xref: /freebsd/sys/cddl/contrib/opensolaris/uts/common/dtrace/dtrace.c (revision 907b59d76938e654f0d040a888e8dfca3de1e222)
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 #ifndef illumos
71 #include <sys/time.h>
72 #endif
73 #include <sys/stat.h>
74 #include <sys/modctl.h>
75 #include <sys/conf.h>
76 #include <sys/systm.h>
77 #ifdef illumos
78 #include <sys/ddi.h>
79 #include <sys/sunddi.h>
80 #endif
81 #include <sys/cpuvar.h>
82 #include <sys/kmem.h>
83 #ifdef illumos
84 #include <sys/strsubr.h>
85 #endif
86 #include <sys/sysmacros.h>
87 #include <sys/dtrace_impl.h>
88 #include <sys/atomic.h>
89 #include <sys/cmn_err.h>
90 #ifdef illumos
91 #include <sys/mutex_impl.h>
92 #include <sys/rwlock_impl.h>
93 #endif
94 #include <sys/ctf_api.h>
95 #ifdef illumos
96 #include <sys/panic.h>
97 #include <sys/priv_impl.h>
98 #endif
99 #include <sys/policy.h>
100 #ifdef illumos
101 #include <sys/cred_impl.h>
102 #include <sys/procfs_isa.h>
103 #endif
104 #include <sys/taskq.h>
105 #ifdef illumos
106 #include <sys/mkdev.h>
107 #include <sys/kdi.h>
108 #endif
109 #include <sys/zone.h>
110 #include <sys/socket.h>
111 #include <netinet/in.h>
112 #include "strtolctype.h"
113 
114 /* FreeBSD includes: */
115 #ifndef illumos
116 #include <sys/callout.h>
117 #include <sys/ctype.h>
118 #include <sys/eventhandler.h>
119 #include <sys/limits.h>
120 #include <sys/linker.h>
121 #include <sys/kdb.h>
122 #include <sys/kernel.h>
123 #include <sys/malloc.h>
124 #include <sys/lock.h>
125 #include <sys/mutex.h>
126 #include <sys/ptrace.h>
127 #include <sys/rwlock.h>
128 #include <sys/sx.h>
129 #include <sys/sysctl.h>
130 
131 #include <sys/dtrace_bsd.h>
132 
133 #include <netinet/in.h>
134 
135 #include "dtrace_cddl.h"
136 #include "dtrace_debug.c"
137 #endif
138 
139 /*
140  * DTrace Tunable Variables
141  *
142  * The following variables may be tuned by adding a line to /etc/system that
143  * includes both the name of the DTrace module ("dtrace") and the name of the
144  * variable.  For example:
145  *
146  *   set dtrace:dtrace_destructive_disallow = 1
147  *
148  * In general, the only variables that one should be tuning this way are those
149  * that affect system-wide DTrace behavior, and for which the default behavior
150  * is undesirable.  Most of these variables are tunable on a per-consumer
151  * basis using DTrace options, and need not be tuned on a system-wide basis.
152  * When tuning these variables, avoid pathological values; while some attempt
153  * is made to verify the integrity of these variables, they are not considered
154  * part of the supported interface to DTrace, and they are therefore not
155  * checked comprehensively.  Further, these variables should not be tuned
156  * dynamically via "mdb -kw" or other means; they should only be tuned via
157  * /etc/system.
158  */
159 int		dtrace_destructive_disallow = 0;
160 dtrace_optval_t	dtrace_nonroot_maxsize = (16 * 1024 * 1024);
161 size_t		dtrace_difo_maxsize = (256 * 1024);
162 dtrace_optval_t	dtrace_dof_maxsize = (8 * 1024 * 1024);
163 size_t		dtrace_statvar_maxsize = (16 * 1024);
164 size_t		dtrace_actions_max = (16 * 1024);
165 size_t		dtrace_retain_max = 1024;
166 dtrace_optval_t	dtrace_helper_actions_max = 128;
167 dtrace_optval_t	dtrace_helper_providers_max = 32;
168 dtrace_optval_t	dtrace_dstate_defsize = (1 * 1024 * 1024);
169 size_t		dtrace_strsize_default = 256;
170 dtrace_optval_t	dtrace_cleanrate_default = 9900990;		/* 101 hz */
171 dtrace_optval_t	dtrace_cleanrate_min = 200000;			/* 5000 hz */
172 dtrace_optval_t	dtrace_cleanrate_max = (uint64_t)60 * NANOSEC;	/* 1/minute */
173 dtrace_optval_t	dtrace_aggrate_default = NANOSEC;		/* 1 hz */
174 dtrace_optval_t	dtrace_statusrate_default = NANOSEC;		/* 1 hz */
175 dtrace_optval_t dtrace_statusrate_max = (hrtime_t)10 * NANOSEC;	 /* 6/minute */
176 dtrace_optval_t	dtrace_switchrate_default = NANOSEC;		/* 1 hz */
177 dtrace_optval_t	dtrace_nspec_default = 1;
178 dtrace_optval_t	dtrace_specsize_default = 32 * 1024;
179 dtrace_optval_t dtrace_stackframes_default = 20;
180 dtrace_optval_t dtrace_ustackframes_default = 20;
181 dtrace_optval_t dtrace_jstackframes_default = 50;
182 dtrace_optval_t dtrace_jstackstrsize_default = 512;
183 int		dtrace_msgdsize_max = 128;
184 hrtime_t	dtrace_chill_max = MSEC2NSEC(500);		/* 500 ms */
185 hrtime_t	dtrace_chill_interval = NANOSEC;		/* 1000 ms */
186 int		dtrace_devdepth_max = 32;
187 int		dtrace_err_verbose;
188 hrtime_t	dtrace_deadman_interval = NANOSEC;
189 hrtime_t	dtrace_deadman_timeout = (hrtime_t)10 * NANOSEC;
190 hrtime_t	dtrace_deadman_user = (hrtime_t)30 * NANOSEC;
191 hrtime_t	dtrace_unregister_defunct_reap = (hrtime_t)60 * NANOSEC;
192 #ifndef illumos
193 int		dtrace_memstr_max = 4096;
194 #endif
195 
196 /*
197  * DTrace External Variables
198  *
199  * As dtrace(7D) is a kernel module, any DTrace variables are obviously
200  * available to DTrace consumers via the backtick (`) syntax.  One of these,
201  * dtrace_zero, is made deliberately so:  it is provided as a source of
202  * well-known, zero-filled memory.  While this variable is not documented,
203  * it is used by some translators as an implementation detail.
204  */
205 const char	dtrace_zero[256] = { 0 };	/* zero-filled memory */
206 
207 /*
208  * DTrace Internal Variables
209  */
210 #ifdef illumos
211 static dev_info_t	*dtrace_devi;		/* device info */
212 #endif
213 #ifdef illumos
214 static vmem_t		*dtrace_arena;		/* probe ID arena */
215 static vmem_t		*dtrace_minor;		/* minor number arena */
216 #else
217 static taskq_t		*dtrace_taskq;		/* task queue */
218 static struct unrhdr	*dtrace_arena;		/* Probe ID number.     */
219 #endif
220 static dtrace_probe_t	**dtrace_probes;	/* array of all probes */
221 static int		dtrace_nprobes;		/* number of probes */
222 static dtrace_provider_t *dtrace_provider;	/* provider list */
223 static dtrace_meta_t	*dtrace_meta_pid;	/* user-land meta provider */
224 static int		dtrace_opens;		/* number of opens */
225 static int		dtrace_helpers;		/* number of helpers */
226 static int		dtrace_getf;		/* number of unpriv getf()s */
227 #ifdef illumos
228 static void		*dtrace_softstate;	/* softstate pointer */
229 #endif
230 static dtrace_hash_t	*dtrace_bymod;		/* probes hashed by module */
231 static dtrace_hash_t	*dtrace_byfunc;		/* probes hashed by function */
232 static dtrace_hash_t	*dtrace_byname;		/* probes hashed by name */
233 static dtrace_toxrange_t *dtrace_toxrange;	/* toxic range array */
234 static int		dtrace_toxranges;	/* number of toxic ranges */
235 static int		dtrace_toxranges_max;	/* size of toxic range array */
236 static dtrace_anon_t	dtrace_anon;		/* anonymous enabling */
237 static kmem_cache_t	*dtrace_state_cache;	/* cache for dynamic state */
238 static uint64_t		dtrace_vtime_references; /* number of vtimestamp refs */
239 static kthread_t	*dtrace_panicked;	/* panicking thread */
240 static dtrace_ecb_t	*dtrace_ecb_create_cache; /* cached created ECB */
241 static dtrace_genid_t	dtrace_probegen;	/* current probe generation */
242 static dtrace_helpers_t *dtrace_deferred_pid;	/* deferred helper list */
243 static dtrace_enabling_t *dtrace_retained;	/* list of retained enablings */
244 static dtrace_genid_t	dtrace_retained_gen;	/* current retained enab gen */
245 static dtrace_dynvar_t	dtrace_dynhash_sink;	/* end of dynamic hash chains */
246 static int		dtrace_dynvar_failclean; /* dynvars failed to clean */
247 #ifndef illumos
248 static struct mtx	dtrace_unr_mtx;
249 MTX_SYSINIT(dtrace_unr_mtx, &dtrace_unr_mtx, "Unique resource identifier", MTX_DEF);
250 static eventhandler_tag	dtrace_kld_load_tag;
251 static eventhandler_tag	dtrace_kld_unload_try_tag;
252 #endif
253 
254 /*
255  * DTrace Locking
256  * DTrace is protected by three (relatively coarse-grained) locks:
257  *
258  * (1) dtrace_lock is required to manipulate essentially any DTrace state,
259  *     including enabling state, probes, ECBs, consumer state, helper state,
260  *     etc.  Importantly, dtrace_lock is _not_ required when in probe context;
261  *     probe context is lock-free -- synchronization is handled via the
262  *     dtrace_sync() cross call mechanism.
263  *
264  * (2) dtrace_provider_lock is required when manipulating provider state, or
265  *     when provider state must be held constant.
266  *
267  * (3) dtrace_meta_lock is required when manipulating meta provider state, or
268  *     when meta provider state must be held constant.
269  *
270  * The lock ordering between these three locks is dtrace_meta_lock before
271  * dtrace_provider_lock before dtrace_lock.  (In particular, there are
272  * several places where dtrace_provider_lock is held by the framework as it
273  * calls into the providers -- which then call back into the framework,
274  * grabbing dtrace_lock.)
275  *
276  * There are two other locks in the mix:  mod_lock and cpu_lock.  With respect
277  * to dtrace_provider_lock and dtrace_lock, cpu_lock continues its historical
278  * role as a coarse-grained lock; it is acquired before both of these locks.
279  * With respect to dtrace_meta_lock, its behavior is stranger:  cpu_lock must
280  * be acquired _between_ dtrace_meta_lock and any other DTrace locks.
281  * mod_lock is similar with respect to dtrace_provider_lock in that it must be
282  * acquired _between_ dtrace_provider_lock and dtrace_lock.
283  */
284 static kmutex_t		dtrace_lock;		/* probe state lock */
285 static kmutex_t		dtrace_provider_lock;	/* provider state lock */
286 static kmutex_t		dtrace_meta_lock;	/* meta-provider state lock */
287 
288 #ifndef illumos
289 /* XXX FreeBSD hacks. */
290 #define cr_suid		cr_svuid
291 #define cr_sgid		cr_svgid
292 #define	ipaddr_t	in_addr_t
293 #define mod_modname	pathname
294 #define vuprintf	vprintf
295 #define ttoproc(_a)	((_a)->td_proc)
296 #define crgetzoneid(_a)	0
297 #define	NCPU		MAXCPU
298 #define SNOCD		0
299 #define CPU_ON_INTR(_a)	0
300 
301 #define PRIV_EFFECTIVE		(1 << 0)
302 #define PRIV_DTRACE_KERNEL	(1 << 1)
303 #define PRIV_DTRACE_PROC	(1 << 2)
304 #define PRIV_DTRACE_USER	(1 << 3)
305 #define PRIV_PROC_OWNER		(1 << 4)
306 #define PRIV_PROC_ZONE		(1 << 5)
307 #define PRIV_ALL		~0
308 
309 SYSCTL_DECL(_debug_dtrace);
310 SYSCTL_DECL(_kern_dtrace);
311 #endif
312 
313 #ifdef illumos
314 #define curcpu	CPU->cpu_id
315 #endif
316 
317 
318 /*
319  * DTrace Provider Variables
320  *
321  * These are the variables relating to DTrace as a provider (that is, the
322  * provider of the BEGIN, END, and ERROR probes).
323  */
324 static dtrace_pattr_t	dtrace_provider_attr = {
325 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
326 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
327 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
328 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
329 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
330 };
331 
332 static void
333 dtrace_nullop(void)
334 {}
335 
336 static dtrace_pops_t	dtrace_provider_ops = {
337 	(void (*)(void *, dtrace_probedesc_t *))dtrace_nullop,
338 	(void (*)(void *, modctl_t *))dtrace_nullop,
339 	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
340 	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
341 	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
342 	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
343 	NULL,
344 	NULL,
345 	NULL,
346 	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop
347 };
348 
349 static dtrace_id_t	dtrace_probeid_begin;	/* special BEGIN probe */
350 static dtrace_id_t	dtrace_probeid_end;	/* special END probe */
351 dtrace_id_t		dtrace_probeid_error;	/* special ERROR probe */
352 
353 /*
354  * DTrace Helper Tracing Variables
355  *
356  * These variables should be set dynamically to enable helper tracing.  The
357  * only variables that should be set are dtrace_helptrace_enable (which should
358  * be set to a non-zero value to allocate helper tracing buffers on the next
359  * open of /dev/dtrace) and dtrace_helptrace_disable (which should be set to a
360  * non-zero value to deallocate helper tracing buffers on the next close of
361  * /dev/dtrace).  When (and only when) helper tracing is disabled, the
362  * buffer size may also be set via dtrace_helptrace_bufsize.
363  */
364 int			dtrace_helptrace_enable = 0;
365 int			dtrace_helptrace_disable = 0;
366 int			dtrace_helptrace_bufsize = 16 * 1024 * 1024;
367 uint32_t		dtrace_helptrace_nlocals;
368 static dtrace_helptrace_t *dtrace_helptrace_buffer;
369 static uint32_t		dtrace_helptrace_next = 0;
370 static int		dtrace_helptrace_wrapped = 0;
371 
372 /*
373  * DTrace Error Hashing
374  *
375  * On DEBUG kernels, DTrace will track the errors that has seen in a hash
376  * table.  This is very useful for checking coverage of tests that are
377  * expected to induce DIF or DOF processing errors, and may be useful for
378  * debugging problems in the DIF code generator or in DOF generation .  The
379  * error hash may be examined with the ::dtrace_errhash MDB dcmd.
380  */
381 #ifdef DEBUG
382 static dtrace_errhash_t	dtrace_errhash[DTRACE_ERRHASHSZ];
383 static const char *dtrace_errlast;
384 static kthread_t *dtrace_errthread;
385 static kmutex_t dtrace_errlock;
386 #endif
387 
388 /*
389  * DTrace Macros and Constants
390  *
391  * These are various macros that are useful in various spots in the
392  * implementation, along with a few random constants that have no meaning
393  * outside of the implementation.  There is no real structure to this cpp
394  * mishmash -- but is there ever?
395  */
396 #define	DTRACE_HASHSTR(hash, probe)	\
397 	dtrace_hash_str(*((char **)((uintptr_t)(probe) + (hash)->dth_stroffs)))
398 
399 #define	DTRACE_HASHNEXT(hash, probe)	\
400 	(dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_nextoffs)
401 
402 #define	DTRACE_HASHPREV(hash, probe)	\
403 	(dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_prevoffs)
404 
405 #define	DTRACE_HASHEQ(hash, lhs, rhs)	\
406 	(strcmp(*((char **)((uintptr_t)(lhs) + (hash)->dth_stroffs)), \
407 	    *((char **)((uintptr_t)(rhs) + (hash)->dth_stroffs))) == 0)
408 
409 #define	DTRACE_AGGHASHSIZE_SLEW		17
410 
411 #define	DTRACE_V4MAPPED_OFFSET		(sizeof (uint32_t) * 3)
412 
413 /*
414  * The key for a thread-local variable consists of the lower 61 bits of the
415  * t_did, plus the 3 bits of the highest active interrupt above LOCK_LEVEL.
416  * We add DIF_VARIABLE_MAX to t_did to assure that the thread key is never
417  * equal to a variable identifier.  This is necessary (but not sufficient) to
418  * assure that global associative arrays never collide with thread-local
419  * variables.  To guarantee that they cannot collide, we must also define the
420  * order for keying dynamic variables.  That order is:
421  *
422  *   [ key0 ] ... [ keyn ] [ variable-key ] [ tls-key ]
423  *
424  * Because the variable-key and the tls-key are in orthogonal spaces, there is
425  * no way for a global variable key signature to match a thread-local key
426  * signature.
427  */
428 #ifdef illumos
429 #define	DTRACE_TLS_THRKEY(where) { \
430 	uint_t intr = 0; \
431 	uint_t actv = CPU->cpu_intr_actv >> (LOCK_LEVEL + 1); \
432 	for (; actv; actv >>= 1) \
433 		intr++; \
434 	ASSERT(intr < (1 << 3)); \
435 	(where) = ((curthread->t_did + DIF_VARIABLE_MAX) & \
436 	    (((uint64_t)1 << 61) - 1)) | ((uint64_t)intr << 61); \
437 }
438 #else
439 #define	DTRACE_TLS_THRKEY(where) { \
440 	solaris_cpu_t *_c = &solaris_cpu[curcpu]; \
441 	uint_t intr = 0; \
442 	uint_t actv = _c->cpu_intr_actv; \
443 	for (; actv; actv >>= 1) \
444 		intr++; \
445 	ASSERT(intr < (1 << 3)); \
446 	(where) = ((curthread->td_tid + DIF_VARIABLE_MAX) & \
447 	    (((uint64_t)1 << 61) - 1)) | ((uint64_t)intr << 61); \
448 }
449 #endif
450 
451 #define	DT_BSWAP_8(x)	((x) & 0xff)
452 #define	DT_BSWAP_16(x)	((DT_BSWAP_8(x) << 8) | DT_BSWAP_8((x) >> 8))
453 #define	DT_BSWAP_32(x)	((DT_BSWAP_16(x) << 16) | DT_BSWAP_16((x) >> 16))
454 #define	DT_BSWAP_64(x)	((DT_BSWAP_32(x) << 32) | DT_BSWAP_32((x) >> 32))
455 
456 #define	DT_MASK_LO 0x00000000FFFFFFFFULL
457 
458 #define	DTRACE_STORE(type, tomax, offset, what) \
459 	*((type *)((uintptr_t)(tomax) + (uintptr_t)offset)) = (type)(what);
460 
461 #ifndef __x86
462 #define	DTRACE_ALIGNCHECK(addr, size, flags)				\
463 	if (addr & (size - 1)) {					\
464 		*flags |= CPU_DTRACE_BADALIGN;				\
465 		cpu_core[curcpu].cpuc_dtrace_illval = addr;	\
466 		return (0);						\
467 	}
468 #else
469 #define	DTRACE_ALIGNCHECK(addr, size, flags)
470 #endif
471 
472 /*
473  * Test whether a range of memory starting at testaddr of size testsz falls
474  * within the range of memory described by addr, sz.  We take care to avoid
475  * problems with overflow and underflow of the unsigned quantities, and
476  * disallow all negative sizes.  Ranges of size 0 are allowed.
477  */
478 #define	DTRACE_INRANGE(testaddr, testsz, baseaddr, basesz) \
479 	((testaddr) - (uintptr_t)(baseaddr) < (basesz) && \
480 	(testaddr) + (testsz) - (uintptr_t)(baseaddr) <= (basesz) && \
481 	(testaddr) + (testsz) >= (testaddr))
482 
483 /*
484  * Test whether alloc_sz bytes will fit in the scratch region.  We isolate
485  * alloc_sz on the righthand side of the comparison in order to avoid overflow
486  * or underflow in the comparison with it.  This is simpler than the INRANGE
487  * check above, because we know that the dtms_scratch_ptr is valid in the
488  * range.  Allocations of size zero are allowed.
489  */
490 #define	DTRACE_INSCRATCH(mstate, alloc_sz) \
491 	((mstate)->dtms_scratch_base + (mstate)->dtms_scratch_size - \
492 	(mstate)->dtms_scratch_ptr >= (alloc_sz))
493 
494 #define	DTRACE_LOADFUNC(bits)						\
495 /*CSTYLED*/								\
496 uint##bits##_t								\
497 dtrace_load##bits(uintptr_t addr)					\
498 {									\
499 	size_t size = bits / NBBY;					\
500 	/*CSTYLED*/							\
501 	uint##bits##_t rval;						\
502 	int i;								\
503 	volatile uint16_t *flags = (volatile uint16_t *)		\
504 	    &cpu_core[curcpu].cpuc_dtrace_flags;			\
505 									\
506 	DTRACE_ALIGNCHECK(addr, size, flags);				\
507 									\
508 	for (i = 0; i < dtrace_toxranges; i++) {			\
509 		if (addr >= dtrace_toxrange[i].dtt_limit)		\
510 			continue;					\
511 									\
512 		if (addr + size <= dtrace_toxrange[i].dtt_base)		\
513 			continue;					\
514 									\
515 		/*							\
516 		 * This address falls within a toxic region; return 0.	\
517 		 */							\
518 		*flags |= CPU_DTRACE_BADADDR;				\
519 		cpu_core[curcpu].cpuc_dtrace_illval = addr;		\
520 		return (0);						\
521 	}								\
522 									\
523 	*flags |= CPU_DTRACE_NOFAULT;					\
524 	/*CSTYLED*/							\
525 	rval = *((volatile uint##bits##_t *)addr);			\
526 	*flags &= ~CPU_DTRACE_NOFAULT;					\
527 									\
528 	return (!(*flags & CPU_DTRACE_FAULT) ? rval : 0);		\
529 }
530 
531 #ifdef _LP64
532 #define	dtrace_loadptr	dtrace_load64
533 #else
534 #define	dtrace_loadptr	dtrace_load32
535 #endif
536 
537 #define	DTRACE_DYNHASH_FREE	0
538 #define	DTRACE_DYNHASH_SINK	1
539 #define	DTRACE_DYNHASH_VALID	2
540 
541 #define	DTRACE_MATCH_NEXT	0
542 #define	DTRACE_MATCH_DONE	1
543 #define	DTRACE_ANCHORED(probe)	((probe)->dtpr_func[0] != '\0')
544 #define	DTRACE_STATE_ALIGN	64
545 
546 #define	DTRACE_FLAGS2FLT(flags)						\
547 	(((flags) & CPU_DTRACE_BADADDR) ? DTRACEFLT_BADADDR :		\
548 	((flags) & CPU_DTRACE_ILLOP) ? DTRACEFLT_ILLOP :		\
549 	((flags) & CPU_DTRACE_DIVZERO) ? DTRACEFLT_DIVZERO :		\
550 	((flags) & CPU_DTRACE_KPRIV) ? DTRACEFLT_KPRIV :		\
551 	((flags) & CPU_DTRACE_UPRIV) ? DTRACEFLT_UPRIV :		\
552 	((flags) & CPU_DTRACE_TUPOFLOW) ?  DTRACEFLT_TUPOFLOW :		\
553 	((flags) & CPU_DTRACE_BADALIGN) ?  DTRACEFLT_BADALIGN :		\
554 	((flags) & CPU_DTRACE_NOSCRATCH) ?  DTRACEFLT_NOSCRATCH :	\
555 	((flags) & CPU_DTRACE_BADSTACK) ?  DTRACEFLT_BADSTACK :		\
556 	DTRACEFLT_UNKNOWN)
557 
558 #define	DTRACEACT_ISSTRING(act)						\
559 	((act)->dta_kind == DTRACEACT_DIFEXPR &&			\
560 	(act)->dta_difo->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING)
561 
562 /* Function prototype definitions: */
563 static size_t dtrace_strlen(const char *, size_t);
564 static dtrace_probe_t *dtrace_probe_lookup_id(dtrace_id_t id);
565 static void dtrace_enabling_provide(dtrace_provider_t *);
566 static int dtrace_enabling_match(dtrace_enabling_t *, int *);
567 static void dtrace_enabling_matchall(void);
568 static void dtrace_enabling_reap(void);
569 static dtrace_state_t *dtrace_anon_grab(void);
570 static uint64_t dtrace_helper(int, dtrace_mstate_t *,
571     dtrace_state_t *, uint64_t, uint64_t);
572 static dtrace_helpers_t *dtrace_helpers_create(proc_t *);
573 static void dtrace_buffer_drop(dtrace_buffer_t *);
574 static int dtrace_buffer_consumed(dtrace_buffer_t *, hrtime_t when);
575 static intptr_t dtrace_buffer_reserve(dtrace_buffer_t *, size_t, size_t,
576     dtrace_state_t *, dtrace_mstate_t *);
577 static int dtrace_state_option(dtrace_state_t *, dtrace_optid_t,
578     dtrace_optval_t);
579 static int dtrace_ecb_create_enable(dtrace_probe_t *, void *);
580 static void dtrace_helper_provider_destroy(dtrace_helper_provider_t *);
581 uint16_t dtrace_load16(uintptr_t);
582 uint32_t dtrace_load32(uintptr_t);
583 uint64_t dtrace_load64(uintptr_t);
584 uint8_t dtrace_load8(uintptr_t);
585 void dtrace_dynvar_clean(dtrace_dstate_t *);
586 dtrace_dynvar_t *dtrace_dynvar(dtrace_dstate_t *, uint_t, dtrace_key_t *,
587     size_t, dtrace_dynvar_op_t, dtrace_mstate_t *, dtrace_vstate_t *);
588 uintptr_t dtrace_dif_varstr(uintptr_t, dtrace_state_t *, dtrace_mstate_t *);
589 static int dtrace_priv_proc(dtrace_state_t *);
590 static void dtrace_getf_barrier(void);
591 
592 /*
593  * DTrace Probe Context Functions
594  *
595  * These functions are called from probe context.  Because probe context is
596  * any context in which C may be called, arbitrarily locks may be held,
597  * interrupts may be disabled, we may be in arbitrary dispatched state, etc.
598  * As a result, functions called from probe context may only call other DTrace
599  * support functions -- they may not interact at all with the system at large.
600  * (Note that the ASSERT macro is made probe-context safe by redefining it in
601  * terms of dtrace_assfail(), a probe-context safe function.) If arbitrary
602  * loads are to be performed from probe context, they _must_ be in terms of
603  * the safe dtrace_load*() variants.
604  *
605  * Some functions in this block are not actually called from probe context;
606  * for these functions, there will be a comment above the function reading
607  * "Note:  not called from probe context."
608  */
609 void
610 dtrace_panic(const char *format, ...)
611 {
612 	va_list alist;
613 
614 	va_start(alist, format);
615 #ifdef __FreeBSD__
616 	vpanic(format, alist);
617 #else
618 	dtrace_vpanic(format, alist);
619 #endif
620 	va_end(alist);
621 }
622 
623 int
624 dtrace_assfail(const char *a, const char *f, int l)
625 {
626 	dtrace_panic("assertion failed: %s, file: %s, line: %d", a, f, l);
627 
628 	/*
629 	 * We just need something here that even the most clever compiler
630 	 * cannot optimize away.
631 	 */
632 	return (a[(uintptr_t)f]);
633 }
634 
635 /*
636  * Atomically increment a specified error counter from probe context.
637  */
638 static void
639 dtrace_error(uint32_t *counter)
640 {
641 	/*
642 	 * Most counters stored to in probe context are per-CPU counters.
643 	 * However, there are some error conditions that are sufficiently
644 	 * arcane that they don't merit per-CPU storage.  If these counters
645 	 * are incremented concurrently on different CPUs, scalability will be
646 	 * adversely affected -- but we don't expect them to be white-hot in a
647 	 * correctly constructed enabling...
648 	 */
649 	uint32_t oval, nval;
650 
651 	do {
652 		oval = *counter;
653 
654 		if ((nval = oval + 1) == 0) {
655 			/*
656 			 * If the counter would wrap, set it to 1 -- assuring
657 			 * that the counter is never zero when we have seen
658 			 * errors.  (The counter must be 32-bits because we
659 			 * aren't guaranteed a 64-bit compare&swap operation.)
660 			 * To save this code both the infamy of being fingered
661 			 * by a priggish news story and the indignity of being
662 			 * the target of a neo-puritan witch trial, we're
663 			 * carefully avoiding any colorful description of the
664 			 * likelihood of this condition -- but suffice it to
665 			 * say that it is only slightly more likely than the
666 			 * overflow of predicate cache IDs, as discussed in
667 			 * dtrace_predicate_create().
668 			 */
669 			nval = 1;
670 		}
671 	} while (dtrace_cas32(counter, oval, nval) != oval);
672 }
673 
674 /*
675  * Use the DTRACE_LOADFUNC macro to define functions for each of loading a
676  * uint8_t, a uint16_t, a uint32_t and a uint64_t.
677  */
678 /* BEGIN CSTYLED */
679 DTRACE_LOADFUNC(8)
680 DTRACE_LOADFUNC(16)
681 DTRACE_LOADFUNC(32)
682 DTRACE_LOADFUNC(64)
683 /* END CSTYLED */
684 
685 static int
686 dtrace_inscratch(uintptr_t dest, size_t size, dtrace_mstate_t *mstate)
687 {
688 	if (dest < mstate->dtms_scratch_base)
689 		return (0);
690 
691 	if (dest + size < dest)
692 		return (0);
693 
694 	if (dest + size > mstate->dtms_scratch_ptr)
695 		return (0);
696 
697 	return (1);
698 }
699 
700 static int
701 dtrace_canstore_statvar(uint64_t addr, size_t sz,
702     dtrace_statvar_t **svars, int nsvars)
703 {
704 	int i;
705 	size_t maxglobalsize, maxlocalsize;
706 
707 	if (nsvars == 0)
708 		return (0);
709 
710 	maxglobalsize = dtrace_statvar_maxsize;
711 	maxlocalsize = (maxglobalsize + sizeof (uint64_t)) * NCPU;
712 
713 	for (i = 0; i < nsvars; i++) {
714 		dtrace_statvar_t *svar = svars[i];
715 		uint8_t scope;
716 		size_t size;
717 
718 		if (svar == NULL || (size = svar->dtsv_size) == 0)
719 			continue;
720 
721 		scope = svar->dtsv_var.dtdv_scope;
722 
723 		/*
724 		 * We verify that our size is valid in the spirit of providing
725 		 * defense in depth:  we want to prevent attackers from using
726 		 * DTrace to escalate an orthogonal kernel heap corruption bug
727 		 * into the ability to store to arbitrary locations in memory.
728 		 */
729 		VERIFY((scope == DIFV_SCOPE_GLOBAL && size < maxglobalsize) ||
730 		    (scope == DIFV_SCOPE_LOCAL && size < maxlocalsize));
731 
732 		if (DTRACE_INRANGE(addr, sz, svar->dtsv_data, svar->dtsv_size))
733 			return (1);
734 	}
735 
736 	return (0);
737 }
738 
739 /*
740  * Check to see if the address is within a memory region to which a store may
741  * be issued.  This includes the DTrace scratch areas, and any DTrace variable
742  * region.  The caller of dtrace_canstore() is responsible for performing any
743  * alignment checks that are needed before stores are actually executed.
744  */
745 static int
746 dtrace_canstore(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
747     dtrace_vstate_t *vstate)
748 {
749 	/*
750 	 * First, check to see if the address is in scratch space...
751 	 */
752 	if (DTRACE_INRANGE(addr, sz, mstate->dtms_scratch_base,
753 	    mstate->dtms_scratch_size))
754 		return (1);
755 
756 	/*
757 	 * Now check to see if it's a dynamic variable.  This check will pick
758 	 * up both thread-local variables and any global dynamically-allocated
759 	 * variables.
760 	 */
761 	if (DTRACE_INRANGE(addr, sz, vstate->dtvs_dynvars.dtds_base,
762 	    vstate->dtvs_dynvars.dtds_size)) {
763 		dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
764 		uintptr_t base = (uintptr_t)dstate->dtds_base +
765 		    (dstate->dtds_hashsize * sizeof (dtrace_dynhash_t));
766 		uintptr_t chunkoffs;
767 		dtrace_dynvar_t *dvar;
768 
769 		/*
770 		 * Before we assume that we can store here, we need to make
771 		 * sure that it isn't in our metadata -- storing to our
772 		 * dynamic variable metadata would corrupt our state.  For
773 		 * the range to not include any dynamic variable metadata,
774 		 * it must:
775 		 *
776 		 *	(1) Start above the hash table that is at the base of
777 		 *	the dynamic variable space
778 		 *
779 		 *	(2) Have a starting chunk offset that is beyond the
780 		 *	dtrace_dynvar_t that is at the base of every chunk
781 		 *
782 		 *	(3) Not span a chunk boundary
783 		 *
784 		 *	(4) Not be in the tuple space of a dynamic variable
785 		 *
786 		 */
787 		if (addr < base)
788 			return (0);
789 
790 		chunkoffs = (addr - base) % dstate->dtds_chunksize;
791 
792 		if (chunkoffs < sizeof (dtrace_dynvar_t))
793 			return (0);
794 
795 		if (chunkoffs + sz > dstate->dtds_chunksize)
796 			return (0);
797 
798 		dvar = (dtrace_dynvar_t *)((uintptr_t)addr - chunkoffs);
799 
800 		if (dvar->dtdv_hashval == DTRACE_DYNHASH_FREE)
801 			return (0);
802 
803 		if (chunkoffs < sizeof (dtrace_dynvar_t) +
804 		    ((dvar->dtdv_tuple.dtt_nkeys - 1) * sizeof (dtrace_key_t)))
805 			return (0);
806 
807 		return (1);
808 	}
809 
810 	/*
811 	 * Finally, check the static local and global variables.  These checks
812 	 * take the longest, so we perform them last.
813 	 */
814 	if (dtrace_canstore_statvar(addr, sz,
815 	    vstate->dtvs_locals, vstate->dtvs_nlocals))
816 		return (1);
817 
818 	if (dtrace_canstore_statvar(addr, sz,
819 	    vstate->dtvs_globals, vstate->dtvs_nglobals))
820 		return (1);
821 
822 	return (0);
823 }
824 
825 
826 /*
827  * Convenience routine to check to see if the address is within a memory
828  * region in which a load may be issued given the user's privilege level;
829  * if not, it sets the appropriate error flags and loads 'addr' into the
830  * illegal value slot.
831  *
832  * DTrace subroutines (DIF_SUBR_*) should use this helper to implement
833  * appropriate memory access protection.
834  */
835 static int
836 dtrace_canload(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
837     dtrace_vstate_t *vstate)
838 {
839 	volatile uintptr_t *illval = &cpu_core[curcpu].cpuc_dtrace_illval;
840 	file_t *fp;
841 
842 	/*
843 	 * If we hold the privilege to read from kernel memory, then
844 	 * everything is readable.
845 	 */
846 	if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
847 		return (1);
848 
849 	/*
850 	 * You can obviously read that which you can store.
851 	 */
852 	if (dtrace_canstore(addr, sz, mstate, vstate))
853 		return (1);
854 
855 	/*
856 	 * We're allowed to read from our own string table.
857 	 */
858 	if (DTRACE_INRANGE(addr, sz, mstate->dtms_difo->dtdo_strtab,
859 	    mstate->dtms_difo->dtdo_strlen))
860 		return (1);
861 
862 	if (vstate->dtvs_state != NULL &&
863 	    dtrace_priv_proc(vstate->dtvs_state)) {
864 		proc_t *p;
865 
866 		/*
867 		 * When we have privileges to the current process, there are
868 		 * several context-related kernel structures that are safe to
869 		 * read, even absent the privilege to read from kernel memory.
870 		 * These reads are safe because these structures contain only
871 		 * state that (1) we're permitted to read, (2) is harmless or
872 		 * (3) contains pointers to additional kernel state that we're
873 		 * not permitted to read (and as such, do not present an
874 		 * opportunity for privilege escalation).  Finally (and
875 		 * critically), because of the nature of their relation with
876 		 * the current thread context, the memory associated with these
877 		 * structures cannot change over the duration of probe context,
878 		 * and it is therefore impossible for this memory to be
879 		 * deallocated and reallocated as something else while it's
880 		 * being operated upon.
881 		 */
882 		if (DTRACE_INRANGE(addr, sz, curthread, sizeof (kthread_t)))
883 			return (1);
884 
885 		if ((p = curthread->t_procp) != NULL && DTRACE_INRANGE(addr,
886 		    sz, curthread->t_procp, sizeof (proc_t))) {
887 			return (1);
888 		}
889 
890 		if (curthread->t_cred != NULL && DTRACE_INRANGE(addr, sz,
891 		    curthread->t_cred, sizeof (cred_t))) {
892 			return (1);
893 		}
894 
895 #ifdef illumos
896 		if (p != NULL && p->p_pidp != NULL && DTRACE_INRANGE(addr, sz,
897 		    &(p->p_pidp->pid_id), sizeof (pid_t))) {
898 			return (1);
899 		}
900 
901 		if (curthread->t_cpu != NULL && DTRACE_INRANGE(addr, sz,
902 		    curthread->t_cpu, offsetof(cpu_t, cpu_pause_thread))) {
903 			return (1);
904 		}
905 #endif
906 	}
907 
908 	if ((fp = mstate->dtms_getf) != NULL) {
909 		uintptr_t psz = sizeof (void *);
910 		vnode_t *vp;
911 		vnodeops_t *op;
912 
913 		/*
914 		 * When getf() returns a file_t, the enabling is implicitly
915 		 * granted the (transient) right to read the returned file_t
916 		 * as well as the v_path and v_op->vnop_name of the underlying
917 		 * vnode.  These accesses are allowed after a successful
918 		 * getf() because the members that they refer to cannot change
919 		 * once set -- and the barrier logic in the kernel's closef()
920 		 * path assures that the file_t and its referenced vode_t
921 		 * cannot themselves be stale (that is, it impossible for
922 		 * either dtms_getf itself or its f_vnode member to reference
923 		 * freed memory).
924 		 */
925 		if (DTRACE_INRANGE(addr, sz, fp, sizeof (file_t)))
926 			return (1);
927 
928 		if ((vp = fp->f_vnode) != NULL) {
929 #ifdef illumos
930 			if (DTRACE_INRANGE(addr, sz, &vp->v_path, psz))
931 				return (1);
932 			if (vp->v_path != NULL && DTRACE_INRANGE(addr, sz,
933 			    vp->v_path, strlen(vp->v_path) + 1)) {
934 				return (1);
935 			}
936 #endif
937 
938 			if (DTRACE_INRANGE(addr, sz, &vp->v_op, psz))
939 				return (1);
940 
941 #ifdef illumos
942 			if ((op = vp->v_op) != NULL &&
943 			    DTRACE_INRANGE(addr, sz, &op->vnop_name, psz)) {
944 				return (1);
945 			}
946 
947 			if (op != NULL && op->vnop_name != NULL &&
948 			    DTRACE_INRANGE(addr, sz, op->vnop_name,
949 			    strlen(op->vnop_name) + 1)) {
950 				return (1);
951 			}
952 #endif
953 		}
954 	}
955 
956 	DTRACE_CPUFLAG_SET(CPU_DTRACE_KPRIV);
957 	*illval = addr;
958 	return (0);
959 }
960 
961 /*
962  * Convenience routine to check to see if a given string is within a memory
963  * region in which a load may be issued given the user's privilege level;
964  * this exists so that we don't need to issue unnecessary dtrace_strlen()
965  * calls in the event that the user has all privileges.
966  */
967 static int
968 dtrace_strcanload(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
969     dtrace_vstate_t *vstate)
970 {
971 	size_t strsz;
972 
973 	/*
974 	 * If we hold the privilege to read from kernel memory, then
975 	 * everything is readable.
976 	 */
977 	if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
978 		return (1);
979 
980 	strsz = 1 + dtrace_strlen((char *)(uintptr_t)addr, sz);
981 	if (dtrace_canload(addr, strsz, mstate, vstate))
982 		return (1);
983 
984 	return (0);
985 }
986 
987 /*
988  * Convenience routine to check to see if a given variable is within a memory
989  * region in which a load may be issued given the user's privilege level.
990  */
991 static int
992 dtrace_vcanload(void *src, dtrace_diftype_t *type, dtrace_mstate_t *mstate,
993     dtrace_vstate_t *vstate)
994 {
995 	size_t sz;
996 	ASSERT(type->dtdt_flags & DIF_TF_BYREF);
997 
998 	/*
999 	 * If we hold the privilege to read from kernel memory, then
1000 	 * everything is readable.
1001 	 */
1002 	if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
1003 		return (1);
1004 
1005 	if (type->dtdt_kind == DIF_TYPE_STRING)
1006 		sz = dtrace_strlen(src,
1007 		    vstate->dtvs_state->dts_options[DTRACEOPT_STRSIZE]) + 1;
1008 	else
1009 		sz = type->dtdt_size;
1010 
1011 	return (dtrace_canload((uintptr_t)src, sz, mstate, vstate));
1012 }
1013 
1014 /*
1015  * Convert a string to a signed integer using safe loads.
1016  *
1017  * NOTE: This function uses various macros from strtolctype.h to manipulate
1018  * digit values, etc -- these have all been checked to ensure they make
1019  * no additional function calls.
1020  */
1021 static int64_t
1022 dtrace_strtoll(char *input, int base, size_t limit)
1023 {
1024 	uintptr_t pos = (uintptr_t)input;
1025 	int64_t val = 0;
1026 	int x;
1027 	boolean_t neg = B_FALSE;
1028 	char c, cc, ccc;
1029 	uintptr_t end = pos + limit;
1030 
1031 	/*
1032 	 * Consume any whitespace preceding digits.
1033 	 */
1034 	while ((c = dtrace_load8(pos)) == ' ' || c == '\t')
1035 		pos++;
1036 
1037 	/*
1038 	 * Handle an explicit sign if one is present.
1039 	 */
1040 	if (c == '-' || c == '+') {
1041 		if (c == '-')
1042 			neg = B_TRUE;
1043 		c = dtrace_load8(++pos);
1044 	}
1045 
1046 	/*
1047 	 * Check for an explicit hexadecimal prefix ("0x" or "0X") and skip it
1048 	 * if present.
1049 	 */
1050 	if (base == 16 && c == '0' && ((cc = dtrace_load8(pos + 1)) == 'x' ||
1051 	    cc == 'X') && isxdigit(ccc = dtrace_load8(pos + 2))) {
1052 		pos += 2;
1053 		c = ccc;
1054 	}
1055 
1056 	/*
1057 	 * Read in contiguous digits until the first non-digit character.
1058 	 */
1059 	for (; pos < end && c != '\0' && lisalnum(c) && (x = DIGIT(c)) < base;
1060 	    c = dtrace_load8(++pos))
1061 		val = val * base + x;
1062 
1063 	return (neg ? -val : val);
1064 }
1065 
1066 /*
1067  * Compare two strings using safe loads.
1068  */
1069 static int
1070 dtrace_strncmp(char *s1, char *s2, size_t limit)
1071 {
1072 	uint8_t c1, c2;
1073 	volatile uint16_t *flags;
1074 
1075 	if (s1 == s2 || limit == 0)
1076 		return (0);
1077 
1078 	flags = (volatile uint16_t *)&cpu_core[curcpu].cpuc_dtrace_flags;
1079 
1080 	do {
1081 		if (s1 == NULL) {
1082 			c1 = '\0';
1083 		} else {
1084 			c1 = dtrace_load8((uintptr_t)s1++);
1085 		}
1086 
1087 		if (s2 == NULL) {
1088 			c2 = '\0';
1089 		} else {
1090 			c2 = dtrace_load8((uintptr_t)s2++);
1091 		}
1092 
1093 		if (c1 != c2)
1094 			return (c1 - c2);
1095 	} while (--limit && c1 != '\0' && !(*flags & CPU_DTRACE_FAULT));
1096 
1097 	return (0);
1098 }
1099 
1100 /*
1101  * Compute strlen(s) for a string using safe memory accesses.  The additional
1102  * len parameter is used to specify a maximum length to ensure completion.
1103  */
1104 static size_t
1105 dtrace_strlen(const char *s, size_t lim)
1106 {
1107 	uint_t len;
1108 
1109 	for (len = 0; len != lim; len++) {
1110 		if (dtrace_load8((uintptr_t)s++) == '\0')
1111 			break;
1112 	}
1113 
1114 	return (len);
1115 }
1116 
1117 /*
1118  * Check if an address falls within a toxic region.
1119  */
1120 static int
1121 dtrace_istoxic(uintptr_t kaddr, size_t size)
1122 {
1123 	uintptr_t taddr, tsize;
1124 	int i;
1125 
1126 	for (i = 0; i < dtrace_toxranges; i++) {
1127 		taddr = dtrace_toxrange[i].dtt_base;
1128 		tsize = dtrace_toxrange[i].dtt_limit - taddr;
1129 
1130 		if (kaddr - taddr < tsize) {
1131 			DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
1132 			cpu_core[curcpu].cpuc_dtrace_illval = kaddr;
1133 			return (1);
1134 		}
1135 
1136 		if (taddr - kaddr < size) {
1137 			DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
1138 			cpu_core[curcpu].cpuc_dtrace_illval = taddr;
1139 			return (1);
1140 		}
1141 	}
1142 
1143 	return (0);
1144 }
1145 
1146 /*
1147  * Copy src to dst using safe memory accesses.  The src is assumed to be unsafe
1148  * memory specified by the DIF program.  The dst is assumed to be safe memory
1149  * that we can store to directly because it is managed by DTrace.  As with
1150  * standard bcopy, overlapping copies are handled properly.
1151  */
1152 static void
1153 dtrace_bcopy(const void *src, void *dst, size_t len)
1154 {
1155 	if (len != 0) {
1156 		uint8_t *s1 = dst;
1157 		const uint8_t *s2 = src;
1158 
1159 		if (s1 <= s2) {
1160 			do {
1161 				*s1++ = dtrace_load8((uintptr_t)s2++);
1162 			} while (--len != 0);
1163 		} else {
1164 			s2 += len;
1165 			s1 += len;
1166 
1167 			do {
1168 				*--s1 = dtrace_load8((uintptr_t)--s2);
1169 			} while (--len != 0);
1170 		}
1171 	}
1172 }
1173 
1174 /*
1175  * Copy src to dst using safe memory accesses, up to either the specified
1176  * length, or the point that a nul byte is encountered.  The src is assumed to
1177  * be unsafe memory specified by the DIF program.  The dst is assumed to be
1178  * safe memory that we can store to directly because it is managed by DTrace.
1179  * Unlike dtrace_bcopy(), overlapping regions are not handled.
1180  */
1181 static void
1182 dtrace_strcpy(const void *src, void *dst, size_t len)
1183 {
1184 	if (len != 0) {
1185 		uint8_t *s1 = dst, c;
1186 		const uint8_t *s2 = src;
1187 
1188 		do {
1189 			*s1++ = c = dtrace_load8((uintptr_t)s2++);
1190 		} while (--len != 0 && c != '\0');
1191 	}
1192 }
1193 
1194 /*
1195  * Copy src to dst, deriving the size and type from the specified (BYREF)
1196  * variable type.  The src is assumed to be unsafe memory specified by the DIF
1197  * program.  The dst is assumed to be DTrace variable memory that is of the
1198  * specified type; we assume that we can store to directly.
1199  */
1200 static void
1201 dtrace_vcopy(void *src, void *dst, dtrace_diftype_t *type)
1202 {
1203 	ASSERT(type->dtdt_flags & DIF_TF_BYREF);
1204 
1205 	if (type->dtdt_kind == DIF_TYPE_STRING) {
1206 		dtrace_strcpy(src, dst, type->dtdt_size);
1207 	} else {
1208 		dtrace_bcopy(src, dst, type->dtdt_size);
1209 	}
1210 }
1211 
1212 /*
1213  * Compare s1 to s2 using safe memory accesses.  The s1 data is assumed to be
1214  * unsafe memory specified by the DIF program.  The s2 data is assumed to be
1215  * safe memory that we can access directly because it is managed by DTrace.
1216  */
1217 static int
1218 dtrace_bcmp(const void *s1, const void *s2, size_t len)
1219 {
1220 	volatile uint16_t *flags;
1221 
1222 	flags = (volatile uint16_t *)&cpu_core[curcpu].cpuc_dtrace_flags;
1223 
1224 	if (s1 == s2)
1225 		return (0);
1226 
1227 	if (s1 == NULL || s2 == NULL)
1228 		return (1);
1229 
1230 	if (s1 != s2 && len != 0) {
1231 		const uint8_t *ps1 = s1;
1232 		const uint8_t *ps2 = s2;
1233 
1234 		do {
1235 			if (dtrace_load8((uintptr_t)ps1++) != *ps2++)
1236 				return (1);
1237 		} while (--len != 0 && !(*flags & CPU_DTRACE_FAULT));
1238 	}
1239 	return (0);
1240 }
1241 
1242 /*
1243  * Zero the specified region using a simple byte-by-byte loop.  Note that this
1244  * is for safe DTrace-managed memory only.
1245  */
1246 static void
1247 dtrace_bzero(void *dst, size_t len)
1248 {
1249 	uchar_t *cp;
1250 
1251 	for (cp = dst; len != 0; len--)
1252 		*cp++ = 0;
1253 }
1254 
1255 static void
1256 dtrace_add_128(uint64_t *addend1, uint64_t *addend2, uint64_t *sum)
1257 {
1258 	uint64_t result[2];
1259 
1260 	result[0] = addend1[0] + addend2[0];
1261 	result[1] = addend1[1] + addend2[1] +
1262 	    (result[0] < addend1[0] || result[0] < addend2[0] ? 1 : 0);
1263 
1264 	sum[0] = result[0];
1265 	sum[1] = result[1];
1266 }
1267 
1268 /*
1269  * Shift the 128-bit value in a by b. If b is positive, shift left.
1270  * If b is negative, shift right.
1271  */
1272 static void
1273 dtrace_shift_128(uint64_t *a, int b)
1274 {
1275 	uint64_t mask;
1276 
1277 	if (b == 0)
1278 		return;
1279 
1280 	if (b < 0) {
1281 		b = -b;
1282 		if (b >= 64) {
1283 			a[0] = a[1] >> (b - 64);
1284 			a[1] = 0;
1285 		} else {
1286 			a[0] >>= b;
1287 			mask = 1LL << (64 - b);
1288 			mask -= 1;
1289 			a[0] |= ((a[1] & mask) << (64 - b));
1290 			a[1] >>= b;
1291 		}
1292 	} else {
1293 		if (b >= 64) {
1294 			a[1] = a[0] << (b - 64);
1295 			a[0] = 0;
1296 		} else {
1297 			a[1] <<= b;
1298 			mask = a[0] >> (64 - b);
1299 			a[1] |= mask;
1300 			a[0] <<= b;
1301 		}
1302 	}
1303 }
1304 
1305 /*
1306  * The basic idea is to break the 2 64-bit values into 4 32-bit values,
1307  * use native multiplication on those, and then re-combine into the
1308  * resulting 128-bit value.
1309  *
1310  * (hi1 << 32 + lo1) * (hi2 << 32 + lo2) =
1311  *     hi1 * hi2 << 64 +
1312  *     hi1 * lo2 << 32 +
1313  *     hi2 * lo1 << 32 +
1314  *     lo1 * lo2
1315  */
1316 static void
1317 dtrace_multiply_128(uint64_t factor1, uint64_t factor2, uint64_t *product)
1318 {
1319 	uint64_t hi1, hi2, lo1, lo2;
1320 	uint64_t tmp[2];
1321 
1322 	hi1 = factor1 >> 32;
1323 	hi2 = factor2 >> 32;
1324 
1325 	lo1 = factor1 & DT_MASK_LO;
1326 	lo2 = factor2 & DT_MASK_LO;
1327 
1328 	product[0] = lo1 * lo2;
1329 	product[1] = hi1 * hi2;
1330 
1331 	tmp[0] = hi1 * lo2;
1332 	tmp[1] = 0;
1333 	dtrace_shift_128(tmp, 32);
1334 	dtrace_add_128(product, tmp, product);
1335 
1336 	tmp[0] = hi2 * lo1;
1337 	tmp[1] = 0;
1338 	dtrace_shift_128(tmp, 32);
1339 	dtrace_add_128(product, tmp, product);
1340 }
1341 
1342 /*
1343  * This privilege check should be used by actions and subroutines to
1344  * verify that the user credentials of the process that enabled the
1345  * invoking ECB match the target credentials
1346  */
1347 static int
1348 dtrace_priv_proc_common_user(dtrace_state_t *state)
1349 {
1350 	cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1351 
1352 	/*
1353 	 * We should always have a non-NULL state cred here, since if cred
1354 	 * is null (anonymous tracing), we fast-path bypass this routine.
1355 	 */
1356 	ASSERT(s_cr != NULL);
1357 
1358 	if ((cr = CRED()) != NULL &&
1359 	    s_cr->cr_uid == cr->cr_uid &&
1360 	    s_cr->cr_uid == cr->cr_ruid &&
1361 	    s_cr->cr_uid == cr->cr_suid &&
1362 	    s_cr->cr_gid == cr->cr_gid &&
1363 	    s_cr->cr_gid == cr->cr_rgid &&
1364 	    s_cr->cr_gid == cr->cr_sgid)
1365 		return (1);
1366 
1367 	return (0);
1368 }
1369 
1370 /*
1371  * This privilege check should be used by actions and subroutines to
1372  * verify that the zone of the process that enabled the invoking ECB
1373  * matches the target credentials
1374  */
1375 static int
1376 dtrace_priv_proc_common_zone(dtrace_state_t *state)
1377 {
1378 #ifdef illumos
1379 	cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1380 
1381 	/*
1382 	 * We should always have a non-NULL state cred here, since if cred
1383 	 * is null (anonymous tracing), we fast-path bypass this routine.
1384 	 */
1385 	ASSERT(s_cr != NULL);
1386 
1387 	if ((cr = CRED()) != NULL && s_cr->cr_zone == cr->cr_zone)
1388 		return (1);
1389 
1390 	return (0);
1391 #else
1392 	return (1);
1393 #endif
1394 }
1395 
1396 /*
1397  * This privilege check should be used by actions and subroutines to
1398  * verify that the process has not setuid or changed credentials.
1399  */
1400 static int
1401 dtrace_priv_proc_common_nocd(void)
1402 {
1403 	proc_t *proc;
1404 
1405 	if ((proc = ttoproc(curthread)) != NULL &&
1406 	    !(proc->p_flag & SNOCD))
1407 		return (1);
1408 
1409 	return (0);
1410 }
1411 
1412 static int
1413 dtrace_priv_proc_destructive(dtrace_state_t *state)
1414 {
1415 	int action = state->dts_cred.dcr_action;
1416 
1417 	if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE) == 0) &&
1418 	    dtrace_priv_proc_common_zone(state) == 0)
1419 		goto bad;
1420 
1421 	if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER) == 0) &&
1422 	    dtrace_priv_proc_common_user(state) == 0)
1423 		goto bad;
1424 
1425 	if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG) == 0) &&
1426 	    dtrace_priv_proc_common_nocd() == 0)
1427 		goto bad;
1428 
1429 	return (1);
1430 
1431 bad:
1432 	cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1433 
1434 	return (0);
1435 }
1436 
1437 static int
1438 dtrace_priv_proc_control(dtrace_state_t *state)
1439 {
1440 	if (state->dts_cred.dcr_action & DTRACE_CRA_PROC_CONTROL)
1441 		return (1);
1442 
1443 	if (dtrace_priv_proc_common_zone(state) &&
1444 	    dtrace_priv_proc_common_user(state) &&
1445 	    dtrace_priv_proc_common_nocd())
1446 		return (1);
1447 
1448 	cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1449 
1450 	return (0);
1451 }
1452 
1453 static int
1454 dtrace_priv_proc(dtrace_state_t *state)
1455 {
1456 	if (state->dts_cred.dcr_action & DTRACE_CRA_PROC)
1457 		return (1);
1458 
1459 	cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1460 
1461 	return (0);
1462 }
1463 
1464 static int
1465 dtrace_priv_kernel(dtrace_state_t *state)
1466 {
1467 	if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL)
1468 		return (1);
1469 
1470 	cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1471 
1472 	return (0);
1473 }
1474 
1475 static int
1476 dtrace_priv_kernel_destructive(dtrace_state_t *state)
1477 {
1478 	if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL_DESTRUCTIVE)
1479 		return (1);
1480 
1481 	cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1482 
1483 	return (0);
1484 }
1485 
1486 /*
1487  * Determine if the dte_cond of the specified ECB allows for processing of
1488  * the current probe to continue.  Note that this routine may allow continued
1489  * processing, but with access(es) stripped from the mstate's dtms_access
1490  * field.
1491  */
1492 static int
1493 dtrace_priv_probe(dtrace_state_t *state, dtrace_mstate_t *mstate,
1494     dtrace_ecb_t *ecb)
1495 {
1496 	dtrace_probe_t *probe = ecb->dte_probe;
1497 	dtrace_provider_t *prov = probe->dtpr_provider;
1498 	dtrace_pops_t *pops = &prov->dtpv_pops;
1499 	int mode = DTRACE_MODE_NOPRIV_DROP;
1500 
1501 	ASSERT(ecb->dte_cond);
1502 
1503 #ifdef illumos
1504 	if (pops->dtps_mode != NULL) {
1505 		mode = pops->dtps_mode(prov->dtpv_arg,
1506 		    probe->dtpr_id, probe->dtpr_arg);
1507 
1508 		ASSERT((mode & DTRACE_MODE_USER) ||
1509 		    (mode & DTRACE_MODE_KERNEL));
1510 		ASSERT((mode & DTRACE_MODE_NOPRIV_RESTRICT) ||
1511 		    (mode & DTRACE_MODE_NOPRIV_DROP));
1512 	}
1513 
1514 	/*
1515 	 * If the dte_cond bits indicate that this consumer is only allowed to
1516 	 * see user-mode firings of this probe, call the provider's dtps_mode()
1517 	 * entry point to check that the probe was fired while in a user
1518 	 * context.  If that's not the case, use the policy specified by the
1519 	 * provider to determine if we drop the probe or merely restrict
1520 	 * operation.
1521 	 */
1522 	if (ecb->dte_cond & DTRACE_COND_USERMODE) {
1523 		ASSERT(mode != DTRACE_MODE_NOPRIV_DROP);
1524 
1525 		if (!(mode & DTRACE_MODE_USER)) {
1526 			if (mode & DTRACE_MODE_NOPRIV_DROP)
1527 				return (0);
1528 
1529 			mstate->dtms_access &= ~DTRACE_ACCESS_ARGS;
1530 		}
1531 	}
1532 #endif
1533 
1534 	/*
1535 	 * This is more subtle than it looks. We have to be absolutely certain
1536 	 * that CRED() isn't going to change out from under us so it's only
1537 	 * legit to examine that structure if we're in constrained situations.
1538 	 * Currently, the only times we'll this check is if a non-super-user
1539 	 * has enabled the profile or syscall providers -- providers that
1540 	 * allow visibility of all processes. For the profile case, the check
1541 	 * above will ensure that we're examining a user context.
1542 	 */
1543 	if (ecb->dte_cond & DTRACE_COND_OWNER) {
1544 		cred_t *cr;
1545 		cred_t *s_cr = state->dts_cred.dcr_cred;
1546 		proc_t *proc;
1547 
1548 		ASSERT(s_cr != NULL);
1549 
1550 		if ((cr = CRED()) == NULL ||
1551 		    s_cr->cr_uid != cr->cr_uid ||
1552 		    s_cr->cr_uid != cr->cr_ruid ||
1553 		    s_cr->cr_uid != cr->cr_suid ||
1554 		    s_cr->cr_gid != cr->cr_gid ||
1555 		    s_cr->cr_gid != cr->cr_rgid ||
1556 		    s_cr->cr_gid != cr->cr_sgid ||
1557 		    (proc = ttoproc(curthread)) == NULL ||
1558 		    (proc->p_flag & SNOCD)) {
1559 			if (mode & DTRACE_MODE_NOPRIV_DROP)
1560 				return (0);
1561 
1562 #ifdef illumos
1563 			mstate->dtms_access &= ~DTRACE_ACCESS_PROC;
1564 #endif
1565 		}
1566 	}
1567 
1568 #ifdef illumos
1569 	/*
1570 	 * If our dte_cond is set to DTRACE_COND_ZONEOWNER and we are not
1571 	 * in our zone, check to see if our mode policy is to restrict rather
1572 	 * than to drop; if to restrict, strip away both DTRACE_ACCESS_PROC
1573 	 * and DTRACE_ACCESS_ARGS
1574 	 */
1575 	if (ecb->dte_cond & DTRACE_COND_ZONEOWNER) {
1576 		cred_t *cr;
1577 		cred_t *s_cr = state->dts_cred.dcr_cred;
1578 
1579 		ASSERT(s_cr != NULL);
1580 
1581 		if ((cr = CRED()) == NULL ||
1582 		    s_cr->cr_zone->zone_id != cr->cr_zone->zone_id) {
1583 			if (mode & DTRACE_MODE_NOPRIV_DROP)
1584 				return (0);
1585 
1586 			mstate->dtms_access &=
1587 			    ~(DTRACE_ACCESS_PROC | DTRACE_ACCESS_ARGS);
1588 		}
1589 	}
1590 #endif
1591 
1592 	return (1);
1593 }
1594 
1595 /*
1596  * Note:  not called from probe context.  This function is called
1597  * asynchronously (and at a regular interval) from outside of probe context to
1598  * clean the dirty dynamic variable lists on all CPUs.  Dynamic variable
1599  * cleaning is explained in detail in <sys/dtrace_impl.h>.
1600  */
1601 void
1602 dtrace_dynvar_clean(dtrace_dstate_t *dstate)
1603 {
1604 	dtrace_dynvar_t *dirty;
1605 	dtrace_dstate_percpu_t *dcpu;
1606 	dtrace_dynvar_t **rinsep;
1607 	int i, j, work = 0;
1608 
1609 	for (i = 0; i < NCPU; i++) {
1610 		dcpu = &dstate->dtds_percpu[i];
1611 		rinsep = &dcpu->dtdsc_rinsing;
1612 
1613 		/*
1614 		 * If the dirty list is NULL, there is no dirty work to do.
1615 		 */
1616 		if (dcpu->dtdsc_dirty == NULL)
1617 			continue;
1618 
1619 		if (dcpu->dtdsc_rinsing != NULL) {
1620 			/*
1621 			 * If the rinsing list is non-NULL, then it is because
1622 			 * this CPU was selected to accept another CPU's
1623 			 * dirty list -- and since that time, dirty buffers
1624 			 * have accumulated.  This is a highly unlikely
1625 			 * condition, but we choose to ignore the dirty
1626 			 * buffers -- they'll be picked up a future cleanse.
1627 			 */
1628 			continue;
1629 		}
1630 
1631 		if (dcpu->dtdsc_clean != NULL) {
1632 			/*
1633 			 * If the clean list is non-NULL, then we're in a
1634 			 * situation where a CPU has done deallocations (we
1635 			 * have a non-NULL dirty list) but no allocations (we
1636 			 * also have a non-NULL clean list).  We can't simply
1637 			 * move the dirty list into the clean list on this
1638 			 * CPU, yet we also don't want to allow this condition
1639 			 * to persist, lest a short clean list prevent a
1640 			 * massive dirty list from being cleaned (which in
1641 			 * turn could lead to otherwise avoidable dynamic
1642 			 * drops).  To deal with this, we look for some CPU
1643 			 * with a NULL clean list, NULL dirty list, and NULL
1644 			 * rinsing list -- and then we borrow this CPU to
1645 			 * rinse our dirty list.
1646 			 */
1647 			for (j = 0; j < NCPU; j++) {
1648 				dtrace_dstate_percpu_t *rinser;
1649 
1650 				rinser = &dstate->dtds_percpu[j];
1651 
1652 				if (rinser->dtdsc_rinsing != NULL)
1653 					continue;
1654 
1655 				if (rinser->dtdsc_dirty != NULL)
1656 					continue;
1657 
1658 				if (rinser->dtdsc_clean != NULL)
1659 					continue;
1660 
1661 				rinsep = &rinser->dtdsc_rinsing;
1662 				break;
1663 			}
1664 
1665 			if (j == NCPU) {
1666 				/*
1667 				 * We were unable to find another CPU that
1668 				 * could accept this dirty list -- we are
1669 				 * therefore unable to clean it now.
1670 				 */
1671 				dtrace_dynvar_failclean++;
1672 				continue;
1673 			}
1674 		}
1675 
1676 		work = 1;
1677 
1678 		/*
1679 		 * Atomically move the dirty list aside.
1680 		 */
1681 		do {
1682 			dirty = dcpu->dtdsc_dirty;
1683 
1684 			/*
1685 			 * Before we zap the dirty list, set the rinsing list.
1686 			 * (This allows for a potential assertion in
1687 			 * dtrace_dynvar():  if a free dynamic variable appears
1688 			 * on a hash chain, either the dirty list or the
1689 			 * rinsing list for some CPU must be non-NULL.)
1690 			 */
1691 			*rinsep = dirty;
1692 			dtrace_membar_producer();
1693 		} while (dtrace_casptr(&dcpu->dtdsc_dirty,
1694 		    dirty, NULL) != dirty);
1695 	}
1696 
1697 	if (!work) {
1698 		/*
1699 		 * We have no work to do; we can simply return.
1700 		 */
1701 		return;
1702 	}
1703 
1704 	dtrace_sync();
1705 
1706 	for (i = 0; i < NCPU; i++) {
1707 		dcpu = &dstate->dtds_percpu[i];
1708 
1709 		if (dcpu->dtdsc_rinsing == NULL)
1710 			continue;
1711 
1712 		/*
1713 		 * We are now guaranteed that no hash chain contains a pointer
1714 		 * into this dirty list; we can make it clean.
1715 		 */
1716 		ASSERT(dcpu->dtdsc_clean == NULL);
1717 		dcpu->dtdsc_clean = dcpu->dtdsc_rinsing;
1718 		dcpu->dtdsc_rinsing = NULL;
1719 	}
1720 
1721 	/*
1722 	 * Before we actually set the state to be DTRACE_DSTATE_CLEAN, make
1723 	 * sure that all CPUs have seen all of the dtdsc_clean pointers.
1724 	 * This prevents a race whereby a CPU incorrectly decides that
1725 	 * the state should be something other than DTRACE_DSTATE_CLEAN
1726 	 * after dtrace_dynvar_clean() has completed.
1727 	 */
1728 	dtrace_sync();
1729 
1730 	dstate->dtds_state = DTRACE_DSTATE_CLEAN;
1731 }
1732 
1733 /*
1734  * Depending on the value of the op parameter, this function looks-up,
1735  * allocates or deallocates an arbitrarily-keyed dynamic variable.  If an
1736  * allocation is requested, this function will return a pointer to a
1737  * dtrace_dynvar_t corresponding to the allocated variable -- or NULL if no
1738  * variable can be allocated.  If NULL is returned, the appropriate counter
1739  * will be incremented.
1740  */
1741 dtrace_dynvar_t *
1742 dtrace_dynvar(dtrace_dstate_t *dstate, uint_t nkeys,
1743     dtrace_key_t *key, size_t dsize, dtrace_dynvar_op_t op,
1744     dtrace_mstate_t *mstate, dtrace_vstate_t *vstate)
1745 {
1746 	uint64_t hashval = DTRACE_DYNHASH_VALID;
1747 	dtrace_dynhash_t *hash = dstate->dtds_hash;
1748 	dtrace_dynvar_t *free, *new_free, *next, *dvar, *start, *prev = NULL;
1749 	processorid_t me = curcpu, cpu = me;
1750 	dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[me];
1751 	size_t bucket, ksize;
1752 	size_t chunksize = dstate->dtds_chunksize;
1753 	uintptr_t kdata, lock, nstate;
1754 	uint_t i;
1755 
1756 	ASSERT(nkeys != 0);
1757 
1758 	/*
1759 	 * Hash the key.  As with aggregations, we use Jenkins' "One-at-a-time"
1760 	 * algorithm.  For the by-value portions, we perform the algorithm in
1761 	 * 16-bit chunks (as opposed to 8-bit chunks).  This speeds things up a
1762 	 * bit, and seems to have only a minute effect on distribution.  For
1763 	 * the by-reference data, we perform "One-at-a-time" iterating (safely)
1764 	 * over each referenced byte.  It's painful to do this, but it's much
1765 	 * better than pathological hash distribution.  The efficacy of the
1766 	 * hashing algorithm (and a comparison with other algorithms) may be
1767 	 * found by running the ::dtrace_dynstat MDB dcmd.
1768 	 */
1769 	for (i = 0; i < nkeys; i++) {
1770 		if (key[i].dttk_size == 0) {
1771 			uint64_t val = key[i].dttk_value;
1772 
1773 			hashval += (val >> 48) & 0xffff;
1774 			hashval += (hashval << 10);
1775 			hashval ^= (hashval >> 6);
1776 
1777 			hashval += (val >> 32) & 0xffff;
1778 			hashval += (hashval << 10);
1779 			hashval ^= (hashval >> 6);
1780 
1781 			hashval += (val >> 16) & 0xffff;
1782 			hashval += (hashval << 10);
1783 			hashval ^= (hashval >> 6);
1784 
1785 			hashval += val & 0xffff;
1786 			hashval += (hashval << 10);
1787 			hashval ^= (hashval >> 6);
1788 		} else {
1789 			/*
1790 			 * This is incredibly painful, but it beats the hell
1791 			 * out of the alternative.
1792 			 */
1793 			uint64_t j, size = key[i].dttk_size;
1794 			uintptr_t base = (uintptr_t)key[i].dttk_value;
1795 
1796 			if (!dtrace_canload(base, size, mstate, vstate))
1797 				break;
1798 
1799 			for (j = 0; j < size; j++) {
1800 				hashval += dtrace_load8(base + j);
1801 				hashval += (hashval << 10);
1802 				hashval ^= (hashval >> 6);
1803 			}
1804 		}
1805 	}
1806 
1807 	if (DTRACE_CPUFLAG_ISSET(CPU_DTRACE_FAULT))
1808 		return (NULL);
1809 
1810 	hashval += (hashval << 3);
1811 	hashval ^= (hashval >> 11);
1812 	hashval += (hashval << 15);
1813 
1814 	/*
1815 	 * There is a remote chance (ideally, 1 in 2^31) that our hashval
1816 	 * comes out to be one of our two sentinel hash values.  If this
1817 	 * actually happens, we set the hashval to be a value known to be a
1818 	 * non-sentinel value.
1819 	 */
1820 	if (hashval == DTRACE_DYNHASH_FREE || hashval == DTRACE_DYNHASH_SINK)
1821 		hashval = DTRACE_DYNHASH_VALID;
1822 
1823 	/*
1824 	 * Yes, it's painful to do a divide here.  If the cycle count becomes
1825 	 * important here, tricks can be pulled to reduce it.  (However, it's
1826 	 * critical that hash collisions be kept to an absolute minimum;
1827 	 * they're much more painful than a divide.)  It's better to have a
1828 	 * solution that generates few collisions and still keeps things
1829 	 * relatively simple.
1830 	 */
1831 	bucket = hashval % dstate->dtds_hashsize;
1832 
1833 	if (op == DTRACE_DYNVAR_DEALLOC) {
1834 		volatile uintptr_t *lockp = &hash[bucket].dtdh_lock;
1835 
1836 		for (;;) {
1837 			while ((lock = *lockp) & 1)
1838 				continue;
1839 
1840 			if (dtrace_casptr((volatile void *)lockp,
1841 			    (volatile void *)lock, (volatile void *)(lock + 1)) == (void *)lock)
1842 				break;
1843 		}
1844 
1845 		dtrace_membar_producer();
1846 	}
1847 
1848 top:
1849 	prev = NULL;
1850 	lock = hash[bucket].dtdh_lock;
1851 
1852 	dtrace_membar_consumer();
1853 
1854 	start = hash[bucket].dtdh_chain;
1855 	ASSERT(start != NULL && (start->dtdv_hashval == DTRACE_DYNHASH_SINK ||
1856 	    start->dtdv_hashval != DTRACE_DYNHASH_FREE ||
1857 	    op != DTRACE_DYNVAR_DEALLOC));
1858 
1859 	for (dvar = start; dvar != NULL; dvar = dvar->dtdv_next) {
1860 		dtrace_tuple_t *dtuple = &dvar->dtdv_tuple;
1861 		dtrace_key_t *dkey = &dtuple->dtt_key[0];
1862 
1863 		if (dvar->dtdv_hashval != hashval) {
1864 			if (dvar->dtdv_hashval == DTRACE_DYNHASH_SINK) {
1865 				/*
1866 				 * We've reached the sink, and therefore the
1867 				 * end of the hash chain; we can kick out of
1868 				 * the loop knowing that we have seen a valid
1869 				 * snapshot of state.
1870 				 */
1871 				ASSERT(dvar->dtdv_next == NULL);
1872 				ASSERT(dvar == &dtrace_dynhash_sink);
1873 				break;
1874 			}
1875 
1876 			if (dvar->dtdv_hashval == DTRACE_DYNHASH_FREE) {
1877 				/*
1878 				 * We've gone off the rails:  somewhere along
1879 				 * the line, one of the members of this hash
1880 				 * chain was deleted.  Note that we could also
1881 				 * detect this by simply letting this loop run
1882 				 * to completion, as we would eventually hit
1883 				 * the end of the dirty list.  However, we
1884 				 * want to avoid running the length of the
1885 				 * dirty list unnecessarily (it might be quite
1886 				 * long), so we catch this as early as
1887 				 * possible by detecting the hash marker.  In
1888 				 * this case, we simply set dvar to NULL and
1889 				 * break; the conditional after the loop will
1890 				 * send us back to top.
1891 				 */
1892 				dvar = NULL;
1893 				break;
1894 			}
1895 
1896 			goto next;
1897 		}
1898 
1899 		if (dtuple->dtt_nkeys != nkeys)
1900 			goto next;
1901 
1902 		for (i = 0; i < nkeys; i++, dkey++) {
1903 			if (dkey->dttk_size != key[i].dttk_size)
1904 				goto next; /* size or type mismatch */
1905 
1906 			if (dkey->dttk_size != 0) {
1907 				if (dtrace_bcmp(
1908 				    (void *)(uintptr_t)key[i].dttk_value,
1909 				    (void *)(uintptr_t)dkey->dttk_value,
1910 				    dkey->dttk_size))
1911 					goto next;
1912 			} else {
1913 				if (dkey->dttk_value != key[i].dttk_value)
1914 					goto next;
1915 			}
1916 		}
1917 
1918 		if (op != DTRACE_DYNVAR_DEALLOC)
1919 			return (dvar);
1920 
1921 		ASSERT(dvar->dtdv_next == NULL ||
1922 		    dvar->dtdv_next->dtdv_hashval != DTRACE_DYNHASH_FREE);
1923 
1924 		if (prev != NULL) {
1925 			ASSERT(hash[bucket].dtdh_chain != dvar);
1926 			ASSERT(start != dvar);
1927 			ASSERT(prev->dtdv_next == dvar);
1928 			prev->dtdv_next = dvar->dtdv_next;
1929 		} else {
1930 			if (dtrace_casptr(&hash[bucket].dtdh_chain,
1931 			    start, dvar->dtdv_next) != start) {
1932 				/*
1933 				 * We have failed to atomically swing the
1934 				 * hash table head pointer, presumably because
1935 				 * of a conflicting allocation on another CPU.
1936 				 * We need to reread the hash chain and try
1937 				 * again.
1938 				 */
1939 				goto top;
1940 			}
1941 		}
1942 
1943 		dtrace_membar_producer();
1944 
1945 		/*
1946 		 * Now set the hash value to indicate that it's free.
1947 		 */
1948 		ASSERT(hash[bucket].dtdh_chain != dvar);
1949 		dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
1950 
1951 		dtrace_membar_producer();
1952 
1953 		/*
1954 		 * Set the next pointer to point at the dirty list, and
1955 		 * atomically swing the dirty pointer to the newly freed dvar.
1956 		 */
1957 		do {
1958 			next = dcpu->dtdsc_dirty;
1959 			dvar->dtdv_next = next;
1960 		} while (dtrace_casptr(&dcpu->dtdsc_dirty, next, dvar) != next);
1961 
1962 		/*
1963 		 * Finally, unlock this hash bucket.
1964 		 */
1965 		ASSERT(hash[bucket].dtdh_lock == lock);
1966 		ASSERT(lock & 1);
1967 		hash[bucket].dtdh_lock++;
1968 
1969 		return (NULL);
1970 next:
1971 		prev = dvar;
1972 		continue;
1973 	}
1974 
1975 	if (dvar == NULL) {
1976 		/*
1977 		 * If dvar is NULL, it is because we went off the rails:
1978 		 * one of the elements that we traversed in the hash chain
1979 		 * was deleted while we were traversing it.  In this case,
1980 		 * we assert that we aren't doing a dealloc (deallocs lock
1981 		 * the hash bucket to prevent themselves from racing with
1982 		 * one another), and retry the hash chain traversal.
1983 		 */
1984 		ASSERT(op != DTRACE_DYNVAR_DEALLOC);
1985 		goto top;
1986 	}
1987 
1988 	if (op != DTRACE_DYNVAR_ALLOC) {
1989 		/*
1990 		 * If we are not to allocate a new variable, we want to
1991 		 * return NULL now.  Before we return, check that the value
1992 		 * of the lock word hasn't changed.  If it has, we may have
1993 		 * seen an inconsistent snapshot.
1994 		 */
1995 		if (op == DTRACE_DYNVAR_NOALLOC) {
1996 			if (hash[bucket].dtdh_lock != lock)
1997 				goto top;
1998 		} else {
1999 			ASSERT(op == DTRACE_DYNVAR_DEALLOC);
2000 			ASSERT(hash[bucket].dtdh_lock == lock);
2001 			ASSERT(lock & 1);
2002 			hash[bucket].dtdh_lock++;
2003 		}
2004 
2005 		return (NULL);
2006 	}
2007 
2008 	/*
2009 	 * We need to allocate a new dynamic variable.  The size we need is the
2010 	 * size of dtrace_dynvar plus the size of nkeys dtrace_key_t's plus the
2011 	 * size of any auxiliary key data (rounded up to 8-byte alignment) plus
2012 	 * the size of any referred-to data (dsize).  We then round the final
2013 	 * size up to the chunksize for allocation.
2014 	 */
2015 	for (ksize = 0, i = 0; i < nkeys; i++)
2016 		ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
2017 
2018 	/*
2019 	 * This should be pretty much impossible, but could happen if, say,
2020 	 * strange DIF specified the tuple.  Ideally, this should be an
2021 	 * assertion and not an error condition -- but that requires that the
2022 	 * chunksize calculation in dtrace_difo_chunksize() be absolutely
2023 	 * bullet-proof.  (That is, it must not be able to be fooled by
2024 	 * malicious DIF.)  Given the lack of backwards branches in DIF,
2025 	 * solving this would presumably not amount to solving the Halting
2026 	 * Problem -- but it still seems awfully hard.
2027 	 */
2028 	if (sizeof (dtrace_dynvar_t) + sizeof (dtrace_key_t) * (nkeys - 1) +
2029 	    ksize + dsize > chunksize) {
2030 		dcpu->dtdsc_drops++;
2031 		return (NULL);
2032 	}
2033 
2034 	nstate = DTRACE_DSTATE_EMPTY;
2035 
2036 	do {
2037 retry:
2038 		free = dcpu->dtdsc_free;
2039 
2040 		if (free == NULL) {
2041 			dtrace_dynvar_t *clean = dcpu->dtdsc_clean;
2042 			void *rval;
2043 
2044 			if (clean == NULL) {
2045 				/*
2046 				 * We're out of dynamic variable space on
2047 				 * this CPU.  Unless we have tried all CPUs,
2048 				 * we'll try to allocate from a different
2049 				 * CPU.
2050 				 */
2051 				switch (dstate->dtds_state) {
2052 				case DTRACE_DSTATE_CLEAN: {
2053 					void *sp = &dstate->dtds_state;
2054 
2055 					if (++cpu >= NCPU)
2056 						cpu = 0;
2057 
2058 					if (dcpu->dtdsc_dirty != NULL &&
2059 					    nstate == DTRACE_DSTATE_EMPTY)
2060 						nstate = DTRACE_DSTATE_DIRTY;
2061 
2062 					if (dcpu->dtdsc_rinsing != NULL)
2063 						nstate = DTRACE_DSTATE_RINSING;
2064 
2065 					dcpu = &dstate->dtds_percpu[cpu];
2066 
2067 					if (cpu != me)
2068 						goto retry;
2069 
2070 					(void) dtrace_cas32(sp,
2071 					    DTRACE_DSTATE_CLEAN, nstate);
2072 
2073 					/*
2074 					 * To increment the correct bean
2075 					 * counter, take another lap.
2076 					 */
2077 					goto retry;
2078 				}
2079 
2080 				case DTRACE_DSTATE_DIRTY:
2081 					dcpu->dtdsc_dirty_drops++;
2082 					break;
2083 
2084 				case DTRACE_DSTATE_RINSING:
2085 					dcpu->dtdsc_rinsing_drops++;
2086 					break;
2087 
2088 				case DTRACE_DSTATE_EMPTY:
2089 					dcpu->dtdsc_drops++;
2090 					break;
2091 				}
2092 
2093 				DTRACE_CPUFLAG_SET(CPU_DTRACE_DROP);
2094 				return (NULL);
2095 			}
2096 
2097 			/*
2098 			 * The clean list appears to be non-empty.  We want to
2099 			 * move the clean list to the free list; we start by
2100 			 * moving the clean pointer aside.
2101 			 */
2102 			if (dtrace_casptr(&dcpu->dtdsc_clean,
2103 			    clean, NULL) != clean) {
2104 				/*
2105 				 * We are in one of two situations:
2106 				 *
2107 				 *  (a)	The clean list was switched to the
2108 				 *	free list by another CPU.
2109 				 *
2110 				 *  (b)	The clean list was added to by the
2111 				 *	cleansing cyclic.
2112 				 *
2113 				 * In either of these situations, we can
2114 				 * just reattempt the free list allocation.
2115 				 */
2116 				goto retry;
2117 			}
2118 
2119 			ASSERT(clean->dtdv_hashval == DTRACE_DYNHASH_FREE);
2120 
2121 			/*
2122 			 * Now we'll move the clean list to our free list.
2123 			 * It's impossible for this to fail:  the only way
2124 			 * the free list can be updated is through this
2125 			 * code path, and only one CPU can own the clean list.
2126 			 * Thus, it would only be possible for this to fail if
2127 			 * this code were racing with dtrace_dynvar_clean().
2128 			 * (That is, if dtrace_dynvar_clean() updated the clean
2129 			 * list, and we ended up racing to update the free
2130 			 * list.)  This race is prevented by the dtrace_sync()
2131 			 * in dtrace_dynvar_clean() -- which flushes the
2132 			 * owners of the clean lists out before resetting
2133 			 * the clean lists.
2134 			 */
2135 			dcpu = &dstate->dtds_percpu[me];
2136 			rval = dtrace_casptr(&dcpu->dtdsc_free, NULL, clean);
2137 			ASSERT(rval == NULL);
2138 			goto retry;
2139 		}
2140 
2141 		dvar = free;
2142 		new_free = dvar->dtdv_next;
2143 	} while (dtrace_casptr(&dcpu->dtdsc_free, free, new_free) != free);
2144 
2145 	/*
2146 	 * We have now allocated a new chunk.  We copy the tuple keys into the
2147 	 * tuple array and copy any referenced key data into the data space
2148 	 * following the tuple array.  As we do this, we relocate dttk_value
2149 	 * in the final tuple to point to the key data address in the chunk.
2150 	 */
2151 	kdata = (uintptr_t)&dvar->dtdv_tuple.dtt_key[nkeys];
2152 	dvar->dtdv_data = (void *)(kdata + ksize);
2153 	dvar->dtdv_tuple.dtt_nkeys = nkeys;
2154 
2155 	for (i = 0; i < nkeys; i++) {
2156 		dtrace_key_t *dkey = &dvar->dtdv_tuple.dtt_key[i];
2157 		size_t kesize = key[i].dttk_size;
2158 
2159 		if (kesize != 0) {
2160 			dtrace_bcopy(
2161 			    (const void *)(uintptr_t)key[i].dttk_value,
2162 			    (void *)kdata, kesize);
2163 			dkey->dttk_value = kdata;
2164 			kdata += P2ROUNDUP(kesize, sizeof (uint64_t));
2165 		} else {
2166 			dkey->dttk_value = key[i].dttk_value;
2167 		}
2168 
2169 		dkey->dttk_size = kesize;
2170 	}
2171 
2172 	ASSERT(dvar->dtdv_hashval == DTRACE_DYNHASH_FREE);
2173 	dvar->dtdv_hashval = hashval;
2174 	dvar->dtdv_next = start;
2175 
2176 	if (dtrace_casptr(&hash[bucket].dtdh_chain, start, dvar) == start)
2177 		return (dvar);
2178 
2179 	/*
2180 	 * The cas has failed.  Either another CPU is adding an element to
2181 	 * this hash chain, or another CPU is deleting an element from this
2182 	 * hash chain.  The simplest way to deal with both of these cases
2183 	 * (though not necessarily the most efficient) is to free our
2184 	 * allocated block and re-attempt it all.  Note that the free is
2185 	 * to the dirty list and _not_ to the free list.  This is to prevent
2186 	 * races with allocators, above.
