xref: /freebsd/sys/cddl/contrib/opensolaris/uts/common/dtrace/dtrace.c (revision cfd6422a5217410fbd66f7a7a8a64d9d85e61229)
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  * $FreeBSD$
22  */
23 
24 /*
25  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
26  * Copyright (c) 2016, Joyent, Inc. All rights reserved.
27  * Copyright (c) 2012, 2014 by Delphix. All rights reserved.
28  */
29 
30 /*
31  * DTrace - Dynamic Tracing for Solaris
32  *
33  * This is the implementation of the Solaris Dynamic Tracing framework
34  * (DTrace).  The user-visible interface to DTrace is described at length in
35  * the "Solaris Dynamic Tracing Guide".  The interfaces between the libdtrace
36  * library, the in-kernel DTrace framework, and the DTrace providers are
37  * described in the block comments in the <sys/dtrace.h> header file.  The
38  * internal architecture of DTrace is described in the block comments in the
39  * <sys/dtrace_impl.h> header file.  The comments contained within the DTrace
40  * implementation very much assume mastery of all of these sources; if one has
41  * an unanswered question about the implementation, one should consult them
42  * first.
43  *
44  * The functions here are ordered roughly as follows:
45  *
46  *   - Probe context functions
47  *   - Probe hashing functions
48  *   - Non-probe context utility functions
49  *   - Matching functions
50  *   - Provider-to-Framework API functions
51  *   - Probe management functions
52  *   - DIF object functions
53  *   - Format functions
54  *   - Predicate functions
55  *   - ECB functions
56  *   - Buffer functions
57  *   - Enabling functions
58  *   - DOF functions
59  *   - Anonymous enabling functions
60  *   - Consumer state functions
61  *   - Helper functions
62  *   - Hook functions
63  *   - Driver cookbook functions
64  *
65  * Each group of functions begins with a block comment labelled the "DTrace
66  * [Group] Functions", allowing one to find each block by searching forward
67  * on capital-f functions.
68  */
69 #include <sys/errno.h>
70 #include <sys/param.h>
71 #include <sys/types.h>
72 #ifndef illumos
73 #include <sys/time.h>
74 #endif
75 #include <sys/stat.h>
76 #include <sys/conf.h>
77 #include <sys/systm.h>
78 #include <sys/endian.h>
79 #ifdef illumos
80 #include <sys/ddi.h>
81 #include <sys/sunddi.h>
82 #endif
83 #include <sys/cpuvar.h>
84 #include <sys/kmem.h>
85 #ifdef illumos
86 #include <sys/strsubr.h>
87 #endif
88 #include <sys/sysmacros.h>
89 #include <sys/dtrace_impl.h>
90 #include <sys/atomic.h>
91 #include <sys/cmn_err.h>
92 #ifdef illumos
93 #include <sys/mutex_impl.h>
94 #include <sys/rwlock_impl.h>
95 #endif
96 #include <sys/ctf_api.h>
97 #ifdef illumos
98 #include <sys/panic.h>
99 #include <sys/priv_impl.h>
100 #endif
101 #ifdef illumos
102 #include <sys/cred_impl.h>
103 #include <sys/procfs_isa.h>
104 #endif
105 #include <sys/taskq.h>
106 #ifdef illumos
107 #include <sys/mkdev.h>
108 #include <sys/kdi.h>
109 #endif
110 #include <sys/zone.h>
111 #include <sys/socket.h>
112 #include <netinet/in.h>
113 #include "strtolctype.h"
114 
115 /* FreeBSD includes: */
116 #ifndef illumos
117 #include <sys/callout.h>
118 #include <sys/ctype.h>
119 #include <sys/eventhandler.h>
120 #include <sys/limits.h>
121 #include <sys/linker.h>
122 #include <sys/kdb.h>
123 #include <sys/jail.h>
124 #include <sys/kernel.h>
125 #include <sys/malloc.h>
126 #include <sys/lock.h>
127 #include <sys/mutex.h>
128 #include <sys/ptrace.h>
129 #include <sys/random.h>
130 #include <sys/rwlock.h>
131 #include <sys/sx.h>
132 #include <sys/sysctl.h>
133 
134 
135 #include <sys/mount.h>
136 #undef AT_UID
137 #undef AT_GID
138 #include <sys/vnode.h>
139 #include <sys/cred.h>
140 
141 #include <sys/dtrace_bsd.h>
142 
143 #include <netinet/in.h>
144 
145 #include "dtrace_cddl.h"
146 #include "dtrace_debug.c"
147 #endif
148 
149 #include "dtrace_xoroshiro128_plus.h"
150 
151 /*
152  * DTrace Tunable Variables
153  *
154  * The following variables may be tuned by adding a line to /etc/system that
155  * includes both the name of the DTrace module ("dtrace") and the name of the
156  * variable.  For example:
157  *
158  *   set dtrace:dtrace_destructive_disallow = 1
159  *
160  * In general, the only variables that one should be tuning this way are those
161  * that affect system-wide DTrace behavior, and for which the default behavior
162  * is undesirable.  Most of these variables are tunable on a per-consumer
163  * basis using DTrace options, and need not be tuned on a system-wide basis.
164  * When tuning these variables, avoid pathological values; while some attempt
165  * is made to verify the integrity of these variables, they are not considered
166  * part of the supported interface to DTrace, and they are therefore not
167  * checked comprehensively.  Further, these variables should not be tuned
168  * dynamically via "mdb -kw" or other means; they should only be tuned via
169  * /etc/system.
170  */
171 int		dtrace_destructive_disallow = 0;
172 #ifndef illumos
173 /* Positive logic version of dtrace_destructive_disallow for loader tunable */
174 int		dtrace_allow_destructive = 1;
175 #endif
176 dtrace_optval_t	dtrace_nonroot_maxsize = (16 * 1024 * 1024);
177 size_t		dtrace_difo_maxsize = (256 * 1024);
178 dtrace_optval_t	dtrace_dof_maxsize = (8 * 1024 * 1024);
179 size_t		dtrace_statvar_maxsize = (16 * 1024);
180 size_t		dtrace_actions_max = (16 * 1024);
181 size_t		dtrace_retain_max = 1024;
182 dtrace_optval_t	dtrace_helper_actions_max = 128;
183 dtrace_optval_t	dtrace_helper_providers_max = 32;
184 dtrace_optval_t	dtrace_dstate_defsize = (1 * 1024 * 1024);
185 size_t		dtrace_strsize_default = 256;
186 dtrace_optval_t	dtrace_cleanrate_default = 9900990;		/* 101 hz */
187 dtrace_optval_t	dtrace_cleanrate_min = 200000;			/* 5000 hz */
188 dtrace_optval_t	dtrace_cleanrate_max = (uint64_t)60 * NANOSEC;	/* 1/minute */
189 dtrace_optval_t	dtrace_aggrate_default = NANOSEC;		/* 1 hz */
190 dtrace_optval_t	dtrace_statusrate_default = NANOSEC;		/* 1 hz */
191 dtrace_optval_t dtrace_statusrate_max = (hrtime_t)10 * NANOSEC;	 /* 6/minute */
192 dtrace_optval_t	dtrace_switchrate_default = NANOSEC;		/* 1 hz */
193 dtrace_optval_t	dtrace_nspec_default = 1;
194 dtrace_optval_t	dtrace_specsize_default = 32 * 1024;
195 dtrace_optval_t dtrace_stackframes_default = 20;
196 dtrace_optval_t dtrace_ustackframes_default = 20;
197 dtrace_optval_t dtrace_jstackframes_default = 50;
198 dtrace_optval_t dtrace_jstackstrsize_default = 512;
199 int		dtrace_msgdsize_max = 128;
200 hrtime_t	dtrace_chill_max = MSEC2NSEC(500);		/* 500 ms */
201 hrtime_t	dtrace_chill_interval = NANOSEC;		/* 1000 ms */
202 int		dtrace_devdepth_max = 32;
203 int		dtrace_err_verbose;
204 hrtime_t	dtrace_deadman_interval = NANOSEC;
205 hrtime_t	dtrace_deadman_timeout = (hrtime_t)10 * NANOSEC;
206 hrtime_t	dtrace_deadman_user = (hrtime_t)30 * NANOSEC;
207 hrtime_t	dtrace_unregister_defunct_reap = (hrtime_t)60 * NANOSEC;
208 #ifndef illumos
209 int		dtrace_memstr_max = 4096;
210 #endif
211 
212 /*
213  * DTrace External Variables
214  *
215  * As dtrace(7D) is a kernel module, any DTrace variables are obviously
216  * available to DTrace consumers via the backtick (`) syntax.  One of these,
217  * dtrace_zero, is made deliberately so:  it is provided as a source of
218  * well-known, zero-filled memory.  While this variable is not documented,
219  * it is used by some translators as an implementation detail.
220  */
221 const char	dtrace_zero[256] = { 0 };	/* zero-filled memory */
222 
223 /*
224  * DTrace Internal Variables
225  */
226 #ifdef illumos
227 static dev_info_t	*dtrace_devi;		/* device info */
228 #endif
229 #ifdef illumos
230 static vmem_t		*dtrace_arena;		/* probe ID arena */
231 static vmem_t		*dtrace_minor;		/* minor number arena */
232 #else
233 static taskq_t		*dtrace_taskq;		/* task queue */
234 static struct unrhdr	*dtrace_arena;		/* Probe ID number.     */
235 #endif
236 static dtrace_probe_t	**dtrace_probes;	/* array of all probes */
237 static int		dtrace_nprobes;		/* number of probes */
238 static dtrace_provider_t *dtrace_provider;	/* provider list */
239 static dtrace_meta_t	*dtrace_meta_pid;	/* user-land meta provider */
240 static int		dtrace_opens;		/* number of opens */
241 static int		dtrace_helpers;		/* number of helpers */
242 static int		dtrace_getf;		/* number of unpriv getf()s */
243 #ifdef illumos
244 static void		*dtrace_softstate;	/* softstate pointer */
245 #endif
246 static dtrace_hash_t	*dtrace_bymod;		/* probes hashed by module */
247 static dtrace_hash_t	*dtrace_byfunc;		/* probes hashed by function */
248 static dtrace_hash_t	*dtrace_byname;		/* probes hashed by name */
249 static dtrace_toxrange_t *dtrace_toxrange;	/* toxic range array */
250 static int		dtrace_toxranges;	/* number of toxic ranges */
251 static int		dtrace_toxranges_max;	/* size of toxic range array */
252 static dtrace_anon_t	dtrace_anon;		/* anonymous enabling */
253 static kmem_cache_t	*dtrace_state_cache;	/* cache for dynamic state */
254 static uint64_t		dtrace_vtime_references; /* number of vtimestamp refs */
255 static kthread_t	*dtrace_panicked;	/* panicking thread */
256 static dtrace_ecb_t	*dtrace_ecb_create_cache; /* cached created ECB */
257 static dtrace_genid_t	dtrace_probegen;	/* current probe generation */
258 static dtrace_helpers_t *dtrace_deferred_pid;	/* deferred helper list */
259 static dtrace_enabling_t *dtrace_retained;	/* list of retained enablings */
260 static dtrace_genid_t	dtrace_retained_gen;	/* current retained enab gen */
261 static dtrace_dynvar_t	dtrace_dynhash_sink;	/* end of dynamic hash chains */
262 static int		dtrace_dynvar_failclean; /* dynvars failed to clean */
263 #ifndef illumos
264 static struct mtx	dtrace_unr_mtx;
265 MTX_SYSINIT(dtrace_unr_mtx, &dtrace_unr_mtx, "Unique resource identifier", MTX_DEF);
266 static eventhandler_tag	dtrace_kld_load_tag;
267 static eventhandler_tag	dtrace_kld_unload_try_tag;
268 #endif
269 
270 /*
271  * DTrace Locking
272  * DTrace is protected by three (relatively coarse-grained) locks:
273  *
274  * (1) dtrace_lock is required to manipulate essentially any DTrace state,
275  *     including enabling state, probes, ECBs, consumer state, helper state,
276  *     etc.  Importantly, dtrace_lock is _not_ required when in probe context;
277  *     probe context is lock-free -- synchronization is handled via the
278  *     dtrace_sync() cross call mechanism.
279  *
280  * (2) dtrace_provider_lock is required when manipulating provider state, or
281  *     when provider state must be held constant.
282  *
283  * (3) dtrace_meta_lock is required when manipulating meta provider state, or
284  *     when meta provider state must be held constant.
285  *
286  * The lock ordering between these three locks is dtrace_meta_lock before
287  * dtrace_provider_lock before dtrace_lock.  (In particular, there are
288  * several places where dtrace_provider_lock is held by the framework as it
289  * calls into the providers -- which then call back into the framework,
290  * grabbing dtrace_lock.)
291  *
292  * There are two other locks in the mix:  mod_lock and cpu_lock.  With respect
293  * to dtrace_provider_lock and dtrace_lock, cpu_lock continues its historical
294  * role as a coarse-grained lock; it is acquired before both of these locks.
295  * With respect to dtrace_meta_lock, its behavior is stranger:  cpu_lock must
296  * be acquired _between_ dtrace_meta_lock and any other DTrace locks.
297  * mod_lock is similar with respect to dtrace_provider_lock in that it must be
298  * acquired _between_ dtrace_provider_lock and dtrace_lock.
299  */
300 static kmutex_t		dtrace_lock;		/* probe state lock */
301 static kmutex_t		dtrace_provider_lock;	/* provider state lock */
302 static kmutex_t		dtrace_meta_lock;	/* meta-provider state lock */
303 
304 #ifndef illumos
305 /* XXX FreeBSD hacks. */
306 #define cr_suid		cr_svuid
307 #define cr_sgid		cr_svgid
308 #define	ipaddr_t	in_addr_t
309 #define mod_modname	pathname
310 #define vuprintf	vprintf
311 #ifndef crgetzoneid
312 #define crgetzoneid(_a)        0
313 #endif
314 #define ttoproc(_a)	((_a)->td_proc)
315 #define SNOCD		0
316 #define CPU_ON_INTR(_a)	0
317 
318 #define PRIV_EFFECTIVE		(1 << 0)
319 #define PRIV_DTRACE_KERNEL	(1 << 1)
320 #define PRIV_DTRACE_PROC	(1 << 2)
321 #define PRIV_DTRACE_USER	(1 << 3)
322 #define PRIV_PROC_OWNER		(1 << 4)
323 #define PRIV_PROC_ZONE		(1 << 5)
324 #define PRIV_ALL		~0
325 
326 SYSCTL_DECL(_debug_dtrace);
327 SYSCTL_DECL(_kern_dtrace);
328 #endif
329 
330 #ifdef illumos
331 #define curcpu	CPU->cpu_id
332 #endif
333 
334 
335 /*
336  * DTrace Provider Variables
337  *
338  * These are the variables relating to DTrace as a provider (that is, the
339  * provider of the BEGIN, END, and ERROR probes).
340  */
341 static dtrace_pattr_t	dtrace_provider_attr = {
342 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
343 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
344 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
345 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
346 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
347 };
348 
349 static void
350 dtrace_nullop(void)
351 {}
352 
353 static dtrace_pops_t dtrace_provider_ops = {
354 	.dtps_provide =	(void (*)(void *, dtrace_probedesc_t *))dtrace_nullop,
355 	.dtps_provide_module =	(void (*)(void *, modctl_t *))dtrace_nullop,
356 	.dtps_enable =	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
357 	.dtps_disable =	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
358 	.dtps_suspend =	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
359 	.dtps_resume =	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
360 	.dtps_getargdesc =	NULL,
361 	.dtps_getargval =	NULL,
362 	.dtps_usermode =	NULL,
363 	.dtps_destroy =	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
364 };
365 
366 static dtrace_id_t	dtrace_probeid_begin;	/* special BEGIN probe */
367 static dtrace_id_t	dtrace_probeid_end;	/* special END probe */
368 dtrace_id_t		dtrace_probeid_error;	/* special ERROR probe */
369 
370 /*
371  * DTrace Helper Tracing Variables
372  *
373  * These variables should be set dynamically to enable helper tracing.  The
374  * only variables that should be set are dtrace_helptrace_enable (which should
375  * be set to a non-zero value to allocate helper tracing buffers on the next
376  * open of /dev/dtrace) and dtrace_helptrace_disable (which should be set to a
377  * non-zero value to deallocate helper tracing buffers on the next close of
378  * /dev/dtrace).  When (and only when) helper tracing is disabled, the
379  * buffer size may also be set via dtrace_helptrace_bufsize.
380  */
381 int			dtrace_helptrace_enable = 0;
382 int			dtrace_helptrace_disable = 0;
383 int			dtrace_helptrace_bufsize = 16 * 1024 * 1024;
384 uint32_t		dtrace_helptrace_nlocals;
385 static dtrace_helptrace_t *dtrace_helptrace_buffer;
386 static uint32_t		dtrace_helptrace_next = 0;
387 static int		dtrace_helptrace_wrapped = 0;
388 
389 /*
390  * DTrace Error Hashing
391  *
392  * On DEBUG kernels, DTrace will track the errors that has seen in a hash
393  * table.  This is very useful for checking coverage of tests that are
394  * expected to induce DIF or DOF processing errors, and may be useful for
395  * debugging problems in the DIF code generator or in DOF generation .  The
396  * error hash may be examined with the ::dtrace_errhash MDB dcmd.
397  */
398 #ifdef DEBUG
399 static dtrace_errhash_t	dtrace_errhash[DTRACE_ERRHASHSZ];
400 static const char *dtrace_errlast;
401 static kthread_t *dtrace_errthread;
402 static kmutex_t dtrace_errlock;
403 #endif
404 
405 /*
406  * DTrace Macros and Constants
407  *
408  * These are various macros that are useful in various spots in the
409  * implementation, along with a few random constants that have no meaning
410  * outside of the implementation.  There is no real structure to this cpp
411  * mishmash -- but is there ever?
412  */
413 #define	DTRACE_HASHSTR(hash, probe)	\
414 	dtrace_hash_str(*((char **)((uintptr_t)(probe) + (hash)->dth_stroffs)))
415 
416 #define	DTRACE_HASHNEXT(hash, probe)	\
417 	(dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_nextoffs)
418 
419 #define	DTRACE_HASHPREV(hash, probe)	\
420 	(dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_prevoffs)
421 
422 #define	DTRACE_HASHEQ(hash, lhs, rhs)	\
423 	(strcmp(*((char **)((uintptr_t)(lhs) + (hash)->dth_stroffs)), \
424 	    *((char **)((uintptr_t)(rhs) + (hash)->dth_stroffs))) == 0)
425 
426 #define	DTRACE_AGGHASHSIZE_SLEW		17
427 
428 #define	DTRACE_V4MAPPED_OFFSET		(sizeof (uint32_t) * 3)
429 
430 /*
431  * The key for a thread-local variable consists of the lower 61 bits of the
432  * t_did, plus the 3 bits of the highest active interrupt above LOCK_LEVEL.
433  * We add DIF_VARIABLE_MAX to t_did to assure that the thread key is never
434  * equal to a variable identifier.  This is necessary (but not sufficient) to
435  * assure that global associative arrays never collide with thread-local
436  * variables.  To guarantee that they cannot collide, we must also define the
437  * order for keying dynamic variables.  That order is:
438  *
439  *   [ key0 ] ... [ keyn ] [ variable-key ] [ tls-key ]
440  *
441  * Because the variable-key and the tls-key are in orthogonal spaces, there is
442  * no way for a global variable key signature to match a thread-local key
443  * signature.
444  */
445 #ifdef illumos
446 #define	DTRACE_TLS_THRKEY(where) { \
447 	uint_t intr = 0; \
448 	uint_t actv = CPU->cpu_intr_actv >> (LOCK_LEVEL + 1); \
449 	for (; actv; actv >>= 1) \
450 		intr++; \
451 	ASSERT(intr < (1 << 3)); \
452 	(where) = ((curthread->t_did + DIF_VARIABLE_MAX) & \
453 	    (((uint64_t)1 << 61) - 1)) | ((uint64_t)intr << 61); \
454 }
455 #else
456 #define	DTRACE_TLS_THRKEY(where) { \
457 	solaris_cpu_t *_c = &solaris_cpu[curcpu]; \
458 	uint_t intr = 0; \
459 	uint_t actv = _c->cpu_intr_actv; \
460 	for (; actv; actv >>= 1) \
461 		intr++; \
462 	ASSERT(intr < (1 << 3)); \
463 	(where) = ((curthread->td_tid + DIF_VARIABLE_MAX) & \
464 	    (((uint64_t)1 << 61) - 1)) | ((uint64_t)intr << 61); \
465 }
466 #endif
467 
468 #define	DT_BSWAP_8(x)	((x) & 0xff)
469 #define	DT_BSWAP_16(x)	((DT_BSWAP_8(x) << 8) | DT_BSWAP_8((x) >> 8))
470 #define	DT_BSWAP_32(x)	((DT_BSWAP_16(x) << 16) | DT_BSWAP_16((x) >> 16))
471 #define	DT_BSWAP_64(x)	((DT_BSWAP_32(x) << 32) | DT_BSWAP_32((x) >> 32))
472 
473 #define	DT_MASK_LO 0x00000000FFFFFFFFULL
474 
475 #define	DTRACE_STORE(type, tomax, offset, what) \
476 	*((type *)((uintptr_t)(tomax) + (uintptr_t)offset)) = (type)(what);
477 
478 #ifndef __x86
479 #define	DTRACE_ALIGNCHECK(addr, size, flags)				\
480 	if (addr & (size - 1)) {					\
481 		*flags |= CPU_DTRACE_BADALIGN;				\
482 		cpu_core[curcpu].cpuc_dtrace_illval = addr;	\
483 		return (0);						\
484 	}
485 #else
486 #define	DTRACE_ALIGNCHECK(addr, size, flags)
487 #endif
488 
489 /*
490  * Test whether a range of memory starting at testaddr of size testsz falls
491  * within the range of memory described by addr, sz.  We take care to avoid
492  * problems with overflow and underflow of the unsigned quantities, and
493  * disallow all negative sizes.  Ranges of size 0 are allowed.
494  */
495 #define	DTRACE_INRANGE(testaddr, testsz, baseaddr, basesz) \
496 	((testaddr) - (uintptr_t)(baseaddr) < (basesz) && \
497 	(testaddr) + (testsz) - (uintptr_t)(baseaddr) <= (basesz) && \
498 	(testaddr) + (testsz) >= (testaddr))
499 
500 #define	DTRACE_RANGE_REMAIN(remp, addr, baseaddr, basesz)		\
501 do {									\
502 	if ((remp) != NULL) {						\
503 		*(remp) = (uintptr_t)(baseaddr) + (basesz) - (addr);	\
504 	}								\
505 } while (0)
506 
507 
508 /*
509  * Test whether alloc_sz bytes will fit in the scratch region.  We isolate
510  * alloc_sz on the righthand side of the comparison in order to avoid overflow
511  * or underflow in the comparison with it.  This is simpler than the INRANGE
512  * check above, because we know that the dtms_scratch_ptr is valid in the
513  * range.  Allocations of size zero are allowed.
514  */
515 #define	DTRACE_INSCRATCH(mstate, alloc_sz) \
516 	((mstate)->dtms_scratch_base + (mstate)->dtms_scratch_size - \
517 	(mstate)->dtms_scratch_ptr >= (alloc_sz))
518 
519 #define	DTRACE_LOADFUNC(bits)						\
520 /*CSTYLED*/								\
521 uint##bits##_t								\
522 dtrace_load##bits(uintptr_t addr)					\
523 {									\
524 	size_t size = bits / NBBY;					\
525 	/*CSTYLED*/							\
526 	uint##bits##_t rval;						\
527 	int i;								\
528 	volatile uint16_t *flags = (volatile uint16_t *)		\
529 	    &cpu_core[curcpu].cpuc_dtrace_flags;			\
530 									\
531 	DTRACE_ALIGNCHECK(addr, size, flags);				\
532 									\
533 	for (i = 0; i < dtrace_toxranges; i++) {			\
534 		if (addr >= dtrace_toxrange[i].dtt_limit)		\
535 			continue;					\
536 									\
537 		if (addr + size <= dtrace_toxrange[i].dtt_base)		\
538 			continue;					\
539 									\
540 		/*							\
541 		 * This address falls within a toxic region; return 0.	\
542 		 */							\
543 		*flags |= CPU_DTRACE_BADADDR;				\
544 		cpu_core[curcpu].cpuc_dtrace_illval = addr;		\
545 		return (0);						\
546 	}								\
547 									\
548 	*flags |= CPU_DTRACE_NOFAULT;					\
549 	/*CSTYLED*/							\
550 	rval = *((volatile uint##bits##_t *)addr);			\
551 	*flags &= ~CPU_DTRACE_NOFAULT;					\
552 									\
553 	return (!(*flags & CPU_DTRACE_FAULT) ? rval : 0);		\
554 }
555 
556 #ifdef _LP64
557 #define	dtrace_loadptr	dtrace_load64
558 #else
559 #define	dtrace_loadptr	dtrace_load32
560 #endif
561 
562 #define	DTRACE_DYNHASH_FREE	0
563 #define	DTRACE_DYNHASH_SINK	1
564 #define	DTRACE_DYNHASH_VALID	2
565 
566 #define	DTRACE_MATCH_NEXT	0
567 #define	DTRACE_MATCH_DONE	1
568 #define	DTRACE_ANCHORED(probe)	((probe)->dtpr_func[0] != '\0')
569 #define	DTRACE_STATE_ALIGN	64
570 
571 #define	DTRACE_FLAGS2FLT(flags)						\
572 	(((flags) & CPU_DTRACE_BADADDR) ? DTRACEFLT_BADADDR :		\
573 	((flags) & CPU_DTRACE_ILLOP) ? DTRACEFLT_ILLOP :		\
574 	((flags) & CPU_DTRACE_DIVZERO) ? DTRACEFLT_DIVZERO :		\
575 	((flags) & CPU_DTRACE_KPRIV) ? DTRACEFLT_KPRIV :		\
576 	((flags) & CPU_DTRACE_UPRIV) ? DTRACEFLT_UPRIV :		\
577 	((flags) & CPU_DTRACE_TUPOFLOW) ?  DTRACEFLT_TUPOFLOW :		\
578 	((flags) & CPU_DTRACE_BADALIGN) ?  DTRACEFLT_BADALIGN :		\
579 	((flags) & CPU_DTRACE_NOSCRATCH) ?  DTRACEFLT_NOSCRATCH :	\
580 	((flags) & CPU_DTRACE_BADSTACK) ?  DTRACEFLT_BADSTACK :		\
581 	DTRACEFLT_UNKNOWN)
582 
583 #define	DTRACEACT_ISSTRING(act)						\
584 	((act)->dta_kind == DTRACEACT_DIFEXPR &&			\
585 	(act)->dta_difo->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING)
586 
587 /* Function prototype definitions: */
588 static size_t dtrace_strlen(const char *, size_t);
589 static dtrace_probe_t *dtrace_probe_lookup_id(dtrace_id_t id);
590 static void dtrace_enabling_provide(dtrace_provider_t *);
591 static int dtrace_enabling_match(dtrace_enabling_t *, int *);
592 static void dtrace_enabling_matchall(void);
593 static void dtrace_enabling_reap(void);
594 static dtrace_state_t *dtrace_anon_grab(void);
595 static uint64_t dtrace_helper(int, dtrace_mstate_t *,
596     dtrace_state_t *, uint64_t, uint64_t);
597 static dtrace_helpers_t *dtrace_helpers_create(proc_t *);
598 static void dtrace_buffer_drop(dtrace_buffer_t *);
599 static int dtrace_buffer_consumed(dtrace_buffer_t *, hrtime_t when);
600 static intptr_t dtrace_buffer_reserve(dtrace_buffer_t *, size_t, size_t,
601     dtrace_state_t *, dtrace_mstate_t *);
602 static int dtrace_state_option(dtrace_state_t *, dtrace_optid_t,
603     dtrace_optval_t);
604 static int dtrace_ecb_create_enable(dtrace_probe_t *, void *);
605 static void dtrace_helper_provider_destroy(dtrace_helper_provider_t *);
606 uint16_t dtrace_load16(uintptr_t);
607 uint32_t dtrace_load32(uintptr_t);
608 uint64_t dtrace_load64(uintptr_t);
609 uint8_t dtrace_load8(uintptr_t);
610 void dtrace_dynvar_clean(dtrace_dstate_t *);
611 dtrace_dynvar_t *dtrace_dynvar(dtrace_dstate_t *, uint_t, dtrace_key_t *,
612     size_t, dtrace_dynvar_op_t, dtrace_mstate_t *, dtrace_vstate_t *);
613 uintptr_t dtrace_dif_varstr(uintptr_t, dtrace_state_t *, dtrace_mstate_t *);
614 static int dtrace_priv_proc(dtrace_state_t *);
615 static void dtrace_getf_barrier(void);
616 static int dtrace_canload_remains(uint64_t, size_t, size_t *,
617     dtrace_mstate_t *, dtrace_vstate_t *);
618 static int dtrace_canstore_remains(uint64_t, size_t, size_t *,
619     dtrace_mstate_t *, dtrace_vstate_t *);
620 
621 /*
622  * DTrace Probe Context Functions
623  *
624  * These functions are called from probe context.  Because probe context is
625  * any context in which C may be called, arbitrarily locks may be held,
626  * interrupts may be disabled, we may be in arbitrary dispatched state, etc.
627  * As a result, functions called from probe context may only call other DTrace
628  * support functions -- they may not interact at all with the system at large.
629  * (Note that the ASSERT macro is made probe-context safe by redefining it in
630  * terms of dtrace_assfail(), a probe-context safe function.) If arbitrary
631  * loads are to be performed from probe context, they _must_ be in terms of
632  * the safe dtrace_load*() variants.
633  *
634  * Some functions in this block are not actually called from probe context;
635  * for these functions, there will be a comment above the function reading
636  * "Note:  not called from probe context."
637  */
638 void
639 dtrace_panic(const char *format, ...)
640 {
641 	va_list alist;
642 
643 	va_start(alist, format);
644 #ifdef __FreeBSD__
645 	vpanic(format, alist);
646 #else
647 	dtrace_vpanic(format, alist);
648 #endif
649 	va_end(alist);
650 }
651 
652 int
653 dtrace_assfail(const char *a, const char *f, int l)
654 {
655 	dtrace_panic("assertion failed: %s, file: %s, line: %d", a, f, l);
656 
657 	/*
658 	 * We just need something here that even the most clever compiler
659 	 * cannot optimize away.
660 	 */
661 	return (a[(uintptr_t)f]);
662 }
663 
664 /*
665  * Atomically increment a specified error counter from probe context.
666  */
667 static void
668 dtrace_error(uint32_t *counter)
669 {
670 	/*
671 	 * Most counters stored to in probe context are per-CPU counters.
672 	 * However, there are some error conditions that are sufficiently
673 	 * arcane that they don't merit per-CPU storage.  If these counters
674 	 * are incremented concurrently on different CPUs, scalability will be
675 	 * adversely affected -- but we don't expect them to be white-hot in a
676 	 * correctly constructed enabling...
677 	 */
678 	uint32_t oval, nval;
679 
680 	do {
681 		oval = *counter;
682 
683 		if ((nval = oval + 1) == 0) {
684 			/*
685 			 * If the counter would wrap, set it to 1 -- assuring
686 			 * that the counter is never zero when we have seen
687 			 * errors.  (The counter must be 32-bits because we
688 			 * aren't guaranteed a 64-bit compare&swap operation.)
689 			 * To save this code both the infamy of being fingered
690 			 * by a priggish news story and the indignity of being
691 			 * the target of a neo-puritan witch trial, we're
692 			 * carefully avoiding any colorful description of the
693 			 * likelihood of this condition -- but suffice it to
694 			 * say that it is only slightly more likely than the
695 			 * overflow of predicate cache IDs, as discussed in
696 			 * dtrace_predicate_create().
697 			 */
698 			nval = 1;
699 		}
700 	} while (dtrace_cas32(counter, oval, nval) != oval);
701 }
702 
703 /*
704  * Use the DTRACE_LOADFUNC macro to define functions for each of loading a
705  * uint8_t, a uint16_t, a uint32_t and a uint64_t.
706  */
707 /* BEGIN CSTYLED */
708 DTRACE_LOADFUNC(8)
709 DTRACE_LOADFUNC(16)
710 DTRACE_LOADFUNC(32)
711 DTRACE_LOADFUNC(64)
712 /* END CSTYLED */
713 
714 static int
715 dtrace_inscratch(uintptr_t dest, size_t size, dtrace_mstate_t *mstate)
716 {
717 	if (dest < mstate->dtms_scratch_base)
718 		return (0);
719 
720 	if (dest + size < dest)
721 		return (0);
722 
723 	if (dest + size > mstate->dtms_scratch_ptr)
724 		return (0);
725 
726 	return (1);
727 }
728 
729 static int
730 dtrace_canstore_statvar(uint64_t addr, size_t sz, size_t *remain,
731     dtrace_statvar_t **svars, int nsvars)
732 {
733 	int i;
734 	size_t maxglobalsize, maxlocalsize;
735 
736 	if (nsvars == 0)
737 		return (0);
738 
739 	maxglobalsize = dtrace_statvar_maxsize + sizeof (uint64_t);
740 	maxlocalsize = maxglobalsize * NCPU;
741 
742 	for (i = 0; i < nsvars; i++) {
743 		dtrace_statvar_t *svar = svars[i];
744 		uint8_t scope;
745 		size_t size;
746 
747 		if (svar == NULL || (size = svar->dtsv_size) == 0)
748 			continue;
749 
750 		scope = svar->dtsv_var.dtdv_scope;
751 
752 		/*
753 		 * We verify that our size is valid in the spirit of providing
754 		 * defense in depth:  we want to prevent attackers from using
755 		 * DTrace to escalate an orthogonal kernel heap corruption bug
756 		 * into the ability to store to arbitrary locations in memory.
757 		 */
758 		VERIFY((scope == DIFV_SCOPE_GLOBAL && size <= maxglobalsize) ||
759 		    (scope == DIFV_SCOPE_LOCAL && size <= maxlocalsize));
760 
761 		if (DTRACE_INRANGE(addr, sz, svar->dtsv_data,
762 		    svar->dtsv_size)) {
763 			DTRACE_RANGE_REMAIN(remain, addr, svar->dtsv_data,
764 			    svar->dtsv_size);
765 			return (1);
766 		}
767 	}
768 
769 	return (0);
770 }
771 
772 /*
773  * Check to see if the address is within a memory region to which a store may
774  * be issued.  This includes the DTrace scratch areas, and any DTrace variable
775  * region.  The caller of dtrace_canstore() is responsible for performing any
776  * alignment checks that are needed before stores are actually executed.
777  */
778 static int
779 dtrace_canstore(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
780     dtrace_vstate_t *vstate)
781 {
782 	return (dtrace_canstore_remains(addr, sz, NULL, mstate, vstate));
783 }
784 
785 /*
786  * Implementation of dtrace_canstore which communicates the upper bound of the
787  * allowed memory region.
788  */
789 static int
790 dtrace_canstore_remains(uint64_t addr, size_t sz, size_t *remain,
791     dtrace_mstate_t *mstate, dtrace_vstate_t *vstate)
792 {
793 	/*
794 	 * First, check to see if the address is in scratch space...
795 	 */
796 	if (DTRACE_INRANGE(addr, sz, mstate->dtms_scratch_base,
797 	    mstate->dtms_scratch_size)) {
798 		DTRACE_RANGE_REMAIN(remain, addr, mstate->dtms_scratch_base,
799 		    mstate->dtms_scratch_size);
800 		return (1);
801 	}
802 
803 	/*
804 	 * Now check to see if it's a dynamic variable.  This check will pick
805 	 * up both thread-local variables and any global dynamically-allocated
806 	 * variables.
807 	 */
808 	if (DTRACE_INRANGE(addr, sz, vstate->dtvs_dynvars.dtds_base,
809 	    vstate->dtvs_dynvars.dtds_size)) {
810 		dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
811 		uintptr_t base = (uintptr_t)dstate->dtds_base +
812 		    (dstate->dtds_hashsize * sizeof (dtrace_dynhash_t));
813 		uintptr_t chunkoffs;
814 		dtrace_dynvar_t *dvar;
815 
816 		/*
817 		 * Before we assume that we can store here, we need to make
818 		 * sure that it isn't in our metadata -- storing to our
819 		 * dynamic variable metadata would corrupt our state.  For
820 		 * the range to not include any dynamic variable metadata,
821 		 * it must:
822 		 *
823 		 *	(1) Start above the hash table that is at the base of
824 		 *	the dynamic variable space
825 		 *
826 		 *	(2) Have a starting chunk offset that is beyond the
827 		 *	dtrace_dynvar_t that is at the base of every chunk
828 		 *
829 		 *	(3) Not span a chunk boundary
830 		 *
831 		 *	(4) Not be in the tuple space of a dynamic variable
832 		 *
833 		 */
834 		if (addr < base)
835 			return (0);
836 
837 		chunkoffs = (addr - base) % dstate->dtds_chunksize;
838 
839 		if (chunkoffs < sizeof (dtrace_dynvar_t))
840 			return (0);
841 
842 		if (chunkoffs + sz > dstate->dtds_chunksize)
843 			return (0);
844 
845 		dvar = (dtrace_dynvar_t *)((uintptr_t)addr - chunkoffs);
846 
847 		if (dvar->dtdv_hashval == DTRACE_DYNHASH_FREE)
848 			return (0);
849 
850 		if (chunkoffs < sizeof (dtrace_dynvar_t) +
851 		    ((dvar->dtdv_tuple.dtt_nkeys - 1) * sizeof (dtrace_key_t)))
852 			return (0);
853 
854 		DTRACE_RANGE_REMAIN(remain, addr, dvar, dstate->dtds_chunksize);
855 		return (1);
856 	}
857 
858 	/*
859 	 * Finally, check the static local and global variables.  These checks
860 	 * take the longest, so we perform them last.
861 	 */
862 	if (dtrace_canstore_statvar(addr, sz, remain,
863 	    vstate->dtvs_locals, vstate->dtvs_nlocals))
864 		return (1);
865 
866 	if (dtrace_canstore_statvar(addr, sz, remain,
867 	    vstate->dtvs_globals, vstate->dtvs_nglobals))
868 		return (1);
869 
870 	return (0);
871 }
872 
873 
874 /*
875  * Convenience routine to check to see if the address is within a memory
876  * region in which a load may be issued given the user's privilege level;
877  * if not, it sets the appropriate error flags and loads 'addr' into the
878  * illegal value slot.
879  *
880  * DTrace subroutines (DIF_SUBR_*) should use this helper to implement
881  * appropriate memory access protection.
882  */
883 static int
884 dtrace_canload(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
885     dtrace_vstate_t *vstate)
886 {
887 	return (dtrace_canload_remains(addr, sz, NULL, mstate, vstate));
888 }
889 
890 /*
891  * Implementation of dtrace_canload which communicates the uppoer bound of the
892  * allowed memory region.
893  */
894 static int
895 dtrace_canload_remains(uint64_t addr, size_t sz, size_t *remain,
896     dtrace_mstate_t *mstate, dtrace_vstate_t *vstate)
897 {
898 	volatile uintptr_t *illval = &cpu_core[curcpu].cpuc_dtrace_illval;
899 	file_t *fp;
900 
901 	/*
902 	 * If we hold the privilege to read from kernel memory, then
903 	 * everything is readable.
904 	 */
905 	if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0) {
906 		DTRACE_RANGE_REMAIN(remain, addr, addr, sz);
907 		return (1);
908 	}
909 
910 	/*
911 	 * You can obviously read that which you can store.
912 	 */
913 	if (dtrace_canstore_remains(addr, sz, remain, mstate, vstate))
914 		return (1);
915 
916 	/*
917 	 * We're allowed to read from our own string table.
918 	 */
919 	if (DTRACE_INRANGE(addr, sz, mstate->dtms_difo->dtdo_strtab,
920 	    mstate->dtms_difo->dtdo_strlen)) {
921 		DTRACE_RANGE_REMAIN(remain, addr,
922 		    mstate->dtms_difo->dtdo_strtab,
923 		    mstate->dtms_difo->dtdo_strlen);
924 		return (1);
925 	}
926 
927 	if (vstate->dtvs_state != NULL &&
928 	    dtrace_priv_proc(vstate->dtvs_state)) {
929 		proc_t *p;
930 
931 		/*
932 		 * When we have privileges to the current process, there are
933 		 * several context-related kernel structures that are safe to
934 		 * read, even absent the privilege to read from kernel memory.
935 		 * These reads are safe because these structures contain only
936 		 * state that (1) we're permitted to read, (2) is harmless or
937 		 * (3) contains pointers to additional kernel state that we're
938 		 * not permitted to read (and as such, do not present an
939 		 * opportunity for privilege escalation).  Finally (and
940 		 * critically), because of the nature of their relation with
941 		 * the current thread context, the memory associated with these
942 		 * structures cannot change over the duration of probe context,
943 		 * and it is therefore impossible for this memory to be
944 		 * deallocated and reallocated as something else while it's
945 		 * being operated upon.
946 		 */
947 		if (DTRACE_INRANGE(addr, sz, curthread, sizeof (kthread_t))) {
948 			DTRACE_RANGE_REMAIN(remain, addr, curthread,
949 			    sizeof (kthread_t));
950 			return (1);
951 		}
952 
953 		if ((p = curthread->t_procp) != NULL && DTRACE_INRANGE(addr,
954 		    sz, curthread->t_procp, sizeof (proc_t))) {
955 			DTRACE_RANGE_REMAIN(remain, addr, curthread->t_procp,
956 			    sizeof (proc_t));
957 			return (1);
958 		}
959 
960 		if (curthread->t_cred != NULL && DTRACE_INRANGE(addr, sz,
961 		    curthread->t_cred, sizeof (cred_t))) {
962 			DTRACE_RANGE_REMAIN(remain, addr, curthread->t_cred,
963 			    sizeof (cred_t));
964 			return (1);
965 		}
966 
967 #ifdef illumos
968 		if (p != NULL && p->p_pidp != NULL && DTRACE_INRANGE(addr, sz,
969 		    &(p->p_pidp->pid_id), sizeof (pid_t))) {
970 			DTRACE_RANGE_REMAIN(remain, addr, &(p->p_pidp->pid_id),
971 			    sizeof (pid_t));
972 			return (1);
973 		}
974 
975 		if (curthread->t_cpu != NULL && DTRACE_INRANGE(addr, sz,
976 		    curthread->t_cpu, offsetof(cpu_t, cpu_pause_thread))) {
977 			DTRACE_RANGE_REMAIN(remain, addr, curthread->t_cpu,
978 			    offsetof(cpu_t, cpu_pause_thread));
979 			return (1);
980 		}
981 #endif
982 	}
983 
984 	if ((fp = mstate->dtms_getf) != NULL) {
985 		uintptr_t psz = sizeof (void *);
986 		vnode_t *vp;
987 		vnodeops_t *op;
988 
989 		/*
990 		 * When getf() returns a file_t, the enabling is implicitly
991 		 * granted the (transient) right to read the returned file_t
992 		 * as well as the v_path and v_op->vnop_name of the underlying
993 		 * vnode.  These accesses are allowed after a successful
994 		 * getf() because the members that they refer to cannot change
995 		 * once set -- and the barrier logic in the kernel's closef()
996 		 * path assures that the file_t and its referenced vode_t
997 		 * cannot themselves be stale (that is, it impossible for
998 		 * either dtms_getf itself or its f_vnode member to reference
999 		 * freed memory).
1000 		 */
1001 		if (DTRACE_INRANGE(addr, sz, fp, sizeof (file_t))) {
1002 			DTRACE_RANGE_REMAIN(remain, addr, fp, sizeof (file_t));
1003 			return (1);
1004 		}
1005 
1006 		if ((vp = fp->f_vnode) != NULL) {
1007 			size_t slen;
1008 #ifdef illumos
1009 			if (DTRACE_INRANGE(addr, sz, &vp->v_path, psz)) {
1010 				DTRACE_RANGE_REMAIN(remain, addr, &vp->v_path,
1011 				    psz);
1012 				return (1);
1013 			}
1014 			slen = strlen(vp->v_path) + 1;
1015 			if (DTRACE_INRANGE(addr, sz, vp->v_path, slen)) {
1016 				DTRACE_RANGE_REMAIN(remain, addr, vp->v_path,
1017 				    slen);
1018 				return (1);
1019 			}
1020 #endif
1021 
1022 			if (DTRACE_INRANGE(addr, sz, &vp->v_op, psz)) {
1023 				DTRACE_RANGE_REMAIN(remain, addr, &vp->v_op,
1024 				    psz);
1025 				return (1);
1026 			}
1027 
1028 #ifdef illumos
1029 			if ((op = vp->v_op) != NULL &&
1030 			    DTRACE_INRANGE(addr, sz, &op->vnop_name, psz)) {
1031 				DTRACE_RANGE_REMAIN(remain, addr,
1032 				    &op->vnop_name, psz);
1033 				return (1);
1034 			}
1035 
1036 			if (op != NULL && op->vnop_name != NULL &&
1037 			    DTRACE_INRANGE(addr, sz, op->vnop_name,
1038 			    (slen = strlen(op->vnop_name) + 1))) {
1039 				DTRACE_RANGE_REMAIN(remain, addr,
1040 				    op->vnop_name, slen);
1041 				return (1);
1042 			}
1043 #endif
1044 		}
1045 	}
1046 
1047 	DTRACE_CPUFLAG_SET(CPU_DTRACE_KPRIV);
1048 	*illval = addr;
1049 	return (0);
1050 }
1051 
1052 /*
1053  * Convenience routine to check to see if a given string is within a memory
1054  * region in which a load may be issued given the user's privilege level;
1055  * this exists so that we don't need to issue unnecessary dtrace_strlen()
1056  * calls in the event that the user has all privileges.
1057  */
1058 static int
1059 dtrace_strcanload(uint64_t addr, size_t sz, size_t *remain,
1060     dtrace_mstate_t *mstate, dtrace_vstate_t *vstate)
1061 {
1062 	size_t rsize;
1063 
1064 	/*
1065 	 * If we hold the privilege to read from kernel memory, then
1066 	 * everything is readable.
1067 	 */
1068 	if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0) {
1069 		DTRACE_RANGE_REMAIN(remain, addr, addr, sz);
1070 		return (1);
1071 	}
1072 
1073 	/*
1074 	 * Even if the caller is uninterested in querying the remaining valid
1075 	 * range, it is required to ensure that the access is allowed.
1076 	 */
1077 	if (remain == NULL) {
1078 		remain = &rsize;
1079 	}
1080 	if (dtrace_canload_remains(addr, 0, remain, mstate, vstate)) {
1081 		size_t strsz;
1082 		/*
1083 		 * Perform the strlen after determining the length of the
1084 		 * memory region which is accessible.  This prevents timing
1085 		 * information from being used to find NULs in memory which is
1086 		 * not accessible to the caller.
1087 		 */
1088 		strsz = 1 + dtrace_strlen((char *)(uintptr_t)addr,
1089 		    MIN(sz, *remain));
1090 		if (strsz <= *remain) {
1091 			return (1);
1092 		}
1093 	}
1094 
1095 	return (0);
1096 }
1097 
1098 /*
1099  * Convenience routine to check to see if a given variable is within a memory
1100  * region in which a load may be issued given the user's privilege level.
1101  */
1102 static int
1103 dtrace_vcanload(void *src, dtrace_diftype_t *type, size_t *remain,
1104     dtrace_mstate_t *mstate, dtrace_vstate_t *vstate)
1105 {
1106 	size_t sz;
1107 	ASSERT(type->dtdt_flags & DIF_TF_BYREF);
1108 
1109 	/*
1110 	 * Calculate the max size before performing any checks since even
1111 	 * DTRACE_ACCESS_KERNEL-credentialed callers expect that this function
1112 	 * return the max length via 'remain'.
1113 	 */
1114 	if (type->dtdt_kind == DIF_TYPE_STRING) {
1115 		dtrace_state_t *state = vstate->dtvs_state;
1116 
1117 		if (state != NULL) {
1118 			sz = state->dts_options[DTRACEOPT_STRSIZE];
1119 		} else {
1120 			/*
1121 			 * In helper context, we have a NULL state; fall back
1122 			 * to using the system-wide default for the string size
1123 			 * in this case.
1124 			 */
1125 			sz = dtrace_strsize_default;
1126 		}
1127 	} else {
1128 		sz = type->dtdt_size;
1129 	}
1130 
1131 	/*
1132 	 * If we hold the privilege to read from kernel memory, then
1133 	 * everything is readable.
1134 	 */
1135 	if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0) {
1136 		DTRACE_RANGE_REMAIN(remain, (uintptr_t)src, src, sz);
1137 		return (1);
1138 	}
1139 
1140 	if (type->dtdt_kind == DIF_TYPE_STRING) {
1141 		return (dtrace_strcanload((uintptr_t)src, sz, remain, mstate,
1142 		    vstate));
1143 	}
1144 	return (dtrace_canload_remains((uintptr_t)src, sz, remain, mstate,
1145 	    vstate));
1146 }
1147 
1148 /*
1149  * Convert a string to a signed integer using safe loads.
1150  *
1151  * NOTE: This function uses various macros from strtolctype.h to manipulate
1152  * digit values, etc -- these have all been checked to ensure they make
1153  * no additional function calls.
1154  */
1155 static int64_t
1156 dtrace_strtoll(char *input, int base, size_t limit)
1157 {
1158 	uintptr_t pos = (uintptr_t)input;
1159 	int64_t val = 0;
1160 	int x;
1161 	boolean_t neg = B_FALSE;
1162 	char c, cc, ccc;
1163 	uintptr_t end = pos + limit;
1164 
1165 	/*
1166 	 * Consume any whitespace preceding digits.
1167 	 */
1168 	while ((c = dtrace_load8(pos)) == ' ' || c == '\t')
1169 		pos++;
1170 
1171 	/*
1172 	 * Handle an explicit sign if one is present.
1173 	 */
1174 	if (c == '-' || c == '+') {
1175 		if (c == '-')
1176 			neg = B_TRUE;
1177 		c = dtrace_load8(++pos);
1178 	}
1179 
1180 	/*
1181 	 * Check for an explicit hexadecimal prefix ("0x" or "0X") and skip it
1182 	 * if present.
1183 	 */
1184 	if (base == 16 && c == '0' && ((cc = dtrace_load8(pos + 1)) == 'x' ||
1185 	    cc == 'X') && isxdigit(ccc = dtrace_load8(pos + 2))) {
1186 		pos += 2;
1187 		c = ccc;
1188 	}
1189 
1190 	/*
1191 	 * Read in contiguous digits until the first non-digit character.
1192 	 */
1193 	for (; pos < end && c != '\0' && lisalnum(c) && (x = DIGIT(c)) < base;
1194 	    c = dtrace_load8(++pos))
1195 		val = val * base + x;
1196 
1197 	return (neg ? -val : val);
1198 }
1199 
1200 /*
1201  * Compare two strings using safe loads.
1202  */
1203 static int
1204 dtrace_strncmp(char *s1, char *s2, size_t limit)
1205 {
1206 	uint8_t c1, c2;
1207 	volatile uint16_t *flags;
1208 
1209 	if (s1 == s2 || limit == 0)
1210 		return (0);
1211 
1212 	flags = (volatile uint16_t *)&cpu_core[curcpu].cpuc_dtrace_flags;
1213 
1214 	do {
1215 		if (s1 == NULL) {
1216 			c1 = '\0';
1217 		} else {
1218 			c1 = dtrace_load8((uintptr_t)s1++);
1219 		}
1220 
1221 		if (s2 == NULL) {
1222 			c2 = '\0';
1223 		} else {
1224 			c2 = dtrace_load8((uintptr_t)s2++);
1225 		}
1226 
1227 		if (c1 != c2)
1228 			return (c1 - c2);
1229 	} while (--limit && c1 != '\0' && !(*flags & CPU_DTRACE_FAULT));
1230 
1231 	return (0);
1232 }
1233 
1234 /*
1235  * Compute strlen(s) for a string using safe memory accesses.  The additional
1236  * len parameter is used to specify a maximum length to ensure completion.
1237  */
1238 static size_t
1239 dtrace_strlen(const char *s, size_t lim)
1240 {
1241 	uint_t len;
1242 
1243 	for (len = 0; len != lim; len++) {
1244 		if (dtrace_load8((uintptr_t)s++) == '\0')
1245 			break;
1246 	}
1247 
1248 	return (len);
1249 }
1250 
1251 /*
1252  * Check if an address falls within a toxic region.
1253  */
1254 static int
1255 dtrace_istoxic(uintptr_t kaddr, size_t size)
1256 {
1257 	uintptr_t taddr, tsize;
1258 	int i;
1259 
1260 	for (i = 0; i < dtrace_toxranges; i++) {
1261 		taddr = dtrace_toxrange[i].dtt_base;
1262 		tsize = dtrace_toxrange[i].dtt_limit - taddr;
1263 
1264 		if (kaddr - taddr < tsize) {
1265 			DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
1266 			cpu_core[curcpu].cpuc_dtrace_illval = kaddr;
1267 			return (1);
1268 		}
1269 
1270 		if (taddr - kaddr < size) {
1271 			DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
1272 			cpu_core[curcpu].cpuc_dtrace_illval = taddr;
1273 			return (1);
1274 		}
1275 	}
1276 
1277 	return (0);
1278 }
1279 
1280 /*
1281  * Copy src to dst using safe memory accesses.  The src is assumed to be unsafe
1282  * memory specified by the DIF program.  The dst is assumed to be safe memory
1283  * that we can store to directly because it is managed by DTrace.  As with
1284  * standard bcopy, overlapping copies are handled properly.
1285  */
1286 static void
1287 dtrace_bcopy(const void *src, void *dst, size_t len)
1288 {
1289 	if (len != 0) {
1290 		uint8_t *s1 = dst;
1291 		const uint8_t *s2 = src;
1292 
1293 		if (s1 <= s2) {
1294 			do {
1295 				*s1++ = dtrace_load8((uintptr_t)s2++);
1296 			} while (--len != 0);
1297 		} else {
1298 			s2 += len;
1299 			s1 += len;
1300 
1301 			do {
1302 				*--s1 = dtrace_load8((uintptr_t)--s2);
1303 			} while (--len != 0);
1304 		}
1305 	}
1306 }
1307 
1308 /*
1309  * Copy src to dst using safe memory accesses, up to either the specified
1310  * length, or the point that a nul byte is encountered.  The src is assumed to
1311  * be unsafe memory specified by the DIF program.  The dst is assumed to be
1312  * safe memory that we can store to directly because it is managed by DTrace.
1313  * Unlike dtrace_bcopy(), overlapping regions are not handled.
1314  */
1315 static void
1316 dtrace_strcpy(const void *src, void *dst, size_t len)
1317 {
1318 	if (len != 0) {
1319 		uint8_t *s1 = dst, c;
1320 		const uint8_t *s2 = src;
1321 
1322 		do {
1323 			*s1++ = c = dtrace_load8((uintptr_t)s2++);
1324 		} while (--len != 0 && c != '\0');
1325 	}
1326 }
1327 
1328 /*
1329  * Copy src to dst, deriving the size and type from the specified (BYREF)
1330  * variable type.  The src is assumed to be unsafe memory specified by the DIF
1331  * program.  The dst is assumed to be DTrace variable memory that is of the
1332  * specified type; we assume that we can store to directly.
1333  */
1334 static void
1335 dtrace_vcopy(void *src, void *dst, dtrace_diftype_t *type, size_t limit)
1336 {
1337 	ASSERT(type->dtdt_flags & DIF_TF_BYREF);
1338 
1339 	if (type->dtdt_kind == DIF_TYPE_STRING) {
1340 		dtrace_strcpy(src, dst, MIN(type->dtdt_size, limit));
1341 	} else {
1342 		dtrace_bcopy(src, dst, MIN(type->dtdt_size, limit));
1343 	}
1344 }
1345 
1346 /*
1347  * Compare s1 to s2 using safe memory accesses.  The s1 data is assumed to be
1348  * unsafe memory specified by the DIF program.  The s2 data is assumed to be
1349  * safe memory that we can access directly because it is managed by DTrace.
1350  */
1351 static int
1352 dtrace_bcmp(const void *s1, const void *s2, size_t len)
1353 {
1354 	volatile uint16_t *flags;
1355 
1356 	flags = (volatile uint16_t *)&cpu_core[curcpu].cpuc_dtrace_flags;
1357 
1358 	if (s1 == s2)
1359 		return (0);
1360 
1361 	if (s1 == NULL || s2 == NULL)
1362 		return (1);
1363 
1364 	if (s1 != s2 && len != 0) {
1365 		const uint8_t *ps1 = s1;
1366 		const uint8_t *ps2 = s2;
1367 
1368 		do {
1369 			if (dtrace_load8((uintptr_t)ps1++) != *ps2++)
1370 				return (1);
1371 		} while (--len != 0 && !(*flags & CPU_DTRACE_FAULT));
1372 	}
1373 	return (0);
1374 }
1375 
1376 /*
1377  * Zero the specified region using a simple byte-by-byte loop.  Note that this
1378  * is for safe DTrace-managed memory only.
1379  */
1380 static void
1381 dtrace_bzero(void *dst, size_t len)
1382 {
1383 	uchar_t *cp;
1384 
1385 	for (cp = dst; len != 0; len--)
1386 		*cp++ = 0;
1387 }
1388 
1389 static void
1390 dtrace_add_128(uint64_t *addend1, uint64_t *addend2, uint64_t *sum)
1391 {
1392 	uint64_t result[2];
1393 
1394 	result[0] = addend1[0] + addend2[0];
1395 	result[1] = addend1[1] + addend2[1] +
1396 	    (result[0] < addend1[0] || result[0] < addend2[0] ? 1 : 0);
1397 
1398 	sum[0] = result[0];
1399 	sum[1] = result[1];
1400 }
1401 
1402 /*
1403  * Shift the 128-bit value in a by b. If b is positive, shift left.
1404  * If b is negative, shift right.
1405  */
1406 static void
1407 dtrace_shift_128(uint64_t *a, int b)
1408 {
1409 	uint64_t mask;
1410 
1411 	if (b == 0)
1412 		return;
1413 
1414 	if (b < 0) {
1415 		b = -b;
1416 		if (b >= 64) {
1417 			a[0] = a[1] >> (b - 64);
1418 			a[1] = 0;
1419 		} else {
1420 			a[0] >>= b;
1421 			mask = 1LL << (64 - b);
1422 			mask -= 1;
1423 			a[0] |= ((a[1] & mask) << (64 - b));
1424 			a[1] >>= b;
1425 		}
1426 	} else {
1427 		if (b >= 64) {
1428 			a[1] = a[0] << (b - 64);
1429 			a[0] = 0;
1430 		} else {
1431 			a[1] <<= b;
1432 			mask = a[0] >> (64 - b);
1433 			a[1] |= mask;
1434 			a[0] <<= b;
1435 		}
1436 	}
1437 }
1438 
1439 /*
1440  * The basic idea is to break the 2 64-bit values into 4 32-bit values,
1441  * use native multiplication on those, and then re-combine into the
1442  * resulting 128-bit value.
1443  *
1444  * (hi1 << 32 + lo1) * (hi2 << 32 + lo2) =
1445  *     hi1 * hi2 << 64 +
1446  *     hi1 * lo2 << 32 +
1447  *     hi2 * lo1 << 32 +
1448  *     lo1 * lo2
1449  */
1450 static void
1451 dtrace_multiply_128(uint64_t factor1, uint64_t factor2, uint64_t *product)
1452 {
1453 	uint64_t hi1, hi2, lo1, lo2;
1454 	uint64_t tmp[2];
1455 
1456 	hi1 = factor1 >> 32;
1457 	hi2 = factor2 >> 32;
1458 
1459 	lo1 = factor1 & DT_MASK_LO;
1460 	lo2 = factor2 & DT_MASK_LO;
1461 
1462 	product[0] = lo1 * lo2;
1463 	product[1] = hi1 * hi2;
1464 
1465 	tmp[0] = hi1 * lo2;
1466 	tmp[1] = 0;
1467 	dtrace_shift_128(tmp, 32);
1468 	dtrace_add_128(product, tmp, product);
1469 
1470 	tmp[0] = hi2 * lo1;
1471 	tmp[1] = 0;
1472 	dtrace_shift_128(tmp, 32);
1473 	dtrace_add_128(product, tmp, product);
1474 }
1475 
1476 /*
1477  * This privilege check should be used by actions and subroutines to
1478  * verify that the user credentials of the process that enabled the
1479  * invoking ECB match the target credentials
1480  */
1481 static int
1482 dtrace_priv_proc_common_user(dtrace_state_t *state)
1483 {
1484 	cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1485 
1486 	/*
1487 	 * We should always have a non-NULL state cred here, since if cred
1488 	 * is null (anonymous tracing), we fast-path bypass this routine.
1489 	 */
1490 	ASSERT(s_cr != NULL);
1491 
1492 	if ((cr = CRED()) != NULL &&
1493 	    s_cr->cr_uid == cr->cr_uid &&
1494 	    s_cr->cr_uid == cr->cr_ruid &&
1495 	    s_cr->cr_uid == cr->cr_suid &&
1496 	    s_cr->cr_gid == cr->cr_gid &&
1497 	    s_cr->cr_gid == cr->cr_rgid &&
1498 	    s_cr->cr_gid == cr->cr_sgid)
1499 		return (1);
1500 
1501 	return (0);
1502 }
1503 
1504 /*
1505  * This privilege check should be used by actions and subroutines to
1506  * verify that the zone of the process that enabled the invoking ECB
1507  * matches the target credentials
1508  */
1509 static int
1510 dtrace_priv_proc_common_zone(dtrace_state_t *state)
1511 {
1512 #ifdef illumos
1513 	cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1514 
1515 	/*
1516 	 * We should always have a non-NULL state cred here, since if cred
1517 	 * is null (anonymous tracing), we fast-path bypass this routine.
1518 	 */
1519 	ASSERT(s_cr != NULL);
1520 
1521 	if ((cr = CRED()) != NULL && s_cr->cr_zone == cr->cr_zone)
1522 		return (1);
1523 
1524 	return (0);
1525 #else
1526 	return (1);
1527 #endif
1528 }
1529 
1530 /*
1531  * This privilege check should be used by actions and subroutines to
1532  * verify that the process has not setuid or changed credentials.
1533  */
1534 static int
1535 dtrace_priv_proc_common_nocd(void)
1536 {
1537 	proc_t *proc;
1538 
1539 	if ((proc = ttoproc(curthread)) != NULL &&
1540 	    !(proc->p_flag & SNOCD))
1541 		return (1);
1542 
1543 	return (0);
1544 }
1545 
1546 static int
1547 dtrace_priv_proc_destructive(dtrace_state_t *state)
1548 {
1549 	int action = state->dts_cred.dcr_action;
1550 
1551 	if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE) == 0) &&
1552 	    dtrace_priv_proc_common_zone(state) == 0)
1553 		goto bad;
1554 
1555 	if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER) == 0) &&
1556 	    dtrace_priv_proc_common_user(state) == 0)
1557 		goto bad;
1558 
1559 	if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG) == 0) &&
1560 	    dtrace_priv_proc_common_nocd() == 0)
1561 		goto bad;
1562 
1563 	return (1);
1564 
1565 bad:
1566 	cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1567 
1568 	return (0);
1569 }
1570 
1571 static int
1572 dtrace_priv_proc_control(dtrace_state_t *state)
1573 {
1574 	if (state->dts_cred.dcr_action & DTRACE_CRA_PROC_CONTROL)
1575 		return (1);
1576 
1577 	if (dtrace_priv_proc_common_zone(state) &&
1578 	    dtrace_priv_proc_common_user(state) &&
1579 	    dtrace_priv_proc_common_nocd())
1580 		return (1);
1581 
1582 	cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1583 
1584 	return (0);
1585 }
1586 
1587 static int
1588 dtrace_priv_proc(dtrace_state_t *state)
1589 {
1590 	if (state->dts_cred.dcr_action & DTRACE_CRA_PROC)
1591 		return (1);
1592 
1593 	cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1594 
1595 	return (0);
1596 }
1597 
1598 static int
1599 dtrace_priv_kernel(dtrace_state_t *state)
1600 {
1601 	if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL)
1602 		return (1);
1603 
1604 	cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1605 
1606 	return (0);
1607 }
1608 
1609 static int
1610 dtrace_priv_kernel_destructive(dtrace_state_t *state)
1611 {
1612 	if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL_DESTRUCTIVE)
1613 		return (1);
1614 
1615 	cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1616 
1617 	return (0);
1618 }
1619 
1620 /*
1621  * Determine if the dte_cond of the specified ECB allows for processing of
1622  * the current probe to continue.  Note that this routine may allow continued
1623  * processing, but with access(es) stripped from the mstate's dtms_access
1624  * field.
1625  */
1626 static int
1627 dtrace_priv_probe(dtrace_state_t *state, dtrace_mstate_t *mstate,
1628     dtrace_ecb_t *ecb)
1629 {
1630 	dtrace_probe_t *probe = ecb->dte_probe;
1631 	dtrace_provider_t *prov = probe->dtpr_provider;
1632 	dtrace_pops_t *pops = &prov->dtpv_pops;
1633 	int mode = DTRACE_MODE_NOPRIV_DROP;
1634 
1635 	ASSERT(ecb->dte_cond);
1636 
1637 #ifdef illumos
1638 	if (pops->dtps_mode != NULL) {
1639 		mode = pops->dtps_mode(prov->dtpv_arg,
1640 		    probe->dtpr_id, probe->dtpr_arg);
1641 
1642 		ASSERT((mode & DTRACE_MODE_USER) ||
1643 		    (mode & DTRACE_MODE_KERNEL));
1644 		ASSERT((mode & DTRACE_MODE_NOPRIV_RESTRICT) ||
1645 		    (mode & DTRACE_MODE_NOPRIV_DROP));
1646 	}
1647 
1648 	/*
1649 	 * If the dte_cond bits indicate that this consumer is only allowed to
1650 	 * see user-mode firings of this probe, call the provider's dtps_mode()
1651 	 * entry point to check that the probe was fired while in a user
1652 	 * context.  If that's not the case, use the policy specified by the
1653 	 * provider to determine if we drop the probe or merely restrict
1654 	 * operation.
1655 	 */
1656 	if (ecb->dte_cond & DTRACE_COND_USERMODE) {
1657 		ASSERT(mode != DTRACE_MODE_NOPRIV_DROP);
1658 
1659 		if (!(mode & DTRACE_MODE_USER)) {
1660 			if (mode & DTRACE_MODE_NOPRIV_DROP)
1661 				return (0);
1662 
1663 			mstate->dtms_access &= ~DTRACE_ACCESS_ARGS;
1664 		}
1665 	}
1666 #endif
1667 
1668 	/*
1669 	 * This is more subtle than it looks. We have to be absolutely certain
1670 	 * that CRED() isn't going to change out from under us so it's only
1671 	 * legit to examine that structure if we're in constrained situations.
1672 	 * Currently, the only times we'll this check is if a non-super-user
1673 	 * has enabled the profile or syscall providers -- providers that
1674 	 * allow visibility of all processes. For the profile case, the check
1675 	 * above will ensure that we're examining a user context.
1676 	 */
1677 	if (ecb->dte_cond & DTRACE_COND_OWNER) {
1678 		cred_t *cr;
1679 		cred_t *s_cr = state->dts_cred.dcr_cred;
1680 		proc_t *proc;
1681 
1682 		ASSERT(s_cr != NULL);
1683 
1684 		if ((cr = CRED()) == NULL ||
1685 		    s_cr->cr_uid != cr->cr_uid ||
1686 		    s_cr->cr_uid != cr->cr_ruid ||
1687 		    s_cr->cr_uid != cr->cr_suid ||
1688 		    s_cr->cr_gid != cr->cr_gid ||
1689 		    s_cr->cr_gid != cr->cr_rgid ||
1690 		    s_cr->cr_gid != cr->cr_sgid ||
1691 		    (proc = ttoproc(curthread)) == NULL ||
1692 		    (proc->p_flag & SNOCD)) {
1693 			if (mode & DTRACE_MODE_NOPRIV_DROP)
1694 				return (0);
1695 
1696 #ifdef illumos
1697 			mstate->dtms_access &= ~DTRACE_ACCESS_PROC;
1698 #endif
1699 		}
1700 	}
1701 
1702 #ifdef illumos
1703 	/*
1704 	 * If our dte_cond is set to DTRACE_COND_ZONEOWNER and we are not
1705 	 * in our zone, check to see if our mode policy is to restrict rather
1706 	 * than to drop; if to restrict, strip away both DTRACE_ACCESS_PROC
1707 	 * and DTRACE_ACCESS_ARGS
1708 	 */
1709 	if (ecb->dte_cond & DTRACE_COND_ZONEOWNER) {
1710 		cred_t *cr;
1711 		cred_t *s_cr = state->dts_cred.dcr_cred;
1712 
1713 		ASSERT(s_cr != NULL);
1714 
1715 		if ((cr = CRED()) == NULL ||
1716 		    s_cr->cr_zone->zone_id != cr->cr_zone->zone_id) {
1717 			if (mode & DTRACE_MODE_NOPRIV_DROP)
1718 				return (0);
1719 
1720 			mstate->dtms_access &=
1721 			    ~(DTRACE_ACCESS_PROC | DTRACE_ACCESS_ARGS);
1722 		}
1723 	}
1724 #endif
1725 
1726 	return (1);
1727 }
1728 
1729 /*
1730  * Note:  not called from probe context.  This function is called
1731  * asynchronously (and at a regular interval) from outside of probe context to
1732  * clean the dirty dynamic variable lists on all CPUs.  Dynamic variable
1733  * cleaning is explained in detail in <sys/dtrace_impl.h>.
1734  */
1735 void
1736 dtrace_dynvar_clean(dtrace_dstate_t *dstate)
1737 {
1738 	dtrace_dynvar_t *dirty;
1739 	dtrace_dstate_percpu_t *dcpu;
1740 	dtrace_dynvar_t **rinsep;
1741 	int i, j, work = 0;
1742 
1743 	for (i = 0; i < NCPU; i++) {
1744 		dcpu = &dstate->dtds_percpu[i];
1745 		rinsep = &dcpu->dtdsc_rinsing;
1746 
1747 		/*
1748 		 * If the dirty list is NULL, there is no dirty work to do.
1749 		 */
1750 		if (dcpu->dtdsc_dirty == NULL)
1751 			continue;
1752 
1753 		if (dcpu->dtdsc_rinsing != NULL) {
1754 			/*
1755 			 * If the rinsing list is non-NULL, then it is because
1756 			 * this CPU was selected to accept another CPU's
1757 			 * dirty list -- and since that time, dirty buffers
1758 			 * have accumulated.  This is a highly unlikely
1759 			 * condition, but we choose to ignore the dirty
1760 			 * buffers -- they'll be picked up a future cleanse.
1761 			 */
1762 			continue;
1763 		}
1764 
1765 		if (dcpu->dtdsc_clean != NULL) {
1766 			/*
1767 			 * If the clean list is non-NULL, then we're in a
1768 			 * situation where a CPU has done deallocations (we
1769 			 * have a non-NULL dirty list) but no allocations (we
1770 			 * also have a non-NULL clean list).  We can't simply
1771 			 * move the dirty list into the clean list on this
1772 			 * CPU, yet we also don't want to allow this condition
1773 			 * to persist, lest a short clean list prevent a
1774 			 * massive dirty list from being cleaned (which in
1775 			 * turn could lead to otherwise avoidable dynamic
1776 			 * drops).  To deal with this, we look for some CPU
1777 			 * with a NULL clean list, NULL dirty list, and NULL
1778 			 * rinsing list -- and then we borrow this CPU to
1779 			 * rinse our dirty list.
1780 			 */
1781 			for (j = 0; j < NCPU; j++) {
1782 				dtrace_dstate_percpu_t *rinser;
1783 
1784 				rinser = &dstate->dtds_percpu[j];
1785 
1786 				if (rinser->dtdsc_rinsing != NULL)
1787 					continue;
1788 
1789 				if (rinser->dtdsc_dirty != NULL)
1790 					continue;
1791 
1792 				if (rinser->dtdsc_clean != NULL)
1793 					continue;
1794 
1795 				rinsep = &rinser->dtdsc_rinsing;
1796 				break;
1797 			}
1798 
1799 			if (j == NCPU) {
1800 				/*
1801 				 * We were unable to find another CPU that
1802 				 * could accept this dirty list -- we are
1803 				 * therefore unable to clean it now.
1804 				 */
1805 				dtrace_dynvar_failclean++;
1806 				continue;
1807 			}
1808 		}
1809 
1810 		work = 1;
1811 
1812 		/*
1813 		 * Atomically move the dirty list aside.
1814 		 */
1815 		do {
1816 			dirty = dcpu->dtdsc_dirty;
1817 
1818 			/*
1819 			 * Before we zap the dirty list, set the rinsing list.
1820 			 * (This allows for a potential assertion in
1821 			 * dtrace_dynvar():  if a free dynamic variable appears
1822 			 * on a hash chain, either the dirty list or the
1823 			 * rinsing list for some CPU must be non-NULL.)
1824 			 */
1825 			*rinsep = dirty;
1826 			dtrace_membar_producer();
1827 		} while (dtrace_casptr(&dcpu->dtdsc_dirty,
1828 		    dirty, NULL) != dirty);
1829 	}
1830 
1831 	if (!work) {
1832 		/*
1833 		 * We have no work to do; we can simply return.
1834 		 */
1835 		return;
1836 	}
1837 
1838 	dtrace_sync();
1839 
1840 	for (i = 0; i < NCPU; i++) {
1841 		dcpu = &dstate->dtds_percpu[i];
1842 
1843 		if (dcpu->dtdsc_rinsing == NULL)
1844 			continue;
1845 
1846 		/*
1847 		 * We are now guaranteed that no hash chain contains a pointer
1848 		 * into this dirty list; we can make it clean.
1849 		 */
1850 		ASSERT(dcpu->dtdsc_clean == NULL);
1851 		dcpu->dtdsc_clean = dcpu->dtdsc_rinsing;
1852 		dcpu->dtdsc_rinsing = NULL;
1853 	}
1854 
1855 	/*
1856 	 * Before we actually set the state to be DTRACE_DSTATE_CLEAN, make
1857 	 * sure that all CPUs have seen all of the dtdsc_clean pointers.
1858 	 * This prevents a race whereby a CPU incorrectly decides that
1859 	 * the state should be something other than DTRACE_DSTATE_CLEAN
1860 	 * after dtrace_dynvar_clean() has completed.
1861 	 */
1862 	dtrace_sync();
1863 
1864 	dstate->dtds_state = DTRACE_DSTATE_CLEAN;
1865 }
1866 
1867 /*
1868  * Depending on the value of the op parameter, this function looks-up,
1869  * allocates or deallocates an arbitrarily-keyed dynamic variable.  If an
1870  * allocation is requested, this function will return a pointer to a
1871  * dtrace_dynvar_t corresponding to the allocated variable -- or NULL if no
1872  * variable can be allocated.  If NULL is returned, the appropriate counter
1873  * will be incremented.
1874  */
1875 dtrace_dynvar_t *
1876 dtrace_dynvar(dtrace_dstate_t *dstate, uint_t nkeys,
1877     dtrace_key_t *key, size_t dsize, dtrace_dynvar_op_t op,
1878     dtrace_mstate_t *mstate, dtrace_vstate_t *vstate)
1879 {
1880 	uint64_t hashval = DTRACE_DYNHASH_VALID;
1881 	dtrace_dynhash_t *hash = dstate->dtds_hash;
1882 	dtrace_dynvar_t *free, *new_free, *next, *dvar, *start, *prev = NULL;
1883 	processorid_t me = curcpu, cpu = me;
1884 	dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[me];
1885 	size_t bucket, ksize;
1886 	size_t chunksize = dstate->dtds_chunksize;
1887 	uintptr_t kdata, lock, nstate;
1888 	uint_t i;
1889 
1890 	ASSERT(nkeys != 0);
1891 
1892 	/*
1893 	 * Hash the key.  As with aggregations, we use Jenkins' "One-at-a-time"
1894 	 * algorithm.  For the by-value portions, we perform the algorithm in
1895 	 * 16-bit chunks (as opposed to 8-bit chunks).  This speeds things up a
1896 	 * bit, and seems to have only a minute effect on distribution.  For
1897 	 * the by-reference data, we perform "One-at-a-time" iterating (safely)
1898 	 * over each referenced byte.  It's painful to do this, but it's much
1899 	 * better than pathological hash distribution.  The efficacy of the
1900 	 * hashing algorithm (and a comparison with other algorithms) may be
1901 	 * found by running the ::dtrace_dynstat MDB dcmd.
1902 	 */
1903 	for (i = 0; i < nkeys; i++) {
1904 		if (key[i].dttk_size == 0) {
1905 			uint64_t val = key[i].dttk_value;
1906 
1907 			hashval += (val >> 48) & 0xffff;
1908 			hashval += (hashval << 10);
1909 			hashval ^= (hashval >> 6);
1910 
1911 			hashval += (val >> 32) & 0xffff;
1912 			hashval += (hashval << 10);
1913 			hashval ^= (hashval >> 6);
1914 
1915 			hashval += (val >> 16) & 0xffff;
1916 			hashval += (hashval << 10);
1917 			hashval ^= (hashval >> 6);
1918 
1919 			hashval += val & 0xffff;
1920 			hashval += (hashval << 10);
1921 			hashval ^= (hashval >> 6);
1922 		} else {
1923 			/*
1924 			 * This is incredibly painful, but it beats the hell
1925 			 * out of the alternative.
1926 			 */
1927 			uint64_t j, size = key[i].dttk_size;
1928 			uintptr_t base = (uintptr_t)key[i].dttk_value;
1929 
1930 			if (!dtrace_canload(base, size, mstate, vstate))
1931 				break;
1932 
1933 			for (j = 0; j < size; j++) {
1934 				hashval += dtrace_load8(base + j);
1935 				hashval += (hashval << 10);
1936 				hashval ^= (hashval >> 6);
1937 			}
1938 		}
1939 	}
1940 
1941 	if (DTRACE_CPUFLAG_ISSET(CPU_DTRACE_FAULT))
1942 		return (NULL);
1943 
1944 	hashval += (hashval << 3);
1945 	hashval ^= (hashval >> 11);
1946 	hashval += (hashval << 15);
1947 
1948 	/*
1949 	 * There is a remote chance (ideally, 1 in 2^31) that our hashval
1950 	 * comes out to be one of our two sentinel hash values.  If this
1951 	 * actually happens, we set the hashval to be a value known to be a
1952 	 * non-sentinel value.
1953 	 */
1954 	if (hashval == DTRACE_DYNHASH_FREE || hashval == DTRACE_DYNHASH_SINK)
1955 		hashval = DTRACE_DYNHASH_VALID;
1956 
1957 	/*
1958 	 * Yes, it's painful to do a divide here.  If the cycle count becomes
1959 	 * important here, tricks can be pulled to reduce it.  (However, it's
1960 	 * critical that hash collisions be kept to an absolute minimum;
1961 	 * they're much more painful than a divide.)  It's better to have a
1962 	 * solution that generates few collisions and still keeps things
1963 	 * relatively simple.
1964 	 */
1965 	bucket = hashval % dstate->dtds_hashsize;
1966 
1967 	if (op == DTRACE_DYNVAR_DEALLOC) {
1968 		volatile uintptr_t *lockp = &hash[bucket].dtdh_lock;
1969 
1970 		for (;;) {
1971 			while ((lock = *lockp) & 1)
1972 				continue;
1973 
1974 			if (dtrace_casptr((volatile void *)lockp,
1975 			    (volatile void *)lock, (volatile void *)(lock + 1)) == (void *)lock)
1976 				break;
1977 		}
1978 
1979 		dtrace_membar_producer();
1980 	}
1981 
1982 top:
1983 	prev = NULL;
1984 	lock = hash[bucket].dtdh_lock;
1985 
1986 	dtrace_membar_consumer();
1987 
1988 	start = hash[bucket].dtdh_chain;
1989 	ASSERT(start != NULL && (start->dtdv_hashval == DTRACE_DYNHASH_SINK ||
1990 	    start->dtdv_hashval != DTRACE_DYNHASH_FREE ||
1991 	    op != DTRACE_DYNVAR_DEALLOC));
1992 
1993 	for (dvar = start; dvar != NULL; dvar = dvar->dtdv_next) {
1994 		dtrace_tuple_t *dtuple = &dvar->dtdv_tuple;
1995 		dtrace_key_t *dkey = &dtuple->dtt_key[0];
1996 
1997 		if (dvar->dtdv_hashval != hashval) {
1998 			if (dvar->dtdv_hashval == DTRACE_DYNHASH_SINK) {
1999 				/*
2000 				 * We've reached the sink, and therefore the
2001 				 * end of the hash chain; we can kick out of
2002 				 * the loop knowing that we have seen a valid
2003 				 * snapshot of state.
2004 				 */
2005 				ASSERT(dvar->dtdv_next == NULL);
2006 				ASSERT(dvar == &dtrace_dynhash_sink);
2007 				break;
2008 			}
2009 
2010 			if (dvar->dtdv_hashval == DTRACE_DYNHASH_FREE) {
2011 				/*
2012 				 * We've gone off the rails:  somewhere along
2013 				 * the line, one of the members of this hash
2014 				 * chain was deleted.  Note that we could also
2015 				 * detect this by simply letting this loop run
2016 				 * to completion, as we would eventually hit
2017 				 * the end of the dirty list.  However, we
2018 				 * want to avoid running the length of the
2019 				 * dirty list unnecessarily (it might be quite
2020 				 * long), so we catch this as early as
2021 				 * possible by detecting the hash marker.  In
2022 				 * this case, we simply set dvar to NULL and
2023 				 * break; the conditional after the loop will
2024 				 * send us back to top.
2025 				 */
2026 				dvar = NULL;
2027 				break;
2028 			}
2029 
2030 			goto next;
2031 		}
2032 
2033 		if (dtuple->dtt_nkeys != nkeys)
2034 			goto next;
2035 
2036 		for (i = 0; i < nkeys; i++, dkey++) {
2037 			if (dkey->dttk_size != key[i].dttk_size)
2038 				goto next; /* size or type mismatch */
2039 
2040 			if (dkey->dttk_size != 0) {
2041 				if (dtrace_bcmp(
2042 				    (void *)(uintptr_t)key[i].dttk_value,
2043 				    (void *)(uintptr_t)dkey->dttk_value,
2044 				    dkey->dttk_size))
2045 					goto next;
2046 			} else {
2047 				if (dkey->dttk_value != key[i].dttk_value)
2048 					goto next;
2049 			}
2050 		}
2051 
2052 		if (op != DTRACE_DYNVAR_DEALLOC)
2053 			return (dvar);
2054 
2055 		ASSERT(dvar->dtdv_next == NULL ||
2056 		    dvar->dtdv_next->dtdv_hashval != DTRACE_DYNHASH_FREE);
2057 
2058 		if (prev != NULL) {
2059 			ASSERT(hash[bucket].dtdh_chain != dvar);
2060 			ASSERT(start != dvar);
2061 			ASSERT(prev->dtdv_next == dvar);
2062 			prev->dtdv_next = dvar->dtdv_next;
2063 		} else {
2064 			if (dtrace_casptr(&hash[bucket].dtdh_chain,
2065 			    start, dvar->dtdv_next) != start) {
2066 				/*
2067 				 * We have failed to atomically swing the
2068 				 * hash table head pointer, presumably because
2069 				 * of a conflicting allocation on another CPU.
2070 				 * We need to reread the hash chain and try
2071 				 * again.
2072 				 */
2073 				goto top;
2074 			}
2075 		}
2076 
2077 		dtrace_membar_producer();
2078 
2079 		/*
2080 		 * Now set the hash value to indicate that it's free.
2081 		 */
2082 		ASSERT(hash[bucket].dtdh_chain != dvar);
2083 		dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
2084 
2085 		dtrace_membar_producer();
2086 
2087 		/*
2088 		 * Set the next pointer to point at the dirty list, and
2089 		 * atomically swing the dirty pointer to the newly freed dvar.
2090 		 */
2091 		do {
2092 			next = dcpu->dtdsc_dirty;
2093 			dvar->dtdv_next = next;
2094 		} while (dtrace_casptr(&dcpu->dtdsc_dirty, next, dvar) != next);
2095 
2096 		/*
2097 		 * Finally, unlock this hash bucket.
2098 		 */
2099 		ASSERT(hash[bucket].dtdh_lock == lock);
2100 		ASSERT(lock & 1);
2101 		hash[bucket].dtdh_lock++;
2102 
2103 		return (NULL);
2104 next:
2105 		prev = dvar;
2106 		continue;
2107 	}
2108 
2109 	if (dvar == NULL) {
2110 		/*
2111 		 * If dvar is NULL, it is because we went off the rails:
2112 		 * one of the elements that we traversed in the hash chain
2113 		 * was deleted while we were traversing it.  In this case,
2114 		 * we assert that we aren't doing a dealloc (deallocs lock
2115 		 * the hash bucket to prevent themselves from racing with
2116 		 * one another), and retry the hash chain traversal.
2117 		 */
2118 		ASSERT(op != DTRACE_DYNVAR_DEALLOC);
2119 		goto top;
2120 	}
2121 
2122 	if (op != DTRACE_DYNVAR_ALLOC) {
2123 		/*
2124 		 * If we are not to allocate a new variable, we want to
2125 		 * return NULL now.  Before we return, check that the value
2126 		 * of the lock word hasn't changed.  If it has, we may have
2127 		 * seen an inconsistent snapshot.
2128 		 */
2129 		if (op == DTRACE_DYNVAR_NOALLOC) {
2130 			if (hash[bucket].dtdh_lock != lock)
2131 				goto top;
2132 		} else {
2133 			ASSERT(op == DTRACE_DYNVAR_DEALLOC);
2134 			ASSERT(hash[bucket].dtdh_lock == lock);
2135 			ASSERT(lock & 1);
2136 			hash[bucket].dtdh_lock++;
2137 		}
2138 
2139 		return (NULL);
2140 	}
2141 
2142 	/*
2143 	 * We need to allocate a new dynamic variable.  The size we need is the
2144 	 * size of dtrace_dynvar plus the size of nkeys dtrace_key_t's plus the
2145 	 * size of any auxiliary key data (rounded up to 8-byte alignment) plus
2146 	 * the size of any referred-to data (dsize).  We then round the final
2147 	 * size up to the chunksize for allocation.
2148 	 */
2149 	for (ksize = 0, i = 0; i < nkeys; i++)
2150 		ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
2151 
2152 	/*
2153 	 * This should be pretty much impossible, but could happen if, say,
2154 	 * strange DIF specified the tuple.  Ideally, this should be an
2155 	 * assertion and not an error condition -- but that requires that the
2156 	 * chunksize calculation in dtrace_difo_chunksize() be absolutely
2157 	 * bullet-proof.  (That is, it must not be able to be fooled by
2158 	 * malicious DIF.)  Given the lack of backwards branches in DIF,
2159 	 * solving this would presumably not amount to solving the Halting
2160 	 * Problem -- but it still seems awfully hard.
2161 	 */
2162 	if (sizeof (dtrace_dynvar_t) + sizeof (dtrace_key_t) * (nkeys - 1) +
2163 	    ksize + dsize > chunksize) {
2164 		dcpu->dtdsc_drops++;
2165 		return (NULL);
2166 	}
2167 
2168 	nstate = DTRACE_DSTATE_EMPTY;
2169 
2170 	do {
2171 retry:
2172 		free = dcpu->dtdsc_free;
2173 
2174 		if (free == NULL) {
2175 			dtrace_dynvar_t *clean = dcpu->dtdsc_clean;
2176 			void *rval;
2177 
2178 			if (clean == NULL) {
2179 				/*
2180 				 * We're out of dynamic variable space on
2181 				 * this CPU.  Unless we have tried all CPUs,
2182 				 * we'll try to allocate from a different
2183 				 * CPU.
