstate, size)) {
5834 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5835 				regs[rd] = 0;
5836 				break;
5837 			}
5838 			base = (char *)mstate->dtms_scratch_ptr;
5839 			end = (char *)mstate->dtms_scratch_ptr + size - 1;
5840 
5841 			/*
5842 			 * Stringify as a dotted decimal quad.
5843 			 */
5844 			*end-- = '\0';
5845 			ptr8 = (uint8_t *)&ip4;
5846 			for (i = 3; i >= 0; i--) {
5847 				val = ptr8[i];
5848 
5849 				if (val == 0) {
5850 					*end-- = '0';
5851 				} else {
5852 					for (; val; val /= 10) {
5853 						*end-- = '0' + (val % 10);
5854 					}
5855 				}
5856 
5857 				if (i > 0)
5858 					*end-- = '.';
5859 			}
5860 			ASSERT(end + 1 >= base);
5861 
5862 		} else if (af == AF_INET6) {
5863 			struct in6_addr ip6;
5864 			int firstzero, tryzero, numzero, v6end;
5865 			uint16_t val;
5866 			const char digits[] = "0123456789abcdef";
5867 
5868 			/*
5869 			 * Stringify using RFC 1884 convention 2 - 16 bit
5870 			 * hexadecimal values with a zero-run compression.
5871 			 * Lower case hexadecimal digits are used.
5872 			 * 	eg, fe80::214:4fff:fe0b:76c8.
5873 			 * The IPv4 embedded form is returned for inet_ntop,
5874 			 * just the IPv4 string is returned for inet_ntoa6.
5875 			 */
5876 
5877 			if (!dtrace_canload(tupregs[argi].dttk_value,
5878 			    sizeof (struct in6_addr), mstate, vstate)) {
5879 				regs[rd] = 0;
5880 				break;
5881 			}
5882 
5883 			/*
5884 			 * Safely load the IPv6 address.
5885 			 */
5886 			dtrace_bcopy(
5887 			    (void *)(uintptr_t)tupregs[argi].dttk_value,
5888 			    (void *)(uintptr_t)&ip6, sizeof (struct in6_addr));
5889 
5890 			/*
5891 			 * Check an IPv6 string will fit in scratch.
5892 			 */
5893 			size = INET6_ADDRSTRLEN;
5894 			if (!DTRACE_INSCRATCH(mstate, size)) {
5895 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5896 				regs[rd] = 0;
5897 				break;
5898 			}
5899 			base = (char *)mstate->dtms_scratch_ptr;
5900 			end = (char *)mstate->dtms_scratch_ptr + size - 1;
5901 			*end-- = '\0';
5902 
5903 			/*
5904 			 * Find the longest run of 16 bit zero values
5905 			 * for the single allowed zero compression - "::".
5906 			 */
5907 			firstzero = -1;
5908 			tryzero = -1;
5909 			numzero = 1;
5910 			for (i = 0; i < sizeof (struct in6_addr); i++) {
5911 #ifdef illumos
5912 				if (ip6._S6_un._S6_u8[i] == 0 &&
5913 #else
5914 				if (ip6.__u6_addr.__u6_addr8[i] == 0 &&
5915 #endif
5916 				    tryzero == -1 && i % 2 == 0) {
5917 					tryzero = i;
5918 					continue;
5919 				}
5920 
5921 				if (tryzero != -1 &&
5922 #ifdef illumos
5923 				    (ip6._S6_un._S6_u8[i] != 0 ||
5924 #else
5925 				    (ip6.__u6_addr.__u6_addr8[i] != 0 ||
5926 #endif
5927 				    i == sizeof (struct in6_addr) - 1)) {
5928 
5929 					if (i - tryzero <= numzero) {
5930 						tryzero = -1;
5931 						continue;
5932 					}
5933 
5934 					firstzero = tryzero;
5935 					numzero = i - i % 2 - tryzero;
5936 					tryzero = -1;
5937 
5938 #ifdef illumos
5939 					if (ip6._S6_un._S6_u8[i] == 0 &&
5940 #else
5941 					if (ip6.__u6_addr.__u6_addr8[i] == 0 &&
5942 #endif
5943 					    i == sizeof (struct in6_addr) - 1)
5944 						numzero += 2;
5945 				}
5946 			}
5947 			ASSERT(firstzero + numzero <= sizeof (struct in6_addr));
5948 
5949 			/*
5950 			 * Check for an IPv4 embedded address.
5951 			 */
5952 			v6end = sizeof (struct in6_addr) - 2;
5953 			if (IN6_IS_ADDR_V4MAPPED(&ip6) ||
5954 			    IN6_IS_ADDR_V4COMPAT(&ip6)) {
5955 				for (i = sizeof (struct in6_addr) - 1;
5956 				    i >= DTRACE_V4MAPPED_OFFSET; i--) {
5957 					ASSERT(end >= base);
5958 
5959 #ifdef illumos
5960 					val = ip6._S6_un._S6_u8[i];
5961 #else
5962 					val = ip6.__u6_addr.__u6_addr8[i];
5963 #endif
5964 
5965 					if (val == 0) {
5966 						*end-- = '0';
5967 					} else {
5968 						for (; val; val /= 10) {
5969 							*end-- = '0' + val % 10;
5970 						}
5971 					}
5972 
5973 					if (i > DTRACE_V4MAPPED_OFFSET)
5974 						*end-- = '.';
5975 				}
5976 
5977 				if (subr == DIF_SUBR_INET_NTOA6)
5978 					goto inetout;
5979 
5980 				/*
5981 				 * Set v6end to skip the IPv4 address that
5982 				 * we have already stringified.
5983 				 */
5984 				v6end = 10;
5985 			}
5986 
5987 			/*
5988 			 * Build the IPv6 string by working through the
5989 			 * address in reverse.
5990 			 */
5991 			for (i = v6end; i >= 0; i -= 2) {
5992 				ASSERT(end >= base);
5993 
5994 				if (i == firstzero + numzero - 2) {
5995 					*end-- = ':';
5996 					*end-- = ':';
5997 					i -= numzero - 2;
5998 					continue;
5999 				}
6000 
6001 				if (i < 14 && i != firstzero - 2)
6002 					*end-- = ':';
6003 
6004 #ifdef illumos
6005 				val = (ip6._S6_un._S6_u8[i] << 8) +
6006 				    ip6._S6_un._S6_u8[i + 1];
6007 #else
6008 				val = (ip6.__u6_addr.__u6_addr8[i] << 8) +
6009 				    ip6.__u6_addr.__u6_addr8[i + 1];
6010 #endif
6011 
6012 				if (val == 0) {
6013 					*end-- = '0';
6014 				} else {
6015 					for (; val; val /= 16) {
6016 						*end-- = digits[val % 16];
6017 					}
6018 				}
6019 			}
6020 			ASSERT(end + 1 >= base);
6021 
6022 		} else {
6023 			/*
6024 			 * The user didn't use AH_INET or AH_INET6.
6025 			 */
6026 			DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
6027 			regs[rd] = 0;
6028 			break;
6029 		}
6030 
6031 inetout:	regs[rd] = (uintptr_t)end + 1;
6032 		mstate->dtms_scratch_ptr += size;
6033 		break;
6034 	}
6035 
6036 	case DIF_SUBR_MEMREF: {
6037 		uintptr_t size = 2 * sizeof(uintptr_t);
6038 		uintptr_t *memref = (uintptr_t *) P2ROUNDUP(mstate->dtms_scratch_ptr, sizeof(uintptr_t));
6039 		size_t scratch_size = ((uintptr_t) memref - mstate->dtms_scratch_ptr) + size;
6040 
6041 		/* address and length */
6042 		memref[0] = tupregs[0].dttk_value;
6043 		memref[1] = tupregs[1].dttk_value;
6044 
6045 		regs[rd] = (uintptr_t) memref;
6046 		mstate->dtms_scratch_ptr += scratch_size;
6047 		break;
6048 	}
6049 
6050 #ifndef illumos
6051 	case DIF_SUBR_MEMSTR: {
6052 		char *str = (char *)mstate->dtms_scratch_ptr;
6053 		uintptr_t mem = tupregs[0].dttk_value;
6054 		char c = tupregs[1].dttk_value;
6055 		size_t size = tupregs[2].dttk_value;
6056 		uint8_t n;
6057 		int i;
6058 
6059 		regs[rd] = 0;
6060 
6061 		if (size == 0)
6062 			break;
6063 
6064 		if (!dtrace_canload(mem, size - 1, mstate, vstate))
6065 			break;
6066 
6067 		if (!DTRACE_INSCRATCH(mstate, size)) {
6068 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
6069 			break;
6070 		}
6071 
6072 		if (dtrace_memstr_max != 0 && size > dtrace_memstr_max) {
6073 			*flags |= CPU_DTRACE_ILLOP;
6074 			break;
6075 		}
6076 
6077 		for (i = 0; i < size - 1; i++) {
6078 			n = dtrace_load8(mem++);
6079 			str[i] = (n == 0) ? c : n;
6080 		}
6081 		str[size - 1] = 0;
6082 
6083 		regs[rd] = (uintptr_t)str;
6084 		mstate->dtms_scratch_ptr += size;
6085 		break;
6086 	}
6087 #endif
6088 	}
6089 }
6090 
6091 /*
6092  * Emulate the execution of DTrace IR instructions specified by the given
6093  * DIF object.  This function is deliberately void of assertions as all of
6094  * the necessary checks are handled by a call to dtrace_difo_validate().
6095  */
6096 static uint64_t
6097 dtrace_dif_emulate(dtrace_difo_t *difo, dtrace_mstate_t *mstate,
6098     dtrace_vstate_t *vstate, dtrace_state_t *state)
6099 {
6100 	const dif_instr_t *text = difo->dtdo_buf;
6101 	const uint_t textlen = difo->dtdo_len;
6102 	const char *strtab = difo->dtdo_strtab;
6103 	const uint64_t *inttab = difo->dtdo_inttab;
6104 
6105 	uint64_t rval = 0;
6106 	dtrace_statvar_t *svar;
6107 	dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
6108 	dtrace_difv_t *v;
6109 	volatile uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
6110 	volatile uintptr_t *illval = &cpu_core[curcpu].cpuc_dtrace_illval;
6111 
6112 	dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
6113 	uint64_t regs[DIF_DIR_NREGS];
6114 	uint64_t *tmp;
6115 
6116 	uint8_t cc_n = 0, cc_z = 0, cc_v = 0, cc_c = 0;
6117 	int64_t cc_r;
6118 	uint_t pc = 0, id, opc = 0;
6119 	uint8_t ttop = 0;
6120 	dif_instr_t instr;
6121 	uint_t r1, r2, rd;
6122 
6123 	/*
6124 	 * We stash the current DIF object into the machine state: we need it
6125 	 * for subsequent access checking.
6126 	 */
6127 	mstate->dtms_difo = difo;
6128 
6129 	regs[DIF_REG_R0] = 0; 		/* %r0 is fixed at zero */
6130 
6131 	while (pc < textlen && !(*flags & CPU_DTRACE_FAULT)) {
6132 		opc = pc;
6133 
6134 		instr = text[pc++];
6135 		r1 = DIF_INSTR_R1(instr);
6136 		r2 = DIF_INSTR_R2(instr);
6137 		rd = DIF_INSTR_RD(instr);
6138 
6139 		switch (DIF_INSTR_OP(instr)) {
6140 		case DIF_OP_OR:
6141 			regs[rd] = regs[r1] | regs[r2];
6142 			break;
6143 		case DIF_OP_XOR:
6144 			regs[rd] = regs[r1] ^ regs[r2];
6145 			break;
6146 		case DIF_OP_AND:
6147 			regs[rd] = regs[r1] & regs[r2];
6148 			break;
6149 		case DIF_OP_SLL:
6150 			regs[rd] = regs[r1] << regs[r2];
6151 			break;
6152 		case DIF_OP_SRL:
6153 			regs[rd] = regs[r1] >> regs[r2];
6154 			break;
6155 		case DIF_OP_SUB:
6156 			regs[rd] = regs[r1] - regs[r2];
6157 			break;
6158 		case DIF_OP_ADD:
6159 			regs[rd] = regs[r1] + regs[r2];
6160 			break;
6161 		case DIF_OP_MUL:
6162 			regs[rd] = regs[r1] * regs[r2];
6163 			break;
6164 		case DIF_OP_SDIV:
6165 			if (regs[r2] == 0) {
6166 				regs[rd] = 0;
6167 				*flags |= CPU_DTRACE_DIVZERO;
6168 			} else {
6169 				regs[rd] = (int64_t)regs[r1] /
6170 				    (int64_t)regs[r2];
6171 			}
6172 			break;
6173 
6174 		case DIF_OP_UDIV:
6175 			if (regs[r2] == 0) {
6176 				regs[rd] = 0;
6177 				*flags |= CPU_DTRACE_DIVZERO;
6178 			} else {
6179 				regs[rd] = regs[r1] / regs[r2];
6180 			}
6181 			break;
6182 
6183 		case DIF_OP_SREM:
6184 			if (regs[r2] == 0) {
6185 				regs[rd] = 0;
6186 				*flags |= CPU_DTRACE_DIVZERO;
6187 			} else {
6188 				regs[rd] = (int64_t)regs[r1] %
6189 				    (int64_t)regs[r2];
6190 			}
6191 			break;
6192 
6193 		case DIF_OP_UREM:
6194 			if (regs[r2] == 0) {
6195 				regs[rd] = 0;
6196 				*flags |= CPU_DTRACE_DIVZERO;
6197 			} else {
6198 				regs[rd] = regs[r1] % regs[r2];
6199 			}
6200 			break;
6201 
6202 		case DIF_OP_NOT:
6203 			regs[rd] = ~regs[r1];
6204 			break;
6205 		case DIF_OP_MOV:
6206 			regs[rd] = regs[r1];
6207 			break;
6208 		case DIF_OP_CMP:
6209 			cc_r = regs[r1] - regs[r2];
6210 			cc_n = cc_r < 0;
6211 			cc_z = cc_r == 0;
6212 			cc_v = 0;
6213 			cc_c = regs[r1] < regs[r2];
6214 			break;
6215 		case DIF_OP_TST:
6216 			cc_n = cc_v = cc_c = 0;
6217 			cc_z = regs[r1] == 0;
6218 			break;
6219 		case DIF_OP_BA:
6220 			pc = DIF_INSTR_LABEL(instr);
6221 			break;
6222 		case DIF_OP_BE:
6223 			if (cc_z)
6224 				pc = DIF_INSTR_LABEL(instr);
6225 			break;
6226 		case DIF_OP_BNE:
6227 			if (cc_z == 0)
6228 				pc = DIF_INSTR_LABEL(instr);
6229 			break;
6230 		case DIF_OP_BG:
6231 			if ((cc_z | (cc_n ^ cc_v)) == 0)
6232 				pc = DIF_INSTR_LABEL(instr);
6233 			break;
6234 		case DIF_OP_BGU:
6235 			if ((cc_c | cc_z) == 0)
6236 				pc = DIF_INSTR_LABEL(instr);
6237 			break;
6238 		case DIF_OP_BGE:
6239 			if ((cc_n ^ cc_v) == 0)
6240 				pc = DIF_INSTR_LABEL(instr);
6241 			break;
6242 		case DIF_OP_BGEU:
6243 			if (cc_c == 0)
6244 				pc = DIF_INSTR_LABEL(instr);
6245 			break;
6246 		case DIF_OP_BL:
6247 			if (cc_n ^ cc_v)
6248 				pc = DIF_INSTR_LABEL(instr);
6249 			break;
6250 		case DIF_OP_BLU:
6251 			if (cc_c)
6252 				pc = DIF_INSTR_LABEL(instr);
6253 			break;
6254 		case DIF_OP_BLE:
6255 			if (cc_z | (cc_n ^ cc_v))
6256 				pc = DIF_INSTR_LABEL(instr);
6257 			break;
6258 		case DIF_OP_BLEU:
6259 			if (cc_c | cc_z)
6260 				pc = DIF_INSTR_LABEL(instr);
6261 			break;
6262 		case DIF_OP_RLDSB:
6263 			if (!dtrace_canload(regs[r1], 1, mstate, vstate))
6264 				break;
6265 			/*FALLTHROUGH*/
6266 		case DIF_OP_LDSB:
6267 			regs[rd] = (int8_t)dtrace_load8(regs[r1]);
6268 			break;
6269 		case DIF_OP_RLDSH:
6270 			if (!dtrace_canload(regs[r1], 2, mstate, vstate))
6271 				break;
6272 			/*FALLTHROUGH*/
6273 		case DIF_OP_LDSH:
6274 			regs[rd] = (int16_t)dtrace_load16(regs[r1]);
6275 			break;
6276 		case DIF_OP_RLDSW:
6277 			if (!dtrace_canload(regs[r1], 4, mstate, vstate))
6278 				break;
6279 			/*FALLTHROUGH*/
6280 		case DIF_OP_LDSW:
6281 			regs[rd] = (int32_t)dtrace_load32(regs[r1]);
6282 			break;
6283 		case DIF_OP_RLDUB:
6284 			if (!dtrace_canload(regs[r1], 1, mstate, vstate))
6285 				break;
6286 			/*FALLTHROUGH*/
6287 		case DIF_OP_LDUB:
6288 			regs[rd] = dtrace_load8(regs[r1]);
6289 			break;
6290 		case DIF_OP_RLDUH:
6291 			if (!dtrace_canload(regs[r1], 2, mstate, vstate))
6292 				break;
6293 			/*FALLTHROUGH*/
6294 		case DIF_OP_LDUH:
6295 			regs[rd] = dtrace_load16(regs[r1]);
6296 			break;
6297 		case DIF_OP_RLDUW:
6298 			if (!dtrace_canload(regs[r1], 4, mstate, vstate))
6299 				break;
6300 			/*FALLTHROUGH*/
6301 		case DIF_OP_LDUW:
6302 			regs[rd] = dtrace_load32(regs[r1]);
6303 			break;
6304 		case DIF_OP_RLDX:
6305 			if (!dtrace_canload(regs[r1], 8, mstate, vstate))
6306 				break;
6307 			/*FALLTHROUGH*/
6308 		case DIF_OP_LDX:
6309 			regs[rd] = dtrace_load64(regs[r1]);
6310 			break;
6311 		case DIF_OP_ULDSB:
6312 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6313 			regs[rd] = (int8_t)
6314 			    dtrace_fuword8((void *)(uintptr_t)regs[r1]);
6315 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6316 			break;
6317 		case DIF_OP_ULDSH:
6318 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6319 			regs[rd] = (int16_t)
6320 			    dtrace_fuword16((void *)(uintptr_t)regs[r1]);
6321 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6322 			break;
6323 		case DIF_OP_ULDSW:
6324 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6325 			regs[rd] = (int32_t)
6326 			    dtrace_fuword32((void *)(uintptr_t)regs[r1]);
6327 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6328 			break;
6329 		case DIF_OP_ULDUB:
6330 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6331 			regs[rd] =
6332 			    dtrace_fuword8((void *)(uintptr_t)regs[r1]);
6333 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6334 			break;
6335 		case DIF_OP_ULDUH:
6336 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6337 			regs[rd] =
6338 			    dtrace_fuword16((void *)(uintptr_t)regs[r1]);
6339 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6340 			break;
6341 		case DIF_OP_ULDUW:
6342 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6343 			regs[rd] =
6344 			    dtrace_fuword32((void *)(uintptr_t)regs[r1]);
6345 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6346 			break;
6347 		case DIF_OP_ULDX:
6348 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6349 			regs[rd] =
6350 			    dtrace_fuword64((void *)(uintptr_t)regs[r1]);
6351 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6352 			break;
6353 		case DIF_OP_RET:
6354 			rval = regs[rd];
6355 			pc = textlen;
6356 			break;
6357 		case DIF_OP_NOP:
6358 			break;
6359 		case DIF_OP_SETX:
6360 			regs[rd] = inttab[DIF_INSTR_INTEGER(instr)];
6361 			break;
6362 		case DIF_OP_SETS:
6363 			regs[rd] = (uint64_t)(uintptr_t)
6364 			    (strtab + DIF_INSTR_STRING(instr));
6365 			break;
6366 		case DIF_OP_SCMP: {
6367 			size_t sz = state->dts_options[DTRACEOPT_STRSIZE];
6368 			uintptr_t s1 = regs[r1];
6369 			uintptr_t s2 = regs[r2];
6370 			size_t lim1, lim2;
6371 
6372 			/*
6373 			 * If one of the strings is NULL then the limit becomes
6374 			 * 0 which compares 0 characters in dtrace_strncmp()
6375 			 * resulting in a false positive.  dtrace_strncmp()
6376 			 * treats a NULL as an empty 1-char string.
6377 			 */
6378 			lim1 = lim2 = 1;
6379 
6380 			if (s1 != 0 &&
6381 			    !dtrace_strcanload(s1, sz, &lim1, mstate, vstate))
6382 				break;
6383 			if (s2 != 0 &&
6384 			    !dtrace_strcanload(s2, sz, &lim2, mstate, vstate))
6385 				break;
6386 
6387 			cc_r = dtrace_strncmp((char *)s1, (char *)s2,
6388 			    MIN(lim1, lim2));
6389 
6390 			cc_n = cc_r < 0;
6391 			cc_z = cc_r == 0;
6392 			cc_v = cc_c = 0;
6393 			break;
6394 		}
6395 		case DIF_OP_LDGA:
6396 			regs[rd] = dtrace_dif_variable(mstate, state,
6397 			    r1, regs[r2]);
6398 			break;
6399 		case DIF_OP_LDGS:
6400 			id = DIF_INSTR_VAR(instr);
6401 
6402 			if (id >= DIF_VAR_OTHER_UBASE) {
6403 				uintptr_t a;
6404 
6405 				id -= DIF_VAR_OTHER_UBASE;
6406 				svar = vstate->dtvs_globals[id];
6407 				ASSERT(svar != NULL);
6408 				v = &svar->dtsv_var;
6409 
6410 				if (!(v->dtdv_type.dtdt_flags & DIF_TF_BYREF)) {
6411 					regs[rd] = svar->dtsv_data;
6412 					break;
6413 				}
6414 
6415 				a = (uintptr_t)svar->dtsv_data;
6416 
6417 				if (*(uint8_t *)a == UINT8_MAX) {
6418 					/*
6419 					 * If the 0th byte is set to UINT8_MAX
6420 					 * then this is to be treated as a
6421 					 * reference to a NULL variable.
6422 					 */
6423 					regs[rd] = 0;
6424 				} else {
6425 					regs[rd] = a + sizeof (uint64_t);
6426 				}
6427 
6428 				break;
6429 			}
6430 
6431 			regs[rd] = dtrace_dif_variable(mstate, state, id, 0);
6432 			break;
6433 
6434 		case DIF_OP_STGS:
6435 			id = DIF_INSTR_VAR(instr);
6436 
6437 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
6438 			id -= DIF_VAR_OTHER_UBASE;
6439 
6440 			VERIFY(id < vstate->dtvs_nglobals);
6441 			svar = vstate->dtvs_globals[id];
6442 			ASSERT(svar != NULL);
6443 			v = &svar->dtsv_var;
6444 
6445 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6446 				uintptr_t a = (uintptr_t)svar->dtsv_data;
6447 				size_t lim;
6448 
6449 				ASSERT(a != 0);
6450 				ASSERT(svar->dtsv_size != 0);
6451 
6452 				if (regs[rd] == 0) {
6453 					*(uint8_t *)a = UINT8_MAX;
6454 					break;
6455 				} else {
6456 					*(uint8_t *)a = 0;
6457 					a += sizeof (uint64_t);
6458 				}
6459 				if (!dtrace_vcanload(
6460 				    (void *)(uintptr_t)regs[rd], &v->dtdv_type,
6461 				    &lim, mstate, vstate))
6462 					break;
6463 
6464 				dtrace_vcopy((void *)(uintptr_t)regs[rd],
6465 				    (void *)a, &v->dtdv_type, lim);
6466 				break;
6467 			}
6468 
6469 			svar->dtsv_data = regs[rd];
6470 			break;
6471 
6472 		case DIF_OP_LDTA:
6473 			/*
6474 			 * There are no DTrace built-in thread-local arrays at
6475 			 * present.  This opcode is saved for future work.
6476 			 */
6477 			*flags |= CPU_DTRACE_ILLOP;
6478 			regs[rd] = 0;
6479 			break;
6480 
6481 		case DIF_OP_LDLS:
6482 			id = DIF_INSTR_VAR(instr);
6483 
6484 			if (id < DIF_VAR_OTHER_UBASE) {
6485 				/*
6486 				 * For now, this has no meaning.
6487 				 */
6488 				regs[rd] = 0;
6489 				break;
6490 			}
6491 
6492 			id -= DIF_VAR_OTHER_UBASE;
6493 
6494 			ASSERT(id < vstate->dtvs_nlocals);
6495 			ASSERT(vstate->dtvs_locals != NULL);
6496 
6497 			svar = vstate->dtvs_locals[id];
6498 			ASSERT(svar != NULL);
6499 			v = &svar->dtsv_var;
6500 
6501 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6502 				uintptr_t a = (uintptr_t)svar->dtsv_data;
6503 				size_t sz = v->dtdv_type.dtdt_size;
6504 				size_t lim;
6505 
6506 				sz += sizeof (uint64_t);
6507 				ASSERT(svar->dtsv_size == NCPU * sz);
6508 				a += curcpu * sz;
6509 
6510 				if (*(uint8_t *)a == UINT8_MAX) {
6511 					/*
6512 					 * If the 0th byte is set to UINT8_MAX
6513 					 * then this is to be treated as a
6514 					 * reference to a NULL variable.
6515 					 */
6516 					regs[rd] = 0;
6517 				} else {
6518 					regs[rd] = a + sizeof (uint64_t);
6519 				}
6520 
6521 				break;
6522 			}
6523 
6524 			ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
6525 			tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
6526 			regs[rd] = tmp[curcpu];
6527 			break;
6528 
6529 		case DIF_OP_STLS:
6530 			id = DIF_INSTR_VAR(instr);
6531 
6532 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
6533 			id -= DIF_VAR_OTHER_UBASE;
6534 			VERIFY(id < vstate->dtvs_nlocals);
6535 
6536 			ASSERT(vstate->dtvs_locals != NULL);
6537 			svar = vstate->dtvs_locals[id];
6538 			ASSERT(svar != NULL);
6539 			v = &svar->dtsv_var;
6540 
6541 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6542 				uintptr_t a = (uintptr_t)svar->dtsv_data;
6543 				size_t sz = v->dtdv_type.dtdt_size;
6544 				size_t lim;
6545 
6546 				sz += sizeof (uint64_t);
6547 				ASSERT(svar->dtsv_size == NCPU * sz);
6548 				a += curcpu * sz;
6549 
6550 				if (regs[rd] == 0) {
6551 					*(uint8_t *)a = UINT8_MAX;
6552 					break;
6553 				} else {
6554 					*(uint8_t *)a = 0;
6555 					a += sizeof (uint64_t);
6556 				}
6557 
6558 				if (!dtrace_vcanload(
6559 				    (void *)(uintptr_t)regs[rd], &v->dtdv_type,
6560 				    &lim, mstate, vstate))
6561 					break;
6562 
6563 				dtrace_vcopy((void *)(uintptr_t)regs[rd],
6564 				    (void *)a, &v->dtdv_type, lim);
6565 				break;
6566 			}
6567 
6568 			ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
6569 			tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
6570 			tmp[curcpu] = regs[rd];
6571 			break;
6572 
6573 		case DIF_OP_LDTS: {
6574 			dtrace_dynvar_t *dvar;
6575 			dtrace_key_t *key;
6576 
6577 			id = DIF_INSTR_VAR(instr);
6578 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
6579 			id -= DIF_VAR_OTHER_UBASE;
6580 			v = &vstate->dtvs_tlocals[id];
6581 
6582 			key = &tupregs[DIF_DTR_NREGS];
6583 			key[0].dttk_value = (uint64_t)id;
6584 			key[0].dttk_size = 0;
6585 			DTRACE_TLS_THRKEY(key[1].dttk_value);
6586 			key[1].dttk_size = 0;
6587 
6588 			dvar = dtrace_dynvar(dstate, 2, key,
6589 			    sizeof (uint64_t), DTRACE_DYNVAR_NOALLOC,
6590 			    mstate, vstate);
6591 
6592 			if (dvar == NULL) {
6593 				regs[rd] = 0;
6594 				break;
6595 			}
6596 
6597 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6598 				regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
6599 			} else {
6600 				regs[rd] = *((uint64_t *)dvar->dtdv_data);
6601 			}
6602 
6603 			break;
6604 		}
6605 
6606 		case DIF_OP_STTS: {
6607 			dtrace_dynvar_t *dvar;
6608 			dtrace_key_t *key;
6609 
6610 			id = DIF_INSTR_VAR(instr);
6611 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
6612 			id -= DIF_VAR_OTHER_UBASE;
6613 			VERIFY(id < vstate->dtvs_ntlocals);
6614 
6615 			key = &tupregs[DIF_DTR_NREGS];
6616 			key[0].dttk_value = (uint64_t)id;
6617 			key[0].dttk_size = 0;
6618 			DTRACE_TLS_THRKEY(key[1].dttk_value);
6619 			key[1].dttk_size = 0;
6620 			v = &vstate->dtvs_tlocals[id];
6621 
6622 			dvar = dtrace_dynvar(dstate, 2, key,
6623 			    v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
6624 			    v->dtdv_type.dtdt_size : sizeof (uint64_t),
6625 			    regs[rd] ? DTRACE_DYNVAR_ALLOC :
6626 			    DTRACE_DYNVAR_DEALLOC, mstate, vstate);
6627 
6628 			/*
6629 			 * Given that we're storing to thread-local data,
6630 			 * we need to flush our predicate cache.
6631 			 */
6632 			curthread->t_predcache = 0;
6633 
6634 			if (dvar == NULL)
6635 				break;
6636 
6637 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6638 				size_t lim;
6639 
6640 				if (!dtrace_vcanload(
6641 				    (void *)(uintptr_t)regs[rd],
6642 				    &v->dtdv_type, &lim, mstate, vstate))
6643 					break;
6644 
6645 				dtrace_vcopy((void *)(uintptr_t)regs[rd],
6646 				    dvar->dtdv_data, &v->dtdv_type, lim);
6647 			} else {
6648 				*((uint64_t *)dvar->dtdv_data) = regs[rd];
6649 			}
6650 
6651 			break;
6652 		}
6653 
6654 		case DIF_OP_SRA:
6655 			regs[rd] = (int64_t)regs[r1] >> regs[r2];
6656 			break;
6657 
6658 		case DIF_OP_CALL:
6659 			dtrace_dif_subr(DIF_INSTR_SUBR(instr), rd,
6660 			    regs, tupregs, ttop, mstate, state);
6661 			break;
6662 
6663 		case DIF_OP_PUSHTR:
6664 			if (ttop == DIF_DTR_NREGS) {
6665 				*flags |= CPU_DTRACE_TUPOFLOW;
6666 				break;
6667 			}
6668 
6669 			if (r1 == DIF_TYPE_STRING) {
6670 				/*
6671 				 * If this is a string type and the size is 0,
6672 				 * we'll use the system-wide default string
6673 				 * size.  Note that we are _not_ looking at
6674 				 * the value of the DTRACEOPT_STRSIZE option;
6675 				 * had this been set, we would expect to have
6676 				 * a non-zero size value in the "pushtr".
6677 				 */
6678 				tupregs[ttop].dttk_size =
6679 				    dtrace_strlen((char *)(uintptr_t)regs[rd],
6680 				    regs[r2] ? regs[r2] :
6681 				    dtrace_strsize_default) + 1;
6682 			} else {
6683 				if (regs[r2] > LONG_MAX) {
6684 					*flags |= CPU_DTRACE_ILLOP;
6685 					break;
6686 				}
6687 
6688 				tupregs[ttop].dttk_size = regs[r2];
6689 			}
6690 
6691 			tupregs[ttop++].dttk_value = regs[rd];
6692 			break;
6693 
6694 		case DIF_OP_PUSHTV:
6695 			if (ttop == DIF_DTR_NREGS) {
6696 				*flags |= CPU_DTRACE_TUPOFLOW;
6697 				break;
6698 			}
6699 
6700 			tupregs[ttop].dttk_value = regs[rd];
6701 			tupregs[ttop++].dttk_size = 0;
6702 			break;
6703 
6704 		case DIF_OP_POPTS:
6705 			if (ttop != 0)
6706 				ttop--;
6707 			break;
6708 
6709 		case DIF_OP_FLUSHTS:
6710 			ttop = 0;
6711 			break;
6712 
6713 		case DIF_OP_LDGAA:
6714 		case DIF_OP_LDTAA: {
6715 			dtrace_dynvar_t *dvar;
6716 			dtrace_key_t *key = tupregs;
6717 			uint_t nkeys = ttop;
6718 
6719 			id = DIF_INSTR_VAR(instr);
6720 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
6721 			id -= DIF_VAR_OTHER_UBASE;
6722 
6723 			key[nkeys].dttk_value = (uint64_t)id;
6724 			key[nkeys++].dttk_size = 0;
6725 
6726 			if (DIF_INSTR_OP(instr) == DIF_OP_LDTAA) {
6727 				DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
6728 				key[nkeys++].dttk_size = 0;
6729 				VERIFY(id < vstate->dtvs_ntlocals);
6730 				v = &vstate->dtvs_tlocals[id];
6731 			} else {
6732 				VERIFY(id < vstate->dtvs_nglobals);
6733 				v = &vstate->dtvs_globals[id]->dtsv_var;
6734 			}
6735 
6736 			dvar = dtrace_dynvar(dstate, nkeys, key,
6737 			    v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
6738 			    v->dtdv_type.dtdt_size : sizeof (uint64_t),
6739 			    DTRACE_DYNVAR_NOALLOC, mstate, vstate);
6740 
6741 			if (dvar == NULL) {
6742 				regs[rd] = 0;
6743 				break;
6744 			}
6745 
6746 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6747 				regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
6748 			} else {
6749 				regs[rd] = *((uint64_t *)dvar->dtdv_data);
6750 			}
6751 
6752 			break;
6753 		}
6754 
6755 		case DIF_OP_STGAA:
6756 		case DIF_OP_STTAA: {
6757 			dtrace_dynvar_t *dvar;
6758 			dtrace_key_t *key = tupregs;
6759 			uint_t nkeys = ttop;
6760 
6761 			id = DIF_INSTR_VAR(instr);
6762 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
6763 			id -= DIF_VAR_OTHER_UBASE;
6764 
6765 			key[nkeys].dttk_value = (uint64_t)id;
6766 			key[nkeys++].dttk_size = 0;
6767 
6768 			if (DIF_INSTR_OP(instr) == DIF_OP_STTAA) {
6769 				DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
6770 				key[nkeys++].dttk_size = 0;
6771 				VERIFY(id < vstate->dtvs_ntlocals);
6772 				v = &vstate->dtvs_tlocals[id];
6773 			} else {
6774 				VERIFY(id < vstate->dtvs_nglobals);
6775 				v = &vstate->dtvs_globals[id]->dtsv_var;
6776 			}
6777 
6778 			dvar = dtrace_dynvar(dstate, nkeys, key,
6779 			    v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
6780 			    v->dtdv_type.dtdt_size : sizeof (uint64_t),
6781 			    regs[rd] ? DTRACE_DYNVAR_ALLOC :
6782 			    DTRACE_DYNVAR_DEALLOC, mstate, vstate);
6783 
6784 			if (dvar == NULL)
6785 				break;
6786 
6787 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6788 				size_t lim;
6789 
6790 				if (!dtrace_vcanload(
6791 				    (void *)(uintptr_t)regs[rd], &v->dtdv_type,
6792 				    &lim, mstate, vstate))
6793 					break;
6794 
6795 				dtrace_vcopy((void *)(uintptr_t)regs[rd],
6796 				    dvar->dtdv_data, &v->dtdv_type, lim);
6797 			} else {
6798 				*((uint64_t *)dvar->dtdv_data) = regs[rd];
6799 			}
6800 
6801 			break;
6802 		}
6803 
6804 		case DIF_OP_ALLOCS: {
6805 			uintptr_t ptr = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
6806 			size_t size = ptr - mstate->dtms_scratch_ptr + regs[r1];
6807 
6808 			/*
6809 			 * Rounding up the user allocation size could have
6810 			 * overflowed large, bogus allocations (like -1ULL) to
6811 			 * 0.
6812 			 */
6813 			if (size < regs[r1] ||
6814 			    !DTRACE_INSCRATCH(mstate, size)) {
6815 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
6816 				regs[rd] = 0;
6817 				break;
6818 			}
6819 
6820 			dtrace_bzero((void *) mstate->dtms_scratch_ptr, size);
6821 			mstate->dtms_scratch_ptr += size;
6822 			regs[rd] = ptr;
6823 			break;
6824 		}
6825 
6826 		case DIF_OP_COPYS:
6827 			if (!dtrace_canstore(regs[rd], regs[r2],
6828 			    mstate, vstate)) {
6829 				*flags |= CPU_DTRACE_BADADDR;
6830 				*illval = regs[rd];
6831 				break;
6832 			}
6833 
6834 			if (!dtrace_canload(regs[r1], regs[r2], mstate, vstate))
6835 				break;
6836 
6837 			dtrace_bcopy((void *)(uintptr_t)regs[r1],
6838 			    (void *)(uintptr_t)regs[rd], (size_t)regs[r2]);
6839 			break;
6840 
6841 		case DIF_OP_STB:
6842 			if (!dtrace_canstore(regs[rd], 1, mstate, vstate)) {
6843 				*flags |= CPU_DTRACE_BADADDR;
6844 				*illval = regs[rd];
6845 				break;
6846 			}
6847 			*((uint8_t *)(uintptr_t)regs[rd]) = (uint8_t)regs[r1];
6848 			break;
6849 
6850 		case DIF_OP_STH:
6851 			if (!dtrace_canstore(regs[rd], 2, mstate, vstate)) {
6852 				*flags |= CPU_DTRACE_BADADDR;
6853 				*illval = regs[rd];
6854 				break;
6855 			}
6856 			if (regs[rd] & 1) {
6857 				*flags |= CPU_DTRACE_BADALIGN;
6858 				*illval = regs[rd];
6859 				break;
6860 			}
6861 			*((uint16_t *)(uintptr_t)regs[rd]) = (uint16_t)regs[r1];
6862 			break;
6863 
6864 		case DIF_OP_STW:
6865 			if (!dtrace_canstore(regs[rd], 4, mstate, vstate)) {
6866 				*flags |= CPU_DTRACE_BADADDR;
6867 				*illval = regs[rd];
6868 				break;
6869 			}
6870 			if (regs[rd] & 3) {
6871 				*flags |= CPU_DTRACE_BADALIGN;
6872 				*illval = regs[rd];
6873 				break;
6874 			}
6875 			*((uint32_t *)(uintptr_t)regs[rd]) = (uint32_t)regs[r1];
6876 			break;
6877 
6878 		case DIF_OP_STX:
6879 			if (!dtrace_canstore(regs[rd], 8, mstate, vstate)) {
6880 				*flags |= CPU_DTRACE_BADADDR;
6881 				*illval = regs[rd];
6882 				break;
6883 			}
6884 			if (regs[rd] & 7) {
6885 				*flags |= CPU_DTRACE_BADALIGN;
6886 				*illval = regs[rd];
6887 				break;
6888 			}
6889 			*((uint64_t *)(uintptr_t)regs[rd]) = regs[r1];
6890 			break;
6891 		}
6892 	}
6893 
6894 	if (!(*flags & CPU_DTRACE_FAULT))
6895 		return (rval);
6896 
6897 	mstate->dtms_fltoffs = opc * sizeof (dif_instr_t);
6898 	mstate->dtms_present |= DTRACE_MSTATE_FLTOFFS;
6899 
6900 	return (0);
6901 }
6902 
6903 static void
6904 dtrace_action_breakpoint(dtrace_ecb_t *ecb)
6905 {
6906 	dtrace_probe_t *probe = ecb->dte_probe;
6907 	dtrace_provider_t *prov = probe->dtpr_provider;
6908 	char c[DTRACE_FULLNAMELEN + 80], *str;
6909 	char *msg = "dtrace: breakpoint action at probe ";
6910 	char *ecbmsg = " (ecb ";
6911 	uintptr_t mask = (0xf << (sizeof (uintptr_t) * NBBY / 4));
6912 	uintptr_t val = (uintptr_t)ecb;
6913 	int shift = (sizeof (uintptr_t) * NBBY) - 4, i = 0;
6914 
6915 	if (dtrace_destructive_disallow)
6916 		return;
6917 
6918 	/*
6919 	 * It's impossible to be taking action on the NULL probe.
6920 	 */
6921 	ASSERT(probe != NULL);
6922 
6923 	/*
6924 	 * This is a poor man's (destitute man's?) sprintf():  we want to
6925 	 * print the provider name, module name, function name and name of
6926 	 * the probe, along with the hex address of the ECB with the breakpoint
6927 	 * action -- all of which we must place in the character buffer by
6928 	 * hand.
6929 	 */
6930 	while (*msg != '\0')
6931 		c[i++] = *msg++;
6932 
6933 	for (str = prov->dtpv_name; *str != '\0'; str++)
6934 		c[i++] = *str;
6935 	c[i++] = ':';
6936 
6937 	for (str = probe->dtpr_mod; *str != '\0'; str++)
6938 		c[i++] = *str;
6939 	c[i++] = ':';
6940 
6941 	for (str = probe->dtpr_func; *str != '\0'; str++)
6942 		c[i++] = *str;
6943 	c[i++] = ':';
6944 
6945 	for (str = probe->dtpr_name; *str != '\0'; str++)
6946 		c[i++] = *str;
6947 
6948 	while (*ecbmsg != '\0')
6949 		c[i++] = *ecbmsg++;
6950 
6951 	while (shift >= 0) {
6952 		mask = (uintptr_t)0xf << shift;
6953 
6954 		if (val >= ((uintptr_t)1 << shift))
6955 			c[i++] = "0123456789abcdef"[(val & mask) >> shift];
6956 		shift -= 4;
6957 	}
6958 
6959 	c[i++] = ')';
6960 	c[i] = '\0';
6961 
6962 #ifdef illumos
6963 	debug_enter(c);
6964 #else
6965 	kdb_enter(KDB_WHY_DTRACE, "breakpoint action");
6966 #endif
6967 }
6968 
6969 static void
6970 dtrace_action_panic(dtrace_ecb_t *ecb)
6971 {
6972 	dtrace_probe_t *probe = ecb->dte_probe;
6973 
6974 	/*
6975 	 * It's impossible to be taking action on the NULL probe.
6976 	 */
6977 	ASSERT(probe != NULL);
6978 
6979 	if (dtrace_destructive_disallow)
6980 		return;
6981 
6982 	if (dtrace_panicked != NULL)
6983 		return;
6984 
6985 	if (dtrace_casptr(&dtrace_panicked, NULL, curthread) != NULL)
6986 		return;
6987 
6988 	/*
6989 	 * We won the right to panic.  (We want to be sure that only one
6990 	 * thread calls panic() from dtrace_probe(), and that panic() is
6991 	 * called exactly once.)
6992 	 */
6993 	dtrace_panic("dtrace: panic action at probe %s:%s:%s:%s (ecb %p)",
6994 	    probe->dtpr_provider->dtpv_name, probe->dtpr_mod,
6995 	    probe->dtpr_func, probe->dtpr_name, (void *)ecb);
6996 }
6997 
6998 static void
6999 dtrace_action_raise(uint64_t sig)
7000 {
7001 	if (dtrace_destructive_disallow)
7002 		return;
7003 
7004 	if (sig >= NSIG) {
7005 		DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
7006 		return;
7007 	}
7008 
7009 #ifdef illumos
7010 	/*
7011 	 * raise() has a queue depth of 1 -- we ignore all subsequent
7012 	 * invocations of the raise() action.
7013 	 */
7014 	if (curthread->t_dtrace_sig == 0)
7015 		curthread->t_dtrace_sig = (uint8_t)sig;
7016 
7017 	curthread->t_sig_check = 1;
7018 	aston(curthread);
7019 #else
7020 	struct proc *p = curproc;
7021 	PROC_LOCK(p);
7022 	kern_psignal(p, sig);
7023 	PROC_UNLOCK(p);
7024 #endif
7025 }
7026 
7027 static void
7028 dtrace_action_stop(void)
7029 {
7030 	if (dtrace_destructive_disallow)
7031 		return;
7032 
7033 #ifdef illumos
7034 	if (!curthread->t_dtrace_stop) {
7035 		curthread->t_dtrace_stop = 1;
7036 		curthread->t_sig_check = 1;
7037 		aston(curthread);
7038 	}
7039 #else
7040 	struct proc *p = curproc;
7041 	PROC_LOCK(p);
7042 	kern_psignal(p, SIGSTOP);
7043 	PROC_UNLOCK(p);
7044 #endif
7045 }
7046 
7047 static void
7048 dtrace_action_chill(dtrace_mstate_t *mstate, hrtime_t val)
7049 {
7050 	hrtime_t now;
7051 	volatile uint16_t *flags;
7052 #ifdef illumos
7053 	cpu_t *cpu = CPU;
7054 #else
7055 	cpu_t *cpu = &solaris_cpu[curcpu];
7056 #endif
7057 
7058 	if (dtrace_destructive_disallow)
7059 		return;
7060 
7061 	flags = (volatile uint16_t *)&cpu_core[curcpu].cpuc_dtrace_flags;
7062 
7063 	now = dtrace_gethrtime();
7064 
7065 	if (now - cpu->cpu_dtrace_chillmark > dtrace_chill_interval) {
7066 		/*
7067 		 * We need to advance the mark to the current time.
7068 		 */
7069 		cpu->cpu_dtrace_chillmark = now;
7070 		cpu->cpu_dtrace_chilled = 0;
7071 	}
7072 
7073 	/*
7074 	 * Now check to see if the requested chill time would take us over
7075 	 * the maximum amount of time allowed in the chill interval.  (Or
7076 	 * worse, if the calculation itself induces overflow.)
7077 	 */
7078 	if (cpu->cpu_dtrace_chilled + val > dtrace_chill_max ||
7079 	    cpu->cpu_dtrace_chilled + val < cpu->cpu_dtrace_chilled) {
7080 		*flags |= CPU_DTRACE_ILLOP;
7081 		return;
7082 	}
7083 
7084 	while (dtrace_gethrtime() - now < val)
7085 		continue;
7086 
7087 	/*
7088 	 * Normally, we assure that the value of the variable "timestamp" does
7089 	 * not change within an ECB.  The presence of chill() represents an
7090 	 * exception to this rule, however.
7091 	 */
7092 	mstate->dtms_present &= ~DTRACE_MSTATE_TIMESTAMP;
7093 	cpu->cpu_dtrace_chilled += val;
7094 }
7095 
7096 static void
7097 dtrace_action_ustack(dtrace_mstate_t *mstate, dtrace_state_t *state,
7098     uint64_t *buf, uint64_t arg)
7099 {
7100 	int nframes = DTRACE_USTACK_NFRAMES(arg);
7101 	int strsize = DTRACE_USTACK_STRSIZE(arg);
7102 	uint64_t *pcs = &buf[1], *fps;
7103 	char *str = (char *)&pcs[nframes];
7104 	int size, offs = 0, i, j;
7105 	size_t rem;
7106 	uintptr_t old = mstate->dtms_scratch_ptr, saved;
7107 	uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
7108 	char *sym;
7109 
7110 	/*
7111 	 * Should be taking a faster path if string space has not been
7112 	 * allocated.
7113 	 */
7114 	ASSERT(strsize != 0);
7115 
7116 	/*
7117 	 * We will first allocate some temporary space for the frame pointers.
7118 	 */
7119 	fps = (uint64_t *)P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
7120 	size = (uintptr_t)fps - mstate->dtms_scratch_ptr +
7121 	    (nframes * sizeof (uint64_t));
7122 
7123 	if (!DTRACE_INSCRATCH(mstate, size)) {
7124 		/*
7125 		 * Not enough room for our frame pointers -- need to indicate
7126 		 * that we ran out of scratch space.
7127 		 */
7128 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
7129 		return;
7130 	}
7131 
7132 	mstate->dtms_scratch_ptr += size;
7133 	saved = mstate->dtms_scratch_ptr;
7134 
7135 	/*
7136 	 * Now get a stack with both program counters and frame pointers.
7137 	 */
7138 	DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
7139 	dtrace_getufpstack(buf, fps, nframes + 1);
7140 	DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
7141 
7142 	/*
7143 	 * If that faulted, we're cooked.
7144 	 */
7145 	if (*flags & CPU_DTRACE_FAULT)
7146 		goto out;
7147 
7148 	/*
7149 	 * Now we want to walk up the stack, calling the USTACK helper.  For
7150 	 * each iteration, we restore the scratch pointer.
7151 	 */
7152 	for (i = 0; i < nframes; i++) {
7153 		mstate->dtms_scratch_ptr = saved;
7154 
7155 		if (offs >= strsize)
7156 			break;
7157 
7158 		sym = (char *)(uintptr_t)dtrace_helper(
7159 		    DTRACE_HELPER_ACTION_USTACK,
7160 		    mstate, state, pcs[i], fps[i]);
7161 
7162 		/*
7163 		 * If we faulted while running the helper, we're going to
7164 		 * clear the fault and null out the corresponding string.
7165 		 */
7166 		if (*flags & CPU_DTRACE_FAULT) {
7167 			*flags &= ~CPU_DTRACE_FAULT;
7168 			str[offs++] = '\0';
7169 			continue;
7170 		}
7171 
7172 		if (sym == NULL) {
7173 			str[offs++] = '\0';
7174 			continue;
7175 		}
7176 
7177 		if (!dtrace_strcanload((uintptr_t)sym, strsize, &rem, mstate,
7178 		    &(state->dts_vstate))) {
7179 			str[offs++] = '\0';
7180 			continue;
7181 		}
7182 
7183 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
7184 
7185 		/*
7186 		 * Now copy in the string that the helper returned to us.
7187 		 */
7188 		for (j = 0; offs + j < strsize && j < rem; j++) {
7189 			if ((str[offs + j] = sym[j]) == '\0')
7190 				break;
7191 		}
7192 
7193 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
7194 
7195 		offs += j + 1;
7196 	}
7197 
7198 	if (offs >= strsize) {
7199 		/*
7200 		 * If we didn't have room for all of the strings, we don't
7201 		 * abort processing -- this needn't be a fatal error -- but we
7202 		 * still want to increment a counter (dts_stkstroverflows) to
7203 		 * allow this condition to be warned about.  (If this is from
7204 		 * a jstack() action, it is easily tuned via jstackstrsize.)
7205 		 */
7206 		dtrace_error(&state->dts_stkstroverflows);
7207 	}
7208 
7209 	while (offs < strsize)
7210 		str[offs++] = '\0';
7211 
7212 out:
7213 	mstate->dtms_scratch_ptr = old;
7214 }
7215 
7216 static void
7217 dtrace_store_by_ref(dtrace_difo_t *dp, caddr_t tomax, size_t size,
7218     size_t *valoffsp, uint64_t *valp, uint64_t end, int intuple, int dtkind)
7219 {
7220 	volatile uint16_t *flags;
7221 	uint64_t val = *valp;
7222 	size_t valoffs = *valoffsp;
7223 
7224 	flags = (volatile uint16_t *)&cpu_core[curcpu].cpuc_dtrace_flags;
7225 	ASSERT(dtkind == DIF_TF_BYREF || dtkind == DIF_TF_BYUREF);
7226 
7227 	/*
7228 	 * If this is a string, we're going to only load until we find the zero
7229 	 * byte -- after which we'll store zero bytes.
7230 	 */
7231 	if (dp->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING) {
7232 		char c = '\0' + 1;
7233 		size_t s;
7234 
7235 		for (s = 0; s < size; s++) {
7236 			if (c != '\0' && dtkind == DIF_TF_BYREF) {
7237 				c = dtrace_load8(val++);
7238 			} else if (c != '\0' && dtkind == DIF_TF_BYUREF) {
7239 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
7240 				c = dtrace_fuword8((void *)(uintptr_t)val++);
7241 				DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
7242 				if (*flags & CPU_DTRACE_FAULT)
7243 					break;
7244 			}
7245 
7246 			DTRACE_STORE(uint8_t, tomax, valoffs++, c);
7247 
7248 			if (c == '\0' && intuple)
7249 				break;
7250 		}
7251 	} else {
7252 		uint8_t c;
7253 		while (valoffs < end) {
7254 			if (dtkind == DIF_TF_BYREF) {
7255 				c = dtrace_load8(val++);
7256 			} else if (dtkind == DIF_TF_BYUREF) {
7257 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
7258 				c = dtrace_fuword8((void *)(uintptr_t)val++);
7259 				DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
7260 				if (*flags & CPU_DTRACE_FAULT)
7261 					break;
7262 			}
7263 
7264 			DTRACE_STORE(uint8_t, tomax,
7265 			    valoffs++, c);
7266 		}
7267 	}
7268 
7269 	*valp = val;
7270 	*valoffsp = valoffs;
7271 }
7272 
7273 /*
7274  * Disables interrupts and sets the per-thread inprobe flag. When DEBUG is
7275  * defined, we also assert that we are not recursing unless the probe ID is an
7276  * error probe.
7277  */
7278 static dtrace_icookie_t
7279 dtrace_probe_enter(dtrace_id_t id)
7280 {
7281 	dtrace_icookie_t cookie;
7282 
7283 	cookie = dtrace_interrupt_disable();
7284 
7285 	/*
7286 	 * Unless this is an ERROR probe, we are not allowed to recurse in
7287 	 * dtrace_probe(). Recursing into DTrace probe usually means that a
7288 	 * function is instrumented that should not have been instrumented or
7289 	 * that the ordering guarantee of the records will be violated,
7290 	 * resulting in unexpected output. If there is an exception to this
7291 	 * assertion, a new case should be added.
7292 	 */
7293 	ASSERT(curthread->t_dtrace_inprobe == 0 ||
7294 	    id == dtrace_probeid_error);
7295 	curthread->t_dtrace_inprobe = 1;
7296 
7297 	return (cookie);
7298 }
7299 
7300 /*
7301  * Clears the per-thread inprobe flag and enables interrupts.
7302  */
7303 static void
7304 dtrace_probe_exit(dtrace_icookie_t cookie)
7305 {
7306 
7307 	curthread->t_dtrace_inprobe = 0;
7308 	dtrace_interrupt_enable(cookie);
7309 }
7310 
7311 /*
7312  * If you're looking for the epicenter of DTrace, you just found it.  This
7313  * is the function called by the provider to fire a probe -- from which all
7314  * subsequent probe-context DTrace activity emanates.
7315  */
7316 void
7317 dtrace_probe(dtrace_id_t id, uintptr_t arg0, uintptr_t arg1,
7318     uintptr_t arg2, uintptr_t arg3, uintptr_t arg4)
7319 {
7320 	processorid_t cpuid;
7321 	dtrace_icookie_t cookie;
7322 	dtrace_probe_t *probe;
7323 	dtrace_mstate_t mstate;
7324 	dtrace_ecb_t *ecb;
7325 	dtrace_action_t *act;
7326 	intptr_t offs;
7327 	size_t size;
7328 	int vtime, onintr;
7329 	volatile uint16_t *flags;
7330 	hrtime_t now;
7331 
7332 	if (KERNEL_PANICKED())
7333 		return;
7334 
7335 #ifdef illumos
7336 	/*
7337 	 * Kick out immediately if this CPU is still being born (in which case
7338 	 * curthread will be set to -1) or the current thread can't allow
7339 	 * probes in its current context.
7340 	 */
7341 	if (((uintptr_t)curthread & 1) || (curthread->t_flag & T_DONTDTRACE))
7342 		return;
7343 #endif
7344 
7345 	cookie = dtrace_probe_enter(id);
7346 	probe = dtrace_probes[id - 1];
7347 	cpuid = curcpu;
7348 	onintr = CPU_ON_INTR(CPU);
7349 
7350 	if (!onintr && probe->dtpr_predcache != DTRACE_CACHEIDNONE &&
7351 	    probe->dtpr_predcache == curthread->t_predcache) {
7352 		/*
7353 		 * We have hit in the predicate cache; we know that
7354 		 * this predicate would evaluate to be false.
7355 		 */
7356 		dtrace_probe_exit(cookie);
7357 		return;
7358 	}
7359 
7360 #ifdef illumos
7361 	if (panic_quiesce) {
7362 #else
7363 	if (KERNEL_PANICKED()) {
7364 #endif
7365 		/*
7366 		 * We don't trace anything if we're panicking.
7367 		 */
7368 		dtrace_probe_exit(cookie);
7369 		return;
7370 	}
7371 
7372 	now = mstate.dtms_timestamp = dtrace_gethrtime();
7373 	mstate.dtms_present = DTRACE_MSTATE_TIMESTAMP;
7374 	vtime = dtrace_vtime_references != 0;
7375 
7376 	if (vtime && curthread->t_dtrace_start)
7377 		curthread->t_dtrace_vtime += now - curthread->t_dtrace_start;
7378 
7379 	mstate.dtms_difo = NULL;
7380 	mstate.dtms_probe = probe;
7381 	mstate.dtms_strtok = 0;
7382 	mstate.dtms_arg[0] = arg0;
7383 	mstate.dtms_arg[1] = arg1;
7384 	mstate.dtms_arg[2] = arg2;
7385 	mstate.dtms_arg[3] = arg3;
7386 	mstate.dtms_arg[4] = arg4;
7387 
7388 	flags = (volatile uint16_t *)&cpu_core[cpuid].cpuc_dtrace_flags;
7389 
7390 	for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
7391 		dtrace_predicate_t *pred = ecb->dte_predicate;
7392 		dtrace_state_t *state = ecb->dte_state;
7393 		dtrace_buffer_t *buf = &state->dts_buffer[cpuid];
7394 		dtrace_buffer_t *aggbuf = &state->dts_aggbuffer[cpuid];
7395 		dtrace_vstate_t *vstate = &state->dts_vstate;
7396 		dtrace_provider_t *prov = probe->dtpr_provider;
7397 		uint64_t tracememsize = 0;
7398 		int committed = 0;
7399 		caddr_t tomax;
7400 
7401 		/*
7402 		 * A little subtlety with the following (seemingly innocuous)
7403 		 * declaration of the automatic 'val':  by looking at the
7404 		 * code, you might think that it could be declared in the
7405 		 * action processing loop, below.  (That is, it's only used in
7406 		 * the action processing loop.)  However, it must be declared
7407 		 * out of that scope because in the case of DIF expression
7408 		 * arguments to aggregating actions, one iteration of the
7409 		 * action loop will use the last iteration's value.
7410 		 */
7411 		uint64_t val = 0;
7412 
7413 		mstate.dtms_present = DTRACE_MSTATE_ARGS | DTRACE_MSTATE_PROBE;
7414 		mstate.dtms_getf = NULL;
7415 
7416 		*flags &= ~CPU_DTRACE_ERROR;
7417 
7418 		if (prov == dtrace_provider) {
7419 			/*
7420 			 * If dtrace itself is the provider of this probe,
7421 			 * we're only going to continue processing the ECB if
7422 			 * arg0 (the dtrace_state_t) is equal to the ECB's
7423 			 * creating state.  (This prevents disjoint consumers
7424 			 * from seeing one another's metaprobes.)
7425 			 */
7426 			if (arg0 != (uint64_t)(uintptr_t)state)
7427 				continue;
7428 		}
7429 
7430 		if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE) {
7431 			/*
7432 			 * We're not currently active.  If our provider isn't
7433 			 * the dtrace pseudo provider, we're not interested.
7434 			 */
7435 			if (prov != dtrace_provider)
7436 				continue;
7437 
7438 			/*
7439 			 * Now we must further check if we are in the BEGIN
7440 			 * probe.  If we are, we will only continue processing
7441 			 * if we're still in WARMUP -- if one BEGIN enabling
7442 			 * has invoked the exit() action, we don't want to
7443 			 * evaluate subsequent BEGIN enablings.
7444 			 */
7445 			if (probe->dtpr_id == dtrace_probeid_begin &&
7446 			    state->dts_activity != DTRACE_ACTIVITY_WARMUP) {
7447 				ASSERT(state->dts_activity ==
7448 				    DTRACE_ACTIVITY_DRAINING);
7449 				continue;
7450 			}
7451 		}
7452 
7453 		if (ecb->dte_cond) {
7454 			/*
7455 			 * If the dte_cond bits indicate that this
7456 			 * consumer is only allowed to see user-mode firings
7457 			 * of this probe, call the provider's dtps_usermode()
7458 			 * entry point to check that the probe was fired
7459 			 * while in a user context. Skip this ECB if that's
7460 			 * not the case.