2187 	 */
2188 	dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
2189 
2190 	dtrace_membar_producer();
2191 
2192 	do {
2193 		free = dcpu->dtdsc_dirty;
2194 		dvar->dtdv_next = free;
2195 	} while (dtrace_casptr(&dcpu->dtdsc_dirty, free, dvar) != free);
2196 
2197 	goto top;
2198 }
2199 
2200 /*ARGSUSED*/
2201 static void
2202 dtrace_aggregate_min(uint64_t *oval, uint64_t nval, uint64_t arg)
2203 {
2204 	if ((int64_t)nval < (int64_t)*oval)
2205 		*oval = nval;
2206 }
2207 
2208 /*ARGSUSED*/
2209 static void
2210 dtrace_aggregate_max(uint64_t *oval, uint64_t nval, uint64_t arg)
2211 {
2212 	if ((int64_t)nval > (int64_t)*oval)
2213 		*oval = nval;
2214 }
2215 
2216 static void
2217 dtrace_aggregate_quantize(uint64_t *quanta, uint64_t nval, uint64_t incr)
2218 {
2219 	int i, zero = DTRACE_QUANTIZE_ZEROBUCKET;
2220 	int64_t val = (int64_t)nval;
2221 
2222 	if (val < 0) {
2223 		for (i = 0; i < zero; i++) {
2224 			if (val <= DTRACE_QUANTIZE_BUCKETVAL(i)) {
2225 				quanta[i] += incr;
2226 				return;
2227 			}
2228 		}
2229 	} else {
2230 		for (i = zero + 1; i < DTRACE_QUANTIZE_NBUCKETS; i++) {
2231 			if (val < DTRACE_QUANTIZE_BUCKETVAL(i)) {
2232 				quanta[i - 1] += incr;
2233 				return;
2234 			}
2235 		}
2236 
2237 		quanta[DTRACE_QUANTIZE_NBUCKETS - 1] += incr;
2238 		return;
2239 	}
2240 
2241 	ASSERT(0);
2242 }
2243 
2244 static void
2245 dtrace_aggregate_lquantize(uint64_t *lquanta, uint64_t nval, uint64_t incr)
2246 {
2247 	uint64_t arg = *lquanta++;
2248 	int32_t base = DTRACE_LQUANTIZE_BASE(arg);
2249 	uint16_t step = DTRACE_LQUANTIZE_STEP(arg);
2250 	uint16_t levels = DTRACE_LQUANTIZE_LEVELS(arg);
2251 	int32_t val = (int32_t)nval, level;
2252 
2253 	ASSERT(step != 0);
2254 	ASSERT(levels != 0);
2255 
2256 	if (val < base) {
2257 		/*
2258 		 * This is an underflow.
2259 		 */
2260 		lquanta[0] += incr;
2261 		return;
2262 	}
2263 
2264 	level = (val - base) / step;
2265 
2266 	if (level < levels) {
2267 		lquanta[level + 1] += incr;
2268 		return;
2269 	}
2270 
2271 	/*
2272 	 * This is an overflow.
2273 	 */
2274 	lquanta[levels + 1] += incr;
2275 }
2276 
2277 static int
2278 dtrace_aggregate_llquantize_bucket(uint16_t factor, uint16_t low,
2279     uint16_t high, uint16_t nsteps, int64_t value)
2280 {
2281 	int64_t this = 1, last, next;
2282 	int base = 1, order;
2283 
2284 	ASSERT(factor <= nsteps);
2285 	ASSERT(nsteps % factor == 0);
2286 
2287 	for (order = 0; order < low; order++)
2288 		this *= factor;
2289 
2290 	/*
2291 	 * If our value is less than our factor taken to the power of the
2292 	 * low order of magnitude, it goes into the zeroth bucket.
2293 	 */
2294 	if (value < (last = this))
2295 		return (0);
2296 
2297 	for (this *= factor; order <= high; order++) {
2298 		int nbuckets = this > nsteps ? nsteps : this;
2299 
2300 		if ((next = this * factor) < this) {
2301 			/*
2302 			 * We should not generally get log/linear quantizations
2303 			 * with a high magnitude that allows 64-bits to
2304 			 * overflow, but we nonetheless protect against this
2305 			 * by explicitly checking for overflow, and clamping
2306 			 * our value accordingly.
2307 			 */
2308 			value = this - 1;
2309 		}
2310 
2311 		if (value < this) {
2312 			/*
2313 			 * If our value lies within this order of magnitude,
2314 			 * determine its position by taking the offset within
2315 			 * the order of magnitude, dividing by the bucket
2316 			 * width, and adding to our (accumulated) base.
2317 			 */
2318 			return (base + (value - last) / (this / nbuckets));
2319 		}
2320 
2321 		base += nbuckets - (nbuckets / factor);
2322 		last = this;
2323 		this = next;
2324 	}
2325 
2326 	/*
2327 	 * Our value is greater than or equal to our factor taken to the
2328 	 * power of one plus the high magnitude -- return the top bucket.
2329 	 */
2330 	return (base);
2331 }
2332 
2333 static void
2334 dtrace_aggregate_llquantize(uint64_t *llquanta, uint64_t nval, uint64_t incr)
2335 {
2336 	uint64_t arg = *llquanta++;
2337 	uint16_t factor = DTRACE_LLQUANTIZE_FACTOR(arg);
2338 	uint16_t low = DTRACE_LLQUANTIZE_LOW(arg);
2339 	uint16_t high = DTRACE_LLQUANTIZE_HIGH(arg);
2340 	uint16_t nsteps = DTRACE_LLQUANTIZE_NSTEP(arg);
2341 
2342 	llquanta[dtrace_aggregate_llquantize_bucket(factor,
2343 	    low, high, nsteps, nval)] += incr;
2344 }
2345 
2346 /*ARGSUSED*/
2347 static void
2348 dtrace_aggregate_avg(uint64_t *data, uint64_t nval, uint64_t arg)
2349 {
2350 	data[0]++;
2351 	data[1] += nval;
2352 }
2353 
2354 /*ARGSUSED*/
2355 static void
2356 dtrace_aggregate_stddev(uint64_t *data, uint64_t nval, uint64_t arg)
2357 {
2358 	int64_t snval = (int64_t)nval;
2359 	uint64_t tmp[2];
2360 
2361 	data[0]++;
2362 	data[1] += nval;
2363 
2364 	/*
2365 	 * What we want to say here is:
2366 	 *
2367 	 * data[2] += nval * nval;
2368 	 *
2369 	 * But given that nval is 64-bit, we could easily overflow, so
2370 	 * we do this as 128-bit arithmetic.
2371 	 */
2372 	if (snval < 0)
2373 		snval = -snval;
2374 
2375 	dtrace_multiply_128((uint64_t)snval, (uint64_t)snval, tmp);
2376 	dtrace_add_128(data + 2, tmp, data + 2);
2377 }
2378 
2379 /*ARGSUSED*/
2380 static void
2381 dtrace_aggregate_count(uint64_t *oval, uint64_t nval, uint64_t arg)
2382 {
2383 	*oval = *oval + 1;
2384 }
2385 
2386 /*ARGSUSED*/
2387 static void
2388 dtrace_aggregate_sum(uint64_t *oval, uint64_t nval, uint64_t arg)
2389 {
2390 	*oval += nval;
2391 }
2392 
2393 /*
2394  * Aggregate given the tuple in the principal data buffer, and the aggregating
2395  * action denoted by the specified dtrace_aggregation_t.  The aggregation
2396  * buffer is specified as the buf parameter.  This routine does not return
2397  * failure; if there is no space in the aggregation buffer, the data will be
2398  * dropped, and a corresponding counter incremented.
2399  */
2400 static void
2401 dtrace_aggregate(dtrace_aggregation_t *agg, dtrace_buffer_t *dbuf,
2402     intptr_t offset, dtrace_buffer_t *buf, uint64_t expr, uint64_t arg)
2403 {
2404 	dtrace_recdesc_t *rec = &agg->dtag_action.dta_rec;
2405 	uint32_t i, ndx, size, fsize;
2406 	uint32_t align = sizeof (uint64_t) - 1;
2407 	dtrace_aggbuffer_t *agb;
2408 	dtrace_aggkey_t *key;
2409 	uint32_t hashval = 0, limit, isstr;
2410 	caddr_t tomax, data, kdata;
2411 	dtrace_actkind_t action;
2412 	dtrace_action_t *act;
2413 	uintptr_t offs;
2414 
2415 	if (buf == NULL)
2416 		return;
2417 
2418 	if (!agg->dtag_hasarg) {
2419 		/*
2420 		 * Currently, only quantize() and lquantize() take additional
2421 		 * arguments, and they have the same semantics:  an increment
2422 		 * value that defaults to 1 when not present.  If additional
2423 		 * aggregating actions take arguments, the setting of the
2424 		 * default argument value will presumably have to become more
2425 		 * sophisticated...
2426 		 */
2427 		arg = 1;
2428 	}
2429 
2430 	action = agg->dtag_action.dta_kind - DTRACEACT_AGGREGATION;
2431 	size = rec->dtrd_offset - agg->dtag_base;
2432 	fsize = size + rec->dtrd_size;
2433 
2434 	ASSERT(dbuf->dtb_tomax != NULL);
2435 	data = dbuf->dtb_tomax + offset + agg->dtag_base;
2436 
2437 	if ((tomax = buf->dtb_tomax) == NULL) {
2438 		dtrace_buffer_drop(buf);
2439 		return;
2440 	}
2441 
2442 	/*
2443 	 * The metastructure is always at the bottom of the buffer.
2444 	 */
2445 	agb = (dtrace_aggbuffer_t *)(tomax + buf->dtb_size -
2446 	    sizeof (dtrace_aggbuffer_t));
2447 
2448 	if (buf->dtb_offset == 0) {
2449 		/*
2450 		 * We just kludge up approximately 1/8th of the size to be
2451 		 * buckets.  If this guess ends up being routinely
2452 		 * off-the-mark, we may need to dynamically readjust this
2453 		 * based on past performance.
2454 		 */
2455 		uintptr_t hashsize = (buf->dtb_size >> 3) / sizeof (uintptr_t);
2456 
2457 		if ((uintptr_t)agb - hashsize * sizeof (dtrace_aggkey_t *) <
2458 		    (uintptr_t)tomax || hashsize == 0) {
2459 			/*
2460 			 * We've been given a ludicrously small buffer;
2461 			 * increment our drop count and leave.
2462 			 */
2463 			dtrace_buffer_drop(buf);
2464 			return;
2465 		}
2466 
2467 		/*
2468 		 * And now, a pathetic attempt to try to get a an odd (or
2469 		 * perchance, a prime) hash size for better hash distribution.
2470 		 */
2471 		if (hashsize > (DTRACE_AGGHASHSIZE_SLEW << 3))
2472 			hashsize -= DTRACE_AGGHASHSIZE_SLEW;
2473 
2474 		agb->dtagb_hashsize = hashsize;
2475 		agb->dtagb_hash = (dtrace_aggkey_t **)((uintptr_t)agb -
2476 		    agb->dtagb_hashsize * sizeof (dtrace_aggkey_t *));
2477 		agb->dtagb_free = (uintptr_t)agb->dtagb_hash;
2478 
2479 		for (i = 0; i < agb->dtagb_hashsize; i++)
2480 			agb->dtagb_hash[i] = NULL;
2481 	}
2482 
2483 	ASSERT(agg->dtag_first != NULL);
2484 	ASSERT(agg->dtag_first->dta_intuple);
2485 
2486 	/*
2487 	 * Calculate the hash value based on the key.  Note that we _don't_
2488 	 * include the aggid in the hashing (but we will store it as part of
2489 	 * the key).  The hashing algorithm is Bob Jenkins' "One-at-a-time"
2490 	 * algorithm: a simple, quick algorithm that has no known funnels, and
2491 	 * gets good distribution in practice.  The efficacy of the hashing
2492 	 * algorithm (and a comparison with other algorithms) may be found by
2493 	 * running the ::dtrace_aggstat MDB dcmd.
2494 	 */
2495 	for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
2496 		i = act->dta_rec.dtrd_offset - agg->dtag_base;
2497 		limit = i + act->dta_rec.dtrd_size;
2498 		ASSERT(limit <= size);
2499 		isstr = DTRACEACT_ISSTRING(act);
2500 
2501 		for (; i < limit; i++) {
2502 			hashval += data[i];
2503 			hashval += (hashval << 10);
2504 			hashval ^= (hashval >> 6);
2505 
2506 			if (isstr && data[i] == '\0')
2507 				break;
2508 		}
2509 	}
2510 
2511 	hashval += (hashval << 3);
2512 	hashval ^= (hashval >> 11);
2513 	hashval += (hashval << 15);
2514 
2515 	/*
2516 	 * Yes, the divide here is expensive -- but it's generally the least
2517 	 * of the performance issues given the amount of data that we iterate
2518 	 * over to compute hash values, compare data, etc.
2519 	 */
2520 	ndx = hashval % agb->dtagb_hashsize;
2521 
2522 	for (key = agb->dtagb_hash[ndx]; key != NULL; key = key->dtak_next) {
2523 		ASSERT((caddr_t)key >= tomax);
2524 		ASSERT((caddr_t)key < tomax + buf->dtb_size);
2525 
2526 		if (hashval != key->dtak_hashval || key->dtak_size != size)
2527 			continue;
2528 
2529 		kdata = key->dtak_data;
2530 		ASSERT(kdata >= tomax && kdata < tomax + buf->dtb_size);
2531 
2532 		for (act = agg->dtag_first; act->dta_intuple;
2533 		    act = act->dta_next) {
2534 			i = act->dta_rec.dtrd_offset - agg->dtag_base;
2535 			limit = i + act->dta_rec.dtrd_size;
2536 			ASSERT(limit <= size);
2537 			isstr = DTRACEACT_ISSTRING(act);
2538 
2539 			for (; i < limit; i++) {
2540 				if (kdata[i] != data[i])
2541 					goto next;
2542 
2543 				if (isstr && data[i] == '\0')
2544 					break;
2545 			}
2546 		}
2547 
2548 		if (action != key->dtak_action) {
2549 			/*
2550 			 * We are aggregating on the same value in the same
2551 			 * aggregation with two different aggregating actions.
2552 			 * (This should have been picked up in the compiler,
2553 			 * so we may be dealing with errant or devious DIF.)
2554 			 * This is an error condition; we indicate as much,
2555 			 * and return.
2556 			 */
2557 			DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
2558 			return;
2559 		}
2560 
2561 		/*
2562 		 * This is a hit:  we need to apply the aggregator to
2563 		 * the value at this key.
2564 		 */
2565 		agg->dtag_aggregate((uint64_t *)(kdata + size), expr, arg);
2566 		return;
2567 next:
2568 		continue;
2569 	}
2570 
2571 	/*
2572 	 * We didn't find it.  We need to allocate some zero-filled space,
2573 	 * link it into the hash table appropriately, and apply the aggregator
2574 	 * to the (zero-filled) value.
2575 	 */
2576 	offs = buf->dtb_offset;
2577 	while (offs & (align - 1))
2578 		offs += sizeof (uint32_t);
2579 
2580 	/*
2581 	 * If we don't have enough room to both allocate a new key _and_
2582 	 * its associated data, increment the drop count and return.
2583 	 */
2584 	if ((uintptr_t)tomax + offs + fsize >
2585 	    agb->dtagb_free - sizeof (dtrace_aggkey_t)) {
2586 		dtrace_buffer_drop(buf);
2587 		return;
2588 	}
2589 
2590 	/*CONSTCOND*/
2591 	ASSERT(!(sizeof (dtrace_aggkey_t) & (sizeof (uintptr_t) - 1)));
2592 	key = (dtrace_aggkey_t *)(agb->dtagb_free - sizeof (dtrace_aggkey_t));
2593 	agb->dtagb_free -= sizeof (dtrace_aggkey_t);
2594 
2595 	key->dtak_data = kdata = tomax + offs;
2596 	buf->dtb_offset = offs + fsize;
2597 
2598 	/*
2599 	 * Now copy the data across.
2600 	 */
2601 	*((dtrace_aggid_t *)kdata) = agg->dtag_id;
2602 
2603 	for (i = sizeof (dtrace_aggid_t); i < size; i++)
2604 		kdata[i] = data[i];
2605 
2606 	/*
2607 	 * Because strings are not zeroed out by default, we need to iterate
2608 	 * looking for actions that store strings, and we need to explicitly
2609 	 * pad these strings out with zeroes.
2610 	 */
2611 	for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
2612 		int nul;
2613 
2614 		if (!DTRACEACT_ISSTRING(act))
2615 			continue;
2616 
2617 		i = act->dta_rec.dtrd_offset - agg->dtag_base;
2618 		limit = i + act->dta_rec.dtrd_size;
2619 		ASSERT(limit <= size);
2620 
2621 		for (nul = 0; i < limit; i++) {
2622 			if (nul) {
2623 				kdata[i] = '\0';
2624 				continue;
2625 			}
2626 
2627 			if (data[i] != '\0')
2628 				continue;
2629 
2630 			nul = 1;
2631 		}
2632 	}
2633 
2634 	for (i = size; i < fsize; i++)
2635 		kdata[i] = 0;
2636 
2637 	key->dtak_hashval = hashval;
2638 	key->dtak_size = size;
2639 	key->dtak_action = action;
2640 	key->dtak_next = agb->dtagb_hash[ndx];
2641 	agb->dtagb_hash[ndx] = key;
2642 
2643 	/*
2644 	 * Finally, apply the aggregator.
2645 	 */
2646 	*((uint64_t *)(key->dtak_data + size)) = agg->dtag_initial;
2647 	agg->dtag_aggregate((uint64_t *)(key->dtak_data + size), expr, arg);
2648 }
2649 
2650 /*
2651  * Given consumer state, this routine finds a speculation in the INACTIVE
2652  * state and transitions it into the ACTIVE state.  If there is no speculation
2653  * in the INACTIVE state, 0 is returned.  In this case, no error counter is
2654  * incremented -- it is up to the caller to take appropriate action.
2655  */
2656 static int
2657 dtrace_speculation(dtrace_state_t *state)
2658 {
2659 	int i = 0;
2660 	dtrace_speculation_state_t current;
2661 	uint32_t *stat = &state->dts_speculations_unavail, count;
2662 
2663 	while (i < state->dts_nspeculations) {
2664 		dtrace_speculation_t *spec = &state->dts_speculations[i];
2665 
2666 		current = spec->dtsp_state;
2667 
2668 		if (current != DTRACESPEC_INACTIVE) {
2669 			if (current == DTRACESPEC_COMMITTINGMANY ||
2670 			    current == DTRACESPEC_COMMITTING ||
2671 			    current == DTRACESPEC_DISCARDING)
2672 				stat = &state->dts_speculations_busy;
2673 			i++;
2674 			continue;
2675 		}
2676 
2677 		if (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2678 		    current, DTRACESPEC_ACTIVE) == current)
2679 			return (i + 1);
2680 	}
2681 
2682 	/*
2683 	 * We couldn't find a speculation.  If we found as much as a single
2684 	 * busy speculation buffer, we'll attribute this failure as "busy"
2685 	 * instead of "unavail".
2686 	 */
2687 	do {
2688 		count = *stat;
2689 	} while (dtrace_cas32(stat, count, count + 1) != count);
2690 
2691 	return (0);
2692 }
2693 
2694 /*
2695  * This routine commits an active speculation.  If the specified speculation
2696  * is not in a valid state to perform a commit(), this routine will silently do
2697  * nothing.  The state of the specified speculation is transitioned according
2698  * to the state transition diagram outlined in <sys/dtrace_impl.h>
2699  */
2700 static void
2701 dtrace_speculation_commit(dtrace_state_t *state, processorid_t cpu,
2702     dtrace_specid_t which)
2703 {
2704 	dtrace_speculation_t *spec;
2705 	dtrace_buffer_t *src, *dest;
2706 	uintptr_t daddr, saddr, dlimit, slimit;
2707 	dtrace_speculation_state_t current, new = 0;
2708 	intptr_t offs;
2709 	uint64_t timestamp;
2710 
2711 	if (which == 0)
2712 		return;
2713 
2714 	if (which > state->dts_nspeculations) {
2715 		cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2716 		return;
2717 	}
2718 
2719 	spec = &state->dts_speculations[which - 1];
2720 	src = &spec->dtsp_buffer[cpu];
2721 	dest = &state->dts_buffer[cpu];
2722 
2723 	do {
2724 		current = spec->dtsp_state;
2725 
2726 		if (current == DTRACESPEC_COMMITTINGMANY)
2727 			break;
2728 
2729 		switch (current) {
2730 		case DTRACESPEC_INACTIVE:
2731 		case DTRACESPEC_DISCARDING:
2732 			return;
2733 
2734 		case DTRACESPEC_COMMITTING:
2735 			/*
2736 			 * This is only possible if we are (a) commit()'ing
2737 			 * without having done a prior speculate() on this CPU
2738 			 * and (b) racing with another commit() on a different
2739 			 * CPU.  There's nothing to do -- we just assert that
2740 			 * our offset is 0.
2741 			 */
2742 			ASSERT(src->dtb_offset == 0);
2743 			return;
2744 
2745 		case DTRACESPEC_ACTIVE:
2746 			new = DTRACESPEC_COMMITTING;
2747 			break;
2748 
2749 		case DTRACESPEC_ACTIVEONE:
2750 			/*
2751 			 * This speculation is active on one CPU.  If our
2752 			 * buffer offset is non-zero, we know that the one CPU
2753 			 * must be us.  Otherwise, we are committing on a
2754 			 * different CPU from the speculate(), and we must
2755 			 * rely on being asynchronously cleaned.
2756 			 */
2757 			if (src->dtb_offset != 0) {
2758 				new = DTRACESPEC_COMMITTING;
2759 				break;
2760 			}
2761 			/*FALLTHROUGH*/
2762 
2763 		case DTRACESPEC_ACTIVEMANY:
2764 			new = DTRACESPEC_COMMITTINGMANY;
2765 			break;
2766 
2767 		default:
2768 			ASSERT(0);
2769 		}
2770 	} while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2771 	    current, new) != current);
2772 
2773 	/*
2774 	 * We have set the state to indicate that we are committing this
2775 	 * speculation.  Now reserve the necessary space in the destination
2776 	 * buffer.
2777 	 */
2778 	if ((offs = dtrace_buffer_reserve(dest, src->dtb_offset,
2779 	    sizeof (uint64_t), state, NULL)) < 0) {
2780 		dtrace_buffer_drop(dest);
2781 		goto out;
2782 	}
2783 
2784 	/*
2785 	 * We have sufficient space to copy the speculative buffer into the
2786 	 * primary buffer.  First, modify the speculative buffer, filling
2787 	 * in the timestamp of all entries with the current time.  The data
2788 	 * must have the commit() time rather than the time it was traced,
2789 	 * so that all entries in the primary buffer are in timestamp order.
2790 	 */
2791 	timestamp = dtrace_gethrtime();
2792 	saddr = (uintptr_t)src->dtb_tomax;
2793 	slimit = saddr + src->dtb_offset;
2794 	while (saddr < slimit) {
2795 		size_t size;
2796 		dtrace_rechdr_t *dtrh = (dtrace_rechdr_t *)saddr;
2797 
2798 		if (dtrh->dtrh_epid == DTRACE_EPIDNONE) {
2799 			saddr += sizeof (dtrace_epid_t);
2800 			continue;
2801 		}
2802 		ASSERT3U(dtrh->dtrh_epid, <=, state->dts_necbs);
2803 		size = state->dts_ecbs[dtrh->dtrh_epid - 1]->dte_size;
2804 
2805 		ASSERT3U(saddr + size, <=, slimit);
2806 		ASSERT3U(size, >=, sizeof (dtrace_rechdr_t));
2807 		ASSERT3U(DTRACE_RECORD_LOAD_TIMESTAMP(dtrh), ==, UINT64_MAX);
2808 
2809 		DTRACE_RECORD_STORE_TIMESTAMP(dtrh, timestamp);
2810 
2811 		saddr += size;
2812 	}
2813 
2814 	/*
2815 	 * Copy the buffer across.  (Note that this is a
2816 	 * highly subobtimal bcopy(); in the unlikely event that this becomes
2817 	 * a serious performance issue, a high-performance DTrace-specific
2818 	 * bcopy() should obviously be invented.)
2819 	 */
2820 	daddr = (uintptr_t)dest->dtb_tomax + offs;
2821 	dlimit = daddr + src->dtb_offset;
2822 	saddr = (uintptr_t)src->dtb_tomax;
2823 
2824 	/*
2825 	 * First, the aligned portion.
2826 	 */
2827 	while (dlimit - daddr >= sizeof (uint64_t)) {
2828 		*((uint64_t *)daddr) = *((uint64_t *)saddr);
2829 
2830 		daddr += sizeof (uint64_t);
2831 		saddr += sizeof (uint64_t);
2832 	}
2833 
2834 	/*
2835 	 * Now any left-over bit...
2836 	 */
2837 	while (dlimit - daddr)
2838 		*((uint8_t *)daddr++) = *((uint8_t *)saddr++);
2839 
2840 	/*
2841 	 * Finally, commit the reserved space in the destination buffer.
2842 	 */
2843 	dest->dtb_offset = offs + src->dtb_offset;
2844 
2845 out:
2846 	/*
2847 	 * If we're lucky enough to be the only active CPU on this speculation
2848 	 * buffer, we can just set the state back to DTRACESPEC_INACTIVE.
2849 	 */
2850 	if (current == DTRACESPEC_ACTIVE ||
2851 	    (current == DTRACESPEC_ACTIVEONE && new == DTRACESPEC_COMMITTING)) {
2852 		uint32_t rval = dtrace_cas32((uint32_t *)&spec->dtsp_state,
2853 		    DTRACESPEC_COMMITTING, DTRACESPEC_INACTIVE);
2854 
2855 		ASSERT(rval == DTRACESPEC_COMMITTING);
2856 	}
2857 
2858 	src->dtb_offset = 0;
2859 	src->dtb_xamot_drops += src->dtb_drops;
2860 	src->dtb_drops = 0;
2861 }
2862 
2863 /*
2864  * This routine discards an active speculation.  If the specified speculation
2865  * is not in a valid state to perform a discard(), this routine will silently
2866  * do nothing.  The state of the specified speculation is transitioned
2867  * according to the state transition diagram outlined in <sys/dtrace_impl.h>
2868  */
2869 static void
2870 dtrace_speculation_discard(dtrace_state_t *state, processorid_t cpu,
2871     dtrace_specid_t which)
2872 {
2873 	dtrace_speculation_t *spec;
2874 	dtrace_speculation_state_t current, new = 0;
2875 	dtrace_buffer_t *buf;
2876 
2877 	if (which == 0)
2878 		return;
2879 
2880 	if (which > state->dts_nspeculations) {
2881 		cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2882 		return;
2883 	}
2884 
2885 	spec = &state->dts_speculations[which - 1];
2886 	buf = &spec->dtsp_buffer[cpu];
2887 
2888 	do {
2889 		current = spec->dtsp_state;
2890 
2891 		switch (current) {
2892 		case DTRACESPEC_INACTIVE:
2893 		case DTRACESPEC_COMMITTINGMANY:
2894 		case DTRACESPEC_COMMITTING:
2895 		case DTRACESPEC_DISCARDING:
2896 			return;
2897 
2898 		case DTRACESPEC_ACTIVE:
2899 		case DTRACESPEC_ACTIVEMANY:
2900 			new = DTRACESPEC_DISCARDING;
2901 			break;
2902 
2903 		case DTRACESPEC_ACTIVEONE:
2904 			if (buf->dtb_offset != 0) {
2905 				new = DTRACESPEC_INACTIVE;
2906 			} else {
2907 				new = DTRACESPEC_DISCARDING;
2908 			}
2909 			break;
2910 
2911 		default:
2912 			ASSERT(0);
2913 		}
2914 	} while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2915 	    current, new) != current);
2916 
2917 	buf->dtb_offset = 0;
2918 	buf->dtb_drops = 0;
2919 }
2920 
2921 /*
2922  * Note:  not called from probe context.  This function is called
2923  * asynchronously from cross call context to clean any speculations that are
2924  * in the COMMITTINGMANY or DISCARDING states.  These speculations may not be
2925  * transitioned back to the INACTIVE state until all CPUs have cleaned the
2926  * speculation.
2927  */
2928 static void
2929 dtrace_speculation_clean_here(dtrace_state_t *state)
2930 {
2931 	dtrace_icookie_t cookie;
2932 	processorid_t cpu = curcpu;
2933 	dtrace_buffer_t *dest = &state->dts_buffer[cpu];
2934 	dtrace_specid_t i;
2935 
2936 	cookie = dtrace_interrupt_disable();
2937 
2938 	if (dest->dtb_tomax == NULL) {
2939 		dtrace_interrupt_enable(cookie);
2940 		return;
2941 	}
2942 
2943 	for (i = 0; i < state->dts_nspeculations; i++) {
2944 		dtrace_speculation_t *spec = &state->dts_speculations[i];
2945 		dtrace_buffer_t *src = &spec->dtsp_buffer[cpu];
2946 
2947 		if (src->dtb_tomax == NULL)
2948 			continue;
2949 
2950 		if (spec->dtsp_state == DTRACESPEC_DISCARDING) {
2951 			src->dtb_offset = 0;
2952 			continue;
2953 		}
2954 
2955 		if (spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
2956 			continue;
2957 
2958 		if (src->dtb_offset == 0)
2959 			continue;
2960 
2961 		dtrace_speculation_commit(state, cpu, i + 1);
2962 	}
2963 
2964 	dtrace_interrupt_enable(cookie);
2965 }
2966 
2967 /*
2968  * Note:  not called from probe context.  This function is called
2969  * asynchronously (and at a regular interval) to clean any speculations that
2970  * are in the COMMITTINGMANY or DISCARDING states.  If it discovers that there
2971  * is work to be done, it cross calls all CPUs to perform that work;
2972  * COMMITMANY and DISCARDING speculations may not be transitioned back to the
2973  * INACTIVE state until they have been cleaned by all CPUs.
2974  */
2975 static void
2976 dtrace_speculation_clean(dtrace_state_t *state)
2977 {
2978 	int work = 0, rv;
2979 	dtrace_specid_t i;
2980 
2981 	for (i = 0; i < state->dts_nspeculations; i++) {
2982 		dtrace_speculation_t *spec = &state->dts_speculations[i];
2983 
2984 		ASSERT(!spec->dtsp_cleaning);
2985 
2986 		if (spec->dtsp_state != DTRACESPEC_DISCARDING &&
2987 		    spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
2988 			continue;
2989 
2990 		work++;
2991 		spec->dtsp_cleaning = 1;
2992 	}
2993 
2994 	if (!work)
2995 		return;
2996 
2997 	dtrace_xcall(DTRACE_CPUALL,
2998 	    (dtrace_xcall_t)dtrace_speculation_clean_here, state);
2999 
3000 	/*
3001 	 * We now know that all CPUs have committed or discarded their
3002 	 * speculation buffers, as appropriate.  We can now set the state
3003 	 * to inactive.
3004 	 */
3005 	for (i = 0; i < state->dts_nspeculations; i++) {
3006 		dtrace_speculation_t *spec = &state->dts_speculations[i];
3007 		dtrace_speculation_state_t current, new;
3008 
3009 		if (!spec->dtsp_cleaning)
3010 			continue;
3011 
3012 		current = spec->dtsp_state;
3013 		ASSERT(current == DTRACESPEC_DISCARDING ||
3014 		    current == DTRACESPEC_COMMITTINGMANY);
3015 
3016 		new = DTRACESPEC_INACTIVE;
3017 
3018 		rv = dtrace_cas32((uint32_t *)&spec->dtsp_state, current, new);
3019 		ASSERT(rv == current);
3020 		spec->dtsp_cleaning = 0;
3021 	}
3022 }
3023 
3024 /*
3025  * Called as part of a speculate() to get the speculative buffer associated
3026  * with a given speculation.  Returns NULL if the specified speculation is not
3027  * in an ACTIVE state.  If the speculation is in the ACTIVEONE state -- and
3028  * the active CPU is not the specified CPU -- the speculation will be
3029  * atomically transitioned into the ACTIVEMANY state.
3030  */
3031 static dtrace_buffer_t *
3032 dtrace_speculation_buffer(dtrace_state_t *state, processorid_t cpuid,
3033     dtrace_specid_t which)
3034 {
3035 	dtrace_speculation_t *spec;
3036 	dtrace_speculation_state_t current, new = 0;
3037 	dtrace_buffer_t *buf;
3038 
3039 	if (which == 0)
3040 		return (NULL);
3041 
3042 	if (which > state->dts_nspeculations) {
3043 		cpu_core[cpuid].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
3044 		return (NULL);
3045 	}
3046 
3047 	spec = &state->dts_speculations[which - 1];
3048 	buf = &spec->dtsp_buffer[cpuid];
3049 
3050 	do {
3051 		current = spec->dtsp_state;
3052 
3053 		switch (current) {
3054 		case DTRACESPEC_INACTIVE:
3055 		case DTRACESPEC_COMMITTINGMANY:
3056 		case DTRACESPEC_DISCARDING:
3057 			return (NULL);
3058 
3059 		case DTRACESPEC_COMMITTING:
3060 			ASSERT(buf->dtb_offset == 0);
3061 			return (NULL);
3062 
3063 		case DTRACESPEC_ACTIVEONE:
3064 			/*
3065 			 * This speculation is currently active on one CPU.
3066 			 * Check the offset in the buffer; if it's non-zero,
3067 			 * that CPU must be us (and we leave the state alone).
3068 			 * If it's zero, assume that we're starting on a new
3069 			 * CPU -- and change the state to indicate that the
3070 			 * speculation is active on more than one CPU.
3071 			 */
3072 			if (buf->dtb_offset != 0)
3073 				return (buf);
3074 
3075 			new = DTRACESPEC_ACTIVEMANY;
3076 			break;
3077 
3078 		case DTRACESPEC_ACTIVEMANY:
3079 			return (buf);
3080 
3081 		case DTRACESPEC_ACTIVE:
3082 			new = DTRACESPEC_ACTIVEONE;
3083 			break;
3084 
3085 		default:
3086 			ASSERT(0);
3087 		}
3088 	} while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
3089 	    current, new) != current);
3090 
3091 	ASSERT(new == DTRACESPEC_ACTIVEONE || new == DTRACESPEC_ACTIVEMANY);
3092 	return (buf);
3093 }
3094 
3095 /*
3096  * Return a string.  In the event that the user lacks the privilege to access
3097  * arbitrary kernel memory, we copy the string out to scratch memory so that we
3098  * don't fail access checking.
3099  *
3100  * dtrace_dif_variable() uses this routine as a helper for various
3101  * builtin values such as 'execname' and 'probefunc.'
3102  */
3103 uintptr_t
3104 dtrace_dif_varstr(uintptr_t addr, dtrace_state_t *state,
3105     dtrace_mstate_t *mstate)
3106 {
3107 	uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3108 	uintptr_t ret;
3109 	size_t strsz;
3110 
3111 	/*
3112 	 * The easy case: this probe is allowed to read all of memory, so
3113 	 * we can just return this as a vanilla pointer.
3114 	 */
3115 	if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
3116 		return (addr);
3117 
3118 	/*
3119 	 * This is the tougher case: we copy the string in question from
3120 	 * kernel memory into scratch memory and return it that way: this
3121 	 * ensures that we won't trip up when access checking tests the
3122 	 * BYREF return value.
3123 	 */
3124 	strsz = dtrace_strlen((char *)addr, size) + 1;
3125 
3126 	if (mstate->dtms_scratch_ptr + strsz >
3127 	    mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
3128 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3129 		return (0);
3130 	}
3131 
3132 	dtrace_strcpy((const void *)addr, (void *)mstate->dtms_scratch_ptr,
3133 	    strsz);
3134 	ret = mstate->dtms_scratch_ptr;
3135 	mstate->dtms_scratch_ptr += strsz;
3136 	return (ret);
3137 }
3138 
3139 /*
3140  * Return a string from a memoy address which is known to have one or
3141  * more concatenated, individually zero terminated, sub-strings.
3142  * In the event that the user lacks the privilege to access
3143  * arbitrary kernel memory, we copy the string out to scratch memory so that we
3144  * don't fail access checking.
3145  *
3146  * dtrace_dif_variable() uses this routine as a helper for various
3147  * builtin values such as 'execargs'.
3148  */
3149 static uintptr_t
3150 dtrace_dif_varstrz(uintptr_t addr, size_t strsz, dtrace_state_t *state,
3151     dtrace_mstate_t *mstate)
3152 {
3153 	char *p;
3154 	size_t i;
3155 	uintptr_t ret;
3156 
3157 	if (mstate->dtms_scratch_ptr + strsz >
3158 	    mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
3159 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3160 		return (0);
3161 	}
3162 
3163 	dtrace_bcopy((const void *)addr, (void *)mstate->dtms_scratch_ptr,
3164 	    strsz);
3165 
3166 	/* Replace sub-string termination characters with a space. */
3167 	for (p = (char *) mstate->dtms_scratch_ptr, i = 0; i < strsz - 1;
3168 	    p++, i++)
3169 		if (*p == '\0')
3170 			*p = ' ';
3171 
3172 	ret = mstate->dtms_scratch_ptr;
3173 	mstate->dtms_scratch_ptr += strsz;
3174 	return (ret);
3175 }
3176 
3177 /*
3178  * This function implements the DIF emulator's variable lookups.  The emulator
3179  * passes a reserved variable identifier and optional built-in array index.
3180  */
3181 static uint64_t
3182 dtrace_dif_variable(dtrace_mstate_t *mstate, dtrace_state_t *state, uint64_t v,
3183     uint64_t ndx)
3184 {
3185 	/*
3186 	 * If we're accessing one of the uncached arguments, we'll turn this
3187 	 * into a reference in the args array.
3188 	 */
3189 	if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9) {
3190 		ndx = v - DIF_VAR_ARG0;
3191 		v = DIF_VAR_ARGS;
3192 	}
3193 
3194 	switch (v) {
3195 	case DIF_VAR_ARGS:
3196 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_ARGS);
3197 		if (ndx >= sizeof (mstate->dtms_arg) /
3198 		    sizeof (mstate->dtms_arg[0])) {
3199 			int aframes = mstate->dtms_probe->dtpr_aframes + 2;
3200 			dtrace_provider_t *pv;
3201 			uint64_t val;
3202 
3203 			pv = mstate->dtms_probe->dtpr_provider;
3204 			if (pv->dtpv_pops.dtps_getargval != NULL)
3205 				val = pv->dtpv_pops.dtps_getargval(pv->dtpv_arg,
3206 				    mstate->dtms_probe->dtpr_id,
3207 				    mstate->dtms_probe->dtpr_arg, ndx, aframes);
3208 			else
3209 				val = dtrace_getarg(ndx, aframes);
3210 
3211 			/*
3212 			 * This is regrettably required to keep the compiler
3213 			 * from tail-optimizing the call to dtrace_getarg().
3214 			 * The condition always evaluates to true, but the
3215 			 * compiler has no way of figuring that out a priori.
3216 			 * (None of this would be necessary if the compiler
3217 			 * could be relied upon to _always_ tail-optimize
3218 			 * the call to dtrace_getarg() -- but it can't.)
3219 			 */
3220 			if (mstate->dtms_probe != NULL)
3221 				return (val);
3222 
3223 			ASSERT(0);
3224 		}
3225 
3226 		return (mstate->dtms_arg[ndx]);
3227 
3228 #ifdef illumos
3229 	case DIF_VAR_UREGS: {
3230 		klwp_t *lwp;
3231 
3232 		if (!dtrace_priv_proc(state))
3233 			return (0);
3234 
3235 		if ((lwp = curthread->t_lwp) == NULL) {
3236 			DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
3237 			cpu_core[curcpu].cpuc_dtrace_illval = NULL;
3238 			return (0);
3239 		}
3240 
3241 		return (dtrace_getreg(lwp->lwp_regs, ndx));
3242 		return (0);
3243 	}
3244 #else
3245 	case DIF_VAR_UREGS: {
3246 		struct trapframe *tframe;
3247 
3248 		if (!dtrace_priv_proc(state))
3249 			return (0);
3250 
3251 		if ((tframe = curthread->td_frame) == NULL) {
3252 			DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
3253 			cpu_core[curcpu].cpuc_dtrace_illval = 0;
3254 			return (0);
3255 		}
3256 
3257 		return (dtrace_getreg(tframe, ndx));
3258 	}
3259 #endif
3260 
3261 	case DIF_VAR_CURTHREAD:
3262 		if (!dtrace_priv_proc(state))
3263 			return (0);
3264 		return ((uint64_t)(uintptr_t)curthread);
3265 
3266 	case DIF_VAR_TIMESTAMP:
3267 		if (!(mstate->dtms_present & DTRACE_MSTATE_TIMESTAMP)) {
3268 			mstate->dtms_timestamp = dtrace_gethrtime();
3269 			mstate->dtms_present |= DTRACE_MSTATE_TIMESTAMP;
3270 		}
3271 		return (mstate->dtms_timestamp);
3272 
3273 	case DIF_VAR_VTIMESTAMP:
3274 		ASSERT(dtrace_vtime_references != 0);
3275 		return (curthread->t_dtrace_vtime);
3276 
3277 	case DIF_VAR_WALLTIMESTAMP:
3278 		if (!(mstate->dtms_present & DTRACE_MSTATE_WALLTIMESTAMP)) {
3279 			mstate->dtms_walltimestamp = dtrace_gethrestime();
3280 			mstate->dtms_present |= DTRACE_MSTATE_WALLTIMESTAMP;
3281 		}
3282 		return (mstate->dtms_walltimestamp);
3283 
3284 #ifdef illumos
3285 	case DIF_VAR_IPL:
3286 		if (!dtrace_priv_kernel(state))
3287 			return (0);
3288 		if (!(mstate->dtms_present & DTRACE_MSTATE_IPL)) {
3289 			mstate->dtms_ipl = dtrace_getipl();
3290 			mstate->dtms_present |= DTRACE_MSTATE_IPL;
3291 		}
3292 		return (mstate->dtms_ipl);
3293 #endif
3294 
3295 	case DIF_VAR_EPID:
3296 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_EPID);
3297 		return (mstate->dtms_epid);
3298 
3299 	case DIF_VAR_ID:
3300 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3301 		return (mstate->dtms_probe->dtpr_id);
3302 
3303 	case DIF_VAR_STACKDEPTH:
3304 		if (!dtrace_priv_kernel(state))
3305 			return (0);
3306 		if (!(mstate->dtms_present & DTRACE_MSTATE_STACKDEPTH)) {
3307 			int aframes = mstate->dtms_probe->dtpr_aframes + 2;
3308 
3309 			mstate->dtms_stackdepth = dtrace_getstackdepth(aframes);
3310 			mstate->dtms_present |= DTRACE_MSTATE_STACKDEPTH;
3311 		}
3312 		return (mstate->dtms_stackdepth);
3313 
3314 	case DIF_VAR_USTACKDEPTH:
3315 		if (!dtrace_priv_proc(state))
3316 			return (0);
3317 		if (!(mstate->dtms_present & DTRACE_MSTATE_USTACKDEPTH)) {
3318 			/*
3319 			 * See comment in DIF_VAR_PID.
3320 			 */
3321 			if (DTRACE_ANCHORED(mstate->dtms_probe) &&
3322 			    CPU_ON_INTR(CPU)) {
3323 				mstate->dtms_ustackdepth = 0;
3324 			} else {
3325 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3326 				mstate->dtms_ustackdepth =
3327 				    dtrace_getustackdepth();
3328 				DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3329 			}
3330 			mstate->dtms_present |= DTRACE_MSTATE_USTACKDEPTH;
3331 		}
3332 		return (mstate->dtms_ustackdepth);
3333 
3334 	case DIF_VAR_CALLER:
3335 		if (!dtrace_priv_kernel(state))
3336 			return (0);
3337 		if (!(mstate->dtms_present & DTRACE_MSTATE_CALLER)) {
3338 			int aframes = mstate->dtms_probe->dtpr_aframes + 2;
3339 
3340 			if (!DTRACE_ANCHORED(mstate->dtms_probe)) {
3341 				/*
3342 				 * If this is an unanchored probe, we are
3343 				 * required to go through the slow path:
3344 				 * dtrace_caller() only guarantees correct
3345 				 * results for anchored probes.
3346 				 */
3347 				pc_t caller[2] = {0, 0};
3348 
3349 				dtrace_getpcstack(caller, 2, aframes,
3350 				    (uint32_t *)(uintptr_t)mstate->dtms_arg[0]);
3351 				mstate->dtms_caller = caller[1];
3352 			} else if ((mstate->dtms_caller =
3353 			    dtrace_caller(aframes)) == -1) {
3354 				/*
3355 				 * We have failed to do this the quick way;
3356 				 * we must resort to the slower approach of
3357 				 * calling dtrace_getpcstack().
3358 				 */
3359 				pc_t caller = 0;
3360 
3361 				dtrace_getpcstack(&caller, 1, aframes, NULL);
3362 				mstate->dtms_caller = caller;
3363 			}
3364 
3365 			mstate->dtms_present |= DTRACE_MSTATE_CALLER;
3366 		}
3367 		return (mstate->dtms_caller);
3368 
3369 	case DIF_VAR_UCALLER:
3370 		if (!dtrace_priv_proc(state))
3371 			return (0);
3372 
3373 		if (!(mstate->dtms_present & DTRACE_MSTATE_UCALLER)) {
3374 			uint64_t ustack[3];
3375 
3376 			/*
3377 			 * dtrace_getupcstack() fills in the first uint64_t
3378 			 * with the current PID.  The second uint64_t will
3379 			 * be the program counter at user-level.  The third
3380 			 * uint64_t will contain the caller, which is what
3381 			 * we're after.
3382 			 */
3383 			ustack[2] = 0;
3384 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3385 			dtrace_getupcstack(ustack, 3);
3386 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3387 			mstate->dtms_ucaller = ustack[2];
3388 			mstate->dtms_present |= DTRACE_MSTATE_UCALLER;
3389 		}
3390 
3391 		return (mstate->dtms_ucaller);
3392 
3393 	case DIF_VAR_PROBEPROV:
3394 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3395 		return (dtrace_dif_varstr(
3396 		    (uintptr_t)mstate->dtms_probe->dtpr_provider->dtpv_name,
3397 		    state, mstate));
3398 
3399 	case DIF_VAR_PROBEMOD:
3400 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3401 		return (dtrace_dif_varstr(
3402 		    (uintptr_t)mstate->dtms_probe->dtpr_mod,
3403 		    state, mstate));
3404 
3405 	case DIF_VAR_PROBEFUNC:
3406 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3407 		return (dtrace_dif_varstr(
3408 		    (uintptr_t)mstate->dtms_probe->dtpr_func,
3409 		    state, mstate));
3410 
3411 	case DIF_VAR_PROBENAME:
3412 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3413 		return (dtrace_dif_varstr(
3414 		    (uintptr_t)mstate->dtms_probe->dtpr_name,
3415 		    state, mstate));
3416 
3417 	case DIF_VAR_PID:
3418 		if (!dtrace_priv_proc(state))
3419 			return (0);
3420 
3421 #ifdef illumos
3422 		/*
3423 		 * Note that we are assuming that an unanchored probe is
3424 		 * always due to a high-level interrupt.  (And we're assuming
3425 		 * that there is only a single high level interrupt.)
3426 		 */
3427 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3428 			return (pid0.pid_id);
3429 
3430 		/*
3431 		 * It is always safe to dereference one's own t_procp pointer:
3432 		 * it always points to a valid, allocated proc structure.
3433 		 * Further, it is always safe to dereference the p_pidp member
3434 		 * of one's own proc structure.  (These are truisms becuase
3435 		 * threads and processes don't clean up their own state --
3436 		 * they leave that task to whomever reaps them.)
3437 		 */
3438 		return ((uint64_t)curthread->t_procp->p_pidp->pid_id);
3439 #else
3440 		return ((uint64_t)curproc->p_pid);
3441 #endif
3442 
3443 	case DIF_VAR_PPID:
3444 		if (!dtrace_priv_proc(state))
3445 			return (0);
3446 
3447 #ifdef illumos
3448 		/*
3449 		 * See comment in DIF_VAR_PID.
3450 		 */
3451 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3452 			return (pid0.pid_id);
3453 
3454 		/*
3455 		 * It is always safe to dereference one's own t_procp pointer:
3456 		 * it always points to a valid, allocated proc structure.
3457 		 * (This is true because threads don't clean up their own
3458 		 * state -- they leave that task to whomever reaps them.)
3459 		 */
3460 		return ((uint64_t)curthread->t_procp->p_ppid);
3461 #else
3462 		if (curproc->p_pid == proc0.p_pid)
3463 			return (curproc->p_pid);
3464 		else
3465 			return (curproc->p_pptr->p_pid);
3466 #endif
3467 
3468 	case DIF_VAR_TID:
3469 #ifdef illumos
3470 		/*
3471 		 * See comment in DIF_VAR_PID.
3472 		 */
3473 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3474 			return (0);
3475 #endif
3476 
3477 		return ((uint64_t)curthread->t_tid);
3478 
3479 	case DIF_VAR_EXECARGS: {
3480 		struct pargs *p_args = curthread->td_proc->p_args;
3481 
3482 		if (p_args == NULL)
3483 			return(0);
3484 
3485 		return (dtrace_dif_varstrz(
3486 		    (uintptr_t) p_args->ar_args, p_args->ar_length, state, mstate));
3487 	}
3488 
3489 	case DIF_VAR_EXECNAME:
3490 #ifdef illumos
3491 		if (!dtrace_priv_proc(state))
3492 			return (0);
3493 
3494 		/*
3495 		 * See comment in DIF_VAR_PID.
3496 		 */
3497 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3498 			return ((uint64_t)(uintptr_t)p0.p_user.u_comm);
3499 
3500 		/*
3501 		 * It is always safe to dereference one's own t_procp pointer:
3502 		 * it always points to a valid, allocated proc structure.
3503 		 * (This is true because threads don't clean up their own
3504 		 * state -- they leave that task to whomever reaps them.)
3505 		 */
3506 		return (dtrace_dif_varstr(
3507 		    (uintptr_t)curthread->t_procp->p_user.u_comm,
3508 		    state, mstate));
3509 #else
3510 		return (dtrace_dif_varstr(
3511 		    (uintptr_t) curthread->td_proc->p_comm, state, mstate));
3512 #endif
3513 
3514 	case DIF_VAR_ZONENAME:
3515 #ifdef illumos
3516 		if (!dtrace_priv_proc(state))
3517 			return (0);
3518 
3519 		/*
3520 		 * See comment in DIF_VAR_PID.
3521 		 */
3522 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3523 			return ((uint64_t)(uintptr_t)p0.p_zone->zone_name);
3524 
3525 		/*
3526 		 * It is always safe to dereference one's own t_procp pointer:
3527 		 * it always points to a valid, allocated proc structure.
3528 		 * (This is true because threads don't clean up their own
3529 		 * state -- they leave that task to whomever reaps them.)
3530 		 */
3531 		return (dtrace_dif_varstr(
3532 		    (uintptr_t)curthread->t_procp->p_zone->zone_name,
3533 		    state, mstate));
3534 #else
3535 		return (0);
3536 #endif
3537 
3538 	case DIF_VAR_UID:
3539 		if (!dtrace_priv_proc(state))
3540 			return (0);
3541 
3542 #ifdef illumos
3543 		/*
3544 		 * See comment in DIF_VAR_PID.
3545 		 */
3546 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3547 			return ((uint64_t)p0.p_cred->cr_uid);
3548 
3549 		/*
3550 		 * It is always safe to dereference one's own t_procp pointer:
3551 		 * it always points to a valid, allocated proc structure.
3552 		 * (This is true because threads don't clean up their own
3553 		 * state -- they leave that task to whomever reaps them.)
3554 		 *
3555 		 * Additionally, it is safe to dereference one's own process
3556 		 * credential, since this is never NULL after process birth.
3557 		 */
3558 		return ((uint64_t)curthread->t_procp->p_cred->cr_uid);
3559 #else
3560 		return ((uint64_t)curthread->td_ucred->cr_uid);
3561 #endif
3562 
3563 	case DIF_VAR_GID:
3564 		if (!dtrace_priv_proc(state))
3565 			return (0);
3566 
3567 #ifdef illumos
3568 		/*
3569 		 * See comment in DIF_VAR_PID.
3570 		 */
3571 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3572 			return ((uint64_t)p0.p_cred->cr_gid);
3573 
3574 		/*
3575 		 * It is always safe to dereference one's own t_procp pointer:
3576 		 * it always points to a valid, allocated proc structure.
3577 		 * (This is true because threads don't clean up their own
3578 		 * state -- they leave that task to whomever reaps them.)
3579 		 *
3580 		 * Additionally, it is safe to dereference one's own process
3581 		 * credential, since this is never NULL after process birth.
3582 		 */
3583 		return ((uint64_t)curthread->t_procp->p_cred->cr_gid);
3584 #else
3585 		return ((uint64_t)curthread->td_ucred->cr_gid);
3586 #endif
3587 
3588 	case DIF_VAR_ERRNO: {
3589 #ifdef illumos
3590 		klwp_t *lwp;
3591 		if (!dtrace_priv_proc(state))
3592 			return (0);
3593 
3594 		/*
3595 		 * See comment in DIF_VAR_PID.
3596 		 */
3597 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3598 			return (0);
3599 
3600 		/*
3601 		 * It is always safe to dereference one's own t_lwp pointer in
3602 		 * the event that this pointer is non-NULL.  (This is true
3603 		 * because threads and lwps don't clean up their own state --
3604 		 * they leave that task to whomever reaps them.)
3605 		 */
3606 		if ((lwp = curthread->t_lwp) == NULL)
3607 			return (0);
3608 
3609 		return ((uint64_t)lwp->lwp_errno);
3610 #else
3611 		return (curthread->td_errno);
3612 #endif
3613 	}
3614 #ifndef illumos
3615 	case DIF_VAR_CPU: {
3616 		return curcpu;
3617 	}
3618 #endif
3619 	default:
3620 		DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
3621 		return (0);
3622 	}
3623 }
3624 
3625 
3626 typedef enum dtrace_json_state {
3627 	DTRACE_JSON_REST = 1,
3628 	DTRACE_JSON_OBJECT,
3629 	DTRACE_JSON_STRING,
3630 	DTRACE_JSON_STRING_ESCAPE,
3631 	DTRACE_JSON_STRING_ESCAPE_UNICODE,
3632 	DTRACE_JSON_COLON,
3633 	DTRACE_JSON_COMMA,
3634 	DTRACE_JSON_VALUE,
3635 	DTRACE_JSON_IDENTIFIER,
3636 	DTRACE_JSON_NUMBER,
3637 	DTRACE_JSON_NUMBER_FRAC,
3638 	DTRACE_JSON_NUMBER_EXP,
3639 	DTRACE_JSON_COLLECT_OBJECT
3640 } dtrace_json_state_t;
3641 
3642 /*
3643  * This function possesses just enough knowledge about JSON to extract a single
3644  * value from a JSON string and store it in the scratch buffer.  It is able
3645  * to extract nested object values, and members of arrays by index.
3646  *
3647  * elemlist is a list of JSON keys, stored as packed NUL-terminated strings, to
3648  * be looked up as we descend into the object tree.  e.g.
3649  *
3650  *    foo[0].bar.baz[32] --> "foo" NUL "0" NUL "bar" NUL "baz" NUL "32" NUL
3651  *       with nelems = 5.
3652  *
3653  * The run time of this function must be bounded above by strsize to limit the
3654  * amount of work done in probe context.  As such, it is implemented as a
3655  * simple state machine, reading one character at a time using safe loads
3656  * until we find the requested element, hit a parsing error or run off the
3657  * end of the object or string.
3658  *
3659  * As there is no way for a subroutine to return an error without interrupting
3660  * clause execution, we simply return NULL in the event of a missing key or any
3661  * other error condition.  Each NULL return in this function is commented with
3662  * the error condition it represents -- parsing or otherwise.
3663  *
3664  * The set of states for the state machine closely matches the JSON
3665  * specification (http://json.org/).  Briefly:
3666  *
3667  *   DTRACE_JSON_REST:
3668  *     Skip whitespace until we find either a top-level Object, moving
3669  *     to DTRACE_JSON_OBJECT; or an Array, moving to DTRACE_JSON_VALUE.
3670  *
3671  *   DTRACE_JSON_OBJECT:
3672  *     Locate the next key String in an Object.  Sets a flag to denote
3673  *     the next String as a key string and moves to DTRACE_JSON_STRING.
3674  *
3675  *   DTRACE_JSON_COLON:
3676  *     Skip whitespace until we find the colon that separates key Strings
3677  *     from their values.  Once found, move to DTRACE_JSON_VALUE.
3678  *
3679  *   DTRACE_JSON_VALUE:
3680  *     Detects the type of the next value (String, Number, Identifier, Object
3681  *     or Array) and routes to the states that process that type.  Here we also
3682  *     deal with the element selector list if we are requested to traverse down
3683  *     into the object tree.
3684  *
3685  *   DTRACE_JSON_COMMA:
3686  *     Skip whitespace until we find the comma that separates key-value pairs
3687  *     in Objects (returning to DTRACE_JSON_OBJECT) or values in Arrays
3688  *     (similarly DTRACE_JSON_VALUE).  All following literal value processing
3689  *     states return to this state at the end of their value, unless otherwise
3690  *     noted.
3691  *
3692  *   DTRACE_JSON_NUMBER, DTRACE_JSON_NUMBER_FRAC, DTRACE_JSON_NUMBER_EXP:
3693  *     Processes a Number literal from the JSON, including any exponent
3694  *     component that may be present.  Numbers are returned as strings, which
3695  *     may be passed to strtoll() if an integer is required.
3696  *
3697  *   DTRACE_JSON_IDENTIFIER:
3698  *     Processes a "true", "false" or "null" literal in the JSON.
3699  *
3700  *   DTRACE_JSON_STRING, DTRACE_JSON_STRING_ESCAPE,
3701  *   DTRACE_JSON_STRING_ESCAPE_UNICODE:
3702  *     Processes a String literal from the JSON, whether the String denotes
3703  *     a key, a value or part of a larger Object.  Handles all escape sequences
3704  *     present in the specification, including four-digit unicode characters,
3705  *     but merely includes the escape sequence without converting it to the
3706  *     actual escaped character.  If the String is flagged as a key, we
3707  *     move to DTRACE_JSON_COLON rather than DTRACE_JSON_COMMA.
3708  *
3709  *   DTRACE_JSON_COLLECT_OBJECT:
3710  *     This state collects an entire Object (or Array), correctly handling
3711  *     embedded strings.  If the full element selector list matches this nested
3712  *     object, we return the Object in full as a string.  If not, we use this
3713  *     state to skip to the next value at this level and continue processing.
3714  *
3715  * NOTE: This function uses various macros from strtolctype.h to manipulate
3716  * digit values, etc -- these have all been checked to ensure they make
3717  * no additional function calls.
3718  */
3719 static char *
3720 dtrace_json(uint64_t size, uintptr_t json, char *elemlist, int nelems,
3721     char *dest)
3722 {
3723 	dtrace_json_state_t state = DTRACE_JSON_REST;
3724 	int64_t array_elem = INT64_MIN;
3725 	int64_t array_pos = 0;
3726 	uint8_t escape_unicount = 0;
3727 	boolean_t string_is_key = B_FALSE;
3728 	boolean_t collect_object = B_FALSE;
3729 	boolean_t found_key = B_FALSE;
3730 	boolean_t in_array = B_FALSE;
3731 	uint32_t braces = 0, brackets = 0;
3732 	char *elem = elemlist;
3733 	char *dd = dest;
3734 	uintptr_t cur;
3735 
3736 	for (cur = json; cur < json + size; cur++) {
3737 		char cc = dtrace_load8(cur);
3738 		if (cc == '\0')
3739 			return (NULL);
3740 
3741 		switch (state) {
3742 		case DTRACE_JSON_REST:
3743 			if (isspace(cc))
3744 				break;
3745 
3746 			if (cc == '{') {
3747 				state = DTRACE_JSON_OBJECT;
3748 				break;
3749 			}
3750 
3751 			if (cc == '[') {
3752 				in_array = B_TRUE;
3753 				array_pos = 0;
3754 				array_elem = dtrace_strtoll(elem, 10, size);
3755 				found_key = array_elem == 0 ? B_TRUE : B_FALSE;
3756 				state = DTRACE_JSON_VALUE;
3757 				break;
3758 			}
3759 
3760 			/*
3761 			 * ERROR: expected to find a top-level object or array.
3762 			 */
3763 			return (NULL);
3764 		case DTRACE_JSON_OBJECT:
3765 			if (isspace(cc))
3766 				break;
3767 
3768 			if (cc == '"') {
3769 				state = DTRACE_JSON_STRING;
3770 				string_is_key = B_TRUE;
3771 				break;
3772 			}
3773 
3774 			/*
3775 			 * ERROR: either the object did not start with a key
3776 			 * string, or we've run off the end of the object
3777 			 * without finding the requested key.
3778 			 */
3779 			return (NULL);
3780 		case DTRACE_JSON_STRING:
3781 			if (cc == '\\') {
3782 				*dd++ = '\\';
3783 				state = DTRACE_JSON_STRING_ESCAPE;
3784 				break;
3785 			}
3786 
3787 			if (cc == '"') {
3788 				if (collect_object) {
3789 					/*
3790 					 * We don't reset the dest here, as
3791 					 * the string is part of a larger
3792 					 * object being collected.
3793 					 */
3794 					*dd++ = cc;
3795 					collect_object = B_FALSE;
3796 					state = DTRACE_JSON_COLLECT_OBJECT;
3797 					break;
3798 				}
3799 				*dd = '\0';
3800 				dd = dest; /* reset string buffer */
3801 				if (string_is_key) {
3802 					if (dtrace_strncmp(dest, elem,
3803 					    size) == 0)
3804 						found_key = B_TRUE;
3805 				} else if (found_key) {
3806 					if (nelems > 1) {
3807 						/*
3808 						 * We expected an object, not
3809 						 * this string.
3810 						 */
3811 						return (NULL);
3812 					}
3813 					return (dest);
3814 				}
3815 				state = string_is_key ? DTRACE_JSON_COLON :
3816 				    DTRACE_JSON_COMMA;
3817 				string_is_key = B_FALSE;
3818 				break;
3819 			}
3820 
3821 			*dd++ = cc;
3822 			break;
3823 		case DTRACE_JSON_STRING_ESCAPE:
3824 			*dd++ = cc;
3825 			if (cc == 'u') {
3826 				escape_unicount = 0;
3827 				state = DTRACE_JSON_STRING_ESCAPE_UNICODE;
3828 			} else {
3829 				state = DTRACE_JSON_STRING;
3830 			}
3831 			break;
3832 		case DTRACE_JSON_STRING_ESCAPE_UNICODE:
3833 			if (!isxdigit(cc)) {
3834 				/*
3835 				 * ERROR: invalid unicode escape, expected
3836 				 * four valid hexidecimal digits.
3837 				 */
3838 				return (NULL);
3839 			}
3840 
3841 			*dd++ = cc;
3842 			if (++escape_unicount == 4)
3843 				state = DTRACE_JSON_STRING;
3844 			break;
3845 		case DTRACE_JSON_COLON:
3846 			if (isspace(cc))
3847 				break;
3848 
3849 			if (cc == ':') {
3850 				state = DTRACE_JSON_VALUE;
3851 				break;
3852 			}
3853 
3854 			/*
3855 			 * ERROR: expected a colon.
3856 			 */
3857 			return (NULL);
3858 		case DTRACE_JSON_COMMA:
3859 			if (isspace(cc))
3860 				break;
3861 
3862 			if (cc == ',') {
3863 				if (in_array) {
3864 					state = DTRACE_JSON_VALUE;
3865 					if (++array_pos == array_elem)
3866 						found_key = B_TRUE;
3867 				} else {
3868 					state = DTRACE_JSON_OBJECT;
3869 				}
3870 				break;
3871 			}
3872 
3873 			/*
3874 			 * ERROR: either we hit an unexpected character, or
3875 			 * we reached the end of the object or array without
3876 			 * finding the requested key.
3877 			 */
3878 			return (NULL);
3879 		case DTRACE_JSON_IDENTIFIER:
3880 			if (islower(cc)) {
3881 				*dd++ = cc;
3882 				break;
3883 			}
3884 
3885 			*dd = '\0';
3886 			dd = dest; /* reset string buffer */
3887 
3888 			if (dtrace_strncmp(dest, "true", 5) == 0 ||
3889 			    dtrace_strncmp(dest, "false", 6) == 0 ||
3890 			    dtrace_strncmp(dest, "null", 5) == 0) {
3891 				if (found_key) {
3892 					if (nelems > 1) {
3893 						/*
3894 						 * ERROR: We expected an object,
3895 						 * not this identifier.
3896 						 */
3897 						return (NULL);
3898 					}
3899 					return (dest);
3900 				} else {
3901 					cur--;
3902 					state = DTRACE_JSON_COMMA;
3903 					break;
3904 				}
3905 			}
3906 
3907 			/*
3908 			 * ERROR: we did not recognise the identifier as one
3909 			 * of those in the JSON specification.
3910 			 */
3911 			return (NULL);
3912 		case DTRACE_JSON_NUMBER:
3913 			if (cc == '.') {
3914 				*dd++ = cc;
3915 				state = DTRACE_JSON_NUMBER_FRAC;
3916 				break;
3917 			}
3918 
3919 			if (cc == 'x' || cc == 'X') {
3920 				/*
3921 				 * ERROR: specification explicitly excludes
3922 				 * hexidecimal or octal numbers.
3923 				 */
3924 				return (NULL);
3925 			}
3926 
3927 			/* FALLTHRU */
3928 		case DTRACE_JSON_NUMBER_FRAC:
3929 			if (cc == 'e' || cc == 'E') {
3930 				*dd++ = cc;
3931 				state = DTRACE_JSON_NUMBER_EXP;
3932 				break;
3933 			}
3934 
3935 			if (cc == '+' || cc == '-') {
3936 				/*
3937 				 * ERROR: expect sign as part of exponent only.
3938 				 */
3939 				return (NULL);
3940 			}
3941 			/* FALLTHRU */
3942 		case DTRACE_JSON_NUMBER_EXP:
3943 			if (isdigit(cc) || cc == '+' || cc == '-') {
3944 				*dd++ = cc;
3945 				break;
3946 			}
3947 
3948 			*dd = '\0';
3949 			dd = dest; /* reset string buffer */
3950 			if (found_key) {
3951 				if (nelems > 1) {
3952 					/*
3953 					 * ERROR: We expected an object, not
3954 					 * this number.
3955 					 */
3956 					return (NULL);
3957 				}
3958 				return (dest);
3959 			}
3960 
3961 			cur--;
3962 			state = DTRACE_JSON_COMMA;
3963 			break;
3964 		case DTRACE_JSON_VALUE:
3965 			if (isspace(cc))
3966 				break;
3967 
3968 			if (cc == '{' || cc == '[') {
3969 				if (nelems > 1 && found_key) {
3970 					in_array = cc == '[' ? B_TRUE : B_FALSE;
3971 					/*
3972 					 * If our element selector directs us
3973 					 * to descend into this nested object,
3974 					 * then move to the next selector
3975 					 * element in the list and restart the
3976 					 * state machine.
3977 					 */
3978 					while (*elem != '\0')
3979 						elem++;
3980 					elem++; /* skip the inter-element NUL */
3981 					nelems--;
3982 					dd = dest;
3983 					if (in_array) {
3984 						state = DTRACE_JSON_VALUE;
3985 						array_pos = 0;
3986 						array_elem = dtrace_strtoll(
3987 						    elem, 10, size);
3988 						found_key = array_elem == 0 ?
3989 						    B_TRUE : B_FALSE;
3990 					} else {
3991 						found_key = B_FALSE;
3992 						state = DTRACE_JSON_OBJECT;
3993 					}
3994 					break;
3995 				}
3996 
3997 				/*
3998 				 * Otherwise, we wish to either skip this
3999 				 * nested object or return it in full.
4000 				 */
4001 				if (cc == '[')
4002 					brackets = 1;
4003 				else
4004 					braces = 1;
4005 				*dd++ = cc;
4006 				state = DTRACE_JSON_COLLECT_OBJECT;
4007 				break;
4008 			}
4009 
4010 			if (cc == '"') {
4011 				state = DTRACE_JSON_STRING;
4012 				break;
4013 			}
4014 
4015 			if (islower(cc)) {
4016 				/*
4017 				 * Here we deal with true, false and null.
4018 				 */
4019 				*dd++ = cc;
4020 				state = DTRACE_JSON_IDENTIFIER;
4021 				break;
4022 			}
4023 
4024 			if (cc == '-' || isdigit(cc)) {
4025 				*dd++ = cc;
4026 				state = DTRACE_JSON_NUMBER;
4027 				break;
4028 			}
4029 
4030 			/*
4031 			 * ERROR: unexpected character at start of value.
4032 			 */
4033 			return (NULL);
4034 		case DTRACE_JSON_COLLECT_OBJECT:
4035 			if (cc == '\0')
4036 				/*
4037 				 * ERROR: unexpected end of input.
4038 				 */
4039 				return (NULL);
4040 
4041 			*dd++ = cc;
4042 			if (cc == '"') {
4043 				collect_object = B_TRUE;
4044 				state = DTRACE_JSON_STRING;
4045 				break;
4046 			}
4047 
4048 			if (cc == ']') {
4049 				if (brackets-- == 0) {
4050 					/*
4051 					 * ERROR: unbalanced brackets.
4052 					 */
4053 					return (NULL);
4054 				}
4055 			} else if (cc == '}') {
4056 				if (braces-- == 0) {
4057 					/*
4058 					 * ERROR: unbalanced braces.
4059 					 */
4060 					return (NULL);
4061 				}
4062 			} else if (cc == '{') {
4063 				braces++;
4064 			} else if (cc == '[') {
4065 				brackets++;
4066 			}
4067 
4068 			if (brackets == 0 && braces == 0) {
4069 				if (found_key) {
4070 					*dd = '\0';
4071 					return (dest);
4072 				}
4073 				dd = dest; /* reset string buffer */
4074 				state = DTRACE_JSON_COMMA;
4075 			}
4076 			break;
4077 		}
4078 	}
4079 	return (NULL);
4080 }
4081 
4082 /*
4083  * Emulate the execution of DTrace ID subroutines invoked by the call opcode.
4084  * Notice that we don't bother validating the proper number of arguments or
4085  * their types in the tuple stack.  This isn't needed because all argument
4086  * interpretation is safe because of our load safety -- the worst that can
4087  * happen is that a bogus program can obtain bogus results.
4088  */
4089 static void
4090 dtrace_dif_subr(uint_t subr, uint_t rd, uint64_t *regs,
4091     dtrace_key_t *tupregs, int nargs,
4092     dtrace_mstate_t *mstate, dtrace_state_t *state)
4093 {
4094 	volatile uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
4095 	volatile uintptr_t *illval = &cpu_core[curcpu].cpuc_dtrace_illval;
4096 	dtrace_vstate_t *vstate = &state->dts_vstate;
4097 
4098 #ifdef illumos
4099 	union {
4100 		mutex_impl_t mi;
4101 		uint64_t mx;
4102 	} m;
4103 
4104 	union {
4105 		krwlock_t ri;
4106 		uintptr_t rw;
4107 	} r;
4108 #else
4109 	struct thread *lowner;
4110 	union {
4111 		struct lock_object *li;
4112 		uintptr_t lx;
4113 	} l;
4114 #endif
4115 
4116 	switch (subr) {
4117 	case DIF_SUBR_RAND:
4118 		regs[rd] = (dtrace_gethrtime() * 2416 + 374441) % 1771875;
4119 		break;
4120 
4121 #ifdef illumos
4122 	case DIF_SUBR_MUTEX_OWNED:
4123 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
4124 		    mstate, vstate)) {
4125 			regs[rd] = 0;
4126 			break;
4127 		}
4128 
4129 		m.mx = dtrace_load64(tupregs[0].dttk_value);
4130 		if (MUTEX_TYPE_ADAPTIVE(&m.mi))
4131 			regs[rd] = MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER;
4132 		else
4133 			regs[rd] = LOCK_HELD(&m.mi.m_spin.m_spinlock);
4134 		break;
4135 
4136 	case DIF_SUBR_MUTEX_OWNER:
4137 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
4138 		    mstate, vstate)) {
4139 			regs[rd] = 0;
4140 			break;
4141 		}
4142 
4143 		m.mx = dtrace_load64(tupregs[0].dttk_value);
4144 		if (MUTEX_TYPE_ADAPTIVE(&m.mi) &&
4145 		    MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER)
4146 			regs[rd] = (uintptr_t)MUTEX_OWNER(&m.mi);
4147 		else
4148 			regs[rd] = 0;
4149 		break;
4150 
4151 	case DIF_SUBR_MUTEX_TYPE_ADAPTIVE:
4152 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
4153 		    mstate, vstate)) {
4154 			regs[rd] = 0;
4155 			break;
4156 		}
4157 
4158 		m.mx = dtrace_load64(tupregs[0].dttk_value);
4159 		regs[rd] = MUTEX_TYPE_ADAPTIVE(&m.mi);
4160 		break;
4161 
4162 	case DIF_SUBR_MUTEX_TYPE_SPIN:
4163 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
4164 		    mstate, vstate)) {
4165 			regs[rd] = 0;
4166 			break;
4167 		}
4168 
4169 		m.mx = dtrace_load64(tupregs[0].dttk_value);
4170 		regs[rd] = MUTEX_TYPE_SPIN(&m.mi);
4171 		break;
4172 
4173 	case DIF_SUBR_RW_READ_HELD: {
4174 		uintptr_t tmp;
4175 
4176 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
4177 		    mstate, vstate)) {
4178 			regs[rd] = 0;
4179 			break;
4180 		}
4181 
4182 		r.rw = dtrace_loadptr(tupregs[0].dttk_value);
4183 		regs[rd] = _RW_READ_HELD(&r.ri, tmp);
4184 		break;
4185 	}
4186 
4187 	case DIF_SUBR_RW_WRITE_HELD:
4188 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
4189 		    mstate, vstate)) {
4190 			regs[rd] = 0;
4191 			break;
4192 		}
4193 
4194 		r.rw = dtrace_loadptr(tupregs[0].dttk_value);
4195 		regs[rd] = _RW_WRITE_HELD(&r.ri);
4196 		break;
4197 
4198 	case DIF_SUBR_RW_ISWRITER:
4199 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
4200 		    mstate, vstate)) {
4201 			regs[rd] = 0;
4202 			break;
4203 		}
4204 
4205 		r.rw = dtrace_loadptr(tupregs[0].dttk_value);
4206 		regs[rd] = _RW_ISWRITER(&r.ri);
4207 		break;
4208 
4209 #else /* !illumos */
4210 	case DIF_SUBR_MUTEX_OWNED:
4211 		if (!dtrace_canload(tupregs[0].dttk_value,
4212 			sizeof (struct lock_object), mstate, vstate)) {
4213 			regs[rd] = 0;
4214 			break;
4215 		}
4216 		l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
4217 		regs[rd] = LOCK_CLASS(l.li)->lc_owner(l.li, &lowner);
4218 		break;
4219 
4220 	case DIF_SUBR_MUTEX_OWNER:
4221 		if (!dtrace_canload(tupregs[0].dttk_value,
4222 			sizeof (struct lock_object), mstate, vstate)) {
4223 			regs[rd] = 0;
4224 			break;
4225 		}
4226 		l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
4227 		LOCK_CLASS(l.li)->lc_owner(l.li, &lowner);
4228 		regs[rd] = (uintptr_t)lowner;
4229 		break;
4230 
4231 	case DIF_SUBR_MUTEX_TYPE_ADAPTIVE:
4232 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (struct mtx),
4233 		    mstate, vstate)) {
4234 			regs[rd] = 0;
4235 			break;
4236 		}
4237 		l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
4238 		/* XXX - should be only LC_SLEEPABLE? */
4239 		regs[rd] = (LOCK_CLASS(l.li)->lc_flags &
4240 		    (LC_SLEEPLOCK | LC_SLEEPABLE)) != 0;
4241 		break;
4242 
4243 	case DIF_SUBR_MUTEX_TYPE_SPIN:
4244 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (struct mtx),
4245 		    mstate, vstate)) {
4246 			regs[rd] = 0;
4247 			break;
4248 		}
4249 		l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
4250 		regs[rd] = (LOCK_CLASS(l.li)->lc_flags & LC_SPINLOCK) != 0;
4251 		break;
4252 
4253 	case DIF_SUBR_RW_READ_HELD:
4254 	case DIF_SUBR_SX_SHARED_HELD:
4255 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
4256 		    mstate, vstate)) {
4257 			regs[rd] = 0;
4258 			break;
4259 		}
4260 		l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
4261 		regs[rd] = LOCK_CLASS(l.li)->lc_owner(l.li, &lowner) &&
4262 		    lowner == NULL;
4263 		break;
4264 
4265 	case DIF_SUBR_RW_WRITE_HELD:
4266 	case DIF_SUBR_SX_EXCLUSIVE_HELD:
4267 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
4268 		    mstate, vstate)) {
4269 			regs[rd] = 0;
4270 			break;
4271 		}
4272 		l.lx = dtrace_loadptr(tupregs[0].dttk_value);
4273 		LOCK_CLASS(l.li)->lc_owner(l.li, &lowner);
4274 		regs[rd] = (lowner == curthread);
4275 		break;
4276 
4277 	case DIF_SUBR_RW_ISWRITER:
4278 	case DIF_SUBR_SX_ISEXCLUSIVE:
4279 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
4280 		    mstate, vstate)) {
4281 			regs[rd] = 0;
4282 			break;
4283 		}
4284 		l.lx = dtrace_loadptr(tupregs[0].dttk_value);
4285 		regs[rd] = LOCK_CLASS(l.li)->lc_owner(l.li, &lowner) &&
4286 		    lowner != NULL;
4287 		break;
4288 #endif /* illumos */
4289 
4290 	case DIF_SUBR_BCOPY: {
4291 		/*
4292 		 * We need to be sure that the destination is in the scratch
4293 		 * region -- no other region is allowed.
4294 		 */
4295 		uintptr_t src = tupregs[0].dttk_value;
4296 		uintptr_t dest = tupregs[1].dttk_value;
4297 		size_t size = tupregs[2].dttk_value;
4298 
4299 		if (!dtrace_inscratch(dest, size, mstate)) {
4300 			*flags |= CPU_DTRACE_BADADDR;
4301 			*illval = regs[rd];
4302 			break;
4303 		}
4304 
4305 		if (!dtrace_canload(src, size, mstate, vstate)) {
4306 			regs[rd] = 0;
4307 			break;
4308 		}
4309 
4310 		dtrace_bcopy((void *)src, (void *)dest, size);
4311 		break;
4312 	}
4313 
4314 	case DIF_SUBR_ALLOCA:
4315 	case DIF_SUBR_COPYIN: {
4316 		uintptr_t dest = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
4317 		uint64_t size =
4318 		    tupregs[subr == DIF_SUBR_ALLOCA ? 0 : 1].dttk_value;
4319 		size_t scratch_size = (dest - mstate->dtms_scratch_ptr) + size;
4320 
4321 		/*
4322 		 * This action doesn't require any credential checks since
4323 		 * probes will not activate in user contexts to which the
4324 		 * enabling user does not have permissions.
4325 		 */
4326 
4327 		/*
4328 		 * Rounding up the user allocation size could have overflowed
4329 		 * a large, bogus allocation (like -1ULL) to 0.
4330 		 */
4331 		if (scratch_size < size ||
4332 		    !DTRACE_INSCRATCH(mstate, scratch_size)) {
4333 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4334 			regs[rd] = 0;
4335 			break;
4336 		}
4337 
4338 		if (subr == DIF_SUBR_COPYIN) {
4339 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4340 			dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
4341 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4342 		}
4343 
4344 		mstate->dtms_scratch_ptr += scratch_size;
4345 		regs[rd] = dest;
4346 		break;
4347 	}
4348 
4349 	case DIF_SUBR_COPYINTO: {
4350 		uint64_t size = tupregs[1].dttk_value;
4351 		uintptr_t dest = tupregs[2].dttk_value;
4352 
4353 		/*
4354 		 * This action doesn't require any credential checks since
4355 		 * probes will not activate in user contexts to which the
4356 		 * enabling user does not have permissions.
4357 		 */
4358 		if (!dtrace_inscratch(dest, size, mstate)) {
4359 			*flags |= CPU_DTRACE_BADADDR;
4360 			*illval = regs[rd];
4361 			break;
4362 		}
4363 
4364 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4365 		dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
4366 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4367 		break;
4368 	}
4369 
4370 	case DIF_SUBR_COPYINSTR: {
4371 		uintptr_t dest = mstate->dtms_scratch_ptr;
4372 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4373 
4374 		if (nargs > 1 && tupregs[1].dttk_value < size)
4375 			size = tupregs[1].dttk_value + 1;
4376 
4377 		/*
4378 		 * This action doesn't require any credential checks since
4379 		 * probes will not activate in user contexts to which the
4380 		 * enabling user does not have permissions.
4381 		 */
4382 		if (!DTRACE_INSCRATCH(mstate, size)) {
4383 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4384 			regs[rd] = 0;
4385 			break;
4386 		}
4387 
4388 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4389 		dtrace_copyinstr(tupregs[0].dttk_value, dest, size, flags);
4390 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4391 
4392 		((char *)dest)[size - 1] = '\0';
4393 		mstate->dtms_scratch_ptr += size;
4394 		regs[rd] = dest;
4395 		break;
4396 	}
4397 
4398 #ifdef illumos
4399 	case DIF_SUBR_MSGSIZE:
4400 	case DIF_SUBR_MSGDSIZE: {
4401 		uintptr_t baddr = tupregs[0].dttk_value, daddr;
4402 		uintptr_t wptr, rptr;
4403 		size_t count = 0;
4404 		int cont = 0;
4405 
4406 		while (baddr != 0 && !(*flags & CPU_DTRACE_FAULT)) {
4407 
4408 			if (!dtrace_canload(baddr, sizeof (mblk_t), mstate,
4409 			    vstate)) {
4410 				regs[rd] = 0;
4411 				break;
4412 			}
4413 
4414 			wptr = dtrace_loadptr(baddr +
4415 			    offsetof(mblk_t, b_wptr));
4416 
4417 			rptr = dtrace_loadptr(baddr +
4418 			    offsetof(mblk_t, b_rptr));
4419 
4420 			if (wptr < rptr) {
4421 				*flags |= CPU_DTRACE_BADADDR;
4422 				*illval = tupregs[0].dttk_value;
4423 				break;
4424 			}
4425 
4426 			daddr = dtrace_loadptr(baddr +
4427 			    offsetof(mblk_t, b_datap));
4428 
4429 			baddr = dtrace_loadptr(baddr +
4430 			    offsetof(mblk_t, b_cont));
4431 
4432 			/*
4433 			 * We want to prevent against denial-of-service here,
4434 			 * so we're only going to search the list for
4435 			 * dtrace_msgdsize_max mblks.
4436 			 */
4437 			if (cont++ > dtrace_msgdsize_max) {
4438 				*flags |= CPU_DTRACE_ILLOP;
4439 				break;
4440 			}
4441 
4442 			if (subr == DIF_SUBR_MSGDSIZE) {
4443 				if (dtrace_load8(daddr +
4444 				    offsetof(dblk_t, db_type)) != M_DATA)
4445 					continue;
4446 			}
4447 
4448 			count += wptr - rptr;
4449 		}
4450 
4451 		if (!(*flags & CPU_DTRACE_FAULT))
4452 			regs[rd] = count;
4453 
4454 		break;
4455 	}
4456 #endif
4457 
4458 	case DIF_SUBR_PROGENYOF: {
4459 		pid_t pid = tupregs[0].dttk_value;
4460 		proc_t *p;
4461 		int rval = 0;
4462 
4463 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4464 
4465 		for (p = curthread->t_procp; p != NULL; p = p->p_parent) {
4466 #ifdef illumos
4467 			if (p->p_pidp->pid_id == pid) {
4468 #else
4469 			if (p->p_pid == pid) {
4470 #endif
4471 				rval = 1;
4472 				break;
4473 			}
4474 		}
4475 
4476 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4477 
4478 		regs[rd] = rval;
4479 		break;
4480 	}
4481 
4482 	case DIF_SUBR_SPECULATION:
4483 		regs[rd] = dtrace_speculation(state);
4484 		break;
4485 
4486 	case DIF_SUBR_COPYOUT: {
4487 		uintptr_t kaddr = tupregs[0].dttk_value;
4488 		uintptr_t uaddr = tupregs[1].dttk_value;
4489 		uint64_t size = tupregs[2].dttk_value;
4490 
4491 		if (!dtrace_destructive_disallow &&
4492 		    dtrace_priv_proc_control(state) &&
4493 		    !dtrace_istoxic(kaddr, size) &&
4494 		    dtrace_canload(kaddr, size, mstate, vstate)) {
4495 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4496 			dtrace_copyout(kaddr, uaddr, size, flags);
4497 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4498 		}
4499 		break;
4500 	}
4501 
4502 	case DIF_SUBR_COPYOUTSTR: {
4503 		uintptr_t kaddr = tupregs[0].dttk_value;
4504 		uintptr_t uaddr = tupregs[1].dttk_value;
4505 		uint64_t size = tupregs[2].dttk_value;
4506 
4507 		if (!dtrace_destructive_disallow &&
4508 		    dtrace_priv_proc_control(state) &&
4509 		    !dtrace_istoxic(kaddr, size) &&
4510 		    dtrace_strcanload(kaddr, size, mstate, vstate)) {
4511 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4512 			dtrace_copyoutstr(kaddr, uaddr, size, flags);
4513 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4514 		}
4515 		break;
4516 	}
4517 
4518 	case DIF_SUBR_STRLEN: {
4519 		size_t sz;
4520 		uintptr_t addr = (uintptr_t)tupregs[0].dttk_value;
4521 		sz = dtrace_strlen((char *)addr,
4522 		    state->dts_options[DTRACEOPT_STRSIZE]);
4523 
4524 		if (!dtrace_canload(addr, sz + 1, mstate, vstate)) {
4525 			regs[rd] = 0;
4526 			break;
4527 		}
4528 
4529 		regs[rd] = sz;
4530 
4531 		break;
4532 	}
4533 
4534 	case DIF_SUBR_STRCHR:
4535 	case DIF_SUBR_STRRCHR: {
4536 		/*
4537 		 * We're going to iterate over the string looking for the
4538 		 * specified character.  We will iterate until we have reached
4539 		 * the string length or we have found the character.  If this
4540 		 * is DIF_SUBR_STRRCHR, we will look for the last occurrence
4541 		 * of the specified character instead of the first.