2184 				 */
2185 				switch (dstate->dtds_state) {
2186 				case DTRACE_DSTATE_CLEAN: {
2187 					void *sp = &dstate->dtds_state;
2188 
2189 					if (++cpu >= NCPU)
2190 						cpu = 0;
2191 
2192 					if (dcpu->dtdsc_dirty != NULL &&
2193 					    nstate == DTRACE_DSTATE_EMPTY)
2194 						nstate = DTRACE_DSTATE_DIRTY;
2195 
2196 					if (dcpu->dtdsc_rinsing != NULL)
2197 						nstate = DTRACE_DSTATE_RINSING;
2198 
2199 					dcpu = &dstate->dtds_percpu[cpu];
2200 
2201 					if (cpu != me)
2202 						goto retry;
2203 
2204 					(void) dtrace_cas32(sp,
2205 					    DTRACE_DSTATE_CLEAN, nstate);
2206 
2207 					/*
2208 					 * To increment the correct bean
2209 					 * counter, take another lap.
2210 					 */
2211 					goto retry;
2212 				}
2213 
2214 				case DTRACE_DSTATE_DIRTY:
2215 					dcpu->dtdsc_dirty_drops++;
2216 					break;
2217 
2218 				case DTRACE_DSTATE_RINSING:
2219 					dcpu->dtdsc_rinsing_drops++;
2220 					break;
2221 
2222 				case DTRACE_DSTATE_EMPTY:
2223 					dcpu->dtdsc_drops++;
2224 					break;
2225 				}
2226 
2227 				DTRACE_CPUFLAG_SET(CPU_DTRACE_DROP);
2228 				return (NULL);
2229 			}
2230 
2231 			/*
2232 			 * The clean list appears to be non-empty.  We want to
2233 			 * move the clean list to the free list; we start by
2234 			 * moving the clean pointer aside.
2235 			 */
2236 			if (dtrace_casptr(&dcpu->dtdsc_clean,
2237 			    clean, NULL) != clean) {
2238 				/*
2239 				 * We are in one of two situations:
2240 				 *
2241 				 *  (a)	The clean list was switched to the
2242 				 *	free list by another CPU.
2243 				 *
2244 				 *  (b)	The clean list was added to by the
2245 				 *	cleansing cyclic.
2246 				 *
2247 				 * In either of these situations, we can
2248 				 * just reattempt the free list allocation.
2249 				 */
2250 				goto retry;
2251 			}
2252 
2253 			ASSERT(clean->dtdv_hashval == DTRACE_DYNHASH_FREE);
2254 
2255 			/*
2256 			 * Now we'll move the clean list to our free list.
2257 			 * It's impossible for this to fail:  the only way
2258 			 * the free list can be updated is through this
2259 			 * code path, and only one CPU can own the clean list.
2260 			 * Thus, it would only be possible for this to fail if
2261 			 * this code were racing with dtrace_dynvar_clean().
2262 			 * (That is, if dtrace_dynvar_clean() updated the clean
2263 			 * list, and we ended up racing to update the free
2264 			 * list.)  This race is prevented by the dtrace_sync()
2265 			 * in dtrace_dynvar_clean() -- which flushes the
2266 			 * owners of the clean lists out before resetting
2267 			 * the clean lists.
2268 			 */
2269 			dcpu = &dstate->dtds_percpu[me];
2270 			rval = dtrace_casptr(&dcpu->dtdsc_free, NULL, clean);
2271 			ASSERT(rval == NULL);
2272 			goto retry;
2273 		}
2274 
2275 		dvar = free;
2276 		new_free = dvar->dtdv_next;
2277 	} while (dtrace_casptr(&dcpu->dtdsc_free, free, new_free) != free);
2278 
2279 	/*
2280 	 * We have now allocated a new chunk.  We copy the tuple keys into the
2281 	 * tuple array and copy any referenced key data into the data space
2282 	 * following the tuple array.  As we do this, we relocate dttk_value
2283 	 * in the final tuple to point to the key data address in the chunk.
2284 	 */
2285 	kdata = (uintptr_t)&dvar->dtdv_tuple.dtt_key[nkeys];
2286 	dvar->dtdv_data = (void *)(kdata + ksize);
2287 	dvar->dtdv_tuple.dtt_nkeys = nkeys;
2288 
2289 	for (i = 0; i < nkeys; i++) {
2290 		dtrace_key_t *dkey = &dvar->dtdv_tuple.dtt_key[i];
2291 		size_t kesize = key[i].dttk_size;
2292 
2293 		if (kesize != 0) {
2294 			dtrace_bcopy(
2295 			    (const void *)(uintptr_t)key[i].dttk_value,
2296 			    (void *)kdata, kesize);
2297 			dkey->dttk_value = kdata;
2298 			kdata += P2ROUNDUP(kesize, sizeof (uint64_t));
2299 		} else {
2300 			dkey->dttk_value = key[i].dttk_value;
2301 		}
2302 
2303 		dkey->dttk_size = kesize;
2304 	}
2305 
2306 	ASSERT(dvar->dtdv_hashval == DTRACE_DYNHASH_FREE);
2307 	dvar->dtdv_hashval = hashval;
2308 	dvar->dtdv_next = start;
2309 
2310 	if (dtrace_casptr(&hash[bucket].dtdh_chain, start, dvar) == start)
2311 		return (dvar);
2312 
2313 	/*
2314 	 * The cas has failed.  Either another CPU is adding an element to
2315 	 * this hash chain, or another CPU is deleting an element from this
2316 	 * hash chain.  The simplest way to deal with both of these cases
2317 	 * (though not necessarily the most efficient) is to free our
2318 	 * allocated block and re-attempt it all.  Note that the free is
2319 	 * to the dirty list and _not_ to the free list.  This is to prevent
2320 	 * races with allocators, above.
2321 	 */
2322 	dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
2323 
2324 	dtrace_membar_producer();
2325 
2326 	do {
2327 		free = dcpu->dtdsc_dirty;
2328 		dvar->dtdv_next = free;
2329 	} while (dtrace_casptr(&dcpu->dtdsc_dirty, free, dvar) != free);
2330 
2331 	goto top;
2332 }
2333 
2334 /*ARGSUSED*/
2335 static void
2336 dtrace_aggregate_min(uint64_t *oval, uint64_t nval, uint64_t arg)
2337 {
2338 	if ((int64_t)nval < (int64_t)*oval)
2339 		*oval = nval;
2340 }
2341 
2342 /*ARGSUSED*/
2343 static void
2344 dtrace_aggregate_max(uint64_t *oval, uint64_t nval, uint64_t arg)
2345 {
2346 	if ((int64_t)nval > (int64_t)*oval)
2347 		*oval = nval;
2348 }
2349 
2350 static void
2351 dtrace_aggregate_quantize(uint64_t *quanta, uint64_t nval, uint64_t incr)
2352 {
2353 	int i, zero = DTRACE_QUANTIZE_ZEROBUCKET;
2354 	int64_t val = (int64_t)nval;
2355 
2356 	if (val < 0) {
2357 		for (i = 0; i < zero; i++) {
2358 			if (val <= DTRACE_QUANTIZE_BUCKETVAL(i)) {
2359 				quanta[i] += incr;
2360 				return;
2361 			}
2362 		}
2363 	} else {
2364 		for (i = zero + 1; i < DTRACE_QUANTIZE_NBUCKETS; i++) {
2365 			if (val < DTRACE_QUANTIZE_BUCKETVAL(i)) {
2366 				quanta[i - 1] += incr;
2367 				return;
2368 			}
2369 		}
2370 
2371 		quanta[DTRACE_QUANTIZE_NBUCKETS - 1] += incr;
2372 		return;
2373 	}
2374 
2375 	ASSERT(0);
2376 }
2377 
2378 static void
2379 dtrace_aggregate_lquantize(uint64_t *lquanta, uint64_t nval, uint64_t incr)
2380 {
2381 	uint64_t arg = *lquanta++;
2382 	int32_t base = DTRACE_LQUANTIZE_BASE(arg);
2383 	uint16_t step = DTRACE_LQUANTIZE_STEP(arg);
2384 	uint16_t levels = DTRACE_LQUANTIZE_LEVELS(arg);
2385 	int32_t val = (int32_t)nval, level;
2386 
2387 	ASSERT(step != 0);
2388 	ASSERT(levels != 0);
2389 
2390 	if (val < base) {
2391 		/*
2392 		 * This is an underflow.
2393 		 */
2394 		lquanta[0] += incr;
2395 		return;
2396 	}
2397 
2398 	level = (val - base) / step;
2399 
2400 	if (level < levels) {
2401 		lquanta[level + 1] += incr;
2402 		return;
2403 	}
2404 
2405 	/*
2406 	 * This is an overflow.
2407 	 */
2408 	lquanta[levels + 1] += incr;
2409 }
2410 
2411 static int
2412 dtrace_aggregate_llquantize_bucket(uint16_t factor, uint16_t low,
2413     uint16_t high, uint16_t nsteps, int64_t value)
2414 {
2415 	int64_t this = 1, last, next;
2416 	int base = 1, order;
2417 
2418 	ASSERT(factor <= nsteps);
2419 	ASSERT(nsteps % factor == 0);
2420 
2421 	for (order = 0; order < low; order++)
2422 		this *= factor;
2423 
2424 	/*
2425 	 * If our value is less than our factor taken to the power of the
2426 	 * low order of magnitude, it goes into the zeroth bucket.
2427 	 */
2428 	if (value < (last = this))
2429 		return (0);
2430 
2431 	for (this *= factor; order <= high; order++) {
2432 		int nbuckets = this > nsteps ? nsteps : this;
2433 
2434 		if ((next = this * factor) < this) {
2435 			/*
2436 			 * We should not generally get log/linear quantizations
2437 			 * with a high magnitude that allows 64-bits to
2438 			 * overflow, but we nonetheless protect against this
2439 			 * by explicitly checking for overflow, and clamping
2440 			 * our value accordingly.
2441 			 */
2442 			value = this - 1;
2443 		}
2444 
2445 		if (value < this) {
2446 			/*
2447 			 * If our value lies within this order of magnitude,
2448 			 * determine its position by taking the offset within
2449 			 * the order of magnitude, dividing by the bucket
2450 			 * width, and adding to our (accumulated) base.
2451 			 */
2452 			return (base + (value - last) / (this / nbuckets));
2453 		}
2454 
2455 		base += nbuckets - (nbuckets / factor);
2456 		last = this;
2457 		this = next;
2458 	}
2459 
2460 	/*
2461 	 * Our value is greater than or equal to our factor taken to the
2462 	 * power of one plus the high magnitude -- return the top bucket.
2463 	 */
2464 	return (base);
2465 }
2466 
2467 static void
2468 dtrace_aggregate_llquantize(uint64_t *llquanta, uint64_t nval, uint64_t incr)
2469 {
2470 	uint64_t arg = *llquanta++;
2471 	uint16_t factor = DTRACE_LLQUANTIZE_FACTOR(arg);
2472 	uint16_t low = DTRACE_LLQUANTIZE_LOW(arg);
2473 	uint16_t high = DTRACE_LLQUANTIZE_HIGH(arg);
2474 	uint16_t nsteps = DTRACE_LLQUANTIZE_NSTEP(arg);
2475 
2476 	llquanta[dtrace_aggregate_llquantize_bucket(factor,
2477 	    low, high, nsteps, nval)] += incr;
2478 }
2479 
2480 /*ARGSUSED*/
2481 static void
2482 dtrace_aggregate_avg(uint64_t *data, uint64_t nval, uint64_t arg)
2483 {
2484 	data[0]++;
2485 	data[1] += nval;
2486 }
2487 
2488 /*ARGSUSED*/
2489 static void
2490 dtrace_aggregate_stddev(uint64_t *data, uint64_t nval, uint64_t arg)
2491 {
2492 	int64_t snval = (int64_t)nval;
2493 	uint64_t tmp[2];
2494 
2495 	data[0]++;
2496 	data[1] += nval;
2497 
2498 	/*
2499 	 * What we want to say here is:
2500 	 *
2501 	 * data[2] += nval * nval;
2502 	 *
2503 	 * But given that nval is 64-bit, we could easily overflow, so
2504 	 * we do this as 128-bit arithmetic.
2505 	 */
2506 	if (snval < 0)
2507 		snval = -snval;
2508 
2509 	dtrace_multiply_128((uint64_t)snval, (uint64_t)snval, tmp);
2510 	dtrace_add_128(data + 2, tmp, data + 2);
2511 }
2512 
2513 /*ARGSUSED*/
2514 static void
2515 dtrace_aggregate_count(uint64_t *oval, uint64_t nval, uint64_t arg)
2516 {
2517 	*oval = *oval + 1;
2518 }
2519 
2520 /*ARGSUSED*/
2521 static void
2522 dtrace_aggregate_sum(uint64_t *oval, uint64_t nval, uint64_t arg)
2523 {
2524 	*oval += nval;
2525 }
2526 
2527 /*
2528  * Aggregate given the tuple in the principal data buffer, and the aggregating
2529  * action denoted by the specified dtrace_aggregation_t.  The aggregation
2530  * buffer is specified as the buf parameter.  This routine does not return
2531  * failure; if there is no space in the aggregation buffer, the data will be
2532  * dropped, and a corresponding counter incremented.
2533  */
2534 static void
2535 dtrace_aggregate(dtrace_aggregation_t *agg, dtrace_buffer_t *dbuf,
2536     intptr_t offset, dtrace_buffer_t *buf, uint64_t expr, uint64_t arg)
2537 {
2538 	dtrace_recdesc_t *rec = &agg->dtag_action.dta_rec;
2539 	uint32_t i, ndx, size, fsize;
2540 	uint32_t align = sizeof (uint64_t) - 1;
2541 	dtrace_aggbuffer_t *agb;
2542 	dtrace_aggkey_t *key;
2543 	uint32_t hashval = 0, limit, isstr;
2544 	caddr_t tomax, data, kdata;
2545 	dtrace_actkind_t action;
2546 	dtrace_action_t *act;
2547 	uintptr_t offs;
2548 
2549 	if (buf == NULL)
2550 		return;
2551 
2552 	if (!agg->dtag_hasarg) {
2553 		/*
2554 		 * Currently, only quantize() and lquantize() take additional
2555 		 * arguments, and they have the same semantics:  an increment
2556 		 * value that defaults to 1 when not present.  If additional
2557 		 * aggregating actions take arguments, the setting of the
2558 		 * default argument value will presumably have to become more
2559 		 * sophisticated...
2560 		 */
2561 		arg = 1;
2562 	}
2563 
2564 	action = agg->dtag_action.dta_kind - DTRACEACT_AGGREGATION;
2565 	size = rec->dtrd_offset - agg->dtag_base;
2566 	fsize = size + rec->dtrd_size;
2567 
2568 	ASSERT(dbuf->dtb_tomax != NULL);
2569 	data = dbuf->dtb_tomax + offset + agg->dtag_base;
2570 
2571 	if ((tomax = buf->dtb_tomax) == NULL) {
2572 		dtrace_buffer_drop(buf);
2573 		return;
2574 	}
2575 
2576 	/*
2577 	 * The metastructure is always at the bottom of the buffer.
2578 	 */
2579 	agb = (dtrace_aggbuffer_t *)(tomax + buf->dtb_size -
2580 	    sizeof (dtrace_aggbuffer_t));
2581 
2582 	if (buf->dtb_offset == 0) {
2583 		/*
2584 		 * We just kludge up approximately 1/8th of the size to be
2585 		 * buckets.  If this guess ends up being routinely
2586 		 * off-the-mark, we may need to dynamically readjust this
2587 		 * based on past performance.
2588 		 */
2589 		uintptr_t hashsize = (buf->dtb_size >> 3) / sizeof (uintptr_t);
2590 
2591 		if ((uintptr_t)agb - hashsize * sizeof (dtrace_aggkey_t *) <
2592 		    (uintptr_t)tomax || hashsize == 0) {
2593 			/*
2594 			 * We've been given a ludicrously small buffer;
2595 			 * increment our drop count and leave.
2596 			 */
2597 			dtrace_buffer_drop(buf);
2598 			return;
2599 		}
2600 
2601 		/*
2602 		 * And now, a pathetic attempt to try to get a an odd (or
2603 		 * perchance, a prime) hash size for better hash distribution.
2604 		 */
2605 		if (hashsize > (DTRACE_AGGHASHSIZE_SLEW << 3))
2606 			hashsize -= DTRACE_AGGHASHSIZE_SLEW;
2607 
2608 		agb->dtagb_hashsize = hashsize;
2609 		agb->dtagb_hash = (dtrace_aggkey_t **)((uintptr_t)agb -
2610 		    agb->dtagb_hashsize * sizeof (dtrace_aggkey_t *));
2611 		agb->dtagb_free = (uintptr_t)agb->dtagb_hash;
2612 
2613 		for (i = 0; i < agb->dtagb_hashsize; i++)
2614 			agb->dtagb_hash[i] = NULL;
2615 	}
2616 
2617 	ASSERT(agg->dtag_first != NULL);
2618 	ASSERT(agg->dtag_first->dta_intuple);
2619 
2620 	/*
2621 	 * Calculate the hash value based on the key.  Note that we _don't_
2622 	 * include the aggid in the hashing (but we will store it as part of
2623 	 * the key).  The hashing algorithm is Bob Jenkins' "One-at-a-time"
2624 	 * algorithm: a simple, quick algorithm that has no known funnels, and
2625 	 * gets good distribution in practice.  The efficacy of the hashing
2626 	 * algorithm (and a comparison with other algorithms) may be found by
2627 	 * running the ::dtrace_aggstat MDB dcmd.
2628 	 */
2629 	for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
2630 		i = act->dta_rec.dtrd_offset - agg->dtag_base;
2631 		limit = i + act->dta_rec.dtrd_size;
2632 		ASSERT(limit <= size);
2633 		isstr = DTRACEACT_ISSTRING(act);
2634 
2635 		for (; i < limit; i++) {
2636 			hashval += data[i];
2637 			hashval += (hashval << 10);
2638 			hashval ^= (hashval >> 6);
2639 
2640 			if (isstr && data[i] == '\0')
2641 				break;
2642 		}
2643 	}
2644 
2645 	hashval += (hashval << 3);
2646 	hashval ^= (hashval >> 11);
2647 	hashval += (hashval << 15);
2648 
2649 	/*
2650 	 * Yes, the divide here is expensive -- but it's generally the least
2651 	 * of the performance issues given the amount of data that we iterate
2652 	 * over to compute hash values, compare data, etc.
2653 	 */
2654 	ndx = hashval % agb->dtagb_hashsize;
2655 
2656 	for (key = agb->dtagb_hash[ndx]; key != NULL; key = key->dtak_next) {
2657 		ASSERT((caddr_t)key >= tomax);
2658 		ASSERT((caddr_t)key < tomax + buf->dtb_size);
2659 
2660 		if (hashval != key->dtak_hashval || key->dtak_size != size)
2661 			continue;
2662 
2663 		kdata = key->dtak_data;
2664 		ASSERT(kdata >= tomax && kdata < tomax + buf->dtb_size);
2665 
2666 		for (act = agg->dtag_first; act->dta_intuple;
2667 		    act = act->dta_next) {
2668 			i = act->dta_rec.dtrd_offset - agg->dtag_base;
2669 			limit = i + act->dta_rec.dtrd_size;
2670 			ASSERT(limit <= size);
2671 			isstr = DTRACEACT_ISSTRING(act);
2672 
2673 			for (; i < limit; i++) {
2674 				if (kdata[i] != data[i])
2675 					goto next;
2676 
2677 				if (isstr && data[i] == '\0')
2678 					break;
2679 			}
2680 		}
2681 
2682 		if (action != key->dtak_action) {
2683 			/*
2684 			 * We are aggregating on the same value in the same
2685 			 * aggregation with two different aggregating actions.
2686 			 * (This should have been picked up in the compiler,
2687 			 * so we may be dealing with errant or devious DIF.)
2688 			 * This is an error condition; we indicate as much,
2689 			 * and return.
2690 			 */
2691 			DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
2692 			return;
2693 		}
2694 
2695 		/*
2696 		 * This is a hit:  we need to apply the aggregator to
2697 		 * the value at this key.
2698 		 */
2699 		agg->dtag_aggregate((uint64_t *)(kdata + size), expr, arg);
2700 		return;
2701 next:
2702 		continue;
2703 	}
2704 
2705 	/*
2706 	 * We didn't find it.  We need to allocate some zero-filled space,
2707 	 * link it into the hash table appropriately, and apply the aggregator
2708 	 * to the (zero-filled) value.
2709 	 */
2710 	offs = buf->dtb_offset;
2711 	while (offs & (align - 1))
2712 		offs += sizeof (uint32_t);
2713 
2714 	/*
2715 	 * If we don't have enough room to both allocate a new key _and_
2716 	 * its associated data, increment the drop count and return.
2717 	 */
2718 	if ((uintptr_t)tomax + offs + fsize >
2719 	    agb->dtagb_free - sizeof (dtrace_aggkey_t)) {
2720 		dtrace_buffer_drop(buf);
2721 		return;
2722 	}
2723 
2724 	/*CONSTCOND*/
2725 	ASSERT(!(sizeof (dtrace_aggkey_t) & (sizeof (uintptr_t) - 1)));
2726 	key = (dtrace_aggkey_t *)(agb->dtagb_free - sizeof (dtrace_aggkey_t));
2727 	agb->dtagb_free -= sizeof (dtrace_aggkey_t);
2728 
2729 	key->dtak_data = kdata = tomax + offs;
2730 	buf->dtb_offset = offs + fsize;
2731 
2732 	/*
2733 	 * Now copy the data across.
2734 	 */
2735 	*((dtrace_aggid_t *)kdata) = agg->dtag_id;
2736 
2737 	for (i = sizeof (dtrace_aggid_t); i < size; i++)
2738 		kdata[i] = data[i];
2739 
2740 	/*
2741 	 * Because strings are not zeroed out by default, we need to iterate
2742 	 * looking for actions that store strings, and we need to explicitly
2743 	 * pad these strings out with zeroes.
2744 	 */
2745 	for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
2746 		int nul;
2747 
2748 		if (!DTRACEACT_ISSTRING(act))
2749 			continue;
2750 
2751 		i = act->dta_rec.dtrd_offset - agg->dtag_base;
2752 		limit = i + act->dta_rec.dtrd_size;
2753 		ASSERT(limit <= size);
2754 
2755 		for (nul = 0; i < limit; i++) {
2756 			if (nul) {
2757 				kdata[i] = '\0';
2758 				continue;
2759 			}
2760 
2761 			if (data[i] != '\0')
2762 				continue;
2763 
2764 			nul = 1;
2765 		}
2766 	}
2767 
2768 	for (i = size; i < fsize; i++)
2769 		kdata[i] = 0;
2770 
2771 	key->dtak_hashval = hashval;
2772 	key->dtak_size = size;
2773 	key->dtak_action = action;
2774 	key->dtak_next = agb->dtagb_hash[ndx];
2775 	agb->dtagb_hash[ndx] = key;
2776 
2777 	/*
2778 	 * Finally, apply the aggregator.
2779 	 */
2780 	*((uint64_t *)(key->dtak_data + size)) = agg->dtag_initial;
2781 	agg->dtag_aggregate((uint64_t *)(key->dtak_data + size), expr, arg);
2782 }
2783 
2784 /*
2785  * Given consumer state, this routine finds a speculation in the INACTIVE
2786  * state and transitions it into the ACTIVE state.  If there is no speculation
2787  * in the INACTIVE state, 0 is returned.  In this case, no error counter is
2788  * incremented -- it is up to the caller to take appropriate action.
2789  */
2790 static int
2791 dtrace_speculation(dtrace_state_t *state)
2792 {
2793 	int i = 0;
2794 	dtrace_speculation_state_t curstate;
2795 	uint32_t *stat = &state->dts_speculations_unavail, count;
2796 
2797 	while (i < state->dts_nspeculations) {
2798 		dtrace_speculation_t *spec = &state->dts_speculations[i];
2799 
2800 		curstate = spec->dtsp_state;
2801 
2802 		if (curstate != DTRACESPEC_INACTIVE) {
2803 			if (curstate == DTRACESPEC_COMMITTINGMANY ||
2804 			    curstate == DTRACESPEC_COMMITTING ||
2805 			    curstate == DTRACESPEC_DISCARDING)
2806 				stat = &state->dts_speculations_busy;
2807 			i++;
2808 			continue;
2809 		}
2810 
2811 		if (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2812 		    curstate, DTRACESPEC_ACTIVE) == curstate)
2813 			return (i + 1);
2814 	}
2815 
2816 	/*
2817 	 * We couldn't find a speculation.  If we found as much as a single
2818 	 * busy speculation buffer, we'll attribute this failure as "busy"
2819 	 * instead of "unavail".
2820 	 */
2821 	do {
2822 		count = *stat;
2823 	} while (dtrace_cas32(stat, count, count + 1) != count);
2824 
2825 	return (0);
2826 }
2827 
2828 /*
2829  * This routine commits an active speculation.  If the specified speculation
2830  * is not in a valid state to perform a commit(), this routine will silently do
2831  * nothing.  The state of the specified speculation is transitioned according
2832  * to the state transition diagram outlined in <sys/dtrace_impl.h>
2833  */
2834 static void
2835 dtrace_speculation_commit(dtrace_state_t *state, processorid_t cpu,
2836     dtrace_specid_t which)
2837 {
2838 	dtrace_speculation_t *spec;
2839 	dtrace_buffer_t *src, *dest;
2840 	uintptr_t daddr, saddr, dlimit, slimit;
2841 	dtrace_speculation_state_t curstate, new = 0;
2842 	intptr_t offs;
2843 	uint64_t timestamp;
2844 
2845 	if (which == 0)
2846 		return;
2847 
2848 	if (which > state->dts_nspeculations) {
2849 		cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2850 		return;
2851 	}
2852 
2853 	spec = &state->dts_speculations[which - 1];
2854 	src = &spec->dtsp_buffer[cpu];
2855 	dest = &state->dts_buffer[cpu];
2856 
2857 	do {
2858 		curstate = spec->dtsp_state;
2859 
2860 		if (curstate == DTRACESPEC_COMMITTINGMANY)
2861 			break;
2862 
2863 		switch (curstate) {
2864 		case DTRACESPEC_INACTIVE:
2865 		case DTRACESPEC_DISCARDING:
2866 			return;
2867 
2868 		case DTRACESPEC_COMMITTING:
2869 			/*
2870 			 * This is only possible if we are (a) commit()'ing
2871 			 * without having done a prior speculate() on this CPU
2872 			 * and (b) racing with another commit() on a different
2873 			 * CPU.  There's nothing to do -- we just assert that
2874 			 * our offset is 0.
2875 			 */
2876 			ASSERT(src->dtb_offset == 0);
2877 			return;
2878 
2879 		case DTRACESPEC_ACTIVE:
2880 			new = DTRACESPEC_COMMITTING;
2881 			break;
2882 
2883 		case DTRACESPEC_ACTIVEONE:
2884 			/*
2885 			 * This speculation is active on one CPU.  If our
2886 			 * buffer offset is non-zero, we know that the one CPU
2887 			 * must be us.  Otherwise, we are committing on a
2888 			 * different CPU from the speculate(), and we must
2889 			 * rely on being asynchronously cleaned.
2890 			 */
2891 			if (src->dtb_offset != 0) {
2892 				new = DTRACESPEC_COMMITTING;
2893 				break;
2894 			}
2895 			/*FALLTHROUGH*/
2896 
2897 		case DTRACESPEC_ACTIVEMANY:
2898 			new = DTRACESPEC_COMMITTINGMANY;
2899 			break;
2900 
2901 		default:
2902 			ASSERT(0);
2903 		}
2904 	} while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2905 	    curstate, new) != curstate);
2906 
2907 	/*
2908 	 * We have set the state to indicate that we are committing this
2909 	 * speculation.  Now reserve the necessary space in the destination
2910 	 * buffer.
2911 	 */
2912 	if ((offs = dtrace_buffer_reserve(dest, src->dtb_offset,
2913 	    sizeof (uint64_t), state, NULL)) < 0) {
2914 		dtrace_buffer_drop(dest);
2915 		goto out;
2916 	}
2917 
2918 	/*
2919 	 * We have sufficient space to copy the speculative buffer into the
2920 	 * primary buffer.  First, modify the speculative buffer, filling
2921 	 * in the timestamp of all entries with the curstate time.  The data
2922 	 * must have the commit() time rather than the time it was traced,
2923 	 * so that all entries in the primary buffer are in timestamp order.
2924 	 */
2925 	timestamp = dtrace_gethrtime();
2926 	saddr = (uintptr_t)src->dtb_tomax;
2927 	slimit = saddr + src->dtb_offset;
2928 	while (saddr < slimit) {
2929 		size_t size;
2930 		dtrace_rechdr_t *dtrh = (dtrace_rechdr_t *)saddr;
2931 
2932 		if (dtrh->dtrh_epid == DTRACE_EPIDNONE) {
2933 			saddr += sizeof (dtrace_epid_t);
2934 			continue;
2935 		}
2936 		ASSERT3U(dtrh->dtrh_epid, <=, state->dts_necbs);
2937 		size = state->dts_ecbs[dtrh->dtrh_epid - 1]->dte_size;
2938 
2939 		ASSERT3U(saddr + size, <=, slimit);
2940 		ASSERT3U(size, >=, sizeof (dtrace_rechdr_t));
2941 		ASSERT3U(DTRACE_RECORD_LOAD_TIMESTAMP(dtrh), ==, UINT64_MAX);
2942 
2943 		DTRACE_RECORD_STORE_TIMESTAMP(dtrh, timestamp);
2944 
2945 		saddr += size;
2946 	}
2947 
2948 	/*
2949 	 * Copy the buffer across.  (Note that this is a
2950 	 * highly subobtimal bcopy(); in the unlikely event that this becomes
2951 	 * a serious performance issue, a high-performance DTrace-specific
2952 	 * bcopy() should obviously be invented.)
2953 	 */
2954 	daddr = (uintptr_t)dest->dtb_tomax + offs;
2955 	dlimit = daddr + src->dtb_offset;
2956 	saddr = (uintptr_t)src->dtb_tomax;
2957 
2958 	/*
2959 	 * First, the aligned portion.
2960 	 */
2961 	while (dlimit - daddr >= sizeof (uint64_t)) {
2962 		*((uint64_t *)daddr) = *((uint64_t *)saddr);
2963 
2964 		daddr += sizeof (uint64_t);
2965 		saddr += sizeof (uint64_t);
2966 	}
2967 
2968 	/*
2969 	 * Now any left-over bit...
2970 	 */
2971 	while (dlimit - daddr)
2972 		*((uint8_t *)daddr++) = *((uint8_t *)saddr++);
2973 
2974 	/*
2975 	 * Finally, commit the reserved space in the destination buffer.
2976 	 */
2977 	dest->dtb_offset = offs + src->dtb_offset;
2978 
2979 out:
2980 	/*
2981 	 * If we're lucky enough to be the only active CPU on this speculation
2982 	 * buffer, we can just set the state back to DTRACESPEC_INACTIVE.
2983 	 */
2984 	if (curstate == DTRACESPEC_ACTIVE ||
2985 	    (curstate == DTRACESPEC_ACTIVEONE && new == DTRACESPEC_COMMITTING)) {
2986 		uint32_t rval = dtrace_cas32((uint32_t *)&spec->dtsp_state,
2987 		    DTRACESPEC_COMMITTING, DTRACESPEC_INACTIVE);
2988 
2989 		ASSERT(rval == DTRACESPEC_COMMITTING);
2990 	}
2991 
2992 	src->dtb_offset = 0;
2993 	src->dtb_xamot_drops += src->dtb_drops;
2994 	src->dtb_drops = 0;
2995 }
2996 
2997 /*
2998  * This routine discards an active speculation.  If the specified speculation
2999  * is not in a valid state to perform a discard(), this routine will silently
3000  * do nothing.  The state of the specified speculation is transitioned
3001  * according to the state transition diagram outlined in <sys/dtrace_impl.h>
3002  */
3003 static void
3004 dtrace_speculation_discard(dtrace_state_t *state, processorid_t cpu,
3005     dtrace_specid_t which)
3006 {
3007 	dtrace_speculation_t *spec;
3008 	dtrace_speculation_state_t curstate, new = 0;
3009 	dtrace_buffer_t *buf;
3010 
3011 	if (which == 0)
3012 		return;
3013 
3014 	if (which > state->dts_nspeculations) {
3015 		cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
3016 		return;
3017 	}
3018 
3019 	spec = &state->dts_speculations[which - 1];
3020 	buf = &spec->dtsp_buffer[cpu];
3021 
3022 	do {
3023 		curstate = spec->dtsp_state;
3024 
3025 		switch (curstate) {
3026 		case DTRACESPEC_INACTIVE:
3027 		case DTRACESPEC_COMMITTINGMANY:
3028 		case DTRACESPEC_COMMITTING:
3029 		case DTRACESPEC_DISCARDING:
3030 			return;
3031 
3032 		case DTRACESPEC_ACTIVE:
3033 		case DTRACESPEC_ACTIVEMANY:
3034 			new = DTRACESPEC_DISCARDING;
3035 			break;
3036 
3037 		case DTRACESPEC_ACTIVEONE:
3038 			if (buf->dtb_offset != 0) {
3039 				new = DTRACESPEC_INACTIVE;
3040 			} else {
3041 				new = DTRACESPEC_DISCARDING;
3042 			}
3043 			break;
3044 
3045 		default:
3046 			ASSERT(0);
3047 		}
3048 	} while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
3049 	    curstate, new) != curstate);
3050 
3051 	buf->dtb_offset = 0;
3052 	buf->dtb_drops = 0;
3053 }
3054 
3055 /*
3056  * Note:  not called from probe context.  This function is called
3057  * asynchronously from cross call context to clean any speculations that are
3058  * in the COMMITTINGMANY or DISCARDING states.  These speculations may not be
3059  * transitioned back to the INACTIVE state until all CPUs have cleaned the
3060  * speculation.
3061  */
3062 static void
3063 dtrace_speculation_clean_here(dtrace_state_t *state)
3064 {
3065 	dtrace_icookie_t cookie;
3066 	processorid_t cpu = curcpu;
3067 	dtrace_buffer_t *dest = &state->dts_buffer[cpu];
3068 	dtrace_specid_t i;
3069 
3070 	cookie = dtrace_interrupt_disable();
3071 
3072 	if (dest->dtb_tomax == NULL) {
3073 		dtrace_interrupt_enable(cookie);
3074 		return;
3075 	}
3076 
3077 	for (i = 0; i < state->dts_nspeculations; i++) {
3078 		dtrace_speculation_t *spec = &state->dts_speculations[i];
3079 		dtrace_buffer_t *src = &spec->dtsp_buffer[cpu];
3080 
3081 		if (src->dtb_tomax == NULL)
3082 			continue;
3083 
3084 		if (spec->dtsp_state == DTRACESPEC_DISCARDING) {
3085 			src->dtb_offset = 0;
3086 			continue;
3087 		}
3088 
3089 		if (spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
3090 			continue;
3091 
3092 		if (src->dtb_offset == 0)
3093 			continue;
3094 
3095 		dtrace_speculation_commit(state, cpu, i + 1);
3096 	}
3097 
3098 	dtrace_interrupt_enable(cookie);
3099 }
3100 
3101 /*
3102  * Note:  not called from probe context.  This function is called
3103  * asynchronously (and at a regular interval) to clean any speculations that
3104  * are in the COMMITTINGMANY or DISCARDING states.  If it discovers that there
3105  * is work to be done, it cross calls all CPUs to perform that work;
3106  * COMMITMANY and DISCARDING speculations may not be transitioned back to the
3107  * INACTIVE state until they have been cleaned by all CPUs.
3108  */
3109 static void
3110 dtrace_speculation_clean(dtrace_state_t *state)
3111 {
3112 	int work = 0, rv;
3113 	dtrace_specid_t i;
3114 
3115 	for (i = 0; i < state->dts_nspeculations; i++) {
3116 		dtrace_speculation_t *spec = &state->dts_speculations[i];
3117 
3118 		ASSERT(!spec->dtsp_cleaning);
3119 
3120 		if (spec->dtsp_state != DTRACESPEC_DISCARDING &&
3121 		    spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
3122 			continue;
3123 
3124 		work++;
3125 		spec->dtsp_cleaning = 1;
3126 	}
3127 
3128 	if (!work)
3129 		return;
3130 
3131 	dtrace_xcall(DTRACE_CPUALL,
3132 	    (dtrace_xcall_t)dtrace_speculation_clean_here, state);
3133 
3134 	/*
3135 	 * We now know that all CPUs have committed or discarded their
3136 	 * speculation buffers, as appropriate.  We can now set the state
3137 	 * to inactive.
3138 	 */
3139 	for (i = 0; i < state->dts_nspeculations; i++) {
3140 		dtrace_speculation_t *spec = &state->dts_speculations[i];
3141 		dtrace_speculation_state_t curstate, new;
3142 
3143 		if (!spec->dtsp_cleaning)
3144 			continue;
3145 
3146 		curstate = spec->dtsp_state;
3147 		ASSERT(curstate == DTRACESPEC_DISCARDING ||
3148 		    curstate == DTRACESPEC_COMMITTINGMANY);
3149 
3150 		new = DTRACESPEC_INACTIVE;
3151 
3152 		rv = dtrace_cas32((uint32_t *)&spec->dtsp_state, curstate, new);
3153 		ASSERT(rv == curstate);
3154 		spec->dtsp_cleaning = 0;
3155 	}
3156 }
3157 
3158 /*
3159  * Called as part of a speculate() to get the speculative buffer associated
3160  * with a given speculation.  Returns NULL if the specified speculation is not
3161  * in an ACTIVE state.  If the speculation is in the ACTIVEONE state -- and
3162  * the active CPU is not the specified CPU -- the speculation will be
3163  * atomically transitioned into the ACTIVEMANY state.
3164  */
3165 static dtrace_buffer_t *
3166 dtrace_speculation_buffer(dtrace_state_t *state, processorid_t cpuid,
3167     dtrace_specid_t which)
3168 {
3169 	dtrace_speculation_t *spec;
3170 	dtrace_speculation_state_t curstate, new = 0;
3171 	dtrace_buffer_t *buf;
3172 
3173 	if (which == 0)
3174 		return (NULL);
3175 
3176 	if (which > state->dts_nspeculations) {
3177 		cpu_core[cpuid].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
3178 		return (NULL);
3179 	}
3180 
3181 	spec = &state->dts_speculations[which - 1];
3182 	buf = &spec->dtsp_buffer[cpuid];
3183 
3184 	do {
3185 		curstate = spec->dtsp_state;
3186 
3187 		switch (curstate) {
3188 		case DTRACESPEC_INACTIVE:
3189 		case DTRACESPEC_COMMITTINGMANY:
3190 		case DTRACESPEC_DISCARDING:
3191 			return (NULL);
3192 
3193 		case DTRACESPEC_COMMITTING:
3194 			ASSERT(buf->dtb_offset == 0);
3195 			return (NULL);
3196 
3197 		case DTRACESPEC_ACTIVEONE:
3198 			/*
3199 			 * This speculation is currently active on one CPU.
3200 			 * Check the offset in the buffer; if it's non-zero,
3201 			 * that CPU must be us (and we leave the state alone).
3202 			 * If it's zero, assume that we're starting on a new
3203 			 * CPU -- and change the state to indicate that the
3204 			 * speculation is active on more than one CPU.
3205 			 */
3206 			if (buf->dtb_offset != 0)
3207 				return (buf);
3208 
3209 			new = DTRACESPEC_ACTIVEMANY;
3210 			break;
3211 
3212 		case DTRACESPEC_ACTIVEMANY:
3213 			return (buf);
3214 
3215 		case DTRACESPEC_ACTIVE:
3216 			new = DTRACESPEC_ACTIVEONE;
3217 			break;
3218 
3219 		default:
3220 			ASSERT(0);
3221 		}
3222 	} while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
3223 	    curstate, new) != curstate);
3224 
3225 	ASSERT(new == DTRACESPEC_ACTIVEONE || new == DTRACESPEC_ACTIVEMANY);
3226 	return (buf);
3227 }
3228 
3229 /*
3230  * Return a string.  In the event that the user lacks the privilege to access
3231  * arbitrary kernel memory, we copy the string out to scratch memory so that we
3232  * don't fail access checking.
3233  *
3234  * dtrace_dif_variable() uses this routine as a helper for various
3235  * builtin values such as 'execname' and 'probefunc.'
3236  */
3237 uintptr_t
3238 dtrace_dif_varstr(uintptr_t addr, dtrace_state_t *state,
3239     dtrace_mstate_t *mstate)
3240 {
3241 	uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3242 	uintptr_t ret;
3243 	size_t strsz;
3244 
3245 	/*
3246 	 * The easy case: this probe is allowed to read all of memory, so
3247 	 * we can just return this as a vanilla pointer.
3248 	 */
3249 	if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
3250 		return (addr);
3251 
3252 	/*
3253 	 * This is the tougher case: we copy the string in question from
3254 	 * kernel memory into scratch memory and return it that way: this
3255 	 * ensures that we won't trip up when access checking tests the
3256 	 * BYREF return value.
3257 	 */
3258 	strsz = dtrace_strlen((char *)addr, size) + 1;
3259 
3260 	if (mstate->dtms_scratch_ptr + strsz >
3261 	    mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
3262 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3263 		return (0);
3264 	}
3265 
3266 	dtrace_strcpy((const void *)addr, (void *)mstate->dtms_scratch_ptr,
3267 	    strsz);
3268 	ret = mstate->dtms_scratch_ptr;
3269 	mstate->dtms_scratch_ptr += strsz;
3270 	return (ret);
3271 }
3272 
3273 /*
3274  * Return a string from a memoy address which is known to have one or
3275  * more concatenated, individually zero terminated, sub-strings.
3276  * In the event that the user lacks the privilege to access
3277  * arbitrary kernel memory, we copy the string out to scratch memory so that we
3278  * don't fail access checking.
3279  *
3280  * dtrace_dif_variable() uses this routine as a helper for various
3281  * builtin values such as 'execargs'.
3282  */
3283 static uintptr_t
3284 dtrace_dif_varstrz(uintptr_t addr, size_t strsz, dtrace_state_t *state,
3285     dtrace_mstate_t *mstate)
3286 {
3287 	char *p;
3288 	size_t i;
3289 	uintptr_t ret;
3290 
3291 	if (mstate->dtms_scratch_ptr + strsz >
3292 	    mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
3293 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3294 		return (0);
3295 	}
3296 
3297 	dtrace_bcopy((const void *)addr, (void *)mstate->dtms_scratch_ptr,
3298 	    strsz);
3299 
3300 	/* Replace sub-string termination characters with a space. */
3301 	for (p = (char *) mstate->dtms_scratch_ptr, i = 0; i < strsz - 1;
3302 	    p++, i++)
3303 		if (*p == '\0')
3304 			*p = ' ';
3305 
3306 	ret = mstate->dtms_scratch_ptr;
3307 	mstate->dtms_scratch_ptr += strsz;
3308 	return (ret);
3309 }
3310 
3311 /*
3312  * This function implements the DIF emulator's variable lookups.  The emulator
3313  * passes a reserved variable identifier and optional built-in array index.
3314  */
3315 static uint64_t
3316 dtrace_dif_variable(dtrace_mstate_t *mstate, dtrace_state_t *state, uint64_t v,
3317     uint64_t ndx)
3318 {
3319 	/*
3320 	 * If we're accessing one of the uncached arguments, we'll turn this
3321 	 * into a reference in the args array.
3322 	 */
3323 	if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9) {
3324 		ndx = v - DIF_VAR_ARG0;
3325 		v = DIF_VAR_ARGS;
3326 	}
3327 
3328 	switch (v) {
3329 	case DIF_VAR_ARGS:
3330 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_ARGS);
3331 		if (ndx >= sizeof (mstate->dtms_arg) /
3332 		    sizeof (mstate->dtms_arg[0])) {
3333 			int aframes = mstate->dtms_probe->dtpr_aframes + 2;
3334 			dtrace_provider_t *pv;
3335 			uint64_t val;
3336 
3337 			pv = mstate->dtms_probe->dtpr_provider;
3338 			if (pv->dtpv_pops.dtps_getargval != NULL)
3339 				val = pv->dtpv_pops.dtps_getargval(pv->dtpv_arg,
3340 				    mstate->dtms_probe->dtpr_id,
3341 				    mstate->dtms_probe->dtpr_arg, ndx, aframes);
3342 			else
3343 				val = dtrace_getarg(ndx, aframes);
3344 
3345 			/*
3346 			 * This is regrettably required to keep the compiler
3347 			 * from tail-optimizing the call to dtrace_getarg().
3348 			 * The condition always evaluates to true, but the
3349 			 * compiler has no way of figuring that out a priori.
3350 			 * (None of this would be necessary if the compiler
3351 			 * could be relied upon to _always_ tail-optimize
3352 			 * the call to dtrace_getarg() -- but it can't.)
3353 			 */
3354 			if (mstate->dtms_probe != NULL)
3355 				return (val);
3356 
3357 			ASSERT(0);
3358 		}
3359 
3360 		return (mstate->dtms_arg[ndx]);
3361 
3362 #ifdef illumos
3363 	case DIF_VAR_UREGS: {
3364 		klwp_t *lwp;
3365 
3366 		if (!dtrace_priv_proc(state))
3367 			return (0);
3368 
3369 		if ((lwp = curthread->t_lwp) == NULL) {
3370 			DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
3371 			cpu_core[curcpu].cpuc_dtrace_illval = NULL;
3372 			return (0);
3373 		}
3374 
3375 		return (dtrace_getreg(lwp->lwp_regs, ndx));
3376 		return (0);
3377 	}
3378 #else
3379 	case DIF_VAR_UREGS: {
3380 		struct trapframe *tframe;
3381 
3382 		if (!dtrace_priv_proc(state))
3383 			return (0);
3384 
3385 		if ((tframe = curthread->td_frame) == NULL) {
3386 			DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
3387 			cpu_core[curcpu].cpuc_dtrace_illval = 0;
3388 			return (0);
3389 		}
3390 
3391 		return (dtrace_getreg(tframe, ndx));
3392 	}
3393 #endif
3394 
3395 	case DIF_VAR_CURTHREAD:
3396 		if (!dtrace_priv_proc(state))
3397 			return (0);
3398 		return ((uint64_t)(uintptr_t)curthread);
3399 
3400 	case DIF_VAR_TIMESTAMP:
3401 		if (!(mstate->dtms_present & DTRACE_MSTATE_TIMESTAMP)) {
3402 			mstate->dtms_timestamp = dtrace_gethrtime();
3403 			mstate->dtms_present |= DTRACE_MSTATE_TIMESTAMP;
3404 		}
3405 		return (mstate->dtms_timestamp);
3406 
3407 	case DIF_VAR_VTIMESTAMP:
3408 		ASSERT(dtrace_vtime_references != 0);
3409 		return (curthread->t_dtrace_vtime);
3410 
3411 	case DIF_VAR_WALLTIMESTAMP:
3412 		if (!(mstate->dtms_present & DTRACE_MSTATE_WALLTIMESTAMP)) {
3413 			mstate->dtms_walltimestamp = dtrace_gethrestime();
3414 			mstate->dtms_present |= DTRACE_MSTATE_WALLTIMESTAMP;
3415 		}
3416 		return (mstate->dtms_walltimestamp);
3417 
3418 #ifdef illumos
3419 	case DIF_VAR_IPL:
3420 		if (!dtrace_priv_kernel(state))
3421 			return (0);
3422 		if (!(mstate->dtms_present & DTRACE_MSTATE_IPL)) {
3423 			mstate->dtms_ipl = dtrace_getipl();
3424 			mstate->dtms_present |= DTRACE_MSTATE_IPL;
3425 		}
3426 		return (mstate->dtms_ipl);
3427 #endif
3428 
3429 	case DIF_VAR_EPID:
3430 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_EPID);
3431 		return (mstate->dtms_epid);
3432 
3433 	case DIF_VAR_ID:
3434 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3435 		return (mstate->dtms_probe->dtpr_id);
3436 
3437 	case DIF_VAR_STACKDEPTH:
3438 		if (!dtrace_priv_kernel(state))
3439 			return (0);
3440 		if (!(mstate->dtms_present & DTRACE_MSTATE_STACKDEPTH)) {
3441 			int aframes = mstate->dtms_probe->dtpr_aframes + 2;
3442 
3443 			mstate->dtms_stackdepth = dtrace_getstackdepth(aframes);
3444 			mstate->dtms_present |= DTRACE_MSTATE_STACKDEPTH;
3445 		}
3446 		return (mstate->dtms_stackdepth);
3447 
3448 	case DIF_VAR_USTACKDEPTH:
3449 		if (!dtrace_priv_proc(state))
3450 			return (0);
3451 		if (!(mstate->dtms_present & DTRACE_MSTATE_USTACKDEPTH)) {
3452 			/*
3453 			 * See comment in DIF_VAR_PID.
3454 			 */
3455 			if (DTRACE_ANCHORED(mstate->dtms_probe) &&
3456 			    CPU_ON_INTR(CPU)) {
3457 				mstate->dtms_ustackdepth = 0;
3458 			} else {
3459 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3460 				mstate->dtms_ustackdepth =
3461 				    dtrace_getustackdepth();
3462 				DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3463 			}
3464 			mstate->dtms_present |= DTRACE_MSTATE_USTACKDEPTH;
3465 		}
3466 		return (mstate->dtms_ustackdepth);
3467 
3468 	case DIF_VAR_CALLER:
3469 		if (!dtrace_priv_kernel(state))
3470 			return (0);
3471 		if (!(mstate->dtms_present & DTRACE_MSTATE_CALLER)) {
3472 			int aframes = mstate->dtms_probe->dtpr_aframes + 2;
3473 
3474 			if (!DTRACE_ANCHORED(mstate->dtms_probe)) {
3475 				/*
3476 				 * If this is an unanchored probe, we are
3477 				 * required to go through the slow path:
3478 				 * dtrace_caller() only guarantees correct
3479 				 * results for anchored probes.
3480 				 */
3481 				pc_t caller[2] = {0, 0};
3482 
3483 				dtrace_getpcstack(caller, 2, aframes,
3484 				    (uint32_t *)(uintptr_t)mstate->dtms_arg[0]);
3485 				mstate->dtms_caller = caller[1];
3486 			} else if ((mstate->dtms_caller =
3487 			    dtrace_caller(aframes)) == -1) {
3488 				/*
3489 				 * We have failed to do this the quick way;
3490 				 * we must resort to the slower approach of
3491 				 * calling dtrace_getpcstack().
3492 				 */
3493 				pc_t caller = 0;
3494 
3495 				dtrace_getpcstack(&caller, 1, aframes, NULL);
3496 				mstate->dtms_caller = caller;
3497 			}
3498 
3499 			mstate->dtms_present |= DTRACE_MSTATE_CALLER;
3500 		}
3501 		return (mstate->dtms_caller);
3502 
3503 	case DIF_VAR_UCALLER:
3504 		if (!dtrace_priv_proc(state))
3505 			return (0);
3506 
3507 		if (!(mstate->dtms_present & DTRACE_MSTATE_UCALLER)) {
3508 			uint64_t ustack[3];
3509 
3510 			/*
3511 			 * dtrace_getupcstack() fills in the first uint64_t
3512 			 * with the current PID.  The second uint64_t will
3513 			 * be the program counter at user-level.  The third
3514 			 * uint64_t will contain the caller, which is what
3515 			 * we're after.
3516 			 */
3517 			ustack[2] = 0;
3518 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3519 			dtrace_getupcstack(ustack, 3);
3520 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3521 			mstate->dtms_ucaller = ustack[2];
3522 			mstate->dtms_present |= DTRACE_MSTATE_UCALLER;
3523 		}
3524 
3525 		return (mstate->dtms_ucaller);
3526 
3527 	case DIF_VAR_PROBEPROV:
3528 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3529 		return (dtrace_dif_varstr(
3530 		    (uintptr_t)mstate->dtms_probe->dtpr_provider->dtpv_name,
3531 		    state, mstate));
3532 
3533 	case DIF_VAR_PROBEMOD:
3534 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3535 		return (dtrace_dif_varstr(
3536 		    (uintptr_t)mstate->dtms_probe->dtpr_mod,
3537 		    state, mstate));
3538 
3539 	case DIF_VAR_PROBEFUNC:
3540 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3541 		return (dtrace_dif_varstr(
3542 		    (uintptr_t)mstate->dtms_probe->dtpr_func,
3543 		    state, mstate));
3544 
3545 	case DIF_VAR_PROBENAME:
3546 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3547 		return (dtrace_dif_varstr(
3548 		    (uintptr_t)mstate->dtms_probe->dtpr_name,
3549 		    state, mstate));
3550 
3551 	case DIF_VAR_PID:
3552 		if (!dtrace_priv_proc(state))
3553 			return (0);
3554 
3555 #ifdef illumos
3556 		/*
3557 		 * Note that we are assuming that an unanchored probe is
3558 		 * always due to a high-level interrupt.  (And we're assuming
3559 		 * that there is only a single high level interrupt.)
3560 		 */
3561 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3562 			return (pid0.pid_id);
3563 
3564 		/*
3565 		 * It is always safe to dereference one's own t_procp pointer:
3566 		 * it always points to a valid, allocated proc structure.
3567 		 * Further, it is always safe to dereference the p_pidp member
3568 		 * of one's own proc structure.  (These are truisms becuase
3569 		 * threads and processes don't clean up their own state --
3570 		 * they leave that task to whomever reaps them.)
3571 		 */
3572 		return ((uint64_t)curthread->t_procp->p_pidp->pid_id);
3573 #else
3574 		return ((uint64_t)curproc->p_pid);
3575 #endif
3576 
3577 	case DIF_VAR_PPID:
3578 		if (!dtrace_priv_proc(state))
3579 			return (0);
3580 
3581 #ifdef illumos
3582 		/*
3583 		 * See comment in DIF_VAR_PID.
3584 		 */
3585 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3586 			return (pid0.pid_id);
3587 
3588 		/*
3589 		 * It is always safe to dereference one's own t_procp pointer:
3590 		 * it always points to a valid, allocated proc structure.
3591 		 * (This is true because threads don't clean up their own
3592 		 * state -- they leave that task to whomever reaps them.)
3593 		 */
3594 		return ((uint64_t)curthread->t_procp->p_ppid);
3595 #else
3596 		if (curproc->p_pid == proc0.p_pid)
3597 			return (curproc->p_pid);
3598 		else
3599 			return (curproc->p_pptr->p_pid);
3600 #endif
3601 
3602 	case DIF_VAR_TID:
3603 #ifdef illumos
3604 		/*
3605 		 * See comment in DIF_VAR_PID.
3606 		 */
3607 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3608 			return (0);
3609 #endif
3610 
3611 		return ((uint64_t)curthread->t_tid);
3612 
3613 	case DIF_VAR_EXECARGS: {
3614 		struct pargs *p_args = curthread->td_proc->p_args;
3615 
3616 		if (p_args == NULL)
3617 			return(0);
3618 
3619 		return (dtrace_dif_varstrz(
3620 		    (uintptr_t) p_args->ar_args, p_args->ar_length, state, mstate));
3621 	}
3622 
3623 	case DIF_VAR_EXECNAME:
3624 #ifdef illumos
3625 		if (!dtrace_priv_proc(state))
3626 			return (0);
3627 
3628 		/*
3629 		 * See comment in DIF_VAR_PID.
3630 		 */
3631 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3632 			return ((uint64_t)(uintptr_t)p0.p_user.u_comm);
3633 
3634 		/*
3635 		 * It is always safe to dereference one's own t_procp pointer:
3636 		 * it always points to a valid, allocated proc structure.
3637 		 * (This is true because threads don't clean up their own
3638 		 * state -- they leave that task to whomever reaps them.)
3639 		 */
3640 		return (dtrace_dif_varstr(
3641 		    (uintptr_t)curthread->t_procp->p_user.u_comm,
3642 		    state, mstate));
3643 #else
3644 		return (dtrace_dif_varstr(
3645 		    (uintptr_t) curthread->td_proc->p_comm, state, mstate));
3646 #endif
3647 
3648 	case DIF_VAR_ZONENAME:
3649 #ifdef illumos
3650 		if (!dtrace_priv_proc(state))
3651 			return (0);
3652 
3653 		/*
3654 		 * See comment in DIF_VAR_PID.
3655 		 */
3656 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3657 			return ((uint64_t)(uintptr_t)p0.p_zone->zone_name);
3658 
3659 		/*
3660 		 * It is always safe to dereference one's own t_procp pointer:
3661 		 * it always points to a valid, allocated proc structure.
3662 		 * (This is true because threads don't clean up their own
3663 		 * state -- they leave that task to whomever reaps them.)
3664 		 */
3665 		return (dtrace_dif_varstr(
3666 		    (uintptr_t)curthread->t_procp->p_zone->zone_name,
3667 		    state, mstate));
3668 #elif defined(__FreeBSD__)
3669 	/*
3670 	 * On FreeBSD, we introduce compatibility to zonename by falling through
3671 	 * into jailname.
3672 	 */
3673 	case DIF_VAR_JAILNAME:
3674 		if (!dtrace_priv_kernel(state))
3675 			return (0);
3676 
3677 		return (dtrace_dif_varstr(
3678 		    (uintptr_t)curthread->td_ucred->cr_prison->pr_name,
3679 		    state, mstate));
3680 
3681 	case DIF_VAR_JID:
3682 		if (!dtrace_priv_kernel(state))
3683 			return (0);
3684 
3685 		return ((uint64_t)curthread->td_ucred->cr_prison->pr_id);
3686 #else
3687 		return (0);
3688 #endif
3689 
3690 	case DIF_VAR_UID:
3691 		if (!dtrace_priv_proc(state))
3692 			return (0);
3693 
3694 #ifdef illumos
3695 		/*
3696 		 * See comment in DIF_VAR_PID.
3697 		 */
3698 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3699 			return ((uint64_t)p0.p_cred->cr_uid);
3700 
3701 		/*
3702 		 * It is always safe to dereference one's own t_procp pointer:
3703 		 * it always points to a valid, allocated proc structure.
3704 		 * (This is true because threads don't clean up their own
3705 		 * state -- they leave that task to whomever reaps them.)
3706 		 *
3707 		 * Additionally, it is safe to dereference one's own process
3708 		 * credential, since this is never NULL after process birth.
3709 		 */
3710 		return ((uint64_t)curthread->t_procp->p_cred->cr_uid);
3711 #else
3712 		return ((uint64_t)curthread->td_ucred->cr_uid);
3713 #endif
3714 
3715 	case DIF_VAR_GID:
3716 		if (!dtrace_priv_proc(state))
3717 			return (0);
3718 
3719 #ifdef illumos
3720 		/*
3721 		 * See comment in DIF_VAR_PID.
3722 		 */
3723 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3724 			return ((uint64_t)p0.p_cred->cr_gid);
3725 
3726 		/*
3727 		 * It is always safe to dereference one's own t_procp pointer:
3728 		 * it always points to a valid, allocated proc structure.
3729 		 * (This is true because threads don't clean up their own
3730 		 * state -- they leave that task to whomever reaps them.)
3731 		 *
3732 		 * Additionally, it is safe to dereference one's own process
3733 		 * credential, since this is never NULL after process birth.
3734 		 */
3735 		return ((uint64_t)curthread->t_procp->p_cred->cr_gid);
3736 #else
3737 		return ((uint64_t)curthread->td_ucred->cr_gid);
3738 #endif
3739 
3740 	case DIF_VAR_ERRNO: {
3741 #ifdef illumos
3742 		klwp_t *lwp;
3743 		if (!dtrace_priv_proc(state))
3744 			return (0);
3745 
3746 		/*
3747 		 * See comment in DIF_VAR_PID.
3748 		 */
3749 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3750 			return (0);
3751 
3752 		/*
3753 		 * It is always safe to dereference one's own t_lwp pointer in
3754 		 * the event that this pointer is non-NULL.  (This is true
3755 		 * because threads and lwps don't clean up their own state --
3756 		 * they leave that task to whomever reaps them.)
3757 		 */
3758 		if ((lwp = curthread->t_lwp) == NULL)
3759 			return (0);
3760 
3761 		return ((uint64_t)lwp->lwp_errno);
3762 #else
3763 		return (curthread->td_errno);
3764 #endif
3765 	}
3766 #ifndef illumos
3767 	case DIF_VAR_CPU: {
3768 		return curcpu;
3769 	}
3770 #endif
3771 	default:
3772 		DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
3773 		return (0);
3774 	}
3775 }
3776 
3777 
3778 typedef enum dtrace_json_state {
3779 	DTRACE_JSON_REST = 1,
3780 	DTRACE_JSON_OBJECT,
3781 	DTRACE_JSON_STRING,
3782 	DTRACE_JSON_STRING_ESCAPE,
3783 	DTRACE_JSON_STRING_ESCAPE_UNICODE,
3784 	DTRACE_JSON_COLON,
3785 	DTRACE_JSON_COMMA,
3786 	DTRACE_JSON_VALUE,
3787 	DTRACE_JSON_IDENTIFIER,
3788 	DTRACE_JSON_NUMBER,
3789 	DTRACE_JSON_NUMBER_FRAC,
3790 	DTRACE_JSON_NUMBER_EXP,
3791 	DTRACE_JSON_COLLECT_OBJECT
3792 } dtrace_json_state_t;
3793 
3794 /*
3795  * This function possesses just enough knowledge about JSON to extract a single
3796  * value from a JSON string and store it in the scratch buffer.  It is able
3797  * to extract nested object values, and members of arrays by index.
3798  *
3799  * elemlist is a list of JSON keys, stored as packed NUL-terminated strings, to
3800  * be looked up as we descend into the object tree.  e.g.
3801  *
3802  *    foo[0].bar.baz[32] --> "foo" NUL "0" NUL "bar" NUL "baz" NUL "32" NUL
3803  *       with nelems = 5.
3804  *
3805  * The run time of this function must be bounded above by strsize to limit the
3806  * amount of work done in probe context.  As such, it is implemented as a
3807  * simple state machine, reading one character at a time using safe loads
3808  * until we find the requested element, hit a parsing error or run off the
3809  * end of the object or string.
3810  *
3811  * As there is no way for a subroutine to return an error without interrupting
3812  * clause execution, we simply return NULL in the event of a missing key or any
3813  * other error condition.  Each NULL return in this function is commented with
3814  * the error condition it represents -- parsing or otherwise.
3815  *
3816  * The set of states for the state machine closely matches the JSON
3817  * specification (http://json.org/).  Briefly:
3818  *
3819  *   DTRACE_JSON_REST:
3820  *     Skip whitespace until we find either a top-level Object, moving
3821  *     to DTRACE_JSON_OBJECT; or an Array, moving to DTRACE_JSON_VALUE.
3822  *
3823  *   DTRACE_JSON_OBJECT:
3824  *     Locate the next key String in an Object.  Sets a flag to denote
3825  *     the next String as a key string and moves to DTRACE_JSON_STRING.
3826  *
3827  *   DTRACE_JSON_COLON:
3828  *     Skip whitespace until we find the colon that separates key Strings
3829  *     from their values.  Once found, move to DTRACE_JSON_VALUE.
3830  *
3831  *   DTRACE_JSON_VALUE:
3832  *     Detects the type of the next value (String, Number, Identifier, Object
3833  *     or Array) and routes to the states that process that type.  Here we also
3834  *     deal with the element selector list if we are requested to traverse down
3835  *     into the object tree.
3836  *
3837  *   DTRACE_JSON_COMMA:
3838  *     Skip whitespace until we find the comma that separates key-value pairs
3839  *     in Objects (returning to DTRACE_JSON_OBJECT) or values in Arrays
3840  *     (similarly DTRACE_JSON_VALUE).  All following literal value processing
3841  *     states return to this state at the end of their value, unless otherwise
3842  *     noted.
3843  *
3844  *   DTRACE_JSON_NUMBER, DTRACE_JSON_NUMBER_FRAC, DTRACE_JSON_NUMBER_EXP:
3845  *     Processes a Number literal from the JSON, including any exponent
3846  *     component that may be present.  Numbers are returned as strings, which
3847  *     may be passed to strtoll() if an integer is required.
3848  *
3849  *   DTRACE_JSON_IDENTIFIER:
3850  *     Processes a "true", "false" or "null" literal in the JSON.
3851  *
3852  *   DTRACE_JSON_STRING, DTRACE_JSON_STRING_ESCAPE,
3853  *   DTRACE_JSON_STRING_ESCAPE_UNICODE:
3854  *     Processes a String literal from the JSON, whether the String denotes
3855  *     a key, a value or part of a larger Object.  Handles all escape sequences
3856  *     present in the specification, including four-digit unicode characters,
3857  *     but merely includes the escape sequence without converting it to the
3858  *     actual escaped character.  If the String is flagged as a key, we
3859  *     move to DTRACE_JSON_COLON rather than DTRACE_JSON_COMMA.
3860  *
3861  *   DTRACE_JSON_COLLECT_OBJECT:
3862  *     This state collects an entire Object (or Array), correctly handling
3863  *     embedded strings.  If the full element selector list matches this nested
3864  *     object, we return the Object in full as a string.  If not, we use this
3865  *     state to skip to the next value at this level and continue processing.
3866  *
3867  * NOTE: This function uses various macros from strtolctype.h to manipulate
3868  * digit values, etc -- these have all been checked to ensure they make
3869  * no additional function calls.
3870  */
3871 static char *
3872 dtrace_json(uint64_t size, uintptr_t json, char *elemlist, int nelems,
3873     char *dest)
3874 {
3875 	dtrace_json_state_t state = DTRACE_JSON_REST;
3876 	int64_t array_elem = INT64_MIN;
3877 	int64_t array_pos = 0;
3878 	uint8_t escape_unicount = 0;
3879 	boolean_t string_is_key = B_FALSE;
3880 	boolean_t collect_object = B_FALSE;
3881 	boolean_t found_key = B_FALSE;
3882 	boolean_t in_array = B_FALSE;
3883 	uint32_t braces = 0, brackets = 0;
3884 	char *elem = elemlist;
3885 	char *dd = dest;
3886 	uintptr_t cur;
3887 
3888 	for (cur = json; cur < json + size; cur++) {
3889 		char cc = dtrace_load8(cur);
3890 		if (cc == '\0')
3891 			return (NULL);
3892 
3893 		switch (state) {
3894 		case DTRACE_JSON_REST:
3895 			if (isspace(cc))
3896 				break;
3897 
3898 			if (cc == '{') {
3899 				state = DTRACE_JSON_OBJECT;
3900 				break;
3901 			}
3902 
3903 			if (cc == '[') {
3904 				in_array = B_TRUE;
3905 				array_pos = 0;
3906 				array_elem = dtrace_strtoll(elem, 10, size);
3907 				found_key = array_elem == 0 ? B_TRUE : B_FALSE;
3908 				state = DTRACE_JSON_VALUE;
3909 				break;
3910 			}
3911 
3912 			/*
3913 			 * ERROR: expected to find a top-level object or array.
3914 			 */
3915 			return (NULL);
3916 		case DTRACE_JSON_OBJECT:
3917 			if (isspace(cc))
3918 				break;
3919 
3920 			if (cc == '"') {
3921 				state = DTRACE_JSON_STRING;
3922 				string_is_key = B_TRUE;
3923 				break;
3924 			}
3925 
3926 			/*
3927 			 * ERROR: either the object did not start with a key
3928 			 * string, or we've run off the end of the object
3929 			 * without finding the requested key.
3930 			 */
3931 			return (NULL);
3932 		case DTRACE_JSON_STRING:
3933 			if (cc == '\\') {
3934 				*dd++ = '\\';
3935 				state = DTRACE_JSON_STRING_ESCAPE;
3936 				break;
3937 			}
3938 
3939 			if (cc == '"') {
3940 				if (collect_object) {
3941 					/*
3942 					 * We don't reset the dest here, as
3943 					 * the string is part of a larger
3944 					 * object being collected.
3945 					 */
3946 					*dd++ = cc;
3947 					collect_object = B_FALSE;
3948 					state = DTRACE_JSON_COLLECT_OBJECT;
3949 					break;
3950 				}
3951 				*dd = '\0';
3952 				dd = dest; /* reset string buffer */
3953 				if (string_is_key) {
3954 					if (dtrace_strncmp(dest, elem,
3955 					    size) == 0)
3956 						found_key = B_TRUE;
3957 				} else if (found_key) {
3958 					if (nelems > 1) {
3959 						/*
3960 						 * We expected an object, not
3961 						 * this string.
3962 						 */
3963 						return (NULL);
3964 					}
3965 					return (dest);
3966 				}
3967 				state = string_is_key ? DTRACE_JSON_COLON :
3968 				    DTRACE_JSON_COMMA;
3969 				string_is_key = B_FALSE;
3970 				break;
3971 			}
3972 
3973 			*dd++ = cc;
3974 			break;
3975 		case DTRACE_JSON_STRING_ESCAPE:
3976 			*dd++ = cc;
3977 			if (cc == 'u') {
3978 				escape_unicount = 0;
3979 				state = DTRACE_JSON_STRING_ESCAPE_UNICODE;
3980 			} else {
3981 				state = DTRACE_JSON_STRING;
3982 			}
3983 			break;
3984 		case DTRACE_JSON_STRING_ESCAPE_UNICODE:
3985 			if (!isxdigit(cc)) {
3986 				/*
3987 				 * ERROR: invalid unicode escape, expected
3988 				 * four valid hexidecimal digits.
3989 				 */
3990 				return (NULL);
3991 			}
3992 
3993 			*dd++ = cc;
3994 			if (++escape_unicount == 4)
3995 				state = DTRACE_JSON_STRING;
3996 			break;
3997 		case DTRACE_JSON_COLON:
3998 			if (isspace(cc))
3999 				break;
4000 
4001 			if (cc == ':') {
4002 				state = DTRACE_JSON_VALUE;
4003 				break;
4004 			}
4005 
4006 			/*
4007 			 * ERROR: expected a colon.
4008 			 */
4009 			return (NULL);
4010 		case DTRACE_JSON_COMMA:
4011 			if (isspace(cc))
4012 				break;
4013 
4014 			if (cc == ',') {
4015 				if (in_array) {
4016 					state = DTRACE_JSON_VALUE;
4017 					if (++array_pos == array_elem)
4018 						found_key = B_TRUE;
4019 				} else {
4020 					state = DTRACE_JSON_OBJECT;
4021 				}
4022 				break;
4023 			}
4024 
4025 			/*
4026 			 * ERROR: either we hit an unexpected character, or
4027 			 * we reached the end of the object or array without
4028 			 * finding the requested key.
4029 			 */
4030 			return (NULL);
4031 		case DTRACE_JSON_IDENTIFIER:
4032 			if (islower(cc)) {
4033 				*dd++ = cc;
4034 				break;
4035 			}
4036 
4037 			*dd = '\0';
4038 			dd = dest; /* reset string buffer */
4039 
4040 			if (dtrace_strncmp(dest, "true", 5) == 0 ||
4041 			    dtrace_strncmp(dest, "false", 6) == 0 ||
4042 			    dtrace_strncmp(dest, "null", 5) == 0) {
4043 				if (found_key) {
4044 					if (nelems > 1) {
4045 						/*
4046 						 * ERROR: We expected an object,
4047 						 * not this identifier.
4048 						 */
4049 						return (NULL);
4050 					}
4051 					return (dest);
4052 				} else {
4053 					cur--;
4054 					state = DTRACE_JSON_COMMA;
4055 					break;
4056 				}
4057 			}
4058 
4059 			/*
4060 			 * ERROR: we did not recognise the identifier as one
4061 			 * of those in the JSON specification.
4062 			 */
4063 			return (NULL);
4064 		case DTRACE_JSON_NUMBER:
4065 			if (cc == '.') {
4066 				*dd++ = cc;
4067 				state = DTRACE_JSON_NUMBER_FRAC;
4068 				break;
4069 			}
4070 
4071 			if (cc == 'x' || cc == 'X') {
4072 				/*
4073 				 * ERROR: specification explicitly excludes
4074 				 * hexidecimal or octal numbers.
4075 				 */
4076 				return (NULL);
4077 			}
4078 
4079 			/* FALLTHRU */
4080 		case DTRACE_JSON_NUMBER_FRAC:
4081 			if (cc == 'e' || cc == 'E') {
4082 				*dd++ = cc;
4083 				state = DTRACE_JSON_NUMBER_EXP;
4084 				break;
4085 			}
4086 
4087 			if (cc == '+' || cc == '-') {
4088 				/*
4089 				 * ERROR: expect sign as part of exponent only.
4090 				 */
4091 				return (NULL);
4092 			}
4093 			/* FALLTHRU */
4094 		case DTRACE_JSON_NUMBER_EXP:
4095 			if (isdigit(cc) || cc == '+' || cc == '-') {
4096 				*dd++ = cc;
4097 				break;
4098 			}
4099 
4100 			*dd = '\0';
4101 			dd = dest; /* reset string buffer */
4102 			if (found_key) {
4103 				if (nelems > 1) {
4104 					/*
4105 					 * ERROR: We expected an object, not
4106 					 * this number.
4107 					 */
4108 					return (NULL);
4109 				}
4110 				return (dest);
4111 			}
4112 
4113 			cur--;
4114 			state = DTRACE_JSON_COMMA;
4115 			break;
4116 		case DTRACE_JSON_VALUE:
4117 			if (isspace(cc))
4118 				break;
4119 
4120 			if (cc == '{' || cc == '[') {
4121 				if (nelems > 1 && found_key) {
4122 					in_array = cc == '[' ? B_TRUE : B_FALSE;
4123 					/*
4124 					 * If our element selector directs us
4125 					 * to descend into this nested object,
4126 					 * then move to the next selector
4127 					 * element in the list and restart the
4128 					 * state machine.
4129 					 */
4130 					while (*elem != '\0')
4131 						elem++;
4132 					elem++; /* skip the inter-element NUL */
4133 					nelems--;
4134 					dd = dest;
4135 					if (in_array) {
4136 						state = DTRACE_JSON_VALUE;
4137 						array_pos = 0;
4138 						array_elem = dtrace_strtoll(
4139 						    elem, 10, size);
4140 						found_key = array_elem == 0 ?
4141 						    B_TRUE : B_FALSE;
4142 					} else {
4143 						found_key = B_FALSE;
4144 						state = DTRACE_JSON_OBJECT;
4145 					}
4146 					break;
4147 				}
4148 
4149 				/*
4150 				 * Otherwise, we wish to either skip this
4151 				 * nested object or return it in full.
4152 				 */
4153 				if (cc == '[')
4154 					brackets = 1;
4155 				else
4156 					braces = 1;
4157 				*dd++ = cc;
4158 				state = DTRACE_JSON_COLLECT_OBJECT;
4159 				break;
4160 			}
4161 
4162 			if (cc == '"') {
4163 				state = DTRACE_JSON_STRING;
4164 				break;
4165 			}
4166 
4167 			if (islower(cc)) {
4168 				/*
4169 				 * Here we deal with true, false and null.
4170 				 */
4171 				*dd++ = cc;
4172 				state = DTRACE_JSON_IDENTIFIER;
4173 				break;
4174 			}
4175 
4176 			if (cc == '-' || isdigit(cc)) {
4177 				*dd++ = cc;
4178 				state = DTRACE_JSON_NUMBER;
4179 				break;
4180 			}
4181 
4182 			/*
4183 			 * ERROR: unexpected character at start of value.
4184 			 */
4185 			return (NULL);
4186 		case DTRACE_JSON_COLLECT_OBJECT:
4187 			if (cc == '\0')
4188 				/*
4189 				 * ERROR: unexpected end of input.
4190 				 */
4191 				return (NULL);
4192 
4193 			*dd++ = cc;
4194 			if (cc == '"') {
4195 				collect_object = B_TRUE;
4196 				state = DTRACE_JSON_STRING;
4197 				break;
4198 			}
4199 
4200 			if (cc == ']') {
4201 				if (brackets-- == 0) {
4202 					/*
4203 					 * ERROR: unbalanced brackets.
4204 					 */
4205 					return (NULL);
4206 				}
4207 			} else if (cc == '}') {
4208 				if (braces-- == 0) {
4209 					/*
4210 					 * ERROR: unbalanced braces.
4211 					 */
4212 					return (NULL);
4213 				}
4214 			} else if (cc == '{') {
4215 				braces++;
4216 			} else if (cc == '[') {
4217 				brackets++;
4218 			}
4219 
4220 			if (brackets == 0 && braces == 0) {
4221 				if (found_key) {
4222 					*dd = '\0';
4223 					return (dest);
4224 				}
4225 				dd = dest; /* reset string buffer */
4226 				state = DTRACE_JSON_COMMA;
4227 			}
4228 			break;
4229 		}
4230 	}
4231 	return (NULL);
4232 }
4233 
4234 /*
4235  * Emulate the execution of DTrace ID subroutines invoked by the call opcode.
4236  * Notice that we don't bother validating the proper number of arguments or
4237  * their types in the tuple stack.  This isn't needed because all argument
4238  * interpretation is safe because of our load safety -- the worst that can
4239  * happen is that a bogus program can obtain bogus results.
4240  */
4241 static void
4242 dtrace_dif_subr(uint_t subr, uint_t rd, uint64_t *regs,
4243     dtrace_key_t *tupregs, int nargs,
4244     dtrace_mstate_t *mstate, dtrace_state_t *state)
4245 {
4246 	volatile uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
4247 	volatile uintptr_t *illval = &cpu_core[curcpu].cpuc_dtrace_illval;
4248 	dtrace_vstate_t *vstate = &state->dts_vstate;
4249 
4250 #ifdef illumos
4251 	union {
4252 		mutex_impl_t mi;
4253 		uint64_t mx;
4254 	} m;
4255 
4256 	union {
4257 		krwlock_t ri;
4258 		uintptr_t rw;
4259 	} r;
4260 #else
4261 	struct thread *lowner;
4262 	union {
4263 		struct lock_object *li;
4264 		uintptr_t lx;
4265 	} l;
4266 #endif
4267 
4268 	switch (subr) {
4269 	case DIF_SUBR_RAND:
4270 		regs[rd] = dtrace_xoroshiro128_plus_next(
4271 		    state->dts_rstate[curcpu]);
4272 		break;
4273 
4274 #ifdef illumos
4275 	case DIF_SUBR_MUTEX_OWNED:
4276 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
4277 		    mstate, vstate)) {
4278 			regs[rd] = 0;
4279 			break;
4280 		}
4281 
4282 		m.mx = dtrace_load64(tupregs[0].dttk_value);
4283 		if (MUTEX_TYPE_ADAPTIVE(&m.mi))
4284 			regs[rd] = MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER;
4285 		else
4286 			regs[rd] = LOCK_HELD(&m.mi.m_spin.m_spinlock);
4287 		break;
4288 
4289 	case DIF_SUBR_MUTEX_OWNER:
4290 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
4291 		    mstate, vstate)) {
4292 			regs[rd] = 0;
4293 			break;
4294 		}
4295 
4296 		m.mx = dtrace_load64(tupregs[0].dttk_value);
4297 		if (MUTEX_TYPE_ADAPTIVE(&m.mi) &&
4298 		    MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER)
4299 			regs[rd] = (uintptr_t)MUTEX_OWNER(&m.mi);
4300 		else
4301 			regs[rd] = 0;
4302 		break;
4303 
4304 	case DIF_SUBR_MUTEX_TYPE_ADAPTIVE:
4305 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
4306 		    mstate, vstate)) {
4307 			regs[rd] = 0;
4308 			break;
4309 		}
4310 
4311 		m.mx = dtrace_load64(tupregs[0].dttk_value);
4312 		regs[rd] = MUTEX_TYPE_ADAPTIVE(&m.mi);
4313 		break;
4314 
4315 	case DIF_SUBR_MUTEX_TYPE_SPIN:
4316 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
4317 		    mstate, vstate)) {
4318 			regs[rd] = 0;
4319 			break;
4320 		}
4321 
4322 		m.mx = dtrace_load64(tupregs[0].dttk_value);
4323 		regs[rd] = MUTEX_TYPE_SPIN(&m.mi);
4324 		break;
4325 
4326 	case DIF_SUBR_RW_READ_HELD: {
4327 		uintptr_t tmp;
4328 
4329 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
4330 		    mstate, vstate)) {
4331 			regs[rd] = 0;
4332 			break;
4333 		}
4334 
4335 		r.rw = dtrace_loadptr(tupregs[0].dttk_value);
4336 		regs[rd] = _RW_READ_HELD(&r.ri, tmp);
4337 		break;
4338 	}
4339 
4340 	case DIF_SUBR_RW_WRITE_HELD:
4341 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
4342 		    mstate, vstate)) {
4343 			regs[rd] = 0;
4344 			break;
4345 		}
4346 
4347 		r.rw = dtrace_loadptr(tupregs[0].dttk_value);
4348 		regs[rd] = _RW_WRITE_HELD(&r.ri);
4349 		break;
4350 
4351 	case DIF_SUBR_RW_ISWRITER:
4352 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
4353 		    mstate, vstate)) {
4354 			regs[rd] = 0;
4355 			break;
4356 		}
4357 
4358 		r.rw = dtrace_loadptr(tupregs[0].dttk_value);
4359 		regs[rd] = _RW_ISWRITER(&r.ri);
4360 		break;
4361 
4362 #else /* !illumos */
4363 	case DIF_SUBR_MUTEX_OWNED:
4364 		if (!dtrace_canload(tupregs[0].dttk_value,
4365 			sizeof (struct lock_object), mstate, vstate)) {
4366 			regs[rd] = 0;
4367 			break;
4368 		}
4369 		l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
4370 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4371 		regs[rd] = LOCK_CLASS(l.li)->lc_owner(l.li, &lowner);
4372 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4373 		break;
4374 
4375 	case DIF_SUBR_MUTEX_OWNER:
4376 		if (!dtrace_canload(tupregs[0].dttk_value,
4377 			sizeof (struct lock_object), mstate, vstate)) {
4378 			regs[rd] = 0;
4379 			break;
4380 		}
4381 		l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
4382 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4383 		LOCK_CLASS(l.li)->lc_owner(l.li, &lowner);
4384 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4385 		regs[rd] = (uintptr_t)lowner;
4386 		break;
4387 
4388 	case DIF_SUBR_MUTEX_TYPE_ADAPTIVE:
4389 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (struct mtx),
4390 		    mstate, vstate)) {
4391 			regs[rd] = 0;
4392 			break;
4393 		}
4394 		l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
4395 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4396 		regs[rd] = (LOCK_CLASS(l.li)->lc_flags & LC_SLEEPLOCK) != 0;
4397 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4398 		break;
4399 
4400 	case DIF_SUBR_MUTEX_TYPE_SPIN:
4401 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (struct mtx),
4402 		    mstate, vstate)) {
4403 			regs[rd] = 0;
4404 			break;
4405 		}
4406 		l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
4407 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4408 		regs[rd] = (LOCK_CLASS(l.li)->lc_flags & LC_SPINLOCK) != 0;
4409 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4410 		break;
4411 
4412 	case DIF_SUBR_RW_READ_HELD:
4413 	case DIF_SUBR_SX_SHARED_HELD:
4414 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
4415 		    mstate, vstate)) {
4416 			regs[rd] = 0;
4417 			break;
4418 		}
4419 		l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
4420 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4421 		regs[rd] = LOCK_CLASS(l.li)->lc_owner(l.li, &lowner) &&
4422 		    lowner == NULL;
4423 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4424 		break;
4425 
4426 	case DIF_SUBR_RW_WRITE_HELD:
4427 	case DIF_SUBR_SX_EXCLUSIVE_HELD:
4428 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
4429 		    mstate, vstate)) {
4430 			regs[rd] = 0;
4431 			break;
4432 		}
4433 		l.lx = dtrace_loadptr(tupregs[0].dttk_value);
4434 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4435 		regs[rd] = LOCK_CLASS(l.li)->lc_owner(l.li, &lowner) &&
4436 		    lowner != NULL;
4437 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4438 		break;
4439 
4440 	case DIF_SUBR_RW_ISWRITER:
4441 	case DIF_SUBR_SX_ISEXCLUSIVE:
4442 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
4443 		    mstate, vstate)) {
4444 			regs[rd] = 0;
4445 			break;
4446 		}
4447 		l.lx = dtrace_loadptr(tupregs[0].dttk_value);
4448 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4449 		LOCK_CLASS(l.li)->lc_owner(l.li, &lowner);
4450 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4451 		regs[rd] = (lowner == curthread);
4452 		break;
4453 #endif /* illumos */
4454 
4455 	case DIF_SUBR_BCOPY: {
4456 		/*
4457 		 * We need to be sure that the destination is in the scratch
4458 		 * region -- no other region is allowed.
4459 		 */
4460 		uintptr_t src = tupregs[0].dttk_value;
4461 		uintptr_t dest = tupregs[1].dttk_value;
4462 		size_t size = tupregs[2].dttk_value;
4463 
4464 		if (!dtrace_inscratch(dest, size, mstate)) {
4465 			*flags |= CPU_DTRACE_BADADDR;
4466 			*illval = regs[rd];
4467 			break;
4468 		}
4469 
4470 		if (!dtrace_canload(src, size, mstate, vstate)) {
4471 			regs[rd] = 0;
4472 			break;
4473 		}
4474 
4475 		dtrace_bcopy((void *)src, (void *)dest, size);
4476 		break;
4477 	}
4478 
4479 	case DIF_SUBR_ALLOCA:
4480 	case DIF_SUBR_COPYIN: {
4481 		uintptr_t dest = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
4482 		uint64_t size =
4483 		    tupregs[subr == DIF_SUBR_ALLOCA ? 0 : 1].dttk_value;
4484 		size_t scratch_size = (dest - mstate->dtms_scratch_ptr) + size;
4485 
4486 		/*
4487 		 * This action doesn't require any credential checks since
4488 		 * probes will not activate in user contexts to which the
4489 		 * enabling user does not have permissions.
4490 		 */
4491 
4492 		/*
4493 		 * Rounding up the user allocation size could have overflowed
4494 		 * a large, bogus allocation (like -1ULL) to 0.
4495 		 */
4496 		if (scratch_size < size ||
4497 		    !DTRACE_INSCRATCH(mstate, scratch_size)) {
4498 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4499 			regs[rd] = 0;
4500 			break;
4501 		}
4502 
4503 		if (subr == DIF_SUBR_COPYIN) {
4504 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4505 			dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
4506 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4507 		}
4508 
4509 		mstate->dtms_scratch_ptr += scratch_size;
4510 		regs[rd] = dest;
4511 		break;
4512 	}
4513 
4514 	case DIF_SUBR_COPYINTO: {
4515 		uint64_t size = tupregs[1].dttk_value;
4516 		uintptr_t dest = tupregs[2].dttk_value;
4517 
4518 		/*
4519 		 * This action doesn't require any credential checks since
4520 		 * probes will not activate in user contexts to which the
4521 		 * enabling user does not have permissions.
4522 		 */
4523 		if (!dtrace_inscratch(dest, size, mstate)) {
4524 			*flags |= CPU_DTRACE_BADADDR;
4525 			*illval = regs[rd];
4526 			break;
4527 		}
4528 
4529 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4530 		dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
4531 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4532 		break;
4533 	}
4534 
4535 	case DIF_SUBR_COPYINSTR: {
4536 		uintptr_t dest = mstate->dtms_scratch_ptr;
4537 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4538 
4539 		if (nargs > 1 && tupregs[1].dttk_value < size)
4540 			size = tupregs[1].dttk_value + 1;
4541 
4542 		/*
4543 		 * This action doesn't require any credential checks since
4544 		 * probes will not activate in user contexts to which the
4545 		 * enabling user does not have permissions.