7461 			 */
7462 			if ((ecb->dte_cond & DTRACE_COND_USERMODE) &&
7463 			    prov->dtpv_pops.dtps_usermode(prov->dtpv_arg,
7464 			    probe->dtpr_id, probe->dtpr_arg) == 0)
7465 				continue;
7466 
7467 #ifdef illumos
7468 			/*
7469 			 * This is more subtle than it looks. We have to be
7470 			 * absolutely certain that CRED() isn't going to
7471 			 * change out from under us so it's only legit to
7472 			 * examine that structure if we're in constrained
7473 			 * situations. Currently, the only times we'll this
7474 			 * check is if a non-super-user has enabled the
7475 			 * profile or syscall providers -- providers that
7476 			 * allow visibility of all processes. For the
7477 			 * profile case, the check above will ensure that
7478 			 * we're examining a user context.
7479 			 */
7480 			if (ecb->dte_cond & DTRACE_COND_OWNER) {
7481 				cred_t *cr;
7482 				cred_t *s_cr =
7483 				    ecb->dte_state->dts_cred.dcr_cred;
7484 				proc_t *proc;
7485 
7486 				ASSERT(s_cr != NULL);
7487 
7488 				if ((cr = CRED()) == NULL ||
7489 				    s_cr->cr_uid != cr->cr_uid ||
7490 				    s_cr->cr_uid != cr->cr_ruid ||
7491 				    s_cr->cr_uid != cr->cr_suid ||
7492 				    s_cr->cr_gid != cr->cr_gid ||
7493 				    s_cr->cr_gid != cr->cr_rgid ||
7494 				    s_cr->cr_gid != cr->cr_sgid ||
7495 				    (proc = ttoproc(curthread)) == NULL ||
7496 				    (proc->p_flag & SNOCD))
7497 					continue;
7498 			}
7499 
7500 			if (ecb->dte_cond & DTRACE_COND_ZONEOWNER) {
7501 				cred_t *cr;
7502 				cred_t *s_cr =
7503 				    ecb->dte_state->dts_cred.dcr_cred;
7504 
7505 				ASSERT(s_cr != NULL);
7506 
7507 				if ((cr = CRED()) == NULL ||
7508 				    s_cr->cr_zone->zone_id !=
7509 				    cr->cr_zone->zone_id)
7510 					continue;
7511 			}
7512 #endif
7513 		}
7514 
7515 		if (now - state->dts_alive > dtrace_deadman_timeout) {
7516 			/*
7517 			 * We seem to be dead.  Unless we (a) have kernel
7518 			 * destructive permissions (b) have explicitly enabled
7519 			 * destructive actions and (c) destructive actions have
7520 			 * not been disabled, we're going to transition into
7521 			 * the KILLED state, from which no further processing
7522 			 * on this state will be performed.
7523 			 */
7524 			if (!dtrace_priv_kernel_destructive(state) ||
7525 			    !state->dts_cred.dcr_destructive ||
7526 			    dtrace_destructive_disallow) {
7527 				void *activity = &state->dts_activity;
7528 				dtrace_activity_t curstate;
7529 
7530 				do {
7531 					curstate = state->dts_activity;
7532 				} while (dtrace_cas32(activity, curstate,
7533 				    DTRACE_ACTIVITY_KILLED) != curstate);
7534 
7535 				continue;
7536 			}
7537 		}
7538 
7539 		if ((offs = dtrace_buffer_reserve(buf, ecb->dte_needed,
7540 		    ecb->dte_alignment, state, &mstate)) < 0)
7541 			continue;
7542 
7543 		tomax = buf->dtb_tomax;
7544 		ASSERT(tomax != NULL);
7545 
7546 		if (ecb->dte_size != 0) {
7547 			dtrace_rechdr_t dtrh;
7548 			if (!(mstate.dtms_present & DTRACE_MSTATE_TIMESTAMP)) {
7549 				mstate.dtms_timestamp = dtrace_gethrtime();
7550 				mstate.dtms_present |= DTRACE_MSTATE_TIMESTAMP;
7551 			}
7552 			ASSERT3U(ecb->dte_size, >=, sizeof (dtrace_rechdr_t));
7553 			dtrh.dtrh_epid = ecb->dte_epid;
7554 			DTRACE_RECORD_STORE_TIMESTAMP(&dtrh,
7555 			    mstate.dtms_timestamp);
7556 			*((dtrace_rechdr_t *)(tomax + offs)) = dtrh;
7557 		}
7558 
7559 		mstate.dtms_epid = ecb->dte_epid;
7560 		mstate.dtms_present |= DTRACE_MSTATE_EPID;
7561 
7562 		if (state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)
7563 			mstate.dtms_access = DTRACE_ACCESS_KERNEL;
7564 		else
7565 			mstate.dtms_access = 0;
7566 
7567 		if (pred != NULL) {
7568 			dtrace_difo_t *dp = pred->dtp_difo;
7569 			uint64_t rval;
7570 
7571 			rval = dtrace_dif_emulate(dp, &mstate, vstate, state);
7572 
7573 			if (!(*flags & CPU_DTRACE_ERROR) && !rval) {
7574 				dtrace_cacheid_t cid = probe->dtpr_predcache;
7575 
7576 				if (cid != DTRACE_CACHEIDNONE && !onintr) {
7577 					/*
7578 					 * Update the predicate cache...
7579 					 */
7580 					ASSERT(cid == pred->dtp_cacheid);
7581 					curthread->t_predcache = cid;
7582 				}
7583 
7584 				continue;
7585 			}
7586 		}
7587 
7588 		for (act = ecb->dte_action; !(*flags & CPU_DTRACE_ERROR) &&
7589 		    act != NULL; act = act->dta_next) {
7590 			size_t valoffs;
7591 			dtrace_difo_t *dp;
7592 			dtrace_recdesc_t *rec = &act->dta_rec;
7593 
7594 			size = rec->dtrd_size;
7595 			valoffs = offs + rec->dtrd_offset;
7596 
7597 			if (DTRACEACT_ISAGG(act->dta_kind)) {
7598 				uint64_t v = 0xbad;
7599 				dtrace_aggregation_t *agg;
7600 
7601 				agg = (dtrace_aggregation_t *)act;
7602 
7603 				if ((dp = act->dta_difo) != NULL)
7604 					v = dtrace_dif_emulate(dp,
7605 					    &mstate, vstate, state);
7606 
7607 				if (*flags & CPU_DTRACE_ERROR)
7608 					continue;
7609 
7610 				/*
7611 				 * Note that we always pass the expression
7612 				 * value from the previous iteration of the
7613 				 * action loop.  This value will only be used
7614 				 * if there is an expression argument to the
7615 				 * aggregating action, denoted by the
7616 				 * dtag_hasarg field.
7617 				 */
7618 				dtrace_aggregate(agg, buf,
7619 				    offs, aggbuf, v, val);
7620 				continue;
7621 			}
7622 
7623 			switch (act->dta_kind) {
7624 			case DTRACEACT_STOP:
7625 				if (dtrace_priv_proc_destructive(state))
7626 					dtrace_action_stop();
7627 				continue;
7628 
7629 			case DTRACEACT_BREAKPOINT:
7630 				if (dtrace_priv_kernel_destructive(state))
7631 					dtrace_action_breakpoint(ecb);
7632 				continue;
7633 
7634 			case DTRACEACT_PANIC:
7635 				if (dtrace_priv_kernel_destructive(state))
7636 					dtrace_action_panic(ecb);
7637 				continue;
7638 
7639 			case DTRACEACT_STACK:
7640 				if (!dtrace_priv_kernel(state))
7641 					continue;
7642 
7643 				dtrace_getpcstack((pc_t *)(tomax + valoffs),
7644 				    size / sizeof (pc_t), probe->dtpr_aframes,
7645 				    DTRACE_ANCHORED(probe) ? NULL :
7646 				    (uint32_t *)arg0);
7647 				continue;
7648 
7649 			case DTRACEACT_JSTACK:
7650 			case DTRACEACT_USTACK:
7651 				if (!dtrace_priv_proc(state))
7652 					continue;
7653 
7654 				/*
7655 				 * See comment in DIF_VAR_PID.
7656 				 */
7657 				if (DTRACE_ANCHORED(mstate.dtms_probe) &&
7658 				    CPU_ON_INTR(CPU)) {
7659 					int depth = DTRACE_USTACK_NFRAMES(
7660 					    rec->dtrd_arg) + 1;
7661 
7662 					dtrace_bzero((void *)(tomax + valoffs),
7663 					    DTRACE_USTACK_STRSIZE(rec->dtrd_arg)
7664 					    + depth * sizeof (uint64_t));
7665 
7666 					continue;
7667 				}
7668 
7669 				if (DTRACE_USTACK_STRSIZE(rec->dtrd_arg) != 0 &&
7670 				    curproc->p_dtrace_helpers != NULL) {
7671 					/*
7672 					 * This is the slow path -- we have
7673 					 * allocated string space, and we're
7674 					 * getting the stack of a process that
7675 					 * has helpers.  Call into a separate
7676 					 * routine to perform this processing.
7677 					 */
7678 					dtrace_action_ustack(&mstate, state,
7679 					    (uint64_t *)(tomax + valoffs),
7680 					    rec->dtrd_arg);
7681 					continue;
7682 				}
7683 
7684 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
7685 				dtrace_getupcstack((uint64_t *)
7686 				    (tomax + valoffs),
7687 				    DTRACE_USTACK_NFRAMES(rec->dtrd_arg) + 1);
7688 				DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
7689 				continue;
7690 
7691 			default:
7692 				break;
7693 			}
7694 
7695 			dp = act->dta_difo;
7696 			ASSERT(dp != NULL);
7697 
7698 			val = dtrace_dif_emulate(dp, &mstate, vstate, state);
7699 
7700 			if (*flags & CPU_DTRACE_ERROR)
7701 				continue;
7702 
7703 			switch (act->dta_kind) {
7704 			case DTRACEACT_SPECULATE: {
7705 				dtrace_rechdr_t *dtrh;
7706 
7707 				ASSERT(buf == &state->dts_buffer[cpuid]);
7708 				buf = dtrace_speculation_buffer(state,
7709 				    cpuid, val);
7710 
7711 				if (buf == NULL) {
7712 					*flags |= CPU_DTRACE_DROP;
7713 					continue;
7714 				}
7715 
7716 				offs = dtrace_buffer_reserve(buf,
7717 				    ecb->dte_needed, ecb->dte_alignment,
7718 				    state, NULL);
7719 
7720 				if (offs < 0) {
7721 					*flags |= CPU_DTRACE_DROP;
7722 					continue;
7723 				}
7724 
7725 				tomax = buf->dtb_tomax;
7726 				ASSERT(tomax != NULL);
7727 
7728 				if (ecb->dte_size == 0)
7729 					continue;
7730 
7731 				ASSERT3U(ecb->dte_size, >=,
7732 				    sizeof (dtrace_rechdr_t));
7733 				dtrh = ((void *)(tomax + offs));
7734 				dtrh->dtrh_epid = ecb->dte_epid;
7735 				/*
7736 				 * When the speculation is committed, all of
7737 				 * the records in the speculative buffer will
7738 				 * have their timestamps set to the commit
7739 				 * time.  Until then, it is set to a sentinel
7740 				 * value, for debugability.
7741 				 */
7742 				DTRACE_RECORD_STORE_TIMESTAMP(dtrh, UINT64_MAX);
7743 				continue;
7744 			}
7745 
7746 			case DTRACEACT_PRINTM: {
7747 				/*
7748 				 * printm() assumes that the DIF returns a
7749 				 * pointer returned by memref(). memref() is a
7750 				 * subroutine that is used to get around the
7751 				 * single-valued returns of DIF and is assumed
7752 				 * to always be allocated in the scratch space.
7753 				 * Therefore, we need to validate that the
7754 				 * pointer given to printm() is in the scratch
7755 				 * space in order to avoid a potential panic.
7756 				 */
7757 				uintptr_t *memref = (uintptr_t *)(uintptr_t) val;
7758 
7759 				if (!DTRACE_INSCRATCHPTR(&mstate,
7760 				    (uintptr_t)memref, 2 * sizeof(uintptr_t))) {
7761 					*flags |= CPU_DTRACE_BADADDR;
7762 					continue;
7763 				}
7764 
7765 				/* Get the size from the memref. */
7766 				size = memref[1];
7767 
7768 				/*
7769 				 * Check if the size exceeds the allocated
7770 				 * buffer size.
7771 				 */
7772 				if (size + sizeof(uintptr_t) > dp->dtdo_rtype.dtdt_size) {
7773 					/* Flag a drop! */
7774 					*flags |= CPU_DTRACE_DROP;
7775 					continue;
7776 				}
7777 
7778 				/* Store the size in the buffer first. */
7779 				DTRACE_STORE(uintptr_t, tomax,
7780 				    valoffs, size);
7781 
7782 				/*
7783 				 * Offset the buffer address to the start
7784 				 * of the data.
7785 				 */
7786 				valoffs += sizeof(uintptr_t);
7787 
7788 				/*
7789 				 * Reset to the memory address rather than
7790 				 * the memref array, then let the BYREF
7791 				 * code below do the work to store the
7792 				 * memory data in the buffer.
7793 				 */
7794 				val = memref[0];
7795 				break;
7796 			}
7797 
7798 			case DTRACEACT_CHILL:
7799 				if (dtrace_priv_kernel_destructive(state))
7800 					dtrace_action_chill(&mstate, val);
7801 				continue;
7802 
7803 			case DTRACEACT_RAISE:
7804 				if (dtrace_priv_proc_destructive(state))
7805 					dtrace_action_raise(val);
7806 				continue;
7807 
7808 			case DTRACEACT_COMMIT:
7809 				ASSERT(!committed);
7810 
7811 				/*
7812 				 * We need to commit our buffer state.
7813 				 */
7814 				if (ecb->dte_size)
7815 					buf->dtb_offset = offs + ecb->dte_size;
7816 				buf = &state->dts_buffer[cpuid];
7817 				dtrace_speculation_commit(state, cpuid, val);
7818 				committed = 1;
7819 				continue;
7820 
7821 			case DTRACEACT_DISCARD:
7822 				dtrace_speculation_discard(state, cpuid, val);
7823 				continue;
7824 
7825 			case DTRACEACT_DIFEXPR:
7826 			case DTRACEACT_LIBACT:
7827 			case DTRACEACT_PRINTF:
7828 			case DTRACEACT_PRINTA:
7829 			case DTRACEACT_SYSTEM:
7830 			case DTRACEACT_FREOPEN:
7831 			case DTRACEACT_TRACEMEM:
7832 				break;
7833 
7834 			case DTRACEACT_TRACEMEM_DYNSIZE:
7835 				tracememsize = val;
7836 				break;
7837 
7838 			case DTRACEACT_SYM:
7839 			case DTRACEACT_MOD:
7840 				if (!dtrace_priv_kernel(state))
7841 					continue;
7842 				break;
7843 
7844 			case DTRACEACT_USYM:
7845 			case DTRACEACT_UMOD:
7846 			case DTRACEACT_UADDR: {
7847 #ifdef illumos
7848 				struct pid *pid = curthread->t_procp->p_pidp;
7849 #endif
7850 
7851 				if (!dtrace_priv_proc(state))
7852 					continue;
7853 
7854 				DTRACE_STORE(uint64_t, tomax,
7855 #ifdef illumos
7856 				    valoffs, (uint64_t)pid->pid_id);
7857 #else
7858 				    valoffs, (uint64_t) curproc->p_pid);
7859 #endif
7860 				DTRACE_STORE(uint64_t, tomax,
7861 				    valoffs + sizeof (uint64_t), val);
7862 
7863 				continue;
7864 			}
7865 
7866 			case DTRACEACT_EXIT: {
7867 				/*
7868 				 * For the exit action, we are going to attempt
7869 				 * to atomically set our activity to be
7870 				 * draining.  If this fails (either because
7871 				 * another CPU has beat us to the exit action,
7872 				 * or because our current activity is something
7873 				 * other than ACTIVE or WARMUP), we will
7874 				 * continue.  This assures that the exit action
7875 				 * can be successfully recorded at most once
7876 				 * when we're in the ACTIVE state.  If we're
7877 				 * encountering the exit() action while in
7878 				 * COOLDOWN, however, we want to honor the new
7879 				 * status code.  (We know that we're the only
7880 				 * thread in COOLDOWN, so there is no race.)
7881 				 */
7882 				void *activity = &state->dts_activity;
7883 				dtrace_activity_t curstate = state->dts_activity;
7884 
7885 				if (curstate == DTRACE_ACTIVITY_COOLDOWN)
7886 					break;
7887 
7888 				if (curstate != DTRACE_ACTIVITY_WARMUP)
7889 					curstate = DTRACE_ACTIVITY_ACTIVE;
7890 
7891 				if (dtrace_cas32(activity, curstate,
7892 				    DTRACE_ACTIVITY_DRAINING) != curstate) {
7893 					*flags |= CPU_DTRACE_DROP;
7894 					continue;
7895 				}
7896 
7897 				break;
7898 			}
7899 
7900 			default:
7901 				ASSERT(0);
7902 			}
7903 
7904 			if (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF ||
7905 			    dp->dtdo_rtype.dtdt_flags & DIF_TF_BYUREF) {
7906 				uintptr_t end = valoffs + size;
7907 
7908 				if (tracememsize != 0 &&
7909 				    valoffs + tracememsize < end) {
7910 					end = valoffs + tracememsize;
7911 					tracememsize = 0;
7912 				}
7913 
7914 				if (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF &&
7915 				    !dtrace_vcanload((void *)(uintptr_t)val,
7916 				    &dp->dtdo_rtype, NULL, &mstate, vstate))
7917 					continue;
7918 
7919 				dtrace_store_by_ref(dp, tomax, size, &valoffs,
7920 				    &val, end, act->dta_intuple,
7921 				    dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF ?
7922 				    DIF_TF_BYREF: DIF_TF_BYUREF);
7923 				continue;
7924 			}
7925 
7926 			switch (size) {
7927 			case 0:
7928 				break;
7929 
7930 			case sizeof (uint8_t):
7931 				DTRACE_STORE(uint8_t, tomax, valoffs, val);
7932 				break;
7933 			case sizeof (uint16_t):
7934 				DTRACE_STORE(uint16_t, tomax, valoffs, val);
7935 				break;
7936 			case sizeof (uint32_t):
7937 				DTRACE_STORE(uint32_t, tomax, valoffs, val);
7938 				break;
7939 			case sizeof (uint64_t):
7940 				DTRACE_STORE(uint64_t, tomax, valoffs, val);
7941 				break;
7942 			default:
7943 				/*
7944 				 * Any other size should have been returned by
7945 				 * reference, not by value.
7946 				 */
7947 				ASSERT(0);
7948 				break;
7949 			}
7950 		}
7951 
7952 		if (*flags & CPU_DTRACE_DROP)
7953 			continue;
7954 
7955 		if (*flags & CPU_DTRACE_FAULT) {
7956 			int ndx;
7957 			dtrace_action_t *err;
7958 
7959 			buf->dtb_errors++;
7960 
7961 			if (probe->dtpr_id == dtrace_probeid_error) {
7962 				/*
7963 				 * There's nothing we can do -- we had an
7964 				 * error on the error probe.  We bump an
7965 				 * error counter to at least indicate that
7966 				 * this condition happened.
7967 				 */
7968 				dtrace_error(&state->dts_dblerrors);
7969 				continue;
7970 			}
7971 
7972 			if (vtime) {
7973 				/*
7974 				 * Before recursing on dtrace_probe(), we
7975 				 * need to explicitly clear out our start
7976 				 * time to prevent it from being accumulated
7977 				 * into t_dtrace_vtime.
7978 				 */
7979 				curthread->t_dtrace_start = 0;
7980 			}
7981 
7982 			/*
7983 			 * Iterate over the actions to figure out which action
7984 			 * we were processing when we experienced the error.
7985 			 * Note that act points _past_ the faulting action; if
7986 			 * act is ecb->dte_action, the fault was in the
7987 			 * predicate, if it's ecb->dte_action->dta_next it's
7988 			 * in action #1, and so on.
7989 			 */
7990 			for (err = ecb->dte_action, ndx = 0;
7991 			    err != act; err = err->dta_next, ndx++)
7992 				continue;
7993 
7994 			dtrace_probe_error(state, ecb->dte_epid, ndx,
7995 			    (mstate.dtms_present & DTRACE_MSTATE_FLTOFFS) ?
7996 			    mstate.dtms_fltoffs : -1, DTRACE_FLAGS2FLT(*flags),
7997 			    cpu_core[cpuid].cpuc_dtrace_illval);
7998 
7999 			continue;
8000 		}
8001 
8002 		if (!committed)
8003 			buf->dtb_offset = offs + ecb->dte_size;
8004 	}
8005 
8006 	if (vtime)
8007 		curthread->t_dtrace_start = dtrace_gethrtime();
8008 
8009 	dtrace_probe_exit(cookie);
8010 }
8011 
8012 /*
8013  * DTrace Probe Hashing Functions
8014  *
8015  * The functions in this section (and indeed, the functions in remaining
8016  * sections) are not _called_ from probe context.  (Any exceptions to this are
8017  * marked with a "Note:".)  Rather, they are called from elsewhere in the
8018  * DTrace framework to look-up probes in, add probes to and remove probes from
8019  * the DTrace probe hashes.  (Each probe is hashed by each element of the
8020  * probe tuple -- allowing for fast lookups, regardless of what was
8021  * specified.)
8022  */
8023 static uint_t
8024 dtrace_hash_str(const char *p)
8025 {
8026 	unsigned int g;
8027 	uint_t hval = 0;
8028 
8029 	while (*p) {
8030 		hval = (hval << 4) + *p++;
8031 		if ((g = (hval & 0xf0000000)) != 0)
8032 			hval ^= g >> 24;
8033 		hval &= ~g;
8034 	}
8035 	return (hval);
8036 }
8037 
8038 static dtrace_hash_t *
8039 dtrace_hash_create(uintptr_t stroffs, uintptr_t nextoffs, uintptr_t prevoffs)
8040 {
8041 	dtrace_hash_t *hash = kmem_zalloc(sizeof (dtrace_hash_t), KM_SLEEP);
8042 
8043 	hash->dth_stroffs = stroffs;
8044 	hash->dth_nextoffs = nextoffs;
8045 	hash->dth_prevoffs = prevoffs;
8046 
8047 	hash->dth_size = 1;
8048 	hash->dth_mask = hash->dth_size - 1;
8049 
8050 	hash->dth_tab = kmem_zalloc(hash->dth_size *
8051 	    sizeof (dtrace_hashbucket_t *), KM_SLEEP);
8052 
8053 	return (hash);
8054 }
8055 
8056 static void
8057 dtrace_hash_destroy(dtrace_hash_t *hash)
8058 {
8059 #ifdef DEBUG
8060 	int i;
8061 
8062 	for (i = 0; i < hash->dth_size; i++)
8063 		ASSERT(hash->dth_tab[i] == NULL);
8064 #endif
8065 
8066 	kmem_free(hash->dth_tab,
8067 	    hash->dth_size * sizeof (dtrace_hashbucket_t *));
8068 	kmem_free(hash, sizeof (dtrace_hash_t));
8069 }
8070 
8071 static void
8072 dtrace_hash_resize(dtrace_hash_t *hash)
8073 {
8074 	int size = hash->dth_size, i, ndx;
8075 	int new_size = hash->dth_size << 1;
8076 	int new_mask = new_size - 1;
8077 	dtrace_hashbucket_t **new_tab, *bucket, *next;
8078 
8079 	ASSERT((new_size & new_mask) == 0);
8080 
8081 	new_tab = kmem_zalloc(new_size * sizeof (void *), KM_SLEEP);
8082 
8083 	for (i = 0; i < size; i++) {
8084 		for (bucket = hash->dth_tab[i]; bucket != NULL; bucket = next) {
8085 			dtrace_probe_t *probe = bucket->dthb_chain;
8086 
8087 			ASSERT(probe != NULL);
8088 			ndx = DTRACE_HASHSTR(hash, probe) & new_mask;
8089 
8090 			next = bucket->dthb_next;
8091 			bucket->dthb_next = new_tab[ndx];
8092 			new_tab[ndx] = bucket;
8093 		}
8094 	}
8095 
8096 	kmem_free(hash->dth_tab, hash->dth_size * sizeof (void *));
8097 	hash->dth_tab = new_tab;
8098 	hash->dth_size = new_size;
8099 	hash->dth_mask = new_mask;
8100 }
8101 
8102 static void
8103 dtrace_hash_add(dtrace_hash_t *hash, dtrace_probe_t *new)
8104 {
8105 	int hashval = DTRACE_HASHSTR(hash, new);
8106 	int ndx = hashval & hash->dth_mask;
8107 	dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
8108 	dtrace_probe_t **nextp, **prevp;
8109 
8110 	for (; bucket != NULL; bucket = bucket->dthb_next) {
8111 		if (DTRACE_HASHEQ(hash, bucket->dthb_chain, new))
8112 			goto add;
8113 	}
8114 
8115 	if ((hash->dth_nbuckets >> 1) > hash->dth_size) {
8116 		dtrace_hash_resize(hash);
8117 		dtrace_hash_add(hash, new);
8118 		return;
8119 	}
8120 
8121 	bucket = kmem_zalloc(sizeof (dtrace_hashbucket_t), KM_SLEEP);
8122 	bucket->dthb_next = hash->dth_tab[ndx];
8123 	hash->dth_tab[ndx] = bucket;
8124 	hash->dth_nbuckets++;
8125 
8126 add:
8127 	nextp = DTRACE_HASHNEXT(hash, new);
8128 	ASSERT(*nextp == NULL && *(DTRACE_HASHPREV(hash, new)) == NULL);
8129 	*nextp = bucket->dthb_chain;
8130 
8131 	if (bucket->dthb_chain != NULL) {
8132 		prevp = DTRACE_HASHPREV(hash, bucket->dthb_chain);
8133 		ASSERT(*prevp == NULL);
8134 		*prevp = new;
8135 	}
8136 
8137 	bucket->dthb_chain = new;
8138 	bucket->dthb_len++;
8139 }
8140 
8141 static dtrace_probe_t *
8142 dtrace_hash_lookup(dtrace_hash_t *hash, dtrace_probe_t *template)
8143 {
8144 	int hashval = DTRACE_HASHSTR(hash, template);
8145 	int ndx = hashval & hash->dth_mask;
8146 	dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
8147 
8148 	for (; bucket != NULL; bucket = bucket->dthb_next) {
8149 		if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
8150 			return (bucket->dthb_chain);
8151 	}
8152 
8153 	return (NULL);
8154 }
8155 
8156 static int
8157 dtrace_hash_collisions(dtrace_hash_t *hash, dtrace_probe_t *template)
8158 {
8159 	int hashval = DTRACE_HASHSTR(hash, template);
8160 	int ndx = hashval & hash->dth_mask;
8161 	dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
8162 
8163 	for (; bucket != NULL; bucket = bucket->dthb_next) {
8164 		if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
8165 			return (bucket->dthb_len);
8166 	}
8167 
8168 	return (0);
8169 }
8170 
8171 static void
8172 dtrace_hash_remove(dtrace_hash_t *hash, dtrace_probe_t *probe)
8173 {
8174 	int ndx = DTRACE_HASHSTR(hash, probe) & hash->dth_mask;
8175 	dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
8176 
8177 	dtrace_probe_t **prevp = DTRACE_HASHPREV(hash, probe);
8178 	dtrace_probe_t **nextp = DTRACE_HASHNEXT(hash, probe);
8179 
8180 	/*
8181 	 * Find the bucket that we're removing this probe from.
8182 	 */
8183 	for (; bucket != NULL; bucket = bucket->dthb_next) {
8184 		if (DTRACE_HASHEQ(hash, bucket->dthb_chain, probe))
8185 			break;
8186 	}
8187 
8188 	ASSERT(bucket != NULL);
8189 
8190 	if (*prevp == NULL) {
8191 		if (*nextp == NULL) {
8192 			/*
8193 			 * The removed probe was the only probe on this
8194 			 * bucket; we need to remove the bucket.
8195 			 */
8196 			dtrace_hashbucket_t *b = hash->dth_tab[ndx];
8197 
8198 			ASSERT(bucket->dthb_chain == probe);
8199 			ASSERT(b != NULL);
8200 
8201 			if (b == bucket) {
8202 				hash->dth_tab[ndx] = bucket->dthb_next;
8203 			} else {
8204 				while (b->dthb_next != bucket)
8205 					b = b->dthb_next;
8206 				b->dthb_next = bucket->dthb_next;
8207 			}
8208 
8209 			ASSERT(hash->dth_nbuckets > 0);
8210 			hash->dth_nbuckets--;
8211 			kmem_free(bucket, sizeof (dtrace_hashbucket_t));
8212 			return;
8213 		}
8214 
8215 		bucket->dthb_chain = *nextp;
8216 	} else {
8217 		*(DTRACE_HASHNEXT(hash, *prevp)) = *nextp;
8218 	}
8219 
8220 	if (*nextp != NULL)
8221 		*(DTRACE_HASHPREV(hash, *nextp)) = *prevp;
8222 }
8223 
8224 /*
8225  * DTrace Utility Functions
8226  *
8227  * These are random utility functions that are _not_ called from probe context.
8228  */
8229 static int
8230 dtrace_badattr(const dtrace_attribute_t *a)
8231 {
8232 	return (a->dtat_name > DTRACE_STABILITY_MAX ||
8233 	    a->dtat_data > DTRACE_STABILITY_MAX ||
8234 	    a->dtat_class > DTRACE_CLASS_MAX);
8235 }
8236 
8237 /*
8238  * Return a duplicate copy of a string.  If the specified string is NULL,
8239  * this function returns a zero-length string.
8240  */
8241 static char *
8242 dtrace_strdup(const char *str)
8243 {
8244 	char *new = kmem_zalloc((str != NULL ? strlen(str) : 0) + 1, KM_SLEEP);
8245 
8246 	if (str != NULL)
8247 		(void) strcpy(new, str);
8248 
8249 	return (new);
8250 }
8251 
8252 #define	DTRACE_ISALPHA(c)	\
8253 	(((c) >= 'a' && (c) <= 'z') || ((c) >= 'A' && (c) <= 'Z'))
8254 
8255 static int
8256 dtrace_badname(const char *s)
8257 {
8258 	char c;
8259 
8260 	if (s == NULL || (c = *s++) == '\0')
8261 		return (0);
8262 
8263 	if (!DTRACE_ISALPHA(c) && c != '-' && c != '_' && c != '.')
8264 		return (1);
8265 
8266 	while ((c = *s++) != '\0') {
8267 		if (!DTRACE_ISALPHA(c) && (c < '0' || c > '9') &&
8268 		    c != '-' && c != '_' && c != '.' && c != '`')
8269 			return (1);
8270 	}
8271 
8272 	return (0);
8273 }
8274 
8275 static void
8276 dtrace_cred2priv(cred_t *cr, uint32_t *privp, uid_t *uidp, zoneid_t *zoneidp)
8277 {
8278 	uint32_t priv;
8279 
8280 #ifdef illumos
8281 	if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
8282 		/*
8283 		 * For DTRACE_PRIV_ALL, the uid and zoneid don't matter.
8284 		 */
8285 		priv = DTRACE_PRIV_ALL;
8286 	} else {
8287 		*uidp = crgetuid(cr);
8288 		*zoneidp = crgetzoneid(cr);
8289 
8290 		priv = 0;
8291 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE))
8292 			priv |= DTRACE_PRIV_KERNEL | DTRACE_PRIV_USER;
8293 		else if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE))
8294 			priv |= DTRACE_PRIV_USER;
8295 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE))
8296 			priv |= DTRACE_PRIV_PROC;
8297 		if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
8298 			priv |= DTRACE_PRIV_OWNER;
8299 		if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
8300 			priv |= DTRACE_PRIV_ZONEOWNER;
8301 	}
8302 #else
8303 	priv = DTRACE_PRIV_ALL;
8304 #endif
8305 
8306 	*privp = priv;
8307 }
8308 
8309 #ifdef DTRACE_ERRDEBUG
8310 static void
8311 dtrace_errdebug(const char *str)
8312 {
8313 	int hval = dtrace_hash_str(str) % DTRACE_ERRHASHSZ;
8314 	int occupied = 0;
8315 
8316 	mutex_enter(&dtrace_errlock);
8317 	dtrace_errlast = str;
8318 	dtrace_errthread = curthread;
8319 
8320 	while (occupied++ < DTRACE_ERRHASHSZ) {
8321 		if (dtrace_errhash[hval].dter_msg == str) {
8322 			dtrace_errhash[hval].dter_count++;
8323 			goto out;
8324 		}
8325 
8326 		if (dtrace_errhash[hval].dter_msg != NULL) {
8327 			hval = (hval + 1) % DTRACE_ERRHASHSZ;
8328 			continue;
8329 		}
8330 
8331 		dtrace_errhash[hval].dter_msg = str;
8332 		dtrace_errhash[hval].dter_count = 1;
8333 		goto out;
8334 	}
8335 
8336 	panic("dtrace: undersized error hash");
8337 out:
8338 	mutex_exit(&dtrace_errlock);
8339 }
8340 #endif
8341 
8342 /*
8343  * DTrace Matching Functions
8344  *
8345  * These functions are used to match groups of probes, given some elements of
8346  * a probe tuple, or some globbed expressions for elements of a probe tuple.
8347  */
8348 static int
8349 dtrace_match_priv(const dtrace_probe_t *prp, uint32_t priv, uid_t uid,
8350     zoneid_t zoneid)
8351 {
8352 	if (priv != DTRACE_PRIV_ALL) {
8353 		uint32_t ppriv = prp->dtpr_provider->dtpv_priv.dtpp_flags;
8354 		uint32_t match = priv & ppriv;
8355 
8356 		/*
8357 		 * No PRIV_DTRACE_* privileges...
8358 		 */
8359 		if ((priv & (DTRACE_PRIV_PROC | DTRACE_PRIV_USER |
8360 		    DTRACE_PRIV_KERNEL)) == 0)
8361 			return (0);
8362 
8363 		/*
8364 		 * No matching bits, but there were bits to match...
8365 		 */
8366 		if (match == 0 && ppriv != 0)
8367 			return (0);
8368 
8369 		/*
8370 		 * Need to have permissions to the process, but don't...
8371 		 */
8372 		if (((ppriv & ~match) & DTRACE_PRIV_OWNER) != 0 &&
8373 		    uid != prp->dtpr_provider->dtpv_priv.dtpp_uid) {
8374 			return (0);
8375 		}
8376 
8377 		/*
8378 		 * Need to be in the same zone unless we possess the
8379 		 * privilege to examine all zones.
8380 		 */
8381 		if (((ppriv & ~match) & DTRACE_PRIV_ZONEOWNER) != 0 &&
8382 		    zoneid != prp->dtpr_provider->dtpv_priv.dtpp_zoneid) {
8383 			return (0);
8384 		}
8385 	}
8386 
8387 	return (1);
8388 }
8389 
8390 /*
8391  * dtrace_match_probe compares a dtrace_probe_t to a pre-compiled key, which
8392  * consists of input pattern strings and an ops-vector to evaluate them.
8393  * This function returns >0 for match, 0 for no match, and <0 for error.
8394  */
8395 static int
8396 dtrace_match_probe(const dtrace_probe_t *prp, const dtrace_probekey_t *pkp,
8397     uint32_t priv, uid_t uid, zoneid_t zoneid)
8398 {
8399 	dtrace_provider_t *pvp = prp->dtpr_provider;
8400 	int rv;
8401 
8402 	if (pvp->dtpv_defunct)
8403 		return (0);
8404 
8405 	if ((rv = pkp->dtpk_pmatch(pvp->dtpv_name, pkp->dtpk_prov, 0)) <= 0)
8406 		return (rv);
8407 
8408 	if ((rv = pkp->dtpk_mmatch(prp->dtpr_mod, pkp->dtpk_mod, 0)) <= 0)
8409 		return (rv);
8410 
8411 	if ((rv = pkp->dtpk_fmatch(prp->dtpr_func, pkp->dtpk_func, 0)) <= 0)
8412 		return (rv);
8413 
8414 	if ((rv = pkp->dtpk_nmatch(prp->dtpr_name, pkp->dtpk_name, 0)) <= 0)
8415 		return (rv);
8416 
8417 	if (dtrace_match_priv(prp, priv, uid, zoneid) == 0)
8418 		return (0);
8419 
8420 	return (rv);
8421 }
8422 
8423 /*
8424  * dtrace_match_glob() is a safe kernel implementation of the gmatch(3GEN)
8425  * interface for matching a glob pattern 'p' to an input string 's'.  Unlike
8426  * libc's version, the kernel version only applies to 8-bit ASCII strings.
8427  * In addition, all of the recursion cases except for '*' matching have been
8428  * unwound.  For '*', we still implement recursive evaluation, but a depth
8429  * counter is maintained and matching is aborted if we recurse too deep.
8430  * The function returns 0 if no match, >0 if match, and <0 if recursion error.
8431  */
8432 static int
8433 dtrace_match_glob(const char *s, const char *p, int depth)
8434 {
8435 	const char *olds;
8436 	char s1, c;
8437 	int gs;
8438 
8439 	if (depth > DTRACE_PROBEKEY_MAXDEPTH)
8440 		return (-1);
8441 
8442 	if (s == NULL)
8443 		s = ""; /* treat NULL as empty string */
8444 
8445 top:
8446 	olds = s;
8447 	s1 = *s++;
8448 
8449 	if (p == NULL)
8450 		return (0);
8451 
8452 	if ((c = *p++) == '\0')
8453 		return (s1 == '\0');
8454 
8455 	switch (c) {
8456 	case '[': {
8457 		int ok = 0, notflag = 0;
8458 		char lc = '\0';
8459 
8460 		if (s1 == '\0')
8461 			return (0);
8462 
8463 		if (*p == '!') {
8464 			notflag = 1;
8465 			p++;
8466 		}
8467 
8468 		if ((c = *p++) == '\0')
8469 			return (0);
8470 
8471 		do {
8472 			if (c == '-' && lc != '\0' && *p != ']') {
8473 				if ((c = *p++) == '\0')
8474 					return (0);
8475 				if (c == '\\' && (c = *p++) == '\0')
8476 					return (0);
8477 
8478 				if (notflag) {
8479 					if (s1 < lc || s1 > c)
8480 						ok++;
8481 					else
8482 						return (0);
8483 				} else if (lc <= s1 && s1 <= c)
8484 					ok++;
8485 
8486 			} else if (c == '\\' && (c = *p++) == '\0')
8487 				return (0);
8488 
8489 			lc = c; /* save left-hand 'c' for next iteration */
8490 
8491 			if (notflag) {
8492 				if (s1 != c)
8493 					ok++;
8494 				else
8495 					return (0);
8496 			} else if (s1 == c)
8497 				ok++;
8498 
8499 			if ((c = *p++) == '\0')
8500 				return (0);
8501 
8502 		} while (c != ']');
8503 
8504 		if (ok)
8505 			goto top;
8506 
8507 		return (0);
8508 	}
8509 
8510 	case '\\':
8511 		if ((c = *p++) == '\0')
8512 			return (0);
8513 		/*FALLTHRU*/
8514 
8515 	default:
8516 		if (c != s1)
8517 			return (0);
8518 		/*FALLTHRU*/
8519 
8520 	case '?':
8521 		if (s1 != '\0')
8522 			goto top;
8523 		return (0);
8524 
8525 	case '*':
8526 		while (*p == '*')
8527 			p++; /* consecutive *'s are identical to a single one */
8528 
8529 		if (*p == '\0')
8530 			return (1);
8531 
8532 		for (s = olds; *s != '\0'; s++) {
8533 			if ((gs = dtrace_match_glob(s, p, depth + 1)) != 0)
8534 				return (gs);
8535 		}
8536 
8537 		return (0);
8538 	}
8539 }
8540 
8541 /*ARGSUSED*/
8542 static int
8543 dtrace_match_string(const char *s, const char *p, int depth)
8544 {
8545 	return (s != NULL && strcmp(s, p) == 0);
8546 }
8547 
8548 /*ARGSUSED*/
8549 static int
8550 dtrace_match_nul(const char *s, const char *p, int depth)
8551 {
8552 	return (1); /* always match the empty pattern */
8553 }
8554 
8555 /*ARGSUSED*/
8556 static int
8557 dtrace_match_nonzero(const char *s, const char *p, int depth)
8558 {
8559 	return (s != NULL && s[0] != '\0');
8560 }
8561 
8562 static int
8563 dtrace_match(const dtrace_probekey_t *pkp, uint32_t priv, uid_t uid,
8564     zoneid_t zoneid, int (*matched)(dtrace_probe_t *, void *), void *arg)
8565 {
8566 	dtrace_probe_t template, *probe;
8567 	dtrace_hash_t *hash = NULL;
8568 	int len, best = INT_MAX, nmatched = 0;
8569 	dtrace_id_t i;
8570 
8571 	ASSERT(MUTEX_HELD(&dtrace_lock));
8572 
8573 	/*
8574 	 * If the probe ID is specified in the key, just lookup by ID and
8575 	 * invoke the match callback once if a matching probe is found.
8576 	 */
8577 	if (pkp->dtpk_id != DTRACE_IDNONE) {
8578 		if ((probe = dtrace_probe_lookup_id(pkp->dtpk_id)) != NULL &&
8579 		    dtrace_match_probe(probe, pkp, priv, uid, zoneid) > 0) {
8580 			(void) (*matched)(probe, arg);
8581 			nmatched++;
8582 		}
8583 		return (nmatched);
8584 	}
8585 
8586 	template.dtpr_mod = (char *)pkp->dtpk_mod;
8587 	template.dtpr_func = (char *)pkp->dtpk_func;
8588 	template.dtpr_name = (char *)pkp->dtpk_name;
8589 
8590 	/*
8591 	 * We want to find the most distinct of the module name, function
8592 	 * name, and name.  So for each one that is not a glob pattern or
8593 	 * empty string, we perform a lookup in the corresponding hash and
8594 	 * use the hash table with the fewest collisions to do our search.
8595 	 */
8596 	if (pkp->dtpk_mmatch == &dtrace_match_string &&
8597 	    (len = dtrace_hash_collisions(dtrace_bymod, &template)) < best) {
8598 		best = len;
8599 		hash = dtrace_bymod;
8600 	}
8601 
8602 	if (pkp->dtpk_fmatch == &dtrace_match_string &&
8603 	    (len = dtrace_hash_collisions(dtrace_byfunc, &template)) < best) {
8604 		best = len;
8605 		hash = dtrace_byfunc;
8606 	}
8607 
8608 	if (pkp->dtpk_nmatch == &dtrace_match_string &&
8609 	    (len = dtrace_hash_collisions(dtrace_byname, &template)) < best) {
8610 		best = len;
8611 		hash = dtrace_byname;
8612 	}
8613 
8614 	/*
8615 	 * If we did not select a hash table, iterate over every probe and
8616 	 * invoke our callback for each one that matches our input probe key.
8617 	 */
8618 	if (hash == NULL) {
8619 		for (i = 0; i < dtrace_nprobes; i++) {
8620 			if ((probe = dtrace_probes[i]) == NULL ||
8621 			    dtrace_match_probe(probe, pkp, priv, uid,
8622 			    zoneid) <= 0)
8623 				continue;
8624 
8625 			nmatched++;
8626 
8627 			if ((*matched)(probe, arg) != DTRACE_MATCH_NEXT)
8628 				break;
8629 		}
8630 
8631 		return (nmatched);
8632 	}
8633 
8634 	/*
8635 	 * If we selected a hash table, iterate over each probe of the same key
8636 	 * name and invoke the callback for every probe that matches the other
8637 	 * attributes of our input probe key.
8638 	 */
8639 	for (probe = dtrace_hash_lookup(hash, &template); probe != NULL;
8640 	    probe = *(DTRACE_HASHNEXT(hash, probe))) {
8641 
8642 		if (dtrace_match_probe(probe, pkp, priv, uid, zoneid) <= 0)
8643 			continue;
8644 
8645 		nmatched++;
8646 
8647 		if ((*matched)(probe, arg) != DTRACE_MATCH_NEXT)
8648 			break;
8649 	}
8650 
8651 	return (nmatched);
8652 }
8653 
8654 /*
8655  * Return the function pointer dtrace_probecmp() should use to compare the
8656  * specified pattern with a string.  For NULL or empty patterns, we select
8657  * dtrace_match_nul().  For glob pattern strings, we use dtrace_match_glob().
8658  * For non-empty non-glob strings, we use dtrace_match_string().
8659  */
8660 static dtrace_probekey_f *
8661 dtrace_probekey_func(const char *p)
8662 {
8663 	char c;
8664 
8665 	if (p == NULL || *p == '\0')
8666 		return (&dtrace_match_nul);
8667 
8668 	while ((c = *p++) != '\0') {
8669 		if (c == '[' || c == '?' || c == '*' || c == '\\')
8670 			return (&dtrace_match_glob);
8671 	}
8672 
8673 	return (&dtrace_match_string);
8674 }
8675 
8676 /*
8677  * Build a probe comparison key for use with dtrace_match_probe() from the
8678  * given probe description.  By convention, a null key only matches anchored
8679  * probes: if each field is the empty string, reset dtpk_fmatch to
8680  * dtrace_match_nonzero().
8681  */
8682 static void
8683 dtrace_probekey(dtrace_probedesc_t *pdp, dtrace_probekey_t *pkp)
8684 {
8685 	pkp->dtpk_prov = pdp->dtpd_provider;
8686 	pkp->dtpk_pmatch = dtrace_probekey_func(pdp->dtpd_provider);
8687 
8688 	pkp->dtpk_mod = pdp->dtpd_mod;
8689 	pkp->dtpk_mmatch = dtrace_probekey_func(pdp->dtpd_mod);
8690 
8691 	pkp->dtpk_func = pdp->dtpd_func;
8692 	pkp->dtpk_fmatch = dtrace_probekey_func(pdp->dtpd_func);
8693 
8694 	pkp->dtpk_name = pdp->dtpd_name;
8695 	pkp->dtpk_nmatch = dtrace_probekey_func(pdp->dtpd_name);
8696 
8697 	pkp->dtpk_id = pdp->dtpd_id;
8698 
8699 	if (pkp->dtpk_id == DTRACE_IDNONE &&
8700 	    pkp->dtpk_pmatch == &dtrace_match_nul &&
8701 	    pkp->dtpk_mmatch == &dtrace_match_nul &&
8702 	    pkp->dtpk_fmatch == &dtrace_match_nul &&
8703 	    pkp->dtpk_nmatch == &dtrace_match_nul)
8704 		pkp->dtpk_fmatch = &dtrace_match_nonzero;
8705 }
8706 
8707 /*
8708  * DTrace Provider-to-Framework API Functions
8709  *
8710  * These functions implement much of the Provider-to-Framework API, as
8711  * described in <sys/dtrace.h>.  The parts of the API not in this section are
8712  * the functions in the API for probe management (found below), and
8713  * dtrace_probe() itself (found above).
8714  */
8715 
8716 /*
8717  * Register the calling provider with the DTrace framework.  This should
8718  * generally be called by DTrace providers in their attach(9E) entry point.
8719  */
8720 int
8721 dtrace_register(const char *name, const dtrace_pattr_t *pap, uint32_t priv,
8722     cred_t *cr, const dtrace_pops_t *pops, void *arg, dtrace_provider_id_t *idp)
8723 {
8724 	dtrace_provider_t *provider;
8725 
8726 	if (name == NULL || pap == NULL || pops == NULL || idp == NULL) {
8727 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
8728 		    "arguments", name ? name : "<NULL>");
8729 		return (EINVAL);
8730 	}
8731 
8732 	if (name[0] == '\0' || dtrace_badname(name)) {
8733 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
8734 		    "provider name", name);
8735 		return (EINVAL);
8736 	}
8737 
8738 	if ((pops->dtps_provide == NULL && pops->dtps_provide_module == NULL) ||
8739 	    pops->dtps_enable == NULL || pops->dtps_disable == NULL ||
8740 	    pops->dtps_destroy == NULL ||
8741 	    ((pops->dtps_resume == NULL) != (pops->dtps_suspend == NULL))) {
8742 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
8743 		    "provider ops", name);
8744 		return (EINVAL);
8745 	}
8746 
8747 	if (dtrace_badattr(&pap->dtpa_provider) ||
8748 	    dtrace_badattr(&pap->dtpa_mod) ||
8749 	    dtrace_badattr(&pap->dtpa_func) ||
8750 	    dtrace_badattr(&pap->dtpa_name) ||
8751 	    dtrace_badattr(&pap->dtpa_args)) {
8752 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
8753 		    "provider attributes", name);
8754 		return (EINVAL);
8755 	}
8756 
8757 	if (priv & ~DTRACE_PRIV_ALL) {
8758 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
8759 		    "privilege attributes", name);
8760 		return (EINVAL);
8761 	}
8762 
8763 	if ((priv & DTRACE_PRIV_KERNEL) &&
8764 	    (priv & (DTRACE_PRIV_USER | DTRACE_PRIV_OWNER)) &&
8765 	    pops->dtps_usermode == NULL) {
8766 		cmn_err(CE_WARN, "failed to register provider '%s': need "
8767 		    "dtps_usermode() op for given privilege attributes", name);
8768 		return (EINVAL);
8769 	}
8770 
8771 	provider = kmem_zalloc(sizeof (dtrace_provider_t), KM_SLEEP);
8772 	provider->dtpv_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
8773 	(void) strcpy(provider->dtpv_name, name);
8774 
8775 	provider->dtpv_attr = *pap;
8776 	provider->dtpv_priv.dtpp_flags = priv;
8777 	if (cr != NULL) {
8778 		provider->dtpv_priv.dtpp_uid = crgetuid(cr);
8779 		provider->dtpv_priv.dtpp_zoneid = crgetzoneid(cr);
8780 	}
8781 	provider->dtpv_pops = *pops;
8782 
8783 	if (pops->dtps_provide == NULL) {
8784 		ASSERT(pops->dtps_provide_module != NULL);
8785 		provider->dtpv_pops.dtps_provide =
8786 		    (void (*)(void *, dtrace_probedesc_t *))dtrace_nullop;
8787 	}
8788 
8789 	if (pops->dtps_provide_module == NULL) {
8790 		ASSERT(pops->dtps_provide != NULL);
8791 		provider->dtpv_pops.dtps_provide_module =
8792 		    (void (*)(void *, modctl_t *))dtrace_nullop;
8793 	}
8794 
8795 	if (pops->dtps_suspend == NULL) {
8796 		ASSERT(pops->dtps_resume == NULL);
8797 		provider->dtpv_pops.dtps_suspend =
8798 		    (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
8799 		provider->dtpv_pops.dtps_resume =
8800 		    (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
8801 	}
8802 
8803 	provider->dtpv_arg = arg;
8804 	*idp = (dtrace_provider_id_t)provider;
8805 
8806 	if (pops == &dtrace_provider_ops) {
8807 		ASSERT(MUTEX_HELD(&dtrace_provider_lock));
8808 		ASSERT(MUTEX_HELD(&dtrace_lock));
8809 		ASSERT(dtrace_anon.dta_enabling == NULL);
8810 
8811 		/*
8812 		 * We make sure that the DTrace provider is at the head of
8813 		 * the provider chain.
8814 		 */
8815 		provider->dtpv_next = dtrace_provider;
8816 		dtrace_provider = provider;
8817 		return (0);
8818 	}
8819 
8820 	mutex_enter(&dtrace_provider_lock);
8821 	mutex_enter(&dtrace_lock);
8822 
8823 	/*
8824 	 * If there is at least one provider registered, we'll add this
8825 	 * provider after the first provider.
8826 	 */
8827 	if (dtrace_provider != NULL) {
8828 		provider->dtpv_next = dtrace_provider->dtpv_next;
8829 		dtrace_provider->dtpv_next = provider;
8830 	} else {
8831 		dtrace_provider = provider;
8832 	}
8833 
8834 	if (dtrace_retained != NULL) {
8835 		dtrace_enabling_provide(provider);
8836 
8837 		/*
8838 		 * Now we need to call dtrace_enabling_matchall() -- which
8839 		 * will acquire cpu_lock and dtrace_lock.  We therefore need
8840 		 * to drop all of our locks before calling into it...
8841 		 */
8842 		mutex_exit(&dtrace_lock);
8843 		mutex_exit(&dtrace_provider_lock);
8844 		dtrace_enabling_matchall();
8845 
8846 		return (0);
8847 	}
8848 
8849 	mutex_exit(&dtrace_lock);
8850 	mutex_exit(&dtrace_provider_lock);
8851 
8852 	return (0);
8853 }
8854 
8855 /*
8856  * Unregister the specified provider from the DTrace framework.  This should
8857  * generally be called by DTrace providers in their detach(9E) entry point.
8858  */
8859 int
8860 dtrace_unregister(dtrace_provider_id_t id)
8861 {
8862 	dtrace_provider_t *old = (dtrace_provider_t *)id;
8863 	dtrace_provider_t *prev = NULL;
8864 	int i, self = 0, noreap = 0;
8865 	dtrace_probe_t *probe, *first = NULL;
8866 
8867 	if (old->dtpv_pops.dtps_enable ==
8868 	    (void (*)(void *, dtrace_id_t, void *))dtrace_nullop) {
8869 		/*
8870 		 * If DTrace itself is the provider, we're called with locks
8871 		 * already held.
8872 		 */
8873 		ASSERT(old == dtrace_provider);
8874 #ifdef illumos
8875 		ASSERT(dtrace_devi != NULL);
8876 #endif
8877 		ASSERT(MUTEX_HELD(&dtrace_provider_lock));
8878 		ASSERT(MUTEX_HELD(&dtrace_lock));
8879 		self = 1;
8880 
8881 		if (dtrace_provider->dtpv_next != NULL) {
8882 			/*
8883 			 * There's another provider here; return failure.
8884 			 */
8885 			return (EBUSY);
8886 		}
8887 	} else {
8888 		mutex_enter(&dtrace_provider_lock);
8889 #ifdef illumos
8890 		mutex_enter(&mod_lock);
8891 #endif
8892 		mutex_enter(&dtrace_lock);
8893 	}
8894 
8895 	/*
8896 	 * If anyone has /dev/dtrace open, or if there are anonymous enabled
8897 	 * probes, we refuse to let providers slither away, unless this
8898 	 * provider has already been explicitly invalidated.
8899 	 */
8900 	if (!old->dtpv_defunct &&
8901 	    (dtrace_opens || (dtrace_anon.dta_state != NULL &&
8902 	    dtrace_anon.dta_state->dts_necbs > 0))) {
8903 		if (!self) {
8904 			mutex_exit(&dtrace_lock);
8905 #ifdef illumos
8906 			mutex_exit(&mod_lock);
8907 #endif
8908 			mutex_exit(&dtrace_provider_lock);
8909 		}
8910 		return (EBUSY);
8911 	}
8912 
8913 	/*
8914 	 * Attempt to destroy the probes associated with this provider.
8915 	 */
8916 	for (i = 0; i < dtrace_nprobes; i++) {
8917 		if ((probe = dtrace_probes[i]) == NULL)
8918 			continue;
8919 
8920 		if (probe->dtpr_provider != old)
8921 			continue;
8922 
8923 		if (probe->dtpr_ecb == NULL)
8924 			continue;
8925 
8926 		/*
8927 		 * If we are trying to unregister a defunct provider, and the
8928 		 * provider was made defunct within the interval dictated by
8929 		 * dtrace_unregister_defunct_reap, we'll (asynchronously)
8930 		 * attempt to reap our enablings.  To denote that the provider
8931 		 * should reattempt to unregister itself at some point in the
8932 		 * future, we will return a differentiable error code (EAGAIN
8933 		 * instead of EBUSY) in this case.
8934 		 */
8935 		if (dtrace_gethrtime() - old->dtpv_defunct >
8936 		    dtrace_unregister_defunct_reap)
8937 			noreap = 1;
8938 
8939 		if (!self) {
8940 			mutex_exit(&dtrace_lock);
8941 #ifdef illumos
8942 			mutex_exit(&mod_lock);
8943 #endif
8944 			mutex_exit(&dtrace_provider_lock);
8945 		}
8946 
8947 		if (noreap)
8948 			return (EBUSY);
8949 
8950 		(void) taskq_dispatch(dtrace_taskq,
8951 		    (task_func_t *)dtrace_enabling_reap, NULL, TQ_SLEEP);
8952 
8953 		return (EAGAIN);
8954 	}
8955 
8956 	/*
8957 	 * All of the probes for this provider are disabled; we can safely
8958 	 * remove all of them from their hash chains and from the probe array.
8959 	 */
8960 	for (i = 0; i < dtrace_nprobes; i++) {
8961 		if ((probe = dtrace_probes[i]) == NULL)
8962 			continue;
8963 
8964 		if (probe->dtpr_provider != old)
8965 			continue;
8966 
8967 		dtrace_probes[i] = NULL;
8968 
8969 		dtrace_hash_remove(dtrace_bymod, probe);
8970 		dtrace_hash_remove(dtrace_byfunc, probe);
8971 		dtrace_hash_remove(dtrace_byname, probe);
8972 
8973 		if (first == NULL) {
8974 			first = probe;
8975 			probe->dtpr_nextmod = NULL;
8976 		} else {
8977 			probe->dtpr_nextmod = first;
8978 			first = probe;
8979 		}
8980 	}
8981 
8982 	/*
8983 	 * The provider's probes have been removed from the hash chains and
8984 	 * from the probe array.  Now issue a dtrace_sync() to be sure that
8985 	 * everyone has cleared out from any probe array processing.
8986 	 */
8987 	dtrace_sync();
8988 
8989 	for (probe = first; probe != NULL; probe = first) {
8990 		first = probe->dtpr_nextmod;
8991 
8992 		old->dtpv_pops.dtps_destroy(old->dtpv_arg, probe->dtpr_id,
8993 		    probe->dtpr_arg);
8994 		kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
8995 		kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
8996 		kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
8997 #ifdef illumos
8998 		vmem_free(dtrace_arena, (void *)(uintptr_t)(probe->dtpr_id), 1);
8999 #else
9000 		free_unr(dtrace_arena, probe->dtpr_id);
9001 #endif
9002 		kmem_free(probe, sizeof (dtrace_probe_t));
9003 	}
9004 
9005 	if ((prev = dtrace_provider) == old) {
9006 #ifdef illumos
9007 		ASSERT(self || dtrace_devi == NULL);
9008 		ASSERT(old->dtpv_next == NULL || dtrace_devi == NULL);
9009 #endif
9010 		dtrace_provider = old->dtpv_next;
9011 	} else {
9012 		while (prev != NULL && prev->dtpv_next != old)
9013 			prev = prev->dtpv_next;
9014 
9015 		if (prev == NULL) {
9016 			panic("attempt to unregister non-existent "
9017 			    "dtrace provider %p\n", (void *)id);
9018 		}
9019 
9020 		prev->dtpv_next = old->dtpv_next;
9021 	}
9022 
9023 	if (!self) {
9024 		mutex_exit(&dtrace_lock);
9025 #ifdef illumos
9026 		mutex_exit(&mod_lock);
9027 #endif
9028 		mutex_exit(&dtrace_provider_lock);
9029 	}
9030 
9031 	kmem_free(old->dtpv_name, strlen(old->dtpv_name) + 1);
9032 	kmem_free(old, sizeof (dtrace_provider_t));
9033 
9034 	return (0);
9035 }
9036 
9037 /*
9038  * Invalidate the specified provider.  All subsequent probe lookups for the
9039  * specified provider will fail, but its probes will not be removed.
9040  */
9041 void
9042 dtrace_invalidate(dtrace_provider_id_t id)
9043 {
9044 	dtrace_provider_t *pvp = (dtrace_provider_t *)id;
9045 
9046 	ASSERT(pvp->dtpv_pops.dtps_enable !=
9047 	    (void (*)(void *, dtrace_id_t, void *))dtrace_nullop);
9048 
9049 	mutex_enter(&dtrace_provider_lock);
9050 	mutex_enter(&dtrace_lock);
9051 
9052 	pvp->dtpv_defunct = dtrace_gethrtime();
9053 
9054 	mutex_exit(&dtrace_lock);
9055 	mutex_exit(&dtrace_provider_lock);
9056 }
9057 
9058 /*
9059  * Indicate whether or not DTrace has attached.
9060  */
9061 int
9062 dtrace_attached(void)
9063 {
9064 	/*
9065 	 * dtrace_provider will be non-NULL iff the DTrace driver has
9066 	 * attached.  (It's non-NULL because DTrace is always itself a
9067 	 * provider.)
9068 	 */
9069 	return (dtrace_provider != NULL);
9070 }
9071 
9072 /*
9073  * Remove all the unenabled probes for the given provider.  This function is
9074  * not unlike dtrace_unregister(), except that it doesn't remove the provider
9075  * -- just as many of its associated probes as it can.
9076  */
9077 int
9078 dtrace_condense(dtrace_provider_id_t id)
9079 {
9080 	dtrace_provider_t *prov = (dtrace_provider_t *)id;
9081 	int i;
9082 	dtrace_probe_t *probe;
9083 
9084 	/*
9085 	 * Make sure this isn't the dtrace provider itself.