4542 		 */
4543 		uintptr_t saddr = tupregs[0].dttk_value;
4544 		uintptr_t addr = tupregs[0].dttk_value;
4545 		uintptr_t limit = addr + state->dts_options[DTRACEOPT_STRSIZE];
4546 		char c, target = (char)tupregs[1].dttk_value;
4547 
4548 		for (regs[rd] = 0; addr < limit; addr++) {
4549 			if ((c = dtrace_load8(addr)) == target) {
4550 				regs[rd] = addr;
4551 
4552 				if (subr == DIF_SUBR_STRCHR)
4553 					break;
4554 			}
4555 
4556 			if (c == '\0')
4557 				break;
4558 		}
4559 
4560 		if (!dtrace_canload(saddr, addr - saddr, mstate, vstate)) {
4561 			regs[rd] = 0;
4562 			break;
4563 		}
4564 
4565 		break;
4566 	}
4567 
4568 	case DIF_SUBR_STRSTR:
4569 	case DIF_SUBR_INDEX:
4570 	case DIF_SUBR_RINDEX: {
4571 		/*
4572 		 * We're going to iterate over the string looking for the
4573 		 * specified string.  We will iterate until we have reached
4574 		 * the string length or we have found the string.  (Yes, this
4575 		 * is done in the most naive way possible -- but considering
4576 		 * that the string we're searching for is likely to be
4577 		 * relatively short, the complexity of Rabin-Karp or similar
4578 		 * hardly seems merited.)
4579 		 */
4580 		char *addr = (char *)(uintptr_t)tupregs[0].dttk_value;
4581 		char *substr = (char *)(uintptr_t)tupregs[1].dttk_value;
4582 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4583 		size_t len = dtrace_strlen(addr, size);
4584 		size_t sublen = dtrace_strlen(substr, size);
4585 		char *limit = addr + len, *orig = addr;
4586 		int notfound = subr == DIF_SUBR_STRSTR ? 0 : -1;
4587 		int inc = 1;
4588 
4589 		regs[rd] = notfound;
4590 
4591 		if (!dtrace_canload((uintptr_t)addr, len + 1, mstate, vstate)) {
4592 			regs[rd] = 0;
4593 			break;
4594 		}
4595 
4596 		if (!dtrace_canload((uintptr_t)substr, sublen + 1, mstate,
4597 		    vstate)) {
4598 			regs[rd] = 0;
4599 			break;
4600 		}
4601 
4602 		/*
4603 		 * strstr() and index()/rindex() have similar semantics if
4604 		 * both strings are the empty string: strstr() returns a
4605 		 * pointer to the (empty) string, and index() and rindex()
4606 		 * both return index 0 (regardless of any position argument).
4607 		 */
4608 		if (sublen == 0 && len == 0) {
4609 			if (subr == DIF_SUBR_STRSTR)
4610 				regs[rd] = (uintptr_t)addr;
4611 			else
4612 				regs[rd] = 0;
4613 			break;
4614 		}
4615 
4616 		if (subr != DIF_SUBR_STRSTR) {
4617 			if (subr == DIF_SUBR_RINDEX) {
4618 				limit = orig - 1;
4619 				addr += len;
4620 				inc = -1;
4621 			}
4622 
4623 			/*
4624 			 * Both index() and rindex() take an optional position
4625 			 * argument that denotes the starting position.
4626 			 */
4627 			if (nargs == 3) {
4628 				int64_t pos = (int64_t)tupregs[2].dttk_value;
4629 
4630 				/*
4631 				 * If the position argument to index() is
4632 				 * negative, Perl implicitly clamps it at
4633 				 * zero.  This semantic is a little surprising
4634 				 * given the special meaning of negative
4635 				 * positions to similar Perl functions like
4636 				 * substr(), but it appears to reflect a
4637 				 * notion that index() can start from a
4638 				 * negative index and increment its way up to
4639 				 * the string.  Given this notion, Perl's
4640 				 * rindex() is at least self-consistent in
4641 				 * that it implicitly clamps positions greater
4642 				 * than the string length to be the string
4643 				 * length.  Where Perl completely loses
4644 				 * coherence, however, is when the specified
4645 				 * substring is the empty string ("").  In
4646 				 * this case, even if the position is
4647 				 * negative, rindex() returns 0 -- and even if
4648 				 * the position is greater than the length,
4649 				 * index() returns the string length.  These
4650 				 * semantics violate the notion that index()
4651 				 * should never return a value less than the
4652 				 * specified position and that rindex() should
4653 				 * never return a value greater than the
4654 				 * specified position.  (One assumes that
4655 				 * these semantics are artifacts of Perl's
4656 				 * implementation and not the results of
4657 				 * deliberate design -- it beggars belief that
4658 				 * even Larry Wall could desire such oddness.)
4659 				 * While in the abstract one would wish for
4660 				 * consistent position semantics across
4661 				 * substr(), index() and rindex() -- or at the
4662 				 * very least self-consistent position
4663 				 * semantics for index() and rindex() -- we
4664 				 * instead opt to keep with the extant Perl
4665 				 * semantics, in all their broken glory.  (Do
4666 				 * we have more desire to maintain Perl's
4667 				 * semantics than Perl does?  Probably.)
4668 				 */
4669 				if (subr == DIF_SUBR_RINDEX) {
4670 					if (pos < 0) {
4671 						if (sublen == 0)
4672 							regs[rd] = 0;
4673 						break;
4674 					}
4675 
4676 					if (pos > len)
4677 						pos = len;
4678 				} else {
4679 					if (pos < 0)
4680 						pos = 0;
4681 
4682 					if (pos >= len) {
4683 						if (sublen == 0)
4684 							regs[rd] = len;
4685 						break;
4686 					}
4687 				}
4688 
4689 				addr = orig + pos;
4690 			}
4691 		}
4692 
4693 		for (regs[rd] = notfound; addr != limit; addr += inc) {
4694 			if (dtrace_strncmp(addr, substr, sublen) == 0) {
4695 				if (subr != DIF_SUBR_STRSTR) {
4696 					/*
4697 					 * As D index() and rindex() are
4698 					 * modeled on Perl (and not on awk),
4699 					 * we return a zero-based (and not a
4700 					 * one-based) index.  (For you Perl
4701 					 * weenies: no, we're not going to add
4702 					 * $[ -- and shouldn't you be at a con
4703 					 * or something?)
4704 					 */
4705 					regs[rd] = (uintptr_t)(addr - orig);
4706 					break;
4707 				}
4708 
4709 				ASSERT(subr == DIF_SUBR_STRSTR);
4710 				regs[rd] = (uintptr_t)addr;
4711 				break;
4712 			}
4713 		}
4714 
4715 		break;
4716 	}
4717 
4718 	case DIF_SUBR_STRTOK: {
4719 		uintptr_t addr = tupregs[0].dttk_value;
4720 		uintptr_t tokaddr = tupregs[1].dttk_value;
4721 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4722 		uintptr_t limit, toklimit = tokaddr + size;
4723 		uint8_t c = 0, tokmap[32];	 /* 256 / 8 */
4724 		char *dest = (char *)mstate->dtms_scratch_ptr;
4725 		int i;
4726 
4727 		/*
4728 		 * Check both the token buffer and (later) the input buffer,
4729 		 * since both could be non-scratch addresses.
4730 		 */
4731 		if (!dtrace_strcanload(tokaddr, size, mstate, vstate)) {
4732 			regs[rd] = 0;
4733 			break;
4734 		}
4735 
4736 		if (!DTRACE_INSCRATCH(mstate, size)) {
4737 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4738 			regs[rd] = 0;
4739 			break;
4740 		}
4741 
4742 		if (addr == 0) {
4743 			/*
4744 			 * If the address specified is NULL, we use our saved
4745 			 * strtok pointer from the mstate.  Note that this
4746 			 * means that the saved strtok pointer is _only_
4747 			 * valid within multiple enablings of the same probe --
4748 			 * it behaves like an implicit clause-local variable.
4749 			 */
4750 			addr = mstate->dtms_strtok;
4751 		} else {
4752 			/*
4753 			 * If the user-specified address is non-NULL we must
4754 			 * access check it.  This is the only time we have
4755 			 * a chance to do so, since this address may reside
4756 			 * in the string table of this clause-- future calls
4757 			 * (when we fetch addr from mstate->dtms_strtok)
4758 			 * would fail this access check.
4759 			 */
4760 			if (!dtrace_strcanload(addr, size, mstate, vstate)) {
4761 				regs[rd] = 0;
4762 				break;
4763 			}
4764 		}
4765 
4766 		/*
4767 		 * First, zero the token map, and then process the token
4768 		 * string -- setting a bit in the map for every character
4769 		 * found in the token string.
4770 		 */
4771 		for (i = 0; i < sizeof (tokmap); i++)
4772 			tokmap[i] = 0;
4773 
4774 		for (; tokaddr < toklimit; tokaddr++) {
4775 			if ((c = dtrace_load8(tokaddr)) == '\0')
4776 				break;
4777 
4778 			ASSERT((c >> 3) < sizeof (tokmap));
4779 			tokmap[c >> 3] |= (1 << (c & 0x7));
4780 		}
4781 
4782 		for (limit = addr + size; addr < limit; addr++) {
4783 			/*
4784 			 * We're looking for a character that is _not_ contained
4785 			 * in the token string.
4786 			 */
4787 			if ((c = dtrace_load8(addr)) == '\0')
4788 				break;
4789 
4790 			if (!(tokmap[c >> 3] & (1 << (c & 0x7))))
4791 				break;
4792 		}
4793 
4794 		if (c == '\0') {
4795 			/*
4796 			 * We reached the end of the string without finding
4797 			 * any character that was not in the token string.
4798 			 * We return NULL in this case, and we set the saved
4799 			 * address to NULL as well.
4800 			 */
4801 			regs[rd] = 0;
4802 			mstate->dtms_strtok = 0;
4803 			break;
4804 		}
4805 
4806 		/*
4807 		 * From here on, we're copying into the destination string.
4808 		 */
4809 		for (i = 0; addr < limit && i < size - 1; addr++) {
4810 			if ((c = dtrace_load8(addr)) == '\0')
4811 				break;
4812 
4813 			if (tokmap[c >> 3] & (1 << (c & 0x7)))
4814 				break;
4815 
4816 			ASSERT(i < size);
4817 			dest[i++] = c;
4818 		}
4819 
4820 		ASSERT(i < size);
4821 		dest[i] = '\0';
4822 		regs[rd] = (uintptr_t)dest;
4823 		mstate->dtms_scratch_ptr += size;
4824 		mstate->dtms_strtok = addr;
4825 		break;
4826 	}
4827 
4828 	case DIF_SUBR_SUBSTR: {
4829 		uintptr_t s = tupregs[0].dttk_value;
4830 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4831 		char *d = (char *)mstate->dtms_scratch_ptr;
4832 		int64_t index = (int64_t)tupregs[1].dttk_value;
4833 		int64_t remaining = (int64_t)tupregs[2].dttk_value;
4834 		size_t len = dtrace_strlen((char *)s, size);
4835 		int64_t i;
4836 
4837 		if (!dtrace_canload(s, len + 1, mstate, vstate)) {
4838 			regs[rd] = 0;
4839 			break;
4840 		}
4841 
4842 		if (!DTRACE_INSCRATCH(mstate, size)) {
4843 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4844 			regs[rd] = 0;
4845 			break;
4846 		}
4847 
4848 		if (nargs <= 2)
4849 			remaining = (int64_t)size;
4850 
4851 		if (index < 0) {
4852 			index += len;
4853 
4854 			if (index < 0 && index + remaining > 0) {
4855 				remaining += index;
4856 				index = 0;
4857 			}
4858 		}
4859 
4860 		if (index >= len || index < 0) {
4861 			remaining = 0;
4862 		} else if (remaining < 0) {
4863 			remaining += len - index;
4864 		} else if (index + remaining > size) {
4865 			remaining = size - index;
4866 		}
4867 
4868 		for (i = 0; i < remaining; i++) {
4869 			if ((d[i] = dtrace_load8(s + index + i)) == '\0')
4870 				break;
4871 		}
4872 
4873 		d[i] = '\0';
4874 
4875 		mstate->dtms_scratch_ptr += size;
4876 		regs[rd] = (uintptr_t)d;
4877 		break;
4878 	}
4879 
4880 	case DIF_SUBR_JSON: {
4881 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4882 		uintptr_t json = tupregs[0].dttk_value;
4883 		size_t jsonlen = dtrace_strlen((char *)json, size);
4884 		uintptr_t elem = tupregs[1].dttk_value;
4885 		size_t elemlen = dtrace_strlen((char *)elem, size);
4886 
4887 		char *dest = (char *)mstate->dtms_scratch_ptr;
4888 		char *elemlist = (char *)mstate->dtms_scratch_ptr + jsonlen + 1;
4889 		char *ee = elemlist;
4890 		int nelems = 1;
4891 		uintptr_t cur;
4892 
4893 		if (!dtrace_canload(json, jsonlen + 1, mstate, vstate) ||
4894 		    !dtrace_canload(elem, elemlen + 1, mstate, vstate)) {
4895 			regs[rd] = 0;
4896 			break;
4897 		}
4898 
4899 		if (!DTRACE_INSCRATCH(mstate, jsonlen + 1 + elemlen + 1)) {
4900 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4901 			regs[rd] = 0;
4902 			break;
4903 		}
4904 
4905 		/*
4906 		 * Read the element selector and split it up into a packed list
4907 		 * of strings.
4908 		 */
4909 		for (cur = elem; cur < elem + elemlen; cur++) {
4910 			char cc = dtrace_load8(cur);
4911 
4912 			if (cur == elem && cc == '[') {
4913 				/*
4914 				 * If the first element selector key is
4915 				 * actually an array index then ignore the
4916 				 * bracket.
4917 				 */
4918 				continue;
4919 			}
4920 
4921 			if (cc == ']')
4922 				continue;
4923 
4924 			if (cc == '.' || cc == '[') {
4925 				nelems++;
4926 				cc = '\0';
4927 			}
4928 
4929 			*ee++ = cc;
4930 		}
4931 		*ee++ = '\0';
4932 
4933 		if ((regs[rd] = (uintptr_t)dtrace_json(size, json, elemlist,
4934 		    nelems, dest)) != 0)
4935 			mstate->dtms_scratch_ptr += jsonlen + 1;
4936 		break;
4937 	}
4938 
4939 	case DIF_SUBR_TOUPPER:
4940 	case DIF_SUBR_TOLOWER: {
4941 		uintptr_t s = tupregs[0].dttk_value;
4942 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4943 		char *dest = (char *)mstate->dtms_scratch_ptr, c;
4944 		size_t len = dtrace_strlen((char *)s, size);
4945 		char lower, upper, convert;
4946 		int64_t i;
4947 
4948 		if (subr == DIF_SUBR_TOUPPER) {
4949 			lower = 'a';
4950 			upper = 'z';
4951 			convert = 'A';
4952 		} else {
4953 			lower = 'A';
4954 			upper = 'Z';
4955 			convert = 'a';
4956 		}
4957 
4958 		if (!dtrace_canload(s, len + 1, mstate, vstate)) {
4959 			regs[rd] = 0;
4960 			break;
4961 		}
4962 
4963 		if (!DTRACE_INSCRATCH(mstate, size)) {
4964 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4965 			regs[rd] = 0;
4966 			break;
4967 		}
4968 
4969 		for (i = 0; i < size - 1; i++) {
4970 			if ((c = dtrace_load8(s + i)) == '\0')
4971 				break;
4972 
4973 			if (c >= lower && c <= upper)
4974 				c = convert + (c - lower);
4975 
4976 			dest[i] = c;
4977 		}
4978 
4979 		ASSERT(i < size);
4980 		dest[i] = '\0';
4981 		regs[rd] = (uintptr_t)dest;
4982 		mstate->dtms_scratch_ptr += size;
4983 		break;
4984 	}
4985 
4986 #ifdef illumos
4987 	case DIF_SUBR_GETMAJOR:
4988 #ifdef _LP64
4989 		regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR64) & MAXMAJ64;
4990 #else
4991 		regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR) & MAXMAJ;
4992 #endif
4993 		break;
4994 
4995 	case DIF_SUBR_GETMINOR:
4996 #ifdef _LP64
4997 		regs[rd] = tupregs[0].dttk_value & MAXMIN64;
4998 #else
4999 		regs[rd] = tupregs[0].dttk_value & MAXMIN;
5000 #endif
5001 		break;
5002 
5003 	case DIF_SUBR_DDI_PATHNAME: {
5004 		/*
5005 		 * This one is a galactic mess.  We are going to roughly
5006 		 * emulate ddi_pathname(), but it's made more complicated
5007 		 * by the fact that we (a) want to include the minor name and
5008 		 * (b) must proceed iteratively instead of recursively.
5009 		 */
5010 		uintptr_t dest = mstate->dtms_scratch_ptr;
5011 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
5012 		char *start = (char *)dest, *end = start + size - 1;
5013 		uintptr_t daddr = tupregs[0].dttk_value;
5014 		int64_t minor = (int64_t)tupregs[1].dttk_value;
5015 		char *s;
5016 		int i, len, depth = 0;
5017 
5018 		/*
5019 		 * Due to all the pointer jumping we do and context we must
5020 		 * rely upon, we just mandate that the user must have kernel
5021 		 * read privileges to use this routine.
5022 		 */
5023 		if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) == 0) {
5024 			*flags |= CPU_DTRACE_KPRIV;
5025 			*illval = daddr;
5026 			regs[rd] = 0;
5027 		}
5028 
5029 		if (!DTRACE_INSCRATCH(mstate, size)) {
5030 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5031 			regs[rd] = 0;
5032 			break;
5033 		}
5034 
5035 		*end = '\0';
5036 
5037 		/*
5038 		 * We want to have a name for the minor.  In order to do this,
5039 		 * we need to walk the minor list from the devinfo.  We want
5040 		 * to be sure that we don't infinitely walk a circular list,
5041 		 * so we check for circularity by sending a scout pointer
5042 		 * ahead two elements for every element that we iterate over;
5043 		 * if the list is circular, these will ultimately point to the
5044 		 * same element.  You may recognize this little trick as the
5045 		 * answer to a stupid interview question -- one that always
5046 		 * seems to be asked by those who had to have it laboriously
5047 		 * explained to them, and who can't even concisely describe
5048 		 * the conditions under which one would be forced to resort to
5049 		 * this technique.  Needless to say, those conditions are
5050 		 * found here -- and probably only here.  Is this the only use
5051 		 * of this infamous trick in shipping, production code?  If it
5052 		 * isn't, it probably should be...
5053 		 */
5054 		if (minor != -1) {
5055 			uintptr_t maddr = dtrace_loadptr(daddr +
5056 			    offsetof(struct dev_info, devi_minor));
5057 
5058 			uintptr_t next = offsetof(struct ddi_minor_data, next);
5059 			uintptr_t name = offsetof(struct ddi_minor_data,
5060 			    d_minor) + offsetof(struct ddi_minor, name);
5061 			uintptr_t dev = offsetof(struct ddi_minor_data,
5062 			    d_minor) + offsetof(struct ddi_minor, dev);
5063 			uintptr_t scout;
5064 
5065 			if (maddr != NULL)
5066 				scout = dtrace_loadptr(maddr + next);
5067 
5068 			while (maddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
5069 				uint64_t m;
5070 #ifdef _LP64
5071 				m = dtrace_load64(maddr + dev) & MAXMIN64;
5072 #else
5073 				m = dtrace_load32(maddr + dev) & MAXMIN;
5074 #endif
5075 				if (m != minor) {
5076 					maddr = dtrace_loadptr(maddr + next);
5077 
5078 					if (scout == NULL)
5079 						continue;
5080 
5081 					scout = dtrace_loadptr(scout + next);
5082 
5083 					if (scout == NULL)
5084 						continue;
5085 
5086 					scout = dtrace_loadptr(scout + next);
5087 
5088 					if (scout == NULL)
5089 						continue;
5090 
5091 					if (scout == maddr) {
5092 						*flags |= CPU_DTRACE_ILLOP;
5093 						break;
5094 					}
5095 
5096 					continue;
5097 				}
5098 
5099 				/*
5100 				 * We have the minor data.  Now we need to
5101 				 * copy the minor's name into the end of the
5102 				 * pathname.
5103 				 */
5104 				s = (char *)dtrace_loadptr(maddr + name);
5105 				len = dtrace_strlen(s, size);
5106 
5107 				if (*flags & CPU_DTRACE_FAULT)
5108 					break;
5109 
5110 				if (len != 0) {
5111 					if ((end -= (len + 1)) < start)
5112 						break;
5113 
5114 					*end = ':';
5115 				}
5116 
5117 				for (i = 1; i <= len; i++)
5118 					end[i] = dtrace_load8((uintptr_t)s++);
5119 				break;
5120 			}
5121 		}
5122 
5123 		while (daddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
5124 			ddi_node_state_t devi_state;
5125 
5126 			devi_state = dtrace_load32(daddr +
5127 			    offsetof(struct dev_info, devi_node_state));
5128 
5129 			if (*flags & CPU_DTRACE_FAULT)
5130 				break;
5131 
5132 			if (devi_state >= DS_INITIALIZED) {
5133 				s = (char *)dtrace_loadptr(daddr +
5134 				    offsetof(struct dev_info, devi_addr));
5135 				len = dtrace_strlen(s, size);
5136 
5137 				if (*flags & CPU_DTRACE_FAULT)
5138 					break;
5139 
5140 				if (len != 0) {
5141 					if ((end -= (len + 1)) < start)
5142 						break;
5143 
5144 					*end = '@';
5145 				}
5146 
5147 				for (i = 1; i <= len; i++)
5148 					end[i] = dtrace_load8((uintptr_t)s++);
5149 			}
5150 
5151 			/*
5152 			 * Now for the node name...
5153 			 */
5154 			s = (char *)dtrace_loadptr(daddr +
5155 			    offsetof(struct dev_info, devi_node_name));
5156 
5157 			daddr = dtrace_loadptr(daddr +
5158 			    offsetof(struct dev_info, devi_parent));
5159 
5160 			/*
5161 			 * If our parent is NULL (that is, if we're the root
5162 			 * node), we're going to use the special path
5163 			 * "devices".
5164 			 */
5165 			if (daddr == 0)
5166 				s = "devices";
5167 
5168 			len = dtrace_strlen(s, size);
5169 			if (*flags & CPU_DTRACE_FAULT)
5170 				break;
5171 
5172 			if ((end -= (len + 1)) < start)
5173 				break;
5174 
5175 			for (i = 1; i <= len; i++)
5176 				end[i] = dtrace_load8((uintptr_t)s++);
5177 			*end = '/';
5178 
5179 			if (depth++ > dtrace_devdepth_max) {
5180 				*flags |= CPU_DTRACE_ILLOP;
5181 				break;
5182 			}
5183 		}
5184 
5185 		if (end < start)
5186 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5187 
5188 		if (daddr == 0) {
5189 			regs[rd] = (uintptr_t)end;
5190 			mstate->dtms_scratch_ptr += size;
5191 		}
5192 
5193 		break;
5194 	}
5195 #endif
5196 
5197 	case DIF_SUBR_STRJOIN: {
5198 		char *d = (char *)mstate->dtms_scratch_ptr;
5199 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
5200 		uintptr_t s1 = tupregs[0].dttk_value;
5201 		uintptr_t s2 = tupregs[1].dttk_value;
5202 		int i = 0;
5203 
5204 		if (!dtrace_strcanload(s1, size, mstate, vstate) ||
5205 		    !dtrace_strcanload(s2, size, mstate, vstate)) {
5206 			regs[rd] = 0;
5207 			break;
5208 		}
5209 
5210 		if (!DTRACE_INSCRATCH(mstate, size)) {
5211 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5212 			regs[rd] = 0;
5213 			break;
5214 		}
5215 
5216 		for (;;) {
5217 			if (i >= size) {
5218 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5219 				regs[rd] = 0;
5220 				break;
5221 			}
5222 
5223 			if ((d[i++] = dtrace_load8(s1++)) == '\0') {
5224 				i--;
5225 				break;
5226 			}
5227 		}
5228 
5229 		for (;;) {
5230 			if (i >= size) {
5231 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5232 				regs[rd] = 0;
5233 				break;
5234 			}
5235 
5236 			if ((d[i++] = dtrace_load8(s2++)) == '\0')
5237 				break;
5238 		}
5239 
5240 		if (i < size) {
5241 			mstate->dtms_scratch_ptr += i;
5242 			regs[rd] = (uintptr_t)d;
5243 		}
5244 
5245 		break;
5246 	}
5247 
5248 	case DIF_SUBR_STRTOLL: {
5249 		uintptr_t s = tupregs[0].dttk_value;
5250 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
5251 		int base = 10;
5252 
5253 		if (nargs > 1) {
5254 			if ((base = tupregs[1].dttk_value) <= 1 ||
5255 			    base > ('z' - 'a' + 1) + ('9' - '0' + 1)) {
5256 				*flags |= CPU_DTRACE_ILLOP;
5257 				break;
5258 			}
5259 		}
5260 
5261 		if (!dtrace_strcanload(s, size, mstate, vstate)) {
5262 			regs[rd] = INT64_MIN;
5263 			break;
5264 		}
5265 
5266 		regs[rd] = dtrace_strtoll((char *)s, base, size);
5267 		break;
5268 	}
5269 
5270 	case DIF_SUBR_LLTOSTR: {
5271 		int64_t i = (int64_t)tupregs[0].dttk_value;
5272 		uint64_t val, digit;
5273 		uint64_t size = 65;	/* enough room for 2^64 in binary */
5274 		char *end = (char *)mstate->dtms_scratch_ptr + size - 1;
5275 		int base = 10;
5276 
5277 		if (nargs > 1) {
5278 			if ((base = tupregs[1].dttk_value) <= 1 ||
5279 			    base > ('z' - 'a' + 1) + ('9' - '0' + 1)) {
5280 				*flags |= CPU_DTRACE_ILLOP;
5281 				break;
5282 			}
5283 		}
5284 
5285 		val = (base == 10 && i < 0) ? i * -1 : i;
5286 
5287 		if (!DTRACE_INSCRATCH(mstate, size)) {
5288 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5289 			regs[rd] = 0;
5290 			break;
5291 		}
5292 
5293 		for (*end-- = '\0'; val; val /= base) {
5294 			if ((digit = val % base) <= '9' - '0') {
5295 				*end-- = '0' + digit;
5296 			} else {
5297 				*end-- = 'a' + (digit - ('9' - '0') - 1);
5298 			}
5299 		}
5300 
5301 		if (i == 0 && base == 16)
5302 			*end-- = '0';
5303 
5304 		if (base == 16)
5305 			*end-- = 'x';
5306 
5307 		if (i == 0 || base == 8 || base == 16)
5308 			*end-- = '0';
5309 
5310 		if (i < 0 && base == 10)
5311 			*end-- = '-';
5312 
5313 		regs[rd] = (uintptr_t)end + 1;
5314 		mstate->dtms_scratch_ptr += size;
5315 		break;
5316 	}
5317 
5318 	case DIF_SUBR_HTONS:
5319 	case DIF_SUBR_NTOHS:
5320 #if BYTE_ORDER == BIG_ENDIAN
5321 		regs[rd] = (uint16_t)tupregs[0].dttk_value;
5322 #else
5323 		regs[rd] = DT_BSWAP_16((uint16_t)tupregs[0].dttk_value);
5324 #endif
5325 		break;
5326 
5327 
5328 	case DIF_SUBR_HTONL:
5329 	case DIF_SUBR_NTOHL:
5330 #if BYTE_ORDER == BIG_ENDIAN
5331 		regs[rd] = (uint32_t)tupregs[0].dttk_value;
5332 #else
5333 		regs[rd] = DT_BSWAP_32((uint32_t)tupregs[0].dttk_value);
5334 #endif
5335 		break;
5336 
5337 
5338 	case DIF_SUBR_HTONLL:
5339 	case DIF_SUBR_NTOHLL:
5340 #if BYTE_ORDER == BIG_ENDIAN
5341 		regs[rd] = (uint64_t)tupregs[0].dttk_value;
5342 #else
5343 		regs[rd] = DT_BSWAP_64((uint64_t)tupregs[0].dttk_value);
5344 #endif
5345 		break;
5346 
5347 
5348 	case DIF_SUBR_DIRNAME:
5349 	case DIF_SUBR_BASENAME: {
5350 		char *dest = (char *)mstate->dtms_scratch_ptr;
5351 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
5352 		uintptr_t src = tupregs[0].dttk_value;
5353 		int i, j, len = dtrace_strlen((char *)src, size);
5354 		int lastbase = -1, firstbase = -1, lastdir = -1;
5355 		int start, end;
5356 
5357 		if (!dtrace_canload(src, len + 1, mstate, vstate)) {
5358 			regs[rd] = 0;
5359 			break;
5360 		}
5361 
5362 		if (!DTRACE_INSCRATCH(mstate, size)) {
5363 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5364 			regs[rd] = 0;
5365 			break;
5366 		}
5367 
5368 		/*
5369 		 * The basename and dirname for a zero-length string is
5370 		 * defined to be "."
5371 		 */
5372 		if (len == 0) {
5373 			len = 1;
5374 			src = (uintptr_t)".";
5375 		}
5376 
5377 		/*
5378 		 * Start from the back of the string, moving back toward the
5379 		 * front until we see a character that isn't a slash.  That
5380 		 * character is the last character in the basename.
5381 		 */
5382 		for (i = len - 1; i >= 0; i--) {
5383 			if (dtrace_load8(src + i) != '/')
5384 				break;
5385 		}
5386 
5387 		if (i >= 0)
5388 			lastbase = i;
5389 
5390 		/*
5391 		 * Starting from the last character in the basename, move
5392 		 * towards the front until we find a slash.  The character
5393 		 * that we processed immediately before that is the first
5394 		 * character in the basename.
5395 		 */
5396 		for (; i >= 0; i--) {
5397 			if (dtrace_load8(src + i) == '/')
5398 				break;
5399 		}
5400 
5401 		if (i >= 0)
5402 			firstbase = i + 1;
5403 
5404 		/*
5405 		 * Now keep going until we find a non-slash character.  That
5406 		 * character is the last character in the dirname.
5407 		 */
5408 		for (; i >= 0; i--) {
5409 			if (dtrace_load8(src + i) != '/')
5410 				break;
5411 		}
5412 
5413 		if (i >= 0)
5414 			lastdir = i;
5415 
5416 		ASSERT(!(lastbase == -1 && firstbase != -1));
5417 		ASSERT(!(firstbase == -1 && lastdir != -1));
5418 
5419 		if (lastbase == -1) {
5420 			/*
5421 			 * We didn't find a non-slash character.  We know that
5422 			 * the length is non-zero, so the whole string must be
5423 			 * slashes.  In either the dirname or the basename
5424 			 * case, we return '/'.
5425 			 */
5426 			ASSERT(firstbase == -1);
5427 			firstbase = lastbase = lastdir = 0;
5428 		}
5429 
5430 		if (firstbase == -1) {
5431 			/*
5432 			 * The entire string consists only of a basename
5433 			 * component.  If we're looking for dirname, we need
5434 			 * to change our string to be just "."; if we're
5435 			 * looking for a basename, we'll just set the first
5436 			 * character of the basename to be 0.
5437 			 */
5438 			if (subr == DIF_SUBR_DIRNAME) {
5439 				ASSERT(lastdir == -1);
5440 				src = (uintptr_t)".";
5441 				lastdir = 0;
5442 			} else {
5443 				firstbase = 0;
5444 			}
5445 		}
5446 
5447 		if (subr == DIF_SUBR_DIRNAME) {
5448 			if (lastdir == -1) {
5449 				/*
5450 				 * We know that we have a slash in the name --
5451 				 * or lastdir would be set to 0, above.  And
5452 				 * because lastdir is -1, we know that this
5453 				 * slash must be the first character.  (That
5454 				 * is, the full string must be of the form
5455 				 * "/basename".)  In this case, the last
5456 				 * character of the directory name is 0.
5457 				 */
5458 				lastdir = 0;
5459 			}
5460 
5461 			start = 0;
5462 			end = lastdir;
5463 		} else {
5464 			ASSERT(subr == DIF_SUBR_BASENAME);
5465 			ASSERT(firstbase != -1 && lastbase != -1);
5466 			start = firstbase;
5467 			end = lastbase;
5468 		}
5469 
5470 		for (i = start, j = 0; i <= end && j < size - 1; i++, j++)
5471 			dest[j] = dtrace_load8(src + i);
5472 
5473 		dest[j] = '\0';
5474 		regs[rd] = (uintptr_t)dest;
5475 		mstate->dtms_scratch_ptr += size;
5476 		break;
5477 	}
5478 
5479 	case DIF_SUBR_GETF: {
5480 		uintptr_t fd = tupregs[0].dttk_value;
5481 		struct filedesc *fdp;
5482 		file_t *fp;
5483 
5484 		if (!dtrace_priv_proc(state)) {
5485 			regs[rd] = 0;
5486 			break;
5487 		}
5488 		fdp = curproc->p_fd;
5489 		FILEDESC_SLOCK(fdp);
5490 		fp = fget_locked(fdp, fd);
5491 		mstate->dtms_getf = fp;
5492 		regs[rd] = (uintptr_t)fp;
5493 		FILEDESC_SUNLOCK(fdp);
5494 		break;
5495 	}
5496 
5497 	case DIF_SUBR_CLEANPATH: {
5498 		char *dest = (char *)mstate->dtms_scratch_ptr, c;
5499 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
5500 		uintptr_t src = tupregs[0].dttk_value;
5501 		int i = 0, j = 0;
5502 #ifdef illumos
5503 		zone_t *z;
5504 #endif
5505 
5506 		if (!dtrace_strcanload(src, size, mstate, vstate)) {
5507 			regs[rd] = 0;
5508 			break;
5509 		}
5510 
5511 		if (!DTRACE_INSCRATCH(mstate, size)) {
5512 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5513 			regs[rd] = 0;
5514 			break;
5515 		}
5516 
5517 		/*
5518 		 * Move forward, loading each character.
5519 		 */
5520 		do {
5521 			c = dtrace_load8(src + i++);
5522 next:
5523 			if (j + 5 >= size)	/* 5 = strlen("/..c\0") */
5524 				break;
5525 
5526 			if (c != '/') {
5527 				dest[j++] = c;
5528 				continue;
5529 			}
5530 
5531 			c = dtrace_load8(src + i++);
5532 
5533 			if (c == '/') {
5534 				/*
5535 				 * We have two slashes -- we can just advance
5536 				 * to the next character.
5537 				 */
5538 				goto next;
5539 			}
5540 
5541 			if (c != '.') {
5542 				/*
5543 				 * This is not "." and it's not ".." -- we can
5544 				 * just store the "/" and this character and
5545 				 * drive on.
5546 				 */
5547 				dest[j++] = '/';
5548 				dest[j++] = c;
5549 				continue;
5550 			}
5551 
5552 			c = dtrace_load8(src + i++);
5553 
5554 			if (c == '/') {
5555 				/*
5556 				 * This is a "/./" component.  We're not going
5557 				 * to store anything in the destination buffer;
5558 				 * we're just going to go to the next component.
5559 				 */
5560 				goto next;
5561 			}
5562 
5563 			if (c != '.') {
5564 				/*
5565 				 * This is not ".." -- we can just store the
5566 				 * "/." and this character and continue
5567 				 * processing.
5568 				 */
5569 				dest[j++] = '/';
5570 				dest[j++] = '.';
5571 				dest[j++] = c;
5572 				continue;
5573 			}
5574 
5575 			c = dtrace_load8(src + i++);
5576 
5577 			if (c != '/' && c != '\0') {
5578 				/*
5579 				 * This is not ".." -- it's "..[mumble]".
5580 				 * We'll store the "/.." and this character
5581 				 * and continue processing.
5582 				 */
5583 				dest[j++] = '/';
5584 				dest[j++] = '.';
5585 				dest[j++] = '.';
5586 				dest[j++] = c;
5587 				continue;
5588 			}
5589 
5590 			/*
5591 			 * This is "/../" or "/..\0".  We need to back up
5592 			 * our destination pointer until we find a "/".
5593 			 */
5594 			i--;
5595 			while (j != 0 && dest[--j] != '/')
5596 				continue;
5597 
5598 			if (c == '\0')
5599 				dest[++j] = '/';
5600 		} while (c != '\0');
5601 
5602 		dest[j] = '\0';
5603 
5604 #ifdef illumos
5605 		if (mstate->dtms_getf != NULL &&
5606 		    !(mstate->dtms_access & DTRACE_ACCESS_KERNEL) &&
5607 		    (z = state->dts_cred.dcr_cred->cr_zone) != kcred->cr_zone) {
5608 			/*
5609 			 * If we've done a getf() as a part of this ECB and we
5610 			 * don't have kernel access (and we're not in the global
5611 			 * zone), check if the path we cleaned up begins with
5612 			 * the zone's root path, and trim it off if so.  Note
5613 			 * that this is an output cleanliness issue, not a
5614 			 * security issue: knowing one's zone root path does
5615 			 * not enable privilege escalation.
5616 			 */
5617 			if (strstr(dest, z->zone_rootpath) == dest)
5618 				dest += strlen(z->zone_rootpath) - 1;
5619 		}
5620 #endif
5621 
5622 		regs[rd] = (uintptr_t)dest;
5623 		mstate->dtms_scratch_ptr += size;
5624 		break;
5625 	}
5626 
5627 	case DIF_SUBR_INET_NTOA:
5628 	case DIF_SUBR_INET_NTOA6:
5629 	case DIF_SUBR_INET_NTOP: {
5630 		size_t size;
5631 		int af, argi, i;
5632 		char *base, *end;
5633 
5634 		if (subr == DIF_SUBR_INET_NTOP) {
5635 			af = (int)tupregs[0].dttk_value;
5636 			argi = 1;
5637 		} else {
5638 			af = subr == DIF_SUBR_INET_NTOA ? AF_INET: AF_INET6;
5639 			argi = 0;
5640 		}
5641 
5642 		if (af == AF_INET) {
5643 			ipaddr_t ip4;
5644 			uint8_t *ptr8, val;
5645 
5646 			if (!dtrace_canload(tupregs[argi].dttk_value,
5647 			    sizeof (ipaddr_t), mstate, vstate)) {
5648 				regs[rd] = 0;
5649 				break;
5650 			}
5651 
5652 			/*
5653 			 * Safely load the IPv4 address.
5654 			 */
5655 			ip4 = dtrace_load32(tupregs[argi].dttk_value);
5656 
5657 			/*
5658 			 * Check an IPv4 string will fit in scratch.
5659 			 */
5660 			size = INET_ADDRSTRLEN;
5661 			if (!DTRACE_INSCRATCH(mstate, size)) {
5662 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5663 				regs[rd] = 0;
5664 				break;
5665 			}
5666 			base = (char *)mstate->dtms_scratch_ptr;
5667 			end = (char *)mstate->dtms_scratch_ptr + size - 1;
5668 
5669 			/*
5670 			 * Stringify as a dotted decimal quad.
5671 			 */
5672 			*end-- = '\0';
5673 			ptr8 = (uint8_t *)&ip4;
5674 			for (i = 3; i >= 0; i--) {
5675 				val = ptr8[i];
5676 
5677 				if (val == 0) {
5678 					*end-- = '0';
5679 				} else {
5680 					for (; val; val /= 10) {
5681 						*end-- = '0' + (val % 10);
5682 					}
5683 				}
5684 
5685 				if (i > 0)
5686 					*end-- = '.';
5687 			}
5688 			ASSERT(end + 1 >= base);
5689 
5690 		} else if (af == AF_INET6) {
5691 			struct in6_addr ip6;
5692 			int firstzero, tryzero, numzero, v6end;
5693 			uint16_t val;
5694 			const char digits[] = "0123456789abcdef";
5695 
5696 			/*
5697 			 * Stringify using RFC 1884 convention 2 - 16 bit
5698 			 * hexadecimal values with a zero-run compression.
5699 			 * Lower case hexadecimal digits are used.
5700 			 * 	eg, fe80::214:4fff:fe0b:76c8.
5701 			 * The IPv4 embedded form is returned for inet_ntop,
5702 			 * just the IPv4 string is returned for inet_ntoa6.
5703 			 */
5704 
5705 			if (!dtrace_canload(tupregs[argi].dttk_value,
5706 			    sizeof (struct in6_addr), mstate, vstate)) {
5707 				regs[rd] = 0;
5708 				break;
5709 			}
5710 
5711 			/*
5712 			 * Safely load the IPv6 address.
5713 			 */
5714 			dtrace_bcopy(
5715 			    (void *)(uintptr_t)tupregs[argi].dttk_value,
5716 			    (void *)(uintptr_t)&ip6, sizeof (struct in6_addr));
5717 
5718 			/*
5719 			 * Check an IPv6 string will fit in scratch.
5720 			 */
5721 			size = INET6_ADDRSTRLEN;
5722 			if (!DTRACE_INSCRATCH(mstate, size)) {
5723 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5724 				regs[rd] = 0;
5725 				break;
5726 			}
5727 			base = (char *)mstate->dtms_scratch_ptr;
5728 			end = (char *)mstate->dtms_scratch_ptr + size - 1;
5729 			*end-- = '\0';
5730 
5731 			/*
5732 			 * Find the longest run of 16 bit zero values
5733 			 * for the single allowed zero compression - "::".
5734 			 */
5735 			firstzero = -1;
5736 			tryzero = -1;
5737 			numzero = 1;
5738 			for (i = 0; i < sizeof (struct in6_addr); i++) {
5739 #ifdef illumos
5740 				if (ip6._S6_un._S6_u8[i] == 0 &&
5741 #else
5742 				if (ip6.__u6_addr.__u6_addr8[i] == 0 &&
5743 #endif
5744 				    tryzero == -1 && i % 2 == 0) {
5745 					tryzero = i;
5746 					continue;
5747 				}
5748 
5749 				if (tryzero != -1 &&
5750 #ifdef illumos
5751 				    (ip6._S6_un._S6_u8[i] != 0 ||
5752 #else
5753 				    (ip6.__u6_addr.__u6_addr8[i] != 0 ||
5754 #endif
5755 				    i == sizeof (struct in6_addr) - 1)) {
5756 
5757 					if (i - tryzero <= numzero) {
5758 						tryzero = -1;
5759 						continue;
5760 					}
5761 
5762 					firstzero = tryzero;
5763 					numzero = i - i % 2 - tryzero;
5764 					tryzero = -1;
5765 
5766 #ifdef illumos
5767 					if (ip6._S6_un._S6_u8[i] == 0 &&
5768 #else
5769 					if (ip6.__u6_addr.__u6_addr8[i] == 0 &&
5770 #endif
5771 					    i == sizeof (struct in6_addr) - 1)
5772 						numzero += 2;
5773 				}
5774 			}
5775 			ASSERT(firstzero + numzero <= sizeof (struct in6_addr));
5776 
5777 			/*
5778 			 * Check for an IPv4 embedded address.
5779 			 */
5780 			v6end = sizeof (struct in6_addr) - 2;
5781 			if (IN6_IS_ADDR_V4MAPPED(&ip6) ||
5782 			    IN6_IS_ADDR_V4COMPAT(&ip6)) {
5783 				for (i = sizeof (struct in6_addr) - 1;
5784 				    i >= DTRACE_V4MAPPED_OFFSET; i--) {
5785 					ASSERT(end >= base);
5786 
5787 #ifdef illumos
5788 					val = ip6._S6_un._S6_u8[i];
5789 #else
5790 					val = ip6.__u6_addr.__u6_addr8[i];
5791 #endif
5792 
5793 					if (val == 0) {
5794 						*end-- = '0';
5795 					} else {
5796 						for (; val; val /= 10) {
5797 							*end-- = '0' + val % 10;
5798 						}
5799 					}
5800 
5801 					if (i > DTRACE_V4MAPPED_OFFSET)
5802 						*end-- = '.';
5803 				}
5804 
5805 				if (subr == DIF_SUBR_INET_NTOA6)
5806 					goto inetout;
5807 
5808 				/*
5809 				 * Set v6end to skip the IPv4 address that
5810 				 * we have already stringified.
5811 				 */
5812 				v6end = 10;
5813 			}
5814 
5815 			/*
5816 			 * Build the IPv6 string by working through the
5817 			 * address in reverse.
5818 			 */
5819 			for (i = v6end; i >= 0; i -= 2) {
5820 				ASSERT(end >= base);
5821 
5822 				if (i == firstzero + numzero - 2) {
5823 					*end-- = ':';
5824 					*end-- = ':';
5825 					i -= numzero - 2;
5826 					continue;
5827 				}
5828 
5829 				if (i < 14 && i != firstzero - 2)
5830 					*end-- = ':';
5831 
5832 #ifdef illumos
5833 				val = (ip6._S6_un._S6_u8[i] << 8) +
5834 				    ip6._S6_un._S6_u8[i + 1];
5835 #else
5836 				val = (ip6.__u6_addr.__u6_addr8[i] << 8) +
5837 				    ip6.__u6_addr.__u6_addr8[i + 1];
5838 #endif
5839 
5840 				if (val == 0) {
5841 					*end-- = '0';
5842 				} else {
5843 					for (; val; val /= 16) {
5844 						*end-- = digits[val % 16];
5845 					}
5846 				}
5847 			}
5848 			ASSERT(end + 1 >= base);
5849 
5850 		} else {
5851 			/*
5852 			 * The user didn't use AH_INET or AH_INET6.
5853 			 */
5854 			DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
5855 			regs[rd] = 0;
5856 			break;
5857 		}
5858 
5859 inetout:	regs[rd] = (uintptr_t)end + 1;
5860 		mstate->dtms_scratch_ptr += size;
5861 		break;
5862 	}
5863 
5864 	case DIF_SUBR_MEMREF: {
5865 		uintptr_t size = 2 * sizeof(uintptr_t);
5866 		uintptr_t *memref = (uintptr_t *) P2ROUNDUP(mstate->dtms_scratch_ptr, sizeof(uintptr_t));
5867 		size_t scratch_size = ((uintptr_t) memref - mstate->dtms_scratch_ptr) + size;
5868 
5869 		/* address and length */
5870 		memref[0] = tupregs[0].dttk_value;
5871 		memref[1] = tupregs[1].dttk_value;
5872 
5873 		regs[rd] = (uintptr_t) memref;
5874 		mstate->dtms_scratch_ptr += scratch_size;
5875 		break;
5876 	}
5877 
5878 #ifndef illumos
5879 	case DIF_SUBR_MEMSTR: {
5880 		char *str = (char *)mstate->dtms_scratch_ptr;
5881 		uintptr_t mem = tupregs[0].dttk_value;
5882 		char c = tupregs[1].dttk_value;
5883 		size_t size = tupregs[2].dttk_value;
5884 		uint8_t n;
5885 		int i;
5886 
5887 		regs[rd] = 0;
5888 
5889 		if (size == 0)
5890 			break;
5891 
5892 		if (!dtrace_canload(mem, size - 1, mstate, vstate))
5893 			break;
5894 
5895 		if (!DTRACE_INSCRATCH(mstate, size)) {
5896 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5897 			break;
5898 		}
5899 
5900 		if (dtrace_memstr_max != 0 && size > dtrace_memstr_max) {
5901 			*flags |= CPU_DTRACE_ILLOP;
5902 			break;
5903 		}
5904 
5905 		for (i = 0; i < size - 1; i++) {
5906 			n = dtrace_load8(mem++);
5907 			str[i] = (n == 0) ? c : n;
5908 		}
5909 		str[size - 1] = 0;
5910 
5911 		regs[rd] = (uintptr_t)str;
5912 		mstate->dtms_scratch_ptr += size;
5913 		break;
5914 	}
5915 #endif
5916 
5917 	case DIF_SUBR_TYPEREF: {
5918 		uintptr_t size = 4 * sizeof(uintptr_t);
5919 		uintptr_t *typeref = (uintptr_t *) P2ROUNDUP(mstate->dtms_scratch_ptr, sizeof(uintptr_t));
5920 		size_t scratch_size = ((uintptr_t) typeref - mstate->dtms_scratch_ptr) + size;
5921 
5922 		/* address, num_elements, type_str, type_len */
5923 		typeref[0] = tupregs[0].dttk_value;
5924 		typeref[1] = tupregs[1].dttk_value;
5925 		typeref[2] = tupregs[2].dttk_value;
5926 		typeref[3] = tupregs[3].dttk_value;
5927 
5928 		regs[rd] = (uintptr_t) typeref;
5929 		mstate->dtms_scratch_ptr += scratch_size;
5930 		break;
5931 	}
5932 	}
5933 }
5934 
5935 /*
5936  * Emulate the execution of DTrace IR instructions specified by the given
5937  * DIF object.  This function is deliberately void of assertions as all of
5938  * the necessary checks are handled by a call to dtrace_difo_validate().
5939  */
5940 static uint64_t
5941 dtrace_dif_emulate(dtrace_difo_t *difo, dtrace_mstate_t *mstate,
5942     dtrace_vstate_t *vstate, dtrace_state_t *state)
5943 {
5944 	const dif_instr_t *text = difo->dtdo_buf;
5945 	const uint_t textlen = difo->dtdo_len;
5946 	const char *strtab = difo->dtdo_strtab;
5947 	const uint64_t *inttab = difo->dtdo_inttab;
5948 
5949 	uint64_t rval = 0;
5950 	dtrace_statvar_t *svar;
5951 	dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
5952 	dtrace_difv_t *v;
5953 	volatile uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
5954 	volatile uintptr_t *illval = &cpu_core[curcpu].cpuc_dtrace_illval;
5955 
5956 	dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
5957 	uint64_t regs[DIF_DIR_NREGS];
5958 	uint64_t *tmp;
5959 
5960 	uint8_t cc_n = 0, cc_z = 0, cc_v = 0, cc_c = 0;
5961 	int64_t cc_r;
5962 	uint_t pc = 0, id, opc = 0;
5963 	uint8_t ttop = 0;
5964 	dif_instr_t instr;
5965 	uint_t r1, r2, rd;
5966 
5967 	/*
5968 	 * We stash the current DIF object into the machine state: we need it
5969 	 * for subsequent access checking.
5970 	 */
5971 	mstate->dtms_difo = difo;
5972 
5973 	regs[DIF_REG_R0] = 0; 		/* %r0 is fixed at zero */
5974 
5975 	while (pc < textlen && !(*flags & CPU_DTRACE_FAULT)) {
5976 		opc = pc;
5977 
5978 		instr = text[pc++];
5979 		r1 = DIF_INSTR_R1(instr);
5980 		r2 = DIF_INSTR_R2(instr);
5981 		rd = DIF_INSTR_RD(instr);
5982 
5983 		switch (DIF_INSTR_OP(instr)) {
5984 		case DIF_OP_OR:
5985 			regs[rd] = regs[r1] | regs[r2];
5986 			break;
5987 		case DIF_OP_XOR:
5988 			regs[rd] = regs[r1] ^ regs[r2];
5989 			break;
5990 		case DIF_OP_AND:
5991 			regs[rd] = regs[r1] & regs[r2];
5992 			break;
5993 		case DIF_OP_SLL:
5994 			regs[rd] = regs[r1] << regs[r2];
5995 			break;
5996 		case DIF_OP_SRL:
5997 			regs[rd] = regs[r1] >> regs[r2];
5998 			break;
5999 		case DIF_OP_SUB:
6000 			regs[rd] = regs[r1] - regs[r2];
6001 			break;
6002 		case DIF_OP_ADD:
6003 			regs[rd] = regs[r1] + regs[r2];
6004 			break;
6005 		case DIF_OP_MUL:
6006 			regs[rd] = regs[r1] * regs[r2];
6007 			break;
6008 		case DIF_OP_SDIV:
6009 			if (regs[r2] == 0) {
6010 				regs[rd] = 0;
6011 				*flags |= CPU_DTRACE_DIVZERO;
6012 			} else {
6013 				regs[rd] = (int64_t)regs[r1] /
6014 				    (int64_t)regs[r2];
6015 			}
6016 			break;
6017 
6018 		case DIF_OP_UDIV:
6019 			if (regs[r2] == 0) {
6020 				regs[rd] = 0;
6021 				*flags |= CPU_DTRACE_DIVZERO;
6022 			} else {
6023 				regs[rd] = regs[r1] / regs[r2];
6024 			}
6025 			break;
6026 
6027 		case DIF_OP_SREM:
6028 			if (regs[r2] == 0) {
6029 				regs[rd] = 0;
6030 				*flags |= CPU_DTRACE_DIVZERO;
6031 			} else {
6032 				regs[rd] = (int64_t)regs[r1] %
6033 				    (int64_t)regs[r2];
6034 			}
6035 			break;
6036 
6037 		case DIF_OP_UREM:
6038 			if (regs[r2] == 0) {
6039 				regs[rd] = 0;
6040 				*flags |= CPU_DTRACE_DIVZERO;
6041 			} else {
6042 				regs[rd] = regs[r1] % regs[r2];
6043 			}
6044 			break;
6045 
6046 		case DIF_OP_NOT:
6047 			regs[rd] = ~regs[r1];
6048 			break;
6049 		case DIF_OP_MOV:
6050 			regs[rd] = regs[r1];
6051 			break;
6052 		case DIF_OP_CMP:
6053 			cc_r = regs[r1] - regs[r2];
6054 			cc_n = cc_r < 0;
6055 			cc_z = cc_r == 0;
6056 			cc_v = 0;
6057 			cc_c = regs[r1] < regs[r2];
6058 			break;
6059 		case DIF_OP_TST:
6060 			cc_n = cc_v = cc_c = 0;
6061 			cc_z = regs[r1] == 0;
6062 			break;
6063 		case DIF_OP_BA:
6064 			pc = DIF_INSTR_LABEL(instr);
6065 			break;
6066 		case DIF_OP_BE:
6067 			if (cc_z)
6068 				pc = DIF_INSTR_LABEL(instr);
6069 			break;
6070 		case DIF_OP_BNE:
6071 			if (cc_z == 0)
6072 				pc = DIF_INSTR_LABEL(instr);
6073 			break;
6074 		case DIF_OP_BG:
6075 			if ((cc_z | (cc_n ^ cc_v)) == 0)
6076 				pc = DIF_INSTR_LABEL(instr);
6077 			break;
6078 		case DIF_OP_BGU:
6079 			if ((cc_c | cc_z) == 0)
6080 				pc = DIF_INSTR_LABEL(instr);
6081 			break;
6082 		case DIF_OP_BGE:
6083 			if ((cc_n ^ cc_v) == 0)
6084 				pc = DIF_INSTR_LABEL(instr);
6085 			break;
6086 		case DIF_OP_BGEU:
6087 			if (cc_c == 0)
6088 				pc = DIF_INSTR_LABEL(instr);
6089 			break;
6090 		case DIF_OP_BL:
6091 			if (cc_n ^ cc_v)
6092 				pc = DIF_INSTR_LABEL(instr);
6093 			break;
6094 		case DIF_OP_BLU:
6095 			if (cc_c)
6096 				pc = DIF_INSTR_LABEL(instr);
6097 			break;
6098 		case DIF_OP_BLE:
6099 			if (cc_z | (cc_n ^ cc_v))
6100 				pc = DIF_INSTR_LABEL(instr);
6101 			break;
6102 		case DIF_OP_BLEU:
6103 			if (cc_c | cc_z)
6104 				pc = DIF_INSTR_LABEL(instr);
6105 			break;
6106 		case DIF_OP_RLDSB:
6107 			if (!dtrace_canload(regs[r1], 1, mstate, vstate))
6108 				break;
6109 			/*FALLTHROUGH*/
6110 		case DIF_OP_LDSB:
6111 			regs[rd] = (int8_t)dtrace_load8(regs[r1]);
6112 			break;
6113 		case DIF_OP_RLDSH:
6114 			if (!dtrace_canload(regs[r1], 2, mstate, vstate))
6115 				break;
6116 			/*FALLTHROUGH*/
6117 		case DIF_OP_LDSH:
6118 			regs[rd] = (int16_t)dtrace_load16(regs[r1]);
6119 			break;
6120 		case DIF_OP_RLDSW:
6121 			if (!dtrace_canload(regs[r1], 4, mstate, vstate))
6122 				break;
6123 			/*FALLTHROUGH*/
6124 		case DIF_OP_LDSW:
6125 			regs[rd] = (int32_t)dtrace_load32(regs[r1]);
6126 			break;
6127 		case DIF_OP_RLDUB:
6128 			if (!dtrace_canload(regs[r1], 1, mstate, vstate))
6129 				break;
6130 			/*FALLTHROUGH*/
6131 		case DIF_OP_LDUB:
6132 			regs[rd] = dtrace_load8(regs[r1]);
6133 			break;
6134 		case DIF_OP_RLDUH:
6135 			if (!dtrace_canload(regs[r1], 2, mstate, vstate))
6136 				break;
6137 			/*FALLTHROUGH*/
6138 		case DIF_OP_LDUH:
6139 			regs[rd] = dtrace_load16(regs[r1]);
6140 			break;
6141 		case DIF_OP_RLDUW:
6142 			if (!dtrace_canload(regs[r1], 4, mstate, vstate))
6143 				break;
6144 			/*FALLTHROUGH*/
6145 		case DIF_OP_LDUW:
6146 			regs[rd] = dtrace_load32(regs[r1]);
6147 			break;
6148 		case DIF_OP_RLDX:
6149 			if (!dtrace_canload(regs[r1], 8, mstate, vstate))
6150 				break;
6151 			/*FALLTHROUGH*/
6152 		case DIF_OP_LDX:
6153 			regs[rd] = dtrace_load64(regs[r1]);
6154 			break;
6155 		case DIF_OP_ULDSB:
6156 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6157 			regs[rd] = (int8_t)
6158 			    dtrace_fuword8((void *)(uintptr_t)regs[r1]);
6159 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6160 			break;
6161 		case DIF_OP_ULDSH:
6162 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6163 			regs[rd] = (int16_t)
6164 			    dtrace_fuword16((void *)(uintptr_t)regs[r1]);
6165 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6166 			break;
6167 		case DIF_OP_ULDSW:
6168 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6169 			regs[rd] = (int32_t)
6170 			    dtrace_fuword32((void *)(uintptr_t)regs[r1]);
6171 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6172 			break;
6173 		case DIF_OP_ULDUB:
6174 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6175 			regs[rd] =
6176 			    dtrace_fuword8((void *)(uintptr_t)regs[r1]);
6177 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6178 			break;
6179 		case DIF_OP_ULDUH:
6180 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6181 			regs[rd] =
6182 			    dtrace_fuword16((void *)(uintptr_t)regs[r1]);
6183 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6184 			break;
6185 		case DIF_OP_ULDUW:
6186 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6187 			regs[rd] =
6188 			    dtrace_fuword32((void *)(uintptr_t)regs[r1]);
6189 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6190 			break;
6191 		case DIF_OP_ULDX:
6192 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6193 			regs[rd] =
6194 			    dtrace_fuword64((void *)(uintptr_t)regs[r1]);
6195 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6196 			break;
6197 		case DIF_OP_RET:
6198 			rval = regs[rd];
6199 			pc = textlen;
6200 			break;
6201 		case DIF_OP_NOP:
6202 			break;
6203 		case DIF_OP_SETX:
6204 			regs[rd] = inttab[DIF_INSTR_INTEGER(instr)];
6205 			break;
6206 		case DIF_OP_SETS:
6207 			regs[rd] = (uint64_t)(uintptr_t)
6208 			    (strtab + DIF_INSTR_STRING(instr));
6209 			break;
6210 		case DIF_OP_SCMP: {
6211 			size_t sz = state->dts_options[DTRACEOPT_STRSIZE];
6212 			uintptr_t s1 = regs[r1];
6213 			uintptr_t s2 = regs[r2];
6214 
6215 			if (s1 != 0 &&
6216 			    !dtrace_strcanload(s1, sz, mstate, vstate))
6217 				break;
6218 			if (s2 != 0 &&
6219 			    !dtrace_strcanload(s2, sz, mstate, vstate))
6220 				break;
6221 
6222 			cc_r = dtrace_strncmp((char *)s1, (char *)s2, sz);
6223 
6224 			cc_n = cc_r < 0;
6225 			cc_z = cc_r == 0;
6226 			cc_v = cc_c = 0;
6227 			break;
6228 		}
6229 		case DIF_OP_LDGA:
6230 			regs[rd] = dtrace_dif_variable(mstate, state,
6231 			    r1, regs[r2]);
6232 			break;
6233 		case DIF_OP_LDGS:
6234 			id = DIF_INSTR_VAR(instr);
6235 
6236 			if (id >= DIF_VAR_OTHER_UBASE) {
6237 				uintptr_t a;
6238 
6239 				id -= DIF_VAR_OTHER_UBASE;
6240 				svar = vstate->dtvs_globals[id];
6241 				ASSERT(svar != NULL);
6242 				v = &svar->dtsv_var;
6243 
6244 				if (!(v->dtdv_type.dtdt_flags & DIF_TF_BYREF)) {
6245 					regs[rd] = svar->dtsv_data;
6246 					break;
6247 				}
6248 
6249 				a = (uintptr_t)svar->dtsv_data;
6250 
6251 				if (*(uint8_t *)a == UINT8_MAX) {
6252 					/*
6253 					 * If the 0th byte is set to UINT8_MAX
6254 					 * then this is to be treated as a
6255 					 * reference to a NULL variable.
6256 					 */
6257 					regs[rd] = 0;
6258 				} else {
6259 					regs[rd] = a + sizeof (uint64_t);
6260 				}
6261 
6262 				break;
6263 			}
6264 
6265 			regs[rd] = dtrace_dif_variable(mstate, state, id, 0);
6266 			break;
6267 
6268 		case DIF_OP_STGS:
6269 			id = DIF_INSTR_VAR(instr);
6270 
6271 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
6272 			id -= DIF_VAR_OTHER_UBASE;
6273 
6274 			VERIFY(id < vstate->dtvs_nglobals);
6275 			svar = vstate->dtvs_globals[id];
6276 			ASSERT(svar != NULL);
6277 			v = &svar->dtsv_var;
6278 
6279 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6280 				uintptr_t a = (uintptr_t)svar->dtsv_data;
6281 
6282 				ASSERT(a != 0);
6283 				ASSERT(svar->dtsv_size != 0);
6284 
6285 				if (regs[rd] == 0) {
6286 					*(uint8_t *)a = UINT8_MAX;
6287 					break;
6288 				} else {
6289 					*(uint8_t *)a = 0;
6290 					a += sizeof (uint64_t);
6291 				}
6292 				if (!dtrace_vcanload(
6293 				    (void *)(uintptr_t)regs[rd], &v->dtdv_type,
6294 				    mstate, vstate))
6295 					break;
6296 
6297 				dtrace_vcopy((void *)(uintptr_t)regs[rd],
6298 				    (void *)a, &v->dtdv_type);
6299 				break;
6300 			}
6301 
6302 			svar->dtsv_data = regs[rd];
6303 			break;
6304 
6305 		case DIF_OP_LDTA:
6306 			/*
6307 			 * There are no DTrace built-in thread-local arrays at
6308 			 * present.  This opcode is saved for future work.
6309 			 */
6310 			*flags |= CPU_DTRACE_ILLOP;
6311 			regs[rd] = 0;
6312 			break;
6313 
6314 		case DIF_OP_LDLS:
6315 			id = DIF_INSTR_VAR(instr);
6316 
6317 			if (id < DIF_VAR_OTHER_UBASE) {
6318 				/*
6319 				 * For now, this has no meaning.
6320 				 */
6321 				regs[rd] = 0;
6322 				break;
6323 			}
6324 
6325 			id -= DIF_VAR_OTHER_UBASE;
6326 
6327 			ASSERT(id < vstate->dtvs_nlocals);
6328 			ASSERT(vstate->dtvs_locals != NULL);
6329 
6330 			svar = vstate->dtvs_locals[id];
6331 			ASSERT(svar != NULL);
6332 			v = &svar->dtsv_var;
6333 
6334 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6335 				uintptr_t a = (uintptr_t)svar->dtsv_data;
6336 				size_t sz = v->dtdv_type.dtdt_size;
6337 
6338 				sz += sizeof (uint64_t);
6339 				ASSERT(svar->dtsv_size == NCPU * sz);
6340 				a += curcpu * sz;
6341 
6342 				if (*(uint8_t *)a == UINT8_MAX) {
6343 					/*
6344 					 * If the 0th byte is set to UINT8_MAX
6345 					 * then this is to be treated as a
6346 					 * reference to a NULL variable.
6347 					 */
6348 					regs[rd] = 0;
6349 				} else {
6350 					regs[rd] = a + sizeof (uint64_t);
6351 				}
6352 
6353 				break;
6354 			}
6355 
6356 			ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
6357 			tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
6358 			regs[rd] = tmp[curcpu];
6359 			break;
6360 
6361 		case DIF_OP_STLS:
6362 			id = DIF_INSTR_VAR(instr);
6363 
6364 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
6365 			id -= DIF_VAR_OTHER_UBASE;
6366 			VERIFY(id < vstate->dtvs_nlocals);
6367 
6368 			ASSERT(vstate->dtvs_locals != NULL);
6369 			svar = vstate->dtvs_locals[id];
6370 			ASSERT(svar != NULL);
6371 			v = &svar->dtsv_var;
6372 
6373 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6374 				uintptr_t a = (uintptr_t)svar->dtsv_data;
6375 				size_t sz = v->dtdv_type.dtdt_size;
6376 
6377 				sz += sizeof (uint64_t);
6378 				ASSERT(svar->dtsv_size == NCPU * sz);
6379 				a += curcpu * sz;
6380 
6381 				if (regs[rd] == 0) {
6382 					*(uint8_t *)a = UINT8_MAX;
6383 					break;
6384 				} else {
6385 					*(uint8_t *)a = 0;
6386 					a += sizeof (uint64_t);
6387 				}
6388 
6389 				if (!dtrace_vcanload(
6390 				    (void *)(uintptr_t)regs[rd], &v->dtdv_type,
6391 				    mstate, vstate))
6392 					break;
6393 
6394 				dtrace_vcopy((void *)(uintptr_t)regs[rd],
6395 				    (void *)a, &v->dtdv_type);
6396 				break;
6397 			}
6398 
6399 			ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
6400 			tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
6401 			tmp[curcpu] = regs[rd];
6402 			break;
6403 
6404 		case DIF_OP_LDTS: {
6405 			dtrace_dynvar_t *dvar;
6406 			dtrace_key_t *key;
6407 
6408 			id = DIF_INSTR_VAR(instr);
6409 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
6410 			id -= DIF_VAR_OTHER_UBASE;
6411 			v = &vstate->dtvs_tlocals[id];
6412 
6413 			key = &tupregs[DIF_DTR_NREGS];
6414 			key[0].dttk_value = (uint64_t)id;
6415 			key[0].dttk_size = 0;
6416 			DTRACE_TLS_THRKEY(key[1].dttk_value);
6417 			key[1].dttk_size = 0;
6418 
6419 			dvar = dtrace_dynvar(dstate, 2, key,
6420 			    sizeof (uint64_t), DTRACE_DYNVAR_NOALLOC,
6421 			    mstate, vstate);
6422 
6423 			if (dvar == NULL) {
6424 				regs[rd] = 0;
6425 				break;
6426 			}
6427 
6428 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6429 				regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
6430 			} else {
6431 				regs[rd] = *((uint64_t *)dvar->dtdv_data);
6432 			}
6433 
6434 			break;
6435 		}
6436 
6437 		case DIF_OP_STTS: {
6438 			dtrace_dynvar_t *dvar;
6439 			dtrace_key_t *key;
6440 
6441 			id = DIF_INSTR_VAR(instr);
6442 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
6443 			id -= DIF_VAR_OTHER_UBASE;
6444 			VERIFY(id < vstate->dtvs_ntlocals);
6445 
6446 			key = &tupregs[DIF_DTR_NREGS];
6447 			key[0].dttk_value = (uint64_t)id;
6448 			key[0].dttk_size = 0;
6449 			DTRACE_TLS_THRKEY(key[1].dttk_value);
6450 			key[1].dttk_size = 0;
6451 			v = &vstate->dtvs_tlocals[id];
6452 
6453 			dvar = dtrace_dynvar(dstate, 2, key,
6454 			    v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
6455 			    v->dtdv_type.dtdt_size : sizeof (uint64_t),
6456 			    regs[rd] ? DTRACE_DYNVAR_ALLOC :
6457 			    DTRACE_DYNVAR_DEALLOC, mstate, vstate);
6458 
6459 			/*
6460 			 * Given that we're storing to thread-local data,
6461 			 * we need to flush our predicate cache.
6462 			 */
6463 			curthread->t_predcache = 0;
6464 
6465 			if (dvar == NULL)
6466 				break;
6467 
6468 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6469 				if (!dtrace_vcanload(
6470 				    (void *)(uintptr_t)regs[rd],
6471 				    &v->dtdv_type, mstate, vstate))
6472 					break;
6473 
6474 				dtrace_vcopy((void *)(uintptr_t)regs[rd],
6475 				    dvar->dtdv_data, &v->dtdv_type);
6476 			} else {
6477 				*((uint64_t *)dvar->dtdv_data) = regs[rd];
6478 			}
6479 
6480 			break;
6481 		}
6482 
6483 		case DIF_OP_SRA:
6484 			regs[rd] = (int64_t)regs[r1] >> regs[r2];
6485 			break;
6486 
6487 		case DIF_OP_CALL:
6488 			dtrace_dif_subr(DIF_INSTR_SUBR(instr), rd,
6489 			    regs, tupregs, ttop, mstate, state);
6490 			break;
6491 
6492 		case DIF_OP_PUSHTR:
6493 			if (ttop == DIF_DTR_NREGS) {
6494 				*flags |= CPU_DTRACE_TUPOFLOW;
6495 				break;
6496 			}
6497 
6498 			if (r1 == DIF_TYPE_STRING) {
6499 				/*
6500 				 * If this is a string type and the size is 0,
6501 				 * we'll use the system-wide default string
6502 				 * size.  Note that we are _not_ looking at
6503 				 * the value of the DTRACEOPT_STRSIZE option;
6504 				 * had this been set, we would expect to have
6505 				 * a non-zero size value in the "pushtr".
6506 				 */
6507 				tupregs[ttop].dttk_size =
6508 				    dtrace_strlen((char *)(uintptr_t)regs[rd],
6509 				    regs[r2] ? regs[r2] :
6510 				    dtrace_strsize_default) + 1;
6511 			} else {
6512 				if (regs[r2] > LONG_MAX) {
6513 					*flags |= CPU_DTRACE_ILLOP;
6514 					break;
6515 				}
6516 
6517 				tupregs[ttop].dttk_size = regs[r2];
6518 			}
6519 
6520 			tupregs[ttop++].dttk_value = regs[rd];
6521 			break;
6522 
6523 		case DIF_OP_PUSHTV:
6524 			if (ttop == DIF_DTR_NREGS) {
6525 				*flags |= CPU_DTRACE_TUPOFLOW;
6526 				break;
6527 			}
6528 
6529 			tupregs[ttop].dttk_value = regs[rd];
6530 			tupregs[ttop++].dttk_size = 0;
6531 			break;
6532 
6533 		case DIF_OP_POPTS:
6534 			if (ttop != 0)
6535 				ttop--;
6536 			break;
6537 
6538 		case DIF_OP_FLUSHTS:
6539 			ttop = 0;
6540 			break;
6541 
6542 		case DIF_OP_LDGAA:
6543 		case DIF_OP_LDTAA: {
6544 			dtrace_dynvar_t *dvar;
6545 			dtrace_key_t *key = tupregs;
6546 			uint_t nkeys = ttop;
6547 
6548 			id = DIF_INSTR_VAR(instr);
6549 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
6550 			id -= DIF_VAR_OTHER_UBASE;
6551 
6552 			key[nkeys].dttk_value = (uint64_t)id;
6553 			key[nkeys++].dttk_size = 0;
6554 
6555 			if (DIF_INSTR_OP(instr) == DIF_OP_LDTAA) {
6556 				DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
6557 				key[nkeys++].dttk_size = 0;
6558 				VERIFY(id < vstate->dtvs_ntlocals);
6559 				v = &vstate->dtvs_tlocals[id];
6560 			} else {
6561 				VERIFY(id < vstate->dtvs_nglobals);
6562 				v = &vstate->dtvs_globals[id]->dtsv_var;
6563 			}
6564 
6565 			dvar = dtrace_dynvar(dstate, nkeys, key,
6566 			    v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
6567 			    v->dtdv_type.dtdt_size : sizeof (uint64_t),
6568 			    DTRACE_DYNVAR_NOALLOC, mstate, vstate);
6569 
6570 			if (dvar == NULL) {
6571 				regs[rd] = 0;
6572 				break;
6573 			}
6574 
6575 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6576 				regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
6577 			} else {
6578 				regs[rd] = *((uint64_t *)dvar->dtdv_data);
6579 			}
6580 
6581 			break;
6582 		}
6583 
6584 		case DIF_OP_STGAA:
6585 		case DIF_OP_STTAA: {
6586 			dtrace_dynvar_t *dvar;
6587 			dtrace_key_t *key = tupregs;
6588 			uint_t nkeys = ttop;
6589 
6590 			id = DIF_INSTR_VAR(instr);
6591 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
6592 			id -= DIF_VAR_OTHER_UBASE;
6593 
6594 			key[nkeys].dttk_value = (uint64_t)id;
6595 			key[nkeys++].dttk_size = 0;
6596 
6597 			if (DIF_INSTR_OP(instr) == DIF_OP_STTAA) {
6598 				DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
6599 				key[nkeys++].dttk_size = 0;
6600 				VERIFY(id < vstate->dtvs_ntlocals);
6601 				v = &vstate->dtvs_tlocals[id];
6602 			} else {
6603 				VERIFY(id < vstate->dtvs_nglobals);
6604 				v = &vstate->dtvs_globals[id]->dtsv_var;
6605 			}
6606 
6607 			dvar = dtrace_dynvar(dstate, nkeys, key,
6608 			    v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
6609 			    v->dtdv_type.dtdt_size : sizeof (uint64_t),
6610 			    regs[rd] ? DTRACE_DYNVAR_ALLOC :
6611 			    DTRACE_DYNVAR_DEALLOC, mstate, vstate);
6612 
6613 			if (dvar == NULL)
6614 				break;
6615 
6616 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6617 				if (!dtrace_vcanload(
6618 				    (void *)(uintptr_t)regs[rd], &v->dtdv_type,
6619 				    mstate, vstate))
6620 					break;
6621 
6622 				dtrace_vcopy((void *)(uintptr_t)regs[rd],
6623 				    dvar->dtdv_data, &v->dtdv_type);
6624 			} else {
6625 				*((uint64_t *)dvar->dtdv_data) = regs[rd];
6626 			}
6627 
6628 			break;
6629 		}
6630 
6631 		case DIF_OP_ALLOCS: {
6632 			uintptr_t ptr = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
6633 			size_t size = ptr - mstate->dtms_scratch_ptr + regs[r1];
6634 
6635 			/*
6636 			 * Rounding up the user allocation size could have
6637 			 * overflowed large, bogus allocations (like -1ULL) to
6638 			 * 0.
6639 			 */
6640 			if (size < regs[r1] ||
6641 			    !DTRACE_INSCRATCH(mstate, size)) {
6642 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
6643 				regs[rd] = 0;
6644 				break;
6645 			}
6646 
6647 			dtrace_bzero((void *) mstate->dtms_scratch_ptr, size);
6648 			mstate->dtms_scratch_ptr += size;
6649 			regs[rd] = ptr;
6650 			break;
6651 		}
6652 
6653 		case DIF_OP_COPYS:
6654 			if (!dtrace_canstore(regs[rd], regs[r2],
6655 			    mstate, vstate)) {
6656 				*flags |= CPU_DTRACE_BADADDR;
6657 				*illval = regs[rd];
6658 				break;
6659 			}
6660 
6661 			if (!dtrace_canload(regs[r1], regs[r2], mstate, vstate))
6662 				break;
6663 
6664 			dtrace_bcopy((void *)(uintptr_t)regs[r1],
6665 			    (void *)(uintptr_t)regs[rd], (size_t)regs[r2]);
6666 			break;
6667 
6668 		case DIF_OP_STB:
6669 			if (!dtrace_canstore(regs[rd], 1, mstate, vstate)) {
6670 				*flags |= CPU_DTRACE_BADADDR;
6671 				*illval = regs[rd];
6672 				break;
6673 			}
6674 			*((uint8_t *)(uintptr_t)regs[rd]) = (uint8_t)regs[r1];
6675 			break;
6676 
6677 		case DIF_OP_STH:
6678 			if (!dtrace_canstore(regs[rd], 2, mstate, vstate)) {
6679 				*flags |= CPU_DTRACE_BADADDR;
6680 				*illval = regs[rd];
6681 				break;
6682 			}
6683 			if (regs[rd] & 1) {
6684 				*flags |= CPU_DTRACE_BADALIGN;
6685 				*illval = regs[rd];
6686 				break;
6687 			}
6688 			*((uint16_t *)(uintptr_t)regs[rd]) = (uint16_t)regs[r1];
6689 			break;
6690 
6691 		case DIF_OP_STW:
6692 			if (!dtrace_canstore(regs[rd], 4, mstate, vstate)) {
6693 				*flags |= CPU_DTRACE_BADADDR;
6694 				*illval = regs[rd];
6695 				break;
6696 			}
6697 			if (regs[rd] & 3) {
6698 				*flags |= CPU_DTRACE_BADALIGN;
6699 				*illval = regs[rd];
6700 				break;
6701 			}
6702 			*((uint32_t *)(uintptr_t)regs[rd]) = (uint32_t)regs[r1];
6703 			break;
6704 
6705 		case DIF_OP_STX:
6706 			if (!dtrace_canstore(regs[rd], 8, mstate, vstate)) {
6707 				*flags |= CPU_DTRACE_BADADDR;
6708 				*illval = regs[rd];
6709 				break;
6710 			}
6711 			if (regs[rd] & 7) {
6712 				*flags |= CPU_DTRACE_BADALIGN;
6713 				*illval = regs[rd];
6714 				break;
6715 			}
6716 			*((uint64_t *)(uintptr_t)regs[rd]) = regs[r1];
6717 			break;
6718 		}
6719 	}
6720 
6721 	if (!(*flags & CPU_DTRACE_FAULT))
6722 		return (rval);
6723 
6724 	mstate->dtms_fltoffs = opc * sizeof (dif_instr_t);
6725 	mstate->dtms_present |= DTRACE_MSTATE_FLTOFFS;
6726 
6727 	return (0);
6728 }
6729 
6730 static void
6731 dtrace_action_breakpoint(dtrace_ecb_t *ecb)
6732 {
6733 	dtrace_probe_t *probe = ecb->dte_probe;
6734 	dtrace_provider_t *prov = probe->dtpr_provider;
6735 	char c[DTRACE_FULLNAMELEN + 80], *str;
6736 	char *msg = "dtrace: breakpoint action at probe ";
6737 	char *ecbmsg = " (ecb ";
6738 	uintptr_t mask = (0xf << (sizeof (uintptr_t) * NBBY / 4));
6739 	uintptr_t val = (uintptr_t)ecb;
6740 	int shift = (sizeof (uintptr_t) * NBBY) - 4, i = 0;
6741 
6742 	if (dtrace_destructive_disallow)
6743 		return;
6744 
6745 	/*
6746 	 * It's impossible to be taking action on the NULL probe.
6747 	 */
6748 	ASSERT(probe != NULL);
6749 
6750 	/*
6751 	 * This is a poor man's (destitute man's?) sprintf():  we want to
6752 	 * print the provider name, module name, function name and name of
6753 	 * the probe, along with the hex address of the ECB with the breakpoint
6754 	 * action -- all of which we must place in the character buffer by
6755 	 * hand.
6756 	 */
6757 	while (*msg != '\0')
6758 		c[i++] = *msg++;
6759 
6760 	for (str = prov->dtpv_name; *str != '\0'; str++)
6761 		c[i++] = *str;
6762 	c[i++] = ':';
6763 
6764 	for (str = probe->dtpr_mod; *str != '\0'; str++)
6765 		c[i++] = *str;
6766 	c[i++] = ':';
6767 
6768 	for (str = probe->dtpr_func; *str != '\0'; str++)
6769 		c[i++] = *str;
6770 	c[i++] = ':';
6771 
6772 	for (str = probe->dtpr_name; *str != '\0'; str++)
6773 		c[i++] = *str;
6774 
6775 	while (*ecbmsg != '\0')
6776 		c[i++] = *ecbmsg++;
6777 
6778 	while (shift >= 0) {
6779 		mask = (uintptr_t)0xf << shift;
6780 
6781 		if (val >= ((uintptr_t)1 << shift))
6782 			c[i++] = "0123456789abcdef"[(val & mask) >> shift];
6783 		shift -= 4;
6784 	}
6785 
6786 	c[i++] = ')';
6787 	c[i] = '\0';
6788 
6789 #ifdef illumos
6790 	debug_enter(c);
6791 #else
6792 	kdb_enter(KDB_WHY_DTRACE, "breakpoint action");
6793 #endif
6794 }
6795 
6796 static void
6797 dtrace_action_panic(dtrace_ecb_t *ecb)
6798 {
6799 	dtrace_probe_t *probe = ecb->dte_probe;
6800 
6801 	/*
6802 	 * It's impossible to be taking action on the NULL probe.
6803 	 */
6804 	ASSERT(probe != NULL);
6805 
6806 	if (dtrace_destructive_disallow)
6807 		return;
6808 
6809 	if (dtrace_panicked != NULL)
6810 		return;
6811 
6812 	if (dtrace_casptr(&dtrace_panicked, NULL, curthread) != NULL)
6813 		return;
6814 
6815 	/*
6816 	 * We won the right to panic.  (We want to be sure that only one
6817 	 * thread calls panic() from dtrace_probe(), and that panic() is
6818 	 * called exactly once.)
6819 	 */
6820 	dtrace_panic("dtrace: panic action at probe %s:%s:%s:%s (ecb %p)",
6821 	    probe->dtpr_provider->dtpv_name, probe->dtpr_mod,
6822 	    probe->dtpr_func, probe->dtpr_name, (void *)ecb);
6823 }
6824 
6825 static void
6826 dtrace_action_raise(uint64_t sig)
6827 {
6828 	if (dtrace_destructive_disallow)
6829 		return;
6830 
6831 	if (sig >= NSIG) {
6832 		DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
6833 		return;
6834 	}
6835 
6836 #ifdef illumos
6837 	/*
6838 	 * raise() has a queue depth of 1 -- we ignore all subsequent
6839 	 * invocations of the raise() action.
6840 	 */
6841 	if (curthread->t_dtrace_sig == 0)
6842 		curthread->t_dtrace_sig = (uint8_t)sig;
6843 
6844 	curthread->t_sig_check = 1;
6845 	aston(curthread);
6846 #else
6847 	struct proc *p = curproc;
6848 	PROC_LOCK(p);
6849 	kern_psignal(p, sig);
6850 	PROC_UNLOCK(p);
6851 #endif
6852 }
6853 
6854 static void
6855 dtrace_action_stop(void)
6856 {
6857 	if (dtrace_destructive_disallow)
6858 		return;
6859 
6860 #ifdef illumos
6861 	if (!curthread->t_dtrace_stop) {
6862 		curthread->t_dtrace_stop = 1;
6863 		curthread->t_sig_check = 1;
6864 		aston(curthread);
6865 	}
6866 #else
6867 	struct proc *p = curproc;
6868 	PROC_LOCK(p);
6869 	kern_psignal(p, SIGSTOP);
6870 	PROC_UNLOCK(p);
6871 #endif
6872 }
6873 
6874 static void
6875 dtrace_action_chill(dtrace_mstate_t *mstate, hrtime_t val)
6876 {
6877 	hrtime_t now;
6878 	volatile uint16_t *flags;
6879 #ifdef illumos
6880 	cpu_t *cpu = CPU;
6881 #else
6882 	cpu_t *cpu = &solaris_cpu[curcpu];
6883 #endif
6884 
6885 	if (dtrace_destructive_disallow)
6886 		return;
6887 
6888 	flags = (volatile uint16_t *)&cpu_core[curcpu].cpuc_dtrace_flags;
6889 
6890 	now = dtrace_gethrtime();
6891 
6892 	if (now - cpu->cpu_dtrace_chillmark > dtrace_chill_interval) {
6893 		/*
6894 		 * We need to advance the mark to the current time.
6895 		 */
6896 		cpu->cpu_dtrace_chillmark = now;
6897 		cpu->cpu_dtrace_chilled = 0;
6898 	}
6899 
6900 	/*
6901 	 * Now check to see if the requested chill time would take us over
6902 	 * the maximum amount of time allowed in the chill interval.  (Or
6903 	 * worse, if the calculation itself induces overflow.)
6904 	 */
6905 	if (cpu->cpu_dtrace_chilled + val > dtrace_chill_max ||
6906 	    cpu->cpu_dtrace_chilled + val < cpu->cpu_dtrace_chilled) {
6907 		*flags |= CPU_DTRACE_ILLOP;
6908 		return;
6909 	}
6910 
6911 	while (dtrace_gethrtime() - now < val)
6912 		continue;
6913 
6914 	/*
6915 	 * Normally, we assure that the value of the variable "timestamp" does
6916 	 * not change within an ECB.  The presence of chill() represents an
6917 	 * exception to this rule, however.
6918 	 */
6919 	mstate->dtms_present &= ~DTRACE_MSTATE_TIMESTAMP;
6920 	cpu->cpu_dtrace_chilled += val;
6921 }
6922 
6923 static void
6924 dtrace_action_ustack(dtrace_mstate_t *mstate, dtrace_state_t *state,
6925     uint64_t *buf, uint64_t arg)
6926 {
6927 	int nframes = DTRACE_USTACK_NFRAMES(arg);
6928 	int strsize = DTRACE_USTACK_STRSIZE(arg);
6929 	uint64_t *pcs = &buf[1], *fps;
6930 	char *str = (char *)&pcs[nframes];
6931 	int size, offs = 0, i, j;
6932 	uintptr_t old = mstate->dtms_scratch_ptr, saved;
6933 	uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
6934 	char *sym;
6935 
6936 	/*
6937 	 * Should be taking a faster path if string space has not been
6938 	 * allocated.