4546 		 */
4547 		if (!DTRACE_INSCRATCH(mstate, size)) {
4548 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4549 			regs[rd] = 0;
4550 			break;
4551 		}
4552 
4553 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4554 		dtrace_copyinstr(tupregs[0].dttk_value, dest, size, flags);
4555 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4556 
4557 		((char *)dest)[size - 1] = '\0';
4558 		mstate->dtms_scratch_ptr += size;
4559 		regs[rd] = dest;
4560 		break;
4561 	}
4562 
4563 #ifdef illumos
4564 	case DIF_SUBR_MSGSIZE:
4565 	case DIF_SUBR_MSGDSIZE: {
4566 		uintptr_t baddr = tupregs[0].dttk_value, daddr;
4567 		uintptr_t wptr, rptr;
4568 		size_t count = 0;
4569 		int cont = 0;
4570 
4571 		while (baddr != 0 && !(*flags & CPU_DTRACE_FAULT)) {
4572 
4573 			if (!dtrace_canload(baddr, sizeof (mblk_t), mstate,
4574 			    vstate)) {
4575 				regs[rd] = 0;
4576 				break;
4577 			}
4578 
4579 			wptr = dtrace_loadptr(baddr +
4580 			    offsetof(mblk_t, b_wptr));
4581 
4582 			rptr = dtrace_loadptr(baddr +
4583 			    offsetof(mblk_t, b_rptr));
4584 
4585 			if (wptr < rptr) {
4586 				*flags |= CPU_DTRACE_BADADDR;
4587 				*illval = tupregs[0].dttk_value;
4588 				break;
4589 			}
4590 
4591 			daddr = dtrace_loadptr(baddr +
4592 			    offsetof(mblk_t, b_datap));
4593 
4594 			baddr = dtrace_loadptr(baddr +
4595 			    offsetof(mblk_t, b_cont));
4596 
4597 			/*
4598 			 * We want to prevent against denial-of-service here,
4599 			 * so we're only going to search the list for
4600 			 * dtrace_msgdsize_max mblks.
4601 			 */
4602 			if (cont++ > dtrace_msgdsize_max) {
4603 				*flags |= CPU_DTRACE_ILLOP;
4604 				break;
4605 			}
4606 
4607 			if (subr == DIF_SUBR_MSGDSIZE) {
4608 				if (dtrace_load8(daddr +
4609 				    offsetof(dblk_t, db_type)) != M_DATA)
4610 					continue;
4611 			}
4612 
4613 			count += wptr - rptr;
4614 		}
4615 
4616 		if (!(*flags & CPU_DTRACE_FAULT))
4617 			regs[rd] = count;
4618 
4619 		break;
4620 	}
4621 #endif
4622 
4623 	case DIF_SUBR_PROGENYOF: {
4624 		pid_t pid = tupregs[0].dttk_value;
4625 		proc_t *p;
4626 		int rval = 0;
4627 
4628 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4629 
4630 		for (p = curthread->t_procp; p != NULL; p = p->p_parent) {
4631 #ifdef illumos
4632 			if (p->p_pidp->pid_id == pid) {
4633 #else
4634 			if (p->p_pid == pid) {
4635 #endif
4636 				rval = 1;
4637 				break;
4638 			}
4639 		}
4640 
4641 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4642 
4643 		regs[rd] = rval;
4644 		break;
4645 	}
4646 
4647 	case DIF_SUBR_SPECULATION:
4648 		regs[rd] = dtrace_speculation(state);
4649 		break;
4650 
4651 	case DIF_SUBR_COPYOUT: {
4652 		uintptr_t kaddr = tupregs[0].dttk_value;
4653 		uintptr_t uaddr = tupregs[1].dttk_value;
4654 		uint64_t size = tupregs[2].dttk_value;
4655 
4656 		if (!dtrace_destructive_disallow &&
4657 		    dtrace_priv_proc_control(state) &&
4658 		    !dtrace_istoxic(kaddr, size) &&
4659 		    dtrace_canload(kaddr, size, mstate, vstate)) {
4660 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4661 			dtrace_copyout(kaddr, uaddr, size, flags);
4662 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4663 		}
4664 		break;
4665 	}
4666 
4667 	case DIF_SUBR_COPYOUTSTR: {
4668 		uintptr_t kaddr = tupregs[0].dttk_value;
4669 		uintptr_t uaddr = tupregs[1].dttk_value;
4670 		uint64_t size = tupregs[2].dttk_value;
4671 		size_t lim;
4672 
4673 		if (!dtrace_destructive_disallow &&
4674 		    dtrace_priv_proc_control(state) &&
4675 		    !dtrace_istoxic(kaddr, size) &&
4676 		    dtrace_strcanload(kaddr, size, &lim, mstate, vstate)) {
4677 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4678 			dtrace_copyoutstr(kaddr, uaddr, lim, flags);
4679 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4680 		}
4681 		break;
4682 	}
4683 
4684 	case DIF_SUBR_STRLEN: {
4685 		size_t size = state->dts_options[DTRACEOPT_STRSIZE];
4686 		uintptr_t addr = (uintptr_t)tupregs[0].dttk_value;
4687 		size_t lim;
4688 
4689 		if (!dtrace_strcanload(addr, size, &lim, mstate, vstate)) {
4690 			regs[rd] = 0;
4691 			break;
4692 		}
4693 
4694 		regs[rd] = dtrace_strlen((char *)addr, lim);
4695 		break;
4696 	}
4697 
4698 	case DIF_SUBR_STRCHR:
4699 	case DIF_SUBR_STRRCHR: {
4700 		/*
4701 		 * We're going to iterate over the string looking for the
4702 		 * specified character.  We will iterate until we have reached
4703 		 * the string length or we have found the character.  If this
4704 		 * is DIF_SUBR_STRRCHR, we will look for the last occurrence
4705 		 * of the specified character instead of the first.
4706 		 */
4707 		uintptr_t addr = tupregs[0].dttk_value;
4708 		uintptr_t addr_limit;
4709 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4710 		size_t lim;
4711 		char c, target = (char)tupregs[1].dttk_value;
4712 
4713 		if (!dtrace_strcanload(addr, size, &lim, mstate, vstate)) {
4714 			regs[rd] = 0;
4715 			break;
4716 		}
4717 		addr_limit = addr + lim;
4718 
4719 		for (regs[rd] = 0; addr < addr_limit; addr++) {
4720 			if ((c = dtrace_load8(addr)) == target) {
4721 				regs[rd] = addr;
4722 
4723 				if (subr == DIF_SUBR_STRCHR)
4724 					break;
4725 			}
4726 
4727 			if (c == '\0')
4728 				break;
4729 		}
4730 		break;
4731 	}
4732 
4733 	case DIF_SUBR_STRSTR:
4734 	case DIF_SUBR_INDEX:
4735 	case DIF_SUBR_RINDEX: {
4736 		/*
4737 		 * We're going to iterate over the string looking for the
4738 		 * specified string.  We will iterate until we have reached
4739 		 * the string length or we have found the string.  (Yes, this
4740 		 * is done in the most naive way possible -- but considering
4741 		 * that the string we're searching for is likely to be
4742 		 * relatively short, the complexity of Rabin-Karp or similar
4743 		 * hardly seems merited.)
4744 		 */
4745 		char *addr = (char *)(uintptr_t)tupregs[0].dttk_value;
4746 		char *substr = (char *)(uintptr_t)tupregs[1].dttk_value;
4747 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4748 		size_t len = dtrace_strlen(addr, size);
4749 		size_t sublen = dtrace_strlen(substr, size);
4750 		char *limit = addr + len, *orig = addr;
4751 		int notfound = subr == DIF_SUBR_STRSTR ? 0 : -1;
4752 		int inc = 1;
4753 
4754 		regs[rd] = notfound;
4755 
4756 		if (!dtrace_canload((uintptr_t)addr, len + 1, mstate, vstate)) {
4757 			regs[rd] = 0;
4758 			break;
4759 		}
4760 
4761 		if (!dtrace_canload((uintptr_t)substr, sublen + 1, mstate,
4762 		    vstate)) {
4763 			regs[rd] = 0;
4764 			break;
4765 		}
4766 
4767 		/*
4768 		 * strstr() and index()/rindex() have similar semantics if
4769 		 * both strings are the empty string: strstr() returns a
4770 		 * pointer to the (empty) string, and index() and rindex()
4771 		 * both return index 0 (regardless of any position argument).
4772 		 */
4773 		if (sublen == 0 && len == 0) {
4774 			if (subr == DIF_SUBR_STRSTR)
4775 				regs[rd] = (uintptr_t)addr;
4776 			else
4777 				regs[rd] = 0;
4778 			break;
4779 		}
4780 
4781 		if (subr != DIF_SUBR_STRSTR) {
4782 			if (subr == DIF_SUBR_RINDEX) {
4783 				limit = orig - 1;
4784 				addr += len;
4785 				inc = -1;
4786 			}
4787 
4788 			/*
4789 			 * Both index() and rindex() take an optional position
4790 			 * argument that denotes the starting position.
4791 			 */
4792 			if (nargs == 3) {
4793 				int64_t pos = (int64_t)tupregs[2].dttk_value;
4794 
4795 				/*
4796 				 * If the position argument to index() is
4797 				 * negative, Perl implicitly clamps it at
4798 				 * zero.  This semantic is a little surprising
4799 				 * given the special meaning of negative
4800 				 * positions to similar Perl functions like
4801 				 * substr(), but it appears to reflect a
4802 				 * notion that index() can start from a
4803 				 * negative index and increment its way up to
4804 				 * the string.  Given this notion, Perl's
4805 				 * rindex() is at least self-consistent in
4806 				 * that it implicitly clamps positions greater
4807 				 * than the string length to be the string
4808 				 * length.  Where Perl completely loses
4809 				 * coherence, however, is when the specified
4810 				 * substring is the empty string ("").  In
4811 				 * this case, even if the position is
4812 				 * negative, rindex() returns 0 -- and even if
4813 				 * the position is greater than the length,
4814 				 * index() returns the string length.  These
4815 				 * semantics violate the notion that index()
4816 				 * should never return a value less than the
4817 				 * specified position and that rindex() should
4818 				 * never return a value greater than the
4819 				 * specified position.  (One assumes that
4820 				 * these semantics are artifacts of Perl's
4821 				 * implementation and not the results of
4822 				 * deliberate design -- it beggars belief that
4823 				 * even Larry Wall could desire such oddness.)
4824 				 * While in the abstract one would wish for
4825 				 * consistent position semantics across
4826 				 * substr(), index() and rindex() -- or at the
4827 				 * very least self-consistent position
4828 				 * semantics for index() and rindex() -- we
4829 				 * instead opt to keep with the extant Perl
4830 				 * semantics, in all their broken glory.  (Do
4831 				 * we have more desire to maintain Perl's
4832 				 * semantics than Perl does?  Probably.)
4833 				 */
4834 				if (subr == DIF_SUBR_RINDEX) {
4835 					if (pos < 0) {
4836 						if (sublen == 0)
4837 							regs[rd] = 0;
4838 						break;
4839 					}
4840 
4841 					if (pos > len)
4842 						pos = len;
4843 				} else {
4844 					if (pos < 0)
4845 						pos = 0;
4846 
4847 					if (pos >= len) {
4848 						if (sublen == 0)
4849 							regs[rd] = len;
4850 						break;
4851 					}
4852 				}
4853 
4854 				addr = orig + pos;
4855 			}
4856 		}
4857 
4858 		for (regs[rd] = notfound; addr != limit; addr += inc) {
4859 			if (dtrace_strncmp(addr, substr, sublen) == 0) {
4860 				if (subr != DIF_SUBR_STRSTR) {
4861 					/*
4862 					 * As D index() and rindex() are
4863 					 * modeled on Perl (and not on awk),
4864 					 * we return a zero-based (and not a
4865 					 * one-based) index.  (For you Perl
4866 					 * weenies: no, we're not going to add
4867 					 * $[ -- and shouldn't you be at a con
4868 					 * or something?)
4869 					 */
4870 					regs[rd] = (uintptr_t)(addr - orig);
4871 					break;
4872 				}
4873 
4874 				ASSERT(subr == DIF_SUBR_STRSTR);
4875 				regs[rd] = (uintptr_t)addr;
4876 				break;
4877 			}
4878 		}
4879 
4880 		break;
4881 	}
4882 
4883 	case DIF_SUBR_STRTOK: {
4884 		uintptr_t addr = tupregs[0].dttk_value;
4885 		uintptr_t tokaddr = tupregs[1].dttk_value;
4886 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4887 		uintptr_t limit, toklimit;
4888 		size_t clim;
4889 		uint8_t c = 0, tokmap[32];	 /* 256 / 8 */
4890 		char *dest = (char *)mstate->dtms_scratch_ptr;
4891 		int i;
4892 
4893 		/*
4894 		 * Check both the token buffer and (later) the input buffer,
4895 		 * since both could be non-scratch addresses.
4896 		 */
4897 		if (!dtrace_strcanload(tokaddr, size, &clim, mstate, vstate)) {
4898 			regs[rd] = 0;
4899 			break;
4900 		}
4901 		toklimit = tokaddr + clim;
4902 
4903 		if (!DTRACE_INSCRATCH(mstate, size)) {
4904 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4905 			regs[rd] = 0;
4906 			break;
4907 		}
4908 
4909 		if (addr == 0) {
4910 			/*
4911 			 * If the address specified is NULL, we use our saved
4912 			 * strtok pointer from the mstate.  Note that this
4913 			 * means that the saved strtok pointer is _only_
4914 			 * valid within multiple enablings of the same probe --
4915 			 * it behaves like an implicit clause-local variable.
4916 			 */
4917 			addr = mstate->dtms_strtok;
4918 			limit = mstate->dtms_strtok_limit;
4919 		} else {
4920 			/*
4921 			 * If the user-specified address is non-NULL we must
4922 			 * access check it.  This is the only time we have
4923 			 * a chance to do so, since this address may reside
4924 			 * in the string table of this clause-- future calls
4925 			 * (when we fetch addr from mstate->dtms_strtok)
4926 			 * would fail this access check.
4927 			 */
4928 			if (!dtrace_strcanload(addr, size, &clim, mstate,
4929 			    vstate)) {
4930 				regs[rd] = 0;
4931 				break;
4932 			}
4933 			limit = addr + clim;
4934 		}
4935 
4936 		/*
4937 		 * First, zero the token map, and then process the token
4938 		 * string -- setting a bit in the map for every character
4939 		 * found in the token string.
4940 		 */
4941 		for (i = 0; i < sizeof (tokmap); i++)
4942 			tokmap[i] = 0;
4943 
4944 		for (; tokaddr < toklimit; tokaddr++) {
4945 			if ((c = dtrace_load8(tokaddr)) == '\0')
4946 				break;
4947 
4948 			ASSERT((c >> 3) < sizeof (tokmap));
4949 			tokmap[c >> 3] |= (1 << (c & 0x7));
4950 		}
4951 
4952 		for (; addr < limit; addr++) {
4953 			/*
4954 			 * We're looking for a character that is _not_
4955 			 * contained in the token string.
4956 			 */
4957 			if ((c = dtrace_load8(addr)) == '\0')
4958 				break;
4959 
4960 			if (!(tokmap[c >> 3] & (1 << (c & 0x7))))
4961 				break;
4962 		}
4963 
4964 		if (c == '\0') {
4965 			/*
4966 			 * We reached the end of the string without finding
4967 			 * any character that was not in the token string.
4968 			 * We return NULL in this case, and we set the saved
4969 			 * address to NULL as well.
4970 			 */
4971 			regs[rd] = 0;
4972 			mstate->dtms_strtok = 0;
4973 			mstate->dtms_strtok_limit = 0;
4974 			break;
4975 		}
4976 
4977 		/*
4978 		 * From here on, we're copying into the destination string.
4979 		 */
4980 		for (i = 0; addr < limit && i < size - 1; addr++) {
4981 			if ((c = dtrace_load8(addr)) == '\0')
4982 				break;
4983 
4984 			if (tokmap[c >> 3] & (1 << (c & 0x7)))
4985 				break;
4986 
4987 			ASSERT(i < size);
4988 			dest[i++] = c;
4989 		}
4990 
4991 		ASSERT(i < size);
4992 		dest[i] = '\0';
4993 		regs[rd] = (uintptr_t)dest;
4994 		mstate->dtms_scratch_ptr += size;
4995 		mstate->dtms_strtok = addr;
4996 		mstate->dtms_strtok_limit = limit;
4997 		break;
4998 	}
4999 
5000 	case DIF_SUBR_SUBSTR: {
5001 		uintptr_t s = tupregs[0].dttk_value;
5002 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
5003 		char *d = (char *)mstate->dtms_scratch_ptr;
5004 		int64_t index = (int64_t)tupregs[1].dttk_value;
5005 		int64_t remaining = (int64_t)tupregs[2].dttk_value;
5006 		size_t len = dtrace_strlen((char *)s, size);
5007 		int64_t i;
5008 
5009 		if (!dtrace_canload(s, len + 1, mstate, vstate)) {
5010 			regs[rd] = 0;
5011 			break;
5012 		}
5013 
5014 		if (!DTRACE_INSCRATCH(mstate, size)) {
5015 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5016 			regs[rd] = 0;
5017 			break;
5018 		}
5019 
5020 		if (nargs <= 2)
5021 			remaining = (int64_t)size;
5022 
5023 		if (index < 0) {
5024 			index += len;
5025 
5026 			if (index < 0 && index + remaining > 0) {
5027 				remaining += index;
5028 				index = 0;
5029 			}
5030 		}
5031 
5032 		if (index >= len || index < 0) {
5033 			remaining = 0;
5034 		} else if (remaining < 0) {
5035 			remaining += len - index;
5036 		} else if (index + remaining > size) {
5037 			remaining = size - index;
5038 		}
5039 
5040 		for (i = 0; i < remaining; i++) {
5041 			if ((d[i] = dtrace_load8(s + index + i)) == '\0')
5042 				break;
5043 		}
5044 
5045 		d[i] = '\0';
5046 
5047 		mstate->dtms_scratch_ptr += size;
5048 		regs[rd] = (uintptr_t)d;
5049 		break;
5050 	}
5051 
5052 	case DIF_SUBR_JSON: {
5053 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
5054 		uintptr_t json = tupregs[0].dttk_value;
5055 		size_t jsonlen = dtrace_strlen((char *)json, size);
5056 		uintptr_t elem = tupregs[1].dttk_value;
5057 		size_t elemlen = dtrace_strlen((char *)elem, size);
5058 
5059 		char *dest = (char *)mstate->dtms_scratch_ptr;
5060 		char *elemlist = (char *)mstate->dtms_scratch_ptr + jsonlen + 1;
5061 		char *ee = elemlist;
5062 		int nelems = 1;
5063 		uintptr_t cur;
5064 
5065 		if (!dtrace_canload(json, jsonlen + 1, mstate, vstate) ||
5066 		    !dtrace_canload(elem, elemlen + 1, mstate, vstate)) {
5067 			regs[rd] = 0;
5068 			break;
5069 		}
5070 
5071 		if (!DTRACE_INSCRATCH(mstate, jsonlen + 1 + elemlen + 1)) {
5072 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5073 			regs[rd] = 0;
5074 			break;
5075 		}
5076 
5077 		/*
5078 		 * Read the element selector and split it up into a packed list
5079 		 * of strings.
5080 		 */
5081 		for (cur = elem; cur < elem + elemlen; cur++) {
5082 			char cc = dtrace_load8(cur);
5083 
5084 			if (cur == elem && cc == '[') {
5085 				/*
5086 				 * If the first element selector key is
5087 				 * actually an array index then ignore the
5088 				 * bracket.
5089 				 */
5090 				continue;
5091 			}
5092 
5093 			if (cc == ']')
5094 				continue;
5095 
5096 			if (cc == '.' || cc == '[') {
5097 				nelems++;
5098 				cc = '\0';
5099 			}
5100 
5101 			*ee++ = cc;
5102 		}
5103 		*ee++ = '\0';
5104 
5105 		if ((regs[rd] = (uintptr_t)dtrace_json(size, json, elemlist,
5106 		    nelems, dest)) != 0)
5107 			mstate->dtms_scratch_ptr += jsonlen + 1;
5108 		break;
5109 	}
5110 
5111 	case DIF_SUBR_TOUPPER:
5112 	case DIF_SUBR_TOLOWER: {
5113 		uintptr_t s = tupregs[0].dttk_value;
5114 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
5115 		char *dest = (char *)mstate->dtms_scratch_ptr, c;
5116 		size_t len = dtrace_strlen((char *)s, size);
5117 		char lower, upper, convert;
5118 		int64_t i;
5119 
5120 		if (subr == DIF_SUBR_TOUPPER) {
5121 			lower = 'a';
5122 			upper = 'z';
5123 			convert = 'A';
5124 		} else {
5125 			lower = 'A';
5126 			upper = 'Z';
5127 			convert = 'a';
5128 		}
5129 
5130 		if (!dtrace_canload(s, len + 1, mstate, vstate)) {
5131 			regs[rd] = 0;
5132 			break;
5133 		}
5134 
5135 		if (!DTRACE_INSCRATCH(mstate, size)) {
5136 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5137 			regs[rd] = 0;
5138 			break;
5139 		}
5140 
5141 		for (i = 0; i < size - 1; i++) {
5142 			if ((c = dtrace_load8(s + i)) == '\0')
5143 				break;
5144 
5145 			if (c >= lower && c <= upper)
5146 				c = convert + (c - lower);
5147 
5148 			dest[i] = c;
5149 		}
5150 
5151 		ASSERT(i < size);
5152 		dest[i] = '\0';
5153 		regs[rd] = (uintptr_t)dest;
5154 		mstate->dtms_scratch_ptr += size;
5155 		break;
5156 	}
5157 
5158 #ifdef illumos
5159 	case DIF_SUBR_GETMAJOR:
5160 #ifdef _LP64
5161 		regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR64) & MAXMAJ64;
5162 #else
5163 		regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR) & MAXMAJ;
5164 #endif
5165 		break;
5166 
5167 	case DIF_SUBR_GETMINOR:
5168 #ifdef _LP64
5169 		regs[rd] = tupregs[0].dttk_value & MAXMIN64;
5170 #else
5171 		regs[rd] = tupregs[0].dttk_value & MAXMIN;
5172 #endif
5173 		break;
5174 
5175 	case DIF_SUBR_DDI_PATHNAME: {
5176 		/*
5177 		 * This one is a galactic mess.  We are going to roughly
5178 		 * emulate ddi_pathname(), but it's made more complicated
5179 		 * by the fact that we (a) want to include the minor name and
5180 		 * (b) must proceed iteratively instead of recursively.
5181 		 */
5182 		uintptr_t dest = mstate->dtms_scratch_ptr;
5183 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
5184 		char *start = (char *)dest, *end = start + size - 1;
5185 		uintptr_t daddr = tupregs[0].dttk_value;
5186 		int64_t minor = (int64_t)tupregs[1].dttk_value;
5187 		char *s;
5188 		int i, len, depth = 0;
5189 
5190 		/*
5191 		 * Due to all the pointer jumping we do and context we must
5192 		 * rely upon, we just mandate that the user must have kernel
5193 		 * read privileges to use this routine.
5194 		 */
5195 		if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) == 0) {
5196 			*flags |= CPU_DTRACE_KPRIV;
5197 			*illval = daddr;
5198 			regs[rd] = 0;
5199 		}
5200 
5201 		if (!DTRACE_INSCRATCH(mstate, size)) {
5202 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5203 			regs[rd] = 0;
5204 			break;
5205 		}
5206 
5207 		*end = '\0';
5208 
5209 		/*
5210 		 * We want to have a name for the minor.  In order to do this,
5211 		 * we need to walk the minor list from the devinfo.  We want
5212 		 * to be sure that we don't infinitely walk a circular list,
5213 		 * so we check for circularity by sending a scout pointer
5214 		 * ahead two elements for every element that we iterate over;
5215 		 * if the list is circular, these will ultimately point to the
5216 		 * same element.  You may recognize this little trick as the
5217 		 * answer to a stupid interview question -- one that always
5218 		 * seems to be asked by those who had to have it laboriously
5219 		 * explained to them, and who can't even concisely describe
5220 		 * the conditions under which one would be forced to resort to
5221 		 * this technique.  Needless to say, those conditions are
5222 		 * found here -- and probably only here.  Is this the only use
5223 		 * of this infamous trick in shipping, production code?  If it
5224 		 * isn't, it probably should be...
5225 		 */
5226 		if (minor != -1) {
5227 			uintptr_t maddr = dtrace_loadptr(daddr +
5228 			    offsetof(struct dev_info, devi_minor));
5229 
5230 			uintptr_t next = offsetof(struct ddi_minor_data, next);
5231 			uintptr_t name = offsetof(struct ddi_minor_data,
5232 			    d_minor) + offsetof(struct ddi_minor, name);
5233 			uintptr_t dev = offsetof(struct ddi_minor_data,
5234 			    d_minor) + offsetof(struct ddi_minor, dev);
5235 			uintptr_t scout;
5236 
5237 			if (maddr != NULL)
5238 				scout = dtrace_loadptr(maddr + next);
5239 
5240 			while (maddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
5241 				uint64_t m;
5242 #ifdef _LP64
5243 				m = dtrace_load64(maddr + dev) & MAXMIN64;
5244 #else
5245 				m = dtrace_load32(maddr + dev) & MAXMIN;
5246 #endif
5247 				if (m != minor) {
5248 					maddr = dtrace_loadptr(maddr + next);
5249 
5250 					if (scout == NULL)
5251 						continue;
5252 
5253 					scout = dtrace_loadptr(scout + next);
5254 
5255 					if (scout == NULL)
5256 						continue;
5257 
5258 					scout = dtrace_loadptr(scout + next);
5259 
5260 					if (scout == NULL)
5261 						continue;
5262 
5263 					if (scout == maddr) {
5264 						*flags |= CPU_DTRACE_ILLOP;
5265 						break;
5266 					}
5267 
5268 					continue;
5269 				}
5270 
5271 				/*
5272 				 * We have the minor data.  Now we need to
5273 				 * copy the minor's name into the end of the
5274 				 * pathname.
5275 				 */
5276 				s = (char *)dtrace_loadptr(maddr + name);
5277 				len = dtrace_strlen(s, size);
5278 
5279 				if (*flags & CPU_DTRACE_FAULT)
5280 					break;
5281 
5282 				if (len != 0) {
5283 					if ((end -= (len + 1)) < start)
5284 						break;
5285 
5286 					*end = ':';
5287 				}
5288 
5289 				for (i = 1; i <= len; i++)
5290 					end[i] = dtrace_load8((uintptr_t)s++);
5291 				break;
5292 			}
5293 		}
5294 
5295 		while (daddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
5296 			ddi_node_state_t devi_state;
5297 
5298 			devi_state = dtrace_load32(daddr +
5299 			    offsetof(struct dev_info, devi_node_state));
5300 
5301 			if (*flags & CPU_DTRACE_FAULT)
5302 				break;
5303 
5304 			if (devi_state >= DS_INITIALIZED) {
5305 				s = (char *)dtrace_loadptr(daddr +
5306 				    offsetof(struct dev_info, devi_addr));
5307 				len = dtrace_strlen(s, size);
5308 
5309 				if (*flags & CPU_DTRACE_FAULT)
5310 					break;
5311 
5312 				if (len != 0) {
5313 					if ((end -= (len + 1)) < start)
5314 						break;
5315 
5316 					*end = '@';
5317 				}
5318 
5319 				for (i = 1; i <= len; i++)
5320 					end[i] = dtrace_load8((uintptr_t)s++);
5321 			}
5322 
5323 			/*
5324 			 * Now for the node name...
5325 			 */
5326 			s = (char *)dtrace_loadptr(daddr +
5327 			    offsetof(struct dev_info, devi_node_name));
5328 
5329 			daddr = dtrace_loadptr(daddr +
5330 			    offsetof(struct dev_info, devi_parent));
5331 
5332 			/*
5333 			 * If our parent is NULL (that is, if we're the root
5334 			 * node), we're going to use the special path
5335 			 * "devices".
5336 			 */
5337 			if (daddr == 0)
5338 				s = "devices";
5339 
5340 			len = dtrace_strlen(s, size);
5341 			if (*flags & CPU_DTRACE_FAULT)
5342 				break;
5343 
5344 			if ((end -= (len + 1)) < start)
5345 				break;
5346 
5347 			for (i = 1; i <= len; i++)
5348 				end[i] = dtrace_load8((uintptr_t)s++);
5349 			*end = '/';
5350 
5351 			if (depth++ > dtrace_devdepth_max) {
5352 				*flags |= CPU_DTRACE_ILLOP;
5353 				break;
5354 			}
5355 		}
5356 
5357 		if (end < start)
5358 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5359 
5360 		if (daddr == 0) {
5361 			regs[rd] = (uintptr_t)end;
5362 			mstate->dtms_scratch_ptr += size;
5363 		}
5364 
5365 		break;
5366 	}
5367 #endif
5368 
5369 	case DIF_SUBR_STRJOIN: {
5370 		char *d = (char *)mstate->dtms_scratch_ptr;
5371 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
5372 		uintptr_t s1 = tupregs[0].dttk_value;
5373 		uintptr_t s2 = tupregs[1].dttk_value;
5374 		int i = 0, j = 0;
5375 		size_t lim1, lim2;
5376 		char c;
5377 
5378 		if (!dtrace_strcanload(s1, size, &lim1, mstate, vstate) ||
5379 		    !dtrace_strcanload(s2, size, &lim2, mstate, vstate)) {
5380 			regs[rd] = 0;
5381 			break;
5382 		}
5383 
5384 		if (!DTRACE_INSCRATCH(mstate, size)) {
5385 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5386 			regs[rd] = 0;
5387 			break;
5388 		}
5389 
5390 		for (;;) {
5391 			if (i >= size) {
5392 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5393 				regs[rd] = 0;
5394 				break;
5395 			}
5396 			c = (i >= lim1) ? '\0' : dtrace_load8(s1++);
5397 			if ((d[i++] = c) == '\0') {
5398 				i--;
5399 				break;
5400 			}
5401 		}
5402 
5403 		for (;;) {
5404 			if (i >= size) {
5405 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5406 				regs[rd] = 0;
5407 				break;
5408 			}
5409 
5410 			c = (j++ >= lim2) ? '\0' : dtrace_load8(s2++);
5411 			if ((d[i++] = c) == '\0')
5412 				break;
5413 		}
5414 
5415 		if (i < size) {
5416 			mstate->dtms_scratch_ptr += i;
5417 			regs[rd] = (uintptr_t)d;
5418 		}
5419 
5420 		break;
5421 	}
5422 
5423 	case DIF_SUBR_STRTOLL: {
5424 		uintptr_t s = tupregs[0].dttk_value;
5425 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
5426 		size_t lim;
5427 		int base = 10;
5428 
5429 		if (nargs > 1) {
5430 			if ((base = tupregs[1].dttk_value) <= 1 ||
5431 			    base > ('z' - 'a' + 1) + ('9' - '0' + 1)) {
5432 				*flags |= CPU_DTRACE_ILLOP;
5433 				break;
5434 			}
5435 		}
5436 
5437 		if (!dtrace_strcanload(s, size, &lim, mstate, vstate)) {
5438 			regs[rd] = INT64_MIN;
5439 			break;
5440 		}
5441 
5442 		regs[rd] = dtrace_strtoll((char *)s, base, lim);
5443 		break;
5444 	}
5445 
5446 	case DIF_SUBR_LLTOSTR: {
5447 		int64_t i = (int64_t)tupregs[0].dttk_value;
5448 		uint64_t val, digit;
5449 		uint64_t size = 65;	/* enough room for 2^64 in binary */
5450 		char *end = (char *)mstate->dtms_scratch_ptr + size - 1;
5451 		int base = 10;
5452 
5453 		if (nargs > 1) {
5454 			if ((base = tupregs[1].dttk_value) <= 1 ||
5455 			    base > ('z' - 'a' + 1) + ('9' - '0' + 1)) {
5456 				*flags |= CPU_DTRACE_ILLOP;
5457 				break;
5458 			}
5459 		}
5460 
5461 		val = (base == 10 && i < 0) ? i * -1 : i;
5462 
5463 		if (!DTRACE_INSCRATCH(mstate, size)) {
5464 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5465 			regs[rd] = 0;
5466 			break;
5467 		}
5468 
5469 		for (*end-- = '\0'; val; val /= base) {
5470 			if ((digit = val % base) <= '9' - '0') {
5471 				*end-- = '0' + digit;
5472 			} else {
5473 				*end-- = 'a' + (digit - ('9' - '0') - 1);
5474 			}
5475 		}
5476 
5477 		if (i == 0 && base == 16)
5478 			*end-- = '0';
5479 
5480 		if (base == 16)
5481 			*end-- = 'x';
5482 
5483 		if (i == 0 || base == 8 || base == 16)
5484 			*end-- = '0';
5485 
5486 		if (i < 0 && base == 10)
5487 			*end-- = '-';
5488 
5489 		regs[rd] = (uintptr_t)end + 1;
5490 		mstate->dtms_scratch_ptr += size;
5491 		break;
5492 	}
5493 
5494 	case DIF_SUBR_HTONS:
5495 	case DIF_SUBR_NTOHS:
5496 #if BYTE_ORDER == BIG_ENDIAN
5497 		regs[rd] = (uint16_t)tupregs[0].dttk_value;
5498 #else
5499 		regs[rd] = DT_BSWAP_16((uint16_t)tupregs[0].dttk_value);
5500 #endif
5501 		break;
5502 
5503 
5504 	case DIF_SUBR_HTONL:
5505 	case DIF_SUBR_NTOHL:
5506 #if BYTE_ORDER == BIG_ENDIAN
5507 		regs[rd] = (uint32_t)tupregs[0].dttk_value;
5508 #else
5509 		regs[rd] = DT_BSWAP_32((uint32_t)tupregs[0].dttk_value);
5510 #endif
5511 		break;
5512 
5513 
5514 	case DIF_SUBR_HTONLL:
5515 	case DIF_SUBR_NTOHLL:
5516 #if BYTE_ORDER == BIG_ENDIAN
5517 		regs[rd] = (uint64_t)tupregs[0].dttk_value;
5518 #else
5519 		regs[rd] = DT_BSWAP_64((uint64_t)tupregs[0].dttk_value);
5520 #endif
5521 		break;
5522 
5523 
5524 	case DIF_SUBR_DIRNAME:
5525 	case DIF_SUBR_BASENAME: {
5526 		char *dest = (char *)mstate->dtms_scratch_ptr;
5527 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
5528 		uintptr_t src = tupregs[0].dttk_value;
5529 		int i, j, len = dtrace_strlen((char *)src, size);
5530 		int lastbase = -1, firstbase = -1, lastdir = -1;
5531 		int start, end;
5532 
5533 		if (!dtrace_canload(src, len + 1, mstate, vstate)) {
5534 			regs[rd] = 0;
5535 			break;
5536 		}
5537 
5538 		if (!DTRACE_INSCRATCH(mstate, size)) {
5539 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5540 			regs[rd] = 0;
5541 			break;
5542 		}
5543 
5544 		/*
5545 		 * The basename and dirname for a zero-length string is
5546 		 * defined to be "."
5547 		 */
5548 		if (len == 0) {
5549 			len = 1;
5550 			src = (uintptr_t)".";
5551 		}
5552 
5553 		/*
5554 		 * Start from the back of the string, moving back toward the
5555 		 * front until we see a character that isn't a slash.  That
5556 		 * character is the last character in the basename.
5557 		 */
5558 		for (i = len - 1; i >= 0; i--) {
5559 			if (dtrace_load8(src + i) != '/')
5560 				break;
5561 		}
5562 
5563 		if (i >= 0)
5564 			lastbase = i;
5565 
5566 		/*
5567 		 * Starting from the last character in the basename, move
5568 		 * towards the front until we find a slash.  The character
5569 		 * that we processed immediately before that is the first
5570 		 * character in the basename.
5571 		 */
5572 		for (; i >= 0; i--) {
5573 			if (dtrace_load8(src + i) == '/')
5574 				break;
5575 		}
5576 
5577 		if (i >= 0)
5578 			firstbase = i + 1;
5579 
5580 		/*
5581 		 * Now keep going until we find a non-slash character.  That
5582 		 * character is the last character in the dirname.
5583 		 */
5584 		for (; i >= 0; i--) {
5585 			if (dtrace_load8(src + i) != '/')
5586 				break;
5587 		}
5588 
5589 		if (i >= 0)
5590 			lastdir = i;
5591 
5592 		ASSERT(!(lastbase == -1 && firstbase != -1));
5593 		ASSERT(!(firstbase == -1 && lastdir != -1));
5594 
5595 		if (lastbase == -1) {
5596 			/*
5597 			 * We didn't find a non-slash character.  We know that
5598 			 * the length is non-zero, so the whole string must be
5599 			 * slashes.  In either the dirname or the basename
5600 			 * case, we return '/'.
5601 			 */
5602 			ASSERT(firstbase == -1);
5603 			firstbase = lastbase = lastdir = 0;
5604 		}
5605 
5606 		if (firstbase == -1) {
5607 			/*
5608 			 * The entire string consists only of a basename
5609 			 * component.  If we're looking for dirname, we need
5610 			 * to change our string to be just "."; if we're
5611 			 * looking for a basename, we'll just set the first
5612 			 * character of the basename to be 0.
5613 			 */
5614 			if (subr == DIF_SUBR_DIRNAME) {
5615 				ASSERT(lastdir == -1);
5616 				src = (uintptr_t)".";
5617 				lastdir = 0;
5618 			} else {
5619 				firstbase = 0;
5620 			}
5621 		}
5622 
5623 		if (subr == DIF_SUBR_DIRNAME) {
5624 			if (lastdir == -1) {
5625 				/*
5626 				 * We know that we have a slash in the name --
5627 				 * or lastdir would be set to 0, above.  And
5628 				 * because lastdir is -1, we know that this
5629 				 * slash must be the first character.  (That
5630 				 * is, the full string must be of the form
5631 				 * "/basename".)  In this case, the last
5632 				 * character of the directory name is 0.
5633 				 */
5634 				lastdir = 0;
5635 			}
5636 
5637 			start = 0;
5638 			end = lastdir;
5639 		} else {
5640 			ASSERT(subr == DIF_SUBR_BASENAME);
5641 			ASSERT(firstbase != -1 && lastbase != -1);
5642 			start = firstbase;
5643 			end = lastbase;
5644 		}
5645 
5646 		for (i = start, j = 0; i <= end && j < size - 1; i++, j++)
5647 			dest[j] = dtrace_load8(src + i);
5648 
5649 		dest[j] = '\0';
5650 		regs[rd] = (uintptr_t)dest;
5651 		mstate->dtms_scratch_ptr += size;
5652 		break;
5653 	}
5654 
5655 	case DIF_SUBR_GETF: {
5656 		uintptr_t fd = tupregs[0].dttk_value;
5657 		struct filedesc *fdp;
5658 		file_t *fp;
5659 
5660 		if (!dtrace_priv_proc(state)) {
5661 			regs[rd] = 0;
5662 			break;
5663 		}
5664 		fdp = curproc->p_fd;
5665 		FILEDESC_SLOCK(fdp);
5666 		fp = fget_locked(fdp, fd);
5667 		mstate->dtms_getf = fp;
5668 		regs[rd] = (uintptr_t)fp;
5669 		FILEDESC_SUNLOCK(fdp);
5670 		break;
5671 	}
5672 
5673 	case DIF_SUBR_CLEANPATH: {
5674 		char *dest = (char *)mstate->dtms_scratch_ptr, c;
5675 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
5676 		uintptr_t src = tupregs[0].dttk_value;
5677 		size_t lim;
5678 		int i = 0, j = 0;
5679 #ifdef illumos
5680 		zone_t *z;
5681 #endif
5682 
5683 		if (!dtrace_strcanload(src, size, &lim, mstate, vstate)) {
5684 			regs[rd] = 0;
5685 			break;
5686 		}
5687 
5688 		if (!DTRACE_INSCRATCH(mstate, size)) {
5689 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5690 			regs[rd] = 0;
5691 			break;
5692 		}
5693 
5694 		/*
5695 		 * Move forward, loading each character.
5696 		 */
5697 		do {
5698 			c = (i >= lim) ? '\0' : dtrace_load8(src + i++);
5699 next:
5700 			if (j + 5 >= size)	/* 5 = strlen("/..c\0") */
5701 				break;
5702 
5703 			if (c != '/') {
5704 				dest[j++] = c;
5705 				continue;
5706 			}
5707 
5708 			c = (i >= lim) ? '\0' : dtrace_load8(src + i++);
5709 
5710 			if (c == '/') {
5711 				/*
5712 				 * We have two slashes -- we can just advance
5713 				 * to the next character.
5714 				 */
5715 				goto next;
5716 			}
5717 
5718 			if (c != '.') {
5719 				/*
5720 				 * This is not "." and it's not ".." -- we can
5721 				 * just store the "/" and this character and
5722 				 * drive on.
5723 				 */
5724 				dest[j++] = '/';
5725 				dest[j++] = c;
5726 				continue;
5727 			}
5728 
5729 			c = (i >= lim) ? '\0' : dtrace_load8(src + i++);
5730 
5731 			if (c == '/') {
5732 				/*
5733 				 * This is a "/./" component.  We're not going
5734 				 * to store anything in the destination buffer;
5735 				 * we're just going to go to the next component.
5736 				 */
5737 				goto next;
5738 			}
5739 
5740 			if (c != '.') {
5741 				/*
5742 				 * This is not ".." -- we can just store the
5743 				 * "/." and this character and continue
5744 				 * processing.
5745 				 */
5746 				dest[j++] = '/';
5747 				dest[j++] = '.';
5748 				dest[j++] = c;
5749 				continue;
5750 			}
5751 
5752 			c = (i >= lim) ? '\0' : dtrace_load8(src + i++);
5753 
5754 			if (c != '/' && c != '\0') {
5755 				/*
5756 				 * This is not ".." -- it's "..[mumble]".
5757 				 * We'll store the "/.." and this character
5758 				 * and continue processing.
5759 				 */
5760 				dest[j++] = '/';
5761 				dest[j++] = '.';
5762 				dest[j++] = '.';
5763 				dest[j++] = c;
5764 				continue;
5765 			}
5766 
5767 			/*
5768 			 * This is "/../" or "/..\0".  We need to back up
5769 			 * our destination pointer until we find a "/".
5770 			 */
5771 			i--;
5772 			while (j != 0 && dest[--j] != '/')
5773 				continue;
5774 
5775 			if (c == '\0')
5776 				dest[++j] = '/';
5777 		} while (c != '\0');
5778 
5779 		dest[j] = '\0';
5780 
5781 #ifdef illumos
5782 		if (mstate->dtms_getf != NULL &&
5783 		    !(mstate->dtms_access & DTRACE_ACCESS_KERNEL) &&
5784 		    (z = state->dts_cred.dcr_cred->cr_zone) != kcred->cr_zone) {
5785 			/*
5786 			 * If we've done a getf() as a part of this ECB and we
5787 			 * don't have kernel access (and we're not in the global
5788 			 * zone), check if the path we cleaned up begins with
5789 			 * the zone's root path, and trim it off if so.  Note
5790 			 * that this is an output cleanliness issue, not a
5791 			 * security issue: knowing one's zone root path does
5792 			 * not enable privilege escalation.
5793 			 */
5794 			if (strstr(dest, z->zone_rootpath) == dest)
5795 				dest += strlen(z->zone_rootpath) - 1;
5796 		}
5797 #endif
5798 
5799 		regs[rd] = (uintptr_t)dest;
5800 		mstate->dtms_scratch_ptr += size;
5801 		break;
5802 	}
5803 
5804 	case DIF_SUBR_INET_NTOA:
5805 	case DIF_SUBR_INET_NTOA6:
5806 	case DIF_SUBR_INET_NTOP: {
5807 		size_t size;
5808 		int af, argi, i;
5809 		char *base, *end;
5810 
5811 		if (subr == DIF_SUBR_INET_NTOP) {
5812 			af = (int)tupregs[0].dttk_value;
5813 			argi = 1;
5814 		} else {
5815 			af = subr == DIF_SUBR_INET_NTOA ? AF_INET: AF_INET6;
5816 			argi = 0;
5817 		}
5818 
5819 		if (af == AF_INET) {
5820 			ipaddr_t ip4;
5821 			uint8_t *ptr8, val;
5822 
5823 			if (!dtrace_canload(tupregs[argi].dttk_value,
5824 			    sizeof (ipaddr_t), mstate, vstate)) {
5825 				regs[rd] = 0;
5826 				break;
5827 			}
5828 
5829 			/*
5830 			 * Safely load the IPv4 address.
5831 			 */
5832 			ip4 = dtrace_load32(tupregs[argi].dttk_value);
5833 
5834 			/*
5835 			 * Check an IPv4 string will fit in scratch.
5836 			 */
5837 			size = INET_ADDRSTRLEN;
5838 			if (!DTRACE_INSCRATCH(mstate, size)) {
5839 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5840 				regs[rd] = 0;
5841 				break;
5842 			}
5843 			base = (char *)mstate->dtms_scratch_ptr;
5844 			end = (char *)mstate->dtms_scratch_ptr + size - 1;
5845 
5846 			/*
5847 			 * Stringify as a dotted decimal quad.
5848 			 */
5849 			*end-- = '\0';
5850 			ptr8 = (uint8_t *)&ip4;
5851 			for (i = 3; i >= 0; i--) {
5852 				val = ptr8[i];
5853 
5854 				if (val == 0) {
5855 					*end-- = '0';
5856 				} else {
5857 					for (; val; val /= 10) {
5858 						*end-- = '0' + (val % 10);
5859 					}
5860 				}
5861 
5862 				if (i > 0)
5863 					*end-- = '.';
5864 			}
5865 			ASSERT(end + 1 >= base);
5866 
5867 		} else if (af == AF_INET6) {
5868 			struct in6_addr ip6;
5869 			int firstzero, tryzero, numzero, v6end;
5870 			uint16_t val;
5871 			const char digits[] = "0123456789abcdef";
5872 
5873 			/*
5874 			 * Stringify using RFC 1884 convention 2 - 16 bit
5875 			 * hexadecimal values with a zero-run compression.
5876 			 * Lower case hexadecimal digits are used.
5877 			 * 	eg, fe80::214:4fff:fe0b:76c8.
5878 			 * The IPv4 embedded form is returned for inet_ntop,
5879 			 * just the IPv4 string is returned for inet_ntoa6.
5880 			 */
5881 
5882 			if (!dtrace_canload(tupregs[argi].dttk_value,
5883 			    sizeof (struct in6_addr), mstate, vstate)) {
5884 				regs[rd] = 0;
5885 				break;
5886 			}
5887 
5888 			/*
5889 			 * Safely load the IPv6 address.
5890 			 */
5891 			dtrace_bcopy(
5892 			    (void *)(uintptr_t)tupregs[argi].dttk_value,
5893 			    (void *)(uintptr_t)&ip6, sizeof (struct in6_addr));
5894 
5895 			/*
5896 			 * Check an IPv6 string will fit in scratch.
5897 			 */
5898 			size = INET6_ADDRSTRLEN;
5899 			if (!DTRACE_INSCRATCH(mstate, size)) {
5900 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5901 				regs[rd] = 0;
5902 				break;
5903 			}
5904 			base = (char *)mstate->dtms_scratch_ptr;
5905 			end = (char *)mstate->dtms_scratch_ptr + size - 1;
5906 			*end-- = '\0';
5907 
5908 			/*
5909 			 * Find the longest run of 16 bit zero values
5910 			 * for the single allowed zero compression - "::".
5911 			 */
5912 			firstzero = -1;
5913 			tryzero = -1;
5914 			numzero = 1;
5915 			for (i = 0; i < sizeof (struct in6_addr); i++) {
5916 #ifdef illumos
5917 				if (ip6._S6_un._S6_u8[i] == 0 &&
5918 #else
5919 				if (ip6.__u6_addr.__u6_addr8[i] == 0 &&
5920 #endif
5921 				    tryzero == -1 && i % 2 == 0) {
5922 					tryzero = i;
5923 					continue;
5924 				}
5925 
5926 				if (tryzero != -1 &&
5927 #ifdef illumos
5928 				    (ip6._S6_un._S6_u8[i] != 0 ||
5929 #else
5930 				    (ip6.__u6_addr.__u6_addr8[i] != 0 ||
5931 #endif
5932 				    i == sizeof (struct in6_addr) - 1)) {
5933 
5934 					if (i - tryzero <= numzero) {
5935 						tryzero = -1;
5936 						continue;
5937 					}
5938 
5939 					firstzero = tryzero;
5940 					numzero = i - i % 2 - tryzero;
5941 					tryzero = -1;
5942 
5943 #ifdef illumos
5944 					if (ip6._S6_un._S6_u8[i] == 0 &&
5945 #else
5946 					if (ip6.__u6_addr.__u6_addr8[i] == 0 &&
5947 #endif
5948 					    i == sizeof (struct in6_addr) - 1)
5949 						numzero += 2;
5950 				}
5951 			}
5952 			ASSERT(firstzero + numzero <= sizeof (struct in6_addr));
5953 
5954 			/*
5955 			 * Check for an IPv4 embedded address.
5956 			 */
5957 			v6end = sizeof (struct in6_addr) - 2;
5958 			if (IN6_IS_ADDR_V4MAPPED(&ip6) ||
5959 			    IN6_IS_ADDR_V4COMPAT(&ip6)) {
5960 				for (i = sizeof (struct in6_addr) - 1;
5961 				    i >= DTRACE_V4MAPPED_OFFSET; i--) {
5962 					ASSERT(end >= base);
5963 
5964 #ifdef illumos
5965 					val = ip6._S6_un._S6_u8[i];
5966 #else
5967 					val = ip6.__u6_addr.__u6_addr8[i];
5968 #endif
5969 
5970 					if (val == 0) {
5971 						*end-- = '0';
5972 					} else {
5973 						for (; val; val /= 10) {
5974 							*end-- = '0' + val % 10;
5975 						}
5976 					}
5977 
5978 					if (i > DTRACE_V4MAPPED_OFFSET)
5979 						*end-- = '.';
5980 				}
5981 
5982 				if (subr == DIF_SUBR_INET_NTOA6)
5983 					goto inetout;
5984 
5985 				/*
5986 				 * Set v6end to skip the IPv4 address that
5987 				 * we have already stringified.
5988 				 */
5989 				v6end = 10;
5990 			}
5991 
5992 			/*
5993 			 * Build the IPv6 string by working through the
5994 			 * address in reverse.
5995 			 */
5996 			for (i = v6end; i >= 0; i -= 2) {
5997 				ASSERT(end >= base);
5998 
5999 				if (i == firstzero + numzero - 2) {
6000 					*end-- = ':';
6001 					*end-- = ':';
6002 					i -= numzero - 2;
6003 					continue;
6004 				}
6005 
6006 				if (i < 14 && i != firstzero - 2)
6007 					*end-- = ':';
6008 
6009 #ifdef illumos
6010 				val = (ip6._S6_un._S6_u8[i] << 8) +
6011 				    ip6._S6_un._S6_u8[i + 1];
6012 #else
6013 				val = (ip6.__u6_addr.__u6_addr8[i] << 8) +
6014 				    ip6.__u6_addr.__u6_addr8[i + 1];
6015 #endif
6016 
6017 				if (val == 0) {
6018 					*end-- = '0';
6019 				} else {
6020 					for (; val; val /= 16) {
6021 						*end-- = digits[val % 16];
6022 					}
6023 				}
6024 			}
6025 			ASSERT(end + 1 >= base);
6026 
6027 		} else {
6028 			/*
6029 			 * The user didn't use AH_INET or AH_INET6.
6030 			 */
6031 			DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
6032 			regs[rd] = 0;
6033 			break;
6034 		}
6035 
6036 inetout:	regs[rd] = (uintptr_t)end + 1;
6037 		mstate->dtms_scratch_ptr += size;
6038 		break;
6039 	}
6040 
6041 	case DIF_SUBR_MEMREF: {
6042 		uintptr_t size = 2 * sizeof(uintptr_t);
6043 		uintptr_t *memref = (uintptr_t *) P2ROUNDUP(mstate->dtms_scratch_ptr, sizeof(uintptr_t));
6044 		size_t scratch_size = ((uintptr_t) memref - mstate->dtms_scratch_ptr) + size;
6045 
6046 		/* address and length */
6047 		memref[0] = tupregs[0].dttk_value;
6048 		memref[1] = tupregs[1].dttk_value;
6049 
6050 		regs[rd] = (uintptr_t) memref;
6051 		mstate->dtms_scratch_ptr += scratch_size;
6052 		break;
6053 	}
6054 
6055 #ifndef illumos
6056 	case DIF_SUBR_MEMSTR: {
6057 		char *str = (char *)mstate->dtms_scratch_ptr;
6058 		uintptr_t mem = tupregs[0].dttk_value;
6059 		char c = tupregs[1].dttk_value;
6060 		size_t size = tupregs[2].dttk_value;
6061 		uint8_t n;
6062 		int i;
6063 
6064 		regs[rd] = 0;
6065 
6066 		if (size == 0)
6067 			break;
6068 
6069 		if (!dtrace_canload(mem, size - 1, mstate, vstate))
6070 			break;
6071 
6072 		if (!DTRACE_INSCRATCH(mstate, size)) {
6073 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
6074 			break;
6075 		}
6076 
6077 		if (dtrace_memstr_max != 0 && size > dtrace_memstr_max) {
6078 			*flags |= CPU_DTRACE_ILLOP;
6079 			break;
6080 		}
6081 
6082 		for (i = 0; i < size - 1; i++) {
6083 			n = dtrace_load8(mem++);
6084 			str[i] = (n == 0) ? c : n;
6085 		}
6086 		str[size - 1] = 0;
6087 
6088 		regs[rd] = (uintptr_t)str;
6089 		mstate->dtms_scratch_ptr += size;
6090 		break;
6091 	}
6092 #endif
6093 	}
6094 }
6095 
6096 /*
6097  * Emulate the execution of DTrace IR instructions specified by the given
6098  * DIF object.  This function is deliberately void of assertions as all of
6099  * the necessary checks are handled by a call to dtrace_difo_validate().
6100  */
6101 static uint64_t
6102 dtrace_dif_emulate(dtrace_difo_t *difo, dtrace_mstate_t *mstate,
6103     dtrace_vstate_t *vstate, dtrace_state_t *state)
6104 {
6105 	const dif_instr_t *text = difo->dtdo_buf;
6106 	const uint_t textlen = difo->dtdo_len;
6107 	const char *strtab = difo->dtdo_strtab;
6108 	const uint64_t *inttab = difo->dtdo_inttab;
6109 
6110 	uint64_t rval = 0;
6111 	dtrace_statvar_t *svar;
6112 	dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
6113 	dtrace_difv_t *v;
6114 	volatile uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
6115 	volatile uintptr_t *illval = &cpu_core[curcpu].cpuc_dtrace_illval;
6116 
6117 	dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
6118 	uint64_t regs[DIF_DIR_NREGS];
6119 	uint64_t *tmp;
6120 
6121 	uint8_t cc_n = 0, cc_z = 0, cc_v = 0, cc_c = 0;
6122 	int64_t cc_r;
6123 	uint_t pc = 0, id, opc = 0;
6124 	uint8_t ttop = 0;
6125 	dif_instr_t instr;
6126 	uint_t r1, r2, rd;
6127 
6128 	/*
6129 	 * We stash the current DIF object into the machine state: we need it
6130 	 * for subsequent access checking.
6131 	 */
6132 	mstate->dtms_difo = difo;
6133 
6134 	regs[DIF_REG_R0] = 0; 		/* %r0 is fixed at zero */
6135 
6136 	while (pc < textlen && !(*flags & CPU_DTRACE_FAULT)) {
6137 		opc = pc;
6138 
6139 		instr = text[pc++];
6140 		r1 = DIF_INSTR_R1(instr);
6141 		r2 = DIF_INSTR_R2(instr);
6142 		rd = DIF_INSTR_RD(instr);
6143 
6144 		switch (DIF_INSTR_OP(instr)) {
6145 		case DIF_OP_OR:
6146 			regs[rd] = regs[r1] | regs[r2];
6147 			break;
6148 		case DIF_OP_XOR:
6149 			regs[rd] = regs[r1] ^ regs[r2];
6150 			break;
6151 		case DIF_OP_AND:
6152 			regs[rd] = regs[r1] & regs[r2];
6153 			break;
6154 		case DIF_OP_SLL:
6155 			regs[rd] = regs[r1] << regs[r2];
6156 			break;
6157 		case DIF_OP_SRL:
6158 			regs[rd] = regs[r1] >> regs[r2];
6159 			break;
6160 		case DIF_OP_SUB:
6161 			regs[rd] = regs[r1] - regs[r2];
6162 			break;
6163 		case DIF_OP_ADD:
6164 			regs[rd] = regs[r1] + regs[r2];
6165 			break;
6166 		case DIF_OP_MUL:
6167 			regs[rd] = regs[r1] * regs[r2];
6168 			break;
6169 		case DIF_OP_SDIV:
6170 			if (regs[r2] == 0) {
6171 				regs[rd] = 0;
6172 				*flags |= CPU_DTRACE_DIVZERO;
6173 			} else {
6174 				regs[rd] = (int64_t)regs[r1] /
6175 				    (int64_t)regs[r2];
6176 			}
6177 			break;
6178 
6179 		case DIF_OP_UDIV:
6180 			if (regs[r2] == 0) {
6181 				regs[rd] = 0;
6182 				*flags |= CPU_DTRACE_DIVZERO;
6183 			} else {
6184 				regs[rd] = regs[r1] / regs[r2];
6185 			}
6186 			break;
6187 
6188 		case DIF_OP_SREM:
6189 			if (regs[r2] == 0) {
6190 				regs[rd] = 0;
6191 				*flags |= CPU_DTRACE_DIVZERO;
6192 			} else {
6193 				regs[rd] = (int64_t)regs[r1] %
6194 				    (int64_t)regs[r2];
6195 			}
6196 			break;
6197 
6198 		case DIF_OP_UREM:
6199 			if (regs[r2] == 0) {
6200 				regs[rd] = 0;
6201 				*flags |= CPU_DTRACE_DIVZERO;
6202 			} else {
6203 				regs[rd] = regs[r1] % regs[r2];
6204 			}
6205 			break;
6206 
6207 		case DIF_OP_NOT:
6208 			regs[rd] = ~regs[r1];
6209 			break;
6210 		case DIF_OP_MOV:
6211 			regs[rd] = regs[r1];
6212 			break;
6213 		case DIF_OP_CMP:
6214 			cc_r = regs[r1] - regs[r2];
6215 			cc_n = cc_r < 0;
6216 			cc_z = cc_r == 0;
6217 			cc_v = 0;
6218 			cc_c = regs[r1] < regs[r2];
6219 			break;
6220 		case DIF_OP_TST:
6221 			cc_n = cc_v = cc_c = 0;
6222 			cc_z = regs[r1] == 0;
6223 			break;
6224 		case DIF_OP_BA:
6225 			pc = DIF_INSTR_LABEL(instr);
6226 			break;
6227 		case DIF_OP_BE:
6228 			if (cc_z)
6229 				pc = DIF_INSTR_LABEL(instr);
6230 			break;
6231 		case DIF_OP_BNE:
6232 			if (cc_z == 0)
6233 				pc = DIF_INSTR_LABEL(instr);
6234 			break;
6235 		case DIF_OP_BG:
6236 			if ((cc_z | (cc_n ^ cc_v)) == 0)
6237 				pc = DIF_INSTR_LABEL(instr);
6238 			break;
6239 		case DIF_OP_BGU:
6240 			if ((cc_c | cc_z) == 0)
6241 				pc = DIF_INSTR_LABEL(instr);
6242 			break;
6243 		case DIF_OP_BGE:
6244 			if ((cc_n ^ cc_v) == 0)
6245 				pc = DIF_INSTR_LABEL(instr);
6246 			break;
6247 		case DIF_OP_BGEU:
6248 			if (cc_c == 0)
6249 				pc = DIF_INSTR_LABEL(instr);
6250 			break;
6251 		case DIF_OP_BL:
6252 			if (cc_n ^ cc_v)
6253 				pc = DIF_INSTR_LABEL(instr);
6254 			break;
6255 		case DIF_OP_BLU:
6256 			if (cc_c)
6257 				pc = DIF_INSTR_LABEL(instr);
6258 			break;
6259 		case DIF_OP_BLE:
6260 			if (cc_z | (cc_n ^ cc_v))
6261 				pc = DIF_INSTR_LABEL(instr);
6262 			break;
6263 		case DIF_OP_BLEU:
6264 			if (cc_c | cc_z)
6265 				pc = DIF_INSTR_LABEL(instr);
6266 			break;
6267 		case DIF_OP_RLDSB:
6268 			if (!dtrace_canload(regs[r1], 1, mstate, vstate))
6269 				break;
6270 			/*FALLTHROUGH*/
6271 		case DIF_OP_LDSB:
6272 			regs[rd] = (int8_t)dtrace_load8(regs[r1]);
6273 			break;
6274 		case DIF_OP_RLDSH:
6275 			if (!dtrace_canload(regs[r1], 2, mstate, vstate))
6276 				break;
6277 			/*FALLTHROUGH*/
6278 		case DIF_OP_LDSH:
6279 			regs[rd] = (int16_t)dtrace_load16(regs[r1]);
6280 			break;
6281 		case DIF_OP_RLDSW:
6282 			if (!dtrace_canload(regs[r1], 4, mstate, vstate))
6283 				break;
6284 			/*FALLTHROUGH*/
6285 		case DIF_OP_LDSW:
6286 			regs[rd] = (int32_t)dtrace_load32(regs[r1]);
6287 			break;
6288 		case DIF_OP_RLDUB:
6289 			if (!dtrace_canload(regs[r1], 1, mstate, vstate))
6290 				break;
6291 			/*FALLTHROUGH*/
6292 		case DIF_OP_LDUB:
6293 			regs[rd] = dtrace_load8(regs[r1]);
6294 			break;
6295 		case DIF_OP_RLDUH:
6296 			if (!dtrace_canload(regs[r1], 2, mstate, vstate))
6297 				break;
6298 			/*FALLTHROUGH*/
6299 		case DIF_OP_LDUH:
6300 			regs[rd] = dtrace_load16(regs[r1]);
6301 			break;
6302 		case DIF_OP_RLDUW:
6303 			if (!dtrace_canload(regs[r1], 4, mstate, vstate))
6304 				break;
6305 			/*FALLTHROUGH*/
6306 		case DIF_OP_LDUW:
6307 			regs[rd] = dtrace_load32(regs[r1]);
6308 			break;
6309 		case DIF_OP_RLDX:
6310 			if (!dtrace_canload(regs[r1], 8, mstate, vstate))
6311 				break;
6312 			/*FALLTHROUGH*/
6313 		case DIF_OP_LDX:
6314 			regs[rd] = dtrace_load64(regs[r1]);
6315 			break;
6316 		case DIF_OP_ULDSB:
6317 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6318 			regs[rd] = (int8_t)
6319 			    dtrace_fuword8((void *)(uintptr_t)regs[r1]);
6320 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6321 			break;
6322 		case DIF_OP_ULDSH:
6323 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6324 			regs[rd] = (int16_t)
6325 			    dtrace_fuword16((void *)(uintptr_t)regs[r1]);
6326 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6327 			break;
6328 		case DIF_OP_ULDSW:
6329 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6330 			regs[rd] = (int32_t)
6331 			    dtrace_fuword32((void *)(uintptr_t)regs[r1]);
6332 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6333 			break;
6334 		case DIF_OP_ULDUB:
6335 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6336 			regs[rd] =
6337 			    dtrace_fuword8((void *)(uintptr_t)regs[r1]);
6338 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6339 			break;
6340 		case DIF_OP_ULDUH:
6341 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6342 			regs[rd] =
6343 			    dtrace_fuword16((void *)(uintptr_t)regs[r1]);
6344 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6345 			break;
6346 		case DIF_OP_ULDUW:
6347 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6348 			regs[rd] =
6349 			    dtrace_fuword32((void *)(uintptr_t)regs[r1]);
6350 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6351 			break;
6352 		case DIF_OP_ULDX:
6353 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6354 			regs[rd] =
6355 			    dtrace_fuword64((void *)(uintptr_t)regs[r1]);
6356 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6357 			break;
6358 		case DIF_OP_RET:
6359 			rval = regs[rd];
6360 			pc = textlen;
6361 			break;
6362 		case DIF_OP_NOP:
6363 			break;
6364 		case DIF_OP_SETX:
6365 			regs[rd] = inttab[DIF_INSTR_INTEGER(instr)];
6366 			break;
6367 		case DIF_OP_SETS:
6368 			regs[rd] = (uint64_t)(uintptr_t)
6369 			    (strtab + DIF_INSTR_STRING(instr));
6370 			break;
6371 		case DIF_OP_SCMP: {
6372 			size_t sz = state->dts_options[DTRACEOPT_STRSIZE];
6373 			uintptr_t s1 = regs[r1];
6374 			uintptr_t s2 = regs[r2];
6375 			size_t lim1, lim2;
6376 
6377 			/*
6378 			 * If one of the strings is NULL then the limit becomes
6379 			 * 0 which compares 0 characters in dtrace_strncmp()
6380 			 * resulting in a false positive.  dtrace_strncmp()
6381 			 * treats a NULL as an empty 1-char string.
6382 			 */
6383 			lim1 = lim2 = 1;
6384 
6385 			if (s1 != 0 &&
6386 			    !dtrace_strcanload(s1, sz, &lim1, mstate, vstate))
6387 				break;
6388 			if (s2 != 0 &&
6389 			    !dtrace_strcanload(s2, sz, &lim2, mstate, vstate))
6390 				break;
6391 
6392 			cc_r = dtrace_strncmp((char *)s1, (char *)s2,
6393 			    MIN(lim1, lim2));
6394 
6395 			cc_n = cc_r < 0;
6396 			cc_z = cc_r == 0;
6397 			cc_v = cc_c = 0;
6398 			break;
6399 		}
6400 		case DIF_OP_LDGA:
6401 			regs[rd] = dtrace_dif_variable(mstate, state,
6402 			    r1, regs[r2]);
6403 			break;
6404 		case DIF_OP_LDGS:
6405 			id = DIF_INSTR_VAR(instr);
6406 
6407 			if (id >= DIF_VAR_OTHER_UBASE) {
6408 				uintptr_t a;
6409 
6410 				id -= DIF_VAR_OTHER_UBASE;
6411 				svar = vstate->dtvs_globals[id];
6412 				ASSERT(svar != NULL);
6413 				v = &svar->dtsv_var;
6414 
6415 				if (!(v->dtdv_type.dtdt_flags & DIF_TF_BYREF)) {
6416 					regs[rd] = svar->dtsv_data;
6417 					break;
6418 				}
6419 
6420 				a = (uintptr_t)svar->dtsv_data;
6421 
6422 				if (*(uint8_t *)a == UINT8_MAX) {
6423 					/*
6424 					 * If the 0th byte is set to UINT8_MAX
6425 					 * then this is to be treated as a
6426 					 * reference to a NULL variable.
6427 					 */
6428 					regs[rd] = 0;
6429 				} else {
6430 					regs[rd] = a + sizeof (uint64_t);
6431 				}
6432 
6433 				break;
6434 			}
6435 
6436 			regs[rd] = dtrace_dif_variable(mstate, state, id, 0);
6437 			break;
6438 
6439 		case DIF_OP_STGS:
6440 			id = DIF_INSTR_VAR(instr);
6441 
6442 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
6443 			id -= DIF_VAR_OTHER_UBASE;
6444 
6445 			VERIFY(id < vstate->dtvs_nglobals);
6446 			svar = vstate->dtvs_globals[id];
6447 			ASSERT(svar != NULL);
6448 			v = &svar->dtsv_var;
6449 
6450 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6451 				uintptr_t a = (uintptr_t)svar->dtsv_data;
6452 				size_t lim;
6453 
6454 				ASSERT(a != 0);
6455 				ASSERT(svar->dtsv_size != 0);
6456 
6457 				if (regs[rd] == 0) {
6458 					*(uint8_t *)a = UINT8_MAX;
6459 					break;
6460 				} else {
6461 					*(uint8_t *)a = 0;
6462 					a += sizeof (uint64_t);
6463 				}
6464 				if (!dtrace_vcanload(
6465 				    (void *)(uintptr_t)regs[rd], &v->dtdv_type,
6466 				    &lim, mstate, vstate))
6467 					break;
6468 
6469 				dtrace_vcopy((void *)(uintptr_t)regs[rd],
6470 				    (void *)a, &v->dtdv_type, lim);
6471 				break;
6472 			}
6473 
6474 			svar->dtsv_data = regs[rd];
6475 			break;
6476 
6477 		case DIF_OP_LDTA:
6478 			/*
6479 			 * There are no DTrace built-in thread-local arrays at
6480 			 * present.  This opcode is saved for future work.
6481 			 */
6482 			*flags |= CPU_DTRACE_ILLOP;
6483 			regs[rd] = 0;
6484 			break;
6485 
6486 		case DIF_OP_LDLS:
6487 			id = DIF_INSTR_VAR(instr);
6488 
6489 			if (id < DIF_VAR_OTHER_UBASE) {
6490 				/*
6491 				 * For now, this has no meaning.
6492 				 */
6493 				regs[rd] = 0;
6494 				break;
6495 			}
6496 
6497 			id -= DIF_VAR_OTHER_UBASE;
6498 
6499 			ASSERT(id < vstate->dtvs_nlocals);
6500 			ASSERT(vstate->dtvs_locals != NULL);
6501 
6502 			svar = vstate->dtvs_locals[id];
6503 			ASSERT(svar != NULL);
6504 			v = &svar->dtsv_var;
6505 
6506 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6507 				uintptr_t a = (uintptr_t)svar->dtsv_data;
6508 				size_t sz = v->dtdv_type.dtdt_size;
6509 				size_t lim;
6510 
6511 				sz += sizeof (uint64_t);
6512 				ASSERT(svar->dtsv_size == NCPU * sz);
6513 				a += curcpu * sz;
6514 
6515 				if (*(uint8_t *)a == UINT8_MAX) {
6516 					/*
6517 					 * If the 0th byte is set to UINT8_MAX
6518 					 * then this is to be treated as a
6519 					 * reference to a NULL variable.
6520 					 */
6521 					regs[rd] = 0;
6522 				} else {
6523 					regs[rd] = a + sizeof (uint64_t);
6524 				}
6525 
6526 				break;
6527 			}
6528 
6529 			ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
6530 			tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
6531 			regs[rd] = tmp[curcpu];
6532 			break;
6533 
6534 		case DIF_OP_STLS:
6535 			id = DIF_INSTR_VAR(instr);
6536 
6537 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
6538 			id -= DIF_VAR_OTHER_UBASE;
6539 			VERIFY(id < vstate->dtvs_nlocals);
6540 
6541 			ASSERT(vstate->dtvs_locals != NULL);
6542 			svar = vstate->dtvs_locals[id];
6543 			ASSERT(svar != NULL);
6544 			v = &svar->dtsv_var;
6545 
6546 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6547 				uintptr_t a = (uintptr_t)svar->dtsv_data;
6548 				size_t sz = v->dtdv_type.dtdt_size;
6549 				size_t lim;
6550 
6551 				sz += sizeof (uint64_t);
6552 				ASSERT(svar->dtsv_size == NCPU * sz);
6553 				a += curcpu * sz;
6554 
6555 				if (regs[rd] == 0) {
6556 					*(uint8_t *)a = UINT8_MAX;
6557 					break;
6558 				} else {
6559 					*(uint8_t *)a = 0;
6560 					a += sizeof (uint64_t);
6561 				}
6562 
6563 				if (!dtrace_vcanload(
6564 				    (void *)(uintptr_t)regs[rd], &v->dtdv_type,
6565 				    &lim, mstate, vstate))
6566 					break;
6567 
6568 				dtrace_vcopy((void *)(uintptr_t)regs[rd],
6569 				    (void *)a, &v->dtdv_type, lim);
6570 				break;
6571 			}
6572 
6573 			ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
6574 			tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
6575 			tmp[curcpu] = regs[rd];
6576 			break;
6577 
6578 		case DIF_OP_LDTS: {
6579 			dtrace_dynvar_t *dvar;
6580 			dtrace_key_t *key;
6581 
6582 			id = DIF_INSTR_VAR(instr);
6583 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
6584 			id -= DIF_VAR_OTHER_UBASE;
6585 			v = &vstate->dtvs_tlocals[id];
6586 
6587 			key = &tupregs[DIF_DTR_NREGS];
6588 			key[0].dttk_value = (uint64_t)id;
6589 			key[0].dttk_size = 0;
6590 			DTRACE_TLS_THRKEY(key[1].dttk_value);
6591 			key[1].dttk_size = 0;
6592 
6593 			dvar = dtrace_dynvar(dstate, 2, key,
6594 			    sizeof (uint64_t), DTRACE_DYNVAR_NOALLOC,
6595 			    mstate, vstate);
6596 
6597 			if (dvar == NULL) {
6598 				regs[rd] = 0;
6599 				break;
6600 			}
6601 
6602 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6603 				regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
6604 			} else {
6605 				regs[rd] = *((uint64_t *)dvar->dtdv_data);
6606 			}
6607 
6608 			break;
6609 		}
6610 
6611 		case DIF_OP_STTS: {
6612 			dtrace_dynvar_t *dvar;
6613 			dtrace_key_t *key;
6614 
6615 			id = DIF_INSTR_VAR(instr);
6616 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
6617 			id -= DIF_VAR_OTHER_UBASE;
6618 			VERIFY(id < vstate->dtvs_ntlocals);
6619 
6620 			key = &tupregs[DIF_DTR_NREGS];
6621 			key[0].dttk_value = (uint64_t)id;
6622 			key[0].dttk_size = 0;
6623 			DTRACE_TLS_THRKEY(key[1].dttk_value);
6624 			key[1].dttk_size = 0;
6625 			v = &vstate->dtvs_tlocals[id];
6626 
6627 			dvar = dtrace_dynvar(dstate, 2, key,
6628 			    v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
6629 			    v->dtdv_type.dtdt_size : sizeof (uint64_t),
6630 			    regs[rd] ? DTRACE_DYNVAR_ALLOC :
6631 			    DTRACE_DYNVAR_DEALLOC, mstate, vstate);
6632 
6633 			/*
6634 			 * Given that we're storing to thread-local data,
6635 			 * we need to flush our predicate cache.
6636 			 */
6637 			curthread->t_predcache = 0;
6638 
6639 			if (dvar == NULL)
6640 				break;
6641 
6642 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6643 				size_t lim;
6644 
6645 				if (!dtrace_vcanload(
6646 				    (void *)(uintptr_t)regs[rd],
6647 				    &v->dtdv_type, &lim, mstate, vstate))
6648 					break;
6649 
6650 				dtrace_vcopy((void *)(uintptr_t)regs[rd],
6651 				    dvar->dtdv_data, &v->dtdv_type, lim);
6652 			} else {
6653 				*((uint64_t *)dvar->dtdv_data) = regs[rd];
6654 			}
6655 
6656 			break;
6657 		}
6658 
6659 		case DIF_OP_SRA:
6660 			regs[rd] = (int64_t)regs[r1] >> regs[r2];
6661 			break;
6662 
6663 		case DIF_OP_CALL:
6664 			dtrace_dif_subr(DIF_INSTR_SUBR(instr), rd,
6665 			    regs, tupregs, ttop, mstate, state);
6666 			break;
6667 
6668 		case DIF_OP_PUSHTR:
6669 			if (ttop == DIF_DTR_NREGS) {
6670 				*flags |= CPU_DTRACE_TUPOFLOW;
6671 				break;
6672 			}
6673 
6674 			if (r1 == DIF_TYPE_STRING) {
6675 				/*
6676 				 * If this is a string type and the size is 0,
6677 				 * we'll use the system-wide default string
6678 				 * size.  Note that we are _not_ looking at
6679 				 * the value of the DTRACEOPT_STRSIZE option;
6680 				 * had this been set, we would expect to have
6681 				 * a non-zero size value in the "pushtr".
6682 				 */
6683 				tupregs[ttop].dttk_size =
6684 				    dtrace_strlen((char *)(uintptr_t)regs[rd],
6685 				    regs[r2] ? regs[r2] :
6686 				    dtrace_strsize_default) + 1;
6687 			} else {
6688 				if (regs[r2] > LONG_MAX) {
6689 					*flags |= CPU_DTRACE_ILLOP;
6690 					break;
6691 				}
6692 
6693 				tupregs[ttop].dttk_size = regs[r2];
6694 			}
6695 
6696 			tupregs[ttop++].dttk_value = regs[rd];
6697 			break;
6698 
6699 		case DIF_OP_PUSHTV:
6700 			if (ttop == DIF_DTR_NREGS) {
6701 				*flags |= CPU_DTRACE_TUPOFLOW;
6702 				break;
6703 			}
6704 
6705 			tupregs[ttop].dttk_value = regs[rd];
6706 			tupregs[ttop++].dttk_size = 0;
6707 			break;
6708 
6709 		case DIF_OP_POPTS:
6710 			if (ttop != 0)
6711 				ttop--;
6712 			break;
6713 
6714 		case DIF_OP_FLUSHTS:
6715 			ttop = 0;
6716 			break;
6717 
6718 		case DIF_OP_LDGAA:
6719 		case DIF_OP_LDTAA: {
6720 			dtrace_dynvar_t *dvar;
6721 			dtrace_key_t *key = tupregs;
6722 			uint_t nkeys = ttop;
6723 
6724 			id = DIF_INSTR_VAR(instr);
6725 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
6726 			id -= DIF_VAR_OTHER_UBASE;
6727 
6728 			key[nkeys].dttk_value = (uint64_t)id;
6729 			key[nkeys++].dttk_size = 0;
6730 
6731 			if (DIF_INSTR_OP(instr) == DIF_OP_LDTAA) {
6732 				DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
6733 				key[nkeys++].dttk_size = 0;
6734 				VERIFY(id < vstate->dtvs_ntlocals);
6735 				v = &vstate->dtvs_tlocals[id];
6736 			} else {
6737 				VERIFY(id < vstate->dtvs_nglobals);
6738 				v = &vstate->dtvs_globals[id]->dtsv_var;
6739 			}
6740 
6741 			dvar = dtrace_dynvar(dstate, nkeys, key,
6742 			    v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
6743 			    v->dtdv_type.dtdt_size : sizeof (uint64_t),
6744 			    DTRACE_DYNVAR_NOALLOC, mstate, vstate);
6745 
6746 			if (dvar == NULL) {
6747 				regs[rd] = 0;
6748 				break;
6749 			}
6750 
6751 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6752 				regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
6753 			} else {
6754 				regs[rd] = *((uint64_t *)dvar->dtdv_data);
6755 			}
6756 
6757 			break;
6758 		}
6759 
6760 		case DIF_OP_STGAA:
6761 		case DIF_OP_STTAA: {
6762 			dtrace_dynvar_t *dvar;
6763 			dtrace_key_t *key = tupregs;
6764 			uint_t nkeys = ttop;
6765 
6766 			id = DIF_INSTR_VAR(instr);
6767 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
6768 			id -= DIF_VAR_OTHER_UBASE;
6769 
6770 			key[nkeys].dttk_value = (uint64_t)id;
6771 			key[nkeys++].dttk_size = 0;
6772 
6773 			if (DIF_INSTR_OP(instr) == DIF_OP_STTAA) {
6774 				DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
6775 				key[nkeys++].dttk_size = 0;
6776 				VERIFY(id < vstate->dtvs_ntlocals);
6777 				v = &vstate->dtvs_tlocals[id];
6778 			} else {
6779 				VERIFY(id < vstate->dtvs_nglobals);
6780 				v = &vstate->dtvs_globals[id]->dtsv_var;
6781 			}
6782 
6783 			dvar = dtrace_dynvar(dstate, nkeys, key,
6784 			    v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
6785 			    v->dtdv_type.dtdt_size : sizeof (uint64_t),
6786 			    regs[rd] ? DTRACE_DYNVAR_ALLOC :
6787 			    DTRACE_DYNVAR_DEALLOC, mstate, vstate);
6788 
6789 			if (dvar == NULL)
6790 				break;
6791 
6792 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6793 				size_t lim;
6794 
6795 				if (!dtrace_vcanload(
6796 				    (void *)(uintptr_t)regs[rd], &v->dtdv_type,
6797 				    &lim, mstate, vstate))
6798 					break;
6799 
6800 				dtrace_vcopy((void *)(uintptr_t)regs[rd],
6801 				    dvar->dtdv_data, &v->dtdv_type, lim);
6802 			} else {
6803 				*((uint64_t *)dvar->dtdv_data) = regs[rd];
6804 			}
6805 
6806 			break;
6807 		}
6808 
6809 		case DIF_OP_ALLOCS: {
6810 			uintptr_t ptr = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
6811 			size_t size = ptr - mstate->dtms_scratch_ptr + regs[r1];
6812 
6813 			/*
6814 			 * Rounding up the user allocation size could have
6815 			 * overflowed large, bogus allocations (like -1ULL) to
6816 			 * 0.
6817 			 */
6818 			if (size < regs[r1] ||
6819 			    !DTRACE_INSCRATCH(mstate, size)) {
6820 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
6821 				regs[rd] = 0;
6822 				break;
6823 			}
6824 
6825 			dtrace_bzero((void *) mstate->dtms_scratch_ptr, size);
6826 			mstate->dtms_scratch_ptr += size;
6827 			regs[rd] = ptr;
6828 			break;
6829 		}
6830 
6831 		case DIF_OP_COPYS:
6832 			if (!dtrace_canstore(regs[rd], regs[r2],
6833 			    mstate, vstate)) {
6834 				*flags |= CPU_DTRACE_BADADDR;
6835 				*illval = regs[rd];
6836 				break;
6837 			}
6838 
6839 			if (!dtrace_canload(regs[r1], regs[r2], mstate, vstate))
6840 				break;
6841 
6842 			dtrace_bcopy((void *)(uintptr_t)regs[r1],
6843 			    (void *)(uintptr_t)regs[rd], (size_t)regs[r2]);
6844 			break;
6845 
6846 		case DIF_OP_STB:
6847 			if (!dtrace_canstore(regs[rd], 1, mstate, vstate)) {
6848 				*flags |= CPU_DTRACE_BADADDR;
6849 				*illval = regs[rd];
6850 				break;
6851 			}
6852 			*((uint8_t *)(uintptr_t)regs[rd]) = (uint8_t)regs[r1];
6853 			break;
6854 
6855 		case DIF_OP_STH:
6856 			if (!dtrace_canstore(regs[rd], 2, mstate, vstate)) {
6857 				*flags |= CPU_DTRACE_BADADDR;
6858 				*illval = regs[rd];
6859 				break;
6860 			}
6861 			if (regs[rd] & 1) {
6862 				*flags |= CPU_DTRACE_BADALIGN;
6863 				*illval = regs[rd];
6864 				break;
6865 			}
6866 			*((uint16_t *)(uintptr_t)regs[rd]) = (uint16_t)regs[r1];
6867 			break;
6868 
6869 		case DIF_OP_STW:
6870 			if (!dtrace_canstore(regs[rd], 4, mstate, vstate)) {
6871 				*flags |= CPU_DTRACE_BADADDR;
6872 				*illval = regs[rd];
6873 				break;
6874 			}
6875 			if (regs[rd] & 3) {
6876 				*flags |= CPU_DTRACE_BADALIGN;
6877 				*illval = regs[rd];
6878 				break;
6879 			}
6880 			*((uint32_t *)(uintptr_t)regs[rd]) = (uint32_t)regs[r1];
6881 			break;
6882 
6883 		case DIF_OP_STX:
6884 			if (!dtrace_canstore(regs[rd], 8, mstate, vstate)) {
6885 				*flags |= CPU_DTRACE_BADADDR;
6886 				*illval = regs[rd];
6887 				break;
6888 			}
6889 			if (regs[rd] & 7) {
6890 				*flags |= CPU_DTRACE_BADALIGN;
6891 				*illval = regs[rd];
6892 				break;
6893 			}
6894 			*((uint64_t *)(uintptr_t)regs[rd]) = regs[r1];
6895 			break;
6896 		}
6897 	}
6898 
6899 	if (!(*flags & CPU_DTRACE_FAULT))
6900 		return (rval);
6901 
6902 	mstate->dtms_fltoffs = opc * sizeof (dif_instr_t);
6903 	mstate->dtms_present |= DTRACE_MSTATE_FLTOFFS;
6904 
6905 	return (0);
6906 }
6907 
6908 static void
6909 dtrace_action_breakpoint(dtrace_ecb_t *ecb)
6910 {
6911 	dtrace_probe_t *probe = ecb->dte_probe;
6912 	dtrace_provider_t *prov = probe->dtpr_provider;
6913 	char c[DTRACE_FULLNAMELEN + 80], *str;
6914 	char *msg = "dtrace: breakpoint action at probe ";
6915 	char *ecbmsg = " (ecb ";
6916 	uintptr_t mask = (0xf << (sizeof (uintptr_t) * NBBY / 4));
6917 	uintptr_t val = (uintptr_t)ecb;
6918 	int shift = (sizeof (uintptr_t) * NBBY) - 4, i = 0;
6919 
6920 	if (dtrace_destructive_disallow)
6921 		return;
6922 
6923 	/*
6924 	 * It's impossible to be taking action on the NULL probe.
6925 	 */
6926 	ASSERT(probe != NULL);
6927 
6928 	/*
6929 	 * This is a poor man's (destitute man's?) sprintf():  we want to
6930 	 * print the provider name, module name, function name and name of
6931 	 * the probe, along with the hex address of the ECB with the breakpoint
6932 	 * action -- all of which we must place in the character buffer by
6933 	 * hand.
6934 	 */
6935 	while (*msg != '\0')
6936 		c[i++] = *msg++;
6937 
6938 	for (str = prov->dtpv_name; *str != '\0'; str++)
6939 		c[i++] = *str;
6940 	c[i++] = ':';
6941 
6942 	for (str = probe->dtpr_mod; *str != '\0'; str++)
6943 		c[i++] = *str;
6944 	c[i++] = ':';
6945 
6946 	for (str = probe->dtpr_func; *str != '\0'; str++)
6947 		c[i++] = *str;
6948 	c[i++] = ':';
6949 
6950 	for (str = probe->dtpr_name; *str != '\0'; str++)
6951 		c[i++] = *str;
6952 
6953 	while (*ecbmsg != '\0')
6954 		c[i++] = *ecbmsg++;
6955 
6956 	while (shift >= 0) {
6957 		mask = (uintptr_t)0xf << shift;
6958 
6959 		if (val >= ((uintptr_t)1 << shift))
6960 			c[i++] = "0123456789abcdef"[(val & mask) >> shift];
6961 		shift -= 4;
6962 	}
6963 
6964 	c[i++] = ')';
6965 	c[i] = '\0';
6966 
6967 #ifdef illumos
6968 	debug_enter(c);
6969 #else
6970 	kdb_enter(KDB_WHY_DTRACE, "breakpoint action");
6971 #endif
6972 }
6973 
6974 static void
6975 dtrace_action_panic(dtrace_ecb_t *ecb)
6976 {
6977 	dtrace_probe_t *probe = ecb->dte_probe;
6978 
6979 	/*
6980 	 * It's impossible to be taking action on the NULL probe.
6981 	 */
6982 	ASSERT(probe != NULL);
6983 
6984 	if (dtrace_destructive_disallow)
6985 		return;
6986 
6987 	if (dtrace_panicked != NULL)
6988 		return;
6989 
6990 	if (dtrace_casptr(&dtrace_panicked, NULL, curthread) != NULL)
6991 		return;
6992 
6993 	/*
6994 	 * We won the right to panic.  (We want to be sure that only one
6995 	 * thread calls panic() from dtrace_probe(), and that panic() is
6996 	 * called exactly once.)
6997 	 */
6998 	dtrace_panic("dtrace: panic action at probe %s:%s:%s:%s (ecb %p)",
6999 	    probe->dtpr_provider->dtpv_name, probe->dtpr_mod,
7000 	    probe->dtpr_func, probe->dtpr_name, (void *)ecb);
7001 }
7002 
7003 static void
7004 dtrace_action_raise(uint64_t sig)
7005 {
7006 	if (dtrace_destructive_disallow)
7007 		return;
7008 
7009 	if (sig >= NSIG) {
7010 		DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
7011 		return;
7012 	}
7013 
7014 #ifdef illumos
7015 	/*
7016 	 * raise() has a queue depth of 1 -- we ignore all subsequent
7017 	 * invocations of the raise() action.