9086 	 */
9087 	ASSERT(prov->dtpv_pops.dtps_enable !=
9088 	    (void (*)(void *, dtrace_id_t, void *))dtrace_nullop);
9089 
9090 	mutex_enter(&dtrace_provider_lock);
9091 	mutex_enter(&dtrace_lock);
9092 
9093 	/*
9094 	 * Attempt to destroy the probes associated with this provider.
9095 	 */
9096 	for (i = 0; i < dtrace_nprobes; i++) {
9097 		if ((probe = dtrace_probes[i]) == NULL)
9098 			continue;
9099 
9100 		if (probe->dtpr_provider != prov)
9101 			continue;
9102 
9103 		if (probe->dtpr_ecb != NULL)
9104 			continue;
9105 
9106 		dtrace_probes[i] = NULL;
9107 
9108 		dtrace_hash_remove(dtrace_bymod, probe);
9109 		dtrace_hash_remove(dtrace_byfunc, probe);
9110 		dtrace_hash_remove(dtrace_byname, probe);
9111 
9112 		prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, i + 1,
9113 		    probe->dtpr_arg);
9114 		kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
9115 		kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
9116 		kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
9117 		kmem_free(probe, sizeof (dtrace_probe_t));
9118 #ifdef illumos
9119 		vmem_free(dtrace_arena, (void *)((uintptr_t)i + 1), 1);
9120 #else
9121 		free_unr(dtrace_arena, i + 1);
9122 #endif
9123 	}
9124 
9125 	mutex_exit(&dtrace_lock);
9126 	mutex_exit(&dtrace_provider_lock);
9127 
9128 	return (0);
9129 }
9130 
9131 /*
9132  * DTrace Probe Management Functions
9133  *
9134  * The functions in this section perform the DTrace probe management,
9135  * including functions to create probes, look-up probes, and call into the
9136  * providers to request that probes be provided.  Some of these functions are
9137  * in the Provider-to-Framework API; these functions can be identified by the
9138  * fact that they are not declared "static".
9139  */
9140 
9141 /*
9142  * Create a probe with the specified module name, function name, and name.
9143  */
9144 dtrace_id_t
9145 dtrace_probe_create(dtrace_provider_id_t prov, const char *mod,
9146     const char *func, const char *name, int aframes, void *arg)
9147 {
9148 	dtrace_probe_t *probe, **probes;
9149 	dtrace_provider_t *provider = (dtrace_provider_t *)prov;
9150 	dtrace_id_t id;
9151 
9152 	if (provider == dtrace_provider) {
9153 		ASSERT(MUTEX_HELD(&dtrace_lock));
9154 	} else {
9155 		mutex_enter(&dtrace_lock);
9156 	}
9157 
9158 #ifdef illumos
9159 	id = (dtrace_id_t)(uintptr_t)vmem_alloc(dtrace_arena, 1,
9160 	    VM_BESTFIT | VM_SLEEP);
9161 #else
9162 	id = alloc_unr(dtrace_arena);
9163 #endif
9164 	probe = kmem_zalloc(sizeof (dtrace_probe_t), KM_SLEEP);
9165 
9166 	probe->dtpr_id = id;
9167 	probe->dtpr_gen = dtrace_probegen++;
9168 	probe->dtpr_mod = dtrace_strdup(mod);
9169 	probe->dtpr_func = dtrace_strdup(func);
9170 	probe->dtpr_name = dtrace_strdup(name);
9171 	probe->dtpr_arg = arg;
9172 	probe->dtpr_aframes = aframes;
9173 	probe->dtpr_provider = provider;
9174 
9175 	dtrace_hash_add(dtrace_bymod, probe);
9176 	dtrace_hash_add(dtrace_byfunc, probe);
9177 	dtrace_hash_add(dtrace_byname, probe);
9178 
9179 	if (id - 1 >= dtrace_nprobes) {
9180 		size_t osize = dtrace_nprobes * sizeof (dtrace_probe_t *);
9181 		size_t nsize = osize << 1;
9182 
9183 		if (nsize == 0) {
9184 			ASSERT(osize == 0);
9185 			ASSERT(dtrace_probes == NULL);
9186 			nsize = sizeof (dtrace_probe_t *);
9187 		}
9188 
9189 		probes = kmem_zalloc(nsize, KM_SLEEP);
9190 
9191 		if (dtrace_probes == NULL) {
9192 			ASSERT(osize == 0);
9193 			dtrace_probes = probes;
9194 			dtrace_nprobes = 1;
9195 		} else {
9196 			dtrace_probe_t **oprobes = dtrace_probes;
9197 
9198 			bcopy(oprobes, probes, osize);
9199 			dtrace_membar_producer();
9200 			dtrace_probes = probes;
9201 
9202 			dtrace_sync();
9203 
9204 			/*
9205 			 * All CPUs are now seeing the new probes array; we can
9206 			 * safely free the old array.
9207 			 */
9208 			kmem_free(oprobes, osize);
9209 			dtrace_nprobes <<= 1;
9210 		}
9211 
9212 		ASSERT(id - 1 < dtrace_nprobes);
9213 	}
9214 
9215 	ASSERT(dtrace_probes[id - 1] == NULL);
9216 	dtrace_probes[id - 1] = probe;
9217 
9218 	if (provider != dtrace_provider)
9219 		mutex_exit(&dtrace_lock);
9220 
9221 	return (id);
9222 }
9223 
9224 static dtrace_probe_t *
9225 dtrace_probe_lookup_id(dtrace_id_t id)
9226 {
9227 	ASSERT(MUTEX_HELD(&dtrace_lock));
9228 
9229 	if (id == 0 || id > dtrace_nprobes)
9230 		return (NULL);
9231 
9232 	return (dtrace_probes[id - 1]);
9233 }
9234 
9235 static int
9236 dtrace_probe_lookup_match(dtrace_probe_t *probe, void *arg)
9237 {
9238 	*((dtrace_id_t *)arg) = probe->dtpr_id;
9239 
9240 	return (DTRACE_MATCH_DONE);
9241 }
9242 
9243 /*
9244  * Look up a probe based on provider and one or more of module name, function
9245  * name and probe name.
9246  */
9247 dtrace_id_t
9248 dtrace_probe_lookup(dtrace_provider_id_t prid, char *mod,
9249     char *func, char *name)
9250 {
9251 	dtrace_probekey_t pkey;
9252 	dtrace_id_t id;
9253 	int match;
9254 
9255 	pkey.dtpk_prov = ((dtrace_provider_t *)prid)->dtpv_name;
9256 	pkey.dtpk_pmatch = &dtrace_match_string;
9257 	pkey.dtpk_mod = mod;
9258 	pkey.dtpk_mmatch = mod ? &dtrace_match_string : &dtrace_match_nul;
9259 	pkey.dtpk_func = func;
9260 	pkey.dtpk_fmatch = func ? &dtrace_match_string : &dtrace_match_nul;
9261 	pkey.dtpk_name = name;
9262 	pkey.dtpk_nmatch = name ? &dtrace_match_string : &dtrace_match_nul;
9263 	pkey.dtpk_id = DTRACE_IDNONE;
9264 
9265 	mutex_enter(&dtrace_lock);
9266 	match = dtrace_match(&pkey, DTRACE_PRIV_ALL, 0, 0,
9267 	    dtrace_probe_lookup_match, &id);
9268 	mutex_exit(&dtrace_lock);
9269 
9270 	ASSERT(match == 1 || match == 0);
9271 	return (match ? id : 0);
9272 }
9273 
9274 /*
9275  * Returns the probe argument associated with the specified probe.
9276  */
9277 void *
9278 dtrace_probe_arg(dtrace_provider_id_t id, dtrace_id_t pid)
9279 {
9280 	dtrace_probe_t *probe;
9281 	void *rval = NULL;
9282 
9283 	mutex_enter(&dtrace_lock);
9284 
9285 	if ((probe = dtrace_probe_lookup_id(pid)) != NULL &&
9286 	    probe->dtpr_provider == (dtrace_provider_t *)id)
9287 		rval = probe->dtpr_arg;
9288 
9289 	mutex_exit(&dtrace_lock);
9290 
9291 	return (rval);
9292 }
9293 
9294 /*
9295  * Copy a probe into a probe description.
9296  */
9297 static void
9298 dtrace_probe_description(const dtrace_probe_t *prp, dtrace_probedesc_t *pdp)
9299 {
9300 	bzero(pdp, sizeof (dtrace_probedesc_t));
9301 	pdp->dtpd_id = prp->dtpr_id;
9302 
9303 	(void) strncpy(pdp->dtpd_provider,
9304 	    prp->dtpr_provider->dtpv_name, DTRACE_PROVNAMELEN - 1);
9305 
9306 	(void) strncpy(pdp->dtpd_mod, prp->dtpr_mod, DTRACE_MODNAMELEN - 1);
9307 	(void) strncpy(pdp->dtpd_func, prp->dtpr_func, DTRACE_FUNCNAMELEN - 1);
9308 	(void) strncpy(pdp->dtpd_name, prp->dtpr_name, DTRACE_NAMELEN - 1);
9309 }
9310 
9311 /*
9312  * Called to indicate that a probe -- or probes -- should be provided by a
9313  * specfied provider.  If the specified description is NULL, the provider will
9314  * be told to provide all of its probes.  (This is done whenever a new
9315  * consumer comes along, or whenever a retained enabling is to be matched.) If
9316  * the specified description is non-NULL, the provider is given the
9317  * opportunity to dynamically provide the specified probe, allowing providers
9318  * to support the creation of probes on-the-fly.  (So-called _autocreated_
9319  * probes.)  If the provider is NULL, the operations will be applied to all
9320  * providers; if the provider is non-NULL the operations will only be applied
9321  * to the specified provider.  The dtrace_provider_lock must be held, and the
9322  * dtrace_lock must _not_ be held -- the provider's dtps_provide() operation
9323  * will need to grab the dtrace_lock when it reenters the framework through
9324  * dtrace_probe_lookup(), dtrace_probe_create(), etc.
9325  */
9326 static void
9327 dtrace_probe_provide(dtrace_probedesc_t *desc, dtrace_provider_t *prv)
9328 {
9329 #ifdef illumos
9330 	modctl_t *ctl;
9331 #endif
9332 	int all = 0;
9333 
9334 	ASSERT(MUTEX_HELD(&dtrace_provider_lock));
9335 
9336 	if (prv == NULL) {
9337 		all = 1;
9338 		prv = dtrace_provider;
9339 	}
9340 
9341 	do {
9342 		/*
9343 		 * First, call the blanket provide operation.
9344 		 */
9345 		prv->dtpv_pops.dtps_provide(prv->dtpv_arg, desc);
9346 
9347 #ifdef illumos
9348 		/*
9349 		 * Now call the per-module provide operation.  We will grab
9350 		 * mod_lock to prevent the list from being modified.  Note
9351 		 * that this also prevents the mod_busy bits from changing.
9352 		 * (mod_busy can only be changed with mod_lock held.)
9353 		 */
9354 		mutex_enter(&mod_lock);
9355 
9356 		ctl = &modules;
9357 		do {
9358 			if (ctl->mod_busy || ctl->mod_mp == NULL)
9359 				continue;
9360 
9361 			prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
9362 
9363 		} while ((ctl = ctl->mod_next) != &modules);
9364 
9365 		mutex_exit(&mod_lock);
9366 #endif
9367 	} while (all && (prv = prv->dtpv_next) != NULL);
9368 }
9369 
9370 #ifdef illumos
9371 /*
9372  * Iterate over each probe, and call the Framework-to-Provider API function
9373  * denoted by offs.
9374  */
9375 static void
9376 dtrace_probe_foreach(uintptr_t offs)
9377 {
9378 	dtrace_provider_t *prov;
9379 	void (*func)(void *, dtrace_id_t, void *);
9380 	dtrace_probe_t *probe;
9381 	dtrace_icookie_t cookie;
9382 	int i;
9383 
9384 	/*
9385 	 * We disable interrupts to walk through the probe array.  This is
9386 	 * safe -- the dtrace_sync() in dtrace_unregister() assures that we
9387 	 * won't see stale data.
9388 	 */
9389 	cookie = dtrace_interrupt_disable();
9390 
9391 	for (i = 0; i < dtrace_nprobes; i++) {
9392 		if ((probe = dtrace_probes[i]) == NULL)
9393 			continue;
9394 
9395 		if (probe->dtpr_ecb == NULL) {
9396 			/*
9397 			 * This probe isn't enabled -- don't call the function.
9398 			 */
9399 			continue;
9400 		}
9401 
9402 		prov = probe->dtpr_provider;
9403 		func = *((void(**)(void *, dtrace_id_t, void *))
9404 		    ((uintptr_t)&prov->dtpv_pops + offs));
9405 
9406 		func(prov->dtpv_arg, i + 1, probe->dtpr_arg);
9407 	}
9408 
9409 	dtrace_interrupt_enable(cookie);
9410 }
9411 #endif
9412 
9413 static int
9414 dtrace_probe_enable(dtrace_probedesc_t *desc, dtrace_enabling_t *enab)
9415 {
9416 	dtrace_probekey_t pkey;
9417 	uint32_t priv;
9418 	uid_t uid;
9419 	zoneid_t zoneid;
9420 
9421 	ASSERT(MUTEX_HELD(&dtrace_lock));
9422 	dtrace_ecb_create_cache = NULL;
9423 
9424 	if (desc == NULL) {
9425 		/*
9426 		 * If we're passed a NULL description, we're being asked to
9427 		 * create an ECB with a NULL probe.
9428 		 */
9429 		(void) dtrace_ecb_create_enable(NULL, enab);
9430 		return (0);
9431 	}
9432 
9433 	dtrace_probekey(desc, &pkey);
9434 	dtrace_cred2priv(enab->dten_vstate->dtvs_state->dts_cred.dcr_cred,
9435 	    &priv, &uid, &zoneid);
9436 
9437 	return (dtrace_match(&pkey, priv, uid, zoneid, dtrace_ecb_create_enable,
9438 	    enab));
9439 }
9440 
9441 /*
9442  * DTrace Helper Provider Functions
9443  */
9444 static void
9445 dtrace_dofattr2attr(dtrace_attribute_t *attr, const dof_attr_t dofattr)
9446 {
9447 	attr->dtat_name = DOF_ATTR_NAME(dofattr);
9448 	attr->dtat_data = DOF_ATTR_DATA(dofattr);
9449 	attr->dtat_class = DOF_ATTR_CLASS(dofattr);
9450 }
9451 
9452 static void
9453 dtrace_dofprov2hprov(dtrace_helper_provdesc_t *hprov,
9454     const dof_provider_t *dofprov, char *strtab)
9455 {
9456 	hprov->dthpv_provname = strtab + dofprov->dofpv_name;
9457 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_provider,
9458 	    dofprov->dofpv_provattr);
9459 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_mod,
9460 	    dofprov->dofpv_modattr);
9461 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_func,
9462 	    dofprov->dofpv_funcattr);
9463 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_name,
9464 	    dofprov->dofpv_nameattr);
9465 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_args,
9466 	    dofprov->dofpv_argsattr);
9467 }
9468 
9469 static void
9470 dtrace_helper_provide_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
9471 {
9472 	uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
9473 	dof_hdr_t *dof = (dof_hdr_t *)daddr;
9474 	dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
9475 	dof_provider_t *provider;
9476 	dof_probe_t *probe;
9477 	uint32_t *off, *enoff;
9478 	uint8_t *arg;
9479 	char *strtab;
9480 	uint_t i, nprobes;
9481 	dtrace_helper_provdesc_t dhpv;
9482 	dtrace_helper_probedesc_t dhpb;
9483 	dtrace_meta_t *meta = dtrace_meta_pid;
9484 	dtrace_mops_t *mops = &meta->dtm_mops;
9485 	void *parg;
9486 
9487 	provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
9488 	str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
9489 	    provider->dofpv_strtab * dof->dofh_secsize);
9490 	prb_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
9491 	    provider->dofpv_probes * dof->dofh_secsize);
9492 	arg_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
9493 	    provider->dofpv_prargs * dof->dofh_secsize);
9494 	off_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
9495 	    provider->dofpv_proffs * dof->dofh_secsize);
9496 
9497 	strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
9498 	off = (uint32_t *)(uintptr_t)(daddr + off_sec->dofs_offset);
9499 	arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
9500 	enoff = NULL;
9501 
9502 	/*
9503 	 * See dtrace_helper_provider_validate().
9504 	 */
9505 	if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
9506 	    provider->dofpv_prenoffs != DOF_SECT_NONE) {
9507 		enoff_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
9508 		    provider->dofpv_prenoffs * dof->dofh_secsize);
9509 		enoff = (uint32_t *)(uintptr_t)(daddr + enoff_sec->dofs_offset);
9510 	}
9511 
9512 	nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
9513 
9514 	/*
9515 	 * Create the provider.
9516 	 */
9517 	dtrace_dofprov2hprov(&dhpv, provider, strtab);
9518 
9519 	if ((parg = mops->dtms_provide_pid(meta->dtm_arg, &dhpv, pid)) == NULL)
9520 		return;
9521 
9522 	meta->dtm_count++;
9523 
9524 	/*
9525 	 * Create the probes.
9526 	 */
9527 	for (i = 0; i < nprobes; i++) {
9528 		probe = (dof_probe_t *)(uintptr_t)(daddr +
9529 		    prb_sec->dofs_offset + i * prb_sec->dofs_entsize);
9530 
9531 		/* See the check in dtrace_helper_provider_validate(). */
9532 		if (strlen(strtab + probe->dofpr_func) >= DTRACE_FUNCNAMELEN)
9533 			continue;
9534 
9535 		dhpb.dthpb_mod = dhp->dofhp_mod;
9536 		dhpb.dthpb_func = strtab + probe->dofpr_func;
9537 		dhpb.dthpb_name = strtab + probe->dofpr_name;
9538 		dhpb.dthpb_base = probe->dofpr_addr;
9539 		dhpb.dthpb_offs = off + probe->dofpr_offidx;
9540 		dhpb.dthpb_noffs = probe->dofpr_noffs;
9541 		if (enoff != NULL) {
9542 			dhpb.dthpb_enoffs = enoff + probe->dofpr_enoffidx;
9543 			dhpb.dthpb_nenoffs = probe->dofpr_nenoffs;
9544 		} else {
9545 			dhpb.dthpb_enoffs = NULL;
9546 			dhpb.dthpb_nenoffs = 0;
9547 		}
9548 		dhpb.dthpb_args = arg + probe->dofpr_argidx;
9549 		dhpb.dthpb_nargc = probe->dofpr_nargc;
9550 		dhpb.dthpb_xargc = probe->dofpr_xargc;
9551 		dhpb.dthpb_ntypes = strtab + probe->dofpr_nargv;
9552 		dhpb.dthpb_xtypes = strtab + probe->dofpr_xargv;
9553 
9554 		mops->dtms_create_probe(meta->dtm_arg, parg, &dhpb);
9555 	}
9556 }
9557 
9558 static void
9559 dtrace_helper_provide(dof_helper_t *dhp, pid_t pid)
9560 {
9561 	uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
9562 	dof_hdr_t *dof = (dof_hdr_t *)daddr;
9563 	int i;
9564 
9565 	ASSERT(MUTEX_HELD(&dtrace_meta_lock));
9566 
9567 	for (i = 0; i < dof->dofh_secnum; i++) {
9568 		dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
9569 		    dof->dofh_secoff + i * dof->dofh_secsize);
9570 
9571 		if (sec->dofs_type != DOF_SECT_PROVIDER)
9572 			continue;
9573 
9574 		dtrace_helper_provide_one(dhp, sec, pid);
9575 	}
9576 
9577 	/*
9578 	 * We may have just created probes, so we must now rematch against
9579 	 * any retained enablings.  Note that this call will acquire both
9580 	 * cpu_lock and dtrace_lock; the fact that we are holding
9581 	 * dtrace_meta_lock now is what defines the ordering with respect to
9582 	 * these three locks.
9583 	 */
9584 	dtrace_enabling_matchall();
9585 }
9586 
9587 static void
9588 dtrace_helper_provider_remove_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
9589 {
9590 	uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
9591 	dof_hdr_t *dof = (dof_hdr_t *)daddr;
9592 	dof_sec_t *str_sec;
9593 	dof_provider_t *provider;
9594 	char *strtab;
9595 	dtrace_helper_provdesc_t dhpv;
9596 	dtrace_meta_t *meta = dtrace_meta_pid;
9597 	dtrace_mops_t *mops = &meta->dtm_mops;
9598 
9599 	provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
9600 	str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
9601 	    provider->dofpv_strtab * dof->dofh_secsize);
9602 
9603 	strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
9604 
9605 	/*
9606 	 * Create the provider.
9607 	 */
9608 	dtrace_dofprov2hprov(&dhpv, provider, strtab);
9609 
9610 	mops->dtms_remove_pid(meta->dtm_arg, &dhpv, pid);
9611 
9612 	meta->dtm_count--;
9613 }
9614 
9615 static void
9616 dtrace_helper_provider_remove(dof_helper_t *dhp, pid_t pid)
9617 {
9618 	uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
9619 	dof_hdr_t *dof = (dof_hdr_t *)daddr;
9620 	int i;
9621 
9622 	ASSERT(MUTEX_HELD(&dtrace_meta_lock));
9623 
9624 	for (i = 0; i < dof->dofh_secnum; i++) {
9625 		dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
9626 		    dof->dofh_secoff + i * dof->dofh_secsize);
9627 
9628 		if (sec->dofs_type != DOF_SECT_PROVIDER)
9629 			continue;
9630 
9631 		dtrace_helper_provider_remove_one(dhp, sec, pid);
9632 	}
9633 }
9634 
9635 /*
9636  * DTrace Meta Provider-to-Framework API Functions
9637  *
9638  * These functions implement the Meta Provider-to-Framework API, as described
9639  * in <sys/dtrace.h>.
9640  */
9641 int
9642 dtrace_meta_register(const char *name, const dtrace_mops_t *mops, void *arg,
9643     dtrace_meta_provider_id_t *idp)
9644 {
9645 	dtrace_meta_t *meta;
9646 	dtrace_helpers_t *help, *next;
9647 	int i;
9648 
9649 	*idp = DTRACE_METAPROVNONE;
9650 
9651 	/*
9652 	 * We strictly don't need the name, but we hold onto it for
9653 	 * debuggability. All hail error queues!
9654 	 */
9655 	if (name == NULL) {
9656 		cmn_err(CE_WARN, "failed to register meta-provider: "
9657 		    "invalid name");
9658 		return (EINVAL);
9659 	}
9660 
9661 	if (mops == NULL ||
9662 	    mops->dtms_create_probe == NULL ||
9663 	    mops->dtms_provide_pid == NULL ||
9664 	    mops->dtms_remove_pid == NULL) {
9665 		cmn_err(CE_WARN, "failed to register meta-register %s: "
9666 		    "invalid ops", name);
9667 		return (EINVAL);
9668 	}
9669 
9670 	meta = kmem_zalloc(sizeof (dtrace_meta_t), KM_SLEEP);
9671 	meta->dtm_mops = *mops;
9672 	meta->dtm_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
9673 	(void) strcpy(meta->dtm_name, name);
9674 	meta->dtm_arg = arg;
9675 
9676 	mutex_enter(&dtrace_meta_lock);
9677 	mutex_enter(&dtrace_lock);
9678 
9679 	if (dtrace_meta_pid != NULL) {
9680 		mutex_exit(&dtrace_lock);
9681 		mutex_exit(&dtrace_meta_lock);
9682 		cmn_err(CE_WARN, "failed to register meta-register %s: "
9683 		    "user-land meta-provider exists", name);
9684 		kmem_free(meta->dtm_name, strlen(meta->dtm_name) + 1);
9685 		kmem_free(meta, sizeof (dtrace_meta_t));
9686 		return (EINVAL);
9687 	}
9688 
9689 	dtrace_meta_pid = meta;
9690 	*idp = (dtrace_meta_provider_id_t)meta;
9691 
9692 	/*
9693 	 * If there are providers and probes ready to go, pass them
9694 	 * off to the new meta provider now.
9695 	 */
9696 
9697 	help = dtrace_deferred_pid;
9698 	dtrace_deferred_pid = NULL;
9699 
9700 	mutex_exit(&dtrace_lock);
9701 
9702 	while (help != NULL) {
9703 		for (i = 0; i < help->dthps_nprovs; i++) {
9704 			dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
9705 			    help->dthps_pid);
9706 		}
9707 
9708 		next = help->dthps_next;
9709 		help->dthps_next = NULL;
9710 		help->dthps_prev = NULL;
9711 		help->dthps_deferred = 0;
9712 		help = next;
9713 	}
9714 
9715 	mutex_exit(&dtrace_meta_lock);
9716 
9717 	return (0);
9718 }
9719 
9720 int
9721 dtrace_meta_unregister(dtrace_meta_provider_id_t id)
9722 {
9723 	dtrace_meta_t **pp, *old = (dtrace_meta_t *)id;
9724 
9725 	mutex_enter(&dtrace_meta_lock);
9726 	mutex_enter(&dtrace_lock);
9727 
9728 	if (old == dtrace_meta_pid) {
9729 		pp = &dtrace_meta_pid;
9730 	} else {
9731 		panic("attempt to unregister non-existent "
9732 		    "dtrace meta-provider %p\n", (void *)old);
9733 	}
9734 
9735 	if (old->dtm_count != 0) {
9736 		mutex_exit(&dtrace_lock);
9737 		mutex_exit(&dtrace_meta_lock);
9738 		return (EBUSY);
9739 	}
9740 
9741 	*pp = NULL;
9742 
9743 	mutex_exit(&dtrace_lock);
9744 	mutex_exit(&dtrace_meta_lock);
9745 
9746 	kmem_free(old->dtm_name, strlen(old->dtm_name) + 1);
9747 	kmem_free(old, sizeof (dtrace_meta_t));
9748 
9749 	return (0);
9750 }
9751 
9752 
9753 /*
9754  * DTrace DIF Object Functions
9755  */
9756 static int
9757 dtrace_difo_err(uint_t pc, const char *format, ...)
9758 {
9759 	if (dtrace_err_verbose) {
9760 		va_list alist;
9761 
9762 		(void) uprintf("dtrace DIF object error: [%u]: ", pc);
9763 		va_start(alist, format);
9764 		(void) vuprintf(format, alist);
9765 		va_end(alist);
9766 	}
9767 
9768 #ifdef DTRACE_ERRDEBUG
9769 	dtrace_errdebug(format);
9770 #endif
9771 	return (1);
9772 }
9773 
9774 /*
9775  * Validate a DTrace DIF object by checking the IR instructions.  The following
9776  * rules are currently enforced by dtrace_difo_validate():
9777  *
9778  * 1. Each instruction must have a valid opcode
9779  * 2. Each register, string, variable, or subroutine reference must be valid
9780  * 3. No instruction can modify register %r0 (must be zero)
9781  * 4. All instruction reserved bits must be set to zero
9782  * 5. The last instruction must be a "ret" instruction
9783  * 6. All branch targets must reference a valid instruction _after_ the branch
9784  */
9785 static int
9786 dtrace_difo_validate(dtrace_difo_t *dp, dtrace_vstate_t *vstate, uint_t nregs,
9787     cred_t *cr)
9788 {
9789 	int err = 0, i;
9790 	int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
9791 	int kcheckload;
9792 	uint_t pc;
9793 	int maxglobal = -1, maxlocal = -1, maxtlocal = -1;
9794 
9795 	kcheckload = cr == NULL ||
9796 	    (vstate->dtvs_state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) == 0;
9797 
9798 	dp->dtdo_destructive = 0;
9799 
9800 	for (pc = 0; pc < dp->dtdo_len && err == 0; pc++) {
9801 		dif_instr_t instr = dp->dtdo_buf[pc];
9802 
9803 		uint_t r1 = DIF_INSTR_R1(instr);
9804 		uint_t r2 = DIF_INSTR_R2(instr);
9805 		uint_t rd = DIF_INSTR_RD(instr);
9806 		uint_t rs = DIF_INSTR_RS(instr);
9807 		uint_t label = DIF_INSTR_LABEL(instr);
9808 		uint_t v = DIF_INSTR_VAR(instr);
9809 		uint_t subr = DIF_INSTR_SUBR(instr);
9810 		uint_t type = DIF_INSTR_TYPE(instr);
9811 		uint_t op = DIF_INSTR_OP(instr);
9812 
9813 		switch (op) {
9814 		case DIF_OP_OR:
9815 		case DIF_OP_XOR:
9816 		case DIF_OP_AND:
9817 		case DIF_OP_SLL:
9818 		case DIF_OP_SRL:
9819 		case DIF_OP_SRA:
9820 		case DIF_OP_SUB:
9821 		case DIF_OP_ADD:
9822 		case DIF_OP_MUL:
9823 		case DIF_OP_SDIV:
9824 		case DIF_OP_UDIV:
9825 		case DIF_OP_SREM:
9826 		case DIF_OP_UREM:
9827 		case DIF_OP_COPYS:
9828 			if (r1 >= nregs)
9829 				err += efunc(pc, "invalid register %u\n", r1);
9830 			if (r2 >= nregs)
9831 				err += efunc(pc, "invalid register %u\n", r2);
9832 			if (rd >= nregs)
9833 				err += efunc(pc, "invalid register %u\n", rd);
9834 			if (rd == 0)
9835 				err += efunc(pc, "cannot write to %%r0\n");
9836 			break;
9837 		case DIF_OP_NOT:
9838 		case DIF_OP_MOV:
9839 		case DIF_OP_ALLOCS:
9840 			if (r1 >= nregs)
9841 				err += efunc(pc, "invalid register %u\n", r1);
9842 			if (r2 != 0)
9843 				err += efunc(pc, "non-zero reserved bits\n");
9844 			if (rd >= nregs)
9845 				err += efunc(pc, "invalid register %u\n", rd);
9846 			if (rd == 0)
9847 				err += efunc(pc, "cannot write to %%r0\n");
9848 			break;
9849 		case DIF_OP_LDSB:
9850 		case DIF_OP_LDSH:
9851 		case DIF_OP_LDSW:
9852 		case DIF_OP_LDUB:
9853 		case DIF_OP_LDUH:
9854 		case DIF_OP_LDUW:
9855 		case DIF_OP_LDX:
9856 			if (r1 >= nregs)
9857 				err += efunc(pc, "invalid register %u\n", r1);
9858 			if (r2 != 0)
9859 				err += efunc(pc, "non-zero reserved bits\n");
9860 			if (rd >= nregs)
9861 				err += efunc(pc, "invalid register %u\n", rd);
9862 			if (rd == 0)
9863 				err += efunc(pc, "cannot write to %%r0\n");
9864 			if (kcheckload)
9865 				dp->dtdo_buf[pc] = DIF_INSTR_LOAD(op +
9866 				    DIF_OP_RLDSB - DIF_OP_LDSB, r1, rd);
9867 			break;
9868 		case DIF_OP_RLDSB:
9869 		case DIF_OP_RLDSH:
9870 		case DIF_OP_RLDSW:
9871 		case DIF_OP_RLDUB:
9872 		case DIF_OP_RLDUH:
9873 		case DIF_OP_RLDUW:
9874 		case DIF_OP_RLDX:
9875 			if (r1 >= nregs)
9876 				err += efunc(pc, "invalid register %u\n", r1);
9877 			if (r2 != 0)
9878 				err += efunc(pc, "non-zero reserved bits\n");
9879 			if (rd >= nregs)
9880 				err += efunc(pc, "invalid register %u\n", rd);
9881 			if (rd == 0)
9882 				err += efunc(pc, "cannot write to %%r0\n");
9883 			break;
9884 		case DIF_OP_ULDSB:
9885 		case DIF_OP_ULDSH:
9886 		case DIF_OP_ULDSW:
9887 		case DIF_OP_ULDUB:
9888 		case DIF_OP_ULDUH:
9889 		case DIF_OP_ULDUW:
9890 		case DIF_OP_ULDX:
9891 			if (r1 >= nregs)
9892 				err += efunc(pc, "invalid register %u\n", r1);
9893 			if (r2 != 0)
9894 				err += efunc(pc, "non-zero reserved bits\n");
9895 			if (rd >= nregs)
9896 				err += efunc(pc, "invalid register %u\n", rd);
9897 			if (rd == 0)
9898 				err += efunc(pc, "cannot write to %%r0\n");
9899 			break;
9900 		case DIF_OP_STB:
9901 		case DIF_OP_STH:
9902 		case DIF_OP_STW:
9903 		case DIF_OP_STX:
9904 			if (r1 >= nregs)
9905 				err += efunc(pc, "invalid register %u\n", r1);
9906 			if (r2 != 0)
9907 				err += efunc(pc, "non-zero reserved bits\n");
9908 			if (rd >= nregs)
9909 				err += efunc(pc, "invalid register %u\n", rd);
9910 			if (rd == 0)
9911 				err += efunc(pc, "cannot write to 0 address\n");
9912 			break;
9913 		case DIF_OP_CMP:
9914 		case DIF_OP_SCMP:
9915 			if (r1 >= nregs)
9916 				err += efunc(pc, "invalid register %u\n", r1);
9917 			if (r2 >= nregs)
9918 				err += efunc(pc, "invalid register %u\n", r2);
9919 			if (rd != 0)
9920 				err += efunc(pc, "non-zero reserved bits\n");
9921 			break;
9922 		case DIF_OP_TST:
9923 			if (r1 >= nregs)
9924 				err += efunc(pc, "invalid register %u\n", r1);
9925 			if (r2 != 0 || rd != 0)
9926 				err += efunc(pc, "non-zero reserved bits\n");
9927 			break;
9928 		case DIF_OP_BA:
9929 		case DIF_OP_BE:
9930 		case DIF_OP_BNE:
9931 		case DIF_OP_BG:
9932 		case DIF_OP_BGU:
9933 		case DIF_OP_BGE:
9934 		case DIF_OP_BGEU:
9935 		case DIF_OP_BL:
9936 		case DIF_OP_BLU:
9937 		case DIF_OP_BLE:
9938 		case DIF_OP_BLEU:
9939 			if (label >= dp->dtdo_len) {
9940 				err += efunc(pc, "invalid branch target %u\n",
9941 				    label);
9942 			}
9943 			if (label <= pc) {
9944 				err += efunc(pc, "backward branch to %u\n",
9945 				    label);
9946 			}
9947 			break;
9948 		case DIF_OP_RET:
9949 			if (r1 != 0 || r2 != 0)
9950 				err += efunc(pc, "non-zero reserved bits\n");
9951 			if (rd >= nregs)
9952 				err += efunc(pc, "invalid register %u\n", rd);
9953 			break;
9954 		case DIF_OP_NOP:
9955 		case DIF_OP_POPTS:
9956 		case DIF_OP_FLUSHTS:
9957 			if (r1 != 0 || r2 != 0 || rd != 0)
9958 				err += efunc(pc, "non-zero reserved bits\n");
9959 			break;
9960 		case DIF_OP_SETX:
9961 			if (DIF_INSTR_INTEGER(instr) >= dp->dtdo_intlen) {
9962 				err += efunc(pc, "invalid integer ref %u\n",
9963 				    DIF_INSTR_INTEGER(instr));
9964 			}
9965 			if (rd >= nregs)
9966 				err += efunc(pc, "invalid register %u\n", rd);
9967 			if (rd == 0)
9968 				err += efunc(pc, "cannot write to %%r0\n");
9969 			break;
9970 		case DIF_OP_SETS:
9971 			if (DIF_INSTR_STRING(instr) >= dp->dtdo_strlen) {
9972 				err += efunc(pc, "invalid string ref %u\n",
9973 				    DIF_INSTR_STRING(instr));
9974 			}
9975 			if (rd >= nregs)
9976 				err += efunc(pc, "invalid register %u\n", rd);
9977 			if (rd == 0)
9978 				err += efunc(pc, "cannot write to %%r0\n");
9979 			break;
9980 		case DIF_OP_LDGA:
9981 		case DIF_OP_LDTA:
9982 			if (r1 > DIF_VAR_ARRAY_MAX)
9983 				err += efunc(pc, "invalid array %u\n", r1);
9984 			if (r2 >= nregs)
9985 				err += efunc(pc, "invalid register %u\n", r2);
9986 			if (rd >= nregs)
9987 				err += efunc(pc, "invalid register %u\n", rd);
9988 			if (rd == 0)
9989 				err += efunc(pc, "cannot write to %%r0\n");
9990 			break;
9991 		case DIF_OP_LDGS:
9992 		case DIF_OP_LDTS:
9993 		case DIF_OP_LDLS:
9994 		case DIF_OP_LDGAA:
9995 		case DIF_OP_LDTAA:
9996 			if (v < DIF_VAR_OTHER_MIN || v > DIF_VAR_OTHER_MAX)
9997 				err += efunc(pc, "invalid variable %u\n", v);
9998 			if (rd >= nregs)
9999 				err += efunc(pc, "invalid register %u\n", rd);
10000 			if (rd == 0)
10001 				err += efunc(pc, "cannot write to %%r0\n");
10002 			break;
10003 		case DIF_OP_STGS:
10004 		case DIF_OP_STTS:
10005 		case DIF_OP_STLS:
10006 		case DIF_OP_STGAA:
10007 		case DIF_OP_STTAA:
10008 			if (v < DIF_VAR_OTHER_UBASE || v > DIF_VAR_OTHER_MAX)
10009 				err += efunc(pc, "invalid variable %u\n", v);
10010 			if (rs >= nregs)
10011 				err += efunc(pc, "invalid register %u\n", rd);
10012 			break;
10013 		case DIF_OP_CALL:
10014 			if (subr > DIF_SUBR_MAX)
10015 				err += efunc(pc, "invalid subr %u\n", subr);
10016 			if (rd >= nregs)
10017 				err += efunc(pc, "invalid register %u\n", rd);
10018 			if (rd == 0)
10019 				err += efunc(pc, "cannot write to %%r0\n");
10020 
10021 			if (subr == DIF_SUBR_COPYOUT ||
10022 			    subr == DIF_SUBR_COPYOUTSTR) {
10023 				dp->dtdo_destructive = 1;
10024 			}
10025 
10026 			if (subr == DIF_SUBR_GETF) {
10027 #ifdef __FreeBSD__
10028 				err += efunc(pc, "getf() not supported");
10029 #else
10030 				/*
10031 				 * If we have a getf() we need to record that
10032 				 * in our state.  Note that our state can be
10033 				 * NULL if this is a helper -- but in that
10034 				 * case, the call to getf() is itself illegal,
10035 				 * and will be caught (slightly later) when
10036 				 * the helper is validated.
10037 				 */
10038 				if (vstate->dtvs_state != NULL)
10039 					vstate->dtvs_state->dts_getf++;
10040 #endif
10041 			}
10042 
10043 			break;
10044 		case DIF_OP_PUSHTR:
10045 			if (type != DIF_TYPE_STRING && type != DIF_TYPE_CTF)
10046 				err += efunc(pc, "invalid ref type %u\n", type);
10047 			if (r2 >= nregs)
10048 				err += efunc(pc, "invalid register %u\n", r2);
10049 			if (rs >= nregs)
10050 				err += efunc(pc, "invalid register %u\n", rs);
10051 			break;
10052 		case DIF_OP_PUSHTV:
10053 			if (type != DIF_TYPE_CTF)
10054 				err += efunc(pc, "invalid val type %u\n", type);
10055 			if (r2 >= nregs)
10056 				err += efunc(pc, "invalid register %u\n", r2);
10057 			if (rs >= nregs)
10058 				err += efunc(pc, "invalid register %u\n", rs);
10059 			break;
10060 		default:
10061 			err += efunc(pc, "invalid opcode %u\n",
10062 			    DIF_INSTR_OP(instr));
10063 		}
10064 	}
10065 
10066 	if (dp->dtdo_len != 0 &&
10067 	    DIF_INSTR_OP(dp->dtdo_buf[dp->dtdo_len - 1]) != DIF_OP_RET) {
10068 		err += efunc(dp->dtdo_len - 1,
10069 		    "expected 'ret' as last DIF instruction\n");
10070 	}
10071 
10072 	if (!(dp->dtdo_rtype.dtdt_flags & (DIF_TF_BYREF | DIF_TF_BYUREF))) {
10073 		/*
10074 		 * If we're not returning by reference, the size must be either
10075 		 * 0 or the size of one of the base types.
10076 		 */
10077 		switch (dp->dtdo_rtype.dtdt_size) {
10078 		case 0:
10079 		case sizeof (uint8_t):
10080 		case sizeof (uint16_t):
10081 		case sizeof (uint32_t):
10082 		case sizeof (uint64_t):
10083 			break;
10084 
10085 		default:
10086 			err += efunc(dp->dtdo_len - 1, "bad return size\n");
10087 		}
10088 	}
10089 
10090 	for (i = 0; i < dp->dtdo_varlen && err == 0; i++) {
10091 		dtrace_difv_t *v = &dp->dtdo_vartab[i], *existing = NULL;
10092 		dtrace_diftype_t *vt, *et;
10093 		uint_t id, ndx;
10094 
10095 		if (v->dtdv_scope != DIFV_SCOPE_GLOBAL &&
10096 		    v->dtdv_scope != DIFV_SCOPE_THREAD &&
10097 		    v->dtdv_scope != DIFV_SCOPE_LOCAL) {
10098 			err += efunc(i, "unrecognized variable scope %d\n",
10099 			    v->dtdv_scope);
10100 			break;
10101 		}
10102 
10103 		if (v->dtdv_kind != DIFV_KIND_ARRAY &&
10104 		    v->dtdv_kind != DIFV_KIND_SCALAR) {
10105 			err += efunc(i, "unrecognized variable type %d\n",
10106 			    v->dtdv_kind);
10107 			break;
10108 		}
10109 
10110 		if ((id = v->dtdv_id) > DIF_VARIABLE_MAX) {
10111 			err += efunc(i, "%d exceeds variable id limit\n", id);
10112 			break;
10113 		}
10114 
10115 		if (id < DIF_VAR_OTHER_UBASE)
10116 			continue;
10117 
10118 		/*
10119 		 * For user-defined variables, we need to check that this
10120 		 * definition is identical to any previous definition that we
10121 		 * encountered.
10122 		 */
10123 		ndx = id - DIF_VAR_OTHER_UBASE;
10124 
10125 		switch (v->dtdv_scope) {
10126 		case DIFV_SCOPE_GLOBAL:
10127 			if (maxglobal == -1 || ndx > maxglobal)
10128 				maxglobal = ndx;
10129 
10130 			if (ndx < vstate->dtvs_nglobals) {
10131 				dtrace_statvar_t *svar;
10132 
10133 				if ((svar = vstate->dtvs_globals[ndx]) != NULL)
10134 					existing = &svar->dtsv_var;
10135 			}
10136 
10137 			break;
10138 
10139 		case DIFV_SCOPE_THREAD:
10140 			if (maxtlocal == -1 || ndx > maxtlocal)
10141 				maxtlocal = ndx;
10142 
10143 			if (ndx < vstate->dtvs_ntlocals)
10144 				existing = &vstate->dtvs_tlocals[ndx];
10145 			break;
10146 
10147 		case DIFV_SCOPE_LOCAL:
10148 			if (maxlocal == -1 || ndx > maxlocal)
10149 				maxlocal = ndx;
10150 
10151 			if (ndx < vstate->dtvs_nlocals) {
10152 				dtrace_statvar_t *svar;
10153 
10154 				if ((svar = vstate->dtvs_locals[ndx]) != NULL)
10155 					existing = &svar->dtsv_var;
10156 			}
10157 
10158 			break;
10159 		}
10160 
10161 		vt = &v->dtdv_type;
10162 
10163 		if (vt->dtdt_flags & DIF_TF_BYREF) {
10164 			if (vt->dtdt_size == 0) {
10165 				err += efunc(i, "zero-sized variable\n");
10166 				break;
10167 			}
10168 
10169 			if ((v->dtdv_scope == DIFV_SCOPE_GLOBAL ||
10170 			    v->dtdv_scope == DIFV_SCOPE_LOCAL) &&
10171 			    vt->dtdt_size > dtrace_statvar_maxsize) {
10172 				err += efunc(i, "oversized by-ref static\n");
10173 				break;
10174 			}
10175 		}
10176 
10177 		if (existing == NULL || existing->dtdv_id == 0)
10178 			continue;
10179 
10180 		ASSERT(existing->dtdv_id == v->dtdv_id);
10181 		ASSERT(existing->dtdv_scope == v->dtdv_scope);
10182 
10183 		if (existing->dtdv_kind != v->dtdv_kind)
10184 			err += efunc(i, "%d changed variable kind\n", id);
10185 
10186 		et = &existing->dtdv_type;
10187 
10188 		if (vt->dtdt_flags != et->dtdt_flags) {
10189 			err += efunc(i, "%d changed variable type flags\n", id);
10190 			break;
10191 		}
10192 
10193 		if (vt->dtdt_size != 0 && vt->dtdt_size != et->dtdt_size) {
10194 			err += efunc(i, "%d changed variable type size\n", id);
10195 			break;
10196 		}
10197 	}
10198 
10199 	for (pc = 0; pc < dp->dtdo_len && err == 0; pc++) {
10200 		dif_instr_t instr = dp->dtdo_buf[pc];
10201 
10202 		uint_t v = DIF_INSTR_VAR(instr);
10203 		uint_t op = DIF_INSTR_OP(instr);
10204 
10205 		switch (op) {
10206 		case DIF_OP_LDGS:
10207 		case DIF_OP_LDGAA:
10208 		case DIF_OP_STGS:
10209 		case DIF_OP_STGAA:
10210 			if (v > DIF_VAR_OTHER_UBASE + maxglobal)
10211 				err += efunc(pc, "invalid variable %u\n", v);
10212 			break;
10213 		case DIF_OP_LDTS:
10214 		case DIF_OP_LDTAA:
10215 		case DIF_OP_STTS:
10216 		case DIF_OP_STTAA:
10217 			if (v > DIF_VAR_OTHER_UBASE + maxtlocal)
10218 				err += efunc(pc, "invalid variable %u\n", v);
10219 			break;
10220 		case DIF_OP_LDLS:
10221 		case DIF_OP_STLS:
10222 			if (v > DIF_VAR_OTHER_UBASE + maxlocal)
10223 				err += efunc(pc, "invalid variable %u\n", v);
10224 			break;
10225 		default:
10226 			break;
10227 		}
10228 	}
10229 
10230 	return (err);
10231 }
10232 
10233 /*
10234  * Validate a DTrace DIF object that it is to be used as a helper.  Helpers
10235  * are much more constrained than normal DIFOs.  Specifically, they may
10236  * not:
10237  *
10238  * 1. Make calls to subroutines other than copyin(), copyinstr() or
10239  *    miscellaneous string routines
10240  * 2. Access DTrace variables other than the args[] array, and the
10241  *    curthread, pid, ppid, tid, execname, zonename, uid and gid variables.
10242  * 3. Have thread-local variables.
10243  * 4. Have dynamic variables.
10244  */
10245 static int
10246 dtrace_difo_validate_helper(dtrace_difo_t *dp)
10247 {
10248 	int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
10249 	int err = 0;
10250 	uint_t pc;
10251 
10252 	for (pc = 0; pc < dp->dtdo_len; pc++) {
10253 		dif_instr_t instr = dp->dtdo_buf[pc];
10254 
10255 		uint_t v = DIF_INSTR_VAR(instr);
10256 		uint_t subr = DIF_INSTR_SUBR(instr);
10257 		uint_t op = DIF_INSTR_OP(instr);
10258 
10259 		switch (op) {
10260 		case DIF_OP_OR:
10261 		case DIF_OP_XOR:
10262 		case DIF_OP_AND:
10263 		case DIF_OP_SLL:
10264 		case DIF_OP_SRL:
10265 		case DIF_OP_SRA:
10266 		case DIF_OP_SUB:
10267 		case DIF_OP_ADD:
10268 		case DIF_OP_MUL:
10269 		case DIF_OP_SDIV:
10270 		case DIF_OP_UDIV:
10271 		case DIF_OP_SREM:
10272 		case DIF_OP_UREM:
10273 		case DIF_OP_COPYS:
10274 		case DIF_OP_NOT:
10275 		case DIF_OP_MOV:
10276 		case DIF_OP_RLDSB:
10277 		case DIF_OP_RLDSH:
10278 		case DIF_OP_RLDSW:
10279 		case DIF_OP_RLDUB:
10280 		case DIF_OP_RLDUH:
10281 		case DIF_OP_RLDUW:
10282 		case DIF_OP_RLDX:
10283 		case DIF_OP_ULDSB:
10284 		case DIF_OP_ULDSH:
10285 		case DIF_OP_ULDSW:
10286 		case DIF_OP_ULDUB:
10287 		case DIF_OP_ULDUH:
10288 		case DIF_OP_ULDUW:
10289 		case DIF_OP_ULDX:
10290 		case DIF_OP_STB:
10291 		case DIF_OP_STH:
10292 		case DIF_OP_STW:
10293 		case DIF_OP_STX:
10294 		case DIF_OP_ALLOCS:
10295 		case DIF_OP_CMP:
10296 		case DIF_OP_SCMP:
10297 		case DIF_OP_TST:
10298 		case DIF_OP_BA:
10299 		case DIF_OP_BE:
10300 		case DIF_OP_BNE:
10301 		case DIF_OP_BG:
10302 		case DIF_OP_BGU:
10303 		case DIF_OP_BGE:
10304 		case DIF_OP_BGEU:
10305 		case DIF_OP_BL:
10306 		case DIF_OP_BLU:
10307 		case DIF_OP_BLE:
10308 		case DIF_OP_BLEU:
10309 		case DIF_OP_RET:
10310 		case DIF_OP_NOP:
10311 		case DIF_OP_POPTS:
10312 		case DIF_OP_FLUSHTS:
10313 		case DIF_OP_SETX:
10314 		case DIF_OP_SETS:
10315 		case DIF_OP_LDGA:
10316 		case DIF_OP_LDLS:
10317 		case DIF_OP_STGS:
10318 		case DIF_OP_STLS:
10319 		case DIF_OP_PUSHTR:
10320 		case DIF_OP_PUSHTV:
10321 			break;
10322 
10323 		case DIF_OP_LDGS:
10324 			if (v >= DIF_VAR_OTHER_UBASE)
10325 				break;
10326 
10327 			if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9)
10328 				break;
10329 
10330 			if (v == DIF_VAR_CURTHREAD || v == DIF_VAR_PID ||
10331 			    v == DIF_VAR_PPID || v == DIF_VAR_TID ||
10332 			    v == DIF_VAR_EXECARGS ||
10333 			    v == DIF_VAR_EXECNAME || v == DIF_VAR_ZONENAME ||
10334 			    v == DIF_VAR_UID || v == DIF_VAR_GID)
10335 				break;
10336 
10337 			err += efunc(pc, "illegal variable %u\n", v);
10338 			break;
10339 
10340 		case DIF_OP_LDTA:
10341 		case DIF_OP_LDTS:
10342 		case DIF_OP_LDGAA:
10343 		case DIF_OP_LDTAA:
10344 			err += efunc(pc, "illegal dynamic variable load\n");
10345 			break;
10346 
10347 		case DIF_OP_STTS:
10348 		case DIF_OP_STGAA:
10349 		case DIF_OP_STTAA:
10350 			err += efunc(pc, "illegal dynamic variable store\n");
10351 			break;
10352 
10353 		case DIF_OP_CALL:
10354 			if (subr == DIF_SUBR_ALLOCA ||
10355 			    subr == DIF_SUBR_BCOPY ||
10356 			    subr == DIF_SUBR_COPYIN ||
10357 			    subr == DIF_SUBR_COPYINTO ||
10358 			    subr == DIF_SUBR_COPYINSTR ||
10359 			    subr == DIF_SUBR_INDEX ||
10360 			    subr == DIF_SUBR_INET_NTOA ||
10361 			    subr == DIF_SUBR_INET_NTOA6 ||
10362 			    subr == DIF_SUBR_INET_NTOP ||
10363 			    subr == DIF_SUBR_JSON ||
10364 			    subr == DIF_SUBR_LLTOSTR ||
10365 			    subr == DIF_SUBR_STRTOLL ||
10366 			    subr == DIF_SUBR_RINDEX ||
10367 			    subr == DIF_SUBR_STRCHR ||
10368 			    subr == DIF_SUBR_STRJOIN ||
10369 			    subr == DIF_SUBR_STRRCHR ||
10370 			    subr == DIF_SUBR_STRSTR ||
10371 			    subr == DIF_SUBR_HTONS ||
10372 			    subr == DIF_SUBR_HTONL ||
10373 			    subr == DIF_SUBR_HTONLL ||
10374 			    subr == DIF_SUBR_NTOHS ||
10375 			    subr == DIF_SUBR_NTOHL ||
10376 			    subr == DIF_SUBR_NTOHLL ||
10377 			    subr == DIF_SUBR_MEMREF)
10378 				break;
10379 #ifdef __FreeBSD__
10380 			if (subr == DIF_SUBR_MEMSTR)
10381 				break;
10382 #endif
10383 
10384 			err += efunc(pc, "invalid subr %u\n", subr);
10385 			break;
10386 
10387 		default:
10388 			err += efunc(pc, "invalid opcode %u\n",
10389 			    DIF_INSTR_OP(instr));
10390 		}
10391 	}
10392 
10393 	return (err);
10394 }
10395 
10396 /*
10397  * Returns 1 if the expression in the DIF object can be cached on a per-thread
10398  * basis; 0 if not.
10399  */
10400 static int
10401 dtrace_difo_cacheable(dtrace_difo_t *dp)
10402 {
10403 	int i;
10404 
10405 	if (dp == NULL)
10406 		return (0);
10407 
10408 	for (i = 0; i < dp->dtdo_varlen; i++) {
10409 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
10410 
10411 		if (v->dtdv_scope != DIFV_SCOPE_GLOBAL)
10412 			continue;
10413 
10414 		switch (v->dtdv_id) {
10415 		case DIF_VAR_CURTHREAD:
10416 		case DIF_VAR_PID:
10417 		case DIF_VAR_TID:
10418 		case DIF_VAR_EXECARGS:
10419 		case DIF_VAR_EXECNAME:
10420 		case DIF_VAR_ZONENAME:
10421 			break;
10422 
10423 		default:
10424 			return (0);
10425 		}
10426 	}
10427 
10428 	/*
10429 	 * This DIF object may be cacheable.  Now we need to look for any
10430 	 * array loading instructions, any memory loading instructions, or
10431 	 * any stores to thread-local variables.
10432 	 */
10433 	for (i = 0; i < dp->dtdo_len; i++) {
10434 		uint_t op = DIF_INSTR_OP(dp->dtdo_buf[i]);
10435 
10436 		if ((op >= DIF_OP_LDSB && op <= DIF_OP_LDX) ||
10437 		    (op >= DIF_OP_ULDSB && op <= DIF_OP_ULDX) ||
10438 		    (op >= DIF_OP_RLDSB && op <= DIF_OP_RLDX) ||
10439 		    op == DIF_OP_LDGA || op == DIF_OP_STTS)
10440 			return (0);
10441 	}
10442 
10443 	return (1);
10444 }
10445 
10446 static void
10447 dtrace_difo_hold(dtrace_difo_t *dp)
10448 {
10449 	int i;
10450 
10451 	ASSERT(MUTEX_HELD(&dtrace_lock));
10452 
10453 	dp->dtdo_refcnt++;
10454 	ASSERT(dp->dtdo_refcnt != 0);
10455 
10456 	/*
10457 	 * We need to check this DIF object for references to the variable
10458 	 * DIF_VAR_VTIMESTAMP.
10459 	 */
10460 	for (i = 0; i < dp->dtdo_varlen; i++) {
10461 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
10462 
10463 		if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
10464 			continue;
10465 
10466 		if (dtrace_vtime_references++ == 0)
10467 			dtrace_vtime_enable();
10468 	}
10469 }
10470 
10471 /*
10472  * This routine calculates the dynamic variable chunksize for a given DIF
10473  * object.  The calculation is not fool-proof, and can probably be tricked by
10474  * malicious DIF -- but it works for all compiler-generated DIF.  Because this
10475  * calculation is likely imperfect, dtrace_dynvar() is able to gracefully fail
10476  * if a dynamic variable size exceeds the chunksize.
10477  */
10478 static void
10479 dtrace_difo_chunksize(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
10480 {
10481 	uint64_t sval = 0;
10482 	dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
10483 	const dif_instr_t *text = dp->dtdo_buf;
10484 	uint_t pc, srd = 0;
10485 	uint_t ttop = 0;
10486 	size_t size, ksize;
10487 	uint_t id, i;
10488 
10489 	for (pc = 0; pc < dp->dtdo_len; pc++) {
10490 		dif_instr_t instr = text[pc];
10491 		uint_t op = DIF_INSTR_OP(instr);
10492 		uint_t rd = DIF_INSTR_RD(instr);
10493 		uint_t r1 = DIF_INSTR_R1(instr);
10494 		uint_t nkeys = 0;
10495 		uchar_t scope = 0;
10496 
10497 		dtrace_key_t *key = tupregs;
10498 
10499 		switch (op) {
10500 		case DIF_OP_SETX:
10501 			sval = dp->dtdo_inttab[DIF_INSTR_INTEGER(instr)];
10502 			srd = rd;
10503 			continue;
10504 
10505 		case DIF_OP_STTS:
10506 			key = &tupregs[DIF_DTR_NREGS];
10507 			key[0].dttk_size = 0;
10508 			key[1].dttk_size = 0;
10509 			nkeys = 2;
10510 			scope = DIFV_SCOPE_THREAD;
10511 			break;
10512 
10513 		case DIF_OP_STGAA:
10514 		case DIF_OP_STTAA:
10515 			nkeys = ttop;
10516 
10517 			if (DIF_INSTR_OP(instr) == DIF_OP_STTAA)
10518 				key[nkeys++].dttk_size = 0;
10519 
10520 			key[nkeys++].dttk_size = 0;
10521 
10522 			if (op == DIF_OP_STTAA) {
10523 				scope = DIFV_SCOPE_THREAD;
10524 			} else {
10525 				scope = DIFV_SCOPE_GLOBAL;
10526 			}
10527 
10528 			break;
10529 
10530 		case DIF_OP_PUSHTR:
10531 			if (ttop == DIF_DTR_NREGS)
10532 				return;
10533 
10534 			if ((srd == 0 || sval == 0) && r1 == DIF_TYPE_STRING) {
10535 				/*
10536 				 * If the register for the size of the "pushtr"
10537 				 * is %r0 (or the value is 0) and the type is
10538 				 * a string, we'll use the system-wide default
10539 				 * string size.
10540 				 */
10541 				tupregs[ttop++].dttk_size =
10542 				    dtrace_strsize_default;
10543 			} else {
10544 				if (srd == 0)
10545 					return;
10546 
10547 				if (sval > LONG_MAX)
10548 					return;
10549 
10550 				tupregs[ttop++].dttk_size = sval;
10551 			}
10552 
10553 			break;
10554 
10555 		case DIF_OP_PUSHTV:
10556 			if (ttop == DIF_DTR_NREGS)
10557 				return;
10558 
10559 			tupregs[ttop++].dttk_size = 0;
10560 			break;
10561 
10562 		case DIF_OP_FLUSHTS:
10563 			ttop = 0;
10564 			break;
10565 
10566 		case DIF_OP_POPTS:
10567 			if (ttop != 0)
10568 				ttop--;
10569 			break;
10570 		}
10571 
10572 		sval = 0;
10573 		srd = 0;
10574 
10575 		if (nkeys == 0)
10576 			continue;
10577 
10578 		/*
10579 		 * We have a dynamic variable allocation; calculate its size.
10580 		 */
10581 		for (ksize = 0, i = 0; i < nkeys; i++)
10582 			ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
10583 
10584 		size = sizeof (dtrace_dynvar_t);
10585 		size += sizeof (dtrace_key_t) * (nkeys - 1);
10586 		size += ksize;
10587 
10588 		/*
10589 		 * Now we need to determine the size of the stored data.
10590 		 */
10591 		id = DIF_INSTR_VAR(instr);
10592 
10593 		for (i = 0; i < dp->dtdo_varlen; i++) {
10594 			dtrace_difv_t *v = &dp->dtdo_vartab[i];
10595 
10596 			if (v->dtdv_id == id && v->dtdv_scope == scope) {
10597 				size += v->dtdv_type.dtdt_size;
10598 				break;
10599 			}
10600 		}
10601 
10602 		if (i == dp->dtdo_varlen)
10603 			return;
10604 
10605 		/*
10606 		 * We have the size.  If this is larger than the chunk size
10607 		 * for our dynamic variable state, reset the chunk size.
10608 		 */
10609 		size = P2ROUNDUP(size, sizeof (uint64_t));
10610 
10611 		/*
10612 		 * Before setting the chunk size, check that we're not going
10613 		 * to set it to a negative value...