6939 	 */
6940 	ASSERT(strsize != 0);
6941 
6942 	/*
6943 	 * We will first allocate some temporary space for the frame pointers.
6944 	 */
6945 	fps = (uint64_t *)P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
6946 	size = (uintptr_t)fps - mstate->dtms_scratch_ptr +
6947 	    (nframes * sizeof (uint64_t));
6948 
6949 	if (!DTRACE_INSCRATCH(mstate, size)) {
6950 		/*
6951 		 * Not enough room for our frame pointers -- need to indicate
6952 		 * that we ran out of scratch space.
6953 		 */
6954 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
6955 		return;
6956 	}
6957 
6958 	mstate->dtms_scratch_ptr += size;
6959 	saved = mstate->dtms_scratch_ptr;
6960 
6961 	/*
6962 	 * Now get a stack with both program counters and frame pointers.
6963 	 */
6964 	DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6965 	dtrace_getufpstack(buf, fps, nframes + 1);
6966 	DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6967 
6968 	/*
6969 	 * If that faulted, we're cooked.
6970 	 */
6971 	if (*flags & CPU_DTRACE_FAULT)
6972 		goto out;
6973 
6974 	/*
6975 	 * Now we want to walk up the stack, calling the USTACK helper.  For
6976 	 * each iteration, we restore the scratch pointer.
6977 	 */
6978 	for (i = 0; i < nframes; i++) {
6979 		mstate->dtms_scratch_ptr = saved;
6980 
6981 		if (offs >= strsize)
6982 			break;
6983 
6984 		sym = (char *)(uintptr_t)dtrace_helper(
6985 		    DTRACE_HELPER_ACTION_USTACK,
6986 		    mstate, state, pcs[i], fps[i]);
6987 
6988 		/*
6989 		 * If we faulted while running the helper, we're going to
6990 		 * clear the fault and null out the corresponding string.
6991 		 */
6992 		if (*flags & CPU_DTRACE_FAULT) {
6993 			*flags &= ~CPU_DTRACE_FAULT;
6994 			str[offs++] = '\0';
6995 			continue;
6996 		}
6997 
6998 		if (sym == NULL) {
6999 			str[offs++] = '\0';
7000 			continue;
7001 		}
7002 
7003 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
7004 
7005 		/*
7006 		 * Now copy in the string that the helper returned to us.
7007 		 */
7008 		for (j = 0; offs + j < strsize; j++) {
7009 			if ((str[offs + j] = sym[j]) == '\0')
7010 				break;
7011 		}
7012 
7013 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
7014 
7015 		offs += j + 1;
7016 	}
7017 
7018 	if (offs >= strsize) {
7019 		/*
7020 		 * If we didn't have room for all of the strings, we don't
7021 		 * abort processing -- this needn't be a fatal error -- but we
7022 		 * still want to increment a counter (dts_stkstroverflows) to
7023 		 * allow this condition to be warned about.  (If this is from
7024 		 * a jstack() action, it is easily tuned via jstackstrsize.)
7025 		 */
7026 		dtrace_error(&state->dts_stkstroverflows);
7027 	}
7028 
7029 	while (offs < strsize)
7030 		str[offs++] = '\0';
7031 
7032 out:
7033 	mstate->dtms_scratch_ptr = old;
7034 }
7035 
7036 static void
7037 dtrace_store_by_ref(dtrace_difo_t *dp, caddr_t tomax, size_t size,
7038     size_t *valoffsp, uint64_t *valp, uint64_t end, int intuple, int dtkind)
7039 {
7040 	volatile uint16_t *flags;
7041 	uint64_t val = *valp;
7042 	size_t valoffs = *valoffsp;
7043 
7044 	flags = (volatile uint16_t *)&cpu_core[curcpu].cpuc_dtrace_flags;
7045 	ASSERT(dtkind == DIF_TF_BYREF || dtkind == DIF_TF_BYUREF);
7046 
7047 	/*
7048 	 * If this is a string, we're going to only load until we find the zero
7049 	 * byte -- after which we'll store zero bytes.
7050 	 */
7051 	if (dp->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING) {
7052 		char c = '\0' + 1;
7053 		size_t s;
7054 
7055 		for (s = 0; s < size; s++) {
7056 			if (c != '\0' && dtkind == DIF_TF_BYREF) {
7057 				c = dtrace_load8(val++);
7058 			} else if (c != '\0' && dtkind == DIF_TF_BYUREF) {
7059 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
7060 				c = dtrace_fuword8((void *)(uintptr_t)val++);
7061 				DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
7062 				if (*flags & CPU_DTRACE_FAULT)
7063 					break;
7064 			}
7065 
7066 			DTRACE_STORE(uint8_t, tomax, valoffs++, c);
7067 
7068 			if (c == '\0' && intuple)
7069 				break;
7070 		}
7071 	} else {
7072 		uint8_t c;
7073 		while (valoffs < end) {
7074 			if (dtkind == DIF_TF_BYREF) {
7075 				c = dtrace_load8(val++);
7076 			} else if (dtkind == DIF_TF_BYUREF) {
7077 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
7078 				c = dtrace_fuword8((void *)(uintptr_t)val++);
7079 				DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
7080 				if (*flags & CPU_DTRACE_FAULT)
7081 					break;
7082 			}
7083 
7084 			DTRACE_STORE(uint8_t, tomax,
7085 			    valoffs++, c);
7086 		}
7087 	}
7088 
7089 	*valp = val;
7090 	*valoffsp = valoffs;
7091 }
7092 
7093 /*
7094  * If you're looking for the epicenter of DTrace, you just found it.  This
7095  * is the function called by the provider to fire a probe -- from which all
7096  * subsequent probe-context DTrace activity emanates.
7097  */
7098 void
7099 dtrace_probe(dtrace_id_t id, uintptr_t arg0, uintptr_t arg1,
7100     uintptr_t arg2, uintptr_t arg3, uintptr_t arg4)
7101 {
7102 	processorid_t cpuid;
7103 	dtrace_icookie_t cookie;
7104 	dtrace_probe_t *probe;
7105 	dtrace_mstate_t mstate;
7106 	dtrace_ecb_t *ecb;
7107 	dtrace_action_t *act;
7108 	intptr_t offs;
7109 	size_t size;
7110 	int vtime, onintr;
7111 	volatile uint16_t *flags;
7112 	hrtime_t now;
7113 
7114 	if (panicstr != NULL)
7115 		return;
7116 
7117 #ifdef illumos
7118 	/*
7119 	 * Kick out immediately if this CPU is still being born (in which case
7120 	 * curthread will be set to -1) or the current thread can't allow
7121 	 * probes in its current context.
7122 	 */
7123 	if (((uintptr_t)curthread & 1) || (curthread->t_flag & T_DONTDTRACE))
7124 		return;
7125 #endif
7126 
7127 	cookie = dtrace_interrupt_disable();
7128 	probe = dtrace_probes[id - 1];
7129 	cpuid = curcpu;
7130 	onintr = CPU_ON_INTR(CPU);
7131 
7132 	if (!onintr && probe->dtpr_predcache != DTRACE_CACHEIDNONE &&
7133 	    probe->dtpr_predcache == curthread->t_predcache) {
7134 		/*
7135 		 * We have hit in the predicate cache; we know that
7136 		 * this predicate would evaluate to be false.
7137 		 */
7138 		dtrace_interrupt_enable(cookie);
7139 		return;
7140 	}
7141 
7142 #ifdef illumos
7143 	if (panic_quiesce) {
7144 #else
7145 	if (panicstr != NULL) {
7146 #endif
7147 		/*
7148 		 * We don't trace anything if we're panicking.
7149 		 */
7150 		dtrace_interrupt_enable(cookie);
7151 		return;
7152 	}
7153 
7154 	now = mstate.dtms_timestamp = dtrace_gethrtime();
7155 	mstate.dtms_present |= DTRACE_MSTATE_TIMESTAMP;
7156 	vtime = dtrace_vtime_references != 0;
7157 
7158 	if (vtime && curthread->t_dtrace_start)
7159 		curthread->t_dtrace_vtime += now - curthread->t_dtrace_start;
7160 
7161 	mstate.dtms_difo = NULL;
7162 	mstate.dtms_probe = probe;
7163 	mstate.dtms_strtok = 0;
7164 	mstate.dtms_arg[0] = arg0;
7165 	mstate.dtms_arg[1] = arg1;
7166 	mstate.dtms_arg[2] = arg2;
7167 	mstate.dtms_arg[3] = arg3;
7168 	mstate.dtms_arg[4] = arg4;
7169 
7170 	flags = (volatile uint16_t *)&cpu_core[cpuid].cpuc_dtrace_flags;
7171 
7172 	for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
7173 		dtrace_predicate_t *pred = ecb->dte_predicate;
7174 		dtrace_state_t *state = ecb->dte_state;
7175 		dtrace_buffer_t *buf = &state->dts_buffer[cpuid];
7176 		dtrace_buffer_t *aggbuf = &state->dts_aggbuffer[cpuid];
7177 		dtrace_vstate_t *vstate = &state->dts_vstate;
7178 		dtrace_provider_t *prov = probe->dtpr_provider;
7179 		uint64_t tracememsize = 0;
7180 		int committed = 0;
7181 		caddr_t tomax;
7182 
7183 		/*
7184 		 * A little subtlety with the following (seemingly innocuous)
7185 		 * declaration of the automatic 'val':  by looking at the
7186 		 * code, you might think that it could be declared in the
7187 		 * action processing loop, below.  (That is, it's only used in
7188 		 * the action processing loop.)  However, it must be declared
7189 		 * out of that scope because in the case of DIF expression
7190 		 * arguments to aggregating actions, one iteration of the
7191 		 * action loop will use the last iteration's value.
7192 		 */
7193 		uint64_t val = 0;
7194 
7195 		mstate.dtms_present = DTRACE_MSTATE_ARGS | DTRACE_MSTATE_PROBE;
7196 		mstate.dtms_getf = NULL;
7197 
7198 		*flags &= ~CPU_DTRACE_ERROR;
7199 
7200 		if (prov == dtrace_provider) {
7201 			/*
7202 			 * If dtrace itself is the provider of this probe,
7203 			 * we're only going to continue processing the ECB if
7204 			 * arg0 (the dtrace_state_t) is equal to the ECB's
7205 			 * creating state.  (This prevents disjoint consumers
7206 			 * from seeing one another's metaprobes.)
7207 			 */
7208 			if (arg0 != (uint64_t)(uintptr_t)state)
7209 				continue;
7210 		}
7211 
7212 		if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE) {
7213 			/*
7214 			 * We're not currently active.  If our provider isn't
7215 			 * the dtrace pseudo provider, we're not interested.
7216 			 */
7217 			if (prov != dtrace_provider)
7218 				continue;
7219 
7220 			/*
7221 			 * Now we must further check if we are in the BEGIN
7222 			 * probe.  If we are, we will only continue processing
7223 			 * if we're still in WARMUP -- if one BEGIN enabling
7224 			 * has invoked the exit() action, we don't want to
7225 			 * evaluate subsequent BEGIN enablings.
7226 			 */
7227 			if (probe->dtpr_id == dtrace_probeid_begin &&
7228 			    state->dts_activity != DTRACE_ACTIVITY_WARMUP) {
7229 				ASSERT(state->dts_activity ==
7230 				    DTRACE_ACTIVITY_DRAINING);
7231 				continue;
7232 			}
7233 		}
7234 
7235 		if (ecb->dte_cond) {
7236 			/*
7237 			 * If the dte_cond bits indicate that this
7238 			 * consumer is only allowed to see user-mode firings
7239 			 * of this probe, call the provider's dtps_usermode()
7240 			 * entry point to check that the probe was fired
7241 			 * while in a user context. Skip this ECB if that's
7242 			 * not the case.
7243 			 */
7244 			if ((ecb->dte_cond & DTRACE_COND_USERMODE) &&
7245 			    prov->dtpv_pops.dtps_usermode(prov->dtpv_arg,
7246 			    probe->dtpr_id, probe->dtpr_arg) == 0)
7247 				continue;
7248 
7249 #ifdef illumos
7250 			/*
7251 			 * This is more subtle than it looks. We have to be
7252 			 * absolutely certain that CRED() isn't going to
7253 			 * change out from under us so it's only legit to
7254 			 * examine that structure if we're in constrained
7255 			 * situations. Currently, the only times we'll this
7256 			 * check is if a non-super-user has enabled the
7257 			 * profile or syscall providers -- providers that
7258 			 * allow visibility of all processes. For the
7259 			 * profile case, the check above will ensure that
7260 			 * we're examining a user context.
7261 			 */
7262 			if (ecb->dte_cond & DTRACE_COND_OWNER) {
7263 				cred_t *cr;
7264 				cred_t *s_cr =
7265 				    ecb->dte_state->dts_cred.dcr_cred;
7266 				proc_t *proc;
7267 
7268 				ASSERT(s_cr != NULL);
7269 
7270 				if ((cr = CRED()) == NULL ||
7271 				    s_cr->cr_uid != cr->cr_uid ||
7272 				    s_cr->cr_uid != cr->cr_ruid ||
7273 				    s_cr->cr_uid != cr->cr_suid ||
7274 				    s_cr->cr_gid != cr->cr_gid ||
7275 				    s_cr->cr_gid != cr->cr_rgid ||
7276 				    s_cr->cr_gid != cr->cr_sgid ||
7277 				    (proc = ttoproc(curthread)) == NULL ||
7278 				    (proc->p_flag & SNOCD))
7279 					continue;
7280 			}
7281 
7282 			if (ecb->dte_cond & DTRACE_COND_ZONEOWNER) {
7283 				cred_t *cr;
7284 				cred_t *s_cr =
7285 				    ecb->dte_state->dts_cred.dcr_cred;
7286 
7287 				ASSERT(s_cr != NULL);
7288 
7289 				if ((cr = CRED()) == NULL ||
7290 				    s_cr->cr_zone->zone_id !=
7291 				    cr->cr_zone->zone_id)
7292 					continue;
7293 			}
7294 #endif
7295 		}
7296 
7297 		if (now - state->dts_alive > dtrace_deadman_timeout) {
7298 			/*
7299 			 * We seem to be dead.  Unless we (a) have kernel
7300 			 * destructive permissions (b) have explicitly enabled
7301 			 * destructive actions and (c) destructive actions have
7302 			 * not been disabled, we're going to transition into
7303 			 * the KILLED state, from which no further processing
7304 			 * on this state will be performed.
7305 			 */
7306 			if (!dtrace_priv_kernel_destructive(state) ||
7307 			    !state->dts_cred.dcr_destructive ||
7308 			    dtrace_destructive_disallow) {
7309 				void *activity = &state->dts_activity;
7310 				dtrace_activity_t current;
7311 
7312 				do {
7313 					current = state->dts_activity;
7314 				} while (dtrace_cas32(activity, current,
7315 				    DTRACE_ACTIVITY_KILLED) != current);
7316 
7317 				continue;
7318 			}
7319 		}
7320 
7321 		if ((offs = dtrace_buffer_reserve(buf, ecb->dte_needed,
7322 		    ecb->dte_alignment, state, &mstate)) < 0)
7323 			continue;
7324 
7325 		tomax = buf->dtb_tomax;
7326 		ASSERT(tomax != NULL);
7327 
7328 		if (ecb->dte_size != 0) {
7329 			dtrace_rechdr_t dtrh;
7330 			if (!(mstate.dtms_present & DTRACE_MSTATE_TIMESTAMP)) {
7331 				mstate.dtms_timestamp = dtrace_gethrtime();
7332 				mstate.dtms_present |= DTRACE_MSTATE_TIMESTAMP;
7333 			}
7334 			ASSERT3U(ecb->dte_size, >=, sizeof (dtrace_rechdr_t));
7335 			dtrh.dtrh_epid = ecb->dte_epid;
7336 			DTRACE_RECORD_STORE_TIMESTAMP(&dtrh,
7337 			    mstate.dtms_timestamp);
7338 			*((dtrace_rechdr_t *)(tomax + offs)) = dtrh;
7339 		}
7340 
7341 		mstate.dtms_epid = ecb->dte_epid;
7342 		mstate.dtms_present |= DTRACE_MSTATE_EPID;
7343 
7344 		if (state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)
7345 			mstate.dtms_access = DTRACE_ACCESS_KERNEL;
7346 		else
7347 			mstate.dtms_access = 0;
7348 
7349 		if (pred != NULL) {
7350 			dtrace_difo_t *dp = pred->dtp_difo;
7351 			uint64_t rval;
7352 
7353 			rval = dtrace_dif_emulate(dp, &mstate, vstate, state);
7354 
7355 			if (!(*flags & CPU_DTRACE_ERROR) && !rval) {
7356 				dtrace_cacheid_t cid = probe->dtpr_predcache;
7357 
7358 				if (cid != DTRACE_CACHEIDNONE && !onintr) {
7359 					/*
7360 					 * Update the predicate cache...
7361 					 */
7362 					ASSERT(cid == pred->dtp_cacheid);
7363 					curthread->t_predcache = cid;
7364 				}
7365 
7366 				continue;
7367 			}
7368 		}
7369 
7370 		for (act = ecb->dte_action; !(*flags & CPU_DTRACE_ERROR) &&
7371 		    act != NULL; act = act->dta_next) {
7372 			size_t valoffs;
7373 			dtrace_difo_t *dp;
7374 			dtrace_recdesc_t *rec = &act->dta_rec;
7375 
7376 			size = rec->dtrd_size;
7377 			valoffs = offs + rec->dtrd_offset;
7378 
7379 			if (DTRACEACT_ISAGG(act->dta_kind)) {
7380 				uint64_t v = 0xbad;
7381 				dtrace_aggregation_t *agg;
7382 
7383 				agg = (dtrace_aggregation_t *)act;
7384 
7385 				if ((dp = act->dta_difo) != NULL)
7386 					v = dtrace_dif_emulate(dp,
7387 					    &mstate, vstate, state);
7388 
7389 				if (*flags & CPU_DTRACE_ERROR)
7390 					continue;
7391 
7392 				/*
7393 				 * Note that we always pass the expression
7394 				 * value from the previous iteration of the
7395 				 * action loop.  This value will only be used
7396 				 * if there is an expression argument to the
7397 				 * aggregating action, denoted by the
7398 				 * dtag_hasarg field.
7399 				 */
7400 				dtrace_aggregate(agg, buf,
7401 				    offs, aggbuf, v, val);
7402 				continue;
7403 			}
7404 
7405 			switch (act->dta_kind) {
7406 			case DTRACEACT_STOP:
7407 				if (dtrace_priv_proc_destructive(state))
7408 					dtrace_action_stop();
7409 				continue;
7410 
7411 			case DTRACEACT_BREAKPOINT:
7412 				if (dtrace_priv_kernel_destructive(state))
7413 					dtrace_action_breakpoint(ecb);
7414 				continue;
7415 
7416 			case DTRACEACT_PANIC:
7417 				if (dtrace_priv_kernel_destructive(state))
7418 					dtrace_action_panic(ecb);
7419 				continue;
7420 
7421 			case DTRACEACT_STACK:
7422 				if (!dtrace_priv_kernel(state))
7423 					continue;
7424 
7425 				dtrace_getpcstack((pc_t *)(tomax + valoffs),
7426 				    size / sizeof (pc_t), probe->dtpr_aframes,
7427 				    DTRACE_ANCHORED(probe) ? NULL :
7428 				    (uint32_t *)arg0);
7429 				continue;
7430 
7431 			case DTRACEACT_JSTACK:
7432 			case DTRACEACT_USTACK:
7433 				if (!dtrace_priv_proc(state))
7434 					continue;
7435 
7436 				/*
7437 				 * See comment in DIF_VAR_PID.
7438 				 */
7439 				if (DTRACE_ANCHORED(mstate.dtms_probe) &&
7440 				    CPU_ON_INTR(CPU)) {
7441 					int depth = DTRACE_USTACK_NFRAMES(
7442 					    rec->dtrd_arg) + 1;
7443 
7444 					dtrace_bzero((void *)(tomax + valoffs),
7445 					    DTRACE_USTACK_STRSIZE(rec->dtrd_arg)
7446 					    + depth * sizeof (uint64_t));
7447 
7448 					continue;
7449 				}
7450 
7451 				if (DTRACE_USTACK_STRSIZE(rec->dtrd_arg) != 0 &&
7452 				    curproc->p_dtrace_helpers != NULL) {
7453 					/*
7454 					 * This is the slow path -- we have
7455 					 * allocated string space, and we're
7456 					 * getting the stack of a process that
7457 					 * has helpers.  Call into a separate
7458 					 * routine to perform this processing.
7459 					 */
7460 					dtrace_action_ustack(&mstate, state,
7461 					    (uint64_t *)(tomax + valoffs),
7462 					    rec->dtrd_arg);
7463 					continue;
7464 				}
7465 
7466 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
7467 				dtrace_getupcstack((uint64_t *)
7468 				    (tomax + valoffs),
7469 				    DTRACE_USTACK_NFRAMES(rec->dtrd_arg) + 1);
7470 				DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
7471 				continue;
7472 
7473 			default:
7474 				break;
7475 			}
7476 
7477 			dp = act->dta_difo;
7478 			ASSERT(dp != NULL);
7479 
7480 			val = dtrace_dif_emulate(dp, &mstate, vstate, state);
7481 
7482 			if (*flags & CPU_DTRACE_ERROR)
7483 				continue;
7484 
7485 			switch (act->dta_kind) {
7486 			case DTRACEACT_SPECULATE: {
7487 				dtrace_rechdr_t *dtrh;
7488 
7489 				ASSERT(buf == &state->dts_buffer[cpuid]);
7490 				buf = dtrace_speculation_buffer(state,
7491 				    cpuid, val);
7492 
7493 				if (buf == NULL) {
7494 					*flags |= CPU_DTRACE_DROP;
7495 					continue;
7496 				}
7497 
7498 				offs = dtrace_buffer_reserve(buf,
7499 				    ecb->dte_needed, ecb->dte_alignment,
7500 				    state, NULL);
7501 
7502 				if (offs < 0) {
7503 					*flags |= CPU_DTRACE_DROP;
7504 					continue;
7505 				}
7506 
7507 				tomax = buf->dtb_tomax;
7508 				ASSERT(tomax != NULL);
7509 
7510 				if (ecb->dte_size == 0)
7511 					continue;
7512 
7513 				ASSERT3U(ecb->dte_size, >=,
7514 				    sizeof (dtrace_rechdr_t));
7515 				dtrh = ((void *)(tomax + offs));
7516 				dtrh->dtrh_epid = ecb->dte_epid;
7517 				/*
7518 				 * When the speculation is committed, all of
7519 				 * the records in the speculative buffer will
7520 				 * have their timestamps set to the commit
7521 				 * time.  Until then, it is set to a sentinel
7522 				 * value, for debugability.
7523 				 */
7524 				DTRACE_RECORD_STORE_TIMESTAMP(dtrh, UINT64_MAX);
7525 				continue;
7526 			}
7527 
7528 			case DTRACEACT_PRINTM: {
7529 				/* The DIF returns a 'memref'. */
7530 				uintptr_t *memref = (uintptr_t *)(uintptr_t) val;
7531 
7532 				/* Get the size from the memref. */
7533 				size = memref[1];
7534 
7535 				/*
7536 				 * Check if the size exceeds the allocated
7537 				 * buffer size.
7538 				 */
7539 				if (size + sizeof(uintptr_t) > dp->dtdo_rtype.dtdt_size) {
7540 					/* Flag a drop! */
7541 					*flags |= CPU_DTRACE_DROP;
7542 					continue;
7543 				}
7544 
7545 				/* Store the size in the buffer first. */
7546 				DTRACE_STORE(uintptr_t, tomax,
7547 				    valoffs, size);
7548 
7549 				/*
7550 				 * Offset the buffer address to the start
7551 				 * of the data.
7552 				 */
7553 				valoffs += sizeof(uintptr_t);
7554 
7555 				/*
7556 				 * Reset to the memory address rather than
7557 				 * the memref array, then let the BYREF
7558 				 * code below do the work to store the
7559 				 * memory data in the buffer.
7560 				 */
7561 				val = memref[0];
7562 				break;
7563 			}
7564 
7565 			case DTRACEACT_PRINTT: {
7566 				/* The DIF returns a 'typeref'. */
7567 				uintptr_t *typeref = (uintptr_t *)(uintptr_t) val;
7568 				char c = '\0' + 1;
7569 				size_t s;
7570 
7571 				/*
7572 				 * Get the type string length and round it
7573 				 * up so that the data that follows is
7574 				 * aligned for easy access.
7575 				 */
7576 				size_t typs = strlen((char *) typeref[2]) + 1;
7577 				typs = roundup(typs,  sizeof(uintptr_t));
7578 
7579 				/*
7580 				 *Get the size from the typeref using the
7581 				 * number of elements and the type size.
7582 				 */
7583 				size = typeref[1] * typeref[3];
7584 
7585 				/*
7586 				 * Check if the size exceeds the allocated
7587 				 * buffer size.
7588 				 */
7589 				if (size + typs + 2 * sizeof(uintptr_t) > dp->dtdo_rtype.dtdt_size) {
7590 					/* Flag a drop! */
7591 					*flags |= CPU_DTRACE_DROP;
7592 
7593 				}
7594 
7595 				/* Store the size in the buffer first. */
7596 				DTRACE_STORE(uintptr_t, tomax,
7597 				    valoffs, size);
7598 				valoffs += sizeof(uintptr_t);
7599 
7600 				/* Store the type size in the buffer. */
7601 				DTRACE_STORE(uintptr_t, tomax,
7602 				    valoffs, typeref[3]);
7603 				valoffs += sizeof(uintptr_t);
7604 
7605 				val = typeref[2];
7606 
7607 				for (s = 0; s < typs; s++) {
7608 					if (c != '\0')
7609 						c = dtrace_load8(val++);
7610 
7611 					DTRACE_STORE(uint8_t, tomax,
7612 					    valoffs++, c);
7613 				}
7614 
7615 				/*
7616 				 * Reset to the memory address rather than
7617 				 * the typeref array, then let the BYREF
7618 				 * code below do the work to store the
7619 				 * memory data in the buffer.
7620 				 */
7621 				val = typeref[0];
7622 				break;
7623 			}
7624 
7625 			case DTRACEACT_CHILL:
7626 				if (dtrace_priv_kernel_destructive(state))
7627 					dtrace_action_chill(&mstate, val);
7628 				continue;
7629 
7630 			case DTRACEACT_RAISE:
7631 				if (dtrace_priv_proc_destructive(state))
7632 					dtrace_action_raise(val);
7633 				continue;
7634 
7635 			case DTRACEACT_COMMIT:
7636 				ASSERT(!committed);
7637 
7638 				/*
7639 				 * We need to commit our buffer state.
7640 				 */
7641 				if (ecb->dte_size)
7642 					buf->dtb_offset = offs + ecb->dte_size;
7643 				buf = &state->dts_buffer[cpuid];
7644 				dtrace_speculation_commit(state, cpuid, val);
7645 				committed = 1;
7646 				continue;
7647 
7648 			case DTRACEACT_DISCARD:
7649 				dtrace_speculation_discard(state, cpuid, val);
7650 				continue;
7651 
7652 			case DTRACEACT_DIFEXPR:
7653 			case DTRACEACT_LIBACT:
7654 			case DTRACEACT_PRINTF:
7655 			case DTRACEACT_PRINTA:
7656 			case DTRACEACT_SYSTEM:
7657 			case DTRACEACT_FREOPEN:
7658 			case DTRACEACT_TRACEMEM:
7659 				break;
7660 
7661 			case DTRACEACT_TRACEMEM_DYNSIZE:
7662 				tracememsize = val;
7663 				break;
7664 
7665 			case DTRACEACT_SYM:
7666 			case DTRACEACT_MOD:
7667 				if (!dtrace_priv_kernel(state))
7668 					continue;
7669 				break;
7670 
7671 			case DTRACEACT_USYM:
7672 			case DTRACEACT_UMOD:
7673 			case DTRACEACT_UADDR: {
7674 #ifdef illumos
7675 				struct pid *pid = curthread->t_procp->p_pidp;
7676 #endif
7677 
7678 				if (!dtrace_priv_proc(state))
7679 					continue;
7680 
7681 				DTRACE_STORE(uint64_t, tomax,
7682 #ifdef illumos
7683 				    valoffs, (uint64_t)pid->pid_id);
7684 #else
7685 				    valoffs, (uint64_t) curproc->p_pid);
7686 #endif
7687 				DTRACE_STORE(uint64_t, tomax,
7688 				    valoffs + sizeof (uint64_t), val);
7689 
7690 				continue;
7691 			}
7692 
7693 			case DTRACEACT_EXIT: {
7694 				/*
7695 				 * For the exit action, we are going to attempt
7696 				 * to atomically set our activity to be
7697 				 * draining.  If this fails (either because
7698 				 * another CPU has beat us to the exit action,
7699 				 * or because our current activity is something
7700 				 * other than ACTIVE or WARMUP), we will
7701 				 * continue.  This assures that the exit action
7702 				 * can be successfully recorded at most once
7703 				 * when we're in the ACTIVE state.  If we're
7704 				 * encountering the exit() action while in
7705 				 * COOLDOWN, however, we want to honor the new
7706 				 * status code.  (We know that we're the only
7707 				 * thread in COOLDOWN, so there is no race.)
7708 				 */
7709 				void *activity = &state->dts_activity;
7710 				dtrace_activity_t current = state->dts_activity;
7711 
7712 				if (current == DTRACE_ACTIVITY_COOLDOWN)
7713 					break;
7714 
7715 				if (current != DTRACE_ACTIVITY_WARMUP)
7716 					current = DTRACE_ACTIVITY_ACTIVE;
7717 
7718 				if (dtrace_cas32(activity, current,
7719 				    DTRACE_ACTIVITY_DRAINING) != current) {
7720 					*flags |= CPU_DTRACE_DROP;
7721 					continue;
7722 				}
7723 
7724 				break;
7725 			}
7726 
7727 			default:
7728 				ASSERT(0);
7729 			}
7730 
7731 			if (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF ||
7732 			    dp->dtdo_rtype.dtdt_flags & DIF_TF_BYUREF) {
7733 				uintptr_t end = valoffs + size;
7734 
7735 				if (tracememsize != 0 &&
7736 				    valoffs + tracememsize < end) {
7737 					end = valoffs + tracememsize;
7738 					tracememsize = 0;
7739 				}
7740 
7741 				if (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF &&
7742 				    !dtrace_vcanload((void *)(uintptr_t)val,
7743 				    &dp->dtdo_rtype, &mstate, vstate))
7744 					continue;
7745 
7746 				dtrace_store_by_ref(dp, tomax, size, &valoffs,
7747 				    &val, end, act->dta_intuple,
7748 				    dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF ?
7749 				    DIF_TF_BYREF: DIF_TF_BYUREF);
7750 				continue;
7751 			}
7752 
7753 			switch (size) {
7754 			case 0:
7755 				break;
7756 
7757 			case sizeof (uint8_t):
7758 				DTRACE_STORE(uint8_t, tomax, valoffs, val);
7759 				break;
7760 			case sizeof (uint16_t):
7761 				DTRACE_STORE(uint16_t, tomax, valoffs, val);
7762 				break;
7763 			case sizeof (uint32_t):
7764 				DTRACE_STORE(uint32_t, tomax, valoffs, val);
7765 				break;
7766 			case sizeof (uint64_t):
7767 				DTRACE_STORE(uint64_t, tomax, valoffs, val);
7768 				break;
7769 			default:
7770 				/*
7771 				 * Any other size should have been returned by
7772 				 * reference, not by value.
7773 				 */
7774 				ASSERT(0);
7775 				break;
7776 			}
7777 		}
7778 
7779 		if (*flags & CPU_DTRACE_DROP)
7780 			continue;
7781 
7782 		if (*flags & CPU_DTRACE_FAULT) {
7783 			int ndx;
7784 			dtrace_action_t *err;
7785 
7786 			buf->dtb_errors++;
7787 
7788 			if (probe->dtpr_id == dtrace_probeid_error) {
7789 				/*
7790 				 * There's nothing we can do -- we had an
7791 				 * error on the error probe.  We bump an
7792 				 * error counter to at least indicate that
7793 				 * this condition happened.
7794 				 */
7795 				dtrace_error(&state->dts_dblerrors);
7796 				continue;
7797 			}
7798 
7799 			if (vtime) {
7800 				/*
7801 				 * Before recursing on dtrace_probe(), we
7802 				 * need to explicitly clear out our start
7803 				 * time to prevent it from being accumulated
7804 				 * into t_dtrace_vtime.
7805 				 */
7806 				curthread->t_dtrace_start = 0;
7807 			}
7808 
7809 			/*
7810 			 * Iterate over the actions to figure out which action
7811 			 * we were processing when we experienced the error.
7812 			 * Note that act points _past_ the faulting action; if
7813 			 * act is ecb->dte_action, the fault was in the
7814 			 * predicate, if it's ecb->dte_action->dta_next it's
7815 			 * in action #1, and so on.
7816 			 */
7817 			for (err = ecb->dte_action, ndx = 0;
7818 			    err != act; err = err->dta_next, ndx++)
7819 				continue;
7820 
7821 			dtrace_probe_error(state, ecb->dte_epid, ndx,
7822 			    (mstate.dtms_present & DTRACE_MSTATE_FLTOFFS) ?
7823 			    mstate.dtms_fltoffs : -1, DTRACE_FLAGS2FLT(*flags),
7824 			    cpu_core[cpuid].cpuc_dtrace_illval);
7825 
7826 			continue;
7827 		}
7828 
7829 		if (!committed)
7830 			buf->dtb_offset = offs + ecb->dte_size;
7831 	}
7832 
7833 	if (vtime)
7834 		curthread->t_dtrace_start = dtrace_gethrtime();
7835 
7836 	dtrace_interrupt_enable(cookie);
7837 }
7838 
7839 /*
7840  * DTrace Probe Hashing Functions
7841  *
7842  * The functions in this section (and indeed, the functions in remaining
7843  * sections) are not _called_ from probe context.  (Any exceptions to this are
7844  * marked with a "Note:".)  Rather, they are called from elsewhere in the
7845  * DTrace framework to look-up probes in, add probes to and remove probes from
7846  * the DTrace probe hashes.  (Each probe is hashed by each element of the
7847  * probe tuple -- allowing for fast lookups, regardless of what was
7848  * specified.)
7849  */
7850 static uint_t
7851 dtrace_hash_str(const char *p)
7852 {
7853 	unsigned int g;
7854 	uint_t hval = 0;
7855 
7856 	while (*p) {
7857 		hval = (hval << 4) + *p++;
7858 		if ((g = (hval & 0xf0000000)) != 0)
7859 			hval ^= g >> 24;
7860 		hval &= ~g;
7861 	}
7862 	return (hval);
7863 }
7864 
7865 static dtrace_hash_t *
7866 dtrace_hash_create(uintptr_t stroffs, uintptr_t nextoffs, uintptr_t prevoffs)
7867 {
7868 	dtrace_hash_t *hash = kmem_zalloc(sizeof (dtrace_hash_t), KM_SLEEP);
7869 
7870 	hash->dth_stroffs = stroffs;
7871 	hash->dth_nextoffs = nextoffs;
7872 	hash->dth_prevoffs = prevoffs;
7873 
7874 	hash->dth_size = 1;
7875 	hash->dth_mask = hash->dth_size - 1;
7876 
7877 	hash->dth_tab = kmem_zalloc(hash->dth_size *
7878 	    sizeof (dtrace_hashbucket_t *), KM_SLEEP);
7879 
7880 	return (hash);
7881 }
7882 
7883 static void
7884 dtrace_hash_destroy(dtrace_hash_t *hash)
7885 {
7886 #ifdef DEBUG
7887 	int i;
7888 
7889 	for (i = 0; i < hash->dth_size; i++)
7890 		ASSERT(hash->dth_tab[i] == NULL);
7891 #endif
7892 
7893 	kmem_free(hash->dth_tab,
7894 	    hash->dth_size * sizeof (dtrace_hashbucket_t *));
7895 	kmem_free(hash, sizeof (dtrace_hash_t));
7896 }
7897 
7898 static void
7899 dtrace_hash_resize(dtrace_hash_t *hash)
7900 {
7901 	int size = hash->dth_size, i, ndx;
7902 	int new_size = hash->dth_size << 1;
7903 	int new_mask = new_size - 1;
7904 	dtrace_hashbucket_t **new_tab, *bucket, *next;
7905 
7906 	ASSERT((new_size & new_mask) == 0);
7907 
7908 	new_tab = kmem_zalloc(new_size * sizeof (void *), KM_SLEEP);
7909 
7910 	for (i = 0; i < size; i++) {
7911 		for (bucket = hash->dth_tab[i]; bucket != NULL; bucket = next) {
7912 			dtrace_probe_t *probe = bucket->dthb_chain;
7913 
7914 			ASSERT(probe != NULL);
7915 			ndx = DTRACE_HASHSTR(hash, probe) & new_mask;
7916 
7917 			next = bucket->dthb_next;
7918 			bucket->dthb_next = new_tab[ndx];
7919 			new_tab[ndx] = bucket;
7920 		}
7921 	}
7922 
7923 	kmem_free(hash->dth_tab, hash->dth_size * sizeof (void *));
7924 	hash->dth_tab = new_tab;
7925 	hash->dth_size = new_size;
7926 	hash->dth_mask = new_mask;
7927 }
7928 
7929 static void
7930 dtrace_hash_add(dtrace_hash_t *hash, dtrace_probe_t *new)
7931 {
7932 	int hashval = DTRACE_HASHSTR(hash, new);
7933 	int ndx = hashval & hash->dth_mask;
7934 	dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
7935 	dtrace_probe_t **nextp, **prevp;
7936 
7937 	for (; bucket != NULL; bucket = bucket->dthb_next) {
7938 		if (DTRACE_HASHEQ(hash, bucket->dthb_chain, new))
7939 			goto add;
7940 	}
7941 
7942 	if ((hash->dth_nbuckets >> 1) > hash->dth_size) {
7943 		dtrace_hash_resize(hash);
7944 		dtrace_hash_add(hash, new);
7945 		return;
7946 	}
7947 
7948 	bucket = kmem_zalloc(sizeof (dtrace_hashbucket_t), KM_SLEEP);
7949 	bucket->dthb_next = hash->dth_tab[ndx];
7950 	hash->dth_tab[ndx] = bucket;
7951 	hash->dth_nbuckets++;
7952 
7953 add:
7954 	nextp = DTRACE_HASHNEXT(hash, new);
7955 	ASSERT(*nextp == NULL && *(DTRACE_HASHPREV(hash, new)) == NULL);
7956 	*nextp = bucket->dthb_chain;
7957 
7958 	if (bucket->dthb_chain != NULL) {
7959 		prevp = DTRACE_HASHPREV(hash, bucket->dthb_chain);
7960 		ASSERT(*prevp == NULL);
7961 		*prevp = new;
7962 	}
7963 
7964 	bucket->dthb_chain = new;
7965 	bucket->dthb_len++;
7966 }
7967 
7968 static dtrace_probe_t *
7969 dtrace_hash_lookup(dtrace_hash_t *hash, dtrace_probe_t *template)
7970 {
7971 	int hashval = DTRACE_HASHSTR(hash, template);
7972 	int ndx = hashval & hash->dth_mask;
7973 	dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
7974 
7975 	for (; bucket != NULL; bucket = bucket->dthb_next) {
7976 		if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
7977 			return (bucket->dthb_chain);
7978 	}
7979 
7980 	return (NULL);
7981 }
7982 
7983 static int
7984 dtrace_hash_collisions(dtrace_hash_t *hash, dtrace_probe_t *template)
7985 {
7986 	int hashval = DTRACE_HASHSTR(hash, template);
7987 	int ndx = hashval & hash->dth_mask;
7988 	dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
7989 
7990 	for (; bucket != NULL; bucket = bucket->dthb_next) {
7991 		if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
7992 			return (bucket->dthb_len);
7993 	}
7994 
7995 	return (0);
7996 }
7997 
7998 static void
7999 dtrace_hash_remove(dtrace_hash_t *hash, dtrace_probe_t *probe)
8000 {
8001 	int ndx = DTRACE_HASHSTR(hash, probe) & hash->dth_mask;
8002 	dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
8003 
8004 	dtrace_probe_t **prevp = DTRACE_HASHPREV(hash, probe);
8005 	dtrace_probe_t **nextp = DTRACE_HASHNEXT(hash, probe);
8006 
8007 	/*
8008 	 * Find the bucket that we're removing this probe from.
8009 	 */
8010 	for (; bucket != NULL; bucket = bucket->dthb_next) {
8011 		if (DTRACE_HASHEQ(hash, bucket->dthb_chain, probe))
8012 			break;
8013 	}
8014 
8015 	ASSERT(bucket != NULL);
8016 
8017 	if (*prevp == NULL) {
8018 		if (*nextp == NULL) {
8019 			/*
8020 			 * The removed probe was the only probe on this
8021 			 * bucket; we need to remove the bucket.
8022 			 */
8023 			dtrace_hashbucket_t *b = hash->dth_tab[ndx];
8024 
8025 			ASSERT(bucket->dthb_chain == probe);
8026 			ASSERT(b != NULL);
8027 
8028 			if (b == bucket) {
8029 				hash->dth_tab[ndx] = bucket->dthb_next;
8030 			} else {
8031 				while (b->dthb_next != bucket)
8032 					b = b->dthb_next;
8033 				b->dthb_next = bucket->dthb_next;
8034 			}
8035 
8036 			ASSERT(hash->dth_nbuckets > 0);
8037 			hash->dth_nbuckets--;
8038 			kmem_free(bucket, sizeof (dtrace_hashbucket_t));
8039 			return;
8040 		}
8041 
8042 		bucket->dthb_chain = *nextp;
8043 	} else {
8044 		*(DTRACE_HASHNEXT(hash, *prevp)) = *nextp;
8045 	}
8046 
8047 	if (*nextp != NULL)
8048 		*(DTRACE_HASHPREV(hash, *nextp)) = *prevp;
8049 }
8050 
8051 /*
8052  * DTrace Utility Functions
8053  *
8054  * These are random utility functions that are _not_ called from probe context.
8055  */
8056 static int
8057 dtrace_badattr(const dtrace_attribute_t *a)
8058 {
8059 	return (a->dtat_name > DTRACE_STABILITY_MAX ||
8060 	    a->dtat_data > DTRACE_STABILITY_MAX ||
8061 	    a->dtat_class > DTRACE_CLASS_MAX);
8062 }
8063 
8064 /*
8065  * Return a duplicate copy of a string.  If the specified string is NULL,
8066  * this function returns a zero-length string.
8067  */
8068 static char *
8069 dtrace_strdup(const char *str)
8070 {
8071 	char *new = kmem_zalloc((str != NULL ? strlen(str) : 0) + 1, KM_SLEEP);
8072 
8073 	if (str != NULL)
8074 		(void) strcpy(new, str);
8075 
8076 	return (new);
8077 }
8078 
8079 #define	DTRACE_ISALPHA(c)	\
8080 	(((c) >= 'a' && (c) <= 'z') || ((c) >= 'A' && (c) <= 'Z'))
8081 
8082 static int
8083 dtrace_badname(const char *s)
8084 {
8085 	char c;
8086 
8087 	if (s == NULL || (c = *s++) == '\0')
8088 		return (0);
8089 
8090 	if (!DTRACE_ISALPHA(c) && c != '-' && c != '_' && c != '.')
8091 		return (1);
8092 
8093 	while ((c = *s++) != '\0') {
8094 		if (!DTRACE_ISALPHA(c) && (c < '0' || c > '9') &&
8095 		    c != '-' && c != '_' && c != '.' && c != '`')
8096 			return (1);
8097 	}
8098 
8099 	return (0);
8100 }
8101 
8102 static void
8103 dtrace_cred2priv(cred_t *cr, uint32_t *privp, uid_t *uidp, zoneid_t *zoneidp)
8104 {
8105 	uint32_t priv;
8106 
8107 #ifdef illumos
8108 	if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
8109 		/*
8110 		 * For DTRACE_PRIV_ALL, the uid and zoneid don't matter.
8111 		 */
8112 		priv = DTRACE_PRIV_ALL;
8113 	} else {
8114 		*uidp = crgetuid(cr);
8115 		*zoneidp = crgetzoneid(cr);
8116 
8117 		priv = 0;
8118 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE))
8119 			priv |= DTRACE_PRIV_KERNEL | DTRACE_PRIV_USER;
8120 		else if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE))
8121 			priv |= DTRACE_PRIV_USER;
8122 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE))
8123 			priv |= DTRACE_PRIV_PROC;
8124 		if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
8125 			priv |= DTRACE_PRIV_OWNER;
8126 		if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
8127 			priv |= DTRACE_PRIV_ZONEOWNER;
8128 	}
8129 #else
8130 	priv = DTRACE_PRIV_ALL;
8131 #endif
8132 
8133 	*privp = priv;
8134 }
8135 
8136 #ifdef DTRACE_ERRDEBUG
8137 static void
8138 dtrace_errdebug(const char *str)
8139 {
8140 	int hval = dtrace_hash_str(str) % DTRACE_ERRHASHSZ;
8141 	int occupied = 0;
8142 
8143 	mutex_enter(&dtrace_errlock);
8144 	dtrace_errlast = str;
8145 	dtrace_errthread = curthread;
8146 
8147 	while (occupied++ < DTRACE_ERRHASHSZ) {
8148 		if (dtrace_errhash[hval].dter_msg == str) {
8149 			dtrace_errhash[hval].dter_count++;
8150 			goto out;
8151 		}
8152 
8153 		if (dtrace_errhash[hval].dter_msg != NULL) {
8154 			hval = (hval + 1) % DTRACE_ERRHASHSZ;
8155 			continue;
8156 		}
8157 
8158 		dtrace_errhash[hval].dter_msg = str;
8159 		dtrace_errhash[hval].dter_count = 1;
8160 		goto out;
8161 	}
8162 
8163 	panic("dtrace: undersized error hash");
8164 out:
8165 	mutex_exit(&dtrace_errlock);
8166 }
8167 #endif
8168 
8169 /*
8170  * DTrace Matching Functions
8171  *
8172  * These functions are used to match groups of probes, given some elements of
8173  * a probe tuple, or some globbed expressions for elements of a probe tuple.
8174  */
8175 static int
8176 dtrace_match_priv(const dtrace_probe_t *prp, uint32_t priv, uid_t uid,
8177     zoneid_t zoneid)
8178 {
8179 	if (priv != DTRACE_PRIV_ALL) {
8180 		uint32_t ppriv = prp->dtpr_provider->dtpv_priv.dtpp_flags;
8181 		uint32_t match = priv & ppriv;
8182 
8183 		/*
8184 		 * No PRIV_DTRACE_* privileges...
8185 		 */
8186 		if ((priv & (DTRACE_PRIV_PROC | DTRACE_PRIV_USER |
8187 		    DTRACE_PRIV_KERNEL)) == 0)
8188 			return (0);
8189 
8190 		/*
8191 		 * No matching bits, but there were bits to match...
8192 		 */
8193 		if (match == 0 && ppriv != 0)
8194 			return (0);
8195 
8196 		/*
8197 		 * Need to have permissions to the process, but don't...
8198 		 */
8199 		if (((ppriv & ~match) & DTRACE_PRIV_OWNER) != 0 &&
8200 		    uid != prp->dtpr_provider->dtpv_priv.dtpp_uid) {
8201 			return (0);
8202 		}
8203 
8204 		/*
8205 		 * Need to be in the same zone unless we possess the
8206 		 * privilege to examine all zones.
8207 		 */
8208 		if (((ppriv & ~match) & DTRACE_PRIV_ZONEOWNER) != 0 &&
8209 		    zoneid != prp->dtpr_provider->dtpv_priv.dtpp_zoneid) {
8210 			return (0);
8211 		}
8212 	}
8213 
8214 	return (1);
8215 }
8216 
8217 /*
8218  * dtrace_match_probe compares a dtrace_probe_t to a pre-compiled key, which
8219  * consists of input pattern strings and an ops-vector to evaluate them.
8220  * This function returns >0 for match, 0 for no match, and <0 for error.
8221  */
8222 static int
8223 dtrace_match_probe(const dtrace_probe_t *prp, const dtrace_probekey_t *pkp,
8224     uint32_t priv, uid_t uid, zoneid_t zoneid)
8225 {
8226 	dtrace_provider_t *pvp = prp->dtpr_provider;
8227 	int rv;
8228 
8229 	if (pvp->dtpv_defunct)
8230 		return (0);
8231 
8232 	if ((rv = pkp->dtpk_pmatch(pvp->dtpv_name, pkp->dtpk_prov, 0)) <= 0)
8233 		return (rv);
8234 
8235 	if ((rv = pkp->dtpk_mmatch(prp->dtpr_mod, pkp->dtpk_mod, 0)) <= 0)
8236 		return (rv);
8237 
8238 	if ((rv = pkp->dtpk_fmatch(prp->dtpr_func, pkp->dtpk_func, 0)) <= 0)
8239 		return (rv);
8240 
8241 	if ((rv = pkp->dtpk_nmatch(prp->dtpr_name, pkp->dtpk_name, 0)) <= 0)
8242 		return (rv);
8243 
8244 	if (dtrace_match_priv(prp, priv, uid, zoneid) == 0)
8245 		return (0);
8246 
8247 	return (rv);
8248 }
8249 
8250 /*
8251  * dtrace_match_glob() is a safe kernel implementation of the gmatch(3GEN)
8252  * interface for matching a glob pattern 'p' to an input string 's'.  Unlike
8253  * libc's version, the kernel version only applies to 8-bit ASCII strings.
8254  * In addition, all of the recursion cases except for '*' matching have been
8255  * unwound.  For '*', we still implement recursive evaluation, but a depth
8256  * counter is maintained and matching is aborted if we recurse too deep.
8257  * The function returns 0 if no match, >0 if match, and <0 if recursion error.
8258  */
8259 static int
8260 dtrace_match_glob(const char *s, const char *p, int depth)
8261 {
8262 	const char *olds;
8263 	char s1, c;
8264 	int gs;
8265 
8266 	if (depth > DTRACE_PROBEKEY_MAXDEPTH)
8267 		return (-1);
8268 
8269 	if (s == NULL)
8270 		s = ""; /* treat NULL as empty string */
8271 
8272 top:
8273 	olds = s;
8274 	s1 = *s++;
8275 
8276 	if (p == NULL)
8277 		return (0);
8278 
8279 	if ((c = *p++) == '\0')
8280 		return (s1 == '\0');
8281 
8282 	switch (c) {
8283 	case '[': {
8284 		int ok = 0, notflag = 0;
8285 		char lc = '\0';
8286 
8287 		if (s1 == '\0')
8288 			return (0);
8289 
8290 		if (*p == '!') {
8291 			notflag = 1;
8292 			p++;
8293 		}
8294 
8295 		if ((c = *p++) == '\0')
8296 			return (0);
8297 
8298 		do {
8299 			if (c == '-' && lc != '\0' && *p != ']') {
8300 				if ((c = *p++) == '\0')
8301 					return (0);
8302 				if (c == '\\' && (c = *p++) == '\0')
8303 					return (0);
8304 
8305 				if (notflag) {
8306 					if (s1 < lc || s1 > c)
8307 						ok++;
8308 					else
8309 						return (0);
8310 				} else if (lc <= s1 && s1 <= c)
8311 					ok++;
8312 
8313 			} else if (c == '\\' && (c = *p++) == '\0')
8314 				return (0);
8315 
8316 			lc = c; /* save left-hand 'c' for next iteration */
8317 
8318 			if (notflag) {
8319 				if (s1 != c)
8320 					ok++;
8321 				else
8322 					return (0);
8323 			} else if (s1 == c)
8324 				ok++;
8325 
8326 			if ((c = *p++) == '\0')
8327 				return (0);
8328 
8329 		} while (c != ']');
8330 
8331 		if (ok)
8332 			goto top;
8333 
8334 		return (0);
8335 	}
8336 
8337 	case '\\':
8338 		if ((c = *p++) == '\0')
8339 			return (0);
8340 		/*FALLTHRU*/
8341 
8342 	default:
8343 		if (c != s1)
8344 			return (0);
8345 		/*FALLTHRU*/
8346 
8347 	case '?':
8348 		if (s1 != '\0')
8349 			goto top;
8350 		return (0);
8351 
8352 	case '*':
8353 		while (*p == '*')
8354 			p++; /* consecutive *'s are identical to a single one */
8355 
8356 		if (*p == '\0')
8357 			return (1);
8358 
8359 		for (s = olds; *s != '\0'; s++) {
8360 			if ((gs = dtrace_match_glob(s, p, depth + 1)) != 0)
8361 				return (gs);
8362 		}
8363 
8364 		return (0);
8365 	}
8366 }
8367 
8368 /*ARGSUSED*/
8369 static int
8370 dtrace_match_string(const char *s, const char *p, int depth)
8371 {
8372 	return (s != NULL && strcmp(s, p) == 0);
8373 }
8374 
8375 /*ARGSUSED*/
8376 static int
8377 dtrace_match_nul(const char *s, const char *p, int depth)
8378 {
8379 	return (1); /* always match the empty pattern */
8380 }
8381 
8382 /*ARGSUSED*/
8383 static int
8384 dtrace_match_nonzero(const char *s, const char *p, int depth)
8385 {
8386 	return (s != NULL && s[0] != '\0');
8387 }
8388 
8389 static int
8390 dtrace_match(const dtrace_probekey_t *pkp, uint32_t priv, uid_t uid,
8391     zoneid_t zoneid, int (*matched)(dtrace_probe_t *, void *), void *arg)
8392 {
8393 	dtrace_probe_t template, *probe;
8394 	dtrace_hash_t *hash = NULL;
8395 	int len, best = INT_MAX, nmatched = 0;
8396 	dtrace_id_t i;
8397 
8398 	ASSERT(MUTEX_HELD(&dtrace_lock));
8399 
8400 	/*
8401 	 * If the probe ID is specified in the key, just lookup by ID and
8402 	 * invoke the match callback once if a matching probe is found.
8403 	 */
8404 	if (pkp->dtpk_id != DTRACE_IDNONE) {
8405 		if ((probe = dtrace_probe_lookup_id(pkp->dtpk_id)) != NULL &&
8406 		    dtrace_match_probe(probe, pkp, priv, uid, zoneid) > 0) {
8407 			(void) (*matched)(probe, arg);
8408 			nmatched++;
8409 		}
8410 		return (nmatched);
8411 	}
8412 
8413 	template.dtpr_mod = (char *)pkp->dtpk_mod;
8414 	template.dtpr_func = (char *)pkp->dtpk_func;
8415 	template.dtpr_name = (char *)pkp->dtpk_name;
8416 
8417 	/*
8418 	 * We want to find the most distinct of the module name, function
8419 	 * name, and name.  So for each one that is not a glob pattern or
8420 	 * empty string, we perform a lookup in the corresponding hash and
8421 	 * use the hash table with the fewest collisions to do our search.
8422 	 */
8423 	if (pkp->dtpk_mmatch == &dtrace_match_string &&
8424 	    (len = dtrace_hash_collisions(dtrace_bymod, &template)) < best) {
8425 		best = len;
8426 		hash = dtrace_bymod;
8427 	}
8428 
8429 	if (pkp->dtpk_fmatch == &dtrace_match_string &&
8430 	    (len = dtrace_hash_collisions(dtrace_byfunc, &template)) < best) {
8431 		best = len;
8432 		hash = dtrace_byfunc;
8433 	}
8434 
8435 	if (pkp->dtpk_nmatch == &dtrace_match_string &&
8436 	    (len = dtrace_hash_collisions(dtrace_byname, &template)) < best) {
8437 		best = len;
8438 		hash = dtrace_byname;
8439 	}
8440 
8441 	/*
8442 	 * If we did not select a hash table, iterate over every probe and
8443 	 * invoke our callback for each one that matches our input probe key.
8444 	 */
8445 	if (hash == NULL) {
8446 		for (i = 0; i < dtrace_nprobes; i++) {
8447 			if ((probe = dtrace_probes[i]) == NULL ||
8448 			    dtrace_match_probe(probe, pkp, priv, uid,
8449 			    zoneid) <= 0)
8450 				continue;
8451 
8452 			nmatched++;
8453 
8454 			if ((*matched)(probe, arg) != DTRACE_MATCH_NEXT)
8455 				break;
8456 		}
8457 
8458 		return (nmatched);
8459 	}
8460 
8461 	/*
8462 	 * If we selected a hash table, iterate over each probe of the same key
8463 	 * name and invoke the callback for every probe that matches the other
8464 	 * attributes of our input probe key.
8465 	 */
8466 	for (probe = dtrace_hash_lookup(hash, &template); probe != NULL;
8467 	    probe = *(DTRACE_HASHNEXT(hash, probe))) {
8468 
8469 		if (dtrace_match_probe(probe, pkp, priv, uid, zoneid) <= 0)
8470 			continue;
8471 
8472 		nmatched++;
8473 
8474 		if ((*matched)(probe, arg) != DTRACE_MATCH_NEXT)
8475 			break;
8476 	}
8477 
8478 	return (nmatched);
8479 }
8480 
8481 /*
8482  * Return the function pointer dtrace_probecmp() should use to compare the
8483  * specified pattern with a string.  For NULL or empty patterns, we select
8484  * dtrace_match_nul().  For glob pattern strings, we use dtrace_match_glob().
8485  * For non-empty non-glob strings, we use dtrace_match_string().
8486  */
8487 static dtrace_probekey_f *
8488 dtrace_probekey_func(const char *p)
8489 {
8490 	char c;
8491 
8492 	if (p == NULL || *p == '\0')
8493 		return (&dtrace_match_nul);
8494 
8495 	while ((c = *p++) != '\0') {
8496 		if (c == '[' || c == '?' || c == '*' || c == '\\')
8497 			return (&dtrace_match_glob);
8498 	}
8499 
8500 	return (&dtrace_match_string);
8501 }
8502 
8503 /*
8504  * Build a probe comparison key for use with dtrace_match_probe() from the
8505  * given probe description.  By convention, a null key only matches anchored
8506  * probes: if each field is the empty string, reset dtpk_fmatch to
8507  * dtrace_match_nonzero().
8508  */
8509 static void
8510 dtrace_probekey(dtrace_probedesc_t *pdp, dtrace_probekey_t *pkp)
8511 {
8512 	pkp->dtpk_prov = pdp->dtpd_provider;
8513 	pkp->dtpk_pmatch = dtrace_probekey_func(pdp->dtpd_provider);
8514 
8515 	pkp->dtpk_mod = pdp->dtpd_mod;
8516 	pkp->dtpk_mmatch = dtrace_probekey_func(pdp->dtpd_mod);
8517 
8518 	pkp->dtpk_func = pdp->dtpd_func;
8519 	pkp->dtpk_fmatch = dtrace_probekey_func(pdp->dtpd_func);
8520 
8521 	pkp->dtpk_name = pdp->dtpd_name;
8522 	pkp->dtpk_nmatch = dtrace_probekey_func(pdp->dtpd_name);
8523 
8524 	pkp->dtpk_id = pdp->dtpd_id;
8525 
8526 	if (pkp->dtpk_id == DTRACE_IDNONE &&
8527 	    pkp->dtpk_pmatch == &dtrace_match_nul &&
8528 	    pkp->dtpk_mmatch == &dtrace_match_nul &&
8529 	    pkp->dtpk_fmatch == &dtrace_match_nul &&
8530 	    pkp->dtpk_nmatch == &dtrace_match_nul)
8531 		pkp->dtpk_fmatch = &dtrace_match_nonzero;
8532 }
8533 
8534 /*
8535  * DTrace Provider-to-Framework API Functions
8536  *
8537  * These functions implement much of the Provider-to-Framework API, as
8538  * described in <sys/dtrace.h>.  The parts of the API not in this section are
8539  * the functions in the API for probe management (found below), and
8540  * dtrace_probe() itself (found above).
8541  */
8542 
8543 /*
8544  * Register the calling provider with the DTrace framework.  This should
8545  * generally be called by DTrace providers in their attach(9E) entry point.
8546  */
8547 int
8548 dtrace_register(const char *name, const dtrace_pattr_t *pap, uint32_t priv,
8549     cred_t *cr, const dtrace_pops_t *pops, void *arg, dtrace_provider_id_t *idp)
8550 {
8551 	dtrace_provider_t *provider;
8552 
8553 	if (name == NULL || pap == NULL || pops == NULL || idp == NULL) {
8554 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
8555 		    "arguments", name ? name : "<NULL>");
8556 		return (EINVAL);
8557 	}
8558 
8559 	if (name[0] == '\0' || dtrace_badname(name)) {
8560 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
8561 		    "provider name", name);
8562 		return (EINVAL);
8563 	}
8564 
8565 	if ((pops->dtps_provide == NULL && pops->dtps_provide_module == NULL) ||
8566 	    pops->dtps_enable == NULL || pops->dtps_disable == NULL ||
8567 	    pops->dtps_destroy == NULL ||
8568 	    ((pops->dtps_resume == NULL) != (pops->dtps_suspend == NULL))) {
8569 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
8570 		    "provider ops", name);
8571 		return (EINVAL);
8572 	}
8573 
8574 	if (dtrace_badattr(&pap->dtpa_provider) ||
8575 	    dtrace_badattr(&pap->dtpa_mod) ||
8576 	    dtrace_badattr(&pap->dtpa_func) ||
8577 	    dtrace_badattr(&pap->dtpa_name) ||
8578 	    dtrace_badattr(&pap->dtpa_args)) {
8579 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
8580 		    "provider attributes", name);
8581 		return (EINVAL);
8582 	}
8583 
8584 	if (priv & ~DTRACE_PRIV_ALL) {
8585 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
8586 		    "privilege attributes", name);
8587 		return (EINVAL);
8588 	}
8589 
8590 	if ((priv & DTRACE_PRIV_KERNEL) &&
8591 	    (priv & (DTRACE_PRIV_USER | DTRACE_PRIV_OWNER)) &&
8592 	    pops->dtps_usermode == NULL) {
8593 		cmn_err(CE_WARN, "failed to register provider '%s': need "
8594 		    "dtps_usermode() op for given privilege attributes", name);
8595 		return (EINVAL);
8596 	}
8597 
8598 	provider = kmem_zalloc(sizeof (dtrace_provider_t), KM_SLEEP);
8599 	provider->dtpv_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
8600 	(void) strcpy(provider->dtpv_name, name);
8601 
8602 	provider->dtpv_attr = *pap;
8603 	provider->dtpv_priv.dtpp_flags = priv;
8604 	if (cr != NULL) {
8605 		provider->dtpv_priv.dtpp_uid = crgetuid(cr);
8606 		provider->dtpv_priv.dtpp_zoneid = crgetzoneid(cr);
8607 	}
8608 	provider->dtpv_pops = *pops;
8609 
8610 	if (pops->dtps_provide == NULL) {
8611 		ASSERT(pops->dtps_provide_module != NULL);
8612 		provider->dtpv_pops.dtps_provide =
8613 		    (void (*)(void *, dtrace_probedesc_t *))dtrace_nullop;
8614 	}
8615 
8616 	if (pops->dtps_provide_module == NULL) {
8617 		ASSERT(pops->dtps_provide != NULL);
8618 		provider->dtpv_pops.dtps_provide_module =
8619 		    (void (*)(void *, modctl_t *))dtrace_nullop;
8620 	}
8621 
8622 	if (pops->dtps_suspend == NULL) {
8623 		ASSERT(pops->dtps_resume == NULL);
8624 		provider->dtpv_pops.dtps_suspend =
8625 		    (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
8626 		provider->dtpv_pops.dtps_resume =
8627 		    (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
8628 	}
8629 
8630 	provider->dtpv_arg = arg;
8631 	*idp = (dtrace_provider_id_t)provider;
8632 
8633 	if (pops == &dtrace_provider_ops) {
8634 		ASSERT(MUTEX_HELD(&dtrace_provider_lock));
8635 		ASSERT(MUTEX_HELD(&dtrace_lock));
8636 		ASSERT(dtrace_anon.dta_enabling == NULL);
8637 
8638 		/*
8639 		 * We make sure that the DTrace provider is at the head of
8640 		 * the provider chain.
8641 		 */
8642 		provider->dtpv_next = dtrace_provider;
8643 		dtrace_provider = provider;
8644 		return (0);
8645 	}
8646 
8647 	mutex_enter(&dtrace_provider_lock);
8648 	mutex_enter(&dtrace_lock);
8649 
8650 	/*
8651 	 * If there is at least one provider registered, we'll add this
8652 	 * provider after the first provider.
8653 	 */
8654 	if (dtrace_provider != NULL) {
8655 		provider->dtpv_next = dtrace_provider->dtpv_next;
8656 		dtrace_provider->dtpv_next = provider;
8657 	} else {
8658 		dtrace_provider = provider;
8659 	}
8660 
8661 	if (dtrace_retained != NULL) {
8662 		dtrace_enabling_provide(provider);
8663 
8664 		/*
8665 		 * Now we need to call dtrace_enabling_matchall() -- which
8666 		 * will acquire cpu_lock and dtrace_lock.  We therefore need
8667 		 * to drop all of our locks before calling into it...
8668 		 */
8669 		mutex_exit(&dtrace_lock);
8670 		mutex_exit(&dtrace_provider_lock);
8671 		dtrace_enabling_matchall();
8672 
8673 		return (0);
8674 	}
8675 
8676 	mutex_exit(&dtrace_lock);
8677 	mutex_exit(&dtrace_provider_lock);
8678 
8679 	return (0);
8680 }
8681 
8682 /*
8683  * Unregister the specified provider from the DTrace framework.  This should
8684  * generally be called by DTrace providers in their detach(9E) entry point.
8685  */
8686 int
8687 dtrace_unregister(dtrace_provider_id_t id)
8688 {
8689 	dtrace_provider_t *old = (dtrace_provider_t *)id;
8690 	dtrace_provider_t *prev = NULL;
8691 	int i, self = 0, noreap = 0;
8692 	dtrace_probe_t *probe, *first = NULL;
8693 
8694 	if (old->dtpv_pops.dtps_enable ==
8695 	    (void (*)(void *, dtrace_id_t, void *))dtrace_nullop) {
8696 		/*
8697 		 * If DTrace itself is the provider, we're called with locks
8698 		 * already held.
8699 		 */
8700 		ASSERT(old == dtrace_provider);
8701 #ifdef illumos
8702 		ASSERT(dtrace_devi != NULL);
8703 #endif
8704 		ASSERT(MUTEX_HELD(&dtrace_provider_lock));
8705 		ASSERT(MUTEX_HELD(&dtrace_lock));
8706 		self = 1;
8707 
8708 		if (dtrace_provider->dtpv_next != NULL) {
8709 			/*
8710 			 * There's another provider here; return failure.
8711 			 */
8712 			return (EBUSY);
8713 		}
8714 	} else {
8715 		mutex_enter(&dtrace_provider_lock);
8716 #ifdef illumos
8717 		mutex_enter(&mod_lock);
8718 #endif
8719 		mutex_enter(&dtrace_lock);
8720 	}
8721 
8722 	/*
8723 	 * If anyone has /dev/dtrace open, or if there are anonymous enabled
8724 	 * probes, we refuse to let providers slither away, unless this
8725 	 * provider has already been explicitly invalidated.
8726 	 */
8727 	if (!old->dtpv_defunct &&
8728 	    (dtrace_opens || (dtrace_anon.dta_state != NULL &&
8729 	    dtrace_anon.dta_state->dts_necbs > 0))) {
8730 		if (!self) {
8731 			mutex_exit(&dtrace_lock);
8732 #ifdef illumos
8733 			mutex_exit(&mod_lock);
8734 #endif
8735 			mutex_exit(&dtrace_provider_lock);
8736 		}
8737 		return (EBUSY);
8738 	}
8739 
8740 	/*
8741 	 * Attempt to destroy the probes associated with this provider.
8742 	 */
8743 	for (i = 0; i < dtrace_nprobes; i++) {
8744 		if ((probe = dtrace_probes[i]) == NULL)
8745 			continue;
8746 
8747 		if (probe->dtpr_provider != old)
8748 			continue;
8749 
8750 		if (probe->dtpr_ecb == NULL)
8751 			continue;
8752 
8753 		/*
8754 		 * If we are trying to unregister a defunct provider, and the
8755 		 * provider was made defunct within the interval dictated by
8756 		 * dtrace_unregister_defunct_reap, we'll (asynchronously)
8757 		 * attempt to reap our enablings.  To denote that the provider
8758 		 * should reattempt to unregister itself at some point in the
8759 		 * future, we will return a differentiable error code (EAGAIN
8760 		 * instead of EBUSY) in this case.
8761 		 */
8762 		if (dtrace_gethrtime() - old->dtpv_defunct >
8763 		    dtrace_unregister_defunct_reap)
8764 			noreap = 1;
8765 
8766 		if (!self) {
8767 			mutex_exit(&dtrace_lock);
8768 #ifdef illumos
8769 			mutex_exit(&mod_lock);
8770 #endif
8771 			mutex_exit(&dtrace_provider_lock);
8772 		}
8773 
8774 		if (noreap)
8775 			return (EBUSY);
8776 
8777 		(void) taskq_dispatch(dtrace_taskq,
8778 		    (task_func_t *)dtrace_enabling_reap, NULL, TQ_SLEEP);
8779 
8780 		return (EAGAIN);
8781 	}
8782 
8783 	/*
8784 	 * All of the probes for this provider are disabled; we can safely
8785 	 * remove all of them from their hash chains and from the probe array.
8786 	 */
8787 	for (i = 0; i < dtrace_nprobes; i++) {
8788 		if ((probe = dtrace_probes[i]) == NULL)
8789 			continue;
8790 
8791 		if (probe->dtpr_provider != old)
8792 			continue;
8793 
8794 		dtrace_probes[i] = NULL;
8795 
8796 		dtrace_hash_remove(dtrace_bymod, probe);
8797 		dtrace_hash_remove(dtrace_byfunc, probe);
8798 		dtrace_hash_remove(dtrace_byname, probe);
8799 
8800 		if (first == NULL) {
8801 			first = probe;
8802 			probe->dtpr_nextmod = NULL;
8803 		} else {
8804 			probe->dtpr_nextmod = first;
8805 			first = probe;
8806 		}
8807 	}
8808 
8809 	/*
8810 	 * The provider's probes have been removed from the hash chains and
8811 	 * from the probe array.  Now issue a dtrace_sync() to be sure that
8812 	 * everyone has cleared out from any probe array processing.
8813 	 */
8814 	dtrace_sync();
8815 
8816 	for (probe = first; probe != NULL; probe = first) {
8817 		first = probe->dtpr_nextmod;
8818 
8819 		old->dtpv_pops.dtps_destroy(old->dtpv_arg, probe->dtpr_id,
8820 		    probe->dtpr_arg);
8821 		kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
8822 		kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
8823 		kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
8824 #ifdef illumos
8825 		vmem_free(dtrace_arena, (void *)(uintptr_t)(probe->dtpr_id), 1);
8826 #else
8827 		free_unr(dtrace_arena, probe->dtpr_id);
8828 #endif
8829 		kmem_free(probe, sizeof (dtrace_probe_t));
8830 	}
8831 
8832 	if ((prev = dtrace_provider) == old) {
8833 #ifdef illumos
8834 		ASSERT(self || dtrace_devi == NULL);
8835 		ASSERT(old->dtpv_next == NULL || dtrace_devi == NULL);
8836 #endif
8837 		dtrace_provider = old->dtpv_next;
8838 	} else {
8839 		while (prev != NULL && prev->dtpv_next != old)
8840 			prev = prev->dtpv_next;
8841 
8842 		if (prev == NULL) {
8843 			panic("attempt to unregister non-existent "
8844 			    "dtrace provider %p\n", (void *)id);
8845 		}
8846 
8847 		prev->dtpv_next = old->dtpv_next;
8848 	}
8849 
8850 	if (!self) {
8851 		mutex_exit(&dtrace_lock);
8852 #ifdef illumos
8853 		mutex_exit(&mod_lock);
8854 #endif
8855 		mutex_exit(&dtrace_provider_lock);
8856 	}
8857 
8858 	kmem_free(old->dtpv_name, strlen(old->dtpv_name) + 1);
8859 	kmem_free(old, sizeof (dtrace_provider_t));
8860 
8861 	return (0);
8862 }
8863 
8864 /*
8865  * Invalidate the specified provider.  All subsequent probe lookups for the
8866  * specified provider will fail, but its probes will not be removed.
8867  */
8868 void
8869 dtrace_invalidate(dtrace_provider_id_t id)
8870 {
8871 	dtrace_provider_t *pvp = (dtrace_provider_t *)id;
8872 
8873 	ASSERT(pvp->dtpv_pops.dtps_enable !=
8874 	    (void (*)(void *, dtrace_id_t, void *))dtrace_nullop);
8875 
8876 	mutex_enter(&dtrace_provider_lock);
8877 	mutex_enter(&dtrace_lock);
8878 
8879 	pvp->dtpv_defunct = dtrace_gethrtime();
8880 
8881 	mutex_exit(&dtrace_lock);
8882 	mutex_exit(&dtrace_provider_lock);
8883 }
8884 
8885 /*
8886  * Indicate whether or not DTrace has attached.
8887  */
8888 int
8889 dtrace_attached(void)
8890 {
8891 	/*
8892 	 * dtrace_provider will be non-NULL iff the DTrace driver has
8893 	 * attached.  (It's non-NULL because DTrace is always itself a
8894 	 * provider.)
8895 	 */
8896 	return (dtrace_provider != NULL);
8897 }
8898 
8899 /*
8900  * Remove all the unenabled probes for the given provider.  This function is
8901  * not unlike dtrace_unregister(), except that it doesn't remove the provider
8902  * -- just as many of its associated probes as it can.
8903  */
8904 int
8905 dtrace_condense(dtrace_provider_id_t id)
8906 {
8907 	dtrace_provider_t *prov = (dtrace_provider_t *)id;
8908 	int i;
8909 	dtrace_probe_t *probe;
8910 
8911 	/*
8912 	 * Make sure this isn't the dtrace provider itself.
8913 	 */
8914 	ASSERT(prov->dtpv_pops.dtps_enable !=
8915 	    (void (*)(void *, dtrace_id_t, void *))dtrace_nullop);
8916 
8917 	mutex_enter(&dtrace_provider_lock);
8918 	mutex_enter(&dtrace_lock);
8919 
8920 	/*
8921 	 * Attempt to destroy the probes associated with this provider.
8922 	 */
8923 	for (i = 0; i < dtrace_nprobes; i++) {
8924 		if ((probe = dtrace_probes[i]) == NULL)
8925 			continue;
8926 
8927 		if (probe->dtpr_provider != prov)
8928 			continue;
8929 
8930 		if (probe->dtpr_ecb != NULL)
8931 			continue;
8932 
8933 		dtrace_probes[i] = NULL;
8934 
8935 		dtrace_hash_remove(dtrace_bymod, probe);
8936 		dtrace_hash_remove(dtrace_byfunc, probe);
8937 		dtrace_hash_remove(dtrace_byname, probe);
8938 
8939 		prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, i + 1,
8940 		    probe->dtpr_arg);
8941 		kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
8942 		kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
8943 		kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
8944 		kmem_free(probe, sizeof (dtrace_probe_t));
8945 #ifdef illumos
8946 		vmem_free(dtrace_arena, (void *)((uintptr_t)i + 1), 1);
8947 #else
8948 		free_unr(dtrace_arena, i + 1);
8949 #endif
8950 	}
8951 
8952 	mutex_exit(&dtrace_lock);
8953 	mutex_exit(&dtrace_provider_lock);
8954 
8955 	return (0);
8956 }
8957 
8958 /*
8959  * DTrace Probe Management Functions
8960  *
8961  * The functions in this section perform the DTrace probe management,
8962  * including functions to create probes, look-up probes, and call into the
8963  * providers to request that probes be provided.  Some of these functions are
8964  * in the Provider-to-Framework API; these functions can be identified by the
8965  * fact that they are not declared "static".
8966  */
8967 
8968 /*
8969  * Create a probe with the specified module name, function name, and name.
8970  */
8971 dtrace_id_t
8972 dtrace_probe_create(dtrace_provider_id_t prov, const char *mod,
8973     const char *func, const char *name, int aframes, void *arg)
8974 {
8975 	dtrace_probe_t *probe, **probes;
8976 	dtrace_provider_t *provider = (dtrace_provider_t *)prov;
8977 	dtrace_id_t id;
8978 
8979 	if (provider == dtrace_provider) {
8980 		ASSERT(MUTEX_HELD(&dtrace_lock));
8981 	} else {
8982 		mutex_enter(&dtrace_lock);
8983 	}
8984 
8985 #ifdef illumos
8986 	id = (dtrace_id_t)(uintptr_t)vmem_alloc(dtrace_arena, 1,
8987 	    VM_BESTFIT | VM_SLEEP);
8988 #else
8989 	id = alloc_unr(dtrace_arena);
8990 #endif
8991 	probe = kmem_zalloc(sizeof (dtrace_probe_t), KM_SLEEP);
8992 
8993 	probe->dtpr_id = id;
8994 	probe->dtpr_gen = dtrace_probegen++;
8995 	probe->dtpr_mod = dtrace_strdup(mod);
8996 	probe->dtpr_func = dtrace_strdup(func);
8997 	probe->dtpr_name = dtrace_strdup(name);
8998 	probe->dtpr_arg = arg;
8999 	probe->dtpr_aframes = aframes;
9000 	probe->dtpr_provider = provider;
9001 
9002 	dtrace_hash_add(dtrace_bymod, probe);
9003 	dtrace_hash_add(dtrace_byfunc, probe);
9004 	dtrace_hash_add(dtrace_byname, probe);
9005 
9006 	if (id - 1 >= dtrace_nprobes) {
9007 		size_t osize = dtrace_nprobes * sizeof (dtrace_probe_t *);
9008 		size_t nsize = osize << 1;
9009 
9010 		if (nsize == 0) {
9011 			ASSERT(osize == 0);
9012 			ASSERT(dtrace_probes == NULL);
9013 			nsize = sizeof (dtrace_probe_t *);
9014 		}
9015 
9016 		probes = kmem_zalloc(nsize, KM_SLEEP);
9017 
9018 		if (dtrace_probes == NULL) {
9019 			ASSERT(osize == 0);
9020 			dtrace_probes = probes;
9021 			dtrace_nprobes = 1;
9022 		} else {
9023 			dtrace_probe_t **oprobes = dtrace_probes;
9024 
9025 			bcopy(oprobes, probes, osize);
9026 			dtrace_membar_producer();
9027 			dtrace_probes = probes;
9028 
9029 			dtrace_sync();
9030 
9031 			/*
9032 			 * All CPUs are now seeing the new probes array; we can
9033 			 * safely free the old array.
9034 			 */
9035 			kmem_free(oprobes, osize);
9036 			dtrace_nprobes <<= 1;
9037 		}
9038 
9039 		ASSERT(id - 1 < dtrace_nprobes);
9040 	}
9041 
9042 	ASSERT(dtrace_probes[id - 1] == NULL);
9043 	dtrace_probes[id - 1] = probe;
9044 
9045 	if (provider != dtrace_provider)
9046 		mutex_exit(&dtrace_lock);
9047 
9048 	return (id);
9049 }
9050 
9051 static dtrace_probe_t *
9052 dtrace_probe_lookup_id(dtrace_id_t id)
9053 {
9054 	ASSERT(MUTEX_HELD(&dtrace_lock));
9055 
9056 	if (id == 0 || id > dtrace_nprobes)
9057 		return (NULL);
9058 
9059 	return (dtrace_probes[id - 1]);
9060 }
9061 
9062 static int
9063 dtrace_probe_lookup_match(dtrace_probe_t *probe, void *arg)
9064 {
9065 	*((dtrace_id_t *)arg) = probe->dtpr_id;
9066 
9067 	return (DTRACE_MATCH_DONE);
9068 }
9069 
9070 /*
9071  * Look up a probe based on provider and one or more of module name, function
9072  * name and probe name.
9073  */
9074 dtrace_id_t
9075 dtrace_probe_lookup(dtrace_provider_id_t prid, char *mod,
9076     char *func, char *name)
9077 {
9078 	dtrace_probekey_t pkey;
9079 	dtrace_id_t id;
9080 	int match;
9081 
9082 	pkey.dtpk_prov = ((dtrace_provider_t *)prid)->dtpv_name;
9083 	pkey.dtpk_pmatch = &dtrace_match_string;
9084 	pkey.dtpk_mod = mod;
9085 	pkey.dtpk_mmatch = mod ? &dtrace_match_string : &dtrace_match_nul;
9086 	pkey.dtpk_func = func;
9087 	pkey.dtpk_fmatch = func ? &dtrace_match_string : &dtrace_match_nul;
9088 	pkey.dtpk_name = name;
9089 	pkey.dtpk_nmatch = name ? &dtrace_match_string : &dtrace_match_nul;
9090 	pkey.dtpk_id = DTRACE_IDNONE;
9091 
9092 	mutex_enter(&dtrace_lock);
9093 	match = dtrace_match(&pkey, DTRACE_PRIV_ALL, 0, 0,
9094 	    dtrace_probe_lookup_match, &id);
9095 	mutex_exit(&dtrace_lock);
9096 
9097 	ASSERT(match == 1 || match == 0);
9098 	return (match ? id : 0);
9099 }
9100 
9101 /*
9102  * Returns the probe argument associated with the specified probe.
9103  */
9104 void *
9105 dtrace_probe_arg(dtrace_provider_id_t id, dtrace_id_t pid)
9106 {
9107 	dtrace_probe_t *probe;
9108 	void *rval = NULL;
9109 
9110 	mutex_enter(&dtrace_lock);
9111 
9112 	if ((probe = dtrace_probe_lookup_id(pid)) != NULL &&
9113 	    probe->dtpr_provider == (dtrace_provider_t *)id)
9114 		rval = probe->dtpr_arg;
9115 
9116 	mutex_exit(&dtrace_lock);
9117 
9118 	return (rval);
9119 }
9120 
9121 /*
9122  * Copy a probe into a probe description.
9123  */
9124 static void
9125 dtrace_probe_description(const dtrace_probe_t *prp, dtrace_probedesc_t *pdp)
9126 {
9127 	bzero(pdp, sizeof (dtrace_probedesc_t));
9128 	pdp->dtpd_id = prp->dtpr_id;
9129 
9130 	(void) strncpy(pdp->dtpd_provider,
9131 	    prp->dtpr_provider->dtpv_name, DTRACE_PROVNAMELEN - 1);
9132 
9133 	(void) strncpy(pdp->dtpd_mod, prp->dtpr_mod, DTRACE_MODNAMELEN - 1);
9134 	(void) strncpy(pdp->dtpd_func, prp->dtpr_func, DTRACE_FUNCNAMELEN - 1);
9135 	(void) strncpy(pdp->dtpd_name, prp->dtpr_name, DTRACE_NAMELEN - 1);
9136 }
9137 
9138 /*
9139  * Called to indicate that a probe -- or probes -- should be provided by a
9140  * specfied provider.  If the specified description is NULL, the provider will
9141  * be told to provide all of its probes.  (This is done whenever a new
9142  * consumer comes along, or whenever a retained enabling is to be matched.) If
9143  * the specified description is non-NULL, the provider is given the
9144  * opportunity to dynamically provide the specified probe, allowing providers
9145  * to support the creation of probes on-the-fly.  (So-called _autocreated_
9146  * probes.)  If the provider is NULL, the operations will be applied to all
9147  * providers; if the provider is non-NULL the operations will only be applied
9148  * to the specified provider.  The dtrace_provider_lock must be held, and the
9149  * dtrace_lock must _not_ be held -- the provider's dtps_provide() operation
9150  * will need to grab the dtrace_lock when it reenters the framework through
9151  * dtrace_probe_lookup(), dtrace_probe_create(), etc.
9152  */
9153 static void
9154 dtrace_probe_provide(dtrace_probedesc_t *desc, dtrace_provider_t *prv)
9155 {
9156 #ifdef illumos
9157 	modctl_t *ctl;
9158 #endif
9159 	int all = 0;
9160 
9161 	ASSERT(MUTEX_HELD(&dtrace_provider_lock));
9162 
9163 	if (prv == NULL) {
9164 		all = 1;
9165 		prv = dtrace_provider;
9166 	}
9167 
9168 	do {
9169 		/*
9170 		 * First, call the blanket provide operation.
9171 		 */
9172 		prv->dtpv_pops.dtps_provide(prv->dtpv_arg, desc);
9173 
9174 #ifdef illumos
9175 		/*
9176 		 * Now call the per-module provide operation.  We will grab
9177 		 * mod_lock to prevent the list from being modified.  Note
9178 		 * that this also prevents the mod_busy bits from changing.
9179 		 * (mod_busy can only be changed with mod_lock held.)
9180 		 */
9181 		mutex_enter(&mod_lock);
9182 
9183 		ctl = &modules;
9184 		do {
9185 			if (ctl->mod_busy || ctl->mod_mp == NULL)
9186 				continue;
9187 
9188 			prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
9189 
9190 		} while ((ctl = ctl->mod_next) != &modules);
9191 
9192 		mutex_exit(&mod_lock);
9193 #endif
9194 	} while (all && (prv = prv->dtpv_next) != NULL);
9195 }
9196 
9197 #ifdef illumos
9198 /*
9199  * Iterate over each probe, and call the Framework-to-Provider API function
9200  * denoted by offs.
9201  */
9202 static void
9203 dtrace_probe_foreach(uintptr_t offs)
9204 {
9205 	dtrace_provider_t *prov;
9206 	void (*func)(void *, dtrace_id_t, void *);
9207 	dtrace_probe_t *probe;
9208 	dtrace_icookie_t cookie;
9209 	int i;
9210 
9211 	/*
9212 	 * We disable interrupts to walk through the probe array.  This is
9213 	 * safe -- the dtrace_sync() in dtrace_unregister() assures that we
9214 	 * won't see stale data.
9215 	 */
9216 	cookie = dtrace_interrupt_disable();
9217 
9218 	for (i = 0; i < dtrace_nprobes; i++) {
9219 		if ((probe = dtrace_probes[i]) == NULL)
9220 			continue;
9221 
9222 		if (probe->dtpr_ecb == NULL) {
9223 			/*
9224 			 * This probe isn't enabled -- don't call the function.