7018 	 */
7019 	if (curthread->t_dtrace_sig == 0)
7020 		curthread->t_dtrace_sig = (uint8_t)sig;
7021 
7022 	curthread->t_sig_check = 1;
7023 	aston(curthread);
7024 #else
7025 	struct proc *p = curproc;
7026 	PROC_LOCK(p);
7027 	kern_psignal(p, sig);
7028 	PROC_UNLOCK(p);
7029 #endif
7030 }
7031 
7032 static void
7033 dtrace_action_stop(void)
7034 {
7035 	if (dtrace_destructive_disallow)
7036 		return;
7037 
7038 #ifdef illumos
7039 	if (!curthread->t_dtrace_stop) {
7040 		curthread->t_dtrace_stop = 1;
7041 		curthread->t_sig_check = 1;
7042 		aston(curthread);
7043 	}
7044 #else
7045 	struct proc *p = curproc;
7046 	PROC_LOCK(p);
7047 	kern_psignal(p, SIGSTOP);
7048 	PROC_UNLOCK(p);
7049 #endif
7050 }
7051 
7052 static void
7053 dtrace_action_chill(dtrace_mstate_t *mstate, hrtime_t val)
7054 {
7055 	hrtime_t now;
7056 	volatile uint16_t *flags;
7057 #ifdef illumos
7058 	cpu_t *cpu = CPU;
7059 #else
7060 	cpu_t *cpu = &solaris_cpu[curcpu];
7061 #endif
7062 
7063 	if (dtrace_destructive_disallow)
7064 		return;
7065 
7066 	flags = (volatile uint16_t *)&cpu_core[curcpu].cpuc_dtrace_flags;
7067 
7068 	now = dtrace_gethrtime();
7069 
7070 	if (now - cpu->cpu_dtrace_chillmark > dtrace_chill_interval) {
7071 		/*
7072 		 * We need to advance the mark to the current time.
7073 		 */
7074 		cpu->cpu_dtrace_chillmark = now;
7075 		cpu->cpu_dtrace_chilled = 0;
7076 	}
7077 
7078 	/*
7079 	 * Now check to see if the requested chill time would take us over
7080 	 * the maximum amount of time allowed in the chill interval.  (Or
7081 	 * worse, if the calculation itself induces overflow.)
7082 	 */
7083 	if (cpu->cpu_dtrace_chilled + val > dtrace_chill_max ||
7084 	    cpu->cpu_dtrace_chilled + val < cpu->cpu_dtrace_chilled) {
7085 		*flags |= CPU_DTRACE_ILLOP;
7086 		return;
7087 	}
7088 
7089 	while (dtrace_gethrtime() - now < val)
7090 		continue;
7091 
7092 	/*
7093 	 * Normally, we assure that the value of the variable "timestamp" does
7094 	 * not change within an ECB.  The presence of chill() represents an
7095 	 * exception to this rule, however.
7096 	 */
7097 	mstate->dtms_present &= ~DTRACE_MSTATE_TIMESTAMP;
7098 	cpu->cpu_dtrace_chilled += val;
7099 }
7100 
7101 static void
7102 dtrace_action_ustack(dtrace_mstate_t *mstate, dtrace_state_t *state,
7103     uint64_t *buf, uint64_t arg)
7104 {
7105 	int nframes = DTRACE_USTACK_NFRAMES(arg);
7106 	int strsize = DTRACE_USTACK_STRSIZE(arg);
7107 	uint64_t *pcs = &buf[1], *fps;
7108 	char *str = (char *)&pcs[nframes];
7109 	int size, offs = 0, i, j;
7110 	size_t rem;
7111 	uintptr_t old = mstate->dtms_scratch_ptr, saved;
7112 	uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
7113 	char *sym;
7114 
7115 	/*
7116 	 * Should be taking a faster path if string space has not been
7117 	 * allocated.
7118 	 */
7119 	ASSERT(strsize != 0);
7120 
7121 	/*
7122 	 * We will first allocate some temporary space for the frame pointers.
7123 	 */
7124 	fps = (uint64_t *)P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
7125 	size = (uintptr_t)fps - mstate->dtms_scratch_ptr +
7126 	    (nframes * sizeof (uint64_t));
7127 
7128 	if (!DTRACE_INSCRATCH(mstate, size)) {
7129 		/*
7130 		 * Not enough room for our frame pointers -- need to indicate
7131 		 * that we ran out of scratch space.
7132 		 */
7133 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
7134 		return;
7135 	}
7136 
7137 	mstate->dtms_scratch_ptr += size;
7138 	saved = mstate->dtms_scratch_ptr;
7139 
7140 	/*
7141 	 * Now get a stack with both program counters and frame pointers.
7142 	 */
7143 	DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
7144 	dtrace_getufpstack(buf, fps, nframes + 1);
7145 	DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
7146 
7147 	/*
7148 	 * If that faulted, we're cooked.
7149 	 */
7150 	if (*flags & CPU_DTRACE_FAULT)
7151 		goto out;
7152 
7153 	/*
7154 	 * Now we want to walk up the stack, calling the USTACK helper.  For
7155 	 * each iteration, we restore the scratch pointer.
7156 	 */
7157 	for (i = 0; i < nframes; i++) {
7158 		mstate->dtms_scratch_ptr = saved;
7159 
7160 		if (offs >= strsize)
7161 			break;
7162 
7163 		sym = (char *)(uintptr_t)dtrace_helper(
7164 		    DTRACE_HELPER_ACTION_USTACK,
7165 		    mstate, state, pcs[i], fps[i]);
7166 
7167 		/*
7168 		 * If we faulted while running the helper, we're going to
7169 		 * clear the fault and null out the corresponding string.
7170 		 */
7171 		if (*flags & CPU_DTRACE_FAULT) {
7172 			*flags &= ~CPU_DTRACE_FAULT;
7173 			str[offs++] = '\0';
7174 			continue;
7175 		}
7176 
7177 		if (sym == NULL) {
7178 			str[offs++] = '\0';
7179 			continue;
7180 		}
7181 
7182 		if (!dtrace_strcanload((uintptr_t)sym, strsize, &rem, mstate,
7183 		    &(state->dts_vstate))) {
7184 			str[offs++] = '\0';
7185 			continue;
7186 		}
7187 
7188 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
7189 
7190 		/*
7191 		 * Now copy in the string that the helper returned to us.
7192 		 */
7193 		for (j = 0; offs + j < strsize && j < rem; j++) {
7194 			if ((str[offs + j] = sym[j]) == '\0')
7195 				break;
7196 		}
7197 
7198 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
7199 
7200 		offs += j + 1;
7201 	}
7202 
7203 	if (offs >= strsize) {
7204 		/*
7205 		 * If we didn't have room for all of the strings, we don't
7206 		 * abort processing -- this needn't be a fatal error -- but we
7207 		 * still want to increment a counter (dts_stkstroverflows) to
7208 		 * allow this condition to be warned about.  (If this is from
7209 		 * a jstack() action, it is easily tuned via jstackstrsize.)
7210 		 */
7211 		dtrace_error(&state->dts_stkstroverflows);
7212 	}
7213 
7214 	while (offs < strsize)
7215 		str[offs++] = '\0';
7216 
7217 out:
7218 	mstate->dtms_scratch_ptr = old;
7219 }
7220 
7221 static void
7222 dtrace_store_by_ref(dtrace_difo_t *dp, caddr_t tomax, size_t size,
7223     size_t *valoffsp, uint64_t *valp, uint64_t end, int intuple, int dtkind)
7224 {
7225 	volatile uint16_t *flags;
7226 	uint64_t val = *valp;
7227 	size_t valoffs = *valoffsp;
7228 
7229 	flags = (volatile uint16_t *)&cpu_core[curcpu].cpuc_dtrace_flags;
7230 	ASSERT(dtkind == DIF_TF_BYREF || dtkind == DIF_TF_BYUREF);
7231 
7232 	/*
7233 	 * If this is a string, we're going to only load until we find the zero
7234 	 * byte -- after which we'll store zero bytes.
7235 	 */
7236 	if (dp->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING) {
7237 		char c = '\0' + 1;
7238 		size_t s;
7239 
7240 		for (s = 0; s < size; s++) {
7241 			if (c != '\0' && dtkind == DIF_TF_BYREF) {
7242 				c = dtrace_load8(val++);
7243 			} else if (c != '\0' && dtkind == DIF_TF_BYUREF) {
7244 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
7245 				c = dtrace_fuword8((void *)(uintptr_t)val++);
7246 				DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
7247 				if (*flags & CPU_DTRACE_FAULT)
7248 					break;
7249 			}
7250 
7251 			DTRACE_STORE(uint8_t, tomax, valoffs++, c);
7252 
7253 			if (c == '\0' && intuple)
7254 				break;
7255 		}
7256 	} else {
7257 		uint8_t c;
7258 		while (valoffs < end) {
7259 			if (dtkind == DIF_TF_BYREF) {
7260 				c = dtrace_load8(val++);
7261 			} else if (dtkind == DIF_TF_BYUREF) {
7262 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
7263 				c = dtrace_fuword8((void *)(uintptr_t)val++);
7264 				DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
7265 				if (*flags & CPU_DTRACE_FAULT)
7266 					break;
7267 			}
7268 
7269 			DTRACE_STORE(uint8_t, tomax,
7270 			    valoffs++, c);
7271 		}
7272 	}
7273 
7274 	*valp = val;
7275 	*valoffsp = valoffs;
7276 }
7277 
7278 /*
7279  * Disables interrupts and sets the per-thread inprobe flag. When DEBUG is
7280  * defined, we also assert that we are not recursing unless the probe ID is an
7281  * error probe.
7282  */
7283 static dtrace_icookie_t
7284 dtrace_probe_enter(dtrace_id_t id)
7285 {
7286 	dtrace_icookie_t cookie;
7287 
7288 	cookie = dtrace_interrupt_disable();
7289 
7290 	/*
7291 	 * Unless this is an ERROR probe, we are not allowed to recurse in
7292 	 * dtrace_probe(). Recursing into DTrace probe usually means that a
7293 	 * function is instrumented that should not have been instrumented or
7294 	 * that the ordering guarantee of the records will be violated,
7295 	 * resulting in unexpected output. If there is an exception to this
7296 	 * assertion, a new case should be added.
7297 	 */
7298 	ASSERT(curthread->t_dtrace_inprobe == 0 ||
7299 	    id == dtrace_probeid_error);
7300 	curthread->t_dtrace_inprobe = 1;
7301 
7302 	return (cookie);
7303 }
7304 
7305 /*
7306  * Clears the per-thread inprobe flag and enables interrupts.
7307  */
7308 static void
7309 dtrace_probe_exit(dtrace_icookie_t cookie)
7310 {
7311 
7312 	curthread->t_dtrace_inprobe = 0;
7313 	dtrace_interrupt_enable(cookie);
7314 }
7315 
7316 /*
7317  * If you're looking for the epicenter of DTrace, you just found it.  This
7318  * is the function called by the provider to fire a probe -- from which all
7319  * subsequent probe-context DTrace activity emanates.
7320  */
7321 void
7322 dtrace_probe(dtrace_id_t id, uintptr_t arg0, uintptr_t arg1,
7323     uintptr_t arg2, uintptr_t arg3, uintptr_t arg4)
7324 {
7325 	processorid_t cpuid;
7326 	dtrace_icookie_t cookie;
7327 	dtrace_probe_t *probe;
7328 	dtrace_mstate_t mstate;
7329 	dtrace_ecb_t *ecb;
7330 	dtrace_action_t *act;
7331 	intptr_t offs;
7332 	size_t size;
7333 	int vtime, onintr;
7334 	volatile uint16_t *flags;
7335 	hrtime_t now;
7336 
7337 	if (panicstr != NULL)
7338 		return;
7339 
7340 #ifdef illumos
7341 	/*
7342 	 * Kick out immediately if this CPU is still being born (in which case
7343 	 * curthread will be set to -1) or the current thread can't allow
7344 	 * probes in its current context.
7345 	 */
7346 	if (((uintptr_t)curthread & 1) || (curthread->t_flag & T_DONTDTRACE))
7347 		return;
7348 #endif
7349 
7350 	cookie = dtrace_probe_enter(id);
7351 	probe = dtrace_probes[id - 1];
7352 	cpuid = curcpu;
7353 	onintr = CPU_ON_INTR(CPU);
7354 
7355 	if (!onintr && probe->dtpr_predcache != DTRACE_CACHEIDNONE &&
7356 	    probe->dtpr_predcache == curthread->t_predcache) {
7357 		/*
7358 		 * We have hit in the predicate cache; we know that
7359 		 * this predicate would evaluate to be false.
7360 		 */
7361 		dtrace_probe_exit(cookie);
7362 		return;
7363 	}
7364 
7365 #ifdef illumos
7366 	if (panic_quiesce) {
7367 #else
7368 	if (panicstr != NULL) {
7369 #endif
7370 		/*
7371 		 * We don't trace anything if we're panicking.
7372 		 */
7373 		dtrace_probe_exit(cookie);
7374 		return;
7375 	}
7376 
7377 	now = mstate.dtms_timestamp = dtrace_gethrtime();
7378 	mstate.dtms_present = DTRACE_MSTATE_TIMESTAMP;
7379 	vtime = dtrace_vtime_references != 0;
7380 
7381 	if (vtime && curthread->t_dtrace_start)
7382 		curthread->t_dtrace_vtime += now - curthread->t_dtrace_start;
7383 
7384 	mstate.dtms_difo = NULL;
7385 	mstate.dtms_probe = probe;
7386 	mstate.dtms_strtok = 0;
7387 	mstate.dtms_arg[0] = arg0;
7388 	mstate.dtms_arg[1] = arg1;
7389 	mstate.dtms_arg[2] = arg2;
7390 	mstate.dtms_arg[3] = arg3;
7391 	mstate.dtms_arg[4] = arg4;
7392 
7393 	flags = (volatile uint16_t *)&cpu_core[cpuid].cpuc_dtrace_flags;
7394 
7395 	for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
7396 		dtrace_predicate_t *pred = ecb->dte_predicate;
7397 		dtrace_state_t *state = ecb->dte_state;
7398 		dtrace_buffer_t *buf = &state->dts_buffer[cpuid];
7399 		dtrace_buffer_t *aggbuf = &state->dts_aggbuffer[cpuid];
7400 		dtrace_vstate_t *vstate = &state->dts_vstate;
7401 		dtrace_provider_t *prov = probe->dtpr_provider;
7402 		uint64_t tracememsize = 0;
7403 		int committed = 0;
7404 		caddr_t tomax;
7405 
7406 		/*
7407 		 * A little subtlety with the following (seemingly innocuous)
7408 		 * declaration of the automatic 'val':  by looking at the
7409 		 * code, you might think that it could be declared in the
7410 		 * action processing loop, below.  (That is, it's only used in
7411 		 * the action processing loop.)  However, it must be declared
7412 		 * out of that scope because in the case of DIF expression
7413 		 * arguments to aggregating actions, one iteration of the
7414 		 * action loop will use the last iteration's value.
7415 		 */
7416 		uint64_t val = 0;
7417 
7418 		mstate.dtms_present = DTRACE_MSTATE_ARGS | DTRACE_MSTATE_PROBE;
7419 		mstate.dtms_getf = NULL;
7420 
7421 		*flags &= ~CPU_DTRACE_ERROR;
7422 
7423 		if (prov == dtrace_provider) {
7424 			/*
7425 			 * If dtrace itself is the provider of this probe,
7426 			 * we're only going to continue processing the ECB if
7427 			 * arg0 (the dtrace_state_t) is equal to the ECB's
7428 			 * creating state.  (This prevents disjoint consumers
7429 			 * from seeing one another's metaprobes.)
7430 			 */
7431 			if (arg0 != (uint64_t)(uintptr_t)state)
7432 				continue;
7433 		}
7434 
7435 		if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE) {
7436 			/*
7437 			 * We're not currently active.  If our provider isn't
7438 			 * the dtrace pseudo provider, we're not interested.
7439 			 */
7440 			if (prov != dtrace_provider)
7441 				continue;
7442 
7443 			/*
7444 			 * Now we must further check if we are in the BEGIN
7445 			 * probe.  If we are, we will only continue processing
7446 			 * if we're still in WARMUP -- if one BEGIN enabling
7447 			 * has invoked the exit() action, we don't want to
7448 			 * evaluate subsequent BEGIN enablings.
7449 			 */
7450 			if (probe->dtpr_id == dtrace_probeid_begin &&
7451 			    state->dts_activity != DTRACE_ACTIVITY_WARMUP) {
7452 				ASSERT(state->dts_activity ==
7453 				    DTRACE_ACTIVITY_DRAINING);
7454 				continue;
7455 			}
7456 		}
7457 
7458 		if (ecb->dte_cond) {
7459 			/*
7460 			 * If the dte_cond bits indicate that this
7461 			 * consumer is only allowed to see user-mode firings
7462 			 * of this probe, call the provider's dtps_usermode()
7463 			 * entry point to check that the probe was fired
7464 			 * while in a user context. Skip this ECB if that's
7465 			 * not the case.
7466 			 */
7467 			if ((ecb->dte_cond & DTRACE_COND_USERMODE) &&
7468 			    prov->dtpv_pops.dtps_usermode(prov->dtpv_arg,
7469 			    probe->dtpr_id, probe->dtpr_arg) == 0)
7470 				continue;
7471 
7472 #ifdef illumos
7473 			/*
7474 			 * This is more subtle than it looks. We have to be
7475 			 * absolutely certain that CRED() isn't going to
7476 			 * change out from under us so it's only legit to
7477 			 * examine that structure if we're in constrained
7478 			 * situations. Currently, the only times we'll this
7479 			 * check is if a non-super-user has enabled the
7480 			 * profile or syscall providers -- providers that
7481 			 * allow visibility of all processes. For the
7482 			 * profile case, the check above will ensure that
7483 			 * we're examining a user context.
7484 			 */
7485 			if (ecb->dte_cond & DTRACE_COND_OWNER) {
7486 				cred_t *cr;
7487 				cred_t *s_cr =
7488 				    ecb->dte_state->dts_cred.dcr_cred;
7489 				proc_t *proc;
7490 
7491 				ASSERT(s_cr != NULL);
7492 
7493 				if ((cr = CRED()) == NULL ||
7494 				    s_cr->cr_uid != cr->cr_uid ||
7495 				    s_cr->cr_uid != cr->cr_ruid ||
7496 				    s_cr->cr_uid != cr->cr_suid ||
7497 				    s_cr->cr_gid != cr->cr_gid ||
7498 				    s_cr->cr_gid != cr->cr_rgid ||
7499 				    s_cr->cr_gid != cr->cr_sgid ||
7500 				    (proc = ttoproc(curthread)) == NULL ||
7501 				    (proc->p_flag & SNOCD))
7502 					continue;
7503 			}
7504 
7505 			if (ecb->dte_cond & DTRACE_COND_ZONEOWNER) {
7506 				cred_t *cr;
7507 				cred_t *s_cr =
7508 				    ecb->dte_state->dts_cred.dcr_cred;
7509 
7510 				ASSERT(s_cr != NULL);
7511 
7512 				if ((cr = CRED()) == NULL ||
7513 				    s_cr->cr_zone->zone_id !=
7514 				    cr->cr_zone->zone_id)
7515 					continue;
7516 			}
7517 #endif
7518 		}
7519 
7520 		if (now - state->dts_alive > dtrace_deadman_timeout) {
7521 			/*
7522 			 * We seem to be dead.  Unless we (a) have kernel
7523 			 * destructive permissions (b) have explicitly enabled
7524 			 * destructive actions and (c) destructive actions have
7525 			 * not been disabled, we're going to transition into
7526 			 * the KILLED state, from which no further processing
7527 			 * on this state will be performed.
7528 			 */
7529 			if (!dtrace_priv_kernel_destructive(state) ||
7530 			    !state->dts_cred.dcr_destructive ||
7531 			    dtrace_destructive_disallow) {
7532 				void *activity = &state->dts_activity;
7533 				dtrace_activity_t curstate;
7534 
7535 				do {
7536 					curstate = state->dts_activity;
7537 				} while (dtrace_cas32(activity, curstate,
7538 				    DTRACE_ACTIVITY_KILLED) != curstate);
7539 
7540 				continue;
7541 			}
7542 		}
7543 
7544 		if ((offs = dtrace_buffer_reserve(buf, ecb->dte_needed,
7545 		    ecb->dte_alignment, state, &mstate)) < 0)
7546 			continue;
7547 
7548 		tomax = buf->dtb_tomax;
7549 		ASSERT(tomax != NULL);
7550 
7551 		if (ecb->dte_size != 0) {
7552 			dtrace_rechdr_t dtrh;
7553 			if (!(mstate.dtms_present & DTRACE_MSTATE_TIMESTAMP)) {
7554 				mstate.dtms_timestamp = dtrace_gethrtime();
7555 				mstate.dtms_present |= DTRACE_MSTATE_TIMESTAMP;
7556 			}
7557 			ASSERT3U(ecb->dte_size, >=, sizeof (dtrace_rechdr_t));
7558 			dtrh.dtrh_epid = ecb->dte_epid;
7559 			DTRACE_RECORD_STORE_TIMESTAMP(&dtrh,
7560 			    mstate.dtms_timestamp);
7561 			*((dtrace_rechdr_t *)(tomax + offs)) = dtrh;
7562 		}
7563 
7564 		mstate.dtms_epid = ecb->dte_epid;
7565 		mstate.dtms_present |= DTRACE_MSTATE_EPID;
7566 
7567 		if (state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)
7568 			mstate.dtms_access = DTRACE_ACCESS_KERNEL;
7569 		else
7570 			mstate.dtms_access = 0;
7571 
7572 		if (pred != NULL) {
7573 			dtrace_difo_t *dp = pred->dtp_difo;
7574 			uint64_t rval;
7575 
7576 			rval = dtrace_dif_emulate(dp, &mstate, vstate, state);
7577 
7578 			if (!(*flags & CPU_DTRACE_ERROR) && !rval) {
7579 				dtrace_cacheid_t cid = probe->dtpr_predcache;
7580 
7581 				if (cid != DTRACE_CACHEIDNONE && !onintr) {
7582 					/*
7583 					 * Update the predicate cache...
7584 					 */
7585 					ASSERT(cid == pred->dtp_cacheid);
7586 					curthread->t_predcache = cid;
7587 				}
7588 
7589 				continue;
7590 			}
7591 		}
7592 
7593 		for (act = ecb->dte_action; !(*flags & CPU_DTRACE_ERROR) &&
7594 		    act != NULL; act = act->dta_next) {
7595 			size_t valoffs;
7596 			dtrace_difo_t *dp;
7597 			dtrace_recdesc_t *rec = &act->dta_rec;
7598 
7599 			size = rec->dtrd_size;
7600 			valoffs = offs + rec->dtrd_offset;
7601 
7602 			if (DTRACEACT_ISAGG(act->dta_kind)) {
7603 				uint64_t v = 0xbad;
7604 				dtrace_aggregation_t *agg;
7605 
7606 				agg = (dtrace_aggregation_t *)act;
7607 
7608 				if ((dp = act->dta_difo) != NULL)
7609 					v = dtrace_dif_emulate(dp,
7610 					    &mstate, vstate, state);
7611 
7612 				if (*flags & CPU_DTRACE_ERROR)
7613 					continue;
7614 
7615 				/*
7616 				 * Note that we always pass the expression
7617 				 * value from the previous iteration of the
7618 				 * action loop.  This value will only be used
7619 				 * if there is an expression argument to the
7620 				 * aggregating action, denoted by the
7621 				 * dtag_hasarg field.
7622 				 */
7623 				dtrace_aggregate(agg, buf,
7624 				    offs, aggbuf, v, val);
7625 				continue;
7626 			}
7627 
7628 			switch (act->dta_kind) {
7629 			case DTRACEACT_STOP:
7630 				if (dtrace_priv_proc_destructive(state))
7631 					dtrace_action_stop();
7632 				continue;
7633 
7634 			case DTRACEACT_BREAKPOINT:
7635 				if (dtrace_priv_kernel_destructive(state))
7636 					dtrace_action_breakpoint(ecb);
7637 				continue;
7638 
7639 			case DTRACEACT_PANIC:
7640 				if (dtrace_priv_kernel_destructive(state))
7641 					dtrace_action_panic(ecb);
7642 				continue;
7643 
7644 			case DTRACEACT_STACK:
7645 				if (!dtrace_priv_kernel(state))
7646 					continue;
7647 
7648 				dtrace_getpcstack((pc_t *)(tomax + valoffs),
7649 				    size / sizeof (pc_t), probe->dtpr_aframes,
7650 				    DTRACE_ANCHORED(probe) ? NULL :
7651 				    (uint32_t *)arg0);
7652 				continue;
7653 
7654 			case DTRACEACT_JSTACK:
7655 			case DTRACEACT_USTACK:
7656 				if (!dtrace_priv_proc(state))
7657 					continue;
7658 
7659 				/*
7660 				 * See comment in DIF_VAR_PID.
7661 				 */
7662 				if (DTRACE_ANCHORED(mstate.dtms_probe) &&
7663 				    CPU_ON_INTR(CPU)) {
7664 					int depth = DTRACE_USTACK_NFRAMES(
7665 					    rec->dtrd_arg) + 1;
7666 
7667 					dtrace_bzero((void *)(tomax + valoffs),
7668 					    DTRACE_USTACK_STRSIZE(rec->dtrd_arg)
7669 					    + depth * sizeof (uint64_t));
7670 
7671 					continue;
7672 				}
7673 
7674 				if (DTRACE_USTACK_STRSIZE(rec->dtrd_arg) != 0 &&
7675 				    curproc->p_dtrace_helpers != NULL) {
7676 					/*
7677 					 * This is the slow path -- we have
7678 					 * allocated string space, and we're
7679 					 * getting the stack of a process that
7680 					 * has helpers.  Call into a separate
7681 					 * routine to perform this processing.
7682 					 */
7683 					dtrace_action_ustack(&mstate, state,
7684 					    (uint64_t *)(tomax + valoffs),
7685 					    rec->dtrd_arg);
7686 					continue;
7687 				}
7688 
7689 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
7690 				dtrace_getupcstack((uint64_t *)
7691 				    (tomax + valoffs),
7692 				    DTRACE_USTACK_NFRAMES(rec->dtrd_arg) + 1);
7693 				DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
7694 				continue;
7695 
7696 			default:
7697 				break;
7698 			}
7699 
7700 			dp = act->dta_difo;
7701 			ASSERT(dp != NULL);
7702 
7703 			val = dtrace_dif_emulate(dp, &mstate, vstate, state);
7704 
7705 			if (*flags & CPU_DTRACE_ERROR)
7706 				continue;
7707 
7708 			switch (act->dta_kind) {
7709 			case DTRACEACT_SPECULATE: {
7710 				dtrace_rechdr_t *dtrh;
7711 
7712 				ASSERT(buf == &state->dts_buffer[cpuid]);
7713 				buf = dtrace_speculation_buffer(state,
7714 				    cpuid, val);
7715 
7716 				if (buf == NULL) {
7717 					*flags |= CPU_DTRACE_DROP;
7718 					continue;
7719 				}
7720 
7721 				offs = dtrace_buffer_reserve(buf,
7722 				    ecb->dte_needed, ecb->dte_alignment,
7723 				    state, NULL);
7724 
7725 				if (offs < 0) {
7726 					*flags |= CPU_DTRACE_DROP;
7727 					continue;
7728 				}
7729 
7730 				tomax = buf->dtb_tomax;
7731 				ASSERT(tomax != NULL);
7732 
7733 				if (ecb->dte_size == 0)
7734 					continue;
7735 
7736 				ASSERT3U(ecb->dte_size, >=,
7737 				    sizeof (dtrace_rechdr_t));
7738 				dtrh = ((void *)(tomax + offs));
7739 				dtrh->dtrh_epid = ecb->dte_epid;
7740 				/*
7741 				 * When the speculation is committed, all of
7742 				 * the records in the speculative buffer will
7743 				 * have their timestamps set to the commit
7744 				 * time.  Until then, it is set to a sentinel
7745 				 * value, for debugability.
7746 				 */
7747 				DTRACE_RECORD_STORE_TIMESTAMP(dtrh, UINT64_MAX);
7748 				continue;
7749 			}
7750 
7751 			case DTRACEACT_PRINTM: {
7752 				/* The DIF returns a 'memref'. */
7753 				uintptr_t *memref = (uintptr_t *)(uintptr_t) val;
7754 
7755 				/* Get the size from the memref. */
7756 				size = memref[1];
7757 
7758 				/*
7759 				 * Check if the size exceeds the allocated
7760 				 * buffer size.
7761 				 */
7762 				if (size + sizeof(uintptr_t) > dp->dtdo_rtype.dtdt_size) {
7763 					/* Flag a drop! */
7764 					*flags |= CPU_DTRACE_DROP;
7765 					continue;
7766 				}
7767 
7768 				/* Store the size in the buffer first. */
7769 				DTRACE_STORE(uintptr_t, tomax,
7770 				    valoffs, size);
7771 
7772 				/*
7773 				 * Offset the buffer address to the start
7774 				 * of the data.
7775 				 */
7776 				valoffs += sizeof(uintptr_t);
7777 
7778 				/*
7779 				 * Reset to the memory address rather than
7780 				 * the memref array, then let the BYREF
7781 				 * code below do the work to store the
7782 				 * memory data in the buffer.
7783 				 */
7784 				val = memref[0];
7785 				break;
7786 			}
7787 
7788 			case DTRACEACT_CHILL:
7789 				if (dtrace_priv_kernel_destructive(state))
7790 					dtrace_action_chill(&mstate, val);
7791 				continue;
7792 
7793 			case DTRACEACT_RAISE:
7794 				if (dtrace_priv_proc_destructive(state))
7795 					dtrace_action_raise(val);
7796 				continue;
7797 
7798 			case DTRACEACT_COMMIT:
7799 				ASSERT(!committed);
7800 
7801 				/*
7802 				 * We need to commit our buffer state.
7803 				 */
7804 				if (ecb->dte_size)
7805 					buf->dtb_offset = offs + ecb->dte_size;
7806 				buf = &state->dts_buffer[cpuid];
7807 				dtrace_speculation_commit(state, cpuid, val);
7808 				committed = 1;
7809 				continue;
7810 
7811 			case DTRACEACT_DISCARD:
7812 				dtrace_speculation_discard(state, cpuid, val);
7813 				continue;
7814 
7815 			case DTRACEACT_DIFEXPR:
7816 			case DTRACEACT_LIBACT:
7817 			case DTRACEACT_PRINTF:
7818 			case DTRACEACT_PRINTA:
7819 			case DTRACEACT_SYSTEM:
7820 			case DTRACEACT_FREOPEN:
7821 			case DTRACEACT_TRACEMEM:
7822 				break;
7823 
7824 			case DTRACEACT_TRACEMEM_DYNSIZE:
7825 				tracememsize = val;
7826 				break;
7827 
7828 			case DTRACEACT_SYM:
7829 			case DTRACEACT_MOD:
7830 				if (!dtrace_priv_kernel(state))
7831 					continue;
7832 				break;
7833 
7834 			case DTRACEACT_USYM:
7835 			case DTRACEACT_UMOD:
7836 			case DTRACEACT_UADDR: {
7837 #ifdef illumos
7838 				struct pid *pid = curthread->t_procp->p_pidp;
7839 #endif
7840 
7841 				if (!dtrace_priv_proc(state))
7842 					continue;
7843 
7844 				DTRACE_STORE(uint64_t, tomax,
7845 #ifdef illumos
7846 				    valoffs, (uint64_t)pid->pid_id);
7847 #else
7848 				    valoffs, (uint64_t) curproc->p_pid);
7849 #endif
7850 				DTRACE_STORE(uint64_t, tomax,
7851 				    valoffs + sizeof (uint64_t), val);
7852 
7853 				continue;
7854 			}
7855 
7856 			case DTRACEACT_EXIT: {
7857 				/*
7858 				 * For the exit action, we are going to attempt
7859 				 * to atomically set our activity to be
7860 				 * draining.  If this fails (either because
7861 				 * another CPU has beat us to the exit action,
7862 				 * or because our current activity is something
7863 				 * other than ACTIVE or WARMUP), we will
7864 				 * continue.  This assures that the exit action
7865 				 * can be successfully recorded at most once
7866 				 * when we're in the ACTIVE state.  If we're
7867 				 * encountering the exit() action while in
7868 				 * COOLDOWN, however, we want to honor the new
7869 				 * status code.  (We know that we're the only
7870 				 * thread in COOLDOWN, so there is no race.)
7871 				 */
7872 				void *activity = &state->dts_activity;
7873 				dtrace_activity_t curstate = state->dts_activity;
7874 
7875 				if (curstate == DTRACE_ACTIVITY_COOLDOWN)
7876 					break;
7877 
7878 				if (curstate != DTRACE_ACTIVITY_WARMUP)
7879 					curstate = DTRACE_ACTIVITY_ACTIVE;
7880 
7881 				if (dtrace_cas32(activity, curstate,
7882 				    DTRACE_ACTIVITY_DRAINING) != curstate) {
7883 					*flags |= CPU_DTRACE_DROP;
7884 					continue;
7885 				}
7886 
7887 				break;
7888 			}
7889 
7890 			default:
7891 				ASSERT(0);
7892 			}
7893 
7894 			if (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF ||
7895 			    dp->dtdo_rtype.dtdt_flags & DIF_TF_BYUREF) {
7896 				uintptr_t end = valoffs + size;
7897 
7898 				if (tracememsize != 0 &&
7899 				    valoffs + tracememsize < end) {
7900 					end = valoffs + tracememsize;
7901 					tracememsize = 0;
7902 				}
7903 
7904 				if (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF &&
7905 				    !dtrace_vcanload((void *)(uintptr_t)val,
7906 				    &dp->dtdo_rtype, NULL, &mstate, vstate))
7907 					continue;
7908 
7909 				dtrace_store_by_ref(dp, tomax, size, &valoffs,
7910 				    &val, end, act->dta_intuple,
7911 				    dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF ?
7912 				    DIF_TF_BYREF: DIF_TF_BYUREF);
7913 				continue;
7914 			}
7915 
7916 			switch (size) {
7917 			case 0:
7918 				break;
7919 
7920 			case sizeof (uint8_t):
7921 				DTRACE_STORE(uint8_t, tomax, valoffs, val);
7922 				break;
7923 			case sizeof (uint16_t):
7924 				DTRACE_STORE(uint16_t, tomax, valoffs, val);
7925 				break;
7926 			case sizeof (uint32_t):
7927 				DTRACE_STORE(uint32_t, tomax, valoffs, val);
7928 				break;
7929 			case sizeof (uint64_t):
7930 				DTRACE_STORE(uint64_t, tomax, valoffs, val);
7931 				break;
7932 			default:
7933 				/*
7934 				 * Any other size should have been returned by
7935 				 * reference, not by value.
7936 				 */
7937 				ASSERT(0);
7938 				break;
7939 			}
7940 		}
7941 
7942 		if (*flags & CPU_DTRACE_DROP)
7943 			continue;
7944 
7945 		if (*flags & CPU_DTRACE_FAULT) {
7946 			int ndx;
7947 			dtrace_action_t *err;
7948 
7949 			buf->dtb_errors++;
7950 
7951 			if (probe->dtpr_id == dtrace_probeid_error) {
7952 				/*
7953 				 * There's nothing we can do -- we had an
7954 				 * error on the error probe.  We bump an
7955 				 * error counter to at least indicate that
7956 				 * this condition happened.
7957 				 */
7958 				dtrace_error(&state->dts_dblerrors);
7959 				continue;
7960 			}
7961 
7962 			if (vtime) {
7963 				/*
7964 				 * Before recursing on dtrace_probe(), we
7965 				 * need to explicitly clear out our start
7966 				 * time to prevent it from being accumulated
7967 				 * into t_dtrace_vtime.
7968 				 */
7969 				curthread->t_dtrace_start = 0;
7970 			}
7971 
7972 			/*
7973 			 * Iterate over the actions to figure out which action
7974 			 * we were processing when we experienced the error.
7975 			 * Note that act points _past_ the faulting action; if
7976 			 * act is ecb->dte_action, the fault was in the
7977 			 * predicate, if it's ecb->dte_action->dta_next it's
7978 			 * in action #1, and so on.
7979 			 */
7980 			for (err = ecb->dte_action, ndx = 0;
7981 			    err != act; err = err->dta_next, ndx++)
7982 				continue;
7983 
7984 			dtrace_probe_error(state, ecb->dte_epid, ndx,
7985 			    (mstate.dtms_present & DTRACE_MSTATE_FLTOFFS) ?
7986 			    mstate.dtms_fltoffs : -1, DTRACE_FLAGS2FLT(*flags),
7987 			    cpu_core[cpuid].cpuc_dtrace_illval);
7988 
7989 			continue;
7990 		}
7991 
7992 		if (!committed)
7993 			buf->dtb_offset = offs + ecb->dte_size;
7994 	}
7995 
7996 	if (vtime)
7997 		curthread->t_dtrace_start = dtrace_gethrtime();
7998 
7999 	dtrace_probe_exit(cookie);
8000 }
8001 
8002 /*
8003  * DTrace Probe Hashing Functions
8004  *
8005  * The functions in this section (and indeed, the functions in remaining
8006  * sections) are not _called_ from probe context.  (Any exceptions to this are
8007  * marked with a "Note:".)  Rather, they are called from elsewhere in the
8008  * DTrace framework to look-up probes in, add probes to and remove probes from
8009  * the DTrace probe hashes.  (Each probe is hashed by each element of the
8010  * probe tuple -- allowing for fast lookups, regardless of what was
8011  * specified.)
8012  */
8013 static uint_t
8014 dtrace_hash_str(const char *p)
8015 {
8016 	unsigned int g;
8017 	uint_t hval = 0;
8018 
8019 	while (*p) {
8020 		hval = (hval << 4) + *p++;
8021 		if ((g = (hval & 0xf0000000)) != 0)
8022 			hval ^= g >> 24;
8023 		hval &= ~g;
8024 	}
8025 	return (hval);
8026 }
8027 
8028 static dtrace_hash_t *
8029 dtrace_hash_create(uintptr_t stroffs, uintptr_t nextoffs, uintptr_t prevoffs)
8030 {
8031 	dtrace_hash_t *hash = kmem_zalloc(sizeof (dtrace_hash_t), KM_SLEEP);
8032 
8033 	hash->dth_stroffs = stroffs;
8034 	hash->dth_nextoffs = nextoffs;
8035 	hash->dth_prevoffs = prevoffs;
8036 
8037 	hash->dth_size = 1;
8038 	hash->dth_mask = hash->dth_size - 1;
8039 
8040 	hash->dth_tab = kmem_zalloc(hash->dth_size *
8041 	    sizeof (dtrace_hashbucket_t *), KM_SLEEP);
8042 
8043 	return (hash);
8044 }
8045 
8046 static void
8047 dtrace_hash_destroy(dtrace_hash_t *hash)
8048 {
8049 #ifdef DEBUG
8050 	int i;
8051 
8052 	for (i = 0; i < hash->dth_size; i++)
8053 		ASSERT(hash->dth_tab[i] == NULL);
8054 #endif
8055 
8056 	kmem_free(hash->dth_tab,
8057 	    hash->dth_size * sizeof (dtrace_hashbucket_t *));
8058 	kmem_free(hash, sizeof (dtrace_hash_t));
8059 }
8060 
8061 static void
8062 dtrace_hash_resize(dtrace_hash_t *hash)
8063 {
8064 	int size = hash->dth_size, i, ndx;
8065 	int new_size = hash->dth_size << 1;
8066 	int new_mask = new_size - 1;
8067 	dtrace_hashbucket_t **new_tab, *bucket, *next;
8068 
8069 	ASSERT((new_size & new_mask) == 0);
8070 
8071 	new_tab = kmem_zalloc(new_size * sizeof (void *), KM_SLEEP);
8072 
8073 	for (i = 0; i < size; i++) {
8074 		for (bucket = hash->dth_tab[i]; bucket != NULL; bucket = next) {
8075 			dtrace_probe_t *probe = bucket->dthb_chain;
8076 
8077 			ASSERT(probe != NULL);
8078 			ndx = DTRACE_HASHSTR(hash, probe) & new_mask;
8079 
8080 			next = bucket->dthb_next;
8081 			bucket->dthb_next = new_tab[ndx];
8082 			new_tab[ndx] = bucket;
8083 		}
8084 	}
8085 
8086 	kmem_free(hash->dth_tab, hash->dth_size * sizeof (void *));
8087 	hash->dth_tab = new_tab;
8088 	hash->dth_size = new_size;
8089 	hash->dth_mask = new_mask;
8090 }
8091 
8092 static void
8093 dtrace_hash_add(dtrace_hash_t *hash, dtrace_probe_t *new)
8094 {
8095 	int hashval = DTRACE_HASHSTR(hash, new);
8096 	int ndx = hashval & hash->dth_mask;
8097 	dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
8098 	dtrace_probe_t **nextp, **prevp;
8099 
8100 	for (; bucket != NULL; bucket = bucket->dthb_next) {
8101 		if (DTRACE_HASHEQ(hash, bucket->dthb_chain, new))
8102 			goto add;
8103 	}
8104 
8105 	if ((hash->dth_nbuckets >> 1) > hash->dth_size) {
8106 		dtrace_hash_resize(hash);
8107 		dtrace_hash_add(hash, new);
8108 		return;
8109 	}
8110 
8111 	bucket = kmem_zalloc(sizeof (dtrace_hashbucket_t), KM_SLEEP);
8112 	bucket->dthb_next = hash->dth_tab[ndx];
8113 	hash->dth_tab[ndx] = bucket;
8114 	hash->dth_nbuckets++;
8115 
8116 add:
8117 	nextp = DTRACE_HASHNEXT(hash, new);
8118 	ASSERT(*nextp == NULL && *(DTRACE_HASHPREV(hash, new)) == NULL);
8119 	*nextp = bucket->dthb_chain;
8120 
8121 	if (bucket->dthb_chain != NULL) {
8122 		prevp = DTRACE_HASHPREV(hash, bucket->dthb_chain);
8123 		ASSERT(*prevp == NULL);
8124 		*prevp = new;
8125 	}
8126 
8127 	bucket->dthb_chain = new;
8128 	bucket->dthb_len++;
8129 }
8130 
8131 static dtrace_probe_t *
8132 dtrace_hash_lookup(dtrace_hash_t *hash, dtrace_probe_t *template)
8133 {
8134 	int hashval = DTRACE_HASHSTR(hash, template);
8135 	int ndx = hashval & hash->dth_mask;
8136 	dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
8137 
8138 	for (; bucket != NULL; bucket = bucket->dthb_next) {
8139 		if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
8140 			return (bucket->dthb_chain);
8141 	}
8142 
8143 	return (NULL);
8144 }
8145 
8146 static int
8147 dtrace_hash_collisions(dtrace_hash_t *hash, dtrace_probe_t *template)
8148 {
8149 	int hashval = DTRACE_HASHSTR(hash, template);
8150 	int ndx = hashval & hash->dth_mask;
8151 	dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
8152 
8153 	for (; bucket != NULL; bucket = bucket->dthb_next) {
8154 		if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
8155 			return (bucket->dthb_len);
8156 	}
8157 
8158 	return (0);
8159 }
8160 
8161 static void
8162 dtrace_hash_remove(dtrace_hash_t *hash, dtrace_probe_t *probe)
8163 {
8164 	int ndx = DTRACE_HASHSTR(hash, probe) & hash->dth_mask;
8165 	dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
8166 
8167 	dtrace_probe_t **prevp = DTRACE_HASHPREV(hash, probe);
8168 	dtrace_probe_t **nextp = DTRACE_HASHNEXT(hash, probe);
8169 
8170 	/*
8171 	 * Find the bucket that we're removing this probe from.
8172 	 */
8173 	for (; bucket != NULL; bucket = bucket->dthb_next) {
8174 		if (DTRACE_HASHEQ(hash, bucket->dthb_chain, probe))
8175 			break;
8176 	}
8177 
8178 	ASSERT(bucket != NULL);
8179 
8180 	if (*prevp == NULL) {
8181 		if (*nextp == NULL) {
8182 			/*
8183 			 * The removed probe was the only probe on this
8184 			 * bucket; we need to remove the bucket.
8185 			 */
8186 			dtrace_hashbucket_t *b = hash->dth_tab[ndx];
8187 
8188 			ASSERT(bucket->dthb_chain == probe);
8189 			ASSERT(b != NULL);
8190 
8191 			if (b == bucket) {
8192 				hash->dth_tab[ndx] = bucket->dthb_next;
8193 			} else {
8194 				while (b->dthb_next != bucket)
8195 					b = b->dthb_next;
8196 				b->dthb_next = bucket->dthb_next;
8197 			}
8198 
8199 			ASSERT(hash->dth_nbuckets > 0);
8200 			hash->dth_nbuckets--;
8201 			kmem_free(bucket, sizeof (dtrace_hashbucket_t));
8202 			return;
8203 		}
8204 
8205 		bucket->dthb_chain = *nextp;
8206 	} else {
8207 		*(DTRACE_HASHNEXT(hash, *prevp)) = *nextp;
8208 	}
8209 
8210 	if (*nextp != NULL)
8211 		*(DTRACE_HASHPREV(hash, *nextp)) = *prevp;
8212 }
8213 
8214 /*
8215  * DTrace Utility Functions
8216  *
8217  * These are random utility functions that are _not_ called from probe context.
8218  */
8219 static int
8220 dtrace_badattr(const dtrace_attribute_t *a)
8221 {
8222 	return (a->dtat_name > DTRACE_STABILITY_MAX ||
8223 	    a->dtat_data > DTRACE_STABILITY_MAX ||
8224 	    a->dtat_class > DTRACE_CLASS_MAX);
8225 }
8226 
8227 /*
8228  * Return a duplicate copy of a string.  If the specified string is NULL,
8229  * this function returns a zero-length string.
8230  */
8231 static char *
8232 dtrace_strdup(const char *str)
8233 {
8234 	char *new = kmem_zalloc((str != NULL ? strlen(str) : 0) + 1, KM_SLEEP);
8235 
8236 	if (str != NULL)
8237 		(void) strcpy(new, str);
8238 
8239 	return (new);
8240 }
8241 
8242 #define	DTRACE_ISALPHA(c)	\
8243 	(((c) >= 'a' && (c) <= 'z') || ((c) >= 'A' && (c) <= 'Z'))
8244 
8245 static int
8246 dtrace_badname(const char *s)
8247 {
8248 	char c;
8249 
8250 	if (s == NULL || (c = *s++) == '\0')
8251 		return (0);
8252 
8253 	if (!DTRACE_ISALPHA(c) && c != '-' && c != '_' && c != '.')
8254 		return (1);
8255 
8256 	while ((c = *s++) != '\0') {
8257 		if (!DTRACE_ISALPHA(c) && (c < '0' || c > '9') &&
8258 		    c != '-' && c != '_' && c != '.' && c != '`')
8259 			return (1);
8260 	}
8261 
8262 	return (0);
8263 }
8264 
8265 static void
8266 dtrace_cred2priv(cred_t *cr, uint32_t *privp, uid_t *uidp, zoneid_t *zoneidp)
8267 {
8268 	uint32_t priv;
8269 
8270 #ifdef illumos
8271 	if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
8272 		/*
8273 		 * For DTRACE_PRIV_ALL, the uid and zoneid don't matter.
8274 		 */
8275 		priv = DTRACE_PRIV_ALL;
8276 	} else {
8277 		*uidp = crgetuid(cr);
8278 		*zoneidp = crgetzoneid(cr);
8279 
8280 		priv = 0;
8281 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE))
8282 			priv |= DTRACE_PRIV_KERNEL | DTRACE_PRIV_USER;
8283 		else if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE))
8284 			priv |= DTRACE_PRIV_USER;
8285 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE))
8286 			priv |= DTRACE_PRIV_PROC;
8287 		if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
8288 			priv |= DTRACE_PRIV_OWNER;
8289 		if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
8290 			priv |= DTRACE_PRIV_ZONEOWNER;
8291 	}
8292 #else
8293 	priv = DTRACE_PRIV_ALL;
8294 #endif
8295 
8296 	*privp = priv;
8297 }
8298 
8299 #ifdef DTRACE_ERRDEBUG
8300 static void
8301 dtrace_errdebug(const char *str)
8302 {
8303 	int hval = dtrace_hash_str(str) % DTRACE_ERRHASHSZ;
8304 	int occupied = 0;
8305 
8306 	mutex_enter(&dtrace_errlock);
8307 	dtrace_errlast = str;
8308 	dtrace_errthread = curthread;
8309 
8310 	while (occupied++ < DTRACE_ERRHASHSZ) {
8311 		if (dtrace_errhash[hval].dter_msg == str) {
8312 			dtrace_errhash[hval].dter_count++;
8313 			goto out;
8314 		}
8315 
8316 		if (dtrace_errhash[hval].dter_msg != NULL) {
8317 			hval = (hval + 1) % DTRACE_ERRHASHSZ;
8318 			continue;
8319 		}
8320 
8321 		dtrace_errhash[hval].dter_msg = str;
8322 		dtrace_errhash[hval].dter_count = 1;
8323 		goto out;
8324 	}
8325 
8326 	panic("dtrace: undersized error hash");
8327 out:
8328 	mutex_exit(&dtrace_errlock);
8329 }
8330 #endif
8331 
8332 /*
8333  * DTrace Matching Functions
8334  *
8335  * These functions are used to match groups of probes, given some elements of
8336  * a probe tuple, or some globbed expressions for elements of a probe tuple.
8337  */
8338 static int
8339 dtrace_match_priv(const dtrace_probe_t *prp, uint32_t priv, uid_t uid,
8340     zoneid_t zoneid)
8341 {
8342 	if (priv != DTRACE_PRIV_ALL) {
8343 		uint32_t ppriv = prp->dtpr_provider->dtpv_priv.dtpp_flags;
8344 		uint32_t match = priv & ppriv;
8345 
8346 		/*
8347 		 * No PRIV_DTRACE_* privileges...
8348 		 */
8349 		if ((priv & (DTRACE_PRIV_PROC | DTRACE_PRIV_USER |
8350 		    DTRACE_PRIV_KERNEL)) == 0)
8351 			return (0);
8352 
8353 		/*
8354 		 * No matching bits, but there were bits to match...
8355 		 */
8356 		if (match == 0 && ppriv != 0)
8357 			return (0);
8358 
8359 		/*
8360 		 * Need to have permissions to the process, but don't...
8361 		 */
8362 		if (((ppriv & ~match) & DTRACE_PRIV_OWNER) != 0 &&
8363 		    uid != prp->dtpr_provider->dtpv_priv.dtpp_uid) {
8364 			return (0);
8365 		}
8366 
8367 		/*
8368 		 * Need to be in the same zone unless we possess the
8369 		 * privilege to examine all zones.
8370 		 */
8371 		if (((ppriv & ~match) & DTRACE_PRIV_ZONEOWNER) != 0 &&
8372 		    zoneid != prp->dtpr_provider->dtpv_priv.dtpp_zoneid) {
8373 			return (0);
8374 		}
8375 	}
8376 
8377 	return (1);
8378 }
8379 
8380 /*
8381  * dtrace_match_probe compares a dtrace_probe_t to a pre-compiled key, which
8382  * consists of input pattern strings and an ops-vector to evaluate them.
8383  * This function returns >0 for match, 0 for no match, and <0 for error.
8384  */
8385 static int
8386 dtrace_match_probe(const dtrace_probe_t *prp, const dtrace_probekey_t *pkp,
8387     uint32_t priv, uid_t uid, zoneid_t zoneid)
8388 {
8389 	dtrace_provider_t *pvp = prp->dtpr_provider;
8390 	int rv;
8391 
8392 	if (pvp->dtpv_defunct)
8393 		return (0);
8394 
8395 	if ((rv = pkp->dtpk_pmatch(pvp->dtpv_name, pkp->dtpk_prov, 0)) <= 0)
8396 		return (rv);
8397 
8398 	if ((rv = pkp->dtpk_mmatch(prp->dtpr_mod, pkp->dtpk_mod, 0)) <= 0)
8399 		return (rv);
8400 
8401 	if ((rv = pkp->dtpk_fmatch(prp->dtpr_func, pkp->dtpk_func, 0)) <= 0)
8402 		return (rv);
8403 
8404 	if ((rv = pkp->dtpk_nmatch(prp->dtpr_name, pkp->dtpk_name, 0)) <= 0)
8405 		return (rv);
8406 
8407 	if (dtrace_match_priv(prp, priv, uid, zoneid) == 0)
8408 		return (0);
8409 
8410 	return (rv);
8411 }
8412 
8413 /*
8414  * dtrace_match_glob() is a safe kernel implementation of the gmatch(3GEN)
8415  * interface for matching a glob pattern 'p' to an input string 's'.  Unlike
8416  * libc's version, the kernel version only applies to 8-bit ASCII strings.
8417  * In addition, all of the recursion cases except for '*' matching have been
8418  * unwound.  For '*', we still implement recursive evaluation, but a depth
8419  * counter is maintained and matching is aborted if we recurse too deep.
8420  * The function returns 0 if no match, >0 if match, and <0 if recursion error.
8421  */
8422 static int
8423 dtrace_match_glob(const char *s, const char *p, int depth)
8424 {
8425 	const char *olds;
8426 	char s1, c;
8427 	int gs;
8428 
8429 	if (depth > DTRACE_PROBEKEY_MAXDEPTH)
8430 		return (-1);
8431 
8432 	if (s == NULL)
8433 		s = ""; /* treat NULL as empty string */
8434 
8435 top:
8436 	olds = s;
8437 	s1 = *s++;
8438 
8439 	if (p == NULL)
8440 		return (0);
8441 
8442 	if ((c = *p++) == '\0')
8443 		return (s1 == '\0');
8444 
8445 	switch (c) {
8446 	case '[': {
8447 		int ok = 0, notflag = 0;
8448 		char lc = '\0';
8449 
8450 		if (s1 == '\0')
8451 			return (0);
8452 
8453 		if (*p == '!') {
8454 			notflag = 1;
8455 			p++;
8456 		}
8457 
8458 		if ((c = *p++) == '\0')
8459 			return (0);
8460 
8461 		do {
8462 			if (c == '-' && lc != '\0' && *p != ']') {
8463 				if ((c = *p++) == '\0')
8464 					return (0);
8465 				if (c == '\\' && (c = *p++) == '\0')
8466 					return (0);
8467 
8468 				if (notflag) {
8469 					if (s1 < lc || s1 > c)
8470 						ok++;
8471 					else
8472 						return (0);
8473 				} else if (lc <= s1 && s1 <= c)
8474 					ok++;
8475 
8476 			} else if (c == '\\' && (c = *p++) == '\0')
8477 				return (0);
8478 
8479 			lc = c; /* save left-hand 'c' for next iteration */
8480 
8481 			if (notflag) {
8482 				if (s1 != c)
8483 					ok++;
8484 				else
8485 					return (0);
8486 			} else if (s1 == c)
8487 				ok++;
8488 
8489 			if ((c = *p++) == '\0')
8490 				return (0);
8491 
8492 		} while (c != ']');
8493 
8494 		if (ok)
8495 			goto top;
8496 
8497 		return (0);
8498 	}
8499 
8500 	case '\\':
8501 		if ((c = *p++) == '\0')
8502 			return (0);
8503 		/*FALLTHRU*/
8504 
8505 	default:
8506 		if (c != s1)
8507 			return (0);
8508 		/*FALLTHRU*/
8509 
8510 	case '?':
8511 		if (s1 != '\0')
8512 			goto top;
8513 		return (0);
8514 
8515 	case '*':
8516 		while (*p == '*')
8517 			p++; /* consecutive *'s are identical to a single one */
8518 
8519 		if (*p == '\0')
8520 			return (1);
8521 
8522 		for (s = olds; *s != '\0'; s++) {
8523 			if ((gs = dtrace_match_glob(s, p, depth + 1)) != 0)
8524 				return (gs);
8525 		}
8526 
8527 		return (0);
8528 	}
8529 }
8530 
8531 /*ARGSUSED*/
8532 static int
8533 dtrace_match_string(const char *s, const char *p, int depth)
8534 {
8535 	return (s != NULL && strcmp(s, p) == 0);
8536 }
8537 
8538 /*ARGSUSED*/
8539 static int
8540 dtrace_match_nul(const char *s, const char *p, int depth)
8541 {
8542 	return (1); /* always match the empty pattern */
8543 }
8544 
8545 /*ARGSUSED*/
8546 static int
8547 dtrace_match_nonzero(const char *s, const char *p, int depth)
8548 {
8549 	return (s != NULL && s[0] != '\0');
8550 }
8551 
8552 static int
8553 dtrace_match(const dtrace_probekey_t *pkp, uint32_t priv, uid_t uid,
8554     zoneid_t zoneid, int (*matched)(dtrace_probe_t *, void *), void *arg)
8555 {
8556 	dtrace_probe_t template, *probe;
8557 	dtrace_hash_t *hash = NULL;
8558 	int len, best = INT_MAX, nmatched = 0;
8559 	dtrace_id_t i;
8560 
8561 	ASSERT(MUTEX_HELD(&dtrace_lock));
8562 
8563 	/*
8564 	 * If the probe ID is specified in the key, just lookup by ID and
8565 	 * invoke the match callback once if a matching probe is found.
8566 	 */
8567 	if (pkp->dtpk_id != DTRACE_IDNONE) {
8568 		if ((probe = dtrace_probe_lookup_id(pkp->dtpk_id)) != NULL &&
8569 		    dtrace_match_probe(probe, pkp, priv, uid, zoneid) > 0) {
8570 			(void) (*matched)(probe, arg);
8571 			nmatched++;
8572 		}
8573 		return (nmatched);
8574 	}
8575 
8576 	template.dtpr_mod = (char *)pkp->dtpk_mod;
8577 	template.dtpr_func = (char *)pkp->dtpk_func;
8578 	template.dtpr_name = (char *)pkp->dtpk_name;
8579 
8580 	/*
8581 	 * We want to find the most distinct of the module name, function
8582 	 * name, and name.  So for each one that is not a glob pattern or
8583 	 * empty string, we perform a lookup in the corresponding hash and
8584 	 * use the hash table with the fewest collisions to do our search.
8585 	 */
8586 	if (pkp->dtpk_mmatch == &dtrace_match_string &&
8587 	    (len = dtrace_hash_collisions(dtrace_bymod, &template)) < best) {
8588 		best = len;
8589 		hash = dtrace_bymod;
8590 	}
8591 
8592 	if (pkp->dtpk_fmatch == &dtrace_match_string &&
8593 	    (len = dtrace_hash_collisions(dtrace_byfunc, &template)) < best) {
8594 		best = len;
8595 		hash = dtrace_byfunc;
8596 	}
8597 
8598 	if (pkp->dtpk_nmatch == &dtrace_match_string &&
8599 	    (len = dtrace_hash_collisions(dtrace_byname, &template)) < best) {
8600 		best = len;
8601 		hash = dtrace_byname;
8602 	}
8603 
8604 	/*
8605 	 * If we did not select a hash table, iterate over every probe and
8606 	 * invoke our callback for each one that matches our input probe key.
8607 	 */
8608 	if (hash == NULL) {
8609 		for (i = 0; i < dtrace_nprobes; i++) {
8610 			if ((probe = dtrace_probes[i]) == NULL ||
8611 			    dtrace_match_probe(probe, pkp, priv, uid,
8612 			    zoneid) <= 0)
8613 				continue;
8614 
8615 			nmatched++;
8616 
8617 			if ((*matched)(probe, arg) != DTRACE_MATCH_NEXT)
8618 				break;
8619 		}
8620 
8621 		return (nmatched);
8622 	}
8623 
8624 	/*
8625 	 * If we selected a hash table, iterate over each probe of the same key
8626 	 * name and invoke the callback for every probe that matches the other
8627 	 * attributes of our input probe key.
8628 	 */
8629 	for (probe = dtrace_hash_lookup(hash, &template); probe != NULL;
8630 	    probe = *(DTRACE_HASHNEXT(hash, probe))) {
8631 
8632 		if (dtrace_match_probe(probe, pkp, priv, uid, zoneid) <= 0)
8633 			continue;
8634 
8635 		nmatched++;
8636 
8637 		if ((*matched)(probe, arg) != DTRACE_MATCH_NEXT)
8638 			break;
8639 	}
8640 
8641 	return (nmatched);
8642 }
8643 
8644 /*
8645  * Return the function pointer dtrace_probecmp() should use to compare the
8646  * specified pattern with a string.  For NULL or empty patterns, we select
8647  * dtrace_match_nul().  For glob pattern strings, we use dtrace_match_glob().
8648  * For non-empty non-glob strings, we use dtrace_match_string().
8649  */
8650 static dtrace_probekey_f *
8651 dtrace_probekey_func(const char *p)
8652 {
8653 	char c;
8654 
8655 	if (p == NULL || *p == '\0')
8656 		return (&dtrace_match_nul);
8657 
8658 	while ((c = *p++) != '\0') {
8659 		if (c == '[' || c == '?' || c == '*' || c == '\\')
8660 			return (&dtrace_match_glob);
8661 	}
8662 
8663 	return (&dtrace_match_string);
8664 }
8665 
8666 /*
8667  * Build a probe comparison key for use with dtrace_match_probe() from the
8668  * given probe description.  By convention, a null key only matches anchored
8669  * probes: if each field is the empty string, reset dtpk_fmatch to
8670  * dtrace_match_nonzero().
8671  */
8672 static void
8673 dtrace_probekey(dtrace_probedesc_t *pdp, dtrace_probekey_t *pkp)
8674 {
8675 	pkp->dtpk_prov = pdp->dtpd_provider;
8676 	pkp->dtpk_pmatch = dtrace_probekey_func(pdp->dtpd_provider);
8677 
8678 	pkp->dtpk_mod = pdp->dtpd_mod;
8679 	pkp->dtpk_mmatch = dtrace_probekey_func(pdp->dtpd_mod);
8680 
8681 	pkp->dtpk_func = pdp->dtpd_func;
8682 	pkp->dtpk_fmatch = dtrace_probekey_func(pdp->dtpd_func);
8683 
8684 	pkp->dtpk_name = pdp->dtpd_name;
8685 	pkp->dtpk_nmatch = dtrace_probekey_func(pdp->dtpd_name);
8686 
8687 	pkp->dtpk_id = pdp->dtpd_id;
8688 
8689 	if (pkp->dtpk_id == DTRACE_IDNONE &&
8690 	    pkp->dtpk_pmatch == &dtrace_match_nul &&
8691 	    pkp->dtpk_mmatch == &dtrace_match_nul &&
8692 	    pkp->dtpk_fmatch == &dtrace_match_nul &&
8693 	    pkp->dtpk_nmatch == &dtrace_match_nul)
8694 		pkp->dtpk_fmatch = &dtrace_match_nonzero;
8695 }
8696 
8697 /*
8698  * DTrace Provider-to-Framework API Functions
8699  *
8700  * These functions implement much of the Provider-to-Framework API, as
8701  * described in <sys/dtrace.h>.  The parts of the API not in this section are
8702  * the functions in the API for probe management (found below), and
8703  * dtrace_probe() itself (found above).
8704  */
8705 
8706 /*
8707  * Register the calling provider with the DTrace framework.  This should
8708  * generally be called by DTrace providers in their attach(9E) entry point.
8709  */
8710 int
8711 dtrace_register(const char *name, const dtrace_pattr_t *pap, uint32_t priv,
8712     cred_t *cr, const dtrace_pops_t *pops, void *arg, dtrace_provider_id_t *idp)
8713 {
8714 	dtrace_provider_t *provider;
8715 
8716 	if (name == NULL || pap == NULL || pops == NULL || idp == NULL) {
8717 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
8718 		    "arguments", name ? name : "<NULL>");
8719 		return (EINVAL);
8720 	}
8721 
8722 	if (name[0] == '\0' || dtrace_badname(name)) {
8723 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
8724 		    "provider name", name);
8725 		return (EINVAL);
8726 	}
8727 
8728 	if ((pops->dtps_provide == NULL && pops->dtps_provide_module == NULL) ||
8729 	    pops->dtps_enable == NULL || pops->dtps_disable == NULL ||
8730 	    pops->dtps_destroy == NULL ||
8731 	    ((pops->dtps_resume == NULL) != (pops->dtps_suspend == NULL))) {
8732 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
8733 		    "provider ops", name);
8734 		return (EINVAL);
8735 	}
8736 
8737 	if (dtrace_badattr(&pap->dtpa_provider) ||
8738 	    dtrace_badattr(&pap->dtpa_mod) ||
8739 	    dtrace_badattr(&pap->dtpa_func) ||
8740 	    dtrace_badattr(&pap->dtpa_name) ||
8741 	    dtrace_badattr(&pap->dtpa_args)) {
8742 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
8743 		    "provider attributes", name);
8744 		return (EINVAL);
8745 	}
8746 
8747 	if (priv & ~DTRACE_PRIV_ALL) {
8748 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
8749 		    "privilege attributes", name);
8750 		return (EINVAL);
8751 	}
8752 
8753 	if ((priv & DTRACE_PRIV_KERNEL) &&
8754 	    (priv & (DTRACE_PRIV_USER | DTRACE_PRIV_OWNER)) &&
8755 	    pops->dtps_usermode == NULL) {
8756 		cmn_err(CE_WARN, "failed to register provider '%s': need "
8757 		    "dtps_usermode() op for given privilege attributes", name);
8758 		return (EINVAL);
8759 	}
8760 
8761 	provider = kmem_zalloc(sizeof (dtrace_provider_t), KM_SLEEP);
8762 	provider->dtpv_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
8763 	(void) strcpy(provider->dtpv_name, name);
8764 
8765 	provider->dtpv_attr = *pap;
8766 	provider->dtpv_priv.dtpp_flags = priv;
8767 	if (cr != NULL) {
8768 		provider->dtpv_priv.dtpp_uid = crgetuid(cr);
8769 		provider->dtpv_priv.dtpp_zoneid = crgetzoneid(cr);
8770 	}
8771 	provider->dtpv_pops = *pops;
8772 
8773 	if (pops->dtps_provide == NULL) {
8774 		ASSERT(pops->dtps_provide_module != NULL);
8775 		provider->dtpv_pops.dtps_provide =
8776 		    (void (*)(void *, dtrace_probedesc_t *))dtrace_nullop;
8777 	}
8778 
8779 	if (pops->dtps_provide_module == NULL) {
8780 		ASSERT(pops->dtps_provide != NULL);
8781 		provider->dtpv_pops.dtps_provide_module =
8782 		    (void (*)(void *, modctl_t *))dtrace_nullop;
8783 	}
8784 
8785 	if (pops->dtps_suspend == NULL) {
8786 		ASSERT(pops->dtps_resume == NULL);
8787 		provider->dtpv_pops.dtps_suspend =
8788 		    (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
8789 		provider->dtpv_pops.dtps_resume =
8790 		    (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
8791 	}
8792 
8793 	provider->dtpv_arg = arg;
8794 	*idp = (dtrace_provider_id_t)provider;
8795 
8796 	if (pops == &dtrace_provider_ops) {
8797 		ASSERT(MUTEX_HELD(&dtrace_provider_lock));
8798 		ASSERT(MUTEX_HELD(&dtrace_lock));
8799 		ASSERT(dtrace_anon.dta_enabling == NULL);
8800 
8801 		/*
8802 		 * We make sure that the DTrace provider is at the head of
8803 		 * the provider chain.
8804 		 */
8805 		provider->dtpv_next = dtrace_provider;
8806 		dtrace_provider = provider;
8807 		return (0);
8808 	}
8809 
8810 	mutex_enter(&dtrace_provider_lock);
8811 	mutex_enter(&dtrace_lock);
8812 
8813 	/*
8814 	 * If there is at least one provider registered, we'll add this
8815 	 * provider after the first provider.
8816 	 */
8817 	if (dtrace_provider != NULL) {
8818 		provider->dtpv_next = dtrace_provider->dtpv_next;
8819 		dtrace_provider->dtpv_next = provider;
8820 	} else {
8821 		dtrace_provider = provider;
8822 	}
8823 
8824 	if (dtrace_retained != NULL) {
8825 		dtrace_enabling_provide(provider);
8826 
8827 		/*
8828 		 * Now we need to call dtrace_enabling_matchall() -- which
8829 		 * will acquire cpu_lock and dtrace_lock.  We therefore need
8830 		 * to drop all of our locks before calling into it...
8831 		 */
8832 		mutex_exit(&dtrace_lock);
8833 		mutex_exit(&dtrace_provider_lock);
8834 		dtrace_enabling_matchall();
8835 
8836 		return (0);
8837 	}
8838 
8839 	mutex_exit(&dtrace_lock);
8840 	mutex_exit(&dtrace_provider_lock);
8841 
8842 	return (0);
8843 }
8844 
8845 /*
8846  * Unregister the specified provider from the DTrace framework.  This should
8847  * generally be called by DTrace providers in their detach(9E) entry point.
8848  */
8849 int
8850 dtrace_unregister(dtrace_provider_id_t id)
8851 {
8852 	dtrace_provider_t *old = (dtrace_provider_t *)id;
8853 	dtrace_provider_t *prev = NULL;
8854 	int i, self = 0, noreap = 0;
8855 	dtrace_probe_t *probe, *first = NULL;
8856 
8857 	if (old->dtpv_pops.dtps_enable ==
8858 	    (void (*)(void *, dtrace_id_t, void *))dtrace_nullop) {
8859 		/*
8860 		 * If DTrace itself is the provider, we're called with locks
8861 		 * already held.
8862 		 */
8863 		ASSERT(old == dtrace_provider);
8864 #ifdef illumos
8865 		ASSERT(dtrace_devi != NULL);
8866 #endif
8867 		ASSERT(MUTEX_HELD(&dtrace_provider_lock));
8868 		ASSERT(MUTEX_HELD(&dtrace_lock));
8869 		self = 1;
8870 
8871 		if (dtrace_provider->dtpv_next != NULL) {
8872 			/*
8873 			 * There's another provider here; return failure.
8874 			 */
8875 			return (EBUSY);
8876 		}
8877 	} else {
8878 		mutex_enter(&dtrace_provider_lock);
8879 #ifdef illumos
8880 		mutex_enter(&mod_lock);
8881 #endif
8882 		mutex_enter(&dtrace_lock);
8883 	}
8884 
8885 	/*
8886 	 * If anyone has /dev/dtrace open, or if there are anonymous enabled
8887 	 * probes, we refuse to let providers slither away, unless this
8888 	 * provider has already been explicitly invalidated.
8889 	 */
8890 	if (!old->dtpv_defunct &&
8891 	    (dtrace_opens || (dtrace_anon.dta_state != NULL &&
8892 	    dtrace_anon.dta_state->dts_necbs > 0))) {
8893 		if (!self) {
8894 			mutex_exit(&dtrace_lock);
8895 #ifdef illumos
8896 			mutex_exit(&mod_lock);
8897 #endif
8898 			mutex_exit(&dtrace_provider_lock);
8899 		}
8900 		return (EBUSY);
8901 	}
8902 
8903 	/*
8904 	 * Attempt to destroy the probes associated with this provider.
8905 	 */
8906 	for (i = 0; i < dtrace_nprobes; i++) {
8907 		if ((probe = dtrace_probes[i]) == NULL)
8908 			continue;
8909 
8910 		if (probe->dtpr_provider != old)
8911 			continue;
8912 
8913 		if (probe->dtpr_ecb == NULL)
8914 			continue;
8915 
8916 		/*
8917 		 * If we are trying to unregister a defunct provider, and the
8918 		 * provider was made defunct within the interval dictated by
8919 		 * dtrace_unregister_defunct_reap, we'll (asynchronously)
8920 		 * attempt to reap our enablings.  To denote that the provider
8921 		 * should reattempt to unregister itself at some point in the
8922 		 * future, we will return a differentiable error code (EAGAIN
8923 		 * instead of EBUSY) in this case.
8924 		 */
8925 		if (dtrace_gethrtime() - old->dtpv_defunct >
8926 		    dtrace_unregister_defunct_reap)
8927 			noreap = 1;
8928 
8929 		if (!self) {
8930 			mutex_exit(&dtrace_lock);
8931 #ifdef illumos
8932 			mutex_exit(&mod_lock);
8933 #endif
8934 			mutex_exit(&dtrace_provider_lock);
8935 		}
8936 
8937 		if (noreap)
8938 			return (EBUSY);
8939 
8940 		(void) taskq_dispatch(dtrace_taskq,
8941 		    (task_func_t *)dtrace_enabling_reap, NULL, TQ_SLEEP);
8942 
8943 		return (EAGAIN);
8944 	}
8945 
8946 	/*
8947 	 * All of the probes for this provider are disabled; we can safely
8948 	 * remove all of them from their hash chains and from the probe array.
8949 	 */
8950 	for (i = 0; i < dtrace_nprobes; i++) {
8951 		if ((probe = dtrace_probes[i]) == NULL)
8952 			continue;
8953 
8954 		if (probe->dtpr_provider != old)
8955 			continue;
8956 
8957 		dtrace_probes[i] = NULL;
8958 
8959 		dtrace_hash_remove(dtrace_bymod, probe);
8960 		dtrace_hash_remove(dtrace_byfunc, probe);
8961 		dtrace_hash_remove(dtrace_byname, probe);
8962 
8963 		if (first == NULL) {
8964 			first = probe;
8965 			probe->dtpr_nextmod = NULL;
8966 		} else {
8967 			probe->dtpr_nextmod = first;
8968 			first = probe;
8969 		}
8970 	}
8971 
8972 	/*
8973 	 * The provider's probes have been removed from the hash chains and
8974 	 * from the probe array.  Now issue a dtrace_sync() to be sure that
8975 	 * everyone has cleared out from any probe array processing.
8976 	 */
8977 	dtrace_sync();
8978 
8979 	for (probe = first; probe != NULL; probe = first) {
8980 		first = probe->dtpr_nextmod;
8981 
8982 		old->dtpv_pops.dtps_destroy(old->dtpv_arg, probe->dtpr_id,
8983 		    probe->dtpr_arg);
8984 		kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
8985 		kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
8986 		kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
8987 #ifdef illumos
8988 		vmem_free(dtrace_arena, (void *)(uintptr_t)(probe->dtpr_id), 1);
8989 #else
8990 		free_unr(dtrace_arena, probe->dtpr_id);
8991 #endif
8992 		kmem_free(probe, sizeof (dtrace_probe_t));
8993 	}
8994 
8995 	if ((prev = dtrace_provider) == old) {
8996 #ifdef illumos
8997 		ASSERT(self || dtrace_devi == NULL);
8998 		ASSERT(old->dtpv_next == NULL || dtrace_devi == NULL);
8999 #endif
9000 		dtrace_provider = old->dtpv_next;
9001 	} else {
9002 		while (prev != NULL && prev->dtpv_next != old)
9003 			prev = prev->dtpv_next;
9004 
9005 		if (prev == NULL) {
9006 			panic("attempt to unregister non-existent "
9007 			    "dtrace provider %p\n", (void *)id);
9008 		}
9009 
9010 		prev->dtpv_next = old->dtpv_next;
9011 	}
9012 
9013 	if (!self) {
9014 		mutex_exit(&dtrace_lock);
9015 #ifdef illumos
9016 		mutex_exit(&mod_lock);
9017 #endif
9018 		mutex_exit(&dtrace_provider_lock);
9019 	}
9020 
9021 	kmem_free(old->dtpv_name, strlen(old->dtpv_name) + 1);
9022 	kmem_free(old, sizeof (dtrace_provider_t));
9023 
9024 	return (0);
9025 }
9026 
9027 /*
9028  * Invalidate the specified provider.  All subsequent probe lookups for the
9029  * specified provider will fail, but its probes will not be removed.
9030  */
9031 void
9032 dtrace_invalidate(dtrace_provider_id_t id)
9033 {
9034 	dtrace_provider_t *pvp = (dtrace_provider_t *)id;
9035 
9036 	ASSERT(pvp->dtpv_pops.dtps_enable !=
9037 	    (void (*)(void *, dtrace_id_t, void *))dtrace_nullop);
9038 
9039 	mutex_enter(&dtrace_provider_lock);
9040 	mutex_enter(&dtrace_lock);
9041 
9042 	pvp->dtpv_defunct = dtrace_gethrtime();
9043 
9044 	mutex_exit(&dtrace_lock);
9045 	mutex_exit(&dtrace_provider_lock);
9046 }
9047 
9048 /*
9049  * Indicate whether or not DTrace has attached.
9050  */
9051 int
9052 dtrace_attached(void)
9053 {
9054 	/*
9055 	 * dtrace_provider will be non-NULL iff the DTrace driver has
9056 	 * attached.  (It's non-NULL because DTrace is always itself a
9057 	 * provider.)
9058 	 */
9059 	return (dtrace_provider != NULL);
9060 }
9061 
9062 /*
9063  * Remove all the unenabled probes for the given provider.  This function is
9064  * not unlike dtrace_unregister(), except that it doesn't remove the provider
9065  * -- just as many of its associated probes as it can.
9066  */
9067 int
9068 dtrace_condense(dtrace_provider_id_t id)
9069 {
9070 	dtrace_provider_t *prov = (dtrace_provider_t *)id;
9071 	int i;
9072 	dtrace_probe_t *probe;
9073 
9074 	/*
9075 	 * Make sure this isn't the dtrace provider itself.
9076 	 */
9077 	ASSERT(prov->dtpv_pops.dtps_enable !=
9078 	    (void (*)(void *, dtrace_id_t, void *))dtrace_nullop);
9079 
9080 	mutex_enter(&dtrace_provider_lock);
9081 	mutex_enter(&dtrace_lock);
9082 
9083 	/*
9084 	 * Attempt to destroy the probes associated with this provider.
9085 	 */
9086 	for (i = 0; i < dtrace_nprobes; i++) {
9087 		if ((probe = dtrace_probes[i]) == NULL)
9088 			continue;
9089 
9090 		if (probe->dtpr_provider != prov)
9091 			continue;
9092 
9093 		if (probe->dtpr_ecb != NULL)
9094 			continue;
9095 
9096 		dtrace_probes[i] = NULL;
9097 
9098 		dtrace_hash_remove(dtrace_bymod, probe);
9099 		dtrace_hash_remove(dtrace_byfunc, probe);
9100 		dtrace_hash_remove(dtrace_byname, probe);
9101 
9102 		prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, i + 1,
9103 		    probe->dtpr_arg);
9104 		kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
9105 		kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
9106 		kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
9107 		kmem_free(probe, sizeof (dtrace_probe_t));
9108 #ifdef illumos
9109 		vmem_free(dtrace_arena, (void *)((uintptr_t)i + 1), 1);
9110 #else
9111 		free_unr(dtrace_arena, i + 1);
9112 #endif
9113 	}
9114 
9115 	mutex_exit(&dtrace_lock);
9116 	mutex_exit(&dtrace_provider_lock);
9117 
9118 	return (0);
9119 }
9120 
9121 /*
9122  * DTrace Probe Management Functions
9123  *
9124  * The functions in this section perform the DTrace probe management,
9125  * including functions to create probes, look-up probes, and call into the
9126  * providers to request that probes be provided.  Some of these functions are
9127  * in the Provider-to-Framework API; these functions can be identified by the
9128  * fact that they are not declared "static".
9129  */
9130 
9131 /*
9132  * Create a probe with the specified module name, function name, and name.
9133  */
9134 dtrace_id_t
9135 dtrace_probe_create(dtrace_provider_id_t prov, const char *mod,
9136     const char *func, const char *name, int aframes, void *arg)
9137 {
9138 	dtrace_probe_t *probe, **probes;
9139 	dtrace_provider_t *provider = (dtrace_provider_t *)prov;
9140 	dtrace_id_t id;
9141 
9142 	if (provider == dtrace_provider) {
9143 		ASSERT(MUTEX_HELD(&dtrace_lock));
9144 	} else {
9145 		mutex_enter(&dtrace_lock);
9146 	}
9147 
9148 #ifdef illumos
9149 	id = (dtrace_id_t)(uintptr_t)vmem_alloc(dtrace_arena, 1,
9150 	    VM_BESTFIT | VM_SLEEP);
9151 #else
9152 	id = alloc_unr(dtrace_arena);
9153 #endif
9154 	probe = kmem_zalloc(sizeof (dtrace_probe_t), KM_SLEEP);
9155 
9156 	probe->dtpr_id = id;
9157 	probe->dtpr_gen = dtrace_probegen++;
9158 	probe->dtpr_mod = dtrace_strdup(mod);
9159 	probe->dtpr_func = dtrace_strdup(func);
9160 	probe->dtpr_name = dtrace_strdup(name);
9161 	probe->dtpr_arg = arg;
9162 	probe->dtpr_aframes = aframes;
9163 	probe->dtpr_provider = provider;
9164 
9165 	dtrace_hash_add(dtrace_bymod, probe);
9166 	dtrace_hash_add(dtrace_byfunc, probe);
9167 	dtrace_hash_add(dtrace_byname, probe);
9168 
9169 	if (id - 1 >= dtrace_nprobes) {
9170 		size_t osize = dtrace_nprobes * sizeof (dtrace_probe_t *);
9171 		size_t nsize = osize << 1;
9172 
9173 		if (nsize == 0) {
9174 			ASSERT(osize == 0);
9175 			ASSERT(dtrace_probes == NULL);
9176 			nsize = sizeof (dtrace_probe_t *);
9177 		}
9178 
9179 		probes = kmem_zalloc(nsize, KM_SLEEP);
9180 
9181 		if (dtrace_probes == NULL) {
9182 			ASSERT(osize == 0);
9183 			dtrace_probes = probes;
9184 			dtrace_nprobes = 1;
9185 		} else {
9186 			dtrace_probe_t **oprobes = dtrace_probes;
9187 
9188 			bcopy(oprobes, probes, osize);
9189 			dtrace_membar_producer();
9190 			dtrace_probes = probes;
9191 
9192 			dtrace_sync();
9193 
9194 			/*
9195 			 * All CPUs are now seeing the new probes array; we can
9196 			 * safely free the old array.
9197 			 */
9198 			kmem_free(oprobes, osize);
9199 			dtrace_nprobes <<= 1;
9200 		}
9201 
9202 		ASSERT(id - 1 < dtrace_nprobes);
9203 	}
9204 
9205 	ASSERT(dtrace_probes[id - 1] == NULL);
9206 	dtrace_probes[id - 1] = probe;
9207 
9208 	if (provider != dtrace_provider)
9209 		mutex_exit(&dtrace_lock);
9210 
9211 	return (id);
9212 }
9213 
9214 static dtrace_probe_t *
9215 dtrace_probe_lookup_id(dtrace_id_t id)
9216 {
9217 	ASSERT(MUTEX_HELD(&dtrace_lock));
9218 
9219 	if (id == 0 || id > dtrace_nprobes)
9220 		return (NULL);
9221 
9222 	return (dtrace_probes[id - 1]);
9223 }
9224 
9225 static int
9226 dtrace_probe_lookup_match(dtrace_probe_t *probe, void *arg)
9227 {
9228 	*((dtrace_id_t *)arg) = probe->dtpr_id;
9229 
9230 	return (DTRACE_MATCH_DONE);
9231 }
9232 
9233 /*
9234  * Look up a probe based on provider and one or more of module name, function
9235  * name and probe name.
9236  */
9237 dtrace_id_t
9238 dtrace_probe_lookup(dtrace_provider_id_t prid, char *mod,
9239     char *func, char *name)
9240 {
9241 	dtrace_probekey_t pkey;
9242 	dtrace_id_t id;
9243 	int match;
9244 
9245 	pkey.dtpk_prov = ((dtrace_provider_t *)prid)->dtpv_name;
9246 	pkey.dtpk_pmatch = &dtrace_match_string;
9247 	pkey.dtpk_mod = mod;
9248 	pkey.dtpk_mmatch = mod ? &dtrace_match_string : &dtrace_match_nul;
9249 	pkey.dtpk_func = func;
9250 	pkey.dtpk_fmatch = func ? &dtrace_match_string : &dtrace_match_nul;
9251 	pkey.dtpk_name = name;
9252 	pkey.dtpk_nmatch = name ? &dtrace_match_string : &dtrace_match_nul;
9253 	pkey.dtpk_id = DTRACE_IDNONE;
9254 
9255 	mutex_enter(&dtrace_lock);
9256 	match = dtrace_match(&pkey, DTRACE_PRIV_ALL, 0, 0,
9257 	    dtrace_probe_lookup_match, &id);
9258 	mutex_exit(&dtrace_lock);
9259 
9260 	ASSERT(match == 1 || match == 0);
9261 	return (match ? id : 0);
9262 }
9263 
9264 /*
9265  * Returns the probe argument associated with the specified probe.
9266  */
9267 void *
9268 dtrace_probe_arg(dtrace_provider_id_t id, dtrace_id_t pid)
9269 {
9270 	dtrace_probe_t *probe;
9271 	void *rval = NULL;
9272 
9273 	mutex_enter(&dtrace_lock);
9274 
9275 	if ((probe = dtrace_probe_lookup_id(pid)) != NULL &&
9276 	    probe->dtpr_provider == (dtrace_provider_t *)id)
9277 		rval = probe->dtpr_arg;
9278 
9279 	mutex_exit(&dtrace_lock);
9280 
9281 	return (rval);
9282 }
9283 
9284 /*
9285  * Copy a probe into a probe description.
9286  */
9287 static void
9288 dtrace_probe_description(const dtrace_probe_t *prp, dtrace_probedesc_t *pdp)
9289 {
9290 	bzero(pdp, sizeof (dtrace_probedesc_t));
9291 	pdp->dtpd_id = prp->dtpr_id;
9292 
9293 	(void) strncpy(pdp->dtpd_provider,
9294 	    prp->dtpr_provider->dtpv_name, DTRACE_PROVNAMELEN - 1);
9295 
9296 	(void) strncpy(pdp->dtpd_mod, prp->dtpr_mod, DTRACE_MODNAMELEN - 1);
9297 	(void) strncpy(pdp->dtpd_func, prp->dtpr_func, DTRACE_FUNCNAMELEN - 1);
9298 	(void) strncpy(pdp->dtpd_name, prp->dtpr_name, DTRACE_NAMELEN - 1);
9299 }
9300 
9301 /*
9302  * Called to indicate that a probe -- or probes -- should be provided by a
9303  * specfied provider.  If the specified description is NULL, the provider will
9304  * be told to provide all of its probes.  (This is done whenever a new
9305  * consumer comes along, or whenever a retained enabling is to be matched.) If
9306  * the specified description is non-NULL, the provider is given the
9307  * opportunity to dynamically provide the specified probe, allowing providers
9308  * to support the creation of probes on-the-fly.  (So-called _autocreated_
9309  * probes.)  If the provider is NULL, the operations will be applied to all
9310  * providers; if the provider is non-NULL the operations will only be applied
9311  * to the specified provider.  The dtrace_provider_lock must be held, and the
9312  * dtrace_lock must _not_ be held -- the provider's dtps_provide() operation
9313  * will need to grab the dtrace_lock when it reenters the framework through
9314  * dtrace_probe_lookup(), dtrace_probe_create(), etc.
9315  */
9316 static void
9317 dtrace_probe_provide(dtrace_probedesc_t *desc, dtrace_provider_t *prv)
9318 {
9319 #ifdef illumos
9320 	modctl_t *ctl;
9321 #endif
9322 	int all = 0;
9323 
9324 	ASSERT(MUTEX_HELD(&dtrace_provider_lock));
9325 
9326 	if (prv == NULL) {
9327 		all = 1;
9328 		prv = dtrace_provider;
9329 	}
9330 
9331 	do {
9332 		/*
9333 		 * First, call the blanket provide operation.
9334 		 */
9335 		prv->dtpv_pops.dtps_provide(prv->dtpv_arg, desc);
9336 
9337 #ifdef illumos
9338 		/*
9339 		 * Now call the per-module provide operation.  We will grab
9340 		 * mod_lock to prevent the list from being modified.  Note
9341 		 * that this also prevents the mod_busy bits from changing.
9342 		 * (mod_busy can only be changed with mod_lock held.)
9343 		 */
9344 		mutex_enter(&mod_lock);
9345 
9346 		ctl = &modules;
9347 		do {
9348 			if (ctl->mod_busy || ctl->mod_mp == NULL)
9349 				continue;
9350 
9351 			prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
9352 
9353 		} while ((ctl = ctl->mod_next) != &modules);
9354 
9355 		mutex_exit(&mod_lock);
9356 #endif
9357 	} while (all && (prv = prv->dtpv_next) != NULL);
9358 }
9359 
9360 #ifdef illumos
9361 /*
9362  * Iterate over each probe, and call the Framework-to-Provider API function
9363  * denoted by offs.