10614 		 */
10615 		if (size > LONG_MAX)
10616 			return;
10617 
10618 		/*
10619 		 * ...and make certain that we didn't badly overflow.
10620 		 */
10621 		if (size < ksize || size < sizeof (dtrace_dynvar_t))
10622 			return;
10623 
10624 		if (size > vstate->dtvs_dynvars.dtds_chunksize)
10625 			vstate->dtvs_dynvars.dtds_chunksize = size;
10626 	}
10627 }
10628 
10629 static void
10630 dtrace_difo_init(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
10631 {
10632 	int i, oldsvars, osz, nsz, otlocals, ntlocals;
10633 	uint_t id;
10634 
10635 	ASSERT(MUTEX_HELD(&dtrace_lock));
10636 	ASSERT(dp->dtdo_buf != NULL && dp->dtdo_len != 0);
10637 
10638 	for (i = 0; i < dp->dtdo_varlen; i++) {
10639 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
10640 		dtrace_statvar_t *svar, ***svarp = NULL;
10641 		size_t dsize = 0;
10642 		uint8_t scope = v->dtdv_scope;
10643 		int *np = NULL;
10644 
10645 		if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
10646 			continue;
10647 
10648 		id -= DIF_VAR_OTHER_UBASE;
10649 
10650 		switch (scope) {
10651 		case DIFV_SCOPE_THREAD:
10652 			while (id >= (otlocals = vstate->dtvs_ntlocals)) {
10653 				dtrace_difv_t *tlocals;
10654 
10655 				if ((ntlocals = (otlocals << 1)) == 0)
10656 					ntlocals = 1;
10657 
10658 				osz = otlocals * sizeof (dtrace_difv_t);
10659 				nsz = ntlocals * sizeof (dtrace_difv_t);
10660 
10661 				tlocals = kmem_zalloc(nsz, KM_SLEEP);
10662 
10663 				if (osz != 0) {
10664 					bcopy(vstate->dtvs_tlocals,
10665 					    tlocals, osz);
10666 					kmem_free(vstate->dtvs_tlocals, osz);
10667 				}
10668 
10669 				vstate->dtvs_tlocals = tlocals;
10670 				vstate->dtvs_ntlocals = ntlocals;
10671 			}
10672 
10673 			vstate->dtvs_tlocals[id] = *v;
10674 			continue;
10675 
10676 		case DIFV_SCOPE_LOCAL:
10677 			np = &vstate->dtvs_nlocals;
10678 			svarp = &vstate->dtvs_locals;
10679 
10680 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
10681 				dsize = NCPU * (v->dtdv_type.dtdt_size +
10682 				    sizeof (uint64_t));
10683 			else
10684 				dsize = NCPU * sizeof (uint64_t);
10685 
10686 			break;
10687 
10688 		case DIFV_SCOPE_GLOBAL:
10689 			np = &vstate->dtvs_nglobals;
10690 			svarp = &vstate->dtvs_globals;
10691 
10692 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
10693 				dsize = v->dtdv_type.dtdt_size +
10694 				    sizeof (uint64_t);
10695 
10696 			break;
10697 
10698 		default:
10699 			ASSERT(0);
10700 		}
10701 
10702 		while (id >= (oldsvars = *np)) {
10703 			dtrace_statvar_t **statics;
10704 			int newsvars, oldsize, newsize;
10705 
10706 			if ((newsvars = (oldsvars << 1)) == 0)
10707 				newsvars = 1;
10708 
10709 			oldsize = oldsvars * sizeof (dtrace_statvar_t *);
10710 			newsize = newsvars * sizeof (dtrace_statvar_t *);
10711 
10712 			statics = kmem_zalloc(newsize, KM_SLEEP);
10713 
10714 			if (oldsize != 0) {
10715 				bcopy(*svarp, statics, oldsize);
10716 				kmem_free(*svarp, oldsize);
10717 			}
10718 
10719 			*svarp = statics;
10720 			*np = newsvars;
10721 		}
10722 
10723 		if ((svar = (*svarp)[id]) == NULL) {
10724 			svar = kmem_zalloc(sizeof (dtrace_statvar_t), KM_SLEEP);
10725 			svar->dtsv_var = *v;
10726 
10727 			if ((svar->dtsv_size = dsize) != 0) {
10728 				svar->dtsv_data = (uint64_t)(uintptr_t)
10729 				    kmem_zalloc(dsize, KM_SLEEP);
10730 			}
10731 
10732 			(*svarp)[id] = svar;
10733 		}
10734 
10735 		svar->dtsv_refcnt++;
10736 	}
10737 
10738 	dtrace_difo_chunksize(dp, vstate);
10739 	dtrace_difo_hold(dp);
10740 }
10741 
10742 static dtrace_difo_t *
10743 dtrace_difo_duplicate(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
10744 {
10745 	dtrace_difo_t *new;
10746 	size_t sz;
10747 
10748 	ASSERT(dp->dtdo_buf != NULL);
10749 	ASSERT(dp->dtdo_refcnt != 0);
10750 
10751 	new = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
10752 
10753 	ASSERT(dp->dtdo_buf != NULL);
10754 	sz = dp->dtdo_len * sizeof (dif_instr_t);
10755 	new->dtdo_buf = kmem_alloc(sz, KM_SLEEP);
10756 	bcopy(dp->dtdo_buf, new->dtdo_buf, sz);
10757 	new->dtdo_len = dp->dtdo_len;
10758 
10759 	if (dp->dtdo_strtab != NULL) {
10760 		ASSERT(dp->dtdo_strlen != 0);
10761 		new->dtdo_strtab = kmem_alloc(dp->dtdo_strlen, KM_SLEEP);
10762 		bcopy(dp->dtdo_strtab, new->dtdo_strtab, dp->dtdo_strlen);
10763 		new->dtdo_strlen = dp->dtdo_strlen;
10764 	}
10765 
10766 	if (dp->dtdo_inttab != NULL) {
10767 		ASSERT(dp->dtdo_intlen != 0);
10768 		sz = dp->dtdo_intlen * sizeof (uint64_t);
10769 		new->dtdo_inttab = kmem_alloc(sz, KM_SLEEP);
10770 		bcopy(dp->dtdo_inttab, new->dtdo_inttab, sz);
10771 		new->dtdo_intlen = dp->dtdo_intlen;
10772 	}
10773 
10774 	if (dp->dtdo_vartab != NULL) {
10775 		ASSERT(dp->dtdo_varlen != 0);
10776 		sz = dp->dtdo_varlen * sizeof (dtrace_difv_t);
10777 		new->dtdo_vartab = kmem_alloc(sz, KM_SLEEP);
10778 		bcopy(dp->dtdo_vartab, new->dtdo_vartab, sz);
10779 		new->dtdo_varlen = dp->dtdo_varlen;
10780 	}
10781 
10782 	dtrace_difo_init(new, vstate);
10783 	return (new);
10784 }
10785 
10786 static void
10787 dtrace_difo_destroy(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
10788 {
10789 	int i;
10790 
10791 	ASSERT(dp->dtdo_refcnt == 0);
10792 
10793 	for (i = 0; i < dp->dtdo_varlen; i++) {
10794 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
10795 		dtrace_statvar_t *svar, **svarp = NULL;
10796 		uint_t id;
10797 		uint8_t scope = v->dtdv_scope;
10798 		int *np = NULL;
10799 
10800 		switch (scope) {
10801 		case DIFV_SCOPE_THREAD:
10802 			continue;
10803 
10804 		case DIFV_SCOPE_LOCAL:
10805 			np = &vstate->dtvs_nlocals;
10806 			svarp = vstate->dtvs_locals;
10807 			break;
10808 
10809 		case DIFV_SCOPE_GLOBAL:
10810 			np = &vstate->dtvs_nglobals;
10811 			svarp = vstate->dtvs_globals;
10812 			break;
10813 
10814 		default:
10815 			ASSERT(0);
10816 		}
10817 
10818 		if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
10819 			continue;
10820 
10821 		id -= DIF_VAR_OTHER_UBASE;
10822 		ASSERT(id < *np);
10823 
10824 		svar = svarp[id];
10825 		ASSERT(svar != NULL);
10826 		ASSERT(svar->dtsv_refcnt > 0);
10827 
10828 		if (--svar->dtsv_refcnt > 0)
10829 			continue;
10830 
10831 		if (svar->dtsv_size != 0) {
10832 			ASSERT(svar->dtsv_data != 0);
10833 			kmem_free((void *)(uintptr_t)svar->dtsv_data,
10834 			    svar->dtsv_size);
10835 		}
10836 
10837 		kmem_free(svar, sizeof (dtrace_statvar_t));
10838 		svarp[id] = NULL;
10839 	}
10840 
10841 	if (dp->dtdo_buf != NULL)
10842 		kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
10843 	if (dp->dtdo_inttab != NULL)
10844 		kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
10845 	if (dp->dtdo_strtab != NULL)
10846 		kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
10847 	if (dp->dtdo_vartab != NULL)
10848 		kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
10849 
10850 	kmem_free(dp, sizeof (dtrace_difo_t));
10851 }
10852 
10853 static void
10854 dtrace_difo_release(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
10855 {
10856 	int i;
10857 
10858 	ASSERT(MUTEX_HELD(&dtrace_lock));
10859 	ASSERT(dp->dtdo_refcnt != 0);
10860 
10861 	for (i = 0; i < dp->dtdo_varlen; i++) {
10862 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
10863 
10864 		if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
10865 			continue;
10866 
10867 		ASSERT(dtrace_vtime_references > 0);
10868 		if (--dtrace_vtime_references == 0)
10869 			dtrace_vtime_disable();
10870 	}
10871 
10872 	if (--dp->dtdo_refcnt == 0)
10873 		dtrace_difo_destroy(dp, vstate);
10874 }
10875 
10876 /*
10877  * DTrace Format Functions
10878  */
10879 static uint16_t
10880 dtrace_format_add(dtrace_state_t *state, char *str)
10881 {
10882 	char *fmt, **new;
10883 	uint16_t ndx, len = strlen(str) + 1;
10884 
10885 	fmt = kmem_zalloc(len, KM_SLEEP);
10886 	bcopy(str, fmt, len);
10887 
10888 	for (ndx = 0; ndx < state->dts_nformats; ndx++) {
10889 		if (state->dts_formats[ndx] == NULL) {
10890 			state->dts_formats[ndx] = fmt;
10891 			return (ndx + 1);
10892 		}
10893 	}
10894 
10895 	if (state->dts_nformats == USHRT_MAX) {
10896 		/*
10897 		 * This is only likely if a denial-of-service attack is being
10898 		 * attempted.  As such, it's okay to fail silently here.
10899 		 */
10900 		kmem_free(fmt, len);
10901 		return (0);
10902 	}
10903 
10904 	/*
10905 	 * For simplicity, we always resize the formats array to be exactly the
10906 	 * number of formats.
10907 	 */
10908 	ndx = state->dts_nformats++;
10909 	new = kmem_alloc((ndx + 1) * sizeof (char *), KM_SLEEP);
10910 
10911 	if (state->dts_formats != NULL) {
10912 		ASSERT(ndx != 0);
10913 		bcopy(state->dts_formats, new, ndx * sizeof (char *));
10914 		kmem_free(state->dts_formats, ndx * sizeof (char *));
10915 	}
10916 
10917 	state->dts_formats = new;
10918 	state->dts_formats[ndx] = fmt;
10919 
10920 	return (ndx + 1);
10921 }
10922 
10923 static void
10924 dtrace_format_remove(dtrace_state_t *state, uint16_t format)
10925 {
10926 	char *fmt;
10927 
10928 	ASSERT(state->dts_formats != NULL);
10929 	ASSERT(format <= state->dts_nformats);
10930 	ASSERT(state->dts_formats[format - 1] != NULL);
10931 
10932 	fmt = state->dts_formats[format - 1];
10933 	kmem_free(fmt, strlen(fmt) + 1);
10934 	state->dts_formats[format - 1] = NULL;
10935 }
10936 
10937 static void
10938 dtrace_format_destroy(dtrace_state_t *state)
10939 {
10940 	int i;
10941 
10942 	if (state->dts_nformats == 0) {
10943 		ASSERT(state->dts_formats == NULL);
10944 		return;
10945 	}
10946 
10947 	ASSERT(state->dts_formats != NULL);
10948 
10949 	for (i = 0; i < state->dts_nformats; i++) {
10950 		char *fmt = state->dts_formats[i];
10951 
10952 		if (fmt == NULL)
10953 			continue;
10954 
10955 		kmem_free(fmt, strlen(fmt) + 1);
10956 	}
10957 
10958 	kmem_free(state->dts_formats, state->dts_nformats * sizeof (char *));
10959 	state->dts_nformats = 0;
10960 	state->dts_formats = NULL;
10961 }
10962 
10963 /*
10964  * DTrace Predicate Functions
10965  */
10966 static dtrace_predicate_t *
10967 dtrace_predicate_create(dtrace_difo_t *dp)
10968 {
10969 	dtrace_predicate_t *pred;
10970 
10971 	ASSERT(MUTEX_HELD(&dtrace_lock));
10972 	ASSERT(dp->dtdo_refcnt != 0);
10973 
10974 	pred = kmem_zalloc(sizeof (dtrace_predicate_t), KM_SLEEP);
10975 	pred->dtp_difo = dp;
10976 	pred->dtp_refcnt = 1;
10977 
10978 	if (!dtrace_difo_cacheable(dp))
10979 		return (pred);
10980 
10981 	if (dtrace_predcache_id == DTRACE_CACHEIDNONE) {
10982 		/*
10983 		 * This is only theoretically possible -- we have had 2^32
10984 		 * cacheable predicates on this machine.  We cannot allow any
10985 		 * more predicates to become cacheable:  as unlikely as it is,
10986 		 * there may be a thread caching a (now stale) predicate cache
10987 		 * ID. (N.B.: the temptation is being successfully resisted to
10988 		 * have this cmn_err() "Holy shit -- we executed this code!")
10989 		 */
10990 		return (pred);
10991 	}
10992 
10993 	pred->dtp_cacheid = dtrace_predcache_id++;
10994 
10995 	return (pred);
10996 }
10997 
10998 static void
10999 dtrace_predicate_hold(dtrace_predicate_t *pred)
11000 {
11001 	ASSERT(MUTEX_HELD(&dtrace_lock));
11002 	ASSERT(pred->dtp_difo != NULL && pred->dtp_difo->dtdo_refcnt != 0);
11003 	ASSERT(pred->dtp_refcnt > 0);
11004 
11005 	pred->dtp_refcnt++;
11006 }
11007 
11008 static void
11009 dtrace_predicate_release(dtrace_predicate_t *pred, dtrace_vstate_t *vstate)
11010 {
11011 	dtrace_difo_t *dp = pred->dtp_difo;
11012 
11013 	ASSERT(MUTEX_HELD(&dtrace_lock));
11014 	ASSERT(dp != NULL && dp->dtdo_refcnt != 0);
11015 	ASSERT(pred->dtp_refcnt > 0);
11016 
11017 	if (--pred->dtp_refcnt == 0) {
11018 		dtrace_difo_release(pred->dtp_difo, vstate);
11019 		kmem_free(pred, sizeof (dtrace_predicate_t));
11020 	}
11021 }
11022 
11023 /*
11024  * DTrace Action Description Functions
11025  */
11026 static dtrace_actdesc_t *
11027 dtrace_actdesc_create(dtrace_actkind_t kind, uint32_t ntuple,
11028     uint64_t uarg, uint64_t arg)
11029 {
11030 	dtrace_actdesc_t *act;
11031 
11032 #ifdef illumos
11033 	ASSERT(!DTRACEACT_ISPRINTFLIKE(kind) || (arg != NULL &&
11034 	    arg >= KERNELBASE) || (arg == NULL && kind == DTRACEACT_PRINTA));
11035 #endif
11036 
11037 	act = kmem_zalloc(sizeof (dtrace_actdesc_t), KM_SLEEP);
11038 	act->dtad_kind = kind;
11039 	act->dtad_ntuple = ntuple;
11040 	act->dtad_uarg = uarg;
11041 	act->dtad_arg = arg;
11042 	act->dtad_refcnt = 1;
11043 
11044 	return (act);
11045 }
11046 
11047 static void
11048 dtrace_actdesc_hold(dtrace_actdesc_t *act)
11049 {
11050 	ASSERT(act->dtad_refcnt >= 1);
11051 	act->dtad_refcnt++;
11052 }
11053 
11054 static void
11055 dtrace_actdesc_release(dtrace_actdesc_t *act, dtrace_vstate_t *vstate)
11056 {
11057 	dtrace_actkind_t kind = act->dtad_kind;
11058 	dtrace_difo_t *dp;
11059 
11060 	ASSERT(act->dtad_refcnt >= 1);
11061 
11062 	if (--act->dtad_refcnt != 0)
11063 		return;
11064 
11065 	if ((dp = act->dtad_difo) != NULL)
11066 		dtrace_difo_release(dp, vstate);
11067 
11068 	if (DTRACEACT_ISPRINTFLIKE(kind)) {
11069 		char *str = (char *)(uintptr_t)act->dtad_arg;
11070 
11071 #ifdef illumos
11072 		ASSERT((str != NULL && (uintptr_t)str >= KERNELBASE) ||
11073 		    (str == NULL && act->dtad_kind == DTRACEACT_PRINTA));
11074 #endif
11075 
11076 		if (str != NULL)
11077 			kmem_free(str, strlen(str) + 1);
11078 	}
11079 
11080 	kmem_free(act, sizeof (dtrace_actdesc_t));
11081 }
11082 
11083 /*
11084  * DTrace ECB Functions
11085  */
11086 static dtrace_ecb_t *
11087 dtrace_ecb_add(dtrace_state_t *state, dtrace_probe_t *probe)
11088 {
11089 	dtrace_ecb_t *ecb;
11090 	dtrace_epid_t epid;
11091 
11092 	ASSERT(MUTEX_HELD(&dtrace_lock));
11093 
11094 	ecb = kmem_zalloc(sizeof (dtrace_ecb_t), KM_SLEEP);
11095 	ecb->dte_predicate = NULL;
11096 	ecb->dte_probe = probe;
11097 
11098 	/*
11099 	 * The default size is the size of the default action: recording
11100 	 * the header.
11101 	 */
11102 	ecb->dte_size = ecb->dte_needed = sizeof (dtrace_rechdr_t);
11103 	ecb->dte_alignment = sizeof (dtrace_epid_t);
11104 
11105 	epid = state->dts_epid++;
11106 
11107 	if (epid - 1 >= state->dts_necbs) {
11108 		dtrace_ecb_t **oecbs = state->dts_ecbs, **ecbs;
11109 		int necbs = state->dts_necbs << 1;
11110 
11111 		ASSERT(epid == state->dts_necbs + 1);
11112 
11113 		if (necbs == 0) {
11114 			ASSERT(oecbs == NULL);
11115 			necbs = 1;
11116 		}
11117 
11118 		ecbs = kmem_zalloc(necbs * sizeof (*ecbs), KM_SLEEP);
11119 
11120 		if (oecbs != NULL)
11121 			bcopy(oecbs, ecbs, state->dts_necbs * sizeof (*ecbs));
11122 
11123 		dtrace_membar_producer();
11124 		state->dts_ecbs = ecbs;
11125 
11126 		if (oecbs != NULL) {
11127 			/*
11128 			 * If this state is active, we must dtrace_sync()
11129 			 * before we can free the old dts_ecbs array:  we're
11130 			 * coming in hot, and there may be active ring
11131 			 * buffer processing (which indexes into the dts_ecbs
11132 			 * array) on another CPU.
11133 			 */
11134 			if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
11135 				dtrace_sync();
11136 
11137 			kmem_free(oecbs, state->dts_necbs * sizeof (*ecbs));
11138 		}
11139 
11140 		dtrace_membar_producer();
11141 		state->dts_necbs = necbs;
11142 	}
11143 
11144 	ecb->dte_state = state;
11145 
11146 	ASSERT(state->dts_ecbs[epid - 1] == NULL);
11147 	dtrace_membar_producer();
11148 	state->dts_ecbs[(ecb->dte_epid = epid) - 1] = ecb;
11149 
11150 	return (ecb);
11151 }
11152 
11153 static void
11154 dtrace_ecb_enable(dtrace_ecb_t *ecb)
11155 {
11156 	dtrace_probe_t *probe = ecb->dte_probe;
11157 
11158 	ASSERT(MUTEX_HELD(&cpu_lock));
11159 	ASSERT(MUTEX_HELD(&dtrace_lock));
11160 	ASSERT(ecb->dte_next == NULL);
11161 
11162 	if (probe == NULL) {
11163 		/*
11164 		 * This is the NULL probe -- there's nothing to do.
11165 		 */
11166 		return;
11167 	}
11168 
11169 	if (probe->dtpr_ecb == NULL) {
11170 		dtrace_provider_t *prov = probe->dtpr_provider;
11171 
11172 		/*
11173 		 * We're the first ECB on this probe.
11174 		 */
11175 		probe->dtpr_ecb = probe->dtpr_ecb_last = ecb;
11176 
11177 		if (ecb->dte_predicate != NULL)
11178 			probe->dtpr_predcache = ecb->dte_predicate->dtp_cacheid;
11179 
11180 		prov->dtpv_pops.dtps_enable(prov->dtpv_arg,
11181 		    probe->dtpr_id, probe->dtpr_arg);
11182 	} else {
11183 		/*
11184 		 * This probe is already active.  Swing the last pointer to
11185 		 * point to the new ECB, and issue a dtrace_sync() to assure
11186 		 * that all CPUs have seen the change.
11187 		 */
11188 		ASSERT(probe->dtpr_ecb_last != NULL);
11189 		probe->dtpr_ecb_last->dte_next = ecb;
11190 		probe->dtpr_ecb_last = ecb;
11191 		probe->dtpr_predcache = 0;
11192 
11193 		dtrace_sync();
11194 	}
11195 }
11196 
11197 static int
11198 dtrace_ecb_resize(dtrace_ecb_t *ecb)
11199 {
11200 	dtrace_action_t *act;
11201 	uint32_t curneeded = UINT32_MAX;
11202 	uint32_t aggbase = UINT32_MAX;
11203 
11204 	/*
11205 	 * If we record anything, we always record the dtrace_rechdr_t.  (And
11206 	 * we always record it first.)
11207 	 */
11208 	ecb->dte_size = sizeof (dtrace_rechdr_t);
11209 	ecb->dte_alignment = sizeof (dtrace_epid_t);
11210 
11211 	for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
11212 		dtrace_recdesc_t *rec = &act->dta_rec;
11213 		ASSERT(rec->dtrd_size > 0 || rec->dtrd_alignment == 1);
11214 
11215 		ecb->dte_alignment = MAX(ecb->dte_alignment,
11216 		    rec->dtrd_alignment);
11217 
11218 		if (DTRACEACT_ISAGG(act->dta_kind)) {
11219 			dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
11220 
11221 			ASSERT(rec->dtrd_size != 0);
11222 			ASSERT(agg->dtag_first != NULL);
11223 			ASSERT(act->dta_prev->dta_intuple);
11224 			ASSERT(aggbase != UINT32_MAX);
11225 			ASSERT(curneeded != UINT32_MAX);
11226 
11227 			agg->dtag_base = aggbase;
11228 
11229 			curneeded = P2ROUNDUP(curneeded, rec->dtrd_alignment);
11230 			rec->dtrd_offset = curneeded;
11231 			if (curneeded + rec->dtrd_size < curneeded)
11232 				return (EINVAL);
11233 			curneeded += rec->dtrd_size;
11234 			ecb->dte_needed = MAX(ecb->dte_needed, curneeded);
11235 
11236 			aggbase = UINT32_MAX;
11237 			curneeded = UINT32_MAX;
11238 		} else if (act->dta_intuple) {
11239 			if (curneeded == UINT32_MAX) {
11240 				/*
11241 				 * This is the first record in a tuple.  Align
11242 				 * curneeded to be at offset 4 in an 8-byte
11243 				 * aligned block.
11244 				 */
11245 				ASSERT(act->dta_prev == NULL ||
11246 				    !act->dta_prev->dta_intuple);
11247 				ASSERT3U(aggbase, ==, UINT32_MAX);
11248 				curneeded = P2PHASEUP(ecb->dte_size,
11249 				    sizeof (uint64_t), sizeof (dtrace_aggid_t));
11250 
11251 				aggbase = curneeded - sizeof (dtrace_aggid_t);
11252 				ASSERT(IS_P2ALIGNED(aggbase,
11253 				    sizeof (uint64_t)));
11254 			}
11255 			curneeded = P2ROUNDUP(curneeded, rec->dtrd_alignment);
11256 			rec->dtrd_offset = curneeded;
11257 			if (curneeded + rec->dtrd_size < curneeded)
11258 				return (EINVAL);
11259 			curneeded += rec->dtrd_size;
11260 		} else {
11261 			/* tuples must be followed by an aggregation */
11262 			ASSERT(act->dta_prev == NULL ||
11263 			    !act->dta_prev->dta_intuple);
11264 
11265 			ecb->dte_size = P2ROUNDUP(ecb->dte_size,
11266 			    rec->dtrd_alignment);
11267 			rec->dtrd_offset = ecb->dte_size;
11268 			if (ecb->dte_size + rec->dtrd_size < ecb->dte_size)
11269 				return (EINVAL);
11270 			ecb->dte_size += rec->dtrd_size;
11271 			ecb->dte_needed = MAX(ecb->dte_needed, ecb->dte_size);
11272 		}
11273 	}
11274 
11275 	if ((act = ecb->dte_action) != NULL &&
11276 	    !(act->dta_kind == DTRACEACT_SPECULATE && act->dta_next == NULL) &&
11277 	    ecb->dte_size == sizeof (dtrace_rechdr_t)) {
11278 		/*
11279 		 * If the size is still sizeof (dtrace_rechdr_t), then all
11280 		 * actions store no data; set the size to 0.
11281 		 */
11282 		ecb->dte_size = 0;
11283 	}
11284 
11285 	ecb->dte_size = P2ROUNDUP(ecb->dte_size, sizeof (dtrace_epid_t));
11286 	ecb->dte_needed = P2ROUNDUP(ecb->dte_needed, (sizeof (dtrace_epid_t)));
11287 	ecb->dte_state->dts_needed = MAX(ecb->dte_state->dts_needed,
11288 	    ecb->dte_needed);
11289 	return (0);
11290 }
11291 
11292 static dtrace_action_t *
11293 dtrace_ecb_aggregation_create(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
11294 {
11295 	dtrace_aggregation_t *agg;
11296 	size_t size = sizeof (uint64_t);
11297 	int ntuple = desc->dtad_ntuple;
11298 	dtrace_action_t *act;
11299 	dtrace_recdesc_t *frec;
11300 	dtrace_aggid_t aggid;
11301 	dtrace_state_t *state = ecb->dte_state;
11302 
11303 	agg = kmem_zalloc(sizeof (dtrace_aggregation_t), KM_SLEEP);
11304 	agg->dtag_ecb = ecb;
11305 
11306 	ASSERT(DTRACEACT_ISAGG(desc->dtad_kind));
11307 
11308 	switch (desc->dtad_kind) {
11309 	case DTRACEAGG_MIN:
11310 		agg->dtag_initial = INT64_MAX;
11311 		agg->dtag_aggregate = dtrace_aggregate_min;
11312 		break;
11313 
11314 	case DTRACEAGG_MAX:
11315 		agg->dtag_initial = INT64_MIN;
11316 		agg->dtag_aggregate = dtrace_aggregate_max;
11317 		break;
11318 
11319 	case DTRACEAGG_COUNT:
11320 		agg->dtag_aggregate = dtrace_aggregate_count;
11321 		break;
11322 
11323 	case DTRACEAGG_QUANTIZE:
11324 		agg->dtag_aggregate = dtrace_aggregate_quantize;
11325 		size = (((sizeof (uint64_t) * NBBY) - 1) * 2 + 1) *
11326 		    sizeof (uint64_t);
11327 		break;
11328 
11329 	case DTRACEAGG_LQUANTIZE: {
11330 		uint16_t step = DTRACE_LQUANTIZE_STEP(desc->dtad_arg);
11331 		uint16_t levels = DTRACE_LQUANTIZE_LEVELS(desc->dtad_arg);
11332 
11333 		agg->dtag_initial = desc->dtad_arg;
11334 		agg->dtag_aggregate = dtrace_aggregate_lquantize;
11335 
11336 		if (step == 0 || levels == 0)
11337 			goto err;
11338 
11339 		size = levels * sizeof (uint64_t) + 3 * sizeof (uint64_t);
11340 		break;
11341 	}
11342 
11343 	case DTRACEAGG_LLQUANTIZE: {
11344 		uint16_t factor = DTRACE_LLQUANTIZE_FACTOR(desc->dtad_arg);
11345 		uint16_t low = DTRACE_LLQUANTIZE_LOW(desc->dtad_arg);
11346 		uint16_t high = DTRACE_LLQUANTIZE_HIGH(desc->dtad_arg);
11347 		uint16_t nsteps = DTRACE_LLQUANTIZE_NSTEP(desc->dtad_arg);
11348 		int64_t v;
11349 
11350 		agg->dtag_initial = desc->dtad_arg;
11351 		agg->dtag_aggregate = dtrace_aggregate_llquantize;
11352 
11353 		if (factor < 2 || low >= high || nsteps < factor)
11354 			goto err;
11355 
11356 		/*
11357 		 * Now check that the number of steps evenly divides a power
11358 		 * of the factor.  (This assures both integer bucket size and
11359 		 * linearity within each magnitude.)
11360 		 */
11361 		for (v = factor; v < nsteps; v *= factor)
11362 			continue;
11363 
11364 		if ((v % nsteps) || (nsteps % factor))
11365 			goto err;
11366 
11367 		size = (dtrace_aggregate_llquantize_bucket(factor,
11368 		    low, high, nsteps, INT64_MAX) + 2) * sizeof (uint64_t);
11369 		break;
11370 	}
11371 
11372 	case DTRACEAGG_AVG:
11373 		agg->dtag_aggregate = dtrace_aggregate_avg;
11374 		size = sizeof (uint64_t) * 2;
11375 		break;
11376 
11377 	case DTRACEAGG_STDDEV:
11378 		agg->dtag_aggregate = dtrace_aggregate_stddev;
11379 		size = sizeof (uint64_t) * 4;
11380 		break;
11381 
11382 	case DTRACEAGG_SUM:
11383 		agg->dtag_aggregate = dtrace_aggregate_sum;
11384 		break;
11385 
11386 	default:
11387 		goto err;
11388 	}
11389 
11390 	agg->dtag_action.dta_rec.dtrd_size = size;
11391 
11392 	if (ntuple == 0)
11393 		goto err;
11394 
11395 	/*
11396 	 * We must make sure that we have enough actions for the n-tuple.
11397 	 */
11398 	for (act = ecb->dte_action_last; act != NULL; act = act->dta_prev) {
11399 		if (DTRACEACT_ISAGG(act->dta_kind))
11400 			break;
11401 
11402 		if (--ntuple == 0) {
11403 			/*
11404 			 * This is the action with which our n-tuple begins.
11405 			 */
11406 			agg->dtag_first = act;
11407 			goto success;
11408 		}
11409 	}
11410 
11411 	/*
11412 	 * This n-tuple is short by ntuple elements.  Return failure.
11413 	 */
11414 	ASSERT(ntuple != 0);
11415 err:
11416 	kmem_free(agg, sizeof (dtrace_aggregation_t));
11417 	return (NULL);
11418 
11419 success:
11420 	/*
11421 	 * If the last action in the tuple has a size of zero, it's actually
11422 	 * an expression argument for the aggregating action.
11423 	 */
11424 	ASSERT(ecb->dte_action_last != NULL);
11425 	act = ecb->dte_action_last;
11426 
11427 	if (act->dta_kind == DTRACEACT_DIFEXPR) {
11428 		ASSERT(act->dta_difo != NULL);
11429 
11430 		if (act->dta_difo->dtdo_rtype.dtdt_size == 0)
11431 			agg->dtag_hasarg = 1;
11432 	}
11433 
11434 	/*
11435 	 * We need to allocate an id for this aggregation.
11436 	 */
11437 #ifdef illumos
11438 	aggid = (dtrace_aggid_t)(uintptr_t)vmem_alloc(state->dts_aggid_arena, 1,
11439 	    VM_BESTFIT | VM_SLEEP);
11440 #else
11441 	aggid = alloc_unr(state->dts_aggid_arena);
11442 #endif
11443 
11444 	if (aggid - 1 >= state->dts_naggregations) {
11445 		dtrace_aggregation_t **oaggs = state->dts_aggregations;
11446 		dtrace_aggregation_t **aggs;
11447 		int naggs = state->dts_naggregations << 1;
11448 		int onaggs = state->dts_naggregations;
11449 
11450 		ASSERT(aggid == state->dts_naggregations + 1);
11451 
11452 		if (naggs == 0) {
11453 			ASSERT(oaggs == NULL);
11454 			naggs = 1;
11455 		}
11456 
11457 		aggs = kmem_zalloc(naggs * sizeof (*aggs), KM_SLEEP);
11458 
11459 		if (oaggs != NULL) {
11460 			bcopy(oaggs, aggs, onaggs * sizeof (*aggs));
11461 			kmem_free(oaggs, onaggs * sizeof (*aggs));
11462 		}
11463 
11464 		state->dts_aggregations = aggs;
11465 		state->dts_naggregations = naggs;
11466 	}
11467 
11468 	ASSERT(state->dts_aggregations[aggid - 1] == NULL);
11469 	state->dts_aggregations[(agg->dtag_id = aggid) - 1] = agg;
11470 
11471 	frec = &agg->dtag_first->dta_rec;
11472 	if (frec->dtrd_alignment < sizeof (dtrace_aggid_t))
11473 		frec->dtrd_alignment = sizeof (dtrace_aggid_t);
11474 
11475 	for (act = agg->dtag_first; act != NULL; act = act->dta_next) {
11476 		ASSERT(!act->dta_intuple);
11477 		act->dta_intuple = 1;
11478 	}
11479 
11480 	return (&agg->dtag_action);
11481 }
11482 
11483 static void
11484 dtrace_ecb_aggregation_destroy(dtrace_ecb_t *ecb, dtrace_action_t *act)
11485 {
11486 	dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
11487 	dtrace_state_t *state = ecb->dte_state;
11488 	dtrace_aggid_t aggid = agg->dtag_id;
11489 
11490 	ASSERT(DTRACEACT_ISAGG(act->dta_kind));
11491 #ifdef illumos
11492 	vmem_free(state->dts_aggid_arena, (void *)(uintptr_t)aggid, 1);
11493 #else
11494 	free_unr(state->dts_aggid_arena, aggid);
11495 #endif
11496 
11497 	ASSERT(state->dts_aggregations[aggid - 1] == agg);
11498 	state->dts_aggregations[aggid - 1] = NULL;
11499 
11500 	kmem_free(agg, sizeof (dtrace_aggregation_t));
11501 }
11502 
11503 static int
11504 dtrace_ecb_action_add(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
11505 {
11506 	dtrace_action_t *action, *last;
11507 	dtrace_difo_t *dp = desc->dtad_difo;
11508 	uint32_t size = 0, align = sizeof (uint8_t), mask;
11509 	uint16_t format = 0;
11510 	dtrace_recdesc_t *rec;
11511 	dtrace_state_t *state = ecb->dte_state;
11512 	dtrace_optval_t *opt = state->dts_options, nframes = 0, strsize;
11513 	uint64_t arg = desc->dtad_arg;
11514 
11515 	ASSERT(MUTEX_HELD(&dtrace_lock));
11516 	ASSERT(ecb->dte_action == NULL || ecb->dte_action->dta_refcnt == 1);
11517 
11518 	if (DTRACEACT_ISAGG(desc->dtad_kind)) {
11519 		/*
11520 		 * If this is an aggregating action, there must be neither
11521 		 * a speculate nor a commit on the action chain.
11522 		 */
11523 		dtrace_action_t *act;
11524 
11525 		for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
11526 			if (act->dta_kind == DTRACEACT_COMMIT)
11527 				return (EINVAL);
11528 
11529 			if (act->dta_kind == DTRACEACT_SPECULATE)
11530 				return (EINVAL);
11531 		}
11532 
11533 		action = dtrace_ecb_aggregation_create(ecb, desc);
11534 
11535 		if (action == NULL)
11536 			return (EINVAL);
11537 	} else {
11538 		if (DTRACEACT_ISDESTRUCTIVE(desc->dtad_kind) ||
11539 		    (desc->dtad_kind == DTRACEACT_DIFEXPR &&
11540 		    dp != NULL && dp->dtdo_destructive)) {
11541 			state->dts_destructive = 1;
11542 		}
11543 
11544 		switch (desc->dtad_kind) {
11545 		case DTRACEACT_PRINTF:
11546 		case DTRACEACT_PRINTA:
11547 		case DTRACEACT_SYSTEM:
11548 		case DTRACEACT_FREOPEN:
11549 		case DTRACEACT_DIFEXPR:
11550 			/*
11551 			 * We know that our arg is a string -- turn it into a
11552 			 * format.
11553 			 */
11554 			if (arg == 0) {
11555 				ASSERT(desc->dtad_kind == DTRACEACT_PRINTA ||
11556 				    desc->dtad_kind == DTRACEACT_DIFEXPR);
11557 				format = 0;
11558 			} else {
11559 				ASSERT(arg != 0);
11560 #ifdef illumos
11561 				ASSERT(arg > KERNELBASE);
11562 #endif
11563 				format = dtrace_format_add(state,
11564 				    (char *)(uintptr_t)arg);
11565 			}
11566 
11567 			/*FALLTHROUGH*/
11568 		case DTRACEACT_LIBACT:
11569 		case DTRACEACT_TRACEMEM:
11570 		case DTRACEACT_TRACEMEM_DYNSIZE:
11571 			if (dp == NULL)
11572 				return (EINVAL);
11573 
11574 			if ((size = dp->dtdo_rtype.dtdt_size) != 0)
11575 				break;
11576 
11577 			if (dp->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING) {
11578 				if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
11579 					return (EINVAL);
11580 
11581 				size = opt[DTRACEOPT_STRSIZE];
11582 			}
11583 
11584 			break;
11585 
11586 		case DTRACEACT_STACK:
11587 			if ((nframes = arg) == 0) {
11588 				nframes = opt[DTRACEOPT_STACKFRAMES];
11589 				ASSERT(nframes > 0);
11590 				arg = nframes;
11591 			}
11592 
11593 			size = nframes * sizeof (pc_t);
11594 			break;
11595 
11596 		case DTRACEACT_JSTACK:
11597 			if ((strsize = DTRACE_USTACK_STRSIZE(arg)) == 0)
11598 				strsize = opt[DTRACEOPT_JSTACKSTRSIZE];
11599 
11600 			if ((nframes = DTRACE_USTACK_NFRAMES(arg)) == 0)
11601 				nframes = opt[DTRACEOPT_JSTACKFRAMES];
11602 
11603 			arg = DTRACE_USTACK_ARG(nframes, strsize);
11604 
11605 			/*FALLTHROUGH*/
11606 		case DTRACEACT_USTACK:
11607 			if (desc->dtad_kind != DTRACEACT_JSTACK &&
11608 			    (nframes = DTRACE_USTACK_NFRAMES(arg)) == 0) {
11609 				strsize = DTRACE_USTACK_STRSIZE(arg);
11610 				nframes = opt[DTRACEOPT_USTACKFRAMES];
11611 				ASSERT(nframes > 0);
11612 				arg = DTRACE_USTACK_ARG(nframes, strsize);
11613 			}
11614 
11615 			/*
11616 			 * Save a slot for the pid.
11617 			 */
11618 			size = (nframes + 1) * sizeof (uint64_t);
11619 			size += DTRACE_USTACK_STRSIZE(arg);
11620 			size = P2ROUNDUP(size, (uint32_t)(sizeof (uintptr_t)));
11621 
11622 			break;
11623 
11624 		case DTRACEACT_SYM:
11625 		case DTRACEACT_MOD:
11626 			if (dp == NULL || ((size = dp->dtdo_rtype.dtdt_size) !=
11627 			    sizeof (uint64_t)) ||
11628 			    (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
11629 				return (EINVAL);
11630 			break;
11631 
11632 		case DTRACEACT_USYM:
11633 		case DTRACEACT_UMOD:
11634 		case DTRACEACT_UADDR:
11635 			if (dp == NULL ||
11636 			    (dp->dtdo_rtype.dtdt_size != sizeof (uint64_t)) ||
11637 			    (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
11638 				return (EINVAL);
11639 
11640 			/*
11641 			 * We have a slot for the pid, plus a slot for the
11642 			 * argument.  To keep things simple (aligned with
11643 			 * bitness-neutral sizing), we store each as a 64-bit
11644 			 * quantity.
11645 			 */
11646 			size = 2 * sizeof (uint64_t);
11647 			break;
11648 
11649 		case DTRACEACT_STOP:
11650 		case DTRACEACT_BREAKPOINT:
11651 		case DTRACEACT_PANIC:
11652 			break;
11653 
11654 		case DTRACEACT_CHILL:
11655 		case DTRACEACT_DISCARD:
11656 		case DTRACEACT_RAISE:
11657 			if (dp == NULL)
11658 				return (EINVAL);
11659 			break;
11660 
11661 		case DTRACEACT_EXIT:
11662 			if (dp == NULL ||
11663 			    (size = dp->dtdo_rtype.dtdt_size) != sizeof (int) ||
11664 			    (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
11665 				return (EINVAL);
11666 			break;
11667 
11668 		case DTRACEACT_SPECULATE:
11669 			if (ecb->dte_size > sizeof (dtrace_rechdr_t))
11670 				return (EINVAL);
11671 
11672 			if (dp == NULL)
11673 				return (EINVAL);
11674 
11675 			state->dts_speculates = 1;
11676 			break;
11677 
11678 		case DTRACEACT_PRINTM:
11679 		    	size = dp->dtdo_rtype.dtdt_size;
11680 			break;
11681 
11682 		case DTRACEACT_COMMIT: {
11683 			dtrace_action_t *act = ecb->dte_action;
11684 
11685 			for (; act != NULL; act = act->dta_next) {
11686 				if (act->dta_kind == DTRACEACT_COMMIT)
11687 					return (EINVAL);
11688 			}
11689 
11690 			if (dp == NULL)
11691 				return (EINVAL);
11692 			break;
11693 		}
11694 
11695 		default:
11696 			return (EINVAL);
11697 		}
11698 
11699 		if (size != 0 || desc->dtad_kind == DTRACEACT_SPECULATE) {
11700 			/*
11701 			 * If this is a data-storing action or a speculate,
11702 			 * we must be sure that there isn't a commit on the
11703 			 * action chain.
11704 			 */
11705 			dtrace_action_t *act = ecb->dte_action;
11706 
11707 			for (; act != NULL; act = act->dta_next) {
11708 				if (act->dta_kind == DTRACEACT_COMMIT)
11709 					return (EINVAL);
11710 			}
11711 		}
11712 
11713 		action = kmem_zalloc(sizeof (dtrace_action_t), KM_SLEEP);
11714 		action->dta_rec.dtrd_size = size;
11715 	}
11716 
11717 	action->dta_refcnt = 1;
11718 	rec = &action->dta_rec;
11719 	size = rec->dtrd_size;
11720 
11721 	for (mask = sizeof (uint64_t) - 1; size != 0 && mask > 0; mask >>= 1) {
11722 		if (!(size & mask)) {
11723 			align = mask + 1;
11724 			break;
11725 		}
11726 	}
11727 
11728 	action->dta_kind = desc->dtad_kind;
11729 
11730 	if ((action->dta_difo = dp) != NULL)
11731 		dtrace_difo_hold(dp);
11732 
11733 	rec->dtrd_action = action->dta_kind;
11734 	rec->dtrd_arg = arg;
11735 	rec->dtrd_uarg = desc->dtad_uarg;
11736 	rec->dtrd_alignment = (uint16_t)align;
11737 	rec->dtrd_format = format;
11738 
11739 	if ((last = ecb->dte_action_last) != NULL) {
11740 		ASSERT(ecb->dte_action != NULL);
11741 		action->dta_prev = last;
11742 		last->dta_next = action;
11743 	} else {
11744 		ASSERT(ecb->dte_action == NULL);
11745 		ecb->dte_action = action;
11746 	}
11747 
11748 	ecb->dte_action_last = action;
11749 
11750 	return (0);
11751 }
11752 
11753 static void
11754 dtrace_ecb_action_remove(dtrace_ecb_t *ecb)
11755 {
11756 	dtrace_action_t *act = ecb->dte_action, *next;
11757 	dtrace_vstate_t *vstate = &ecb->dte_state->dts_vstate;
11758 	dtrace_difo_t *dp;
11759 	uint16_t format;
11760 
11761 	if (act != NULL && act->dta_refcnt > 1) {
11762 		ASSERT(act->dta_next == NULL || act->dta_next->dta_refcnt == 1);
11763 		act->dta_refcnt--;
11764 	} else {
11765 		for (; act != NULL; act = next) {
11766 			next = act->dta_next;
11767 			ASSERT(next != NULL || act == ecb->dte_action_last);
11768 			ASSERT(act->dta_refcnt == 1);
11769 
11770 			if ((format = act->dta_rec.dtrd_format) != 0)
11771 				dtrace_format_remove(ecb->dte_state, format);
11772 
11773 			if ((dp = act->dta_difo) != NULL)
11774 				dtrace_difo_release(dp, vstate);
11775 
11776 			if (DTRACEACT_ISAGG(act->dta_kind)) {
11777 				dtrace_ecb_aggregation_destroy(ecb, act);
11778 			} else {
11779 				kmem_free(act, sizeof (dtrace_action_t));
11780 			}
11781 		}
11782 	}
11783 
11784 	ecb->dte_action = NULL;
11785 	ecb->dte_action_last = NULL;
11786 	ecb->dte_size = 0;
11787 }
11788 
11789 static void
11790 dtrace_ecb_disable(dtrace_ecb_t *ecb)
11791 {
11792 	/*
11793 	 * We disable the ECB by removing it from its probe.
11794 	 */
11795 	dtrace_ecb_t *pecb, *prev = NULL;
11796 	dtrace_probe_t *probe = ecb->dte_probe;
11797 
11798 	ASSERT(MUTEX_HELD(&dtrace_lock));
11799 
11800 	if (probe == NULL) {
11801 		/*
11802 		 * This is the NULL probe; there is nothing to disable.
11803 		 */
11804 		return;
11805 	}
11806 
11807 	for (pecb = probe->dtpr_ecb; pecb != NULL; pecb = pecb->dte_next) {
11808 		if (pecb == ecb)
11809 			break;
11810 		prev = pecb;
11811 	}
11812 
11813 	ASSERT(pecb != NULL);
11814 
11815 	if (prev == NULL) {
11816 		probe->dtpr_ecb = ecb->dte_next;
11817 	} else {
11818 		prev->dte_next = ecb->dte_next;
11819 	}
11820 
11821 	if (ecb == probe->dtpr_ecb_last) {
11822 		ASSERT(ecb->dte_next == NULL);
11823 		probe->dtpr_ecb_last = prev;
11824 	}
11825 
11826 	/*
11827 	 * The ECB has been disconnected from the probe; now sync to assure
11828 	 * that all CPUs have seen the change before returning.
11829 	 */
11830 	dtrace_sync();
11831 
11832 	if (probe->dtpr_ecb == NULL) {
11833 		/*
11834 		 * That was the last ECB on the probe; clear the predicate
11835 		 * cache ID for the probe, disable it and sync one more time
11836 		 * to assure that we'll never hit it again.
11837 		 */
11838 		dtrace_provider_t *prov = probe->dtpr_provider;
11839 
11840 		ASSERT(ecb->dte_next == NULL);
11841 		ASSERT(probe->dtpr_ecb_last == NULL);
11842 		probe->dtpr_predcache = DTRACE_CACHEIDNONE;
11843 		prov->dtpv_pops.dtps_disable(prov->dtpv_arg,
11844 		    probe->dtpr_id, probe->dtpr_arg);
11845 		dtrace_sync();
11846 	} else {
11847 		/*
11848 		 * There is at least one ECB remaining on the probe.  If there
11849 		 * is _exactly_ one, set the probe's predicate cache ID to be
11850 		 * the predicate cache ID of the remaining ECB.
11851 		 */
11852 		ASSERT(probe->dtpr_ecb_last != NULL);
11853 		ASSERT(probe->dtpr_predcache == DTRACE_CACHEIDNONE);
11854 
11855 		if (probe->dtpr_ecb == probe->dtpr_ecb_last) {
11856 			dtrace_predicate_t *p = probe->dtpr_ecb->dte_predicate;
11857 
11858 			ASSERT(probe->dtpr_ecb->dte_next == NULL);
11859 
11860 			if (p != NULL)
11861 				probe->dtpr_predcache = p->dtp_cacheid;
11862 		}
11863 
11864 		ecb->dte_next = NULL;
11865 	}
11866 }
11867 
11868 static void
11869 dtrace_ecb_destroy(dtrace_ecb_t *ecb)
11870 {
11871 	dtrace_state_t *state = ecb->dte_state;
11872 	dtrace_vstate_t *vstate = &state->dts_vstate;
11873 	dtrace_predicate_t *pred;
11874 	dtrace_epid_t epid = ecb->dte_epid;
11875 
11876 	ASSERT(MUTEX_HELD(&dtrace_lock));
11877 	ASSERT(ecb->dte_next == NULL);
11878 	ASSERT(ecb->dte_probe == NULL || ecb->dte_probe->dtpr_ecb != ecb);
11879 
11880 	if ((pred = ecb->dte_predicate) != NULL)
11881 		dtrace_predicate_release(pred, vstate);
11882 
11883 	dtrace_ecb_action_remove(ecb);
11884 
11885 	ASSERT(state->dts_ecbs[epid - 1] == ecb);
11886 	state->dts_ecbs[epid - 1] = NULL;
11887 
11888 	kmem_free(ecb, sizeof (dtrace_ecb_t));
11889 }
11890 
11891 static dtrace_ecb_t *
11892 dtrace_ecb_create(dtrace_state_t *state, dtrace_probe_t *probe,
11893     dtrace_enabling_t *enab)
11894 {
11895 	dtrace_ecb_t *ecb;
11896 	dtrace_predicate_t *pred;
11897 	dtrace_actdesc_t *act;
11898 	dtrace_provider_t *prov;
11899 	dtrace_ecbdesc_t *desc = enab->dten_current;
11900 
11901 	ASSERT(MUTEX_HELD(&dtrace_lock));
11902 	ASSERT(state != NULL);
11903 
11904 	ecb = dtrace_ecb_add(state, probe);
11905 	ecb->dte_uarg = desc->dted_uarg;
11906 
11907 	if ((pred = desc->dted_pred.dtpdd_predicate) != NULL) {
11908 		dtrace_predicate_hold(pred);
11909 		ecb->dte_predicate = pred;
11910 	}
11911 
11912 	if (probe != NULL) {
11913 		/*
11914 		 * If the provider shows more leg than the consumer is old
11915 		 * enough to see, we need to enable the appropriate implicit
11916 		 * predicate bits to prevent the ecb from activating at
11917 		 * revealing times.
11918 		 *
11919 		 * Providers specifying DTRACE_PRIV_USER at register time
11920 		 * are stating that they need the /proc-style privilege
11921 		 * model to be enforced, and this is what DTRACE_COND_OWNER
11922 		 * and DTRACE_COND_ZONEOWNER will then do at probe time.
11923 		 */
11924 		prov = probe->dtpr_provider;
11925 		if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLPROC) &&
11926 		    (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
11927 			ecb->dte_cond |= DTRACE_COND_OWNER;
11928 
11929 		if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLZONE) &&
11930 		    (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
11931 			ecb->dte_cond |= DTRACE_COND_ZONEOWNER;
11932 
11933 		/*
11934 		 * If the provider shows us kernel innards and the user
11935 		 * is lacking sufficient privilege, enable the
11936 		 * DTRACE_COND_USERMODE implicit predicate.
11937 		 */
11938 		if (!(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) &&
11939 		    (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_KERNEL))
11940 			ecb->dte_cond |= DTRACE_COND_USERMODE;
11941 	}
11942 
11943 	if (dtrace_ecb_create_cache != NULL) {
11944 		/*
11945 		 * If we have a cached ecb, we'll use its action list instead
11946 		 * of creating our own (saving both time and space).
11947 		 */
11948 		dtrace_ecb_t *cached = dtrace_ecb_create_cache;
11949 		dtrace_action_t *act = cached->dte_action;
11950 
11951 		if (act != NULL) {
11952 			ASSERT(act->dta_refcnt > 0);
11953 			act->dta_refcnt++;
11954 			ecb->dte_action = act;
11955 			ecb->dte_action_last = cached->dte_action_last;
11956 			ecb->dte_needed = cached->dte_needed;
11957 			ecb->dte_size = cached->dte_size;
11958 			ecb->dte_alignment = cached->dte_alignment;
11959 		}
11960 
11961 		return (ecb);
11962 	}
11963 
11964 	for (act = desc->dted_action; act != NULL; act = act->dtad_next) {
11965 		if ((enab->dten_error = dtrace_ecb_action_add(ecb, act)) != 0) {
11966 			dtrace_ecb_destroy(ecb);
11967 			return (NULL);
11968 		}
11969 	}
11970 
11971 	if ((enab->dten_error = dtrace_ecb_resize(ecb)) != 0) {
11972 		dtrace_ecb_destroy(ecb);
11973 		return (NULL);
11974 	}
11975 
11976 	return (dtrace_ecb_create_cache = ecb);
11977 }
11978 
11979 static int
11980 dtrace_ecb_create_enable(dtrace_probe_t *probe, void *arg)
11981 {
11982 	dtrace_ecb_t *ecb;
11983 	dtrace_enabling_t *enab = arg;
11984 	dtrace_state_t *state = enab->dten_vstate->dtvs_state;
11985 
11986 	ASSERT(state != NULL);
11987 
11988 	if (probe != NULL && probe->dtpr_gen < enab->dten_probegen) {
11989 		/*
11990 		 * This probe was created in a generation for which this
11991 		 * enabling has previously created ECBs; we don't want to
11992 		 * enable it again, so just kick out.
11993 		 */
11994 		return (DTRACE_MATCH_NEXT);
11995 	}
11996 
11997 	if ((ecb = dtrace_ecb_create(state, probe, enab)) == NULL)
11998 		return (DTRACE_MATCH_DONE);
11999 
12000 	dtrace_ecb_enable(ecb);
12001 	return (DTRACE_MATCH_NEXT);
12002 }
12003 
12004 static dtrace_ecb_t *
12005 dtrace_epid2ecb(dtrace_state_t *state, dtrace_epid_t id)
12006 {
12007 	dtrace_ecb_t *ecb;
12008 
12009 	ASSERT(MUTEX_HELD(&dtrace_lock));
12010 
12011 	if (id == 0 || id > state->dts_necbs)
12012 		return (NULL);
12013 
12014 	ASSERT(state->dts_necbs > 0 && state->dts_ecbs != NULL);
12015 	ASSERT((ecb = state->dts_ecbs[id - 1]) == NULL || ecb->dte_epid == id);
12016 
12017 	return (state->dts_ecbs[id - 1]);
12018 }
12019 
12020 static dtrace_aggregation_t *
12021 dtrace_aggid2agg(dtrace_state_t *state, dtrace_aggid_t id)
12022 {
12023 	dtrace_aggregation_t *agg;
12024 
12025 	ASSERT(MUTEX_HELD(&dtrace_lock));
12026 
12027 	if (id == 0 || id > state->dts_naggregations)
12028 		return (NULL);
12029 
12030 	ASSERT(state->dts_naggregations > 0 && state->dts_aggregations != NULL);
12031 	ASSERT((agg = state->dts_aggregations[id - 1]) == NULL ||
12032 	    agg->dtag_id == id);
12033 
12034 	return (state->dts_aggregations[id - 1]);
12035 }
12036 
12037 /*
12038  * DTrace Buffer Functions
12039  *
12040  * The following functions manipulate DTrace buffers.  Most of these functions
12041  * are called in the context of establishing or processing consumer state;
12042  * exceptions are explicitly noted.
12043  */
12044 
12045 /*
12046  * Note:  called from cross call context.  This function switches the two
12047  * buffers on a given CPU.  The atomicity of this operation is assured by
12048  * disabling interrupts while the actual switch takes place; the disabling of
12049  * interrupts serializes the execution with any execution of dtrace_probe() on
12050  * the same CPU.
12051  */
12052 static void
12053 dtrace_buffer_switch(dtrace_buffer_t *buf)
12054 {
12055 	caddr_t tomax = buf->dtb_tomax;
12056 	caddr_t xamot = buf->dtb_xamot;
12057 	dtrace_icookie_t cookie;
12058 	hrtime_t now;
12059 
12060 	ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
12061 
12062 	cookie = dtrace_interrupt_disable();
12063 	now = dtrace_gethrtime();
12064 	buf->dtb_tomax = xamot;
12065 	buf->dtb_xamot = tomax;
12066 	buf->dtb_xamot_drops = buf->dtb_drops;
12067 	buf->dtb_xamot_offset = buf->dtb_offset;
12068 	buf->dtb_xamot_errors = buf->dtb_errors;
12069 	buf->dtb_xamot_flags = buf->dtb_flags;
12070 	buf->dtb_offset = 0;
12071 	buf->dtb_drops = 0;
12072 	buf->dtb_errors = 0;
12073 	buf->dtb_flags &= ~(DTRACEBUF_ERROR | DTRACEBUF_DROPPED);
12074 	buf->dtb_interval = now - buf->dtb_switched;
12075 	buf->dtb_switched = now;
12076 	dtrace_interrupt_enable(cookie);
12077 }
12078 
12079 /*
12080  * Note:  called from cross call context.  This function activates a buffer
12081  * on a CPU.  As with dtrace_buffer_switch(), the atomicity of the operation
12082  * is guaranteed by the disabling of interrupts.
12083  */
12084 static void
12085 dtrace_buffer_activate(dtrace_state_t *state)
12086 {
12087 	dtrace_buffer_t *buf;
12088 	dtrace_icookie_t cookie = dtrace_interrupt_disable();
12089 
12090 	buf = &state->dts_buffer[curcpu];
12091 
12092 	if (buf->dtb_tomax != NULL) {
12093 		/*
12094 		 * We might like to assert that the buffer is marked inactive,
12095 		 * but this isn't necessarily true:  the buffer for the CPU
12096 		 * that processes the BEGIN probe has its buffer activated
12097 		 * manually.  In this case, we take the (harmless) action
12098 		 * re-clearing the bit INACTIVE bit.
12099 		 */
12100 		buf->dtb_flags &= ~DTRACEBUF_INACTIVE;
12101 	}
12102 
12103 	dtrace_interrupt_enable(cookie);
12104 }
12105 
12106 #ifdef __FreeBSD__
12107 /*
12108  * Activate the specified per-CPU buffer.  This is used instead of
12109  * dtrace_buffer_activate() when APs have not yet started, i.e. when
12110  * activating anonymous state.
12111  */
12112 static void
12113 dtrace_buffer_activate_cpu(dtrace_state_t *state, int cpu)
12114 {
12115 
12116 	if (state->dts_buffer[cpu].dtb_tomax != NULL)
12117 		state->dts_buffer[cpu].dtb_flags &= ~DTRACEBUF_INACTIVE;
12118 }
12119 #endif
12120 
12121 static int
12122 dtrace_buffer_alloc(dtrace_buffer_t *bufs, size_t size, int flags,
12123     processorid_t cpu, int *factor)
12124 {
12125 #ifdef illumos
12126 	cpu_t *cp;
12127 #endif
12128 	dtrace_buffer_t *buf;
12129 	int allocated = 0, desired = 0;
12130 
12131 #ifdef illumos
12132 	ASSERT(MUTEX_HELD(&cpu_lock));
12133 	ASSERT(MUTEX_HELD(&dtrace_lock));
12134 
12135 	*factor = 1;
12136 
12137 	if (size > dtrace_nonroot_maxsize &&
12138 	    !PRIV_POLICY_CHOICE(CRED(), PRIV_ALL, B_FALSE))
12139 		return (EFBIG);
12140 
12141 	cp = cpu_list;
12142 
12143 	do {
12144 		if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
12145 			continue;
12146 
12147 		buf = &bufs[cp->cpu_id];
12148 
12149 		/*
12150 		 * If there is already a buffer allocated for this CPU, it
12151 		 * is only possible that this is a DR event.  In this case,
12152 		 */
12153 		if (buf->dtb_tomax != NULL) {
12154 			ASSERT(buf->dtb_size == size);
12155 			continue;
12156 		}
12157 
12158 		ASSERT(buf->dtb_xamot == NULL);
12159 
12160 		if ((buf->dtb_tomax = kmem_zalloc(size,
12161 		    KM_NOSLEEP | KM_NORMALPRI)) == NULL)
12162 			goto err;
12163 
12164 		buf->dtb_size = size;
12165 		buf->dtb_flags = flags;
12166 		buf->dtb_offset = 0;
12167 		buf->dtb_drops = 0;
12168 
12169 		if (flags & DTRACEBUF_NOSWITCH)
12170 			continue;
12171 
12172 		if ((buf->dtb_xamot = kmem_zalloc(size,
12173 		    KM_NOSLEEP | KM_NORMALPRI)) == NULL)
12174 			goto err;
12175 	} while ((cp = cp->cpu_next) != cpu_list);
12176 
12177 	return (0);
12178 
12179 err:
12180 	cp = cpu_list;
12181 
12182 	do {
12183 		if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
12184 			continue;
12185 
12186 		buf = &bufs[cp->cpu_id];
12187 		desired += 2;
12188 
12189 		if (buf->dtb_xamot != NULL) {
12190 			ASSERT(buf->dtb_tomax != NULL);
12191 			ASSERT(buf->dtb_size == size);
12192 			kmem_free(buf->dtb_xamot, size);
12193 			allocated++;
12194 		}
12195 
12196 		if (buf->dtb_tomax != NULL) {
12197 			ASSERT(buf->dtb_size == size);
12198 			kmem_free(buf->dtb_tomax, size);
12199 			allocated++;
12200 		}
12201 
12202 		buf->dtb_tomax = NULL;
12203 		buf->dtb_xamot = NULL;
12204 		buf->dtb_size = 0;
12205 	} while ((cp = cp->cpu_next) != cpu_list);
12206 #else
12207 	int i;
12208 
12209 	*factor = 1;
12210 #if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \
12211     defined(__mips__) || defined(__powerpc__) || defined(__riscv)
12212 	/*
12213 	 * FreeBSD isn't good at limiting the amount of memory we
12214 	 * ask to malloc, so let's place a limit here before trying
12215 	 * to do something that might well end in tears at bedtime.