9225 			 */
9226 			continue;
9227 		}
9228 
9229 		prov = probe->dtpr_provider;
9230 		func = *((void(**)(void *, dtrace_id_t, void *))
9231 		    ((uintptr_t)&prov->dtpv_pops + offs));
9232 
9233 		func(prov->dtpv_arg, i + 1, probe->dtpr_arg);
9234 	}
9235 
9236 	dtrace_interrupt_enable(cookie);
9237 }
9238 #endif
9239 
9240 static int
9241 dtrace_probe_enable(dtrace_probedesc_t *desc, dtrace_enabling_t *enab)
9242 {
9243 	dtrace_probekey_t pkey;
9244 	uint32_t priv;
9245 	uid_t uid;
9246 	zoneid_t zoneid;
9247 
9248 	ASSERT(MUTEX_HELD(&dtrace_lock));
9249 	dtrace_ecb_create_cache = NULL;
9250 
9251 	if (desc == NULL) {
9252 		/*
9253 		 * If we're passed a NULL description, we're being asked to
9254 		 * create an ECB with a NULL probe.
9255 		 */
9256 		(void) dtrace_ecb_create_enable(NULL, enab);
9257 		return (0);
9258 	}
9259 
9260 	dtrace_probekey(desc, &pkey);
9261 	dtrace_cred2priv(enab->dten_vstate->dtvs_state->dts_cred.dcr_cred,
9262 	    &priv, &uid, &zoneid);
9263 
9264 	return (dtrace_match(&pkey, priv, uid, zoneid, dtrace_ecb_create_enable,
9265 	    enab));
9266 }
9267 
9268 /*
9269  * DTrace Helper Provider Functions
9270  */
9271 static void
9272 dtrace_dofattr2attr(dtrace_attribute_t *attr, const dof_attr_t dofattr)
9273 {
9274 	attr->dtat_name = DOF_ATTR_NAME(dofattr);
9275 	attr->dtat_data = DOF_ATTR_DATA(dofattr);
9276 	attr->dtat_class = DOF_ATTR_CLASS(dofattr);
9277 }
9278 
9279 static void
9280 dtrace_dofprov2hprov(dtrace_helper_provdesc_t *hprov,
9281     const dof_provider_t *dofprov, char *strtab)
9282 {
9283 	hprov->dthpv_provname = strtab + dofprov->dofpv_name;
9284 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_provider,
9285 	    dofprov->dofpv_provattr);
9286 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_mod,
9287 	    dofprov->dofpv_modattr);
9288 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_func,
9289 	    dofprov->dofpv_funcattr);
9290 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_name,
9291 	    dofprov->dofpv_nameattr);
9292 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_args,
9293 	    dofprov->dofpv_argsattr);
9294 }
9295 
9296 static void
9297 dtrace_helper_provide_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
9298 {
9299 	uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
9300 	dof_hdr_t *dof = (dof_hdr_t *)daddr;
9301 	dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
9302 	dof_provider_t *provider;
9303 	dof_probe_t *probe;
9304 	uint32_t *off, *enoff;
9305 	uint8_t *arg;
9306 	char *strtab;
9307 	uint_t i, nprobes;
9308 	dtrace_helper_provdesc_t dhpv;
9309 	dtrace_helper_probedesc_t dhpb;
9310 	dtrace_meta_t *meta = dtrace_meta_pid;
9311 	dtrace_mops_t *mops = &meta->dtm_mops;
9312 	void *parg;
9313 
9314 	provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
9315 	str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
9316 	    provider->dofpv_strtab * dof->dofh_secsize);
9317 	prb_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
9318 	    provider->dofpv_probes * dof->dofh_secsize);
9319 	arg_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
9320 	    provider->dofpv_prargs * dof->dofh_secsize);
9321 	off_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
9322 	    provider->dofpv_proffs * dof->dofh_secsize);
9323 
9324 	strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
9325 	off = (uint32_t *)(uintptr_t)(daddr + off_sec->dofs_offset);
9326 	arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
9327 	enoff = NULL;
9328 
9329 	/*
9330 	 * See dtrace_helper_provider_validate().
9331 	 */
9332 	if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
9333 	    provider->dofpv_prenoffs != DOF_SECT_NONE) {
9334 		enoff_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
9335 		    provider->dofpv_prenoffs * dof->dofh_secsize);
9336 		enoff = (uint32_t *)(uintptr_t)(daddr + enoff_sec->dofs_offset);
9337 	}
9338 
9339 	nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
9340 
9341 	/*
9342 	 * Create the provider.
9343 	 */
9344 	dtrace_dofprov2hprov(&dhpv, provider, strtab);
9345 
9346 	if ((parg = mops->dtms_provide_pid(meta->dtm_arg, &dhpv, pid)) == NULL)
9347 		return;
9348 
9349 	meta->dtm_count++;
9350 
9351 	/*
9352 	 * Create the probes.
9353 	 */
9354 	for (i = 0; i < nprobes; i++) {
9355 		probe = (dof_probe_t *)(uintptr_t)(daddr +
9356 		    prb_sec->dofs_offset + i * prb_sec->dofs_entsize);
9357 
9358 		/* See the check in dtrace_helper_provider_validate(). */
9359 		if (strlen(strtab + probe->dofpr_func) >= DTRACE_FUNCNAMELEN)
9360 			continue;
9361 
9362 		dhpb.dthpb_mod = dhp->dofhp_mod;
9363 		dhpb.dthpb_func = strtab + probe->dofpr_func;
9364 		dhpb.dthpb_name = strtab + probe->dofpr_name;
9365 		dhpb.dthpb_base = probe->dofpr_addr;
9366 		dhpb.dthpb_offs = off + probe->dofpr_offidx;
9367 		dhpb.dthpb_noffs = probe->dofpr_noffs;
9368 		if (enoff != NULL) {
9369 			dhpb.dthpb_enoffs = enoff + probe->dofpr_enoffidx;
9370 			dhpb.dthpb_nenoffs = probe->dofpr_nenoffs;
9371 		} else {
9372 			dhpb.dthpb_enoffs = NULL;
9373 			dhpb.dthpb_nenoffs = 0;
9374 		}
9375 		dhpb.dthpb_args = arg + probe->dofpr_argidx;
9376 		dhpb.dthpb_nargc = probe->dofpr_nargc;
9377 		dhpb.dthpb_xargc = probe->dofpr_xargc;
9378 		dhpb.dthpb_ntypes = strtab + probe->dofpr_nargv;
9379 		dhpb.dthpb_xtypes = strtab + probe->dofpr_xargv;
9380 
9381 		mops->dtms_create_probe(meta->dtm_arg, parg, &dhpb);
9382 	}
9383 }
9384 
9385 static void
9386 dtrace_helper_provide(dof_helper_t *dhp, pid_t pid)
9387 {
9388 	uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
9389 	dof_hdr_t *dof = (dof_hdr_t *)daddr;
9390 	int i;
9391 
9392 	ASSERT(MUTEX_HELD(&dtrace_meta_lock));
9393 
9394 	for (i = 0; i < dof->dofh_secnum; i++) {
9395 		dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
9396 		    dof->dofh_secoff + i * dof->dofh_secsize);
9397 
9398 		if (sec->dofs_type != DOF_SECT_PROVIDER)
9399 			continue;
9400 
9401 		dtrace_helper_provide_one(dhp, sec, pid);
9402 	}
9403 
9404 	/*
9405 	 * We may have just created probes, so we must now rematch against
9406 	 * any retained enablings.  Note that this call will acquire both
9407 	 * cpu_lock and dtrace_lock; the fact that we are holding
9408 	 * dtrace_meta_lock now is what defines the ordering with respect to
9409 	 * these three locks.
9410 	 */
9411 	dtrace_enabling_matchall();
9412 }
9413 
9414 static void
9415 dtrace_helper_provider_remove_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
9416 {
9417 	uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
9418 	dof_hdr_t *dof = (dof_hdr_t *)daddr;
9419 	dof_sec_t *str_sec;
9420 	dof_provider_t *provider;
9421 	char *strtab;
9422 	dtrace_helper_provdesc_t dhpv;
9423 	dtrace_meta_t *meta = dtrace_meta_pid;
9424 	dtrace_mops_t *mops = &meta->dtm_mops;
9425 
9426 	provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
9427 	str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
9428 	    provider->dofpv_strtab * dof->dofh_secsize);
9429 
9430 	strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
9431 
9432 	/*
9433 	 * Create the provider.
9434 	 */
9435 	dtrace_dofprov2hprov(&dhpv, provider, strtab);
9436 
9437 	mops->dtms_remove_pid(meta->dtm_arg, &dhpv, pid);
9438 
9439 	meta->dtm_count--;
9440 }
9441 
9442 static void
9443 dtrace_helper_provider_remove(dof_helper_t *dhp, pid_t pid)
9444 {
9445 	uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
9446 	dof_hdr_t *dof = (dof_hdr_t *)daddr;
9447 	int i;
9448 
9449 	ASSERT(MUTEX_HELD(&dtrace_meta_lock));
9450 
9451 	for (i = 0; i < dof->dofh_secnum; i++) {
9452 		dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
9453 		    dof->dofh_secoff + i * dof->dofh_secsize);
9454 
9455 		if (sec->dofs_type != DOF_SECT_PROVIDER)
9456 			continue;
9457 
9458 		dtrace_helper_provider_remove_one(dhp, sec, pid);
9459 	}
9460 }
9461 
9462 /*
9463  * DTrace Meta Provider-to-Framework API Functions
9464  *
9465  * These functions implement the Meta Provider-to-Framework API, as described
9466  * in <sys/dtrace.h>.
9467  */
9468 int
9469 dtrace_meta_register(const char *name, const dtrace_mops_t *mops, void *arg,
9470     dtrace_meta_provider_id_t *idp)
9471 {
9472 	dtrace_meta_t *meta;
9473 	dtrace_helpers_t *help, *next;
9474 	int i;
9475 
9476 	*idp = DTRACE_METAPROVNONE;
9477 
9478 	/*
9479 	 * We strictly don't need the name, but we hold onto it for
9480 	 * debuggability. All hail error queues!
9481 	 */
9482 	if (name == NULL) {
9483 		cmn_err(CE_WARN, "failed to register meta-provider: "
9484 		    "invalid name");
9485 		return (EINVAL);
9486 	}
9487 
9488 	if (mops == NULL ||
9489 	    mops->dtms_create_probe == NULL ||
9490 	    mops->dtms_provide_pid == NULL ||
9491 	    mops->dtms_remove_pid == NULL) {
9492 		cmn_err(CE_WARN, "failed to register meta-register %s: "
9493 		    "invalid ops", name);
9494 		return (EINVAL);
9495 	}
9496 
9497 	meta = kmem_zalloc(sizeof (dtrace_meta_t), KM_SLEEP);
9498 	meta->dtm_mops = *mops;
9499 	meta->dtm_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
9500 	(void) strcpy(meta->dtm_name, name);
9501 	meta->dtm_arg = arg;
9502 
9503 	mutex_enter(&dtrace_meta_lock);
9504 	mutex_enter(&dtrace_lock);
9505 
9506 	if (dtrace_meta_pid != NULL) {
9507 		mutex_exit(&dtrace_lock);
9508 		mutex_exit(&dtrace_meta_lock);
9509 		cmn_err(CE_WARN, "failed to register meta-register %s: "
9510 		    "user-land meta-provider exists", name);
9511 		kmem_free(meta->dtm_name, strlen(meta->dtm_name) + 1);
9512 		kmem_free(meta, sizeof (dtrace_meta_t));
9513 		return (EINVAL);
9514 	}
9515 
9516 	dtrace_meta_pid = meta;
9517 	*idp = (dtrace_meta_provider_id_t)meta;
9518 
9519 	/*
9520 	 * If there are providers and probes ready to go, pass them
9521 	 * off to the new meta provider now.
9522 	 */
9523 
9524 	help = dtrace_deferred_pid;
9525 	dtrace_deferred_pid = NULL;
9526 
9527 	mutex_exit(&dtrace_lock);
9528 
9529 	while (help != NULL) {
9530 		for (i = 0; i < help->dthps_nprovs; i++) {
9531 			dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
9532 			    help->dthps_pid);
9533 		}
9534 
9535 		next = help->dthps_next;
9536 		help->dthps_next = NULL;
9537 		help->dthps_prev = NULL;
9538 		help->dthps_deferred = 0;
9539 		help = next;
9540 	}
9541 
9542 	mutex_exit(&dtrace_meta_lock);
9543 
9544 	return (0);
9545 }
9546 
9547 int
9548 dtrace_meta_unregister(dtrace_meta_provider_id_t id)
9549 {
9550 	dtrace_meta_t **pp, *old = (dtrace_meta_t *)id;
9551 
9552 	mutex_enter(&dtrace_meta_lock);
9553 	mutex_enter(&dtrace_lock);
9554 
9555 	if (old == dtrace_meta_pid) {
9556 		pp = &dtrace_meta_pid;
9557 	} else {
9558 		panic("attempt to unregister non-existent "
9559 		    "dtrace meta-provider %p\n", (void *)old);
9560 	}
9561 
9562 	if (old->dtm_count != 0) {
9563 		mutex_exit(&dtrace_lock);
9564 		mutex_exit(&dtrace_meta_lock);
9565 		return (EBUSY);
9566 	}
9567 
9568 	*pp = NULL;
9569 
9570 	mutex_exit(&dtrace_lock);
9571 	mutex_exit(&dtrace_meta_lock);
9572 
9573 	kmem_free(old->dtm_name, strlen(old->dtm_name) + 1);
9574 	kmem_free(old, sizeof (dtrace_meta_t));
9575 
9576 	return (0);
9577 }
9578 
9579 
9580 /*
9581  * DTrace DIF Object Functions
9582  */
9583 static int
9584 dtrace_difo_err(uint_t pc, const char *format, ...)
9585 {
9586 	if (dtrace_err_verbose) {
9587 		va_list alist;
9588 
9589 		(void) uprintf("dtrace DIF object error: [%u]: ", pc);
9590 		va_start(alist, format);
9591 		(void) vuprintf(format, alist);
9592 		va_end(alist);
9593 	}
9594 
9595 #ifdef DTRACE_ERRDEBUG
9596 	dtrace_errdebug(format);
9597 #endif
9598 	return (1);
9599 }
9600 
9601 /*
9602  * Validate a DTrace DIF object by checking the IR instructions.  The following
9603  * rules are currently enforced by dtrace_difo_validate():
9604  *
9605  * 1. Each instruction must have a valid opcode
9606  * 2. Each register, string, variable, or subroutine reference must be valid
9607  * 3. No instruction can modify register %r0 (must be zero)
9608  * 4. All instruction reserved bits must be set to zero
9609  * 5. The last instruction must be a "ret" instruction
9610  * 6. All branch targets must reference a valid instruction _after_ the branch
9611  */
9612 static int
9613 dtrace_difo_validate(dtrace_difo_t *dp, dtrace_vstate_t *vstate, uint_t nregs,
9614     cred_t *cr)
9615 {
9616 	int err = 0, i;
9617 	int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
9618 	int kcheckload;
9619 	uint_t pc;
9620 	int maxglobal = -1, maxlocal = -1, maxtlocal = -1;
9621 
9622 	kcheckload = cr == NULL ||
9623 	    (vstate->dtvs_state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) == 0;
9624 
9625 	dp->dtdo_destructive = 0;
9626 
9627 	for (pc = 0; pc < dp->dtdo_len && err == 0; pc++) {
9628 		dif_instr_t instr = dp->dtdo_buf[pc];
9629 
9630 		uint_t r1 = DIF_INSTR_R1(instr);
9631 		uint_t r2 = DIF_INSTR_R2(instr);
9632 		uint_t rd = DIF_INSTR_RD(instr);
9633 		uint_t rs = DIF_INSTR_RS(instr);
9634 		uint_t label = DIF_INSTR_LABEL(instr);
9635 		uint_t v = DIF_INSTR_VAR(instr);
9636 		uint_t subr = DIF_INSTR_SUBR(instr);
9637 		uint_t type = DIF_INSTR_TYPE(instr);
9638 		uint_t op = DIF_INSTR_OP(instr);
9639 
9640 		switch (op) {
9641 		case DIF_OP_OR:
9642 		case DIF_OP_XOR:
9643 		case DIF_OP_AND:
9644 		case DIF_OP_SLL:
9645 		case DIF_OP_SRL:
9646 		case DIF_OP_SRA:
9647 		case DIF_OP_SUB:
9648 		case DIF_OP_ADD:
9649 		case DIF_OP_MUL:
9650 		case DIF_OP_SDIV:
9651 		case DIF_OP_UDIV:
9652 		case DIF_OP_SREM:
9653 		case DIF_OP_UREM:
9654 		case DIF_OP_COPYS:
9655 			if (r1 >= nregs)
9656 				err += efunc(pc, "invalid register %u\n", r1);
9657 			if (r2 >= nregs)
9658 				err += efunc(pc, "invalid register %u\n", r2);
9659 			if (rd >= nregs)
9660 				err += efunc(pc, "invalid register %u\n", rd);
9661 			if (rd == 0)
9662 				err += efunc(pc, "cannot write to %r0\n");
9663 			break;
9664 		case DIF_OP_NOT:
9665 		case DIF_OP_MOV:
9666 		case DIF_OP_ALLOCS:
9667 			if (r1 >= nregs)
9668 				err += efunc(pc, "invalid register %u\n", r1);
9669 			if (r2 != 0)
9670 				err += efunc(pc, "non-zero reserved bits\n");
9671 			if (rd >= nregs)
9672 				err += efunc(pc, "invalid register %u\n", rd);
9673 			if (rd == 0)
9674 				err += efunc(pc, "cannot write to %r0\n");
9675 			break;
9676 		case DIF_OP_LDSB:
9677 		case DIF_OP_LDSH:
9678 		case DIF_OP_LDSW:
9679 		case DIF_OP_LDUB:
9680 		case DIF_OP_LDUH:
9681 		case DIF_OP_LDUW:
9682 		case DIF_OP_LDX:
9683 			if (r1 >= nregs)
9684 				err += efunc(pc, "invalid register %u\n", r1);
9685 			if (r2 != 0)
9686 				err += efunc(pc, "non-zero reserved bits\n");
9687 			if (rd >= nregs)
9688 				err += efunc(pc, "invalid register %u\n", rd);
9689 			if (rd == 0)
9690 				err += efunc(pc, "cannot write to %r0\n");
9691 			if (kcheckload)
9692 				dp->dtdo_buf[pc] = DIF_INSTR_LOAD(op +
9693 				    DIF_OP_RLDSB - DIF_OP_LDSB, r1, rd);
9694 			break;
9695 		case DIF_OP_RLDSB:
9696 		case DIF_OP_RLDSH:
9697 		case DIF_OP_RLDSW:
9698 		case DIF_OP_RLDUB:
9699 		case DIF_OP_RLDUH:
9700 		case DIF_OP_RLDUW:
9701 		case DIF_OP_RLDX:
9702 			if (r1 >= nregs)
9703 				err += efunc(pc, "invalid register %u\n", r1);
9704 			if (r2 != 0)
9705 				err += efunc(pc, "non-zero reserved bits\n");
9706 			if (rd >= nregs)
9707 				err += efunc(pc, "invalid register %u\n", rd);
9708 			if (rd == 0)
9709 				err += efunc(pc, "cannot write to %r0\n");
9710 			break;
9711 		case DIF_OP_ULDSB:
9712 		case DIF_OP_ULDSH:
9713 		case DIF_OP_ULDSW:
9714 		case DIF_OP_ULDUB:
9715 		case DIF_OP_ULDUH:
9716 		case DIF_OP_ULDUW:
9717 		case DIF_OP_ULDX:
9718 			if (r1 >= nregs)
9719 				err += efunc(pc, "invalid register %u\n", r1);
9720 			if (r2 != 0)
9721 				err += efunc(pc, "non-zero reserved bits\n");
9722 			if (rd >= nregs)
9723 				err += efunc(pc, "invalid register %u\n", rd);
9724 			if (rd == 0)
9725 				err += efunc(pc, "cannot write to %r0\n");
9726 			break;
9727 		case DIF_OP_STB:
9728 		case DIF_OP_STH:
9729 		case DIF_OP_STW:
9730 		case DIF_OP_STX:
9731 			if (r1 >= nregs)
9732 				err += efunc(pc, "invalid register %u\n", r1);
9733 			if (r2 != 0)
9734 				err += efunc(pc, "non-zero reserved bits\n");
9735 			if (rd >= nregs)
9736 				err += efunc(pc, "invalid register %u\n", rd);
9737 			if (rd == 0)
9738 				err += efunc(pc, "cannot write to 0 address\n");
9739 			break;
9740 		case DIF_OP_CMP:
9741 		case DIF_OP_SCMP:
9742 			if (r1 >= nregs)
9743 				err += efunc(pc, "invalid register %u\n", r1);
9744 			if (r2 >= nregs)
9745 				err += efunc(pc, "invalid register %u\n", r2);
9746 			if (rd != 0)
9747 				err += efunc(pc, "non-zero reserved bits\n");
9748 			break;
9749 		case DIF_OP_TST:
9750 			if (r1 >= nregs)
9751 				err += efunc(pc, "invalid register %u\n", r1);
9752 			if (r2 != 0 || rd != 0)
9753 				err += efunc(pc, "non-zero reserved bits\n");
9754 			break;
9755 		case DIF_OP_BA:
9756 		case DIF_OP_BE:
9757 		case DIF_OP_BNE:
9758 		case DIF_OP_BG:
9759 		case DIF_OP_BGU:
9760 		case DIF_OP_BGE:
9761 		case DIF_OP_BGEU:
9762 		case DIF_OP_BL:
9763 		case DIF_OP_BLU:
9764 		case DIF_OP_BLE:
9765 		case DIF_OP_BLEU:
9766 			if (label >= dp->dtdo_len) {
9767 				err += efunc(pc, "invalid branch target %u\n",
9768 				    label);
9769 			}
9770 			if (label <= pc) {
9771 				err += efunc(pc, "backward branch to %u\n",
9772 				    label);
9773 			}
9774 			break;
9775 		case DIF_OP_RET:
9776 			if (r1 != 0 || r2 != 0)
9777 				err += efunc(pc, "non-zero reserved bits\n");
9778 			if (rd >= nregs)
9779 				err += efunc(pc, "invalid register %u\n", rd);
9780 			break;
9781 		case DIF_OP_NOP:
9782 		case DIF_OP_POPTS:
9783 		case DIF_OP_FLUSHTS:
9784 			if (r1 != 0 || r2 != 0 || rd != 0)
9785 				err += efunc(pc, "non-zero reserved bits\n");
9786 			break;
9787 		case DIF_OP_SETX:
9788 			if (DIF_INSTR_INTEGER(instr) >= dp->dtdo_intlen) {
9789 				err += efunc(pc, "invalid integer ref %u\n",
9790 				    DIF_INSTR_INTEGER(instr));
9791 			}
9792 			if (rd >= nregs)
9793 				err += efunc(pc, "invalid register %u\n", rd);
9794 			if (rd == 0)
9795 				err += efunc(pc, "cannot write to %r0\n");
9796 			break;
9797 		case DIF_OP_SETS:
9798 			if (DIF_INSTR_STRING(instr) >= dp->dtdo_strlen) {
9799 				err += efunc(pc, "invalid string ref %u\n",
9800 				    DIF_INSTR_STRING(instr));
9801 			}
9802 			if (rd >= nregs)
9803 				err += efunc(pc, "invalid register %u\n", rd);
9804 			if (rd == 0)
9805 				err += efunc(pc, "cannot write to %r0\n");
9806 			break;
9807 		case DIF_OP_LDGA:
9808 		case DIF_OP_LDTA:
9809 			if (r1 > DIF_VAR_ARRAY_MAX)
9810 				err += efunc(pc, "invalid array %u\n", r1);
9811 			if (r2 >= nregs)
9812 				err += efunc(pc, "invalid register %u\n", r2);
9813 			if (rd >= nregs)
9814 				err += efunc(pc, "invalid register %u\n", rd);
9815 			if (rd == 0)
9816 				err += efunc(pc, "cannot write to %r0\n");
9817 			break;
9818 		case DIF_OP_LDGS:
9819 		case DIF_OP_LDTS:
9820 		case DIF_OP_LDLS:
9821 		case DIF_OP_LDGAA:
9822 		case DIF_OP_LDTAA:
9823 			if (v < DIF_VAR_OTHER_MIN || v > DIF_VAR_OTHER_MAX)
9824 				err += efunc(pc, "invalid variable %u\n", v);
9825 			if (rd >= nregs)
9826 				err += efunc(pc, "invalid register %u\n", rd);
9827 			if (rd == 0)
9828 				err += efunc(pc, "cannot write to %r0\n");
9829 			break;
9830 		case DIF_OP_STGS:
9831 		case DIF_OP_STTS:
9832 		case DIF_OP_STLS:
9833 		case DIF_OP_STGAA:
9834 		case DIF_OP_STTAA:
9835 			if (v < DIF_VAR_OTHER_UBASE || v > DIF_VAR_OTHER_MAX)
9836 				err += efunc(pc, "invalid variable %u\n", v);
9837 			if (rs >= nregs)
9838 				err += efunc(pc, "invalid register %u\n", rd);
9839 			break;
9840 		case DIF_OP_CALL:
9841 			if (subr > DIF_SUBR_MAX)
9842 				err += efunc(pc, "invalid subr %u\n", subr);
9843 			if (rd >= nregs)
9844 				err += efunc(pc, "invalid register %u\n", rd);
9845 			if (rd == 0)
9846 				err += efunc(pc, "cannot write to %r0\n");
9847 
9848 			if (subr == DIF_SUBR_COPYOUT ||
9849 			    subr == DIF_SUBR_COPYOUTSTR) {
9850 				dp->dtdo_destructive = 1;
9851 			}
9852 
9853 			if (subr == DIF_SUBR_GETF) {
9854 				/*
9855 				 * If we have a getf() we need to record that
9856 				 * in our state.  Note that our state can be
9857 				 * NULL if this is a helper -- but in that
9858 				 * case, the call to getf() is itself illegal,
9859 				 * and will be caught (slightly later) when
9860 				 * the helper is validated.
9861 				 */
9862 				if (vstate->dtvs_state != NULL)
9863 					vstate->dtvs_state->dts_getf++;
9864 			}
9865 
9866 			break;
9867 		case DIF_OP_PUSHTR:
9868 			if (type != DIF_TYPE_STRING && type != DIF_TYPE_CTF)
9869 				err += efunc(pc, "invalid ref type %u\n", type);
9870 			if (r2 >= nregs)
9871 				err += efunc(pc, "invalid register %u\n", r2);
9872 			if (rs >= nregs)
9873 				err += efunc(pc, "invalid register %u\n", rs);
9874 			break;
9875 		case DIF_OP_PUSHTV:
9876 			if (type != DIF_TYPE_CTF)
9877 				err += efunc(pc, "invalid val type %u\n", type);
9878 			if (r2 >= nregs)
9879 				err += efunc(pc, "invalid register %u\n", r2);
9880 			if (rs >= nregs)
9881 				err += efunc(pc, "invalid register %u\n", rs);
9882 			break;
9883 		default:
9884 			err += efunc(pc, "invalid opcode %u\n",
9885 			    DIF_INSTR_OP(instr));
9886 		}
9887 	}
9888 
9889 	if (dp->dtdo_len != 0 &&
9890 	    DIF_INSTR_OP(dp->dtdo_buf[dp->dtdo_len - 1]) != DIF_OP_RET) {
9891 		err += efunc(dp->dtdo_len - 1,
9892 		    "expected 'ret' as last DIF instruction\n");
9893 	}
9894 
9895 	if (!(dp->dtdo_rtype.dtdt_flags & (DIF_TF_BYREF | DIF_TF_BYUREF))) {
9896 		/*
9897 		 * If we're not returning by reference, the size must be either
9898 		 * 0 or the size of one of the base types.
9899 		 */
9900 		switch (dp->dtdo_rtype.dtdt_size) {
9901 		case 0:
9902 		case sizeof (uint8_t):
9903 		case sizeof (uint16_t):
9904 		case sizeof (uint32_t):
9905 		case sizeof (uint64_t):
9906 			break;
9907 
9908 		default:
9909 			err += efunc(dp->dtdo_len - 1, "bad return size\n");
9910 		}
9911 	}
9912 
9913 	for (i = 0; i < dp->dtdo_varlen && err == 0; i++) {
9914 		dtrace_difv_t *v = &dp->dtdo_vartab[i], *existing = NULL;
9915 		dtrace_diftype_t *vt, *et;
9916 		uint_t id, ndx;
9917 
9918 		if (v->dtdv_scope != DIFV_SCOPE_GLOBAL &&
9919 		    v->dtdv_scope != DIFV_SCOPE_THREAD &&
9920 		    v->dtdv_scope != DIFV_SCOPE_LOCAL) {
9921 			err += efunc(i, "unrecognized variable scope %d\n",
9922 			    v->dtdv_scope);
9923 			break;
9924 		}
9925 
9926 		if (v->dtdv_kind != DIFV_KIND_ARRAY &&
9927 		    v->dtdv_kind != DIFV_KIND_SCALAR) {
9928 			err += efunc(i, "unrecognized variable type %d\n",
9929 			    v->dtdv_kind);
9930 			break;
9931 		}
9932 
9933 		if ((id = v->dtdv_id) > DIF_VARIABLE_MAX) {
9934 			err += efunc(i, "%d exceeds variable id limit\n", id);
9935 			break;
9936 		}
9937 
9938 		if (id < DIF_VAR_OTHER_UBASE)
9939 			continue;
9940 
9941 		/*
9942 		 * For user-defined variables, we need to check that this
9943 		 * definition is identical to any previous definition that we
9944 		 * encountered.
9945 		 */
9946 		ndx = id - DIF_VAR_OTHER_UBASE;
9947 
9948 		switch (v->dtdv_scope) {
9949 		case DIFV_SCOPE_GLOBAL:
9950 			if (maxglobal == -1 || ndx > maxglobal)
9951 				maxglobal = ndx;
9952 
9953 			if (ndx < vstate->dtvs_nglobals) {
9954 				dtrace_statvar_t *svar;
9955 
9956 				if ((svar = vstate->dtvs_globals[ndx]) != NULL)
9957 					existing = &svar->dtsv_var;
9958 			}
9959 
9960 			break;
9961 
9962 		case DIFV_SCOPE_THREAD:
9963 			if (maxtlocal == -1 || ndx > maxtlocal)
9964 				maxtlocal = ndx;
9965 
9966 			if (ndx < vstate->dtvs_ntlocals)
9967 				existing = &vstate->dtvs_tlocals[ndx];
9968 			break;
9969 
9970 		case DIFV_SCOPE_LOCAL:
9971 			if (maxlocal == -1 || ndx > maxlocal)
9972 				maxlocal = ndx;
9973 
9974 			if (ndx < vstate->dtvs_nlocals) {
9975 				dtrace_statvar_t *svar;
9976 
9977 				if ((svar = vstate->dtvs_locals[ndx]) != NULL)
9978 					existing = &svar->dtsv_var;
9979 			}
9980 
9981 			break;
9982 		}
9983 
9984 		vt = &v->dtdv_type;
9985 
9986 		if (vt->dtdt_flags & DIF_TF_BYREF) {
9987 			if (vt->dtdt_size == 0) {
9988 				err += efunc(i, "zero-sized variable\n");
9989 				break;
9990 			}
9991 
9992 			if ((v->dtdv_scope == DIFV_SCOPE_GLOBAL ||
9993 			    v->dtdv_scope == DIFV_SCOPE_LOCAL) &&
9994 			    vt->dtdt_size > dtrace_statvar_maxsize) {
9995 				err += efunc(i, "oversized by-ref static\n");
9996 				break;
9997 			}
9998 		}
9999 
10000 		if (existing == NULL || existing->dtdv_id == 0)
10001 			continue;
10002 
10003 		ASSERT(existing->dtdv_id == v->dtdv_id);
10004 		ASSERT(existing->dtdv_scope == v->dtdv_scope);
10005 
10006 		if (existing->dtdv_kind != v->dtdv_kind)
10007 			err += efunc(i, "%d changed variable kind\n", id);
10008 
10009 		et = &existing->dtdv_type;
10010 
10011 		if (vt->dtdt_flags != et->dtdt_flags) {
10012 			err += efunc(i, "%d changed variable type flags\n", id);
10013 			break;
10014 		}
10015 
10016 		if (vt->dtdt_size != 0 && vt->dtdt_size != et->dtdt_size) {
10017 			err += efunc(i, "%d changed variable type size\n", id);
10018 			break;
10019 		}
10020 	}
10021 
10022 	for (pc = 0; pc < dp->dtdo_len && err == 0; pc++) {
10023 		dif_instr_t instr = dp->dtdo_buf[pc];
10024 
10025 		uint_t v = DIF_INSTR_VAR(instr);
10026 		uint_t op = DIF_INSTR_OP(instr);
10027 
10028 		switch (op) {
10029 		case DIF_OP_LDGS:
10030 		case DIF_OP_LDGAA:
10031 		case DIF_OP_STGS:
10032 		case DIF_OP_STGAA:
10033 			if (v > DIF_VAR_OTHER_UBASE + maxglobal)
10034 				err += efunc(pc, "invalid variable %u\n", v);
10035 			break;
10036 		case DIF_OP_LDTS:
10037 		case DIF_OP_LDTAA:
10038 		case DIF_OP_STTS:
10039 		case DIF_OP_STTAA:
10040 			if (v > DIF_VAR_OTHER_UBASE + maxtlocal)
10041 				err += efunc(pc, "invalid variable %u\n", v);
10042 			break;
10043 		case DIF_OP_LDLS:
10044 		case DIF_OP_STLS:
10045 			if (v > DIF_VAR_OTHER_UBASE + maxlocal)
10046 				err += efunc(pc, "invalid variable %u\n", v);
10047 			break;
10048 		default:
10049 			break;
10050 		}
10051 	}
10052 
10053 	return (err);
10054 }
10055 
10056 /*
10057  * Validate a DTrace DIF object that it is to be used as a helper.  Helpers
10058  * are much more constrained than normal DIFOs.  Specifically, they may
10059  * not:
10060  *
10061  * 1. Make calls to subroutines other than copyin(), copyinstr() or
10062  *    miscellaneous string routines
10063  * 2. Access DTrace variables other than the args[] array, and the
10064  *    curthread, pid, ppid, tid, execname, zonename, uid and gid variables.
10065  * 3. Have thread-local variables.
10066  * 4. Have dynamic variables.
10067  */
10068 static int
10069 dtrace_difo_validate_helper(dtrace_difo_t *dp)
10070 {
10071 	int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
10072 	int err = 0;
10073 	uint_t pc;
10074 
10075 	for (pc = 0; pc < dp->dtdo_len; pc++) {
10076 		dif_instr_t instr = dp->dtdo_buf[pc];
10077 
10078 		uint_t v = DIF_INSTR_VAR(instr);
10079 		uint_t subr = DIF_INSTR_SUBR(instr);
10080 		uint_t op = DIF_INSTR_OP(instr);
10081 
10082 		switch (op) {
10083 		case DIF_OP_OR:
10084 		case DIF_OP_XOR:
10085 		case DIF_OP_AND:
10086 		case DIF_OP_SLL:
10087 		case DIF_OP_SRL:
10088 		case DIF_OP_SRA:
10089 		case DIF_OP_SUB:
10090 		case DIF_OP_ADD:
10091 		case DIF_OP_MUL:
10092 		case DIF_OP_SDIV:
10093 		case DIF_OP_UDIV:
10094 		case DIF_OP_SREM:
10095 		case DIF_OP_UREM:
10096 		case DIF_OP_COPYS:
10097 		case DIF_OP_NOT:
10098 		case DIF_OP_MOV:
10099 		case DIF_OP_RLDSB:
10100 		case DIF_OP_RLDSH:
10101 		case DIF_OP_RLDSW:
10102 		case DIF_OP_RLDUB:
10103 		case DIF_OP_RLDUH:
10104 		case DIF_OP_RLDUW:
10105 		case DIF_OP_RLDX:
10106 		case DIF_OP_ULDSB:
10107 		case DIF_OP_ULDSH:
10108 		case DIF_OP_ULDSW:
10109 		case DIF_OP_ULDUB:
10110 		case DIF_OP_ULDUH:
10111 		case DIF_OP_ULDUW:
10112 		case DIF_OP_ULDX:
10113 		case DIF_OP_STB:
10114 		case DIF_OP_STH:
10115 		case DIF_OP_STW:
10116 		case DIF_OP_STX:
10117 		case DIF_OP_ALLOCS:
10118 		case DIF_OP_CMP:
10119 		case DIF_OP_SCMP:
10120 		case DIF_OP_TST:
10121 		case DIF_OP_BA:
10122 		case DIF_OP_BE:
10123 		case DIF_OP_BNE:
10124 		case DIF_OP_BG:
10125 		case DIF_OP_BGU:
10126 		case DIF_OP_BGE:
10127 		case DIF_OP_BGEU:
10128 		case DIF_OP_BL:
10129 		case DIF_OP_BLU:
10130 		case DIF_OP_BLE:
10131 		case DIF_OP_BLEU:
10132 		case DIF_OP_RET:
10133 		case DIF_OP_NOP:
10134 		case DIF_OP_POPTS:
10135 		case DIF_OP_FLUSHTS:
10136 		case DIF_OP_SETX:
10137 		case DIF_OP_SETS:
10138 		case DIF_OP_LDGA:
10139 		case DIF_OP_LDLS:
10140 		case DIF_OP_STGS:
10141 		case DIF_OP_STLS:
10142 		case DIF_OP_PUSHTR:
10143 		case DIF_OP_PUSHTV:
10144 			break;
10145 
10146 		case DIF_OP_LDGS:
10147 			if (v >= DIF_VAR_OTHER_UBASE)
10148 				break;
10149 
10150 			if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9)
10151 				break;
10152 
10153 			if (v == DIF_VAR_CURTHREAD || v == DIF_VAR_PID ||
10154 			    v == DIF_VAR_PPID || v == DIF_VAR_TID ||
10155 			    v == DIF_VAR_EXECARGS ||
10156 			    v == DIF_VAR_EXECNAME || v == DIF_VAR_ZONENAME ||
10157 			    v == DIF_VAR_UID || v == DIF_VAR_GID)
10158 				break;
10159 
10160 			err += efunc(pc, "illegal variable %u\n", v);
10161 			break;
10162 
10163 		case DIF_OP_LDTA:
10164 		case DIF_OP_LDTS:
10165 		case DIF_OP_LDGAA:
10166 		case DIF_OP_LDTAA:
10167 			err += efunc(pc, "illegal dynamic variable load\n");
10168 			break;
10169 
10170 		case DIF_OP_STTS:
10171 		case DIF_OP_STGAA:
10172 		case DIF_OP_STTAA:
10173 			err += efunc(pc, "illegal dynamic variable store\n");
10174 			break;
10175 
10176 		case DIF_OP_CALL:
10177 			if (subr == DIF_SUBR_ALLOCA ||
10178 			    subr == DIF_SUBR_BCOPY ||
10179 			    subr == DIF_SUBR_COPYIN ||
10180 			    subr == DIF_SUBR_COPYINTO ||
10181 			    subr == DIF_SUBR_COPYINSTR ||
10182 			    subr == DIF_SUBR_INDEX ||
10183 			    subr == DIF_SUBR_INET_NTOA ||
10184 			    subr == DIF_SUBR_INET_NTOA6 ||
10185 			    subr == DIF_SUBR_INET_NTOP ||
10186 			    subr == DIF_SUBR_JSON ||
10187 			    subr == DIF_SUBR_LLTOSTR ||
10188 			    subr == DIF_SUBR_STRTOLL ||
10189 			    subr == DIF_SUBR_RINDEX ||
10190 			    subr == DIF_SUBR_STRCHR ||
10191 			    subr == DIF_SUBR_STRJOIN ||
10192 			    subr == DIF_SUBR_STRRCHR ||
10193 			    subr == DIF_SUBR_STRSTR ||
10194 			    subr == DIF_SUBR_HTONS ||
10195 			    subr == DIF_SUBR_HTONL ||
10196 			    subr == DIF_SUBR_HTONLL ||
10197 			    subr == DIF_SUBR_NTOHS ||
10198 			    subr == DIF_SUBR_NTOHL ||
10199 			    subr == DIF_SUBR_NTOHLL ||
10200 			    subr == DIF_SUBR_MEMREF ||
10201 #ifndef illumos
10202 			    subr == DIF_SUBR_MEMSTR ||
10203 #endif
10204 			    subr == DIF_SUBR_TYPEREF)
10205 				break;
10206 
10207 			err += efunc(pc, "invalid subr %u\n", subr);
10208 			break;
10209 
10210 		default:
10211 			err += efunc(pc, "invalid opcode %u\n",
10212 			    DIF_INSTR_OP(instr));
10213 		}
10214 	}
10215 
10216 	return (err);
10217 }
10218 
10219 /*
10220  * Returns 1 if the expression in the DIF object can be cached on a per-thread
10221  * basis; 0 if not.
10222  */
10223 static int
10224 dtrace_difo_cacheable(dtrace_difo_t *dp)
10225 {
10226 	int i;
10227 
10228 	if (dp == NULL)
10229 		return (0);
10230 
10231 	for (i = 0; i < dp->dtdo_varlen; i++) {
10232 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
10233 
10234 		if (v->dtdv_scope != DIFV_SCOPE_GLOBAL)
10235 			continue;
10236 
10237 		switch (v->dtdv_id) {
10238 		case DIF_VAR_CURTHREAD:
10239 		case DIF_VAR_PID:
10240 		case DIF_VAR_TID:
10241 		case DIF_VAR_EXECARGS:
10242 		case DIF_VAR_EXECNAME:
10243 		case DIF_VAR_ZONENAME:
10244 			break;
10245 
10246 		default:
10247 			return (0);
10248 		}
10249 	}
10250 
10251 	/*
10252 	 * This DIF object may be cacheable.  Now we need to look for any
10253 	 * array loading instructions, any memory loading instructions, or
10254 	 * any stores to thread-local variables.
10255 	 */
10256 	for (i = 0; i < dp->dtdo_len; i++) {
10257 		uint_t op = DIF_INSTR_OP(dp->dtdo_buf[i]);
10258 
10259 		if ((op >= DIF_OP_LDSB && op <= DIF_OP_LDX) ||
10260 		    (op >= DIF_OP_ULDSB && op <= DIF_OP_ULDX) ||
10261 		    (op >= DIF_OP_RLDSB && op <= DIF_OP_RLDX) ||
10262 		    op == DIF_OP_LDGA || op == DIF_OP_STTS)
10263 			return (0);
10264 	}
10265 
10266 	return (1);
10267 }
10268 
10269 static void
10270 dtrace_difo_hold(dtrace_difo_t *dp)
10271 {
10272 	int i;
10273 
10274 	ASSERT(MUTEX_HELD(&dtrace_lock));
10275 
10276 	dp->dtdo_refcnt++;
10277 	ASSERT(dp->dtdo_refcnt != 0);
10278 
10279 	/*
10280 	 * We need to check this DIF object for references to the variable
10281 	 * DIF_VAR_VTIMESTAMP.
10282 	 */
10283 	for (i = 0; i < dp->dtdo_varlen; i++) {
10284 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
10285 
10286 		if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
10287 			continue;
10288 
10289 		if (dtrace_vtime_references++ == 0)
10290 			dtrace_vtime_enable();
10291 	}
10292 }
10293 
10294 /*
10295  * This routine calculates the dynamic variable chunksize for a given DIF
10296  * object.  The calculation is not fool-proof, and can probably be tricked by
10297  * malicious DIF -- but it works for all compiler-generated DIF.  Because this
10298  * calculation is likely imperfect, dtrace_dynvar() is able to gracefully fail
10299  * if a dynamic variable size exceeds the chunksize.
10300  */
10301 static void
10302 dtrace_difo_chunksize(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
10303 {
10304 	uint64_t sval = 0;
10305 	dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
10306 	const dif_instr_t *text = dp->dtdo_buf;
10307 	uint_t pc, srd = 0;
10308 	uint_t ttop = 0;
10309 	size_t size, ksize;
10310 	uint_t id, i;
10311 
10312 	for (pc = 0; pc < dp->dtdo_len; pc++) {
10313 		dif_instr_t instr = text[pc];
10314 		uint_t op = DIF_INSTR_OP(instr);
10315 		uint_t rd = DIF_INSTR_RD(instr);
10316 		uint_t r1 = DIF_INSTR_R1(instr);
10317 		uint_t nkeys = 0;
10318 		uchar_t scope = 0;
10319 
10320 		dtrace_key_t *key = tupregs;
10321 
10322 		switch (op) {
10323 		case DIF_OP_SETX:
10324 			sval = dp->dtdo_inttab[DIF_INSTR_INTEGER(instr)];
10325 			srd = rd;
10326 			continue;
10327 
10328 		case DIF_OP_STTS:
10329 			key = &tupregs[DIF_DTR_NREGS];
10330 			key[0].dttk_size = 0;
10331 			key[1].dttk_size = 0;
10332 			nkeys = 2;
10333 			scope = DIFV_SCOPE_THREAD;
10334 			break;
10335 
10336 		case DIF_OP_STGAA:
10337 		case DIF_OP_STTAA:
10338 			nkeys = ttop;
10339 
10340 			if (DIF_INSTR_OP(instr) == DIF_OP_STTAA)
10341 				key[nkeys++].dttk_size = 0;
10342 
10343 			key[nkeys++].dttk_size = 0;
10344 
10345 			if (op == DIF_OP_STTAA) {
10346 				scope = DIFV_SCOPE_THREAD;
10347 			} else {
10348 				scope = DIFV_SCOPE_GLOBAL;
10349 			}
10350 
10351 			break;
10352 
10353 		case DIF_OP_PUSHTR:
10354 			if (ttop == DIF_DTR_NREGS)
10355 				return;
10356 
10357 			if ((srd == 0 || sval == 0) && r1 == DIF_TYPE_STRING) {
10358 				/*
10359 				 * If the register for the size of the "pushtr"
10360 				 * is %r0 (or the value is 0) and the type is
10361 				 * a string, we'll use the system-wide default
10362 				 * string size.
10363 				 */
10364 				tupregs[ttop++].dttk_size =
10365 				    dtrace_strsize_default;
10366 			} else {
10367 				if (srd == 0)
10368 					return;
10369 
10370 				if (sval > LONG_MAX)
10371 					return;
10372 
10373 				tupregs[ttop++].dttk_size = sval;
10374 			}
10375 
10376 			break;
10377 
10378 		case DIF_OP_PUSHTV:
10379 			if (ttop == DIF_DTR_NREGS)
10380 				return;
10381 
10382 			tupregs[ttop++].dttk_size = 0;
10383 			break;
10384 
10385 		case DIF_OP_FLUSHTS:
10386 			ttop = 0;
10387 			break;
10388 
10389 		case DIF_OP_POPTS:
10390 			if (ttop != 0)
10391 				ttop--;
10392 			break;
10393 		}
10394 
10395 		sval = 0;
10396 		srd = 0;
10397 
10398 		if (nkeys == 0)
10399 			continue;
10400 
10401 		/*
10402 		 * We have a dynamic variable allocation; calculate its size.
10403 		 */
10404 		for (ksize = 0, i = 0; i < nkeys; i++)
10405 			ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
10406 
10407 		size = sizeof (dtrace_dynvar_t);
10408 		size += sizeof (dtrace_key_t) * (nkeys - 1);
10409 		size += ksize;
10410 
10411 		/*
10412 		 * Now we need to determine the size of the stored data.
10413 		 */
10414 		id = DIF_INSTR_VAR(instr);
10415 
10416 		for (i = 0; i < dp->dtdo_varlen; i++) {
10417 			dtrace_difv_t *v = &dp->dtdo_vartab[i];
10418 
10419 			if (v->dtdv_id == id && v->dtdv_scope == scope) {
10420 				size += v->dtdv_type.dtdt_size;
10421 				break;
10422 			}
10423 		}
10424 
10425 		if (i == dp->dtdo_varlen)
10426 			return;
10427 
10428 		/*
10429 		 * We have the size.  If this is larger than the chunk size
10430 		 * for our dynamic variable state, reset the chunk size.
10431 		 */
10432 		size = P2ROUNDUP(size, sizeof (uint64_t));
10433 
10434 		/*
10435 		 * Before setting the chunk size, check that we're not going
10436 		 * to set it to a negative value...
10437 		 */
10438 		if (size > LONG_MAX)
10439 			return;
10440 
10441 		/*
10442 		 * ...and make certain that we didn't badly overflow.
10443 		 */
10444 		if (size < ksize || size < sizeof (dtrace_dynvar_t))
10445 			return;
10446 
10447 		if (size > vstate->dtvs_dynvars.dtds_chunksize)
10448 			vstate->dtvs_dynvars.dtds_chunksize = size;
10449 	}
10450 }
10451 
10452 static void
10453 dtrace_difo_init(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
10454 {
10455 	int i, oldsvars, osz, nsz, otlocals, ntlocals;
10456 	uint_t id;
10457 
10458 	ASSERT(MUTEX_HELD(&dtrace_lock));
10459 	ASSERT(dp->dtdo_buf != NULL && dp->dtdo_len != 0);
10460 
10461 	for (i = 0; i < dp->dtdo_varlen; i++) {
10462 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
10463 		dtrace_statvar_t *svar, ***svarp = NULL;
10464 		size_t dsize = 0;
10465 		uint8_t scope = v->dtdv_scope;
10466 		int *np = NULL;
10467 
10468 		if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
10469 			continue;
10470 
10471 		id -= DIF_VAR_OTHER_UBASE;
10472 
10473 		switch (scope) {
10474 		case DIFV_SCOPE_THREAD:
10475 			while (id >= (otlocals = vstate->dtvs_ntlocals)) {
10476 				dtrace_difv_t *tlocals;
10477 
10478 				if ((ntlocals = (otlocals << 1)) == 0)
10479 					ntlocals = 1;
10480 
10481 				osz = otlocals * sizeof (dtrace_difv_t);
10482 				nsz = ntlocals * sizeof (dtrace_difv_t);
10483 
10484 				tlocals = kmem_zalloc(nsz, KM_SLEEP);
10485 
10486 				if (osz != 0) {
10487 					bcopy(vstate->dtvs_tlocals,
10488 					    tlocals, osz);
10489 					kmem_free(vstate->dtvs_tlocals, osz);
10490 				}
10491 
10492 				vstate->dtvs_tlocals = tlocals;
10493 				vstate->dtvs_ntlocals = ntlocals;
10494 			}
10495 
10496 			vstate->dtvs_tlocals[id] = *v;
10497 			continue;
10498 
10499 		case DIFV_SCOPE_LOCAL:
10500 			np = &vstate->dtvs_nlocals;
10501 			svarp = &vstate->dtvs_locals;
10502 
10503 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
10504 				dsize = NCPU * (v->dtdv_type.dtdt_size +
10505 				    sizeof (uint64_t));
10506 			else
10507 				dsize = NCPU * sizeof (uint64_t);
10508 
10509 			break;
10510 
10511 		case DIFV_SCOPE_GLOBAL:
10512 			np = &vstate->dtvs_nglobals;
10513 			svarp = &vstate->dtvs_globals;
10514 
10515 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
10516 				dsize = v->dtdv_type.dtdt_size +
10517 				    sizeof (uint64_t);
10518 
10519 			break;
10520 
10521 		default:
10522 			ASSERT(0);
10523 		}
10524 
10525 		while (id >= (oldsvars = *np)) {
10526 			dtrace_statvar_t **statics;
10527 			int newsvars, oldsize, newsize;
10528 
10529 			if ((newsvars = (oldsvars << 1)) == 0)
10530 				newsvars = 1;
10531 
10532 			oldsize = oldsvars * sizeof (dtrace_statvar_t *);
10533 			newsize = newsvars * sizeof (dtrace_statvar_t *);
10534 
10535 			statics = kmem_zalloc(newsize, KM_SLEEP);
10536 
10537 			if (oldsize != 0) {
10538 				bcopy(*svarp, statics, oldsize);
10539 				kmem_free(*svarp, oldsize);
10540 			}
10541 
10542 			*svarp = statics;
10543 			*np = newsvars;
10544 		}
10545 
10546 		if ((svar = (*svarp)[id]) == NULL) {
10547 			svar = kmem_zalloc(sizeof (dtrace_statvar_t), KM_SLEEP);
10548 			svar->dtsv_var = *v;
10549 
10550 			if ((svar->dtsv_size = dsize) != 0) {
10551 				svar->dtsv_data = (uint64_t)(uintptr_t)
10552 				    kmem_zalloc(dsize, KM_SLEEP);
10553 			}
10554 
10555 			(*svarp)[id] = svar;
10556 		}
10557 
10558 		svar->dtsv_refcnt++;
10559 	}
10560 
10561 	dtrace_difo_chunksize(dp, vstate);
10562 	dtrace_difo_hold(dp);
10563 }
10564 
10565 static dtrace_difo_t *
10566 dtrace_difo_duplicate(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
10567 {
10568 	dtrace_difo_t *new;
10569 	size_t sz;
10570 
10571 	ASSERT(dp->dtdo_buf != NULL);
10572 	ASSERT(dp->dtdo_refcnt != 0);
10573 
10574 	new = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
10575 
10576 	ASSERT(dp->dtdo_buf != NULL);
10577 	sz = dp->dtdo_len * sizeof (dif_instr_t);
10578 	new->dtdo_buf = kmem_alloc(sz, KM_SLEEP);
10579 	bcopy(dp->dtdo_buf, new->dtdo_buf, sz);
10580 	new->dtdo_len = dp->dtdo_len;
10581 
10582 	if (dp->dtdo_strtab != NULL) {
10583 		ASSERT(dp->dtdo_strlen != 0);
10584 		new->dtdo_strtab = kmem_alloc(dp->dtdo_strlen, KM_SLEEP);
10585 		bcopy(dp->dtdo_strtab, new->dtdo_strtab, dp->dtdo_strlen);
10586 		new->dtdo_strlen = dp->dtdo_strlen;
10587 	}
10588 
10589 	if (dp->dtdo_inttab != NULL) {
10590 		ASSERT(dp->dtdo_intlen != 0);
10591 		sz = dp->dtdo_intlen * sizeof (uint64_t);
10592 		new->dtdo_inttab = kmem_alloc(sz, KM_SLEEP);
10593 		bcopy(dp->dtdo_inttab, new->dtdo_inttab, sz);
10594 		new->dtdo_intlen = dp->dtdo_intlen;
10595 	}
10596 
10597 	if (dp->dtdo_vartab != NULL) {
10598 		ASSERT(dp->dtdo_varlen != 0);
10599 		sz = dp->dtdo_varlen * sizeof (dtrace_difv_t);
10600 		new->dtdo_vartab = kmem_alloc(sz, KM_SLEEP);
10601 		bcopy(dp->dtdo_vartab, new->dtdo_vartab, sz);
10602 		new->dtdo_varlen = dp->dtdo_varlen;
10603 	}
10604 
10605 	dtrace_difo_init(new, vstate);
10606 	return (new);
10607 }
10608 
10609 static void
10610 dtrace_difo_destroy(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
10611 {
10612 	int i;
10613 
10614 	ASSERT(dp->dtdo_refcnt == 0);
10615 
10616 	for (i = 0; i < dp->dtdo_varlen; i++) {
10617 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
10618 		dtrace_statvar_t *svar, **svarp = NULL;
10619 		uint_t id;
10620 		uint8_t scope = v->dtdv_scope;
10621 		int *np = NULL;
10622 
10623 		switch (scope) {
10624 		case DIFV_SCOPE_THREAD:
10625 			continue;
10626 
10627 		case DIFV_SCOPE_LOCAL:
10628 			np = &vstate->dtvs_nlocals;
10629 			svarp = vstate->dtvs_locals;
10630 			break;
10631 
10632 		case DIFV_SCOPE_GLOBAL:
10633 			np = &vstate->dtvs_nglobals;
10634 			svarp = vstate->dtvs_globals;
10635 			break;
10636 
10637 		default:
10638 			ASSERT(0);
10639 		}
10640 
10641 		if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
10642 			continue;
10643 
10644 		id -= DIF_VAR_OTHER_UBASE;
10645 		ASSERT(id < *np);
10646 
10647 		svar = svarp[id];
10648 		ASSERT(svar != NULL);
10649 		ASSERT(svar->dtsv_refcnt > 0);
10650 
10651 		if (--svar->dtsv_refcnt > 0)
10652 			continue;
10653 
10654 		if (svar->dtsv_size != 0) {
10655 			ASSERT(svar->dtsv_data != 0);
10656 			kmem_free((void *)(uintptr_t)svar->dtsv_data,
10657 			    svar->dtsv_size);
10658 		}
10659 
10660 		kmem_free(svar, sizeof (dtrace_statvar_t));
10661 		svarp[id] = NULL;
10662 	}
10663 
10664 	if (dp->dtdo_buf != NULL)
10665 		kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
10666 	if (dp->dtdo_inttab != NULL)
10667 		kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
10668 	if (dp->dtdo_strtab != NULL)
10669 		kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
10670 	if (dp->dtdo_vartab != NULL)
10671 		kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
10672 
10673 	kmem_free(dp, sizeof (dtrace_difo_t));
10674 }
10675 
10676 static void
10677 dtrace_difo_release(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
10678 {
10679 	int i;
10680 
10681 	ASSERT(MUTEX_HELD(&dtrace_lock));
10682 	ASSERT(dp->dtdo_refcnt != 0);
10683 
10684 	for (i = 0; i < dp->dtdo_varlen; i++) {
10685 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
10686 
10687 		if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
10688 			continue;
10689 
10690 		ASSERT(dtrace_vtime_references > 0);
10691 		if (--dtrace_vtime_references == 0)
10692 			dtrace_vtime_disable();
10693 	}
10694 
10695 	if (--dp->dtdo_refcnt == 0)
10696 		dtrace_difo_destroy(dp, vstate);
10697 }
10698 
10699 /*
10700  * DTrace Format Functions
10701  */
10702 static uint16_t
10703 dtrace_format_add(dtrace_state_t *state, char *str)
10704 {
10705 	char *fmt, **new;
10706 	uint16_t ndx, len = strlen(str) + 1;
10707 
10708 	fmt = kmem_zalloc(len, KM_SLEEP);
10709 	bcopy(str, fmt, len);
10710 
10711 	for (ndx = 0; ndx < state->dts_nformats; ndx++) {
10712 		if (state->dts_formats[ndx] == NULL) {
10713 			state->dts_formats[ndx] = fmt;
10714 			return (ndx + 1);
10715 		}
10716 	}
10717 
10718 	if (state->dts_nformats == USHRT_MAX) {
10719 		/*
10720 		 * This is only likely if a denial-of-service attack is being
10721 		 * attempted.  As such, it's okay to fail silently here.
10722 		 */
10723 		kmem_free(fmt, len);
10724 		return (0);
10725 	}
10726 
10727 	/*
10728 	 * For simplicity, we always resize the formats array to be exactly the
10729 	 * number of formats.
10730 	 */
10731 	ndx = state->dts_nformats++;
10732 	new = kmem_alloc((ndx + 1) * sizeof (char *), KM_SLEEP);
10733 
10734 	if (state->dts_formats != NULL) {
10735 		ASSERT(ndx != 0);
10736 		bcopy(state->dts_formats, new, ndx * sizeof (char *));
10737 		kmem_free(state->dts_formats, ndx * sizeof (char *));
10738 	}
10739 
10740 	state->dts_formats = new;
10741 	state->dts_formats[ndx] = fmt;
10742 
10743 	return (ndx + 1);
10744 }
10745 
10746 static void
10747 dtrace_format_remove(dtrace_state_t *state, uint16_t format)
10748 {
10749 	char *fmt;
10750 
10751 	ASSERT(state->dts_formats != NULL);
10752 	ASSERT(format <= state->dts_nformats);
10753 	ASSERT(state->dts_formats[format - 1] != NULL);
10754 
10755 	fmt = state->dts_formats[format - 1];
10756 	kmem_free(fmt, strlen(fmt) + 1);
10757 	state->dts_formats[format - 1] = NULL;
10758 }
10759 
10760 static void
10761 dtrace_format_destroy(dtrace_state_t *state)
10762 {
10763 	int i;
10764 
10765 	if (state->dts_nformats == 0) {
10766 		ASSERT(state->dts_formats == NULL);
10767 		return;
10768 	}
10769 
10770 	ASSERT(state->dts_formats != NULL);
10771 
10772 	for (i = 0; i < state->dts_nformats; i++) {
10773 		char *fmt = state->dts_formats[i];
10774 
10775 		if (fmt == NULL)
10776 			continue;
10777 
10778 		kmem_free(fmt, strlen(fmt) + 1);
10779 	}
10780 
10781 	kmem_free(state->dts_formats, state->dts_nformats * sizeof (char *));
10782 	state->dts_nformats = 0;
10783 	state->dts_formats = NULL;
10784 }
10785 
10786 /*
10787  * DTrace Predicate Functions
10788  */
10789 static dtrace_predicate_t *
10790 dtrace_predicate_create(dtrace_difo_t *dp)
10791 {
10792 	dtrace_predicate_t *pred;
10793 
10794 	ASSERT(MUTEX_HELD(&dtrace_lock));
10795 	ASSERT(dp->dtdo_refcnt != 0);
10796 
10797 	pred = kmem_zalloc(sizeof (dtrace_predicate_t), KM_SLEEP);
10798 	pred->dtp_difo = dp;
10799 	pred->dtp_refcnt = 1;
10800 
10801 	if (!dtrace_difo_cacheable(dp))
10802 		return (pred);
10803 
10804 	if (dtrace_predcache_id == DTRACE_CACHEIDNONE) {
10805 		/*
10806 		 * This is only theoretically possible -- we have had 2^32
10807 		 * cacheable predicates on this machine.  We cannot allow any
10808 		 * more predicates to become cacheable:  as unlikely as it is,
10809 		 * there may be a thread caching a (now stale) predicate cache
10810 		 * ID. (N.B.: the temptation is being successfully resisted to
10811 		 * have this cmn_err() "Holy shit -- we executed this code!")
10812 		 */
10813 		return (pred);
10814 	}
10815 
10816 	pred->dtp_cacheid = dtrace_predcache_id++;
10817 
10818 	return (pred);
10819 }
10820 
10821 static void
10822 dtrace_predicate_hold(dtrace_predicate_t *pred)
10823 {
10824 	ASSERT(MUTEX_HELD(&dtrace_lock));
10825 	ASSERT(pred->dtp_difo != NULL && pred->dtp_difo->dtdo_refcnt != 0);
10826 	ASSERT(pred->dtp_refcnt > 0);
10827 
10828 	pred->dtp_refcnt++;
10829 }
10830 
10831 static void
10832 dtrace_predicate_release(dtrace_predicate_t *pred, dtrace_vstate_t *vstate)
10833 {
10834 	dtrace_difo_t *dp = pred->dtp_difo;
10835 
10836 	ASSERT(MUTEX_HELD(&dtrace_lock));
10837 	ASSERT(dp != NULL && dp->dtdo_refcnt != 0);
10838 	ASSERT(pred->dtp_refcnt > 0);
10839 
10840 	if (--pred->dtp_refcnt == 0) {
10841 		dtrace_difo_release(pred->dtp_difo, vstate);
10842 		kmem_free(pred, sizeof (dtrace_predicate_t));
10843 	}
10844 }
10845 
10846 /*
10847  * DTrace Action Description Functions
10848  */
10849 static dtrace_actdesc_t *
10850 dtrace_actdesc_create(dtrace_actkind_t kind, uint32_t ntuple,
10851     uint64_t uarg, uint64_t arg)
10852 {
10853 	dtrace_actdesc_t *act;
10854 
10855 #ifdef illumos
10856 	ASSERT(!DTRACEACT_ISPRINTFLIKE(kind) || (arg != NULL &&
10857 	    arg >= KERNELBASE) || (arg == NULL && kind == DTRACEACT_PRINTA));
10858 #endif
10859 
10860 	act = kmem_zalloc(sizeof (dtrace_actdesc_t), KM_SLEEP);
10861 	act->dtad_kind = kind;
10862 	act->dtad_ntuple = ntuple;
10863 	act->dtad_uarg = uarg;
10864 	act->dtad_arg = arg;
10865 	act->dtad_refcnt = 1;
10866 
10867 	return (act);
10868 }
10869 
10870 static void
10871 dtrace_actdesc_hold(dtrace_actdesc_t *act)
10872 {
10873 	ASSERT(act->dtad_refcnt >= 1);
10874 	act->dtad_refcnt++;
10875 }
10876 
10877 static void
10878 dtrace_actdesc_release(dtrace_actdesc_t *act, dtrace_vstate_t *vstate)
10879 {
10880 	dtrace_actkind_t kind = act->dtad_kind;
10881 	dtrace_difo_t *dp;
10882 
10883 	ASSERT(act->dtad_refcnt >= 1);
10884 
10885 	if (--act->dtad_refcnt != 0)
10886 		return;
10887 
10888 	if ((dp = act->dtad_difo) != NULL)
10889 		dtrace_difo_release(dp, vstate);
10890 
10891 	if (DTRACEACT_ISPRINTFLIKE(kind)) {
10892 		char *str = (char *)(uintptr_t)act->dtad_arg;
10893 
10894 #ifdef illumos
10895 		ASSERT((str != NULL && (uintptr_t)str >= KERNELBASE) ||
10896 		    (str == NULL && act->dtad_kind == DTRACEACT_PRINTA));
10897 #endif
10898 
10899 		if (str != NULL)
10900 			kmem_free(str, strlen(str) + 1);
10901 	}
10902 
10903 	kmem_free(act, sizeof (dtrace_actdesc_t));
10904 }
10905 
10906 /*
10907  * DTrace ECB Functions
10908  */
10909 static dtrace_ecb_t *
10910 dtrace_ecb_add(dtrace_state_t *state, dtrace_probe_t *probe)
10911 {
10912 	dtrace_ecb_t *ecb;
10913 	dtrace_epid_t epid;
10914 
10915 	ASSERT(MUTEX_HELD(&dtrace_lock));
10916 
10917 	ecb = kmem_zalloc(sizeof (dtrace_ecb_t), KM_SLEEP);
10918 	ecb->dte_predicate = NULL;
10919 	ecb->dte_probe = probe;
10920 
10921 	/*
10922 	 * The default size is the size of the default action: recording
10923 	 * the header.
10924 	 */
10925 	ecb->dte_size = ecb->dte_needed = sizeof (dtrace_rechdr_t);
10926 	ecb->dte_alignment = sizeof (dtrace_epid_t);
10927 
10928 	epid = state->dts_epid++;
10929 
10930 	if (epid - 1 >= state->dts_necbs) {
10931 		dtrace_ecb_t **oecbs = state->dts_ecbs, **ecbs;
10932 		int necbs = state->dts_necbs << 1;
10933 
10934 		ASSERT(epid == state->dts_necbs + 1);
10935 
10936 		if (necbs == 0) {
10937 			ASSERT(oecbs == NULL);
10938 			necbs = 1;
10939 		}
10940 
10941 		ecbs = kmem_zalloc(necbs * sizeof (*ecbs), KM_SLEEP);
10942 
10943 		if (oecbs != NULL)
10944 			bcopy(oecbs, ecbs, state->dts_necbs * sizeof (*ecbs));
10945 
10946 		dtrace_membar_producer();
10947 		state->dts_ecbs = ecbs;
10948 
10949 		if (oecbs != NULL) {
10950 			/*
10951 			 * If this state is active, we must dtrace_sync()
10952 			 * before we can free the old dts_ecbs array:  we're
10953 			 * coming in hot, and there may be active ring
10954 			 * buffer processing (which indexes into the dts_ecbs
10955 			 * array) on another CPU.
10956 			 */
10957 			if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
10958 				dtrace_sync();
10959 
10960 			kmem_free(oecbs, state->dts_necbs * sizeof (*ecbs));
10961 		}
10962 
10963 		dtrace_membar_producer();
10964 		state->dts_necbs = necbs;
10965 	}
10966 
10967 	ecb->dte_state = state;
10968 
10969 	ASSERT(state->dts_ecbs[epid - 1] == NULL);
10970 	dtrace_membar_producer();
10971 	state->dts_ecbs[(ecb->dte_epid = epid) - 1] = ecb;
10972 
10973 	return (ecb);
10974 }
10975 
10976 static void
10977 dtrace_ecb_enable(dtrace_ecb_t *ecb)
10978 {
10979 	dtrace_probe_t *probe = ecb->dte_probe;
10980 
10981 	ASSERT(MUTEX_HELD(&cpu_lock));
10982 	ASSERT(MUTEX_HELD(&dtrace_lock));
10983 	ASSERT(ecb->dte_next == NULL);
10984 
10985 	if (probe == NULL) {
10986 		/*
10987 		 * This is the NULL probe -- there's nothing to do.
10988 		 */
10989 		return;
10990 	}
10991 
10992 	if (probe->dtpr_ecb == NULL) {
10993 		dtrace_provider_t *prov = probe->dtpr_provider;
10994 
10995 		/*
10996 		 * We're the first ECB on this probe.
10997 		 */
10998 		probe->dtpr_ecb = probe->dtpr_ecb_last = ecb;
10999 
11000 		if (ecb->dte_predicate != NULL)
11001 			probe->dtpr_predcache = ecb->dte_predicate->dtp_cacheid;
11002 
11003 		prov->dtpv_pops.dtps_enable(prov->dtpv_arg,
11004 		    probe->dtpr_id, probe->dtpr_arg);
11005 	} else {
11006 		/*
11007 		 * This probe is already active.  Swing the last pointer to
11008 		 * point to the new ECB, and issue a dtrace_sync() to assure
11009 		 * that all CPUs have seen the change.
11010 		 */
11011 		ASSERT(probe->dtpr_ecb_last != NULL);
11012 		probe->dtpr_ecb_last->dte_next = ecb;
11013 		probe->dtpr_ecb_last = ecb;
11014 		probe->dtpr_predcache = 0;
11015 
11016 		dtrace_sync();
11017 	}
11018 }
11019 
11020 static void
11021 dtrace_ecb_resize(dtrace_ecb_t *ecb)
11022 {
11023 	dtrace_action_t *act;
11024 	uint32_t curneeded = UINT32_MAX;
11025 	uint32_t aggbase = UINT32_MAX;
11026 
11027 	/*
11028 	 * If we record anything, we always record the dtrace_rechdr_t.  (And
11029 	 * we always record it first.)
11030 	 */
11031 	ecb->dte_size = sizeof (dtrace_rechdr_t);
11032 	ecb->dte_alignment = sizeof (dtrace_epid_t);
11033 
11034 	for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
11035 		dtrace_recdesc_t *rec = &act->dta_rec;
11036 		ASSERT(rec->dtrd_size > 0 || rec->dtrd_alignment == 1);
11037 
11038 		ecb->dte_alignment = MAX(ecb->dte_alignment,
11039 		    rec->dtrd_alignment);
11040 
11041 		if (DTRACEACT_ISAGG(act->dta_kind)) {
11042 			dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
11043 
11044 			ASSERT(rec->dtrd_size != 0);
11045 			ASSERT(agg->dtag_first != NULL);
11046 			ASSERT(act->dta_prev->dta_intuple);
11047 			ASSERT(aggbase != UINT32_MAX);
11048 			ASSERT(curneeded != UINT32_MAX);
11049 
11050 			agg->dtag_base = aggbase;
11051 
11052 			curneeded = P2ROUNDUP(curneeded, rec->dtrd_alignment);
11053 			rec->dtrd_offset = curneeded;
11054 			curneeded += rec->dtrd_size;
11055 			ecb->dte_needed = MAX(ecb->dte_needed, curneeded);
11056 
11057 			aggbase = UINT32_MAX;
11058 			curneeded = UINT32_MAX;
11059 		} else if (act->dta_intuple) {
11060 			if (curneeded == UINT32_MAX) {
11061 				/*
11062 				 * This is the first record in a tuple.  Align
11063 				 * curneeded to be at offset 4 in an 8-byte
11064 				 * aligned block.
11065 				 */
11066 				ASSERT(act->dta_prev == NULL ||
11067 				    !act->dta_prev->dta_intuple);
11068 				ASSERT3U(aggbase, ==, UINT32_MAX);
11069 				curneeded = P2PHASEUP(ecb->dte_size,
11070 				    sizeof (uint64_t), sizeof (dtrace_aggid_t));
11071 
11072 				aggbase = curneeded - sizeof (dtrace_aggid_t);
11073 				ASSERT(IS_P2ALIGNED(aggbase,
11074 				    sizeof (uint64_t)));
11075 			}
11076 			curneeded = P2ROUNDUP(curneeded, rec->dtrd_alignment);
11077 			rec->dtrd_offset = curneeded;
11078 			curneeded += rec->dtrd_size;
11079 		} else {
11080 			/* tuples must be followed by an aggregation */
11081 			ASSERT(act->dta_prev == NULL ||
11082 			    !act->dta_prev->dta_intuple);
11083 
11084 			ecb->dte_size = P2ROUNDUP(ecb->dte_size,
11085 			    rec->dtrd_alignment);
11086 			rec->dtrd_offset = ecb->dte_size;
11087 			ecb->dte_size += rec->dtrd_size;
11088 			ecb->dte_needed = MAX(ecb->dte_needed, ecb->dte_size);
11089 		}
11090 	}
11091 
11092 	if ((act = ecb->dte_action) != NULL &&
11093 	    !(act->dta_kind == DTRACEACT_SPECULATE && act->dta_next == NULL) &&
11094 	    ecb->dte_size == sizeof (dtrace_rechdr_t)) {
11095 		/*
11096 		 * If the size is still sizeof (dtrace_rechdr_t), then all
11097 		 * actions store no data; set the size to 0.
11098 		 */
11099 		ecb->dte_size = 0;
11100 	}
11101 
11102 	ecb->dte_size = P2ROUNDUP(ecb->dte_size, sizeof (dtrace_epid_t));
11103 	ecb->dte_needed = P2ROUNDUP(ecb->dte_needed, (sizeof (dtrace_epid_t)));
11104 	ecb->dte_state->dts_needed = MAX(ecb->dte_state->dts_needed,
11105 	    ecb->dte_needed);
11106 }
11107 
11108 static dtrace_action_t *
11109 dtrace_ecb_aggregation_create(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
11110 {
11111 	dtrace_aggregation_t *agg;
11112 	size_t size = sizeof (uint64_t);
11113 	int ntuple = desc->dtad_ntuple;
11114 	dtrace_action_t *act;
11115 	dtrace_recdesc_t *frec;
11116 	dtrace_aggid_t aggid;
11117 	dtrace_state_t *state = ecb->dte_state;
11118 
11119 	agg = kmem_zalloc(sizeof (dtrace_aggregation_t), KM_SLEEP);
11120 	agg->dtag_ecb = ecb;
11121 
11122 	ASSERT(DTRACEACT_ISAGG(desc->dtad_kind));
11123 
11124 	switch (desc->dtad_kind) {
11125 	case DTRACEAGG_MIN:
11126 		agg->dtag_initial = INT64_MAX;
11127 		agg->dtag_aggregate = dtrace_aggregate_min;
11128 		break;
11129 
11130 	case DTRACEAGG_MAX:
11131 		agg->dtag_initial = INT64_MIN;
11132 		agg->dtag_aggregate = dtrace_aggregate_max;
11133 		break;
11134 
11135 	case DTRACEAGG_COUNT:
11136 		agg->dtag_aggregate = dtrace_aggregate_count;
11137 		break;
11138 
11139 	case DTRACEAGG_QUANTIZE:
11140 		agg->dtag_aggregate = dtrace_aggregate_quantize;
11141 		size = (((sizeof (uint64_t) * NBBY) - 1) * 2 + 1) *
11142 		    sizeof (uint64_t);
11143 		break;
11144 
11145 	case DTRACEAGG_LQUANTIZE: {
11146 		uint16_t step = DTRACE_LQUANTIZE_STEP(desc->dtad_arg);
11147 		uint16_t levels = DTRACE_LQUANTIZE_LEVELS(desc->dtad_arg);
11148 
11149 		agg->dtag_initial = desc->dtad_arg;
11150 		agg->dtag_aggregate = dtrace_aggregate_lquantize;
11151 
11152 		if (step == 0 || levels == 0)
11153 			goto err;
11154 
11155 		size = levels * sizeof (uint64_t) + 3 * sizeof (uint64_t);
11156 		break;
11157 	}
11158 
11159 	case DTRACEAGG_LLQUANTIZE: {
11160 		uint16_t factor = DTRACE_LLQUANTIZE_FACTOR(desc->dtad_arg);
11161 		uint16_t low = DTRACE_LLQUANTIZE_LOW(desc->dtad_arg);
11162 		uint16_t high = DTRACE_LLQUANTIZE_HIGH(desc->dtad_arg);
11163 		uint16_t nsteps = DTRACE_LLQUANTIZE_NSTEP(desc->dtad_arg);
11164 		int64_t v;
11165 
11166 		agg->dtag_initial = desc->dtad_arg;
11167 		agg->dtag_aggregate = dtrace_aggregate_llquantize;
11168 
11169 		if (factor < 2 || low >= high || nsteps < factor)
11170 			goto err;
11171 
11172 		/*
11173 		 * Now check that the number of steps evenly divides a power
11174 		 * of the factor.  (This assures both integer bucket size and
11175 		 * linearity within each magnitude.)
11176 		 */
11177 		for (v = factor; v < nsteps; v *= factor)
11178 			continue;
11179 
11180 		if ((v % nsteps) || (nsteps % factor))
11181 			goto err;
11182 
11183 		size = (dtrace_aggregate_llquantize_bucket(factor,
11184 		    low, high, nsteps, INT64_MAX) + 2) * sizeof (uint64_t);
11185 		break;
11186 	}
11187 
11188 	case DTRACEAGG_AVG:
11189 		agg->dtag_aggregate = dtrace_aggregate_avg;
11190 		size = sizeof (uint64_t) * 2;
11191 		break;
11192 
11193 	case DTRACEAGG_STDDEV:
11194 		agg->dtag_aggregate = dtrace_aggregate_stddev;
11195 		size = sizeof (uint64_t) * 4;
11196 		break;
11197 
11198 	case DTRACEAGG_SUM:
11199 		agg->dtag_aggregate = dtrace_aggregate_sum;
11200 		break;
11201 
11202 	default:
11203 		goto err;
11204 	}
11205 
11206 	agg->dtag_action.dta_rec.dtrd_size = size;
11207 
11208 	if (ntuple == 0)
11209 		goto err;
11210 
11211 	/*
11212 	 * We must make sure that we have enough actions for the n-tuple.
11213 	 */
11214 	for (act = ecb->dte_action_last; act != NULL; act = act->dta_prev) {
11215 		if (DTRACEACT_ISAGG(act->dta_kind))
11216 			break;
11217 
11218 		if (--ntuple == 0) {
11219 			/*
11220 			 * This is the action with which our n-tuple begins.
11221 			 */
11222 			agg->dtag_first = act;
11223 			goto success;
11224 		}
11225 	}
11226 
11227 	/*
11228 	 * This n-tuple is short by ntuple elements.  Return failure.
11229 	 */
11230 	ASSERT(ntuple != 0);
11231 err:
11232 	kmem_free(agg, sizeof (dtrace_aggregation_t));
11233 	return (NULL);
11234 
11235 success:
11236 	/*
11237 	 * If the last action in the tuple has a size of zero, it's actually
11238 	 * an expression argument for the aggregating action.
11239 	 */
11240 	ASSERT(ecb->dte_action_last != NULL);
11241 	act = ecb->dte_action_last;
11242 
11243 	if (act->dta_kind == DTRACEACT_DIFEXPR) {
11244 		ASSERT(act->dta_difo != NULL);
11245 
11246 		if (act->dta_difo->dtdo_rtype.dtdt_size == 0)
11247 			agg->dtag_hasarg = 1;
11248 	}
11249 
11250 	/*
11251 	 * We need to allocate an id for this aggregation.
11252 	 */
11253 #ifdef illumos
11254 	aggid = (dtrace_aggid_t)(uintptr_t)vmem_alloc(state->dts_aggid_arena, 1,
11255 	    VM_BESTFIT | VM_SLEEP);
11256 #else
11257 	aggid = alloc_unr(state->dts_aggid_arena);
11258 #endif
11259 
11260 	if (aggid - 1 >= state->dts_naggregations) {
11261 		dtrace_aggregation_t **oaggs = state->dts_aggregations;
11262 		dtrace_aggregation_t **aggs;
11263 		int naggs = state->dts_naggregations << 1;
11264 		int onaggs = state->dts_naggregations;
11265 
11266 		ASSERT(aggid == state->dts_naggregations + 1);
11267 
11268 		if (naggs == 0) {
11269 			ASSERT(oaggs == NULL);
11270 			naggs = 1;
11271 		}
11272 
11273 		aggs = kmem_zalloc(naggs * sizeof (*aggs), KM_SLEEP);
11274 
11275 		if (oaggs != NULL) {
11276 			bcopy(oaggs, aggs, onaggs * sizeof (*aggs));
11277 			kmem_free(oaggs, onaggs * sizeof (*aggs));
11278 		}
11279 
11280 		state->dts_aggregations = aggs;
11281 		state->dts_naggregations = naggs;
11282 	}
11283 
11284 	ASSERT(state->dts_aggregations[aggid - 1] == NULL);
11285 	state->dts_aggregations[(agg->dtag_id = aggid) - 1] = agg;
11286 
11287 	frec = &agg->dtag_first->dta_rec;
11288 	if (frec->dtrd_alignment < sizeof (dtrace_aggid_t))
11289 		frec->dtrd_alignment = sizeof (dtrace_aggid_t);
11290 
11291 	for (act = agg->dtag_first; act != NULL; act = act->dta_next) {
11292 		ASSERT(!act->dta_intuple);
11293 		act->dta_intuple = 1;
11294 	}
11295 
11296 	return (&agg->dtag_action);
11297 }
11298 
11299 static void
11300 dtrace_ecb_aggregation_destroy(dtrace_ecb_t *ecb, dtrace_action_t *act)
11301 {
11302 	dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
11303 	dtrace_state_t *state = ecb->dte_state;
11304 	dtrace_aggid_t aggid = agg->dtag_id;
11305 
11306 	ASSERT(DTRACEACT_ISAGG(act->dta_kind));
11307 #ifdef illumos
11308 	vmem_free(state->dts_aggid_arena, (void *)(uintptr_t)aggid, 1);
11309 #else
11310 	free_unr(state->dts_aggid_arena, aggid);
11311 #endif
11312 
11313 	ASSERT(state->dts_aggregations[aggid - 1] == agg);
11314 	state->dts_aggregations[aggid - 1] = NULL;
11315 
11316 	kmem_free(agg, sizeof (dtrace_aggregation_t));
11317 }
11318 
11319 static int
11320 dtrace_ecb_action_add(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
11321 {
11322 	dtrace_action_t *action, *last;
11323 	dtrace_difo_t *dp = desc->dtad_difo;
11324 	uint32_t size = 0, align = sizeof (uint8_t), mask;
11325 	uint16_t format = 0;
11326 	dtrace_recdesc_t *rec;
11327 	dtrace_state_t *state = ecb->dte_state;
11328 	dtrace_optval_t *opt = state->dts_options, nframes = 0, strsize;
11329 	uint64_t arg = desc->dtad_arg;
11330 
11331 	ASSERT(MUTEX_HELD(&dtrace_lock));
11332 	ASSERT(ecb->dte_action == NULL || ecb->dte_action->dta_refcnt == 1);
11333 
11334 	if (DTRACEACT_ISAGG(desc->dtad_kind)) {
11335 		/*
11336 		 * If this is an aggregating action, there must be neither
11337 		 * a speculate nor a commit on the action chain.
11338 		 */
11339 		dtrace_action_t *act;
11340 
11341 		for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
11342 			if (act->dta_kind == DTRACEACT_COMMIT)
11343 				return (EINVAL);
11344 
11345 			if (act->dta_kind == DTRACEACT_SPECULATE)
11346 				return (EINVAL);
11347 		}
11348 
11349 		action = dtrace_ecb_aggregation_create(ecb, desc);
11350 
11351 		if (action == NULL)
11352 			return (EINVAL);
11353 	} else {
11354 		if (DTRACEACT_ISDESTRUCTIVE(desc->dtad_kind) ||
11355 		    (desc->dtad_kind == DTRACEACT_DIFEXPR &&
11356 		    dp != NULL && dp->dtdo_destructive)) {
11357 			state->dts_destructive = 1;
11358 		}
11359 
11360 		switch (desc->dtad_kind) {
11361 		case DTRACEACT_PRINTF:
11362 		case DTRACEACT_PRINTA:
11363 		case DTRACEACT_SYSTEM:
11364 		case DTRACEACT_FREOPEN:
11365 		case DTRACEACT_DIFEXPR:
11366 			/*
11367 			 * We know that our arg is a string -- turn it into a
11368 			 * format.
11369 			 */
11370 			if (arg == 0) {
11371 				ASSERT(desc->dtad_kind == DTRACEACT_PRINTA ||
11372 				    desc->dtad_kind == DTRACEACT_DIFEXPR);
11373 				format = 0;
11374 			} else {
11375 				ASSERT(arg != 0);
11376 #ifdef illumos
11377 				ASSERT(arg > KERNELBASE);
11378 #endif
11379 				format = dtrace_format_add(state,
11380 				    (char *)(uintptr_t)arg);
11381 			}
11382 
11383 			/*FALLTHROUGH*/
11384 		case DTRACEACT_LIBACT:
11385 		case DTRACEACT_TRACEMEM:
11386 		case DTRACEACT_TRACEMEM_DYNSIZE:
11387 			if (dp == NULL)
11388 				return (EINVAL);
11389 
11390 			if ((size = dp->dtdo_rtype.dtdt_size) != 0)
11391 				break;
11392 
11393 			if (dp->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING) {
11394 				if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
11395 					return (EINVAL);
11396 
11397 				size = opt[DTRACEOPT_STRSIZE];
11398 			}
11399 
11400 			break;
11401 
11402 		case DTRACEACT_STACK:
11403 			if ((nframes = arg) == 0) {
11404 				nframes = opt[DTRACEOPT_STACKFRAMES];
11405 				ASSERT(nframes > 0);
11406 				arg = nframes;
11407 			}
11408 
11409 			size = nframes * sizeof (pc_t);
11410 			break;
11411 
11412 		case DTRACEACT_JSTACK:
11413 			if ((strsize = DTRACE_USTACK_STRSIZE(arg)) == 0)
11414 				strsize = opt[DTRACEOPT_JSTACKSTRSIZE];
11415 
11416 			if ((nframes = DTRACE_USTACK_NFRAMES(arg)) == 0)
11417 				nframes = opt[DTRACEOPT_JSTACKFRAMES];
11418 
11419 			arg = DTRACE_USTACK_ARG(nframes, strsize);
11420 
11421 			/*FALLTHROUGH*/
11422 		case DTRACEACT_USTACK:
11423 			if (desc->dtad_kind != DTRACEACT_JSTACK &&
11424 			    (nframes = DTRACE_USTACK_NFRAMES(arg)) == 0) {
11425 				strsize = DTRACE_USTACK_STRSIZE(arg);
11426 				nframes = opt[DTRACEOPT_USTACKFRAMES];
11427 				ASSERT(nframes > 0);
11428 				arg = DTRACE_USTACK_ARG(nframes, strsize);
11429 			}
11430 
11431 			/*
11432 			 * Save a slot for the pid.