9364  */
9365 static void
9366 dtrace_probe_foreach(uintptr_t offs)
9367 {
9368 	dtrace_provider_t *prov;
9369 	void (*func)(void *, dtrace_id_t, void *);
9370 	dtrace_probe_t *probe;
9371 	dtrace_icookie_t cookie;
9372 	int i;
9373 
9374 	/*
9375 	 * We disable interrupts to walk through the probe array.  This is
9376 	 * safe -- the dtrace_sync() in dtrace_unregister() assures that we
9377 	 * won't see stale data.
9378 	 */
9379 	cookie = dtrace_interrupt_disable();
9380 
9381 	for (i = 0; i < dtrace_nprobes; i++) {
9382 		if ((probe = dtrace_probes[i]) == NULL)
9383 			continue;
9384 
9385 		if (probe->dtpr_ecb == NULL) {
9386 			/*
9387 			 * This probe isn't enabled -- don't call the function.
9388 			 */
9389 			continue;
9390 		}
9391 
9392 		prov = probe->dtpr_provider;
9393 		func = *((void(**)(void *, dtrace_id_t, void *))
9394 		    ((uintptr_t)&prov->dtpv_pops + offs));
9395 
9396 		func(prov->dtpv_arg, i + 1, probe->dtpr_arg);
9397 	}
9398 
9399 	dtrace_interrupt_enable(cookie);
9400 }
9401 #endif
9402 
9403 static int
9404 dtrace_probe_enable(dtrace_probedesc_t *desc, dtrace_enabling_t *enab)
9405 {
9406 	dtrace_probekey_t pkey;
9407 	uint32_t priv;
9408 	uid_t uid;
9409 	zoneid_t zoneid;
9410 
9411 	ASSERT(MUTEX_HELD(&dtrace_lock));
9412 	dtrace_ecb_create_cache = NULL;
9413 
9414 	if (desc == NULL) {
9415 		/*
9416 		 * If we're passed a NULL description, we're being asked to
9417 		 * create an ECB with a NULL probe.
9418 		 */
9419 		(void) dtrace_ecb_create_enable(NULL, enab);
9420 		return (0);
9421 	}
9422 
9423 	dtrace_probekey(desc, &pkey);
9424 	dtrace_cred2priv(enab->dten_vstate->dtvs_state->dts_cred.dcr_cred,
9425 	    &priv, &uid, &zoneid);
9426 
9427 	return (dtrace_match(&pkey, priv, uid, zoneid, dtrace_ecb_create_enable,
9428 	    enab));
9429 }
9430 
9431 /*
9432  * DTrace Helper Provider Functions
9433  */
9434 static void
9435 dtrace_dofattr2attr(dtrace_attribute_t *attr, const dof_attr_t dofattr)
9436 {
9437 	attr->dtat_name = DOF_ATTR_NAME(dofattr);
9438 	attr->dtat_data = DOF_ATTR_DATA(dofattr);
9439 	attr->dtat_class = DOF_ATTR_CLASS(dofattr);
9440 }
9441 
9442 static void
9443 dtrace_dofprov2hprov(dtrace_helper_provdesc_t *hprov,
9444     const dof_provider_t *dofprov, char *strtab)
9445 {
9446 	hprov->dthpv_provname = strtab + dofprov->dofpv_name;
9447 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_provider,
9448 	    dofprov->dofpv_provattr);
9449 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_mod,
9450 	    dofprov->dofpv_modattr);
9451 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_func,
9452 	    dofprov->dofpv_funcattr);
9453 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_name,
9454 	    dofprov->dofpv_nameattr);
9455 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_args,
9456 	    dofprov->dofpv_argsattr);
9457 }
9458 
9459 static void
9460 dtrace_helper_provide_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
9461 {
9462 	uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
9463 	dof_hdr_t *dof = (dof_hdr_t *)daddr;
9464 	dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
9465 	dof_provider_t *provider;
9466 	dof_probe_t *probe;
9467 	uint32_t *off, *enoff;
9468 	uint8_t *arg;
9469 	char *strtab;
9470 	uint_t i, nprobes;
9471 	dtrace_helper_provdesc_t dhpv;
9472 	dtrace_helper_probedesc_t dhpb;
9473 	dtrace_meta_t *meta = dtrace_meta_pid;
9474 	dtrace_mops_t *mops = &meta->dtm_mops;
9475 	void *parg;
9476 
9477 	provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
9478 	str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
9479 	    provider->dofpv_strtab * dof->dofh_secsize);
9480 	prb_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
9481 	    provider->dofpv_probes * dof->dofh_secsize);
9482 	arg_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
9483 	    provider->dofpv_prargs * dof->dofh_secsize);
9484 	off_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
9485 	    provider->dofpv_proffs * dof->dofh_secsize);
9486 
9487 	strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
9488 	off = (uint32_t *)(uintptr_t)(daddr + off_sec->dofs_offset);
9489 	arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
9490 	enoff = NULL;
9491 
9492 	/*
9493 	 * See dtrace_helper_provider_validate().
9494 	 */
9495 	if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
9496 	    provider->dofpv_prenoffs != DOF_SECT_NONE) {
9497 		enoff_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
9498 		    provider->dofpv_prenoffs * dof->dofh_secsize);
9499 		enoff = (uint32_t *)(uintptr_t)(daddr + enoff_sec->dofs_offset);
9500 	}
9501 
9502 	nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
9503 
9504 	/*
9505 	 * Create the provider.
9506 	 */
9507 	dtrace_dofprov2hprov(&dhpv, provider, strtab);
9508 
9509 	if ((parg = mops->dtms_provide_pid(meta->dtm_arg, &dhpv, pid)) == NULL)
9510 		return;
9511 
9512 	meta->dtm_count++;
9513 
9514 	/*
9515 	 * Create the probes.
9516 	 */
9517 	for (i = 0; i < nprobes; i++) {
9518 		probe = (dof_probe_t *)(uintptr_t)(daddr +
9519 		    prb_sec->dofs_offset + i * prb_sec->dofs_entsize);
9520 
9521 		/* See the check in dtrace_helper_provider_validate(). */
9522 		if (strlen(strtab + probe->dofpr_func) >= DTRACE_FUNCNAMELEN)
9523 			continue;
9524 
9525 		dhpb.dthpb_mod = dhp->dofhp_mod;
9526 		dhpb.dthpb_func = strtab + probe->dofpr_func;
9527 		dhpb.dthpb_name = strtab + probe->dofpr_name;
9528 		dhpb.dthpb_base = probe->dofpr_addr;
9529 		dhpb.dthpb_offs = off + probe->dofpr_offidx;
9530 		dhpb.dthpb_noffs = probe->dofpr_noffs;
9531 		if (enoff != NULL) {
9532 			dhpb.dthpb_enoffs = enoff + probe->dofpr_enoffidx;
9533 			dhpb.dthpb_nenoffs = probe->dofpr_nenoffs;
9534 		} else {
9535 			dhpb.dthpb_enoffs = NULL;
9536 			dhpb.dthpb_nenoffs = 0;
9537 		}
9538 		dhpb.dthpb_args = arg + probe->dofpr_argidx;
9539 		dhpb.dthpb_nargc = probe->dofpr_nargc;
9540 		dhpb.dthpb_xargc = probe->dofpr_xargc;
9541 		dhpb.dthpb_ntypes = strtab + probe->dofpr_nargv;
9542 		dhpb.dthpb_xtypes = strtab + probe->dofpr_xargv;
9543 
9544 		mops->dtms_create_probe(meta->dtm_arg, parg, &dhpb);
9545 	}
9546 }
9547 
9548 static void
9549 dtrace_helper_provide(dof_helper_t *dhp, pid_t pid)
9550 {
9551 	uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
9552 	dof_hdr_t *dof = (dof_hdr_t *)daddr;
9553 	int i;
9554 
9555 	ASSERT(MUTEX_HELD(&dtrace_meta_lock));
9556 
9557 	for (i = 0; i < dof->dofh_secnum; i++) {
9558 		dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
9559 		    dof->dofh_secoff + i * dof->dofh_secsize);
9560 
9561 		if (sec->dofs_type != DOF_SECT_PROVIDER)
9562 			continue;
9563 
9564 		dtrace_helper_provide_one(dhp, sec, pid);
9565 	}
9566 
9567 	/*
9568 	 * We may have just created probes, so we must now rematch against
9569 	 * any retained enablings.  Note that this call will acquire both
9570 	 * cpu_lock and dtrace_lock; the fact that we are holding
9571 	 * dtrace_meta_lock now is what defines the ordering with respect to
9572 	 * these three locks.
9573 	 */
9574 	dtrace_enabling_matchall();
9575 }
9576 
9577 static void
9578 dtrace_helper_provider_remove_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
9579 {
9580 	uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
9581 	dof_hdr_t *dof = (dof_hdr_t *)daddr;
9582 	dof_sec_t *str_sec;
9583 	dof_provider_t *provider;
9584 	char *strtab;
9585 	dtrace_helper_provdesc_t dhpv;
9586 	dtrace_meta_t *meta = dtrace_meta_pid;
9587 	dtrace_mops_t *mops = &meta->dtm_mops;
9588 
9589 	provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
9590 	str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
9591 	    provider->dofpv_strtab * dof->dofh_secsize);
9592 
9593 	strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
9594 
9595 	/*
9596 	 * Create the provider.
9597 	 */
9598 	dtrace_dofprov2hprov(&dhpv, provider, strtab);
9599 
9600 	mops->dtms_remove_pid(meta->dtm_arg, &dhpv, pid);
9601 
9602 	meta->dtm_count--;
9603 }
9604 
9605 static void
9606 dtrace_helper_provider_remove(dof_helper_t *dhp, pid_t pid)
9607 {
9608 	uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
9609 	dof_hdr_t *dof = (dof_hdr_t *)daddr;
9610 	int i;
9611 
9612 	ASSERT(MUTEX_HELD(&dtrace_meta_lock));
9613 
9614 	for (i = 0; i < dof->dofh_secnum; i++) {
9615 		dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
9616 		    dof->dofh_secoff + i * dof->dofh_secsize);
9617 
9618 		if (sec->dofs_type != DOF_SECT_PROVIDER)
9619 			continue;
9620 
9621 		dtrace_helper_provider_remove_one(dhp, sec, pid);
9622 	}
9623 }
9624 
9625 /*
9626  * DTrace Meta Provider-to-Framework API Functions
9627  *
9628  * These functions implement the Meta Provider-to-Framework API, as described
9629  * in <sys/dtrace.h>.
9630  */
9631 int
9632 dtrace_meta_register(const char *name, const dtrace_mops_t *mops, void *arg,
9633     dtrace_meta_provider_id_t *idp)
9634 {
9635 	dtrace_meta_t *meta;
9636 	dtrace_helpers_t *help, *next;
9637 	int i;
9638 
9639 	*idp = DTRACE_METAPROVNONE;
9640 
9641 	/*
9642 	 * We strictly don't need the name, but we hold onto it for
9643 	 * debuggability. All hail error queues!
9644 	 */
9645 	if (name == NULL) {
9646 		cmn_err(CE_WARN, "failed to register meta-provider: "
9647 		    "invalid name");
9648 		return (EINVAL);
9649 	}
9650 
9651 	if (mops == NULL ||
9652 	    mops->dtms_create_probe == NULL ||
9653 	    mops->dtms_provide_pid == NULL ||
9654 	    mops->dtms_remove_pid == NULL) {
9655 		cmn_err(CE_WARN, "failed to register meta-register %s: "
9656 		    "invalid ops", name);
9657 		return (EINVAL);
9658 	}
9659 
9660 	meta = kmem_zalloc(sizeof (dtrace_meta_t), KM_SLEEP);
9661 	meta->dtm_mops = *mops;
9662 	meta->dtm_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
9663 	(void) strcpy(meta->dtm_name, name);
9664 	meta->dtm_arg = arg;
9665 
9666 	mutex_enter(&dtrace_meta_lock);
9667 	mutex_enter(&dtrace_lock);
9668 
9669 	if (dtrace_meta_pid != NULL) {
9670 		mutex_exit(&dtrace_lock);
9671 		mutex_exit(&dtrace_meta_lock);
9672 		cmn_err(CE_WARN, "failed to register meta-register %s: "
9673 		    "user-land meta-provider exists", name);
9674 		kmem_free(meta->dtm_name, strlen(meta->dtm_name) + 1);
9675 		kmem_free(meta, sizeof (dtrace_meta_t));
9676 		return (EINVAL);
9677 	}
9678 
9679 	dtrace_meta_pid = meta;
9680 	*idp = (dtrace_meta_provider_id_t)meta;
9681 
9682 	/*
9683 	 * If there are providers and probes ready to go, pass them
9684 	 * off to the new meta provider now.
9685 	 */
9686 
9687 	help = dtrace_deferred_pid;
9688 	dtrace_deferred_pid = NULL;
9689 
9690 	mutex_exit(&dtrace_lock);
9691 
9692 	while (help != NULL) {
9693 		for (i = 0; i < help->dthps_nprovs; i++) {
9694 			dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
9695 			    help->dthps_pid);
9696 		}
9697 
9698 		next = help->dthps_next;
9699 		help->dthps_next = NULL;
9700 		help->dthps_prev = NULL;
9701 		help->dthps_deferred = 0;
9702 		help = next;
9703 	}
9704 
9705 	mutex_exit(&dtrace_meta_lock);
9706 
9707 	return (0);
9708 }
9709 
9710 int
9711 dtrace_meta_unregister(dtrace_meta_provider_id_t id)
9712 {
9713 	dtrace_meta_t **pp, *old = (dtrace_meta_t *)id;
9714 
9715 	mutex_enter(&dtrace_meta_lock);
9716 	mutex_enter(&dtrace_lock);
9717 
9718 	if (old == dtrace_meta_pid) {
9719 		pp = &dtrace_meta_pid;
9720 	} else {
9721 		panic("attempt to unregister non-existent "
9722 		    "dtrace meta-provider %p\n", (void *)old);
9723 	}
9724 
9725 	if (old->dtm_count != 0) {
9726 		mutex_exit(&dtrace_lock);
9727 		mutex_exit(&dtrace_meta_lock);
9728 		return (EBUSY);
9729 	}
9730 
9731 	*pp = NULL;
9732 
9733 	mutex_exit(&dtrace_lock);
9734 	mutex_exit(&dtrace_meta_lock);
9735 
9736 	kmem_free(old->dtm_name, strlen(old->dtm_name) + 1);
9737 	kmem_free(old, sizeof (dtrace_meta_t));
9738 
9739 	return (0);
9740 }
9741 
9742 
9743 /*
9744  * DTrace DIF Object Functions
9745  */
9746 static int
9747 dtrace_difo_err(uint_t pc, const char *format, ...)
9748 {
9749 	if (dtrace_err_verbose) {
9750 		va_list alist;
9751 
9752 		(void) uprintf("dtrace DIF object error: [%u]: ", pc);
9753 		va_start(alist, format);
9754 		(void) vuprintf(format, alist);
9755 		va_end(alist);
9756 	}
9757 
9758 #ifdef DTRACE_ERRDEBUG
9759 	dtrace_errdebug(format);
9760 #endif
9761 	return (1);
9762 }
9763 
9764 /*
9765  * Validate a DTrace DIF object by checking the IR instructions.  The following
9766  * rules are currently enforced by dtrace_difo_validate():
9767  *
9768  * 1. Each instruction must have a valid opcode
9769  * 2. Each register, string, variable, or subroutine reference must be valid
9770  * 3. No instruction can modify register %r0 (must be zero)
9771  * 4. All instruction reserved bits must be set to zero
9772  * 5. The last instruction must be a "ret" instruction
9773  * 6. All branch targets must reference a valid instruction _after_ the branch
9774  */
9775 static int
9776 dtrace_difo_validate(dtrace_difo_t *dp, dtrace_vstate_t *vstate, uint_t nregs,
9777     cred_t *cr)
9778 {
9779 	int err = 0, i;
9780 	int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
9781 	int kcheckload;
9782 	uint_t pc;
9783 	int maxglobal = -1, maxlocal = -1, maxtlocal = -1;
9784 
9785 	kcheckload = cr == NULL ||
9786 	    (vstate->dtvs_state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) == 0;
9787 
9788 	dp->dtdo_destructive = 0;
9789 
9790 	for (pc = 0; pc < dp->dtdo_len && err == 0; pc++) {
9791 		dif_instr_t instr = dp->dtdo_buf[pc];
9792 
9793 		uint_t r1 = DIF_INSTR_R1(instr);
9794 		uint_t r2 = DIF_INSTR_R2(instr);
9795 		uint_t rd = DIF_INSTR_RD(instr);
9796 		uint_t rs = DIF_INSTR_RS(instr);
9797 		uint_t label = DIF_INSTR_LABEL(instr);
9798 		uint_t v = DIF_INSTR_VAR(instr);
9799 		uint_t subr = DIF_INSTR_SUBR(instr);
9800 		uint_t type = DIF_INSTR_TYPE(instr);
9801 		uint_t op = DIF_INSTR_OP(instr);
9802 
9803 		switch (op) {
9804 		case DIF_OP_OR:
9805 		case DIF_OP_XOR:
9806 		case DIF_OP_AND:
9807 		case DIF_OP_SLL:
9808 		case DIF_OP_SRL:
9809 		case DIF_OP_SRA:
9810 		case DIF_OP_SUB:
9811 		case DIF_OP_ADD:
9812 		case DIF_OP_MUL:
9813 		case DIF_OP_SDIV:
9814 		case DIF_OP_UDIV:
9815 		case DIF_OP_SREM:
9816 		case DIF_OP_UREM:
9817 		case DIF_OP_COPYS:
9818 			if (r1 >= nregs)
9819 				err += efunc(pc, "invalid register %u\n", r1);
9820 			if (r2 >= nregs)
9821 				err += efunc(pc, "invalid register %u\n", r2);
9822 			if (rd >= nregs)
9823 				err += efunc(pc, "invalid register %u\n", rd);
9824 			if (rd == 0)
9825 				err += efunc(pc, "cannot write to %r0\n");
9826 			break;
9827 		case DIF_OP_NOT:
9828 		case DIF_OP_MOV:
9829 		case DIF_OP_ALLOCS:
9830 			if (r1 >= nregs)
9831 				err += efunc(pc, "invalid register %u\n", r1);
9832 			if (r2 != 0)
9833 				err += efunc(pc, "non-zero reserved bits\n");
9834 			if (rd >= nregs)
9835 				err += efunc(pc, "invalid register %u\n", rd);
9836 			if (rd == 0)
9837 				err += efunc(pc, "cannot write to %r0\n");
9838 			break;
9839 		case DIF_OP_LDSB:
9840 		case DIF_OP_LDSH:
9841 		case DIF_OP_LDSW:
9842 		case DIF_OP_LDUB:
9843 		case DIF_OP_LDUH:
9844 		case DIF_OP_LDUW:
9845 		case DIF_OP_LDX:
9846 			if (r1 >= nregs)
9847 				err += efunc(pc, "invalid register %u\n", r1);
9848 			if (r2 != 0)
9849 				err += efunc(pc, "non-zero reserved bits\n");
9850 			if (rd >= nregs)
9851 				err += efunc(pc, "invalid register %u\n", rd);
9852 			if (rd == 0)
9853 				err += efunc(pc, "cannot write to %r0\n");
9854 			if (kcheckload)
9855 				dp->dtdo_buf[pc] = DIF_INSTR_LOAD(op +
9856 				    DIF_OP_RLDSB - DIF_OP_LDSB, r1, rd);
9857 			break;
9858 		case DIF_OP_RLDSB:
9859 		case DIF_OP_RLDSH:
9860 		case DIF_OP_RLDSW:
9861 		case DIF_OP_RLDUB:
9862 		case DIF_OP_RLDUH:
9863 		case DIF_OP_RLDUW:
9864 		case DIF_OP_RLDX:
9865 			if (r1 >= nregs)
9866 				err += efunc(pc, "invalid register %u\n", r1);
9867 			if (r2 != 0)
9868 				err += efunc(pc, "non-zero reserved bits\n");
9869 			if (rd >= nregs)
9870 				err += efunc(pc, "invalid register %u\n", rd);
9871 			if (rd == 0)
9872 				err += efunc(pc, "cannot write to %r0\n");
9873 			break;
9874 		case DIF_OP_ULDSB:
9875 		case DIF_OP_ULDSH:
9876 		case DIF_OP_ULDSW:
9877 		case DIF_OP_ULDUB:
9878 		case DIF_OP_ULDUH:
9879 		case DIF_OP_ULDUW:
9880 		case DIF_OP_ULDX:
9881 			if (r1 >= nregs)
9882 				err += efunc(pc, "invalid register %u\n", r1);
9883 			if (r2 != 0)
9884 				err += efunc(pc, "non-zero reserved bits\n");
9885 			if (rd >= nregs)
9886 				err += efunc(pc, "invalid register %u\n", rd);
9887 			if (rd == 0)
9888 				err += efunc(pc, "cannot write to %r0\n");
9889 			break;
9890 		case DIF_OP_STB:
9891 		case DIF_OP_STH:
9892 		case DIF_OP_STW:
9893 		case DIF_OP_STX:
9894 			if (r1 >= nregs)
9895 				err += efunc(pc, "invalid register %u\n", r1);
9896 			if (r2 != 0)
9897 				err += efunc(pc, "non-zero reserved bits\n");
9898 			if (rd >= nregs)
9899 				err += efunc(pc, "invalid register %u\n", rd);
9900 			if (rd == 0)
9901 				err += efunc(pc, "cannot write to 0 address\n");
9902 			break;
9903 		case DIF_OP_CMP:
9904 		case DIF_OP_SCMP:
9905 			if (r1 >= nregs)
9906 				err += efunc(pc, "invalid register %u\n", r1);
9907 			if (r2 >= nregs)
9908 				err += efunc(pc, "invalid register %u\n", r2);
9909 			if (rd != 0)
9910 				err += efunc(pc, "non-zero reserved bits\n");
9911 			break;
9912 		case DIF_OP_TST:
9913 			if (r1 >= nregs)
9914 				err += efunc(pc, "invalid register %u\n", r1);
9915 			if (r2 != 0 || rd != 0)
9916 				err += efunc(pc, "non-zero reserved bits\n");
9917 			break;
9918 		case DIF_OP_BA:
9919 		case DIF_OP_BE:
9920 		case DIF_OP_BNE:
9921 		case DIF_OP_BG:
9922 		case DIF_OP_BGU:
9923 		case DIF_OP_BGE:
9924 		case DIF_OP_BGEU:
9925 		case DIF_OP_BL:
9926 		case DIF_OP_BLU:
9927 		case DIF_OP_BLE:
9928 		case DIF_OP_BLEU:
9929 			if (label >= dp->dtdo_len) {
9930 				err += efunc(pc, "invalid branch target %u\n",
9931 				    label);
9932 			}
9933 			if (label <= pc) {
9934 				err += efunc(pc, "backward branch to %u\n",
9935 				    label);
9936 			}
9937 			break;
9938 		case DIF_OP_RET:
9939 			if (r1 != 0 || r2 != 0)
9940 				err += efunc(pc, "non-zero reserved bits\n");
9941 			if (rd >= nregs)
9942 				err += efunc(pc, "invalid register %u\n", rd);
9943 			break;
9944 		case DIF_OP_NOP:
9945 		case DIF_OP_POPTS:
9946 		case DIF_OP_FLUSHTS:
9947 			if (r1 != 0 || r2 != 0 || rd != 0)
9948 				err += efunc(pc, "non-zero reserved bits\n");
9949 			break;
9950 		case DIF_OP_SETX:
9951 			if (DIF_INSTR_INTEGER(instr) >= dp->dtdo_intlen) {
9952 				err += efunc(pc, "invalid integer ref %u\n",
9953 				    DIF_INSTR_INTEGER(instr));
9954 			}
9955 			if (rd >= nregs)
9956 				err += efunc(pc, "invalid register %u\n", rd);
9957 			if (rd == 0)
9958 				err += efunc(pc, "cannot write to %r0\n");
9959 			break;
9960 		case DIF_OP_SETS:
9961 			if (DIF_INSTR_STRING(instr) >= dp->dtdo_strlen) {
9962 				err += efunc(pc, "invalid string ref %u\n",
9963 				    DIF_INSTR_STRING(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_LDGA:
9971 		case DIF_OP_LDTA:
9972 			if (r1 > DIF_VAR_ARRAY_MAX)
9973 				err += efunc(pc, "invalid array %u\n", r1);
9974 			if (r2 >= nregs)
9975 				err += efunc(pc, "invalid register %u\n", r2);
9976 			if (rd >= nregs)
9977 				err += efunc(pc, "invalid register %u\n", rd);
9978 			if (rd == 0)
9979 				err += efunc(pc, "cannot write to %r0\n");
9980 			break;
9981 		case DIF_OP_LDGS:
9982 		case DIF_OP_LDTS:
9983 		case DIF_OP_LDLS:
9984 		case DIF_OP_LDGAA:
9985 		case DIF_OP_LDTAA:
9986 			if (v < DIF_VAR_OTHER_MIN || v > DIF_VAR_OTHER_MAX)
9987 				err += efunc(pc, "invalid variable %u\n", v);
9988 			if (rd >= nregs)
9989 				err += efunc(pc, "invalid register %u\n", rd);
9990 			if (rd == 0)
9991 				err += efunc(pc, "cannot write to %r0\n");
9992 			break;
9993 		case DIF_OP_STGS:
9994 		case DIF_OP_STTS:
9995 		case DIF_OP_STLS:
9996 		case DIF_OP_STGAA:
9997 		case DIF_OP_STTAA:
9998 			if (v < DIF_VAR_OTHER_UBASE || v > DIF_VAR_OTHER_MAX)
9999 				err += efunc(pc, "invalid variable %u\n", v);
10000 			if (rs >= nregs)
10001 				err += efunc(pc, "invalid register %u\n", rd);
10002 			break;
10003 		case DIF_OP_CALL:
10004 			if (subr > DIF_SUBR_MAX)
10005 				err += efunc(pc, "invalid subr %u\n", subr);
10006 			if (rd >= nregs)
10007 				err += efunc(pc, "invalid register %u\n", rd);
10008 			if (rd == 0)
10009 				err += efunc(pc, "cannot write to %r0\n");
10010 
10011 			if (subr == DIF_SUBR_COPYOUT ||
10012 			    subr == DIF_SUBR_COPYOUTSTR) {
10013 				dp->dtdo_destructive = 1;
10014 			}
10015 
10016 			if (subr == DIF_SUBR_GETF) {
10017 				/*
10018 				 * If we have a getf() we need to record that
10019 				 * in our state.  Note that our state can be
10020 				 * NULL if this is a helper -- but in that
10021 				 * case, the call to getf() is itself illegal,
10022 				 * and will be caught (slightly later) when
10023 				 * the helper is validated.
10024 				 */
10025 				if (vstate->dtvs_state != NULL)
10026 					vstate->dtvs_state->dts_getf++;
10027 			}
10028 
10029 			break;
10030 		case DIF_OP_PUSHTR:
10031 			if (type != DIF_TYPE_STRING && type != DIF_TYPE_CTF)
10032 				err += efunc(pc, "invalid ref type %u\n", type);
10033 			if (r2 >= nregs)
10034 				err += efunc(pc, "invalid register %u\n", r2);
10035 			if (rs >= nregs)
10036 				err += efunc(pc, "invalid register %u\n", rs);
10037 			break;
10038 		case DIF_OP_PUSHTV:
10039 			if (type != DIF_TYPE_CTF)
10040 				err += efunc(pc, "invalid val type %u\n", type);
10041 			if (r2 >= nregs)
10042 				err += efunc(pc, "invalid register %u\n", r2);
10043 			if (rs >= nregs)
10044 				err += efunc(pc, "invalid register %u\n", rs);
10045 			break;
10046 		default:
10047 			err += efunc(pc, "invalid opcode %u\n",
10048 			    DIF_INSTR_OP(instr));
10049 		}
10050 	}
10051 
10052 	if (dp->dtdo_len != 0 &&
10053 	    DIF_INSTR_OP(dp->dtdo_buf[dp->dtdo_len - 1]) != DIF_OP_RET) {
10054 		err += efunc(dp->dtdo_len - 1,
10055 		    "expected 'ret' as last DIF instruction\n");
10056 	}
10057 
10058 	if (!(dp->dtdo_rtype.dtdt_flags & (DIF_TF_BYREF | DIF_TF_BYUREF))) {
10059 		/*
10060 		 * If we're not returning by reference, the size must be either
10061 		 * 0 or the size of one of the base types.
10062 		 */
10063 		switch (dp->dtdo_rtype.dtdt_size) {
10064 		case 0:
10065 		case sizeof (uint8_t):
10066 		case sizeof (uint16_t):
10067 		case sizeof (uint32_t):
10068 		case sizeof (uint64_t):
10069 			break;
10070 
10071 		default:
10072 			err += efunc(dp->dtdo_len - 1, "bad return size\n");
10073 		}
10074 	}
10075 
10076 	for (i = 0; i < dp->dtdo_varlen && err == 0; i++) {
10077 		dtrace_difv_t *v = &dp->dtdo_vartab[i], *existing = NULL;
10078 		dtrace_diftype_t *vt, *et;
10079 		uint_t id, ndx;
10080 
10081 		if (v->dtdv_scope != DIFV_SCOPE_GLOBAL &&
10082 		    v->dtdv_scope != DIFV_SCOPE_THREAD &&
10083 		    v->dtdv_scope != DIFV_SCOPE_LOCAL) {
10084 			err += efunc(i, "unrecognized variable scope %d\n",
10085 			    v->dtdv_scope);
10086 			break;
10087 		}
10088 
10089 		if (v->dtdv_kind != DIFV_KIND_ARRAY &&
10090 		    v->dtdv_kind != DIFV_KIND_SCALAR) {
10091 			err += efunc(i, "unrecognized variable type %d\n",
10092 			    v->dtdv_kind);
10093 			break;
10094 		}
10095 
10096 		if ((id = v->dtdv_id) > DIF_VARIABLE_MAX) {
10097 			err += efunc(i, "%d exceeds variable id limit\n", id);
10098 			break;
10099 		}
10100 
10101 		if (id < DIF_VAR_OTHER_UBASE)
10102 			continue;
10103 
10104 		/*
10105 		 * For user-defined variables, we need to check that this
10106 		 * definition is identical to any previous definition that we
10107 		 * encountered.
10108 		 */
10109 		ndx = id - DIF_VAR_OTHER_UBASE;
10110 
10111 		switch (v->dtdv_scope) {
10112 		case DIFV_SCOPE_GLOBAL:
10113 			if (maxglobal == -1 || ndx > maxglobal)
10114 				maxglobal = ndx;
10115 
10116 			if (ndx < vstate->dtvs_nglobals) {
10117 				dtrace_statvar_t *svar;
10118 
10119 				if ((svar = vstate->dtvs_globals[ndx]) != NULL)
10120 					existing = &svar->dtsv_var;
10121 			}
10122 
10123 			break;
10124 
10125 		case DIFV_SCOPE_THREAD:
10126 			if (maxtlocal == -1 || ndx > maxtlocal)
10127 				maxtlocal = ndx;
10128 
10129 			if (ndx < vstate->dtvs_ntlocals)
10130 				existing = &vstate->dtvs_tlocals[ndx];
10131 			break;
10132 
10133 		case DIFV_SCOPE_LOCAL:
10134 			if (maxlocal == -1 || ndx > maxlocal)
10135 				maxlocal = ndx;
10136 
10137 			if (ndx < vstate->dtvs_nlocals) {
10138 				dtrace_statvar_t *svar;
10139 
10140 				if ((svar = vstate->dtvs_locals[ndx]) != NULL)
10141 					existing = &svar->dtsv_var;
10142 			}
10143 
10144 			break;
10145 		}
10146 
10147 		vt = &v->dtdv_type;
10148 
10149 		if (vt->dtdt_flags & DIF_TF_BYREF) {
10150 			if (vt->dtdt_size == 0) {
10151 				err += efunc(i, "zero-sized variable\n");
10152 				break;
10153 			}
10154 
10155 			if ((v->dtdv_scope == DIFV_SCOPE_GLOBAL ||
10156 			    v->dtdv_scope == DIFV_SCOPE_LOCAL) &&
10157 			    vt->dtdt_size > dtrace_statvar_maxsize) {
10158 				err += efunc(i, "oversized by-ref static\n");
10159 				break;
10160 			}
10161 		}
10162 
10163 		if (existing == NULL || existing->dtdv_id == 0)
10164 			continue;
10165 
10166 		ASSERT(existing->dtdv_id == v->dtdv_id);
10167 		ASSERT(existing->dtdv_scope == v->dtdv_scope);
10168 
10169 		if (existing->dtdv_kind != v->dtdv_kind)
10170 			err += efunc(i, "%d changed variable kind\n", id);
10171 
10172 		et = &existing->dtdv_type;
10173 
10174 		if (vt->dtdt_flags != et->dtdt_flags) {
10175 			err += efunc(i, "%d changed variable type flags\n", id);
10176 			break;
10177 		}
10178 
10179 		if (vt->dtdt_size != 0 && vt->dtdt_size != et->dtdt_size) {
10180 			err += efunc(i, "%d changed variable type size\n", id);
10181 			break;
10182 		}
10183 	}
10184 
10185 	for (pc = 0; pc < dp->dtdo_len && err == 0; pc++) {
10186 		dif_instr_t instr = dp->dtdo_buf[pc];
10187 
10188 		uint_t v = DIF_INSTR_VAR(instr);
10189 		uint_t op = DIF_INSTR_OP(instr);
10190 
10191 		switch (op) {
10192 		case DIF_OP_LDGS:
10193 		case DIF_OP_LDGAA:
10194 		case DIF_OP_STGS:
10195 		case DIF_OP_STGAA:
10196 			if (v > DIF_VAR_OTHER_UBASE + maxglobal)
10197 				err += efunc(pc, "invalid variable %u\n", v);
10198 			break;
10199 		case DIF_OP_LDTS:
10200 		case DIF_OP_LDTAA:
10201 		case DIF_OP_STTS:
10202 		case DIF_OP_STTAA:
10203 			if (v > DIF_VAR_OTHER_UBASE + maxtlocal)
10204 				err += efunc(pc, "invalid variable %u\n", v);
10205 			break;
10206 		case DIF_OP_LDLS:
10207 		case DIF_OP_STLS:
10208 			if (v > DIF_VAR_OTHER_UBASE + maxlocal)
10209 				err += efunc(pc, "invalid variable %u\n", v);
10210 			break;
10211 		default:
10212 			break;
10213 		}
10214 	}
10215 
10216 	return (err);
10217 }
10218 
10219 /*
10220  * Validate a DTrace DIF object that it is to be used as a helper.  Helpers
10221  * are much more constrained than normal DIFOs.  Specifically, they may
10222  * not:
10223  *
10224  * 1. Make calls to subroutines other than copyin(), copyinstr() or
10225  *    miscellaneous string routines
10226  * 2. Access DTrace variables other than the args[] array, and the
10227  *    curthread, pid, ppid, tid, execname, zonename, uid and gid variables.
10228  * 3. Have thread-local variables.
10229  * 4. Have dynamic variables.
10230  */
10231 static int
10232 dtrace_difo_validate_helper(dtrace_difo_t *dp)
10233 {
10234 	int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
10235 	int err = 0;
10236 	uint_t pc;
10237 
10238 	for (pc = 0; pc < dp->dtdo_len; pc++) {
10239 		dif_instr_t instr = dp->dtdo_buf[pc];
10240 
10241 		uint_t v = DIF_INSTR_VAR(instr);
10242 		uint_t subr = DIF_INSTR_SUBR(instr);
10243 		uint_t op = DIF_INSTR_OP(instr);
10244 
10245 		switch (op) {
10246 		case DIF_OP_OR:
10247 		case DIF_OP_XOR:
10248 		case DIF_OP_AND:
10249 		case DIF_OP_SLL:
10250 		case DIF_OP_SRL:
10251 		case DIF_OP_SRA:
10252 		case DIF_OP_SUB:
10253 		case DIF_OP_ADD:
10254 		case DIF_OP_MUL:
10255 		case DIF_OP_SDIV:
10256 		case DIF_OP_UDIV:
10257 		case DIF_OP_SREM:
10258 		case DIF_OP_UREM:
10259 		case DIF_OP_COPYS:
10260 		case DIF_OP_NOT:
10261 		case DIF_OP_MOV:
10262 		case DIF_OP_RLDSB:
10263 		case DIF_OP_RLDSH:
10264 		case DIF_OP_RLDSW:
10265 		case DIF_OP_RLDUB:
10266 		case DIF_OP_RLDUH:
10267 		case DIF_OP_RLDUW:
10268 		case DIF_OP_RLDX:
10269 		case DIF_OP_ULDSB:
10270 		case DIF_OP_ULDSH:
10271 		case DIF_OP_ULDSW:
10272 		case DIF_OP_ULDUB:
10273 		case DIF_OP_ULDUH:
10274 		case DIF_OP_ULDUW:
10275 		case DIF_OP_ULDX:
10276 		case DIF_OP_STB:
10277 		case DIF_OP_STH:
10278 		case DIF_OP_STW:
10279 		case DIF_OP_STX:
10280 		case DIF_OP_ALLOCS:
10281 		case DIF_OP_CMP:
10282 		case DIF_OP_SCMP:
10283 		case DIF_OP_TST:
10284 		case DIF_OP_BA:
10285 		case DIF_OP_BE:
10286 		case DIF_OP_BNE:
10287 		case DIF_OP_BG:
10288 		case DIF_OP_BGU:
10289 		case DIF_OP_BGE:
10290 		case DIF_OP_BGEU:
10291 		case DIF_OP_BL:
10292 		case DIF_OP_BLU:
10293 		case DIF_OP_BLE:
10294 		case DIF_OP_BLEU:
10295 		case DIF_OP_RET:
10296 		case DIF_OP_NOP:
10297 		case DIF_OP_POPTS:
10298 		case DIF_OP_FLUSHTS:
10299 		case DIF_OP_SETX:
10300 		case DIF_OP_SETS:
10301 		case DIF_OP_LDGA:
10302 		case DIF_OP_LDLS:
10303 		case DIF_OP_STGS:
10304 		case DIF_OP_STLS:
10305 		case DIF_OP_PUSHTR:
10306 		case DIF_OP_PUSHTV:
10307 			break;
10308 
10309 		case DIF_OP_LDGS:
10310 			if (v >= DIF_VAR_OTHER_UBASE)
10311 				break;
10312 
10313 			if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9)
10314 				break;
10315 
10316 			if (v == DIF_VAR_CURTHREAD || v == DIF_VAR_PID ||
10317 			    v == DIF_VAR_PPID || v == DIF_VAR_TID ||
10318 			    v == DIF_VAR_EXECARGS ||
10319 			    v == DIF_VAR_EXECNAME || v == DIF_VAR_ZONENAME ||
10320 			    v == DIF_VAR_UID || v == DIF_VAR_GID)
10321 				break;
10322 
10323 			err += efunc(pc, "illegal variable %u\n", v);
10324 			break;
10325 
10326 		case DIF_OP_LDTA:
10327 		case DIF_OP_LDTS:
10328 		case DIF_OP_LDGAA:
10329 		case DIF_OP_LDTAA:
10330 			err += efunc(pc, "illegal dynamic variable load\n");
10331 			break;
10332 
10333 		case DIF_OP_STTS:
10334 		case DIF_OP_STGAA:
10335 		case DIF_OP_STTAA:
10336 			err += efunc(pc, "illegal dynamic variable store\n");
10337 			break;
10338 
10339 		case DIF_OP_CALL:
10340 			if (subr == DIF_SUBR_ALLOCA ||
10341 			    subr == DIF_SUBR_BCOPY ||
10342 			    subr == DIF_SUBR_COPYIN ||
10343 			    subr == DIF_SUBR_COPYINTO ||
10344 			    subr == DIF_SUBR_COPYINSTR ||
10345 			    subr == DIF_SUBR_INDEX ||
10346 			    subr == DIF_SUBR_INET_NTOA ||
10347 			    subr == DIF_SUBR_INET_NTOA6 ||
10348 			    subr == DIF_SUBR_INET_NTOP ||
10349 			    subr == DIF_SUBR_JSON ||
10350 			    subr == DIF_SUBR_LLTOSTR ||
10351 			    subr == DIF_SUBR_STRTOLL ||
10352 			    subr == DIF_SUBR_RINDEX ||
10353 			    subr == DIF_SUBR_STRCHR ||
10354 			    subr == DIF_SUBR_STRJOIN ||
10355 			    subr == DIF_SUBR_STRRCHR ||
10356 			    subr == DIF_SUBR_STRSTR ||
10357 			    subr == DIF_SUBR_HTONS ||
10358 			    subr == DIF_SUBR_HTONL ||
10359 			    subr == DIF_SUBR_HTONLL ||
10360 			    subr == DIF_SUBR_NTOHS ||
10361 			    subr == DIF_SUBR_NTOHL ||
10362 			    subr == DIF_SUBR_NTOHLL ||
10363 			    subr == DIF_SUBR_MEMREF)
10364 				break;
10365 #ifdef __FreeBSD__
10366 			if (subr == DIF_SUBR_MEMSTR)
10367 				break;
10368 #endif
10369 
10370 			err += efunc(pc, "invalid subr %u\n", subr);
10371 			break;
10372 
10373 		default:
10374 			err += efunc(pc, "invalid opcode %u\n",
10375 			    DIF_INSTR_OP(instr));
10376 		}
10377 	}
10378 
10379 	return (err);
10380 }
10381 
10382 /*
10383  * Returns 1 if the expression in the DIF object can be cached on a per-thread
10384  * basis; 0 if not.
10385  */
10386 static int
10387 dtrace_difo_cacheable(dtrace_difo_t *dp)
10388 {
10389 	int i;
10390 
10391 	if (dp == NULL)
10392 		return (0);
10393 
10394 	for (i = 0; i < dp->dtdo_varlen; i++) {
10395 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
10396 
10397 		if (v->dtdv_scope != DIFV_SCOPE_GLOBAL)
10398 			continue;
10399 
10400 		switch (v->dtdv_id) {
10401 		case DIF_VAR_CURTHREAD:
10402 		case DIF_VAR_PID:
10403 		case DIF_VAR_TID:
10404 		case DIF_VAR_EXECARGS:
10405 		case DIF_VAR_EXECNAME:
10406 		case DIF_VAR_ZONENAME:
10407 			break;
10408 
10409 		default:
10410 			return (0);
10411 		}
10412 	}
10413 
10414 	/*
10415 	 * This DIF object may be cacheable.  Now we need to look for any
10416 	 * array loading instructions, any memory loading instructions, or
10417 	 * any stores to thread-local variables.
10418 	 */
10419 	for (i = 0; i < dp->dtdo_len; i++) {
10420 		uint_t op = DIF_INSTR_OP(dp->dtdo_buf[i]);
10421 
10422 		if ((op >= DIF_OP_LDSB && op <= DIF_OP_LDX) ||
10423 		    (op >= DIF_OP_ULDSB && op <= DIF_OP_ULDX) ||
10424 		    (op >= DIF_OP_RLDSB && op <= DIF_OP_RLDX) ||
10425 		    op == DIF_OP_LDGA || op == DIF_OP_STTS)
10426 			return (0);
10427 	}
10428 
10429 	return (1);
10430 }
10431 
10432 static void
10433 dtrace_difo_hold(dtrace_difo_t *dp)
10434 {
10435 	int i;
10436 
10437 	ASSERT(MUTEX_HELD(&dtrace_lock));
10438 
10439 	dp->dtdo_refcnt++;
10440 	ASSERT(dp->dtdo_refcnt != 0);
10441 
10442 	/*
10443 	 * We need to check this DIF object for references to the variable
10444 	 * DIF_VAR_VTIMESTAMP.
10445 	 */
10446 	for (i = 0; i < dp->dtdo_varlen; i++) {
10447 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
10448 
10449 		if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
10450 			continue;
10451 
10452 		if (dtrace_vtime_references++ == 0)
10453 			dtrace_vtime_enable();
10454 	}
10455 }
10456 
10457 /*
10458  * This routine calculates the dynamic variable chunksize for a given DIF
10459  * object.  The calculation is not fool-proof, and can probably be tricked by
10460  * malicious DIF -- but it works for all compiler-generated DIF.  Because this
10461  * calculation is likely imperfect, dtrace_dynvar() is able to gracefully fail
10462  * if a dynamic variable size exceeds the chunksize.
10463  */
10464 static void
10465 dtrace_difo_chunksize(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
10466 {
10467 	uint64_t sval = 0;
10468 	dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
10469 	const dif_instr_t *text = dp->dtdo_buf;
10470 	uint_t pc, srd = 0;
10471 	uint_t ttop = 0;
10472 	size_t size, ksize;
10473 	uint_t id, i;
10474 
10475 	for (pc = 0; pc < dp->dtdo_len; pc++) {
10476 		dif_instr_t instr = text[pc];
10477 		uint_t op = DIF_INSTR_OP(instr);
10478 		uint_t rd = DIF_INSTR_RD(instr);
10479 		uint_t r1 = DIF_INSTR_R1(instr);
10480 		uint_t nkeys = 0;
10481 		uchar_t scope = 0;
10482 
10483 		dtrace_key_t *key = tupregs;
10484 
10485 		switch (op) {
10486 		case DIF_OP_SETX:
10487 			sval = dp->dtdo_inttab[DIF_INSTR_INTEGER(instr)];
10488 			srd = rd;
10489 			continue;
10490 
10491 		case DIF_OP_STTS:
10492 			key = &tupregs[DIF_DTR_NREGS];
10493 			key[0].dttk_size = 0;
10494 			key[1].dttk_size = 0;
10495 			nkeys = 2;
10496 			scope = DIFV_SCOPE_THREAD;
10497 			break;
10498 
10499 		case DIF_OP_STGAA:
10500 		case DIF_OP_STTAA:
10501 			nkeys = ttop;
10502 
10503 			if (DIF_INSTR_OP(instr) == DIF_OP_STTAA)
10504 				key[nkeys++].dttk_size = 0;
10505 
10506 			key[nkeys++].dttk_size = 0;
10507 
10508 			if (op == DIF_OP_STTAA) {
10509 				scope = DIFV_SCOPE_THREAD;
10510 			} else {
10511 				scope = DIFV_SCOPE_GLOBAL;
10512 			}
10513 
10514 			break;
10515 
10516 		case DIF_OP_PUSHTR:
10517 			if (ttop == DIF_DTR_NREGS)
10518 				return;
10519 
10520 			if ((srd == 0 || sval == 0) && r1 == DIF_TYPE_STRING) {
10521 				/*
10522 				 * If the register for the size of the "pushtr"
10523 				 * is %r0 (or the value is 0) and the type is
10524 				 * a string, we'll use the system-wide default
10525 				 * string size.
10526 				 */
10527 				tupregs[ttop++].dttk_size =
10528 				    dtrace_strsize_default;
10529 			} else {
10530 				if (srd == 0)
10531 					return;
10532 
10533 				if (sval > LONG_MAX)
10534 					return;
10535 
10536 				tupregs[ttop++].dttk_size = sval;
10537 			}
10538 
10539 			break;
10540 
10541 		case DIF_OP_PUSHTV:
10542 			if (ttop == DIF_DTR_NREGS)
10543 				return;
10544 
10545 			tupregs[ttop++].dttk_size = 0;
10546 			break;
10547 
10548 		case DIF_OP_FLUSHTS:
10549 			ttop = 0;
10550 			break;
10551 
10552 		case DIF_OP_POPTS:
10553 			if (ttop != 0)
10554 				ttop--;
10555 			break;
10556 		}
10557 
10558 		sval = 0;
10559 		srd = 0;
10560 
10561 		if (nkeys == 0)
10562 			continue;
10563 
10564 		/*
10565 		 * We have a dynamic variable allocation; calculate its size.
10566 		 */
10567 		for (ksize = 0, i = 0; i < nkeys; i++)
10568 			ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
10569 
10570 		size = sizeof (dtrace_dynvar_t);
10571 		size += sizeof (dtrace_key_t) * (nkeys - 1);
10572 		size += ksize;
10573 
10574 		/*
10575 		 * Now we need to determine the size of the stored data.
10576 		 */
10577 		id = DIF_INSTR_VAR(instr);
10578 
10579 		for (i = 0; i < dp->dtdo_varlen; i++) {
10580 			dtrace_difv_t *v = &dp->dtdo_vartab[i];
10581 
10582 			if (v->dtdv_id == id && v->dtdv_scope == scope) {
10583 				size += v->dtdv_type.dtdt_size;
10584 				break;
10585 			}
10586 		}
10587 
10588 		if (i == dp->dtdo_varlen)
10589 			return;
10590 
10591 		/*
10592 		 * We have the size.  If this is larger than the chunk size
10593 		 * for our dynamic variable state, reset the chunk size.
10594 		 */
10595 		size = P2ROUNDUP(size, sizeof (uint64_t));
10596 
10597 		/*
10598 		 * Before setting the chunk size, check that we're not going
10599 		 * to set it to a negative value...
10600 		 */
10601 		if (size > LONG_MAX)
10602 			return;
10603 
10604 		/*
10605 		 * ...and make certain that we didn't badly overflow.
10606 		 */
10607 		if (size < ksize || size < sizeof (dtrace_dynvar_t))
10608 			return;
10609 
10610 		if (size > vstate->dtvs_dynvars.dtds_chunksize)
10611 			vstate->dtvs_dynvars.dtds_chunksize = size;
10612 	}
10613 }
10614 
10615 static void
10616 dtrace_difo_init(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
10617 {
10618 	int i, oldsvars, osz, nsz, otlocals, ntlocals;
10619 	uint_t id;
10620 
10621 	ASSERT(MUTEX_HELD(&dtrace_lock));
10622 	ASSERT(dp->dtdo_buf != NULL && dp->dtdo_len != 0);
10623 
10624 	for (i = 0; i < dp->dtdo_varlen; i++) {
10625 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
10626 		dtrace_statvar_t *svar, ***svarp = NULL;
10627 		size_t dsize = 0;
10628 		uint8_t scope = v->dtdv_scope;
10629 		int *np = NULL;
10630 
10631 		if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
10632 			continue;
10633 
10634 		id -= DIF_VAR_OTHER_UBASE;
10635 
10636 		switch (scope) {
10637 		case DIFV_SCOPE_THREAD:
10638 			while (id >= (otlocals = vstate->dtvs_ntlocals)) {
10639 				dtrace_difv_t *tlocals;
10640 
10641 				if ((ntlocals = (otlocals << 1)) == 0)
10642 					ntlocals = 1;
10643 
10644 				osz = otlocals * sizeof (dtrace_difv_t);
10645 				nsz = ntlocals * sizeof (dtrace_difv_t);
10646 
10647 				tlocals = kmem_zalloc(nsz, KM_SLEEP);
10648 
10649 				if (osz != 0) {
10650 					bcopy(vstate->dtvs_tlocals,
10651 					    tlocals, osz);
10652 					kmem_free(vstate->dtvs_tlocals, osz);
10653 				}
10654 
10655 				vstate->dtvs_tlocals = tlocals;
10656 				vstate->dtvs_ntlocals = ntlocals;
10657 			}
10658 
10659 			vstate->dtvs_tlocals[id] = *v;
10660 			continue;
10661 
10662 		case DIFV_SCOPE_LOCAL:
10663 			np = &vstate->dtvs_nlocals;
10664 			svarp = &vstate->dtvs_locals;
10665 
10666 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
10667 				dsize = NCPU * (v->dtdv_type.dtdt_size +
10668 				    sizeof (uint64_t));
10669 			else
10670 				dsize = NCPU * sizeof (uint64_t);
10671 
10672 			break;
10673 
10674 		case DIFV_SCOPE_GLOBAL:
10675 			np = &vstate->dtvs_nglobals;
10676 			svarp = &vstate->dtvs_globals;
10677 
10678 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
10679 				dsize = v->dtdv_type.dtdt_size +
10680 				    sizeof (uint64_t);
10681 
10682 			break;
10683 
10684 		default:
10685 			ASSERT(0);
10686 		}
10687 
10688 		while (id >= (oldsvars = *np)) {
10689 			dtrace_statvar_t **statics;
10690 			int newsvars, oldsize, newsize;
10691 
10692 			if ((newsvars = (oldsvars << 1)) == 0)
10693 				newsvars = 1;
10694 
10695 			oldsize = oldsvars * sizeof (dtrace_statvar_t *);
10696 			newsize = newsvars * sizeof (dtrace_statvar_t *);
10697 
10698 			statics = kmem_zalloc(newsize, KM_SLEEP);
10699 
10700 			if (oldsize != 0) {
10701 				bcopy(*svarp, statics, oldsize);
10702 				kmem_free(*svarp, oldsize);
10703 			}
10704 
10705 			*svarp = statics;
10706 			*np = newsvars;
10707 		}
10708 
10709 		if ((svar = (*svarp)[id]) == NULL) {
10710 			svar = kmem_zalloc(sizeof (dtrace_statvar_t), KM_SLEEP);
10711 			svar->dtsv_var = *v;
10712 
10713 			if ((svar->dtsv_size = dsize) != 0) {
10714 				svar->dtsv_data = (uint64_t)(uintptr_t)
10715 				    kmem_zalloc(dsize, KM_SLEEP);
10716 			}
10717 
10718 			(*svarp)[id] = svar;
10719 		}
10720 
10721 		svar->dtsv_refcnt++;
10722 	}
10723 
10724 	dtrace_difo_chunksize(dp, vstate);
10725 	dtrace_difo_hold(dp);
10726 }
10727 
10728 static dtrace_difo_t *
10729 dtrace_difo_duplicate(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
10730 {
10731 	dtrace_difo_t *new;
10732 	size_t sz;
10733 
10734 	ASSERT(dp->dtdo_buf != NULL);
10735 	ASSERT(dp->dtdo_refcnt != 0);
10736 
10737 	new = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
10738 
10739 	ASSERT(dp->dtdo_buf != NULL);
10740 	sz = dp->dtdo_len * sizeof (dif_instr_t);
10741 	new->dtdo_buf = kmem_alloc(sz, KM_SLEEP);
10742 	bcopy(dp->dtdo_buf, new->dtdo_buf, sz);
10743 	new->dtdo_len = dp->dtdo_len;
10744 
10745 	if (dp->dtdo_strtab != NULL) {
10746 		ASSERT(dp->dtdo_strlen != 0);
10747 		new->dtdo_strtab = kmem_alloc(dp->dtdo_strlen, KM_SLEEP);
10748 		bcopy(dp->dtdo_strtab, new->dtdo_strtab, dp->dtdo_strlen);
10749 		new->dtdo_strlen = dp->dtdo_strlen;
10750 	}
10751 
10752 	if (dp->dtdo_inttab != NULL) {
10753 		ASSERT(dp->dtdo_intlen != 0);
10754 		sz = dp->dtdo_intlen * sizeof (uint64_t);
10755 		new->dtdo_inttab = kmem_alloc(sz, KM_SLEEP);
10756 		bcopy(dp->dtdo_inttab, new->dtdo_inttab, sz);
10757 		new->dtdo_intlen = dp->dtdo_intlen;
10758 	}
10759 
10760 	if (dp->dtdo_vartab != NULL) {
10761 		ASSERT(dp->dtdo_varlen != 0);
10762 		sz = dp->dtdo_varlen * sizeof (dtrace_difv_t);
10763 		new->dtdo_vartab = kmem_alloc(sz, KM_SLEEP);
10764 		bcopy(dp->dtdo_vartab, new->dtdo_vartab, sz);
10765 		new->dtdo_varlen = dp->dtdo_varlen;
10766 	}
10767 
10768 	dtrace_difo_init(new, vstate);
10769 	return (new);
10770 }
10771 
10772 static void
10773 dtrace_difo_destroy(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
10774 {
10775 	int i;
10776 
10777 	ASSERT(dp->dtdo_refcnt == 0);
10778 
10779 	for (i = 0; i < dp->dtdo_varlen; i++) {
10780 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
10781 		dtrace_statvar_t *svar, **svarp = NULL;
10782 		uint_t id;
10783 		uint8_t scope = v->dtdv_scope;
10784 		int *np = NULL;
10785 
10786 		switch (scope) {
10787 		case DIFV_SCOPE_THREAD:
10788 			continue;
10789 
10790 		case DIFV_SCOPE_LOCAL:
10791 			np = &vstate->dtvs_nlocals;
10792 			svarp = vstate->dtvs_locals;
10793 			break;
10794 
10795 		case DIFV_SCOPE_GLOBAL:
10796 			np = &vstate->dtvs_nglobals;
10797 			svarp = vstate->dtvs_globals;
10798 			break;
10799 
10800 		default:
10801 			ASSERT(0);
10802 		}
10803 
10804 		if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
10805 			continue;
10806 
10807 		id -= DIF_VAR_OTHER_UBASE;
10808 		ASSERT(id < *np);
10809 
10810 		svar = svarp[id];
10811 		ASSERT(svar != NULL);
10812 		ASSERT(svar->dtsv_refcnt > 0);
10813 
10814 		if (--svar->dtsv_refcnt > 0)
10815 			continue;
10816 
10817 		if (svar->dtsv_size != 0) {
10818 			ASSERT(svar->dtsv_data != 0);
10819 			kmem_free((void *)(uintptr_t)svar->dtsv_data,
10820 			    svar->dtsv_size);
10821 		}
10822 
10823 		kmem_free(svar, sizeof (dtrace_statvar_t));
10824 		svarp[id] = NULL;
10825 	}
10826 
10827 	if (dp->dtdo_buf != NULL)
10828 		kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
10829 	if (dp->dtdo_inttab != NULL)
10830 		kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
10831 	if (dp->dtdo_strtab != NULL)
10832 		kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
10833 	if (dp->dtdo_vartab != NULL)
10834 		kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
10835 
10836 	kmem_free(dp, sizeof (dtrace_difo_t));
10837 }
10838 
10839 static void
10840 dtrace_difo_release(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
10841 {
10842 	int i;
10843 
10844 	ASSERT(MUTEX_HELD(&dtrace_lock));
10845 	ASSERT(dp->dtdo_refcnt != 0);
10846 
10847 	for (i = 0; i < dp->dtdo_varlen; i++) {
10848 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
10849 
10850 		if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
10851 			continue;
10852 
10853 		ASSERT(dtrace_vtime_references > 0);
10854 		if (--dtrace_vtime_references == 0)
10855 			dtrace_vtime_disable();
10856 	}
10857 
10858 	if (--dp->dtdo_refcnt == 0)
10859 		dtrace_difo_destroy(dp, vstate);
10860 }
10861 
10862 /*
10863  * DTrace Format Functions
10864  */
10865 static uint16_t
10866 dtrace_format_add(dtrace_state_t *state, char *str)
10867 {
10868 	char *fmt, **new;
10869 	uint16_t ndx, len = strlen(str) + 1;
10870 
10871 	fmt = kmem_zalloc(len, KM_SLEEP);
10872 	bcopy(str, fmt, len);
10873 
10874 	for (ndx = 0; ndx < state->dts_nformats; ndx++) {
10875 		if (state->dts_formats[ndx] == NULL) {
10876 			state->dts_formats[ndx] = fmt;
10877 			return (ndx + 1);
10878 		}
10879 	}
10880 
10881 	if (state->dts_nformats == USHRT_MAX) {
10882 		/*
10883 		 * This is only likely if a denial-of-service attack is being
10884 		 * attempted.  As such, it's okay to fail silently here.
10885 		 */
10886 		kmem_free(fmt, len);
10887 		return (0);
10888 	}
10889 
10890 	/*
10891 	 * For simplicity, we always resize the formats array to be exactly the
10892 	 * number of formats.
10893 	 */
10894 	ndx = state->dts_nformats++;
10895 	new = kmem_alloc((ndx + 1) * sizeof (char *), KM_SLEEP);
10896 
10897 	if (state->dts_formats != NULL) {
10898 		ASSERT(ndx != 0);
10899 		bcopy(state->dts_formats, new, ndx * sizeof (char *));
10900 		kmem_free(state->dts_formats, ndx * sizeof (char *));
10901 	}
10902 
10903 	state->dts_formats = new;
10904 	state->dts_formats[ndx] = fmt;
10905 
10906 	return (ndx + 1);
10907 }
10908 
10909 static void
10910 dtrace_format_remove(dtrace_state_t *state, uint16_t format)
10911 {
10912 	char *fmt;
10913 
10914 	ASSERT(state->dts_formats != NULL);
10915 	ASSERT(format <= state->dts_nformats);
10916 	ASSERT(state->dts_formats[format - 1] != NULL);
10917 
10918 	fmt = state->dts_formats[format - 1];
10919 	kmem_free(fmt, strlen(fmt) + 1);
10920 	state->dts_formats[format - 1] = NULL;
10921 }
10922 
10923 static void
10924 dtrace_format_destroy(dtrace_state_t *state)
10925 {
10926 	int i;
10927 
10928 	if (state->dts_nformats == 0) {
10929 		ASSERT(state->dts_formats == NULL);
10930 		return;
10931 	}
10932 
10933 	ASSERT(state->dts_formats != NULL);
10934 
10935 	for (i = 0; i < state->dts_nformats; i++) {
10936 		char *fmt = state->dts_formats[i];
10937 
10938 		if (fmt == NULL)
10939 			continue;
10940 
10941 		kmem_free(fmt, strlen(fmt) + 1);
10942 	}
10943 
10944 	kmem_free(state->dts_formats, state->dts_nformats * sizeof (char *));
10945 	state->dts_nformats = 0;
10946 	state->dts_formats = NULL;
10947 }
10948 
10949 /*
10950  * DTrace Predicate Functions
10951  */
10952 static dtrace_predicate_t *
10953 dtrace_predicate_create(dtrace_difo_t *dp)
10954 {
10955 	dtrace_predicate_t *pred;
10956 
10957 	ASSERT(MUTEX_HELD(&dtrace_lock));
10958 	ASSERT(dp->dtdo_refcnt != 0);
10959 
10960 	pred = kmem_zalloc(sizeof (dtrace_predicate_t), KM_SLEEP);
10961 	pred->dtp_difo = dp;
10962 	pred->dtp_refcnt = 1;
10963 
10964 	if (!dtrace_difo_cacheable(dp))
10965 		return (pred);
10966 
10967 	if (dtrace_predcache_id == DTRACE_CACHEIDNONE) {
10968 		/*
10969 		 * This is only theoretically possible -- we have had 2^32
10970 		 * cacheable predicates on this machine.  We cannot allow any
10971 		 * more predicates to become cacheable:  as unlikely as it is,
10972 		 * there may be a thread caching a (now stale) predicate cache
10973 		 * ID. (N.B.: the temptation is being successfully resisted to
10974 		 * have this cmn_err() "Holy shit -- we executed this code!")
10975 		 */
10976 		return (pred);
10977 	}
10978 
10979 	pred->dtp_cacheid = dtrace_predcache_id++;
10980 
10981 	return (pred);
10982 }
10983 
10984 static void
10985 dtrace_predicate_hold(dtrace_predicate_t *pred)
10986 {
10987 	ASSERT(MUTEX_HELD(&dtrace_lock));
10988 	ASSERT(pred->dtp_difo != NULL && pred->dtp_difo->dtdo_refcnt != 0);
10989 	ASSERT(pred->dtp_refcnt > 0);
10990 
10991 	pred->dtp_refcnt++;
10992 }
10993 
10994 static void
10995 dtrace_predicate_release(dtrace_predicate_t *pred, dtrace_vstate_t *vstate)
10996 {
10997 	dtrace_difo_t *dp = pred->dtp_difo;
10998 
10999 	ASSERT(MUTEX_HELD(&dtrace_lock));
11000 	ASSERT(dp != NULL && dp->dtdo_refcnt != 0);
11001 	ASSERT(pred->dtp_refcnt > 0);
11002 
11003 	if (--pred->dtp_refcnt == 0) {
11004 		dtrace_difo_release(pred->dtp_difo, vstate);
11005 		kmem_free(pred, sizeof (dtrace_predicate_t));
11006 	}
11007 }
11008 
11009 /*
11010  * DTrace Action Description Functions
11011  */
11012 static dtrace_actdesc_t *
11013 dtrace_actdesc_create(dtrace_actkind_t kind, uint32_t ntuple,
11014     uint64_t uarg, uint64_t arg)
11015 {
11016 	dtrace_actdesc_t *act;
11017 
11018 #ifdef illumos
11019 	ASSERT(!DTRACEACT_ISPRINTFLIKE(kind) || (arg != NULL &&
11020 	    arg >= KERNELBASE) || (arg == NULL && kind == DTRACEACT_PRINTA));
11021 #endif
11022 
11023 	act = kmem_zalloc(sizeof (dtrace_actdesc_t), KM_SLEEP);
11024 	act->dtad_kind = kind;
11025 	act->dtad_ntuple = ntuple;
11026 	act->dtad_uarg = uarg;
11027 	act->dtad_arg = arg;
11028 	act->dtad_refcnt = 1;
11029 
11030 	return (act);
11031 }
11032 
11033 static void
11034 dtrace_actdesc_hold(dtrace_actdesc_t *act)
11035 {
11036 	ASSERT(act->dtad_refcnt >= 1);
11037 	act->dtad_refcnt++;
11038 }
11039 
11040 static void
11041 dtrace_actdesc_release(dtrace_actdesc_t *act, dtrace_vstate_t *vstate)
11042 {
11043 	dtrace_actkind_t kind = act->dtad_kind;
11044 	dtrace_difo_t *dp;
11045 
11046 	ASSERT(act->dtad_refcnt >= 1);
11047 
11048 	if (--act->dtad_refcnt != 0)
11049 		return;
11050 
11051 	if ((dp = act->dtad_difo) != NULL)
11052 		dtrace_difo_release(dp, vstate);
11053 
11054 	if (DTRACEACT_ISPRINTFLIKE(kind)) {
11055 		char *str = (char *)(uintptr_t)act->dtad_arg;
11056 
11057 #ifdef illumos
11058 		ASSERT((str != NULL && (uintptr_t)str >= KERNELBASE) ||
11059 		    (str == NULL && act->dtad_kind == DTRACEACT_PRINTA));
11060 #endif
11061 
11062 		if (str != NULL)
11063 			kmem_free(str, strlen(str) + 1);
11064 	}
11065 
11066 	kmem_free(act, sizeof (dtrace_actdesc_t));
11067 }
11068 
11069 /*
11070  * DTrace ECB Functions
11071  */
11072 static dtrace_ecb_t *
11073 dtrace_ecb_add(dtrace_state_t *state, dtrace_probe_t *probe)
11074 {
11075 	dtrace_ecb_t *ecb;
11076 	dtrace_epid_t epid;
11077 
11078 	ASSERT(MUTEX_HELD(&dtrace_lock));
11079 
11080 	ecb = kmem_zalloc(sizeof (dtrace_ecb_t), KM_SLEEP);
11081 	ecb->dte_predicate = NULL;
11082 	ecb->dte_probe = probe;
11083 
11084 	/*
11085 	 * The default size is the size of the default action: recording
11086 	 * the header.
11087 	 */
11088 	ecb->dte_size = ecb->dte_needed = sizeof (dtrace_rechdr_t);
11089 	ecb->dte_alignment = sizeof (dtrace_epid_t);
11090 
11091 	epid = state->dts_epid++;
11092 
11093 	if (epid - 1 >= state->dts_necbs) {
11094 		dtrace_ecb_t **oecbs = state->dts_ecbs, **ecbs;
11095 		int necbs = state->dts_necbs << 1;
11096 
11097 		ASSERT(epid == state->dts_necbs + 1);
11098 
11099 		if (necbs == 0) {
11100 			ASSERT(oecbs == NULL);
11101 			necbs = 1;
11102 		}
11103 
11104 		ecbs = kmem_zalloc(necbs * sizeof (*ecbs), KM_SLEEP);
11105 
11106 		if (oecbs != NULL)
11107 			bcopy(oecbs, ecbs, state->dts_necbs * sizeof (*ecbs));
11108 
11109 		dtrace_membar_producer();
11110 		state->dts_ecbs = ecbs;
11111 
11112 		if (oecbs != NULL) {
11113 			/*
11114 			 * If this state is active, we must dtrace_sync()
11115 			 * before we can free the old dts_ecbs array:  we're
11116 			 * coming in hot, and there may be active ring
11117 			 * buffer processing (which indexes into the dts_ecbs
11118 			 * array) on another CPU.
11119 			 */
11120 			if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
11121 				dtrace_sync();
11122 
11123 			kmem_free(oecbs, state->dts_necbs * sizeof (*ecbs));
11124 		}
11125 
11126 		dtrace_membar_producer();
11127 		state->dts_necbs = necbs;
11128 	}
11129 
11130 	ecb->dte_state = state;
11131 
11132 	ASSERT(state->dts_ecbs[epid - 1] == NULL);
11133 	dtrace_membar_producer();
11134 	state->dts_ecbs[(ecb->dte_epid = epid) - 1] = ecb;
11135 
11136 	return (ecb);
11137 }
11138 
11139 static void
11140 dtrace_ecb_enable(dtrace_ecb_t *ecb)
11141 {
11142 	dtrace_probe_t *probe = ecb->dte_probe;
11143 
11144 	ASSERT(MUTEX_HELD(&cpu_lock));
11145 	ASSERT(MUTEX_HELD(&dtrace_lock));
11146 	ASSERT(ecb->dte_next == NULL);
11147 
11148 	if (probe == NULL) {
11149 		/*
11150 		 * This is the NULL probe -- there's nothing to do.
11151 		 */
11152 		return;
11153 	}
11154 
11155 	if (probe->dtpr_ecb == NULL) {
11156 		dtrace_provider_t *prov = probe->dtpr_provider;
11157 
11158 		/*
11159 		 * We're the first ECB on this probe.
11160 		 */
11161 		probe->dtpr_ecb = probe->dtpr_ecb_last = ecb;
11162 
11163 		if (ecb->dte_predicate != NULL)
11164 			probe->dtpr_predcache = ecb->dte_predicate->dtp_cacheid;
11165 
11166 		prov->dtpv_pops.dtps_enable(prov->dtpv_arg,
11167 		    probe->dtpr_id, probe->dtpr_arg);
11168 	} else {
11169 		/*
11170 		 * This probe is already active.  Swing the last pointer to
11171 		 * point to the new ECB, and issue a dtrace_sync() to assure
11172 		 * that all CPUs have seen the change.
11173 		 */
11174 		ASSERT(probe->dtpr_ecb_last != NULL);
11175 		probe->dtpr_ecb_last->dte_next = ecb;
11176 		probe->dtpr_ecb_last = ecb;
11177 		probe->dtpr_predcache = 0;
11178 
11179 		dtrace_sync();
11180 	}
11181 }
11182 
11183 static int
11184 dtrace_ecb_resize(dtrace_ecb_t *ecb)
11185 {
11186 	dtrace_action_t *act;
11187 	uint32_t curneeded = UINT32_MAX;
11188 	uint32_t aggbase = UINT32_MAX;
11189 
11190 	/*
11191 	 * If we record anything, we always record the dtrace_rechdr_t.  (And
11192 	 * we always record it first.)
11193 	 */
11194 	ecb->dte_size = sizeof (dtrace_rechdr_t);
11195 	ecb->dte_alignment = sizeof (dtrace_epid_t);
11196 
11197 	for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
11198 		dtrace_recdesc_t *rec = &act->dta_rec;
11199 		ASSERT(rec->dtrd_size > 0 || rec->dtrd_alignment == 1);
11200 
11201 		ecb->dte_alignment = MAX(ecb->dte_alignment,
11202 		    rec->dtrd_alignment);
11203 
11204 		if (DTRACEACT_ISAGG(act->dta_kind)) {
11205 			dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
11206 
11207 			ASSERT(rec->dtrd_size != 0);
11208 			ASSERT(agg->dtag_first != NULL);
11209 			ASSERT(act->dta_prev->dta_intuple);
11210 			ASSERT(aggbase != UINT32_MAX);
11211 			ASSERT(curneeded != UINT32_MAX);
11212 
11213 			agg->dtag_base = aggbase;
11214 
11215 			curneeded = P2ROUNDUP(curneeded, rec->dtrd_alignment);
11216 			rec->dtrd_offset = curneeded;
11217 			if (curneeded + rec->dtrd_size < curneeded)
11218 				return (EINVAL);
11219 			curneeded += rec->dtrd_size;
11220 			ecb->dte_needed = MAX(ecb->dte_needed, curneeded);
11221 
11222 			aggbase = UINT32_MAX;
11223 			curneeded = UINT32_MAX;
11224 		} else if (act->dta_intuple) {
11225 			if (curneeded == UINT32_MAX) {
11226 				/*
11227 				 * This is the first record in a tuple.  Align
11228 				 * curneeded to be at offset 4 in an 8-byte
11229 				 * aligned block.
11230 				 */
11231 				ASSERT(act->dta_prev == NULL ||
11232 				    !act->dta_prev->dta_intuple);
11233 				ASSERT3U(aggbase, ==, UINT32_MAX);
11234 				curneeded = P2PHASEUP(ecb->dte_size,
11235 				    sizeof (uint64_t), sizeof (dtrace_aggid_t));
11236 
11237 				aggbase = curneeded - sizeof (dtrace_aggid_t);
11238 				ASSERT(IS_P2ALIGNED(aggbase,
11239 				    sizeof (uint64_t)));
11240 			}
11241 			curneeded = P2ROUNDUP(curneeded, rec->dtrd_alignment);
11242 			rec->dtrd_offset = curneeded;
11243 			if (curneeded + rec->dtrd_size < curneeded)
11244 				return (EINVAL);
11245 			curneeded += rec->dtrd_size;
11246 		} else {
11247 			/* tuples must be followed by an aggregation */
11248 			ASSERT(act->dta_prev == NULL ||
11249 			    !act->dta_prev->dta_intuple);
11250 
11251 			ecb->dte_size = P2ROUNDUP(ecb->dte_size,
11252 			    rec->dtrd_alignment);
11253 			rec->dtrd_offset = ecb->dte_size;
11254 			if (ecb->dte_size + rec->dtrd_size < ecb->dte_size)
11255 				return (EINVAL);
11256 			ecb->dte_size += rec->dtrd_size;
11257 			ecb->dte_needed = MAX(ecb->dte_needed, ecb->dte_size);
11258 		}
11259 	}
11260 
11261 	if ((act = ecb->dte_action) != NULL &&
11262 	    !(act->dta_kind == DTRACEACT_SPECULATE && act->dta_next == NULL) &&
11263 	    ecb->dte_size == sizeof (dtrace_rechdr_t)) {
11264 		/*
11265 		 * If the size is still sizeof (dtrace_rechdr_t), then all
11266 		 * actions store no data; set the size to 0.
11267 		 */
11268 		ecb->dte_size = 0;
11269 	}
11270 
11271 	ecb->dte_size = P2ROUNDUP(ecb->dte_size, sizeof (dtrace_epid_t));
11272 	ecb->dte_needed = P2ROUNDUP(ecb->dte_needed, (sizeof (dtrace_epid_t)));
11273 	ecb->dte_state->dts_needed = MAX(ecb->dte_state->dts_needed,
11274 	    ecb->dte_needed);
11275 	return (0);
11276 }
11277 
11278 static dtrace_action_t *
11279 dtrace_ecb_aggregation_create(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
11280 {
11281 	dtrace_aggregation_t *agg;
11282 	size_t size = sizeof (uint64_t);
11283 	int ntuple = desc->dtad_ntuple;
11284 	dtrace_action_t *act;
11285 	dtrace_recdesc_t *frec;
11286 	dtrace_aggid_t aggid;
11287 	dtrace_state_t *state = ecb->dte_state;
11288 
11289 	agg = kmem_zalloc(sizeof (dtrace_aggregation_t), KM_SLEEP);
11290 	agg->dtag_ecb = ecb;
11291 
11292 	ASSERT(DTRACEACT_ISAGG(desc->dtad_kind));
11293 
11294 	switch (desc->dtad_kind) {
11295 	case DTRACEAGG_MIN:
11296 		agg->dtag_initial = INT64_MAX;
11297 		agg->dtag_aggregate = dtrace_aggregate_min;
11298 		break;
11299 
11300 	case DTRACEAGG_MAX:
11301 		agg->dtag_initial = INT64_MIN;
11302 		agg->dtag_aggregate = dtrace_aggregate_max;
11303 		break;
11304 
11305 	case DTRACEAGG_COUNT:
11306 		agg->dtag_aggregate = dtrace_aggregate_count;
11307 		break;
11308 
11309 	case DTRACEAGG_QUANTIZE:
11310 		agg->dtag_aggregate = dtrace_aggregate_quantize;
11311 		size = (((sizeof (uint64_t) * NBBY) - 1) * 2 + 1) *
11312 		    sizeof (uint64_t);
11313 		break;
11314 
11315 	case DTRACEAGG_LQUANTIZE: {
11316 		uint16_t step = DTRACE_LQUANTIZE_STEP(desc->dtad_arg);
11317 		uint16_t levels = DTRACE_LQUANTIZE_LEVELS(desc->dtad_arg);
11318 
11319 		agg->dtag_initial = desc->dtad_arg;
11320 		agg->dtag_aggregate = dtrace_aggregate_lquantize;
11321 
11322 		if (step == 0 || levels == 0)
11323 			goto err;
11324 
11325 		size = levels * sizeof (uint64_t) + 3 * sizeof (uint64_t);
11326 		break;
11327 	}
11328 
11329 	case DTRACEAGG_LLQUANTIZE: {
11330 		uint16_t factor = DTRACE_LLQUANTIZE_FACTOR(desc->dtad_arg);
11331 		uint16_t low = DTRACE_LLQUANTIZE_LOW(desc->dtad_arg);
11332 		uint16_t high = DTRACE_LLQUANTIZE_HIGH(desc->dtad_arg);
11333 		uint16_t nsteps = DTRACE_LLQUANTIZE_NSTEP(desc->dtad_arg);
11334 		int64_t v;
11335 
11336 		agg->dtag_initial = desc->dtad_arg;
11337 		agg->dtag_aggregate = dtrace_aggregate_llquantize;
11338 
11339 		if (factor < 2 || low >= high || nsteps < factor)
11340 			goto err;
11341 
11342 		/*
11343 		 * Now check that the number of steps evenly divides a power
11344 		 * of the factor.  (This assures both integer bucket size and
11345 		 * linearity within each magnitude.)
11346 		 */
11347 		for (v = factor; v < nsteps; v *= factor)
11348 			continue;
11349 
11350 		if ((v % nsteps) || (nsteps % factor))
11351 			goto err;
11352 
11353 		size = (dtrace_aggregate_llquantize_bucket(factor,
11354 		    low, high, nsteps, INT64_MAX) + 2) * sizeof (uint64_t);
11355 		break;
11356 	}
11357 
11358 	case DTRACEAGG_AVG:
11359 		agg->dtag_aggregate = dtrace_aggregate_avg;
11360 		size = sizeof (uint64_t) * 2;
11361 		break;
11362 
11363 	case DTRACEAGG_STDDEV:
11364 		agg->dtag_aggregate = dtrace_aggregate_stddev;
11365 		size = sizeof (uint64_t) * 4;
11366 		break;
11367 
11368 	case DTRACEAGG_SUM:
11369 		agg->dtag_aggregate = dtrace_aggregate_sum;
11370 		break;
11371 
11372 	default:
11373 		goto err;
11374 	}
11375 
11376 	agg->dtag_action.dta_rec.dtrd_size = size;
11377 
11378 	if (ntuple == 0)
11379 		goto err;
11380 
11381 	/*
11382 	 * We must make sure that we have enough actions for the n-tuple.
11383 	 */
11384 	for (act = ecb->dte_action_last; act != NULL; act = act->dta_prev) {
11385 		if (DTRACEACT_ISAGG(act->dta_kind))
11386 			break;
11387 
11388 		if (--ntuple == 0) {
11389 			/*
11390 			 * This is the action with which our n-tuple begins.
11391 			 */
11392 			agg->dtag_first = act;
11393 			goto success;
11394 		}
11395 	}
11396 
11397 	/*
11398 	 * This n-tuple is short by ntuple elements.  Return failure.
11399 	 */
11400 	ASSERT(ntuple != 0);
11401 err:
11402 	kmem_free(agg, sizeof (dtrace_aggregation_t));
11403 	return (NULL);
11404 
11405 success:
11406 	/*
11407 	 * If the last action in the tuple has a size of zero, it's actually
11408 	 * an expression argument for the aggregating action.
11409 	 */
11410 	ASSERT(ecb->dte_action_last != NULL);
11411 	act = ecb->dte_action_last;
11412 
11413 	if (act->dta_kind == DTRACEACT_DIFEXPR) {
11414 		ASSERT(act->dta_difo != NULL);
11415 
11416 		if (act->dta_difo->dtdo_rtype.dtdt_size == 0)
11417 			agg->dtag_hasarg = 1;
11418 	}
11419 
11420 	/*
11421 	 * We need to allocate an id for this aggregation.
11422 	 */
11423 #ifdef illumos
11424 	aggid = (dtrace_aggid_t)(uintptr_t)vmem_alloc(state->dts_aggid_arena, 1,
11425 	    VM_BESTFIT | VM_SLEEP);
11426 #else
11427 	aggid = alloc_unr(state->dts_aggid_arena);
11428 #endif
11429 
11430 	if (aggid - 1 >= state->dts_naggregations) {
11431 		dtrace_aggregation_t **oaggs = state->dts_aggregations;
11432 		dtrace_aggregation_t **aggs;
11433 		int naggs = state->dts_naggregations << 1;
11434 		int onaggs = state->dts_naggregations;
11435 
11436 		ASSERT(aggid == state->dts_naggregations + 1);
11437 
11438 		if (naggs == 0) {
11439 			ASSERT(oaggs == NULL);
11440 			naggs = 1;
11441 		}
11442 
11443 		aggs = kmem_zalloc(naggs * sizeof (*aggs), KM_SLEEP);
11444 
11445 		if (oaggs != NULL) {
11446 			bcopy(oaggs, aggs, onaggs * sizeof (*aggs));
11447 			kmem_free(oaggs, onaggs * sizeof (*aggs));
11448 		}
11449 
11450 		state->dts_aggregations = aggs;
11451 		state->dts_naggregations = naggs;
11452 	}
11453 
11454 	ASSERT(state->dts_aggregations[aggid - 1] == NULL);
11455 	state->dts_aggregations[(agg->dtag_id = aggid) - 1] = agg;
11456 
11457 	frec = &agg->dtag_first->dta_rec;
11458 	if (frec->dtrd_alignment < sizeof (dtrace_aggid_t))
11459 		frec->dtrd_alignment = sizeof (dtrace_aggid_t);
11460 
11461 	for (act = agg->dtag_first; act != NULL; act = act->dta_next) {
11462 		ASSERT(!act->dta_intuple);
11463 		act->dta_intuple = 1;
11464 	}
11465 
11466 	return (&agg->dtag_action);
11467 }
11468 
11469 static void
11470 dtrace_ecb_aggregation_destroy(dtrace_ecb_t *ecb, dtrace_action_t *act)
11471 {
11472 	dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
11473 	dtrace_state_t *state = ecb->dte_state;
11474 	dtrace_aggid_t aggid = agg->dtag_id;
11475 
11476 	ASSERT(DTRACEACT_ISAGG(act->dta_kind));
11477 #ifdef illumos
11478 	vmem_free(state->dts_aggid_arena, (void *)(uintptr_t)aggid, 1);
11479 #else
11480 	free_unr(state->dts_aggid_arena, aggid);
11481 #endif
11482 
11483 	ASSERT(state->dts_aggregations[aggid - 1] == agg);
11484 	state->dts_aggregations[aggid - 1] = NULL;
11485 
11486 	kmem_free(agg, sizeof (dtrace_aggregation_t));
11487 }
11488 
11489 static int
11490 dtrace_ecb_action_add(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
11491 {
11492 	dtrace_action_t *action, *last;
11493 	dtrace_difo_t *dp = desc->dtad_difo;
11494 	uint32_t size = 0, align = sizeof (uint8_t), mask;
11495 	uint16_t format = 0;
11496 	dtrace_recdesc_t *rec;
11497 	dtrace_state_t *state = ecb->dte_state;
11498 	dtrace_optval_t *opt = state->dts_options, nframes = 0, strsize;
11499 	uint64_t arg = desc->dtad_arg;
11500 
11501 	ASSERT(MUTEX_HELD(&dtrace_lock));
11502 	ASSERT(ecb->dte_action == NULL || ecb->dte_action->dta_refcnt == 1);
11503 
11504 	if (DTRACEACT_ISAGG(desc->dtad_kind)) {
11505 		/*
11506 		 * If this is an aggregating action, there must be neither
11507 		 * a speculate nor a commit on the action chain.
11508 		 */
11509 		dtrace_action_t *act;
11510 
11511 		for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
11512 			if (act->dta_kind == DTRACEACT_COMMIT)
11513 				return (EINVAL);
11514 
11515 			if (act->dta_kind == DTRACEACT_SPECULATE)
11516 				return (EINVAL);
11517 		}
11518 
11519 		action = dtrace_ecb_aggregation_create(ecb, desc);
11520 
11521 		if (action == NULL)
11522 			return (EINVAL);
11523 	} else {
11524 		if (DTRACEACT_ISDESTRUCTIVE(desc->dtad_kind) ||
11525 		    (desc->dtad_kind == DTRACEACT_DIFEXPR &&
11526 		    dp != NULL && dp->dtdo_destructive)) {
11527 			state->dts_destructive = 1;
11528 		}
11529 
11530 		switch (desc->dtad_kind) {
11531 		case DTRACEACT_PRINTF:
11532 		case DTRACEACT_PRINTA:
11533 		case DTRACEACT_SYSTEM:
11534 		case DTRACEACT_FREOPEN:
11535 		case DTRACEACT_DIFEXPR:
11536 			/*
11537 			 * We know that our arg is a string -- turn it into a
11538 			 * format.
11539 			 */
11540 			if (arg == 0) {
11541 				ASSERT(desc->dtad_kind == DTRACEACT_PRINTA ||
11542 				    desc->dtad_kind == DTRACEACT_DIFEXPR);
11543 				format = 0;
11544 			} else {
11545 				ASSERT(arg != 0);
11546 #ifdef illumos
11547 				ASSERT(arg > KERNELBASE);
11548 #endif
11549 				format = dtrace_format_add(state,
11550 				    (char *)(uintptr_t)arg);
11551 			}
11552 
11553 			/*FALLTHROUGH*/
11554 		case DTRACEACT_LIBACT:
11555 		case DTRACEACT_TRACEMEM:
11556 		case DTRACEACT_TRACEMEM_DYNSIZE:
11557 			if (dp == NULL)
11558 				return (EINVAL);
11559 
11560 			if ((size = dp->dtdo_rtype.dtdt_size) != 0)
11561 				break;
11562 
11563 			if (dp->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING) {
11564 				if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
11565 					return (EINVAL);
11566 
11567 				size = opt[DTRACEOPT_STRSIZE];
11568 			}
11569 
11570 			break;
11571 
11572 		case DTRACEACT_STACK:
11573 			if ((nframes = arg) == 0) {
11574 				nframes = opt[DTRACEOPT_STACKFRAMES];
11575 				ASSERT(nframes > 0);
11576 				arg = nframes;
11577 			}
11578 
11579 			size = nframes * sizeof (pc_t);
11580 			break;
11581 
11582 		case DTRACEACT_JSTACK:
11583 			if ((strsize = DTRACE_USTACK_STRSIZE(arg)) == 0)
11584 				strsize = opt[DTRACEOPT_JSTACKSTRSIZE];
11585 
11586 			if ((nframes = DTRACE_USTACK_NFRAMES(arg)) == 0)
11587 				nframes = opt[DTRACEOPT_JSTACKFRAMES];
11588 
11589 			arg = DTRACE_USTACK_ARG(nframes, strsize);
11590 
11591 			/*FALLTHROUGH*/
11592 		case DTRACEACT_USTACK:
11593 			if (desc->dtad_kind != DTRACEACT_JSTACK &&
11594 			    (nframes = DTRACE_USTACK_NFRAMES(arg)) == 0) {
11595 				strsize = DTRACE_USTACK_STRSIZE(arg);
11596 				nframes = opt[DTRACEOPT_USTACKFRAMES];
11597 				ASSERT(nframes > 0);
11598 				arg = DTRACE_USTACK_ARG(nframes, strsize);
11599 			}
11600 
11601 			/*
11602 			 * Save a slot for the pid.