12216 	 */
12217 	int bufsize_percpu_frac = dtrace_bufsize_max_frac * mp_ncpus;
12218 	if (size > physmem * PAGE_SIZE / bufsize_percpu_frac)
12219 		return (ENOMEM);
12220 #endif
12221 
12222 	ASSERT(MUTEX_HELD(&dtrace_lock));
12223 	CPU_FOREACH(i) {
12224 		if (cpu != DTRACE_CPUALL && cpu != i)
12225 			continue;
12226 
12227 		buf = &bufs[i];
12228 
12229 		/*
12230 		 * If there is already a buffer allocated for this CPU, it
12231 		 * is only possible that this is a DR event.  In this case,
12232 		 * the buffer size must match our specified size.
12233 		 */
12234 		if (buf->dtb_tomax != NULL) {
12235 			ASSERT(buf->dtb_size == size);
12236 			continue;
12237 		}
12238 
12239 		ASSERT(buf->dtb_xamot == NULL);
12240 
12241 		if ((buf->dtb_tomax = kmem_zalloc(size,
12242 		    KM_NOSLEEP | KM_NORMALPRI)) == NULL)
12243 			goto err;
12244 
12245 		buf->dtb_size = size;
12246 		buf->dtb_flags = flags;
12247 		buf->dtb_offset = 0;
12248 		buf->dtb_drops = 0;
12249 
12250 		if (flags & DTRACEBUF_NOSWITCH)
12251 			continue;
12252 
12253 		if ((buf->dtb_xamot = kmem_zalloc(size,
12254 		    KM_NOSLEEP | KM_NORMALPRI)) == NULL)
12255 			goto err;
12256 	}
12257 
12258 	return (0);
12259 
12260 err:
12261 	/*
12262 	 * Error allocating memory, so free the buffers that were
12263 	 * allocated before the failed allocation.
12264 	 */
12265 	CPU_FOREACH(i) {
12266 		if (cpu != DTRACE_CPUALL && cpu != i)
12267 			continue;
12268 
12269 		buf = &bufs[i];
12270 		desired += 2;
12271 
12272 		if (buf->dtb_xamot != NULL) {
12273 			ASSERT(buf->dtb_tomax != NULL);
12274 			ASSERT(buf->dtb_size == size);
12275 			kmem_free(buf->dtb_xamot, size);
12276 			allocated++;
12277 		}
12278 
12279 		if (buf->dtb_tomax != NULL) {
12280 			ASSERT(buf->dtb_size == size);
12281 			kmem_free(buf->dtb_tomax, size);
12282 			allocated++;
12283 		}
12284 
12285 		buf->dtb_tomax = NULL;
12286 		buf->dtb_xamot = NULL;
12287 		buf->dtb_size = 0;
12288 
12289 	}
12290 #endif
12291 	*factor = desired / (allocated > 0 ? allocated : 1);
12292 
12293 	return (ENOMEM);
12294 }
12295 
12296 /*
12297  * Note:  called from probe context.  This function just increments the drop
12298  * count on a buffer.  It has been made a function to allow for the
12299  * possibility of understanding the source of mysterious drop counts.  (A
12300  * problem for which one may be particularly disappointed that DTrace cannot
12301  * be used to understand DTrace.)
12302  */
12303 static void
12304 dtrace_buffer_drop(dtrace_buffer_t *buf)
12305 {
12306 	buf->dtb_drops++;
12307 }
12308 
12309 /*
12310  * Note:  called from probe context.  This function is called to reserve space
12311  * in a buffer.  If mstate is non-NULL, sets the scratch base and size in the
12312  * mstate.  Returns the new offset in the buffer, or a negative value if an
12313  * error has occurred.
12314  */
12315 static intptr_t
12316 dtrace_buffer_reserve(dtrace_buffer_t *buf, size_t needed, size_t align,
12317     dtrace_state_t *state, dtrace_mstate_t *mstate)
12318 {
12319 	intptr_t offs = buf->dtb_offset, soffs;
12320 	intptr_t woffs;
12321 	caddr_t tomax;
12322 	size_t total;
12323 
12324 	if (buf->dtb_flags & DTRACEBUF_INACTIVE)
12325 		return (-1);
12326 
12327 	if ((tomax = buf->dtb_tomax) == NULL) {
12328 		dtrace_buffer_drop(buf);
12329 		return (-1);
12330 	}
12331 
12332 	if (!(buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL))) {
12333 		while (offs & (align - 1)) {
12334 			/*
12335 			 * Assert that our alignment is off by a number which
12336 			 * is itself sizeof (uint32_t) aligned.
12337 			 */
12338 			ASSERT(!((align - (offs & (align - 1))) &
12339 			    (sizeof (uint32_t) - 1)));
12340 			DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
12341 			offs += sizeof (uint32_t);
12342 		}
12343 
12344 		if ((soffs = offs + needed) > buf->dtb_size) {
12345 			dtrace_buffer_drop(buf);
12346 			return (-1);
12347 		}
12348 
12349 		if (mstate == NULL)
12350 			return (offs);
12351 
12352 		mstate->dtms_scratch_base = (uintptr_t)tomax + soffs;
12353 		mstate->dtms_scratch_size = buf->dtb_size - soffs;
12354 		mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
12355 
12356 		return (offs);
12357 	}
12358 
12359 	if (buf->dtb_flags & DTRACEBUF_FILL) {
12360 		if (state->dts_activity != DTRACE_ACTIVITY_COOLDOWN &&
12361 		    (buf->dtb_flags & DTRACEBUF_FULL))
12362 			return (-1);
12363 		goto out;
12364 	}
12365 
12366 	total = needed + (offs & (align - 1));
12367 
12368 	/*
12369 	 * For a ring buffer, life is quite a bit more complicated.  Before
12370 	 * we can store any padding, we need to adjust our wrapping offset.
12371 	 * (If we've never before wrapped or we're not about to, no adjustment
12372 	 * is required.)
12373 	 */
12374 	if ((buf->dtb_flags & DTRACEBUF_WRAPPED) ||
12375 	    offs + total > buf->dtb_size) {
12376 		woffs = buf->dtb_xamot_offset;
12377 
12378 		if (offs + total > buf->dtb_size) {
12379 			/*
12380 			 * We can't fit in the end of the buffer.  First, a
12381 			 * sanity check that we can fit in the buffer at all.
12382 			 */
12383 			if (total > buf->dtb_size) {
12384 				dtrace_buffer_drop(buf);
12385 				return (-1);
12386 			}
12387 
12388 			/*
12389 			 * We're going to be storing at the top of the buffer,
12390 			 * so now we need to deal with the wrapped offset.  We
12391 			 * only reset our wrapped offset to 0 if it is
12392 			 * currently greater than the current offset.  If it
12393 			 * is less than the current offset, it is because a
12394 			 * previous allocation induced a wrap -- but the
12395 			 * allocation didn't subsequently take the space due
12396 			 * to an error or false predicate evaluation.  In this
12397 			 * case, we'll just leave the wrapped offset alone: if
12398 			 * the wrapped offset hasn't been advanced far enough
12399 			 * for this allocation, it will be adjusted in the
12400 			 * lower loop.
12401 			 */
12402 			if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
12403 				if (woffs >= offs)
12404 					woffs = 0;
12405 			} else {
12406 				woffs = 0;
12407 			}
12408 
12409 			/*
12410 			 * Now we know that we're going to be storing to the
12411 			 * top of the buffer and that there is room for us
12412 			 * there.  We need to clear the buffer from the current
12413 			 * offset to the end (there may be old gunk there).
12414 			 */
12415 			while (offs < buf->dtb_size)
12416 				tomax[offs++] = 0;
12417 
12418 			/*
12419 			 * We need to set our offset to zero.  And because we
12420 			 * are wrapping, we need to set the bit indicating as
12421 			 * much.  We can also adjust our needed space back
12422 			 * down to the space required by the ECB -- we know
12423 			 * that the top of the buffer is aligned.
12424 			 */
12425 			offs = 0;
12426 			total = needed;
12427 			buf->dtb_flags |= DTRACEBUF_WRAPPED;
12428 		} else {
12429 			/*
12430 			 * There is room for us in the buffer, so we simply
12431 			 * need to check the wrapped offset.
12432 			 */
12433 			if (woffs < offs) {
12434 				/*
12435 				 * The wrapped offset is less than the offset.
12436 				 * This can happen if we allocated buffer space
12437 				 * that induced a wrap, but then we didn't
12438 				 * subsequently take the space due to an error
12439 				 * or false predicate evaluation.  This is
12440 				 * okay; we know that _this_ allocation isn't
12441 				 * going to induce a wrap.  We still can't
12442 				 * reset the wrapped offset to be zero,
12443 				 * however: the space may have been trashed in
12444 				 * the previous failed probe attempt.  But at
12445 				 * least the wrapped offset doesn't need to
12446 				 * be adjusted at all...
12447 				 */
12448 				goto out;
12449 			}
12450 		}
12451 
12452 		while (offs + total > woffs) {
12453 			dtrace_epid_t epid = *(uint32_t *)(tomax + woffs);
12454 			size_t size;
12455 
12456 			if (epid == DTRACE_EPIDNONE) {
12457 				size = sizeof (uint32_t);
12458 			} else {
12459 				ASSERT3U(epid, <=, state->dts_necbs);
12460 				ASSERT(state->dts_ecbs[epid - 1] != NULL);
12461 
12462 				size = state->dts_ecbs[epid - 1]->dte_size;
12463 			}
12464 
12465 			ASSERT(woffs + size <= buf->dtb_size);
12466 			ASSERT(size != 0);
12467 
12468 			if (woffs + size == buf->dtb_size) {
12469 				/*
12470 				 * We've reached the end of the buffer; we want
12471 				 * to set the wrapped offset to 0 and break
12472 				 * out.  However, if the offs is 0, then we're
12473 				 * in a strange edge-condition:  the amount of
12474 				 * space that we want to reserve plus the size
12475 				 * of the record that we're overwriting is
12476 				 * greater than the size of the buffer.  This
12477 				 * is problematic because if we reserve the
12478 				 * space but subsequently don't consume it (due
12479 				 * to a failed predicate or error) the wrapped
12480 				 * offset will be 0 -- yet the EPID at offset 0
12481 				 * will not be committed.  This situation is
12482 				 * relatively easy to deal with:  if we're in
12483 				 * this case, the buffer is indistinguishable
12484 				 * from one that hasn't wrapped; we need only
12485 				 * finish the job by clearing the wrapped bit,
12486 				 * explicitly setting the offset to be 0, and
12487 				 * zero'ing out the old data in the buffer.
12488 				 */
12489 				if (offs == 0) {
12490 					buf->dtb_flags &= ~DTRACEBUF_WRAPPED;
12491 					buf->dtb_offset = 0;
12492 					woffs = total;
12493 
12494 					while (woffs < buf->dtb_size)
12495 						tomax[woffs++] = 0;
12496 				}
12497 
12498 				woffs = 0;
12499 				break;
12500 			}
12501 
12502 			woffs += size;
12503 		}
12504 
12505 		/*
12506 		 * We have a wrapped offset.  It may be that the wrapped offset
12507 		 * has become zero -- that's okay.
12508 		 */
12509 		buf->dtb_xamot_offset = woffs;
12510 	}
12511 
12512 out:
12513 	/*
12514 	 * Now we can plow the buffer with any necessary padding.
12515 	 */
12516 	while (offs & (align - 1)) {
12517 		/*
12518 		 * Assert that our alignment is off by a number which
12519 		 * is itself sizeof (uint32_t) aligned.
12520 		 */
12521 		ASSERT(!((align - (offs & (align - 1))) &
12522 		    (sizeof (uint32_t) - 1)));
12523 		DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
12524 		offs += sizeof (uint32_t);
12525 	}
12526 
12527 	if (buf->dtb_flags & DTRACEBUF_FILL) {
12528 		if (offs + needed > buf->dtb_size - state->dts_reserve) {
12529 			buf->dtb_flags |= DTRACEBUF_FULL;
12530 			return (-1);
12531 		}
12532 	}
12533 
12534 	if (mstate == NULL)
12535 		return (offs);
12536 
12537 	/*
12538 	 * For ring buffers and fill buffers, the scratch space is always
12539 	 * the inactive buffer.
12540 	 */
12541 	mstate->dtms_scratch_base = (uintptr_t)buf->dtb_xamot;
12542 	mstate->dtms_scratch_size = buf->dtb_size;
12543 	mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
12544 
12545 	return (offs);
12546 }
12547 
12548 static void
12549 dtrace_buffer_polish(dtrace_buffer_t *buf)
12550 {
12551 	ASSERT(buf->dtb_flags & DTRACEBUF_RING);
12552 	ASSERT(MUTEX_HELD(&dtrace_lock));
12553 
12554 	if (!(buf->dtb_flags & DTRACEBUF_WRAPPED))
12555 		return;
12556 
12557 	/*
12558 	 * We need to polish the ring buffer.  There are three cases:
12559 	 *
12560 	 * - The first (and presumably most common) is that there is no gap
12561 	 *   between the buffer offset and the wrapped offset.  In this case,
12562 	 *   there is nothing in the buffer that isn't valid data; we can
12563 	 *   mark the buffer as polished and return.
12564 	 *
12565 	 * - The second (less common than the first but still more common
12566 	 *   than the third) is that there is a gap between the buffer offset
12567 	 *   and the wrapped offset, and the wrapped offset is larger than the
12568 	 *   buffer offset.  This can happen because of an alignment issue, or
12569 	 *   can happen because of a call to dtrace_buffer_reserve() that
12570 	 *   didn't subsequently consume the buffer space.  In this case,
12571 	 *   we need to zero the data from the buffer offset to the wrapped
12572 	 *   offset.
12573 	 *
12574 	 * - The third (and least common) is that there is a gap between the
12575 	 *   buffer offset and the wrapped offset, but the wrapped offset is
12576 	 *   _less_ than the buffer offset.  This can only happen because a
12577 	 *   call to dtrace_buffer_reserve() induced a wrap, but the space
12578 	 *   was not subsequently consumed.  In this case, we need to zero the
12579 	 *   space from the offset to the end of the buffer _and_ from the
12580 	 *   top of the buffer to the wrapped offset.
12581 	 */
12582 	if (buf->dtb_offset < buf->dtb_xamot_offset) {
12583 		bzero(buf->dtb_tomax + buf->dtb_offset,
12584 		    buf->dtb_xamot_offset - buf->dtb_offset);
12585 	}
12586 
12587 	if (buf->dtb_offset > buf->dtb_xamot_offset) {
12588 		bzero(buf->dtb_tomax + buf->dtb_offset,
12589 		    buf->dtb_size - buf->dtb_offset);
12590 		bzero(buf->dtb_tomax, buf->dtb_xamot_offset);
12591 	}
12592 }
12593 
12594 /*
12595  * This routine determines if data generated at the specified time has likely
12596  * been entirely consumed at user-level.  This routine is called to determine
12597  * if an ECB on a defunct probe (but for an active enabling) can be safely
12598  * disabled and destroyed.
12599  */
12600 static int
12601 dtrace_buffer_consumed(dtrace_buffer_t *bufs, hrtime_t when)
12602 {
12603 	int i;
12604 
12605 	for (i = 0; i < NCPU; i++) {
12606 		dtrace_buffer_t *buf = &bufs[i];
12607 
12608 		if (buf->dtb_size == 0)
12609 			continue;
12610 
12611 		if (buf->dtb_flags & DTRACEBUF_RING)
12612 			return (0);
12613 
12614 		if (!buf->dtb_switched && buf->dtb_offset != 0)
12615 			return (0);
12616 
12617 		if (buf->dtb_switched - buf->dtb_interval < when)
12618 			return (0);
12619 	}
12620 
12621 	return (1);
12622 }
12623 
12624 static void
12625 dtrace_buffer_free(dtrace_buffer_t *bufs)
12626 {
12627 	int i;
12628 
12629 	for (i = 0; i < NCPU; i++) {
12630 		dtrace_buffer_t *buf = &bufs[i];
12631 
12632 		if (buf->dtb_tomax == NULL) {
12633 			ASSERT(buf->dtb_xamot == NULL);
12634 			ASSERT(buf->dtb_size == 0);
12635 			continue;
12636 		}
12637 
12638 		if (buf->dtb_xamot != NULL) {
12639 			ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
12640 			kmem_free(buf->dtb_xamot, buf->dtb_size);
12641 		}
12642 
12643 		kmem_free(buf->dtb_tomax, buf->dtb_size);
12644 		buf->dtb_size = 0;
12645 		buf->dtb_tomax = NULL;
12646 		buf->dtb_xamot = NULL;
12647 	}
12648 }
12649 
12650 /*
12651  * DTrace Enabling Functions
12652  */
12653 static dtrace_enabling_t *
12654 dtrace_enabling_create(dtrace_vstate_t *vstate)
12655 {
12656 	dtrace_enabling_t *enab;
12657 
12658 	enab = kmem_zalloc(sizeof (dtrace_enabling_t), KM_SLEEP);
12659 	enab->dten_vstate = vstate;
12660 
12661 	return (enab);
12662 }
12663 
12664 static void
12665 dtrace_enabling_add(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb)
12666 {
12667 	dtrace_ecbdesc_t **ndesc;
12668 	size_t osize, nsize;
12669 
12670 	/*
12671 	 * We can't add to enablings after we've enabled them, or after we've
12672 	 * retained them.
12673 	 */
12674 	ASSERT(enab->dten_probegen == 0);
12675 	ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
12676 
12677 	if (enab->dten_ndesc < enab->dten_maxdesc) {
12678 		enab->dten_desc[enab->dten_ndesc++] = ecb;
12679 		return;
12680 	}
12681 
12682 	osize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
12683 
12684 	if (enab->dten_maxdesc == 0) {
12685 		enab->dten_maxdesc = 1;
12686 	} else {
12687 		enab->dten_maxdesc <<= 1;
12688 	}
12689 
12690 	ASSERT(enab->dten_ndesc < enab->dten_maxdesc);
12691 
12692 	nsize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
12693 	ndesc = kmem_zalloc(nsize, KM_SLEEP);
12694 	bcopy(enab->dten_desc, ndesc, osize);
12695 	if (enab->dten_desc != NULL)
12696 		kmem_free(enab->dten_desc, osize);
12697 
12698 	enab->dten_desc = ndesc;
12699 	enab->dten_desc[enab->dten_ndesc++] = ecb;
12700 }
12701 
12702 static void
12703 dtrace_enabling_addlike(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb,
12704     dtrace_probedesc_t *pd)
12705 {
12706 	dtrace_ecbdesc_t *new;
12707 	dtrace_predicate_t *pred;
12708 	dtrace_actdesc_t *act;
12709 
12710 	/*
12711 	 * We're going to create a new ECB description that matches the
12712 	 * specified ECB in every way, but has the specified probe description.
12713 	 */
12714 	new = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
12715 
12716 	if ((pred = ecb->dted_pred.dtpdd_predicate) != NULL)
12717 		dtrace_predicate_hold(pred);
12718 
12719 	for (act = ecb->dted_action; act != NULL; act = act->dtad_next)
12720 		dtrace_actdesc_hold(act);
12721 
12722 	new->dted_action = ecb->dted_action;
12723 	new->dted_pred = ecb->dted_pred;
12724 	new->dted_probe = *pd;
12725 	new->dted_uarg = ecb->dted_uarg;
12726 
12727 	dtrace_enabling_add(enab, new);
12728 }
12729 
12730 static void
12731 dtrace_enabling_dump(dtrace_enabling_t *enab)
12732 {
12733 	int i;
12734 
12735 	for (i = 0; i < enab->dten_ndesc; i++) {
12736 		dtrace_probedesc_t *desc = &enab->dten_desc[i]->dted_probe;
12737 
12738 #ifdef __FreeBSD__
12739 		printf("dtrace: enabling probe %d (%s:%s:%s:%s)\n", i,
12740 		    desc->dtpd_provider, desc->dtpd_mod,
12741 		    desc->dtpd_func, desc->dtpd_name);
12742 #else
12743 		cmn_err(CE_NOTE, "enabling probe %d (%s:%s:%s:%s)", i,
12744 		    desc->dtpd_provider, desc->dtpd_mod,
12745 		    desc->dtpd_func, desc->dtpd_name);
12746 #endif
12747 	}
12748 }
12749 
12750 static void
12751 dtrace_enabling_destroy(dtrace_enabling_t *enab)
12752 {
12753 	int i;
12754 	dtrace_ecbdesc_t *ep;
12755 	dtrace_vstate_t *vstate = enab->dten_vstate;
12756 
12757 	ASSERT(MUTEX_HELD(&dtrace_lock));
12758 
12759 	for (i = 0; i < enab->dten_ndesc; i++) {
12760 		dtrace_actdesc_t *act, *next;
12761 		dtrace_predicate_t *pred;
12762 
12763 		ep = enab->dten_desc[i];
12764 
12765 		if ((pred = ep->dted_pred.dtpdd_predicate) != NULL)
12766 			dtrace_predicate_release(pred, vstate);
12767 
12768 		for (act = ep->dted_action; act != NULL; act = next) {
12769 			next = act->dtad_next;
12770 			dtrace_actdesc_release(act, vstate);
12771 		}
12772 
12773 		kmem_free(ep, sizeof (dtrace_ecbdesc_t));
12774 	}
12775 
12776 	if (enab->dten_desc != NULL)
12777 		kmem_free(enab->dten_desc,
12778 		    enab->dten_maxdesc * sizeof (dtrace_enabling_t *));
12779 
12780 	/*
12781 	 * If this was a retained enabling, decrement the dts_nretained count
12782 	 * and take it off of the dtrace_retained list.
12783 	 */
12784 	if (enab->dten_prev != NULL || enab->dten_next != NULL ||
12785 	    dtrace_retained == enab) {
12786 		ASSERT(enab->dten_vstate->dtvs_state != NULL);
12787 		ASSERT(enab->dten_vstate->dtvs_state->dts_nretained > 0);
12788 		enab->dten_vstate->dtvs_state->dts_nretained--;
12789 		dtrace_retained_gen++;
12790 	}
12791 
12792 	if (enab->dten_prev == NULL) {
12793 		if (dtrace_retained == enab) {
12794 			dtrace_retained = enab->dten_next;
12795 
12796 			if (dtrace_retained != NULL)
12797 				dtrace_retained->dten_prev = NULL;
12798 		}
12799 	} else {
12800 		ASSERT(enab != dtrace_retained);
12801 		ASSERT(dtrace_retained != NULL);
12802 		enab->dten_prev->dten_next = enab->dten_next;
12803 	}
12804 
12805 	if (enab->dten_next != NULL) {
12806 		ASSERT(dtrace_retained != NULL);
12807 		enab->dten_next->dten_prev = enab->dten_prev;
12808 	}
12809 
12810 	kmem_free(enab, sizeof (dtrace_enabling_t));
12811 }
12812 
12813 static int
12814 dtrace_enabling_retain(dtrace_enabling_t *enab)
12815 {
12816 	dtrace_state_t *state;
12817 
12818 	ASSERT(MUTEX_HELD(&dtrace_lock));
12819 	ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
12820 	ASSERT(enab->dten_vstate != NULL);
12821 
12822 	state = enab->dten_vstate->dtvs_state;
12823 	ASSERT(state != NULL);
12824 
12825 	/*
12826 	 * We only allow each state to retain dtrace_retain_max enablings.
12827 	 */
12828 	if (state->dts_nretained >= dtrace_retain_max)
12829 		return (ENOSPC);
12830 
12831 	state->dts_nretained++;
12832 	dtrace_retained_gen++;
12833 
12834 	if (dtrace_retained == NULL) {
12835 		dtrace_retained = enab;
12836 		return (0);
12837 	}
12838 
12839 	enab->dten_next = dtrace_retained;
12840 	dtrace_retained->dten_prev = enab;
12841 	dtrace_retained = enab;
12842 
12843 	return (0);
12844 }
12845 
12846 static int
12847 dtrace_enabling_replicate(dtrace_state_t *state, dtrace_probedesc_t *match,
12848     dtrace_probedesc_t *create)
12849 {
12850 	dtrace_enabling_t *new, *enab;
12851 	int found = 0, err = ENOENT;
12852 
12853 	ASSERT(MUTEX_HELD(&dtrace_lock));
12854 	ASSERT(strlen(match->dtpd_provider) < DTRACE_PROVNAMELEN);
12855 	ASSERT(strlen(match->dtpd_mod) < DTRACE_MODNAMELEN);
12856 	ASSERT(strlen(match->dtpd_func) < DTRACE_FUNCNAMELEN);
12857 	ASSERT(strlen(match->dtpd_name) < DTRACE_NAMELEN);
12858 
12859 	new = dtrace_enabling_create(&state->dts_vstate);
12860 
12861 	/*
12862 	 * Iterate over all retained enablings, looking for enablings that
12863 	 * match the specified state.
12864 	 */
12865 	for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
12866 		int i;
12867 
12868 		/*
12869 		 * dtvs_state can only be NULL for helper enablings -- and
12870 		 * helper enablings can't be retained.
12871 		 */
12872 		ASSERT(enab->dten_vstate->dtvs_state != NULL);
12873 
12874 		if (enab->dten_vstate->dtvs_state != state)
12875 			continue;
12876 
12877 		/*
12878 		 * Now iterate over each probe description; we're looking for
12879 		 * an exact match to the specified probe description.
12880 		 */
12881 		for (i = 0; i < enab->dten_ndesc; i++) {
12882 			dtrace_ecbdesc_t *ep = enab->dten_desc[i];
12883 			dtrace_probedesc_t *pd = &ep->dted_probe;
12884 
12885 			if (strcmp(pd->dtpd_provider, match->dtpd_provider))
12886 				continue;
12887 
12888 			if (strcmp(pd->dtpd_mod, match->dtpd_mod))
12889 				continue;
12890 
12891 			if (strcmp(pd->dtpd_func, match->dtpd_func))
12892 				continue;
12893 
12894 			if (strcmp(pd->dtpd_name, match->dtpd_name))
12895 				continue;
12896 
12897 			/*
12898 			 * We have a winning probe!  Add it to our growing
12899 			 * enabling.
12900 			 */
12901 			found = 1;
12902 			dtrace_enabling_addlike(new, ep, create);
12903 		}
12904 	}
12905 
12906 	if (!found || (err = dtrace_enabling_retain(new)) != 0) {
12907 		dtrace_enabling_destroy(new);
12908 		return (err);
12909 	}
12910 
12911 	return (0);
12912 }
12913 
12914 static void
12915 dtrace_enabling_retract(dtrace_state_t *state)
12916 {
12917 	dtrace_enabling_t *enab, *next;
12918 
12919 	ASSERT(MUTEX_HELD(&dtrace_lock));
12920 
12921 	/*
12922 	 * Iterate over all retained enablings, destroy the enablings retained
12923 	 * for the specified state.
12924 	 */
12925 	for (enab = dtrace_retained; enab != NULL; enab = next) {
12926 		next = enab->dten_next;
12927 
12928 		/*
12929 		 * dtvs_state can only be NULL for helper enablings -- and
12930 		 * helper enablings can't be retained.
12931 		 */
12932 		ASSERT(enab->dten_vstate->dtvs_state != NULL);
12933 
12934 		if (enab->dten_vstate->dtvs_state == state) {
12935 			ASSERT(state->dts_nretained > 0);
12936 			dtrace_enabling_destroy(enab);
12937 		}
12938 	}
12939 
12940 	ASSERT(state->dts_nretained == 0);
12941 }
12942 
12943 static int
12944 dtrace_enabling_match(dtrace_enabling_t *enab, int *nmatched)
12945 {
12946 	int i = 0;
12947 	int matched = 0;
12948 
12949 	ASSERT(MUTEX_HELD(&cpu_lock));
12950 	ASSERT(MUTEX_HELD(&dtrace_lock));
12951 
12952 	for (i = 0; i < enab->dten_ndesc; i++) {
12953 		dtrace_ecbdesc_t *ep = enab->dten_desc[i];
12954 
12955 		enab->dten_current = ep;
12956 		enab->dten_error = 0;
12957 
12958 		matched += dtrace_probe_enable(&ep->dted_probe, enab);
12959 
12960 		if (enab->dten_error != 0) {
12961 			/*
12962 			 * If we get an error half-way through enabling the
12963 			 * probes, we kick out -- perhaps with some number of
12964 			 * them enabled.  Leaving enabled probes enabled may
12965 			 * be slightly confusing for user-level, but we expect
12966 			 * that no one will attempt to actually drive on in
12967 			 * the face of such errors.  If this is an anonymous
12968 			 * enabling (indicated with a NULL nmatched pointer),
12969 			 * we cmn_err() a message.  We aren't expecting to
12970 			 * get such an error -- such as it can exist at all,
12971 			 * it would be a result of corrupted DOF in the driver
12972 			 * properties.
12973 			 */
12974 			if (nmatched == NULL) {
12975 				cmn_err(CE_WARN, "dtrace_enabling_match() "
12976 				    "error on %p: %d", (void *)ep,
12977 				    enab->dten_error);
12978 			}
12979 
12980 			return (enab->dten_error);
12981 		}
12982 	}
12983 
12984 	enab->dten_probegen = dtrace_probegen;
12985 	if (nmatched != NULL)
12986 		*nmatched = matched;
12987 
12988 	return (0);
12989 }
12990 
12991 static void
12992 dtrace_enabling_matchall(void)
12993 {
12994 	dtrace_enabling_t *enab;
12995 
12996 	mutex_enter(&cpu_lock);
12997 	mutex_enter(&dtrace_lock);
12998 
12999 	/*
13000 	 * Iterate over all retained enablings to see if any probes match
13001 	 * against them.  We only perform this operation on enablings for which
13002 	 * we have sufficient permissions by virtue of being in the global zone
13003 	 * or in the same zone as the DTrace client.  Because we can be called
13004 	 * after dtrace_detach() has been called, we cannot assert that there
13005 	 * are retained enablings.  We can safely load from dtrace_retained,
13006 	 * however:  the taskq_destroy() at the end of dtrace_detach() will
13007 	 * block pending our completion.
13008 	 */
13009 	for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
13010 #ifdef illumos
13011 		cred_t *cr = enab->dten_vstate->dtvs_state->dts_cred.dcr_cred;
13012 
13013 		if (INGLOBALZONE(curproc) ||
13014 		    cr != NULL && getzoneid() == crgetzoneid(cr))
13015 #endif
13016 			(void) dtrace_enabling_match(enab, NULL);
13017 	}
13018 
13019 	mutex_exit(&dtrace_lock);
13020 	mutex_exit(&cpu_lock);
13021 }
13022 
13023 /*
13024  * If an enabling is to be enabled without having matched probes (that is, if
13025  * dtrace_state_go() is to be called on the underlying dtrace_state_t), the
13026  * enabling must be _primed_ by creating an ECB for every ECB description.
13027  * This must be done to assure that we know the number of speculations, the
13028  * number of aggregations, the minimum buffer size needed, etc. before we
13029  * transition out of DTRACE_ACTIVITY_INACTIVE.  To do this without actually
13030  * enabling any probes, we create ECBs for every ECB decription, but with a
13031  * NULL probe -- which is exactly what this function does.
13032  */
13033 static void
13034 dtrace_enabling_prime(dtrace_state_t *state)
13035 {
13036 	dtrace_enabling_t *enab;
13037 	int i;
13038 
13039 	for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
13040 		ASSERT(enab->dten_vstate->dtvs_state != NULL);
13041 
13042 		if (enab->dten_vstate->dtvs_state != state)
13043 			continue;
13044 
13045 		/*
13046 		 * We don't want to prime an enabling more than once, lest
13047 		 * we allow a malicious user to induce resource exhaustion.
13048 		 * (The ECBs that result from priming an enabling aren't
13049 		 * leaked -- but they also aren't deallocated until the
13050 		 * consumer state is destroyed.)
13051 		 */
13052 		if (enab->dten_primed)
13053 			continue;
13054 
13055 		for (i = 0; i < enab->dten_ndesc; i++) {
13056 			enab->dten_current = enab->dten_desc[i];
13057 			(void) dtrace_probe_enable(NULL, enab);
13058 		}
13059 
13060 		enab->dten_primed = 1;
13061 	}
13062 }
13063 
13064 /*
13065  * Called to indicate that probes should be provided due to retained
13066  * enablings.  This is implemented in terms of dtrace_probe_provide(), but it
13067  * must take an initial lap through the enabling calling the dtps_provide()
13068  * entry point explicitly to allow for autocreated probes.
13069  */
13070 static void
13071 dtrace_enabling_provide(dtrace_provider_t *prv)
13072 {
13073 	int i, all = 0;
13074 	dtrace_probedesc_t desc;
13075 	dtrace_genid_t gen;
13076 
13077 	ASSERT(MUTEX_HELD(&dtrace_lock));
13078 	ASSERT(MUTEX_HELD(&dtrace_provider_lock));
13079 
13080 	if (prv == NULL) {
13081 		all = 1;
13082 		prv = dtrace_provider;
13083 	}
13084 
13085 	do {
13086 		dtrace_enabling_t *enab;
13087 		void *parg = prv->dtpv_arg;
13088 
13089 retry:
13090 		gen = dtrace_retained_gen;
13091 		for (enab = dtrace_retained; enab != NULL;
13092 		    enab = enab->dten_next) {
13093 			for (i = 0; i < enab->dten_ndesc; i++) {
13094 				desc = enab->dten_desc[i]->dted_probe;
13095 				mutex_exit(&dtrace_lock);
13096 				prv->dtpv_pops.dtps_provide(parg, &desc);
13097 				mutex_enter(&dtrace_lock);
13098 				/*
13099 				 * Process the retained enablings again if
13100 				 * they have changed while we weren't holding
13101 				 * dtrace_lock.
13102 				 */
13103 				if (gen != dtrace_retained_gen)
13104 					goto retry;
13105 			}
13106 		}
13107 	} while (all && (prv = prv->dtpv_next) != NULL);
13108 
13109 	mutex_exit(&dtrace_lock);
13110 	dtrace_probe_provide(NULL, all ? NULL : prv);
13111 	mutex_enter(&dtrace_lock);
13112 }
13113 
13114 /*
13115  * Called to reap ECBs that are attached to probes from defunct providers.
13116  */
13117 static void
13118 dtrace_enabling_reap(void)
13119 {
13120 	dtrace_provider_t *prov;
13121 	dtrace_probe_t *probe;
13122 	dtrace_ecb_t *ecb;
13123 	hrtime_t when;
13124 	int i;
13125 
13126 	mutex_enter(&cpu_lock);
13127 	mutex_enter(&dtrace_lock);
13128 
13129 	for (i = 0; i < dtrace_nprobes; i++) {
13130 		if ((probe = dtrace_probes[i]) == NULL)
13131 			continue;
13132 
13133 		if (probe->dtpr_ecb == NULL)
13134 			continue;
13135 
13136 		prov = probe->dtpr_provider;
13137 
13138 		if ((when = prov->dtpv_defunct) == 0)
13139 			continue;
13140 
13141 		/*
13142 		 * We have ECBs on a defunct provider:  we want to reap these
13143 		 * ECBs to allow the provider to unregister.  The destruction
13144 		 * of these ECBs must be done carefully:  if we destroy the ECB
13145 		 * and the consumer later wishes to consume an EPID that
13146 		 * corresponds to the destroyed ECB (and if the EPID metadata
13147 		 * has not been previously consumed), the consumer will abort
13148 		 * processing on the unknown EPID.  To reduce (but not, sadly,
13149 		 * eliminate) the possibility of this, we will only destroy an
13150 		 * ECB for a defunct provider if, for the state that
13151 		 * corresponds to the ECB:
13152 		 *
13153 		 *  (a)	There is no speculative tracing (which can effectively
13154 		 *	cache an EPID for an arbitrary amount of time).
13155 		 *
13156 		 *  (b)	The principal buffers have been switched twice since the
13157 		 *	provider became defunct.
13158 		 *
13159 		 *  (c)	The aggregation buffers are of zero size or have been
13160 		 *	switched twice since the provider became defunct.
13161 		 *
13162 		 * We use dts_speculates to determine (a) and call a function
13163 		 * (dtrace_buffer_consumed()) to determine (b) and (c).  Note
13164 		 * that as soon as we've been unable to destroy one of the ECBs
13165 		 * associated with the probe, we quit trying -- reaping is only
13166 		 * fruitful in as much as we can destroy all ECBs associated
13167 		 * with the defunct provider's probes.
13168 		 */
13169 		while ((ecb = probe->dtpr_ecb) != NULL) {
13170 			dtrace_state_t *state = ecb->dte_state;
13171 			dtrace_buffer_t *buf = state->dts_buffer;
13172 			dtrace_buffer_t *aggbuf = state->dts_aggbuffer;
13173 
13174 			if (state->dts_speculates)
13175 				break;
13176 
13177 			if (!dtrace_buffer_consumed(buf, when))
13178 				break;
13179 
13180 			if (!dtrace_buffer_consumed(aggbuf, when))
13181 				break;
13182 
13183 			dtrace_ecb_disable(ecb);
13184 			ASSERT(probe->dtpr_ecb != ecb);
13185 			dtrace_ecb_destroy(ecb);
13186 		}
13187 	}
13188 
13189 	mutex_exit(&dtrace_lock);
13190 	mutex_exit(&cpu_lock);
13191 }
13192 
13193 /*
13194  * DTrace DOF Functions
13195  */
13196 /*ARGSUSED*/
13197 static void
13198 dtrace_dof_error(dof_hdr_t *dof, const char *str)
13199 {
13200 	if (dtrace_err_verbose)
13201 		cmn_err(CE_WARN, "failed to process DOF: %s", str);
13202 
13203 #ifdef DTRACE_ERRDEBUG
13204 	dtrace_errdebug(str);
13205 #endif
13206 }
13207 
13208 /*
13209  * Create DOF out of a currently enabled state.  Right now, we only create
13210  * DOF containing the run-time options -- but this could be expanded to create
13211  * complete DOF representing the enabled state.
13212  */
13213 static dof_hdr_t *
13214 dtrace_dof_create(dtrace_state_t *state)
13215 {
13216 	dof_hdr_t *dof;
13217 	dof_sec_t *sec;
13218 	dof_optdesc_t *opt;
13219 	int i, len = sizeof (dof_hdr_t) +
13220 	    roundup(sizeof (dof_sec_t), sizeof (uint64_t)) +
13221 	    sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
13222 
13223 	ASSERT(MUTEX_HELD(&dtrace_lock));
13224 
13225 	dof = kmem_zalloc(len, KM_SLEEP);
13226 	dof->dofh_ident[DOF_ID_MAG0] = DOF_MAG_MAG0;
13227 	dof->dofh_ident[DOF_ID_MAG1] = DOF_MAG_MAG1;
13228 	dof->dofh_ident[DOF_ID_MAG2] = DOF_MAG_MAG2;
13229 	dof->dofh_ident[DOF_ID_MAG3] = DOF_MAG_MAG3;
13230 
13231 	dof->dofh_ident[DOF_ID_MODEL] = DOF_MODEL_NATIVE;
13232 	dof->dofh_ident[DOF_ID_ENCODING] = DOF_ENCODE_NATIVE;
13233 	dof->dofh_ident[DOF_ID_VERSION] = DOF_VERSION;
13234 	dof->dofh_ident[DOF_ID_DIFVERS] = DIF_VERSION;
13235 	dof->dofh_ident[DOF_ID_DIFIREG] = DIF_DIR_NREGS;
13236 	dof->dofh_ident[DOF_ID_DIFTREG] = DIF_DTR_NREGS;
13237 
13238 	dof->dofh_flags = 0;
13239 	dof->dofh_hdrsize = sizeof (dof_hdr_t);
13240 	dof->dofh_secsize = sizeof (dof_sec_t);
13241 	dof->dofh_secnum = 1;	/* only DOF_SECT_OPTDESC */
13242 	dof->dofh_secoff = sizeof (dof_hdr_t);
13243 	dof->dofh_loadsz = len;
13244 	dof->dofh_filesz = len;
13245 	dof->dofh_pad = 0;
13246 
13247 	/*
13248 	 * Fill in the option section header...
13249 	 */
13250 	sec = (dof_sec_t *)((uintptr_t)dof + sizeof (dof_hdr_t));
13251 	sec->dofs_type = DOF_SECT_OPTDESC;
13252 	sec->dofs_align = sizeof (uint64_t);
13253 	sec->dofs_flags = DOF_SECF_LOAD;
13254 	sec->dofs_entsize = sizeof (dof_optdesc_t);
13255 
13256 	opt = (dof_optdesc_t *)((uintptr_t)sec +
13257 	    roundup(sizeof (dof_sec_t), sizeof (uint64_t)));
13258 
13259 	sec->dofs_offset = (uintptr_t)opt - (uintptr_t)dof;
13260 	sec->dofs_size = sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
13261 
13262 	for (i = 0; i < DTRACEOPT_MAX; i++) {
13263 		opt[i].dofo_option = i;
13264 		opt[i].dofo_strtab = DOF_SECIDX_NONE;
13265 		opt[i].dofo_value = state->dts_options[i];
13266 	}
13267 
13268 	return (dof);
13269 }
13270 
13271 static dof_hdr_t *
13272 dtrace_dof_copyin(uintptr_t uarg, int *errp)
13273 {
13274 	dof_hdr_t hdr, *dof;
13275 
13276 	ASSERT(!MUTEX_HELD(&dtrace_lock));
13277 
13278 	/*
13279 	 * First, we're going to copyin() the sizeof (dof_hdr_t).
13280 	 */
13281 	if (copyin((void *)uarg, &hdr, sizeof (hdr)) != 0) {
13282 		dtrace_dof_error(NULL, "failed to copyin DOF header");
13283 		*errp = EFAULT;
13284 		return (NULL);
13285 	}
13286 
13287 	/*
13288 	 * Now we'll allocate the entire DOF and copy it in -- provided
13289 	 * that the length isn't outrageous.
13290 	 */
13291 	if (hdr.dofh_loadsz >= dtrace_dof_maxsize) {
13292 		dtrace_dof_error(&hdr, "load size exceeds maximum");
13293 		*errp = E2BIG;
13294 		return (NULL);
13295 	}
13296 
13297 	if (hdr.dofh_loadsz < sizeof (hdr)) {
13298 		dtrace_dof_error(&hdr, "invalid load size");
13299 		*errp = EINVAL;
13300 		return (NULL);
13301 	}
13302 
13303 	dof = kmem_alloc(hdr.dofh_loadsz, KM_SLEEP);
13304 
13305 	if (copyin((void *)uarg, dof, hdr.dofh_loadsz) != 0 ||
13306 	    dof->dofh_loadsz != hdr.dofh_loadsz) {
13307 		kmem_free(dof, hdr.dofh_loadsz);
13308 		*errp = EFAULT;
13309 		return (NULL);
13310 	}
13311 
13312 	return (dof);
13313 }
13314 
13315 #ifdef __FreeBSD__
13316 static dof_hdr_t *
13317 dtrace_dof_copyin_proc(struct proc *p, uintptr_t uarg, int *errp)
13318 {
13319 	dof_hdr_t hdr, *dof;
13320 	struct thread *td;
13321 	size_t loadsz;
13322 
13323 	ASSERT(!MUTEX_HELD(&dtrace_lock));
13324 
13325 	td = curthread;
13326 
13327 	/*
13328 	 * First, we're going to copyin() the sizeof (dof_hdr_t).
13329 	 */
13330 	if (proc_readmem(td, p, uarg, &hdr, sizeof(hdr)) != sizeof(hdr)) {
13331 		dtrace_dof_error(NULL, "failed to copyin DOF header");
13332 		*errp = EFAULT;
13333 		return (NULL);
13334 	}
13335 
13336 	/*
13337 	 * Now we'll allocate the entire DOF and copy it in -- provided
13338 	 * that the length isn't outrageous.
13339 	 */
13340 	if (hdr.dofh_loadsz >= dtrace_dof_maxsize) {
13341 		dtrace_dof_error(&hdr, "load size exceeds maximum");
13342 		*errp = E2BIG;
13343 		return (NULL);
13344 	}
13345 	loadsz = (size_t)hdr.dofh_loadsz;
13346 
13347 	if (loadsz < sizeof (hdr)) {
13348 		dtrace_dof_error(&hdr, "invalid load size");
13349 		*errp = EINVAL;
13350 		return (NULL);
13351 	}
13352 
13353 	dof = kmem_alloc(loadsz, KM_SLEEP);
13354 
13355 	if (proc_readmem(td, p, uarg, dof, loadsz) != loadsz ||
13356 	    dof->dofh_loadsz != loadsz) {
13357 		kmem_free(dof, hdr.dofh_loadsz);
13358 		*errp = EFAULT;
13359 		return (NULL);
13360 	}
13361 
13362 	return (dof);
13363 }
13364 
13365 static __inline uchar_t
13366 dtrace_dof_char(char c)
13367 {
13368 
13369 	switch (c) {
13370 	case '0':
13371 	case '1':
13372 	case '2':
13373 	case '3':
13374 	case '4':
13375 	case '5':
13376 	case '6':
13377 	case '7':
13378 	case '8':
13379 	case '9':
13380 		return (c - '0');
13381 	case 'A':
13382 	case 'B':
13383 	case 'C':
13384 	case 'D':
13385 	case 'E':
13386 	case 'F':
13387 		return (c - 'A' + 10);
13388 	case 'a':
13389 	case 'b':
13390 	case 'c':
13391 	case 'd':
13392 	case 'e':
13393 	case 'f':
13394 		return (c - 'a' + 10);
13395 	}
13396 	/* Should not reach here. */
13397 	return (UCHAR_MAX);
13398 }
13399 #endif /* __FreeBSD__ */
13400 
13401 static dof_hdr_t *
13402 dtrace_dof_property(const char *name)
13403 {
13404 #ifdef __FreeBSD__
13405 	uint8_t *dofbuf;
13406 	u_char *data, *eol;
13407 	caddr_t doffile;
13408 	size_t bytes, len, i;
13409 	dof_hdr_t *dof;
13410 	u_char c1, c2;
13411 
13412 	dof = NULL;
13413 
13414 	doffile = preload_search_by_type("dtrace_dof");
13415 	if (doffile == NULL)
13416 		return (NULL);
13417 
13418 	data = preload_fetch_addr(doffile);
13419 	len = preload_fetch_size(doffile);
13420 	for (;;) {
13421 		/* Look for the end of the line. All lines end in a newline. */
13422 		eol = memchr(data, '\n', len);
13423 		if (eol == NULL)
13424 			return (NULL);
13425 
13426 		if (strncmp(name, data, strlen(name)) == 0)
13427 			break;
13428 
13429 		eol++; /* skip past the newline */
13430 		len -= eol - data;
13431 		data = eol;
13432 	}
13433 
13434 	/* We've found the data corresponding to the specified key. */
13435 
13436 	data += strlen(name) + 1; /* skip past the '=' */
13437 	len = eol - data;
13438 	if (len % 2 != 0) {
13439 		dtrace_dof_error(NULL, "invalid DOF encoding length");
13440 		goto doferr;
13441 	}
13442 	bytes = len / 2;
13443 	if (bytes < sizeof(dof_hdr_t)) {
13444 		dtrace_dof_error(NULL, "truncated header");
13445 		goto doferr;
13446 	}
13447 
13448 	/*
13449 	 * Each byte is represented by the two ASCII characters in its hex
13450 	 * representation.
13451 	 */
13452 	dofbuf = malloc(bytes, M_SOLARIS, M_WAITOK);
13453 	for (i = 0; i < bytes; i++) {
13454 		c1 = dtrace_dof_char(data[i * 2]);
13455 		c2 = dtrace_dof_char(data[i * 2 + 1]);
13456 		if (c1 == UCHAR_MAX || c2 == UCHAR_MAX) {
13457 			dtrace_dof_error(NULL, "invalid hex char in DOF");
13458 			goto doferr;
13459 		}
13460 		dofbuf[i] = c1 * 16 + c2;
13461 	}
13462 
13463 	dof = (dof_hdr_t *)dofbuf;
13464 	if (bytes < dof->dofh_loadsz) {
13465 		dtrace_dof_error(NULL, "truncated DOF");
13466 		goto doferr;
13467 	}
13468 
13469 	if (dof->dofh_loadsz >= dtrace_dof_maxsize) {
13470 		dtrace_dof_error(NULL, "oversized DOF");
13471 		goto doferr;
13472 	}
13473 
13474 	return (dof);
13475 
13476 doferr:
13477 	free(dof, M_SOLARIS);
13478 	return (NULL);
13479 #else /* __FreeBSD__ */
13480 	uchar_t *buf;
13481 	uint64_t loadsz;
13482 	unsigned int len, i;
13483 	dof_hdr_t *dof;
13484 
13485 	/*
13486 	 * Unfortunately, array of values in .conf files are always (and
13487 	 * only) interpreted to be integer arrays.  We must read our DOF
13488 	 * as an integer array, and then squeeze it into a byte array.
13489 	 */
13490 	if (ddi_prop_lookup_int_array(DDI_DEV_T_ANY, dtrace_devi, 0,
13491 	    (char *)name, (int **)&buf, &len) != DDI_PROP_SUCCESS)
13492 		return (NULL);
13493 
13494 	for (i = 0; i < len; i++)
13495 		buf[i] = (uchar_t)(((int *)buf)[i]);
13496 
13497 	if (len < sizeof (dof_hdr_t)) {
13498 		ddi_prop_free(buf);
13499 		dtrace_dof_error(NULL, "truncated header");
13500 		return (NULL);
13501 	}
13502 
13503 	if (len < (loadsz = ((dof_hdr_t *)buf)->dofh_loadsz)) {
13504 		ddi_prop_free(buf);
13505 		dtrace_dof_error(NULL, "truncated DOF");
13506 		return (NULL);
13507 	}
13508 
13509 	if (loadsz >= dtrace_dof_maxsize) {
13510 		ddi_prop_free(buf);
13511 		dtrace_dof_error(NULL, "oversized DOF");
13512 		return (NULL);
13513 	}
13514 
13515 	dof = kmem_alloc(loadsz, KM_SLEEP);
13516 	bcopy(buf, dof, loadsz);
13517 	ddi_prop_free(buf);
13518 
13519 	return (dof);
13520 #endif /* !__FreeBSD__ */
13521 }
13522 
13523 static void
13524 dtrace_dof_destroy(dof_hdr_t *dof)
13525 {
13526 	kmem_free(dof, dof->dofh_loadsz);
13527 }
13528 
13529 /*
13530  * Return the dof_sec_t pointer corresponding to a given section index.  If the
13531  * index is not valid, dtrace_dof_error() is called and NULL is returned.  If
13532  * a type other than DOF_SECT_NONE is specified, the header is checked against
13533  * this type and NULL is returned if the types do not match.
13534  */
13535 static dof_sec_t *
13536 dtrace_dof_sect(dof_hdr_t *dof, uint32_t type, dof_secidx_t i)
13537 {
13538 	dof_sec_t *sec = (dof_sec_t *)(uintptr_t)
13539 	    ((uintptr_t)dof + dof->dofh_secoff + i * dof->dofh_secsize);
13540 
13541 	if (i >= dof->dofh_secnum) {
13542 		dtrace_dof_error(dof, "referenced section index is invalid");
13543 		return (NULL);
13544 	}
13545 
13546 	if (!(sec->dofs_flags & DOF_SECF_LOAD)) {
13547 		dtrace_dof_error(dof, "referenced section is not loadable");
13548 		return (NULL);
13549 	}
13550 
13551 	if (type != DOF_SECT_NONE && type != sec->dofs_type) {
13552 		dtrace_dof_error(dof, "referenced section is the wrong type");
13553 		return (NULL);
13554 	}
13555 
13556 	return (sec);
13557 }
13558 
13559 static dtrace_probedesc_t *
13560 dtrace_dof_probedesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_probedesc_t *desc)
13561 {
13562 	dof_probedesc_t *probe;
13563 	dof_sec_t *strtab;
13564 	uintptr_t daddr = (uintptr_t)dof;
13565 	uintptr_t str;
13566 	size_t size;
13567 
13568 	if (sec->dofs_type != DOF_SECT_PROBEDESC) {
13569 		dtrace_dof_error(dof, "invalid probe section");
13570 		return (NULL);
13571 	}
13572 
13573 	if (sec->dofs_align != sizeof (dof_secidx_t)) {
13574 		dtrace_dof_error(dof, "bad alignment in probe description");
13575 		return (NULL);
13576 	}
13577 
13578 	if (sec->dofs_offset + sizeof (dof_probedesc_t) > dof->dofh_loadsz) {
13579 		dtrace_dof_error(dof, "truncated probe description");
13580 		return (NULL);
13581 	}
13582 
13583 	probe = (dof_probedesc_t *)(uintptr_t)(daddr + sec->dofs_offset);
13584 	strtab = dtrace_dof_sect(dof, DOF_SECT_STRTAB, probe->dofp_strtab);
13585 
13586 	if (strtab == NULL)
13587 		return (NULL);
13588 
13589 	str = daddr + strtab->dofs_offset;
13590 	size = strtab->dofs_size;
13591 
13592 	if (probe->dofp_provider >= strtab->dofs_size) {
13593 		dtrace_dof_error(dof, "corrupt probe provider");
13594 		return (NULL);
13595 	}
13596 
13597 	(void) strncpy(desc->dtpd_provider,
13598 	    (char *)(str + probe->dofp_provider),
13599 	    MIN(DTRACE_PROVNAMELEN - 1, size - probe->dofp_provider));
13600 
13601 	if (probe->dofp_mod >= strtab->dofs_size) {
13602 		dtrace_dof_error(dof, "corrupt probe module");
13603 		return (NULL);
13604 	}
13605 
13606 	(void) strncpy(desc->dtpd_mod, (char *)(str + probe->dofp_mod),
13607 	    MIN(DTRACE_MODNAMELEN - 1, size - probe->dofp_mod));
13608 
13609 	if (probe->dofp_func >= strtab->dofs_size) {
13610 		dtrace_dof_error(dof, "corrupt probe function");
13611 		return (NULL);
13612 	}
13613 
13614 	(void) strncpy(desc->dtpd_func, (char *)(str + probe->dofp_func),
13615 	    MIN(DTRACE_FUNCNAMELEN - 1, size - probe->dofp_func));
13616 
13617 	if (probe->dofp_name >= strtab->dofs_size) {
13618 		dtrace_dof_error(dof, "corrupt probe name");
13619 		return (NULL);
13620 	}
13621 
13622 	(void) strncpy(desc->dtpd_name, (char *)(str + probe->dofp_name),
13623 	    MIN(DTRACE_NAMELEN - 1, size - probe->dofp_name));
13624 
13625 	return (desc);
13626 }
13627 
13628 static dtrace_difo_t *
13629 dtrace_dof_difo(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
13630     cred_t *cr)
13631 {
13632 	dtrace_difo_t *dp;
13633 	size_t ttl = 0;
13634 	dof_difohdr_t *dofd;
13635 	uintptr_t daddr = (uintptr_t)dof;
13636 	size_t max = dtrace_difo_maxsize;
13637 	int i, l, n;
13638 
13639 	static const struct {
13640 		int section;
13641 		int bufoffs;
13642 		int lenoffs;
13643 		int entsize;
13644 		int align;
13645 		const char *msg;
13646 	} difo[] = {
13647 		{ DOF_SECT_DIF, offsetof(dtrace_difo_t, dtdo_buf),
13648 		offsetof(dtrace_difo_t, dtdo_len), sizeof (dif_instr_t),
13649 		sizeof (dif_instr_t), "multiple DIF sections" },
13650 
13651 		{ DOF_SECT_INTTAB, offsetof(dtrace_difo_t, dtdo_inttab),
13652 		offsetof(dtrace_difo_t, dtdo_intlen), sizeof (uint64_t),
13653 		sizeof (uint64_t), "multiple integer tables" },
13654 
13655 		{ DOF_SECT_STRTAB, offsetof(dtrace_difo_t, dtdo_strtab),
13656 		offsetof(dtrace_difo_t, dtdo_strlen), 0,
13657 		sizeof (char), "multiple string tables" },
13658 
13659 		{ DOF_SECT_VARTAB, offsetof(dtrace_difo_t, dtdo_vartab),
13660 		offsetof(dtrace_difo_t, dtdo_varlen), sizeof (dtrace_difv_t),
13661 		sizeof (uint_t), "multiple variable tables" },
13662 
13663 		{ DOF_SECT_NONE, 0, 0, 0, 0, NULL }
13664 	};
13665 
13666 	if (sec->dofs_type != DOF_SECT_DIFOHDR) {
13667 		dtrace_dof_error(dof, "invalid DIFO header section");
13668 		return (NULL);
13669 	}
13670 
13671 	if (sec->dofs_align != sizeof (dof_secidx_t)) {
13672 		dtrace_dof_error(dof, "bad alignment in DIFO header");
13673 		return (NULL);
13674 	}
13675 
13676 	if (sec->dofs_size < sizeof (dof_difohdr_t) ||
13677 	    sec->dofs_size % sizeof (dof_secidx_t)) {
13678 		dtrace_dof_error(dof, "bad size in DIFO header");
13679 		return (NULL);
13680 	}
13681 
13682 	dofd = (dof_difohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
13683 	n = (sec->dofs_size - sizeof (*dofd)) / sizeof (dof_secidx_t) + 1;
13684 
13685 	dp = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
13686 	dp->dtdo_rtype = dofd->dofd_rtype;
13687 
13688 	for (l = 0; l < n; l++) {
13689 		dof_sec_t *subsec;
13690 		void **bufp;
13691 		uint32_t *lenp;
13692 
13693 		if ((subsec = dtrace_dof_sect(dof, DOF_SECT_NONE,
13694 		    dofd->dofd_links[l])) == NULL)
13695 			goto err; /* invalid section link */
13696 
13697 		if (ttl + subsec->dofs_size > max) {
13698 			dtrace_dof_error(dof, "exceeds maximum size");
13699 			goto err;
13700 		}
13701 
13702 		ttl += subsec->dofs_size;
13703 
13704 		for (i = 0; difo[i].section != DOF_SECT_NONE; i++) {
13705 			if (subsec->dofs_type != difo[i].section)
13706 				continue;
13707 
13708 			if (!(subsec->dofs_flags & DOF_SECF_LOAD)) {
13709 				dtrace_dof_error(dof, "section not loaded");
13710 				goto err;
13711 			}
13712 
13713 			if (subsec->dofs_align != difo[i].align) {
13714 				dtrace_dof_error(dof, "bad alignment");
13715 				goto err;
13716 			}
13717 
13718 			bufp = (void **)((uintptr_t)dp + difo[i].bufoffs);
13719 			lenp = (uint32_t *)((uintptr_t)dp + difo[i].lenoffs);
13720 
13721 			if (*bufp != NULL) {
13722 				dtrace_dof_error(dof, difo[i].msg);
13723 				goto err;
13724 			}
13725 
13726 			if (difo[i].entsize != subsec->dofs_entsize) {
13727 				dtrace_dof_error(dof, "entry size mismatch");
13728 				goto err;
13729 			}
13730 
13731 			if (subsec->dofs_entsize != 0 &&
13732 			    (subsec->dofs_size % subsec->dofs_entsize) != 0) {
13733 				dtrace_dof_error(dof, "corrupt entry size");
13734 				goto err;
13735 			}
13736 
13737 			*lenp = subsec->dofs_size;
13738 			*bufp = kmem_alloc(subsec->dofs_size, KM_SLEEP);
13739 			bcopy((char *)(uintptr_t)(daddr + subsec->dofs_offset),
13740 			    *bufp, subsec->dofs_size);
13741 
13742 			if (subsec->dofs_entsize != 0)
13743 				*lenp /= subsec->dofs_entsize;
13744 
13745 			break;
13746 		}
13747 
13748 		/*
13749 		 * If we encounter a loadable DIFO sub-section that is not
13750 		 * known to us, assume this is a broken program and fail.