11433 			 */
11434 			size = (nframes + 1) * sizeof (uint64_t);
11435 			size += DTRACE_USTACK_STRSIZE(arg);
11436 			size = P2ROUNDUP(size, (uint32_t)(sizeof (uintptr_t)));
11437 
11438 			break;
11439 
11440 		case DTRACEACT_SYM:
11441 		case DTRACEACT_MOD:
11442 			if (dp == NULL || ((size = dp->dtdo_rtype.dtdt_size) !=
11443 			    sizeof (uint64_t)) ||
11444 			    (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
11445 				return (EINVAL);
11446 			break;
11447 
11448 		case DTRACEACT_USYM:
11449 		case DTRACEACT_UMOD:
11450 		case DTRACEACT_UADDR:
11451 			if (dp == NULL ||
11452 			    (dp->dtdo_rtype.dtdt_size != sizeof (uint64_t)) ||
11453 			    (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
11454 				return (EINVAL);
11455 
11456 			/*
11457 			 * We have a slot for the pid, plus a slot for the
11458 			 * argument.  To keep things simple (aligned with
11459 			 * bitness-neutral sizing), we store each as a 64-bit
11460 			 * quantity.
11461 			 */
11462 			size = 2 * sizeof (uint64_t);
11463 			break;
11464 
11465 		case DTRACEACT_STOP:
11466 		case DTRACEACT_BREAKPOINT:
11467 		case DTRACEACT_PANIC:
11468 			break;
11469 
11470 		case DTRACEACT_CHILL:
11471 		case DTRACEACT_DISCARD:
11472 		case DTRACEACT_RAISE:
11473 			if (dp == NULL)
11474 				return (EINVAL);
11475 			break;
11476 
11477 		case DTRACEACT_EXIT:
11478 			if (dp == NULL ||
11479 			    (size = dp->dtdo_rtype.dtdt_size) != sizeof (int) ||
11480 			    (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
11481 				return (EINVAL);
11482 			break;
11483 
11484 		case DTRACEACT_SPECULATE:
11485 			if (ecb->dte_size > sizeof (dtrace_rechdr_t))
11486 				return (EINVAL);
11487 
11488 			if (dp == NULL)
11489 				return (EINVAL);
11490 
11491 			state->dts_speculates = 1;
11492 			break;
11493 
11494 		case DTRACEACT_PRINTM:
11495 		    	size = dp->dtdo_rtype.dtdt_size;
11496 			break;
11497 
11498 		case DTRACEACT_PRINTT:
11499 		    	size = dp->dtdo_rtype.dtdt_size;
11500 			break;
11501 
11502 		case DTRACEACT_COMMIT: {
11503 			dtrace_action_t *act = ecb->dte_action;
11504 
11505 			for (; act != NULL; act = act->dta_next) {
11506 				if (act->dta_kind == DTRACEACT_COMMIT)
11507 					return (EINVAL);
11508 			}
11509 
11510 			if (dp == NULL)
11511 				return (EINVAL);
11512 			break;
11513 		}
11514 
11515 		default:
11516 			return (EINVAL);
11517 		}
11518 
11519 		if (size != 0 || desc->dtad_kind == DTRACEACT_SPECULATE) {
11520 			/*
11521 			 * If this is a data-storing action or a speculate,
11522 			 * we must be sure that there isn't a commit on the
11523 			 * action chain.
11524 			 */
11525 			dtrace_action_t *act = ecb->dte_action;
11526 
11527 			for (; act != NULL; act = act->dta_next) {
11528 				if (act->dta_kind == DTRACEACT_COMMIT)
11529 					return (EINVAL);
11530 			}
11531 		}
11532 
11533 		action = kmem_zalloc(sizeof (dtrace_action_t), KM_SLEEP);
11534 		action->dta_rec.dtrd_size = size;
11535 	}
11536 
11537 	action->dta_refcnt = 1;
11538 	rec = &action->dta_rec;
11539 	size = rec->dtrd_size;
11540 
11541 	for (mask = sizeof (uint64_t) - 1; size != 0 && mask > 0; mask >>= 1) {
11542 		if (!(size & mask)) {
11543 			align = mask + 1;
11544 			break;
11545 		}
11546 	}
11547 
11548 	action->dta_kind = desc->dtad_kind;
11549 
11550 	if ((action->dta_difo = dp) != NULL)
11551 		dtrace_difo_hold(dp);
11552 
11553 	rec->dtrd_action = action->dta_kind;
11554 	rec->dtrd_arg = arg;
11555 	rec->dtrd_uarg = desc->dtad_uarg;
11556 	rec->dtrd_alignment = (uint16_t)align;
11557 	rec->dtrd_format = format;
11558 
11559 	if ((last = ecb->dte_action_last) != NULL) {
11560 		ASSERT(ecb->dte_action != NULL);
11561 		action->dta_prev = last;
11562 		last->dta_next = action;
11563 	} else {
11564 		ASSERT(ecb->dte_action == NULL);
11565 		ecb->dte_action = action;
11566 	}
11567 
11568 	ecb->dte_action_last = action;
11569 
11570 	return (0);
11571 }
11572 
11573 static void
11574 dtrace_ecb_action_remove(dtrace_ecb_t *ecb)
11575 {
11576 	dtrace_action_t *act = ecb->dte_action, *next;
11577 	dtrace_vstate_t *vstate = &ecb->dte_state->dts_vstate;
11578 	dtrace_difo_t *dp;
11579 	uint16_t format;
11580 
11581 	if (act != NULL && act->dta_refcnt > 1) {
11582 		ASSERT(act->dta_next == NULL || act->dta_next->dta_refcnt == 1);
11583 		act->dta_refcnt--;
11584 	} else {
11585 		for (; act != NULL; act = next) {
11586 			next = act->dta_next;
11587 			ASSERT(next != NULL || act == ecb->dte_action_last);
11588 			ASSERT(act->dta_refcnt == 1);
11589 
11590 			if ((format = act->dta_rec.dtrd_format) != 0)
11591 				dtrace_format_remove(ecb->dte_state, format);
11592 
11593 			if ((dp = act->dta_difo) != NULL)
11594 				dtrace_difo_release(dp, vstate);
11595 
11596 			if (DTRACEACT_ISAGG(act->dta_kind)) {
11597 				dtrace_ecb_aggregation_destroy(ecb, act);
11598 			} else {
11599 				kmem_free(act, sizeof (dtrace_action_t));
11600 			}
11601 		}
11602 	}
11603 
11604 	ecb->dte_action = NULL;
11605 	ecb->dte_action_last = NULL;
11606 	ecb->dte_size = 0;
11607 }
11608 
11609 static void
11610 dtrace_ecb_disable(dtrace_ecb_t *ecb)
11611 {
11612 	/*
11613 	 * We disable the ECB by removing it from its probe.
11614 	 */
11615 	dtrace_ecb_t *pecb, *prev = NULL;
11616 	dtrace_probe_t *probe = ecb->dte_probe;
11617 
11618 	ASSERT(MUTEX_HELD(&dtrace_lock));
11619 
11620 	if (probe == NULL) {
11621 		/*
11622 		 * This is the NULL probe; there is nothing to disable.
11623 		 */
11624 		return;
11625 	}
11626 
11627 	for (pecb = probe->dtpr_ecb; pecb != NULL; pecb = pecb->dte_next) {
11628 		if (pecb == ecb)
11629 			break;
11630 		prev = pecb;
11631 	}
11632 
11633 	ASSERT(pecb != NULL);
11634 
11635 	if (prev == NULL) {
11636 		probe->dtpr_ecb = ecb->dte_next;
11637 	} else {
11638 		prev->dte_next = ecb->dte_next;
11639 	}
11640 
11641 	if (ecb == probe->dtpr_ecb_last) {
11642 		ASSERT(ecb->dte_next == NULL);
11643 		probe->dtpr_ecb_last = prev;
11644 	}
11645 
11646 	/*
11647 	 * The ECB has been disconnected from the probe; now sync to assure
11648 	 * that all CPUs have seen the change before returning.
11649 	 */
11650 	dtrace_sync();
11651 
11652 	if (probe->dtpr_ecb == NULL) {
11653 		/*
11654 		 * That was the last ECB on the probe; clear the predicate
11655 		 * cache ID for the probe, disable it and sync one more time
11656 		 * to assure that we'll never hit it again.
11657 		 */
11658 		dtrace_provider_t *prov = probe->dtpr_provider;
11659 
11660 		ASSERT(ecb->dte_next == NULL);
11661 		ASSERT(probe->dtpr_ecb_last == NULL);
11662 		probe->dtpr_predcache = DTRACE_CACHEIDNONE;
11663 		prov->dtpv_pops.dtps_disable(prov->dtpv_arg,
11664 		    probe->dtpr_id, probe->dtpr_arg);
11665 		dtrace_sync();
11666 	} else {
11667 		/*
11668 		 * There is at least one ECB remaining on the probe.  If there
11669 		 * is _exactly_ one, set the probe's predicate cache ID to be
11670 		 * the predicate cache ID of the remaining ECB.
11671 		 */
11672 		ASSERT(probe->dtpr_ecb_last != NULL);
11673 		ASSERT(probe->dtpr_predcache == DTRACE_CACHEIDNONE);
11674 
11675 		if (probe->dtpr_ecb == probe->dtpr_ecb_last) {
11676 			dtrace_predicate_t *p = probe->dtpr_ecb->dte_predicate;
11677 
11678 			ASSERT(probe->dtpr_ecb->dte_next == NULL);
11679 
11680 			if (p != NULL)
11681 				probe->dtpr_predcache = p->dtp_cacheid;
11682 		}
11683 
11684 		ecb->dte_next = NULL;
11685 	}
11686 }
11687 
11688 static void
11689 dtrace_ecb_destroy(dtrace_ecb_t *ecb)
11690 {
11691 	dtrace_state_t *state = ecb->dte_state;
11692 	dtrace_vstate_t *vstate = &state->dts_vstate;
11693 	dtrace_predicate_t *pred;
11694 	dtrace_epid_t epid = ecb->dte_epid;
11695 
11696 	ASSERT(MUTEX_HELD(&dtrace_lock));
11697 	ASSERT(ecb->dte_next == NULL);
11698 	ASSERT(ecb->dte_probe == NULL || ecb->dte_probe->dtpr_ecb != ecb);
11699 
11700 	if ((pred = ecb->dte_predicate) != NULL)
11701 		dtrace_predicate_release(pred, vstate);
11702 
11703 	dtrace_ecb_action_remove(ecb);
11704 
11705 	ASSERT(state->dts_ecbs[epid - 1] == ecb);
11706 	state->dts_ecbs[epid - 1] = NULL;
11707 
11708 	kmem_free(ecb, sizeof (dtrace_ecb_t));
11709 }
11710 
11711 static dtrace_ecb_t *
11712 dtrace_ecb_create(dtrace_state_t *state, dtrace_probe_t *probe,
11713     dtrace_enabling_t *enab)
11714 {
11715 	dtrace_ecb_t *ecb;
11716 	dtrace_predicate_t *pred;
11717 	dtrace_actdesc_t *act;
11718 	dtrace_provider_t *prov;
11719 	dtrace_ecbdesc_t *desc = enab->dten_current;
11720 
11721 	ASSERT(MUTEX_HELD(&dtrace_lock));
11722 	ASSERT(state != NULL);
11723 
11724 	ecb = dtrace_ecb_add(state, probe);
11725 	ecb->dte_uarg = desc->dted_uarg;
11726 
11727 	if ((pred = desc->dted_pred.dtpdd_predicate) != NULL) {
11728 		dtrace_predicate_hold(pred);
11729 		ecb->dte_predicate = pred;
11730 	}
11731 
11732 	if (probe != NULL) {
11733 		/*
11734 		 * If the provider shows more leg than the consumer is old
11735 		 * enough to see, we need to enable the appropriate implicit
11736 		 * predicate bits to prevent the ecb from activating at
11737 		 * revealing times.
11738 		 *
11739 		 * Providers specifying DTRACE_PRIV_USER at register time
11740 		 * are stating that they need the /proc-style privilege
11741 		 * model to be enforced, and this is what DTRACE_COND_OWNER
11742 		 * and DTRACE_COND_ZONEOWNER will then do at probe time.
11743 		 */
11744 		prov = probe->dtpr_provider;
11745 		if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLPROC) &&
11746 		    (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
11747 			ecb->dte_cond |= DTRACE_COND_OWNER;
11748 
11749 		if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLZONE) &&
11750 		    (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
11751 			ecb->dte_cond |= DTRACE_COND_ZONEOWNER;
11752 
11753 		/*
11754 		 * If the provider shows us kernel innards and the user
11755 		 * is lacking sufficient privilege, enable the
11756 		 * DTRACE_COND_USERMODE implicit predicate.
11757 		 */
11758 		if (!(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) &&
11759 		    (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_KERNEL))
11760 			ecb->dte_cond |= DTRACE_COND_USERMODE;
11761 	}
11762 
11763 	if (dtrace_ecb_create_cache != NULL) {
11764 		/*
11765 		 * If we have a cached ecb, we'll use its action list instead
11766 		 * of creating our own (saving both time and space).
11767 		 */
11768 		dtrace_ecb_t *cached = dtrace_ecb_create_cache;
11769 		dtrace_action_t *act = cached->dte_action;
11770 
11771 		if (act != NULL) {
11772 			ASSERT(act->dta_refcnt > 0);
11773 			act->dta_refcnt++;
11774 			ecb->dte_action = act;
11775 			ecb->dte_action_last = cached->dte_action_last;
11776 			ecb->dte_needed = cached->dte_needed;
11777 			ecb->dte_size = cached->dte_size;
11778 			ecb->dte_alignment = cached->dte_alignment;
11779 		}
11780 
11781 		return (ecb);
11782 	}
11783 
11784 	for (act = desc->dted_action; act != NULL; act = act->dtad_next) {
11785 		if ((enab->dten_error = dtrace_ecb_action_add(ecb, act)) != 0) {
11786 			dtrace_ecb_destroy(ecb);
11787 			return (NULL);
11788 		}
11789 	}
11790 
11791 	dtrace_ecb_resize(ecb);
11792 
11793 	return (dtrace_ecb_create_cache = ecb);
11794 }
11795 
11796 static int
11797 dtrace_ecb_create_enable(dtrace_probe_t *probe, void *arg)
11798 {
11799 	dtrace_ecb_t *ecb;
11800 	dtrace_enabling_t *enab = arg;
11801 	dtrace_state_t *state = enab->dten_vstate->dtvs_state;
11802 
11803 	ASSERT(state != NULL);
11804 
11805 	if (probe != NULL && probe->dtpr_gen < enab->dten_probegen) {
11806 		/*
11807 		 * This probe was created in a generation for which this
11808 		 * enabling has previously created ECBs; we don't want to
11809 		 * enable it again, so just kick out.
11810 		 */
11811 		return (DTRACE_MATCH_NEXT);
11812 	}
11813 
11814 	if ((ecb = dtrace_ecb_create(state, probe, enab)) == NULL)
11815 		return (DTRACE_MATCH_DONE);
11816 
11817 	dtrace_ecb_enable(ecb);
11818 	return (DTRACE_MATCH_NEXT);
11819 }
11820 
11821 static dtrace_ecb_t *
11822 dtrace_epid2ecb(dtrace_state_t *state, dtrace_epid_t id)
11823 {
11824 	dtrace_ecb_t *ecb;
11825 
11826 	ASSERT(MUTEX_HELD(&dtrace_lock));
11827 
11828 	if (id == 0 || id > state->dts_necbs)
11829 		return (NULL);
11830 
11831 	ASSERT(state->dts_necbs > 0 && state->dts_ecbs != NULL);
11832 	ASSERT((ecb = state->dts_ecbs[id - 1]) == NULL || ecb->dte_epid == id);
11833 
11834 	return (state->dts_ecbs[id - 1]);
11835 }
11836 
11837 static dtrace_aggregation_t *
11838 dtrace_aggid2agg(dtrace_state_t *state, dtrace_aggid_t id)
11839 {
11840 	dtrace_aggregation_t *agg;
11841 
11842 	ASSERT(MUTEX_HELD(&dtrace_lock));
11843 
11844 	if (id == 0 || id > state->dts_naggregations)
11845 		return (NULL);
11846 
11847 	ASSERT(state->dts_naggregations > 0 && state->dts_aggregations != NULL);
11848 	ASSERT((agg = state->dts_aggregations[id - 1]) == NULL ||
11849 	    agg->dtag_id == id);
11850 
11851 	return (state->dts_aggregations[id - 1]);
11852 }
11853 
11854 /*
11855  * DTrace Buffer Functions
11856  *
11857  * The following functions manipulate DTrace buffers.  Most of these functions
11858  * are called in the context of establishing or processing consumer state;
11859  * exceptions are explicitly noted.
11860  */
11861 
11862 /*
11863  * Note:  called from cross call context.  This function switches the two
11864  * buffers on a given CPU.  The atomicity of this operation is assured by
11865  * disabling interrupts while the actual switch takes place; the disabling of
11866  * interrupts serializes the execution with any execution of dtrace_probe() on
11867  * the same CPU.
11868  */
11869 static void
11870 dtrace_buffer_switch(dtrace_buffer_t *buf)
11871 {
11872 	caddr_t tomax = buf->dtb_tomax;
11873 	caddr_t xamot = buf->dtb_xamot;
11874 	dtrace_icookie_t cookie;
11875 	hrtime_t now;
11876 
11877 	ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
11878 	ASSERT(!(buf->dtb_flags & DTRACEBUF_RING));
11879 
11880 	cookie = dtrace_interrupt_disable();
11881 	now = dtrace_gethrtime();
11882 	buf->dtb_tomax = xamot;
11883 	buf->dtb_xamot = tomax;
11884 	buf->dtb_xamot_drops = buf->dtb_drops;
11885 	buf->dtb_xamot_offset = buf->dtb_offset;
11886 	buf->dtb_xamot_errors = buf->dtb_errors;
11887 	buf->dtb_xamot_flags = buf->dtb_flags;
11888 	buf->dtb_offset = 0;
11889 	buf->dtb_drops = 0;
11890 	buf->dtb_errors = 0;
11891 	buf->dtb_flags &= ~(DTRACEBUF_ERROR | DTRACEBUF_DROPPED);
11892 	buf->dtb_interval = now - buf->dtb_switched;
11893 	buf->dtb_switched = now;
11894 	dtrace_interrupt_enable(cookie);
11895 }
11896 
11897 /*
11898  * Note:  called from cross call context.  This function activates a buffer
11899  * on a CPU.  As with dtrace_buffer_switch(), the atomicity of the operation
11900  * is guaranteed by the disabling of interrupts.
11901  */
11902 static void
11903 dtrace_buffer_activate(dtrace_state_t *state)
11904 {
11905 	dtrace_buffer_t *buf;
11906 	dtrace_icookie_t cookie = dtrace_interrupt_disable();
11907 
11908 	buf = &state->dts_buffer[curcpu];
11909 
11910 	if (buf->dtb_tomax != NULL) {
11911 		/*
11912 		 * We might like to assert that the buffer is marked inactive,
11913 		 * but this isn't necessarily true:  the buffer for the CPU
11914 		 * that processes the BEGIN probe has its buffer activated
11915 		 * manually.  In this case, we take the (harmless) action
11916 		 * re-clearing the bit INACTIVE bit.
11917 		 */
11918 		buf->dtb_flags &= ~DTRACEBUF_INACTIVE;
11919 	}
11920 
11921 	dtrace_interrupt_enable(cookie);
11922 }
11923 
11924 #ifdef __FreeBSD__
11925 /*
11926  * Activate the specified per-CPU buffer.  This is used instead of
11927  * dtrace_buffer_activate() when APs have not yet started, i.e. when
11928  * activating anonymous state.
11929  */
11930 static void
11931 dtrace_buffer_activate_cpu(dtrace_state_t *state, int cpu)
11932 {
11933 
11934 	if (state->dts_buffer[cpu].dtb_tomax != NULL)
11935 		state->dts_buffer[cpu].dtb_flags &= ~DTRACEBUF_INACTIVE;
11936 }
11937 #endif
11938 
11939 static int
11940 dtrace_buffer_alloc(dtrace_buffer_t *bufs, size_t size, int flags,
11941     processorid_t cpu, int *factor)
11942 {
11943 #ifdef illumos
11944 	cpu_t *cp;
11945 #endif
11946 	dtrace_buffer_t *buf;
11947 	int allocated = 0, desired = 0;
11948 
11949 #ifdef illumos
11950 	ASSERT(MUTEX_HELD(&cpu_lock));
11951 	ASSERT(MUTEX_HELD(&dtrace_lock));
11952 
11953 	*factor = 1;
11954 
11955 	if (size > dtrace_nonroot_maxsize &&
11956 	    !PRIV_POLICY_CHOICE(CRED(), PRIV_ALL, B_FALSE))
11957 		return (EFBIG);
11958 
11959 	cp = cpu_list;
11960 
11961 	do {
11962 		if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
11963 			continue;
11964 
11965 		buf = &bufs[cp->cpu_id];
11966 
11967 		/*
11968 		 * If there is already a buffer allocated for this CPU, it
11969 		 * is only possible that this is a DR event.  In this case,
11970 		 */
11971 		if (buf->dtb_tomax != NULL) {
11972 			ASSERT(buf->dtb_size == size);
11973 			continue;
11974 		}
11975 
11976 		ASSERT(buf->dtb_xamot == NULL);
11977 
11978 		if ((buf->dtb_tomax = kmem_zalloc(size,
11979 		    KM_NOSLEEP | KM_NORMALPRI)) == NULL)
11980 			goto err;
11981 
11982 		buf->dtb_size = size;
11983 		buf->dtb_flags = flags;
11984 		buf->dtb_offset = 0;
11985 		buf->dtb_drops = 0;
11986 
11987 		if (flags & DTRACEBUF_NOSWITCH)
11988 			continue;
11989 
11990 		if ((buf->dtb_xamot = kmem_zalloc(size,
11991 		    KM_NOSLEEP | KM_NORMALPRI)) == NULL)
11992 			goto err;
11993 	} while ((cp = cp->cpu_next) != cpu_list);
11994 
11995 	return (0);
11996 
11997 err:
11998 	cp = cpu_list;
11999 
12000 	do {
12001 		if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
12002 			continue;
12003 
12004 		buf = &bufs[cp->cpu_id];
12005 		desired += 2;
12006 
12007 		if (buf->dtb_xamot != NULL) {
12008 			ASSERT(buf->dtb_tomax != NULL);
12009 			ASSERT(buf->dtb_size == size);
12010 			kmem_free(buf->dtb_xamot, size);
12011 			allocated++;
12012 		}
12013 
12014 		if (buf->dtb_tomax != NULL) {
12015 			ASSERT(buf->dtb_size == size);
12016 			kmem_free(buf->dtb_tomax, size);
12017 			allocated++;
12018 		}
12019 
12020 		buf->dtb_tomax = NULL;
12021 		buf->dtb_xamot = NULL;
12022 		buf->dtb_size = 0;
12023 	} while ((cp = cp->cpu_next) != cpu_list);
12024 #else
12025 	int i;
12026 
12027 	*factor = 1;
12028 #if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \
12029     defined(__mips__) || defined(__powerpc__) || defined(__riscv__)
12030 	/*
12031 	 * FreeBSD isn't good at limiting the amount of memory we
12032 	 * ask to malloc, so let's place a limit here before trying
12033 	 * to do something that might well end in tears at bedtime.
12034 	 */
12035 	if (size > physmem * PAGE_SIZE / (128 * (mp_maxid + 1)))
12036 		return (ENOMEM);
12037 #endif
12038 
12039 	ASSERT(MUTEX_HELD(&dtrace_lock));
12040 	CPU_FOREACH(i) {
12041 		if (cpu != DTRACE_CPUALL && cpu != i)
12042 			continue;
12043 
12044 		buf = &bufs[i];
12045 
12046 		/*
12047 		 * If there is already a buffer allocated for this CPU, it
12048 		 * is only possible that this is a DR event.  In this case,
12049 		 * the buffer size must match our specified size.
12050 		 */
12051 		if (buf->dtb_tomax != NULL) {
12052 			ASSERT(buf->dtb_size == size);
12053 			continue;
12054 		}
12055 
12056 		ASSERT(buf->dtb_xamot == NULL);
12057 
12058 		if ((buf->dtb_tomax = kmem_zalloc(size,
12059 		    KM_NOSLEEP | KM_NORMALPRI)) == NULL)
12060 			goto err;
12061 
12062 		buf->dtb_size = size;
12063 		buf->dtb_flags = flags;
12064 		buf->dtb_offset = 0;
12065 		buf->dtb_drops = 0;
12066 
12067 		if (flags & DTRACEBUF_NOSWITCH)
12068 			continue;
12069 
12070 		if ((buf->dtb_xamot = kmem_zalloc(size,
12071 		    KM_NOSLEEP | KM_NORMALPRI)) == NULL)
12072 			goto err;
12073 	}
12074 
12075 	return (0);
12076 
12077 err:
12078 	/*
12079 	 * Error allocating memory, so free the buffers that were
12080 	 * allocated before the failed allocation.
12081 	 */
12082 	CPU_FOREACH(i) {
12083 		if (cpu != DTRACE_CPUALL && cpu != i)
12084 			continue;
12085 
12086 		buf = &bufs[i];
12087 		desired += 2;
12088 
12089 		if (buf->dtb_xamot != NULL) {
12090 			ASSERT(buf->dtb_tomax != NULL);
12091 			ASSERT(buf->dtb_size == size);
12092 			kmem_free(buf->dtb_xamot, size);
12093 			allocated++;
12094 		}
12095 
12096 		if (buf->dtb_tomax != NULL) {
12097 			ASSERT(buf->dtb_size == size);
12098 			kmem_free(buf->dtb_tomax, size);
12099 			allocated++;
12100 		}
12101 
12102 		buf->dtb_tomax = NULL;
12103 		buf->dtb_xamot = NULL;
12104 		buf->dtb_size = 0;
12105 
12106 	}
12107 #endif
12108 	*factor = desired / (allocated > 0 ? allocated : 1);
12109 
12110 	return (ENOMEM);
12111 }
12112 
12113 /*
12114  * Note:  called from probe context.  This function just increments the drop
12115  * count on a buffer.  It has been made a function to allow for the
12116  * possibility of understanding the source of mysterious drop counts.  (A
12117  * problem for which one may be particularly disappointed that DTrace cannot
12118  * be used to understand DTrace.)
12119  */
12120 static void
12121 dtrace_buffer_drop(dtrace_buffer_t *buf)
12122 {
12123 	buf->dtb_drops++;
12124 }
12125 
12126 /*
12127  * Note:  called from probe context.  This function is called to reserve space
12128  * in a buffer.  If mstate is non-NULL, sets the scratch base and size in the
12129  * mstate.  Returns the new offset in the buffer, or a negative value if an
12130  * error has occurred.
12131  */
12132 static intptr_t
12133 dtrace_buffer_reserve(dtrace_buffer_t *buf, size_t needed, size_t align,
12134     dtrace_state_t *state, dtrace_mstate_t *mstate)
12135 {
12136 	intptr_t offs = buf->dtb_offset, soffs;
12137 	intptr_t woffs;
12138 	caddr_t tomax;
12139 	size_t total;
12140 
12141 	if (buf->dtb_flags & DTRACEBUF_INACTIVE)
12142 		return (-1);
12143 
12144 	if ((tomax = buf->dtb_tomax) == NULL) {
12145 		dtrace_buffer_drop(buf);
12146 		return (-1);
12147 	}
12148 
12149 	if (!(buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL))) {
12150 		while (offs & (align - 1)) {
12151 			/*
12152 			 * Assert that our alignment is off by a number which
12153 			 * is itself sizeof (uint32_t) aligned.
12154 			 */
12155 			ASSERT(!((align - (offs & (align - 1))) &
12156 			    (sizeof (uint32_t) - 1)));
12157 			DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
12158 			offs += sizeof (uint32_t);
12159 		}
12160 
12161 		if ((soffs = offs + needed) > buf->dtb_size) {
12162 			dtrace_buffer_drop(buf);
12163 			return (-1);
12164 		}
12165 
12166 		if (mstate == NULL)
12167 			return (offs);
12168 
12169 		mstate->dtms_scratch_base = (uintptr_t)tomax + soffs;
12170 		mstate->dtms_scratch_size = buf->dtb_size - soffs;
12171 		mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
12172 
12173 		return (offs);
12174 	}
12175 
12176 	if (buf->dtb_flags & DTRACEBUF_FILL) {
12177 		if (state->dts_activity != DTRACE_ACTIVITY_COOLDOWN &&
12178 		    (buf->dtb_flags & DTRACEBUF_FULL))
12179 			return (-1);
12180 		goto out;
12181 	}
12182 
12183 	total = needed + (offs & (align - 1));
12184 
12185 	/*
12186 	 * For a ring buffer, life is quite a bit more complicated.  Before
12187 	 * we can store any padding, we need to adjust our wrapping offset.
12188 	 * (If we've never before wrapped or we're not about to, no adjustment
12189 	 * is required.)
12190 	 */
12191 	if ((buf->dtb_flags & DTRACEBUF_WRAPPED) ||
12192 	    offs + total > buf->dtb_size) {
12193 		woffs = buf->dtb_xamot_offset;
12194 
12195 		if (offs + total > buf->dtb_size) {
12196 			/*
12197 			 * We can't fit in the end of the buffer.  First, a
12198 			 * sanity check that we can fit in the buffer at all.
12199 			 */
12200 			if (total > buf->dtb_size) {
12201 				dtrace_buffer_drop(buf);
12202 				return (-1);
12203 			}
12204 
12205 			/*
12206 			 * We're going to be storing at the top of the buffer,
12207 			 * so now we need to deal with the wrapped offset.  We
12208 			 * only reset our wrapped offset to 0 if it is
12209 			 * currently greater than the current offset.  If it
12210 			 * is less than the current offset, it is because a
12211 			 * previous allocation induced a wrap -- but the
12212 			 * allocation didn't subsequently take the space due
12213 			 * to an error or false predicate evaluation.  In this
12214 			 * case, we'll just leave the wrapped offset alone: if
12215 			 * the wrapped offset hasn't been advanced far enough
12216 			 * for this allocation, it will be adjusted in the
12217 			 * lower loop.
12218 			 */
12219 			if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
12220 				if (woffs >= offs)
12221 					woffs = 0;
12222 			} else {
12223 				woffs = 0;
12224 			}
12225 
12226 			/*
12227 			 * Now we know that we're going to be storing to the
12228 			 * top of the buffer and that there is room for us
12229 			 * there.  We need to clear the buffer from the current
12230 			 * offset to the end (there may be old gunk there).
12231 			 */
12232 			while (offs < buf->dtb_size)
12233 				tomax[offs++] = 0;
12234 
12235 			/*
12236 			 * We need to set our offset to zero.  And because we
12237 			 * are wrapping, we need to set the bit indicating as
12238 			 * much.  We can also adjust our needed space back
12239 			 * down to the space required by the ECB -- we know
12240 			 * that the top of the buffer is aligned.
12241 			 */
12242 			offs = 0;
12243 			total = needed;
12244 			buf->dtb_flags |= DTRACEBUF_WRAPPED;
12245 		} else {
12246 			/*
12247 			 * There is room for us in the buffer, so we simply
12248 			 * need to check the wrapped offset.
12249 			 */
12250 			if (woffs < offs) {
12251 				/*
12252 				 * The wrapped offset is less than the offset.
12253 				 * This can happen if we allocated buffer space
12254 				 * that induced a wrap, but then we didn't
12255 				 * subsequently take the space due to an error
12256 				 * or false predicate evaluation.  This is
12257 				 * okay; we know that _this_ allocation isn't
12258 				 * going to induce a wrap.  We still can't
12259 				 * reset the wrapped offset to be zero,
12260 				 * however: the space may have been trashed in
12261 				 * the previous failed probe attempt.  But at
12262 				 * least the wrapped offset doesn't need to
12263 				 * be adjusted at all...
12264 				 */
12265 				goto out;
12266 			}
12267 		}
12268 
12269 		while (offs + total > woffs) {
12270 			dtrace_epid_t epid = *(uint32_t *)(tomax + woffs);
12271 			size_t size;
12272 
12273 			if (epid == DTRACE_EPIDNONE) {
12274 				size = sizeof (uint32_t);
12275 			} else {
12276 				ASSERT3U(epid, <=, state->dts_necbs);
12277 				ASSERT(state->dts_ecbs[epid - 1] != NULL);
12278 
12279 				size = state->dts_ecbs[epid - 1]->dte_size;
12280 			}
12281 
12282 			ASSERT(woffs + size <= buf->dtb_size);
12283 			ASSERT(size != 0);
12284 
12285 			if (woffs + size == buf->dtb_size) {
12286 				/*
12287 				 * We've reached the end of the buffer; we want
12288 				 * to set the wrapped offset to 0 and break
12289 				 * out.  However, if the offs is 0, then we're
12290 				 * in a strange edge-condition:  the amount of
12291 				 * space that we want to reserve plus the size
12292 				 * of the record that we're overwriting is
12293 				 * greater than the size of the buffer.  This
12294 				 * is problematic because if we reserve the
12295 				 * space but subsequently don't consume it (due
12296 				 * to a failed predicate or error) the wrapped
12297 				 * offset will be 0 -- yet the EPID at offset 0
12298 				 * will not be committed.  This situation is
12299 				 * relatively easy to deal with:  if we're in
12300 				 * this case, the buffer is indistinguishable
12301 				 * from one that hasn't wrapped; we need only
12302 				 * finish the job by clearing the wrapped bit,
12303 				 * explicitly setting the offset to be 0, and
12304 				 * zero'ing out the old data in the buffer.
12305 				 */
12306 				if (offs == 0) {
12307 					buf->dtb_flags &= ~DTRACEBUF_WRAPPED;
12308 					buf->dtb_offset = 0;
12309 					woffs = total;
12310 
12311 					while (woffs < buf->dtb_size)
12312 						tomax[woffs++] = 0;
12313 				}
12314 
12315 				woffs = 0;
12316 				break;
12317 			}
12318 
12319 			woffs += size;
12320 		}
12321 
12322 		/*
12323 		 * We have a wrapped offset.  It may be that the wrapped offset
12324 		 * has become zero -- that's okay.
12325 		 */
12326 		buf->dtb_xamot_offset = woffs;
12327 	}
12328 
12329 out:
12330 	/*
12331 	 * Now we can plow the buffer with any necessary padding.
12332 	 */
12333 	while (offs & (align - 1)) {
12334 		/*
12335 		 * Assert that our alignment is off by a number which
12336 		 * is itself sizeof (uint32_t) aligned.
12337 		 */
12338 		ASSERT(!((align - (offs & (align - 1))) &
12339 		    (sizeof (uint32_t) - 1)));
12340 		DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
12341 		offs += sizeof (uint32_t);
12342 	}
12343 
12344 	if (buf->dtb_flags & DTRACEBUF_FILL) {
12345 		if (offs + needed > buf->dtb_size - state->dts_reserve) {
12346 			buf->dtb_flags |= DTRACEBUF_FULL;
12347 			return (-1);
12348 		}
12349 	}
12350 
12351 	if (mstate == NULL)
12352 		return (offs);
12353 
12354 	/*
12355 	 * For ring buffers and fill buffers, the scratch space is always
12356 	 * the inactive buffer.
12357 	 */
12358 	mstate->dtms_scratch_base = (uintptr_t)buf->dtb_xamot;
12359 	mstate->dtms_scratch_size = buf->dtb_size;
12360 	mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
12361 
12362 	return (offs);
12363 }
12364 
12365 static void
12366 dtrace_buffer_polish(dtrace_buffer_t *buf)
12367 {
12368 	ASSERT(buf->dtb_flags & DTRACEBUF_RING);
12369 	ASSERT(MUTEX_HELD(&dtrace_lock));
12370 
12371 	if (!(buf->dtb_flags & DTRACEBUF_WRAPPED))
12372 		return;
12373 
12374 	/*
12375 	 * We need to polish the ring buffer.  There are three cases:
12376 	 *
12377 	 * - The first (and presumably most common) is that there is no gap
12378 	 *   between the buffer offset and the wrapped offset.  In this case,
12379 	 *   there is nothing in the buffer that isn't valid data; we can
12380 	 *   mark the buffer as polished and return.
12381 	 *
12382 	 * - The second (less common than the first but still more common
12383 	 *   than the third) is that there is a gap between the buffer offset
12384 	 *   and the wrapped offset, and the wrapped offset is larger than the
12385 	 *   buffer offset.  This can happen because of an alignment issue, or
12386 	 *   can happen because of a call to dtrace_buffer_reserve() that
12387 	 *   didn't subsequently consume the buffer space.  In this case,
12388 	 *   we need to zero the data from the buffer offset to the wrapped
12389 	 *   offset.
12390 	 *
12391 	 * - The third (and least common) is that there is a gap between the
12392 	 *   buffer offset and the wrapped offset, but the wrapped offset is
12393 	 *   _less_ than the buffer offset.  This can only happen because a
12394 	 *   call to dtrace_buffer_reserve() induced a wrap, but the space
12395 	 *   was not subsequently consumed.  In this case, we need to zero the
12396 	 *   space from the offset to the end of the buffer _and_ from the
12397 	 *   top of the buffer to the wrapped offset.
12398 	 */
12399 	if (buf->dtb_offset < buf->dtb_xamot_offset) {
12400 		bzero(buf->dtb_tomax + buf->dtb_offset,
12401 		    buf->dtb_xamot_offset - buf->dtb_offset);
12402 	}
12403 
12404 	if (buf->dtb_offset > buf->dtb_xamot_offset) {
12405 		bzero(buf->dtb_tomax + buf->dtb_offset,
12406 		    buf->dtb_size - buf->dtb_offset);
12407 		bzero(buf->dtb_tomax, buf->dtb_xamot_offset);
12408 	}
12409 }
12410 
12411 /*
12412  * This routine determines if data generated at the specified time has likely
12413  * been entirely consumed at user-level.  This routine is called to determine
12414  * if an ECB on a defunct probe (but for an active enabling) can be safely
12415  * disabled and destroyed.
12416  */
12417 static int
12418 dtrace_buffer_consumed(dtrace_buffer_t *bufs, hrtime_t when)
12419 {
12420 	int i;
12421 
12422 	for (i = 0; i < NCPU; i++) {
12423 		dtrace_buffer_t *buf = &bufs[i];
12424 
12425 		if (buf->dtb_size == 0)
12426 			continue;
12427 
12428 		if (buf->dtb_flags & DTRACEBUF_RING)
12429 			return (0);
12430 
12431 		if (!buf->dtb_switched && buf->dtb_offset != 0)
12432 			return (0);
12433 
12434 		if (buf->dtb_switched - buf->dtb_interval < when)
12435 			return (0);
12436 	}
12437 
12438 	return (1);
12439 }
12440 
12441 static void
12442 dtrace_buffer_free(dtrace_buffer_t *bufs)
12443 {
12444 	int i;
12445 
12446 	for (i = 0; i < NCPU; i++) {
12447 		dtrace_buffer_t *buf = &bufs[i];
12448 
12449 		if (buf->dtb_tomax == NULL) {
12450 			ASSERT(buf->dtb_xamot == NULL);
12451 			ASSERT(buf->dtb_size == 0);
12452 			continue;
12453 		}
12454 
12455 		if (buf->dtb_xamot != NULL) {
12456 			ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
12457 			kmem_free(buf->dtb_xamot, buf->dtb_size);
12458 		}
12459 
12460 		kmem_free(buf->dtb_tomax, buf->dtb_size);
12461 		buf->dtb_size = 0;
12462 		buf->dtb_tomax = NULL;
12463 		buf->dtb_xamot = NULL;
12464 	}
12465 }
12466 
12467 /*
12468  * DTrace Enabling Functions
12469  */
12470 static dtrace_enabling_t *
12471 dtrace_enabling_create(dtrace_vstate_t *vstate)
12472 {
12473 	dtrace_enabling_t *enab;
12474 
12475 	enab = kmem_zalloc(sizeof (dtrace_enabling_t), KM_SLEEP);
12476 	enab->dten_vstate = vstate;
12477 
12478 	return (enab);
12479 }
12480 
12481 static void
12482 dtrace_enabling_add(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb)
12483 {
12484 	dtrace_ecbdesc_t **ndesc;
12485 	size_t osize, nsize;
12486 
12487 	/*
12488 	 * We can't add to enablings after we've enabled them, or after we've
12489 	 * retained them.
12490 	 */
12491 	ASSERT(enab->dten_probegen == 0);
12492 	ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
12493 
12494 	if (enab->dten_ndesc < enab->dten_maxdesc) {
12495 		enab->dten_desc[enab->dten_ndesc++] = ecb;
12496 		return;
12497 	}
12498 
12499 	osize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
12500 
12501 	if (enab->dten_maxdesc == 0) {
12502 		enab->dten_maxdesc = 1;
12503 	} else {
12504 		enab->dten_maxdesc <<= 1;
12505 	}
12506 
12507 	ASSERT(enab->dten_ndesc < enab->dten_maxdesc);
12508 
12509 	nsize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
12510 	ndesc = kmem_zalloc(nsize, KM_SLEEP);
12511 	bcopy(enab->dten_desc, ndesc, osize);
12512 	if (enab->dten_desc != NULL)
12513 		kmem_free(enab->dten_desc, osize);
12514 
12515 	enab->dten_desc = ndesc;
12516 	enab->dten_desc[enab->dten_ndesc++] = ecb;
12517 }
12518 
12519 static void
12520 dtrace_enabling_addlike(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb,
12521     dtrace_probedesc_t *pd)
12522 {
12523 	dtrace_ecbdesc_t *new;
12524 	dtrace_predicate_t *pred;
12525 	dtrace_actdesc_t *act;
12526 
12527 	/*
12528 	 * We're going to create a new ECB description that matches the
12529 	 * specified ECB in every way, but has the specified probe description.
12530 	 */
12531 	new = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
12532 
12533 	if ((pred = ecb->dted_pred.dtpdd_predicate) != NULL)
12534 		dtrace_predicate_hold(pred);
12535 
12536 	for (act = ecb->dted_action; act != NULL; act = act->dtad_next)
12537 		dtrace_actdesc_hold(act);
12538 
12539 	new->dted_action = ecb->dted_action;
12540 	new->dted_pred = ecb->dted_pred;
12541 	new->dted_probe = *pd;
12542 	new->dted_uarg = ecb->dted_uarg;
12543 
12544 	dtrace_enabling_add(enab, new);
12545 }
12546 
12547 static void
12548 dtrace_enabling_dump(dtrace_enabling_t *enab)
12549 {
12550 	int i;
12551 
12552 	for (i = 0; i < enab->dten_ndesc; i++) {
12553 		dtrace_probedesc_t *desc = &enab->dten_desc[i]->dted_probe;
12554 
12555 #ifdef __FreeBSD__
12556 		printf("dtrace: enabling probe %d (%s:%s:%s:%s)\n", i,
12557 		    desc->dtpd_provider, desc->dtpd_mod,
12558 		    desc->dtpd_func, desc->dtpd_name);
12559 #else
12560 		cmn_err(CE_NOTE, "enabling probe %d (%s:%s:%s:%s)", i,
12561 		    desc->dtpd_provider, desc->dtpd_mod,
12562 		    desc->dtpd_func, desc->dtpd_name);
12563 #endif
12564 	}
12565 }
12566 
12567 static void
12568 dtrace_enabling_destroy(dtrace_enabling_t *enab)
12569 {
12570 	int i;
12571 	dtrace_ecbdesc_t *ep;
12572 	dtrace_vstate_t *vstate = enab->dten_vstate;
12573 
12574 	ASSERT(MUTEX_HELD(&dtrace_lock));
12575 
12576 	for (i = 0; i < enab->dten_ndesc; i++) {
12577 		dtrace_actdesc_t *act, *next;
12578 		dtrace_predicate_t *pred;
12579 
12580 		ep = enab->dten_desc[i];
12581 
12582 		if ((pred = ep->dted_pred.dtpdd_predicate) != NULL)
12583 			dtrace_predicate_release(pred, vstate);
12584 
12585 		for (act = ep->dted_action; act != NULL; act = next) {
12586 			next = act->dtad_next;
12587 			dtrace_actdesc_release(act, vstate);
12588 		}
12589 
12590 		kmem_free(ep, sizeof (dtrace_ecbdesc_t));
12591 	}
12592 
12593 	if (enab->dten_desc != NULL)
12594 		kmem_free(enab->dten_desc,
12595 		    enab->dten_maxdesc * sizeof (dtrace_enabling_t *));
12596 
12597 	/*
12598 	 * If this was a retained enabling, decrement the dts_nretained count
12599 	 * and take it off of the dtrace_retained list.
12600 	 */
12601 	if (enab->dten_prev != NULL || enab->dten_next != NULL ||
12602 	    dtrace_retained == enab) {
12603 		ASSERT(enab->dten_vstate->dtvs_state != NULL);
12604 		ASSERT(enab->dten_vstate->dtvs_state->dts_nretained > 0);
12605 		enab->dten_vstate->dtvs_state->dts_nretained--;
12606 		dtrace_retained_gen++;
12607 	}
12608 
12609 	if (enab->dten_prev == NULL) {
12610 		if (dtrace_retained == enab) {
12611 			dtrace_retained = enab->dten_next;
12612 
12613 			if (dtrace_retained != NULL)
12614 				dtrace_retained->dten_prev = NULL;
12615 		}
12616 	} else {
12617 		ASSERT(enab != dtrace_retained);
12618 		ASSERT(dtrace_retained != NULL);
12619 		enab->dten_prev->dten_next = enab->dten_next;
12620 	}
12621 
12622 	if (enab->dten_next != NULL) {
12623 		ASSERT(dtrace_retained != NULL);
12624 		enab->dten_next->dten_prev = enab->dten_prev;
12625 	}
12626 
12627 	kmem_free(enab, sizeof (dtrace_enabling_t));
12628 }
12629 
12630 static int
12631 dtrace_enabling_retain(dtrace_enabling_t *enab)
12632 {
12633 	dtrace_state_t *state;
12634 
12635 	ASSERT(MUTEX_HELD(&dtrace_lock));
12636 	ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
12637 	ASSERT(enab->dten_vstate != NULL);
12638 
12639 	state = enab->dten_vstate->dtvs_state;
12640 	ASSERT(state != NULL);
12641 
12642 	/*
12643 	 * We only allow each state to retain dtrace_retain_max enablings.
12644 	 */
12645 	if (state->dts_nretained >= dtrace_retain_max)
12646 		return (ENOSPC);
12647 
12648 	state->dts_nretained++;
12649 	dtrace_retained_gen++;
12650 
12651 	if (dtrace_retained == NULL) {
12652 		dtrace_retained = enab;
12653 		return (0);
12654 	}
12655 
12656 	enab->dten_next = dtrace_retained;
12657 	dtrace_retained->dten_prev = enab;
12658 	dtrace_retained = enab;
12659 
12660 	return (0);
12661 }
12662 
12663 static int
12664 dtrace_enabling_replicate(dtrace_state_t *state, dtrace_probedesc_t *match,
12665     dtrace_probedesc_t *create)
12666 {
12667 	dtrace_enabling_t *new, *enab;
12668 	int found = 0, err = ENOENT;
12669 
12670 	ASSERT(MUTEX_HELD(&dtrace_lock));
12671 	ASSERT(strlen(match->dtpd_provider) < DTRACE_PROVNAMELEN);
12672 	ASSERT(strlen(match->dtpd_mod) < DTRACE_MODNAMELEN);
12673 	ASSERT(strlen(match->dtpd_func) < DTRACE_FUNCNAMELEN);
12674 	ASSERT(strlen(match->dtpd_name) < DTRACE_NAMELEN);
12675 
12676 	new = dtrace_enabling_create(&state->dts_vstate);
12677 
12678 	/*
12679 	 * Iterate over all retained enablings, looking for enablings that
12680 	 * match the specified state.
12681 	 */
12682 	for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
12683 		int i;
12684 
12685 		/*
12686 		 * dtvs_state can only be NULL for helper enablings -- and
12687 		 * helper enablings can't be retained.
12688 		 */
12689 		ASSERT(enab->dten_vstate->dtvs_state != NULL);
12690 
12691 		if (enab->dten_vstate->dtvs_state != state)
12692 			continue;
12693 
12694 		/*
12695 		 * Now iterate over each probe description; we're looking for
12696 		 * an exact match to the specified probe description.
12697 		 */
12698 		for (i = 0; i < enab->dten_ndesc; i++) {
12699 			dtrace_ecbdesc_t *ep = enab->dten_desc[i];
12700 			dtrace_probedesc_t *pd = &ep->dted_probe;
12701 
12702 			if (strcmp(pd->dtpd_provider, match->dtpd_provider))
12703 				continue;
12704 
12705 			if (strcmp(pd->dtpd_mod, match->dtpd_mod))
12706 				continue;
12707 
12708 			if (strcmp(pd->dtpd_func, match->dtpd_func))
12709 				continue;
12710 
12711 			if (strcmp(pd->dtpd_name, match->dtpd_name))
12712 				continue;
12713 
12714 			/*
12715 			 * We have a winning probe!  Add it to our growing
12716 			 * enabling.
12717 			 */
12718 			found = 1;
12719 			dtrace_enabling_addlike(new, ep, create);
12720 		}
12721 	}
12722 
12723 	if (!found || (err = dtrace_enabling_retain(new)) != 0) {
12724 		dtrace_enabling_destroy(new);
12725 		return (err);
12726 	}
12727 
12728 	return (0);
12729 }
12730 
12731 static void
12732 dtrace_enabling_retract(dtrace_state_t *state)
12733 {
12734 	dtrace_enabling_t *enab, *next;
12735 
12736 	ASSERT(MUTEX_HELD(&dtrace_lock));
12737 
12738 	/*
12739 	 * Iterate over all retained enablings, destroy the enablings retained
12740 	 * for the specified state.
12741 	 */
12742 	for (enab = dtrace_retained; enab != NULL; enab = next) {
12743 		next = enab->dten_next;
12744 
12745 		/*
12746 		 * dtvs_state can only be NULL for helper enablings -- and
12747 		 * helper enablings can't be retained.
12748 		 */
12749 		ASSERT(enab->dten_vstate->dtvs_state != NULL);
12750 
12751 		if (enab->dten_vstate->dtvs_state == state) {
12752 			ASSERT(state->dts_nretained > 0);
12753 			dtrace_enabling_destroy(enab);
12754 		}
12755 	}
12756 
12757 	ASSERT(state->dts_nretained == 0);
12758 }
12759 
12760 static int
12761 dtrace_enabling_match(dtrace_enabling_t *enab, int *nmatched)
12762 {
12763 	int i = 0;
12764 	int matched = 0;
12765 
12766 	ASSERT(MUTEX_HELD(&cpu_lock));
12767 	ASSERT(MUTEX_HELD(&dtrace_lock));
12768 
12769 	for (i = 0; i < enab->dten_ndesc; i++) {
12770 		dtrace_ecbdesc_t *ep = enab->dten_desc[i];
12771 
12772 		enab->dten_current = ep;
12773 		enab->dten_error = 0;
12774 
12775 		matched += dtrace_probe_enable(&ep->dted_probe, enab);
12776 
12777 		if (enab->dten_error != 0) {
12778 			/*
12779 			 * If we get an error half-way through enabling the
12780 			 * probes, we kick out -- perhaps with some number of
12781 			 * them enabled.  Leaving enabled probes enabled may
12782 			 * be slightly confusing for user-level, but we expect
12783 			 * that no one will attempt to actually drive on in
12784 			 * the face of such errors.  If this is an anonymous
12785 			 * enabling (indicated with a NULL nmatched pointer),
12786 			 * we cmn_err() a message.  We aren't expecting to
12787 			 * get such an error -- such as it can exist at all,
12788 			 * it would be a result of corrupted DOF in the driver
12789 			 * properties.
12790 			 */
12791 			if (nmatched == NULL) {
12792 				cmn_err(CE_WARN, "dtrace_enabling_match() "
12793 				    "error on %p: %d", (void *)ep,
12794 				    enab->dten_error);
12795 			}
12796 
12797 			return (enab->dten_error);
12798 		}
12799 	}
12800 
12801 	enab->dten_probegen = dtrace_probegen;
12802 	if (nmatched != NULL)
12803 		*nmatched = matched;
12804 
12805 	return (0);
12806 }
12807 
12808 static void
12809 dtrace_enabling_matchall(void)
12810 {
12811 	dtrace_enabling_t *enab;
12812 
12813 	mutex_enter(&cpu_lock);
12814 	mutex_enter(&dtrace_lock);
12815 
12816 	/*
12817 	 * Iterate over all retained enablings to see if any probes match
12818 	 * against them.  We only perform this operation on enablings for which
12819 	 * we have sufficient permissions by virtue of being in the global zone
12820 	 * or in the same zone as the DTrace client.  Because we can be called
12821 	 * after dtrace_detach() has been called, we cannot assert that there
12822 	 * are retained enablings.  We can safely load from dtrace_retained,
12823 	 * however:  the taskq_destroy() at the end of dtrace_detach() will
12824 	 * block pending our completion.
12825 	 */
12826 	for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
12827 #ifdef illumos
12828 		cred_t *cr = enab->dten_vstate->dtvs_state->dts_cred.dcr_cred;
12829 
12830 		if (INGLOBALZONE(curproc) ||
12831 		    cr != NULL && getzoneid() == crgetzoneid(cr))
12832 #endif
12833 			(void) dtrace_enabling_match(enab, NULL);
12834 	}
12835 
12836 	mutex_exit(&dtrace_lock);
12837 	mutex_exit(&cpu_lock);
12838 }
12839 
12840 /*
12841  * If an enabling is to be enabled without having matched probes (that is, if
12842  * dtrace_state_go() is to be called on the underlying dtrace_state_t), the
12843  * enabling must be _primed_ by creating an ECB for every ECB description.
12844  * This must be done to assure that we know the number of speculations, the
12845  * number of aggregations, the minimum buffer size needed, etc. before we
12846  * transition out of DTRACE_ACTIVITY_INACTIVE.  To do this without actually
12847  * enabling any probes, we create ECBs for every ECB decription, but with a
12848  * NULL probe -- which is exactly what this function does.
12849  */
12850 static void
12851 dtrace_enabling_prime(dtrace_state_t *state)
12852 {
12853 	dtrace_enabling_t *enab;
12854 	int i;
12855 
12856 	for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
12857 		ASSERT(enab->dten_vstate->dtvs_state != NULL);
12858 
12859 		if (enab->dten_vstate->dtvs_state != state)
12860 			continue;
12861 
12862 		/*
12863 		 * We don't want to prime an enabling more than once, lest
12864 		 * we allow a malicious user to induce resource exhaustion.
12865 		 * (The ECBs that result from priming an enabling aren't
12866 		 * leaked -- but they also aren't deallocated until the
12867 		 * consumer state is destroyed.)
12868 		 */
12869 		if (enab->dten_primed)
12870 			continue;
12871 
12872 		for (i = 0; i < enab->dten_ndesc; i++) {
12873 			enab->dten_current = enab->dten_desc[i];
12874 			(void) dtrace_probe_enable(NULL, enab);
12875 		}
12876 
12877 		enab->dten_primed = 1;
12878 	}
12879 }
12880 
12881 /*
12882  * Called to indicate that probes should be provided due to retained
12883  * enablings.  This is implemented in terms of dtrace_probe_provide(), but it
12884  * must take an initial lap through the enabling calling the dtps_provide()
12885  * entry point explicitly to allow for autocreated probes.
12886  */
12887 static void
12888 dtrace_enabling_provide(dtrace_provider_t *prv)
12889 {
12890 	int i, all = 0;
12891 	dtrace_probedesc_t desc;
12892 	dtrace_genid_t gen;
12893 
12894 	ASSERT(MUTEX_HELD(&dtrace_lock));
12895 	ASSERT(MUTEX_HELD(&dtrace_provider_lock));
12896 
12897 	if (prv == NULL) {
12898 		all = 1;
12899 		prv = dtrace_provider;
12900 	}
12901 
12902 	do {
12903 		dtrace_enabling_t *enab;
12904 		void *parg = prv->dtpv_arg;
12905 
12906 retry:
12907 		gen = dtrace_retained_gen;
12908 		for (enab = dtrace_retained; enab != NULL;
12909 		    enab = enab->dten_next) {
12910 			for (i = 0; i < enab->dten_ndesc; i++) {
12911 				desc = enab->dten_desc[i]->dted_probe;
12912 				mutex_exit(&dtrace_lock);
12913 				prv->dtpv_pops.dtps_provide(parg, &desc);
12914 				mutex_enter(&dtrace_lock);
12915 				/*
12916 				 * Process the retained enablings again if
12917 				 * they have changed while we weren't holding
12918 				 * dtrace_lock.
12919 				 */
12920 				if (gen != dtrace_retained_gen)
12921 					goto retry;
12922 			}
12923 		}
12924 	} while (all && (prv = prv->dtpv_next) != NULL);
12925 
12926 	mutex_exit(&dtrace_lock);
12927 	dtrace_probe_provide(NULL, all ? NULL : prv);
12928 	mutex_enter(&dtrace_lock);
12929 }
12930 
12931 /*
12932  * Called to reap ECBs that are attached to probes from defunct providers.
12933  */
12934 static void
12935 dtrace_enabling_reap(void)
12936 {
12937 	dtrace_provider_t *prov;
12938 	dtrace_probe_t *probe;
12939 	dtrace_ecb_t *ecb;
12940 	hrtime_t when;
12941 	int i;
12942 
12943 	mutex_enter(&cpu_lock);
12944 	mutex_enter(&dtrace_lock);
12945 
12946 	for (i = 0; i < dtrace_nprobes; i++) {
12947 		if ((probe = dtrace_probes[i]) == NULL)
12948 			continue;
12949 
12950 		if (probe->dtpr_ecb == NULL)
12951 			continue;
12952 
12953 		prov = probe->dtpr_provider;
12954 
12955 		if ((when = prov->dtpv_defunct) == 0)
12956 			continue;
12957 
12958 		/*
12959 		 * We have ECBs on a defunct provider:  we want to reap these
12960 		 * ECBs to allow the provider to unregister.  The destruction
12961 		 * of these ECBs must be done carefully:  if we destroy the ECB
12962 		 * and the consumer later wishes to consume an EPID that
12963 		 * corresponds to the destroyed ECB (and if the EPID metadata
12964 		 * has not been previously consumed), the consumer will abort
12965 		 * processing on the unknown EPID.  To reduce (but not, sadly,
12966 		 * eliminate) the possibility of this, we will only destroy an
12967 		 * ECB for a defunct provider if, for the state that
12968 		 * corresponds to the ECB:
12969 		 *
12970 		 *  (a)	There is no speculative tracing (which can effectively
12971 		 *	cache an EPID for an arbitrary amount of time).
12972 		 *
12973 		 *  (b)	The principal buffers have been switched twice since the
12974 		 *	provider became defunct.
12975 		 *
12976 		 *  (c)	The aggregation buffers are of zero size or have been
12977 		 *	switched twice since the provider became defunct.
12978 		 *
12979 		 * We use dts_speculates to determine (a) and call a function
12980 		 * (dtrace_buffer_consumed()) to determine (b) and (c).  Note
12981 		 * that as soon as we've been unable to destroy one of the ECBs
12982 		 * associated with the probe, we quit trying -- reaping is only
12983 		 * fruitful in as much as we can destroy all ECBs associated
12984 		 * with the defunct provider's probes.
12985 		 */
12986 		while ((ecb = probe->dtpr_ecb) != NULL) {
12987 			dtrace_state_t *state = ecb->dte_state;
12988 			dtrace_buffer_t *buf = state->dts_buffer;
12989 			dtrace_buffer_t *aggbuf = state->dts_aggbuffer;
12990 
12991 			if (state->dts_speculates)
12992 				break;
12993 
12994 			if (!dtrace_buffer_consumed(buf, when))
12995 				break;
12996 
12997 			if (!dtrace_buffer_consumed(aggbuf, when))
12998 				break;
12999 
13000 			dtrace_ecb_disable(ecb);
13001 			ASSERT(probe->dtpr_ecb != ecb);
13002 			dtrace_ecb_destroy(ecb);
13003 		}
13004 	}
13005 
13006 	mutex_exit(&dtrace_lock);
13007 	mutex_exit(&cpu_lock);
13008 }
13009 
13010 /*
13011  * DTrace DOF Functions
13012  */
13013 /*ARGSUSED*/
13014 static void
13015 dtrace_dof_error(dof_hdr_t *dof, const char *str)
13016 {
13017 	if (dtrace_err_verbose)
13018 		cmn_err(CE_WARN, "failed to process DOF: %s", str);
13019 
13020 #ifdef DTRACE_ERRDEBUG
13021 	dtrace_errdebug(str);
13022 #endif
13023 }
13024 
13025 /*
13026  * Create DOF out of a currently enabled state.  Right now, we only create
13027  * DOF containing the run-time options -- but this could be expanded to create
13028  * complete DOF representing the enabled state.
13029  */
13030 static dof_hdr_t *
13031 dtrace_dof_create(dtrace_state_t *state)
13032 {
13033 	dof_hdr_t *dof;
13034 	dof_sec_t *sec;
13035 	dof_optdesc_t *opt;
13036 	int i, len = sizeof (dof_hdr_t) +
13037 	    roundup(sizeof (dof_sec_t), sizeof (uint64_t)) +
13038 	    sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
13039 
13040 	ASSERT(MUTEX_HELD(&dtrace_lock));
13041 
13042 	dof = kmem_zalloc(len, KM_SLEEP);
13043 	dof->dofh_ident[DOF_ID_MAG0] = DOF_MAG_MAG0;
13044 	dof->dofh_ident[DOF_ID_MAG1] = DOF_MAG_MAG1;
13045 	dof->dofh_ident[DOF_ID_MAG2] = DOF_MAG_MAG2;
13046 	dof->dofh_ident[DOF_ID_MAG3] = DOF_MAG_MAG3;
13047 
13048 	dof->dofh_ident[DOF_ID_MODEL] = DOF_MODEL_NATIVE;
13049 	dof->dofh_ident[DOF_ID_ENCODING] = DOF_ENCODE_NATIVE;
13050 	dof->dofh_ident[DOF_ID_VERSION] = DOF_VERSION;
13051 	dof->dofh_ident[DOF_ID_DIFVERS] = DIF_VERSION;
13052 	dof->dofh_ident[DOF_ID_DIFIREG] = DIF_DIR_NREGS;
13053 	dof->dofh_ident[DOF_ID_DIFTREG] = DIF_DTR_NREGS;
13054 
13055 	dof->dofh_flags = 0;
13056 	dof->dofh_hdrsize = sizeof (dof_hdr_t);
13057 	dof->dofh_secsize = sizeof (dof_sec_t);
13058 	dof->dofh_secnum = 1;	/* only DOF_SECT_OPTDESC */
13059 	dof->dofh_secoff = sizeof (dof_hdr_t);
13060 	dof->dofh_loadsz = len;
13061 	dof->dofh_filesz = len;
13062 	dof->dofh_pad = 0;
13063 
13064 	/*
13065 	 * Fill in the option section header...
13066 	 */
13067 	sec = (dof_sec_t *)((uintptr_t)dof + sizeof (dof_hdr_t));
13068 	sec->dofs_type = DOF_SECT_OPTDESC;
13069 	sec->dofs_align = sizeof (uint64_t);
13070 	sec->dofs_flags = DOF_SECF_LOAD;
13071 	sec->dofs_entsize = sizeof (dof_optdesc_t);
13072 
13073 	opt = (dof_optdesc_t *)((uintptr_t)sec +
13074 	    roundup(sizeof (dof_sec_t), sizeof (uint64_t)));
13075 
13076 	sec->dofs_offset = (uintptr_t)opt - (uintptr_t)dof;
13077 	sec->dofs_size = sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
13078 
13079 	for (i = 0; i < DTRACEOPT_MAX; i++) {
13080 		opt[i].dofo_option = i;
13081 		opt[i].dofo_strtab = DOF_SECIDX_NONE;
13082 		opt[i].dofo_value = state->dts_options[i];
13083 	}
13084 
13085 	return (dof);
13086 }
13087 
13088 static dof_hdr_t *
13089 dtrace_dof_copyin(uintptr_t uarg, int *errp)
13090 {
13091 	dof_hdr_t hdr, *dof;
13092 
13093 	ASSERT(!MUTEX_HELD(&dtrace_lock));
13094 
13095 	/*
13096 	 * First, we're going to copyin() the sizeof (dof_hdr_t).
13097 	 */
13098 	if (copyin((void *)uarg, &hdr, sizeof (hdr)) != 0) {
13099 		dtrace_dof_error(NULL, "failed to copyin DOF header");
13100 		*errp = EFAULT;
13101 		return (NULL);
13102 	}
13103 
13104 	/*
13105 	 * Now we'll allocate the entire DOF and copy it in -- provided
13106 	 * that the length isn't outrageous.
13107 	 */
13108 	if (hdr.dofh_loadsz >= dtrace_dof_maxsize) {
13109 		dtrace_dof_error(&hdr, "load size exceeds maximum");
13110 		*errp = E2BIG;
13111 		return (NULL);
13112 	}
13113 
13114 	if (hdr.dofh_loadsz < sizeof (hdr)) {
13115 		dtrace_dof_error(&hdr, "invalid load size");
13116 		*errp = EINVAL;
13117 		return (NULL);
13118 	}
13119 
13120 	dof = kmem_alloc(hdr.dofh_loadsz, KM_SLEEP);
13121 
13122 	if (copyin((void *)uarg, dof, hdr.dofh_loadsz) != 0 ||
13123 	    dof->dofh_loadsz != hdr.dofh_loadsz) {
13124 		kmem_free(dof, hdr.dofh_loadsz);
13125 		*errp = EFAULT;
13126 		return (NULL);
13127 	}
13128 
13129 	return (dof);
13130 }
13131 
13132 #ifdef __FreeBSD__
13133 static dof_hdr_t *
13134 dtrace_dof_copyin_proc(struct proc *p, uintptr_t uarg, int *errp)
13135 {
13136 	dof_hdr_t hdr, *dof;
13137 	struct thread *td;
13138 	size_t loadsz;
13139 
13140 	ASSERT(!MUTEX_HELD(&dtrace_lock));
13141 
13142 	td = curthread;
13143 
13144 	/*
13145 	 * First, we're going to copyin() the sizeof (dof_hdr_t).
13146 	 */
13147 	if (proc_readmem(td, p, uarg, &hdr, sizeof(hdr)) != sizeof(hdr)) {
13148 		dtrace_dof_error(NULL, "failed to copyin DOF header");
13149 		*errp = EFAULT;
13150 		return (NULL);
13151 	}
13152 
13153 	/*
13154 	 * Now we'll allocate the entire DOF and copy it in -- provided
13155 	 * that the length isn't outrageous.
13156 	 */
13157 	if (hdr.dofh_loadsz >= dtrace_dof_maxsize) {
13158 		dtrace_dof_error(&hdr, "load size exceeds maximum");
13159 		*errp = E2BIG;
13160 		return (NULL);
13161 	}
13162 	loadsz = (size_t)hdr.dofh_loadsz;
13163 
13164 	if (loadsz < sizeof (hdr)) {
13165 		dtrace_dof_error(&hdr, "invalid load size");
13166 		*errp = EINVAL;
13167 		return (NULL);
13168 	}
13169 
13170 	dof = kmem_alloc(loadsz, KM_SLEEP);
13171 
13172 	if (proc_readmem(td, p, uarg, dof, loadsz) != loadsz ||
13173 	    dof->dofh_loadsz != loadsz) {
13174 		kmem_free(dof, hdr.dofh_loadsz);
13175 		*errp = EFAULT;
13176 		return (NULL);
13177 	}
13178 
13179 	return (dof);
13180 }
13181 
13182 static __inline uchar_t
13183 dtrace_dof_char(char c)
13184 {
13185 
13186 	switch (c) {
13187 	case '0':
13188 	case '1':
13189 	case '2':
13190 	case '3':
13191 	case '4':
13192 	case '5':
13193 	case '6':
13194 	case '7':
13195 	case '8':
13196 	case '9':
13197 		return (c - '0');
13198 	case 'A':
13199 	case 'B':
13200 	case 'C':
13201 	case 'D':
13202 	case 'E':
13203 	case 'F':
13204 		return (c - 'A' + 10);
13205 	case 'a':
13206 	case 'b':
13207 	case 'c':
13208 	case 'd':
13209 	case 'e':
13210 	case 'f':
13211 		return (c - 'a' + 10);
13212 	}
13213 	/* Should not reach here. */
13214 	return (UCHAR_MAX);
13215 }
13216 #endif /* __FreeBSD__ */
13217 
13218 static dof_hdr_t *
13219 dtrace_dof_property(const char *name)
13220 {
13221 #ifdef __FreeBSD__
13222 	uint8_t *dofbuf;
13223 	u_char *data, *eol;
13224 	caddr_t doffile;
13225 	size_t bytes, len, i;
13226 	dof_hdr_t *dof;
13227 	u_char c1, c2;
13228 
13229 	dof = NULL;
13230 
13231 	doffile = preload_search_by_type("dtrace_dof");
13232 	if (doffile == NULL)
13233 		return (NULL);
13234 
13235 	data = preload_fetch_addr(doffile);
13236 	len = preload_fetch_size(doffile);
13237 	for (;;) {
13238 		/* Look for the end of the line. All lines end in a newline. */
13239 		eol = memchr(data, '\n', len);
13240 		if (eol == NULL)
13241 			return (NULL);
13242 
13243 		if (strncmp(name, data, strlen(name)) == 0)
13244 			break;
13245 
13246 		eol++; /* skip past the newline */
13247 		len -= eol - data;
13248 		data = eol;
13249 	}
13250 
13251 	/* We've found the data corresponding to the specified key. */
13252 
13253 	data += strlen(name) + 1; /* skip past the '=' */
13254 	len = eol - data;
13255 	bytes = len / 2;
13256 
13257 	if (bytes < sizeof(dof_hdr_t)) {
13258 		dtrace_dof_error(NULL, "truncated header");
13259 		goto doferr;
13260 	}
13261 
13262 	/*
13263 	 * Each byte is represented by the two ASCII characters in its hex
13264 	 * representation.
13265 	 */
13266 	dofbuf = malloc(bytes, M_SOLARIS, M_WAITOK);
13267 	for (i = 0; i < bytes; i++) {
13268 		c1 = dtrace_dof_char(data[i * 2]);
13269 		c2 = dtrace_dof_char(data[i * 2 + 1]);
13270 		if (c1 == UCHAR_MAX || c2 == UCHAR_MAX) {
13271 			dtrace_dof_error(NULL, "invalid hex char in DOF");
13272 			goto doferr;
13273 		}
13274 		dofbuf[i] = c1 * 16 + c2;
13275 	}
13276 
13277 	dof = (dof_hdr_t *)dofbuf;
13278 	if (bytes < dof->dofh_loadsz) {
13279 		dtrace_dof_error(NULL, "truncated DOF");
13280 		goto doferr;
13281 	}
13282 
13283 	if (dof->dofh_loadsz >= dtrace_dof_maxsize) {
13284 		dtrace_dof_error(NULL, "oversized DOF");
13285 		goto doferr;
13286 	}
13287 
13288 	return (dof);
13289 
13290 doferr:
13291 	free(dof, M_SOLARIS);
13292 	return (NULL);
13293 #else /* __FreeBSD__ */
13294 	uchar_t *buf;
13295 	uint64_t loadsz;
13296 	unsigned int len, i;
13297 	dof_hdr_t *dof;
13298 
13299 	/*
13300 	 * Unfortunately, array of values in .conf files are always (and
13301 	 * only) interpreted to be integer arrays.  We must read our DOF
13302 	 * as an integer array, and then squeeze it into a byte array.
13303 	 */
13304 	if (ddi_prop_lookup_int_array(DDI_DEV_T_ANY, dtrace_devi, 0,
13305 	    (char *)name, (int **)&buf, &len) != DDI_PROP_SUCCESS)
13306 		return (NULL);
13307 
13308 	for (i = 0; i < len; i++)
13309 		buf[i] = (uchar_t)(((int *)buf)[i]);
13310 
13311 	if (len < sizeof (dof_hdr_t)) {
13312 		ddi_prop_free(buf);
13313 		dtrace_dof_error(NULL, "truncated header");
13314 		return (NULL);
13315 	}
13316 
13317 	if (len < (loadsz = ((dof_hdr_t *)buf)->dofh_loadsz)) {
13318 		ddi_prop_free(buf);
13319 		dtrace_dof_error(NULL, "truncated DOF");
13320 		return (NULL);
13321 	}
13322 
13323 	if (loadsz >= dtrace_dof_maxsize) {
13324 		ddi_prop_free(buf);
13325 		dtrace_dof_error(NULL, "oversized DOF");
13326 		return (NULL);
13327 	}
13328 
13329 	dof = kmem_alloc(loadsz, KM_SLEEP);
13330 	bcopy(buf, dof, loadsz);
13331 	ddi_prop_free(buf);
13332 
13333 	return (dof);
13334 #endif /* !__FreeBSD__ */
13335 }
13336 
13337 static void
13338 dtrace_dof_destroy(dof_hdr_t *dof)
13339 {
13340 	kmem_free(dof, dof->dofh_loadsz);
13341 }
13342 
13343 /*
13344  * Return the dof_sec_t pointer corresponding to a given section index.  If the
13345  * index is not valid, dtrace_dof_error() is called and NULL is returned.  If
13346  * a type other than DOF_SECT_NONE is specified, the header is checked against
13347  * this type and NULL is returned if the types do not match.
13348  */
13349 static dof_sec_t *
13350 dtrace_dof_sect(dof_hdr_t *dof, uint32_t type, dof_secidx_t i)
13351 {
13352 	dof_sec_t *sec = (dof_sec_t *)(uintptr_t)
13353 	    ((uintptr_t)dof + dof->dofh_secoff + i * dof->dofh_secsize);
13354 
13355 	if (i >= dof->dofh_secnum) {
13356 		dtrace_dof_error(dof, "referenced section index is invalid");
13357 		return (NULL);
13358 	}
13359 
13360 	if (!(sec->dofs_flags & DOF_SECF_LOAD)) {
13361 		dtrace_dof_error(dof, "referenced section is not loadable");
13362 		return (NULL);
13363 	}
13364 
13365 	if (type != DOF_SECT_NONE && type != sec->dofs_type) {
13366 		dtrace_dof_error(dof, "referenced section is the wrong type");
13367 		return (NULL);
13368 	}
13369 
13370 	return (sec);
13371 }
13372 
13373 static dtrace_probedesc_t *
13374 dtrace_dof_probedesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_probedesc_t *desc)
13375 {
13376 	dof_probedesc_t *probe;
13377 	dof_sec_t *strtab;
13378 	uintptr_t daddr = (uintptr_t)dof;
13379 	uintptr_t str;
13380 	size_t size;
13381 
13382 	if (sec->dofs_type != DOF_SECT_PROBEDESC) {
13383 		dtrace_dof_error(dof, "invalid probe section");
13384 		return (NULL);
13385 	}
13386 
13387 	if (sec->dofs_align != sizeof (dof_secidx_t)) {
13388 		dtrace_dof_error(dof, "bad alignment in probe description");
13389 		return (NULL);
13390 	}
13391 
13392 	if (sec->dofs_offset + sizeof (dof_probedesc_t) > dof->dofh_loadsz) {
13393 		dtrace_dof_error(dof, "truncated probe description");
13394 		return (NULL);
13395 	}
13396 
13397 	probe = (dof_probedesc_t *)(uintptr_t)(daddr + sec->dofs_offset);
13398 	strtab = dtrace_dof_sect(dof, DOF_SECT_STRTAB, probe->dofp_strtab);
13399 
13400 	if (strtab == NULL)
13401 		return (NULL);
13402 
13403 	str = daddr + strtab->dofs_offset;
13404 	size = strtab->dofs_size;
13405 
13406 	if (probe->dofp_provider >= strtab->dofs_size) {
13407 		dtrace_dof_error(dof, "corrupt probe provider");
13408 		return (NULL);
13409 	}
13410 
13411 	(void) strncpy(desc->dtpd_provider,
13412 	    (char *)(str + probe->dofp_provider),
13413 	    MIN(DTRACE_PROVNAMELEN - 1, size - probe->dofp_provider));
13414 
13415 	if (probe->dofp_mod >= strtab->dofs_size) {
13416 		dtrace_dof_error(dof, "corrupt probe module");
13417 		return (NULL);
13418 	}
13419 
13420 	(void) strncpy(desc->dtpd_mod, (char *)(str + probe->dofp_mod),
13421 	    MIN(DTRACE_MODNAMELEN - 1, size - probe->dofp_mod));
13422 
13423 	if (probe->dofp_func >= strtab->dofs_size) {
13424 		dtrace_dof_error(dof, "corrupt probe function");
13425 		return (NULL);
13426 	}
13427 
13428 	(void) strncpy(desc->dtpd_func, (char *)(str + probe->dofp_func),
13429 	    MIN(DTRACE_FUNCNAMELEN - 1, size - probe->dofp_func));
13430 
13431 	if (probe->dofp_name >= strtab->dofs_size) {
13432 		dtrace_dof_error(dof, "corrupt probe name");
13433 		return (NULL);
13434 	}
13435 
13436 	(void) strncpy(desc->dtpd_name, (char *)(str + probe->dofp_name),
13437 	    MIN(DTRACE_NAMELEN - 1, size - probe->dofp_name));
13438 
13439 	return (desc);
13440 }
13441 
13442 static dtrace_difo_t *
13443 dtrace_dof_difo(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
13444     cred_t *cr)
13445 {
13446 	dtrace_difo_t *dp;
13447 	size_t ttl = 0;
13448 	dof_difohdr_t *dofd;
13449 	uintptr_t daddr = (uintptr_t)dof;
13450 	size_t max = dtrace_difo_maxsize;
13451 	int i, l, n;
13452 
13453 	static const struct {
13454 		int section;
13455 		int bufoffs;
13456 		int lenoffs;
13457 		int entsize;
13458 		int align;
13459 		const char *msg;
13460 	} difo[] = {
13461 		{ DOF_SECT_DIF, offsetof(dtrace_difo_t, dtdo_buf),
13462 		offsetof(dtrace_difo_t, dtdo_len), sizeof (dif_instr_t),
13463 		sizeof (dif_instr_t), "multiple DIF sections" },
13464 
13465 		{ DOF_SECT_INTTAB, offsetof(dtrace_difo_t, dtdo_inttab),
13466 		offsetof(dtrace_difo_t, dtdo_intlen), sizeof (uint64_t),
13467 		sizeof (uint64_t), "multiple integer tables" },
13468 
13469 		{ DOF_SECT_STRTAB, offsetof(dtrace_difo_t, dtdo_strtab),
13470 		offsetof(dtrace_difo_t, dtdo_strlen), 0,
13471 		sizeof (char), "multiple string tables" },
13472 
13473 		{ DOF_SECT_VARTAB, offsetof(dtrace_difo_t, dtdo_vartab),
13474 		offsetof(dtrace_difo_t, dtdo_varlen), sizeof (dtrace_difv_t),
13475 		sizeof (uint_t), "multiple variable tables" },
13476 
13477 		{ DOF_SECT_NONE, 0, 0, 0, 0, NULL }
13478 	};
13479 
13480 	if (sec->dofs_type != DOF_SECT_DIFOHDR) {
13481 		dtrace_dof_error(dof, "invalid DIFO header section");
13482 		return (NULL);
13483 	}
13484 
13485 	if (sec->dofs_align != sizeof (dof_secidx_t)) {
13486 		dtrace_dof_error(dof, "bad alignment in DIFO header");
13487 		return (NULL);
13488 	}
13489 
13490 	if (sec->dofs_size < sizeof (dof_difohdr_t) ||
13491 	    sec->dofs_size % sizeof (dof_secidx_t)) {
13492 		dtrace_dof_error(dof, "bad size in DIFO header");
13493 		return (NULL);
13494 	}
13495 
13496 	dofd = (dof_difohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
13497 	n = (sec->dofs_size - sizeof (*dofd)) / sizeof (dof_secidx_t) + 1;
13498 
13499 	dp = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
13500 	dp->dtdo_rtype = dofd->dofd_rtype;
13501 
13502 	for (l = 0; l < n; l++) {
13503 		dof_sec_t *subsec;
13504 		void **bufp;
13505 		uint32_t *lenp;
13506 
13507 		if ((subsec = dtrace_dof_sect(dof, DOF_SECT_NONE,
13508 		    dofd->dofd_links[l])) == NULL)
13509 			goto err; /* invalid section link */
13510 
13511 		if (ttl + subsec->dofs_size > max) {
13512 			dtrace_dof_error(dof, "exceeds maximum size");
13513 			goto err;
13514 		}
13515 
13516 		ttl += subsec->dofs_size;
13517 
13518 		for (i = 0; difo[i].section != DOF_SECT_NONE; i++) {
13519 			if (subsec->dofs_type != difo[i].section)
13520 				continue;
13521 
13522 			if (!(subsec->dofs_flags & DOF_SECF_LOAD)) {
13523 				dtrace_dof_error(dof, "section not loaded");
13524 				goto err;
13525 			}
13526 
13527 			if (subsec->dofs_align != difo[i].align) {
13528 				dtrace_dof_error(dof, "bad alignment");
13529 				goto err;
13530 			}
13531 
13532 			bufp = (void **)((uintptr_t)dp + difo[i].bufoffs);
13533 			lenp = (uint32_t *)((uintptr_t)dp + difo[i].lenoffs);
13534 
13535 			if (*bufp != NULL) {
13536 				dtrace_dof_error(dof, difo[i].msg);
13537 				goto err;
13538 			}
13539 
13540 			if (difo[i].entsize != subsec->dofs_entsize) {
13541 				dtrace_dof_error(dof, "entry size mismatch");
13542 				goto err;
13543 			}
13544 
13545 			if (subsec->dofs_entsize != 0 &&
13546 			    (subsec->dofs_size % subsec->dofs_entsize) != 0) {
13547 				dtrace_dof_error(dof, "corrupt entry size");
13548 				goto err;
13549 			}
13550 
13551 			*lenp = subsec->dofs_size;
13552 			*bufp = kmem_alloc(subsec->dofs_size, KM_SLEEP);
13553 			bcopy((char *)(uintptr_t)(daddr + subsec->dofs_offset),
13554 			    *bufp, subsec->dofs_size);
13555 
13556 			if (subsec->dofs_entsize != 0)
13557 				*lenp /= subsec->dofs_entsize;
13558 
13559 			break;
13560 		}
13561 
13562 		/*
13563 		 * If we encounter a loadable DIFO sub-section that is not
13564 		 * known to us, assume this is a broken program and fail.
13565 		 */
13566 		if (difo[i].section == DOF_SECT_NONE &&
13567 		    (subsec->dofs_flags & DOF_SECF_LOAD)) {
13568 			dtrace_dof_error(dof, "unrecognized DIFO subsection");
13569 			goto err;
13570 		}
13571 	}
13572 
13573 	if (dp->dtdo_buf == NULL) {
13574 		/*
13575 		 * We can't have a DIF object without DIF text.
13576 		 */
13577 		dtrace_dof_error(dof, "missing DIF text");
13578 		goto err;
13579 	}
13580 
13581 	/*
13582 	 * Before we validate the DIF object, run through the variable table
13583 	 * looking for the strings -- if any of their size are under, we'll set
13584 	 * their size to be the system-wide default string size.  Note that
13585 	 * this should _not_ happen if the "strsize" option has been set --
13586 	 * in this case, the compiler should have set the size to reflect the
13587 	 * setting of the option.