11603 			 */
11604 			size = (nframes + 1) * sizeof (uint64_t);
11605 			size += DTRACE_USTACK_STRSIZE(arg);
11606 			size = P2ROUNDUP(size, (uint32_t)(sizeof (uintptr_t)));
11607 
11608 			break;
11609 
11610 		case DTRACEACT_SYM:
11611 		case DTRACEACT_MOD:
11612 			if (dp == NULL || ((size = dp->dtdo_rtype.dtdt_size) !=
11613 			    sizeof (uint64_t)) ||
11614 			    (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
11615 				return (EINVAL);
11616 			break;
11617 
11618 		case DTRACEACT_USYM:
11619 		case DTRACEACT_UMOD:
11620 		case DTRACEACT_UADDR:
11621 			if (dp == NULL ||
11622 			    (dp->dtdo_rtype.dtdt_size != sizeof (uint64_t)) ||
11623 			    (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
11624 				return (EINVAL);
11625 
11626 			/*
11627 			 * We have a slot for the pid, plus a slot for the
11628 			 * argument.  To keep things simple (aligned with
11629 			 * bitness-neutral sizing), we store each as a 64-bit
11630 			 * quantity.
11631 			 */
11632 			size = 2 * sizeof (uint64_t);
11633 			break;
11634 
11635 		case DTRACEACT_STOP:
11636 		case DTRACEACT_BREAKPOINT:
11637 		case DTRACEACT_PANIC:
11638 			break;
11639 
11640 		case DTRACEACT_CHILL:
11641 		case DTRACEACT_DISCARD:
11642 		case DTRACEACT_RAISE:
11643 			if (dp == NULL)
11644 				return (EINVAL);
11645 			break;
11646 
11647 		case DTRACEACT_EXIT:
11648 			if (dp == NULL ||
11649 			    (size = dp->dtdo_rtype.dtdt_size) != sizeof (int) ||
11650 			    (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
11651 				return (EINVAL);
11652 			break;
11653 
11654 		case DTRACEACT_SPECULATE:
11655 			if (ecb->dte_size > sizeof (dtrace_rechdr_t))
11656 				return (EINVAL);
11657 
11658 			if (dp == NULL)
11659 				return (EINVAL);
11660 
11661 			state->dts_speculates = 1;
11662 			break;
11663 
11664 		case DTRACEACT_PRINTM:
11665 		    	size = dp->dtdo_rtype.dtdt_size;
11666 			break;
11667 
11668 		case DTRACEACT_COMMIT: {
11669 			dtrace_action_t *act = ecb->dte_action;
11670 
11671 			for (; act != NULL; act = act->dta_next) {
11672 				if (act->dta_kind == DTRACEACT_COMMIT)
11673 					return (EINVAL);
11674 			}
11675 
11676 			if (dp == NULL)
11677 				return (EINVAL);
11678 			break;
11679 		}
11680 
11681 		default:
11682 			return (EINVAL);
11683 		}
11684 
11685 		if (size != 0 || desc->dtad_kind == DTRACEACT_SPECULATE) {
11686 			/*
11687 			 * If this is a data-storing action or a speculate,
11688 			 * we must be sure that there isn't a commit on the
11689 			 * action chain.
11690 			 */
11691 			dtrace_action_t *act = ecb->dte_action;
11692 
11693 			for (; act != NULL; act = act->dta_next) {
11694 				if (act->dta_kind == DTRACEACT_COMMIT)
11695 					return (EINVAL);
11696 			}
11697 		}
11698 
11699 		action = kmem_zalloc(sizeof (dtrace_action_t), KM_SLEEP);
11700 		action->dta_rec.dtrd_size = size;
11701 	}
11702 
11703 	action->dta_refcnt = 1;
11704 	rec = &action->dta_rec;
11705 	size = rec->dtrd_size;
11706 
11707 	for (mask = sizeof (uint64_t) - 1; size != 0 && mask > 0; mask >>= 1) {
11708 		if (!(size & mask)) {
11709 			align = mask + 1;
11710 			break;
11711 		}
11712 	}
11713 
11714 	action->dta_kind = desc->dtad_kind;
11715 
11716 	if ((action->dta_difo = dp) != NULL)
11717 		dtrace_difo_hold(dp);
11718 
11719 	rec->dtrd_action = action->dta_kind;
11720 	rec->dtrd_arg = arg;
11721 	rec->dtrd_uarg = desc->dtad_uarg;
11722 	rec->dtrd_alignment = (uint16_t)align;
11723 	rec->dtrd_format = format;
11724 
11725 	if ((last = ecb->dte_action_last) != NULL) {
11726 		ASSERT(ecb->dte_action != NULL);
11727 		action->dta_prev = last;
11728 		last->dta_next = action;
11729 	} else {
11730 		ASSERT(ecb->dte_action == NULL);
11731 		ecb->dte_action = action;
11732 	}
11733 
11734 	ecb->dte_action_last = action;
11735 
11736 	return (0);
11737 }
11738 
11739 static void
11740 dtrace_ecb_action_remove(dtrace_ecb_t *ecb)
11741 {
11742 	dtrace_action_t *act = ecb->dte_action, *next;
11743 	dtrace_vstate_t *vstate = &ecb->dte_state->dts_vstate;
11744 	dtrace_difo_t *dp;
11745 	uint16_t format;
11746 
11747 	if (act != NULL && act->dta_refcnt > 1) {
11748 		ASSERT(act->dta_next == NULL || act->dta_next->dta_refcnt == 1);
11749 		act->dta_refcnt--;
11750 	} else {
11751 		for (; act != NULL; act = next) {
11752 			next = act->dta_next;
11753 			ASSERT(next != NULL || act == ecb->dte_action_last);
11754 			ASSERT(act->dta_refcnt == 1);
11755 
11756 			if ((format = act->dta_rec.dtrd_format) != 0)
11757 				dtrace_format_remove(ecb->dte_state, format);
11758 
11759 			if ((dp = act->dta_difo) != NULL)
11760 				dtrace_difo_release(dp, vstate);
11761 
11762 			if (DTRACEACT_ISAGG(act->dta_kind)) {
11763 				dtrace_ecb_aggregation_destroy(ecb, act);
11764 			} else {
11765 				kmem_free(act, sizeof (dtrace_action_t));
11766 			}
11767 		}
11768 	}
11769 
11770 	ecb->dte_action = NULL;
11771 	ecb->dte_action_last = NULL;
11772 	ecb->dte_size = 0;
11773 }
11774 
11775 static void
11776 dtrace_ecb_disable(dtrace_ecb_t *ecb)
11777 {
11778 	/*
11779 	 * We disable the ECB by removing it from its probe.
11780 	 */
11781 	dtrace_ecb_t *pecb, *prev = NULL;
11782 	dtrace_probe_t *probe = ecb->dte_probe;
11783 
11784 	ASSERT(MUTEX_HELD(&dtrace_lock));
11785 
11786 	if (probe == NULL) {
11787 		/*
11788 		 * This is the NULL probe; there is nothing to disable.
11789 		 */
11790 		return;
11791 	}
11792 
11793 	for (pecb = probe->dtpr_ecb; pecb != NULL; pecb = pecb->dte_next) {
11794 		if (pecb == ecb)
11795 			break;
11796 		prev = pecb;
11797 	}
11798 
11799 	ASSERT(pecb != NULL);
11800 
11801 	if (prev == NULL) {
11802 		probe->dtpr_ecb = ecb->dte_next;
11803 	} else {
11804 		prev->dte_next = ecb->dte_next;
11805 	}
11806 
11807 	if (ecb == probe->dtpr_ecb_last) {
11808 		ASSERT(ecb->dte_next == NULL);
11809 		probe->dtpr_ecb_last = prev;
11810 	}
11811 
11812 	/*
11813 	 * The ECB has been disconnected from the probe; now sync to assure
11814 	 * that all CPUs have seen the change before returning.
11815 	 */
11816 	dtrace_sync();
11817 
11818 	if (probe->dtpr_ecb == NULL) {
11819 		/*
11820 		 * That was the last ECB on the probe; clear the predicate
11821 		 * cache ID for the probe, disable it and sync one more time
11822 		 * to assure that we'll never hit it again.
11823 		 */
11824 		dtrace_provider_t *prov = probe->dtpr_provider;
11825 
11826 		ASSERT(ecb->dte_next == NULL);
11827 		ASSERT(probe->dtpr_ecb_last == NULL);
11828 		probe->dtpr_predcache = DTRACE_CACHEIDNONE;
11829 		prov->dtpv_pops.dtps_disable(prov->dtpv_arg,
11830 		    probe->dtpr_id, probe->dtpr_arg);
11831 		dtrace_sync();
11832 	} else {
11833 		/*
11834 		 * There is at least one ECB remaining on the probe.  If there
11835 		 * is _exactly_ one, set the probe's predicate cache ID to be
11836 		 * the predicate cache ID of the remaining ECB.
11837 		 */
11838 		ASSERT(probe->dtpr_ecb_last != NULL);
11839 		ASSERT(probe->dtpr_predcache == DTRACE_CACHEIDNONE);
11840 
11841 		if (probe->dtpr_ecb == probe->dtpr_ecb_last) {
11842 			dtrace_predicate_t *p = probe->dtpr_ecb->dte_predicate;
11843 
11844 			ASSERT(probe->dtpr_ecb->dte_next == NULL);
11845 
11846 			if (p != NULL)
11847 				probe->dtpr_predcache = p->dtp_cacheid;
11848 		}
11849 
11850 		ecb->dte_next = NULL;
11851 	}
11852 }
11853 
11854 static void
11855 dtrace_ecb_destroy(dtrace_ecb_t *ecb)
11856 {
11857 	dtrace_state_t *state = ecb->dte_state;
11858 	dtrace_vstate_t *vstate = &state->dts_vstate;
11859 	dtrace_predicate_t *pred;
11860 	dtrace_epid_t epid = ecb->dte_epid;
11861 
11862 	ASSERT(MUTEX_HELD(&dtrace_lock));
11863 	ASSERT(ecb->dte_next == NULL);
11864 	ASSERT(ecb->dte_probe == NULL || ecb->dte_probe->dtpr_ecb != ecb);
11865 
11866 	if ((pred = ecb->dte_predicate) != NULL)
11867 		dtrace_predicate_release(pred, vstate);
11868 
11869 	dtrace_ecb_action_remove(ecb);
11870 
11871 	ASSERT(state->dts_ecbs[epid - 1] == ecb);
11872 	state->dts_ecbs[epid - 1] = NULL;
11873 
11874 	kmem_free(ecb, sizeof (dtrace_ecb_t));
11875 }
11876 
11877 static dtrace_ecb_t *
11878 dtrace_ecb_create(dtrace_state_t *state, dtrace_probe_t *probe,
11879     dtrace_enabling_t *enab)
11880 {
11881 	dtrace_ecb_t *ecb;
11882 	dtrace_predicate_t *pred;
11883 	dtrace_actdesc_t *act;
11884 	dtrace_provider_t *prov;
11885 	dtrace_ecbdesc_t *desc = enab->dten_current;
11886 
11887 	ASSERT(MUTEX_HELD(&dtrace_lock));
11888 	ASSERT(state != NULL);
11889 
11890 	ecb = dtrace_ecb_add(state, probe);
11891 	ecb->dte_uarg = desc->dted_uarg;
11892 
11893 	if ((pred = desc->dted_pred.dtpdd_predicate) != NULL) {
11894 		dtrace_predicate_hold(pred);
11895 		ecb->dte_predicate = pred;
11896 	}
11897 
11898 	if (probe != NULL) {
11899 		/*
11900 		 * If the provider shows more leg than the consumer is old
11901 		 * enough to see, we need to enable the appropriate implicit
11902 		 * predicate bits to prevent the ecb from activating at
11903 		 * revealing times.
11904 		 *
11905 		 * Providers specifying DTRACE_PRIV_USER at register time
11906 		 * are stating that they need the /proc-style privilege
11907 		 * model to be enforced, and this is what DTRACE_COND_OWNER
11908 		 * and DTRACE_COND_ZONEOWNER will then do at probe time.
11909 		 */
11910 		prov = probe->dtpr_provider;
11911 		if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLPROC) &&
11912 		    (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
11913 			ecb->dte_cond |= DTRACE_COND_OWNER;
11914 
11915 		if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLZONE) &&
11916 		    (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
11917 			ecb->dte_cond |= DTRACE_COND_ZONEOWNER;
11918 
11919 		/*
11920 		 * If the provider shows us kernel innards and the user
11921 		 * is lacking sufficient privilege, enable the
11922 		 * DTRACE_COND_USERMODE implicit predicate.
11923 		 */
11924 		if (!(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) &&
11925 		    (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_KERNEL))
11926 			ecb->dte_cond |= DTRACE_COND_USERMODE;
11927 	}
11928 
11929 	if (dtrace_ecb_create_cache != NULL) {
11930 		/*
11931 		 * If we have a cached ecb, we'll use its action list instead
11932 		 * of creating our own (saving both time and space).
11933 		 */
11934 		dtrace_ecb_t *cached = dtrace_ecb_create_cache;
11935 		dtrace_action_t *act = cached->dte_action;
11936 
11937 		if (act != NULL) {
11938 			ASSERT(act->dta_refcnt > 0);
11939 			act->dta_refcnt++;
11940 			ecb->dte_action = act;
11941 			ecb->dte_action_last = cached->dte_action_last;
11942 			ecb->dte_needed = cached->dte_needed;
11943 			ecb->dte_size = cached->dte_size;
11944 			ecb->dte_alignment = cached->dte_alignment;
11945 		}
11946 
11947 		return (ecb);
11948 	}
11949 
11950 	for (act = desc->dted_action; act != NULL; act = act->dtad_next) {
11951 		if ((enab->dten_error = dtrace_ecb_action_add(ecb, act)) != 0) {
11952 			dtrace_ecb_destroy(ecb);
11953 			return (NULL);
11954 		}
11955 	}
11956 
11957 	if ((enab->dten_error = dtrace_ecb_resize(ecb)) != 0) {
11958 		dtrace_ecb_destroy(ecb);
11959 		return (NULL);
11960 	}
11961 
11962 	return (dtrace_ecb_create_cache = ecb);
11963 }
11964 
11965 static int
11966 dtrace_ecb_create_enable(dtrace_probe_t *probe, void *arg)
11967 {
11968 	dtrace_ecb_t *ecb;
11969 	dtrace_enabling_t *enab = arg;
11970 	dtrace_state_t *state = enab->dten_vstate->dtvs_state;
11971 
11972 	ASSERT(state != NULL);
11973 
11974 	if (probe != NULL && probe->dtpr_gen < enab->dten_probegen) {
11975 		/*
11976 		 * This probe was created in a generation for which this
11977 		 * enabling has previously created ECBs; we don't want to
11978 		 * enable it again, so just kick out.
11979 		 */
11980 		return (DTRACE_MATCH_NEXT);
11981 	}
11982 
11983 	if ((ecb = dtrace_ecb_create(state, probe, enab)) == NULL)
11984 		return (DTRACE_MATCH_DONE);
11985 
11986 	dtrace_ecb_enable(ecb);
11987 	return (DTRACE_MATCH_NEXT);
11988 }
11989 
11990 static dtrace_ecb_t *
11991 dtrace_epid2ecb(dtrace_state_t *state, dtrace_epid_t id)
11992 {
11993 	dtrace_ecb_t *ecb;
11994 
11995 	ASSERT(MUTEX_HELD(&dtrace_lock));
11996 
11997 	if (id == 0 || id > state->dts_necbs)
11998 		return (NULL);
11999 
12000 	ASSERT(state->dts_necbs > 0 && state->dts_ecbs != NULL);
12001 	ASSERT((ecb = state->dts_ecbs[id - 1]) == NULL || ecb->dte_epid == id);
12002 
12003 	return (state->dts_ecbs[id - 1]);
12004 }
12005 
12006 static dtrace_aggregation_t *
12007 dtrace_aggid2agg(dtrace_state_t *state, dtrace_aggid_t id)
12008 {
12009 	dtrace_aggregation_t *agg;
12010 
12011 	ASSERT(MUTEX_HELD(&dtrace_lock));
12012 
12013 	if (id == 0 || id > state->dts_naggregations)
12014 		return (NULL);
12015 
12016 	ASSERT(state->dts_naggregations > 0 && state->dts_aggregations != NULL);
12017 	ASSERT((agg = state->dts_aggregations[id - 1]) == NULL ||
12018 	    agg->dtag_id == id);
12019 
12020 	return (state->dts_aggregations[id - 1]);
12021 }
12022 
12023 /*
12024  * DTrace Buffer Functions
12025  *
12026  * The following functions manipulate DTrace buffers.  Most of these functions
12027  * are called in the context of establishing or processing consumer state;
12028  * exceptions are explicitly noted.
12029  */
12030 
12031 /*
12032  * Note:  called from cross call context.  This function switches the two
12033  * buffers on a given CPU.  The atomicity of this operation is assured by
12034  * disabling interrupts while the actual switch takes place; the disabling of
12035  * interrupts serializes the execution with any execution of dtrace_probe() on
12036  * the same CPU.
12037  */
12038 static void
12039 dtrace_buffer_switch(dtrace_buffer_t *buf)
12040 {
12041 	caddr_t tomax = buf->dtb_tomax;
12042 	caddr_t xamot = buf->dtb_xamot;
12043 	dtrace_icookie_t cookie;
12044 	hrtime_t now;
12045 
12046 	ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
12047 	ASSERT(!(buf->dtb_flags & DTRACEBUF_RING));
12048 
12049 	cookie = dtrace_interrupt_disable();
12050 	now = dtrace_gethrtime();
12051 	buf->dtb_tomax = xamot;
12052 	buf->dtb_xamot = tomax;
12053 	buf->dtb_xamot_drops = buf->dtb_drops;
12054 	buf->dtb_xamot_offset = buf->dtb_offset;
12055 	buf->dtb_xamot_errors = buf->dtb_errors;
12056 	buf->dtb_xamot_flags = buf->dtb_flags;
12057 	buf->dtb_offset = 0;
12058 	buf->dtb_drops = 0;
12059 	buf->dtb_errors = 0;
12060 	buf->dtb_flags &= ~(DTRACEBUF_ERROR | DTRACEBUF_DROPPED);
12061 	buf->dtb_interval = now - buf->dtb_switched;
12062 	buf->dtb_switched = now;
12063 	dtrace_interrupt_enable(cookie);
12064 }
12065 
12066 /*
12067  * Note:  called from cross call context.  This function activates a buffer
12068  * on a CPU.  As with dtrace_buffer_switch(), the atomicity of the operation
12069  * is guaranteed by the disabling of interrupts.
12070  */
12071 static void
12072 dtrace_buffer_activate(dtrace_state_t *state)
12073 {
12074 	dtrace_buffer_t *buf;
12075 	dtrace_icookie_t cookie = dtrace_interrupt_disable();
12076 
12077 	buf = &state->dts_buffer[curcpu];
12078 
12079 	if (buf->dtb_tomax != NULL) {
12080 		/*
12081 		 * We might like to assert that the buffer is marked inactive,
12082 		 * but this isn't necessarily true:  the buffer for the CPU
12083 		 * that processes the BEGIN probe has its buffer activated
12084 		 * manually.  In this case, we take the (harmless) action
12085 		 * re-clearing the bit INACTIVE bit.
12086 		 */
12087 		buf->dtb_flags &= ~DTRACEBUF_INACTIVE;
12088 	}
12089 
12090 	dtrace_interrupt_enable(cookie);
12091 }
12092 
12093 #ifdef __FreeBSD__
12094 /*
12095  * Activate the specified per-CPU buffer.  This is used instead of
12096  * dtrace_buffer_activate() when APs have not yet started, i.e. when
12097  * activating anonymous state.
12098  */
12099 static void
12100 dtrace_buffer_activate_cpu(dtrace_state_t *state, int cpu)
12101 {
12102 
12103 	if (state->dts_buffer[cpu].dtb_tomax != NULL)
12104 		state->dts_buffer[cpu].dtb_flags &= ~DTRACEBUF_INACTIVE;
12105 }
12106 #endif
12107 
12108 static int
12109 dtrace_buffer_alloc(dtrace_buffer_t *bufs, size_t size, int flags,
12110     processorid_t cpu, int *factor)
12111 {
12112 #ifdef illumos
12113 	cpu_t *cp;
12114 #endif
12115 	dtrace_buffer_t *buf;
12116 	int allocated = 0, desired = 0;
12117 
12118 #ifdef illumos
12119 	ASSERT(MUTEX_HELD(&cpu_lock));
12120 	ASSERT(MUTEX_HELD(&dtrace_lock));
12121 
12122 	*factor = 1;
12123 
12124 	if (size > dtrace_nonroot_maxsize &&
12125 	    !PRIV_POLICY_CHOICE(CRED(), PRIV_ALL, B_FALSE))
12126 		return (EFBIG);
12127 
12128 	cp = cpu_list;
12129 
12130 	do {
12131 		if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
12132 			continue;
12133 
12134 		buf = &bufs[cp->cpu_id];
12135 
12136 		/*
12137 		 * If there is already a buffer allocated for this CPU, it
12138 		 * is only possible that this is a DR event.  In this case,
12139 		 */
12140 		if (buf->dtb_tomax != NULL) {
12141 			ASSERT(buf->dtb_size == size);
12142 			continue;
12143 		}
12144 
12145 		ASSERT(buf->dtb_xamot == NULL);
12146 
12147 		if ((buf->dtb_tomax = kmem_zalloc(size,
12148 		    KM_NOSLEEP | KM_NORMALPRI)) == NULL)
12149 			goto err;
12150 
12151 		buf->dtb_size = size;
12152 		buf->dtb_flags = flags;
12153 		buf->dtb_offset = 0;
12154 		buf->dtb_drops = 0;
12155 
12156 		if (flags & DTRACEBUF_NOSWITCH)
12157 			continue;
12158 
12159 		if ((buf->dtb_xamot = kmem_zalloc(size,
12160 		    KM_NOSLEEP | KM_NORMALPRI)) == NULL)
12161 			goto err;
12162 	} while ((cp = cp->cpu_next) != cpu_list);
12163 
12164 	return (0);
12165 
12166 err:
12167 	cp = cpu_list;
12168 
12169 	do {
12170 		if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
12171 			continue;
12172 
12173 		buf = &bufs[cp->cpu_id];
12174 		desired += 2;
12175 
12176 		if (buf->dtb_xamot != NULL) {
12177 			ASSERT(buf->dtb_tomax != NULL);
12178 			ASSERT(buf->dtb_size == size);
12179 			kmem_free(buf->dtb_xamot, size);
12180 			allocated++;
12181 		}
12182 
12183 		if (buf->dtb_tomax != NULL) {
12184 			ASSERT(buf->dtb_size == size);
12185 			kmem_free(buf->dtb_tomax, size);
12186 			allocated++;
12187 		}
12188 
12189 		buf->dtb_tomax = NULL;
12190 		buf->dtb_xamot = NULL;
12191 		buf->dtb_size = 0;
12192 	} while ((cp = cp->cpu_next) != cpu_list);
12193 #else
12194 	int i;
12195 
12196 	*factor = 1;
12197 #if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \
12198     defined(__mips__) || defined(__powerpc__) || defined(__riscv)
12199 	/*
12200 	 * FreeBSD isn't good at limiting the amount of memory we
12201 	 * ask to malloc, so let's place a limit here before trying
12202 	 * to do something that might well end in tears at bedtime.
12203 	 */
12204 	if (size > physmem * PAGE_SIZE / (128 * (mp_maxid + 1)))
12205 		return (ENOMEM);
12206 #endif
12207 
12208 	ASSERT(MUTEX_HELD(&dtrace_lock));
12209 	CPU_FOREACH(i) {
12210 		if (cpu != DTRACE_CPUALL && cpu != i)
12211 			continue;
12212 
12213 		buf = &bufs[i];
12214 
12215 		/*
12216 		 * If there is already a buffer allocated for this CPU, it
12217 		 * is only possible that this is a DR event.  In this case,
12218 		 * the buffer size must match our specified size.
12219 		 */
12220 		if (buf->dtb_tomax != NULL) {
12221 			ASSERT(buf->dtb_size == size);
12222 			continue;
12223 		}
12224 
12225 		ASSERT(buf->dtb_xamot == NULL);
12226 
12227 		if ((buf->dtb_tomax = kmem_zalloc(size,
12228 		    KM_NOSLEEP | KM_NORMALPRI)) == NULL)
12229 			goto err;
12230 
12231 		buf->dtb_size = size;
12232 		buf->dtb_flags = flags;
12233 		buf->dtb_offset = 0;
12234 		buf->dtb_drops = 0;
12235 
12236 		if (flags & DTRACEBUF_NOSWITCH)
12237 			continue;
12238 
12239 		if ((buf->dtb_xamot = kmem_zalloc(size,
12240 		    KM_NOSLEEP | KM_NORMALPRI)) == NULL)
12241 			goto err;
12242 	}
12243 
12244 	return (0);
12245 
12246 err:
12247 	/*
12248 	 * Error allocating memory, so free the buffers that were
12249 	 * allocated before the failed allocation.
12250 	 */
12251 	CPU_FOREACH(i) {
12252 		if (cpu != DTRACE_CPUALL && cpu != i)
12253 			continue;
12254 
12255 		buf = &bufs[i];
12256 		desired += 2;
12257 
12258 		if (buf->dtb_xamot != NULL) {
12259 			ASSERT(buf->dtb_tomax != NULL);
12260 			ASSERT(buf->dtb_size == size);
12261 			kmem_free(buf->dtb_xamot, size);
12262 			allocated++;
12263 		}
12264 
12265 		if (buf->dtb_tomax != NULL) {
12266 			ASSERT(buf->dtb_size == size);
12267 			kmem_free(buf->dtb_tomax, size);
12268 			allocated++;
12269 		}
12270 
12271 		buf->dtb_tomax = NULL;
12272 		buf->dtb_xamot = NULL;
12273 		buf->dtb_size = 0;
12274 
12275 	}
12276 #endif
12277 	*factor = desired / (allocated > 0 ? allocated : 1);
12278 
12279 	return (ENOMEM);
12280 }
12281 
12282 /*
12283  * Note:  called from probe context.  This function just increments the drop
12284  * count on a buffer.  It has been made a function to allow for the
12285  * possibility of understanding the source of mysterious drop counts.  (A
12286  * problem for which one may be particularly disappointed that DTrace cannot
12287  * be used to understand DTrace.)
12288  */
12289 static void
12290 dtrace_buffer_drop(dtrace_buffer_t *buf)
12291 {
12292 	buf->dtb_drops++;
12293 }
12294 
12295 /*
12296  * Note:  called from probe context.  This function is called to reserve space
12297  * in a buffer.  If mstate is non-NULL, sets the scratch base and size in the
12298  * mstate.  Returns the new offset in the buffer, or a negative value if an
12299  * error has occurred.
12300  */
12301 static intptr_t
12302 dtrace_buffer_reserve(dtrace_buffer_t *buf, size_t needed, size_t align,
12303     dtrace_state_t *state, dtrace_mstate_t *mstate)
12304 {
12305 	intptr_t offs = buf->dtb_offset, soffs;
12306 	intptr_t woffs;
12307 	caddr_t tomax;
12308 	size_t total;
12309 
12310 	if (buf->dtb_flags & DTRACEBUF_INACTIVE)
12311 		return (-1);
12312 
12313 	if ((tomax = buf->dtb_tomax) == NULL) {
12314 		dtrace_buffer_drop(buf);
12315 		return (-1);
12316 	}
12317 
12318 	if (!(buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL))) {
12319 		while (offs & (align - 1)) {
12320 			/*
12321 			 * Assert that our alignment is off by a number which
12322 			 * is itself sizeof (uint32_t) aligned.
12323 			 */
12324 			ASSERT(!((align - (offs & (align - 1))) &
12325 			    (sizeof (uint32_t) - 1)));
12326 			DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
12327 			offs += sizeof (uint32_t);
12328 		}
12329 
12330 		if ((soffs = offs + needed) > buf->dtb_size) {
12331 			dtrace_buffer_drop(buf);
12332 			return (-1);
12333 		}
12334 
12335 		if (mstate == NULL)
12336 			return (offs);
12337 
12338 		mstate->dtms_scratch_base = (uintptr_t)tomax + soffs;
12339 		mstate->dtms_scratch_size = buf->dtb_size - soffs;
12340 		mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
12341 
12342 		return (offs);
12343 	}
12344 
12345 	if (buf->dtb_flags & DTRACEBUF_FILL) {
12346 		if (state->dts_activity != DTRACE_ACTIVITY_COOLDOWN &&
12347 		    (buf->dtb_flags & DTRACEBUF_FULL))
12348 			return (-1);
12349 		goto out;
12350 	}
12351 
12352 	total = needed + (offs & (align - 1));
12353 
12354 	/*
12355 	 * For a ring buffer, life is quite a bit more complicated.  Before
12356 	 * we can store any padding, we need to adjust our wrapping offset.
12357 	 * (If we've never before wrapped or we're not about to, no adjustment
12358 	 * is required.)
12359 	 */
12360 	if ((buf->dtb_flags & DTRACEBUF_WRAPPED) ||
12361 	    offs + total > buf->dtb_size) {
12362 		woffs = buf->dtb_xamot_offset;
12363 
12364 		if (offs + total > buf->dtb_size) {
12365 			/*
12366 			 * We can't fit in the end of the buffer.  First, a
12367 			 * sanity check that we can fit in the buffer at all.
12368 			 */
12369 			if (total > buf->dtb_size) {
12370 				dtrace_buffer_drop(buf);
12371 				return (-1);
12372 			}
12373 
12374 			/*
12375 			 * We're going to be storing at the top of the buffer,
12376 			 * so now we need to deal with the wrapped offset.  We
12377 			 * only reset our wrapped offset to 0 if it is
12378 			 * currently greater than the current offset.  If it
12379 			 * is less than the current offset, it is because a
12380 			 * previous allocation induced a wrap -- but the
12381 			 * allocation didn't subsequently take the space due
12382 			 * to an error or false predicate evaluation.  In this
12383 			 * case, we'll just leave the wrapped offset alone: if
12384 			 * the wrapped offset hasn't been advanced far enough
12385 			 * for this allocation, it will be adjusted in the
12386 			 * lower loop.
12387 			 */
12388 			if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
12389 				if (woffs >= offs)
12390 					woffs = 0;
12391 			} else {
12392 				woffs = 0;
12393 			}
12394 
12395 			/*
12396 			 * Now we know that we're going to be storing to the
12397 			 * top of the buffer and that there is room for us
12398 			 * there.  We need to clear the buffer from the current
12399 			 * offset to the end (there may be old gunk there).
12400 			 */
12401 			while (offs < buf->dtb_size)
12402 				tomax[offs++] = 0;
12403 
12404 			/*
12405 			 * We need to set our offset to zero.  And because we
12406 			 * are wrapping, we need to set the bit indicating as
12407 			 * much.  We can also adjust our needed space back
12408 			 * down to the space required by the ECB -- we know
12409 			 * that the top of the buffer is aligned.
12410 			 */
12411 			offs = 0;
12412 			total = needed;
12413 			buf->dtb_flags |= DTRACEBUF_WRAPPED;
12414 		} else {
12415 			/*
12416 			 * There is room for us in the buffer, so we simply
12417 			 * need to check the wrapped offset.
12418 			 */
12419 			if (woffs < offs) {
12420 				/*
12421 				 * The wrapped offset is less than the offset.
12422 				 * This can happen if we allocated buffer space
12423 				 * that induced a wrap, but then we didn't
12424 				 * subsequently take the space due to an error
12425 				 * or false predicate evaluation.  This is
12426 				 * okay; we know that _this_ allocation isn't
12427 				 * going to induce a wrap.  We still can't
12428 				 * reset the wrapped offset to be zero,
12429 				 * however: the space may have been trashed in
12430 				 * the previous failed probe attempt.  But at
12431 				 * least the wrapped offset doesn't need to
12432 				 * be adjusted at all...
12433 				 */
12434 				goto out;
12435 			}
12436 		}
12437 
12438 		while (offs + total > woffs) {
12439 			dtrace_epid_t epid = *(uint32_t *)(tomax + woffs);
12440 			size_t size;
12441 
12442 			if (epid == DTRACE_EPIDNONE) {
12443 				size = sizeof (uint32_t);
12444 			} else {
12445 				ASSERT3U(epid, <=, state->dts_necbs);
12446 				ASSERT(state->dts_ecbs[epid - 1] != NULL);
12447 
12448 				size = state->dts_ecbs[epid - 1]->dte_size;
12449 			}
12450 
12451 			ASSERT(woffs + size <= buf->dtb_size);
12452 			ASSERT(size != 0);
12453 
12454 			if (woffs + size == buf->dtb_size) {
12455 				/*
12456 				 * We've reached the end of the buffer; we want
12457 				 * to set the wrapped offset to 0 and break
12458 				 * out.  However, if the offs is 0, then we're
12459 				 * in a strange edge-condition:  the amount of
12460 				 * space that we want to reserve plus the size
12461 				 * of the record that we're overwriting is
12462 				 * greater than the size of the buffer.  This
12463 				 * is problematic because if we reserve the
12464 				 * space but subsequently don't consume it (due
12465 				 * to a failed predicate or error) the wrapped
12466 				 * offset will be 0 -- yet the EPID at offset 0
12467 				 * will not be committed.  This situation is
12468 				 * relatively easy to deal with:  if we're in
12469 				 * this case, the buffer is indistinguishable
12470 				 * from one that hasn't wrapped; we need only
12471 				 * finish the job by clearing the wrapped bit,
12472 				 * explicitly setting the offset to be 0, and
12473 				 * zero'ing out the old data in the buffer.
12474 				 */
12475 				if (offs == 0) {
12476 					buf->dtb_flags &= ~DTRACEBUF_WRAPPED;
12477 					buf->dtb_offset = 0;
12478 					woffs = total;
12479 
12480 					while (woffs < buf->dtb_size)
12481 						tomax[woffs++] = 0;
12482 				}
12483 
12484 				woffs = 0;
12485 				break;
12486 			}
12487 
12488 			woffs += size;
12489 		}
12490 
12491 		/*
12492 		 * We have a wrapped offset.  It may be that the wrapped offset
12493 		 * has become zero -- that's okay.
12494 		 */
12495 		buf->dtb_xamot_offset = woffs;
12496 	}
12497 
12498 out:
12499 	/*
12500 	 * Now we can plow the buffer with any necessary padding.
12501 	 */
12502 	while (offs & (align - 1)) {
12503 		/*
12504 		 * Assert that our alignment is off by a number which
12505 		 * is itself sizeof (uint32_t) aligned.
12506 		 */
12507 		ASSERT(!((align - (offs & (align - 1))) &
12508 		    (sizeof (uint32_t) - 1)));
12509 		DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
12510 		offs += sizeof (uint32_t);
12511 	}
12512 
12513 	if (buf->dtb_flags & DTRACEBUF_FILL) {
12514 		if (offs + needed > buf->dtb_size - state->dts_reserve) {
12515 			buf->dtb_flags |= DTRACEBUF_FULL;
12516 			return (-1);
12517 		}
12518 	}
12519 
12520 	if (mstate == NULL)
12521 		return (offs);
12522 
12523 	/*
12524 	 * For ring buffers and fill buffers, the scratch space is always
12525 	 * the inactive buffer.
12526 	 */
12527 	mstate->dtms_scratch_base = (uintptr_t)buf->dtb_xamot;
12528 	mstate->dtms_scratch_size = buf->dtb_size;
12529 	mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
12530 
12531 	return (offs);
12532 }
12533 
12534 static void
12535 dtrace_buffer_polish(dtrace_buffer_t *buf)
12536 {
12537 	ASSERT(buf->dtb_flags & DTRACEBUF_RING);
12538 	ASSERT(MUTEX_HELD(&dtrace_lock));
12539 
12540 	if (!(buf->dtb_flags & DTRACEBUF_WRAPPED))
12541 		return;
12542 
12543 	/*
12544 	 * We need to polish the ring buffer.  There are three cases:
12545 	 *
12546 	 * - The first (and presumably most common) is that there is no gap
12547 	 *   between the buffer offset and the wrapped offset.  In this case,
12548 	 *   there is nothing in the buffer that isn't valid data; we can
12549 	 *   mark the buffer as polished and return.
12550 	 *
12551 	 * - The second (less common than the first but still more common
12552 	 *   than the third) is that there is a gap between the buffer offset
12553 	 *   and the wrapped offset, and the wrapped offset is larger than the
12554 	 *   buffer offset.  This can happen because of an alignment issue, or
12555 	 *   can happen because of a call to dtrace_buffer_reserve() that
12556 	 *   didn't subsequently consume the buffer space.  In this case,
12557 	 *   we need to zero the data from the buffer offset to the wrapped
12558 	 *   offset.
12559 	 *
12560 	 * - The third (and least common) is that there is a gap between the
12561 	 *   buffer offset and the wrapped offset, but the wrapped offset is
12562 	 *   _less_ than the buffer offset.  This can only happen because a
12563 	 *   call to dtrace_buffer_reserve() induced a wrap, but the space
12564 	 *   was not subsequently consumed.  In this case, we need to zero the
12565 	 *   space from the offset to the end of the buffer _and_ from the
12566 	 *   top of the buffer to the wrapped offset.
12567 	 */
12568 	if (buf->dtb_offset < buf->dtb_xamot_offset) {
12569 		bzero(buf->dtb_tomax + buf->dtb_offset,
12570 		    buf->dtb_xamot_offset - buf->dtb_offset);
12571 	}
12572 
12573 	if (buf->dtb_offset > buf->dtb_xamot_offset) {
12574 		bzero(buf->dtb_tomax + buf->dtb_offset,
12575 		    buf->dtb_size - buf->dtb_offset);
12576 		bzero(buf->dtb_tomax, buf->dtb_xamot_offset);
12577 	}
12578 }
12579 
12580 /*
12581  * This routine determines if data generated at the specified time has likely
12582  * been entirely consumed at user-level.  This routine is called to determine
12583  * if an ECB on a defunct probe (but for an active enabling) can be safely
12584  * disabled and destroyed.
12585  */
12586 static int
12587 dtrace_buffer_consumed(dtrace_buffer_t *bufs, hrtime_t when)
12588 {
12589 	int i;
12590 
12591 	for (i = 0; i < NCPU; i++) {
12592 		dtrace_buffer_t *buf = &bufs[i];
12593 
12594 		if (buf->dtb_size == 0)
12595 			continue;
12596 
12597 		if (buf->dtb_flags & DTRACEBUF_RING)
12598 			return (0);
12599 
12600 		if (!buf->dtb_switched && buf->dtb_offset != 0)
12601 			return (0);
12602 
12603 		if (buf->dtb_switched - buf->dtb_interval < when)
12604 			return (0);
12605 	}
12606 
12607 	return (1);
12608 }
12609 
12610 static void
12611 dtrace_buffer_free(dtrace_buffer_t *bufs)
12612 {
12613 	int i;
12614 
12615 	for (i = 0; i < NCPU; i++) {
12616 		dtrace_buffer_t *buf = &bufs[i];
12617 
12618 		if (buf->dtb_tomax == NULL) {
12619 			ASSERT(buf->dtb_xamot == NULL);
12620 			ASSERT(buf->dtb_size == 0);
12621 			continue;
12622 		}
12623 
12624 		if (buf->dtb_xamot != NULL) {
12625 			ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
12626 			kmem_free(buf->dtb_xamot, buf->dtb_size);
12627 		}
12628 
12629 		kmem_free(buf->dtb_tomax, buf->dtb_size);
12630 		buf->dtb_size = 0;
12631 		buf->dtb_tomax = NULL;
12632 		buf->dtb_xamot = NULL;
12633 	}
12634 }
12635 
12636 /*
12637  * DTrace Enabling Functions
12638  */
12639 static dtrace_enabling_t *
12640 dtrace_enabling_create(dtrace_vstate_t *vstate)
12641 {
12642 	dtrace_enabling_t *enab;
12643 
12644 	enab = kmem_zalloc(sizeof (dtrace_enabling_t), KM_SLEEP);
12645 	enab->dten_vstate = vstate;
12646 
12647 	return (enab);
12648 }
12649 
12650 static void
12651 dtrace_enabling_add(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb)
12652 {
12653 	dtrace_ecbdesc_t **ndesc;
12654 	size_t osize, nsize;
12655 
12656 	/*
12657 	 * We can't add to enablings after we've enabled them, or after we've
12658 	 * retained them.
12659 	 */
12660 	ASSERT(enab->dten_probegen == 0);
12661 	ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
12662 
12663 	if (enab->dten_ndesc < enab->dten_maxdesc) {
12664 		enab->dten_desc[enab->dten_ndesc++] = ecb;
12665 		return;
12666 	}
12667 
12668 	osize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
12669 
12670 	if (enab->dten_maxdesc == 0) {
12671 		enab->dten_maxdesc = 1;
12672 	} else {
12673 		enab->dten_maxdesc <<= 1;
12674 	}
12675 
12676 	ASSERT(enab->dten_ndesc < enab->dten_maxdesc);
12677 
12678 	nsize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
12679 	ndesc = kmem_zalloc(nsize, KM_SLEEP);
12680 	bcopy(enab->dten_desc, ndesc, osize);
12681 	if (enab->dten_desc != NULL)
12682 		kmem_free(enab->dten_desc, osize);
12683 
12684 	enab->dten_desc = ndesc;
12685 	enab->dten_desc[enab->dten_ndesc++] = ecb;
12686 }
12687 
12688 static void
12689 dtrace_enabling_addlike(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb,
12690     dtrace_probedesc_t *pd)
12691 {
12692 	dtrace_ecbdesc_t *new;
12693 	dtrace_predicate_t *pred;
12694 	dtrace_actdesc_t *act;
12695 
12696 	/*
12697 	 * We're going to create a new ECB description that matches the
12698 	 * specified ECB in every way, but has the specified probe description.
12699 	 */
12700 	new = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
12701 
12702 	if ((pred = ecb->dted_pred.dtpdd_predicate) != NULL)
12703 		dtrace_predicate_hold(pred);
12704 
12705 	for (act = ecb->dted_action; act != NULL; act = act->dtad_next)
12706 		dtrace_actdesc_hold(act);
12707 
12708 	new->dted_action = ecb->dted_action;
12709 	new->dted_pred = ecb->dted_pred;
12710 	new->dted_probe = *pd;
12711 	new->dted_uarg = ecb->dted_uarg;
12712 
12713 	dtrace_enabling_add(enab, new);
12714 }
12715 
12716 static void
12717 dtrace_enabling_dump(dtrace_enabling_t *enab)
12718 {
12719 	int i;
12720 
12721 	for (i = 0; i < enab->dten_ndesc; i++) {
12722 		dtrace_probedesc_t *desc = &enab->dten_desc[i]->dted_probe;
12723 
12724 #ifdef __FreeBSD__
12725 		printf("dtrace: enabling probe %d (%s:%s:%s:%s)\n", i,
12726 		    desc->dtpd_provider, desc->dtpd_mod,
12727 		    desc->dtpd_func, desc->dtpd_name);
12728 #else
12729 		cmn_err(CE_NOTE, "enabling probe %d (%s:%s:%s:%s)", i,
12730 		    desc->dtpd_provider, desc->dtpd_mod,
12731 		    desc->dtpd_func, desc->dtpd_name);
12732 #endif
12733 	}
12734 }
12735 
12736 static void
12737 dtrace_enabling_destroy(dtrace_enabling_t *enab)
12738 {
12739 	int i;
12740 	dtrace_ecbdesc_t *ep;
12741 	dtrace_vstate_t *vstate = enab->dten_vstate;
12742 
12743 	ASSERT(MUTEX_HELD(&dtrace_lock));
12744 
12745 	for (i = 0; i < enab->dten_ndesc; i++) {
12746 		dtrace_actdesc_t *act, *next;
12747 		dtrace_predicate_t *pred;
12748 
12749 		ep = enab->dten_desc[i];
12750 
12751 		if ((pred = ep->dted_pred.dtpdd_predicate) != NULL)
12752 			dtrace_predicate_release(pred, vstate);
12753 
12754 		for (act = ep->dted_action; act != NULL; act = next) {
12755 			next = act->dtad_next;
12756 			dtrace_actdesc_release(act, vstate);
12757 		}
12758 
12759 		kmem_free(ep, sizeof (dtrace_ecbdesc_t));
12760 	}
12761 
12762 	if (enab->dten_desc != NULL)
12763 		kmem_free(enab->dten_desc,
12764 		    enab->dten_maxdesc * sizeof (dtrace_enabling_t *));
12765 
12766 	/*
12767 	 * If this was a retained enabling, decrement the dts_nretained count
12768 	 * and take it off of the dtrace_retained list.
12769 	 */
12770 	if (enab->dten_prev != NULL || enab->dten_next != NULL ||
12771 	    dtrace_retained == enab) {
12772 		ASSERT(enab->dten_vstate->dtvs_state != NULL);
12773 		ASSERT(enab->dten_vstate->dtvs_state->dts_nretained > 0);
12774 		enab->dten_vstate->dtvs_state->dts_nretained--;
12775 		dtrace_retained_gen++;
12776 	}
12777 
12778 	if (enab->dten_prev == NULL) {
12779 		if (dtrace_retained == enab) {
12780 			dtrace_retained = enab->dten_next;
12781 
12782 			if (dtrace_retained != NULL)
12783 				dtrace_retained->dten_prev = NULL;
12784 		}
12785 	} else {
12786 		ASSERT(enab != dtrace_retained);
12787 		ASSERT(dtrace_retained != NULL);
12788 		enab->dten_prev->dten_next = enab->dten_next;
12789 	}
12790 
12791 	if (enab->dten_next != NULL) {
12792 		ASSERT(dtrace_retained != NULL);
12793 		enab->dten_next->dten_prev = enab->dten_prev;
12794 	}
12795 
12796 	kmem_free(enab, sizeof (dtrace_enabling_t));
12797 }
12798 
12799 static int
12800 dtrace_enabling_retain(dtrace_enabling_t *enab)
12801 {
12802 	dtrace_state_t *state;
12803 
12804 	ASSERT(MUTEX_HELD(&dtrace_lock));
12805 	ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
12806 	ASSERT(enab->dten_vstate != NULL);
12807 
12808 	state = enab->dten_vstate->dtvs_state;
12809 	ASSERT(state != NULL);
12810 
12811 	/*
12812 	 * We only allow each state to retain dtrace_retain_max enablings.
12813 	 */
12814 	if (state->dts_nretained >= dtrace_retain_max)
12815 		return (ENOSPC);
12816 
12817 	state->dts_nretained++;
12818 	dtrace_retained_gen++;
12819 
12820 	if (dtrace_retained == NULL) {
12821 		dtrace_retained = enab;
12822 		return (0);
12823 	}
12824 
12825 	enab->dten_next = dtrace_retained;
12826 	dtrace_retained->dten_prev = enab;
12827 	dtrace_retained = enab;
12828 
12829 	return (0);
12830 }
12831 
12832 static int
12833 dtrace_enabling_replicate(dtrace_state_t *state, dtrace_probedesc_t *match,
12834     dtrace_probedesc_t *create)
12835 {
12836 	dtrace_enabling_t *new, *enab;
12837 	int found = 0, err = ENOENT;
12838 
12839 	ASSERT(MUTEX_HELD(&dtrace_lock));
12840 	ASSERT(strlen(match->dtpd_provider) < DTRACE_PROVNAMELEN);
12841 	ASSERT(strlen(match->dtpd_mod) < DTRACE_MODNAMELEN);
12842 	ASSERT(strlen(match->dtpd_func) < DTRACE_FUNCNAMELEN);
12843 	ASSERT(strlen(match->dtpd_name) < DTRACE_NAMELEN);
12844 
12845 	new = dtrace_enabling_create(&state->dts_vstate);
12846 
12847 	/*
12848 	 * Iterate over all retained enablings, looking for enablings that
12849 	 * match the specified state.
12850 	 */
12851 	for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
12852 		int i;
12853 
12854 		/*
12855 		 * dtvs_state can only be NULL for helper enablings -- and
12856 		 * helper enablings can't be retained.
12857 		 */
12858 		ASSERT(enab->dten_vstate->dtvs_state != NULL);
12859 
12860 		if (enab->dten_vstate->dtvs_state != state)
12861 			continue;
12862 
12863 		/*
12864 		 * Now iterate over each probe description; we're looking for
12865 		 * an exact match to the specified probe description.
12866 		 */
12867 		for (i = 0; i < enab->dten_ndesc; i++) {
12868 			dtrace_ecbdesc_t *ep = enab->dten_desc[i];
12869 			dtrace_probedesc_t *pd = &ep->dted_probe;
12870 
12871 			if (strcmp(pd->dtpd_provider, match->dtpd_provider))
12872 				continue;
12873 
12874 			if (strcmp(pd->dtpd_mod, match->dtpd_mod))
12875 				continue;
12876 
12877 			if (strcmp(pd->dtpd_func, match->dtpd_func))
12878 				continue;
12879 
12880 			if (strcmp(pd->dtpd_name, match->dtpd_name))
12881 				continue;
12882 
12883 			/*
12884 			 * We have a winning probe!  Add it to our growing
12885 			 * enabling.
12886 			 */
12887 			found = 1;
12888 			dtrace_enabling_addlike(new, ep, create);
12889 		}
12890 	}
12891 
12892 	if (!found || (err = dtrace_enabling_retain(new)) != 0) {
12893 		dtrace_enabling_destroy(new);
12894 		return (err);
12895 	}
12896 
12897 	return (0);
12898 }
12899 
12900 static void
12901 dtrace_enabling_retract(dtrace_state_t *state)
12902 {
12903 	dtrace_enabling_t *enab, *next;
12904 
12905 	ASSERT(MUTEX_HELD(&dtrace_lock));
12906 
12907 	/*
12908 	 * Iterate over all retained enablings, destroy the enablings retained
12909 	 * for the specified state.
12910 	 */
12911 	for (enab = dtrace_retained; enab != NULL; enab = next) {
12912 		next = enab->dten_next;
12913 
12914 		/*
12915 		 * dtvs_state can only be NULL for helper enablings -- and
12916 		 * helper enablings can't be retained.
12917 		 */
12918 		ASSERT(enab->dten_vstate->dtvs_state != NULL);
12919 
12920 		if (enab->dten_vstate->dtvs_state == state) {
12921 			ASSERT(state->dts_nretained > 0);
12922 			dtrace_enabling_destroy(enab);
12923 		}
12924 	}
12925 
12926 	ASSERT(state->dts_nretained == 0);
12927 }
12928 
12929 static int
12930 dtrace_enabling_match(dtrace_enabling_t *enab, int *nmatched)
12931 {
12932 	int i = 0;
12933 	int matched = 0;
12934 
12935 	ASSERT(MUTEX_HELD(&cpu_lock));
12936 	ASSERT(MUTEX_HELD(&dtrace_lock));
12937 
12938 	for (i = 0; i < enab->dten_ndesc; i++) {
12939 		dtrace_ecbdesc_t *ep = enab->dten_desc[i];
12940 
12941 		enab->dten_current = ep;
12942 		enab->dten_error = 0;
12943 
12944 		matched += dtrace_probe_enable(&ep->dted_probe, enab);
12945 
12946 		if (enab->dten_error != 0) {
12947 			/*
12948 			 * If we get an error half-way through enabling the
12949 			 * probes, we kick out -- perhaps with some number of
12950 			 * them enabled.  Leaving enabled probes enabled may
12951 			 * be slightly confusing for user-level, but we expect
12952 			 * that no one will attempt to actually drive on in
12953 			 * the face of such errors.  If this is an anonymous
12954 			 * enabling (indicated with a NULL nmatched pointer),
12955 			 * we cmn_err() a message.  We aren't expecting to
12956 			 * get such an error -- such as it can exist at all,
12957 			 * it would be a result of corrupted DOF in the driver
12958 			 * properties.
12959 			 */
12960 			if (nmatched == NULL) {
12961 				cmn_err(CE_WARN, "dtrace_enabling_match() "
12962 				    "error on %p: %d", (void *)ep,
12963 				    enab->dten_error);
12964 			}
12965 
12966 			return (enab->dten_error);
12967 		}
12968 	}
12969 
12970 	enab->dten_probegen = dtrace_probegen;
12971 	if (nmatched != NULL)
12972 		*nmatched = matched;
12973 
12974 	return (0);
12975 }
12976 
12977 static void
12978 dtrace_enabling_matchall(void)
12979 {
12980 	dtrace_enabling_t *enab;
12981 
12982 	mutex_enter(&cpu_lock);
12983 	mutex_enter(&dtrace_lock);
12984 
12985 	/*
12986 	 * Iterate over all retained enablings to see if any probes match
12987 	 * against them.  We only perform this operation on enablings for which
12988 	 * we have sufficient permissions by virtue of being in the global zone
12989 	 * or in the same zone as the DTrace client.  Because we can be called
12990 	 * after dtrace_detach() has been called, we cannot assert that there
12991 	 * are retained enablings.  We can safely load from dtrace_retained,
12992 	 * however:  the taskq_destroy() at the end of dtrace_detach() will
12993 	 * block pending our completion.
12994 	 */
12995 	for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
12996 #ifdef illumos
12997 		cred_t *cr = enab->dten_vstate->dtvs_state->dts_cred.dcr_cred;
12998 
12999 		if (INGLOBALZONE(curproc) ||
13000 		    cr != NULL && getzoneid() == crgetzoneid(cr))
13001 #endif
13002 			(void) dtrace_enabling_match(enab, NULL);
13003 	}
13004 
13005 	mutex_exit(&dtrace_lock);
13006 	mutex_exit(&cpu_lock);
13007 }
13008 
13009 /*
13010  * If an enabling is to be enabled without having matched probes (that is, if
13011  * dtrace_state_go() is to be called on the underlying dtrace_state_t), the
13012  * enabling must be _primed_ by creating an ECB for every ECB description.
13013  * This must be done to assure that we know the number of speculations, the
13014  * number of aggregations, the minimum buffer size needed, etc. before we
13015  * transition out of DTRACE_ACTIVITY_INACTIVE.  To do this without actually
13016  * enabling any probes, we create ECBs for every ECB decription, but with a
13017  * NULL probe -- which is exactly what this function does.
13018  */
13019 static void
13020 dtrace_enabling_prime(dtrace_state_t *state)
13021 {
13022 	dtrace_enabling_t *enab;
13023 	int i;
13024 
13025 	for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
13026 		ASSERT(enab->dten_vstate->dtvs_state != NULL);
13027 
13028 		if (enab->dten_vstate->dtvs_state != state)
13029 			continue;
13030 
13031 		/*
13032 		 * We don't want to prime an enabling more than once, lest
13033 		 * we allow a malicious user to induce resource exhaustion.
13034 		 * (The ECBs that result from priming an enabling aren't
13035 		 * leaked -- but they also aren't deallocated until the
13036 		 * consumer state is destroyed.)
13037 		 */
13038 		if (enab->dten_primed)
13039 			continue;
13040 
13041 		for (i = 0; i < enab->dten_ndesc; i++) {
13042 			enab->dten_current = enab->dten_desc[i];
13043 			(void) dtrace_probe_enable(NULL, enab);
13044 		}
13045 
13046 		enab->dten_primed = 1;
13047 	}
13048 }
13049 
13050 /*
13051  * Called to indicate that probes should be provided due to retained
13052  * enablings.  This is implemented in terms of dtrace_probe_provide(), but it
13053  * must take an initial lap through the enabling calling the dtps_provide()
13054  * entry point explicitly to allow for autocreated probes.
13055  */
13056 static void
13057 dtrace_enabling_provide(dtrace_provider_t *prv)
13058 {
13059 	int i, all = 0;
13060 	dtrace_probedesc_t desc;
13061 	dtrace_genid_t gen;
13062 
13063 	ASSERT(MUTEX_HELD(&dtrace_lock));
13064 	ASSERT(MUTEX_HELD(&dtrace_provider_lock));
13065 
13066 	if (prv == NULL) {
13067 		all = 1;
13068 		prv = dtrace_provider;
13069 	}
13070 
13071 	do {
13072 		dtrace_enabling_t *enab;
13073 		void *parg = prv->dtpv_arg;
13074 
13075 retry:
13076 		gen = dtrace_retained_gen;
13077 		for (enab = dtrace_retained; enab != NULL;
13078 		    enab = enab->dten_next) {
13079 			for (i = 0; i < enab->dten_ndesc; i++) {
13080 				desc = enab->dten_desc[i]->dted_probe;
13081 				mutex_exit(&dtrace_lock);
13082 				prv->dtpv_pops.dtps_provide(parg, &desc);
13083 				mutex_enter(&dtrace_lock);
13084 				/*
13085 				 * Process the retained enablings again if
13086 				 * they have changed while we weren't holding
13087 				 * dtrace_lock.
13088 				 */
13089 				if (gen != dtrace_retained_gen)
13090 					goto retry;
13091 			}
13092 		}
13093 	} while (all && (prv = prv->dtpv_next) != NULL);
13094 
13095 	mutex_exit(&dtrace_lock);
13096 	dtrace_probe_provide(NULL, all ? NULL : prv);
13097 	mutex_enter(&dtrace_lock);
13098 }
13099 
13100 /*
13101  * Called to reap ECBs that are attached to probes from defunct providers.
13102  */
13103 static void
13104 dtrace_enabling_reap(void)
13105 {
13106 	dtrace_provider_t *prov;
13107 	dtrace_probe_t *probe;
13108 	dtrace_ecb_t *ecb;
13109 	hrtime_t when;
13110 	int i;
13111 
13112 	mutex_enter(&cpu_lock);
13113 	mutex_enter(&dtrace_lock);
13114 
13115 	for (i = 0; i < dtrace_nprobes; i++) {
13116 		if ((probe = dtrace_probes[i]) == NULL)
13117 			continue;
13118 
13119 		if (probe->dtpr_ecb == NULL)
13120 			continue;
13121 
13122 		prov = probe->dtpr_provider;
13123 
13124 		if ((when = prov->dtpv_defunct) == 0)
13125 			continue;
13126 
13127 		/*
13128 		 * We have ECBs on a defunct provider:  we want to reap these
13129 		 * ECBs to allow the provider to unregister.  The destruction
13130 		 * of these ECBs must be done carefully:  if we destroy the ECB
13131 		 * and the consumer later wishes to consume an EPID that
13132 		 * corresponds to the destroyed ECB (and if the EPID metadata
13133 		 * has not been previously consumed), the consumer will abort
13134 		 * processing on the unknown EPID.  To reduce (but not, sadly,
13135 		 * eliminate) the possibility of this, we will only destroy an
13136 		 * ECB for a defunct provider if, for the state that
13137 		 * corresponds to the ECB:
13138 		 *
13139 		 *  (a)	There is no speculative tracing (which can effectively
13140 		 *	cache an EPID for an arbitrary amount of time).
13141 		 *
13142 		 *  (b)	The principal buffers have been switched twice since the
13143 		 *	provider became defunct.
13144 		 *
13145 		 *  (c)	The aggregation buffers are of zero size or have been
13146 		 *	switched twice since the provider became defunct.
13147 		 *
13148 		 * We use dts_speculates to determine (a) and call a function
13149 		 * (dtrace_buffer_consumed()) to determine (b) and (c).  Note
13150 		 * that as soon as we've been unable to destroy one of the ECBs
13151 		 * associated with the probe, we quit trying -- reaping is only
13152 		 * fruitful in as much as we can destroy all ECBs associated
13153 		 * with the defunct provider's probes.
13154 		 */
13155 		while ((ecb = probe->dtpr_ecb) != NULL) {
13156 			dtrace_state_t *state = ecb->dte_state;
13157 			dtrace_buffer_t *buf = state->dts_buffer;
13158 			dtrace_buffer_t *aggbuf = state->dts_aggbuffer;
13159 
13160 			if (state->dts_speculates)
13161 				break;
13162 
13163 			if (!dtrace_buffer_consumed(buf, when))
13164 				break;
13165 
13166 			if (!dtrace_buffer_consumed(aggbuf, when))
13167 				break;
13168 
13169 			dtrace_ecb_disable(ecb);
13170 			ASSERT(probe->dtpr_ecb != ecb);
13171 			dtrace_ecb_destroy(ecb);
13172 		}
13173 	}
13174 
13175 	mutex_exit(&dtrace_lock);
13176 	mutex_exit(&cpu_lock);
13177 }
13178 
13179 /*
13180  * DTrace DOF Functions
13181  */
13182 /*ARGSUSED*/
13183 static void
13184 dtrace_dof_error(dof_hdr_t *dof, const char *str)
13185 {
13186 	if (dtrace_err_verbose)
13187 		cmn_err(CE_WARN, "failed to process DOF: %s", str);
13188 
13189 #ifdef DTRACE_ERRDEBUG
13190 	dtrace_errdebug(str);
13191 #endif
13192 }
13193 
13194 /*
13195  * Create DOF out of a currently enabled state.  Right now, we only create
13196  * DOF containing the run-time options -- but this could be expanded to create
13197  * complete DOF representing the enabled state.
13198  */
13199 static dof_hdr_t *
13200 dtrace_dof_create(dtrace_state_t *state)
13201 {
13202 	dof_hdr_t *dof;
13203 	dof_sec_t *sec;
13204 	dof_optdesc_t *opt;
13205 	int i, len = sizeof (dof_hdr_t) +
13206 	    roundup(sizeof (dof_sec_t), sizeof (uint64_t)) +
13207 	    sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
13208 
13209 	ASSERT(MUTEX_HELD(&dtrace_lock));
13210 
13211 	dof = kmem_zalloc(len, KM_SLEEP);
13212 	dof->dofh_ident[DOF_ID_MAG0] = DOF_MAG_MAG0;
13213 	dof->dofh_ident[DOF_ID_MAG1] = DOF_MAG_MAG1;
13214 	dof->dofh_ident[DOF_ID_MAG2] = DOF_MAG_MAG2;
13215 	dof->dofh_ident[DOF_ID_MAG3] = DOF_MAG_MAG3;
13216 
13217 	dof->dofh_ident[DOF_ID_MODEL] = DOF_MODEL_NATIVE;
13218 	dof->dofh_ident[DOF_ID_ENCODING] = DOF_ENCODE_NATIVE;
13219 	dof->dofh_ident[DOF_ID_VERSION] = DOF_VERSION;
13220 	dof->dofh_ident[DOF_ID_DIFVERS] = DIF_VERSION;
13221 	dof->dofh_ident[DOF_ID_DIFIREG] = DIF_DIR_NREGS;
13222 	dof->dofh_ident[DOF_ID_DIFTREG] = DIF_DTR_NREGS;
13223 
13224 	dof->dofh_flags = 0;
13225 	dof->dofh_hdrsize = sizeof (dof_hdr_t);
13226 	dof->dofh_secsize = sizeof (dof_sec_t);
13227 	dof->dofh_secnum = 1;	/* only DOF_SECT_OPTDESC */
13228 	dof->dofh_secoff = sizeof (dof_hdr_t);
13229 	dof->dofh_loadsz = len;
13230 	dof->dofh_filesz = len;
13231 	dof->dofh_pad = 0;
13232 
13233 	/*
13234 	 * Fill in the option section header...
13235 	 */
13236 	sec = (dof_sec_t *)((uintptr_t)dof + sizeof (dof_hdr_t));
13237 	sec->dofs_type = DOF_SECT_OPTDESC;
13238 	sec->dofs_align = sizeof (uint64_t);
13239 	sec->dofs_flags = DOF_SECF_LOAD;
13240 	sec->dofs_entsize = sizeof (dof_optdesc_t);
13241 
13242 	opt = (dof_optdesc_t *)((uintptr_t)sec +
13243 	    roundup(sizeof (dof_sec_t), sizeof (uint64_t)));
13244 
13245 	sec->dofs_offset = (uintptr_t)opt - (uintptr_t)dof;
13246 	sec->dofs_size = sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
13247 
13248 	for (i = 0; i < DTRACEOPT_MAX; i++) {
13249 		opt[i].dofo_option = i;
13250 		opt[i].dofo_strtab = DOF_SECIDX_NONE;
13251 		opt[i].dofo_value = state->dts_options[i];
13252 	}
13253 
13254 	return (dof);
13255 }
13256 
13257 static dof_hdr_t *
13258 dtrace_dof_copyin(uintptr_t uarg, int *errp)
13259 {
13260 	dof_hdr_t hdr, *dof;
13261 
13262 	ASSERT(!MUTEX_HELD(&dtrace_lock));
13263 
13264 	/*
13265 	 * First, we're going to copyin() the sizeof (dof_hdr_t).
13266 	 */
13267 	if (copyin((void *)uarg, &hdr, sizeof (hdr)) != 0) {
13268 		dtrace_dof_error(NULL, "failed to copyin DOF header");
13269 		*errp = EFAULT;
13270 		return (NULL);
13271 	}
13272 
13273 	/*
13274 	 * Now we'll allocate the entire DOF and copy it in -- provided
13275 	 * that the length isn't outrageous.
13276 	 */
13277 	if (hdr.dofh_loadsz >= dtrace_dof_maxsize) {
13278 		dtrace_dof_error(&hdr, "load size exceeds maximum");
13279 		*errp = E2BIG;
13280 		return (NULL);
13281 	}
13282 
13283 	if (hdr.dofh_loadsz < sizeof (hdr)) {
13284 		dtrace_dof_error(&hdr, "invalid load size");
13285 		*errp = EINVAL;
13286 		return (NULL);
13287 	}
13288 
13289 	dof = kmem_alloc(hdr.dofh_loadsz, KM_SLEEP);
13290 
13291 	if (copyin((void *)uarg, dof, hdr.dofh_loadsz) != 0 ||
13292 	    dof->dofh_loadsz != hdr.dofh_loadsz) {
13293 		kmem_free(dof, hdr.dofh_loadsz);
13294 		*errp = EFAULT;
13295 		return (NULL);
13296 	}
13297 
13298 	return (dof);
13299 }
13300 
13301 #ifdef __FreeBSD__
13302 static dof_hdr_t *
13303 dtrace_dof_copyin_proc(struct proc *p, uintptr_t uarg, int *errp)
13304 {
13305 	dof_hdr_t hdr, *dof;
13306 	struct thread *td;
13307 	size_t loadsz;
13308 
13309 	ASSERT(!MUTEX_HELD(&dtrace_lock));
13310 
13311 	td = curthread;
13312 
13313 	/*
13314 	 * First, we're going to copyin() the sizeof (dof_hdr_t).
13315 	 */
13316 	if (proc_readmem(td, p, uarg, &hdr, sizeof(hdr)) != sizeof(hdr)) {
13317 		dtrace_dof_error(NULL, "failed to copyin DOF header");
13318 		*errp = EFAULT;
13319 		return (NULL);
13320 	}
13321 
13322 	/*
13323 	 * Now we'll allocate the entire DOF and copy it in -- provided
13324 	 * that the length isn't outrageous.
13325 	 */
13326 	if (hdr.dofh_loadsz >= dtrace_dof_maxsize) {
13327 		dtrace_dof_error(&hdr, "load size exceeds maximum");
13328 		*errp = E2BIG;
13329 		return (NULL);
13330 	}
13331 	loadsz = (size_t)hdr.dofh_loadsz;
13332 
13333 	if (loadsz < sizeof (hdr)) {
13334 		dtrace_dof_error(&hdr, "invalid load size");
13335 		*errp = EINVAL;
13336 		return (NULL);
13337 	}
13338 
13339 	dof = kmem_alloc(loadsz, KM_SLEEP);
13340 
13341 	if (proc_readmem(td, p, uarg, dof, loadsz) != loadsz ||
13342 	    dof->dofh_loadsz != loadsz) {
13343 		kmem_free(dof, hdr.dofh_loadsz);
13344 		*errp = EFAULT;
13345 		return (NULL);
13346 	}
13347 
13348 	return (dof);
13349 }
13350 
13351 static __inline uchar_t
13352 dtrace_dof_char(char c)
13353 {
13354 
13355 	switch (c) {
13356 	case '0':
13357 	case '1':
13358 	case '2':
13359 	case '3':
13360 	case '4':
13361 	case '5':
13362 	case '6':
13363 	case '7':
13364 	case '8':
13365 	case '9':
13366 		return (c - '0');
13367 	case 'A':
13368 	case 'B':
13369 	case 'C':
13370 	case 'D':
13371 	case 'E':
13372 	case 'F':
13373 		return (c - 'A' + 10);
13374 	case 'a':
13375 	case 'b':
13376 	case 'c':
13377 	case 'd':
13378 	case 'e':
13379 	case 'f':
13380 		return (c - 'a' + 10);
13381 	}
13382 	/* Should not reach here. */
13383 	return (UCHAR_MAX);
13384 }
13385 #endif /* __FreeBSD__ */
13386 
13387 static dof_hdr_t *
13388 dtrace_dof_property(const char *name)
13389 {
13390 #ifdef __FreeBSD__
13391 	uint8_t *dofbuf;
13392 	u_char *data, *eol;
13393 	caddr_t doffile;
13394 	size_t bytes, len, i;
13395 	dof_hdr_t *dof;
13396 	u_char c1, c2;
13397 
13398 	dof = NULL;
13399 
13400 	doffile = preload_search_by_type("dtrace_dof");
13401 	if (doffile == NULL)
13402 		return (NULL);
13403 
13404 	data = preload_fetch_addr(doffile);
13405 	len = preload_fetch_size(doffile);
13406 	for (;;) {
13407 		/* Look for the end of the line. All lines end in a newline. */
13408 		eol = memchr(data, '\n', len);
13409 		if (eol == NULL)
13410 			return (NULL);
13411 
13412 		if (strncmp(name, data, strlen(name)) == 0)
13413 			break;
13414 
13415 		eol++; /* skip past the newline */
13416 		len -= eol - data;
13417 		data = eol;
13418 	}
13419 
13420 	/* We've found the data corresponding to the specified key. */
13421 
13422 	data += strlen(name) + 1; /* skip past the '=' */
13423 	len = eol - data;
13424 	if (len % 2 != 0) {
13425 		dtrace_dof_error(NULL, "invalid DOF encoding length");
13426 		goto doferr;
13427 	}
13428 	bytes = len / 2;
13429 	if (bytes < sizeof(dof_hdr_t)) {
13430 		dtrace_dof_error(NULL, "truncated header");
13431 		goto doferr;
13432 	}
13433 
13434 	/*
13435 	 * Each byte is represented by the two ASCII characters in its hex
13436 	 * representation.
13437 	 */
13438 	dofbuf = malloc(bytes, M_SOLARIS, M_WAITOK);
13439 	for (i = 0; i < bytes; i++) {
13440 		c1 = dtrace_dof_char(data[i * 2]);
13441 		c2 = dtrace_dof_char(data[i * 2 + 1]);
13442 		if (c1 == UCHAR_MAX || c2 == UCHAR_MAX) {
13443 			dtrace_dof_error(NULL, "invalid hex char in DOF");
13444 			goto doferr;
13445 		}
13446 		dofbuf[i] = c1 * 16 + c2;
13447 	}
13448 
13449 	dof = (dof_hdr_t *)dofbuf;
13450 	if (bytes < dof->dofh_loadsz) {
13451 		dtrace_dof_error(NULL, "truncated DOF");
13452 		goto doferr;
13453 	}
13454 
13455 	if (dof->dofh_loadsz >= dtrace_dof_maxsize) {
13456 		dtrace_dof_error(NULL, "oversized DOF");
13457 		goto doferr;
13458 	}
13459 
13460 	return (dof);
13461 
13462 doferr:
13463 	free(dof, M_SOLARIS);
13464 	return (NULL);
13465 #else /* __FreeBSD__ */
13466 	uchar_t *buf;
13467 	uint64_t loadsz;
13468 	unsigned int len, i;
13469 	dof_hdr_t *dof;
13470 
13471 	/*
13472 	 * Unfortunately, array of values in .conf files are always (and
13473 	 * only) interpreted to be integer arrays.  We must read our DOF
13474 	 * as an integer array, and then squeeze it into a byte array.
13475 	 */
13476 	if (ddi_prop_lookup_int_array(DDI_DEV_T_ANY, dtrace_devi, 0,
13477 	    (char *)name, (int **)&buf, &len) != DDI_PROP_SUCCESS)
13478 		return (NULL);
13479 
13480 	for (i = 0; i < len; i++)
13481 		buf[i] = (uchar_t)(((int *)buf)[i]);
13482 
13483 	if (len < sizeof (dof_hdr_t)) {
13484 		ddi_prop_free(buf);
13485 		dtrace_dof_error(NULL, "truncated header");
13486 		return (NULL);
13487 	}
13488 
13489 	if (len < (loadsz = ((dof_hdr_t *)buf)->dofh_loadsz)) {
13490 		ddi_prop_free(buf);
13491 		dtrace_dof_error(NULL, "truncated DOF");
13492 		return (NULL);
13493 	}
13494 
13495 	if (loadsz >= dtrace_dof_maxsize) {
13496 		ddi_prop_free(buf);
13497 		dtrace_dof_error(NULL, "oversized DOF");
13498 		return (NULL);
13499 	}
13500 
13501 	dof = kmem_alloc(loadsz, KM_SLEEP);
13502 	bcopy(buf, dof, loadsz);
13503 	ddi_prop_free(buf);
13504 
13505 	return (dof);
13506 #endif /* !__FreeBSD__ */
13507 }
13508 
13509 static void
13510 dtrace_dof_destroy(dof_hdr_t *dof)
13511 {
13512 	kmem_free(dof, dof->dofh_loadsz);
13513 }
13514 
13515 /*
13516  * Return the dof_sec_t pointer corresponding to a given section index.  If the
13517  * index is not valid, dtrace_dof_error() is called and NULL is returned.  If
13518  * a type other than DOF_SECT_NONE is specified, the header is checked against
13519  * this type and NULL is returned if the types do not match.
13520  */
13521 static dof_sec_t *
13522 dtrace_dof_sect(dof_hdr_t *dof, uint32_t type, dof_secidx_t i)
13523 {
13524 	dof_sec_t *sec = (dof_sec_t *)(uintptr_t)
13525 	    ((uintptr_t)dof + dof->dofh_secoff + i * dof->dofh_secsize);
13526 
13527 	if (i >= dof->dofh_secnum) {
13528 		dtrace_dof_error(dof, "referenced section index is invalid");
13529 		return (NULL);
13530 	}
13531 
13532 	if (!(sec->dofs_flags & DOF_SECF_LOAD)) {
13533 		dtrace_dof_error(dof, "referenced section is not loadable");
13534 		return (NULL);
13535 	}
13536 
13537 	if (type != DOF_SECT_NONE && type != sec->dofs_type) {
13538 		dtrace_dof_error(dof, "referenced section is the wrong type");
13539 		return (NULL);
13540 	}
13541 
13542 	return (sec);
13543 }
13544 
13545 static dtrace_probedesc_t *
13546 dtrace_dof_probedesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_probedesc_t *desc)
13547 {
13548 	dof_probedesc_t *probe;
13549 	dof_sec_t *strtab;
13550 	uintptr_t daddr = (uintptr_t)dof;
13551 	uintptr_t str;
13552 	size_t size;
13553 
13554 	if (sec->dofs_type != DOF_SECT_PROBEDESC) {
13555 		dtrace_dof_error(dof, "invalid probe section");
13556 		return (NULL);
13557 	}
13558 
13559 	if (sec->dofs_align != sizeof (dof_secidx_t)) {
13560 		dtrace_dof_error(dof, "bad alignment in probe description");
13561 		return (NULL);
13562 	}
13563 
13564 	if (sec->dofs_offset + sizeof (dof_probedesc_t) > dof->dofh_loadsz) {
13565 		dtrace_dof_error(dof, "truncated probe description");
13566 		return (NULL);
13567 	}
13568 
13569 	probe = (dof_probedesc_t *)(uintptr_t)(daddr + sec->dofs_offset);
13570 	strtab = dtrace_dof_sect(dof, DOF_SECT_STRTAB, probe->dofp_strtab);
13571 
13572 	if (strtab == NULL)
13573 		return (NULL);
13574 
13575 	str = daddr + strtab->dofs_offset;
13576 	size = strtab->dofs_size;
13577 
13578 	if (probe->dofp_provider >= strtab->dofs_size) {
13579 		dtrace_dof_error(dof, "corrupt probe provider");
13580 		return (NULL);
13581 	}
13582 
13583 	(void) strncpy(desc->dtpd_provider,
13584 	    (char *)(str + probe->dofp_provider),
13585 	    MIN(DTRACE_PROVNAMELEN - 1, size - probe->dofp_provider));
13586 
13587 	if (probe->dofp_mod >= strtab->dofs_size) {
13588 		dtrace_dof_error(dof, "corrupt probe module");
13589 		return (NULL);
13590 	}
13591 
13592 	(void) strncpy(desc->dtpd_mod, (char *)(str + probe->dofp_mod),
13593 	    MIN(DTRACE_MODNAMELEN - 1, size - probe->dofp_mod));
13594 
13595 	if (probe->dofp_func >= strtab->dofs_size) {
13596 		dtrace_dof_error(dof, "corrupt probe function");
13597 		return (NULL);
13598 	}
13599 
13600 	(void) strncpy(desc->dtpd_func, (char *)(str + probe->dofp_func),
13601 	    MIN(DTRACE_FUNCNAMELEN - 1, size - probe->dofp_func));
13602 
13603 	if (probe->dofp_name >= strtab->dofs_size) {
13604 		dtrace_dof_error(dof, "corrupt probe name");
13605 		return (NULL);
13606 	}
13607 
13608 	(void) strncpy(desc->dtpd_name, (char *)(str + probe->dofp_name),
13609 	    MIN(DTRACE_NAMELEN - 1, size - probe->dofp_name));
13610 
13611 	return (desc);
13612 }
13613 
13614 static dtrace_difo_t *
13615 dtrace_dof_difo(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
13616     cred_t *cr)
13617 {
13618 	dtrace_difo_t *dp;
13619 	size_t ttl = 0;
13620 	dof_difohdr_t *dofd;
13621 	uintptr_t daddr = (uintptr_t)dof;
13622 	size_t max = dtrace_difo_maxsize;
13623 	int i, l, n;
13624 
13625 	static const struct {
13626 		int section;
13627 		int bufoffs;
13628 		int lenoffs;
13629 		int entsize;
13630 		int align;
13631 		const char *msg;
13632 	} difo[] = {
13633 		{ DOF_SECT_DIF, offsetof(dtrace_difo_t, dtdo_buf),
13634 		offsetof(dtrace_difo_t, dtdo_len), sizeof (dif_instr_t),
13635 		sizeof (dif_instr_t), "multiple DIF sections" },
13636 
13637 		{ DOF_SECT_INTTAB, offsetof(dtrace_difo_t, dtdo_inttab),
13638 		offsetof(dtrace_difo_t, dtdo_intlen), sizeof (uint64_t),
13639 		sizeof (uint64_t), "multiple integer tables" },
13640 
13641 		{ DOF_SECT_STRTAB, offsetof(dtrace_difo_t, dtdo_strtab),
13642 		offsetof(dtrace_difo_t, dtdo_strlen), 0,
13643 		sizeof (char), "multiple string tables" },
13644 
13645 		{ DOF_SECT_VARTAB, offsetof(dtrace_difo_t, dtdo_vartab),
13646 		offsetof(dtrace_difo_t, dtdo_varlen), sizeof (dtrace_difv_t),
13647 		sizeof (uint_t), "multiple variable tables" },
13648 
13649 		{ DOF_SECT_NONE, 0, 0, 0, 0, NULL }
13650 	};
13651 
13652 	if (sec->dofs_type != DOF_SECT_DIFOHDR) {
13653 		dtrace_dof_error(dof, "invalid DIFO header section");
13654 		return (NULL);
13655 	}
13656 
13657 	if (sec->dofs_align != sizeof (dof_secidx_t)) {
13658 		dtrace_dof_error(dof, "bad alignment in DIFO header");
13659 		return (NULL);
13660 	}
13661 
13662 	if (sec->dofs_size < sizeof (dof_difohdr_t) ||
13663 	    sec->dofs_size % sizeof (dof_secidx_t)) {
13664 		dtrace_dof_error(dof, "bad size in DIFO header");
13665 		return (NULL);
13666 	}
13667 
13668 	dofd = (dof_difohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
13669 	n = (sec->dofs_size - sizeof (*dofd)) / sizeof (dof_secidx_t) + 1;
13670 
13671 	dp = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
13672 	dp->dtdo_rtype = dofd->dofd_rtype;
13673 
13674 	for (l = 0; l < n; l++) {
13675 		dof_sec_t *subsec;
13676 		void **bufp;
13677 		uint32_t *lenp;
13678 
13679 		if ((subsec = dtrace_dof_sect(dof, DOF_SECT_NONE,
13680 		    dofd->dofd_links[l])) == NULL)
13681 			goto err; /* invalid section link */
13682 
13683 		if (ttl + subsec->dofs_size > max) {
13684 			dtrace_dof_error(dof, "exceeds maximum size");
13685 			goto err;
13686 		}
13687 
13688 		ttl += subsec->dofs_size;
13689 
13690 		for (i = 0; difo[i].section != DOF_SECT_NONE; i++) {
13691 			if (subsec->dofs_type != difo[i].section)
13692 				continue;
13693 
13694 			if (!(subsec->dofs_flags & DOF_SECF_LOAD)) {
13695 				dtrace_dof_error(dof, "section not loaded");
13696 				goto err;
13697 			}
13698 
13699 			if (subsec->dofs_align != difo[i].align) {
13700 				dtrace_dof_error(dof, "bad alignment");
13701 				goto err;
13702 			}
13703 
13704 			bufp = (void **)((uintptr_t)dp + difo[i].bufoffs);
13705 			lenp = (uint32_t *)((uintptr_t)dp + difo[i].lenoffs);
13706 
13707 			if (*bufp != NULL) {
13708 				dtrace_dof_error(dof, difo[i].msg);
13709 				goto err;
13710 			}
13711 
13712 			if (difo[i].entsize != subsec->dofs_entsize) {
13713 				dtrace_dof_error(dof, "entry size mismatch");
13714 				goto err;
13715 			}
13716 
13717 			if (subsec->dofs_entsize != 0 &&
13718 			    (subsec->dofs_size % subsec->dofs_entsize) != 0) {
13719 				dtrace_dof_error(dof, "corrupt entry size");
13720 				goto err;
13721 			}
13722 
13723 			*lenp = subsec->dofs_size;
13724 			*bufp = kmem_alloc(subsec->dofs_size, KM_SLEEP);
13725 			bcopy((char *)(uintptr_t)(daddr + subsec->dofs_offset),
13726 			    *bufp, subsec->dofs_size);
13727 
13728 			if (subsec->dofs_entsize != 0)
13729 				*lenp /= subsec->dofs_entsize;
13730 
13731 			break;
13732 		}
13733 
13734 		/*
13735 		 * If we encounter a loadable DIFO sub-section that is not
13736 		 * known to us, assume this is a broken program and fail.
13737 		 */
13738 		if (difo[i].section == DOF_SECT_NONE &&
13739 		    (subsec->dofs_flags & DOF_SECF_LOAD)) {
13740 			dtrace_dof_error(dof, "unrecognized DIFO subsection");
13741 			goto err;
13742 		}
13743 	}
13744 
13745 	if (dp->dtdo_buf == NULL) {
13746 		/*
13747 		 * We can't have a DIF object without DIF text.
13748 		 */
13749 		dtrace_dof_error(dof, "missing DIF text");
13750 		goto err;
13751 	}
13752 
13753 	/*
13754 	 * Before we validate the DIF object, run through the variable table
13755 	 * looking for the strings -- if any of their size are under, we'll set
13756 	 * their size to be the system-wide default string size.  Note that
13757 	 * this should _not_ happen if the "strsize" option has been set --
13758 	 * in this case, the compiler should have set the size to reflect the
13759 	 * setting of the option.
13760 	 */
13761 	for (i = 0; i < dp->dtdo_varlen; i++) {
13762 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
13763 		dtrace_diftype_t *t = &v->dtdv_type;
13764 
13765 		if (v->dtdv_id < DIF_VAR_OTHER_UBASE)
13766 			continue;
13767 
13768 		if (t->dtdt_kind == DIF_TYPE_STRING && t->dtdt_size == 0)
13769 			t->dtdt_size = dtrace_strsize_default;
13770 	}
13771 
13772 	if (dtrace_difo_validate(dp, vstate, DIF_DIR_NREGS, cr) != 0)
13773 		goto err;
13774 
13775 	dtrace_difo_init(dp, vstate);
13776 	return (dp);
13777 
13778 err:
13779 	kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
13780 	kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
13781 	kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
13782 	kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
13783 
13784 	kmem_free(dp, sizeof (dtrace_difo_t));
13785 	return (NULL);
13786 }
13787 
13788 static dtrace_predicate_t *
13789 dtrace_dof_predicate(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
13790     cred_t *cr)
13791 {
13792 	dtrace_difo_t *dp;
13793 
13794 	if ((dp = dtrace_dof_difo(dof, sec, vstate, cr)) == NULL)
13795 		return (NULL);
13796 
13797 	return (dtrace_predicate_create(dp));
13798 }
13799 
13800 static dtrace_actdesc_t *
13801 dtrace_dof_actdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
13802     cred_t *cr)
13803 {
13804 	dtrace_actdesc_t *act, *first = NULL, *last = NULL, *next;
13805 	dof_actdesc_t *desc;
13806 	dof_sec_t *difosec;
13807 	size_t offs;
13808 	uintptr_t daddr = (uintptr_t)dof;
13809 	uint64_t arg;
13810 	dtrace_actkind_t kind;
13811 
13812 	if (sec->dofs_type != DOF_SECT_ACTDESC) {
13813 		dtrace_dof_error(dof, "invalid action section");
13814 		return (NULL);
13815 	}
13816 
13817 	if (sec->dofs_offset + sizeof (dof_actdesc_t) > dof->dofh_loadsz) {
13818 		dtrace_dof_error(dof, "truncated action description");
13819 		return (NULL);
13820 	}
13821 
13822 	if (sec->dofs_align != sizeof (uint64_t)) {
13823 		dtrace_dof_error(dof, "bad alignment in action description");
13824 		return (NULL);
13825 	}
13826 
13827 	if (sec->dofs_size < sec->dofs_entsize) {
13828 		dtrace_dof_error(dof, "section entry size exceeds total size");
13829 		return (NULL);
13830 	}
13831 
13832 	if (sec->dofs_entsize != sizeof (dof_actdesc_t)) {
13833 		dtrace_dof_error(dof, "bad entry size in action description");
13834 		return (NULL);
13835 	}
13836 
13837 	if (sec->dofs_size / sec->dofs_entsize > dtrace_actions_max) {
13838 		dtrace_dof_error(dof, "actions exceed dtrace_actions_max");
13839 		return (NULL);
13840 	}
13841 
13842 	for (offs = 0; offs < sec->dofs_size; offs += sec->dofs_entsize) {
13843 		desc = (dof_actdesc_t *)(daddr +
13844 		    (uintptr_t)sec->dofs_offset + offs);
13845 		kind = (dtrace_actkind_t)desc->dofa_kind;
13846 
13847 		if ((DTRACEACT_ISPRINTFLIKE(kind) &&
13848 		    (kind != DTRACEACT_PRINTA ||
13849 		    desc->dofa_strtab != DOF_SECIDX_NONE)) ||
13850 		    (kind == DTRACEACT_DIFEXPR &&
13851 		    desc->dofa_strtab != DOF_SECIDX_NONE)) {
13852 			dof_sec_t *strtab;
13853 			char *str, *fmt;
13854 			uint64_t i;
13855 
13856 			/*
13857 			 * The argument to these actions is an index into the
13858 			 * DOF string table.  For printf()-like actions, this
13859 			 * is the format string.  For print(), this is the
13860 			 * CTF type of the expression result.