13751 		 */
13752 		if (difo[i].section == DOF_SECT_NONE &&
13753 		    (subsec->dofs_flags & DOF_SECF_LOAD)) {
13754 			dtrace_dof_error(dof, "unrecognized DIFO subsection");
13755 			goto err;
13756 		}
13757 	}
13758 
13759 	if (dp->dtdo_buf == NULL) {
13760 		/*
13761 		 * We can't have a DIF object without DIF text.
13762 		 */
13763 		dtrace_dof_error(dof, "missing DIF text");
13764 		goto err;
13765 	}
13766 
13767 	/*
13768 	 * Before we validate the DIF object, run through the variable table
13769 	 * looking for the strings -- if any of their size are under, we'll set
13770 	 * their size to be the system-wide default string size.  Note that
13771 	 * this should _not_ happen if the "strsize" option has been set --
13772 	 * in this case, the compiler should have set the size to reflect the
13773 	 * setting of the option.
13774 	 */
13775 	for (i = 0; i < dp->dtdo_varlen; i++) {
13776 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
13777 		dtrace_diftype_t *t = &v->dtdv_type;
13778 
13779 		if (v->dtdv_id < DIF_VAR_OTHER_UBASE)
13780 			continue;
13781 
13782 		if (t->dtdt_kind == DIF_TYPE_STRING && t->dtdt_size == 0)
13783 			t->dtdt_size = dtrace_strsize_default;
13784 	}
13785 
13786 	if (dtrace_difo_validate(dp, vstate, DIF_DIR_NREGS, cr) != 0)
13787 		goto err;
13788 
13789 	dtrace_difo_init(dp, vstate);
13790 	return (dp);
13791 
13792 err:
13793 	kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
13794 	kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
13795 	kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
13796 	kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
13797 
13798 	kmem_free(dp, sizeof (dtrace_difo_t));
13799 	return (NULL);
13800 }
13801 
13802 static dtrace_predicate_t *
13803 dtrace_dof_predicate(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
13804     cred_t *cr)
13805 {
13806 	dtrace_difo_t *dp;
13807 
13808 	if ((dp = dtrace_dof_difo(dof, sec, vstate, cr)) == NULL)
13809 		return (NULL);
13810 
13811 	return (dtrace_predicate_create(dp));
13812 }
13813 
13814 static dtrace_actdesc_t *
13815 dtrace_dof_actdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
13816     cred_t *cr)
13817 {
13818 	dtrace_actdesc_t *act, *first = NULL, *last = NULL, *next;
13819 	dof_actdesc_t *desc;
13820 	dof_sec_t *difosec;
13821 	size_t offs;
13822 	uintptr_t daddr = (uintptr_t)dof;
13823 	uint64_t arg;
13824 	dtrace_actkind_t kind;
13825 
13826 	if (sec->dofs_type != DOF_SECT_ACTDESC) {
13827 		dtrace_dof_error(dof, "invalid action section");
13828 		return (NULL);
13829 	}
13830 
13831 	if (sec->dofs_offset + sizeof (dof_actdesc_t) > dof->dofh_loadsz) {
13832 		dtrace_dof_error(dof, "truncated action description");
13833 		return (NULL);
13834 	}
13835 
13836 	if (sec->dofs_align != sizeof (uint64_t)) {
13837 		dtrace_dof_error(dof, "bad alignment in action description");
13838 		return (NULL);
13839 	}
13840 
13841 	if (sec->dofs_size < sec->dofs_entsize) {
13842 		dtrace_dof_error(dof, "section entry size exceeds total size");
13843 		return (NULL);
13844 	}
13845 
13846 	if (sec->dofs_entsize != sizeof (dof_actdesc_t)) {
13847 		dtrace_dof_error(dof, "bad entry size in action description");
13848 		return (NULL);
13849 	}
13850 
13851 	if (sec->dofs_size / sec->dofs_entsize > dtrace_actions_max) {
13852 		dtrace_dof_error(dof, "actions exceed dtrace_actions_max");
13853 		return (NULL);
13854 	}
13855 
13856 	for (offs = 0; offs < sec->dofs_size; offs += sec->dofs_entsize) {
13857 		desc = (dof_actdesc_t *)(daddr +
13858 		    (uintptr_t)sec->dofs_offset + offs);
13859 		kind = (dtrace_actkind_t)desc->dofa_kind;
13860 
13861 		if ((DTRACEACT_ISPRINTFLIKE(kind) &&
13862 		    (kind != DTRACEACT_PRINTA ||
13863 		    desc->dofa_strtab != DOF_SECIDX_NONE)) ||
13864 		    (kind == DTRACEACT_DIFEXPR &&
13865 		    desc->dofa_strtab != DOF_SECIDX_NONE)) {
13866 			dof_sec_t *strtab;
13867 			char *str, *fmt;
13868 			uint64_t i;
13869 
13870 			/*
13871 			 * The argument to these actions is an index into the
13872 			 * DOF string table.  For printf()-like actions, this
13873 			 * is the format string.  For print(), this is the
13874 			 * CTF type of the expression result.
13875 			 */
13876 			if ((strtab = dtrace_dof_sect(dof,
13877 			    DOF_SECT_STRTAB, desc->dofa_strtab)) == NULL)
13878 				goto err;
13879 
13880 			str = (char *)((uintptr_t)dof +
13881 			    (uintptr_t)strtab->dofs_offset);
13882 
13883 			for (i = desc->dofa_arg; i < strtab->dofs_size; i++) {
13884 				if (str[i] == '\0')
13885 					break;
13886 			}
13887 
13888 			if (i >= strtab->dofs_size) {
13889 				dtrace_dof_error(dof, "bogus format string");
13890 				goto err;
13891 			}
13892 
13893 			if (i == desc->dofa_arg) {
13894 				dtrace_dof_error(dof, "empty format string");
13895 				goto err;
13896 			}
13897 
13898 			i -= desc->dofa_arg;
13899 			fmt = kmem_alloc(i + 1, KM_SLEEP);
13900 			bcopy(&str[desc->dofa_arg], fmt, i + 1);
13901 			arg = (uint64_t)(uintptr_t)fmt;
13902 		} else {
13903 			if (kind == DTRACEACT_PRINTA) {
13904 				ASSERT(desc->dofa_strtab == DOF_SECIDX_NONE);
13905 				arg = 0;
13906 			} else {
13907 				arg = desc->dofa_arg;
13908 			}
13909 		}
13910 
13911 		act = dtrace_actdesc_create(kind, desc->dofa_ntuple,
13912 		    desc->dofa_uarg, arg);
13913 
13914 		if (last != NULL) {
13915 			last->dtad_next = act;
13916 		} else {
13917 			first = act;
13918 		}
13919 
13920 		last = act;
13921 
13922 		if (desc->dofa_difo == DOF_SECIDX_NONE)
13923 			continue;
13924 
13925 		if ((difosec = dtrace_dof_sect(dof,
13926 		    DOF_SECT_DIFOHDR, desc->dofa_difo)) == NULL)
13927 			goto err;
13928 
13929 		act->dtad_difo = dtrace_dof_difo(dof, difosec, vstate, cr);
13930 
13931 		if (act->dtad_difo == NULL)
13932 			goto err;
13933 	}
13934 
13935 	ASSERT(first != NULL);
13936 	return (first);
13937 
13938 err:
13939 	for (act = first; act != NULL; act = next) {
13940 		next = act->dtad_next;
13941 		dtrace_actdesc_release(act, vstate);
13942 	}
13943 
13944 	return (NULL);
13945 }
13946 
13947 static dtrace_ecbdesc_t *
13948 dtrace_dof_ecbdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
13949     cred_t *cr)
13950 {
13951 	dtrace_ecbdesc_t *ep;
13952 	dof_ecbdesc_t *ecb;
13953 	dtrace_probedesc_t *desc;
13954 	dtrace_predicate_t *pred = NULL;
13955 
13956 	if (sec->dofs_size < sizeof (dof_ecbdesc_t)) {
13957 		dtrace_dof_error(dof, "truncated ECB description");
13958 		return (NULL);
13959 	}
13960 
13961 	if (sec->dofs_align != sizeof (uint64_t)) {
13962 		dtrace_dof_error(dof, "bad alignment in ECB description");
13963 		return (NULL);
13964 	}
13965 
13966 	ecb = (dof_ecbdesc_t *)((uintptr_t)dof + (uintptr_t)sec->dofs_offset);
13967 	sec = dtrace_dof_sect(dof, DOF_SECT_PROBEDESC, ecb->dofe_probes);
13968 
13969 	if (sec == NULL)
13970 		return (NULL);
13971 
13972 	ep = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
13973 	ep->dted_uarg = ecb->dofe_uarg;
13974 	desc = &ep->dted_probe;
13975 
13976 	if (dtrace_dof_probedesc(dof, sec, desc) == NULL)
13977 		goto err;
13978 
13979 	if (ecb->dofe_pred != DOF_SECIDX_NONE) {
13980 		if ((sec = dtrace_dof_sect(dof,
13981 		    DOF_SECT_DIFOHDR, ecb->dofe_pred)) == NULL)
13982 			goto err;
13983 
13984 		if ((pred = dtrace_dof_predicate(dof, sec, vstate, cr)) == NULL)
13985 			goto err;
13986 
13987 		ep->dted_pred.dtpdd_predicate = pred;
13988 	}
13989 
13990 	if (ecb->dofe_actions != DOF_SECIDX_NONE) {
13991 		if ((sec = dtrace_dof_sect(dof,
13992 		    DOF_SECT_ACTDESC, ecb->dofe_actions)) == NULL)
13993 			goto err;
13994 
13995 		ep->dted_action = dtrace_dof_actdesc(dof, sec, vstate, cr);
13996 
13997 		if (ep->dted_action == NULL)
13998 			goto err;
13999 	}
14000 
14001 	return (ep);
14002 
14003 err:
14004 	if (pred != NULL)
14005 		dtrace_predicate_release(pred, vstate);
14006 	kmem_free(ep, sizeof (dtrace_ecbdesc_t));
14007 	return (NULL);
14008 }
14009 
14010 /*
14011  * Apply the relocations from the specified 'sec' (a DOF_SECT_URELHDR) to the
14012  * specified DOF.  SETX relocations are computed using 'ubase', the base load
14013  * address of the object containing the DOF, and DOFREL relocations are relative
14014  * to the relocation offset within the DOF.
14015  */
14016 static int
14017 dtrace_dof_relocate(dof_hdr_t *dof, dof_sec_t *sec, uint64_t ubase,
14018     uint64_t udaddr)
14019 {
14020 	uintptr_t daddr = (uintptr_t)dof;
14021 	uintptr_t ts_end;
14022 	dof_relohdr_t *dofr =
14023 	    (dof_relohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
14024 	dof_sec_t *ss, *rs, *ts;
14025 	dof_relodesc_t *r;
14026 	uint_t i, n;
14027 
14028 	if (sec->dofs_size < sizeof (dof_relohdr_t) ||
14029 	    sec->dofs_align != sizeof (dof_secidx_t)) {
14030 		dtrace_dof_error(dof, "invalid relocation header");
14031 		return (-1);
14032 	}
14033 
14034 	ss = dtrace_dof_sect(dof, DOF_SECT_STRTAB, dofr->dofr_strtab);
14035 	rs = dtrace_dof_sect(dof, DOF_SECT_RELTAB, dofr->dofr_relsec);
14036 	ts = dtrace_dof_sect(dof, DOF_SECT_NONE, dofr->dofr_tgtsec);
14037 	ts_end = (uintptr_t)ts + sizeof (dof_sec_t);
14038 
14039 	if (ss == NULL || rs == NULL || ts == NULL)
14040 		return (-1); /* dtrace_dof_error() has been called already */
14041 
14042 	if (rs->dofs_entsize < sizeof (dof_relodesc_t) ||
14043 	    rs->dofs_align != sizeof (uint64_t)) {
14044 		dtrace_dof_error(dof, "invalid relocation section");
14045 		return (-1);
14046 	}
14047 
14048 	r = (dof_relodesc_t *)(uintptr_t)(daddr + rs->dofs_offset);
14049 	n = rs->dofs_size / rs->dofs_entsize;
14050 
14051 	for (i = 0; i < n; i++) {
14052 		uintptr_t taddr = daddr + ts->dofs_offset + r->dofr_offset;
14053 
14054 		switch (r->dofr_type) {
14055 		case DOF_RELO_NONE:
14056 			break;
14057 		case DOF_RELO_SETX:
14058 		case DOF_RELO_DOFREL:
14059 			if (r->dofr_offset >= ts->dofs_size || r->dofr_offset +
14060 			    sizeof (uint64_t) > ts->dofs_size) {
14061 				dtrace_dof_error(dof, "bad relocation offset");
14062 				return (-1);
14063 			}
14064 
14065 			if (taddr >= (uintptr_t)ts && taddr < ts_end) {
14066 				dtrace_dof_error(dof, "bad relocation offset");
14067 				return (-1);
14068 			}
14069 
14070 			if (!IS_P2ALIGNED(taddr, sizeof (uint64_t))) {
14071 				dtrace_dof_error(dof, "misaligned setx relo");
14072 				return (-1);
14073 			}
14074 
14075 			if (r->dofr_type == DOF_RELO_SETX)
14076 				*(uint64_t *)taddr += ubase;
14077 			else
14078 				*(uint64_t *)taddr +=
14079 				    udaddr + ts->dofs_offset + r->dofr_offset;
14080 			break;
14081 		default:
14082 			dtrace_dof_error(dof, "invalid relocation type");
14083 			return (-1);
14084 		}
14085 
14086 		r = (dof_relodesc_t *)((uintptr_t)r + rs->dofs_entsize);
14087 	}
14088 
14089 	return (0);
14090 }
14091 
14092 /*
14093  * The dof_hdr_t passed to dtrace_dof_slurp() should be a partially validated
14094  * header:  it should be at the front of a memory region that is at least
14095  * sizeof (dof_hdr_t) in size -- and then at least dof_hdr.dofh_loadsz in
14096  * size.  It need not be validated in any other way.
14097  */
14098 static int
14099 dtrace_dof_slurp(dof_hdr_t *dof, dtrace_vstate_t *vstate, cred_t *cr,
14100     dtrace_enabling_t **enabp, uint64_t ubase, uint64_t udaddr, int noprobes)
14101 {
14102 	uint64_t len = dof->dofh_loadsz, seclen;
14103 	uintptr_t daddr = (uintptr_t)dof;
14104 	dtrace_ecbdesc_t *ep;
14105 	dtrace_enabling_t *enab;
14106 	uint_t i;
14107 
14108 	ASSERT(MUTEX_HELD(&dtrace_lock));
14109 	ASSERT(dof->dofh_loadsz >= sizeof (dof_hdr_t));
14110 
14111 	/*
14112 	 * Check the DOF header identification bytes.  In addition to checking
14113 	 * valid settings, we also verify that unused bits/bytes are zeroed so
14114 	 * we can use them later without fear of regressing existing binaries.
14115 	 */
14116 	if (bcmp(&dof->dofh_ident[DOF_ID_MAG0],
14117 	    DOF_MAG_STRING, DOF_MAG_STRLEN) != 0) {
14118 		dtrace_dof_error(dof, "DOF magic string mismatch");
14119 		return (-1);
14120 	}
14121 
14122 	if (dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_ILP32 &&
14123 	    dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_LP64) {
14124 		dtrace_dof_error(dof, "DOF has invalid data model");
14125 		return (-1);
14126 	}
14127 
14128 	if (dof->dofh_ident[DOF_ID_ENCODING] != DOF_ENCODE_NATIVE) {
14129 		dtrace_dof_error(dof, "DOF encoding mismatch");
14130 		return (-1);
14131 	}
14132 
14133 	if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
14134 	    dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_2) {
14135 		dtrace_dof_error(dof, "DOF version mismatch");
14136 		return (-1);
14137 	}
14138 
14139 	if (dof->dofh_ident[DOF_ID_DIFVERS] != DIF_VERSION_2) {
14140 		dtrace_dof_error(dof, "DOF uses unsupported instruction set");
14141 		return (-1);
14142 	}
14143 
14144 	if (dof->dofh_ident[DOF_ID_DIFIREG] > DIF_DIR_NREGS) {
14145 		dtrace_dof_error(dof, "DOF uses too many integer registers");
14146 		return (-1);
14147 	}
14148 
14149 	if (dof->dofh_ident[DOF_ID_DIFTREG] > DIF_DTR_NREGS) {
14150 		dtrace_dof_error(dof, "DOF uses too many tuple registers");
14151 		return (-1);
14152 	}
14153 
14154 	for (i = DOF_ID_PAD; i < DOF_ID_SIZE; i++) {
14155 		if (dof->dofh_ident[i] != 0) {
14156 			dtrace_dof_error(dof, "DOF has invalid ident byte set");
14157 			return (-1);
14158 		}
14159 	}
14160 
14161 	if (dof->dofh_flags & ~DOF_FL_VALID) {
14162 		dtrace_dof_error(dof, "DOF has invalid flag bits set");
14163 		return (-1);
14164 	}
14165 
14166 	if (dof->dofh_secsize == 0) {
14167 		dtrace_dof_error(dof, "zero section header size");
14168 		return (-1);
14169 	}
14170 
14171 	/*
14172 	 * Check that the section headers don't exceed the amount of DOF
14173 	 * data.  Note that we cast the section size and number of sections
14174 	 * to uint64_t's to prevent possible overflow in the multiplication.
14175 	 */
14176 	seclen = (uint64_t)dof->dofh_secnum * (uint64_t)dof->dofh_secsize;
14177 
14178 	if (dof->dofh_secoff > len || seclen > len ||
14179 	    dof->dofh_secoff + seclen > len) {
14180 		dtrace_dof_error(dof, "truncated section headers");
14181 		return (-1);
14182 	}
14183 
14184 	if (!IS_P2ALIGNED(dof->dofh_secoff, sizeof (uint64_t))) {
14185 		dtrace_dof_error(dof, "misaligned section headers");
14186 		return (-1);
14187 	}
14188 
14189 	if (!IS_P2ALIGNED(dof->dofh_secsize, sizeof (uint64_t))) {
14190 		dtrace_dof_error(dof, "misaligned section size");
14191 		return (-1);
14192 	}
14193 
14194 	/*
14195 	 * Take an initial pass through the section headers to be sure that
14196 	 * the headers don't have stray offsets.  If the 'noprobes' flag is
14197 	 * set, do not permit sections relating to providers, probes, or args.
14198 	 */
14199 	for (i = 0; i < dof->dofh_secnum; i++) {
14200 		dof_sec_t *sec = (dof_sec_t *)(daddr +
14201 		    (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
14202 
14203 		if (noprobes) {
14204 			switch (sec->dofs_type) {
14205 			case DOF_SECT_PROVIDER:
14206 			case DOF_SECT_PROBES:
14207 			case DOF_SECT_PRARGS:
14208 			case DOF_SECT_PROFFS:
14209 				dtrace_dof_error(dof, "illegal sections "
14210 				    "for enabling");
14211 				return (-1);
14212 			}
14213 		}
14214 
14215 		if (DOF_SEC_ISLOADABLE(sec->dofs_type) &&
14216 		    !(sec->dofs_flags & DOF_SECF_LOAD)) {
14217 			dtrace_dof_error(dof, "loadable section with load "
14218 			    "flag unset");
14219 			return (-1);
14220 		}
14221 
14222 		if (!(sec->dofs_flags & DOF_SECF_LOAD))
14223 			continue; /* just ignore non-loadable sections */
14224 
14225 		if (!ISP2(sec->dofs_align)) {
14226 			dtrace_dof_error(dof, "bad section alignment");
14227 			return (-1);
14228 		}
14229 
14230 		if (sec->dofs_offset & (sec->dofs_align - 1)) {
14231 			dtrace_dof_error(dof, "misaligned section");
14232 			return (-1);
14233 		}
14234 
14235 		if (sec->dofs_offset > len || sec->dofs_size > len ||
14236 		    sec->dofs_offset + sec->dofs_size > len) {
14237 			dtrace_dof_error(dof, "corrupt section header");
14238 			return (-1);
14239 		}
14240 
14241 		if (sec->dofs_type == DOF_SECT_STRTAB && *((char *)daddr +
14242 		    sec->dofs_offset + sec->dofs_size - 1) != '\0') {
14243 			dtrace_dof_error(dof, "non-terminating string table");
14244 			return (-1);
14245 		}
14246 	}
14247 
14248 	/*
14249 	 * Take a second pass through the sections and locate and perform any
14250 	 * relocations that are present.  We do this after the first pass to
14251 	 * be sure that all sections have had their headers validated.
14252 	 */
14253 	for (i = 0; i < dof->dofh_secnum; i++) {
14254 		dof_sec_t *sec = (dof_sec_t *)(daddr +
14255 		    (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
14256 
14257 		if (!(sec->dofs_flags & DOF_SECF_LOAD))
14258 			continue; /* skip sections that are not loadable */
14259 
14260 		switch (sec->dofs_type) {
14261 		case DOF_SECT_URELHDR:
14262 			if (dtrace_dof_relocate(dof, sec, ubase, udaddr) != 0)
14263 				return (-1);
14264 			break;
14265 		}
14266 	}
14267 
14268 	if ((enab = *enabp) == NULL)
14269 		enab = *enabp = dtrace_enabling_create(vstate);
14270 
14271 	for (i = 0; i < dof->dofh_secnum; i++) {
14272 		dof_sec_t *sec = (dof_sec_t *)(daddr +
14273 		    (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
14274 
14275 		if (sec->dofs_type != DOF_SECT_ECBDESC)
14276 			continue;
14277 
14278 		if ((ep = dtrace_dof_ecbdesc(dof, sec, vstate, cr)) == NULL) {
14279 			dtrace_enabling_destroy(enab);
14280 			*enabp = NULL;
14281 			return (-1);
14282 		}
14283 
14284 		dtrace_enabling_add(enab, ep);
14285 	}
14286 
14287 	return (0);
14288 }
14289 
14290 /*
14291  * Process DOF for any options.  This routine assumes that the DOF has been
14292  * at least processed by dtrace_dof_slurp().
14293  */
14294 static int
14295 dtrace_dof_options(dof_hdr_t *dof, dtrace_state_t *state)
14296 {
14297 	int i, rval;
14298 	uint32_t entsize;
14299 	size_t offs;
14300 	dof_optdesc_t *desc;
14301 
14302 	for (i = 0; i < dof->dofh_secnum; i++) {
14303 		dof_sec_t *sec = (dof_sec_t *)((uintptr_t)dof +
14304 		    (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
14305 
14306 		if (sec->dofs_type != DOF_SECT_OPTDESC)
14307 			continue;
14308 
14309 		if (sec->dofs_align != sizeof (uint64_t)) {
14310 			dtrace_dof_error(dof, "bad alignment in "
14311 			    "option description");
14312 			return (EINVAL);
14313 		}
14314 
14315 		if ((entsize = sec->dofs_entsize) == 0) {
14316 			dtrace_dof_error(dof, "zeroed option entry size");
14317 			return (EINVAL);
14318 		}
14319 
14320 		if (entsize < sizeof (dof_optdesc_t)) {
14321 			dtrace_dof_error(dof, "bad option entry size");
14322 			return (EINVAL);
14323 		}
14324 
14325 		for (offs = 0; offs < sec->dofs_size; offs += entsize) {
14326 			desc = (dof_optdesc_t *)((uintptr_t)dof +
14327 			    (uintptr_t)sec->dofs_offset + offs);
14328 
14329 			if (desc->dofo_strtab != DOF_SECIDX_NONE) {
14330 				dtrace_dof_error(dof, "non-zero option string");
14331 				return (EINVAL);
14332 			}
14333 
14334 			if (desc->dofo_value == DTRACEOPT_UNSET) {
14335 				dtrace_dof_error(dof, "unset option");
14336 				return (EINVAL);
14337 			}
14338 
14339 			if ((rval = dtrace_state_option(state,
14340 			    desc->dofo_option, desc->dofo_value)) != 0) {
14341 				dtrace_dof_error(dof, "rejected option");
14342 				return (rval);
14343 			}
14344 		}
14345 	}
14346 
14347 	return (0);
14348 }
14349 
14350 /*
14351  * DTrace Consumer State Functions
14352  */
14353 static int
14354 dtrace_dstate_init(dtrace_dstate_t *dstate, size_t size)
14355 {
14356 	size_t hashsize, maxper, min, chunksize = dstate->dtds_chunksize;
14357 	void *base;
14358 	uintptr_t limit;
14359 	dtrace_dynvar_t *dvar, *next, *start;
14360 	int i;
14361 
14362 	ASSERT(MUTEX_HELD(&dtrace_lock));
14363 	ASSERT(dstate->dtds_base == NULL && dstate->dtds_percpu == NULL);
14364 
14365 	bzero(dstate, sizeof (dtrace_dstate_t));
14366 
14367 	if ((dstate->dtds_chunksize = chunksize) == 0)
14368 		dstate->dtds_chunksize = DTRACE_DYNVAR_CHUNKSIZE;
14369 
14370 	VERIFY(dstate->dtds_chunksize < LONG_MAX);
14371 
14372 	if (size < (min = dstate->dtds_chunksize + sizeof (dtrace_dynhash_t)))
14373 		size = min;
14374 
14375 	if ((base = kmem_zalloc(size, KM_NOSLEEP | KM_NORMALPRI)) == NULL)
14376 		return (ENOMEM);
14377 
14378 	dstate->dtds_size = size;
14379 	dstate->dtds_base = base;
14380 	dstate->dtds_percpu = kmem_cache_alloc(dtrace_state_cache, KM_SLEEP);
14381 	bzero(dstate->dtds_percpu, NCPU * sizeof (dtrace_dstate_percpu_t));
14382 
14383 	hashsize = size / (dstate->dtds_chunksize + sizeof (dtrace_dynhash_t));
14384 
14385 	if (hashsize != 1 && (hashsize & 1))
14386 		hashsize--;
14387 
14388 	dstate->dtds_hashsize = hashsize;
14389 	dstate->dtds_hash = dstate->dtds_base;
14390 
14391 	/*
14392 	 * Set all of our hash buckets to point to the single sink, and (if
14393 	 * it hasn't already been set), set the sink's hash value to be the
14394 	 * sink sentinel value.  The sink is needed for dynamic variable
14395 	 * lookups to know that they have iterated over an entire, valid hash
14396 	 * chain.
14397 	 */
14398 	for (i = 0; i < hashsize; i++)
14399 		dstate->dtds_hash[i].dtdh_chain = &dtrace_dynhash_sink;
14400 
14401 	if (dtrace_dynhash_sink.dtdv_hashval != DTRACE_DYNHASH_SINK)
14402 		dtrace_dynhash_sink.dtdv_hashval = DTRACE_DYNHASH_SINK;
14403 
14404 	/*
14405 	 * Determine number of active CPUs.  Divide free list evenly among
14406 	 * active CPUs.
14407 	 */
14408 	start = (dtrace_dynvar_t *)
14409 	    ((uintptr_t)base + hashsize * sizeof (dtrace_dynhash_t));
14410 	limit = (uintptr_t)base + size;
14411 
14412 	VERIFY((uintptr_t)start < limit);
14413 	VERIFY((uintptr_t)start >= (uintptr_t)base);
14414 
14415 	maxper = (limit - (uintptr_t)start) / NCPU;
14416 	maxper = (maxper / dstate->dtds_chunksize) * dstate->dtds_chunksize;
14417 
14418 #ifndef illumos
14419 	CPU_FOREACH(i) {
14420 #else
14421 	for (i = 0; i < NCPU; i++) {
14422 #endif
14423 		dstate->dtds_percpu[i].dtdsc_free = dvar = start;
14424 
14425 		/*
14426 		 * If we don't even have enough chunks to make it once through
14427 		 * NCPUs, we're just going to allocate everything to the first
14428 		 * CPU.  And if we're on the last CPU, we're going to allocate
14429 		 * whatever is left over.  In either case, we set the limit to
14430 		 * be the limit of the dynamic variable space.
14431 		 */
14432 		if (maxper == 0 || i == NCPU - 1) {
14433 			limit = (uintptr_t)base + size;
14434 			start = NULL;
14435 		} else {
14436 			limit = (uintptr_t)start + maxper;
14437 			start = (dtrace_dynvar_t *)limit;
14438 		}
14439 
14440 		VERIFY(limit <= (uintptr_t)base + size);
14441 
14442 		for (;;) {
14443 			next = (dtrace_dynvar_t *)((uintptr_t)dvar +
14444 			    dstate->dtds_chunksize);
14445 
14446 			if ((uintptr_t)next + dstate->dtds_chunksize >= limit)
14447 				break;
14448 
14449 			VERIFY((uintptr_t)dvar >= (uintptr_t)base &&
14450 			    (uintptr_t)dvar <= (uintptr_t)base + size);
14451 			dvar->dtdv_next = next;
14452 			dvar = next;
14453 		}
14454 
14455 		if (maxper == 0)
14456 			break;
14457 	}
14458 
14459 	return (0);
14460 }
14461 
14462 static void
14463 dtrace_dstate_fini(dtrace_dstate_t *dstate)
14464 {
14465 	ASSERT(MUTEX_HELD(&cpu_lock));
14466 
14467 	if (dstate->dtds_base == NULL)
14468 		return;
14469 
14470 	kmem_free(dstate->dtds_base, dstate->dtds_size);
14471 	kmem_cache_free(dtrace_state_cache, dstate->dtds_percpu);
14472 }
14473 
14474 static void
14475 dtrace_vstate_fini(dtrace_vstate_t *vstate)
14476 {
14477 	/*
14478 	 * Logical XOR, where are you?
14479 	 */
14480 	ASSERT((vstate->dtvs_nglobals == 0) ^ (vstate->dtvs_globals != NULL));
14481 
14482 	if (vstate->dtvs_nglobals > 0) {
14483 		kmem_free(vstate->dtvs_globals, vstate->dtvs_nglobals *
14484 		    sizeof (dtrace_statvar_t *));
14485 	}
14486 
14487 	if (vstate->dtvs_ntlocals > 0) {
14488 		kmem_free(vstate->dtvs_tlocals, vstate->dtvs_ntlocals *
14489 		    sizeof (dtrace_difv_t));
14490 	}
14491 
14492 	ASSERT((vstate->dtvs_nlocals == 0) ^ (vstate->dtvs_locals != NULL));
14493 
14494 	if (vstate->dtvs_nlocals > 0) {
14495 		kmem_free(vstate->dtvs_locals, vstate->dtvs_nlocals *
14496 		    sizeof (dtrace_statvar_t *));
14497 	}
14498 }
14499 
14500 #ifdef illumos
14501 static void
14502 dtrace_state_clean(dtrace_state_t *state)
14503 {
14504 	if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE)
14505 		return;
14506 
14507 	dtrace_dynvar_clean(&state->dts_vstate.dtvs_dynvars);
14508 	dtrace_speculation_clean(state);
14509 }
14510 
14511 static void
14512 dtrace_state_deadman(dtrace_state_t *state)
14513 {
14514 	hrtime_t now;
14515 
14516 	dtrace_sync();
14517 
14518 	now = dtrace_gethrtime();
14519 
14520 	if (state != dtrace_anon.dta_state &&
14521 	    now - state->dts_laststatus >= dtrace_deadman_user)
14522 		return;
14523 
14524 	/*
14525 	 * We must be sure that dts_alive never appears to be less than the
14526 	 * value upon entry to dtrace_state_deadman(), and because we lack a
14527 	 * dtrace_cas64(), we cannot store to it atomically.  We thus instead
14528 	 * store INT64_MAX to it, followed by a memory barrier, followed by
14529 	 * the new value.  This assures that dts_alive never appears to be
14530 	 * less than its true value, regardless of the order in which the
14531 	 * stores to the underlying storage are issued.
14532 	 */
14533 	state->dts_alive = INT64_MAX;
14534 	dtrace_membar_producer();
14535 	state->dts_alive = now;
14536 }
14537 #else	/* !illumos */
14538 static void
14539 dtrace_state_clean(void *arg)
14540 {
14541 	dtrace_state_t *state = arg;
14542 	dtrace_optval_t *opt = state->dts_options;
14543 
14544 	if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE)
14545 		return;
14546 
14547 	dtrace_dynvar_clean(&state->dts_vstate.dtvs_dynvars);
14548 	dtrace_speculation_clean(state);
14549 
14550 	callout_reset(&state->dts_cleaner, hz * opt[DTRACEOPT_CLEANRATE] / NANOSEC,
14551 	    dtrace_state_clean, state);
14552 }
14553 
14554 static void
14555 dtrace_state_deadman(void *arg)
14556 {
14557 	dtrace_state_t *state = arg;
14558 	hrtime_t now;
14559 
14560 	dtrace_sync();
14561 
14562 	dtrace_debug_output();
14563 
14564 	now = dtrace_gethrtime();
14565 
14566 	if (state != dtrace_anon.dta_state &&
14567 	    now - state->dts_laststatus >= dtrace_deadman_user)
14568 		return;
14569 
14570 	/*
14571 	 * We must be sure that dts_alive never appears to be less than the
14572 	 * value upon entry to dtrace_state_deadman(), and because we lack a
14573 	 * dtrace_cas64(), we cannot store to it atomically.  We thus instead
14574 	 * store INT64_MAX to it, followed by a memory barrier, followed by
14575 	 * the new value.  This assures that dts_alive never appears to be
14576 	 * less than its true value, regardless of the order in which the
14577 	 * stores to the underlying storage are issued.
14578 	 */
14579 	state->dts_alive = INT64_MAX;
14580 	dtrace_membar_producer();
14581 	state->dts_alive = now;
14582 
14583 	callout_reset(&state->dts_deadman, hz * dtrace_deadman_interval / NANOSEC,
14584 	    dtrace_state_deadman, state);
14585 }
14586 #endif	/* illumos */
14587 
14588 static dtrace_state_t *
14589 #ifdef illumos
14590 dtrace_state_create(dev_t *devp, cred_t *cr)
14591 #else
14592 dtrace_state_create(struct cdev *dev, struct ucred *cred __unused)
14593 #endif
14594 {
14595 #ifdef illumos
14596 	minor_t minor;
14597 	major_t major;
14598 #else
14599 	cred_t *cr = NULL;
14600 	int m = 0;
14601 #endif
14602 	char c[30];
14603 	dtrace_state_t *state;
14604 	dtrace_optval_t *opt;
14605 	int bufsize = NCPU * sizeof (dtrace_buffer_t), i;
14606 	int cpu_it;
14607 
14608 	ASSERT(MUTEX_HELD(&dtrace_lock));
14609 	ASSERT(MUTEX_HELD(&cpu_lock));
14610 
14611 #ifdef illumos
14612 	minor = (minor_t)(uintptr_t)vmem_alloc(dtrace_minor, 1,
14613 	    VM_BESTFIT | VM_SLEEP);
14614 
14615 	if (ddi_soft_state_zalloc(dtrace_softstate, minor) != DDI_SUCCESS) {
14616 		vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
14617 		return (NULL);
14618 	}
14619 
14620 	state = ddi_get_soft_state(dtrace_softstate, minor);
14621 #else
14622 	if (dev != NULL) {
14623 		cr = dev->si_cred;
14624 		m = dev2unit(dev);
14625 	}
14626 
14627 	/* Allocate memory for the state. */
14628 	state = kmem_zalloc(sizeof(dtrace_state_t), KM_SLEEP);
14629 #endif
14630 
14631 	state->dts_epid = DTRACE_EPIDNONE + 1;
14632 
14633 	(void) snprintf(c, sizeof (c), "dtrace_aggid_%d", m);
14634 #ifdef illumos
14635 	state->dts_aggid_arena = vmem_create(c, (void *)1, UINT32_MAX, 1,
14636 	    NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
14637 
14638 	if (devp != NULL) {
14639 		major = getemajor(*devp);
14640 	} else {
14641 		major = ddi_driver_major(dtrace_devi);
14642 	}
14643 
14644 	state->dts_dev = makedevice(major, minor);
14645 
14646 	if (devp != NULL)
14647 		*devp = state->dts_dev;
14648 #else
14649 	state->dts_aggid_arena = new_unrhdr(1, INT_MAX, &dtrace_unr_mtx);
14650 	state->dts_dev = dev;
14651 #endif
14652 
14653 	/*
14654 	 * We allocate NCPU buffers.  On the one hand, this can be quite
14655 	 * a bit of memory per instance (nearly 36K on a Starcat).  On the
14656 	 * other hand, it saves an additional memory reference in the probe
14657 	 * path.
14658 	 */
14659 	state->dts_buffer = kmem_zalloc(bufsize, KM_SLEEP);
14660 	state->dts_aggbuffer = kmem_zalloc(bufsize, KM_SLEEP);
14661 
14662 	/*
14663          * Allocate and initialise the per-process per-CPU random state.
14664 	 * SI_SUB_RANDOM < SI_SUB_DTRACE_ANON therefore entropy device is
14665          * assumed to be seeded at this point (if from Fortuna seed file).
14666 	 */
14667 	arc4random_buf(&state->dts_rstate[0], 2 * sizeof(uint64_t));
14668 	for (cpu_it = 1; cpu_it < NCPU; cpu_it++) {
14669 		/*
14670 		 * Each CPU is assigned a 2^64 period, non-overlapping
14671 		 * subsequence.
14672 		 */
14673 		dtrace_xoroshiro128_plus_jump(state->dts_rstate[cpu_it-1],
14674 		    state->dts_rstate[cpu_it]);
14675 	}
14676 
14677 #ifdef illumos
14678 	state->dts_cleaner = CYCLIC_NONE;
14679 	state->dts_deadman = CYCLIC_NONE;
14680 #else
14681 	callout_init(&state->dts_cleaner, 1);
14682 	callout_init(&state->dts_deadman, 1);
14683 #endif
14684 	state->dts_vstate.dtvs_state = state;
14685 
14686 	for (i = 0; i < DTRACEOPT_MAX; i++)
14687 		state->dts_options[i] = DTRACEOPT_UNSET;
14688 
14689 	/*
14690 	 * Set the default options.
14691 	 */
14692 	opt = state->dts_options;
14693 	opt[DTRACEOPT_BUFPOLICY] = DTRACEOPT_BUFPOLICY_SWITCH;
14694 	opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_AUTO;
14695 	opt[DTRACEOPT_NSPEC] = dtrace_nspec_default;
14696 	opt[DTRACEOPT_SPECSIZE] = dtrace_specsize_default;
14697 	opt[DTRACEOPT_CPU] = (dtrace_optval_t)DTRACE_CPUALL;
14698 	opt[DTRACEOPT_STRSIZE] = dtrace_strsize_default;
14699 	opt[DTRACEOPT_STACKFRAMES] = dtrace_stackframes_default;
14700 	opt[DTRACEOPT_USTACKFRAMES] = dtrace_ustackframes_default;
14701 	opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_default;
14702 	opt[DTRACEOPT_AGGRATE] = dtrace_aggrate_default;
14703 	opt[DTRACEOPT_SWITCHRATE] = dtrace_switchrate_default;
14704 	opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_default;
14705 	opt[DTRACEOPT_JSTACKFRAMES] = dtrace_jstackframes_default;
14706 	opt[DTRACEOPT_JSTACKSTRSIZE] = dtrace_jstackstrsize_default;
14707 
14708 	state->dts_activity = DTRACE_ACTIVITY_INACTIVE;
14709 
14710 	/*
14711 	 * Depending on the user credentials, we set flag bits which alter probe
14712 	 * visibility or the amount of destructiveness allowed.  In the case of
14713 	 * actual anonymous tracing, or the possession of all privileges, all of
14714 	 * the normal checks are bypassed.
14715 	 */
14716 	if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
14717 		state->dts_cred.dcr_visible = DTRACE_CRV_ALL;
14718 		state->dts_cred.dcr_action = DTRACE_CRA_ALL;
14719 	} else {
14720 		/*
14721 		 * Set up the credentials for this instantiation.  We take a
14722 		 * hold on the credential to prevent it from disappearing on
14723 		 * us; this in turn prevents the zone_t referenced by this
14724 		 * credential from disappearing.  This means that we can
14725 		 * examine the credential and the zone from probe context.
14726 		 */
14727 		crhold(cr);
14728 		state->dts_cred.dcr_cred = cr;
14729 
14730 		/*
14731 		 * CRA_PROC means "we have *some* privilege for dtrace" and
14732 		 * unlocks the use of variables like pid, zonename, etc.
14733 		 */
14734 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE) ||
14735 		    PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
14736 			state->dts_cred.dcr_action |= DTRACE_CRA_PROC;
14737 		}
14738 
14739 		/*
14740 		 * dtrace_user allows use of syscall and profile providers.
14741 		 * If the user also has proc_owner and/or proc_zone, we
14742 		 * extend the scope to include additional visibility and
14743 		 * destructive power.
14744 		 */
14745 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE)) {
14746 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE)) {
14747 				state->dts_cred.dcr_visible |=
14748 				    DTRACE_CRV_ALLPROC;
14749 
14750 				state->dts_cred.dcr_action |=
14751 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
14752 			}
14753 
14754 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE)) {
14755 				state->dts_cred.dcr_visible |=
14756 				    DTRACE_CRV_ALLZONE;
14757 
14758 				state->dts_cred.dcr_action |=
14759 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
14760 			}
14761 
14762 			/*
14763 			 * If we have all privs in whatever zone this is,
14764 			 * we can do destructive things to processes which
14765 			 * have altered credentials.
14766 			 */
14767 #ifdef illumos
14768 			if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
14769 			    cr->cr_zone->zone_privset)) {
14770 				state->dts_cred.dcr_action |=
14771 				    DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
14772 			}
14773 #endif
14774 		}
14775 
14776 		/*
14777 		 * Holding the dtrace_kernel privilege also implies that
14778 		 * the user has the dtrace_user privilege from a visibility
14779 		 * perspective.  But without further privileges, some
14780 		 * destructive actions are not available.
14781 		 */
14782 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE)) {
14783 			/*
14784 			 * Make all probes in all zones visible.  However,
14785 			 * this doesn't mean that all actions become available
14786 			 * to all zones.
14787 			 */
14788 			state->dts_cred.dcr_visible |= DTRACE_CRV_KERNEL |
14789 			    DTRACE_CRV_ALLPROC | DTRACE_CRV_ALLZONE;
14790 
14791 			state->dts_cred.dcr_action |= DTRACE_CRA_KERNEL |
14792 			    DTRACE_CRA_PROC;
14793 			/*
14794 			 * Holding proc_owner means that destructive actions
14795 			 * for *this* zone are allowed.
14796 			 */
14797 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
14798 				state->dts_cred.dcr_action |=
14799 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
14800 
14801 			/*
14802 			 * Holding proc_zone means that destructive actions
14803 			 * for this user/group ID in all zones is allowed.
14804 			 */
14805 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
14806 				state->dts_cred.dcr_action |=
14807 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
14808 
14809 #ifdef illumos
14810 			/*
14811 			 * If we have all privs in whatever zone this is,
14812 			 * we can do destructive things to processes which
14813 			 * have altered credentials.
14814 			 */
14815 			if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
14816 			    cr->cr_zone->zone_privset)) {
14817 				state->dts_cred.dcr_action |=
14818 				    DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
14819 			}
14820 #endif
14821 		}
14822 
14823 		/*
14824 		 * Holding the dtrace_proc privilege gives control over fasttrap
14825 		 * and pid providers.  We need to grant wider destructive
14826 		 * privileges in the event that the user has proc_owner and/or
14827 		 * proc_zone.
14828 		 */
14829 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
14830 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
14831 				state->dts_cred.dcr_action |=
14832 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
14833 
14834 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
14835 				state->dts_cred.dcr_action |=
14836 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
14837 		}
14838 	}
14839 
14840 	return (state);
14841 }
14842 
14843 static int
14844 dtrace_state_buffer(dtrace_state_t *state, dtrace_buffer_t *buf, int which)
14845 {
14846 	dtrace_optval_t *opt = state->dts_options, size;
14847 	processorid_t cpu = 0;
14848 	int flags = 0, rval, factor, divisor = 1;
14849 
14850 	ASSERT(MUTEX_HELD(&dtrace_lock));
14851 	ASSERT(MUTEX_HELD(&cpu_lock));
14852 	ASSERT(which < DTRACEOPT_MAX);
14853 	ASSERT(state->dts_activity == DTRACE_ACTIVITY_INACTIVE ||
14854 	    (state == dtrace_anon.dta_state &&
14855 	    state->dts_activity == DTRACE_ACTIVITY_ACTIVE));
14856 
14857 	if (opt[which] == DTRACEOPT_UNSET || opt[which] == 0)
14858 		return (0);
14859 
14860 	if (opt[DTRACEOPT_CPU] != DTRACEOPT_UNSET)
14861 		cpu = opt[DTRACEOPT_CPU];
14862 
14863 	if (which == DTRACEOPT_SPECSIZE)
14864 		flags |= DTRACEBUF_NOSWITCH;
14865 
14866 	if (which == DTRACEOPT_BUFSIZE) {
14867 		if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_RING)
14868 			flags |= DTRACEBUF_RING | DTRACEBUF_NOSWITCH;
14869 
14870 		if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_FILL)
14871 			flags |= DTRACEBUF_FILL | DTRACEBUF_NOSWITCH;
14872 
14873 		if (state != dtrace_anon.dta_state ||
14874 		    state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
14875 			flags |= DTRACEBUF_INACTIVE;
14876 	}
14877 
14878 	for (size = opt[which]; size >= sizeof (uint64_t); size /= divisor) {
14879 		/*
14880 		 * The size must be 8-byte aligned.  If the size is not 8-byte
14881 		 * aligned, drop it down by the difference.
14882 		 */
14883 		if (size & (sizeof (uint64_t) - 1))
14884 			size -= size & (sizeof (uint64_t) - 1);
14885 
14886 		if (size < state->dts_reserve) {
14887 			/*
14888 			 * Buffers always must be large enough to accommodate
14889 			 * their prereserved space.  We return E2BIG instead
14890 			 * of ENOMEM in this case to allow for user-level
14891 			 * software to differentiate the cases.
14892 			 */
14893 			return (E2BIG);
14894 		}
14895 
14896 		rval = dtrace_buffer_alloc(buf, size, flags, cpu, &factor);
14897 
14898 		if (rval != ENOMEM) {
14899 			opt[which] = size;
14900 			return (rval);
14901 		}
14902 
14903 		if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
14904 			return (rval);
14905 
14906 		for (divisor = 2; divisor < factor; divisor <<= 1)
14907 			continue;
14908 	}
14909 
14910 	return (ENOMEM);
14911 }
14912 
14913 static int
14914 dtrace_state_buffers(dtrace_state_t *state)
14915 {
14916 	dtrace_speculation_t *spec = state->dts_speculations;
14917 	int rval, i;
14918 
14919 	if ((rval = dtrace_state_buffer(state, state->dts_buffer,
14920 	    DTRACEOPT_BUFSIZE)) != 0)
14921 		return (rval);
14922 
14923 	if ((rval = dtrace_state_buffer(state, state->dts_aggbuffer,
14924 	    DTRACEOPT_AGGSIZE)) != 0)
14925 		return (rval);
14926 
14927 	for (i = 0; i < state->dts_nspeculations; i++) {
14928 		if ((rval = dtrace_state_buffer(state,
14929 		    spec[i].dtsp_buffer, DTRACEOPT_SPECSIZE)) != 0)
14930 			return (rval);
14931 	}
14932 
14933 	return (0);
14934 }
14935 
14936 static void
14937 dtrace_state_prereserve(dtrace_state_t *state)
14938 {
14939 	dtrace_ecb_t *ecb;
14940 	dtrace_probe_t *probe;
14941 
14942 	state->dts_reserve = 0;
14943 
14944 	if (state->dts_options[DTRACEOPT_BUFPOLICY] != DTRACEOPT_BUFPOLICY_FILL)
14945 		return;
14946 
14947 	/*
14948 	 * If our buffer policy is a "fill" buffer policy, we need to set the
14949 	 * prereserved space to be the space required by the END probes.
14950 	 */
14951 	probe = dtrace_probes[dtrace_probeid_end - 1];
14952 	ASSERT(probe != NULL);
14953 
14954 	for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
14955 		if (ecb->dte_state != state)
14956 			continue;
14957 
14958 		state->dts_reserve += ecb->dte_needed + ecb->dte_alignment;
14959 	}
14960 }
14961 
14962 static int
14963 dtrace_state_go(dtrace_state_t *state, processorid_t *cpu)
14964 {
14965 	dtrace_optval_t *opt = state->dts_options, sz, nspec;
14966 	dtrace_speculation_t *spec;
14967 	dtrace_buffer_t *buf;
14968 #ifdef illumos
14969 	cyc_handler_t hdlr;
14970 	cyc_time_t when;
14971 #endif
14972 	int rval = 0, i, bufsize = NCPU * sizeof (dtrace_buffer_t);
14973 	dtrace_icookie_t cookie;
14974 
14975 	mutex_enter(&cpu_lock);
14976 	mutex_enter(&dtrace_lock);
14977 
14978 	if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
14979 		rval = EBUSY;
14980 		goto out;
14981 	}
14982 
14983 	/*
14984 	 * Before we can perform any checks, we must prime all of the
14985 	 * retained enablings that correspond to this state.
14986 	 */
14987 	dtrace_enabling_prime(state);
14988 
14989 	if (state->dts_destructive && !state->dts_cred.dcr_destructive) {
14990 		rval = EACCES;
14991 		goto out;
14992 	}
14993 
14994 	dtrace_state_prereserve(state);
14995 
14996 	/*
14997 	 * Now we want to do is try to allocate our speculations.
14998 	 * We do not automatically resize the number of speculations; if
14999 	 * this fails, we will fail the operation.
15000 	 */
15001 	nspec = opt[DTRACEOPT_NSPEC];
15002 	ASSERT(nspec != DTRACEOPT_UNSET);
15003 
15004 	if (nspec > INT_MAX) {
15005 		rval = ENOMEM;
15006 		goto out;
15007 	}
15008 
15009 	spec = kmem_zalloc(nspec * sizeof (dtrace_speculation_t),
15010 	    KM_NOSLEEP | KM_NORMALPRI);
15011 
15012 	if (spec == NULL) {
15013 		rval = ENOMEM;
15014 		goto out;
15015 	}
15016 
15017 	state->dts_speculations = spec;
15018 	state->dts_nspeculations = (int)nspec;
15019 
15020 	for (i = 0; i < nspec; i++) {
15021 		if ((buf = kmem_zalloc(bufsize,
15022 		    KM_NOSLEEP | KM_NORMALPRI)) == NULL) {
15023 			rval = ENOMEM;
15024 			goto err;
15025 		}
15026 
15027 		spec[i].dtsp_buffer = buf;
15028 	}
15029 
15030 	if (opt[DTRACEOPT_GRABANON] != DTRACEOPT_UNSET) {
15031 		if (dtrace_anon.dta_state == NULL) {
15032 			rval = ENOENT;
15033 			goto out;
15034 		}
15035 
15036 		if (state->dts_necbs != 0) {
15037 			rval = EALREADY;
15038 			goto out;
15039 		}
15040 
15041 		state->dts_anon = dtrace_anon_grab();
15042 		ASSERT(state->dts_anon != NULL);
15043 		state = state->dts_anon;
15044 
15045 		/*
15046 		 * We want "grabanon" to be set in the grabbed state, so we'll
15047 		 * copy that option value from the grabbing state into the
15048 		 * grabbed state.
15049 		 */
15050 		state->dts_options[DTRACEOPT_GRABANON] =
15051 		    opt[DTRACEOPT_GRABANON];
15052 
15053 		*cpu = dtrace_anon.dta_beganon;
15054 
15055 		/*
15056 		 * If the anonymous state is active (as it almost certainly
15057 		 * is if the anonymous enabling ultimately matched anything),
15058 		 * we don't allow any further option processing -- but we
15059 		 * don't return failure.
15060 		 */
15061 		if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
15062 			goto out;
15063 	}
15064 
15065 	if (opt[DTRACEOPT_AGGSIZE] != DTRACEOPT_UNSET &&
15066 	    opt[DTRACEOPT_AGGSIZE] != 0) {
15067 		if (state->dts_aggregations == NULL) {
15068 			/*
15069 			 * We're not going to create an aggregation buffer
15070 			 * because we don't have any ECBs that contain
15071 			 * aggregations -- set this option to 0.
15072 			 */
15073 			opt[DTRACEOPT_AGGSIZE] = 0;
15074 		} else {
15075 			/*
15076 			 * If we have an aggregation buffer, we must also have
15077 			 * a buffer to use as scratch.
15078 			 */
15079 			if (opt[DTRACEOPT_BUFSIZE] == DTRACEOPT_UNSET ||
15080 			    opt[DTRACEOPT_BUFSIZE] < state->dts_needed) {
15081 				opt[DTRACEOPT_BUFSIZE] = state->dts_needed;
15082 			}
15083 		}
15084 	}
15085 
15086 	if (opt[DTRACEOPT_SPECSIZE] != DTRACEOPT_UNSET &&
15087 	    opt[DTRACEOPT_SPECSIZE] != 0) {
15088 		if (!state->dts_speculates) {
15089 			/*
15090 			 * We're not going to create speculation buffers
15091 			 * because we don't have any ECBs that actually
15092 			 * speculate -- set the speculation size to 0.
15093 			 */
15094 			opt[DTRACEOPT_SPECSIZE] = 0;
15095 		}
15096 	}
15097 
15098 	/*
15099 	 * The bare minimum size for any buffer that we're actually going to
15100 	 * do anything to is sizeof (uint64_t).
15101 	 */
15102 	sz = sizeof (uint64_t);
15103 
15104 	if ((state->dts_needed != 0 && opt[DTRACEOPT_BUFSIZE] < sz) ||
15105 	    (state->dts_speculates && opt[DTRACEOPT_SPECSIZE] < sz) ||
15106 	    (state->dts_aggregations != NULL && opt[DTRACEOPT_AGGSIZE] < sz)) {
15107 		/*
15108 		 * A buffer size has been explicitly set to 0 (or to a size
15109 		 * that will be adjusted to 0) and we need the space -- we
15110 		 * need to return failure.  We return ENOSPC to differentiate
15111 		 * it from failing to allocate a buffer due to failure to meet
15112 		 * the reserve (for which we return E2BIG).
15113 		 */
15114 		rval = ENOSPC;
15115 		goto out;
15116 	}
15117 
15118 	if ((rval = dtrace_state_buffers(state)) != 0)
15119 		goto err;
15120 
15121 	if ((sz = opt[DTRACEOPT_DYNVARSIZE]) == DTRACEOPT_UNSET)
15122 		sz = dtrace_dstate_defsize;
15123 
15124 	do {
15125 		rval = dtrace_dstate_init(&state->dts_vstate.dtvs_dynvars, sz);
15126 
15127 		if (rval == 0)
15128 			break;
15129 
15130 		if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
15131 			goto err;
15132 	} while (sz >>= 1);
15133 
15134 	opt[DTRACEOPT_DYNVARSIZE] = sz;
15135 
15136 	if (rval != 0)
15137 		goto err;
15138 
15139 	if (opt[DTRACEOPT_STATUSRATE] > dtrace_statusrate_max)
15140 		opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_max;
15141 
15142 	if (opt[DTRACEOPT_CLEANRATE] == 0)
15143 		opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
15144 
15145 	if (opt[DTRACEOPT_CLEANRATE] < dtrace_cleanrate_min)
15146 		opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_min;
15147 
15148 	if (opt[DTRACEOPT_CLEANRATE] > dtrace_cleanrate_max)
15149 		opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
15150 
15151 	state->dts_alive = state->dts_laststatus = dtrace_gethrtime();
15152 #ifdef illumos
15153 	hdlr.cyh_func = (cyc_func_t)dtrace_state_clean;
15154 	hdlr.cyh_arg = state;
15155 	hdlr.cyh_level = CY_LOW_LEVEL;
15156 
15157 	when.cyt_when = 0;
15158 	when.cyt_interval = opt[DTRACEOPT_CLEANRATE];
15159 
15160 	state->dts_cleaner = cyclic_add(&hdlr, &when);
15161 
15162 	hdlr.cyh_func = (cyc_func_t)dtrace_state_deadman;
15163 	hdlr.cyh_arg = state;
15164 	hdlr.cyh_level = CY_LOW_LEVEL;
15165 
15166 	when.cyt_when = 0;
15167 	when.cyt_interval = dtrace_deadman_interval;
15168 
15169 	state->dts_deadman = cyclic_add(&hdlr, &when);
15170 #else
15171 	callout_reset(&state->dts_cleaner, hz * opt[DTRACEOPT_CLEANRATE] / NANOSEC,
15172 	    dtrace_state_clean, state);
15173 	callout_reset(&state->dts_deadman, hz * dtrace_deadman_interval / NANOSEC,
15174 	    dtrace_state_deadman, state);
15175 #endif
15176 
15177 	state->dts_activity = DTRACE_ACTIVITY_WARMUP;
15178 
15179 #ifdef illumos
15180 	if (state->dts_getf != 0 &&
15181 	    !(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)) {
15182 		/*
15183 		 * We don't have kernel privs but we have at least one call
15184 		 * to getf(); we need to bump our zone's count, and (if
15185 		 * this is the first enabling to have an unprivileged call
15186 		 * to getf()) we need to hook into closef().
15187 		 */
15188 		state->dts_cred.dcr_cred->cr_zone->zone_dtrace_getf++;
15189 
15190 		if (dtrace_getf++ == 0) {
15191 			ASSERT(dtrace_closef == NULL);
15192 			dtrace_closef = dtrace_getf_barrier;
15193 		}
15194 	}
15195 #endif
15196 
15197 	/*
15198 	 * Now it's time to actually fire the BEGIN probe.  We need to disable
15199 	 * interrupts here both to record the CPU on which we fired the BEGIN
15200 	 * probe (the data from this CPU will be processed first at user
15201 	 * level) and to manually activate the buffer for this CPU.
15202 	 */
15203 	cookie = dtrace_interrupt_disable();
15204 	*cpu = curcpu;
15205 	ASSERT(state->dts_buffer[*cpu].dtb_flags & DTRACEBUF_INACTIVE);
15206 	state->dts_buffer[*cpu].dtb_flags &= ~DTRACEBUF_INACTIVE;
15207 
15208 	dtrace_probe(dtrace_probeid_begin,
15209 	    (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
15210 	dtrace_interrupt_enable(cookie);
15211 	/*
15212 	 * We may have had an exit action from a BEGIN probe; only change our
15213 	 * state to ACTIVE if we're still in WARMUP.
15214 	 */
15215 	ASSERT(state->dts_activity == DTRACE_ACTIVITY_WARMUP ||
15216 	    state->dts_activity == DTRACE_ACTIVITY_DRAINING);
15217 
15218 	if (state->dts_activity == DTRACE_ACTIVITY_WARMUP)
15219 		state->dts_activity = DTRACE_ACTIVITY_ACTIVE;
15220 
15221 #ifdef __FreeBSD__
15222 	/*
15223 	 * We enable anonymous tracing before APs are started, so we must
15224 	 * activate buffers using the current CPU.
15225 	 */
15226 	if (state == dtrace_anon.dta_state)
15227 		for (int i = 0; i < NCPU; i++)
15228 			dtrace_buffer_activate_cpu(state, i);
15229 	else
15230 		dtrace_xcall(DTRACE_CPUALL,
15231 		    (dtrace_xcall_t)dtrace_buffer_activate, state);
15232 #else
15233 	/*
15234 	 * Regardless of whether or not now we're in ACTIVE or DRAINING, we
15235 	 * want each CPU to transition its principal buffer out of the
15236 	 * INACTIVE state.  Doing this assures that no CPU will suddenly begin
15237 	 * processing an ECB halfway down a probe's ECB chain; all CPUs will
15238 	 * atomically transition from processing none of a state's ECBs to
15239 	 * processing all of them.