13588 	 */
13589 	for (i = 0; i < dp->dtdo_varlen; i++) {
13590 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
13591 		dtrace_diftype_t *t = &v->dtdv_type;
13592 
13593 		if (v->dtdv_id < DIF_VAR_OTHER_UBASE)
13594 			continue;
13595 
13596 		if (t->dtdt_kind == DIF_TYPE_STRING && t->dtdt_size == 0)
13597 			t->dtdt_size = dtrace_strsize_default;
13598 	}
13599 
13600 	if (dtrace_difo_validate(dp, vstate, DIF_DIR_NREGS, cr) != 0)
13601 		goto err;
13602 
13603 	dtrace_difo_init(dp, vstate);
13604 	return (dp);
13605 
13606 err:
13607 	kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
13608 	kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
13609 	kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
13610 	kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
13611 
13612 	kmem_free(dp, sizeof (dtrace_difo_t));
13613 	return (NULL);
13614 }
13615 
13616 static dtrace_predicate_t *
13617 dtrace_dof_predicate(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
13618     cred_t *cr)
13619 {
13620 	dtrace_difo_t *dp;
13621 
13622 	if ((dp = dtrace_dof_difo(dof, sec, vstate, cr)) == NULL)
13623 		return (NULL);
13624 
13625 	return (dtrace_predicate_create(dp));
13626 }
13627 
13628 static dtrace_actdesc_t *
13629 dtrace_dof_actdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
13630     cred_t *cr)
13631 {
13632 	dtrace_actdesc_t *act, *first = NULL, *last = NULL, *next;
13633 	dof_actdesc_t *desc;
13634 	dof_sec_t *difosec;
13635 	size_t offs;
13636 	uintptr_t daddr = (uintptr_t)dof;
13637 	uint64_t arg;
13638 	dtrace_actkind_t kind;
13639 
13640 	if (sec->dofs_type != DOF_SECT_ACTDESC) {
13641 		dtrace_dof_error(dof, "invalid action section");
13642 		return (NULL);
13643 	}
13644 
13645 	if (sec->dofs_offset + sizeof (dof_actdesc_t) > dof->dofh_loadsz) {
13646 		dtrace_dof_error(dof, "truncated action description");
13647 		return (NULL);
13648 	}
13649 
13650 	if (sec->dofs_align != sizeof (uint64_t)) {
13651 		dtrace_dof_error(dof, "bad alignment in action description");
13652 		return (NULL);
13653 	}
13654 
13655 	if (sec->dofs_size < sec->dofs_entsize) {
13656 		dtrace_dof_error(dof, "section entry size exceeds total size");
13657 		return (NULL);
13658 	}
13659 
13660 	if (sec->dofs_entsize != sizeof (dof_actdesc_t)) {
13661 		dtrace_dof_error(dof, "bad entry size in action description");
13662 		return (NULL);
13663 	}
13664 
13665 	if (sec->dofs_size / sec->dofs_entsize > dtrace_actions_max) {
13666 		dtrace_dof_error(dof, "actions exceed dtrace_actions_max");
13667 		return (NULL);
13668 	}
13669 
13670 	for (offs = 0; offs < sec->dofs_size; offs += sec->dofs_entsize) {
13671 		desc = (dof_actdesc_t *)(daddr +
13672 		    (uintptr_t)sec->dofs_offset + offs);
13673 		kind = (dtrace_actkind_t)desc->dofa_kind;
13674 
13675 		if ((DTRACEACT_ISPRINTFLIKE(kind) &&
13676 		    (kind != DTRACEACT_PRINTA ||
13677 		    desc->dofa_strtab != DOF_SECIDX_NONE)) ||
13678 		    (kind == DTRACEACT_DIFEXPR &&
13679 		    desc->dofa_strtab != DOF_SECIDX_NONE)) {
13680 			dof_sec_t *strtab;
13681 			char *str, *fmt;
13682 			uint64_t i;
13683 
13684 			/*
13685 			 * The argument to these actions is an index into the
13686 			 * DOF string table.  For printf()-like actions, this
13687 			 * is the format string.  For print(), this is the
13688 			 * CTF type of the expression result.
13689 			 */
13690 			if ((strtab = dtrace_dof_sect(dof,
13691 			    DOF_SECT_STRTAB, desc->dofa_strtab)) == NULL)
13692 				goto err;
13693 
13694 			str = (char *)((uintptr_t)dof +
13695 			    (uintptr_t)strtab->dofs_offset);
13696 
13697 			for (i = desc->dofa_arg; i < strtab->dofs_size; i++) {
13698 				if (str[i] == '\0')
13699 					break;
13700 			}
13701 
13702 			if (i >= strtab->dofs_size) {
13703 				dtrace_dof_error(dof, "bogus format string");
13704 				goto err;
13705 			}
13706 
13707 			if (i == desc->dofa_arg) {
13708 				dtrace_dof_error(dof, "empty format string");
13709 				goto err;
13710 			}
13711 
13712 			i -= desc->dofa_arg;
13713 			fmt = kmem_alloc(i + 1, KM_SLEEP);
13714 			bcopy(&str[desc->dofa_arg], fmt, i + 1);
13715 			arg = (uint64_t)(uintptr_t)fmt;
13716 		} else {
13717 			if (kind == DTRACEACT_PRINTA) {
13718 				ASSERT(desc->dofa_strtab == DOF_SECIDX_NONE);
13719 				arg = 0;
13720 			} else {
13721 				arg = desc->dofa_arg;
13722 			}
13723 		}
13724 
13725 		act = dtrace_actdesc_create(kind, desc->dofa_ntuple,
13726 		    desc->dofa_uarg, arg);
13727 
13728 		if (last != NULL) {
13729 			last->dtad_next = act;
13730 		} else {
13731 			first = act;
13732 		}
13733 
13734 		last = act;
13735 
13736 		if (desc->dofa_difo == DOF_SECIDX_NONE)
13737 			continue;
13738 
13739 		if ((difosec = dtrace_dof_sect(dof,
13740 		    DOF_SECT_DIFOHDR, desc->dofa_difo)) == NULL)
13741 			goto err;
13742 
13743 		act->dtad_difo = dtrace_dof_difo(dof, difosec, vstate, cr);
13744 
13745 		if (act->dtad_difo == NULL)
13746 			goto err;
13747 	}
13748 
13749 	ASSERT(first != NULL);
13750 	return (first);
13751 
13752 err:
13753 	for (act = first; act != NULL; act = next) {
13754 		next = act->dtad_next;
13755 		dtrace_actdesc_release(act, vstate);
13756 	}
13757 
13758 	return (NULL);
13759 }
13760 
13761 static dtrace_ecbdesc_t *
13762 dtrace_dof_ecbdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
13763     cred_t *cr)
13764 {
13765 	dtrace_ecbdesc_t *ep;
13766 	dof_ecbdesc_t *ecb;
13767 	dtrace_probedesc_t *desc;
13768 	dtrace_predicate_t *pred = NULL;
13769 
13770 	if (sec->dofs_size < sizeof (dof_ecbdesc_t)) {
13771 		dtrace_dof_error(dof, "truncated ECB description");
13772 		return (NULL);
13773 	}
13774 
13775 	if (sec->dofs_align != sizeof (uint64_t)) {
13776 		dtrace_dof_error(dof, "bad alignment in ECB description");
13777 		return (NULL);
13778 	}
13779 
13780 	ecb = (dof_ecbdesc_t *)((uintptr_t)dof + (uintptr_t)sec->dofs_offset);
13781 	sec = dtrace_dof_sect(dof, DOF_SECT_PROBEDESC, ecb->dofe_probes);
13782 
13783 	if (sec == NULL)
13784 		return (NULL);
13785 
13786 	ep = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
13787 	ep->dted_uarg = ecb->dofe_uarg;
13788 	desc = &ep->dted_probe;
13789 
13790 	if (dtrace_dof_probedesc(dof, sec, desc) == NULL)
13791 		goto err;
13792 
13793 	if (ecb->dofe_pred != DOF_SECIDX_NONE) {
13794 		if ((sec = dtrace_dof_sect(dof,
13795 		    DOF_SECT_DIFOHDR, ecb->dofe_pred)) == NULL)
13796 			goto err;
13797 
13798 		if ((pred = dtrace_dof_predicate(dof, sec, vstate, cr)) == NULL)
13799 			goto err;
13800 
13801 		ep->dted_pred.dtpdd_predicate = pred;
13802 	}
13803 
13804 	if (ecb->dofe_actions != DOF_SECIDX_NONE) {
13805 		if ((sec = dtrace_dof_sect(dof,
13806 		    DOF_SECT_ACTDESC, ecb->dofe_actions)) == NULL)
13807 			goto err;
13808 
13809 		ep->dted_action = dtrace_dof_actdesc(dof, sec, vstate, cr);
13810 
13811 		if (ep->dted_action == NULL)
13812 			goto err;
13813 	}
13814 
13815 	return (ep);
13816 
13817 err:
13818 	if (pred != NULL)
13819 		dtrace_predicate_release(pred, vstate);
13820 	kmem_free(ep, sizeof (dtrace_ecbdesc_t));
13821 	return (NULL);
13822 }
13823 
13824 /*
13825  * Apply the relocations from the specified 'sec' (a DOF_SECT_URELHDR) to the
13826  * specified DOF.  At present, this amounts to simply adding 'ubase' to the
13827  * site of any user SETX relocations to account for load object base address.
13828  * In the future, if we need other relocations, this function can be extended.
13829  */
13830 static int
13831 dtrace_dof_relocate(dof_hdr_t *dof, dof_sec_t *sec, uint64_t ubase)
13832 {
13833 	uintptr_t daddr = (uintptr_t)dof;
13834 	dof_relohdr_t *dofr =
13835 	    (dof_relohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
13836 	dof_sec_t *ss, *rs, *ts;
13837 	dof_relodesc_t *r;
13838 	uint_t i, n;
13839 
13840 	if (sec->dofs_size < sizeof (dof_relohdr_t) ||
13841 	    sec->dofs_align != sizeof (dof_secidx_t)) {
13842 		dtrace_dof_error(dof, "invalid relocation header");
13843 		return (-1);
13844 	}
13845 
13846 	ss = dtrace_dof_sect(dof, DOF_SECT_STRTAB, dofr->dofr_strtab);
13847 	rs = dtrace_dof_sect(dof, DOF_SECT_RELTAB, dofr->dofr_relsec);
13848 	ts = dtrace_dof_sect(dof, DOF_SECT_NONE, dofr->dofr_tgtsec);
13849 
13850 	if (ss == NULL || rs == NULL || ts == NULL)
13851 		return (-1); /* dtrace_dof_error() has been called already */
13852 
13853 	if (rs->dofs_entsize < sizeof (dof_relodesc_t) ||
13854 	    rs->dofs_align != sizeof (uint64_t)) {
13855 		dtrace_dof_error(dof, "invalid relocation section");
13856 		return (-1);
13857 	}
13858 
13859 	r = (dof_relodesc_t *)(uintptr_t)(daddr + rs->dofs_offset);
13860 	n = rs->dofs_size / rs->dofs_entsize;
13861 
13862 	for (i = 0; i < n; i++) {
13863 		uintptr_t taddr = daddr + ts->dofs_offset + r->dofr_offset;
13864 
13865 		switch (r->dofr_type) {
13866 		case DOF_RELO_NONE:
13867 			break;
13868 		case DOF_RELO_SETX:
13869 			if (r->dofr_offset >= ts->dofs_size || r->dofr_offset +
13870 			    sizeof (uint64_t) > ts->dofs_size) {
13871 				dtrace_dof_error(dof, "bad relocation offset");
13872 				return (-1);
13873 			}
13874 
13875 			if (!IS_P2ALIGNED(taddr, sizeof (uint64_t))) {
13876 				dtrace_dof_error(dof, "misaligned setx relo");
13877 				return (-1);
13878 			}
13879 
13880 			*(uint64_t *)taddr += ubase;
13881 			break;
13882 		default:
13883 			dtrace_dof_error(dof, "invalid relocation type");
13884 			return (-1);
13885 		}
13886 
13887 		r = (dof_relodesc_t *)((uintptr_t)r + rs->dofs_entsize);
13888 	}
13889 
13890 	return (0);
13891 }
13892 
13893 /*
13894  * The dof_hdr_t passed to dtrace_dof_slurp() should be a partially validated
13895  * header:  it should be at the front of a memory region that is at least
13896  * sizeof (dof_hdr_t) in size -- and then at least dof_hdr.dofh_loadsz in
13897  * size.  It need not be validated in any other way.
13898  */
13899 static int
13900 dtrace_dof_slurp(dof_hdr_t *dof, dtrace_vstate_t *vstate, cred_t *cr,
13901     dtrace_enabling_t **enabp, uint64_t ubase, int noprobes)
13902 {
13903 	uint64_t len = dof->dofh_loadsz, seclen;
13904 	uintptr_t daddr = (uintptr_t)dof;
13905 	dtrace_ecbdesc_t *ep;
13906 	dtrace_enabling_t *enab;
13907 	uint_t i;
13908 
13909 	ASSERT(MUTEX_HELD(&dtrace_lock));
13910 	ASSERT(dof->dofh_loadsz >= sizeof (dof_hdr_t));
13911 
13912 	/*
13913 	 * Check the DOF header identification bytes.  In addition to checking
13914 	 * valid settings, we also verify that unused bits/bytes are zeroed so
13915 	 * we can use them later without fear of regressing existing binaries.
13916 	 */
13917 	if (bcmp(&dof->dofh_ident[DOF_ID_MAG0],
13918 	    DOF_MAG_STRING, DOF_MAG_STRLEN) != 0) {
13919 		dtrace_dof_error(dof, "DOF magic string mismatch");
13920 		return (-1);
13921 	}
13922 
13923 	if (dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_ILP32 &&
13924 	    dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_LP64) {
13925 		dtrace_dof_error(dof, "DOF has invalid data model");
13926 		return (-1);
13927 	}
13928 
13929 	if (dof->dofh_ident[DOF_ID_ENCODING] != DOF_ENCODE_NATIVE) {
13930 		dtrace_dof_error(dof, "DOF encoding mismatch");
13931 		return (-1);
13932 	}
13933 
13934 	if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
13935 	    dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_2) {
13936 		dtrace_dof_error(dof, "DOF version mismatch");
13937 		return (-1);
13938 	}
13939 
13940 	if (dof->dofh_ident[DOF_ID_DIFVERS] != DIF_VERSION_2) {
13941 		dtrace_dof_error(dof, "DOF uses unsupported instruction set");
13942 		return (-1);
13943 	}
13944 
13945 	if (dof->dofh_ident[DOF_ID_DIFIREG] > DIF_DIR_NREGS) {
13946 		dtrace_dof_error(dof, "DOF uses too many integer registers");
13947 		return (-1);
13948 	}
13949 
13950 	if (dof->dofh_ident[DOF_ID_DIFTREG] > DIF_DTR_NREGS) {
13951 		dtrace_dof_error(dof, "DOF uses too many tuple registers");
13952 		return (-1);
13953 	}
13954 
13955 	for (i = DOF_ID_PAD; i < DOF_ID_SIZE; i++) {
13956 		if (dof->dofh_ident[i] != 0) {
13957 			dtrace_dof_error(dof, "DOF has invalid ident byte set");
13958 			return (-1);
13959 		}
13960 	}
13961 
13962 	if (dof->dofh_flags & ~DOF_FL_VALID) {
13963 		dtrace_dof_error(dof, "DOF has invalid flag bits set");
13964 		return (-1);
13965 	}
13966 
13967 	if (dof->dofh_secsize == 0) {
13968 		dtrace_dof_error(dof, "zero section header size");
13969 		return (-1);
13970 	}
13971 
13972 	/*
13973 	 * Check that the section headers don't exceed the amount of DOF
13974 	 * data.  Note that we cast the section size and number of sections
13975 	 * to uint64_t's to prevent possible overflow in the multiplication.
13976 	 */
13977 	seclen = (uint64_t)dof->dofh_secnum * (uint64_t)dof->dofh_secsize;
13978 
13979 	if (dof->dofh_secoff > len || seclen > len ||
13980 	    dof->dofh_secoff + seclen > len) {
13981 		dtrace_dof_error(dof, "truncated section headers");
13982 		return (-1);
13983 	}
13984 
13985 	if (!IS_P2ALIGNED(dof->dofh_secoff, sizeof (uint64_t))) {
13986 		dtrace_dof_error(dof, "misaligned section headers");
13987 		return (-1);
13988 	}
13989 
13990 	if (!IS_P2ALIGNED(dof->dofh_secsize, sizeof (uint64_t))) {
13991 		dtrace_dof_error(dof, "misaligned section size");
13992 		return (-1);
13993 	}
13994 
13995 	/*
13996 	 * Take an initial pass through the section headers to be sure that
13997 	 * the headers don't have stray offsets.  If the 'noprobes' flag is
13998 	 * set, do not permit sections relating to providers, probes, or args.
13999 	 */
14000 	for (i = 0; i < dof->dofh_secnum; i++) {
14001 		dof_sec_t *sec = (dof_sec_t *)(daddr +
14002 		    (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
14003 
14004 		if (noprobes) {
14005 			switch (sec->dofs_type) {
14006 			case DOF_SECT_PROVIDER:
14007 			case DOF_SECT_PROBES:
14008 			case DOF_SECT_PRARGS:
14009 			case DOF_SECT_PROFFS:
14010 				dtrace_dof_error(dof, "illegal sections "
14011 				    "for enabling");
14012 				return (-1);
14013 			}
14014 		}
14015 
14016 		if (DOF_SEC_ISLOADABLE(sec->dofs_type) &&
14017 		    !(sec->dofs_flags & DOF_SECF_LOAD)) {
14018 			dtrace_dof_error(dof, "loadable section with load "
14019 			    "flag unset");
14020 			return (-1);
14021 		}
14022 
14023 		if (!(sec->dofs_flags & DOF_SECF_LOAD))
14024 			continue; /* just ignore non-loadable sections */
14025 
14026 		if (!ISP2(sec->dofs_align)) {
14027 			dtrace_dof_error(dof, "bad section alignment");
14028 			return (-1);
14029 		}
14030 
14031 		if (sec->dofs_offset & (sec->dofs_align - 1)) {
14032 			dtrace_dof_error(dof, "misaligned section");
14033 			return (-1);
14034 		}
14035 
14036 		if (sec->dofs_offset > len || sec->dofs_size > len ||
14037 		    sec->dofs_offset + sec->dofs_size > len) {
14038 			dtrace_dof_error(dof, "corrupt section header");
14039 			return (-1);
14040 		}
14041 
14042 		if (sec->dofs_type == DOF_SECT_STRTAB && *((char *)daddr +
14043 		    sec->dofs_offset + sec->dofs_size - 1) != '\0') {
14044 			dtrace_dof_error(dof, "non-terminating string table");
14045 			return (-1);
14046 		}
14047 	}
14048 
14049 	/*
14050 	 * Take a second pass through the sections and locate and perform any
14051 	 * relocations that are present.  We do this after the first pass to
14052 	 * be sure that all sections have had their headers validated.
14053 	 */
14054 	for (i = 0; i < dof->dofh_secnum; i++) {
14055 		dof_sec_t *sec = (dof_sec_t *)(daddr +
14056 		    (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
14057 
14058 		if (!(sec->dofs_flags & DOF_SECF_LOAD))
14059 			continue; /* skip sections that are not loadable */
14060 
14061 		switch (sec->dofs_type) {
14062 		case DOF_SECT_URELHDR:
14063 			if (dtrace_dof_relocate(dof, sec, ubase) != 0)
14064 				return (-1);
14065 			break;
14066 		}
14067 	}
14068 
14069 	if ((enab = *enabp) == NULL)
14070 		enab = *enabp = dtrace_enabling_create(vstate);
14071 
14072 	for (i = 0; i < dof->dofh_secnum; i++) {
14073 		dof_sec_t *sec = (dof_sec_t *)(daddr +
14074 		    (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
14075 
14076 		if (sec->dofs_type != DOF_SECT_ECBDESC)
14077 			continue;
14078 
14079 		if ((ep = dtrace_dof_ecbdesc(dof, sec, vstate, cr)) == NULL) {
14080 			dtrace_enabling_destroy(enab);
14081 			*enabp = NULL;
14082 			return (-1);
14083 		}
14084 
14085 		dtrace_enabling_add(enab, ep);
14086 	}
14087 
14088 	return (0);
14089 }
14090 
14091 /*
14092  * Process DOF for any options.  This routine assumes that the DOF has been
14093  * at least processed by dtrace_dof_slurp().
14094  */
14095 static int
14096 dtrace_dof_options(dof_hdr_t *dof, dtrace_state_t *state)
14097 {
14098 	int i, rval;
14099 	uint32_t entsize;
14100 	size_t offs;
14101 	dof_optdesc_t *desc;
14102 
14103 	for (i = 0; i < dof->dofh_secnum; i++) {
14104 		dof_sec_t *sec = (dof_sec_t *)((uintptr_t)dof +
14105 		    (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
14106 
14107 		if (sec->dofs_type != DOF_SECT_OPTDESC)
14108 			continue;
14109 
14110 		if (sec->dofs_align != sizeof (uint64_t)) {
14111 			dtrace_dof_error(dof, "bad alignment in "
14112 			    "option description");
14113 			return (EINVAL);
14114 		}
14115 
14116 		if ((entsize = sec->dofs_entsize) == 0) {
14117 			dtrace_dof_error(dof, "zeroed option entry size");
14118 			return (EINVAL);
14119 		}
14120 
14121 		if (entsize < sizeof (dof_optdesc_t)) {
14122 			dtrace_dof_error(dof, "bad option entry size");
14123 			return (EINVAL);
14124 		}
14125 
14126 		for (offs = 0; offs < sec->dofs_size; offs += entsize) {
14127 			desc = (dof_optdesc_t *)((uintptr_t)dof +
14128 			    (uintptr_t)sec->dofs_offset + offs);
14129 
14130 			if (desc->dofo_strtab != DOF_SECIDX_NONE) {
14131 				dtrace_dof_error(dof, "non-zero option string");
14132 				return (EINVAL);
14133 			}
14134 
14135 			if (desc->dofo_value == DTRACEOPT_UNSET) {
14136 				dtrace_dof_error(dof, "unset option");
14137 				return (EINVAL);
14138 			}
14139 
14140 			if ((rval = dtrace_state_option(state,
14141 			    desc->dofo_option, desc->dofo_value)) != 0) {
14142 				dtrace_dof_error(dof, "rejected option");
14143 				return (rval);
14144 			}
14145 		}
14146 	}
14147 
14148 	return (0);
14149 }
14150 
14151 /*
14152  * DTrace Consumer State Functions
14153  */
14154 static int
14155 dtrace_dstate_init(dtrace_dstate_t *dstate, size_t size)
14156 {
14157 	size_t hashsize, maxper, min, chunksize = dstate->dtds_chunksize;
14158 	void *base;
14159 	uintptr_t limit;
14160 	dtrace_dynvar_t *dvar, *next, *start;
14161 	int i;
14162 
14163 	ASSERT(MUTEX_HELD(&dtrace_lock));
14164 	ASSERT(dstate->dtds_base == NULL && dstate->dtds_percpu == NULL);
14165 
14166 	bzero(dstate, sizeof (dtrace_dstate_t));
14167 
14168 	if ((dstate->dtds_chunksize = chunksize) == 0)
14169 		dstate->dtds_chunksize = DTRACE_DYNVAR_CHUNKSIZE;
14170 
14171 	VERIFY(dstate->dtds_chunksize < LONG_MAX);
14172 
14173 	if (size < (min = dstate->dtds_chunksize + sizeof (dtrace_dynhash_t)))
14174 		size = min;
14175 
14176 	if ((base = kmem_zalloc(size, KM_NOSLEEP | KM_NORMALPRI)) == NULL)
14177 		return (ENOMEM);
14178 
14179 	dstate->dtds_size = size;
14180 	dstate->dtds_base = base;
14181 	dstate->dtds_percpu = kmem_cache_alloc(dtrace_state_cache, KM_SLEEP);
14182 	bzero(dstate->dtds_percpu, NCPU * sizeof (dtrace_dstate_percpu_t));
14183 
14184 	hashsize = size / (dstate->dtds_chunksize + sizeof (dtrace_dynhash_t));
14185 
14186 	if (hashsize != 1 && (hashsize & 1))
14187 		hashsize--;
14188 
14189 	dstate->dtds_hashsize = hashsize;
14190 	dstate->dtds_hash = dstate->dtds_base;
14191 
14192 	/*
14193 	 * Set all of our hash buckets to point to the single sink, and (if
14194 	 * it hasn't already been set), set the sink's hash value to be the
14195 	 * sink sentinel value.  The sink is needed for dynamic variable
14196 	 * lookups to know that they have iterated over an entire, valid hash
14197 	 * chain.
14198 	 */
14199 	for (i = 0; i < hashsize; i++)
14200 		dstate->dtds_hash[i].dtdh_chain = &dtrace_dynhash_sink;
14201 
14202 	if (dtrace_dynhash_sink.dtdv_hashval != DTRACE_DYNHASH_SINK)
14203 		dtrace_dynhash_sink.dtdv_hashval = DTRACE_DYNHASH_SINK;
14204 
14205 	/*
14206 	 * Determine number of active CPUs.  Divide free list evenly among
14207 	 * active CPUs.
14208 	 */
14209 	start = (dtrace_dynvar_t *)
14210 	    ((uintptr_t)base + hashsize * sizeof (dtrace_dynhash_t));
14211 	limit = (uintptr_t)base + size;
14212 
14213 	VERIFY((uintptr_t)start < limit);
14214 	VERIFY((uintptr_t)start >= (uintptr_t)base);
14215 
14216 	maxper = (limit - (uintptr_t)start) / NCPU;
14217 	maxper = (maxper / dstate->dtds_chunksize) * dstate->dtds_chunksize;
14218 
14219 #ifndef illumos
14220 	CPU_FOREACH(i) {
14221 #else
14222 	for (i = 0; i < NCPU; i++) {
14223 #endif
14224 		dstate->dtds_percpu[i].dtdsc_free = dvar = start;
14225 
14226 		/*
14227 		 * If we don't even have enough chunks to make it once through
14228 		 * NCPUs, we're just going to allocate everything to the first
14229 		 * CPU.  And if we're on the last CPU, we're going to allocate
14230 		 * whatever is left over.  In either case, we set the limit to
14231 		 * be the limit of the dynamic variable space.
14232 		 */
14233 		if (maxper == 0 || i == NCPU - 1) {
14234 			limit = (uintptr_t)base + size;
14235 			start = NULL;
14236 		} else {
14237 			limit = (uintptr_t)start + maxper;
14238 			start = (dtrace_dynvar_t *)limit;
14239 		}
14240 
14241 		VERIFY(limit <= (uintptr_t)base + size);
14242 
14243 		for (;;) {
14244 			next = (dtrace_dynvar_t *)((uintptr_t)dvar +
14245 			    dstate->dtds_chunksize);
14246 
14247 			if ((uintptr_t)next + dstate->dtds_chunksize >= limit)
14248 				break;
14249 
14250 			VERIFY((uintptr_t)dvar >= (uintptr_t)base &&
14251 			    (uintptr_t)dvar <= (uintptr_t)base + size);
14252 			dvar->dtdv_next = next;
14253 			dvar = next;
14254 		}
14255 
14256 		if (maxper == 0)
14257 			break;
14258 	}
14259 
14260 	return (0);
14261 }
14262 
14263 static void
14264 dtrace_dstate_fini(dtrace_dstate_t *dstate)
14265 {
14266 	ASSERT(MUTEX_HELD(&cpu_lock));
14267 
14268 	if (dstate->dtds_base == NULL)
14269 		return;
14270 
14271 	kmem_free(dstate->dtds_base, dstate->dtds_size);
14272 	kmem_cache_free(dtrace_state_cache, dstate->dtds_percpu);
14273 }
14274 
14275 static void
14276 dtrace_vstate_fini(dtrace_vstate_t *vstate)
14277 {
14278 	/*
14279 	 * Logical XOR, where are you?
14280 	 */
14281 	ASSERT((vstate->dtvs_nglobals == 0) ^ (vstate->dtvs_globals != NULL));
14282 
14283 	if (vstate->dtvs_nglobals > 0) {
14284 		kmem_free(vstate->dtvs_globals, vstate->dtvs_nglobals *
14285 		    sizeof (dtrace_statvar_t *));
14286 	}
14287 
14288 	if (vstate->dtvs_ntlocals > 0) {
14289 		kmem_free(vstate->dtvs_tlocals, vstate->dtvs_ntlocals *
14290 		    sizeof (dtrace_difv_t));
14291 	}
14292 
14293 	ASSERT((vstate->dtvs_nlocals == 0) ^ (vstate->dtvs_locals != NULL));
14294 
14295 	if (vstate->dtvs_nlocals > 0) {
14296 		kmem_free(vstate->dtvs_locals, vstate->dtvs_nlocals *
14297 		    sizeof (dtrace_statvar_t *));
14298 	}
14299 }
14300 
14301 #ifdef illumos
14302 static void
14303 dtrace_state_clean(dtrace_state_t *state)
14304 {
14305 	if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE)
14306 		return;
14307 
14308 	dtrace_dynvar_clean(&state->dts_vstate.dtvs_dynvars);
14309 	dtrace_speculation_clean(state);
14310 }
14311 
14312 static void
14313 dtrace_state_deadman(dtrace_state_t *state)
14314 {
14315 	hrtime_t now;
14316 
14317 	dtrace_sync();
14318 
14319 	now = dtrace_gethrtime();
14320 
14321 	if (state != dtrace_anon.dta_state &&
14322 	    now - state->dts_laststatus >= dtrace_deadman_user)
14323 		return;
14324 
14325 	/*
14326 	 * We must be sure that dts_alive never appears to be less than the
14327 	 * value upon entry to dtrace_state_deadman(), and because we lack a
14328 	 * dtrace_cas64(), we cannot store to it atomically.  We thus instead
14329 	 * store INT64_MAX to it, followed by a memory barrier, followed by
14330 	 * the new value.  This assures that dts_alive never appears to be
14331 	 * less than its true value, regardless of the order in which the
14332 	 * stores to the underlying storage are issued.
14333 	 */
14334 	state->dts_alive = INT64_MAX;
14335 	dtrace_membar_producer();
14336 	state->dts_alive = now;
14337 }
14338 #else	/* !illumos */
14339 static void
14340 dtrace_state_clean(void *arg)
14341 {
14342 	dtrace_state_t *state = arg;
14343 	dtrace_optval_t *opt = state->dts_options;
14344 
14345 	if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE)
14346 		return;
14347 
14348 	dtrace_dynvar_clean(&state->dts_vstate.dtvs_dynvars);
14349 	dtrace_speculation_clean(state);
14350 
14351 	callout_reset(&state->dts_cleaner, hz * opt[DTRACEOPT_CLEANRATE] / NANOSEC,
14352 	    dtrace_state_clean, state);
14353 }
14354 
14355 static void
14356 dtrace_state_deadman(void *arg)
14357 {
14358 	dtrace_state_t *state = arg;
14359 	hrtime_t now;
14360 
14361 	dtrace_sync();
14362 
14363 	dtrace_debug_output();
14364 
14365 	now = dtrace_gethrtime();
14366 
14367 	if (state != dtrace_anon.dta_state &&
14368 	    now - state->dts_laststatus >= dtrace_deadman_user)
14369 		return;
14370 
14371 	/*
14372 	 * We must be sure that dts_alive never appears to be less than the
14373 	 * value upon entry to dtrace_state_deadman(), and because we lack a
14374 	 * dtrace_cas64(), we cannot store to it atomically.  We thus instead
14375 	 * store INT64_MAX to it, followed by a memory barrier, followed by
14376 	 * the new value.  This assures that dts_alive never appears to be
14377 	 * less than its true value, regardless of the order in which the
14378 	 * stores to the underlying storage are issued.
14379 	 */
14380 	state->dts_alive = INT64_MAX;
14381 	dtrace_membar_producer();
14382 	state->dts_alive = now;
14383 
14384 	callout_reset(&state->dts_deadman, hz * dtrace_deadman_interval / NANOSEC,
14385 	    dtrace_state_deadman, state);
14386 }
14387 #endif	/* illumos */
14388 
14389 static dtrace_state_t *
14390 #ifdef illumos
14391 dtrace_state_create(dev_t *devp, cred_t *cr)
14392 #else
14393 dtrace_state_create(struct cdev *dev, struct ucred *cred __unused)
14394 #endif
14395 {
14396 #ifdef illumos
14397 	minor_t minor;
14398 	major_t major;
14399 #else
14400 	cred_t *cr = NULL;
14401 	int m = 0;
14402 #endif
14403 	char c[30];
14404 	dtrace_state_t *state;
14405 	dtrace_optval_t *opt;
14406 	int bufsize = NCPU * sizeof (dtrace_buffer_t), i;
14407 
14408 	ASSERT(MUTEX_HELD(&dtrace_lock));
14409 	ASSERT(MUTEX_HELD(&cpu_lock));
14410 
14411 #ifdef illumos
14412 	minor = (minor_t)(uintptr_t)vmem_alloc(dtrace_minor, 1,
14413 	    VM_BESTFIT | VM_SLEEP);
14414 
14415 	if (ddi_soft_state_zalloc(dtrace_softstate, minor) != DDI_SUCCESS) {
14416 		vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
14417 		return (NULL);
14418 	}
14419 
14420 	state = ddi_get_soft_state(dtrace_softstate, minor);
14421 #else
14422 	if (dev != NULL) {
14423 		cr = dev->si_cred;
14424 		m = dev2unit(dev);
14425 	}
14426 
14427 	/* Allocate memory for the state. */
14428 	state = kmem_zalloc(sizeof(dtrace_state_t), KM_SLEEP);
14429 #endif
14430 
14431 	state->dts_epid = DTRACE_EPIDNONE + 1;
14432 
14433 	(void) snprintf(c, sizeof (c), "dtrace_aggid_%d", m);
14434 #ifdef illumos
14435 	state->dts_aggid_arena = vmem_create(c, (void *)1, UINT32_MAX, 1,
14436 	    NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
14437 
14438 	if (devp != NULL) {
14439 		major = getemajor(*devp);
14440 	} else {
14441 		major = ddi_driver_major(dtrace_devi);
14442 	}
14443 
14444 	state->dts_dev = makedevice(major, minor);
14445 
14446 	if (devp != NULL)
14447 		*devp = state->dts_dev;
14448 #else
14449 	state->dts_aggid_arena = new_unrhdr(1, INT_MAX, &dtrace_unr_mtx);
14450 	state->dts_dev = dev;
14451 #endif
14452 
14453 	/*
14454 	 * We allocate NCPU buffers.  On the one hand, this can be quite
14455 	 * a bit of memory per instance (nearly 36K on a Starcat).  On the
14456 	 * other hand, it saves an additional memory reference in the probe
14457 	 * path.
14458 	 */
14459 	state->dts_buffer = kmem_zalloc(bufsize, KM_SLEEP);
14460 	state->dts_aggbuffer = kmem_zalloc(bufsize, KM_SLEEP);
14461 
14462 #ifdef illumos
14463 	state->dts_cleaner = CYCLIC_NONE;
14464 	state->dts_deadman = CYCLIC_NONE;
14465 #else
14466 	callout_init(&state->dts_cleaner, 1);
14467 	callout_init(&state->dts_deadman, 1);
14468 #endif
14469 	state->dts_vstate.dtvs_state = state;
14470 
14471 	for (i = 0; i < DTRACEOPT_MAX; i++)
14472 		state->dts_options[i] = DTRACEOPT_UNSET;
14473 
14474 	/*
14475 	 * Set the default options.
14476 	 */
14477 	opt = state->dts_options;
14478 	opt[DTRACEOPT_BUFPOLICY] = DTRACEOPT_BUFPOLICY_SWITCH;
14479 	opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_AUTO;
14480 	opt[DTRACEOPT_NSPEC] = dtrace_nspec_default;
14481 	opt[DTRACEOPT_SPECSIZE] = dtrace_specsize_default;
14482 	opt[DTRACEOPT_CPU] = (dtrace_optval_t)DTRACE_CPUALL;
14483 	opt[DTRACEOPT_STRSIZE] = dtrace_strsize_default;
14484 	opt[DTRACEOPT_STACKFRAMES] = dtrace_stackframes_default;
14485 	opt[DTRACEOPT_USTACKFRAMES] = dtrace_ustackframes_default;
14486 	opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_default;
14487 	opt[DTRACEOPT_AGGRATE] = dtrace_aggrate_default;
14488 	opt[DTRACEOPT_SWITCHRATE] = dtrace_switchrate_default;
14489 	opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_default;
14490 	opt[DTRACEOPT_JSTACKFRAMES] = dtrace_jstackframes_default;
14491 	opt[DTRACEOPT_JSTACKSTRSIZE] = dtrace_jstackstrsize_default;
14492 
14493 	state->dts_activity = DTRACE_ACTIVITY_INACTIVE;
14494 
14495 	/*
14496 	 * Depending on the user credentials, we set flag bits which alter probe
14497 	 * visibility or the amount of destructiveness allowed.  In the case of
14498 	 * actual anonymous tracing, or the possession of all privileges, all of
14499 	 * the normal checks are bypassed.
14500 	 */
14501 	if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
14502 		state->dts_cred.dcr_visible = DTRACE_CRV_ALL;
14503 		state->dts_cred.dcr_action = DTRACE_CRA_ALL;
14504 	} else {
14505 		/*
14506 		 * Set up the credentials for this instantiation.  We take a
14507 		 * hold on the credential to prevent it from disappearing on
14508 		 * us; this in turn prevents the zone_t referenced by this
14509 		 * credential from disappearing.  This means that we can
14510 		 * examine the credential and the zone from probe context.
14511 		 */
14512 		crhold(cr);
14513 		state->dts_cred.dcr_cred = cr;
14514 
14515 		/*
14516 		 * CRA_PROC means "we have *some* privilege for dtrace" and
14517 		 * unlocks the use of variables like pid, zonename, etc.
14518 		 */
14519 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE) ||
14520 		    PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
14521 			state->dts_cred.dcr_action |= DTRACE_CRA_PROC;
14522 		}
14523 
14524 		/*
14525 		 * dtrace_user allows use of syscall and profile providers.
14526 		 * If the user also has proc_owner and/or proc_zone, we
14527 		 * extend the scope to include additional visibility and
14528 		 * destructive power.
14529 		 */
14530 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE)) {
14531 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE)) {
14532 				state->dts_cred.dcr_visible |=
14533 				    DTRACE_CRV_ALLPROC;
14534 
14535 				state->dts_cred.dcr_action |=
14536 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
14537 			}
14538 
14539 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE)) {
14540 				state->dts_cred.dcr_visible |=
14541 				    DTRACE_CRV_ALLZONE;
14542 
14543 				state->dts_cred.dcr_action |=
14544 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
14545 			}
14546 
14547 			/*
14548 			 * If we have all privs in whatever zone this is,
14549 			 * we can do destructive things to processes which
14550 			 * have altered credentials.
14551 			 */
14552 #ifdef illumos
14553 			if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
14554 			    cr->cr_zone->zone_privset)) {
14555 				state->dts_cred.dcr_action |=
14556 				    DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
14557 			}
14558 #endif
14559 		}
14560 
14561 		/*
14562 		 * Holding the dtrace_kernel privilege also implies that
14563 		 * the user has the dtrace_user privilege from a visibility
14564 		 * perspective.  But without further privileges, some
14565 		 * destructive actions are not available.
14566 		 */
14567 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE)) {
14568 			/*
14569 			 * Make all probes in all zones visible.  However,
14570 			 * this doesn't mean that all actions become available
14571 			 * to all zones.
14572 			 */
14573 			state->dts_cred.dcr_visible |= DTRACE_CRV_KERNEL |
14574 			    DTRACE_CRV_ALLPROC | DTRACE_CRV_ALLZONE;
14575 
14576 			state->dts_cred.dcr_action |= DTRACE_CRA_KERNEL |
14577 			    DTRACE_CRA_PROC;
14578 			/*
14579 			 * Holding proc_owner means that destructive actions
14580 			 * for *this* zone are allowed.
14581 			 */
14582 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
14583 				state->dts_cred.dcr_action |=
14584 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
14585 
14586 			/*
14587 			 * Holding proc_zone means that destructive actions
14588 			 * for this user/group ID in all zones is allowed.
14589 			 */
14590 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
14591 				state->dts_cred.dcr_action |=
14592 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
14593 
14594 #ifdef illumos
14595 			/*
14596 			 * If we have all privs in whatever zone this is,
14597 			 * we can do destructive things to processes which
14598 			 * have altered credentials.
14599 			 */
14600 			if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
14601 			    cr->cr_zone->zone_privset)) {
14602 				state->dts_cred.dcr_action |=
14603 				    DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
14604 			}
14605 #endif
14606 		}
14607 
14608 		/*
14609 		 * Holding the dtrace_proc privilege gives control over fasttrap
14610 		 * and pid providers.  We need to grant wider destructive
14611 		 * privileges in the event that the user has proc_owner and/or
14612 		 * proc_zone.
14613 		 */
14614 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
14615 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
14616 				state->dts_cred.dcr_action |=
14617 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
14618 
14619 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
14620 				state->dts_cred.dcr_action |=
14621 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
14622 		}
14623 	}
14624 
14625 	return (state);
14626 }
14627 
14628 static int
14629 dtrace_state_buffer(dtrace_state_t *state, dtrace_buffer_t *buf, int which)
14630 {
14631 	dtrace_optval_t *opt = state->dts_options, size;
14632 	processorid_t cpu = 0;;
14633 	int flags = 0, rval, factor, divisor = 1;
14634 
14635 	ASSERT(MUTEX_HELD(&dtrace_lock));
14636 	ASSERT(MUTEX_HELD(&cpu_lock));
14637 	ASSERT(which < DTRACEOPT_MAX);
14638 	ASSERT(state->dts_activity == DTRACE_ACTIVITY_INACTIVE ||
14639 	    (state == dtrace_anon.dta_state &&
14640 	    state->dts_activity == DTRACE_ACTIVITY_ACTIVE));
14641 
14642 	if (opt[which] == DTRACEOPT_UNSET || opt[which] == 0)
14643 		return (0);
14644 
14645 	if (opt[DTRACEOPT_CPU] != DTRACEOPT_UNSET)
14646 		cpu = opt[DTRACEOPT_CPU];
14647 
14648 	if (which == DTRACEOPT_SPECSIZE)
14649 		flags |= DTRACEBUF_NOSWITCH;
14650 
14651 	if (which == DTRACEOPT_BUFSIZE) {
14652 		if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_RING)
14653 			flags |= DTRACEBUF_RING;
14654 
14655 		if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_FILL)
14656 			flags |= DTRACEBUF_FILL;
14657 
14658 		if (state != dtrace_anon.dta_state ||
14659 		    state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
14660 			flags |= DTRACEBUF_INACTIVE;
14661 	}
14662 
14663 	for (size = opt[which]; size >= sizeof (uint64_t); size /= divisor) {
14664 		/*
14665 		 * The size must be 8-byte aligned.  If the size is not 8-byte
14666 		 * aligned, drop it down by the difference.
14667 		 */
14668 		if (size & (sizeof (uint64_t) - 1))
14669 			size -= size & (sizeof (uint64_t) - 1);
14670 
14671 		if (size < state->dts_reserve) {
14672 			/*
14673 			 * Buffers always must be large enough to accommodate
14674 			 * their prereserved space.  We return E2BIG instead
14675 			 * of ENOMEM in this case to allow for user-level
14676 			 * software to differentiate the cases.
14677 			 */
14678 			return (E2BIG);
14679 		}
14680 
14681 		rval = dtrace_buffer_alloc(buf, size, flags, cpu, &factor);
14682 
14683 		if (rval != ENOMEM) {
14684 			opt[which] = size;
14685 			return (rval);
14686 		}
14687 
14688 		if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
14689 			return (rval);
14690 
14691 		for (divisor = 2; divisor < factor; divisor <<= 1)
14692 			continue;
14693 	}
14694 
14695 	return (ENOMEM);
14696 }
14697 
14698 static int
14699 dtrace_state_buffers(dtrace_state_t *state)
14700 {
14701 	dtrace_speculation_t *spec = state->dts_speculations;
14702 	int rval, i;
14703 
14704 	if ((rval = dtrace_state_buffer(state, state->dts_buffer,
14705 	    DTRACEOPT_BUFSIZE)) != 0)
14706 		return (rval);
14707 
14708 	if ((rval = dtrace_state_buffer(state, state->dts_aggbuffer,
14709 	    DTRACEOPT_AGGSIZE)) != 0)
14710 		return (rval);
14711 
14712 	for (i = 0; i < state->dts_nspeculations; i++) {
14713 		if ((rval = dtrace_state_buffer(state,
14714 		    spec[i].dtsp_buffer, DTRACEOPT_SPECSIZE)) != 0)
14715 			return (rval);
14716 	}
14717 
14718 	return (0);
14719 }
14720 
14721 static void
14722 dtrace_state_prereserve(dtrace_state_t *state)
14723 {
14724 	dtrace_ecb_t *ecb;
14725 	dtrace_probe_t *probe;
14726 
14727 	state->dts_reserve = 0;
14728 
14729 	if (state->dts_options[DTRACEOPT_BUFPOLICY] != DTRACEOPT_BUFPOLICY_FILL)
14730 		return;
14731 
14732 	/*
14733 	 * If our buffer policy is a "fill" buffer policy, we need to set the
14734 	 * prereserved space to be the space required by the END probes.
14735 	 */
14736 	probe = dtrace_probes[dtrace_probeid_end - 1];
14737 	ASSERT(probe != NULL);
14738 
14739 	for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
14740 		if (ecb->dte_state != state)
14741 			continue;
14742 
14743 		state->dts_reserve += ecb->dte_needed + ecb->dte_alignment;
14744 	}
14745 }
14746 
14747 static int
14748 dtrace_state_go(dtrace_state_t *state, processorid_t *cpu)
14749 {
14750 	dtrace_optval_t *opt = state->dts_options, sz, nspec;
14751 	dtrace_speculation_t *spec;
14752 	dtrace_buffer_t *buf;
14753 #ifdef illumos
14754 	cyc_handler_t hdlr;
14755 	cyc_time_t when;
14756 #endif
14757 	int rval = 0, i, bufsize = NCPU * sizeof (dtrace_buffer_t);
14758 	dtrace_icookie_t cookie;
14759 
14760 	mutex_enter(&cpu_lock);
14761 	mutex_enter(&dtrace_lock);
14762 
14763 	if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
14764 		rval = EBUSY;
14765 		goto out;
14766 	}
14767 
14768 	/*
14769 	 * Before we can perform any checks, we must prime all of the
14770 	 * retained enablings that correspond to this state.
14771 	 */
14772 	dtrace_enabling_prime(state);
14773 
14774 	if (state->dts_destructive && !state->dts_cred.dcr_destructive) {
14775 		rval = EACCES;
14776 		goto out;
14777 	}
14778 
14779 	dtrace_state_prereserve(state);
14780 
14781 	/*
14782 	 * Now we want to do is try to allocate our speculations.
14783 	 * We do not automatically resize the number of speculations; if
14784 	 * this fails, we will fail the operation.
14785 	 */
14786 	nspec = opt[DTRACEOPT_NSPEC];
14787 	ASSERT(nspec != DTRACEOPT_UNSET);
14788 
14789 	if (nspec > INT_MAX) {
14790 		rval = ENOMEM;
14791 		goto out;
14792 	}
14793 
14794 	spec = kmem_zalloc(nspec * sizeof (dtrace_speculation_t),
14795 	    KM_NOSLEEP | KM_NORMALPRI);
14796 
14797 	if (spec == NULL) {
14798 		rval = ENOMEM;
14799 		goto out;
14800 	}
14801 
14802 	state->dts_speculations = spec;
14803 	state->dts_nspeculations = (int)nspec;
14804 
14805 	for (i = 0; i < nspec; i++) {
14806 		if ((buf = kmem_zalloc(bufsize,
14807 		    KM_NOSLEEP | KM_NORMALPRI)) == NULL) {
14808 			rval = ENOMEM;
14809 			goto err;
14810 		}
14811 
14812 		spec[i].dtsp_buffer = buf;
14813 	}
14814 
14815 	if (opt[DTRACEOPT_GRABANON] != DTRACEOPT_UNSET) {
14816 		if (dtrace_anon.dta_state == NULL) {
14817 			rval = ENOENT;
14818 			goto out;
14819 		}
14820 
14821 		if (state->dts_necbs != 0) {
14822 			rval = EALREADY;
14823 			goto out;
14824 		}
14825 
14826 		state->dts_anon = dtrace_anon_grab();
14827 		ASSERT(state->dts_anon != NULL);
14828 		state = state->dts_anon;
14829 
14830 		/*
14831 		 * We want "grabanon" to be set in the grabbed state, so we'll
14832 		 * copy that option value from the grabbing state into the
14833 		 * grabbed state.
14834 		 */
14835 		state->dts_options[DTRACEOPT_GRABANON] =
14836 		    opt[DTRACEOPT_GRABANON];
14837 
14838 		*cpu = dtrace_anon.dta_beganon;
14839 
14840 		/*
14841 		 * If the anonymous state is active (as it almost certainly
14842 		 * is if the anonymous enabling ultimately matched anything),
14843 		 * we don't allow any further option processing -- but we
14844 		 * don't return failure.
14845 		 */
14846 		if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
14847 			goto out;
14848 	}
14849 
14850 	if (opt[DTRACEOPT_AGGSIZE] != DTRACEOPT_UNSET &&
14851 	    opt[DTRACEOPT_AGGSIZE] != 0) {
14852 		if (state->dts_aggregations == NULL) {
14853 			/*
14854 			 * We're not going to create an aggregation buffer
14855 			 * because we don't have any ECBs that contain
14856 			 * aggregations -- set this option to 0.
14857 			 */
14858 			opt[DTRACEOPT_AGGSIZE] = 0;
14859 		} else {
14860 			/*
14861 			 * If we have an aggregation buffer, we must also have
14862 			 * a buffer to use as scratch.
14863 			 */
14864 			if (opt[DTRACEOPT_BUFSIZE] == DTRACEOPT_UNSET ||
14865 			    opt[DTRACEOPT_BUFSIZE] < state->dts_needed) {
14866 				opt[DTRACEOPT_BUFSIZE] = state->dts_needed;
14867 			}
14868 		}
14869 	}
14870 
14871 	if (opt[DTRACEOPT_SPECSIZE] != DTRACEOPT_UNSET &&
14872 	    opt[DTRACEOPT_SPECSIZE] != 0) {
14873 		if (!state->dts_speculates) {
14874 			/*
14875 			 * We're not going to create speculation buffers
14876 			 * because we don't have any ECBs that actually
14877 			 * speculate -- set the speculation size to 0.
14878 			 */
14879 			opt[DTRACEOPT_SPECSIZE] = 0;
14880 		}
14881 	}
14882 
14883 	/*
14884 	 * The bare minimum size for any buffer that we're actually going to
14885 	 * do anything to is sizeof (uint64_t).
14886 	 */
14887 	sz = sizeof (uint64_t);
14888 
14889 	if ((state->dts_needed != 0 && opt[DTRACEOPT_BUFSIZE] < sz) ||
14890 	    (state->dts_speculates && opt[DTRACEOPT_SPECSIZE] < sz) ||
14891 	    (state->dts_aggregations != NULL && opt[DTRACEOPT_AGGSIZE] < sz)) {
14892 		/*
14893 		 * A buffer size has been explicitly set to 0 (or to a size
14894 		 * that will be adjusted to 0) and we need the space -- we
14895 		 * need to return failure.  We return ENOSPC to differentiate
14896 		 * it from failing to allocate a buffer due to failure to meet
14897 		 * the reserve (for which we return E2BIG).
14898 		 */
14899 		rval = ENOSPC;
14900 		goto out;
14901 	}
14902 
14903 	if ((rval = dtrace_state_buffers(state)) != 0)
14904 		goto err;
14905 
14906 	if ((sz = opt[DTRACEOPT_DYNVARSIZE]) == DTRACEOPT_UNSET)
14907 		sz = dtrace_dstate_defsize;
14908 
14909 	do {
14910 		rval = dtrace_dstate_init(&state->dts_vstate.dtvs_dynvars, sz);
14911 
14912 		if (rval == 0)
14913 			break;
14914 
14915 		if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
14916 			goto err;
14917 	} while (sz >>= 1);
14918 
14919 	opt[DTRACEOPT_DYNVARSIZE] = sz;
14920 
14921 	if (rval != 0)
14922 		goto err;
14923 
14924 	if (opt[DTRACEOPT_STATUSRATE] > dtrace_statusrate_max)
14925 		opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_max;
14926 
14927 	if (opt[DTRACEOPT_CLEANRATE] == 0)
14928 		opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
14929 
14930 	if (opt[DTRACEOPT_CLEANRATE] < dtrace_cleanrate_min)
14931 		opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_min;
14932 
14933 	if (opt[DTRACEOPT_CLEANRATE] > dtrace_cleanrate_max)
14934 		opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
14935 
14936 	state->dts_alive = state->dts_laststatus = dtrace_gethrtime();
14937 #ifdef illumos
14938 	hdlr.cyh_func = (cyc_func_t)dtrace_state_clean;
14939 	hdlr.cyh_arg = state;
14940 	hdlr.cyh_level = CY_LOW_LEVEL;
14941 
14942 	when.cyt_when = 0;
14943 	when.cyt_interval = opt[DTRACEOPT_CLEANRATE];
14944 
14945 	state->dts_cleaner = cyclic_add(&hdlr, &when);
14946 
14947 	hdlr.cyh_func = (cyc_func_t)dtrace_state_deadman;
14948 	hdlr.cyh_arg = state;
14949 	hdlr.cyh_level = CY_LOW_LEVEL;
14950 
14951 	when.cyt_when = 0;
14952 	when.cyt_interval = dtrace_deadman_interval;
14953 
14954 	state->dts_deadman = cyclic_add(&hdlr, &when);
14955 #else
14956 	callout_reset(&state->dts_cleaner, hz * opt[DTRACEOPT_CLEANRATE] / NANOSEC,
14957 	    dtrace_state_clean, state);
14958 	callout_reset(&state->dts_deadman, hz * dtrace_deadman_interval / NANOSEC,
14959 	    dtrace_state_deadman, state);
14960 #endif
14961 
14962 	state->dts_activity = DTRACE_ACTIVITY_WARMUP;
14963 
14964 #ifdef illumos
14965 	if (state->dts_getf != 0 &&
14966 	    !(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)) {
14967 		/*
14968 		 * We don't have kernel privs but we have at least one call
14969 		 * to getf(); we need to bump our zone's count, and (if
14970 		 * this is the first enabling to have an unprivileged call
14971 		 * to getf()) we need to hook into closef().
14972 		 */
14973 		state->dts_cred.dcr_cred->cr_zone->zone_dtrace_getf++;
14974 
14975 		if (dtrace_getf++ == 0) {
14976 			ASSERT(dtrace_closef == NULL);
14977 			dtrace_closef = dtrace_getf_barrier;
14978 		}
14979 	}
14980 #endif
14981 
14982 	/*
14983 	 * Now it's time to actually fire the BEGIN probe.  We need to disable
14984 	 * interrupts here both to record the CPU on which we fired the BEGIN
14985 	 * probe (the data from this CPU will be processed first at user
14986 	 * level) and to manually activate the buffer for this CPU.
14987 	 */
14988 	cookie = dtrace_interrupt_disable();
14989 	*cpu = curcpu;
14990 	ASSERT(state->dts_buffer[*cpu].dtb_flags & DTRACEBUF_INACTIVE);
14991 	state->dts_buffer[*cpu].dtb_flags &= ~DTRACEBUF_INACTIVE;
14992 
14993 	dtrace_probe(dtrace_probeid_begin,
14994 	    (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
14995 	dtrace_interrupt_enable(cookie);
14996 	/*
14997 	 * We may have had an exit action from a BEGIN probe; only change our
14998 	 * state to ACTIVE if we're still in WARMUP.
14999 	 */
15000 	ASSERT(state->dts_activity == DTRACE_ACTIVITY_WARMUP ||
15001 	    state->dts_activity == DTRACE_ACTIVITY_DRAINING);
15002 
15003 	if (state->dts_activity == DTRACE_ACTIVITY_WARMUP)
15004 		state->dts_activity = DTRACE_ACTIVITY_ACTIVE;
15005 
15006 #ifdef __FreeBSD__
15007 	/*
15008 	 * We enable anonymous tracing before APs are started, so we must
15009 	 * activate buffers using the current CPU.
15010 	 */
15011 	if (state == dtrace_anon.dta_state)
15012 		for (int i = 0; i < NCPU; i++)
15013 			dtrace_buffer_activate_cpu(state, i);
15014 	else
15015 		dtrace_xcall(DTRACE_CPUALL,
15016 		    (dtrace_xcall_t)dtrace_buffer_activate, state);
15017 #else
15018 	/*
15019 	 * Regardless of whether or not now we're in ACTIVE or DRAINING, we
15020 	 * want each CPU to transition its principal buffer out of the
15021 	 * INACTIVE state.  Doing this assures that no CPU will suddenly begin
15022 	 * processing an ECB halfway down a probe's ECB chain; all CPUs will
15023 	 * atomically transition from processing none of a state's ECBs to
15024 	 * processing all of them.
15025 	 */
15026 	dtrace_xcall(DTRACE_CPUALL,
15027 	    (dtrace_xcall_t)dtrace_buffer_activate, state);
15028 #endif
15029 	goto out;
15030 
15031 err:
15032 	dtrace_buffer_free(state->dts_buffer);
15033 	dtrace_buffer_free(state->dts_aggbuffer);
15034 
15035 	if ((nspec = state->dts_nspeculations) == 0) {
15036 		ASSERT(state->dts_speculations == NULL);
15037 		goto out;
15038 	}
15039 
15040 	spec = state->dts_speculations;
15041 	ASSERT(spec != NULL);
15042 
15043 	for (i = 0; i < state->dts_nspeculations; i++) {
15044 		if ((buf = spec[i].dtsp_buffer) == NULL)
15045 			break;
15046 
15047 		dtrace_buffer_free(buf);
15048 		kmem_free(buf, bufsize);
15049 	}
15050 
15051 	kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
15052 	state->dts_nspeculations = 0;
15053 	state->dts_speculations = NULL;
15054 
15055 out:
15056 	mutex_exit(&dtrace_lock);
15057 	mutex_exit(&cpu_lock);
15058 
15059 	return (rval);
15060 }
15061 
15062 static int
15063 dtrace_state_stop(dtrace_state_t *state, processorid_t *cpu)
15064 {
15065 	dtrace_icookie_t cookie;
15066 
15067 	ASSERT(MUTEX_HELD(&dtrace_lock));
15068 
15069 	if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE &&
15070 	    state->dts_activity != DTRACE_ACTIVITY_DRAINING)
15071 		return (EINVAL);
15072 
15073 	/*
15074 	 * We'll set the activity to DTRACE_ACTIVITY_DRAINING, and issue a sync
15075 	 * to be sure that every CPU has seen it.  See below for the details
15076 	 * on why this is done.
15077 	 */
15078 	state->dts_activity = DTRACE_ACTIVITY_DRAINING;
15079 	dtrace_sync();
15080 
15081 	/*
15082 	 * By this point, it is impossible for any CPU to be still processing
15083 	 * with DTRACE_ACTIVITY_ACTIVE.  We can thus set our activity to
15084 	 * DTRACE_ACTIVITY_COOLDOWN and know that we're not racing with any
15085 	 * other CPU in dtrace_buffer_reserve().  This allows dtrace_probe()
15086 	 * and callees to know that the activity is DTRACE_ACTIVITY_COOLDOWN
15087 	 * iff we're in the END probe.
15088 	 */
15089 	state->dts_activity = DTRACE_ACTIVITY_COOLDOWN;
15090 	dtrace_sync();
15091 	ASSERT(state->dts_activity == DTRACE_ACTIVITY_COOLDOWN);
15092 
15093 	/*
15094 	 * Finally, we can release the reserve and call the END probe.  We
15095 	 * disable interrupts across calling the END probe to allow us to
15096 	 * return the CPU on which we actually called the END probe.  This
15097 	 * allows user-land to be sure that this CPU's principal buffer is
15098 	 * processed last.
15099 	 */
15100 	state->dts_reserve = 0;
15101 
15102 	cookie = dtrace_interrupt_disable();
15103 	*cpu = curcpu;
15104 	dtrace_probe(dtrace_probeid_end,
15105 	    (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
15106 	dtrace_interrupt_enable(cookie);
15107 
15108 	state->dts_activity = DTRACE_ACTIVITY_STOPPED;
15109 	dtrace_sync();
15110 
15111 #ifdef illumos
15112 	if (state->dts_getf != 0 &&
15113 	    !(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)) {
15114 		/*
15115 		 * We don't have kernel privs but we have at least one call
15116 		 * to getf(); we need to lower our zone's count, and (if
15117 		 * this is the last enabling to have an unprivileged call
15118 		 * to getf()) we need to clear the closef() hook.
15119 		 */
15120 		ASSERT(state->dts_cred.dcr_cred->cr_zone->zone_dtrace_getf > 0);
15121 		ASSERT(dtrace_closef == dtrace_getf_barrier);
15122 		ASSERT(dtrace_getf > 0);
15123 
15124 		state->dts_cred.dcr_cred->cr_zone->zone_dtrace_getf--;
15125 
15126 		if (--dtrace_getf == 0)
15127 			dtrace_closef = NULL;
15128 	}
15129 #endif
15130 
15131 	return (0);
15132 }
15133 
15134 static int
15135 dtrace_state_option(dtrace_state_t *state, dtrace_optid_t option,
15136     dtrace_optval_t val)
15137 {
15138 	ASSERT(MUTEX_HELD(&dtrace_lock));
15139 
15140 	if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
15141 		return (EBUSY);
15142 
15143 	if (option >= DTRACEOPT_MAX)
15144 		return (EINVAL);
15145 
15146 	if (option != DTRACEOPT_CPU && val < 0)
15147 		return (EINVAL);
15148 
15149 	switch (option) {
15150 	case DTRACEOPT_DESTRUCTIVE:
15151 		if (dtrace_destructive_disallow)
15152 			return (EACCES);
15153 
15154 		state->dts_cred.dcr_destructive = 1;
15155 		break;
15156 
15157 	case DTRACEOPT_BUFSIZE:
15158 	case DTRACEOPT_DYNVARSIZE:
15159 	case DTRACEOPT_AGGSIZE:
15160 	case DTRACEOPT_SPECSIZE:
15161 	case DTRACEOPT_STRSIZE:
15162 		if (val < 0)
15163 			return (EINVAL);
15164 
15165 		if (val >= LONG_MAX) {
15166 			/*
15167 			 * If this is an otherwise negative value, set it to
15168 			 * the highest multiple of 128m less than LONG_MAX.
15169 			 * Technically, we're adjusting the size without
15170 			 * regard to the buffer resizing policy, but in fact,
15171 			 * this has no effect -- if we set the buffer size to
15172 			 * ~LONG_MAX and the buffer policy is ultimately set to
15173 			 * be "manual", the buffer allocation is guaranteed to
15174 			 * fail, if only because the allocation requires two
15175 			 * buffers.  (We set the the size to the highest
15176 			 * multiple of 128m because it ensures that the size
15177 			 * will remain a multiple of a megabyte when
15178 			 * repeatedly halved -- all the way down to 15m.)
15179 			 */
15180 			val = LONG_MAX - (1 << 27) + 1;
15181 		}
15182 	}
15183 
15184 	state->dts_options[option] = val;
15185 
15186 	return (0);
15187 }
15188 
15189 static void
15190 dtrace_state_destroy(dtrace_state_t *state)
15191 {
15192 	dtrace_ecb_t *ecb;
15193 	dtrace_vstate_t *vstate = &state->dts_vstate;
15194 #ifdef illumos
15195 	minor_t minor = getminor(state->dts_dev);
15196 #endif
15197 	int i, bufsize = NCPU * sizeof (dtrace_buffer_t);
15198 	dtrace_speculation_t *spec = state->dts_speculations;
15199 	int nspec = state->dts_nspeculations;
15200 	uint32_t match;
15201 
15202 	ASSERT(MUTEX_HELD(&dtrace_lock));
15203 	ASSERT(MUTEX_HELD(&cpu_lock));
15204 
15205 	/*
15206 	 * First, retract any retained enablings for this state.
15207 	 */
15208 	dtrace_enabling_retract(state);
15209 	ASSERT(state->dts_nretained == 0);
15210 
15211 	if (state->dts_activity == DTRACE_ACTIVITY_ACTIVE ||
15212 	    state->dts_activity == DTRACE_ACTIVITY_DRAINING) {
15213 		/*
15214 		 * We have managed to come into dtrace_state_destroy() on a
15215 		 * hot enabling -- almost certainly because of a disorderly
15216 		 * shutdown of a consumer.  (That is, a consumer that is
15217 		 * exiting without having called dtrace_stop().) In this case,
15218 		 * we're going to set our activity to be KILLED, and then
15219 		 * issue a sync to be sure that everyone is out of probe
15220 		 * context before we start blowing away ECBs.
15221 		 */
15222 		state->dts_activity = DTRACE_ACTIVITY_KILLED;
15223 		dtrace_sync();
15224 	}
15225 
15226 	/*
15227 	 * Release the credential hold we took in dtrace_state_create().
15228 	 */
15229 	if (state->dts_cred.dcr_cred != NULL)
15230 		crfree(state->dts_cred.dcr_cred);
15231 
15232 	/*
15233 	 * Now we can safely disable and destroy any enabled probes.  Because
15234 	 * any DTRACE_PRIV_KERNEL probes may actually be slowing our progress
15235 	 * (especially if they're all enabled), we take two passes through the
15236 	 * ECBs:  in the first, we disable just DTRACE_PRIV_KERNEL probes, and
15237 	 * in the second we disable whatever is left over.
15238 	 */
15239 	for (match = DTRACE_PRIV_KERNEL; ; match = 0) {
15240 		for (i = 0; i < state->dts_necbs; i++) {
15241 			if ((ecb = state->dts_ecbs[i]) == NULL)
15242 				continue;
15243 
15244 			if (match && ecb->dte_probe != NULL) {
15245 				dtrace_probe_t *probe = ecb->dte_probe;
15246 				dtrace_provider_t *prov = probe->dtpr_provider;
15247 
15248 				if (!(prov->dtpv_priv.dtpp_flags & match))
15249 					continue;
15250 			}
15251 
15252 			dtrace_ecb_disable(ecb);
15253 			dtrace_ecb_destroy(ecb);
15254 		}
15255 
15256 		if (!match)
15257 			break;
15258 	}
15259 
15260 	/*
15261 	 * Before we free the buffers, perform one more sync to assure that
15262 	 * every CPU is out of probe context.
15263 	 */
15264 	dtrace_sync();
15265 
15266 	dtrace_buffer_free(state->dts_buffer);
15267 	dtrace_buffer_free(state->dts_aggbuffer);
15268 
15269 	for (i = 0; i < nspec; i++)
15270 		dtrace_buffer_free(spec[i].dtsp_buffer);
15271 
15272 #ifdef illumos
15273 	if (state->dts_cleaner != CYCLIC_NONE)
15274 		cyclic_remove(state->dts_cleaner);
15275 
15276 	if (state->dts_deadman != CYCLIC_NONE)
15277 		cyclic_remove(state->dts_deadman);
15278 #else
15279 	callout_stop(&state->dts_cleaner);
15280 	callout_drain(&state->dts_cleaner);
15281 	callout_stop(&state->dts_deadman);
15282 	callout_drain(&state->dts_deadman);
15283 #endif
15284 
15285 	dtrace_dstate_fini(&vstate->dtvs_dynvars);
15286 	dtrace_vstate_fini(vstate);
15287 	if (state->dts_ecbs != NULL)
15288 		kmem_free(state->dts_ecbs, state->dts_necbs * sizeof (dtrace_ecb_t *));
15289 
15290 	if (state->dts_aggregations != NULL) {
15291 #ifdef DEBUG
15292 		for (i = 0; i < state->dts_naggregations; i++)
15293 			ASSERT(state->dts_aggregations[i] == NULL);
15294 #endif
15295 		ASSERT(state->dts_naggregations > 0);
15296 		kmem_free(state->dts_aggregations,
15297 		    state->dts_naggregations * sizeof (dtrace_aggregation_t *));
15298 	}
15299 
15300 	kmem_free(state->dts_buffer, bufsize);
15301 	kmem_free(state->dts_aggbuffer, bufsize);
15302 
15303 	for (i = 0; i < nspec; i++)
15304 		kmem_free(spec[i].dtsp_buffer, bufsize);
15305 
15306 	if (spec != NULL)
15307 		kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
15308 
15309 	dtrace_format_destroy(state);
15310 
15311 	if (state->dts_aggid_arena != NULL) {
15312 #ifdef illumos
15313 		vmem_destroy(state->dts_aggid_arena);
15314 #else
15315 		delete_unrhdr(state->dts_aggid_arena);
15316 #endif
15317 		state->dts_aggid_arena = NULL;
15318 	}
15319 #ifdef illumos
15320 	ddi_soft_state_free(dtrace_softstate, minor);
15321 	vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
15322 #endif
15323 }
15324 
15325 /*
15326  * DTrace Anonymous Enabling Functions
15327  */
15328 static dtrace_state_t *
15329 dtrace_anon_grab(void)
15330 {
15331 	dtrace_state_t *state;
15332 
15333 	ASSERT(MUTEX_HELD(&dtrace_lock));
15334 
15335 	if ((state = dtrace_anon.dta_state) == NULL) {
15336 		ASSERT(dtrace_anon.dta_enabling == NULL);
15337 		return (NULL);
15338 	}
15339 
15340 	ASSERT(dtrace_anon.dta_enabling != NULL);
15341 	ASSERT(dtrace_retained != NULL);
15342 
15343 	dtrace_enabling_destroy(dtrace_anon.dta_enabling);
15344 	dtrace_anon.dta_enabling = NULL;
15345 	dtrace_anon.dta_state = NULL;
15346 
15347 	return (state);
15348 }
15349 
15350 static void
15351 dtrace_anon_property(void)
15352 {
15353 	int i, rv;
15354 	dtrace_state_t *state;
15355 	dof_hdr_t *dof;
15356 	char c[32];		/* enough for "dof-data-" + digits */
15357 
15358 	ASSERT(MUTEX_HELD(&dtrace_lock));
15359 	ASSERT(MUTEX_HELD(&cpu_lock));
15360 
15361 	for (i = 0; ; i++) {
15362 		(void) snprintf(c, sizeof (c), "dof-data-%d", i);
15363 
15364 		dtrace_err_verbose = 1;
15365 
15366 		if ((dof = dtrace_dof_property(c)) == NULL) {
15367 			dtrace_err_verbose = 0;
15368 			break;
15369 		}
15370 
15371 #ifdef illumos
15372 		/*
15373 		 * We want to create anonymous state, so we need to transition
15374 		 * the kernel debugger to indicate that DTrace is active.  If
15375 		 * this fails (e.g. because the debugger has modified text in
15376 		 * some way), we won't continue with the processing.
15377 		 */
15378 		if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
15379 			cmn_err(CE_NOTE, "kernel debugger active; anonymous "
15380 			    "enabling ignored.");
15381 			dtrace_dof_destroy(dof);
15382 			break;
15383 		}
15384 #endif
15385 
15386 		/*
15387 		 * If we haven't allocated an anonymous state, we'll do so now.
15388 		 */
15389 		if ((state = dtrace_anon.dta_state) == NULL) {
15390 			state = dtrace_state_create(NULL, NULL);
15391 			dtrace_anon.dta_state = state;
15392 
15393 			if (state == NULL) {
15394 				/*
15395 				 * This basically shouldn't happen:  the only
15396 				 * failure mode from dtrace_state_create() is a
15397 				 * failure of ddi_soft_state_zalloc() that
15398 				 * itself should never happen.  Still, the
15399 				 * interface allows for a failure mode, and
15400 				 * we want to fail as gracefully as possible:
15401 				 * we'll emit an error message and cease
15402 				 * processing anonymous state in this case.
15403 				 */
15404 				cmn_err(CE_WARN, "failed to create "
15405 				    "anonymous state");
15406 				dtrace_dof_destroy(dof);
15407 				break;
15408 			}
15409 		}
15410 
15411 		rv = dtrace_dof_slurp(dof, &state->dts_vstate, CRED(),
15412 		    &dtrace_anon.dta_enabling, 0, B_TRUE);
15413 
15414 		if (rv == 0)
15415 			rv = dtrace_dof_options(dof, state);
15416 
15417 		dtrace_err_verbose = 0;
15418 		dtrace_dof_destroy(dof);
15419 
15420 		if (rv != 0) {
15421 			/*
15422 			 * This is malformed DOF; chuck any anonymous state
15423 			 * that we created.
15424 			 */
15425 			ASSERT(dtrace_anon.dta_enabling == NULL);
15426 			dtrace_state_destroy(state);
15427 			dtrace_anon.dta_state = NULL;
15428 			break;
15429 		}
15430 
15431 		ASSERT(dtrace_anon.dta_enabling != NULL);
15432 	}
15433 
15434 	if (dtrace_anon.dta_enabling != NULL) {
15435 		int rval;
15436 
15437 		/*
15438 		 * dtrace_enabling_retain() can only fail because we are
15439 		 * trying to retain more enablings than are allowed -- but
15440 		 * we only have one anonymous enabling, and we are guaranteed
15441 		 * to be allowed at least one retained enabling; we assert
15442 		 * that dtrace_enabling_retain() returns success.
15443 		 */
15444 		rval = dtrace_enabling_retain(dtrace_anon.dta_enabling);
15445 		ASSERT(rval == 0);
15446 
15447 		dtrace_enabling_dump(dtrace_anon.dta_enabling);
15448 	}
15449 }
15450 
15451 /*
15452  * DTrace Helper Functions
15453  */
15454 static void
15455 dtrace_helper_trace(dtrace_helper_action_t *helper,
15456     dtrace_mstate_t *mstate, dtrace_vstate_t *vstate, int where)
15457 {
15458 	uint32_t size, next, nnext, i;
15459 	dtrace_helptrace_t *ent, *buffer;
15460 	uint16_t flags = cpu_core[curcpu].cpuc_dtrace_flags;
15461 
15462 	if ((buffer = dtrace_helptrace_buffer) == NULL)
15463 		return;
15464 
15465 	ASSERT(vstate->dtvs_nlocals <= dtrace_helptrace_nlocals);
15466 
15467 	/*
15468 	 * What would a tracing framework be without its own tracing
15469 	 * framework?  (Well, a hell of a lot simpler, for starters...)
15470 	 */
15471 	size = sizeof (dtrace_helptrace_t) + dtrace_helptrace_nlocals *
15472 	    sizeof (uint64_t) - sizeof (uint64_t);
15473 
15474 	/*
15475 	 * Iterate until we can allocate a slot in the trace buffer.
15476 	 */
15477 	do {
15478 		next = dtrace_helptrace_next;
15479 
15480 		if (next + size < dtrace_helptrace_bufsize) {
15481 			nnext = next + size;
15482 		} else {
15483 			nnext = size;
15484 		}
15485 	} while (dtrace_cas32(&dtrace_helptrace_next, next, nnext) != next);
15486 
15487 	/*
15488 	 * We have our slot; fill it in.
15489 	 */
15490 	if (nnext == size) {
15491 		dtrace_helptrace_wrapped++;
15492 		next = 0;
15493 	}
15494 
15495 	ent = (dtrace_helptrace_t *)((uintptr_t)buffer + next);
15496 	ent->dtht_helper = helper;
15497 	ent->dtht_where = where;
15498 	ent->dtht_nlocals = vstate->dtvs_nlocals;
15499 
15500 	ent->dtht_fltoffs = (mstate->dtms_present & DTRACE_MSTATE_FLTOFFS) ?
15501 	    mstate->dtms_fltoffs : -1;
15502 	ent->dtht_fault = DTRACE_FLAGS2FLT(flags);
15503 	ent->dtht_illval = cpu_core[curcpu].cpuc_dtrace_illval;
15504 
15505 	for (i = 0; i < vstate->dtvs_nlocals; i++) {
15506 		dtrace_statvar_t *svar;
15507 
15508 		if ((svar = vstate->dtvs_locals[i]) == NULL)
15509 			continue;
15510 
15511 		ASSERT(svar->dtsv_size >= NCPU * sizeof (uint64_t));
15512 		ent->dtht_locals[i] =
15513 		    ((uint64_t *)(uintptr_t)svar->dtsv_data)[curcpu];
15514 	}
15515 }
15516 
15517 static uint64_t
15518 dtrace_helper(int which, dtrace_mstate_t *mstate,
15519     dtrace_state_t *state, uint64_t arg0, uint64_t arg1)
15520 {
15521 	uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
15522 	uint64_t sarg0 = mstate->dtms_arg[0];
15523 	uint64_t sarg1 = mstate->dtms_arg[1];
15524 	uint64_t rval = 0;
15525 	dtrace_helpers_t *helpers = curproc->p_dtrace_helpers;
15526 	dtrace_helper_action_t *helper;
15527 	dtrace_vstate_t *vstate;
15528 	dtrace_difo_t *pred;
15529 	int i, trace = dtrace_helptrace_buffer != NULL;
15530 
15531 	ASSERT(which >= 0 && which < DTRACE_NHELPER_ACTIONS);
15532 
15533 	if (helpers == NULL)
15534 		return (0);
15535 
15536 	if ((helper = helpers->dthps_actions[which]) == NULL)
15537 		return (0);
15538 
15539 	vstate = &helpers->dthps_vstate;
15540 	mstate->dtms_arg[0] = arg0;
15541 	mstate->dtms_arg[1] = arg1;
15542 
15543 	/*
15544 	 * Now iterate over each helper.  If its predicate evaluates to 'true',
15545 	 * we'll call the corresponding actions.  Note that the below calls
15546 	 * to dtrace_dif_emulate() may set faults in machine state.  This is
15547 	 * okay:  our caller (the outer dtrace_dif_emulate()) will simply plow
15548 	 * the stored DIF offset with its own (which is the desired behavior).
15549 	 * Also, note the calls to dtrace_dif_emulate() may allocate scratch
15550 	 * from machine state; this is okay, too.
15551 	 */
15552 	for (; helper != NULL; helper = helper->dtha_next) {
15553 		if ((pred = helper->dtha_predicate) != NULL) {
15554 			if (trace)
15555 				dtrace_helper_trace(helper, mstate, vstate, 0);
15556 
15557 			if (!dtrace_dif_emulate(pred, mstate, vstate, state))
15558 				goto next;
15559 
15560 			if (*flags & CPU_DTRACE_FAULT)
15561 				goto err;
15562 		}
15563 
15564 		for (i = 0; i < helper->dtha_nactions; i++) {
15565 			if (trace)
15566 				dtrace_helper_trace(helper,
15567 				    mstate, vstate, i + 1);
15568 
15569 			rval = dtrace_dif_emulate(helper->dtha_actions[i],
15570 			    mstate, vstate, state);
15571 
15572 			if (*flags & CPU_DTRACE_FAULT)
15573 				goto err;
15574 		}
15575 
15576 next:
15577 		if (trace)
15578 			dtrace_helper_trace(helper, mstate, vstate,
15579 			    DTRACE_HELPTRACE_NEXT);
15580 	}
15581 
15582 	if (trace)
15583 		dtrace_helper_trace(helper, mstate, vstate,
15584 		    DTRACE_HELPTRACE_DONE);
15585 
15586 	/*
15587 	 * Restore the arg0 that we saved upon entry.
15588 	 */
15589 	mstate->dtms_arg[0] = sarg0;
15590 	mstate->dtms_arg[1] = sarg1;
15591 
15592 	return (rval);
15593 
15594 err:
15595 	if (trace)
15596 		dtrace_helper_trace(helper, mstate, vstate,
15597 		    DTRACE_HELPTRACE_ERR);
15598 
15599 	/*
15600 	 * Restore the arg0 that we saved upon entry.
15601 	 */
15602 	mstate->dtms_arg[0] = sarg0;
15603 	mstate->dtms_arg[1] = sarg1;
15604 
15605 	return (0);
15606 }
15607 
15608 static void
15609 dtrace_helper_action_destroy(dtrace_helper_action_t *helper,
15610     dtrace_vstate_t *vstate)
15611 {
15612 	int i;
15613 
15614 	if (helper->dtha_predicate != NULL)
15615 		dtrace_difo_release(helper->dtha_predicate, vstate);
15616 
15617 	for (i = 0; i < helper->dtha_nactions; i++) {
15618 		ASSERT(helper->dtha_actions[i] != NULL);
15619 		dtrace_difo_release(helper->dtha_actions[i], vstate);
15620 	}
15621 
15622 	kmem_free(helper->dtha_actions,
15623 	    helper->dtha_nactions * sizeof (dtrace_difo_t *));
15624 	kmem_free(helper, sizeof (dtrace_helper_action_t));
15625 }
15626 
15627 static int
15628 dtrace_helper_destroygen(dtrace_helpers_t *help, int gen)
15629 {
15630 	proc_t *p = curproc;
15631 	dtrace_vstate_t *vstate;
15632 	int i;
15633 
15634 	if (help == NULL)
15635 		help = p->p_dtrace_helpers;
15636 
15637 	ASSERT(MUTEX_HELD(&dtrace_lock));
15638 
15639 	if (help == NULL || gen > help->dthps_generation)
15640 		return (EINVAL);
15641 
15642 	vstate = &help->dthps_vstate;
15643 
15644 	for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
15645 		dtrace_helper_action_t *last = NULL, *h, *next;
15646 
15647 		for (h = help->dthps_actions[i]; h != NULL; h = next) {
15648 			next = h->dtha_next;
15649 
15650 			if (h->dtha_generation == gen) {
15651 				if (last != NULL) {
15652 					last->dtha_next = next;
15653 				} else {
15654 					help->dthps_actions[i] = next;
15655 				}
15656 
15657 				dtrace_helper_action_destroy(h, vstate);
15658 			} else {
15659 				last = h;
15660 			}
15661 		}
15662 	}
15663 
15664 	/*
15665 	 * Interate until we've cleared out all helper providers with the
15666 	 * given generation number.
15667 	 */
15668 	for (;;) {
15669 		dtrace_helper_provider_t *prov;
15670 
15671 		/*
15672 		 * Look for a helper provider with the right generation. We
15673 		 * have to start back at the beginning of the list each time
15674 		 * because we drop dtrace_lock. It's unlikely that we'll make
15675 		 * more than two passes.
15676 		 */
15677 		for (i = 0; i < help->dthps_nprovs; i++) {
15678 			prov = help->dthps_provs[i];
15679 
15680 			if (prov->dthp_generation == gen)
15681 				break;
15682 		}
15683 
15684 		/*
15685 		 * If there were no matches, we're done.
15686 		 */
15687 		if (i == help->dthps_nprovs)
15688 			break;
15689 
15690 		/*
15691 		 * Move the last helper provider into this slot.
15692 		 */
15693 		help->dthps_nprovs--;
15694 		help->dthps_provs[i] = help->dthps_provs[help->dthps_nprovs];
15695 		help->dthps_provs[help->dthps_nprovs] = NULL;
15696 
15697 		mutex_exit(&dtrace_lock);
15698 
15699 		/*
15700 		 * If we have a meta provider, remove this helper provider.
15701 		 */
15702 		mutex_enter(&dtrace_meta_lock);
15703 		if (dtrace_meta_pid != NULL) {
15704 			ASSERT(dtrace_deferred_pid == NULL);
15705 			dtrace_helper_provider_remove(&prov->dthp_prov,
15706 			    p->p_pid);
15707 		}
15708 		mutex_exit(&dtrace_meta_lock);
15709 
15710 		dtrace_helper_provider_destroy(prov);
15711 
15712 		mutex_enter(&dtrace_lock);
15713 	}
15714 
15715 	return (0);
15716 }
15717 
15718 static int
15719 dtrace_helper_validate(dtrace_helper_action_t *helper)
15720 {
15721 	int err = 0, i;
15722 	dtrace_difo_t *dp;
15723 
15724 	if ((dp = helper->dtha_predicate) != NULL)
15725 		err += dtrace_difo_validate_helper(dp);
15726 
15727 	for (i = 0; i < helper->dtha_nactions; i++)
15728 		err += dtrace_difo_validate_helper(helper->dtha_actions[i]);
15729 
15730 	return (err == 0);
15731 }
15732 
15733 static int
15734 dtrace_helper_action_add(int which, dtrace_ecbdesc_t *ep,
15735     dtrace_helpers_t *help)
15736 {
15737 	dtrace_helper_action_t *helper, *last;
15738 	dtrace_actdesc_t *act;
15739 	dtrace_vstate_t *vstate;
15740 	dtrace_predicate_t *pred;
15741 	int count = 0, nactions = 0, i;
15742 
15743 	if (which < 0 || which >= DTRACE_NHELPER_ACTIONS)
15744 		return (EINVAL);
15745 
15746 	last = help->dthps_actions[which];
15747 	vstate = &help->dthps_vstate;
15748 
15749 	for (count = 0; last != NULL; last = last->dtha_next) {
15750 		count++;
15751 		if (last->dtha_next == NULL)
15752 			break;
15753 	}
15754 
15755 	/*
15756 	 * If we already have dtrace_helper_actions_max helper actions for this
15757 	 * helper action type, we'll refuse to add a new one.
15758 	 */
15759 	if (count >= dtrace_helper_actions_max)
15760 		return (ENOSPC);
15761 
15762 	helper = kmem_zalloc(sizeof (dtrace_helper_action_t), KM_SLEEP);
15763 	helper->dtha_generation = help->dthps_generation;
15764 
15765 	if ((pred = ep->dted_pred.dtpdd_predicate) != NULL) {
15766 		ASSERT(pred->dtp_difo != NULL);
15767 		dtrace_difo_hold(pred->dtp_difo);
15768 		helper->dtha_predicate = pred->dtp_difo;
15769 	}
15770 
15771 	for (act = ep->dted_action; act != NULL; act = act->dtad_next) {
15772 		if (act->dtad_kind != DTRACEACT_DIFEXPR)
15773 			goto err;
15774 
15775 		if (act->dtad_difo == NULL)
15776 			goto err;
15777 
15778 		nactions++;
15779 	}
15780 
15781 	helper->dtha_actions = kmem_zalloc(sizeof (dtrace_difo_t *) *
15782 	    (helper->dtha_nactions = nactions), KM_SLEEP);
15783 
15784 	for (act = ep->dted_action, i = 0; act != NULL; act = act->dtad_next) {
15785 		dtrace_difo_hold(act->dtad_difo);
15786 		helper->dtha_actions[i++] = act->dtad_difo;
15787 	}
15788 
15789 	if (!dtrace_helper_validate(helper))
15790 		goto err;
15791 
15792 	if (last == NULL) {
15793 		help->dthps_actions[which] = helper;
15794 	} else {
15795 		last->dtha_next = helper;
15796 	}
15797 
15798 	if (vstate->dtvs_nlocals > dtrace_helptrace_nlocals) {
15799 		dtrace_helptrace_nlocals = vstate->dtvs_nlocals;
15800 		dtrace_helptrace_next = 0;
15801 	}
15802 
15803 	return (0);
15804 err:
15805 	dtrace_helper_action_destroy(helper, vstate);
15806 	return (EINVAL);
15807 }
15808 
15809 static void
15810 dtrace_helper_provider_register(proc_t *p, dtrace_helpers_t *help,
15811     dof_helper_t *dofhp)
15812 {
15813 	ASSERT(MUTEX_NOT_HELD(&dtrace_lock));
15814 
15815 	mutex_enter(&dtrace_meta_lock);
15816 	mutex_enter(&dtrace_lock);
15817 
15818 	if (!dtrace_attached() || dtrace_meta_pid == NULL) {
15819 		/*
15820 		 * If the dtrace module is loaded but not attached, or if
15821 		 * there aren't isn't a meta provider registered to deal with
15822 		 * these provider descriptions, we need to postpone creating
15823 		 * the actual providers until later.