13861 			 */
13862 			if ((strtab = dtrace_dof_sect(dof,
13863 			    DOF_SECT_STRTAB, desc->dofa_strtab)) == NULL)
13864 				goto err;
13865 
13866 			str = (char *)((uintptr_t)dof +
13867 			    (uintptr_t)strtab->dofs_offset);
13868 
13869 			for (i = desc->dofa_arg; i < strtab->dofs_size; i++) {
13870 				if (str[i] == '\0')
13871 					break;
13872 			}
13873 
13874 			if (i >= strtab->dofs_size) {
13875 				dtrace_dof_error(dof, "bogus format string");
13876 				goto err;
13877 			}
13878 
13879 			if (i == desc->dofa_arg) {
13880 				dtrace_dof_error(dof, "empty format string");
13881 				goto err;
13882 			}
13883 
13884 			i -= desc->dofa_arg;
13885 			fmt = kmem_alloc(i + 1, KM_SLEEP);
13886 			bcopy(&str[desc->dofa_arg], fmt, i + 1);
13887 			arg = (uint64_t)(uintptr_t)fmt;
13888 		} else {
13889 			if (kind == DTRACEACT_PRINTA) {
13890 				ASSERT(desc->dofa_strtab == DOF_SECIDX_NONE);
13891 				arg = 0;
13892 			} else {
13893 				arg = desc->dofa_arg;
13894 			}
13895 		}
13896 
13897 		act = dtrace_actdesc_create(kind, desc->dofa_ntuple,
13898 		    desc->dofa_uarg, arg);
13899 
13900 		if (last != NULL) {
13901 			last->dtad_next = act;
13902 		} else {
13903 			first = act;
13904 		}
13905 
13906 		last = act;
13907 
13908 		if (desc->dofa_difo == DOF_SECIDX_NONE)
13909 			continue;
13910 
13911 		if ((difosec = dtrace_dof_sect(dof,
13912 		    DOF_SECT_DIFOHDR, desc->dofa_difo)) == NULL)
13913 			goto err;
13914 
13915 		act->dtad_difo = dtrace_dof_difo(dof, difosec, vstate, cr);
13916 
13917 		if (act->dtad_difo == NULL)
13918 			goto err;
13919 	}
13920 
13921 	ASSERT(first != NULL);
13922 	return (first);
13923 
13924 err:
13925 	for (act = first; act != NULL; act = next) {
13926 		next = act->dtad_next;
13927 		dtrace_actdesc_release(act, vstate);
13928 	}
13929 
13930 	return (NULL);
13931 }
13932 
13933 static dtrace_ecbdesc_t *
13934 dtrace_dof_ecbdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
13935     cred_t *cr)
13936 {
13937 	dtrace_ecbdesc_t *ep;
13938 	dof_ecbdesc_t *ecb;
13939 	dtrace_probedesc_t *desc;
13940 	dtrace_predicate_t *pred = NULL;
13941 
13942 	if (sec->dofs_size < sizeof (dof_ecbdesc_t)) {
13943 		dtrace_dof_error(dof, "truncated ECB description");
13944 		return (NULL);
13945 	}
13946 
13947 	if (sec->dofs_align != sizeof (uint64_t)) {
13948 		dtrace_dof_error(dof, "bad alignment in ECB description");
13949 		return (NULL);
13950 	}
13951 
13952 	ecb = (dof_ecbdesc_t *)((uintptr_t)dof + (uintptr_t)sec->dofs_offset);
13953 	sec = dtrace_dof_sect(dof, DOF_SECT_PROBEDESC, ecb->dofe_probes);
13954 
13955 	if (sec == NULL)
13956 		return (NULL);
13957 
13958 	ep = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
13959 	ep->dted_uarg = ecb->dofe_uarg;
13960 	desc = &ep->dted_probe;
13961 
13962 	if (dtrace_dof_probedesc(dof, sec, desc) == NULL)
13963 		goto err;
13964 
13965 	if (ecb->dofe_pred != DOF_SECIDX_NONE) {
13966 		if ((sec = dtrace_dof_sect(dof,
13967 		    DOF_SECT_DIFOHDR, ecb->dofe_pred)) == NULL)
13968 			goto err;
13969 
13970 		if ((pred = dtrace_dof_predicate(dof, sec, vstate, cr)) == NULL)
13971 			goto err;
13972 
13973 		ep->dted_pred.dtpdd_predicate = pred;
13974 	}
13975 
13976 	if (ecb->dofe_actions != DOF_SECIDX_NONE) {
13977 		if ((sec = dtrace_dof_sect(dof,
13978 		    DOF_SECT_ACTDESC, ecb->dofe_actions)) == NULL)
13979 			goto err;
13980 
13981 		ep->dted_action = dtrace_dof_actdesc(dof, sec, vstate, cr);
13982 
13983 		if (ep->dted_action == NULL)
13984 			goto err;
13985 	}
13986 
13987 	return (ep);
13988 
13989 err:
13990 	if (pred != NULL)
13991 		dtrace_predicate_release(pred, vstate);
13992 	kmem_free(ep, sizeof (dtrace_ecbdesc_t));
13993 	return (NULL);
13994 }
13995 
13996 /*
13997  * Apply the relocations from the specified 'sec' (a DOF_SECT_URELHDR) to the
13998  * specified DOF.  SETX relocations are computed using 'ubase', the base load
13999  * address of the object containing the DOF, and DOFREL relocations are relative
14000  * to the relocation offset within the DOF.
14001  */
14002 static int
14003 dtrace_dof_relocate(dof_hdr_t *dof, dof_sec_t *sec, uint64_t ubase,
14004     uint64_t udaddr)
14005 {
14006 	uintptr_t daddr = (uintptr_t)dof;
14007 	uintptr_t ts_end;
14008 	dof_relohdr_t *dofr =
14009 	    (dof_relohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
14010 	dof_sec_t *ss, *rs, *ts;
14011 	dof_relodesc_t *r;
14012 	uint_t i, n;
14013 
14014 	if (sec->dofs_size < sizeof (dof_relohdr_t) ||
14015 	    sec->dofs_align != sizeof (dof_secidx_t)) {
14016 		dtrace_dof_error(dof, "invalid relocation header");
14017 		return (-1);
14018 	}
14019 
14020 	ss = dtrace_dof_sect(dof, DOF_SECT_STRTAB, dofr->dofr_strtab);
14021 	rs = dtrace_dof_sect(dof, DOF_SECT_RELTAB, dofr->dofr_relsec);
14022 	ts = dtrace_dof_sect(dof, DOF_SECT_NONE, dofr->dofr_tgtsec);
14023 	ts_end = (uintptr_t)ts + sizeof (dof_sec_t);
14024 
14025 	if (ss == NULL || rs == NULL || ts == NULL)
14026 		return (-1); /* dtrace_dof_error() has been called already */
14027 
14028 	if (rs->dofs_entsize < sizeof (dof_relodesc_t) ||
14029 	    rs->dofs_align != sizeof (uint64_t)) {
14030 		dtrace_dof_error(dof, "invalid relocation section");
14031 		return (-1);
14032 	}
14033 
14034 	r = (dof_relodesc_t *)(uintptr_t)(daddr + rs->dofs_offset);
14035 	n = rs->dofs_size / rs->dofs_entsize;
14036 
14037 	for (i = 0; i < n; i++) {
14038 		uintptr_t taddr = daddr + ts->dofs_offset + r->dofr_offset;
14039 
14040 		switch (r->dofr_type) {
14041 		case DOF_RELO_NONE:
14042 			break;
14043 		case DOF_RELO_SETX:
14044 		case DOF_RELO_DOFREL:
14045 			if (r->dofr_offset >= ts->dofs_size || r->dofr_offset +
14046 			    sizeof (uint64_t) > ts->dofs_size) {
14047 				dtrace_dof_error(dof, "bad relocation offset");
14048 				return (-1);
14049 			}
14050 
14051 			if (taddr >= (uintptr_t)ts && taddr < ts_end) {
14052 				dtrace_dof_error(dof, "bad relocation offset");
14053 				return (-1);
14054 			}
14055 
14056 			if (!IS_P2ALIGNED(taddr, sizeof (uint64_t))) {
14057 				dtrace_dof_error(dof, "misaligned setx relo");
14058 				return (-1);
14059 			}
14060 
14061 			if (r->dofr_type == DOF_RELO_SETX)
14062 				*(uint64_t *)taddr += ubase;
14063 			else
14064 				*(uint64_t *)taddr +=
14065 				    udaddr + ts->dofs_offset + r->dofr_offset;
14066 			break;
14067 		default:
14068 			dtrace_dof_error(dof, "invalid relocation type");
14069 			return (-1);
14070 		}
14071 
14072 		r = (dof_relodesc_t *)((uintptr_t)r + rs->dofs_entsize);
14073 	}
14074 
14075 	return (0);
14076 }
14077 
14078 /*
14079  * The dof_hdr_t passed to dtrace_dof_slurp() should be a partially validated
14080  * header:  it should be at the front of a memory region that is at least
14081  * sizeof (dof_hdr_t) in size -- and then at least dof_hdr.dofh_loadsz in
14082  * size.  It need not be validated in any other way.
14083  */
14084 static int
14085 dtrace_dof_slurp(dof_hdr_t *dof, dtrace_vstate_t *vstate, cred_t *cr,
14086     dtrace_enabling_t **enabp, uint64_t ubase, uint64_t udaddr, int noprobes)
14087 {
14088 	uint64_t len = dof->dofh_loadsz, seclen;
14089 	uintptr_t daddr = (uintptr_t)dof;
14090 	dtrace_ecbdesc_t *ep;
14091 	dtrace_enabling_t *enab;
14092 	uint_t i;
14093 
14094 	ASSERT(MUTEX_HELD(&dtrace_lock));
14095 	ASSERT(dof->dofh_loadsz >= sizeof (dof_hdr_t));
14096 
14097 	/*
14098 	 * Check the DOF header identification bytes.  In addition to checking
14099 	 * valid settings, we also verify that unused bits/bytes are zeroed so
14100 	 * we can use them later without fear of regressing existing binaries.
14101 	 */
14102 	if (bcmp(&dof->dofh_ident[DOF_ID_MAG0],
14103 	    DOF_MAG_STRING, DOF_MAG_STRLEN) != 0) {
14104 		dtrace_dof_error(dof, "DOF magic string mismatch");
14105 		return (-1);
14106 	}
14107 
14108 	if (dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_ILP32 &&
14109 	    dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_LP64) {
14110 		dtrace_dof_error(dof, "DOF has invalid data model");
14111 		return (-1);
14112 	}
14113 
14114 	if (dof->dofh_ident[DOF_ID_ENCODING] != DOF_ENCODE_NATIVE) {
14115 		dtrace_dof_error(dof, "DOF encoding mismatch");
14116 		return (-1);
14117 	}
14118 
14119 	if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
14120 	    dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_2) {
14121 		dtrace_dof_error(dof, "DOF version mismatch");
14122 		return (-1);
14123 	}
14124 
14125 	if (dof->dofh_ident[DOF_ID_DIFVERS] != DIF_VERSION_2) {
14126 		dtrace_dof_error(dof, "DOF uses unsupported instruction set");
14127 		return (-1);
14128 	}
14129 
14130 	if (dof->dofh_ident[DOF_ID_DIFIREG] > DIF_DIR_NREGS) {
14131 		dtrace_dof_error(dof, "DOF uses too many integer registers");
14132 		return (-1);
14133 	}
14134 
14135 	if (dof->dofh_ident[DOF_ID_DIFTREG] > DIF_DTR_NREGS) {
14136 		dtrace_dof_error(dof, "DOF uses too many tuple registers");
14137 		return (-1);
14138 	}
14139 
14140 	for (i = DOF_ID_PAD; i < DOF_ID_SIZE; i++) {
14141 		if (dof->dofh_ident[i] != 0) {
14142 			dtrace_dof_error(dof, "DOF has invalid ident byte set");
14143 			return (-1);
14144 		}
14145 	}
14146 
14147 	if (dof->dofh_flags & ~DOF_FL_VALID) {
14148 		dtrace_dof_error(dof, "DOF has invalid flag bits set");
14149 		return (-1);
14150 	}
14151 
14152 	if (dof->dofh_secsize == 0) {
14153 		dtrace_dof_error(dof, "zero section header size");
14154 		return (-1);
14155 	}
14156 
14157 	/*
14158 	 * Check that the section headers don't exceed the amount of DOF
14159 	 * data.  Note that we cast the section size and number of sections
14160 	 * to uint64_t's to prevent possible overflow in the multiplication.
14161 	 */
14162 	seclen = (uint64_t)dof->dofh_secnum * (uint64_t)dof->dofh_secsize;
14163 
14164 	if (dof->dofh_secoff > len || seclen > len ||
14165 	    dof->dofh_secoff + seclen > len) {
14166 		dtrace_dof_error(dof, "truncated section headers");
14167 		return (-1);
14168 	}
14169 
14170 	if (!IS_P2ALIGNED(dof->dofh_secoff, sizeof (uint64_t))) {
14171 		dtrace_dof_error(dof, "misaligned section headers");
14172 		return (-1);
14173 	}
14174 
14175 	if (!IS_P2ALIGNED(dof->dofh_secsize, sizeof (uint64_t))) {
14176 		dtrace_dof_error(dof, "misaligned section size");
14177 		return (-1);
14178 	}
14179 
14180 	/*
14181 	 * Take an initial pass through the section headers to be sure that
14182 	 * the headers don't have stray offsets.  If the 'noprobes' flag is
14183 	 * set, do not permit sections relating to providers, probes, or args.
14184 	 */
14185 	for (i = 0; i < dof->dofh_secnum; i++) {
14186 		dof_sec_t *sec = (dof_sec_t *)(daddr +
14187 		    (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
14188 
14189 		if (noprobes) {
14190 			switch (sec->dofs_type) {
14191 			case DOF_SECT_PROVIDER:
14192 			case DOF_SECT_PROBES:
14193 			case DOF_SECT_PRARGS:
14194 			case DOF_SECT_PROFFS:
14195 				dtrace_dof_error(dof, "illegal sections "
14196 				    "for enabling");
14197 				return (-1);
14198 			}
14199 		}
14200 
14201 		if (DOF_SEC_ISLOADABLE(sec->dofs_type) &&
14202 		    !(sec->dofs_flags & DOF_SECF_LOAD)) {
14203 			dtrace_dof_error(dof, "loadable section with load "
14204 			    "flag unset");
14205 			return (-1);
14206 		}
14207 
14208 		if (!(sec->dofs_flags & DOF_SECF_LOAD))
14209 			continue; /* just ignore non-loadable sections */
14210 
14211 		if (!ISP2(sec->dofs_align)) {
14212 			dtrace_dof_error(dof, "bad section alignment");
14213 			return (-1);
14214 		}
14215 
14216 		if (sec->dofs_offset & (sec->dofs_align - 1)) {
14217 			dtrace_dof_error(dof, "misaligned section");
14218 			return (-1);
14219 		}
14220 
14221 		if (sec->dofs_offset > len || sec->dofs_size > len ||
14222 		    sec->dofs_offset + sec->dofs_size > len) {
14223 			dtrace_dof_error(dof, "corrupt section header");
14224 			return (-1);
14225 		}
14226 
14227 		if (sec->dofs_type == DOF_SECT_STRTAB && *((char *)daddr +
14228 		    sec->dofs_offset + sec->dofs_size - 1) != '\0') {
14229 			dtrace_dof_error(dof, "non-terminating string table");
14230 			return (-1);
14231 		}
14232 	}
14233 
14234 	/*
14235 	 * Take a second pass through the sections and locate and perform any
14236 	 * relocations that are present.  We do this after the first pass to
14237 	 * be sure that all sections have had their headers validated.
14238 	 */
14239 	for (i = 0; i < dof->dofh_secnum; i++) {
14240 		dof_sec_t *sec = (dof_sec_t *)(daddr +
14241 		    (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
14242 
14243 		if (!(sec->dofs_flags & DOF_SECF_LOAD))
14244 			continue; /* skip sections that are not loadable */
14245 
14246 		switch (sec->dofs_type) {
14247 		case DOF_SECT_URELHDR:
14248 			if (dtrace_dof_relocate(dof, sec, ubase, udaddr) != 0)
14249 				return (-1);
14250 			break;
14251 		}
14252 	}
14253 
14254 	if ((enab = *enabp) == NULL)
14255 		enab = *enabp = dtrace_enabling_create(vstate);
14256 
14257 	for (i = 0; i < dof->dofh_secnum; i++) {
14258 		dof_sec_t *sec = (dof_sec_t *)(daddr +
14259 		    (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
14260 
14261 		if (sec->dofs_type != DOF_SECT_ECBDESC)
14262 			continue;
14263 
14264 		if ((ep = dtrace_dof_ecbdesc(dof, sec, vstate, cr)) == NULL) {
14265 			dtrace_enabling_destroy(enab);
14266 			*enabp = NULL;
14267 			return (-1);
14268 		}
14269 
14270 		dtrace_enabling_add(enab, ep);
14271 	}
14272 
14273 	return (0);
14274 }
14275 
14276 /*
14277  * Process DOF for any options.  This routine assumes that the DOF has been
14278  * at least processed by dtrace_dof_slurp().
14279  */
14280 static int
14281 dtrace_dof_options(dof_hdr_t *dof, dtrace_state_t *state)
14282 {
14283 	int i, rval;
14284 	uint32_t entsize;
14285 	size_t offs;
14286 	dof_optdesc_t *desc;
14287 
14288 	for (i = 0; i < dof->dofh_secnum; i++) {
14289 		dof_sec_t *sec = (dof_sec_t *)((uintptr_t)dof +
14290 		    (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
14291 
14292 		if (sec->dofs_type != DOF_SECT_OPTDESC)
14293 			continue;
14294 
14295 		if (sec->dofs_align != sizeof (uint64_t)) {
14296 			dtrace_dof_error(dof, "bad alignment in "
14297 			    "option description");
14298 			return (EINVAL);
14299 		}
14300 
14301 		if ((entsize = sec->dofs_entsize) == 0) {
14302 			dtrace_dof_error(dof, "zeroed option entry size");
14303 			return (EINVAL);
14304 		}
14305 
14306 		if (entsize < sizeof (dof_optdesc_t)) {
14307 			dtrace_dof_error(dof, "bad option entry size");
14308 			return (EINVAL);
14309 		}
14310 
14311 		for (offs = 0; offs < sec->dofs_size; offs += entsize) {
14312 			desc = (dof_optdesc_t *)((uintptr_t)dof +
14313 			    (uintptr_t)sec->dofs_offset + offs);
14314 
14315 			if (desc->dofo_strtab != DOF_SECIDX_NONE) {
14316 				dtrace_dof_error(dof, "non-zero option string");
14317 				return (EINVAL);
14318 			}
14319 
14320 			if (desc->dofo_value == DTRACEOPT_UNSET) {
14321 				dtrace_dof_error(dof, "unset option");
14322 				return (EINVAL);
14323 			}
14324 
14325 			if ((rval = dtrace_state_option(state,
14326 			    desc->dofo_option, desc->dofo_value)) != 0) {
14327 				dtrace_dof_error(dof, "rejected option");
14328 				return (rval);
14329 			}
14330 		}
14331 	}
14332 
14333 	return (0);
14334 }
14335 
14336 /*
14337  * DTrace Consumer State Functions
14338  */
14339 static int
14340 dtrace_dstate_init(dtrace_dstate_t *dstate, size_t size)
14341 {
14342 	size_t hashsize, maxper, min, chunksize = dstate->dtds_chunksize;
14343 	void *base;
14344 	uintptr_t limit;
14345 	dtrace_dynvar_t *dvar, *next, *start;
14346 	int i;
14347 
14348 	ASSERT(MUTEX_HELD(&dtrace_lock));
14349 	ASSERT(dstate->dtds_base == NULL && dstate->dtds_percpu == NULL);
14350 
14351 	bzero(dstate, sizeof (dtrace_dstate_t));
14352 
14353 	if ((dstate->dtds_chunksize = chunksize) == 0)
14354 		dstate->dtds_chunksize = DTRACE_DYNVAR_CHUNKSIZE;
14355 
14356 	VERIFY(dstate->dtds_chunksize < LONG_MAX);
14357 
14358 	if (size < (min = dstate->dtds_chunksize + sizeof (dtrace_dynhash_t)))
14359 		size = min;
14360 
14361 	if ((base = kmem_zalloc(size, KM_NOSLEEP | KM_NORMALPRI)) == NULL)
14362 		return (ENOMEM);
14363 
14364 	dstate->dtds_size = size;
14365 	dstate->dtds_base = base;
14366 	dstate->dtds_percpu = kmem_cache_alloc(dtrace_state_cache, KM_SLEEP);
14367 	bzero(dstate->dtds_percpu, NCPU * sizeof (dtrace_dstate_percpu_t));
14368 
14369 	hashsize = size / (dstate->dtds_chunksize + sizeof (dtrace_dynhash_t));
14370 
14371 	if (hashsize != 1 && (hashsize & 1))
14372 		hashsize--;
14373 
14374 	dstate->dtds_hashsize = hashsize;
14375 	dstate->dtds_hash = dstate->dtds_base;
14376 
14377 	/*
14378 	 * Set all of our hash buckets to point to the single sink, and (if
14379 	 * it hasn't already been set), set the sink's hash value to be the
14380 	 * sink sentinel value.  The sink is needed for dynamic variable
14381 	 * lookups to know that they have iterated over an entire, valid hash
14382 	 * chain.
14383 	 */
14384 	for (i = 0; i < hashsize; i++)
14385 		dstate->dtds_hash[i].dtdh_chain = &dtrace_dynhash_sink;
14386 
14387 	if (dtrace_dynhash_sink.dtdv_hashval != DTRACE_DYNHASH_SINK)
14388 		dtrace_dynhash_sink.dtdv_hashval = DTRACE_DYNHASH_SINK;
14389 
14390 	/*
14391 	 * Determine number of active CPUs.  Divide free list evenly among
14392 	 * active CPUs.
14393 	 */
14394 	start = (dtrace_dynvar_t *)
14395 	    ((uintptr_t)base + hashsize * sizeof (dtrace_dynhash_t));
14396 	limit = (uintptr_t)base + size;
14397 
14398 	VERIFY((uintptr_t)start < limit);
14399 	VERIFY((uintptr_t)start >= (uintptr_t)base);
14400 
14401 	maxper = (limit - (uintptr_t)start) / NCPU;
14402 	maxper = (maxper / dstate->dtds_chunksize) * dstate->dtds_chunksize;
14403 
14404 #ifndef illumos
14405 	CPU_FOREACH(i) {
14406 #else
14407 	for (i = 0; i < NCPU; i++) {
14408 #endif
14409 		dstate->dtds_percpu[i].dtdsc_free = dvar = start;
14410 
14411 		/*
14412 		 * If we don't even have enough chunks to make it once through
14413 		 * NCPUs, we're just going to allocate everything to the first
14414 		 * CPU.  And if we're on the last CPU, we're going to allocate
14415 		 * whatever is left over.  In either case, we set the limit to
14416 		 * be the limit of the dynamic variable space.
14417 		 */
14418 		if (maxper == 0 || i == NCPU - 1) {
14419 			limit = (uintptr_t)base + size;
14420 			start = NULL;
14421 		} else {
14422 			limit = (uintptr_t)start + maxper;
14423 			start = (dtrace_dynvar_t *)limit;
14424 		}
14425 
14426 		VERIFY(limit <= (uintptr_t)base + size);
14427 
14428 		for (;;) {
14429 			next = (dtrace_dynvar_t *)((uintptr_t)dvar +
14430 			    dstate->dtds_chunksize);
14431 
14432 			if ((uintptr_t)next + dstate->dtds_chunksize >= limit)
14433 				break;
14434 
14435 			VERIFY((uintptr_t)dvar >= (uintptr_t)base &&
14436 			    (uintptr_t)dvar <= (uintptr_t)base + size);
14437 			dvar->dtdv_next = next;
14438 			dvar = next;
14439 		}
14440 
14441 		if (maxper == 0)
14442 			break;
14443 	}
14444 
14445 	return (0);
14446 }
14447 
14448 static void
14449 dtrace_dstate_fini(dtrace_dstate_t *dstate)
14450 {
14451 	ASSERT(MUTEX_HELD(&cpu_lock));
14452 
14453 	if (dstate->dtds_base == NULL)
14454 		return;
14455 
14456 	kmem_free(dstate->dtds_base, dstate->dtds_size);
14457 	kmem_cache_free(dtrace_state_cache, dstate->dtds_percpu);
14458 }
14459 
14460 static void
14461 dtrace_vstate_fini(dtrace_vstate_t *vstate)
14462 {
14463 	/*
14464 	 * Logical XOR, where are you?
14465 	 */
14466 	ASSERT((vstate->dtvs_nglobals == 0) ^ (vstate->dtvs_globals != NULL));
14467 
14468 	if (vstate->dtvs_nglobals > 0) {
14469 		kmem_free(vstate->dtvs_globals, vstate->dtvs_nglobals *
14470 		    sizeof (dtrace_statvar_t *));
14471 	}
14472 
14473 	if (vstate->dtvs_ntlocals > 0) {
14474 		kmem_free(vstate->dtvs_tlocals, vstate->dtvs_ntlocals *
14475 		    sizeof (dtrace_difv_t));
14476 	}
14477 
14478 	ASSERT((vstate->dtvs_nlocals == 0) ^ (vstate->dtvs_locals != NULL));
14479 
14480 	if (vstate->dtvs_nlocals > 0) {
14481 		kmem_free(vstate->dtvs_locals, vstate->dtvs_nlocals *
14482 		    sizeof (dtrace_statvar_t *));
14483 	}
14484 }
14485 
14486 #ifdef illumos
14487 static void
14488 dtrace_state_clean(dtrace_state_t *state)
14489 {
14490 	if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE)
14491 		return;
14492 
14493 	dtrace_dynvar_clean(&state->dts_vstate.dtvs_dynvars);
14494 	dtrace_speculation_clean(state);
14495 }
14496 
14497 static void
14498 dtrace_state_deadman(dtrace_state_t *state)
14499 {
14500 	hrtime_t now;
14501 
14502 	dtrace_sync();
14503 
14504 	now = dtrace_gethrtime();
14505 
14506 	if (state != dtrace_anon.dta_state &&
14507 	    now - state->dts_laststatus >= dtrace_deadman_user)
14508 		return;
14509 
14510 	/*
14511 	 * We must be sure that dts_alive never appears to be less than the
14512 	 * value upon entry to dtrace_state_deadman(), and because we lack a
14513 	 * dtrace_cas64(), we cannot store to it atomically.  We thus instead
14514 	 * store INT64_MAX to it, followed by a memory barrier, followed by
14515 	 * the new value.  This assures that dts_alive never appears to be
14516 	 * less than its true value, regardless of the order in which the
14517 	 * stores to the underlying storage are issued.
14518 	 */
14519 	state->dts_alive = INT64_MAX;
14520 	dtrace_membar_producer();
14521 	state->dts_alive = now;
14522 }
14523 #else	/* !illumos */
14524 static void
14525 dtrace_state_clean(void *arg)
14526 {
14527 	dtrace_state_t *state = arg;
14528 	dtrace_optval_t *opt = state->dts_options;
14529 
14530 	if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE)
14531 		return;
14532 
14533 	dtrace_dynvar_clean(&state->dts_vstate.dtvs_dynvars);
14534 	dtrace_speculation_clean(state);
14535 
14536 	callout_reset(&state->dts_cleaner, hz * opt[DTRACEOPT_CLEANRATE] / NANOSEC,
14537 	    dtrace_state_clean, state);
14538 }
14539 
14540 static void
14541 dtrace_state_deadman(void *arg)
14542 {
14543 	dtrace_state_t *state = arg;
14544 	hrtime_t now;
14545 
14546 	dtrace_sync();
14547 
14548 	dtrace_debug_output();
14549 
14550 	now = dtrace_gethrtime();
14551 
14552 	if (state != dtrace_anon.dta_state &&
14553 	    now - state->dts_laststatus >= dtrace_deadman_user)
14554 		return;
14555 
14556 	/*
14557 	 * We must be sure that dts_alive never appears to be less than the
14558 	 * value upon entry to dtrace_state_deadman(), and because we lack a
14559 	 * dtrace_cas64(), we cannot store to it atomically.  We thus instead
14560 	 * store INT64_MAX to it, followed by a memory barrier, followed by
14561 	 * the new value.  This assures that dts_alive never appears to be
14562 	 * less than its true value, regardless of the order in which the
14563 	 * stores to the underlying storage are issued.
14564 	 */
14565 	state->dts_alive = INT64_MAX;
14566 	dtrace_membar_producer();
14567 	state->dts_alive = now;
14568 
14569 	callout_reset(&state->dts_deadman, hz * dtrace_deadman_interval / NANOSEC,
14570 	    dtrace_state_deadman, state);
14571 }
14572 #endif	/* illumos */
14573 
14574 static dtrace_state_t *
14575 #ifdef illumos
14576 dtrace_state_create(dev_t *devp, cred_t *cr)
14577 #else
14578 dtrace_state_create(struct cdev *dev, struct ucred *cred __unused)
14579 #endif
14580 {
14581 #ifdef illumos
14582 	minor_t minor;
14583 	major_t major;
14584 #else
14585 	cred_t *cr = NULL;
14586 	int m = 0;
14587 #endif
14588 	char c[30];
14589 	dtrace_state_t *state;
14590 	dtrace_optval_t *opt;
14591 	int bufsize = NCPU * sizeof (dtrace_buffer_t), i;
14592 	int cpu_it;
14593 
14594 	ASSERT(MUTEX_HELD(&dtrace_lock));
14595 	ASSERT(MUTEX_HELD(&cpu_lock));
14596 
14597 #ifdef illumos
14598 	minor = (minor_t)(uintptr_t)vmem_alloc(dtrace_minor, 1,
14599 	    VM_BESTFIT | VM_SLEEP);
14600 
14601 	if (ddi_soft_state_zalloc(dtrace_softstate, minor) != DDI_SUCCESS) {
14602 		vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
14603 		return (NULL);
14604 	}
14605 
14606 	state = ddi_get_soft_state(dtrace_softstate, minor);
14607 #else
14608 	if (dev != NULL) {
14609 		cr = dev->si_cred;
14610 		m = dev2unit(dev);
14611 	}
14612 
14613 	/* Allocate memory for the state. */
14614 	state = kmem_zalloc(sizeof(dtrace_state_t), KM_SLEEP);
14615 #endif
14616 
14617 	state->dts_epid = DTRACE_EPIDNONE + 1;
14618 
14619 	(void) snprintf(c, sizeof (c), "dtrace_aggid_%d", m);
14620 #ifdef illumos
14621 	state->dts_aggid_arena = vmem_create(c, (void *)1, UINT32_MAX, 1,
14622 	    NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
14623 
14624 	if (devp != NULL) {
14625 		major = getemajor(*devp);
14626 	} else {
14627 		major = ddi_driver_major(dtrace_devi);
14628 	}
14629 
14630 	state->dts_dev = makedevice(major, minor);
14631 
14632 	if (devp != NULL)
14633 		*devp = state->dts_dev;
14634 #else
14635 	state->dts_aggid_arena = new_unrhdr(1, INT_MAX, &dtrace_unr_mtx);
14636 	state->dts_dev = dev;
14637 #endif
14638 
14639 	/*
14640 	 * We allocate NCPU buffers.  On the one hand, this can be quite
14641 	 * a bit of memory per instance (nearly 36K on a Starcat).  On the
14642 	 * other hand, it saves an additional memory reference in the probe
14643 	 * path.
14644 	 */
14645 	state->dts_buffer = kmem_zalloc(bufsize, KM_SLEEP);
14646 	state->dts_aggbuffer = kmem_zalloc(bufsize, KM_SLEEP);
14647 
14648 	/*
14649          * Allocate and initialise the per-process per-CPU random state.
14650 	 * SI_SUB_RANDOM < SI_SUB_DTRACE_ANON therefore entropy device is
14651          * assumed to be seeded at this point (if from Fortuna seed file).
14652 	 */
14653 	arc4random_buf(&state->dts_rstate[0], 2 * sizeof(uint64_t));
14654 	for (cpu_it = 1; cpu_it < NCPU; cpu_it++) {
14655 		/*
14656 		 * Each CPU is assigned a 2^64 period, non-overlapping
14657 		 * subsequence.
14658 		 */
14659 		dtrace_xoroshiro128_plus_jump(state->dts_rstate[cpu_it-1],
14660 		    state->dts_rstate[cpu_it]);
14661 	}
14662 
14663 #ifdef illumos
14664 	state->dts_cleaner = CYCLIC_NONE;
14665 	state->dts_deadman = CYCLIC_NONE;
14666 #else
14667 	callout_init(&state->dts_cleaner, 1);
14668 	callout_init(&state->dts_deadman, 1);
14669 #endif
14670 	state->dts_vstate.dtvs_state = state;
14671 
14672 	for (i = 0; i < DTRACEOPT_MAX; i++)
14673 		state->dts_options[i] = DTRACEOPT_UNSET;
14674 
14675 	/*
14676 	 * Set the default options.
14677 	 */
14678 	opt = state->dts_options;
14679 	opt[DTRACEOPT_BUFPOLICY] = DTRACEOPT_BUFPOLICY_SWITCH;
14680 	opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_AUTO;
14681 	opt[DTRACEOPT_NSPEC] = dtrace_nspec_default;
14682 	opt[DTRACEOPT_SPECSIZE] = dtrace_specsize_default;
14683 	opt[DTRACEOPT_CPU] = (dtrace_optval_t)DTRACE_CPUALL;
14684 	opt[DTRACEOPT_STRSIZE] = dtrace_strsize_default;
14685 	opt[DTRACEOPT_STACKFRAMES] = dtrace_stackframes_default;
14686 	opt[DTRACEOPT_USTACKFRAMES] = dtrace_ustackframes_default;
14687 	opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_default;
14688 	opt[DTRACEOPT_AGGRATE] = dtrace_aggrate_default;
14689 	opt[DTRACEOPT_SWITCHRATE] = dtrace_switchrate_default;
14690 	opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_default;
14691 	opt[DTRACEOPT_JSTACKFRAMES] = dtrace_jstackframes_default;
14692 	opt[DTRACEOPT_JSTACKSTRSIZE] = dtrace_jstackstrsize_default;
14693 
14694 	state->dts_activity = DTRACE_ACTIVITY_INACTIVE;
14695 
14696 	/*
14697 	 * Depending on the user credentials, we set flag bits which alter probe
14698 	 * visibility or the amount of destructiveness allowed.  In the case of
14699 	 * actual anonymous tracing, or the possession of all privileges, all of
14700 	 * the normal checks are bypassed.
14701 	 */
14702 	if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
14703 		state->dts_cred.dcr_visible = DTRACE_CRV_ALL;
14704 		state->dts_cred.dcr_action = DTRACE_CRA_ALL;
14705 	} else {
14706 		/*
14707 		 * Set up the credentials for this instantiation.  We take a
14708 		 * hold on the credential to prevent it from disappearing on
14709 		 * us; this in turn prevents the zone_t referenced by this
14710 		 * credential from disappearing.  This means that we can
14711 		 * examine the credential and the zone from probe context.
14712 		 */
14713 		crhold(cr);
14714 		state->dts_cred.dcr_cred = cr;
14715 
14716 		/*
14717 		 * CRA_PROC means "we have *some* privilege for dtrace" and
14718 		 * unlocks the use of variables like pid, zonename, etc.
14719 		 */
14720 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE) ||
14721 		    PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
14722 			state->dts_cred.dcr_action |= DTRACE_CRA_PROC;
14723 		}
14724 
14725 		/*
14726 		 * dtrace_user allows use of syscall and profile providers.
14727 		 * If the user also has proc_owner and/or proc_zone, we
14728 		 * extend the scope to include additional visibility and
14729 		 * destructive power.
14730 		 */
14731 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE)) {
14732 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE)) {
14733 				state->dts_cred.dcr_visible |=
14734 				    DTRACE_CRV_ALLPROC;
14735 
14736 				state->dts_cred.dcr_action |=
14737 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
14738 			}
14739 
14740 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE)) {
14741 				state->dts_cred.dcr_visible |=
14742 				    DTRACE_CRV_ALLZONE;
14743 
14744 				state->dts_cred.dcr_action |=
14745 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
14746 			}
14747 
14748 			/*
14749 			 * If we have all privs in whatever zone this is,
14750 			 * we can do destructive things to processes which
14751 			 * have altered credentials.
14752 			 */
14753 #ifdef illumos
14754 			if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
14755 			    cr->cr_zone->zone_privset)) {
14756 				state->dts_cred.dcr_action |=
14757 				    DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
14758 			}
14759 #endif
14760 		}
14761 
14762 		/*
14763 		 * Holding the dtrace_kernel privilege also implies that
14764 		 * the user has the dtrace_user privilege from a visibility
14765 		 * perspective.  But without further privileges, some
14766 		 * destructive actions are not available.
14767 		 */
14768 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE)) {
14769 			/*
14770 			 * Make all probes in all zones visible.  However,
14771 			 * this doesn't mean that all actions become available
14772 			 * to all zones.
14773 			 */
14774 			state->dts_cred.dcr_visible |= DTRACE_CRV_KERNEL |
14775 			    DTRACE_CRV_ALLPROC | DTRACE_CRV_ALLZONE;
14776 
14777 			state->dts_cred.dcr_action |= DTRACE_CRA_KERNEL |
14778 			    DTRACE_CRA_PROC;
14779 			/*
14780 			 * Holding proc_owner means that destructive actions
14781 			 * for *this* zone are allowed.
14782 			 */
14783 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
14784 				state->dts_cred.dcr_action |=
14785 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
14786 
14787 			/*
14788 			 * Holding proc_zone means that destructive actions
14789 			 * for this user/group ID in all zones is allowed.
14790 			 */
14791 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
14792 				state->dts_cred.dcr_action |=
14793 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
14794 
14795 #ifdef illumos
14796 			/*
14797 			 * If we have all privs in whatever zone this is,
14798 			 * we can do destructive things to processes which
14799 			 * have altered credentials.
14800 			 */
14801 			if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
14802 			    cr->cr_zone->zone_privset)) {
14803 				state->dts_cred.dcr_action |=
14804 				    DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
14805 			}
14806 #endif
14807 		}
14808 
14809 		/*
14810 		 * Holding the dtrace_proc privilege gives control over fasttrap
14811 		 * and pid providers.  We need to grant wider destructive
14812 		 * privileges in the event that the user has proc_owner and/or
14813 		 * proc_zone.
14814 		 */
14815 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
14816 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
14817 				state->dts_cred.dcr_action |=
14818 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
14819 
14820 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
14821 				state->dts_cred.dcr_action |=
14822 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
14823 		}
14824 	}
14825 
14826 	return (state);
14827 }
14828 
14829 static int
14830 dtrace_state_buffer(dtrace_state_t *state, dtrace_buffer_t *buf, int which)
14831 {
14832 	dtrace_optval_t *opt = state->dts_options, size;
14833 	processorid_t cpu = 0;;
14834 	int flags = 0, rval, factor, divisor = 1;
14835 
14836 	ASSERT(MUTEX_HELD(&dtrace_lock));
14837 	ASSERT(MUTEX_HELD(&cpu_lock));
14838 	ASSERT(which < DTRACEOPT_MAX);
14839 	ASSERT(state->dts_activity == DTRACE_ACTIVITY_INACTIVE ||
14840 	    (state == dtrace_anon.dta_state &&
14841 	    state->dts_activity == DTRACE_ACTIVITY_ACTIVE));
14842 
14843 	if (opt[which] == DTRACEOPT_UNSET || opt[which] == 0)
14844 		return (0);
14845 
14846 	if (opt[DTRACEOPT_CPU] != DTRACEOPT_UNSET)
14847 		cpu = opt[DTRACEOPT_CPU];
14848 
14849 	if (which == DTRACEOPT_SPECSIZE)
14850 		flags |= DTRACEBUF_NOSWITCH;
14851 
14852 	if (which == DTRACEOPT_BUFSIZE) {
14853 		if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_RING)
14854 			flags |= DTRACEBUF_RING;
14855 
14856 		if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_FILL)
14857 			flags |= DTRACEBUF_FILL;
14858 
14859 		if (state != dtrace_anon.dta_state ||
14860 		    state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
14861 			flags |= DTRACEBUF_INACTIVE;
14862 	}
14863 
14864 	for (size = opt[which]; size >= sizeof (uint64_t); size /= divisor) {
14865 		/*
14866 		 * The size must be 8-byte aligned.  If the size is not 8-byte
14867 		 * aligned, drop it down by the difference.
14868 		 */
14869 		if (size & (sizeof (uint64_t) - 1))
14870 			size -= size & (sizeof (uint64_t) - 1);
14871 
14872 		if (size < state->dts_reserve) {
14873 			/*
14874 			 * Buffers always must be large enough to accommodate
14875 			 * their prereserved space.  We return E2BIG instead
14876 			 * of ENOMEM in this case to allow for user-level
14877 			 * software to differentiate the cases.
14878 			 */
14879 			return (E2BIG);
14880 		}
14881 
14882 		rval = dtrace_buffer_alloc(buf, size, flags, cpu, &factor);
14883 
14884 		if (rval != ENOMEM) {
14885 			opt[which] = size;
14886 			return (rval);
14887 		}
14888 
14889 		if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
14890 			return (rval);
14891 
14892 		for (divisor = 2; divisor < factor; divisor <<= 1)
14893 			continue;
14894 	}
14895 
14896 	return (ENOMEM);
14897 }
14898 
14899 static int
14900 dtrace_state_buffers(dtrace_state_t *state)
14901 {
14902 	dtrace_speculation_t *spec = state->dts_speculations;
14903 	int rval, i;
14904 
14905 	if ((rval = dtrace_state_buffer(state, state->dts_buffer,
14906 	    DTRACEOPT_BUFSIZE)) != 0)
14907 		return (rval);
14908 
14909 	if ((rval = dtrace_state_buffer(state, state->dts_aggbuffer,
14910 	    DTRACEOPT_AGGSIZE)) != 0)
14911 		return (rval);
14912 
14913 	for (i = 0; i < state->dts_nspeculations; i++) {
14914 		if ((rval = dtrace_state_buffer(state,
14915 		    spec[i].dtsp_buffer, DTRACEOPT_SPECSIZE)) != 0)
14916 			return (rval);
14917 	}
14918 
14919 	return (0);
14920 }
14921 
14922 static void
14923 dtrace_state_prereserve(dtrace_state_t *state)
14924 {
14925 	dtrace_ecb_t *ecb;
14926 	dtrace_probe_t *probe;
14927 
14928 	state->dts_reserve = 0;
14929 
14930 	if (state->dts_options[DTRACEOPT_BUFPOLICY] != DTRACEOPT_BUFPOLICY_FILL)
14931 		return;
14932 
14933 	/*
14934 	 * If our buffer policy is a "fill" buffer policy, we need to set the
14935 	 * prereserved space to be the space required by the END probes.
14936 	 */
14937 	probe = dtrace_probes[dtrace_probeid_end - 1];
14938 	ASSERT(probe != NULL);
14939 
14940 	for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
14941 		if (ecb->dte_state != state)
14942 			continue;
14943 
14944 		state->dts_reserve += ecb->dte_needed + ecb->dte_alignment;
14945 	}
14946 }
14947 
14948 static int
14949 dtrace_state_go(dtrace_state_t *state, processorid_t *cpu)
14950 {
14951 	dtrace_optval_t *opt = state->dts_options, sz, nspec;
14952 	dtrace_speculation_t *spec;
14953 	dtrace_buffer_t *buf;
14954 #ifdef illumos
14955 	cyc_handler_t hdlr;
14956 	cyc_time_t when;
14957 #endif
14958 	int rval = 0, i, bufsize = NCPU * sizeof (dtrace_buffer_t);
14959 	dtrace_icookie_t cookie;
14960 
14961 	mutex_enter(&cpu_lock);
14962 	mutex_enter(&dtrace_lock);
14963 
14964 	if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
14965 		rval = EBUSY;
14966 		goto out;
14967 	}
14968 
14969 	/*
14970 	 * Before we can perform any checks, we must prime all of the
14971 	 * retained enablings that correspond to this state.
14972 	 */
14973 	dtrace_enabling_prime(state);
14974 
14975 	if (state->dts_destructive && !state->dts_cred.dcr_destructive) {
14976 		rval = EACCES;
14977 		goto out;
14978 	}
14979 
14980 	dtrace_state_prereserve(state);
14981 
14982 	/*
14983 	 * Now we want to do is try to allocate our speculations.
14984 	 * We do not automatically resize the number of speculations; if
14985 	 * this fails, we will fail the operation.
14986 	 */
14987 	nspec = opt[DTRACEOPT_NSPEC];
14988 	ASSERT(nspec != DTRACEOPT_UNSET);
14989 
14990 	if (nspec > INT_MAX) {
14991 		rval = ENOMEM;
14992 		goto out;
14993 	}
14994 
14995 	spec = kmem_zalloc(nspec * sizeof (dtrace_speculation_t),
14996 	    KM_NOSLEEP | KM_NORMALPRI);
14997 
14998 	if (spec == NULL) {
14999 		rval = ENOMEM;
15000 		goto out;
15001 	}
15002 
15003 	state->dts_speculations = spec;
15004 	state->dts_nspeculations = (int)nspec;
15005 
15006 	for (i = 0; i < nspec; i++) {
15007 		if ((buf = kmem_zalloc(bufsize,
15008 		    KM_NOSLEEP | KM_NORMALPRI)) == NULL) {
15009 			rval = ENOMEM;
15010 			goto err;
15011 		}
15012 
15013 		spec[i].dtsp_buffer = buf;
15014 	}
15015 
15016 	if (opt[DTRACEOPT_GRABANON] != DTRACEOPT_UNSET) {
15017 		if (dtrace_anon.dta_state == NULL) {
15018 			rval = ENOENT;
15019 			goto out;
15020 		}
15021 
15022 		if (state->dts_necbs != 0) {
15023 			rval = EALREADY;
15024 			goto out;
15025 		}
15026 
15027 		state->dts_anon = dtrace_anon_grab();
15028 		ASSERT(state->dts_anon != NULL);
15029 		state = state->dts_anon;
15030 
15031 		/*
15032 		 * We want "grabanon" to be set in the grabbed state, so we'll
15033 		 * copy that option value from the grabbing state into the
15034 		 * grabbed state.
15035 		 */
15036 		state->dts_options[DTRACEOPT_GRABANON] =
15037 		    opt[DTRACEOPT_GRABANON];
15038 
15039 		*cpu = dtrace_anon.dta_beganon;
15040 
15041 		/*
15042 		 * If the anonymous state is active (as it almost certainly
15043 		 * is if the anonymous enabling ultimately matched anything),
15044 		 * we don't allow any further option processing -- but we
15045 		 * don't return failure.
15046 		 */
15047 		if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
15048 			goto out;
15049 	}
15050 
15051 	if (opt[DTRACEOPT_AGGSIZE] != DTRACEOPT_UNSET &&
15052 	    opt[DTRACEOPT_AGGSIZE] != 0) {
15053 		if (state->dts_aggregations == NULL) {
15054 			/*
15055 			 * We're not going to create an aggregation buffer
15056 			 * because we don't have any ECBs that contain
15057 			 * aggregations -- set this option to 0.
15058 			 */
15059 			opt[DTRACEOPT_AGGSIZE] = 0;
15060 		} else {
15061 			/*
15062 			 * If we have an aggregation buffer, we must also have
15063 			 * a buffer to use as scratch.
15064 			 */
15065 			if (opt[DTRACEOPT_BUFSIZE] == DTRACEOPT_UNSET ||
15066 			    opt[DTRACEOPT_BUFSIZE] < state->dts_needed) {
15067 				opt[DTRACEOPT_BUFSIZE] = state->dts_needed;
15068 			}
15069 		}
15070 	}
15071 
15072 	if (opt[DTRACEOPT_SPECSIZE] != DTRACEOPT_UNSET &&
15073 	    opt[DTRACEOPT_SPECSIZE] != 0) {
15074 		if (!state->dts_speculates) {
15075 			/*
15076 			 * We're not going to create speculation buffers
15077 			 * because we don't have any ECBs that actually
15078 			 * speculate -- set the speculation size to 0.
15079 			 */
15080 			opt[DTRACEOPT_SPECSIZE] = 0;
15081 		}
15082 	}
15083 
15084 	/*
15085 	 * The bare minimum size for any buffer that we're actually going to
15086 	 * do anything to is sizeof (uint64_t).
15087 	 */
15088 	sz = sizeof (uint64_t);
15089 
15090 	if ((state->dts_needed != 0 && opt[DTRACEOPT_BUFSIZE] < sz) ||
15091 	    (state->dts_speculates && opt[DTRACEOPT_SPECSIZE] < sz) ||
15092 	    (state->dts_aggregations != NULL && opt[DTRACEOPT_AGGSIZE] < sz)) {
15093 		/*
15094 		 * A buffer size has been explicitly set to 0 (or to a size
15095 		 * that will be adjusted to 0) and we need the space -- we
15096 		 * need to return failure.  We return ENOSPC to differentiate
15097 		 * it from failing to allocate a buffer due to failure to meet
15098 		 * the reserve (for which we return E2BIG).
15099 		 */
15100 		rval = ENOSPC;
15101 		goto out;
15102 	}
15103 
15104 	if ((rval = dtrace_state_buffers(state)) != 0)
15105 		goto err;
15106 
15107 	if ((sz = opt[DTRACEOPT_DYNVARSIZE]) == DTRACEOPT_UNSET)
15108 		sz = dtrace_dstate_defsize;
15109 
15110 	do {
15111 		rval = dtrace_dstate_init(&state->dts_vstate.dtvs_dynvars, sz);
15112 
15113 		if (rval == 0)
15114 			break;
15115 
15116 		if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
15117 			goto err;
15118 	} while (sz >>= 1);
15119 
15120 	opt[DTRACEOPT_DYNVARSIZE] = sz;
15121 
15122 	if (rval != 0)
15123 		goto err;
15124 
15125 	if (opt[DTRACEOPT_STATUSRATE] > dtrace_statusrate_max)
15126 		opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_max;
15127 
15128 	if (opt[DTRACEOPT_CLEANRATE] == 0)
15129 		opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
15130 
15131 	if (opt[DTRACEOPT_CLEANRATE] < dtrace_cleanrate_min)
15132 		opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_min;
15133 
15134 	if (opt[DTRACEOPT_CLEANRATE] > dtrace_cleanrate_max)
15135 		opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
15136 
15137 	state->dts_alive = state->dts_laststatus = dtrace_gethrtime();
15138 #ifdef illumos
15139 	hdlr.cyh_func = (cyc_func_t)dtrace_state_clean;
15140 	hdlr.cyh_arg = state;
15141 	hdlr.cyh_level = CY_LOW_LEVEL;
15142 
15143 	when.cyt_when = 0;
15144 	when.cyt_interval = opt[DTRACEOPT_CLEANRATE];
15145 
15146 	state->dts_cleaner = cyclic_add(&hdlr, &when);
15147 
15148 	hdlr.cyh_func = (cyc_func_t)dtrace_state_deadman;
15149 	hdlr.cyh_arg = state;
15150 	hdlr.cyh_level = CY_LOW_LEVEL;
15151 
15152 	when.cyt_when = 0;
15153 	when.cyt_interval = dtrace_deadman_interval;
15154 
15155 	state->dts_deadman = cyclic_add(&hdlr, &when);
15156 #else
15157 	callout_reset(&state->dts_cleaner, hz * opt[DTRACEOPT_CLEANRATE] / NANOSEC,
15158 	    dtrace_state_clean, state);
15159 	callout_reset(&state->dts_deadman, hz * dtrace_deadman_interval / NANOSEC,
15160 	    dtrace_state_deadman, state);
15161 #endif
15162 
15163 	state->dts_activity = DTRACE_ACTIVITY_WARMUP;
15164 
15165 #ifdef illumos
15166 	if (state->dts_getf != 0 &&
15167 	    !(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)) {
15168 		/*
15169 		 * We don't have kernel privs but we have at least one call
15170 		 * to getf(); we need to bump our zone's count, and (if
15171 		 * this is the first enabling to have an unprivileged call
15172 		 * to getf()) we need to hook into closef().
15173 		 */
15174 		state->dts_cred.dcr_cred->cr_zone->zone_dtrace_getf++;
15175 
15176 		if (dtrace_getf++ == 0) {
15177 			ASSERT(dtrace_closef == NULL);
15178 			dtrace_closef = dtrace_getf_barrier;
15179 		}
15180 	}
15181 #endif
15182 
15183 	/*
15184 	 * Now it's time to actually fire the BEGIN probe.  We need to disable
15185 	 * interrupts here both to record the CPU on which we fired the BEGIN
15186 	 * probe (the data from this CPU will be processed first at user
15187 	 * level) and to manually activate the buffer for this CPU.
15188 	 */
15189 	cookie = dtrace_interrupt_disable();
15190 	*cpu = curcpu;
15191 	ASSERT(state->dts_buffer[*cpu].dtb_flags & DTRACEBUF_INACTIVE);
15192 	state->dts_buffer[*cpu].dtb_flags &= ~DTRACEBUF_INACTIVE;
15193 
15194 	dtrace_probe(dtrace_probeid_begin,
15195 	    (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
15196 	dtrace_interrupt_enable(cookie);
15197 	/*
15198 	 * We may have had an exit action from a BEGIN probe; only change our
15199 	 * state to ACTIVE if we're still in WARMUP.
15200 	 */
15201 	ASSERT(state->dts_activity == DTRACE_ACTIVITY_WARMUP ||
15202 	    state->dts_activity == DTRACE_ACTIVITY_DRAINING);
15203 
15204 	if (state->dts_activity == DTRACE_ACTIVITY_WARMUP)
15205 		state->dts_activity = DTRACE_ACTIVITY_ACTIVE;
15206 
15207 #ifdef __FreeBSD__
15208 	/*
15209 	 * We enable anonymous tracing before APs are started, so we must
15210 	 * activate buffers using the current CPU.
15211 	 */
15212 	if (state == dtrace_anon.dta_state)
15213 		for (int i = 0; i < NCPU; i++)
15214 			dtrace_buffer_activate_cpu(state, i);
15215 	else
15216 		dtrace_xcall(DTRACE_CPUALL,
15217 		    (dtrace_xcall_t)dtrace_buffer_activate, state);
15218 #else
15219 	/*
15220 	 * Regardless of whether or not now we're in ACTIVE or DRAINING, we
15221 	 * want each CPU to transition its principal buffer out of the
15222 	 * INACTIVE state.  Doing this assures that no CPU will suddenly begin
15223 	 * processing an ECB halfway down a probe's ECB chain; all CPUs will
15224 	 * atomically transition from processing none of a state's ECBs to
15225 	 * processing all of them.
15226 	 */
15227 	dtrace_xcall(DTRACE_CPUALL,
15228 	    (dtrace_xcall_t)dtrace_buffer_activate, state);
15229 #endif
15230 	goto out;
15231 
15232 err:
15233 	dtrace_buffer_free(state->dts_buffer);
15234 	dtrace_buffer_free(state->dts_aggbuffer);
15235 
15236 	if ((nspec = state->dts_nspeculations) == 0) {
15237 		ASSERT(state->dts_speculations == NULL);
15238 		goto out;
15239 	}
15240 
15241 	spec = state->dts_speculations;
15242 	ASSERT(spec != NULL);
15243 
15244 	for (i = 0; i < state->dts_nspeculations; i++) {
15245 		if ((buf = spec[i].dtsp_buffer) == NULL)
15246 			break;
15247 
15248 		dtrace_buffer_free(buf);
15249 		kmem_free(buf, bufsize);
15250 	}
15251 
15252 	kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
15253 	state->dts_nspeculations = 0;
15254 	state->dts_speculations = NULL;
15255 
15256 out:
15257 	mutex_exit(&dtrace_lock);
15258 	mutex_exit(&cpu_lock);
15259 
15260 	return (rval);
15261 }
15262 
15263 static int
15264 dtrace_state_stop(dtrace_state_t *state, processorid_t *cpu)
15265 {
15266 	dtrace_icookie_t cookie;
15267 
15268 	ASSERT(MUTEX_HELD(&dtrace_lock));
15269 
15270 	if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE &&
15271 	    state->dts_activity != DTRACE_ACTIVITY_DRAINING)
15272 		return (EINVAL);
15273 
15274 	/*
15275 	 * We'll set the activity to DTRACE_ACTIVITY_DRAINING, and issue a sync
15276 	 * to be sure that every CPU has seen it.  See below for the details
15277 	 * on why this is done.
15278 	 */
15279 	state->dts_activity = DTRACE_ACTIVITY_DRAINING;
15280 	dtrace_sync();
15281 
15282 	/*
15283 	 * By this point, it is impossible for any CPU to be still processing
15284 	 * with DTRACE_ACTIVITY_ACTIVE.  We can thus set our activity to
15285 	 * DTRACE_ACTIVITY_COOLDOWN and know that we're not racing with any
15286 	 * other CPU in dtrace_buffer_reserve().  This allows dtrace_probe()
15287 	 * and callees to know that the activity is DTRACE_ACTIVITY_COOLDOWN
15288 	 * iff we're in the END probe.
15289 	 */
15290 	state->dts_activity = DTRACE_ACTIVITY_COOLDOWN;
15291 	dtrace_sync();
15292 	ASSERT(state->dts_activity == DTRACE_ACTIVITY_COOLDOWN);
15293 
15294 	/*
15295 	 * Finally, we can release the reserve and call the END probe.  We
15296 	 * disable interrupts across calling the END probe to allow us to
15297 	 * return the CPU on which we actually called the END probe.  This
15298 	 * allows user-land to be sure that this CPU's principal buffer is
15299 	 * processed last.
15300 	 */
15301 	state->dts_reserve = 0;
15302 
15303 	cookie = dtrace_interrupt_disable();
15304 	*cpu = curcpu;
15305 	dtrace_probe(dtrace_probeid_end,
15306 	    (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
15307 	dtrace_interrupt_enable(cookie);
15308 
15309 	state->dts_activity = DTRACE_ACTIVITY_STOPPED;
15310 	dtrace_sync();
15311 
15312 #ifdef illumos
15313 	if (state->dts_getf != 0 &&
15314 	    !(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)) {
15315 		/*
15316 		 * We don't have kernel privs but we have at least one call
15317 		 * to getf(); we need to lower our zone's count, and (if
15318 		 * this is the last enabling to have an unprivileged call
15319 		 * to getf()) we need to clear the closef() hook.
15320 		 */
15321 		ASSERT(state->dts_cred.dcr_cred->cr_zone->zone_dtrace_getf > 0);
15322 		ASSERT(dtrace_closef == dtrace_getf_barrier);
15323 		ASSERT(dtrace_getf > 0);
15324 
15325 		state->dts_cred.dcr_cred->cr_zone->zone_dtrace_getf--;
15326 
15327 		if (--dtrace_getf == 0)
15328 			dtrace_closef = NULL;
15329 	}
15330 #endif
15331 
15332 	return (0);
15333 }
15334 
15335 static int
15336 dtrace_state_option(dtrace_state_t *state, dtrace_optid_t option,
15337     dtrace_optval_t val)
15338 {
15339 	ASSERT(MUTEX_HELD(&dtrace_lock));
15340 
15341 	if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
15342 		return (EBUSY);
15343 
15344 	if (option >= DTRACEOPT_MAX)
15345 		return (EINVAL);
15346 
15347 	if (option != DTRACEOPT_CPU && val < 0)
15348 		return (EINVAL);
15349 
15350 	switch (option) {
15351 	case DTRACEOPT_DESTRUCTIVE:
15352 		if (dtrace_destructive_disallow)
15353 			return (EACCES);
15354 
15355 		state->dts_cred.dcr_destructive = 1;
15356 		break;
15357 
15358 	case DTRACEOPT_BUFSIZE:
15359 	case DTRACEOPT_DYNVARSIZE:
15360 	case DTRACEOPT_AGGSIZE:
15361 	case DTRACEOPT_SPECSIZE:
15362 	case DTRACEOPT_STRSIZE:
15363 		if (val < 0)
15364 			return (EINVAL);
15365 
15366 		if (val >= LONG_MAX) {
15367 			/*
15368 			 * If this is an otherwise negative value, set it to
15369 			 * the highest multiple of 128m less than LONG_MAX.
15370 			 * Technically, we're adjusting the size without
15371 			 * regard to the buffer resizing policy, but in fact,
15372 			 * this has no effect -- if we set the buffer size to
15373 			 * ~LONG_MAX and the buffer policy is ultimately set to
15374 			 * be "manual", the buffer allocation is guaranteed to
15375 			 * fail, if only because the allocation requires two
15376 			 * buffers.  (We set the the size to the highest
15377 			 * multiple of 128m because it ensures that the size
15378 			 * will remain a multiple of a megabyte when
15379 			 * repeatedly halved -- all the way down to 15m.)
15380 			 */
15381 			val = LONG_MAX - (1 << 27) + 1;
15382 		}
15383 	}
15384 
15385 	state->dts_options[option] = val;
15386 
15387 	return (0);
15388 }
15389 
15390 static void
15391 dtrace_state_destroy(dtrace_state_t *state)
15392 {
15393 	dtrace_ecb_t *ecb;
15394 	dtrace_vstate_t *vstate = &state->dts_vstate;
15395 #ifdef illumos
15396 	minor_t minor = getminor(state->dts_dev);
15397 #endif
15398 	int i, bufsize = NCPU * sizeof (dtrace_buffer_t);
15399 	dtrace_speculation_t *spec = state->dts_speculations;
15400 	int nspec = state->dts_nspeculations;
15401 	uint32_t match;
15402 
15403 	ASSERT(MUTEX_HELD(&dtrace_lock));
15404 	ASSERT(MUTEX_HELD(&cpu_lock));
15405 
15406 	/*
15407 	 * First, retract any retained enablings for this state.
15408 	 */
15409 	dtrace_enabling_retract(state);
15410 	ASSERT(state->dts_nretained == 0);
15411 
15412 	if (state->dts_activity == DTRACE_ACTIVITY_ACTIVE ||
15413 	    state->dts_activity == DTRACE_ACTIVITY_DRAINING) {
15414 		/*
15415 		 * We have managed to come into dtrace_state_destroy() on a
15416 		 * hot enabling -- almost certainly because of a disorderly
15417 		 * shutdown of a consumer.  (That is, a consumer that is
15418 		 * exiting without having called dtrace_stop().) In this case,
15419 		 * we're going to set our activity to be KILLED, and then
15420 		 * issue a sync to be sure that everyone is out of probe
15421 		 * context before we start blowing away ECBs.
15422 		 */
15423 		state->dts_activity = DTRACE_ACTIVITY_KILLED;
15424 		dtrace_sync();
15425 	}
15426 
15427 	/*
15428 	 * Release the credential hold we took in dtrace_state_create().
15429 	 */
15430 	if (state->dts_cred.dcr_cred != NULL)
15431 		crfree(state->dts_cred.dcr_cred);
15432 
15433 	/*
15434 	 * Now we can safely disable and destroy any enabled probes.  Because
15435 	 * any DTRACE_PRIV_KERNEL probes may actually be slowing our progress
15436 	 * (especially if they're all enabled), we take two passes through the
15437 	 * ECBs:  in the first, we disable just DTRACE_PRIV_KERNEL probes, and
15438 	 * in the second we disable whatever is left over.
15439 	 */
15440 	for (match = DTRACE_PRIV_KERNEL; ; match = 0) {
15441 		for (i = 0; i < state->dts_necbs; i++) {
15442 			if ((ecb = state->dts_ecbs[i]) == NULL)
15443 				continue;
15444 
15445 			if (match && ecb->dte_probe != NULL) {
15446 				dtrace_probe_t *probe = ecb->dte_probe;
15447 				dtrace_provider_t *prov = probe->dtpr_provider;
15448 
15449 				if (!(prov->dtpv_priv.dtpp_flags & match))
15450 					continue;
15451 			}
15452 
15453 			dtrace_ecb_disable(ecb);
15454 			dtrace_ecb_destroy(ecb);
15455 		}
15456 
15457 		if (!match)
15458 			break;
15459 	}
15460 
15461 	/*
15462 	 * Before we free the buffers, perform one more sync to assure that
15463 	 * every CPU is out of probe context.
15464 	 */
15465 	dtrace_sync();
15466 
15467 	dtrace_buffer_free(state->dts_buffer);
15468 	dtrace_buffer_free(state->dts_aggbuffer);
15469 
15470 	for (i = 0; i < nspec; i++)
15471 		dtrace_buffer_free(spec[i].dtsp_buffer);
15472 
15473 #ifdef illumos
15474 	if (state->dts_cleaner != CYCLIC_NONE)
15475 		cyclic_remove(state->dts_cleaner);
15476 
15477 	if (state->dts_deadman != CYCLIC_NONE)
15478 		cyclic_remove(state->dts_deadman);
15479 #else
15480 	callout_stop(&state->dts_cleaner);
15481 	callout_drain(&state->dts_cleaner);
15482 	callout_stop(&state->dts_deadman);
15483 	callout_drain(&state->dts_deadman);
15484 #endif
15485 
15486 	dtrace_dstate_fini(&vstate->dtvs_dynvars);
15487 	dtrace_vstate_fini(vstate);
15488 	if (state->dts_ecbs != NULL)
15489 		kmem_free(state->dts_ecbs, state->dts_necbs * sizeof (dtrace_ecb_t *));
15490 
15491 	if (state->dts_aggregations != NULL) {
15492 #ifdef DEBUG
15493 		for (i = 0; i < state->dts_naggregations; i++)
15494 			ASSERT(state->dts_aggregations[i] == NULL);
15495 #endif
15496 		ASSERT(state->dts_naggregations > 0);
15497 		kmem_free(state->dts_aggregations,
15498 		    state->dts_naggregations * sizeof (dtrace_aggregation_t *));
15499 	}
15500 
15501 	kmem_free(state->dts_buffer, bufsize);
15502 	kmem_free(state->dts_aggbuffer, bufsize);
15503 
15504 	for (i = 0; i < nspec; i++)
15505 		kmem_free(spec[i].dtsp_buffer, bufsize);
15506 
15507 	if (spec != NULL)
15508 		kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
15509 
15510 	dtrace_format_destroy(state);
15511 
15512 	if (state->dts_aggid_arena != NULL) {
15513 #ifdef illumos
15514 		vmem_destroy(state->dts_aggid_arena);
15515 #else
15516 		delete_unrhdr(state->dts_aggid_arena);
15517 #endif
15518 		state->dts_aggid_arena = NULL;
15519 	}
15520 #ifdef illumos
15521 	ddi_soft_state_free(dtrace_softstate, minor);
15522 	vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
15523 #endif
15524 }
15525 
15526 /*
15527  * DTrace Anonymous Enabling Functions
15528  */
15529 static dtrace_state_t *
15530 dtrace_anon_grab(void)
15531 {
15532 	dtrace_state_t *state;
15533 
15534 	ASSERT(MUTEX_HELD(&dtrace_lock));
15535 
15536 	if ((state = dtrace_anon.dta_state) == NULL) {
15537 		ASSERT(dtrace_anon.dta_enabling == NULL);
15538 		return (NULL);
15539 	}
15540 
15541 	ASSERT(dtrace_anon.dta_enabling != NULL);
15542 	ASSERT(dtrace_retained != NULL);
15543 
15544 	dtrace_enabling_destroy(dtrace_anon.dta_enabling);
15545 	dtrace_anon.dta_enabling = NULL;
15546 	dtrace_anon.dta_state = NULL;
15547 
15548 	return (state);
15549 }
15550 
15551 static void
15552 dtrace_anon_property(void)
15553 {
15554 	int i, rv;
15555 	dtrace_state_t *state;
15556 	dof_hdr_t *dof;
15557 	char c[32];		/* enough for "dof-data-" + digits */
15558 
15559 	ASSERT(MUTEX_HELD(&dtrace_lock));
15560 	ASSERT(MUTEX_HELD(&cpu_lock));
15561 
15562 	for (i = 0; ; i++) {
15563 		(void) snprintf(c, sizeof (c), "dof-data-%d", i);
15564 
15565 		dtrace_err_verbose = 1;
15566 
15567 		if ((dof = dtrace_dof_property(c)) == NULL) {
15568 			dtrace_err_verbose = 0;
15569 			break;
15570 		}
15571 
15572 #ifdef illumos
15573 		/*
15574 		 * We want to create anonymous state, so we need to transition
15575 		 * the kernel debugger to indicate that DTrace is active.  If
15576 		 * this fails (e.g. because the debugger has modified text in
15577 		 * some way), we won't continue with the processing.
15578 		 */
15579 		if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
15580 			cmn_err(CE_NOTE, "kernel debugger active; anonymous "
15581 			    "enabling ignored.");
15582 			dtrace_dof_destroy(dof);
15583 			break;
15584 		}
15585 #endif
15586 
15587 		/*
15588 		 * If we haven't allocated an anonymous state, we'll do so now.
15589 		 */
15590 		if ((state = dtrace_anon.dta_state) == NULL) {
15591 			state = dtrace_state_create(NULL, NULL);
15592 			dtrace_anon.dta_state = state;
15593 
15594 			if (state == NULL) {
15595 				/*
15596 				 * This basically shouldn't happen:  the only
15597 				 * failure mode from dtrace_state_create() is a
15598 				 * failure of ddi_soft_state_zalloc() that
15599 				 * itself should never happen.  Still, the
15600 				 * interface allows for a failure mode, and
15601 				 * we want to fail as gracefully as possible:
15602 				 * we'll emit an error message and cease
15603 				 * processing anonymous state in this case.
15604 				 */
15605 				cmn_err(CE_WARN, "failed to create "
15606 				    "anonymous state");
15607 				dtrace_dof_destroy(dof);
15608 				break;
15609 			}
15610 		}
15611 
15612 		rv = dtrace_dof_slurp(dof, &state->dts_vstate, CRED(),
15613 		    &dtrace_anon.dta_enabling, 0, 0, B_TRUE);
15614 
15615 		if (rv == 0)
15616 			rv = dtrace_dof_options(dof, state);
15617 
15618 		dtrace_err_verbose = 0;
15619 		dtrace_dof_destroy(dof);
15620 
15621 		if (rv != 0) {
15622 			/*
15623 			 * This is malformed DOF; chuck any anonymous state
15624 			 * that we created.
15625 			 */
15626 			ASSERT(dtrace_anon.dta_enabling == NULL);
15627 			dtrace_state_destroy(state);
15628 			dtrace_anon.dta_state = NULL;
15629 			break;
15630 		}
15631 
15632 		ASSERT(dtrace_anon.dta_enabling != NULL);
15633 	}
15634 
15635 	if (dtrace_anon.dta_enabling != NULL) {
15636 		int rval;
15637 
15638 		/*
15639 		 * dtrace_enabling_retain() can only fail because we are
15640 		 * trying to retain more enablings than are allowed -- but
15641 		 * we only have one anonymous enabling, and we are guaranteed
15642 		 * to be allowed at least one retained enabling; we assert
15643 		 * that dtrace_enabling_retain() returns success.
15644 		 */
15645 		rval = dtrace_enabling_retain(dtrace_anon.dta_enabling);
15646 		ASSERT(rval == 0);
15647 
15648 		dtrace_enabling_dump(dtrace_anon.dta_enabling);
15649 	}
15650 }
15651 
15652 /*
15653  * DTrace Helper Functions
15654  */
15655 static void
15656 dtrace_helper_trace(dtrace_helper_action_t *helper,
15657     dtrace_mstate_t *mstate, dtrace_vstate_t *vstate, int where)
15658 {
15659 	uint32_t size, next, nnext, i;
15660 	dtrace_helptrace_t *ent, *buffer;
15661 	uint16_t flags = cpu_core[curcpu].cpuc_dtrace_flags;
15662 
15663 	if ((buffer = dtrace_helptrace_buffer) == NULL)
15664 		return;
15665 
15666 	ASSERT(vstate->dtvs_nlocals <= dtrace_helptrace_nlocals);
15667 
15668 	/*
15669 	 * What would a tracing framework be without its own tracing
15670 	 * framework?  (Well, a hell of a lot simpler, for starters...)
15671 	 */
15672 	size = sizeof (dtrace_helptrace_t) + dtrace_helptrace_nlocals *
15673 	    sizeof (uint64_t) - sizeof (uint64_t);
15674 
15675 	/*
15676 	 * Iterate until we can allocate a slot in the trace buffer.
15677 	 */
15678 	do {
15679 		next = dtrace_helptrace_next;
15680 
15681 		if (next + size < dtrace_helptrace_bufsize) {
15682 			nnext = next + size;
15683 		} else {
15684 			nnext = size;
15685 		}
15686 	} while (dtrace_cas32(&dtrace_helptrace_next, next, nnext) != next);
15687 
15688 	/*
15689 	 * We have our slot; fill it in.
15690 	 */
15691 	if (nnext == size) {
15692 		dtrace_helptrace_wrapped++;
15693 		next = 0;
15694 	}
15695 
15696 	ent = (dtrace_helptrace_t *)((uintptr_t)buffer + next);
15697 	ent->dtht_helper = helper;
15698 	ent->dtht_where = where;
15699 	ent->dtht_nlocals = vstate->dtvs_nlocals;
15700 
15701 	ent->dtht_fltoffs = (mstate->dtms_present & DTRACE_MSTATE_FLTOFFS) ?
15702 	    mstate->dtms_fltoffs : -1;
15703 	ent->dtht_fault = DTRACE_FLAGS2FLT(flags);
15704 	ent->dtht_illval = cpu_core[curcpu].cpuc_dtrace_illval;
15705 
15706 	for (i = 0; i < vstate->dtvs_nlocals; i++) {
15707 		dtrace_statvar_t *svar;
15708 
15709 		if ((svar = vstate->dtvs_locals[i]) == NULL)
15710 			continue;
15711 
15712 		ASSERT(svar->dtsv_size >= NCPU * sizeof (uint64_t));
15713 		ent->dtht_locals[i] =
15714 		    ((uint64_t *)(uintptr_t)svar->dtsv_data)[curcpu];
15715 	}
15716 }
15717 
15718 static uint64_t
15719 dtrace_helper(int which, dtrace_mstate_t *mstate,
15720     dtrace_state_t *state, uint64_t arg0, uint64_t arg1)
15721 {
15722 	uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
15723 	uint64_t sarg0 = mstate->dtms_arg[0];
15724 	uint64_t sarg1 = mstate->dtms_arg[1];
15725 	uint64_t rval = 0;
15726 	dtrace_helpers_t *helpers = curproc->p_dtrace_helpers;
15727 	dtrace_helper_action_t *helper;
15728 	dtrace_vstate_t *vstate;
15729 	dtrace_difo_t *pred;
15730 	int i, trace = dtrace_helptrace_buffer != NULL;
15731 
15732 	ASSERT(which >= 0 && which < DTRACE_NHELPER_ACTIONS);
15733 
15734 	if (helpers == NULL)
15735 		return (0);
15736 
15737 	if ((helper = helpers->dthps_actions[which]) == NULL)
15738 		return (0);
15739 
15740 	vstate = &helpers->dthps_vstate;
15741 	mstate->dtms_arg[0] = arg0;
15742 	mstate->dtms_arg[1] = arg1;
15743 
15744 	/*
15745 	 * Now iterate over each helper.  If its predicate evaluates to 'true',
15746 	 * we'll call the corresponding actions.  Note that the below calls
15747 	 * to dtrace_dif_emulate() may set faults in machine state.  This is
15748 	 * okay:  our caller (the outer dtrace_dif_emulate()) will simply plow
15749 	 * the stored DIF offset with its own (which is the desired behavior).
15750 	 * Also, note the calls to dtrace_dif_emulate() may allocate scratch
15751 	 * from machine state; this is okay, too.
15752 	 */
15753 	for (; helper != NULL; helper = helper->dtha_next) {
15754 		if ((pred = helper->dtha_predicate) != NULL) {
15755 			if (trace)
15756 				dtrace_helper_trace(helper, mstate, vstate, 0);
15757 
15758 			if (!dtrace_dif_emulate(pred, mstate, vstate, state))
15759 				goto next;
15760 
15761 			if (*flags & CPU_DTRACE_FAULT)
15762 				goto err;
15763 		}
15764 
15765 		for (i = 0; i < helper->dtha_nactions; i++) {
15766 			if (trace)
15767 				dtrace_helper_trace(helper,
15768 				    mstate, vstate, i + 1);
15769 
15770 			rval = dtrace_dif_emulate(helper->dtha_actions[i],
15771 			    mstate, vstate, state);
15772 
15773 			if (*flags & CPU_DTRACE_FAULT)
15774 				goto err;
15775 		}
15776 
15777 next:
15778 		if (trace)
15779 			dtrace_helper_trace(helper, mstate, vstate,
15780 			    DTRACE_HELPTRACE_NEXT);
15781 	}
15782 
15783 	if (trace)
15784 		dtrace_helper_trace(helper, mstate, vstate,
15785 		    DTRACE_HELPTRACE_DONE);
15786 
15787 	/*
15788 	 * Restore the arg0 that we saved upon entry.
15789 	 */
15790 	mstate->dtms_arg[0] = sarg0;
15791 	mstate->dtms_arg[1] = sarg1;
15792 
15793 	return (rval);
15794 
15795 err:
15796 	if (trace)
15797 		dtrace_helper_trace(helper, mstate, vstate,
15798 		    DTRACE_HELPTRACE_ERR);
15799 
15800 	/*
15801 	 * Restore the arg0 that we saved upon entry.
15802 	 */
15803 	mstate->dtms_arg[0] = sarg0;
15804 	mstate->dtms_arg[1] = sarg1;
15805 
15806 	return (0);
15807 }
15808 
15809 static void
15810 dtrace_helper_action_destroy(dtrace_helper_action_t *helper,
15811     dtrace_vstate_t *vstate)
15812 {
15813 	int i;
15814 
15815 	if (helper->dtha_predicate != NULL)
15816 		dtrace_difo_release(helper->dtha_predicate, vstate);
15817 
15818 	for (i = 0; i < helper->dtha_nactions; i++) {
15819 		ASSERT(helper->dtha_actions[i] != NULL);
15820 		dtrace_difo_release(helper->dtha_actions[i], vstate);
15821 	}
15822 
15823 	kmem_free(helper->dtha_actions,
15824 	    helper->dtha_nactions * sizeof (dtrace_difo_t *));
15825 	kmem_free(helper, sizeof (dtrace_helper_action_t));
15826 }
15827 
15828 static int
15829 dtrace_helper_destroygen(dtrace_helpers_t *help, int gen)
15830 {
15831 	proc_t *p = curproc;
15832 	dtrace_vstate_t *vstate;
15833 	int i;
15834 
15835 	if (help == NULL)
15836 		help = p->p_dtrace_helpers;
15837 
15838 	ASSERT(MUTEX_HELD(&dtrace_lock));
15839 
15840 	if (help == NULL || gen > help->dthps_generation)
15841 		return (EINVAL);
15842 
15843 	vstate = &help->dthps_vstate;
15844 
15845 	for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
15846 		dtrace_helper_action_t *last = NULL, *h, *next;
15847 
15848 		for (h = help->dthps_actions[i]; h != NULL; h = next) {
15849 			next = h->dtha_next;
15850 
15851 			if (h->dtha_generation == gen) {
15852 				if (last != NULL) {
15853 					last->dtha_next = next;
15854 				} else {
15855 					help->dthps_actions[i] = next;
15856 				}
15857 
15858 				dtrace_helper_action_destroy(h, vstate);
15859 			} else {
15860 				last = h;
15861 			}
15862 		}
15863 	}
15864 
15865 	/*
15866 	 * Interate until we've cleared out all helper providers with the
15867 	 * given generation number.
15868 	 */
15869 	for (;;) {
15870 		dtrace_helper_provider_t *prov;
15871 
15872 		/*
15873 		 * Look for a helper provider with the right generation. We
15874 		 * have to start back at the beginning of the list each time
15875 		 * because we drop dtrace_lock. It's unlikely that we'll make
15876 		 * more than two passes.
15877 		 */
15878 		for (i = 0; i < help->dthps_nprovs; i++) {
15879 			prov = help->dthps_provs[i];
15880 
15881 			if (prov->dthp_generation == gen)
15882 				break;
15883 		}
15884 
15885 		/*
15886 		 * If there were no matches, we're done.
15887 		 */
15888 		if (i == help->dthps_nprovs)
15889 			break;
15890 
15891 		/*
15892 		 * Move the last helper provider into this slot.
15893 		 */
15894 		help->dthps_nprovs--;
15895 		help->dthps_provs[i] = help->dthps_provs[help->dthps_nprovs];
15896 		help->dthps_provs[help->dthps_nprovs] = NULL;
15897 
15898 		mutex_exit(&dtrace_lock);
15899 
15900 		/*
15901 		 * If we have a meta provider, remove this helper provider.
15902 		 */
15903 		mutex_enter(&dtrace_meta_lock);
15904 		if (dtrace_meta_pid != NULL) {
15905 			ASSERT(dtrace_deferred_pid == NULL);
15906 			dtrace_helper_provider_remove(&prov->dthp_prov,
15907 			    p->p_pid);
15908 		}
15909 		mutex_exit(&dtrace_meta_lock);
15910 
15911 		dtrace_helper_provider_destroy(prov);
15912 
15913 		mutex_enter(&dtrace_lock);
15914 	}
15915 
15916 	return (0);
15917 }
15918 
15919 static int
15920 dtrace_helper_validate(dtrace_helper_action_t *helper)
15921 {
15922 	int err = 0, i;
15923 	dtrace_difo_t *dp;
15924 
15925 	if ((dp = helper->dtha_predicate) != NULL)
15926 		err += dtrace_difo_validate_helper(dp);
15927 
15928 	for (i = 0; i < helper->dtha_nactions; i++)
15929 		err += dtrace_difo_validate_helper(helper->dtha_actions[i]);
15930 
15931 	return (err == 0);
15932 }
15933 
15934 static int
15935 dtrace_helper_action_add(int which, dtrace_ecbdesc_t *ep,
15936     dtrace_helpers_t *help)
15937 {
15938 	dtrace_helper_action_t *helper, *last;
15939 	dtrace_actdesc_t *act;
15940 	dtrace_vstate_t *vstate;
15941 	dtrace_predicate_t *pred;
15942 	int count = 0, nactions = 0, i;
15943 
15944 	if (which < 0 || which >= DTRACE_NHELPER_ACTIONS)
15945 		return (EINVAL);
15946 
15947 	last = help->dthps_actions[which];
15948 	vstate = &help->dthps_vstate;
15949 
15950 	for (count = 0; last != NULL; last = last->dtha_next) {
15951 		count++;
15952 		if (last->dtha_next == NULL)
15953 			break;
15954 	}
15955 
15956 	/*
15957 	 * If we already have dtrace_helper_actions_max helper actions for this
15958 	 * helper action type, we'll refuse to add a new one.
15959 	 */
15960 	if (count >= dtrace_helper_actions_max)
15961 		return (ENOSPC);
15962 
15963 	helper = kmem_zalloc(sizeof (dtrace_helper_action_t), KM_SLEEP);
15964 	helper->dtha_generation = help->dthps_generation;
15965 
15966 	if ((pred = ep->dted_pred.dtpdd_predicate) != NULL) {
15967 		ASSERT(pred->dtp_difo != NULL);
15968 		dtrace_difo_hold(pred->dtp_difo);
15969 		helper->dtha_predicate = pred->dtp_difo;
15970 	}
15971 
15972 	for (act = ep->dted_action; act != NULL; act = act->dtad_next) {
15973 		if (act->dtad_kind != DTRACEACT_DIFEXPR)
15974 			goto err;
15975 
15976 		if (act->dtad_difo == NULL)
15977 			goto err;
15978 
15979 		nactions++;
15980 	}
15981 
15982 	helper->dtha_actions = kmem_zalloc(sizeof (dtrace_difo_t *) *
15983 	    (helper->dtha_nactions = nactions), KM_SLEEP);
15984 
15985 	for (act = ep->dted_action, i = 0; act != NULL; act = act->dtad_next) {
15986 		dtrace_difo_hold(act->dtad_difo);
15987 		helper->dtha_actions[i++] = act->dtad_difo;
15988 	}
15989 
15990 	if (!dtrace_helper_validate(helper))
15991 		goto err;
15992 
15993 	if (last == NULL) {
15994 		help->dthps_actions[which] = helper;
15995 	} else {
15996 		last->dtha_next = helper;
15997 	}
15998 
15999 	if (vstate->dtvs_nlocals > dtrace_helptrace_nlocals) {
16000 		dtrace_helptrace_nlocals = vstate->dtvs_nlocals;
16001 		dtrace_helptrace_next = 0;
16002 	}
16003 
16004 	return (0);
16005 err:
16006 	dtrace_helper_action_destroy(helper, vstate);
16007 	return (EINVAL);
16008 }
16009 
16010 static void
16011 dtrace_helper_provider_register(proc_t *p, dtrace_helpers_t *help,
16012     dof_helper_t *dofhp)
16013 {
16014 	ASSERT(MUTEX_NOT_HELD(&dtrace_lock));
16015 
16016 	mutex_enter(&dtrace_meta_lock);
16017 	mutex_enter(&dtrace_lock);
16018 
16019 	if (!dtrace_attached() || dtrace_meta_pid == NULL) {
16020 		/*
16021 		 * If the dtrace module is loaded but not attached, or if
16022 		 * there aren't isn't a meta provider registered to deal with
16023 		 * these provider descriptions, we need to postpone creating
16024 		 * the actual providers until later.
16025 		 */
16026 
16027 		if (help->dthps_next == NULL && help->dthps_prev == NULL &&
16028 		    dtrace_deferred_pid != help) {
16029 			help->dthps_deferred = 1;
16030 			help->dthps_pid = p->p_pid;
16031 			help->dthps_next = dtrace_deferred_pid;
16032 			help->dthps_prev = NULL;
16033 			if (dtrace_deferred_pid != NULL)
16034 				dtrace_deferred_pid->dthps_prev = help;
16035 			dtrace_deferred_pid = help;
16036 		}
16037 
16038 		mutex_exit(&dtrace_lock);
16039 
16040 	} else if (dofhp != NULL) {
16041 		/*
16042 		 * If the dtrace module is loaded and we have a particular
16043 		 * helper provider description, pass that off to the
16044 		 * meta provider.
16045 		 */
16046 
16047 		mutex_exit(&dtrace_lock);
16048 
16049 		dtrace_helper_provide(dofhp, p->p_pid);
16050 
16051 	} else {
16052 		/*
16053 		 * Otherwise, just pass all the helper provider descriptions
16054 		 * off to the meta provider.
16055 		 */
16056 
16057 		int i;
16058 		mutex_exit(&dtrace_lock);
16059 
16060 		for (i = 0; i < help->dthps_nprovs; i++) {
16061 			dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
16062 			    p->p_pid);
16063 		}
16064 	}
16065 
16066 	mutex_exit(&dtrace_meta_lock);
16067 }
16068 
16069 static int
16070 dtrace_helper_provider_add(dof_helper_t *dofhp, dtrace_helpers_t *help, int gen)
16071 {
16072 	dtrace_helper_provider_t *hprov, **tmp_provs;
16073 	uint_t tmp_maxprovs, i;
16074 
16075 	ASSERT(MUTEX_HELD(&dtrace_lock));
16076 	ASSERT(help != NULL);
16077 
16078 	/*
16079 	 * If we already have dtrace_helper_providers_max helper providers,
16080 	 * we're refuse to add a new one.
16081 	 */
16082 	if (help->dthps_nprovs >= dtrace_helper_providers_max)
16083 		return (ENOSPC);
16084 
16085 	/*
16086 	 * Check to make sure this isn't a duplicate.
16087 	 */
16088 	for (i = 0; i < help->dthps_nprovs; i++) {
16089 		if (dofhp->dofhp_addr ==
16090 		    help->dthps_provs[i]->dthp_prov.dofhp_addr)
16091 			return (EALREADY);
16092 	}
16093 
16094 	hprov = kmem_zalloc(sizeof (dtrace_helper_provider_t), KM_SLEEP);
16095 	hprov->dthp_prov = *dofhp;
16096 	hprov->dthp_ref = 1;
16097 	hprov->dthp_generation = gen;
16098 
16099 	/*
16100 	 * Allocate a bigger table for helper providers if it's already full.