15240 	 */
15241 	dtrace_xcall(DTRACE_CPUALL,
15242 	    (dtrace_xcall_t)dtrace_buffer_activate, state);
15243 #endif
15244 	goto out;
15245 
15246 err:
15247 	dtrace_buffer_free(state->dts_buffer);
15248 	dtrace_buffer_free(state->dts_aggbuffer);
15249 
15250 	if ((nspec = state->dts_nspeculations) == 0) {
15251 		ASSERT(state->dts_speculations == NULL);
15252 		goto out;
15253 	}
15254 
15255 	spec = state->dts_speculations;
15256 	ASSERT(spec != NULL);
15257 
15258 	for (i = 0; i < state->dts_nspeculations; i++) {
15259 		if ((buf = spec[i].dtsp_buffer) == NULL)
15260 			break;
15261 
15262 		dtrace_buffer_free(buf);
15263 		kmem_free(buf, bufsize);
15264 	}
15265 
15266 	kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
15267 	state->dts_nspeculations = 0;
15268 	state->dts_speculations = NULL;
15269 
15270 out:
15271 	mutex_exit(&dtrace_lock);
15272 	mutex_exit(&cpu_lock);
15273 
15274 	return (rval);
15275 }
15276 
15277 static int
15278 dtrace_state_stop(dtrace_state_t *state, processorid_t *cpu)
15279 {
15280 	dtrace_icookie_t cookie;
15281 
15282 	ASSERT(MUTEX_HELD(&dtrace_lock));
15283 
15284 	if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE &&
15285 	    state->dts_activity != DTRACE_ACTIVITY_DRAINING)
15286 		return (EINVAL);
15287 
15288 	/*
15289 	 * We'll set the activity to DTRACE_ACTIVITY_DRAINING, and issue a sync
15290 	 * to be sure that every CPU has seen it.  See below for the details
15291 	 * on why this is done.
15292 	 */
15293 	state->dts_activity = DTRACE_ACTIVITY_DRAINING;
15294 	dtrace_sync();
15295 
15296 	/*
15297 	 * By this point, it is impossible for any CPU to be still processing
15298 	 * with DTRACE_ACTIVITY_ACTIVE.  We can thus set our activity to
15299 	 * DTRACE_ACTIVITY_COOLDOWN and know that we're not racing with any
15300 	 * other CPU in dtrace_buffer_reserve().  This allows dtrace_probe()
15301 	 * and callees to know that the activity is DTRACE_ACTIVITY_COOLDOWN
15302 	 * iff we're in the END probe.
15303 	 */
15304 	state->dts_activity = DTRACE_ACTIVITY_COOLDOWN;
15305 	dtrace_sync();
15306 	ASSERT(state->dts_activity == DTRACE_ACTIVITY_COOLDOWN);
15307 
15308 	/*
15309 	 * Finally, we can release the reserve and call the END probe.  We
15310 	 * disable interrupts across calling the END probe to allow us to
15311 	 * return the CPU on which we actually called the END probe.  This
15312 	 * allows user-land to be sure that this CPU's principal buffer is
15313 	 * processed last.
15314 	 */
15315 	state->dts_reserve = 0;
15316 
15317 	cookie = dtrace_interrupt_disable();
15318 	*cpu = curcpu;
15319 	dtrace_probe(dtrace_probeid_end,
15320 	    (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
15321 	dtrace_interrupt_enable(cookie);
15322 
15323 	state->dts_activity = DTRACE_ACTIVITY_STOPPED;
15324 	dtrace_sync();
15325 
15326 #ifdef illumos
15327 	if (state->dts_getf != 0 &&
15328 	    !(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)) {
15329 		/*
15330 		 * We don't have kernel privs but we have at least one call
15331 		 * to getf(); we need to lower our zone's count, and (if
15332 		 * this is the last enabling to have an unprivileged call
15333 		 * to getf()) we need to clear the closef() hook.
15334 		 */
15335 		ASSERT(state->dts_cred.dcr_cred->cr_zone->zone_dtrace_getf > 0);
15336 		ASSERT(dtrace_closef == dtrace_getf_barrier);
15337 		ASSERT(dtrace_getf > 0);
15338 
15339 		state->dts_cred.dcr_cred->cr_zone->zone_dtrace_getf--;
15340 
15341 		if (--dtrace_getf == 0)
15342 			dtrace_closef = NULL;
15343 	}
15344 #endif
15345 
15346 	return (0);
15347 }
15348 
15349 static int
15350 dtrace_state_option(dtrace_state_t *state, dtrace_optid_t option,
15351     dtrace_optval_t val)
15352 {
15353 	ASSERT(MUTEX_HELD(&dtrace_lock));
15354 
15355 	if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
15356 		return (EBUSY);
15357 
15358 	if (option >= DTRACEOPT_MAX)
15359 		return (EINVAL);
15360 
15361 	if (option != DTRACEOPT_CPU && val < 0)
15362 		return (EINVAL);
15363 
15364 	switch (option) {
15365 	case DTRACEOPT_DESTRUCTIVE:
15366 		if (dtrace_destructive_disallow)
15367 			return (EACCES);
15368 
15369 		state->dts_cred.dcr_destructive = 1;
15370 		break;
15371 
15372 	case DTRACEOPT_BUFSIZE:
15373 	case DTRACEOPT_DYNVARSIZE:
15374 	case DTRACEOPT_AGGSIZE:
15375 	case DTRACEOPT_SPECSIZE:
15376 	case DTRACEOPT_STRSIZE:
15377 		if (val < 0)
15378 			return (EINVAL);
15379 
15380 		if (val >= LONG_MAX) {
15381 			/*
15382 			 * If this is an otherwise negative value, set it to
15383 			 * the highest multiple of 128m less than LONG_MAX.
15384 			 * Technically, we're adjusting the size without
15385 			 * regard to the buffer resizing policy, but in fact,
15386 			 * this has no effect -- if we set the buffer size to
15387 			 * ~LONG_MAX and the buffer policy is ultimately set to
15388 			 * be "manual", the buffer allocation is guaranteed to
15389 			 * fail, if only because the allocation requires two
15390 			 * buffers.  (We set the the size to the highest
15391 			 * multiple of 128m because it ensures that the size
15392 			 * will remain a multiple of a megabyte when
15393 			 * repeatedly halved -- all the way down to 15m.)
15394 			 */
15395 			val = LONG_MAX - (1 << 27) + 1;
15396 		}
15397 	}
15398 
15399 	state->dts_options[option] = val;
15400 
15401 	return (0);
15402 }
15403 
15404 static void
15405 dtrace_state_destroy(dtrace_state_t *state)
15406 {
15407 	dtrace_ecb_t *ecb;
15408 	dtrace_vstate_t *vstate = &state->dts_vstate;
15409 #ifdef illumos
15410 	minor_t minor = getminor(state->dts_dev);
15411 #endif
15412 	int i, bufsize = NCPU * sizeof (dtrace_buffer_t);
15413 	dtrace_speculation_t *spec = state->dts_speculations;
15414 	int nspec = state->dts_nspeculations;
15415 	uint32_t match;
15416 
15417 	ASSERT(MUTEX_HELD(&dtrace_lock));
15418 	ASSERT(MUTEX_HELD(&cpu_lock));
15419 
15420 	/*
15421 	 * First, retract any retained enablings for this state.
15422 	 */
15423 	dtrace_enabling_retract(state);
15424 	ASSERT(state->dts_nretained == 0);
15425 
15426 	if (state->dts_activity == DTRACE_ACTIVITY_ACTIVE ||
15427 	    state->dts_activity == DTRACE_ACTIVITY_DRAINING) {
15428 		/*
15429 		 * We have managed to come into dtrace_state_destroy() on a
15430 		 * hot enabling -- almost certainly because of a disorderly
15431 		 * shutdown of a consumer.  (That is, a consumer that is
15432 		 * exiting without having called dtrace_stop().) In this case,
15433 		 * we're going to set our activity to be KILLED, and then
15434 		 * issue a sync to be sure that everyone is out of probe
15435 		 * context before we start blowing away ECBs.
15436 		 */
15437 		state->dts_activity = DTRACE_ACTIVITY_KILLED;
15438 		dtrace_sync();
15439 	}
15440 
15441 	/*
15442 	 * Release the credential hold we took in dtrace_state_create().
15443 	 */
15444 	if (state->dts_cred.dcr_cred != NULL)
15445 		crfree(state->dts_cred.dcr_cred);
15446 
15447 	/*
15448 	 * Now we can safely disable and destroy any enabled probes.  Because
15449 	 * any DTRACE_PRIV_KERNEL probes may actually be slowing our progress
15450 	 * (especially if they're all enabled), we take two passes through the
15451 	 * ECBs:  in the first, we disable just DTRACE_PRIV_KERNEL probes, and
15452 	 * in the second we disable whatever is left over.
15453 	 */
15454 	for (match = DTRACE_PRIV_KERNEL; ; match = 0) {
15455 		for (i = 0; i < state->dts_necbs; i++) {
15456 			if ((ecb = state->dts_ecbs[i]) == NULL)
15457 				continue;
15458 
15459 			if (match && ecb->dte_probe != NULL) {
15460 				dtrace_probe_t *probe = ecb->dte_probe;
15461 				dtrace_provider_t *prov = probe->dtpr_provider;
15462 
15463 				if (!(prov->dtpv_priv.dtpp_flags & match))
15464 					continue;
15465 			}
15466 
15467 			dtrace_ecb_disable(ecb);
15468 			dtrace_ecb_destroy(ecb);
15469 		}
15470 
15471 		if (!match)
15472 			break;
15473 	}
15474 
15475 	/*
15476 	 * Before we free the buffers, perform one more sync to assure that
15477 	 * every CPU is out of probe context.
15478 	 */
15479 	dtrace_sync();
15480 
15481 	dtrace_buffer_free(state->dts_buffer);
15482 	dtrace_buffer_free(state->dts_aggbuffer);
15483 
15484 	for (i = 0; i < nspec; i++)
15485 		dtrace_buffer_free(spec[i].dtsp_buffer);
15486 
15487 #ifdef illumos
15488 	if (state->dts_cleaner != CYCLIC_NONE)
15489 		cyclic_remove(state->dts_cleaner);
15490 
15491 	if (state->dts_deadman != CYCLIC_NONE)
15492 		cyclic_remove(state->dts_deadman);
15493 #else
15494 	callout_stop(&state->dts_cleaner);
15495 	callout_drain(&state->dts_cleaner);
15496 	callout_stop(&state->dts_deadman);
15497 	callout_drain(&state->dts_deadman);
15498 #endif
15499 
15500 	dtrace_dstate_fini(&vstate->dtvs_dynvars);
15501 	dtrace_vstate_fini(vstate);
15502 	if (state->dts_ecbs != NULL)
15503 		kmem_free(state->dts_ecbs, state->dts_necbs * sizeof (dtrace_ecb_t *));
15504 
15505 	if (state->dts_aggregations != NULL) {
15506 #ifdef DEBUG
15507 		for (i = 0; i < state->dts_naggregations; i++)
15508 			ASSERT(state->dts_aggregations[i] == NULL);
15509 #endif
15510 		ASSERT(state->dts_naggregations > 0);
15511 		kmem_free(state->dts_aggregations,
15512 		    state->dts_naggregations * sizeof (dtrace_aggregation_t *));
15513 	}
15514 
15515 	kmem_free(state->dts_buffer, bufsize);
15516 	kmem_free(state->dts_aggbuffer, bufsize);
15517 
15518 	for (i = 0; i < nspec; i++)
15519 		kmem_free(spec[i].dtsp_buffer, bufsize);
15520 
15521 	if (spec != NULL)
15522 		kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
15523 
15524 	dtrace_format_destroy(state);
15525 
15526 	if (state->dts_aggid_arena != NULL) {
15527 #ifdef illumos
15528 		vmem_destroy(state->dts_aggid_arena);
15529 #else
15530 		delete_unrhdr(state->dts_aggid_arena);
15531 #endif
15532 		state->dts_aggid_arena = NULL;
15533 	}
15534 #ifdef illumos
15535 	ddi_soft_state_free(dtrace_softstate, minor);
15536 	vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
15537 #endif
15538 }
15539 
15540 /*
15541  * DTrace Anonymous Enabling Functions
15542  */
15543 static dtrace_state_t *
15544 dtrace_anon_grab(void)
15545 {
15546 	dtrace_state_t *state;
15547 
15548 	ASSERT(MUTEX_HELD(&dtrace_lock));
15549 
15550 	if ((state = dtrace_anon.dta_state) == NULL) {
15551 		ASSERT(dtrace_anon.dta_enabling == NULL);
15552 		return (NULL);
15553 	}
15554 
15555 	ASSERT(dtrace_anon.dta_enabling != NULL);
15556 	ASSERT(dtrace_retained != NULL);
15557 
15558 	dtrace_enabling_destroy(dtrace_anon.dta_enabling);
15559 	dtrace_anon.dta_enabling = NULL;
15560 	dtrace_anon.dta_state = NULL;
15561 
15562 	return (state);
15563 }
15564 
15565 static void
15566 dtrace_anon_property(void)
15567 {
15568 	int i, rv;
15569 	dtrace_state_t *state;
15570 	dof_hdr_t *dof;
15571 	char c[32];		/* enough for "dof-data-" + digits */
15572 
15573 	ASSERT(MUTEX_HELD(&dtrace_lock));
15574 	ASSERT(MUTEX_HELD(&cpu_lock));
15575 
15576 	for (i = 0; ; i++) {
15577 		(void) snprintf(c, sizeof (c), "dof-data-%d", i);
15578 
15579 		dtrace_err_verbose = 1;
15580 
15581 		if ((dof = dtrace_dof_property(c)) == NULL) {
15582 			dtrace_err_verbose = 0;
15583 			break;
15584 		}
15585 
15586 #ifdef illumos
15587 		/*
15588 		 * We want to create anonymous state, so we need to transition
15589 		 * the kernel debugger to indicate that DTrace is active.  If
15590 		 * this fails (e.g. because the debugger has modified text in
15591 		 * some way), we won't continue with the processing.
15592 		 */
15593 		if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
15594 			cmn_err(CE_NOTE, "kernel debugger active; anonymous "
15595 			    "enabling ignored.");
15596 			dtrace_dof_destroy(dof);
15597 			break;
15598 		}
15599 #endif
15600 
15601 		/*
15602 		 * If we haven't allocated an anonymous state, we'll do so now.
15603 		 */
15604 		if ((state = dtrace_anon.dta_state) == NULL) {
15605 			state = dtrace_state_create(NULL, NULL);
15606 			dtrace_anon.dta_state = state;
15607 
15608 			if (state == NULL) {
15609 				/*
15610 				 * This basically shouldn't happen:  the only
15611 				 * failure mode from dtrace_state_create() is a
15612 				 * failure of ddi_soft_state_zalloc() that
15613 				 * itself should never happen.  Still, the
15614 				 * interface allows for a failure mode, and
15615 				 * we want to fail as gracefully as possible:
15616 				 * we'll emit an error message and cease
15617 				 * processing anonymous state in this case.
15618 				 */
15619 				cmn_err(CE_WARN, "failed to create "
15620 				    "anonymous state");
15621 				dtrace_dof_destroy(dof);
15622 				break;
15623 			}
15624 		}
15625 
15626 		rv = dtrace_dof_slurp(dof, &state->dts_vstate, CRED(),
15627 		    &dtrace_anon.dta_enabling, 0, 0, B_TRUE);
15628 
15629 		if (rv == 0)
15630 			rv = dtrace_dof_options(dof, state);
15631 
15632 		dtrace_err_verbose = 0;
15633 		dtrace_dof_destroy(dof);
15634 
15635 		if (rv != 0) {
15636 			/*
15637 			 * This is malformed DOF; chuck any anonymous state
15638 			 * that we created.
15639 			 */
15640 			ASSERT(dtrace_anon.dta_enabling == NULL);
15641 			dtrace_state_destroy(state);
15642 			dtrace_anon.dta_state = NULL;
15643 			break;
15644 		}
15645 
15646 		ASSERT(dtrace_anon.dta_enabling != NULL);
15647 	}
15648 
15649 	if (dtrace_anon.dta_enabling != NULL) {
15650 		int rval;
15651 
15652 		/*
15653 		 * dtrace_enabling_retain() can only fail because we are
15654 		 * trying to retain more enablings than are allowed -- but
15655 		 * we only have one anonymous enabling, and we are guaranteed
15656 		 * to be allowed at least one retained enabling; we assert
15657 		 * that dtrace_enabling_retain() returns success.
15658 		 */
15659 		rval = dtrace_enabling_retain(dtrace_anon.dta_enabling);
15660 		ASSERT(rval == 0);
15661 
15662 		dtrace_enabling_dump(dtrace_anon.dta_enabling);
15663 	}
15664 }
15665 
15666 /*
15667  * DTrace Helper Functions
15668  */
15669 static void
15670 dtrace_helper_trace(dtrace_helper_action_t *helper,
15671     dtrace_mstate_t *mstate, dtrace_vstate_t *vstate, int where)
15672 {
15673 	uint32_t size, next, nnext, i;
15674 	dtrace_helptrace_t *ent, *buffer;
15675 	uint16_t flags = cpu_core[curcpu].cpuc_dtrace_flags;
15676 
15677 	if ((buffer = dtrace_helptrace_buffer) == NULL)
15678 		return;
15679 
15680 	ASSERT(vstate->dtvs_nlocals <= dtrace_helptrace_nlocals);
15681 
15682 	/*
15683 	 * What would a tracing framework be without its own tracing
15684 	 * framework?  (Well, a hell of a lot simpler, for starters...)
15685 	 */
15686 	size = sizeof (dtrace_helptrace_t) + dtrace_helptrace_nlocals *
15687 	    sizeof (uint64_t) - sizeof (uint64_t);
15688 
15689 	/*
15690 	 * Iterate until we can allocate a slot in the trace buffer.
15691 	 */
15692 	do {
15693 		next = dtrace_helptrace_next;
15694 
15695 		if (next + size < dtrace_helptrace_bufsize) {
15696 			nnext = next + size;
15697 		} else {
15698 			nnext = size;
15699 		}
15700 	} while (dtrace_cas32(&dtrace_helptrace_next, next, nnext) != next);
15701 
15702 	/*
15703 	 * We have our slot; fill it in.
15704 	 */
15705 	if (nnext == size) {
15706 		dtrace_helptrace_wrapped++;
15707 		next = 0;
15708 	}
15709 
15710 	ent = (dtrace_helptrace_t *)((uintptr_t)buffer + next);
15711 	ent->dtht_helper = helper;
15712 	ent->dtht_where = where;
15713 	ent->dtht_nlocals = vstate->dtvs_nlocals;
15714 
15715 	ent->dtht_fltoffs = (mstate->dtms_present & DTRACE_MSTATE_FLTOFFS) ?
15716 	    mstate->dtms_fltoffs : -1;
15717 	ent->dtht_fault = DTRACE_FLAGS2FLT(flags);
15718 	ent->dtht_illval = cpu_core[curcpu].cpuc_dtrace_illval;
15719 
15720 	for (i = 0; i < vstate->dtvs_nlocals; i++) {
15721 		dtrace_statvar_t *svar;
15722 
15723 		if ((svar = vstate->dtvs_locals[i]) == NULL)
15724 			continue;
15725 
15726 		ASSERT(svar->dtsv_size >= NCPU * sizeof (uint64_t));
15727 		ent->dtht_locals[i] =
15728 		    ((uint64_t *)(uintptr_t)svar->dtsv_data)[curcpu];
15729 	}
15730 }
15731 
15732 static uint64_t
15733 dtrace_helper(int which, dtrace_mstate_t *mstate,
15734     dtrace_state_t *state, uint64_t arg0, uint64_t arg1)
15735 {
15736 	uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
15737 	uint64_t sarg0 = mstate->dtms_arg[0];
15738 	uint64_t sarg1 = mstate->dtms_arg[1];
15739 	uint64_t rval = 0;
15740 	dtrace_helpers_t *helpers = curproc->p_dtrace_helpers;
15741 	dtrace_helper_action_t *helper;
15742 	dtrace_vstate_t *vstate;
15743 	dtrace_difo_t *pred;
15744 	int i, trace = dtrace_helptrace_buffer != NULL;
15745 
15746 	ASSERT(which >= 0 && which < DTRACE_NHELPER_ACTIONS);
15747 
15748 	if (helpers == NULL)
15749 		return (0);
15750 
15751 	if ((helper = helpers->dthps_actions[which]) == NULL)
15752 		return (0);
15753 
15754 	vstate = &helpers->dthps_vstate;
15755 	mstate->dtms_arg[0] = arg0;
15756 	mstate->dtms_arg[1] = arg1;
15757 
15758 	/*
15759 	 * Now iterate over each helper.  If its predicate evaluates to 'true',
15760 	 * we'll call the corresponding actions.  Note that the below calls
15761 	 * to dtrace_dif_emulate() may set faults in machine state.  This is
15762 	 * okay:  our caller (the outer dtrace_dif_emulate()) will simply plow
15763 	 * the stored DIF offset with its own (which is the desired behavior).
15764 	 * Also, note the calls to dtrace_dif_emulate() may allocate scratch
15765 	 * from machine state; this is okay, too.
15766 	 */
15767 	for (; helper != NULL; helper = helper->dtha_next) {
15768 		if ((pred = helper->dtha_predicate) != NULL) {
15769 			if (trace)
15770 				dtrace_helper_trace(helper, mstate, vstate, 0);
15771 
15772 			if (!dtrace_dif_emulate(pred, mstate, vstate, state))
15773 				goto next;
15774 
15775 			if (*flags & CPU_DTRACE_FAULT)
15776 				goto err;
15777 		}
15778 
15779 		for (i = 0; i < helper->dtha_nactions; i++) {
15780 			if (trace)
15781 				dtrace_helper_trace(helper,
15782 				    mstate, vstate, i + 1);
15783 
15784 			rval = dtrace_dif_emulate(helper->dtha_actions[i],
15785 			    mstate, vstate, state);
15786 
15787 			if (*flags & CPU_DTRACE_FAULT)
15788 				goto err;
15789 		}
15790 
15791 next:
15792 		if (trace)
15793 			dtrace_helper_trace(helper, mstate, vstate,
15794 			    DTRACE_HELPTRACE_NEXT);
15795 	}
15796 
15797 	if (trace)
15798 		dtrace_helper_trace(helper, mstate, vstate,
15799 		    DTRACE_HELPTRACE_DONE);
15800 
15801 	/*
15802 	 * Restore the arg0 that we saved upon entry.
15803 	 */
15804 	mstate->dtms_arg[0] = sarg0;
15805 	mstate->dtms_arg[1] = sarg1;
15806 
15807 	return (rval);
15808 
15809 err:
15810 	if (trace)
15811 		dtrace_helper_trace(helper, mstate, vstate,
15812 		    DTRACE_HELPTRACE_ERR);
15813 
15814 	/*
15815 	 * Restore the arg0 that we saved upon entry.
15816 	 */
15817 	mstate->dtms_arg[0] = sarg0;
15818 	mstate->dtms_arg[1] = sarg1;
15819 
15820 	return (0);
15821 }
15822 
15823 static void
15824 dtrace_helper_action_destroy(dtrace_helper_action_t *helper,
15825     dtrace_vstate_t *vstate)
15826 {
15827 	int i;
15828 
15829 	if (helper->dtha_predicate != NULL)
15830 		dtrace_difo_release(helper->dtha_predicate, vstate);
15831 
15832 	for (i = 0; i < helper->dtha_nactions; i++) {
15833 		ASSERT(helper->dtha_actions[i] != NULL);
15834 		dtrace_difo_release(helper->dtha_actions[i], vstate);
15835 	}
15836 
15837 	kmem_free(helper->dtha_actions,
15838 	    helper->dtha_nactions * sizeof (dtrace_difo_t *));
15839 	kmem_free(helper, sizeof (dtrace_helper_action_t));
15840 }
15841 
15842 static int
15843 dtrace_helper_destroygen(dtrace_helpers_t *help, int gen)
15844 {
15845 	proc_t *p = curproc;
15846 	dtrace_vstate_t *vstate;
15847 	int i;
15848 
15849 	if (help == NULL)
15850 		help = p->p_dtrace_helpers;
15851 
15852 	ASSERT(MUTEX_HELD(&dtrace_lock));
15853 
15854 	if (help == NULL || gen > help->dthps_generation)
15855 		return (EINVAL);
15856 
15857 	vstate = &help->dthps_vstate;
15858 
15859 	for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
15860 		dtrace_helper_action_t *last = NULL, *h, *next;
15861 
15862 		for (h = help->dthps_actions[i]; h != NULL; h = next) {
15863 			next = h->dtha_next;
15864 
15865 			if (h->dtha_generation == gen) {
15866 				if (last != NULL) {
15867 					last->dtha_next = next;
15868 				} else {
15869 					help->dthps_actions[i] = next;
15870 				}
15871 
15872 				dtrace_helper_action_destroy(h, vstate);
15873 			} else {
15874 				last = h;
15875 			}
15876 		}
15877 	}
15878 
15879 	/*
15880 	 * Interate until we've cleared out all helper providers with the
15881 	 * given generation number.
15882 	 */
15883 	for (;;) {
15884 		dtrace_helper_provider_t *prov;
15885 
15886 		/*
15887 		 * Look for a helper provider with the right generation. We
15888 		 * have to start back at the beginning of the list each time
15889 		 * because we drop dtrace_lock. It's unlikely that we'll make
15890 		 * more than two passes.
15891 		 */
15892 		for (i = 0; i < help->dthps_nprovs; i++) {
15893 			prov = help->dthps_provs[i];
15894 
15895 			if (prov->dthp_generation == gen)
15896 				break;
15897 		}
15898 
15899 		/*
15900 		 * If there were no matches, we're done.
15901 		 */
15902 		if (i == help->dthps_nprovs)
15903 			break;
15904 
15905 		/*
15906 		 * Move the last helper provider into this slot.
15907 		 */
15908 		help->dthps_nprovs--;
15909 		help->dthps_provs[i] = help->dthps_provs[help->dthps_nprovs];
15910 		help->dthps_provs[help->dthps_nprovs] = NULL;
15911 
15912 		mutex_exit(&dtrace_lock);
15913 
15914 		/*
15915 		 * If we have a meta provider, remove this helper provider.
15916 		 */
15917 		mutex_enter(&dtrace_meta_lock);
15918 		if (dtrace_meta_pid != NULL) {
15919 			ASSERT(dtrace_deferred_pid == NULL);
15920 			dtrace_helper_provider_remove(&prov->dthp_prov,
15921 			    p->p_pid);
15922 		}
15923 		mutex_exit(&dtrace_meta_lock);
15924 
15925 		dtrace_helper_provider_destroy(prov);
15926 
15927 		mutex_enter(&dtrace_lock);
15928 	}
15929 
15930 	return (0);
15931 }
15932 
15933 static int
15934 dtrace_helper_validate(dtrace_helper_action_t *helper)
15935 {
15936 	int err = 0, i;
15937 	dtrace_difo_t *dp;
15938 
15939 	if ((dp = helper->dtha_predicate) != NULL)
15940 		err += dtrace_difo_validate_helper(dp);
15941 
15942 	for (i = 0; i < helper->dtha_nactions; i++)
15943 		err += dtrace_difo_validate_helper(helper->dtha_actions[i]);
15944 
15945 	return (err == 0);
15946 }
15947 
15948 static int
15949 dtrace_helper_action_add(int which, dtrace_ecbdesc_t *ep,
15950     dtrace_helpers_t *help)
15951 {
15952 	dtrace_helper_action_t *helper, *last;
15953 	dtrace_actdesc_t *act;
15954 	dtrace_vstate_t *vstate;
15955 	dtrace_predicate_t *pred;
15956 	int count = 0, nactions = 0, i;
15957 
15958 	if (which < 0 || which >= DTRACE_NHELPER_ACTIONS)
15959 		return (EINVAL);
15960 
15961 	last = help->dthps_actions[which];
15962 	vstate = &help->dthps_vstate;
15963 
15964 	for (count = 0; last != NULL; last = last->dtha_next) {
15965 		count++;
15966 		if (last->dtha_next == NULL)
15967 			break;
15968 	}
15969 
15970 	/*
15971 	 * If we already have dtrace_helper_actions_max helper actions for this
15972 	 * helper action type, we'll refuse to add a new one.
15973 	 */
15974 	if (count >= dtrace_helper_actions_max)
15975 		return (ENOSPC);
15976 
15977 	helper = kmem_zalloc(sizeof (dtrace_helper_action_t), KM_SLEEP);
15978 	helper->dtha_generation = help->dthps_generation;
15979 
15980 	if ((pred = ep->dted_pred.dtpdd_predicate) != NULL) {
15981 		ASSERT(pred->dtp_difo != NULL);
15982 		dtrace_difo_hold(pred->dtp_difo);
15983 		helper->dtha_predicate = pred->dtp_difo;
15984 	}
15985 
15986 	for (act = ep->dted_action; act != NULL; act = act->dtad_next) {
15987 		if (act->dtad_kind != DTRACEACT_DIFEXPR)
15988 			goto err;
15989 
15990 		if (act->dtad_difo == NULL)
15991 			goto err;
15992 
15993 		nactions++;
15994 	}
15995 
15996 	helper->dtha_actions = kmem_zalloc(sizeof (dtrace_difo_t *) *
15997 	    (helper->dtha_nactions = nactions), KM_SLEEP);
15998 
15999 	for (act = ep->dted_action, i = 0; act != NULL; act = act->dtad_next) {
16000 		dtrace_difo_hold(act->dtad_difo);
16001 		helper->dtha_actions[i++] = act->dtad_difo;
16002 	}
16003 
16004 	if (!dtrace_helper_validate(helper))
16005 		goto err;
16006 
16007 	if (last == NULL) {
16008 		help->dthps_actions[which] = helper;
16009 	} else {
16010 		last->dtha_next = helper;
16011 	}
16012 
16013 	if (vstate->dtvs_nlocals > dtrace_helptrace_nlocals) {
16014 		dtrace_helptrace_nlocals = vstate->dtvs_nlocals;
16015 		dtrace_helptrace_next = 0;
16016 	}
16017 
16018 	return (0);
16019 err:
16020 	dtrace_helper_action_destroy(helper, vstate);
16021 	return (EINVAL);
16022 }
16023 
16024 static void
16025 dtrace_helper_provider_register(proc_t *p, dtrace_helpers_t *help,
16026     dof_helper_t *dofhp)
16027 {
16028 	ASSERT(MUTEX_NOT_HELD(&dtrace_lock));
16029 
16030 	mutex_enter(&dtrace_meta_lock);
16031 	mutex_enter(&dtrace_lock);
16032 
16033 	if (!dtrace_attached() || dtrace_meta_pid == NULL) {
16034 		/*
16035 		 * If the dtrace module is loaded but not attached, or if
16036 		 * there aren't isn't a meta provider registered to deal with
16037 		 * these provider descriptions, we need to postpone creating
16038 		 * the actual providers until later.
16039 		 */
16040 
16041 		if (help->dthps_next == NULL && help->dthps_prev == NULL &&
16042 		    dtrace_deferred_pid != help) {
16043 			help->dthps_deferred = 1;
16044 			help->dthps_pid = p->p_pid;
16045 			help->dthps_next = dtrace_deferred_pid;
16046 			help->dthps_prev = NULL;
16047 			if (dtrace_deferred_pid != NULL)
16048 				dtrace_deferred_pid->dthps_prev = help;
16049 			dtrace_deferred_pid = help;
16050 		}
16051 
16052 		mutex_exit(&dtrace_lock);
16053 
16054 	} else if (dofhp != NULL) {
16055 		/*
16056 		 * If the dtrace module is loaded and we have a particular
16057 		 * helper provider description, pass that off to the
16058 		 * meta provider.
16059 		 */
16060 
16061 		mutex_exit(&dtrace_lock);
16062 
16063 		dtrace_helper_provide(dofhp, p->p_pid);
16064 
16065 	} else {
16066 		/*
16067 		 * Otherwise, just pass all the helper provider descriptions
16068 		 * off to the meta provider.
16069 		 */
16070 
16071 		int i;
16072 		mutex_exit(&dtrace_lock);
16073 
16074 		for (i = 0; i < help->dthps_nprovs; i++) {
16075 			dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
16076 			    p->p_pid);
16077 		}
16078 	}
16079 
16080 	mutex_exit(&dtrace_meta_lock);
16081 }
16082 
16083 static int
16084 dtrace_helper_provider_add(dof_helper_t *dofhp, dtrace_helpers_t *help, int gen)
16085 {
16086 	dtrace_helper_provider_t *hprov, **tmp_provs;
16087 	uint_t tmp_maxprovs, i;
16088 
16089 	ASSERT(MUTEX_HELD(&dtrace_lock));
16090 	ASSERT(help != NULL);
16091 
16092 	/*
16093 	 * If we already have dtrace_helper_providers_max helper providers,
16094 	 * we're refuse to add a new one.
16095 	 */
16096 	if (help->dthps_nprovs >= dtrace_helper_providers_max)
16097 		return (ENOSPC);
16098 
16099 	/*
16100 	 * Check to make sure this isn't a duplicate.
16101 	 */
16102 	for (i = 0; i < help->dthps_nprovs; i++) {
16103 		if (dofhp->dofhp_addr ==
16104 		    help->dthps_provs[i]->dthp_prov.dofhp_addr)
16105 			return (EALREADY);
16106 	}
16107 
16108 	hprov = kmem_zalloc(sizeof (dtrace_helper_provider_t), KM_SLEEP);
16109 	hprov->dthp_prov = *dofhp;
16110 	hprov->dthp_ref = 1;
16111 	hprov->dthp_generation = gen;
16112 
16113 	/*
16114 	 * Allocate a bigger table for helper providers if it's already full.
16115 	 */
16116 	if (help->dthps_maxprovs == help->dthps_nprovs) {
16117 		tmp_maxprovs = help->dthps_maxprovs;
16118 		tmp_provs = help->dthps_provs;
16119 
16120 		if (help->dthps_maxprovs == 0)
16121 			help->dthps_maxprovs = 2;
16122 		else
16123 			help->dthps_maxprovs *= 2;
16124 		if (help->dthps_maxprovs > dtrace_helper_providers_max)
16125 			help->dthps_maxprovs = dtrace_helper_providers_max;
16126 
16127 		ASSERT(tmp_maxprovs < help->dthps_maxprovs);
16128 
16129 		help->dthps_provs = kmem_zalloc(help->dthps_maxprovs *
16130 		    sizeof (dtrace_helper_provider_t *), KM_SLEEP);
16131 
16132 		if (tmp_provs != NULL) {
16133 			bcopy(tmp_provs, help->dthps_provs, tmp_maxprovs *
16134 			    sizeof (dtrace_helper_provider_t *));
16135 			kmem_free(tmp_provs, tmp_maxprovs *
16136 			    sizeof (dtrace_helper_provider_t *));
16137 		}
16138 	}
16139 
16140 	help->dthps_provs[help->dthps_nprovs] = hprov;
16141 	help->dthps_nprovs++;
16142 
16143 	return (0);
16144 }
16145 
16146 static void
16147 dtrace_helper_provider_destroy(dtrace_helper_provider_t *hprov)
16148 {
16149 	mutex_enter(&dtrace_lock);
16150 
16151 	if (--hprov->dthp_ref == 0) {
16152 		dof_hdr_t *dof;
16153 		mutex_exit(&dtrace_lock);
16154 		dof = (dof_hdr_t *)(uintptr_t)hprov->dthp_prov.dofhp_dof;
16155 		dtrace_dof_destroy(dof);
16156 		kmem_free(hprov, sizeof (dtrace_helper_provider_t));
16157 	} else {
16158 		mutex_exit(&dtrace_lock);
16159 	}
16160 }
16161 
16162 static int
16163 dtrace_helper_provider_validate(dof_hdr_t *dof, dof_sec_t *sec)
16164 {
16165 	uintptr_t daddr = (uintptr_t)dof;
16166 	dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
16167 	dof_provider_t *provider;
16168 	dof_probe_t *probe;
16169 	uint8_t *arg;
16170 	char *strtab, *typestr;
16171 	dof_stridx_t typeidx;
16172 	size_t typesz;
16173 	uint_t nprobes, j, k;
16174 
16175 	ASSERT(sec->dofs_type == DOF_SECT_PROVIDER);
16176 
16177 	if (sec->dofs_offset & (sizeof (uint_t) - 1)) {
16178 		dtrace_dof_error(dof, "misaligned section offset");
16179 		return (-1);
16180 	}
16181 
16182 	/*
16183 	 * The section needs to be large enough to contain the DOF provider
16184 	 * structure appropriate for the given version.
16185 	 */
16186 	if (sec->dofs_size <
16187 	    ((dof->dofh_ident[DOF_ID_VERSION] == DOF_VERSION_1) ?
16188 	    offsetof(dof_provider_t, dofpv_prenoffs) :
16189 	    sizeof (dof_provider_t))) {
16190 		dtrace_dof_error(dof, "provider section too small");
16191 		return (-1);
16192 	}
16193 
16194 	provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
16195 	str_sec = dtrace_dof_sect(dof, DOF_SECT_STRTAB, provider->dofpv_strtab);
16196 	prb_sec = dtrace_dof_sect(dof, DOF_SECT_PROBES, provider->dofpv_probes);
16197 	arg_sec = dtrace_dof_sect(dof, DOF_SECT_PRARGS, provider->dofpv_prargs);
16198 	off_sec = dtrace_dof_sect(dof, DOF_SECT_PROFFS, provider->dofpv_proffs);
16199 
16200 	if (str_sec == NULL || prb_sec == NULL ||
16201 	    arg_sec == NULL || off_sec == NULL)
16202 		return (-1);
16203 
16204 	enoff_sec = NULL;
16205 
16206 	if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
16207 	    provider->dofpv_prenoffs != DOF_SECT_NONE &&
16208 	    (enoff_sec = dtrace_dof_sect(dof, DOF_SECT_PRENOFFS,
16209 	    provider->dofpv_prenoffs)) == NULL)
16210 		return (-1);
16211 
16212 	strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
16213 
16214 	if (provider->dofpv_name >= str_sec->dofs_size ||
16215 	    strlen(strtab + provider->dofpv_name) >= DTRACE_PROVNAMELEN) {
16216 		dtrace_dof_error(dof, "invalid provider name");
16217 		return (-1);
16218 	}
16219 
16220 	if (prb_sec->dofs_entsize == 0 ||
16221 	    prb_sec->dofs_entsize > prb_sec->dofs_size) {
16222 		dtrace_dof_error(dof, "invalid entry size");
16223 		return (-1);
16224 	}
16225 
16226 	if (prb_sec->dofs_entsize & (sizeof (uintptr_t) - 1)) {
16227 		dtrace_dof_error(dof, "misaligned entry size");
16228 		return (-1);
16229 	}
16230 
16231 	if (off_sec->dofs_entsize != sizeof (uint32_t)) {
16232 		dtrace_dof_error(dof, "invalid entry size");
16233 		return (-1);
16234 	}
16235 
16236 	if (off_sec->dofs_offset & (sizeof (uint32_t) - 1)) {
16237 		dtrace_dof_error(dof, "misaligned section offset");
16238 		return (-1);
16239 	}
16240 
16241 	if (arg_sec->dofs_entsize != sizeof (uint8_t)) {
16242 		dtrace_dof_error(dof, "invalid entry size");
16243 		return (-1);
16244 	}
16245 
16246 	arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
16247 
16248 	nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
16249 
16250 	/*
16251 	 * Take a pass through the probes to check for errors.
16252 	 */
16253 	for (j = 0; j < nprobes; j++) {
16254 		probe = (dof_probe_t *)(uintptr_t)(daddr +
16255 		    prb_sec->dofs_offset + j * prb_sec->dofs_entsize);
16256 
16257 		if (probe->dofpr_func >= str_sec->dofs_size) {
16258 			dtrace_dof_error(dof, "invalid function name");
16259 			return (-1);
16260 		}
16261 
16262 		if (strlen(strtab + probe->dofpr_func) >= DTRACE_FUNCNAMELEN) {
16263 			dtrace_dof_error(dof, "function name too long");
16264 			/*
16265 			 * Keep going if the function name is too long.
16266 			 * Unlike provider and probe names, we cannot reasonably
16267 			 * impose restrictions on function names, since they're
16268 			 * a property of the code being instrumented. We will
16269 			 * skip this probe in dtrace_helper_provide_one().
16270 			 */
16271 		}
16272 
16273 		if (probe->dofpr_name >= str_sec->dofs_size ||
16274 		    strlen(strtab + probe->dofpr_name) >= DTRACE_NAMELEN) {
16275 			dtrace_dof_error(dof, "invalid probe name");
16276 			return (-1);
16277 		}
16278 
16279 		/*
16280 		 * The offset count must not wrap the index, and the offsets
16281 		 * must also not overflow the section's data.
16282 		 */
16283 		if (probe->dofpr_offidx + probe->dofpr_noffs <
16284 		    probe->dofpr_offidx ||
16285 		    (probe->dofpr_offidx + probe->dofpr_noffs) *
16286 		    off_sec->dofs_entsize > off_sec->dofs_size) {
16287 			dtrace_dof_error(dof, "invalid probe offset");
16288 			return (-1);
16289 		}
16290 
16291 		if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1) {
16292 			/*
16293 			 * If there's no is-enabled offset section, make sure
16294 			 * there aren't any is-enabled offsets. Otherwise
16295 			 * perform the same checks as for probe offsets
16296 			 * (immediately above).
16297 			 */
16298 			if (enoff_sec == NULL) {
16299 				if (probe->dofpr_enoffidx != 0 ||
16300 				    probe->dofpr_nenoffs != 0) {
16301 					dtrace_dof_error(dof, "is-enabled "
16302 					    "offsets with null section");
16303 					return (-1);
16304 				}
16305 			} else if (probe->dofpr_enoffidx +
16306 			    probe->dofpr_nenoffs < probe->dofpr_enoffidx ||
16307 			    (probe->dofpr_enoffidx + probe->dofpr_nenoffs) *
16308 			    enoff_sec->dofs_entsize > enoff_sec->dofs_size) {
16309 				dtrace_dof_error(dof, "invalid is-enabled "
16310 				    "offset");
16311 				return (-1);
16312 			}
16313 
16314 			if (probe->dofpr_noffs + probe->dofpr_nenoffs == 0) {
16315 				dtrace_dof_error(dof, "zero probe and "
16316 				    "is-enabled offsets");
16317 				return (-1);
16318 			}
16319 		} else if (probe->dofpr_noffs == 0) {
16320 			dtrace_dof_error(dof, "zero probe offsets");
16321 			return (-1);
16322 		}
16323 
16324 		if (probe->dofpr_argidx + probe->dofpr_xargc <
16325 		    probe->dofpr_argidx ||
16326 		    (probe->dofpr_argidx + probe->dofpr_xargc) *
16327 		    arg_sec->dofs_entsize > arg_sec->dofs_size) {
16328 			dtrace_dof_error(dof, "invalid args");
16329 			return (-1);
16330 		}
16331 
16332 		typeidx = probe->dofpr_nargv;
16333 		typestr = strtab + probe->dofpr_nargv;
16334 		for (k = 0; k < probe->dofpr_nargc; k++) {
16335 			if (typeidx >= str_sec->dofs_size) {
16336 				dtrace_dof_error(dof, "bad "
16337 				    "native argument type");
16338 				return (-1);
16339 			}
16340 
16341 			typesz = strlen(typestr) + 1;
16342 			if (typesz > DTRACE_ARGTYPELEN) {
16343 				dtrace_dof_error(dof, "native "
16344 				    "argument type too long");
16345 				return (-1);
16346 			}
16347 			typeidx += typesz;
16348 			typestr += typesz;
16349 		}
16350 
16351 		typeidx = probe->dofpr_xargv;
16352 		typestr = strtab + probe->dofpr_xargv;
16353 		for (k = 0; k < probe->dofpr_xargc; k++) {
16354 			if (arg[probe->dofpr_argidx + k] > probe->dofpr_nargc) {
16355 				dtrace_dof_error(dof, "bad "
16356 				    "native argument index");
16357 				return (-1);
16358 			}
16359 
16360 			if (typeidx >= str_sec->dofs_size) {
16361 				dtrace_dof_error(dof, "bad "
16362 				    "translated argument type");
16363 				return (-1);
16364 			}
16365 
16366 			typesz = strlen(typestr) + 1;
16367 			if (typesz > DTRACE_ARGTYPELEN) {
16368 				dtrace_dof_error(dof, "translated argument "
16369 				    "type too long");
16370 				return (-1);
16371 			}
16372 
16373 			typeidx += typesz;
16374 			typestr += typesz;
16375 		}
16376 	}
16377 
16378 	return (0);
16379 }
16380 
16381 static int
16382 dtrace_helper_slurp(dof_hdr_t *dof, dof_helper_t *dhp, struct proc *p)
16383 {
16384 	dtrace_helpers_t *help;
16385 	dtrace_vstate_t *vstate;
16386 	dtrace_enabling_t *enab = NULL;
16387 	int i, gen, rv, nhelpers = 0, nprovs = 0, destroy = 1;
16388 	uintptr_t daddr = (uintptr_t)dof;
16389 
16390 	ASSERT(MUTEX_HELD(&dtrace_lock));
16391 
16392 	if ((help = p->p_dtrace_helpers) == NULL)
16393 		help = dtrace_helpers_create(p);
16394 
16395 	vstate = &help->dthps_vstate;
16396 
16397 	if ((rv = dtrace_dof_slurp(dof, vstate, NULL, &enab, dhp->dofhp_addr,
16398 	    dhp->dofhp_dof, B_FALSE)) != 0) {
16399 		dtrace_dof_destroy(dof);
16400 		return (rv);
16401 	}
16402 
16403 	/*
16404 	 * Look for helper providers and validate their descriptions.
16405 	 */
16406 	for (i = 0; i < dof->dofh_secnum; i++) {
16407 		dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
16408 		    dof->dofh_secoff + i * dof->dofh_secsize);
16409 
16410 		if (sec->dofs_type != DOF_SECT_PROVIDER)
16411 			continue;
16412 
16413 		if (dtrace_helper_provider_validate(dof, sec) != 0) {
16414 			dtrace_enabling_destroy(enab);
16415 			dtrace_dof_destroy(dof);
16416 			return (-1);
16417 		}
16418 
16419 		nprovs++;
16420 	}
16421 
16422 	/*
16423 	 * Now we need to walk through the ECB descriptions in the enabling.
16424 	 */
16425 	for (i = 0; i < enab->dten_ndesc; i++) {
16426 		dtrace_ecbdesc_t *ep = enab->dten_desc[i];
16427 		dtrace_probedesc_t *desc = &ep->dted_probe;
16428 
16429 		if (strcmp(desc->dtpd_provider, "dtrace") != 0)
16430 			continue;
16431 
16432 		if (strcmp(desc->dtpd_mod, "helper") != 0)
16433 			continue;
16434 
16435 		if (strcmp(desc->dtpd_func, "ustack") != 0)
16436 			continue;
16437 
16438 		if ((rv = dtrace_helper_action_add(DTRACE_HELPER_ACTION_USTACK,
16439 		    ep, help)) != 0) {
16440 			/*
16441 			 * Adding this helper action failed -- we are now going
16442 			 * to rip out the entire generation and return failure.
16443 			 */
16444 			(void) dtrace_helper_destroygen(help,
16445 			    help->dthps_generation);
16446 			dtrace_enabling_destroy(enab);
16447 			dtrace_dof_destroy(dof);
16448 			return (-1);
16449 		}
16450 
16451 		nhelpers++;
16452 	}
16453 
16454 	if (nhelpers < enab->dten_ndesc)
16455 		dtrace_dof_error(dof, "unmatched helpers");
16456 
16457 	gen = help->dthps_generation++;
16458 	dtrace_enabling_destroy(enab);
16459 
16460 	if (nprovs > 0) {
16461 		/*
16462 		 * Now that this is in-kernel, we change the sense of the
16463 		 * members:  dofhp_dof denotes the in-kernel copy of the DOF
16464 		 * and dofhp_addr denotes the address at user-level.
16465 		 */
16466 		dhp->dofhp_addr = dhp->dofhp_dof;
16467 		dhp->dofhp_dof = (uint64_t)(uintptr_t)dof;
16468 
16469 		if (dtrace_helper_provider_add(dhp, help, gen) == 0) {
16470 			mutex_exit(&dtrace_lock);
16471 			dtrace_helper_provider_register(p, help, dhp);
16472 			mutex_enter(&dtrace_lock);
16473 
16474 			destroy = 0;
16475 		}
16476 	}
16477 
16478 	if (destroy)
16479 		dtrace_dof_destroy(dof);
16480 
16481 	return (gen);
16482 }
16483 
16484 static dtrace_helpers_t *
16485 dtrace_helpers_create(proc_t *p)
16486 {
16487 	dtrace_helpers_t *help;
16488 
16489 	ASSERT(MUTEX_HELD(&dtrace_lock));
16490 	ASSERT(p->p_dtrace_helpers == NULL);
16491 
16492 	help = kmem_zalloc(sizeof (dtrace_helpers_t), KM_SLEEP);
16493 	help->dthps_actions = kmem_zalloc(sizeof (dtrace_helper_action_t *) *
16494 	    DTRACE_NHELPER_ACTIONS, KM_SLEEP);
16495 
16496 	p->p_dtrace_helpers = help;
16497 	dtrace_helpers++;
16498 
16499 	return (help);
16500 }
16501 
16502 #ifdef illumos
16503 static
16504 #endif
16505 void
16506 dtrace_helpers_destroy(proc_t *p)
16507 {
16508 	dtrace_helpers_t *help;
16509 	dtrace_vstate_t *vstate;
16510 #ifdef illumos
16511 	proc_t *p = curproc;
16512 #endif
16513 	int i;
16514 
16515 	mutex_enter(&dtrace_lock);
16516 
16517 	ASSERT(p->p_dtrace_helpers != NULL);
16518 	ASSERT(dtrace_helpers > 0);
16519 
16520 	help = p->p_dtrace_helpers;
16521 	vstate = &help->dthps_vstate;
16522 
16523 	/*
16524 	 * We're now going to lose the help from this process.
16525 	 */
16526 	p->p_dtrace_helpers = NULL;
16527 	dtrace_sync();
16528 
16529 	/*
16530 	 * Destory the helper actions.
16531 	 */
16532 	for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
16533 		dtrace_helper_action_t *h, *next;
16534 
16535 		for (h = help->dthps_actions[i]; h != NULL; h = next) {
16536 			next = h->dtha_next;
16537 			dtrace_helper_action_destroy(h, vstate);
16538 			h = next;
16539 		}
16540 	}
16541 
16542 	mutex_exit(&dtrace_lock);
16543 
16544 	/*
16545 	 * Destroy the helper providers.
16546 	 */
16547 	if (help->dthps_maxprovs > 0) {
16548 		mutex_enter(&dtrace_meta_lock);
16549 		if (dtrace_meta_pid != NULL) {
16550 			ASSERT(dtrace_deferred_pid == NULL);
16551 
16552 			for (i = 0; i < help->dthps_nprovs; i++) {
16553 				dtrace_helper_provider_remove(
16554 				    &help->dthps_provs[i]->dthp_prov, p->p_pid);
16555 			}
16556 		} else {
16557 			mutex_enter(&dtrace_lock);
16558 			ASSERT(help->dthps_deferred == 0 ||
16559 			    help->dthps_next != NULL ||
16560 			    help->dthps_prev != NULL ||
16561 			    help == dtrace_deferred_pid);
16562 
16563 			/*
16564 			 * Remove the helper from the deferred list.
16565 			 */
16566 			if (help->dthps_next != NULL)
16567 				help->dthps_next->dthps_prev = help->dthps_prev;
16568 			if (help->dthps_prev != NULL)
16569 				help->dthps_prev->dthps_next = help->dthps_next;
16570 			if (dtrace_deferred_pid == help) {
16571 				dtrace_deferred_pid = help->dthps_next;
16572 				ASSERT(help->dthps_prev == NULL);
16573 			}
16574 
16575 			mutex_exit(&dtrace_lock);
16576 		}
16577 
16578 		mutex_exit(&dtrace_meta_lock);
16579 
16580 		for (i = 0; i < help->dthps_nprovs; i++) {
16581 			dtrace_helper_provider_destroy(help->dthps_provs[i]);
16582 		}
16583 
16584 		kmem_free(help->dthps_provs, help->dthps_maxprovs *
16585 		    sizeof (dtrace_helper_provider_t *));
16586 	}
16587 
16588 	mutex_enter(&dtrace_lock);
16589 
16590 	dtrace_vstate_fini(&help->dthps_vstate);
16591 	kmem_free(help->dthps_actions,
16592 	    sizeof (dtrace_helper_action_t *) * DTRACE_NHELPER_ACTIONS);
16593 	kmem_free(help, sizeof (dtrace_helpers_t));
16594 
16595 	--dtrace_helpers;
16596 	mutex_exit(&dtrace_lock);
16597 }
16598 
16599 #ifdef illumos
16600 static
16601 #endif
16602 void
16603 dtrace_helpers_duplicate(proc_t *from, proc_t *to)
16604 {
16605 	dtrace_helpers_t *help, *newhelp;
16606 	dtrace_helper_action_t *helper, *new, *last;
16607 	dtrace_difo_t *dp;
16608 	dtrace_vstate_t *vstate;
16609 	int i, j, sz, hasprovs = 0;
16610 
16611 	mutex_enter(&dtrace_lock);
16612 	ASSERT(from->p_dtrace_helpers != NULL);
16613 	ASSERT(dtrace_helpers > 0);
16614 
16615 	help = from->p_dtrace_helpers;
16616 	newhelp = dtrace_helpers_create(to);
16617 	ASSERT(to->p_dtrace_helpers != NULL);
16618 
16619 	newhelp->dthps_generation = help->dthps_generation;
16620 	vstate = &newhelp->dthps_vstate;
16621 
16622 	/*
16623 	 * Duplicate the helper actions.
16624 	 */
16625 	for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
16626 		if ((helper = help->dthps_actions[i]) == NULL)
16627 			continue;
16628 
16629 		for (last = NULL; helper != NULL; helper = helper->dtha_next) {
16630 			new = kmem_zalloc(sizeof (dtrace_helper_action_t),
16631 			    KM_SLEEP);
16632 			new->dtha_generation = helper->dtha_generation;
16633 
16634 			if ((dp = helper->dtha_predicate) != NULL) {
16635 				dp = dtrace_difo_duplicate(dp, vstate);
16636 				new->dtha_predicate = dp;
16637 			}
16638 
16639 			new->dtha_nactions = helper->dtha_nactions;
16640 			sz = sizeof (dtrace_difo_t *) * new->dtha_nactions;
16641 			new->dtha_actions = kmem_alloc(sz, KM_SLEEP);
16642 
16643 			for (j = 0; j < new->dtha_nactions; j++) {
16644 				dtrace_difo_t *dp = helper->dtha_actions[j];
16645 
16646 				ASSERT(dp != NULL);
16647 				dp = dtrace_difo_duplicate(dp, vstate);
16648 				new->dtha_actions[j] = dp;
16649 			}
16650 
16651 			if (last != NULL) {
16652 				last->dtha_next = new;
16653 			} else {
16654 				newhelp->dthps_actions[i] = new;
16655 			}
16656 
16657 			last = new;
16658 		}
16659 	}
16660 
16661 	/*
16662 	 * Duplicate the helper providers and register them with the
16663 	 * DTrace framework.
16664 	 */
16665 	if (help->dthps_nprovs > 0) {
16666 		newhelp->dthps_nprovs = help->dthps_nprovs;
16667 		newhelp->dthps_maxprovs = help->dthps_nprovs;
16668 		newhelp->dthps_provs = kmem_alloc(newhelp->dthps_nprovs *
16669 		    sizeof (dtrace_helper_provider_t *), KM_SLEEP);
16670 		for (i = 0; i < newhelp->dthps_nprovs; i++) {
16671 			newhelp->dthps_provs[i] = help->dthps_provs[i];
16672 			newhelp->dthps_provs[i]->dthp_ref++;
16673 		}
16674 
16675 		hasprovs = 1;
16676 	}
16677 
16678 	mutex_exit(&dtrace_lock);
16679 
16680 	if (hasprovs)
16681 		dtrace_helper_provider_register(to, newhelp, NULL);
16682 }
16683 
16684 /*
16685  * DTrace Hook Functions
16686  */
16687 static void
16688 dtrace_module_loaded(modctl_t *ctl)
16689 {
16690 	dtrace_provider_t *prv;
16691 
16692 	mutex_enter(&dtrace_provider_lock);
16693 #ifdef illumos
16694 	mutex_enter(&mod_lock);
16695 #endif
16696 
16697 #ifdef illumos
16698 	ASSERT(ctl->mod_busy);
16699 #endif
16700 
16701 	/*
16702 	 * We're going to call each providers per-module provide operation
16703 	 * specifying only this module.
16704 	 */
16705 	for (prv = dtrace_provider; prv != NULL; prv = prv->dtpv_next)
16706 		prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
16707 
16708 #ifdef illumos
16709 	mutex_exit(&mod_lock);
16710 #endif
16711 	mutex_exit(&dtrace_provider_lock);
16712 
16713 	/*
16714 	 * If we have any retained enablings, we need to match against them.
16715 	 * Enabling probes requires that cpu_lock be held, and we cannot hold
16716 	 * cpu_lock here -- it is legal for cpu_lock to be held when loading a
16717 	 * module.  (In particular, this happens when loading scheduling
16718 	 * classes.)  So if we have any retained enablings, we need to dispatch
16719 	 * our task queue to do the match for us.
16720 	 */
16721 	mutex_enter(&dtrace_lock);
16722 
16723 	if (dtrace_retained == NULL) {
16724 		mutex_exit(&dtrace_lock);
16725 		return;
16726 	}
16727 
16728 	(void) taskq_dispatch(dtrace_taskq,
16729 	    (task_func_t *)dtrace_enabling_matchall, NULL, TQ_SLEEP);
16730 
16731 	mutex_exit(&dtrace_lock);
16732 
16733 	/*
16734 	 * And now, for a little heuristic sleaze:  in general, we want to
16735 	 * match modules as soon as they load.  However, we cannot guarantee
16736 	 * this, because it would lead us to the lock ordering violation
16737 	 * outlined above.  The common case, of course, is that cpu_lock is
16738 	 * _not_ held -- so we delay here for a clock tick, hoping that that's
16739 	 * long enough for the task queue to do its work.  If it's not, it's
16740 	 * not a serious problem -- it just means that the module that we
16741 	 * just loaded may not be immediately instrumentable.
16742 	 */
16743 	delay(1);
16744 }
16745 
16746 static void
16747 #ifdef illumos
16748 dtrace_module_unloaded(modctl_t *ctl)
16749 #else
16750 dtrace_module_unloaded(modctl_t *ctl, int *error)
16751 #endif
16752 {
16753 	dtrace_probe_t template, *probe, *first, *next;
16754 	dtrace_provider_t *prov;
16755 #ifndef illumos
16756 	char modname[DTRACE_MODNAMELEN];
16757 	size_t len;
16758 #endif
16759 
16760 #ifdef illumos
16761 	template.dtpr_mod = ctl->mod_modname;
16762 #else
16763 	/* Handle the fact that ctl->filename may end in ".ko". */
16764 	strlcpy(modname, ctl->filename, sizeof(modname));
16765 	len = strlen(ctl->filename);
16766 	if (len > 3 && strcmp(modname + len - 3, ".ko") == 0)
16767 		modname[len - 3] = '\0';
16768 	template.dtpr_mod = modname;
16769 #endif
16770 
16771 	mutex_enter(&dtrace_provider_lock);
16772 #ifdef illumos
16773 	mutex_enter(&mod_lock);
16774 #endif
16775 	mutex_enter(&dtrace_lock);
16776 
16777 #ifndef illumos
16778 	if (ctl->nenabled > 0) {
16779 		/* Don't allow unloads if a probe is enabled. */
16780 		mutex_exit(&dtrace_provider_lock);
16781 		mutex_exit(&dtrace_lock);
16782 		*error = -1;
16783 		printf(
16784 	"kldunload: attempt to unload module that has DTrace probes enabled\n");
16785 		return;
16786 	}
16787 #endif
16788 
16789 	if (dtrace_bymod == NULL) {
16790 		/*
16791 		 * The DTrace module is loaded (obviously) but not attached;
16792 		 * we don't have any work to do.
16793 		 */
16794 		mutex_exit(&dtrace_provider_lock);
16795 #ifdef illumos
16796 		mutex_exit(&mod_lock);
16797 #endif
16798 		mutex_exit(&dtrace_lock);
16799 		return;
16800 	}
16801 
16802 	for (probe = first = dtrace_hash_lookup(dtrace_bymod, &template);
16803 	    probe != NULL; probe = probe->dtpr_nextmod) {
16804 		if (probe->dtpr_ecb != NULL) {
16805 			mutex_exit(&dtrace_provider_lock);
16806 #ifdef illumos
16807 			mutex_exit(&mod_lock);
16808 #endif
16809 			mutex_exit(&dtrace_lock);
16810 
16811 			/*
16812 			 * This shouldn't _actually_ be possible -- we're
16813 			 * unloading a module that has an enabled probe in it.