16101 	 */
16102 	if (help->dthps_maxprovs == help->dthps_nprovs) {
16103 		tmp_maxprovs = help->dthps_maxprovs;
16104 		tmp_provs = help->dthps_provs;
16105 
16106 		if (help->dthps_maxprovs == 0)
16107 			help->dthps_maxprovs = 2;
16108 		else
16109 			help->dthps_maxprovs *= 2;
16110 		if (help->dthps_maxprovs > dtrace_helper_providers_max)
16111 			help->dthps_maxprovs = dtrace_helper_providers_max;
16112 
16113 		ASSERT(tmp_maxprovs < help->dthps_maxprovs);
16114 
16115 		help->dthps_provs = kmem_zalloc(help->dthps_maxprovs *
16116 		    sizeof (dtrace_helper_provider_t *), KM_SLEEP);
16117 
16118 		if (tmp_provs != NULL) {
16119 			bcopy(tmp_provs, help->dthps_provs, tmp_maxprovs *
16120 			    sizeof (dtrace_helper_provider_t *));
16121 			kmem_free(tmp_provs, tmp_maxprovs *
16122 			    sizeof (dtrace_helper_provider_t *));
16123 		}
16124 	}
16125 
16126 	help->dthps_provs[help->dthps_nprovs] = hprov;
16127 	help->dthps_nprovs++;
16128 
16129 	return (0);
16130 }
16131 
16132 static void
16133 dtrace_helper_provider_destroy(dtrace_helper_provider_t *hprov)
16134 {
16135 	mutex_enter(&dtrace_lock);
16136 
16137 	if (--hprov->dthp_ref == 0) {
16138 		dof_hdr_t *dof;
16139 		mutex_exit(&dtrace_lock);
16140 		dof = (dof_hdr_t *)(uintptr_t)hprov->dthp_prov.dofhp_dof;
16141 		dtrace_dof_destroy(dof);
16142 		kmem_free(hprov, sizeof (dtrace_helper_provider_t));
16143 	} else {
16144 		mutex_exit(&dtrace_lock);
16145 	}
16146 }
16147 
16148 static int
16149 dtrace_helper_provider_validate(dof_hdr_t *dof, dof_sec_t *sec)
16150 {
16151 	uintptr_t daddr = (uintptr_t)dof;
16152 	dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
16153 	dof_provider_t *provider;
16154 	dof_probe_t *probe;
16155 	uint8_t *arg;
16156 	char *strtab, *typestr;
16157 	dof_stridx_t typeidx;
16158 	size_t typesz;
16159 	uint_t nprobes, j, k;
16160 
16161 	ASSERT(sec->dofs_type == DOF_SECT_PROVIDER);
16162 
16163 	if (sec->dofs_offset & (sizeof (uint_t) - 1)) {
16164 		dtrace_dof_error(dof, "misaligned section offset");
16165 		return (-1);
16166 	}
16167 
16168 	/*
16169 	 * The section needs to be large enough to contain the DOF provider
16170 	 * structure appropriate for the given version.
16171 	 */
16172 	if (sec->dofs_size <
16173 	    ((dof->dofh_ident[DOF_ID_VERSION] == DOF_VERSION_1) ?
16174 	    offsetof(dof_provider_t, dofpv_prenoffs) :
16175 	    sizeof (dof_provider_t))) {
16176 		dtrace_dof_error(dof, "provider section too small");
16177 		return (-1);
16178 	}
16179 
16180 	provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
16181 	str_sec = dtrace_dof_sect(dof, DOF_SECT_STRTAB, provider->dofpv_strtab);
16182 	prb_sec = dtrace_dof_sect(dof, DOF_SECT_PROBES, provider->dofpv_probes);
16183 	arg_sec = dtrace_dof_sect(dof, DOF_SECT_PRARGS, provider->dofpv_prargs);
16184 	off_sec = dtrace_dof_sect(dof, DOF_SECT_PROFFS, provider->dofpv_proffs);
16185 
16186 	if (str_sec == NULL || prb_sec == NULL ||
16187 	    arg_sec == NULL || off_sec == NULL)
16188 		return (-1);
16189 
16190 	enoff_sec = NULL;
16191 
16192 	if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
16193 	    provider->dofpv_prenoffs != DOF_SECT_NONE &&
16194 	    (enoff_sec = dtrace_dof_sect(dof, DOF_SECT_PRENOFFS,
16195 	    provider->dofpv_prenoffs)) == NULL)
16196 		return (-1);
16197 
16198 	strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
16199 
16200 	if (provider->dofpv_name >= str_sec->dofs_size ||
16201 	    strlen(strtab + provider->dofpv_name) >= DTRACE_PROVNAMELEN) {
16202 		dtrace_dof_error(dof, "invalid provider name");
16203 		return (-1);
16204 	}
16205 
16206 	if (prb_sec->dofs_entsize == 0 ||
16207 	    prb_sec->dofs_entsize > prb_sec->dofs_size) {
16208 		dtrace_dof_error(dof, "invalid entry size");
16209 		return (-1);
16210 	}
16211 
16212 	if (prb_sec->dofs_entsize & (sizeof (uintptr_t) - 1)) {
16213 		dtrace_dof_error(dof, "misaligned entry size");
16214 		return (-1);
16215 	}
16216 
16217 	if (off_sec->dofs_entsize != sizeof (uint32_t)) {
16218 		dtrace_dof_error(dof, "invalid entry size");
16219 		return (-1);
16220 	}
16221 
16222 	if (off_sec->dofs_offset & (sizeof (uint32_t) - 1)) {
16223 		dtrace_dof_error(dof, "misaligned section offset");
16224 		return (-1);
16225 	}
16226 
16227 	if (arg_sec->dofs_entsize != sizeof (uint8_t)) {
16228 		dtrace_dof_error(dof, "invalid entry size");
16229 		return (-1);
16230 	}
16231 
16232 	arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
16233 
16234 	nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
16235 
16236 	/*
16237 	 * Take a pass through the probes to check for errors.
16238 	 */
16239 	for (j = 0; j < nprobes; j++) {
16240 		probe = (dof_probe_t *)(uintptr_t)(daddr +
16241 		    prb_sec->dofs_offset + j * prb_sec->dofs_entsize);
16242 
16243 		if (probe->dofpr_func >= str_sec->dofs_size) {
16244 			dtrace_dof_error(dof, "invalid function name");
16245 			return (-1);
16246 		}
16247 
16248 		if (strlen(strtab + probe->dofpr_func) >= DTRACE_FUNCNAMELEN) {
16249 			dtrace_dof_error(dof, "function name too long");
16250 			/*
16251 			 * Keep going if the function name is too long.
16252 			 * Unlike provider and probe names, we cannot reasonably
16253 			 * impose restrictions on function names, since they're
16254 			 * a property of the code being instrumented. We will
16255 			 * skip this probe in dtrace_helper_provide_one().
16256 			 */
16257 		}
16258 
16259 		if (probe->dofpr_name >= str_sec->dofs_size ||
16260 		    strlen(strtab + probe->dofpr_name) >= DTRACE_NAMELEN) {
16261 			dtrace_dof_error(dof, "invalid probe name");
16262 			return (-1);
16263 		}
16264 
16265 		/*
16266 		 * The offset count must not wrap the index, and the offsets
16267 		 * must also not overflow the section's data.
16268 		 */
16269 		if (probe->dofpr_offidx + probe->dofpr_noffs <
16270 		    probe->dofpr_offidx ||
16271 		    (probe->dofpr_offidx + probe->dofpr_noffs) *
16272 		    off_sec->dofs_entsize > off_sec->dofs_size) {
16273 			dtrace_dof_error(dof, "invalid probe offset");
16274 			return (-1);
16275 		}
16276 
16277 		if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1) {
16278 			/*
16279 			 * If there's no is-enabled offset section, make sure
16280 			 * there aren't any is-enabled offsets. Otherwise
16281 			 * perform the same checks as for probe offsets
16282 			 * (immediately above).
16283 			 */
16284 			if (enoff_sec == NULL) {
16285 				if (probe->dofpr_enoffidx != 0 ||
16286 				    probe->dofpr_nenoffs != 0) {
16287 					dtrace_dof_error(dof, "is-enabled "
16288 					    "offsets with null section");
16289 					return (-1);
16290 				}
16291 			} else if (probe->dofpr_enoffidx +
16292 			    probe->dofpr_nenoffs < probe->dofpr_enoffidx ||
16293 			    (probe->dofpr_enoffidx + probe->dofpr_nenoffs) *
16294 			    enoff_sec->dofs_entsize > enoff_sec->dofs_size) {
16295 				dtrace_dof_error(dof, "invalid is-enabled "
16296 				    "offset");
16297 				return (-1);
16298 			}
16299 
16300 			if (probe->dofpr_noffs + probe->dofpr_nenoffs == 0) {
16301 				dtrace_dof_error(dof, "zero probe and "
16302 				    "is-enabled offsets");
16303 				return (-1);
16304 			}
16305 		} else if (probe->dofpr_noffs == 0) {
16306 			dtrace_dof_error(dof, "zero probe offsets");
16307 			return (-1);
16308 		}
16309 
16310 		if (probe->dofpr_argidx + probe->dofpr_xargc <
16311 		    probe->dofpr_argidx ||
16312 		    (probe->dofpr_argidx + probe->dofpr_xargc) *
16313 		    arg_sec->dofs_entsize > arg_sec->dofs_size) {
16314 			dtrace_dof_error(dof, "invalid args");
16315 			return (-1);
16316 		}
16317 
16318 		typeidx = probe->dofpr_nargv;
16319 		typestr = strtab + probe->dofpr_nargv;
16320 		for (k = 0; k < probe->dofpr_nargc; k++) {
16321 			if (typeidx >= str_sec->dofs_size) {
16322 				dtrace_dof_error(dof, "bad "
16323 				    "native argument type");
16324 				return (-1);
16325 			}
16326 
16327 			typesz = strlen(typestr) + 1;
16328 			if (typesz > DTRACE_ARGTYPELEN) {
16329 				dtrace_dof_error(dof, "native "
16330 				    "argument type too long");
16331 				return (-1);
16332 			}
16333 			typeidx += typesz;
16334 			typestr += typesz;
16335 		}
16336 
16337 		typeidx = probe->dofpr_xargv;
16338 		typestr = strtab + probe->dofpr_xargv;
16339 		for (k = 0; k < probe->dofpr_xargc; k++) {
16340 			if (arg[probe->dofpr_argidx + k] > probe->dofpr_nargc) {
16341 				dtrace_dof_error(dof, "bad "
16342 				    "native argument index");
16343 				return (-1);
16344 			}
16345 
16346 			if (typeidx >= str_sec->dofs_size) {
16347 				dtrace_dof_error(dof, "bad "
16348 				    "translated argument type");
16349 				return (-1);
16350 			}
16351 
16352 			typesz = strlen(typestr) + 1;
16353 			if (typesz > DTRACE_ARGTYPELEN) {
16354 				dtrace_dof_error(dof, "translated argument "
16355 				    "type too long");
16356 				return (-1);
16357 			}
16358 
16359 			typeidx += typesz;
16360 			typestr += typesz;
16361 		}
16362 	}
16363 
16364 	return (0);
16365 }
16366 
16367 static int
16368 dtrace_helper_slurp(dof_hdr_t *dof, dof_helper_t *dhp, struct proc *p)
16369 {
16370 	dtrace_helpers_t *help;
16371 	dtrace_vstate_t *vstate;
16372 	dtrace_enabling_t *enab = NULL;
16373 	int i, gen, rv, nhelpers = 0, nprovs = 0, destroy = 1;
16374 	uintptr_t daddr = (uintptr_t)dof;
16375 
16376 	ASSERT(MUTEX_HELD(&dtrace_lock));
16377 
16378 	if ((help = p->p_dtrace_helpers) == NULL)
16379 		help = dtrace_helpers_create(p);
16380 
16381 	vstate = &help->dthps_vstate;
16382 
16383 	if ((rv = dtrace_dof_slurp(dof, vstate, NULL, &enab, dhp->dofhp_addr,
16384 	    dhp->dofhp_dof, B_FALSE)) != 0) {
16385 		dtrace_dof_destroy(dof);
16386 		return (rv);
16387 	}
16388 
16389 	/*
16390 	 * Look for helper providers and validate their descriptions.
16391 	 */
16392 	for (i = 0; i < dof->dofh_secnum; i++) {
16393 		dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
16394 		    dof->dofh_secoff + i * dof->dofh_secsize);
16395 
16396 		if (sec->dofs_type != DOF_SECT_PROVIDER)
16397 			continue;
16398 
16399 		if (dtrace_helper_provider_validate(dof, sec) != 0) {
16400 			dtrace_enabling_destroy(enab);
16401 			dtrace_dof_destroy(dof);
16402 			return (-1);
16403 		}
16404 
16405 		nprovs++;
16406 	}
16407 
16408 	/*
16409 	 * Now we need to walk through the ECB descriptions in the enabling.
16410 	 */
16411 	for (i = 0; i < enab->dten_ndesc; i++) {
16412 		dtrace_ecbdesc_t *ep = enab->dten_desc[i];
16413 		dtrace_probedesc_t *desc = &ep->dted_probe;
16414 
16415 		if (strcmp(desc->dtpd_provider, "dtrace") != 0)
16416 			continue;
16417 
16418 		if (strcmp(desc->dtpd_mod, "helper") != 0)
16419 			continue;
16420 
16421 		if (strcmp(desc->dtpd_func, "ustack") != 0)
16422 			continue;
16423 
16424 		if ((rv = dtrace_helper_action_add(DTRACE_HELPER_ACTION_USTACK,
16425 		    ep, help)) != 0) {
16426 			/*
16427 			 * Adding this helper action failed -- we are now going
16428 			 * to rip out the entire generation and return failure.
16429 			 */
16430 			(void) dtrace_helper_destroygen(help,
16431 			    help->dthps_generation);
16432 			dtrace_enabling_destroy(enab);
16433 			dtrace_dof_destroy(dof);
16434 			return (-1);
16435 		}
16436 
16437 		nhelpers++;
16438 	}
16439 
16440 	if (nhelpers < enab->dten_ndesc)
16441 		dtrace_dof_error(dof, "unmatched helpers");
16442 
16443 	gen = help->dthps_generation++;
16444 	dtrace_enabling_destroy(enab);
16445 
16446 	if (nprovs > 0) {
16447 		/*
16448 		 * Now that this is in-kernel, we change the sense of the
16449 		 * members:  dofhp_dof denotes the in-kernel copy of the DOF
16450 		 * and dofhp_addr denotes the address at user-level.
16451 		 */
16452 		dhp->dofhp_addr = dhp->dofhp_dof;
16453 		dhp->dofhp_dof = (uint64_t)(uintptr_t)dof;
16454 
16455 		if (dtrace_helper_provider_add(dhp, help, gen) == 0) {
16456 			mutex_exit(&dtrace_lock);
16457 			dtrace_helper_provider_register(p, help, dhp);
16458 			mutex_enter(&dtrace_lock);
16459 
16460 			destroy = 0;
16461 		}
16462 	}
16463 
16464 	if (destroy)
16465 		dtrace_dof_destroy(dof);
16466 
16467 	return (gen);
16468 }
16469 
16470 static dtrace_helpers_t *
16471 dtrace_helpers_create(proc_t *p)
16472 {
16473 	dtrace_helpers_t *help;
16474 
16475 	ASSERT(MUTEX_HELD(&dtrace_lock));
16476 	ASSERT(p->p_dtrace_helpers == NULL);
16477 
16478 	help = kmem_zalloc(sizeof (dtrace_helpers_t), KM_SLEEP);
16479 	help->dthps_actions = kmem_zalloc(sizeof (dtrace_helper_action_t *) *
16480 	    DTRACE_NHELPER_ACTIONS, KM_SLEEP);
16481 
16482 	p->p_dtrace_helpers = help;
16483 	dtrace_helpers++;
16484 
16485 	return (help);
16486 }
16487 
16488 #ifdef illumos
16489 static
16490 #endif
16491 void
16492 dtrace_helpers_destroy(proc_t *p)
16493 {
16494 	dtrace_helpers_t *help;
16495 	dtrace_vstate_t *vstate;
16496 #ifdef illumos
16497 	proc_t *p = curproc;
16498 #endif
16499 	int i;
16500 
16501 	mutex_enter(&dtrace_lock);
16502 
16503 	ASSERT(p->p_dtrace_helpers != NULL);
16504 	ASSERT(dtrace_helpers > 0);
16505 
16506 	help = p->p_dtrace_helpers;
16507 	vstate = &help->dthps_vstate;
16508 
16509 	/*
16510 	 * We're now going to lose the help from this process.
16511 	 */
16512 	p->p_dtrace_helpers = NULL;
16513 	dtrace_sync();
16514 
16515 	/*
16516 	 * Destory the helper actions.
16517 	 */
16518 	for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
16519 		dtrace_helper_action_t *h, *next;
16520 
16521 		for (h = help->dthps_actions[i]; h != NULL; h = next) {
16522 			next = h->dtha_next;
16523 			dtrace_helper_action_destroy(h, vstate);
16524 			h = next;
16525 		}
16526 	}
16527 
16528 	mutex_exit(&dtrace_lock);
16529 
16530 	/*
16531 	 * Destroy the helper providers.
16532 	 */
16533 	if (help->dthps_maxprovs > 0) {
16534 		mutex_enter(&dtrace_meta_lock);
16535 		if (dtrace_meta_pid != NULL) {
16536 			ASSERT(dtrace_deferred_pid == NULL);
16537 
16538 			for (i = 0; i < help->dthps_nprovs; i++) {
16539 				dtrace_helper_provider_remove(
16540 				    &help->dthps_provs[i]->dthp_prov, p->p_pid);
16541 			}
16542 		} else {
16543 			mutex_enter(&dtrace_lock);
16544 			ASSERT(help->dthps_deferred == 0 ||
16545 			    help->dthps_next != NULL ||
16546 			    help->dthps_prev != NULL ||
16547 			    help == dtrace_deferred_pid);
16548 
16549 			/*
16550 			 * Remove the helper from the deferred list.
16551 			 */
16552 			if (help->dthps_next != NULL)
16553 				help->dthps_next->dthps_prev = help->dthps_prev;
16554 			if (help->dthps_prev != NULL)
16555 				help->dthps_prev->dthps_next = help->dthps_next;
16556 			if (dtrace_deferred_pid == help) {
16557 				dtrace_deferred_pid = help->dthps_next;
16558 				ASSERT(help->dthps_prev == NULL);
16559 			}
16560 
16561 			mutex_exit(&dtrace_lock);
16562 		}
16563 
16564 		mutex_exit(&dtrace_meta_lock);
16565 
16566 		for (i = 0; i < help->dthps_nprovs; i++) {
16567 			dtrace_helper_provider_destroy(help->dthps_provs[i]);
16568 		}
16569 
16570 		kmem_free(help->dthps_provs, help->dthps_maxprovs *
16571 		    sizeof (dtrace_helper_provider_t *));
16572 	}
16573 
16574 	mutex_enter(&dtrace_lock);
16575 
16576 	dtrace_vstate_fini(&help->dthps_vstate);
16577 	kmem_free(help->dthps_actions,
16578 	    sizeof (dtrace_helper_action_t *) * DTRACE_NHELPER_ACTIONS);
16579 	kmem_free(help, sizeof (dtrace_helpers_t));
16580 
16581 	--dtrace_helpers;
16582 	mutex_exit(&dtrace_lock);
16583 }
16584 
16585 #ifdef illumos
16586 static
16587 #endif
16588 void
16589 dtrace_helpers_duplicate(proc_t *from, proc_t *to)
16590 {
16591 	dtrace_helpers_t *help, *newhelp;
16592 	dtrace_helper_action_t *helper, *new, *last;
16593 	dtrace_difo_t *dp;
16594 	dtrace_vstate_t *vstate;
16595 	int i, j, sz, hasprovs = 0;
16596 
16597 	mutex_enter(&dtrace_lock);
16598 	ASSERT(from->p_dtrace_helpers != NULL);
16599 	ASSERT(dtrace_helpers > 0);
16600 
16601 	help = from->p_dtrace_helpers;
16602 	newhelp = dtrace_helpers_create(to);
16603 	ASSERT(to->p_dtrace_helpers != NULL);
16604 
16605 	newhelp->dthps_generation = help->dthps_generation;
16606 	vstate = &newhelp->dthps_vstate;
16607 
16608 	/*
16609 	 * Duplicate the helper actions.
16610 	 */
16611 	for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
16612 		if ((helper = help->dthps_actions[i]) == NULL)
16613 			continue;
16614 
16615 		for (last = NULL; helper != NULL; helper = helper->dtha_next) {
16616 			new = kmem_zalloc(sizeof (dtrace_helper_action_t),
16617 			    KM_SLEEP);
16618 			new->dtha_generation = helper->dtha_generation;
16619 
16620 			if ((dp = helper->dtha_predicate) != NULL) {
16621 				dp = dtrace_difo_duplicate(dp, vstate);
16622 				new->dtha_predicate = dp;
16623 			}
16624 
16625 			new->dtha_nactions = helper->dtha_nactions;
16626 			sz = sizeof (dtrace_difo_t *) * new->dtha_nactions;
16627 			new->dtha_actions = kmem_alloc(sz, KM_SLEEP);
16628 
16629 			for (j = 0; j < new->dtha_nactions; j++) {
16630 				dtrace_difo_t *dp = helper->dtha_actions[j];
16631 
16632 				ASSERT(dp != NULL);
16633 				dp = dtrace_difo_duplicate(dp, vstate);
16634 				new->dtha_actions[j] = dp;
16635 			}
16636 
16637 			if (last != NULL) {
16638 				last->dtha_next = new;
16639 			} else {
16640 				newhelp->dthps_actions[i] = new;
16641 			}
16642 
16643 			last = new;
16644 		}
16645 	}
16646 
16647 	/*
16648 	 * Duplicate the helper providers and register them with the
16649 	 * DTrace framework.
16650 	 */
16651 	if (help->dthps_nprovs > 0) {
16652 		newhelp->dthps_nprovs = help->dthps_nprovs;
16653 		newhelp->dthps_maxprovs = help->dthps_nprovs;
16654 		newhelp->dthps_provs = kmem_alloc(newhelp->dthps_nprovs *
16655 		    sizeof (dtrace_helper_provider_t *), KM_SLEEP);
16656 		for (i = 0; i < newhelp->dthps_nprovs; i++) {
16657 			newhelp->dthps_provs[i] = help->dthps_provs[i];
16658 			newhelp->dthps_provs[i]->dthp_ref++;
16659 		}
16660 
16661 		hasprovs = 1;
16662 	}
16663 
16664 	mutex_exit(&dtrace_lock);
16665 
16666 	if (hasprovs)
16667 		dtrace_helper_provider_register(to, newhelp, NULL);
16668 }
16669 
16670 /*
16671  * DTrace Hook Functions
16672  */
16673 static void
16674 dtrace_module_loaded(modctl_t *ctl)
16675 {
16676 	dtrace_provider_t *prv;
16677 
16678 	mutex_enter(&dtrace_provider_lock);
16679 #ifdef illumos
16680 	mutex_enter(&mod_lock);
16681 #endif
16682 
16683 #ifdef illumos
16684 	ASSERT(ctl->mod_busy);
16685 #endif
16686 
16687 	/*
16688 	 * We're going to call each providers per-module provide operation
16689 	 * specifying only this module.
16690 	 */
16691 	for (prv = dtrace_provider; prv != NULL; prv = prv->dtpv_next)
16692 		prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
16693 
16694 #ifdef illumos
16695 	mutex_exit(&mod_lock);
16696 #endif
16697 	mutex_exit(&dtrace_provider_lock);
16698 
16699 	/*
16700 	 * If we have any retained enablings, we need to match against them.
16701 	 * Enabling probes requires that cpu_lock be held, and we cannot hold
16702 	 * cpu_lock here -- it is legal for cpu_lock to be held when loading a
16703 	 * module.  (In particular, this happens when loading scheduling
16704 	 * classes.)  So if we have any retained enablings, we need to dispatch
16705 	 * our task queue to do the match for us.
16706 	 */
16707 	mutex_enter(&dtrace_lock);
16708 
16709 	if (dtrace_retained == NULL) {
16710 		mutex_exit(&dtrace_lock);
16711 		return;
16712 	}
16713 
16714 	(void) taskq_dispatch(dtrace_taskq,
16715 	    (task_func_t *)dtrace_enabling_matchall, NULL, TQ_SLEEP);
16716 
16717 	mutex_exit(&dtrace_lock);
16718 
16719 	/*
16720 	 * And now, for a little heuristic sleaze:  in general, we want to
16721 	 * match modules as soon as they load.  However, we cannot guarantee
16722 	 * this, because it would lead us to the lock ordering violation
16723 	 * outlined above.  The common case, of course, is that cpu_lock is
16724 	 * _not_ held -- so we delay here for a clock tick, hoping that that's
16725 	 * long enough for the task queue to do its work.  If it's not, it's
16726 	 * not a serious problem -- it just means that the module that we
16727 	 * just loaded may not be immediately instrumentable.
16728 	 */
16729 	delay(1);
16730 }
16731 
16732 static void
16733 #ifdef illumos
16734 dtrace_module_unloaded(modctl_t *ctl)
16735 #else
16736 dtrace_module_unloaded(modctl_t *ctl, int *error)
16737 #endif
16738 {
16739 	dtrace_probe_t template, *probe, *first, *next;
16740 	dtrace_provider_t *prov;
16741 #ifndef illumos
16742 	char modname[DTRACE_MODNAMELEN];
16743 	size_t len;
16744 #endif
16745 
16746 #ifdef illumos
16747 	template.dtpr_mod = ctl->mod_modname;
16748 #else
16749 	/* Handle the fact that ctl->filename may end in ".ko". */
16750 	strlcpy(modname, ctl->filename, sizeof(modname));
16751 	len = strlen(ctl->filename);
16752 	if (len > 3 && strcmp(modname + len - 3, ".ko") == 0)
16753 		modname[len - 3] = '\0';
16754 	template.dtpr_mod = modname;
16755 #endif
16756 
16757 	mutex_enter(&dtrace_provider_lock);
16758 #ifdef illumos
16759 	mutex_enter(&mod_lock);
16760 #endif
16761 	mutex_enter(&dtrace_lock);
16762 
16763 #ifndef illumos
16764 	if (ctl->nenabled > 0) {
16765 		/* Don't allow unloads if a probe is enabled. */
16766 		mutex_exit(&dtrace_provider_lock);
16767 		mutex_exit(&dtrace_lock);
16768 		*error = -1;
16769 		printf(
16770 	"kldunload: attempt to unload module that has DTrace probes enabled\n");
16771 		return;
16772 	}
16773 #endif
16774 
16775 	if (dtrace_bymod == NULL) {
16776 		/*
16777 		 * The DTrace module is loaded (obviously) but not attached;
16778 		 * we don't have any work to do.
16779 		 */
16780 		mutex_exit(&dtrace_provider_lock);
16781 #ifdef illumos
16782 		mutex_exit(&mod_lock);
16783 #endif
16784 		mutex_exit(&dtrace_lock);
16785 		return;
16786 	}
16787 
16788 	for (probe = first = dtrace_hash_lookup(dtrace_bymod, &template);
16789 	    probe != NULL; probe = probe->dtpr_nextmod) {
16790 		if (probe->dtpr_ecb != NULL) {
16791 			mutex_exit(&dtrace_provider_lock);
16792 #ifdef illumos
16793 			mutex_exit(&mod_lock);
16794 #endif
16795 			mutex_exit(&dtrace_lock);
16796 
16797 			/*
16798 			 * This shouldn't _actually_ be possible -- we're
16799 			 * unloading a module that has an enabled probe in it.
16800 			 * (It's normally up to the provider to make sure that
16801 			 * this can't happen.)  However, because dtps_enable()
16802 			 * doesn't have a failure mode, there can be an
16803 			 * enable/unload race.  Upshot:  we don't want to
16804 			 * assert, but we're not going to disable the
16805 			 * probe, either.
16806 			 */
16807 			if (dtrace_err_verbose) {
16808 #ifdef illumos
16809 				cmn_err(CE_WARN, "unloaded module '%s' had "
16810 				    "enabled probes", ctl->mod_modname);
16811 #else
16812 				cmn_err(CE_WARN, "unloaded module '%s' had "
16813 				    "enabled probes", modname);
16814 #endif
16815 			}
16816 
16817 			return;
16818 		}
16819 	}
16820 
16821 	probe = first;
16822 
16823 	for (first = NULL; probe != NULL; probe = next) {
16824 		ASSERT(dtrace_probes[probe->dtpr_id - 1] == probe);
16825 
16826 		dtrace_probes[probe->dtpr_id - 1] = NULL;
16827 
16828 		next = probe->dtpr_nextmod;
16829 		dtrace_hash_remove(dtrace_bymod, probe);
16830 		dtrace_hash_remove(dtrace_byfunc, probe);
16831 		dtrace_hash_remove(dtrace_byname, probe);
16832 
16833 		if (first == NULL) {
16834 			first = probe;
16835 			probe->dtpr_nextmod = NULL;
16836 		} else {
16837 			probe->dtpr_nextmod = first;
16838 			first = probe;
16839 		}
16840 	}
16841 
16842 	/*
16843 	 * We've removed all of the module's probes from the hash chains and
16844 	 * from the probe array.  Now issue a dtrace_sync() to be sure that
16845 	 * everyone has cleared out from any probe array processing.
16846 	 */
16847 	dtrace_sync();
16848 
16849 	for (probe = first; probe != NULL; probe = first) {
16850 		first = probe->dtpr_nextmod;
16851 		prov = probe->dtpr_provider;
16852 		prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, probe->dtpr_id,
16853 		    probe->dtpr_arg);
16854 		kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
16855 		kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
16856 		kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
16857 #ifdef illumos
16858 		vmem_free(dtrace_arena, (void *)(uintptr_t)probe->dtpr_id, 1);
16859 #else
16860 		free_unr(dtrace_arena, probe->dtpr_id);
16861 #endif
16862 		kmem_free(probe, sizeof (dtrace_probe_t));
16863 	}
16864 
16865 	mutex_exit(&dtrace_lock);
16866 #ifdef illumos
16867 	mutex_exit(&mod_lock);
16868 #endif
16869 	mutex_exit(&dtrace_provider_lock);
16870 }
16871 
16872 #ifndef illumos
16873 static void
16874 dtrace_kld_load(void *arg __unused, linker_file_t lf)
16875 {
16876 
16877 	dtrace_module_loaded(lf);
16878 }
16879 
16880 static void
16881 dtrace_kld_unload_try(void *arg __unused, linker_file_t lf, int *error)
16882 {
16883 
16884 	if (*error != 0)
16885 		/* We already have an error, so don't do anything. */
16886 		return;
16887 	dtrace_module_unloaded(lf, error);
16888 }
16889 #endif
16890 
16891 #ifdef illumos
16892 static void
16893 dtrace_suspend(void)
16894 {
16895 	dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_suspend));
16896 }
16897 
16898 static void
16899 dtrace_resume(void)
16900 {
16901 	dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_resume));
16902 }
16903 #endif
16904 
16905 static int
16906 dtrace_cpu_setup(cpu_setup_t what, processorid_t cpu)
16907 {
16908 	ASSERT(MUTEX_HELD(&cpu_lock));
16909 	mutex_enter(&dtrace_lock);
16910 
16911 	switch (what) {
16912 	case CPU_CONFIG: {
16913 		dtrace_state_t *state;
16914 		dtrace_optval_t *opt, rs, c;
16915 
16916 		/*
16917 		 * For now, we only allocate a new buffer for anonymous state.
16918 		 */
16919 		if ((state = dtrace_anon.dta_state) == NULL)
16920 			break;
16921 
16922 		if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
16923 			break;
16924 
16925 		opt = state->dts_options;
16926 		c = opt[DTRACEOPT_CPU];
16927 
16928 		if (c != DTRACE_CPUALL && c != DTRACEOPT_UNSET && c != cpu)
16929 			break;
16930 
16931 		/*
16932 		 * Regardless of what the actual policy is, we're going to
16933 		 * temporarily set our resize policy to be manual.  We're
16934 		 * also going to temporarily set our CPU option to denote
16935 		 * the newly configured CPU.
16936 		 */
16937 		rs = opt[DTRACEOPT_BUFRESIZE];
16938 		opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_MANUAL;
16939 		opt[DTRACEOPT_CPU] = (dtrace_optval_t)cpu;
16940 
16941 		(void) dtrace_state_buffers(state);
16942 
16943 		opt[DTRACEOPT_BUFRESIZE] = rs;
16944 		opt[DTRACEOPT_CPU] = c;
16945 
16946 		break;
16947 	}
16948 
16949 	case CPU_UNCONFIG:
16950 		/*
16951 		 * We don't free the buffer in the CPU_UNCONFIG case.  (The
16952 		 * buffer will be freed when the consumer exits.)
16953 		 */
16954 		break;
16955 
16956 	default:
16957 		break;
16958 	}
16959 
16960 	mutex_exit(&dtrace_lock);
16961 	return (0);
16962 }
16963 
16964 #ifdef illumos
16965 static void
16966 dtrace_cpu_setup_initial(processorid_t cpu)
16967 {
16968 	(void) dtrace_cpu_setup(CPU_CONFIG, cpu);
16969 }
16970 #endif
16971 
16972 static void
16973 dtrace_toxrange_add(uintptr_t base, uintptr_t limit)
16974 {
16975 	if (dtrace_toxranges >= dtrace_toxranges_max) {
16976 		int osize, nsize;
16977 		dtrace_toxrange_t *range;
16978 
16979 		osize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
16980 
16981 		if (osize == 0) {
16982 			ASSERT(dtrace_toxrange == NULL);
16983 			ASSERT(dtrace_toxranges_max == 0);
16984 			dtrace_toxranges_max = 1;
16985 		} else {
16986 			dtrace_toxranges_max <<= 1;
16987 		}
16988 
16989 		nsize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
16990 		range = kmem_zalloc(nsize, KM_SLEEP);
16991 
16992 		if (dtrace_toxrange != NULL) {
16993 			ASSERT(osize != 0);
16994 			bcopy(dtrace_toxrange, range, osize);
16995 			kmem_free(dtrace_toxrange, osize);
16996 		}
16997 
16998 		dtrace_toxrange = range;
16999 	}
17000 
17001 	ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_base == 0);
17002 	ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_limit == 0);
17003 
17004 	dtrace_toxrange[dtrace_toxranges].dtt_base = base;
17005 	dtrace_toxrange[dtrace_toxranges].dtt_limit = limit;
17006 	dtrace_toxranges++;
17007 }
17008 
17009 static void
17010 dtrace_getf_barrier()
17011 {
17012 #ifdef illumos
17013 	/*
17014 	 * When we have unprivileged (that is, non-DTRACE_CRV_KERNEL) enablings
17015 	 * that contain calls to getf(), this routine will be called on every
17016 	 * closef() before either the underlying vnode is released or the
17017 	 * file_t itself is freed.  By the time we are here, it is essential
17018 	 * that the file_t can no longer be accessed from a call to getf()
17019 	 * in probe context -- that assures that a dtrace_sync() can be used
17020 	 * to clear out any enablings referring to the old structures.
17021 	 */
17022 	if (curthread->t_procp->p_zone->zone_dtrace_getf != 0 ||
17023 	    kcred->cr_zone->zone_dtrace_getf != 0)
17024 		dtrace_sync();
17025 #endif
17026 }
17027 
17028 /*
17029  * DTrace Driver Cookbook Functions
17030  */
17031 #ifdef illumos
17032 /*ARGSUSED*/
17033 static int
17034 dtrace_attach(dev_info_t *devi, ddi_attach_cmd_t cmd)
17035 {
17036 	dtrace_provider_id_t id;
17037 	dtrace_state_t *state = NULL;
17038 	dtrace_enabling_t *enab;
17039 
17040 	mutex_enter(&cpu_lock);
17041 	mutex_enter(&dtrace_provider_lock);
17042 	mutex_enter(&dtrace_lock);
17043 
17044 	if (ddi_soft_state_init(&dtrace_softstate,
17045 	    sizeof (dtrace_state_t), 0) != 0) {
17046 		cmn_err(CE_NOTE, "/dev/dtrace failed to initialize soft state");
17047 		mutex_exit(&cpu_lock);
17048 		mutex_exit(&dtrace_provider_lock);
17049 		mutex_exit(&dtrace_lock);
17050 		return (DDI_FAILURE);
17051 	}
17052 
17053 	if (ddi_create_minor_node(devi, DTRACEMNR_DTRACE, S_IFCHR,
17054 	    DTRACEMNRN_DTRACE, DDI_PSEUDO, NULL) == DDI_FAILURE ||
17055 	    ddi_create_minor_node(devi, DTRACEMNR_HELPER, S_IFCHR,
17056 	    DTRACEMNRN_HELPER, DDI_PSEUDO, NULL) == DDI_FAILURE) {
17057 		cmn_err(CE_NOTE, "/dev/dtrace couldn't create minor nodes");
17058 		ddi_remove_minor_node(devi, NULL);
17059 		ddi_soft_state_fini(&dtrace_softstate);
17060 		mutex_exit(&cpu_lock);
17061 		mutex_exit(&dtrace_provider_lock);
17062 		mutex_exit(&dtrace_lock);
17063 		return (DDI_FAILURE);
17064 	}
17065 
17066 	ddi_report_dev(devi);
17067 	dtrace_devi = devi;
17068 
17069 	dtrace_modload = dtrace_module_loaded;
17070 	dtrace_modunload = dtrace_module_unloaded;
17071 	dtrace_cpu_init = dtrace_cpu_setup_initial;
17072 	dtrace_helpers_cleanup = dtrace_helpers_destroy;
17073 	dtrace_helpers_fork = dtrace_helpers_duplicate;
17074 	dtrace_cpustart_init = dtrace_suspend;
17075 	dtrace_cpustart_fini = dtrace_resume;
17076 	dtrace_debugger_init = dtrace_suspend;
17077 	dtrace_debugger_fini = dtrace_resume;
17078 
17079 	register_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
17080 
17081 	ASSERT(MUTEX_HELD(&cpu_lock));
17082 
17083 	dtrace_arena = vmem_create("dtrace", (void *)1, UINT32_MAX, 1,
17084 	    NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
17085 	dtrace_minor = vmem_create("dtrace_minor", (void *)DTRACEMNRN_CLONE,
17086 	    UINT32_MAX - DTRACEMNRN_CLONE, 1, NULL, NULL, NULL, 0,
17087 	    VM_SLEEP | VMC_IDENTIFIER);
17088 	dtrace_taskq = taskq_create("dtrace_taskq", 1, maxclsyspri,
17089 	    1, INT_MAX, 0);
17090 
17091 	dtrace_state_cache = kmem_cache_create("dtrace_state_cache",
17092 	    sizeof (dtrace_dstate_percpu_t) * NCPU, DTRACE_STATE_ALIGN,
17093 	    NULL, NULL, NULL, NULL, NULL, 0);
17094 
17095 	ASSERT(MUTEX_HELD(&cpu_lock));
17096 	dtrace_bymod = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_mod),
17097 	    offsetof(dtrace_probe_t, dtpr_nextmod),
17098 	    offsetof(dtrace_probe_t, dtpr_prevmod));
17099 
17100 	dtrace_byfunc = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_func),
17101 	    offsetof(dtrace_probe_t, dtpr_nextfunc),
17102 	    offsetof(dtrace_probe_t, dtpr_prevfunc));
17103 
17104 	dtrace_byname = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_name),
17105 	    offsetof(dtrace_probe_t, dtpr_nextname),
17106 	    offsetof(dtrace_probe_t, dtpr_prevname));
17107 
17108 	if (dtrace_retain_max < 1) {
17109 		cmn_err(CE_WARN, "illegal value (%lu) for dtrace_retain_max; "
17110 		    "setting to 1", dtrace_retain_max);
17111 		dtrace_retain_max = 1;
17112 	}
17113 
17114 	/*
17115 	 * Now discover our toxic ranges.
17116 	 */
17117 	dtrace_toxic_ranges(dtrace_toxrange_add);
17118 
17119 	/*
17120 	 * Before we register ourselves as a provider to our own framework,
17121 	 * we would like to assert that dtrace_provider is NULL -- but that's
17122 	 * not true if we were loaded as a dependency of a DTrace provider.
17123 	 * Once we've registered, we can assert that dtrace_provider is our
17124 	 * pseudo provider.
17125 	 */
17126 	(void) dtrace_register("dtrace", &dtrace_provider_attr,
17127 	    DTRACE_PRIV_NONE, 0, &dtrace_provider_ops, NULL, &id);
17128 
17129 	ASSERT(dtrace_provider != NULL);
17130 	ASSERT((dtrace_provider_id_t)dtrace_provider == id);
17131 
17132 	dtrace_probeid_begin = dtrace_probe_create((dtrace_provider_id_t)
17133 	    dtrace_provider, NULL, NULL, "BEGIN", 0, NULL);
17134 	dtrace_probeid_end = dtrace_probe_create((dtrace_provider_id_t)
17135 	    dtrace_provider, NULL, NULL, "END", 0, NULL);
17136 	dtrace_probeid_error = dtrace_probe_create((dtrace_provider_id_t)
17137 	    dtrace_provider, NULL, NULL, "ERROR", 1, NULL);
17138 
17139 	dtrace_anon_property();
17140 	mutex_exit(&cpu_lock);
17141 
17142 	/*
17143 	 * If there are already providers, we must ask them to provide their
17144 	 * probes, and then match any anonymous enabling against them.  Note
17145 	 * that there should be no other retained enablings at this time:
17146 	 * the only retained enablings at this time should be the anonymous
17147 	 * enabling.
17148 	 */
17149 	if (dtrace_anon.dta_enabling != NULL) {
17150 		ASSERT(dtrace_retained == dtrace_anon.dta_enabling);
17151 
17152 		dtrace_enabling_provide(NULL);
17153 		state = dtrace_anon.dta_state;
17154 
17155 		/*
17156 		 * We couldn't hold cpu_lock across the above call to
17157 		 * dtrace_enabling_provide(), but we must hold it to actually
17158 		 * enable the probes.  We have to drop all of our locks, pick
17159 		 * up cpu_lock, and regain our locks before matching the
17160 		 * retained anonymous enabling.
17161 		 */
17162 		mutex_exit(&dtrace_lock);
17163 		mutex_exit(&dtrace_provider_lock);
17164 
17165 		mutex_enter(&cpu_lock);
17166 		mutex_enter(&dtrace_provider_lock);
17167 		mutex_enter(&dtrace_lock);
17168 
17169 		if ((enab = dtrace_anon.dta_enabling) != NULL)
17170 			(void) dtrace_enabling_match(enab, NULL);
17171 
17172 		mutex_exit(&cpu_lock);
17173 	}
17174 
17175 	mutex_exit(&dtrace_lock);
17176 	mutex_exit(&dtrace_provider_lock);
17177 
17178 	if (state != NULL) {
17179 		/*
17180 		 * If we created any anonymous state, set it going now.
17181 		 */
17182 		(void) dtrace_state_go(state, &dtrace_anon.dta_beganon);
17183 	}
17184 
17185 	return (DDI_SUCCESS);
17186 }
17187 #endif	/* illumos */
17188 
17189 #ifndef illumos
17190 static void dtrace_dtr(void *);
17191 #endif
17192 
17193 /*ARGSUSED*/
17194 static int
17195 #ifdef illumos
17196 dtrace_open(dev_t *devp, int flag, int otyp, cred_t *cred_p)
17197 #else
17198 dtrace_open(struct cdev *dev, int oflags, int devtype, struct thread *td)
17199 #endif
17200 {
17201 	dtrace_state_t *state;
17202 	uint32_t priv;
17203 	uid_t uid;
17204 	zoneid_t zoneid;
17205 
17206 #ifdef illumos
17207 	if (getminor(*devp) == DTRACEMNRN_HELPER)
17208 		return (0);
17209 
17210 	/*
17211 	 * If this wasn't an open with the "helper" minor, then it must be
17212 	 * the "dtrace" minor.
17213 	 */
17214 	if (getminor(*devp) == DTRACEMNRN_DTRACE)
17215 		return (ENXIO);
17216 #else
17217 	cred_t *cred_p = NULL;
17218 	cred_p = dev->si_cred;
17219 
17220 	/*
17221 	 * If no DTRACE_PRIV_* bits are set in the credential, then the
17222 	 * caller lacks sufficient permission to do anything with DTrace.
17223 	 */
17224 	dtrace_cred2priv(cred_p, &priv, &uid, &zoneid);
17225 	if (priv == DTRACE_PRIV_NONE) {
17226 #endif
17227 
17228 		return (EACCES);
17229 	}
17230 
17231 	/*
17232 	 * Ask all providers to provide all their probes.
17233 	 */
17234 	mutex_enter(&dtrace_provider_lock);
17235 	dtrace_probe_provide(NULL, NULL);
17236 	mutex_exit(&dtrace_provider_lock);
17237 
17238 	mutex_enter(&cpu_lock);
17239 	mutex_enter(&dtrace_lock);
17240 	dtrace_opens++;
17241 	dtrace_membar_producer();
17242 
17243 #ifdef illumos
17244 	/*
17245 	 * If the kernel debugger is active (that is, if the kernel debugger
17246 	 * modified text in some way), we won't allow the open.
17247 	 */
17248 	if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
17249 		dtrace_opens--;
17250 		mutex_exit(&cpu_lock);
17251 		mutex_exit(&dtrace_lock);
17252 		return (EBUSY);
17253 	}
17254 
17255 	if (dtrace_helptrace_enable && dtrace_helptrace_buffer == NULL) {
17256 		/*
17257 		 * If DTrace helper tracing is enabled, we need to allocate the
17258 		 * trace buffer and initialize the values.
17259 		 */
17260 		dtrace_helptrace_buffer =
17261 		    kmem_zalloc(dtrace_helptrace_bufsize, KM_SLEEP);
17262 		dtrace_helptrace_next = 0;
17263 		dtrace_helptrace_wrapped = 0;
17264 		dtrace_helptrace_enable = 0;
17265 	}
17266 
17267 	state = dtrace_state_create(devp, cred_p);
17268 #else
17269 	state = dtrace_state_create(dev, NULL);
17270 	devfs_set_cdevpriv(state, dtrace_dtr);
17271 #endif
17272 
17273 	mutex_exit(&cpu_lock);
17274 
17275 	if (state == NULL) {
17276 #ifdef illumos
17277 		if (--dtrace_opens == 0 && dtrace_anon.dta_enabling == NULL)
17278 			(void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
17279 #else
17280 		--dtrace_opens;
17281 #endif
17282 		mutex_exit(&dtrace_lock);
17283 		return (EAGAIN);
17284 	}
17285 
17286 	mutex_exit(&dtrace_lock);
17287 
17288 	return (0);
17289 }
17290 
17291 /*ARGSUSED*/
17292 #ifdef illumos
17293 static int
17294 dtrace_close(dev_t dev, int flag, int otyp, cred_t *cred_p)
17295 #else
17296 static void
17297 dtrace_dtr(void *data)
17298 #endif
17299 {
17300 #ifdef illumos
17301 	minor_t minor = getminor(dev);
17302 	dtrace_state_t *state;
17303 #endif
17304 	dtrace_helptrace_t *buf = NULL;
17305 
17306 #ifdef illumos
17307 	if (minor == DTRACEMNRN_HELPER)
17308 		return (0);
17309 
17310 	state = ddi_get_soft_state(dtrace_softstate, minor);
17311 #else
17312 	dtrace_state_t *state = data;
17313 #endif
17314 
17315 	mutex_enter(&cpu_lock);
17316 	mutex_enter(&dtrace_lock);
17317 
17318 #ifdef illumos
17319 	if (state->dts_anon)
17320 #else
17321 	if (state != NULL && state->dts_anon)
17322 #endif
17323 	{
17324 		/*
17325 		 * There is anonymous state. Destroy that first.
17326 		 */
17327 		ASSERT(dtrace_anon.dta_state == NULL);
17328 		dtrace_state_destroy(state->dts_anon);
17329 	}
17330 
17331 	if (dtrace_helptrace_disable) {
17332 		/*
17333 		 * If we have been told to disable helper tracing, set the
17334 		 * buffer to NULL before calling into dtrace_state_destroy();
17335 		 * we take advantage of its dtrace_sync() to know that no
17336 		 * CPU is in probe context with enabled helper tracing
17337 		 * after it returns.
17338 		 */
17339 		buf = dtrace_helptrace_buffer;
17340 		dtrace_helptrace_buffer = NULL;
17341 	}
17342 
17343 #ifdef illumos
17344 	dtrace_state_destroy(state);
17345 #else
17346 	if (state != NULL) {
17347 		dtrace_state_destroy(state);
17348 		kmem_free(state, 0);
17349 	}
17350 #endif
17351 	ASSERT(dtrace_opens > 0);
17352 
17353 #ifdef illumos
17354 	/*
17355 	 * Only relinquish control of the kernel debugger interface when there
17356 	 * are no consumers and no anonymous enablings.
17357 	 */
17358 	if (--dtrace_opens == 0 && dtrace_anon.dta_enabling == NULL)
17359 		(void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
17360 #else
17361 	--dtrace_opens;
17362 #endif
17363 
17364 	if (buf != NULL) {
17365 		kmem_free(buf, dtrace_helptrace_bufsize);
17366 		dtrace_helptrace_disable = 0;
17367 	}
17368 
17369 	mutex_exit(&dtrace_lock);
17370 	mutex_exit(&cpu_lock);
17371 
17372 #ifdef illumos
17373 	return (0);
17374 #endif
17375 }
17376 
17377 #ifdef illumos
17378 /*ARGSUSED*/
17379 static int
17380 dtrace_ioctl_helper(int cmd, intptr_t arg, int *rv)
17381 {
17382 	int rval;
17383 	dof_helper_t help, *dhp = NULL;
17384 
17385 	switch (cmd) {
17386 	case DTRACEHIOC_ADDDOF:
17387 		if (copyin((void *)arg, &help, sizeof (help)) != 0) {
17388 			dtrace_dof_error(NULL, "failed to copyin DOF helper");
17389 			return (EFAULT);
17390 		}
17391 
17392 		dhp = &help;
17393 		arg = (intptr_t)help.dofhp_dof;
17394 		/*FALLTHROUGH*/
17395 
17396 	case DTRACEHIOC_ADD: {
17397 		dof_hdr_t *dof = dtrace_dof_copyin(arg, &rval);
17398 
17399 		if (dof == NULL)
17400 			return (rval);
17401 
17402 		mutex_enter(&dtrace_lock);
17403 
17404 		/*
17405 		 * dtrace_helper_slurp() takes responsibility for the dof --
17406 		 * it may free it now or it may save it and free it later.
17407 		 */
17408 		if ((rval = dtrace_helper_slurp(dof, dhp)) != -1) {
17409 			*rv = rval;
17410 			rval = 0;
17411 		} else {
17412 			rval = EINVAL;
17413 		}
17414 
17415 		mutex_exit(&dtrace_lock);
17416 		return (rval);
17417 	}
17418 
17419 	case DTRACEHIOC_REMOVE: {
17420 		mutex_enter(&dtrace_lock);
17421 		rval = dtrace_helper_destroygen(NULL, arg);
17422 		mutex_exit(&dtrace_lock);
17423 
17424 		return (rval);
17425 	}
17426 
17427 	default:
17428 		break;
17429 	}
17430 
17431 	return (ENOTTY);
17432 }
17433 
17434 /*ARGSUSED*/
17435 static int
17436 dtrace_ioctl(dev_t dev, int cmd, intptr_t arg, int md, cred_t *cr, int *rv)
17437 {
17438 	minor_t minor = getminor(dev);
17439 	dtrace_state_t *state;
17440 	int rval;
17441 
17442 	if (minor == DTRACEMNRN_HELPER)
17443 		return (dtrace_ioctl_helper(cmd, arg, rv));
17444 
17445 	state = ddi_get_soft_state(dtrace_softstate, minor);
17446 
17447 	if (state->dts_anon) {
17448 		ASSERT(dtrace_anon.dta_state == NULL);
17449 		state = state->dts_anon;
17450 	}
17451 
17452 	switch (cmd) {
17453 	case DTRACEIOC_PROVIDER: {
17454 		dtrace_providerdesc_t pvd;
17455 		dtrace_provider_t *pvp;
17456 
17457 		if (copyin((void *)arg, &pvd, sizeof (pvd)) != 0)
17458 			return (EFAULT);
17459 
17460 		pvd.dtvd_name[DTRACE_PROVNAMELEN - 1] = '\0';
17461 		mutex_enter(&dtrace_provider_lock);
17462 
17463 		for (pvp = dtrace_provider; pvp != NULL; pvp = pvp->dtpv_next) {
17464 			if (strcmp(pvp->dtpv_name, pvd.dtvd_name) == 0)
17465 				break;
17466 		}
17467 
17468 		mutex_exit(&dtrace_provider_lock);
17469 
17470 		if (pvp == NULL)
17471 			return (ESRCH);
17472 
17473 		bcopy(&pvp->dtpv_priv, &pvd.dtvd_priv, sizeof (dtrace_ppriv_t));
17474 		bcopy(&pvp->dtpv_attr, &pvd.dtvd_attr, sizeof (dtrace_pattr_t));
17475 
17476 		if (copyout(&pvd, (void *)arg, sizeof (pvd)) != 0)
17477 			return (EFAULT);
17478 
17479 		return (0);
17480 	}
17481 
17482 	case DTRACEIOC_EPROBE: {
17483 		dtrace_eprobedesc_t epdesc;
17484 		dtrace_ecb_t *ecb;
17485 		dtrace_action_t *act;
17486 		void *buf;
17487 		size_t size;
17488 		uintptr_t dest;
17489 		int nrecs;
17490 
17491 		if (copyin((void *)arg, &epdesc, sizeof (epdesc)) != 0)
17492 			return (EFAULT);
17493 
17494 		mutex_enter(&dtrace_lock);
17495 
17496 		if ((ecb = dtrace_epid2ecb(state, epdesc.dtepd_epid)) == NULL) {
17497 			mutex_exit(&dtrace_lock);
17498 			return (EINVAL);
17499 		}
17500 
17501 		if (ecb->dte_probe == NULL) {
17502 			mutex_exit(&dtrace_lock);
17503 			return (EINVAL);
17504 		}
17505 
17506 		epdesc.dtepd_probeid = ecb->dte_probe->dtpr_id;
17507 		epdesc.dtepd_uarg = ecb->dte_uarg;
17508 		epdesc.dtepd_size = ecb->dte_size;
17509 
17510 		nrecs = epdesc.dtepd_nrecs;
17511 		epdesc.dtepd_nrecs = 0;
17512 		for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
17513 			if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
17514 				continue;
17515 
17516 			epdesc.dtepd_nrecs++;
17517 		}
17518 
17519 		/*
17520 		 * Now that we have the size, we need to allocate a temporary
17521 		 * buffer in which to store the complete description.  We need
17522 		 * the temporary buffer to be able to drop dtrace_lock()
17523 		 * across the copyout(), below.
17524 		 */
17525 		size = sizeof (dtrace_eprobedesc_t) +
17526 		    (epdesc.dtepd_nrecs * sizeof (dtrace_recdesc_t));
17527 
17528 		buf = kmem_alloc(size, KM_SLEEP);
17529 		dest = (uintptr_t)buf;
17530 
17531 		bcopy(&epdesc, (void *)dest, sizeof (epdesc));
17532 		dest += offsetof(dtrace_eprobedesc_t, dtepd_rec[0]);
17533 
17534 		for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
17535 			if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
17536 				continue;
17537 
17538 			if (nrecs-- == 0)
17539 				break;
17540 
17541 			bcopy(&act->dta_rec, (void *)dest,
17542 			    sizeof (dtrace_recdesc_t));
17543 			dest += sizeof (dtrace_recdesc_t);
17544 		}
17545 
17546 		mutex_exit(&dtrace_lock);
17547 
17548 		if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
17549 			kmem_free(buf, size);
17550 			return (EFAULT);
17551 		}
17552 
17553 		kmem_free(buf, size);
17554 		return (0);
17555 	}
17556 
17557 	case DTRACEIOC_AGGDESC: {
17558 		dtrace_aggdesc_t aggdesc;
17559 		dtrace_action_t *act;
17560 		dtrace_aggregation_t *agg;
17561 		int nrecs;
17562 		uint32_t offs;
17563 		dtrace_recdesc_t *lrec;
17564 		void *buf;
17565 		size_t size;
17566 		uintptr_t dest;
17567 
17568 		if (copyin((void *)arg, &aggdesc, sizeof (aggdesc)) != 0)
17569 			return (EFAULT);
17570 
17571 		mutex_enter(&dtrace_lock);
17572 
17573 		if ((agg = dtrace_aggid2agg(state, aggdesc.dtagd_id)) == NULL) {
17574 			mutex_exit(&dtrace_lock);
17575 			return (EINVAL);
17576 		}
17577 
17578 		aggdesc.dtagd_epid = agg->dtag_ecb->dte_epid;
17579 
17580 		nrecs = aggdesc.dtagd_nrecs;
17581 		aggdesc.dtagd_nrecs = 0;
17582 
17583 		offs = agg->dtag_base;
17584 		lrec = &agg->dtag_action.dta_rec;
17585 		aggdesc.dtagd_size = lrec->dtrd_offset + lrec->dtrd_size - offs;
17586 
17587 		for (act = agg->dtag_first; ; act = act->dta_next) {
17588 			ASSERT(act->dta_intuple ||
17589 			    DTRACEACT_ISAGG(act->dta_kind));
17590 
17591 			/*
17592 			 * If this action has a record size of zero, it
17593 			 * denotes an argument to the aggregating action.
17594 			 * Because the presence of this record doesn't (or
17595 			 * shouldn't) affect the way the data is interpreted,
17596 			 * we don't copy it out to save user-level the
17597 			 * confusion of dealing with a zero-length record.
17598 			 */
17599 			if (act->dta_rec.dtrd_size == 0) {
17600 				ASSERT(agg->dtag_hasarg);
17601 				continue;
17602 			}
17603 
17604 			aggdesc.dtagd_nrecs++;
17605 
17606 			if (act == &agg->dtag_action)
17607 				break;
17608 		}
17609 
17610 		/*
17611 		 * Now that we have the size, we need to allocate a temporary
17612 		 * buffer in which to store the complete description.  We need
17613 		 * the temporary buffer to be able to drop dtrace_lock()
17614 		 * across the copyout(), below.
17615 		 */
17616 		size = sizeof (dtrace_aggdesc_t) +
17617 		    (aggdesc.dtagd_nrecs * sizeof (dtrace_recdesc_t));
17618 
17619 		buf = kmem_alloc(size, KM_SLEEP);
17620 		dest = (uintptr_t)buf;
17621 
17622 		bcopy(&aggdesc, (void *)dest, sizeof (aggdesc));
17623 		dest += offsetof(dtrace_aggdesc_t, dtagd_rec[0]);
17624 
17625 		for (act = agg->dtag_first; ; act = act->dta_next) {
17626 			dtrace_recdesc_t rec = act->dta_rec;
17627 
17628 			/*
17629 			 * See the comment in the above loop for why we pass
17630 			 * over zero-length records.
17631 			 */
17632 			if (rec.dtrd_size == 0) {
17633 				ASSERT(agg->dtag_hasarg);
17634 				continue;
17635 			}
17636 
17637 			if (nrecs-- == 0)
17638 				break;
17639 
17640 			rec.dtrd_offset -= offs;
17641 			bcopy(&rec, (void *)dest, sizeof (rec));
17642 			dest += sizeof (dtrace_recdesc_t);
17643 
17644 			if (act == &agg->dtag_action)
17645 				break;
17646 		}
17647 
17648 		mutex_exit(&dtrace_lock);
17649 
17650 		if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
17651 			kmem_free(buf, size);
17652 			return (EFAULT);
17653 		}
17654 
17655 		kmem_free(buf, size);
17656 		return (0);
17657 	}
17658 
17659 	case DTRACEIOC_ENABLE: {
17660 		dof_hdr_t *dof;
17661 		dtrace_enabling_t *enab = NULL;
17662 		dtrace_vstate_t *vstate;
17663 		int err = 0;
17664 
17665 		*rv = 0;
17666 
17667 		/*
17668 		 * If a NULL argument has been passed, we take this as our
17669 		 * cue to reevaluate our enablings.
17670 		 */
17671 		if (arg == NULL) {
17672 			dtrace_enabling_matchall();
17673 
17674 			return (0);
17675 		}
17676 
17677 		if ((dof = dtrace_dof_copyin(arg, &rval)) == NULL)
17678 			return (rval);
17679 
17680 		mutex_enter(&cpu_lock);
17681 		mutex_enter(&dtrace_lock);
17682 		vstate = &state->dts_vstate;
17683 
17684 		if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
17685 			mutex_exit(&dtrace_lock);
17686 			mutex_exit(&cpu_lock);
17687 			dtrace_dof_destroy(dof);
17688 			return (EBUSY);
17689 		}
17690 
17691 		if (dtrace_dof_slurp(dof, vstate, cr, &enab, 0, B_TRUE) != 0) {
17692 			mutex_exit(&dtrace_lock);
17693 			mutex_exit(&cpu_lock);
17694 			dtrace_dof_destroy(dof);
17695 			return (EINVAL);
17696 		}
17697 
17698 		if ((rval = dtrace_dof_options(dof, state)) != 0) {
17699 			dtrace_enabling_destroy(enab);
17700 			mutex_exit(&dtrace_lock);
17701 			mutex_exit(&cpu_lock);
17702 			dtrace_dof_destroy(dof);
17703 			return (rval);
17704 		}
17705 
17706 		if ((err = dtrace_enabling_match(enab, rv)) == 0) {
17707 			err = dtrace_enabling_retain(enab);
17708 		} else {
17709 			dtrace_enabling_destroy(enab);
17710 		}
17711 
17712 		mutex_exit(&cpu_lock);
17713 		mutex_exit(&dtrace_lock);
17714 		dtrace_dof_destroy(dof);
17715 
17716 		return (err);
17717 	}
17718 
17719 	case DTRACEIOC_REPLICATE: {
17720 		dtrace_repldesc_t desc;
17721 		dtrace_probedesc_t *match = &desc.dtrpd_match;
17722 		dtrace_probedesc_t *create = &desc.dtrpd_create;
17723 		int err;
17724 
17725 		if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
17726 			return (EFAULT);
17727 
17728 		match->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
17729 		match->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
17730 		match->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
17731 		match->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
17732 
17733 		create->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
17734 		create->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
17735 		create->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
17736 		create->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
17737 
17738 		mutex_enter(&dtrace_lock);
17739 		err = dtrace_enabling_replicate(state, match, create);
17740 		mutex_exit(&dtrace_lock);
17741 
17742 		return (err);
17743 	}
17744 
17745 	case DTRACEIOC_PROBEMATCH:
17746 	case DTRACEIOC_PROBES: {
17747 		dtrace_probe_t *probe = NULL;
17748 		dtrace_probedesc_t desc;
17749 		dtrace_probekey_t pkey;
17750 		dtrace_id_t i;
17751 		int m = 0;
17752 		uint32_t priv;
17753 		uid_t uid;
17754 		zoneid_t zoneid;
17755 
17756 		if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
17757 			return (EFAULT);
17758 
17759 		desc.dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
17760 		desc.dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
17761 		desc.dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
17762 		desc.dtpd_name[DTRACE_NAMELEN - 1] = '\0';
17763 
17764 		/*
17765 		 * Before we attempt to match this probe, we want to give
17766 		 * all providers the opportunity to provide it.
17767 		 */
17768 		if (desc.dtpd_id == DTRACE_IDNONE) {
17769 			mutex_enter(&dtrace_provider_lock);
17770 			dtrace_probe_provide(&desc, NULL);
17771 			mutex_exit(&dtrace_provider_lock);
17772 			desc.dtpd_id++;
17773 		}
17774 
17775 		if (cmd == DTRACEIOC_PROBEMATCH)  {
17776 			dtrace_probekey(&desc, &pkey);
17777 			pkey.dtpk_id = DTRACE_IDNONE;
17778 		}
17779 
17780 		dtrace_cred2priv(cr, &priv, &uid, &zoneid);
17781 
17782 		mutex_enter(&dtrace_lock);
17783 
17784 		if (cmd == DTRACEIOC_PROBEMATCH) {
17785 			for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
17786 				if ((probe = dtrace_probes[i - 1]) != NULL &&
17787 				    (m = dtrace_match_probe(probe, &pkey,
17788 				    priv, uid, zoneid)) != 0)
17789 					break;
17790 			}
17791 
17792 			if (m < 0) {
17793 				mutex_exit(&dtrace_lock);
17794 				return (EINVAL);
17795 			}
17796 
17797 		} else {
17798 			for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
17799 				if ((probe = dtrace_probes[i - 1]) != NULL &&
17800 				    dtrace_match_priv(probe, priv, uid, zoneid))
17801 					break;
17802 			}
17803 		}
17804 
17805 		if (probe == NULL) {
17806 			mutex_exit(&dtrace_lock);
17807 			return (ESRCH);
17808 		}
17809 
17810 		dtrace_probe_description(probe, &desc);
17811 		mutex_exit(&dtrace_lock);
17812 
17813 		if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
17814 			return (EFAULT);
17815 
17816 		return (0);
17817 	}
17818 
17819 	case DTRACEIOC_PROBEARG: {
17820 		dtrace_argdesc_t desc;
17821 		dtrace_probe_t *probe;
17822 		dtrace_provider_t *prov;
17823 
17824 		if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
17825 			return (EFAULT);
17826 
17827 		if (desc.dtargd_id == DTRACE_IDNONE)
17828 			return (EINVAL);
17829 
17830 		if (desc.dtargd_ndx == DTRACE_ARGNONE)
17831 			return (EINVAL);
17832 
17833 		mutex_enter(&dtrace_provider_lock);
17834 		mutex_enter(&mod_lock);
17835 		mutex_enter(&dtrace_lock);
17836 
17837 		if (desc.dtargd_id > dtrace_nprobes) {
17838 			mutex_exit(&dtrace_lock);
17839 			mutex_exit(&mod_lock);
17840 			mutex_exit(&dtrace_provider_lock);
17841 			return (EINVAL);
17842 		}
17843 
17844 		if ((probe = dtrace_probes[desc.dtargd_id - 1]) == NULL) {
17845 			mutex_exit(&dtrace_lock);
17846 			mutex_exit(&mod_lock);
17847 			mutex_exit(&dtrace_provider_lock);
17848 			return (EINVAL);
17849 		}
17850 
17851 		mutex_exit(&dtrace_lock);
17852 
17853 		prov = probe->dtpr_provider;
17854 
17855 		if (prov->dtpv_pops.dtps_getargdesc == NULL) {
17856 			/*
17857 			 * There isn't any typed information for this probe.
17858 			 * Set the argument number to DTRACE_ARGNONE.
17859 			 */
17860 			desc.dtargd_ndx = DTRACE_ARGNONE;
17861 		} else {
17862 			desc.dtargd_native[0] = '\0';
17863 			desc.dtargd_xlate[0] = '\0';
17864 			desc.dtargd_mapping = desc.dtargd_ndx;
17865 
17866 			prov->dtpv_pops.dtps_getargdesc(prov->dtpv_arg,
17867 			    probe->dtpr_id, probe->dtpr_arg, &desc);
17868 		}
17869 
17870 		mutex_exit(&mod_lock);
17871 		mutex_exit(&dtrace_provider_lock);
17872 
17873 		if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
17874 			return (EFAULT);
17875 
17876 		return (0);
17877 	}
17878 
17879 	case DTRACEIOC_GO: {
17880 		processorid_t cpuid;
17881 		rval = dtrace_state_go(state, &cpuid);
17882 
17883 		if (rval != 0)
17884 			return (rval);
17885 
17886 		if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
17887 			return (EFAULT);
17888 
17889 		return (0);
17890 	}
17891 
17892 	case DTRACEIOC_STOP: {
17893 		processorid_t cpuid;
17894 
17895 		mutex_enter(&dtrace_lock);
17896 		rval = dtrace_state_stop(state, &cpuid);
17897 		mutex_exit(&dtrace_lock);
17898 
17899 		if (rval != 0)
17900 			return (rval);
17901 
17902 		if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
17903 			return (EFAULT);
17904 
17905 		return (0);
17906 	}
17907 
17908 	case DTRACEIOC_DOFGET: {
17909 		dof_hdr_t hdr, *dof;
17910 		uint64_t len;
17911 
17912 		if (copyin((void *)arg, &hdr, sizeof (hdr)) != 0)
17913 			return (EFAULT);
17914 
17915 		mutex_enter(&dtrace_lock);
17916 		dof = dtrace_dof_create(state);
17917 		mutex_exit(&dtrace_lock);
17918 
17919 		len = MIN(hdr.dofh_loadsz, dof->dofh_loadsz);
17920 		rval = copyout(dof, (void *)arg, len);
17921 		dtrace_dof_destroy(dof);
17922 
17923 		return (rval == 0 ? 0 : EFAULT);
17924 	}
17925 
17926 	case DTRACEIOC_AGGSNAP:
17927 	case DTRACEIOC_BUFSNAP: {
17928 		dtrace_bufdesc_t desc;
17929 		caddr_t cached;
17930 		dtrace_buffer_t *buf;
17931 
17932 		if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
17933 			return (EFAULT);
17934 
17935 		if (desc.dtbd_cpu < 0 || desc.dtbd_cpu >= NCPU)
17936 			return (EINVAL);
17937 
17938 		mutex_enter(&dtrace_lock);
17939 
17940 		if (cmd == DTRACEIOC_BUFSNAP) {
17941 			buf = &state->dts_buffer[desc.dtbd_cpu];
17942 		} else {
17943 			buf = &state->dts_aggbuffer[desc.dtbd_cpu];
17944 		}
17945 
17946 		if (buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL)) {
17947 			size_t sz = buf->dtb_offset;
17948 
17949 			if (state->dts_activity != DTRACE_ACTIVITY_STOPPED) {
17950 				mutex_exit(&dtrace_lock);
17951 				return (EBUSY);
17952 			}
17953 
17954 			/*
17955 			 * If this buffer has already been consumed, we're
17956 			 * going to indicate that there's nothing left here
17957 			 * to consume.
17958 			 */
17959 			if (buf->dtb_flags & DTRACEBUF_CONSUMED) {
17960 				mutex_exit(&dtrace_lock);
17961 
17962 				desc.dtbd_size = 0;
17963 				desc.dtbd_drops = 0;
17964 				desc.dtbd_errors = 0;
17965 				desc.dtbd_oldest = 0;
17966 				sz = sizeof (desc);
17967 
17968 				if (copyout(&desc, (void *)arg, sz) != 0)
17969 					return (EFAULT);
17970 
17971 				return (0);
17972 			}
17973 
17974 			/*
17975 			 * If this is a ring buffer that has wrapped, we want
17976 			 * to copy the whole thing out.
17977 			 */
17978 			if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
17979 				dtrace_buffer_polish(buf);
17980 				sz = buf->dtb_size;
17981 			}
17982 
17983 			if (copyout(buf->dtb_tomax, desc.dtbd_data, sz) != 0) {
17984 				mutex_exit(&dtrace_lock);
17985 				return (EFAULT);
17986 			}
17987 
17988 			desc.dtbd_size = sz;
17989 			desc.dtbd_drops = buf->dtb_drops;
17990 			desc.dtbd_errors = buf->dtb_errors;
17991 			desc.dtbd_oldest = buf->dtb_xamot_offset;
17992 			desc.dtbd_timestamp = dtrace_gethrtime();
17993 
17994 			mutex_exit(&dtrace_lock);
17995 
17996 			if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
17997 				return (EFAULT);
17998 
17999 			buf->dtb_flags |= DTRACEBUF_CONSUMED;
18000 
18001 			return (0);
18002 		}
18003 
18004 		if (buf->dtb_tomax == NULL) {
18005 			ASSERT(buf->dtb_xamot == NULL);
18006 			mutex_exit(&dtrace_lock);
18007 			return (ENOENT);
18008 		}
18009 
18010 		cached = buf->dtb_tomax;
18011 		ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
18012 
18013 		dtrace_xcall(desc.dtbd_cpu,
18014 		    (dtrace_xcall_t)dtrace_buffer_switch, buf);
18015 
18016 		state->dts_errors += buf->dtb_xamot_errors;
18017 
18018 		/*
18019 		 * If the buffers did not actually switch, then the cross call
18020 		 * did not take place -- presumably because the given CPU is
18021 		 * not in the ready set.  If this is the case, we'll return
18022 		 * ENOENT.
18023 		 */
18024 		if (buf->dtb_tomax == cached) {
18025 			ASSERT(buf->dtb_xamot != cached);
18026 			mutex_exit(&dtrace_lock);
18027 			return (ENOENT);
18028 		}
18029 
18030 		ASSERT(cached == buf->dtb_xamot);
18031 
18032 		/*
18033 		 * We have our snapshot; now copy it out.
18034 		 */
18035 		if (copyout(buf->dtb_xamot, desc.dtbd_data,
18036 		    buf->dtb_xamot_offset) != 0) {
18037 			mutex_exit(&dtrace_lock);
18038 			return (EFAULT);
18039 		}
18040 
18041 		desc.dtbd_size = buf->dtb_xamot_offset;
18042 		desc.dtbd_drops = buf->dtb_xamot_drops;
18043 		desc.dtbd_errors = buf->dtb_xamot_errors;
18044 		desc.dtbd_oldest = 0;
18045 		desc.dtbd_timestamp = buf->dtb_switched;
18046 
18047 		mutex_exit(&dtrace_lock);
18048 
18049 		/*
18050 		 * Finally, copy out the buffer description.
18051 		 */
18052 		if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
18053 			return (EFAULT);
18054 
18055 		return (0);
18056 	}
18057 
18058 	case DTRACEIOC_CONF: {
18059 		dtrace_conf_t conf;
18060 
18061 		bzero(&conf, sizeof (conf));
18062 		conf.dtc_difversion = DIF_VERSION;
18063 		conf.dtc_difintregs = DIF_DIR_NREGS;
18064 		conf.dtc_diftupregs = DIF_DTR_NREGS;
18065 		conf.dtc_ctfmodel = CTF_MODEL_NATIVE;
18066 
18067 		if (copyout(&conf, (void *)arg, sizeof (conf)) != 0)
18068 			return (EFAULT);
18069 
18070 		return (0);
18071 	}
18072 
18073 	case DTRACEIOC_STATUS: {
18074 		dtrace_status_t stat;
18075 		dtrace_dstate_t *dstate;
18076 		int i, j;
18077 		uint64_t nerrs;
18078 
18079 		/*
18080 		 * See the comment in dtrace_state_deadman() for the reason
18081 		 * for setting dts_laststatus to INT64_MAX before setting
18082 		 * it to the correct value.
18083 		 */
18084 		state->dts_laststatus = INT64_MAX;
18085 		dtrace_membar_producer();
18086 		state->dts_laststatus = dtrace_gethrtime();
18087 
18088 		bzero(&stat, sizeof (stat));
18089 
18090 		mutex_enter(&dtrace_lock);
18091 
18092 		if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE) {
18093 			mutex_exit(&dtrace_lock);
18094 			return (ENOENT);
18095 		}
18096 
18097 		if (state->dts_activity == DTRACE_ACTIVITY_DRAINING)
18098 			stat.dtst_exiting = 1;
18099 
18100 		nerrs = state->dts_errors;
18101 		dstate = &state->dts_vstate.dtvs_dynvars;
18102 
18103 		for (i = 0; i < NCPU; i++) {
18104 			dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[i];
18105 
18106 			stat.dtst_dyndrops += dcpu->dtdsc_drops;
18107 			stat.dtst_dyndrops_dirty += dcpu->dtdsc_dirty_drops;
18108 			stat.dtst_dyndrops_rinsing += dcpu->dtdsc_rinsing_drops;
18109 
18110 			if (state->dts_buffer[i].dtb_flags & DTRACEBUF_FULL)
18111 				stat.dtst_filled++;
18112 
18113 			nerrs += state->dts_buffer[i].dtb_errors;
18114 
18115 			for (j = 0; j < state->dts_nspeculations; j++) {
18116 				dtrace_speculation_t *spec;
18117 				dtrace_buffer_t *buf;
18118 
18119 				spec = &state->dts_speculations[j];
18120 				buf = &spec->dtsp_buffer[i];
18121 				stat.dtst_specdrops += buf->dtb_xamot_drops;
18122 			}
18123 		}
18124 
18125 		stat.dtst_specdrops_busy = state->dts_speculations_busy;
18126 		stat.dtst_specdrops_unavail = state->dts_speculations_unavail;
18127 		stat.dtst_stkstroverflows = state->dts_stkstroverflows;
18128 		stat.dtst_dblerrors = state->dts_dblerrors;
18129 		stat.dtst_killed =
18130 		    (state->dts_activity == DTRACE_ACTIVITY_KILLED);
18131 		stat.dtst_errors = nerrs;
18132 
18133 		mutex_exit(&dtrace_lock);
18134 
18135 		if (copyout(&stat, (void *)arg, sizeof (stat)) != 0)
18136 			return (EFAULT);
18137 
18138 		return (0);
18139 	}
18140 
18141 	case DTRACEIOC_FORMAT: {
18142 		dtrace_fmtdesc_t fmt;
18143 		char *str;
18144 		int len;
18145 
18146 		if (copyin((void *)arg, &fmt, sizeof (fmt)) != 0)
18147 			return (EFAULT);
18148 
18149 		mutex_enter(&dtrace_lock);
18150 
18151 		if (fmt.dtfd_format == 0 ||
18152 		    fmt.dtfd_format > state->dts_nformats) {
18153 			mutex_exit(&dtrace_lock);
18154 			return (EINVAL);
18155 		}
18156 
18157 		/*
18158 		 * Format strings are allocated contiguously and they are
18159 		 * never freed; if a format index is less than the number
18160 		 * of formats, we can assert that the format map is non-NULL
18161 		 * and that the format for the specified index is non-NULL.
18162 		 */
18163 		ASSERT(state->dts_formats != NULL);
18164 		str = state->dts_formats[fmt.dtfd_format - 1];
18165 		ASSERT(str != NULL);
18166 
18167 		len = strlen(str) + 1;
18168 
18169 		if (len > fmt.dtfd_length) {
18170 			fmt.dtfd_length = len;
18171 
18172 			if (copyout(&fmt, (void *)arg, sizeof (fmt)) != 0) {
18173 				mutex_exit(&dtrace_lock);
18174 				return (EINVAL);
18175 			}
18176 		} else {
18177 			if (copyout(str, fmt.dtfd_string, len) != 0) {
18178 				mutex_exit(&dtrace_lock);
18179 				return (EINVAL);
18180 			}
18181 		}
18182 
18183 		mutex_exit(&dtrace_lock);
18184 		return (0);
18185 	}
18186 
18187 	default:
18188 		break;
18189 	}
18190 
18191 	return (ENOTTY);
18192 }
18193 
18194 /*ARGSUSED*/
18195 static int
18196 dtrace_detach(dev_info_t *dip, ddi_detach_cmd_t cmd)
18197 {
18198 	dtrace_state_t *state;
18199 
18200 	switch (cmd) {
18201 	case DDI_DETACH:
18202 		break;
18203 
18204 	case DDI_SUSPEND:
18205 		return (DDI_SUCCESS);
18206 
18207 	default:
18208 		return (DDI_FAILURE);
18209 	}
18210 
18211 	mutex_enter(&cpu_lock);
18212 	mutex_enter(&dtrace_provider_lock);
18213 	mutex_enter(&dtrace_lock);
18214 
18215 	ASSERT(dtrace_opens == 0);
18216 
18217 	if (dtrace_helpers > 0) {
18218 		mutex_exit(&dtrace_provider_lock);
18219 		mutex_exit(&dtrace_lock);
18220 		mutex_exit(&cpu_lock);
18221 		return (DDI_FAILURE);
18222 	}
18223 
18224 	if (dtrace_unregister((dtrace_provider_id_t)dtrace_provider) != 0) {
18225 		mutex_exit(&dtrace_provider_lock);
18226 		mutex_exit(&dtrace_lock);
18227 		mutex_exit(&cpu_lock);
18228 		return (DDI_FAILURE);
18229 	}
18230 
18231 	dtrace_provider = NULL;
18232 
18233 	if ((state = dtrace_anon_grab()) != NULL) {
18234 		/*
18235 		 * If there were ECBs on this state, the provider should
18236 		 * have not been allowed to detach; assert that there is
18237 		 * none.
18238 		 */
18239 		ASSERT(state->dts_necbs == 0);
18240 		dtrace_state_destroy(state);
18241 
18242 		/*
18243 		 * If we're being detached with anonymous state, we need to
18244 		 * indicate to the kernel debugger that DTrace is now inactive.
18245 		 */
18246 		(void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
18247 	}
18248 
18249 	bzero(&dtrace_anon, sizeof (dtrace_anon_t));
18250 	unregister_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
18251 	dtrace_cpu_init = NULL;
18252 	dtrace_helpers_cleanup = NULL;
18253 	dtrace_helpers_fork = NULL;
18254 	dtrace_cpustart_init = NULL;
18255 	dtrace_cpustart_fini = NULL;
18256 	dtrace_debugger_init = NULL;
18257 	dtrace_debugger_fini = NULL;
18258 	dtrace_modload = NULL;
18259 	dtrace_modunload = NULL;
18260 
18261 	ASSERT(dtrace_getf == 0);
18262 	ASSERT(dtrace_closef == NULL);
18263 
18264 	mutex_exit(&cpu_lock);
18265 
18266 	kmem_free(dtrace_probes, dtrace_nprobes * sizeof (dtrace_probe_t *));
18267 	dtrace_probes = NULL;
18268 	dtrace_nprobes = 0;
18269 
18270 	dtrace_hash_destroy(dtrace_bymod);
18271 	dtrace_hash_destroy(dtrace_byfunc);
18272 	dtrace_hash_destroy(dtrace_byname);
18273 	dtrace_bymod = NULL;
18274 	dtrace_byfunc = NULL;
18275 	dtrace_byname = NULL;
18276 
18277 	kmem_cache_destroy(dtrace_state_cache);
18278 	vmem_destroy(dtrace_minor);
18279 	vmem_destroy(dtrace_arena);
18280 
18281 	if (dtrace_toxrange != NULL) {
18282 		kmem_free(dtrace_toxrange,
18283 		    dtrace_toxranges_max * sizeof (dtrace_toxrange_t));
18284 		dtrace_toxrange = NULL;
18285 		dtrace_toxranges = 0;
18286 		dtrace_toxranges_max = 0;
18287 	}
18288 
18289 	ddi_remove_minor_node(dtrace_devi, NULL);
18290 	dtrace_devi = NULL;
18291 
18292 	ddi_soft_state_fini(&dtrace_softstate);
18293 
18294 	ASSERT(dtrace_vtime_references == 0);
18295 	ASSERT(dtrace_opens == 0);
18296 	ASSERT(dtrace_retained == NULL);
18297 
18298 	mutex_exit(&dtrace_lock);
18299 	mutex_exit(&dtrace_provider_lock);
18300 
18301 	/*
18302 	 * We don't destroy the task queue until after we have dropped our
18303 	 * locks (taskq_destroy() may block on running tasks).  To prevent
18304 	 * attempting to do work after we have effectively detached but before
18305 	 * the task queue has been destroyed, all tasks dispatched via the
18306 	 * task queue must check that DTrace is still attached before
18307 	 * performing any operation.
18308 	 */
18309 	taskq_destroy(dtrace_taskq);
18310 	dtrace_taskq = NULL;
18311 
18312 	return (DDI_SUCCESS);
18313 }
18314 #endif
18315 
18316 #ifdef illumos
18317 /*ARGSUSED*/
18318 static int
18319 dtrace_info(dev_info_t *dip, ddi_info_cmd_t infocmd, void *arg, void **result)
18320 {
18321 	int error;
18322 
18323 	switch (infocmd) {
18324 	case DDI_INFO_DEVT2DEVINFO:
18325 		*result = (void *)dtrace_devi;
18326 		error = DDI_SUCCESS;
18327 		break;
18328 	case DDI_INFO_DEVT2INSTANCE:
18329 		*result = (void *)0;
18330 		error = DDI_SUCCESS;
18331 		break;
18332 	default:
18333 		error = DDI_FAILURE;
18334 	}
18335 	return (error);
18336 }
18337 #endif
18338 
18339 #ifdef illumos
18340 static struct cb_ops dtrace_cb_ops = {
18341 	dtrace_open,		/* open */
18342 	dtrace_close,		/* close */
18343 	nulldev,		/* strategy */
18344 	nulldev,		/* print */
18345 	nodev,			/* dump */
18346 	nodev,			/* read */
18347 	nodev,			/* write */
18348 	dtrace_ioctl,		/* ioctl */
18349 	nodev,			/* devmap */
18350 	nodev,			/* mmap */
18351 	nodev,			/* segmap */
18352 	nochpoll,		/* poll */
18353 	ddi_prop_op,		/* cb_prop_op */
18354 	0,			/* streamtab  */
18355 	D_NEW | D_MP		/* Driver compatibility flag */
18356 };
18357 
18358 static struct dev_ops dtrace_ops = {
18359 	DEVO_REV,		/* devo_rev */
18360 	0,			/* refcnt */
18361 	dtrace_info,		/* get_dev_info */
18362 	nulldev,		/* identify */
18363 	nulldev,		/* probe */
18364 	dtrace_attach,		/* attach */
18365 	dtrace_detach,		/* detach */
18366 	nodev,			/* reset */
18367 	&dtrace_cb_ops,		/* driver operations */
18368 	NULL,			/* bus operations */
18369 	nodev			/* dev power */
18370 };
18371 
18372 static struct modldrv modldrv = {
18373 	&mod_driverops,		/* module type (this is a pseudo driver) */
18374 	"Dynamic Tracing",	/* name of module */
18375 	&dtrace_ops,		/* driver ops */
18376 };
18377 
18378 static struct modlinkage modlinkage = {
18379 	MODREV_1,
18380 	(void *)&modldrv,
18381 	NULL
18382 };
18383 
18384 int
18385 _init(void)
18386 {
18387 	return (mod_install(&modlinkage));
18388 }
18389 
18390 int
18391 _info(struct modinfo *modinfop)
18392 {
18393 	return (mod_info(&modlinkage, modinfop));
18394 }
18395 
18396 int
18397 _fini(void)
18398 {
18399 	return (mod_remove(&modlinkage));
18400 }
18401 #else
18402 
18403 static d_ioctl_t	dtrace_ioctl;
18404 static d_ioctl_t	dtrace_ioctl_helper;
18405 static void		dtrace_load(void *);
18406 static int		dtrace_unload(void);
18407 static struct cdev	*dtrace_dev;
18408 static struct cdev	*helper_dev;
18409 
18410 void dtrace_invop_init(void);
18411 void dtrace_invop_uninit(void);
18412 
18413 static struct cdevsw dtrace_cdevsw = {
18414 	.d_version	= D_VERSION,
18415 	.d_ioctl	= dtrace_ioctl,
18416 	.d_open		= dtrace_open,
18417 	.d_name		= "dtrace",
18418 };
18419 
18420 static struct cdevsw helper_cdevsw = {
18421 	.d_version	= D_VERSION,
18422 	.d_ioctl	= dtrace_ioctl_helper,
18423 	.d_name		= "helper",
18424 };
18425 
18426 #include <dtrace_anon.c>
18427 #include <dtrace_ioctl.c>
18428 #include <dtrace_load.c>
18429 #include <dtrace_modevent.c>
18430 #include <dtrace_sysctl.c>
18431 #include <dtrace_unload.c>
18432 #include <dtrace_vtime.c>
18433 #include <dtrace_hacks.c>
18434 #include <dtrace_isa.c>
18435 
18436 SYSINIT(dtrace_load, SI_SUB_DTRACE, SI_ORDER_FIRST, dtrace_load, NULL);
18437 SYSUNINIT(dtrace_unload, SI_SUB_DTRACE, SI_ORDER_FIRST, dtrace_unload, NULL);
18438 SYSINIT(dtrace_anon_init, SI_SUB_DTRACE_ANON, SI_ORDER_FIRST, dtrace_anon_init, NULL);
18439 
18440 DEV_MODULE(dtrace, dtrace_modevent, NULL);
18441 MODULE_VERSION(dtrace, 1);
18442 MODULE_DEPEND(dtrace, opensolaris, 1, 1, 1);
18443 #endif
18444