16814 			 * (It's normally up to the provider to make sure that
16815 			 * this can't happen.)  However, because dtps_enable()
16816 			 * doesn't have a failure mode, there can be an
16817 			 * enable/unload race.  Upshot:  we don't want to
16818 			 * assert, but we're not going to disable the
16819 			 * probe, either.
16820 			 */
16821 			if (dtrace_err_verbose) {
16822 #ifdef illumos
16823 				cmn_err(CE_WARN, "unloaded module '%s' had "
16824 				    "enabled probes", ctl->mod_modname);
16825 #else
16826 				cmn_err(CE_WARN, "unloaded module '%s' had "
16827 				    "enabled probes", modname);
16828 #endif
16829 			}
16830 
16831 			return;
16832 		}
16833 	}
16834 
16835 	probe = first;
16836 
16837 	for (first = NULL; probe != NULL; probe = next) {
16838 		ASSERT(dtrace_probes[probe->dtpr_id - 1] == probe);
16839 
16840 		dtrace_probes[probe->dtpr_id - 1] = NULL;
16841 
16842 		next = probe->dtpr_nextmod;
16843 		dtrace_hash_remove(dtrace_bymod, probe);
16844 		dtrace_hash_remove(dtrace_byfunc, probe);
16845 		dtrace_hash_remove(dtrace_byname, probe);
16846 
16847 		if (first == NULL) {
16848 			first = probe;
16849 			probe->dtpr_nextmod = NULL;
16850 		} else {
16851 			probe->dtpr_nextmod = first;
16852 			first = probe;
16853 		}
16854 	}
16855 
16856 	/*
16857 	 * We've removed all of the module's probes from the hash chains and
16858 	 * from the probe array.  Now issue a dtrace_sync() to be sure that
16859 	 * everyone has cleared out from any probe array processing.
16860 	 */
16861 	dtrace_sync();
16862 
16863 	for (probe = first; probe != NULL; probe = first) {
16864 		first = probe->dtpr_nextmod;
16865 		prov = probe->dtpr_provider;
16866 		prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, probe->dtpr_id,
16867 		    probe->dtpr_arg);
16868 		kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
16869 		kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
16870 		kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
16871 #ifdef illumos
16872 		vmem_free(dtrace_arena, (void *)(uintptr_t)probe->dtpr_id, 1);
16873 #else
16874 		free_unr(dtrace_arena, probe->dtpr_id);
16875 #endif
16876 		kmem_free(probe, sizeof (dtrace_probe_t));
16877 	}
16878 
16879 	mutex_exit(&dtrace_lock);
16880 #ifdef illumos
16881 	mutex_exit(&mod_lock);
16882 #endif
16883 	mutex_exit(&dtrace_provider_lock);
16884 }
16885 
16886 #ifndef illumos
16887 static void
16888 dtrace_kld_load(void *arg __unused, linker_file_t lf)
16889 {
16890 
16891 	dtrace_module_loaded(lf);
16892 }
16893 
16894 static void
16895 dtrace_kld_unload_try(void *arg __unused, linker_file_t lf, int *error)
16896 {
16897 
16898 	if (*error != 0)
16899 		/* We already have an error, so don't do anything. */
16900 		return;
16901 	dtrace_module_unloaded(lf, error);
16902 }
16903 #endif
16904 
16905 #ifdef illumos
16906 static void
16907 dtrace_suspend(void)
16908 {
16909 	dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_suspend));
16910 }
16911 
16912 static void
16913 dtrace_resume(void)
16914 {
16915 	dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_resume));
16916 }
16917 #endif
16918 
16919 static int
16920 dtrace_cpu_setup(cpu_setup_t what, processorid_t cpu)
16921 {
16922 	ASSERT(MUTEX_HELD(&cpu_lock));
16923 	mutex_enter(&dtrace_lock);
16924 
16925 	switch (what) {
16926 	case CPU_CONFIG: {
16927 		dtrace_state_t *state;
16928 		dtrace_optval_t *opt, rs, c;
16929 
16930 		/*
16931 		 * For now, we only allocate a new buffer for anonymous state.
16932 		 */
16933 		if ((state = dtrace_anon.dta_state) == NULL)
16934 			break;
16935 
16936 		if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
16937 			break;
16938 
16939 		opt = state->dts_options;
16940 		c = opt[DTRACEOPT_CPU];
16941 
16942 		if (c != DTRACE_CPUALL && c != DTRACEOPT_UNSET && c != cpu)
16943 			break;
16944 
16945 		/*
16946 		 * Regardless of what the actual policy is, we're going to
16947 		 * temporarily set our resize policy to be manual.  We're
16948 		 * also going to temporarily set our CPU option to denote
16949 		 * the newly configured CPU.
16950 		 */
16951 		rs = opt[DTRACEOPT_BUFRESIZE];
16952 		opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_MANUAL;
16953 		opt[DTRACEOPT_CPU] = (dtrace_optval_t)cpu;
16954 
16955 		(void) dtrace_state_buffers(state);
16956 
16957 		opt[DTRACEOPT_BUFRESIZE] = rs;
16958 		opt[DTRACEOPT_CPU] = c;
16959 
16960 		break;
16961 	}
16962 
16963 	case CPU_UNCONFIG:
16964 		/*
16965 		 * We don't free the buffer in the CPU_UNCONFIG case.  (The
16966 		 * buffer will be freed when the consumer exits.)
16967 		 */
16968 		break;
16969 
16970 	default:
16971 		break;
16972 	}
16973 
16974 	mutex_exit(&dtrace_lock);
16975 	return (0);
16976 }
16977 
16978 #ifdef illumos
16979 static void
16980 dtrace_cpu_setup_initial(processorid_t cpu)
16981 {
16982 	(void) dtrace_cpu_setup(CPU_CONFIG, cpu);
16983 }
16984 #endif
16985 
16986 static void
16987 dtrace_toxrange_add(uintptr_t base, uintptr_t limit)
16988 {
16989 	if (dtrace_toxranges >= dtrace_toxranges_max) {
16990 		int osize, nsize;
16991 		dtrace_toxrange_t *range;
16992 
16993 		osize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
16994 
16995 		if (osize == 0) {
16996 			ASSERT(dtrace_toxrange == NULL);
16997 			ASSERT(dtrace_toxranges_max == 0);
16998 			dtrace_toxranges_max = 1;
16999 		} else {
17000 			dtrace_toxranges_max <<= 1;
17001 		}
17002 
17003 		nsize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
17004 		range = kmem_zalloc(nsize, KM_SLEEP);
17005 
17006 		if (dtrace_toxrange != NULL) {
17007 			ASSERT(osize != 0);
17008 			bcopy(dtrace_toxrange, range, osize);
17009 			kmem_free(dtrace_toxrange, osize);
17010 		}
17011 
17012 		dtrace_toxrange = range;
17013 	}
17014 
17015 	ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_base == 0);
17016 	ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_limit == 0);
17017 
17018 	dtrace_toxrange[dtrace_toxranges].dtt_base = base;
17019 	dtrace_toxrange[dtrace_toxranges].dtt_limit = limit;
17020 	dtrace_toxranges++;
17021 }
17022 
17023 static void
17024 dtrace_getf_barrier(void)
17025 {
17026 #ifdef illumos
17027 	/*
17028 	 * When we have unprivileged (that is, non-DTRACE_CRV_KERNEL) enablings
17029 	 * that contain calls to getf(), this routine will be called on every
17030 	 * closef() before either the underlying vnode is released or the
17031 	 * file_t itself is freed.  By the time we are here, it is essential
17032 	 * that the file_t can no longer be accessed from a call to getf()
17033 	 * in probe context -- that assures that a dtrace_sync() can be used
17034 	 * to clear out any enablings referring to the old structures.
17035 	 */
17036 	if (curthread->t_procp->p_zone->zone_dtrace_getf != 0 ||
17037 	    kcred->cr_zone->zone_dtrace_getf != 0)
17038 		dtrace_sync();
17039 #endif
17040 }
17041 
17042 /*
17043  * DTrace Driver Cookbook Functions
17044  */
17045 #ifdef illumos
17046 /*ARGSUSED*/
17047 static int
17048 dtrace_attach(dev_info_t *devi, ddi_attach_cmd_t cmd)
17049 {
17050 	dtrace_provider_id_t id;
17051 	dtrace_state_t *state = NULL;
17052 	dtrace_enabling_t *enab;
17053 
17054 	mutex_enter(&cpu_lock);
17055 	mutex_enter(&dtrace_provider_lock);
17056 	mutex_enter(&dtrace_lock);
17057 
17058 	if (ddi_soft_state_init(&dtrace_softstate,
17059 	    sizeof (dtrace_state_t), 0) != 0) {
17060 		cmn_err(CE_NOTE, "/dev/dtrace failed to initialize soft state");
17061 		mutex_exit(&cpu_lock);
17062 		mutex_exit(&dtrace_provider_lock);
17063 		mutex_exit(&dtrace_lock);
17064 		return (DDI_FAILURE);
17065 	}
17066 
17067 	if (ddi_create_minor_node(devi, DTRACEMNR_DTRACE, S_IFCHR,
17068 	    DTRACEMNRN_DTRACE, DDI_PSEUDO, NULL) == DDI_FAILURE ||
17069 	    ddi_create_minor_node(devi, DTRACEMNR_HELPER, S_IFCHR,
17070 	    DTRACEMNRN_HELPER, DDI_PSEUDO, NULL) == DDI_FAILURE) {
17071 		cmn_err(CE_NOTE, "/dev/dtrace couldn't create minor nodes");
17072 		ddi_remove_minor_node(devi, NULL);
17073 		ddi_soft_state_fini(&dtrace_softstate);
17074 		mutex_exit(&cpu_lock);
17075 		mutex_exit(&dtrace_provider_lock);
17076 		mutex_exit(&dtrace_lock);
17077 		return (DDI_FAILURE);
17078 	}
17079 
17080 	ddi_report_dev(devi);
17081 	dtrace_devi = devi;
17082 
17083 	dtrace_modload = dtrace_module_loaded;
17084 	dtrace_modunload = dtrace_module_unloaded;
17085 	dtrace_cpu_init = dtrace_cpu_setup_initial;
17086 	dtrace_helpers_cleanup = dtrace_helpers_destroy;
17087 	dtrace_helpers_fork = dtrace_helpers_duplicate;
17088 	dtrace_cpustart_init = dtrace_suspend;
17089 	dtrace_cpustart_fini = dtrace_resume;
17090 	dtrace_debugger_init = dtrace_suspend;
17091 	dtrace_debugger_fini = dtrace_resume;
17092 
17093 	register_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
17094 
17095 	ASSERT(MUTEX_HELD(&cpu_lock));
17096 
17097 	dtrace_arena = vmem_create("dtrace", (void *)1, UINT32_MAX, 1,
17098 	    NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
17099 	dtrace_minor = vmem_create("dtrace_minor", (void *)DTRACEMNRN_CLONE,
17100 	    UINT32_MAX - DTRACEMNRN_CLONE, 1, NULL, NULL, NULL, 0,
17101 	    VM_SLEEP | VMC_IDENTIFIER);
17102 	dtrace_taskq = taskq_create("dtrace_taskq", 1, maxclsyspri,
17103 	    1, INT_MAX, 0);
17104 
17105 	dtrace_state_cache = kmem_cache_create("dtrace_state_cache",
17106 	    sizeof (dtrace_dstate_percpu_t) * NCPU, DTRACE_STATE_ALIGN,
17107 	    NULL, NULL, NULL, NULL, NULL, 0);
17108 
17109 	ASSERT(MUTEX_HELD(&cpu_lock));
17110 	dtrace_bymod = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_mod),
17111 	    offsetof(dtrace_probe_t, dtpr_nextmod),
17112 	    offsetof(dtrace_probe_t, dtpr_prevmod));
17113 
17114 	dtrace_byfunc = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_func),
17115 	    offsetof(dtrace_probe_t, dtpr_nextfunc),
17116 	    offsetof(dtrace_probe_t, dtpr_prevfunc));
17117 
17118 	dtrace_byname = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_name),
17119 	    offsetof(dtrace_probe_t, dtpr_nextname),
17120 	    offsetof(dtrace_probe_t, dtpr_prevname));
17121 
17122 	if (dtrace_retain_max < 1) {
17123 		cmn_err(CE_WARN, "illegal value (%zu) for dtrace_retain_max; "
17124 		    "setting to 1", dtrace_retain_max);
17125 		dtrace_retain_max = 1;
17126 	}
17127 
17128 	/*
17129 	 * Now discover our toxic ranges.
17130 	 */
17131 	dtrace_toxic_ranges(dtrace_toxrange_add);
17132 
17133 	/*
17134 	 * Before we register ourselves as a provider to our own framework,
17135 	 * we would like to assert that dtrace_provider is NULL -- but that's
17136 	 * not true if we were loaded as a dependency of a DTrace provider.
17137 	 * Once we've registered, we can assert that dtrace_provider is our
17138 	 * pseudo provider.
17139 	 */
17140 	(void) dtrace_register("dtrace", &dtrace_provider_attr,
17141 	    DTRACE_PRIV_NONE, 0, &dtrace_provider_ops, NULL, &id);
17142 
17143 	ASSERT(dtrace_provider != NULL);
17144 	ASSERT((dtrace_provider_id_t)dtrace_provider == id);
17145 
17146 	dtrace_probeid_begin = dtrace_probe_create((dtrace_provider_id_t)
17147 	    dtrace_provider, NULL, NULL, "BEGIN", 0, NULL);
17148 	dtrace_probeid_end = dtrace_probe_create((dtrace_provider_id_t)
17149 	    dtrace_provider, NULL, NULL, "END", 0, NULL);
17150 	dtrace_probeid_error = dtrace_probe_create((dtrace_provider_id_t)
17151 	    dtrace_provider, NULL, NULL, "ERROR", 1, NULL);
17152 
17153 	dtrace_anon_property();
17154 	mutex_exit(&cpu_lock);
17155 
17156 	/*
17157 	 * If there are already providers, we must ask them to provide their
17158 	 * probes, and then match any anonymous enabling against them.  Note
17159 	 * that there should be no other retained enablings at this time:
17160 	 * the only retained enablings at this time should be the anonymous
17161 	 * enabling.
17162 	 */
17163 	if (dtrace_anon.dta_enabling != NULL) {
17164 		ASSERT(dtrace_retained == dtrace_anon.dta_enabling);
17165 
17166 		dtrace_enabling_provide(NULL);
17167 		state = dtrace_anon.dta_state;
17168 
17169 		/*
17170 		 * We couldn't hold cpu_lock across the above call to
17171 		 * dtrace_enabling_provide(), but we must hold it to actually
17172 		 * enable the probes.  We have to drop all of our locks, pick
17173 		 * up cpu_lock, and regain our locks before matching the
17174 		 * retained anonymous enabling.
17175 		 */
17176 		mutex_exit(&dtrace_lock);
17177 		mutex_exit(&dtrace_provider_lock);
17178 
17179 		mutex_enter(&cpu_lock);
17180 		mutex_enter(&dtrace_provider_lock);
17181 		mutex_enter(&dtrace_lock);
17182 
17183 		if ((enab = dtrace_anon.dta_enabling) != NULL)
17184 			(void) dtrace_enabling_match(enab, NULL);
17185 
17186 		mutex_exit(&cpu_lock);
17187 	}
17188 
17189 	mutex_exit(&dtrace_lock);
17190 	mutex_exit(&dtrace_provider_lock);
17191 
17192 	if (state != NULL) {
17193 		/*
17194 		 * If we created any anonymous state, set it going now.
17195 		 */
17196 		(void) dtrace_state_go(state, &dtrace_anon.dta_beganon);
17197 	}
17198 
17199 	return (DDI_SUCCESS);
17200 }
17201 #endif	/* illumos */
17202 
17203 #ifndef illumos
17204 static void dtrace_dtr(void *);
17205 #endif
17206 
17207 /*ARGSUSED*/
17208 static int
17209 #ifdef illumos
17210 dtrace_open(dev_t *devp, int flag, int otyp, cred_t *cred_p)
17211 #else
17212 dtrace_open(struct cdev *dev, int oflags, int devtype, struct thread *td)
17213 #endif
17214 {
17215 	dtrace_state_t *state;
17216 	uint32_t priv;
17217 	uid_t uid;
17218 	zoneid_t zoneid;
17219 
17220 #ifdef illumos
17221 	if (getminor(*devp) == DTRACEMNRN_HELPER)
17222 		return (0);
17223 
17224 	/*
17225 	 * If this wasn't an open with the "helper" minor, then it must be
17226 	 * the "dtrace" minor.
17227 	 */
17228 	if (getminor(*devp) == DTRACEMNRN_DTRACE)
17229 		return (ENXIO);
17230 #else
17231 	cred_t *cred_p = NULL;
17232 	cred_p = dev->si_cred;
17233 
17234 	/*
17235 	 * If no DTRACE_PRIV_* bits are set in the credential, then the
17236 	 * caller lacks sufficient permission to do anything with DTrace.
17237 	 */
17238 	dtrace_cred2priv(cred_p, &priv, &uid, &zoneid);
17239 	if (priv == DTRACE_PRIV_NONE) {
17240 #endif
17241 
17242 		return (EACCES);
17243 	}
17244 
17245 	/*
17246 	 * Ask all providers to provide all their probes.
17247 	 */
17248 	mutex_enter(&dtrace_provider_lock);
17249 	dtrace_probe_provide(NULL, NULL);
17250 	mutex_exit(&dtrace_provider_lock);
17251 
17252 	mutex_enter(&cpu_lock);
17253 	mutex_enter(&dtrace_lock);
17254 	dtrace_opens++;
17255 	dtrace_membar_producer();
17256 
17257 #ifdef illumos
17258 	/*
17259 	 * If the kernel debugger is active (that is, if the kernel debugger
17260 	 * modified text in some way), we won't allow the open.
17261 	 */
17262 	if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
17263 		dtrace_opens--;
17264 		mutex_exit(&cpu_lock);
17265 		mutex_exit(&dtrace_lock);
17266 		return (EBUSY);
17267 	}
17268 
17269 	if (dtrace_helptrace_enable && dtrace_helptrace_buffer == NULL) {
17270 		/*
17271 		 * If DTrace helper tracing is enabled, we need to allocate the
17272 		 * trace buffer and initialize the values.
17273 		 */
17274 		dtrace_helptrace_buffer =
17275 		    kmem_zalloc(dtrace_helptrace_bufsize, KM_SLEEP);
17276 		dtrace_helptrace_next = 0;
17277 		dtrace_helptrace_wrapped = 0;
17278 		dtrace_helptrace_enable = 0;
17279 	}
17280 
17281 	state = dtrace_state_create(devp, cred_p);
17282 #else
17283 	state = dtrace_state_create(dev, NULL);
17284 	devfs_set_cdevpriv(state, dtrace_dtr);
17285 #endif
17286 
17287 	mutex_exit(&cpu_lock);
17288 
17289 	if (state == NULL) {
17290 #ifdef illumos
17291 		if (--dtrace_opens == 0 && dtrace_anon.dta_enabling == NULL)
17292 			(void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
17293 #else
17294 		--dtrace_opens;
17295 #endif
17296 		mutex_exit(&dtrace_lock);
17297 		return (EAGAIN);
17298 	}
17299 
17300 	mutex_exit(&dtrace_lock);
17301 
17302 	return (0);
17303 }
17304 
17305 /*ARGSUSED*/
17306 #ifdef illumos
17307 static int
17308 dtrace_close(dev_t dev, int flag, int otyp, cred_t *cred_p)
17309 #else
17310 static void
17311 dtrace_dtr(void *data)
17312 #endif
17313 {
17314 #ifdef illumos
17315 	minor_t minor = getminor(dev);
17316 	dtrace_state_t *state;
17317 #endif
17318 	dtrace_helptrace_t *buf = NULL;
17319 
17320 #ifdef illumos
17321 	if (minor == DTRACEMNRN_HELPER)
17322 		return (0);
17323 
17324 	state = ddi_get_soft_state(dtrace_softstate, minor);
17325 #else
17326 	dtrace_state_t *state = data;
17327 #endif
17328 
17329 	mutex_enter(&cpu_lock);
17330 	mutex_enter(&dtrace_lock);
17331 
17332 #ifdef illumos
17333 	if (state->dts_anon)
17334 #else
17335 	if (state != NULL && state->dts_anon)
17336 #endif
17337 	{
17338 		/*
17339 		 * There is anonymous state. Destroy that first.
17340 		 */
17341 		ASSERT(dtrace_anon.dta_state == NULL);
17342 		dtrace_state_destroy(state->dts_anon);
17343 	}
17344 
17345 	if (dtrace_helptrace_disable) {
17346 		/*
17347 		 * If we have been told to disable helper tracing, set the
17348 		 * buffer to NULL before calling into dtrace_state_destroy();
17349 		 * we take advantage of its dtrace_sync() to know that no
17350 		 * CPU is in probe context with enabled helper tracing
17351 		 * after it returns.
17352 		 */
17353 		buf = dtrace_helptrace_buffer;
17354 		dtrace_helptrace_buffer = NULL;
17355 	}
17356 
17357 #ifdef illumos
17358 	dtrace_state_destroy(state);
17359 #else
17360 	if (state != NULL) {
17361 		dtrace_state_destroy(state);
17362 		kmem_free(state, 0);
17363 	}
17364 #endif
17365 	ASSERT(dtrace_opens > 0);
17366 
17367 #ifdef illumos
17368 	/*
17369 	 * Only relinquish control of the kernel debugger interface when there
17370 	 * are no consumers and no anonymous enablings.
17371 	 */
17372 	if (--dtrace_opens == 0 && dtrace_anon.dta_enabling == NULL)
17373 		(void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
17374 #else
17375 	--dtrace_opens;
17376 #endif
17377 
17378 	if (buf != NULL) {
17379 		kmem_free(buf, dtrace_helptrace_bufsize);
17380 		dtrace_helptrace_disable = 0;
17381 	}
17382 
17383 	mutex_exit(&dtrace_lock);
17384 	mutex_exit(&cpu_lock);
17385 
17386 #ifdef illumos
17387 	return (0);
17388 #endif
17389 }
17390 
17391 #ifdef illumos
17392 /*ARGSUSED*/
17393 static int
17394 dtrace_ioctl_helper(int cmd, intptr_t arg, int *rv)
17395 {
17396 	int rval;
17397 	dof_helper_t help, *dhp = NULL;
17398 
17399 	switch (cmd) {
17400 	case DTRACEHIOC_ADDDOF:
17401 		if (copyin((void *)arg, &help, sizeof (help)) != 0) {
17402 			dtrace_dof_error(NULL, "failed to copyin DOF helper");
17403 			return (EFAULT);
17404 		}
17405 
17406 		dhp = &help;
17407 		arg = (intptr_t)help.dofhp_dof;
17408 		/*FALLTHROUGH*/
17409 
17410 	case DTRACEHIOC_ADD: {
17411 		dof_hdr_t *dof = dtrace_dof_copyin(arg, &rval);
17412 
17413 		if (dof == NULL)
17414 			return (rval);
17415 
17416 		mutex_enter(&dtrace_lock);
17417 
17418 		/*
17419 		 * dtrace_helper_slurp() takes responsibility for the dof --
17420 		 * it may free it now or it may save it and free it later.
17421 		 */
17422 		if ((rval = dtrace_helper_slurp(dof, dhp)) != -1) {
17423 			*rv = rval;
17424 			rval = 0;
17425 		} else {
17426 			rval = EINVAL;
17427 		}
17428 
17429 		mutex_exit(&dtrace_lock);
17430 		return (rval);
17431 	}
17432 
17433 	case DTRACEHIOC_REMOVE: {
17434 		mutex_enter(&dtrace_lock);
17435 		rval = dtrace_helper_destroygen(NULL, arg);
17436 		mutex_exit(&dtrace_lock);
17437 
17438 		return (rval);
17439 	}
17440 
17441 	default:
17442 		break;
17443 	}
17444 
17445 	return (ENOTTY);
17446 }
17447 
17448 /*ARGSUSED*/
17449 static int
17450 dtrace_ioctl(dev_t dev, int cmd, intptr_t arg, int md, cred_t *cr, int *rv)
17451 {
17452 	minor_t minor = getminor(dev);
17453 	dtrace_state_t *state;
17454 	int rval;
17455 
17456 	if (minor == DTRACEMNRN_HELPER)
17457 		return (dtrace_ioctl_helper(cmd, arg, rv));
17458 
17459 	state = ddi_get_soft_state(dtrace_softstate, minor);
17460 
17461 	if (state->dts_anon) {
17462 		ASSERT(dtrace_anon.dta_state == NULL);
17463 		state = state->dts_anon;
17464 	}
17465 
17466 	switch (cmd) {
17467 	case DTRACEIOC_PROVIDER: {
17468 		dtrace_providerdesc_t pvd;
17469 		dtrace_provider_t *pvp;
17470 
17471 		if (copyin((void *)arg, &pvd, sizeof (pvd)) != 0)
17472 			return (EFAULT);
17473 
17474 		pvd.dtvd_name[DTRACE_PROVNAMELEN - 1] = '\0';
17475 		mutex_enter(&dtrace_provider_lock);
17476 
17477 		for (pvp = dtrace_provider; pvp != NULL; pvp = pvp->dtpv_next) {
17478 			if (strcmp(pvp->dtpv_name, pvd.dtvd_name) == 0)
17479 				break;
17480 		}
17481 
17482 		mutex_exit(&dtrace_provider_lock);
17483 
17484 		if (pvp == NULL)
17485 			return (ESRCH);
17486 
17487 		bcopy(&pvp->dtpv_priv, &pvd.dtvd_priv, sizeof (dtrace_ppriv_t));
17488 		bcopy(&pvp->dtpv_attr, &pvd.dtvd_attr, sizeof (dtrace_pattr_t));
17489 
17490 		if (copyout(&pvd, (void *)arg, sizeof (pvd)) != 0)
17491 			return (EFAULT);
17492 
17493 		return (0);
17494 	}
17495 
17496 	case DTRACEIOC_EPROBE: {
17497 		dtrace_eprobedesc_t epdesc;
17498 		dtrace_ecb_t *ecb;
17499 		dtrace_action_t *act;
17500 		void *buf;
17501 		size_t size;
17502 		uintptr_t dest;
17503 		int nrecs;
17504 
17505 		if (copyin((void *)arg, &epdesc, sizeof (epdesc)) != 0)
17506 			return (EFAULT);
17507 
17508 		mutex_enter(&dtrace_lock);
17509 
17510 		if ((ecb = dtrace_epid2ecb(state, epdesc.dtepd_epid)) == NULL) {
17511 			mutex_exit(&dtrace_lock);
17512 			return (EINVAL);
17513 		}
17514 
17515 		if (ecb->dte_probe == NULL) {
17516 			mutex_exit(&dtrace_lock);
17517 			return (EINVAL);
17518 		}
17519 
17520 		epdesc.dtepd_probeid = ecb->dte_probe->dtpr_id;
17521 		epdesc.dtepd_uarg = ecb->dte_uarg;
17522 		epdesc.dtepd_size = ecb->dte_size;
17523 
17524 		nrecs = epdesc.dtepd_nrecs;
17525 		epdesc.dtepd_nrecs = 0;
17526 		for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
17527 			if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
17528 				continue;
17529 
17530 			epdesc.dtepd_nrecs++;
17531 		}
17532 
17533 		/*
17534 		 * Now that we have the size, we need to allocate a temporary
17535 		 * buffer in which to store the complete description.  We need
17536 		 * the temporary buffer to be able to drop dtrace_lock()
17537 		 * across the copyout(), below.
17538 		 */
17539 		size = sizeof (dtrace_eprobedesc_t) +
17540 		    (epdesc.dtepd_nrecs * sizeof (dtrace_recdesc_t));
17541 
17542 		buf = kmem_alloc(size, KM_SLEEP);
17543 		dest = (uintptr_t)buf;
17544 
17545 		bcopy(&epdesc, (void *)dest, sizeof (epdesc));
17546 		dest += offsetof(dtrace_eprobedesc_t, dtepd_rec[0]);
17547 
17548 		for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
17549 			if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
17550 				continue;
17551 
17552 			if (nrecs-- == 0)
17553 				break;
17554 
17555 			bcopy(&act->dta_rec, (void *)dest,
17556 			    sizeof (dtrace_recdesc_t));
17557 			dest += sizeof (dtrace_recdesc_t);
17558 		}
17559 
17560 		mutex_exit(&dtrace_lock);
17561 
17562 		if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
17563 			kmem_free(buf, size);
17564 			return (EFAULT);
17565 		}
17566 
17567 		kmem_free(buf, size);
17568 		return (0);
17569 	}
17570 
17571 	case DTRACEIOC_AGGDESC: {
17572 		dtrace_aggdesc_t aggdesc;
17573 		dtrace_action_t *act;
17574 		dtrace_aggregation_t *agg;
17575 		int nrecs;
17576 		uint32_t offs;
17577 		dtrace_recdesc_t *lrec;
17578 		void *buf;
17579 		size_t size;
17580 		uintptr_t dest;
17581 
17582 		if (copyin((void *)arg, &aggdesc, sizeof (aggdesc)) != 0)
17583 			return (EFAULT);
17584 
17585 		mutex_enter(&dtrace_lock);
17586 
17587 		if ((agg = dtrace_aggid2agg(state, aggdesc.dtagd_id)) == NULL) {
17588 			mutex_exit(&dtrace_lock);
17589 			return (EINVAL);
17590 		}
17591 
17592 		aggdesc.dtagd_epid = agg->dtag_ecb->dte_epid;
17593 
17594 		nrecs = aggdesc.dtagd_nrecs;
17595 		aggdesc.dtagd_nrecs = 0;
17596 
17597 		offs = agg->dtag_base;
17598 		lrec = &agg->dtag_action.dta_rec;
17599 		aggdesc.dtagd_size = lrec->dtrd_offset + lrec->dtrd_size - offs;
17600 
17601 		for (act = agg->dtag_first; ; act = act->dta_next) {
17602 			ASSERT(act->dta_intuple ||
17603 			    DTRACEACT_ISAGG(act->dta_kind));
17604 
17605 			/*
17606 			 * If this action has a record size of zero, it
17607 			 * denotes an argument to the aggregating action.
17608 			 * Because the presence of this record doesn't (or
17609 			 * shouldn't) affect the way the data is interpreted,
17610 			 * we don't copy it out to save user-level the
17611 			 * confusion of dealing with a zero-length record.
17612 			 */
17613 			if (act->dta_rec.dtrd_size == 0) {
17614 				ASSERT(agg->dtag_hasarg);
17615 				continue;
17616 			}
17617 
17618 			aggdesc.dtagd_nrecs++;
17619 
17620 			if (act == &agg->dtag_action)
17621 				break;
17622 		}
17623 
17624 		/*
17625 		 * Now that we have the size, we need to allocate a temporary
17626 		 * buffer in which to store the complete description.  We need
17627 		 * the temporary buffer to be able to drop dtrace_lock()
17628 		 * across the copyout(), below.
17629 		 */
17630 		size = sizeof (dtrace_aggdesc_t) +
17631 		    (aggdesc.dtagd_nrecs * sizeof (dtrace_recdesc_t));
17632 
17633 		buf = kmem_alloc(size, KM_SLEEP);
17634 		dest = (uintptr_t)buf;
17635 
17636 		bcopy(&aggdesc, (void *)dest, sizeof (aggdesc));
17637 		dest += offsetof(dtrace_aggdesc_t, dtagd_rec[0]);
17638 
17639 		for (act = agg->dtag_first; ; act = act->dta_next) {
17640 			dtrace_recdesc_t rec = act->dta_rec;
17641 
17642 			/*
17643 			 * See the comment in the above loop for why we pass
17644 			 * over zero-length records.
17645 			 */
17646 			if (rec.dtrd_size == 0) {
17647 				ASSERT(agg->dtag_hasarg);
17648 				continue;
17649 			}
17650 
17651 			if (nrecs-- == 0)
17652 				break;
17653 
17654 			rec.dtrd_offset -= offs;
17655 			bcopy(&rec, (void *)dest, sizeof (rec));
17656 			dest += sizeof (dtrace_recdesc_t);
17657 
17658 			if (act == &agg->dtag_action)
17659 				break;
17660 		}
17661 
17662 		mutex_exit(&dtrace_lock);
17663 
17664 		if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
17665 			kmem_free(buf, size);
17666 			return (EFAULT);
17667 		}
17668 
17669 		kmem_free(buf, size);
17670 		return (0);
17671 	}
17672 
17673 	case DTRACEIOC_ENABLE: {
17674 		dof_hdr_t *dof;
17675 		dtrace_enabling_t *enab = NULL;
17676 		dtrace_vstate_t *vstate;
17677 		int err = 0;
17678 
17679 		*rv = 0;
17680 
17681 		/*
17682 		 * If a NULL argument has been passed, we take this as our
17683 		 * cue to reevaluate our enablings.
17684 		 */
17685 		if (arg == NULL) {
17686 			dtrace_enabling_matchall();
17687 
17688 			return (0);
17689 		}
17690 
17691 		if ((dof = dtrace_dof_copyin(arg, &rval)) == NULL)
17692 			return (rval);
17693 
17694 		mutex_enter(&cpu_lock);
17695 		mutex_enter(&dtrace_lock);
17696 		vstate = &state->dts_vstate;
17697 
17698 		if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
17699 			mutex_exit(&dtrace_lock);
17700 			mutex_exit(&cpu_lock);
17701 			dtrace_dof_destroy(dof);
17702 			return (EBUSY);
17703 		}
17704 
17705 		if (dtrace_dof_slurp(dof, vstate, cr, &enab, 0, B_TRUE) != 0) {
17706 			mutex_exit(&dtrace_lock);
17707 			mutex_exit(&cpu_lock);
17708 			dtrace_dof_destroy(dof);
17709 			return (EINVAL);
17710 		}
17711 
17712 		if ((rval = dtrace_dof_options(dof, state)) != 0) {
17713 			dtrace_enabling_destroy(enab);
17714 			mutex_exit(&dtrace_lock);
17715 			mutex_exit(&cpu_lock);
17716 			dtrace_dof_destroy(dof);
17717 			return (rval);
17718 		}
17719 
17720 		if ((err = dtrace_enabling_match(enab, rv)) == 0) {
17721 			err = dtrace_enabling_retain(enab);
17722 		} else {
17723 			dtrace_enabling_destroy(enab);
17724 		}
17725 
17726 		mutex_exit(&cpu_lock);
17727 		mutex_exit(&dtrace_lock);
17728 		dtrace_dof_destroy(dof);
17729 
17730 		return (err);
17731 	}
17732 
17733 	case DTRACEIOC_REPLICATE: {
17734 		dtrace_repldesc_t desc;
17735 		dtrace_probedesc_t *match = &desc.dtrpd_match;
17736 		dtrace_probedesc_t *create = &desc.dtrpd_create;
17737 		int err;
17738 
17739 		if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
17740 			return (EFAULT);
17741 
17742 		match->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
17743 		match->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
17744 		match->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
17745 		match->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
17746 
17747 		create->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
17748 		create->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
17749 		create->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
17750 		create->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
17751 
17752 		mutex_enter(&dtrace_lock);
17753 		err = dtrace_enabling_replicate(state, match, create);
17754 		mutex_exit(&dtrace_lock);
17755 
17756 		return (err);
17757 	}
17758 
17759 	case DTRACEIOC_PROBEMATCH:
17760 	case DTRACEIOC_PROBES: {
17761 		dtrace_probe_t *probe = NULL;
17762 		dtrace_probedesc_t desc;
17763 		dtrace_probekey_t pkey;
17764 		dtrace_id_t i;
17765 		int m = 0;
17766 		uint32_t priv;
17767 		uid_t uid;
17768 		zoneid_t zoneid;
17769 
17770 		if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
17771 			return (EFAULT);
17772 
17773 		desc.dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
17774 		desc.dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
17775 		desc.dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
17776 		desc.dtpd_name[DTRACE_NAMELEN - 1] = '\0';
17777 
17778 		/*
17779 		 * Before we attempt to match this probe, we want to give
17780 		 * all providers the opportunity to provide it.
17781 		 */
17782 		if (desc.dtpd_id == DTRACE_IDNONE) {
17783 			mutex_enter(&dtrace_provider_lock);
17784 			dtrace_probe_provide(&desc, NULL);
17785 			mutex_exit(&dtrace_provider_lock);
17786 			desc.dtpd_id++;
17787 		}
17788 
17789 		if (cmd == DTRACEIOC_PROBEMATCH)  {
17790 			dtrace_probekey(&desc, &pkey);
17791 			pkey.dtpk_id = DTRACE_IDNONE;
17792 		}
17793 
17794 		dtrace_cred2priv(cr, &priv, &uid, &zoneid);
17795 
17796 		mutex_enter(&dtrace_lock);
17797 
17798 		if (cmd == DTRACEIOC_PROBEMATCH) {
17799 			for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
17800 				if ((probe = dtrace_probes[i - 1]) != NULL &&
17801 				    (m = dtrace_match_probe(probe, &pkey,
17802 				    priv, uid, zoneid)) != 0)
17803 					break;
17804 			}
17805 
17806 			if (m < 0) {
17807 				mutex_exit(&dtrace_lock);
17808 				return (EINVAL);
17809 			}
17810 
17811 		} else {
17812 			for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
17813 				if ((probe = dtrace_probes[i - 1]) != NULL &&
17814 				    dtrace_match_priv(probe, priv, uid, zoneid))
17815 					break;
17816 			}
17817 		}
17818 
17819 		if (probe == NULL) {
17820 			mutex_exit(&dtrace_lock);
17821 			return (ESRCH);
17822 		}
17823 
17824 		dtrace_probe_description(probe, &desc);
17825 		mutex_exit(&dtrace_lock);
17826 
17827 		if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
17828 			return (EFAULT);
17829 
17830 		return (0);
17831 	}
17832 
17833 	case DTRACEIOC_PROBEARG: {
17834 		dtrace_argdesc_t desc;
17835 		dtrace_probe_t *probe;
17836 		dtrace_provider_t *prov;
17837 
17838 		if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
17839 			return (EFAULT);
17840 
17841 		if (desc.dtargd_id == DTRACE_IDNONE)
17842 			return (EINVAL);
17843 
17844 		if (desc.dtargd_ndx == DTRACE_ARGNONE)
17845 			return (EINVAL);
17846 
17847 		mutex_enter(&dtrace_provider_lock);
17848 		mutex_enter(&mod_lock);
17849 		mutex_enter(&dtrace_lock);
17850 
17851 		if (desc.dtargd_id > dtrace_nprobes) {
17852 			mutex_exit(&dtrace_lock);
17853 			mutex_exit(&mod_lock);
17854 			mutex_exit(&dtrace_provider_lock);
17855 			return (EINVAL);
17856 		}
17857 
17858 		if ((probe = dtrace_probes[desc.dtargd_id - 1]) == NULL) {
17859 			mutex_exit(&dtrace_lock);
17860 			mutex_exit(&mod_lock);
17861 			mutex_exit(&dtrace_provider_lock);
17862 			return (EINVAL);
17863 		}
17864 
17865 		mutex_exit(&dtrace_lock);
17866 
17867 		prov = probe->dtpr_provider;
17868 
17869 		if (prov->dtpv_pops.dtps_getargdesc == NULL) {
17870 			/*
17871 			 * There isn't any typed information for this probe.
17872 			 * Set the argument number to DTRACE_ARGNONE.
17873 			 */
17874 			desc.dtargd_ndx = DTRACE_ARGNONE;
17875 		} else {
17876 			desc.dtargd_native[0] = '\0';
17877 			desc.dtargd_xlate[0] = '\0';
17878 			desc.dtargd_mapping = desc.dtargd_ndx;
17879 
17880 			prov->dtpv_pops.dtps_getargdesc(prov->dtpv_arg,
17881 			    probe->dtpr_id, probe->dtpr_arg, &desc);
17882 		}
17883 
17884 		mutex_exit(&mod_lock);
17885 		mutex_exit(&dtrace_provider_lock);
17886 
17887 		if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
17888 			return (EFAULT);
17889 
17890 		return (0);
17891 	}
17892 
17893 	case DTRACEIOC_GO: {
17894 		processorid_t cpuid;
17895 		rval = dtrace_state_go(state, &cpuid);
17896 
17897 		if (rval != 0)
17898 			return (rval);
17899 
17900 		if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
17901 			return (EFAULT);
17902 
17903 		return (0);
17904 	}
17905 
17906 	case DTRACEIOC_STOP: {
17907 		processorid_t cpuid;
17908 
17909 		mutex_enter(&dtrace_lock);
17910 		rval = dtrace_state_stop(state, &cpuid);
17911 		mutex_exit(&dtrace_lock);
17912 
17913 		if (rval != 0)
17914 			return (rval);
17915 
17916 		if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
17917 			return (EFAULT);
17918 
17919 		return (0);
17920 	}
17921 
17922 	case DTRACEIOC_DOFGET: {
17923 		dof_hdr_t hdr, *dof;
17924 		uint64_t len;
17925 
17926 		if (copyin((void *)arg, &hdr, sizeof (hdr)) != 0)
17927 			return (EFAULT);
17928 
17929 		mutex_enter(&dtrace_lock);
17930 		dof = dtrace_dof_create(state);
17931 		mutex_exit(&dtrace_lock);
17932 
17933 		len = MIN(hdr.dofh_loadsz, dof->dofh_loadsz);
17934 		rval = copyout(dof, (void *)arg, len);
17935 		dtrace_dof_destroy(dof);
17936 
17937 		return (rval == 0 ? 0 : EFAULT);
17938 	}
17939 
17940 	case DTRACEIOC_AGGSNAP:
17941 	case DTRACEIOC_BUFSNAP: {
17942 		dtrace_bufdesc_t desc;
17943 		caddr_t cached;
17944 		dtrace_buffer_t *buf;
17945 
17946 		if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
17947 			return (EFAULT);
17948 
17949 		if (desc.dtbd_cpu < 0 || desc.dtbd_cpu >= NCPU)
17950 			return (EINVAL);
17951 
17952 		mutex_enter(&dtrace_lock);
17953 
17954 		if (cmd == DTRACEIOC_BUFSNAP) {
17955 			buf = &state->dts_buffer[desc.dtbd_cpu];
17956 		} else {
17957 			buf = &state->dts_aggbuffer[desc.dtbd_cpu];
17958 		}
17959 
17960 		if (buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL)) {
17961 			size_t sz = buf->dtb_offset;
17962 
17963 			if (state->dts_activity != DTRACE_ACTIVITY_STOPPED) {
17964 				mutex_exit(&dtrace_lock);
17965 				return (EBUSY);
17966 			}
17967 
17968 			/*
17969 			 * If this buffer has already been consumed, we're
17970 			 * going to indicate that there's nothing left here
17971 			 * to consume.
17972 			 */
17973 			if (buf->dtb_flags & DTRACEBUF_CONSUMED) {
17974 				mutex_exit(&dtrace_lock);
17975 
17976 				desc.dtbd_size = 0;
17977 				desc.dtbd_drops = 0;
17978 				desc.dtbd_errors = 0;
17979 				desc.dtbd_oldest = 0;
17980 				sz = sizeof (desc);
17981 
17982 				if (copyout(&desc, (void *)arg, sz) != 0)
17983 					return (EFAULT);
17984 
17985 				return (0);
17986 			}
17987 
17988 			/*
17989 			 * If this is a ring buffer that has wrapped, we want
17990 			 * to copy the whole thing out.
17991 			 */
17992 			if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
17993 				dtrace_buffer_polish(buf);
17994 				sz = buf->dtb_size;
17995 			}
17996 
17997 			if (copyout(buf->dtb_tomax, desc.dtbd_data, sz) != 0) {
17998 				mutex_exit(&dtrace_lock);
17999 				return (EFAULT);
18000 			}
18001 
18002 			desc.dtbd_size = sz;
18003 			desc.dtbd_drops = buf->dtb_drops;
18004 			desc.dtbd_errors = buf->dtb_errors;
18005 			desc.dtbd_oldest = buf->dtb_xamot_offset;
18006 			desc.dtbd_timestamp = dtrace_gethrtime();
18007 
18008 			mutex_exit(&dtrace_lock);
18009 
18010 			if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
18011 				return (EFAULT);
18012 
18013 			buf->dtb_flags |= DTRACEBUF_CONSUMED;
18014 
18015 			return (0);
18016 		}
18017 
18018 		if (buf->dtb_tomax == NULL) {
18019 			ASSERT(buf->dtb_xamot == NULL);
18020 			mutex_exit(&dtrace_lock);
18021 			return (ENOENT);
18022 		}
18023 
18024 		cached = buf->dtb_tomax;
18025 		ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
18026 
18027 		dtrace_xcall(desc.dtbd_cpu,
18028 		    (dtrace_xcall_t)dtrace_buffer_switch, buf);
18029 
18030 		state->dts_errors += buf->dtb_xamot_errors;
18031 
18032 		/*
18033 		 * If the buffers did not actually switch, then the cross call
18034 		 * did not take place -- presumably because the given CPU is
18035 		 * not in the ready set.  If this is the case, we'll return
18036 		 * ENOENT.
18037 		 */
18038 		if (buf->dtb_tomax == cached) {
18039 			ASSERT(buf->dtb_xamot != cached);
18040 			mutex_exit(&dtrace_lock);
18041 			return (ENOENT);
18042 		}
18043 
18044 		ASSERT(cached == buf->dtb_xamot);
18045 
18046 		/*
18047 		 * We have our snapshot; now copy it out.
18048 		 */
18049 		if (copyout(buf->dtb_xamot, desc.dtbd_data,
18050 		    buf->dtb_xamot_offset) != 0) {
18051 			mutex_exit(&dtrace_lock);
18052 			return (EFAULT);
18053 		}
18054 
18055 		desc.dtbd_size = buf->dtb_xamot_offset;
18056 		desc.dtbd_drops = buf->dtb_xamot_drops;
18057 		desc.dtbd_errors = buf->dtb_xamot_errors;
18058 		desc.dtbd_oldest = 0;
18059 		desc.dtbd_timestamp = buf->dtb_switched;
18060 
18061 		mutex_exit(&dtrace_lock);
18062 
18063 		/*
18064 		 * Finally, copy out the buffer description.
18065 		 */
18066 		if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
18067 			return (EFAULT);
18068 
18069 		return (0);
18070 	}
18071 
18072 	case DTRACEIOC_CONF: {
18073 		dtrace_conf_t conf;
18074 
18075 		bzero(&conf, sizeof (conf));
18076 		conf.dtc_difversion = DIF_VERSION;
18077 		conf.dtc_difintregs = DIF_DIR_NREGS;
18078 		conf.dtc_diftupregs = DIF_DTR_NREGS;
18079 		conf.dtc_ctfmodel = CTF_MODEL_NATIVE;
18080 
18081 		if (copyout(&conf, (void *)arg, sizeof (conf)) != 0)
18082 			return (EFAULT);
18083 
18084 		return (0);
18085 	}
18086 
18087 	case DTRACEIOC_STATUS: {
18088 		dtrace_status_t stat;
18089 		dtrace_dstate_t *dstate;
18090 		int i, j;
18091 		uint64_t nerrs;
18092 
18093 		/*
18094 		 * See the comment in dtrace_state_deadman() for the reason
18095 		 * for setting dts_laststatus to INT64_MAX before setting
18096 		 * it to the correct value.
18097 		 */
18098 		state->dts_laststatus = INT64_MAX;
18099 		dtrace_membar_producer();
18100 		state->dts_laststatus = dtrace_gethrtime();
18101 
18102 		bzero(&stat, sizeof (stat));
18103 
18104 		mutex_enter(&dtrace_lock);
18105 
18106 		if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE) {
18107 			mutex_exit(&dtrace_lock);
18108 			return (ENOENT);
18109 		}
18110 
18111 		if (state->dts_activity == DTRACE_ACTIVITY_DRAINING)
18112 			stat.dtst_exiting = 1;
18113 
18114 		nerrs = state->dts_errors;
18115 		dstate = &state->dts_vstate.dtvs_dynvars;
18116 
18117 		for (i = 0; i < NCPU; i++) {
18118 			dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[i];
18119 
18120 			stat.dtst_dyndrops += dcpu->dtdsc_drops;
18121 			stat.dtst_dyndrops_dirty += dcpu->dtdsc_dirty_drops;
18122 			stat.dtst_dyndrops_rinsing += dcpu->dtdsc_rinsing_drops;
18123 
18124 			if (state->dts_buffer[i].dtb_flags & DTRACEBUF_FULL)
18125 				stat.dtst_filled++;
18126 
18127 			nerrs += state->dts_buffer[i].dtb_errors;
18128 
18129 			for (j = 0; j < state->dts_nspeculations; j++) {
18130 				dtrace_speculation_t *spec;
18131 				dtrace_buffer_t *buf;
18132 
18133 				spec = &state->dts_speculations[j];
18134 				buf = &spec->dtsp_buffer[i];
18135 				stat.dtst_specdrops += buf->dtb_xamot_drops;
18136 			}
18137 		}
18138 
18139 		stat.dtst_specdrops_busy = state->dts_speculations_busy;
18140 		stat.dtst_specdrops_unavail = state->dts_speculations_unavail;
18141 		stat.dtst_stkstroverflows = state->dts_stkstroverflows;
18142 		stat.dtst_dblerrors = state->dts_dblerrors;
18143 		stat.dtst_killed =
18144 		    (state->dts_activity == DTRACE_ACTIVITY_KILLED);
18145 		stat.dtst_errors = nerrs;
18146 
18147 		mutex_exit(&dtrace_lock);
18148 
18149 		if (copyout(&stat, (void *)arg, sizeof (stat)) != 0)
18150 			return (EFAULT);
18151 
18152 		return (0);
18153 	}
18154 
18155 	case DTRACEIOC_FORMAT: {
18156 		dtrace_fmtdesc_t fmt;
18157 		char *str;
18158 		int len;
18159 
18160 		if (copyin((void *)arg, &fmt, sizeof (fmt)) != 0)
18161 			return (EFAULT);
18162 
18163 		mutex_enter(&dtrace_lock);
18164 
18165 		if (fmt.dtfd_format == 0 ||
18166 		    fmt.dtfd_format > state->dts_nformats) {
18167 			mutex_exit(&dtrace_lock);
18168 			return (EINVAL);
18169 		}
18170 
18171 		/*
18172 		 * Format strings are allocated contiguously and they are
18173 		 * never freed; if a format index is less than the number
18174 		 * of formats, we can assert that the format map is non-NULL
18175 		 * and that the format for the specified index is non-NULL.
18176 		 */
18177 		ASSERT(state->dts_formats != NULL);
18178 		str = state->dts_formats[fmt.dtfd_format - 1];
18179 		ASSERT(str != NULL);
18180 
18181 		len = strlen(str) + 1;
18182 
18183 		if (len > fmt.dtfd_length) {
18184 			fmt.dtfd_length = len;
18185 
18186 			if (copyout(&fmt, (void *)arg, sizeof (fmt)) != 0) {
18187 				mutex_exit(&dtrace_lock);
18188 				return (EINVAL);
18189 			}
18190 		} else {
18191 			if (copyout(str, fmt.dtfd_string, len) != 0) {
18192 				mutex_exit(&dtrace_lock);
18193 				return (EINVAL);
18194 			}
18195 		}
18196 
18197 		mutex_exit(&dtrace_lock);
18198 		return (0);
18199 	}
18200 
18201 	default:
18202 		break;
18203 	}
18204 
18205 	return (ENOTTY);
18206 }
18207 
18208 /*ARGSUSED*/
18209 static int
18210 dtrace_detach(dev_info_t *dip, ddi_detach_cmd_t cmd)
18211 {
18212 	dtrace_state_t *state;
18213 
18214 	switch (cmd) {
18215 	case DDI_DETACH:
18216 		break;
18217 
18218 	case DDI_SUSPEND:
18219 		return (DDI_SUCCESS);
18220 
18221 	default:
18222 		return (DDI_FAILURE);
18223 	}
18224 
18225 	mutex_enter(&cpu_lock);
18226 	mutex_enter(&dtrace_provider_lock);
18227 	mutex_enter(&dtrace_lock);
18228 
18229 	ASSERT(dtrace_opens == 0);
18230 
18231 	if (dtrace_helpers > 0) {
18232 		mutex_exit(&dtrace_provider_lock);
18233 		mutex_exit(&dtrace_lock);
18234 		mutex_exit(&cpu_lock);
18235 		return (DDI_FAILURE);
18236 	}
18237 
18238 	if (dtrace_unregister((dtrace_provider_id_t)dtrace_provider) != 0) {
18239 		mutex_exit(&dtrace_provider_lock);
18240 		mutex_exit(&dtrace_lock);
18241 		mutex_exit(&cpu_lock);
18242 		return (DDI_FAILURE);
18243 	}
18244 
18245 	dtrace_provider = NULL;
18246 
18247 	if ((state = dtrace_anon_grab()) != NULL) {
18248 		/*
18249 		 * If there were ECBs on this state, the provider should
18250 		 * have not been allowed to detach; assert that there is
18251 		 * none.
18252 		 */
18253 		ASSERT(state->dts_necbs == 0);
18254 		dtrace_state_destroy(state);
18255 
18256 		/*
18257 		 * If we're being detached with anonymous state, we need to
18258 		 * indicate to the kernel debugger that DTrace is now inactive.
18259 		 */
18260 		(void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
18261 	}
18262 
18263 	bzero(&dtrace_anon, sizeof (dtrace_anon_t));
18264 	unregister_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
18265 	dtrace_cpu_init = NULL;
18266 	dtrace_helpers_cleanup = NULL;
18267 	dtrace_helpers_fork = NULL;
18268 	dtrace_cpustart_init = NULL;
18269 	dtrace_cpustart_fini = NULL;
18270 	dtrace_debugger_init = NULL;
18271 	dtrace_debugger_fini = NULL;
18272 	dtrace_modload = NULL;
18273 	dtrace_modunload = NULL;
18274 
18275 	ASSERT(dtrace_getf == 0);
18276 	ASSERT(dtrace_closef == NULL);
18277 
18278 	mutex_exit(&cpu_lock);
18279 
18280 	kmem_free(dtrace_probes, dtrace_nprobes * sizeof (dtrace_probe_t *));
18281 	dtrace_probes = NULL;
18282 	dtrace_nprobes = 0;
18283 
18284 	dtrace_hash_destroy(dtrace_bymod);
18285 	dtrace_hash_destroy(dtrace_byfunc);
18286 	dtrace_hash_destroy(dtrace_byname);
18287 	dtrace_bymod = NULL;
18288 	dtrace_byfunc = NULL;
18289 	dtrace_byname = NULL;
18290 
18291 	kmem_cache_destroy(dtrace_state_cache);
18292 	vmem_destroy(dtrace_minor);
18293 	vmem_destroy(dtrace_arena);
18294 
18295 	if (dtrace_toxrange != NULL) {
18296 		kmem_free(dtrace_toxrange,
18297 		    dtrace_toxranges_max * sizeof (dtrace_toxrange_t));
18298 		dtrace_toxrange = NULL;
18299 		dtrace_toxranges = 0;
18300 		dtrace_toxranges_max = 0;
18301 	}
18302 
18303 	ddi_remove_minor_node(dtrace_devi, NULL);
18304 	dtrace_devi = NULL;
18305 
18306 	ddi_soft_state_fini(&dtrace_softstate);
18307 
18308 	ASSERT(dtrace_vtime_references == 0);
18309 	ASSERT(dtrace_opens == 0);
18310 	ASSERT(dtrace_retained == NULL);
18311 
18312 	mutex_exit(&dtrace_lock);
18313 	mutex_exit(&dtrace_provider_lock);
18314 
18315 	/*
18316 	 * We don't destroy the task queue until after we have dropped our
18317 	 * locks (taskq_destroy() may block on running tasks).  To prevent
18318 	 * attempting to do work after we have effectively detached but before
18319 	 * the task queue has been destroyed, all tasks dispatched via the
18320 	 * task queue must check that DTrace is still attached before
18321 	 * performing any operation.
18322 	 */
18323 	taskq_destroy(dtrace_taskq);
18324 	dtrace_taskq = NULL;
18325 
18326 	return (DDI_SUCCESS);
18327 }
18328 #endif
18329 
18330 #ifdef illumos
18331 /*ARGSUSED*/
18332 static int
18333 dtrace_info(dev_info_t *dip, ddi_info_cmd_t infocmd, void *arg, void **result)
18334 {
18335 	int error;
18336 
18337 	switch (infocmd) {
18338 	case DDI_INFO_DEVT2DEVINFO:
18339 		*result = (void *)dtrace_devi;
18340 		error = DDI_SUCCESS;
18341 		break;
18342 	case DDI_INFO_DEVT2INSTANCE:
18343 		*result = (void *)0;
18344 		error = DDI_SUCCESS;
18345 		break;
18346 	default:
18347 		error = DDI_FAILURE;
18348 	}
18349 	return (error);
18350 }
18351 #endif
18352 
18353 #ifdef illumos
18354 static struct cb_ops dtrace_cb_ops = {
18355 	dtrace_open,		/* open */
18356 	dtrace_close,		/* close */
18357 	nulldev,		/* strategy */
18358 	nulldev,		/* print */
18359 	nodev,			/* dump */
18360 	nodev,			/* read */
18361 	nodev,			/* write */
18362 	dtrace_ioctl,		/* ioctl */
18363 	nodev,			/* devmap */
18364 	nodev,			/* mmap */
18365 	nodev,			/* segmap */
18366 	nochpoll,		/* poll */
18367 	ddi_prop_op,		/* cb_prop_op */
18368 	0,			/* streamtab  */
18369 	D_NEW | D_MP		/* Driver compatibility flag */
18370 };
18371 
18372 static struct dev_ops dtrace_ops = {
18373 	DEVO_REV,		/* devo_rev */
18374 	0,			/* refcnt */
18375 	dtrace_info,		/* get_dev_info */
18376 	nulldev,		/* identify */
18377 	nulldev,		/* probe */
18378 	dtrace_attach,		/* attach */
18379 	dtrace_detach,		/* detach */
18380 	nodev,			/* reset */
18381 	&dtrace_cb_ops,		/* driver operations */
18382 	NULL,			/* bus operations */
18383 	nodev			/* dev power */
18384 };
18385 
18386 static struct modldrv modldrv = {
18387 	&mod_driverops,		/* module type (this is a pseudo driver) */
18388 	"Dynamic Tracing",	/* name of module */
18389 	&dtrace_ops,		/* driver ops */
18390 };
18391 
18392 static struct modlinkage modlinkage = {
18393 	MODREV_1,
18394 	(void *)&modldrv,
18395 	NULL
18396 };
18397 
18398 int
18399 _init(void)
18400 {
18401 	return (mod_install(&modlinkage));
18402 }
18403 
18404 int
18405 _info(struct modinfo *modinfop)
18406 {
18407 	return (mod_info(&modlinkage, modinfop));
18408 }
18409 
18410 int
18411 _fini(void)
18412 {
18413 	return (mod_remove(&modlinkage));
18414 }
18415 #else
18416 
18417 static d_ioctl_t	dtrace_ioctl;
18418 static d_ioctl_t	dtrace_ioctl_helper;
18419 static void		dtrace_load(void *);
18420 static int		dtrace_unload(void);
18421 static struct cdev	*dtrace_dev;
18422 static struct cdev	*helper_dev;
18423 
18424 void dtrace_invop_init(void);
18425 void dtrace_invop_uninit(void);
18426 
18427 static struct cdevsw dtrace_cdevsw = {
18428 	.d_version	= D_VERSION,
18429 	.d_ioctl	= dtrace_ioctl,
18430 	.d_open		= dtrace_open,
18431 	.d_name		= "dtrace",
18432 };
18433 
18434 static struct cdevsw helper_cdevsw = {
18435 	.d_version	= D_VERSION,
18436 	.d_ioctl	= dtrace_ioctl_helper,
18437 	.d_name		= "helper",
18438 };
18439 
18440 #include <dtrace_anon.c>
18441 #include <dtrace_ioctl.c>
18442 #include <dtrace_load.c>
18443 #include <dtrace_modevent.c>
18444 #include <dtrace_sysctl.c>
18445 #include <dtrace_unload.c>
18446 #include <dtrace_vtime.c>
18447 #include <dtrace_hacks.c>
18448 #include <dtrace_isa.c>
18449 
18450 SYSINIT(dtrace_load, SI_SUB_DTRACE, SI_ORDER_FIRST, dtrace_load, NULL);
18451 SYSUNINIT(dtrace_unload, SI_SUB_DTRACE, SI_ORDER_FIRST, dtrace_unload, NULL);
18452 SYSINIT(dtrace_anon_init, SI_SUB_DTRACE_ANON, SI_ORDER_FIRST, dtrace_anon_init, NULL);
18453 
18454 DEV_MODULE(dtrace, dtrace_modevent, NULL);
18455 MODULE_VERSION(dtrace, 1);
18456 MODULE_DEPEND(dtrace, opensolaris, 1, 1, 1);
18457 #endif
18458