15824 		 */
15825 
15826 		if (help->dthps_next == NULL && help->dthps_prev == NULL &&
15827 		    dtrace_deferred_pid != help) {
15828 			help->dthps_deferred = 1;
15829 			help->dthps_pid = p->p_pid;
15830 			help->dthps_next = dtrace_deferred_pid;
15831 			help->dthps_prev = NULL;
15832 			if (dtrace_deferred_pid != NULL)
15833 				dtrace_deferred_pid->dthps_prev = help;
15834 			dtrace_deferred_pid = help;
15835 		}
15836 
15837 		mutex_exit(&dtrace_lock);
15838 
15839 	} else if (dofhp != NULL) {
15840 		/*
15841 		 * If the dtrace module is loaded and we have a particular
15842 		 * helper provider description, pass that off to the
15843 		 * meta provider.
15844 		 */
15845 
15846 		mutex_exit(&dtrace_lock);
15847 
15848 		dtrace_helper_provide(dofhp, p->p_pid);
15849 
15850 	} else {
15851 		/*
15852 		 * Otherwise, just pass all the helper provider descriptions
15853 		 * off to the meta provider.
15854 		 */
15855 
15856 		int i;
15857 		mutex_exit(&dtrace_lock);
15858 
15859 		for (i = 0; i < help->dthps_nprovs; i++) {
15860 			dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
15861 			    p->p_pid);
15862 		}
15863 	}
15864 
15865 	mutex_exit(&dtrace_meta_lock);
15866 }
15867 
15868 static int
15869 dtrace_helper_provider_add(dof_helper_t *dofhp, dtrace_helpers_t *help, int gen)
15870 {
15871 	dtrace_helper_provider_t *hprov, **tmp_provs;
15872 	uint_t tmp_maxprovs, i;
15873 
15874 	ASSERT(MUTEX_HELD(&dtrace_lock));
15875 	ASSERT(help != NULL);
15876 
15877 	/*
15878 	 * If we already have dtrace_helper_providers_max helper providers,
15879 	 * we're refuse to add a new one.
15880 	 */
15881 	if (help->dthps_nprovs >= dtrace_helper_providers_max)
15882 		return (ENOSPC);
15883 
15884 	/*
15885 	 * Check to make sure this isn't a duplicate.
15886 	 */
15887 	for (i = 0; i < help->dthps_nprovs; i++) {
15888 		if (dofhp->dofhp_addr ==
15889 		    help->dthps_provs[i]->dthp_prov.dofhp_addr)
15890 			return (EALREADY);
15891 	}
15892 
15893 	hprov = kmem_zalloc(sizeof (dtrace_helper_provider_t), KM_SLEEP);
15894 	hprov->dthp_prov = *dofhp;
15895 	hprov->dthp_ref = 1;
15896 	hprov->dthp_generation = gen;
15897 
15898 	/*
15899 	 * Allocate a bigger table for helper providers if it's already full.
15900 	 */
15901 	if (help->dthps_maxprovs == help->dthps_nprovs) {
15902 		tmp_maxprovs = help->dthps_maxprovs;
15903 		tmp_provs = help->dthps_provs;
15904 
15905 		if (help->dthps_maxprovs == 0)
15906 			help->dthps_maxprovs = 2;
15907 		else
15908 			help->dthps_maxprovs *= 2;
15909 		if (help->dthps_maxprovs > dtrace_helper_providers_max)
15910 			help->dthps_maxprovs = dtrace_helper_providers_max;
15911 
15912 		ASSERT(tmp_maxprovs < help->dthps_maxprovs);
15913 
15914 		help->dthps_provs = kmem_zalloc(help->dthps_maxprovs *
15915 		    sizeof (dtrace_helper_provider_t *), KM_SLEEP);
15916 
15917 		if (tmp_provs != NULL) {
15918 			bcopy(tmp_provs, help->dthps_provs, tmp_maxprovs *
15919 			    sizeof (dtrace_helper_provider_t *));
15920 			kmem_free(tmp_provs, tmp_maxprovs *
15921 			    sizeof (dtrace_helper_provider_t *));
15922 		}
15923 	}
15924 
15925 	help->dthps_provs[help->dthps_nprovs] = hprov;
15926 	help->dthps_nprovs++;
15927 
15928 	return (0);
15929 }
15930 
15931 static void
15932 dtrace_helper_provider_destroy(dtrace_helper_provider_t *hprov)
15933 {
15934 	mutex_enter(&dtrace_lock);
15935 
15936 	if (--hprov->dthp_ref == 0) {
15937 		dof_hdr_t *dof;
15938 		mutex_exit(&dtrace_lock);
15939 		dof = (dof_hdr_t *)(uintptr_t)hprov->dthp_prov.dofhp_dof;
15940 		dtrace_dof_destroy(dof);
15941 		kmem_free(hprov, sizeof (dtrace_helper_provider_t));
15942 	} else {
15943 		mutex_exit(&dtrace_lock);
15944 	}
15945 }
15946 
15947 static int
15948 dtrace_helper_provider_validate(dof_hdr_t *dof, dof_sec_t *sec)
15949 {
15950 	uintptr_t daddr = (uintptr_t)dof;
15951 	dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
15952 	dof_provider_t *provider;
15953 	dof_probe_t *probe;
15954 	uint8_t *arg;
15955 	char *strtab, *typestr;
15956 	dof_stridx_t typeidx;
15957 	size_t typesz;
15958 	uint_t nprobes, j, k;
15959 
15960 	ASSERT(sec->dofs_type == DOF_SECT_PROVIDER);
15961 
15962 	if (sec->dofs_offset & (sizeof (uint_t) - 1)) {
15963 		dtrace_dof_error(dof, "misaligned section offset");
15964 		return (-1);
15965 	}
15966 
15967 	/*
15968 	 * The section needs to be large enough to contain the DOF provider
15969 	 * structure appropriate for the given version.
15970 	 */
15971 	if (sec->dofs_size <
15972 	    ((dof->dofh_ident[DOF_ID_VERSION] == DOF_VERSION_1) ?
15973 	    offsetof(dof_provider_t, dofpv_prenoffs) :
15974 	    sizeof (dof_provider_t))) {
15975 		dtrace_dof_error(dof, "provider section too small");
15976 		return (-1);
15977 	}
15978 
15979 	provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
15980 	str_sec = dtrace_dof_sect(dof, DOF_SECT_STRTAB, provider->dofpv_strtab);
15981 	prb_sec = dtrace_dof_sect(dof, DOF_SECT_PROBES, provider->dofpv_probes);
15982 	arg_sec = dtrace_dof_sect(dof, DOF_SECT_PRARGS, provider->dofpv_prargs);
15983 	off_sec = dtrace_dof_sect(dof, DOF_SECT_PROFFS, provider->dofpv_proffs);
15984 
15985 	if (str_sec == NULL || prb_sec == NULL ||
15986 	    arg_sec == NULL || off_sec == NULL)
15987 		return (-1);
15988 
15989 	enoff_sec = NULL;
15990 
15991 	if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
15992 	    provider->dofpv_prenoffs != DOF_SECT_NONE &&
15993 	    (enoff_sec = dtrace_dof_sect(dof, DOF_SECT_PRENOFFS,
15994 	    provider->dofpv_prenoffs)) == NULL)
15995 		return (-1);
15996 
15997 	strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
15998 
15999 	if (provider->dofpv_name >= str_sec->dofs_size ||
16000 	    strlen(strtab + provider->dofpv_name) >= DTRACE_PROVNAMELEN) {
16001 		dtrace_dof_error(dof, "invalid provider name");
16002 		return (-1);
16003 	}
16004 
16005 	if (prb_sec->dofs_entsize == 0 ||
16006 	    prb_sec->dofs_entsize > prb_sec->dofs_size) {
16007 		dtrace_dof_error(dof, "invalid entry size");
16008 		return (-1);
16009 	}
16010 
16011 	if (prb_sec->dofs_entsize & (sizeof (uintptr_t) - 1)) {
16012 		dtrace_dof_error(dof, "misaligned entry size");
16013 		return (-1);
16014 	}
16015 
16016 	if (off_sec->dofs_entsize != sizeof (uint32_t)) {
16017 		dtrace_dof_error(dof, "invalid entry size");
16018 		return (-1);
16019 	}
16020 
16021 	if (off_sec->dofs_offset & (sizeof (uint32_t) - 1)) {
16022 		dtrace_dof_error(dof, "misaligned section offset");
16023 		return (-1);
16024 	}
16025 
16026 	if (arg_sec->dofs_entsize != sizeof (uint8_t)) {
16027 		dtrace_dof_error(dof, "invalid entry size");
16028 		return (-1);
16029 	}
16030 
16031 	arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
16032 
16033 	nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
16034 
16035 	/*
16036 	 * Take a pass through the probes to check for errors.
16037 	 */
16038 	for (j = 0; j < nprobes; j++) {
16039 		probe = (dof_probe_t *)(uintptr_t)(daddr +
16040 		    prb_sec->dofs_offset + j * prb_sec->dofs_entsize);
16041 
16042 		if (probe->dofpr_func >= str_sec->dofs_size) {
16043 			dtrace_dof_error(dof, "invalid function name");
16044 			return (-1);
16045 		}
16046 
16047 		if (strlen(strtab + probe->dofpr_func) >= DTRACE_FUNCNAMELEN) {
16048 			dtrace_dof_error(dof, "function name too long");
16049 			/*
16050 			 * Keep going if the function name is too long.
16051 			 * Unlike provider and probe names, we cannot reasonably
16052 			 * impose restrictions on function names, since they're
16053 			 * a property of the code being instrumented. We will
16054 			 * skip this probe in dtrace_helper_provide_one().
16055 			 */
16056 		}
16057 
16058 		if (probe->dofpr_name >= str_sec->dofs_size ||
16059 		    strlen(strtab + probe->dofpr_name) >= DTRACE_NAMELEN) {
16060 			dtrace_dof_error(dof, "invalid probe name");
16061 			return (-1);
16062 		}
16063 
16064 		/*
16065 		 * The offset count must not wrap the index, and the offsets
16066 		 * must also not overflow the section's data.
16067 		 */
16068 		if (probe->dofpr_offidx + probe->dofpr_noffs <
16069 		    probe->dofpr_offidx ||
16070 		    (probe->dofpr_offidx + probe->dofpr_noffs) *
16071 		    off_sec->dofs_entsize > off_sec->dofs_size) {
16072 			dtrace_dof_error(dof, "invalid probe offset");
16073 			return (-1);
16074 		}
16075 
16076 		if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1) {
16077 			/*
16078 			 * If there's no is-enabled offset section, make sure
16079 			 * there aren't any is-enabled offsets. Otherwise
16080 			 * perform the same checks as for probe offsets
16081 			 * (immediately above).
16082 			 */
16083 			if (enoff_sec == NULL) {
16084 				if (probe->dofpr_enoffidx != 0 ||
16085 				    probe->dofpr_nenoffs != 0) {
16086 					dtrace_dof_error(dof, "is-enabled "
16087 					    "offsets with null section");
16088 					return (-1);
16089 				}
16090 			} else if (probe->dofpr_enoffidx +
16091 			    probe->dofpr_nenoffs < probe->dofpr_enoffidx ||
16092 			    (probe->dofpr_enoffidx + probe->dofpr_nenoffs) *
16093 			    enoff_sec->dofs_entsize > enoff_sec->dofs_size) {
16094 				dtrace_dof_error(dof, "invalid is-enabled "
16095 				    "offset");
16096 				return (-1);
16097 			}
16098 
16099 			if (probe->dofpr_noffs + probe->dofpr_nenoffs == 0) {
16100 				dtrace_dof_error(dof, "zero probe and "
16101 				    "is-enabled offsets");
16102 				return (-1);
16103 			}
16104 		} else if (probe->dofpr_noffs == 0) {
16105 			dtrace_dof_error(dof, "zero probe offsets");
16106 			return (-1);
16107 		}
16108 
16109 		if (probe->dofpr_argidx + probe->dofpr_xargc <
16110 		    probe->dofpr_argidx ||
16111 		    (probe->dofpr_argidx + probe->dofpr_xargc) *
16112 		    arg_sec->dofs_entsize > arg_sec->dofs_size) {
16113 			dtrace_dof_error(dof, "invalid args");
16114 			return (-1);
16115 		}
16116 
16117 		typeidx = probe->dofpr_nargv;
16118 		typestr = strtab + probe->dofpr_nargv;
16119 		for (k = 0; k < probe->dofpr_nargc; k++) {
16120 			if (typeidx >= str_sec->dofs_size) {
16121 				dtrace_dof_error(dof, "bad "
16122 				    "native argument type");
16123 				return (-1);
16124 			}
16125 
16126 			typesz = strlen(typestr) + 1;
16127 			if (typesz > DTRACE_ARGTYPELEN) {
16128 				dtrace_dof_error(dof, "native "
16129 				    "argument type too long");
16130 				return (-1);
16131 			}
16132 			typeidx += typesz;
16133 			typestr += typesz;
16134 		}
16135 
16136 		typeidx = probe->dofpr_xargv;
16137 		typestr = strtab + probe->dofpr_xargv;
16138 		for (k = 0; k < probe->dofpr_xargc; k++) {
16139 			if (arg[probe->dofpr_argidx + k] > probe->dofpr_nargc) {
16140 				dtrace_dof_error(dof, "bad "
16141 				    "native argument index");
16142 				return (-1);
16143 			}
16144 
16145 			if (typeidx >= str_sec->dofs_size) {
16146 				dtrace_dof_error(dof, "bad "
16147 				    "translated argument type");
16148 				return (-1);
16149 			}
16150 
16151 			typesz = strlen(typestr) + 1;
16152 			if (typesz > DTRACE_ARGTYPELEN) {
16153 				dtrace_dof_error(dof, "translated argument "
16154 				    "type too long");
16155 				return (-1);
16156 			}
16157 
16158 			typeidx += typesz;
16159 			typestr += typesz;
16160 		}
16161 	}
16162 
16163 	return (0);
16164 }
16165 
16166 static int
16167 #ifdef __FreeBSD__
16168 dtrace_helper_slurp(dof_hdr_t *dof, dof_helper_t *dhp, struct proc *p)
16169 #else
16170 dtrace_helper_slurp(dof_hdr_t *dof, dof_helper_t *dhp)
16171 #endif
16172 {
16173 	dtrace_helpers_t *help;
16174 	dtrace_vstate_t *vstate;
16175 	dtrace_enabling_t *enab = NULL;
16176 #ifndef __FreeBSD__
16177 	proc_t *p = curproc;
16178 #endif
16179 	int i, gen, rv, nhelpers = 0, nprovs = 0, destroy = 1;
16180 	uintptr_t daddr = (uintptr_t)dof;
16181 
16182 	ASSERT(MUTEX_HELD(&dtrace_lock));
16183 
16184 	if ((help = p->p_dtrace_helpers) == NULL)
16185 		help = dtrace_helpers_create(p);
16186 
16187 	vstate = &help->dthps_vstate;
16188 
16189 	if ((rv = dtrace_dof_slurp(dof, vstate, NULL, &enab,
16190 	    dhp != NULL ? dhp->dofhp_addr : 0, B_FALSE)) != 0) {
16191 		dtrace_dof_destroy(dof);
16192 		return (rv);
16193 	}
16194 
16195 	/*
16196 	 * Look for helper providers and validate their descriptions.
16197 	 */
16198 	if (dhp != NULL) {
16199 		for (i = 0; i < dof->dofh_secnum; i++) {
16200 			dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
16201 			    dof->dofh_secoff + i * dof->dofh_secsize);
16202 
16203 			if (sec->dofs_type != DOF_SECT_PROVIDER)
16204 				continue;
16205 
16206 			if (dtrace_helper_provider_validate(dof, sec) != 0) {
16207 				dtrace_enabling_destroy(enab);
16208 				dtrace_dof_destroy(dof);
16209 				return (-1);
16210 			}
16211 
16212 			nprovs++;
16213 		}
16214 	}
16215 
16216 	/*
16217 	 * Now we need to walk through the ECB descriptions in the enabling.
16218 	 */
16219 	for (i = 0; i < enab->dten_ndesc; i++) {
16220 		dtrace_ecbdesc_t *ep = enab->dten_desc[i];
16221 		dtrace_probedesc_t *desc = &ep->dted_probe;
16222 
16223 		if (strcmp(desc->dtpd_provider, "dtrace") != 0)
16224 			continue;
16225 
16226 		if (strcmp(desc->dtpd_mod, "helper") != 0)
16227 			continue;
16228 
16229 		if (strcmp(desc->dtpd_func, "ustack") != 0)
16230 			continue;
16231 
16232 		if ((rv = dtrace_helper_action_add(DTRACE_HELPER_ACTION_USTACK,
16233 		    ep, help)) != 0) {
16234 			/*
16235 			 * Adding this helper action failed -- we are now going
16236 			 * to rip out the entire generation and return failure.
16237 			 */
16238 			(void) dtrace_helper_destroygen(help,
16239 			    help->dthps_generation);
16240 			dtrace_enabling_destroy(enab);
16241 			dtrace_dof_destroy(dof);
16242 			return (-1);
16243 		}
16244 
16245 		nhelpers++;
16246 	}
16247 
16248 	if (nhelpers < enab->dten_ndesc)
16249 		dtrace_dof_error(dof, "unmatched helpers");
16250 
16251 	gen = help->dthps_generation++;
16252 	dtrace_enabling_destroy(enab);
16253 
16254 	if (dhp != NULL && nprovs > 0) {
16255 		/*
16256 		 * Now that this is in-kernel, we change the sense of the
16257 		 * members:  dofhp_dof denotes the in-kernel copy of the DOF
16258 		 * and dofhp_addr denotes the address at user-level.
16259 		 */
16260 		dhp->dofhp_addr = dhp->dofhp_dof;
16261 		dhp->dofhp_dof = (uint64_t)(uintptr_t)dof;
16262 
16263 		if (dtrace_helper_provider_add(dhp, help, gen) == 0) {
16264 			mutex_exit(&dtrace_lock);
16265 			dtrace_helper_provider_register(p, help, dhp);
16266 			mutex_enter(&dtrace_lock);
16267 
16268 			destroy = 0;
16269 		}
16270 	}
16271 
16272 	if (destroy)
16273 		dtrace_dof_destroy(dof);
16274 
16275 	return (gen);
16276 }
16277 
16278 static dtrace_helpers_t *
16279 dtrace_helpers_create(proc_t *p)
16280 {
16281 	dtrace_helpers_t *help;
16282 
16283 	ASSERT(MUTEX_HELD(&dtrace_lock));
16284 	ASSERT(p->p_dtrace_helpers == NULL);
16285 
16286 	help = kmem_zalloc(sizeof (dtrace_helpers_t), KM_SLEEP);
16287 	help->dthps_actions = kmem_zalloc(sizeof (dtrace_helper_action_t *) *
16288 	    DTRACE_NHELPER_ACTIONS, KM_SLEEP);
16289 
16290 	p->p_dtrace_helpers = help;
16291 	dtrace_helpers++;
16292 
16293 	return (help);
16294 }
16295 
16296 #ifdef illumos
16297 static
16298 #endif
16299 void
16300 dtrace_helpers_destroy(proc_t *p)
16301 {
16302 	dtrace_helpers_t *help;
16303 	dtrace_vstate_t *vstate;
16304 #ifdef illumos
16305 	proc_t *p = curproc;
16306 #endif
16307 	int i;
16308 
16309 	mutex_enter(&dtrace_lock);
16310 
16311 	ASSERT(p->p_dtrace_helpers != NULL);
16312 	ASSERT(dtrace_helpers > 0);
16313 
16314 	help = p->p_dtrace_helpers;
16315 	vstate = &help->dthps_vstate;
16316 
16317 	/*
16318 	 * We're now going to lose the help from this process.
16319 	 */
16320 	p->p_dtrace_helpers = NULL;
16321 	dtrace_sync();
16322 
16323 	/*
16324 	 * Destory the helper actions.
16325 	 */
16326 	for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
16327 		dtrace_helper_action_t *h, *next;
16328 
16329 		for (h = help->dthps_actions[i]; h != NULL; h = next) {
16330 			next = h->dtha_next;
16331 			dtrace_helper_action_destroy(h, vstate);
16332 			h = next;
16333 		}
16334 	}
16335 
16336 	mutex_exit(&dtrace_lock);
16337 
16338 	/*
16339 	 * Destroy the helper providers.
16340 	 */
16341 	if (help->dthps_maxprovs > 0) {
16342 		mutex_enter(&dtrace_meta_lock);
16343 		if (dtrace_meta_pid != NULL) {
16344 			ASSERT(dtrace_deferred_pid == NULL);
16345 
16346 			for (i = 0; i < help->dthps_nprovs; i++) {
16347 				dtrace_helper_provider_remove(
16348 				    &help->dthps_provs[i]->dthp_prov, p->p_pid);
16349 			}
16350 		} else {
16351 			mutex_enter(&dtrace_lock);
16352 			ASSERT(help->dthps_deferred == 0 ||
16353 			    help->dthps_next != NULL ||
16354 			    help->dthps_prev != NULL ||
16355 			    help == dtrace_deferred_pid);
16356 
16357 			/*
16358 			 * Remove the helper from the deferred list.
16359 			 */
16360 			if (help->dthps_next != NULL)
16361 				help->dthps_next->dthps_prev = help->dthps_prev;
16362 			if (help->dthps_prev != NULL)
16363 				help->dthps_prev->dthps_next = help->dthps_next;
16364 			if (dtrace_deferred_pid == help) {
16365 				dtrace_deferred_pid = help->dthps_next;
16366 				ASSERT(help->dthps_prev == NULL);
16367 			}
16368 
16369 			mutex_exit(&dtrace_lock);
16370 		}
16371 
16372 		mutex_exit(&dtrace_meta_lock);
16373 
16374 		for (i = 0; i < help->dthps_nprovs; i++) {
16375 			dtrace_helper_provider_destroy(help->dthps_provs[i]);
16376 		}
16377 
16378 		kmem_free(help->dthps_provs, help->dthps_maxprovs *
16379 		    sizeof (dtrace_helper_provider_t *));
16380 	}
16381 
16382 	mutex_enter(&dtrace_lock);
16383 
16384 	dtrace_vstate_fini(&help->dthps_vstate);
16385 	kmem_free(help->dthps_actions,
16386 	    sizeof (dtrace_helper_action_t *) * DTRACE_NHELPER_ACTIONS);
16387 	kmem_free(help, sizeof (dtrace_helpers_t));
16388 
16389 	--dtrace_helpers;
16390 	mutex_exit(&dtrace_lock);
16391 }
16392 
16393 #ifdef illumos
16394 static
16395 #endif
16396 void
16397 dtrace_helpers_duplicate(proc_t *from, proc_t *to)
16398 {
16399 	dtrace_helpers_t *help, *newhelp;
16400 	dtrace_helper_action_t *helper, *new, *last;
16401 	dtrace_difo_t *dp;
16402 	dtrace_vstate_t *vstate;
16403 	int i, j, sz, hasprovs = 0;
16404 
16405 	mutex_enter(&dtrace_lock);
16406 	ASSERT(from->p_dtrace_helpers != NULL);
16407 	ASSERT(dtrace_helpers > 0);
16408 
16409 	help = from->p_dtrace_helpers;
16410 	newhelp = dtrace_helpers_create(to);
16411 	ASSERT(to->p_dtrace_helpers != NULL);
16412 
16413 	newhelp->dthps_generation = help->dthps_generation;
16414 	vstate = &newhelp->dthps_vstate;
16415 
16416 	/*
16417 	 * Duplicate the helper actions.
16418 	 */
16419 	for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
16420 		if ((helper = help->dthps_actions[i]) == NULL)
16421 			continue;
16422 
16423 		for (last = NULL; helper != NULL; helper = helper->dtha_next) {
16424 			new = kmem_zalloc(sizeof (dtrace_helper_action_t),
16425 			    KM_SLEEP);
16426 			new->dtha_generation = helper->dtha_generation;
16427 
16428 			if ((dp = helper->dtha_predicate) != NULL) {
16429 				dp = dtrace_difo_duplicate(dp, vstate);
16430 				new->dtha_predicate = dp;
16431 			}
16432 
16433 			new->dtha_nactions = helper->dtha_nactions;
16434 			sz = sizeof (dtrace_difo_t *) * new->dtha_nactions;
16435 			new->dtha_actions = kmem_alloc(sz, KM_SLEEP);
16436 
16437 			for (j = 0; j < new->dtha_nactions; j++) {
16438 				dtrace_difo_t *dp = helper->dtha_actions[j];
16439 
16440 				ASSERT(dp != NULL);
16441 				dp = dtrace_difo_duplicate(dp, vstate);
16442 				new->dtha_actions[j] = dp;
16443 			}
16444 
16445 			if (last != NULL) {
16446 				last->dtha_next = new;
16447 			} else {
16448 				newhelp->dthps_actions[i] = new;
16449 			}
16450 
16451 			last = new;
16452 		}
16453 	}
16454 
16455 	/*
16456 	 * Duplicate the helper providers and register them with the
16457 	 * DTrace framework.
16458 	 */
16459 	if (help->dthps_nprovs > 0) {
16460 		newhelp->dthps_nprovs = help->dthps_nprovs;
16461 		newhelp->dthps_maxprovs = help->dthps_nprovs;
16462 		newhelp->dthps_provs = kmem_alloc(newhelp->dthps_nprovs *
16463 		    sizeof (dtrace_helper_provider_t *), KM_SLEEP);
16464 		for (i = 0; i < newhelp->dthps_nprovs; i++) {
16465 			newhelp->dthps_provs[i] = help->dthps_provs[i];
16466 			newhelp->dthps_provs[i]->dthp_ref++;
16467 		}
16468 
16469 		hasprovs = 1;
16470 	}
16471 
16472 	mutex_exit(&dtrace_lock);
16473 
16474 	if (hasprovs)
16475 		dtrace_helper_provider_register(to, newhelp, NULL);
16476 }
16477 
16478 /*
16479  * DTrace Hook Functions
16480  */
16481 static void
16482 dtrace_module_loaded(modctl_t *ctl)
16483 {
16484 	dtrace_provider_t *prv;
16485 
16486 	mutex_enter(&dtrace_provider_lock);
16487 #ifdef illumos
16488 	mutex_enter(&mod_lock);
16489 #endif
16490 
16491 #ifdef illumos
16492 	ASSERT(ctl->mod_busy);
16493 #endif
16494 
16495 	/*
16496 	 * We're going to call each providers per-module provide operation
16497 	 * specifying only this module.
16498 	 */
16499 	for (prv = dtrace_provider; prv != NULL; prv = prv->dtpv_next)
16500 		prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
16501 
16502 #ifdef illumos
16503 	mutex_exit(&mod_lock);
16504 #endif
16505 	mutex_exit(&dtrace_provider_lock);
16506 
16507 	/*
16508 	 * If we have any retained enablings, we need to match against them.
16509 	 * Enabling probes requires that cpu_lock be held, and we cannot hold
16510 	 * cpu_lock here -- it is legal for cpu_lock to be held when loading a
16511 	 * module.  (In particular, this happens when loading scheduling
16512 	 * classes.)  So if we have any retained enablings, we need to dispatch
16513 	 * our task queue to do the match for us.
16514 	 */
16515 	mutex_enter(&dtrace_lock);
16516 
16517 	if (dtrace_retained == NULL) {
16518 		mutex_exit(&dtrace_lock);
16519 		return;
16520 	}
16521 
16522 	(void) taskq_dispatch(dtrace_taskq,
16523 	    (task_func_t *)dtrace_enabling_matchall, NULL, TQ_SLEEP);
16524 
16525 	mutex_exit(&dtrace_lock);
16526 
16527 	/*
16528 	 * And now, for a little heuristic sleaze:  in general, we want to
16529 	 * match modules as soon as they load.  However, we cannot guarantee
16530 	 * this, because it would lead us to the lock ordering violation
16531 	 * outlined above.  The common case, of course, is that cpu_lock is
16532 	 * _not_ held -- so we delay here for a clock tick, hoping that that's
16533 	 * long enough for the task queue to do its work.  If it's not, it's
16534 	 * not a serious problem -- it just means that the module that we
16535 	 * just loaded may not be immediately instrumentable.
16536 	 */
16537 	delay(1);
16538 }
16539 
16540 static void
16541 #ifdef illumos
16542 dtrace_module_unloaded(modctl_t *ctl)
16543 #else
16544 dtrace_module_unloaded(modctl_t *ctl, int *error)
16545 #endif
16546 {
16547 	dtrace_probe_t template, *probe, *first, *next;
16548 	dtrace_provider_t *prov;
16549 #ifndef illumos
16550 	char modname[DTRACE_MODNAMELEN];
16551 	size_t len;
16552 #endif
16553 
16554 #ifdef illumos
16555 	template.dtpr_mod = ctl->mod_modname;
16556 #else
16557 	/* Handle the fact that ctl->filename may end in ".ko". */
16558 	strlcpy(modname, ctl->filename, sizeof(modname));
16559 	len = strlen(ctl->filename);
16560 	if (len > 3 && strcmp(modname + len - 3, ".ko") == 0)
16561 		modname[len - 3] = '\0';
16562 	template.dtpr_mod = modname;
16563 #endif
16564 
16565 	mutex_enter(&dtrace_provider_lock);
16566 #ifdef illumos
16567 	mutex_enter(&mod_lock);
16568 #endif
16569 	mutex_enter(&dtrace_lock);
16570 
16571 #ifndef illumos
16572 	if (ctl->nenabled > 0) {
16573 		/* Don't allow unloads if a probe is enabled. */
16574 		mutex_exit(&dtrace_provider_lock);
16575 		mutex_exit(&dtrace_lock);
16576 		*error = -1;
16577 		printf(
16578 	"kldunload: attempt to unload module that has DTrace probes enabled\n");
16579 		return;
16580 	}
16581 #endif
16582 
16583 	if (dtrace_bymod == NULL) {
16584 		/*
16585 		 * The DTrace module is loaded (obviously) but not attached;
16586 		 * we don't have any work to do.
16587 		 */
16588 		mutex_exit(&dtrace_provider_lock);
16589 #ifdef illumos
16590 		mutex_exit(&mod_lock);
16591 #endif
16592 		mutex_exit(&dtrace_lock);
16593 		return;
16594 	}
16595 
16596 	for (probe = first = dtrace_hash_lookup(dtrace_bymod, &template);
16597 	    probe != NULL; probe = probe->dtpr_nextmod) {
16598 		if (probe->dtpr_ecb != NULL) {
16599 			mutex_exit(&dtrace_provider_lock);
16600 #ifdef illumos
16601 			mutex_exit(&mod_lock);
16602 #endif
16603 			mutex_exit(&dtrace_lock);
16604 
16605 			/*
16606 			 * This shouldn't _actually_ be possible -- we're
16607 			 * unloading a module that has an enabled probe in it.
16608 			 * (It's normally up to the provider to make sure that
16609 			 * this can't happen.)  However, because dtps_enable()
16610 			 * doesn't have a failure mode, there can be an
16611 			 * enable/unload race.  Upshot:  we don't want to
16612 			 * assert, but we're not going to disable the
16613 			 * probe, either.
16614 			 */
16615 			if (dtrace_err_verbose) {
16616 #ifdef illumos
16617 				cmn_err(CE_WARN, "unloaded module '%s' had "
16618 				    "enabled probes", ctl->mod_modname);
16619 #else
16620 				cmn_err(CE_WARN, "unloaded module '%s' had "
16621 				    "enabled probes", modname);
16622 #endif
16623 			}
16624 
16625 			return;
16626 		}
16627 	}
16628 
16629 	probe = first;
16630 
16631 	for (first = NULL; probe != NULL; probe = next) {
16632 		ASSERT(dtrace_probes[probe->dtpr_id - 1] == probe);
16633 
16634 		dtrace_probes[probe->dtpr_id - 1] = NULL;
16635 
16636 		next = probe->dtpr_nextmod;
16637 		dtrace_hash_remove(dtrace_bymod, probe);
16638 		dtrace_hash_remove(dtrace_byfunc, probe);
16639 		dtrace_hash_remove(dtrace_byname, probe);
16640 
16641 		if (first == NULL) {
16642 			first = probe;
16643 			probe->dtpr_nextmod = NULL;
16644 		} else {
16645 			probe->dtpr_nextmod = first;
16646 			first = probe;
16647 		}
16648 	}
16649 
16650 	/*
16651 	 * We've removed all of the module's probes from the hash chains and
16652 	 * from the probe array.  Now issue a dtrace_sync() to be sure that
16653 	 * everyone has cleared out from any probe array processing.
16654 	 */
16655 	dtrace_sync();
16656 
16657 	for (probe = first; probe != NULL; probe = first) {
16658 		first = probe->dtpr_nextmod;
16659 		prov = probe->dtpr_provider;
16660 		prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, probe->dtpr_id,
16661 		    probe->dtpr_arg);
16662 		kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
16663 		kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
16664 		kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
16665 #ifdef illumos
16666 		vmem_free(dtrace_arena, (void *)(uintptr_t)probe->dtpr_id, 1);
16667 #else
16668 		free_unr(dtrace_arena, probe->dtpr_id);
16669 #endif
16670 		kmem_free(probe, sizeof (dtrace_probe_t));
16671 	}
16672 
16673 	mutex_exit(&dtrace_lock);
16674 #ifdef illumos
16675 	mutex_exit(&mod_lock);
16676 #endif
16677 	mutex_exit(&dtrace_provider_lock);
16678 }
16679 
16680 #ifndef illumos
16681 static void
16682 dtrace_kld_load(void *arg __unused, linker_file_t lf)
16683 {
16684 
16685 	dtrace_module_loaded(lf);
16686 }
16687 
16688 static void
16689 dtrace_kld_unload_try(void *arg __unused, linker_file_t lf, int *error)
16690 {
16691 
16692 	if (*error != 0)
16693 		/* We already have an error, so don't do anything. */
16694 		return;
16695 	dtrace_module_unloaded(lf, error);
16696 }
16697 #endif
16698 
16699 #ifdef illumos
16700 static void
16701 dtrace_suspend(void)
16702 {
16703 	dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_suspend));
16704 }
16705 
16706 static void
16707 dtrace_resume(void)
16708 {
16709 	dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_resume));
16710 }
16711 #endif
16712 
16713 static int
16714 dtrace_cpu_setup(cpu_setup_t what, processorid_t cpu)
16715 {
16716 	ASSERT(MUTEX_HELD(&cpu_lock));
16717 	mutex_enter(&dtrace_lock);
16718 
16719 	switch (what) {
16720 	case CPU_CONFIG: {
16721 		dtrace_state_t *state;
16722 		dtrace_optval_t *opt, rs, c;
16723 
16724 		/*
16725 		 * For now, we only allocate a new buffer for anonymous state.
16726 		 */
16727 		if ((state = dtrace_anon.dta_state) == NULL)
16728 			break;
16729 
16730 		if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
16731 			break;
16732 
16733 		opt = state->dts_options;
16734 		c = opt[DTRACEOPT_CPU];
16735 
16736 		if (c != DTRACE_CPUALL && c != DTRACEOPT_UNSET && c != cpu)
16737 			break;
16738 
16739 		/*
16740 		 * Regardless of what the actual policy is, we're going to
16741 		 * temporarily set our resize policy to be manual.  We're
16742 		 * also going to temporarily set our CPU option to denote
16743 		 * the newly configured CPU.
16744 		 */
16745 		rs = opt[DTRACEOPT_BUFRESIZE];
16746 		opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_MANUAL;
16747 		opt[DTRACEOPT_CPU] = (dtrace_optval_t)cpu;
16748 
16749 		(void) dtrace_state_buffers(state);
16750 
16751 		opt[DTRACEOPT_BUFRESIZE] = rs;
16752 		opt[DTRACEOPT_CPU] = c;
16753 
16754 		break;
16755 	}
16756 
16757 	case CPU_UNCONFIG:
16758 		/*
16759 		 * We don't free the buffer in the CPU_UNCONFIG case.  (The
16760 		 * buffer will be freed when the consumer exits.)
16761 		 */
16762 		break;
16763 
16764 	default:
16765 		break;
16766 	}
16767 
16768 	mutex_exit(&dtrace_lock);
16769 	return (0);
16770 }
16771 
16772 #ifdef illumos
16773 static void
16774 dtrace_cpu_setup_initial(processorid_t cpu)
16775 {
16776 	(void) dtrace_cpu_setup(CPU_CONFIG, cpu);
16777 }
16778 #endif
16779 
16780 static void
16781 dtrace_toxrange_add(uintptr_t base, uintptr_t limit)
16782 {
16783 	if (dtrace_toxranges >= dtrace_toxranges_max) {
16784 		int osize, nsize;
16785 		dtrace_toxrange_t *range;
16786 
16787 		osize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
16788 
16789 		if (osize == 0) {
16790 			ASSERT(dtrace_toxrange == NULL);
16791 			ASSERT(dtrace_toxranges_max == 0);
16792 			dtrace_toxranges_max = 1;
16793 		} else {
16794 			dtrace_toxranges_max <<= 1;
16795 		}
16796 
16797 		nsize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
16798 		range = kmem_zalloc(nsize, KM_SLEEP);
16799 
16800 		if (dtrace_toxrange != NULL) {
16801 			ASSERT(osize != 0);
16802 			bcopy(dtrace_toxrange, range, osize);
16803 			kmem_free(dtrace_toxrange, osize);
16804 		}
16805 
16806 		dtrace_toxrange = range;
16807 	}
16808 
16809 	ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_base == 0);
16810 	ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_limit == 0);
16811 
16812 	dtrace_toxrange[dtrace_toxranges].dtt_base = base;
16813 	dtrace_toxrange[dtrace_toxranges].dtt_limit = limit;
16814 	dtrace_toxranges++;
16815 }
16816 
16817 static void
16818 dtrace_getf_barrier()
16819 {
16820 #ifdef illumos
16821 	/*
16822 	 * When we have unprivileged (that is, non-DTRACE_CRV_KERNEL) enablings
16823 	 * that contain calls to getf(), this routine will be called on every
16824 	 * closef() before either the underlying vnode is released or the
16825 	 * file_t itself is freed.  By the time we are here, it is essential
16826 	 * that the file_t can no longer be accessed from a call to getf()
16827 	 * in probe context -- that assures that a dtrace_sync() can be used
16828 	 * to clear out any enablings referring to the old structures.
16829 	 */
16830 	if (curthread->t_procp->p_zone->zone_dtrace_getf != 0 ||
16831 	    kcred->cr_zone->zone_dtrace_getf != 0)
16832 		dtrace_sync();
16833 #endif
16834 }
16835 
16836 /*
16837  * DTrace Driver Cookbook Functions
16838  */
16839 #ifdef illumos
16840 /*ARGSUSED*/
16841 static int
16842 dtrace_attach(dev_info_t *devi, ddi_attach_cmd_t cmd)
16843 {
16844 	dtrace_provider_id_t id;
16845 	dtrace_state_t *state = NULL;
16846 	dtrace_enabling_t *enab;
16847 
16848 	mutex_enter(&cpu_lock);
16849 	mutex_enter(&dtrace_provider_lock);
16850 	mutex_enter(&dtrace_lock);
16851 
16852 	if (ddi_soft_state_init(&dtrace_softstate,
16853 	    sizeof (dtrace_state_t), 0) != 0) {
16854 		cmn_err(CE_NOTE, "/dev/dtrace failed to initialize soft state");
16855 		mutex_exit(&cpu_lock);
16856 		mutex_exit(&dtrace_provider_lock);
16857 		mutex_exit(&dtrace_lock);
16858 		return (DDI_FAILURE);
16859 	}
16860 
16861 	if (ddi_create_minor_node(devi, DTRACEMNR_DTRACE, S_IFCHR,
16862 	    DTRACEMNRN_DTRACE, DDI_PSEUDO, NULL) == DDI_FAILURE ||
16863 	    ddi_create_minor_node(devi, DTRACEMNR_HELPER, S_IFCHR,
16864 	    DTRACEMNRN_HELPER, DDI_PSEUDO, NULL) == DDI_FAILURE) {
16865 		cmn_err(CE_NOTE, "/dev/dtrace couldn't create minor nodes");
16866 		ddi_remove_minor_node(devi, NULL);
16867 		ddi_soft_state_fini(&dtrace_softstate);
16868 		mutex_exit(&cpu_lock);
16869 		mutex_exit(&dtrace_provider_lock);
16870 		mutex_exit(&dtrace_lock);
16871 		return (DDI_FAILURE);
16872 	}
16873 
16874 	ddi_report_dev(devi);
16875 	dtrace_devi = devi;
16876 
16877 	dtrace_modload = dtrace_module_loaded;
16878 	dtrace_modunload = dtrace_module_unloaded;
16879 	dtrace_cpu_init = dtrace_cpu_setup_initial;
16880 	dtrace_helpers_cleanup = dtrace_helpers_destroy;
16881 	dtrace_helpers_fork = dtrace_helpers_duplicate;
16882 	dtrace_cpustart_init = dtrace_suspend;
16883 	dtrace_cpustart_fini = dtrace_resume;
16884 	dtrace_debugger_init = dtrace_suspend;
16885 	dtrace_debugger_fini = dtrace_resume;
16886 
16887 	register_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
16888 
16889 	ASSERT(MUTEX_HELD(&cpu_lock));
16890 
16891 	dtrace_arena = vmem_create("dtrace", (void *)1, UINT32_MAX, 1,
16892 	    NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
16893 	dtrace_minor = vmem_create("dtrace_minor", (void *)DTRACEMNRN_CLONE,
16894 	    UINT32_MAX - DTRACEMNRN_CLONE, 1, NULL, NULL, NULL, 0,
16895 	    VM_SLEEP | VMC_IDENTIFIER);
16896 	dtrace_taskq = taskq_create("dtrace_taskq", 1, maxclsyspri,
16897 	    1, INT_MAX, 0);
16898 
16899 	dtrace_state_cache = kmem_cache_create("dtrace_state_cache",
16900 	    sizeof (dtrace_dstate_percpu_t) * NCPU, DTRACE_STATE_ALIGN,
16901 	    NULL, NULL, NULL, NULL, NULL, 0);
16902 
16903 	ASSERT(MUTEX_HELD(&cpu_lock));
16904 	dtrace_bymod = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_mod),
16905 	    offsetof(dtrace_probe_t, dtpr_nextmod),
16906 	    offsetof(dtrace_probe_t, dtpr_prevmod));
16907 
16908 	dtrace_byfunc = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_func),
16909 	    offsetof(dtrace_probe_t, dtpr_nextfunc),
16910 	    offsetof(dtrace_probe_t, dtpr_prevfunc));
16911 
16912 	dtrace_byname = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_name),
16913 	    offsetof(dtrace_probe_t, dtpr_nextname),
16914 	    offsetof(dtrace_probe_t, dtpr_prevname));
16915 
16916 	if (dtrace_retain_max < 1) {
16917 		cmn_err(CE_WARN, "illegal value (%lu) for dtrace_retain_max; "
16918 		    "setting to 1", dtrace_retain_max);
16919 		dtrace_retain_max = 1;
16920 	}
16921 
16922 	/*
16923 	 * Now discover our toxic ranges.
16924 	 */
16925 	dtrace_toxic_ranges(dtrace_toxrange_add);
16926 
16927 	/*
16928 	 * Before we register ourselves as a provider to our own framework,
16929 	 * we would like to assert that dtrace_provider is NULL -- but that's
16930 	 * not true if we were loaded as a dependency of a DTrace provider.
16931 	 * Once we've registered, we can assert that dtrace_provider is our
16932 	 * pseudo provider.
16933 	 */
16934 	(void) dtrace_register("dtrace", &dtrace_provider_attr,
16935 	    DTRACE_PRIV_NONE, 0, &dtrace_provider_ops, NULL, &id);
16936 
16937 	ASSERT(dtrace_provider != NULL);
16938 	ASSERT((dtrace_provider_id_t)dtrace_provider == id);
16939 
16940 	dtrace_probeid_begin = dtrace_probe_create((dtrace_provider_id_t)
16941 	    dtrace_provider, NULL, NULL, "BEGIN", 0, NULL);
16942 	dtrace_probeid_end = dtrace_probe_create((dtrace_provider_id_t)
16943 	    dtrace_provider, NULL, NULL, "END", 0, NULL);
16944 	dtrace_probeid_error = dtrace_probe_create((dtrace_provider_id_t)
16945 	    dtrace_provider, NULL, NULL, "ERROR", 1, NULL);
16946 
16947 	dtrace_anon_property();
16948 	mutex_exit(&cpu_lock);
16949 
16950 	/*
16951 	 * If there are already providers, we must ask them to provide their
16952 	 * probes, and then match any anonymous enabling against them.  Note
16953 	 * that there should be no other retained enablings at this time:
16954 	 * the only retained enablings at this time should be the anonymous
16955 	 * enabling.
16956 	 */
16957 	if (dtrace_anon.dta_enabling != NULL) {
16958 		ASSERT(dtrace_retained == dtrace_anon.dta_enabling);
16959 
16960 		dtrace_enabling_provide(NULL);
16961 		state = dtrace_anon.dta_state;
16962 
16963 		/*
16964 		 * We couldn't hold cpu_lock across the above call to
16965 		 * dtrace_enabling_provide(), but we must hold it to actually
16966 		 * enable the probes.  We have to drop all of our locks, pick
16967 		 * up cpu_lock, and regain our locks before matching the
16968 		 * retained anonymous enabling.
16969 		 */
16970 		mutex_exit(&dtrace_lock);
16971 		mutex_exit(&dtrace_provider_lock);
16972 
16973 		mutex_enter(&cpu_lock);
16974 		mutex_enter(&dtrace_provider_lock);
16975 		mutex_enter(&dtrace_lock);
16976 
16977 		if ((enab = dtrace_anon.dta_enabling) != NULL)
16978 			(void) dtrace_enabling_match(enab, NULL);
16979 
16980 		mutex_exit(&cpu_lock);
16981 	}
16982 
16983 	mutex_exit(&dtrace_lock);
16984 	mutex_exit(&dtrace_provider_lock);
16985 
16986 	if (state != NULL) {
16987 		/*
16988 		 * If we created any anonymous state, set it going now.
16989 		 */
16990 		(void) dtrace_state_go(state, &dtrace_anon.dta_beganon);
16991 	}
16992 
16993 	return (DDI_SUCCESS);
16994 }
16995 #endif	/* illumos */
16996 
16997 #ifndef illumos
16998 static void dtrace_dtr(void *);
16999 #endif
17000 
17001 /*ARGSUSED*/
17002 static int
17003 #ifdef illumos
17004 dtrace_open(dev_t *devp, int flag, int otyp, cred_t *cred_p)
17005 #else
17006 dtrace_open(struct cdev *dev, int oflags, int devtype, struct thread *td)
17007 #endif
17008 {
17009 	dtrace_state_t *state;
17010 	uint32_t priv;
17011 	uid_t uid;
17012 	zoneid_t zoneid;
17013 
17014 #ifdef illumos
17015 	if (getminor(*devp) == DTRACEMNRN_HELPER)
17016 		return (0);
17017 
17018 	/*
17019 	 * If this wasn't an open with the "helper" minor, then it must be
17020 	 * the "dtrace" minor.
17021 	 */
17022 	if (getminor(*devp) == DTRACEMNRN_DTRACE)
17023 		return (ENXIO);
17024 #else
17025 	cred_t *cred_p = NULL;
17026 	cred_p = dev->si_cred;
17027 
17028 	/*
17029 	 * If no DTRACE_PRIV_* bits are set in the credential, then the
17030 	 * caller lacks sufficient permission to do anything with DTrace.
17031 	 */
17032 	dtrace_cred2priv(cred_p, &priv, &uid, &zoneid);
17033 	if (priv == DTRACE_PRIV_NONE) {
17034 #endif
17035 
17036 		return (EACCES);
17037 	}
17038 
17039 	/*
17040 	 * Ask all providers to provide all their probes.
17041 	 */
17042 	mutex_enter(&dtrace_provider_lock);
17043 	dtrace_probe_provide(NULL, NULL);
17044 	mutex_exit(&dtrace_provider_lock);
17045 
17046 	mutex_enter(&cpu_lock);
17047 	mutex_enter(&dtrace_lock);
17048 	dtrace_opens++;
17049 	dtrace_membar_producer();
17050 
17051 #ifdef illumos
17052 	/*
17053 	 * If the kernel debugger is active (that is, if the kernel debugger
17054 	 * modified text in some way), we won't allow the open.
17055 	 */
17056 	if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
17057 		dtrace_opens--;
17058 		mutex_exit(&cpu_lock);
17059 		mutex_exit(&dtrace_lock);
17060 		return (EBUSY);
17061 	}
17062 
17063 	if (dtrace_helptrace_enable && dtrace_helptrace_buffer == NULL) {
17064 		/*
17065 		 * If DTrace helper tracing is enabled, we need to allocate the
17066 		 * trace buffer and initialize the values.
17067 		 */
17068 		dtrace_helptrace_buffer =
17069 		    kmem_zalloc(dtrace_helptrace_bufsize, KM_SLEEP);
17070 		dtrace_helptrace_next = 0;
17071 		dtrace_helptrace_wrapped = 0;
17072 		dtrace_helptrace_enable = 0;
17073 	}
17074 
17075 	state = dtrace_state_create(devp, cred_p);
17076 #else
17077 	state = dtrace_state_create(dev, NULL);
17078 	devfs_set_cdevpriv(state, dtrace_dtr);
17079 #endif
17080 
17081 	mutex_exit(&cpu_lock);
17082 
17083 	if (state == NULL) {
17084 #ifdef illumos
17085 		if (--dtrace_opens == 0 && dtrace_anon.dta_enabling == NULL)
17086 			(void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
17087 #else
17088 		--dtrace_opens;
17089 #endif
17090 		mutex_exit(&dtrace_lock);
17091 		return (EAGAIN);
17092 	}
17093 
17094 	mutex_exit(&dtrace_lock);
17095 
17096 	return (0);
17097 }
17098 
17099 /*ARGSUSED*/
17100 #ifdef illumos
17101 static int
17102 dtrace_close(dev_t dev, int flag, int otyp, cred_t *cred_p)
17103 #else
17104 static void
17105 dtrace_dtr(void *data)
17106 #endif
17107 {
17108 #ifdef illumos
17109 	minor_t minor = getminor(dev);
17110 	dtrace_state_t *state;
17111 #endif
17112 	dtrace_helptrace_t *buf = NULL;
17113 
17114 #ifdef illumos
17115 	if (minor == DTRACEMNRN_HELPER)
17116 		return (0);
17117 
17118 	state = ddi_get_soft_state(dtrace_softstate, minor);
17119 #else
17120 	dtrace_state_t *state = data;
17121 #endif
17122 
17123 	mutex_enter(&cpu_lock);
17124 	mutex_enter(&dtrace_lock);
17125 
17126 #ifdef illumos
17127 	if (state->dts_anon)
17128 #else
17129 	if (state != NULL && state->dts_anon)
17130 #endif
17131 	{
17132 		/*
17133 		 * There is anonymous state. Destroy that first.
17134 		 */
17135 		ASSERT(dtrace_anon.dta_state == NULL);
17136 		dtrace_state_destroy(state->dts_anon);
17137 	}
17138 
17139 	if (dtrace_helptrace_disable) {
17140 		/*
17141 		 * If we have been told to disable helper tracing, set the
17142 		 * buffer to NULL before calling into dtrace_state_destroy();
17143 		 * we take advantage of its dtrace_sync() to know that no
17144 		 * CPU is in probe context with enabled helper tracing
17145 		 * after it returns.
17146 		 */
17147 		buf = dtrace_helptrace_buffer;
17148 		dtrace_helptrace_buffer = NULL;
17149 	}
17150 
17151 #ifdef illumos
17152 	dtrace_state_destroy(state);
17153 #else
17154 	if (state != NULL) {
17155 		dtrace_state_destroy(state);
17156 		kmem_free(state, 0);
17157 	}
17158 #endif
17159 	ASSERT(dtrace_opens > 0);
17160 
17161 #ifdef illumos
17162 	/*
17163 	 * Only relinquish control of the kernel debugger interface when there
17164 	 * are no consumers and no anonymous enablings.
17165 	 */
17166 	if (--dtrace_opens == 0 && dtrace_anon.dta_enabling == NULL)
17167 		(void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
17168 #else
17169 	--dtrace_opens;
17170 #endif
17171 
17172 	if (buf != NULL) {
17173 		kmem_free(buf, dtrace_helptrace_bufsize);
17174 		dtrace_helptrace_disable = 0;
17175 	}
17176 
17177 	mutex_exit(&dtrace_lock);
17178 	mutex_exit(&cpu_lock);
17179 
17180 #ifdef illumos
17181 	return (0);
17182 #endif
17183 }
17184 
17185 #ifdef illumos
17186 /*ARGSUSED*/
17187 static int
17188 dtrace_ioctl_helper(int cmd, intptr_t arg, int *rv)
17189 {
17190 	int rval;
17191 	dof_helper_t help, *dhp = NULL;
17192 
17193 	switch (cmd) {
17194 	case DTRACEHIOC_ADDDOF:
17195 		if (copyin((void *)arg, &help, sizeof (help)) != 0) {
17196 			dtrace_dof_error(NULL, "failed to copyin DOF helper");
17197 			return (EFAULT);
17198 		}
17199 
17200 		dhp = &help;
17201 		arg = (intptr_t)help.dofhp_dof;
17202 		/*FALLTHROUGH*/
17203 
17204 	case DTRACEHIOC_ADD: {
17205 		dof_hdr_t *dof = dtrace_dof_copyin(arg, &rval);
17206 
17207 		if (dof == NULL)
17208 			return (rval);
17209 
17210 		mutex_enter(&dtrace_lock);
17211 
17212 		/*
17213 		 * dtrace_helper_slurp() takes responsibility for the dof --
17214 		 * it may free it now or it may save it and free it later.
17215 		 */
17216 		if ((rval = dtrace_helper_slurp(dof, dhp)) != -1) {
17217 			*rv = rval;
17218 			rval = 0;
17219 		} else {
17220 			rval = EINVAL;
17221 		}
17222 
17223 		mutex_exit(&dtrace_lock);
17224 		return (rval);
17225 	}
17226 
17227 	case DTRACEHIOC_REMOVE: {
17228 		mutex_enter(&dtrace_lock);
17229 		rval = dtrace_helper_destroygen(NULL, arg);
17230 		mutex_exit(&dtrace_lock);
17231 
17232 		return (rval);
17233 	}
17234 
17235 	default:
17236 		break;
17237 	}
17238 
17239 	return (ENOTTY);
17240 }
17241 
17242 /*ARGSUSED*/
17243 static int
17244 dtrace_ioctl(dev_t dev, int cmd, intptr_t arg, int md, cred_t *cr, int *rv)
17245 {
17246 	minor_t minor = getminor(dev);
17247 	dtrace_state_t *state;
17248 	int rval;
17249 
17250 	if (minor == DTRACEMNRN_HELPER)
17251 		return (dtrace_ioctl_helper(cmd, arg, rv));
17252 
17253 	state = ddi_get_soft_state(dtrace_softstate, minor);
17254 
17255 	if (state->dts_anon) {
17256 		ASSERT(dtrace_anon.dta_state == NULL);
17257 		state = state->dts_anon;
17258 	}
17259 
17260 	switch (cmd) {
17261 	case DTRACEIOC_PROVIDER: {
17262 		dtrace_providerdesc_t pvd;
17263 		dtrace_provider_t *pvp;
17264 
17265 		if (copyin((void *)arg, &pvd, sizeof (pvd)) != 0)
17266 			return (EFAULT);
17267 
17268 		pvd.dtvd_name[DTRACE_PROVNAMELEN - 1] = '\0';
17269 		mutex_enter(&dtrace_provider_lock);
17270 
17271 		for (pvp = dtrace_provider; pvp != NULL; pvp = pvp->dtpv_next) {
17272 			if (strcmp(pvp->dtpv_name, pvd.dtvd_name) == 0)
17273 				break;
17274 		}
17275 
17276 		mutex_exit(&dtrace_provider_lock);
17277 
17278 		if (pvp == NULL)
17279 			return (ESRCH);
17280 
17281 		bcopy(&pvp->dtpv_priv, &pvd.dtvd_priv, sizeof (dtrace_ppriv_t));
17282 		bcopy(&pvp->dtpv_attr, &pvd.dtvd_attr, sizeof (dtrace_pattr_t));
17283 
17284 		if (copyout(&pvd, (void *)arg, sizeof (pvd)) != 0)
17285 			return (EFAULT);
17286 
17287 		return (0);
17288 	}
17289 
17290 	case DTRACEIOC_EPROBE: {
17291 		dtrace_eprobedesc_t epdesc;
17292 		dtrace_ecb_t *ecb;
17293 		dtrace_action_t *act;
17294 		void *buf;
17295 		size_t size;
17296 		uintptr_t dest;
17297 		int nrecs;
17298 
17299 		if (copyin((void *)arg, &epdesc, sizeof (epdesc)) != 0)
17300 			return (EFAULT);
17301 
17302 		mutex_enter(&dtrace_lock);
17303 
17304 		if ((ecb = dtrace_epid2ecb(state, epdesc.dtepd_epid)) == NULL) {
17305 			mutex_exit(&dtrace_lock);
17306 			return (EINVAL);
17307 		}
17308 
17309 		if (ecb->dte_probe == NULL) {
17310 			mutex_exit(&dtrace_lock);
17311 			return (EINVAL);
17312 		}
17313 
17314 		epdesc.dtepd_probeid = ecb->dte_probe->dtpr_id;
17315 		epdesc.dtepd_uarg = ecb->dte_uarg;
17316 		epdesc.dtepd_size = ecb->dte_size;
17317 
17318 		nrecs = epdesc.dtepd_nrecs;
17319 		epdesc.dtepd_nrecs = 0;
17320 		for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
17321 			if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
17322 				continue;
17323 
17324 			epdesc.dtepd_nrecs++;
17325 		}
17326 
17327 		/*
17328 		 * Now that we have the size, we need to allocate a temporary
17329 		 * buffer in which to store the complete description.  We need
17330 		 * the temporary buffer to be able to drop dtrace_lock()
17331 		 * across the copyout(), below.
17332 		 */
17333 		size = sizeof (dtrace_eprobedesc_t) +
17334 		    (epdesc.dtepd_nrecs * sizeof (dtrace_recdesc_t));
17335 
17336 		buf = kmem_alloc(size, KM_SLEEP);
17337 		dest = (uintptr_t)buf;
17338 
17339 		bcopy(&epdesc, (void *)dest, sizeof (epdesc));
17340 		dest += offsetof(dtrace_eprobedesc_t, dtepd_rec[0]);
17341 
17342 		for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
17343 			if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
17344 				continue;
17345 
17346 			if (nrecs-- == 0)
17347 				break;
17348 
17349 			bcopy(&act->dta_rec, (void *)dest,
17350 			    sizeof (dtrace_recdesc_t));
17351 			dest += sizeof (dtrace_recdesc_t);
17352 		}
17353 
17354 		mutex_exit(&dtrace_lock);
17355 
17356 		if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
17357 			kmem_free(buf, size);
17358 			return (EFAULT);
17359 		}
17360 
17361 		kmem_free(buf, size);
17362 		return (0);
17363 	}
17364 
17365 	case DTRACEIOC_AGGDESC: {
17366 		dtrace_aggdesc_t aggdesc;
17367 		dtrace_action_t *act;
17368 		dtrace_aggregation_t *agg;
17369 		int nrecs;
17370 		uint32_t offs;
17371 		dtrace_recdesc_t *lrec;
17372 		void *buf;
17373 		size_t size;
17374 		uintptr_t dest;
17375 
17376 		if (copyin((void *)arg, &aggdesc, sizeof (aggdesc)) != 0)
17377 			return (EFAULT);
17378 
17379 		mutex_enter(&dtrace_lock);
17380 
17381 		if ((agg = dtrace_aggid2agg(state, aggdesc.dtagd_id)) == NULL) {
17382 			mutex_exit(&dtrace_lock);
17383 			return (EINVAL);
17384 		}
17385 
17386 		aggdesc.dtagd_epid = agg->dtag_ecb->dte_epid;
17387 
17388 		nrecs = aggdesc.dtagd_nrecs;
17389 		aggdesc.dtagd_nrecs = 0;
17390 
17391 		offs = agg->dtag_base;
17392 		lrec = &agg->dtag_action.dta_rec;
17393 		aggdesc.dtagd_size = lrec->dtrd_offset + lrec->dtrd_size - offs;
17394 
17395 		for (act = agg->dtag_first; ; act = act->dta_next) {
17396 			ASSERT(act->dta_intuple ||
17397 			    DTRACEACT_ISAGG(act->dta_kind));
17398 
17399 			/*
17400 			 * If this action has a record size of zero, it
17401 			 * denotes an argument to the aggregating action.
17402 			 * Because the presence of this record doesn't (or
17403 			 * shouldn't) affect the way the data is interpreted,
17404 			 * we don't copy it out to save user-level the
17405 			 * confusion of dealing with a zero-length record.
17406 			 */
17407 			if (act->dta_rec.dtrd_size == 0) {
17408 				ASSERT(agg->dtag_hasarg);
17409 				continue;
17410 			}
17411 
17412 			aggdesc.dtagd_nrecs++;
17413 
17414 			if (act == &agg->dtag_action)
17415 				break;
17416 		}
17417 
17418 		/*
17419 		 * Now that we have the size, we need to allocate a temporary
17420 		 * buffer in which to store the complete description.  We need
17421 		 * the temporary buffer to be able to drop dtrace_lock()
17422 		 * across the copyout(), below.
17423 		 */
17424 		size = sizeof (dtrace_aggdesc_t) +
17425 		    (aggdesc.dtagd_nrecs * sizeof (dtrace_recdesc_t));
17426 
17427 		buf = kmem_alloc(size, KM_SLEEP);
17428 		dest = (uintptr_t)buf;
17429 
17430 		bcopy(&aggdesc, (void *)dest, sizeof (aggdesc));
17431 		dest += offsetof(dtrace_aggdesc_t, dtagd_rec[0]);
17432 
17433 		for (act = agg->dtag_first; ; act = act->dta_next) {
17434 			dtrace_recdesc_t rec = act->dta_rec;
17435 
17436 			/*
17437 			 * See the comment in the above loop for why we pass
17438 			 * over zero-length records.
17439 			 */
17440 			if (rec.dtrd_size == 0) {
17441 				ASSERT(agg->dtag_hasarg);
17442 				continue;
17443 			}
17444 
17445 			if (nrecs-- == 0)
17446 				break;
17447 
17448 			rec.dtrd_offset -= offs;
17449 			bcopy(&rec, (void *)dest, sizeof (rec));
17450 			dest += sizeof (dtrace_recdesc_t);
17451 
17452 			if (act == &agg->dtag_action)
17453 				break;
17454 		}
17455 
17456 		mutex_exit(&dtrace_lock);
17457 
17458 		if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
17459 			kmem_free(buf, size);
17460 			return (EFAULT);
17461 		}
17462 
17463 		kmem_free(buf, size);
17464 		return (0);
17465 	}
17466 
17467 	case DTRACEIOC_ENABLE: {
17468 		dof_hdr_t *dof;
17469 		dtrace_enabling_t *enab = NULL;
17470 		dtrace_vstate_t *vstate;
17471 		int err = 0;
17472 
17473 		*rv = 0;
17474 
17475 		/*
17476 		 * If a NULL argument has been passed, we take this as our
17477 		 * cue to reevaluate our enablings.
17478 		 */
17479 		if (arg == NULL) {
17480 			dtrace_enabling_matchall();
17481 
17482 			return (0);
17483 		}
17484 
17485 		if ((dof = dtrace_dof_copyin(arg, &rval)) == NULL)
17486 			return (rval);
17487 
17488 		mutex_enter(&cpu_lock);
17489 		mutex_enter(&dtrace_lock);
17490 		vstate = &state->dts_vstate;
17491 
17492 		if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
17493 			mutex_exit(&dtrace_lock);
17494 			mutex_exit(&cpu_lock);
17495 			dtrace_dof_destroy(dof);
17496 			return (EBUSY);
17497 		}
17498 
17499 		if (dtrace_dof_slurp(dof, vstate, cr, &enab, 0, B_TRUE) != 0) {
17500 			mutex_exit(&dtrace_lock);
17501 			mutex_exit(&cpu_lock);
17502 			dtrace_dof_destroy(dof);
17503 			return (EINVAL);
17504 		}
17505 
17506 		if ((rval = dtrace_dof_options(dof, state)) != 0) {
17507 			dtrace_enabling_destroy(enab);
17508 			mutex_exit(&dtrace_lock);
17509 			mutex_exit(&cpu_lock);
17510 			dtrace_dof_destroy(dof);
17511 			return (rval);
17512 		}
17513 
17514 		if ((err = dtrace_enabling_match(enab, rv)) == 0) {
17515 			err = dtrace_enabling_retain(enab);
17516 		} else {
17517 			dtrace_enabling_destroy(enab);
17518 		}
17519 
17520 		mutex_exit(&cpu_lock);
17521 		mutex_exit(&dtrace_lock);
17522 		dtrace_dof_destroy(dof);
17523 
17524 		return (err);
17525 	}
17526 
17527 	case DTRACEIOC_REPLICATE: {
17528 		dtrace_repldesc_t desc;
17529 		dtrace_probedesc_t *match = &desc.dtrpd_match;
17530 		dtrace_probedesc_t *create = &desc.dtrpd_create;
17531 		int err;
17532 
17533 		if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
17534 			return (EFAULT);
17535 
17536 		match->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
17537 		match->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
17538 		match->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
17539 		match->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
17540 
17541 		create->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
17542 		create->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
17543 		create->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
17544 		create->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
17545 
17546 		mutex_enter(&dtrace_lock);
17547 		err = dtrace_enabling_replicate(state, match, create);
17548 		mutex_exit(&dtrace_lock);
17549 
17550 		return (err);
17551 	}
17552 
17553 	case DTRACEIOC_PROBEMATCH:
17554 	case DTRACEIOC_PROBES: {
17555 		dtrace_probe_t *probe = NULL;
17556 		dtrace_probedesc_t desc;
17557 		dtrace_probekey_t pkey;
17558 		dtrace_id_t i;
17559 		int m = 0;
17560 		uint32_t priv;
17561 		uid_t uid;
17562 		zoneid_t zoneid;
17563 
17564 		if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
17565 			return (EFAULT);
17566 
17567 		desc.dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
17568 		desc.dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
17569 		desc.dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
17570 		desc.dtpd_name[DTRACE_NAMELEN - 1] = '\0';
17571 
17572 		/*
17573 		 * Before we attempt to match this probe, we want to give
17574 		 * all providers the opportunity to provide it.
17575 		 */
17576 		if (desc.dtpd_id == DTRACE_IDNONE) {
17577 			mutex_enter(&dtrace_provider_lock);
17578 			dtrace_probe_provide(&desc, NULL);
17579 			mutex_exit(&dtrace_provider_lock);
17580 			desc.dtpd_id++;
17581 		}
17582 
17583 		if (cmd == DTRACEIOC_PROBEMATCH)  {
17584 			dtrace_probekey(&desc, &pkey);
17585 			pkey.dtpk_id = DTRACE_IDNONE;
17586 		}
17587 
17588 		dtrace_cred2priv(cr, &priv, &uid, &zoneid);
17589 
17590 		mutex_enter(&dtrace_lock);
17591 
17592 		if (cmd == DTRACEIOC_PROBEMATCH) {
17593 			for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
17594 				if ((probe = dtrace_probes[i - 1]) != NULL &&
17595 				    (m = dtrace_match_probe(probe, &pkey,
17596 				    priv, uid, zoneid)) != 0)
17597 					break;
17598 			}
17599 
17600 			if (m < 0) {
17601 				mutex_exit(&dtrace_lock);
17602 				return (EINVAL);
17603 			}
17604 
17605 		} else {
17606 			for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
17607 				if ((probe = dtrace_probes[i - 1]) != NULL &&
17608 				    dtrace_match_priv(probe, priv, uid, zoneid))
17609 					break;
17610 			}
17611 		}
17612 
17613 		if (probe == NULL) {
17614 			mutex_exit(&dtrace_lock);
17615 			return (ESRCH);
17616 		}
17617 
17618 		dtrace_probe_description(probe, &desc);
17619 		mutex_exit(&dtrace_lock);
17620 
17621 		if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
17622 			return (EFAULT);
17623 
17624 		return (0);
17625 	}
17626 
17627 	case DTRACEIOC_PROBEARG: {
17628 		dtrace_argdesc_t desc;
17629 		dtrace_probe_t *probe;
17630 		dtrace_provider_t *prov;
17631 
17632 		if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
17633 			return (EFAULT);
17634 
17635 		if (desc.dtargd_id == DTRACE_IDNONE)
17636 			return (EINVAL);
17637 
17638 		if (desc.dtargd_ndx == DTRACE_ARGNONE)
17639 			return (EINVAL);
17640 
17641 		mutex_enter(&dtrace_provider_lock);
17642 		mutex_enter(&mod_lock);
17643 		mutex_enter(&dtrace_lock);
17644 
17645 		if (desc.dtargd_id > dtrace_nprobes) {
17646 			mutex_exit(&dtrace_lock);
17647 			mutex_exit(&mod_lock);
17648 			mutex_exit(&dtrace_provider_lock);
17649 			return (EINVAL);
17650 		}
17651 
17652 		if ((probe = dtrace_probes[desc.dtargd_id - 1]) == NULL) {
17653 			mutex_exit(&dtrace_lock);
17654 			mutex_exit(&mod_lock);
17655 			mutex_exit(&dtrace_provider_lock);
17656 			return (EINVAL);
17657 		}
17658 
17659 		mutex_exit(&dtrace_lock);
17660 
17661 		prov = probe->dtpr_provider;
17662 
17663 		if (prov->dtpv_pops.dtps_getargdesc == NULL) {
17664 			/*
17665 			 * There isn't any typed information for this probe.
17666 			 * Set the argument number to DTRACE_ARGNONE.
17667 			 */
17668 			desc.dtargd_ndx = DTRACE_ARGNONE;
17669 		} else {
17670 			desc.dtargd_native[0] = '\0';
17671 			desc.dtargd_xlate[0] = '\0';
17672 			desc.dtargd_mapping = desc.dtargd_ndx;
17673 
17674 			prov->dtpv_pops.dtps_getargdesc(prov->dtpv_arg,
17675 			    probe->dtpr_id, probe->dtpr_arg, &desc);
17676 		}
17677 
17678 		mutex_exit(&mod_lock);
17679 		mutex_exit(&dtrace_provider_lock);
17680 
17681 		if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
17682 			return (EFAULT);
17683 
17684 		return (0);
17685 	}
17686 
17687 	case DTRACEIOC_GO: {
17688 		processorid_t cpuid;
17689 		rval = dtrace_state_go(state, &cpuid);
17690 
17691 		if (rval != 0)
17692 			return (rval);
17693 
17694 		if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
17695 			return (EFAULT);
17696 
17697 		return (0);
17698 	}
17699 
17700 	case DTRACEIOC_STOP: {
17701 		processorid_t cpuid;
17702 
17703 		mutex_enter(&dtrace_lock);
17704 		rval = dtrace_state_stop(state, &cpuid);
17705 		mutex_exit(&dtrace_lock);
17706 
17707 		if (rval != 0)
17708 			return (rval);
17709 
17710 		if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
17711 			return (EFAULT);
17712 
17713 		return (0);
17714 	}
17715 
17716 	case DTRACEIOC_DOFGET: {
17717 		dof_hdr_t hdr, *dof;
17718 		uint64_t len;
17719 
17720 		if (copyin((void *)arg, &hdr, sizeof (hdr)) != 0)
17721 			return (EFAULT);
17722 
17723 		mutex_enter(&dtrace_lock);
17724 		dof = dtrace_dof_create(state);
17725 		mutex_exit(&dtrace_lock);
17726 
17727 		len = MIN(hdr.dofh_loadsz, dof->dofh_loadsz);
17728 		rval = copyout(dof, (void *)arg, len);
17729 		dtrace_dof_destroy(dof);
17730 
17731 		return (rval == 0 ? 0 : EFAULT);
17732 	}
17733 
17734 	case DTRACEIOC_AGGSNAP:
17735 	case DTRACEIOC_BUFSNAP: {
17736 		dtrace_bufdesc_t desc;
17737 		caddr_t cached;
17738 		dtrace_buffer_t *buf;
17739 
17740 		if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
17741 			return (EFAULT);
17742 
17743 		if (desc.dtbd_cpu < 0 || desc.dtbd_cpu >= NCPU)
17744 			return (EINVAL);
17745 
17746 		mutex_enter(&dtrace_lock);
17747 
17748 		if (cmd == DTRACEIOC_BUFSNAP) {
17749 			buf = &state->dts_buffer[desc.dtbd_cpu];
17750 		} else {
17751 			buf = &state->dts_aggbuffer[desc.dtbd_cpu];
17752 		}
17753 
17754 		if (buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL)) {
17755 			size_t sz = buf->dtb_offset;
17756 
17757 			if (state->dts_activity != DTRACE_ACTIVITY_STOPPED) {
17758 				mutex_exit(&dtrace_lock);
17759 				return (EBUSY);
17760 			}
17761 
17762 			/*
17763 			 * If this buffer has already been consumed, we're
17764 			 * going to indicate that there's nothing left here
17765 			 * to consume.
17766 			 */
17767 			if (buf->dtb_flags & DTRACEBUF_CONSUMED) {
17768 				mutex_exit(&dtrace_lock);
17769 
17770 				desc.dtbd_size = 0;
17771 				desc.dtbd_drops = 0;
17772 				desc.dtbd_errors = 0;
17773 				desc.dtbd_oldest = 0;
17774 				sz = sizeof (desc);
17775 
17776 				if (copyout(&desc, (void *)arg, sz) != 0)
17777 					return (EFAULT);
17778 
17779 				return (0);
17780 			}
17781 
17782 			/*
17783 			 * If this is a ring buffer that has wrapped, we want
17784 			 * to copy the whole thing out.
17785 			 */
17786 			if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
17787 				dtrace_buffer_polish(buf);
17788 				sz = buf->dtb_size;
17789 			}
17790 
17791 			if (copyout(buf->dtb_tomax, desc.dtbd_data, sz) != 0) {
17792 				mutex_exit(&dtrace_lock);
17793 				return (EFAULT);
17794 			}
17795 
17796 			desc.dtbd_size = sz;
17797 			desc.dtbd_drops = buf->dtb_drops;
17798 			desc.dtbd_errors = buf->dtb_errors;
17799 			desc.dtbd_oldest = buf->dtb_xamot_offset;
17800 			desc.dtbd_timestamp = dtrace_gethrtime();
17801 
17802 			mutex_exit(&dtrace_lock);
17803 
17804 			if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
17805 				return (EFAULT);
17806 
17807 			buf->dtb_flags |= DTRACEBUF_CONSUMED;
17808 
17809 			return (0);
17810 		}
17811 
17812 		if (buf->dtb_tomax == NULL) {
17813 			ASSERT(buf->dtb_xamot == NULL);
17814 			mutex_exit(&dtrace_lock);
17815 			return (ENOENT);
17816 		}
17817 
17818 		cached = buf->dtb_tomax;
17819 		ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
17820 
17821 		dtrace_xcall(desc.dtbd_cpu,
17822 		    (dtrace_xcall_t)dtrace_buffer_switch, buf);
17823 
17824 		state->dts_errors += buf->dtb_xamot_errors;
17825 
17826 		/*
17827 		 * If the buffers did not actually switch, then the cross call
17828 		 * did not take place -- presumably because the given CPU is
17829 		 * not in the ready set.  If this is the case, we'll return
17830 		 * ENOENT.
17831 		 */
17832 		if (buf->dtb_tomax == cached) {
17833 			ASSERT(buf->dtb_xamot != cached);
17834 			mutex_exit(&dtrace_lock);
17835 			return (ENOENT);
17836 		}
17837 
17838 		ASSERT(cached == buf->dtb_xamot);
17839 
17840 		/*
17841 		 * We have our snapshot; now copy it out.
17842 		 */
17843 		if (copyout(buf->dtb_xamot, desc.dtbd_data,
17844 		    buf->dtb_xamot_offset) != 0) {
17845 			mutex_exit(&dtrace_lock);
17846 			return (EFAULT);
17847 		}
17848 
17849 		desc.dtbd_size = buf->dtb_xamot_offset;
17850 		desc.dtbd_drops = buf->dtb_xamot_drops;
17851 		desc.dtbd_errors = buf->dtb_xamot_errors;
17852 		desc.dtbd_oldest = 0;
17853 		desc.dtbd_timestamp = buf->dtb_switched;
17854 
17855 		mutex_exit(&dtrace_lock);
17856 
17857 		/*
17858 		 * Finally, copy out the buffer description.
17859 		 */
17860 		if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
17861 			return (EFAULT);
17862 
17863 		return (0);
17864 	}
17865 
17866 	case DTRACEIOC_CONF: {
17867 		dtrace_conf_t conf;
17868 
17869 		bzero(&conf, sizeof (conf));
17870 		conf.dtc_difversion = DIF_VERSION;
17871 		conf.dtc_difintregs = DIF_DIR_NREGS;
17872 		conf.dtc_diftupregs = DIF_DTR_NREGS;
17873 		conf.dtc_ctfmodel = CTF_MODEL_NATIVE;
17874 
17875 		if (copyout(&conf, (void *)arg, sizeof (conf)) != 0)
17876 			return (EFAULT);
17877 
17878 		return (0);
17879 	}
17880 
17881 	case DTRACEIOC_STATUS: {
17882 		dtrace_status_t stat;
17883 		dtrace_dstate_t *dstate;
17884 		int i, j;
17885 		uint64_t nerrs;
17886 
17887 		/*
17888 		 * See the comment in dtrace_state_deadman() for the reason
17889 		 * for setting dts_laststatus to INT64_MAX before setting
17890 		 * it to the correct value.
17891 		 */
17892 		state->dts_laststatus = INT64_MAX;
17893 		dtrace_membar_producer();
17894 		state->dts_laststatus = dtrace_gethrtime();
17895 
17896 		bzero(&stat, sizeof (stat));
17897 
17898 		mutex_enter(&dtrace_lock);
17899 
17900 		if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE) {
17901 			mutex_exit(&dtrace_lock);
17902 			return (ENOENT);
17903 		}
17904 
17905 		if (state->dts_activity == DTRACE_ACTIVITY_DRAINING)
17906 			stat.dtst_exiting = 1;
17907 
17908 		nerrs = state->dts_errors;
17909 		dstate = &state->dts_vstate.dtvs_dynvars;
17910 
17911 		for (i = 0; i < NCPU; i++) {
17912 			dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[i];
17913 
17914 			stat.dtst_dyndrops += dcpu->dtdsc_drops;
17915 			stat.dtst_dyndrops_dirty += dcpu->dtdsc_dirty_drops;
17916 			stat.dtst_dyndrops_rinsing += dcpu->dtdsc_rinsing_drops;
17917 
17918 			if (state->dts_buffer[i].dtb_flags & DTRACEBUF_FULL)
17919 				stat.dtst_filled++;
17920 
17921 			nerrs += state->dts_buffer[i].dtb_errors;
17922 
17923 			for (j = 0; j < state->dts_nspeculations; j++) {
17924 				dtrace_speculation_t *spec;
17925 				dtrace_buffer_t *buf;
17926 
17927 				spec = &state->dts_speculations[j];
17928 				buf = &spec->dtsp_buffer[i];
17929 				stat.dtst_specdrops += buf->dtb_xamot_drops;
17930 			}
17931 		}
17932 
17933 		stat.dtst_specdrops_busy = state->dts_speculations_busy;
17934 		stat.dtst_specdrops_unavail = state->dts_speculations_unavail;
17935 		stat.dtst_stkstroverflows = state->dts_stkstroverflows;
17936 		stat.dtst_dblerrors = state->dts_dblerrors;
17937 		stat.dtst_killed =
17938 		    (state->dts_activity == DTRACE_ACTIVITY_KILLED);
17939 		stat.dtst_errors = nerrs;
17940 
17941 		mutex_exit(&dtrace_lock);
17942 
17943 		if (copyout(&stat, (void *)arg, sizeof (stat)) != 0)
17944 			return (EFAULT);
17945 
17946 		return (0);
17947 	}
17948 
17949 	case DTRACEIOC_FORMAT: {
17950 		dtrace_fmtdesc_t fmt;
17951 		char *str;
17952 		int len;
17953 
17954 		if (copyin((void *)arg, &fmt, sizeof (fmt)) != 0)
17955 			return (EFAULT);
17956 
17957 		mutex_enter(&dtrace_lock);
17958 
17959 		if (fmt.dtfd_format == 0 ||
17960 		    fmt.dtfd_format > state->dts_nformats) {
17961 			mutex_exit(&dtrace_lock);
17962 			return (EINVAL);
17963 		}
17964 
17965 		/*
17966 		 * Format strings are allocated contiguously and they are
17967 		 * never freed; if a format index is less than the number
17968 		 * of formats, we can assert that the format map is non-NULL
17969 		 * and that the format for the specified index is non-NULL.
17970 		 */
17971 		ASSERT(state->dts_formats != NULL);
17972 		str = state->dts_formats[fmt.dtfd_format - 1];
17973 		ASSERT(str != NULL);
17974 
17975 		len = strlen(str) + 1;
17976 
17977 		if (len > fmt.dtfd_length) {
17978 			fmt.dtfd_length = len;
17979 
17980 			if (copyout(&fmt, (void *)arg, sizeof (fmt)) != 0) {
17981 				mutex_exit(&dtrace_lock);
17982 				return (EINVAL);
17983 			}
17984 		} else {
17985 			if (copyout(str, fmt.dtfd_string, len) != 0) {
17986 				mutex_exit(&dtrace_lock);
17987 				return (EINVAL);
17988 			}
17989 		}
17990 
17991 		mutex_exit(&dtrace_lock);
17992 		return (0);
17993 	}
17994 
17995 	default:
17996 		break;
17997 	}
17998 
17999 	return (ENOTTY);
18000 }
18001 
18002 /*ARGSUSED*/
18003 static int
18004 dtrace_detach(dev_info_t *dip, ddi_detach_cmd_t cmd)
18005 {
18006 	dtrace_state_t *state;
18007 
18008 	switch (cmd) {
18009 	case DDI_DETACH:
18010 		break;
18011 
18012 	case DDI_SUSPEND:
18013 		return (DDI_SUCCESS);
18014 
18015 	default:
18016 		return (DDI_FAILURE);
18017 	}
18018 
18019 	mutex_enter(&cpu_lock);
18020 	mutex_enter(&dtrace_provider_lock);
18021 	mutex_enter(&dtrace_lock);
18022 
18023 	ASSERT(dtrace_opens == 0);
18024 
18025 	if (dtrace_helpers > 0) {
18026 		mutex_exit(&dtrace_provider_lock);
18027 		mutex_exit(&dtrace_lock);
18028 		mutex_exit(&cpu_lock);
18029 		return (DDI_FAILURE);
18030 	}
18031 
18032 	if (dtrace_unregister((dtrace_provider_id_t)dtrace_provider) != 0) {
18033 		mutex_exit(&dtrace_provider_lock);
18034 		mutex_exit(&dtrace_lock);
18035 		mutex_exit(&cpu_lock);
18036 		return (DDI_FAILURE);
18037 	}
18038 
18039 	dtrace_provider = NULL;
18040 
18041 	if ((state = dtrace_anon_grab()) != NULL) {
18042 		/*
18043 		 * If there were ECBs on this state, the provider should
18044 		 * have not been allowed to detach; assert that there is
18045 		 * none.
18046 		 */
18047 		ASSERT(state->dts_necbs == 0);
18048 		dtrace_state_destroy(state);
18049 
18050 		/*
18051 		 * If we're being detached with anonymous state, we need to
18052 		 * indicate to the kernel debugger that DTrace is now inactive.
18053 		 */
18054 		(void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
18055 	}
18056 
18057 	bzero(&dtrace_anon, sizeof (dtrace_anon_t));
18058 	unregister_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
18059 	dtrace_cpu_init = NULL;
18060 	dtrace_helpers_cleanup = NULL;
18061 	dtrace_helpers_fork = NULL;
18062 	dtrace_cpustart_init = NULL;
18063 	dtrace_cpustart_fini = NULL;
18064 	dtrace_debugger_init = NULL;
18065 	dtrace_debugger_fini = NULL;
18066 	dtrace_modload = NULL;
18067 	dtrace_modunload = NULL;
18068 
18069 	ASSERT(dtrace_getf == 0);
18070 	ASSERT(dtrace_closef == NULL);
18071 
18072 	mutex_exit(&cpu_lock);
18073 
18074 	kmem_free(dtrace_probes, dtrace_nprobes * sizeof (dtrace_probe_t *));
18075 	dtrace_probes = NULL;
18076 	dtrace_nprobes = 0;
18077 
18078 	dtrace_hash_destroy(dtrace_bymod);
18079 	dtrace_hash_destroy(dtrace_byfunc);
18080 	dtrace_hash_destroy(dtrace_byname);
18081 	dtrace_bymod = NULL;
18082 	dtrace_byfunc = NULL;
18083 	dtrace_byname = NULL;
18084 
18085 	kmem_cache_destroy(dtrace_state_cache);
18086 	vmem_destroy(dtrace_minor);
18087 	vmem_destroy(dtrace_arena);
18088 
18089 	if (dtrace_toxrange != NULL) {
18090 		kmem_free(dtrace_toxrange,
18091 		    dtrace_toxranges_max * sizeof (dtrace_toxrange_t));
18092 		dtrace_toxrange = NULL;
18093 		dtrace_toxranges = 0;
18094 		dtrace_toxranges_max = 0;
18095 	}
18096 
18097 	ddi_remove_minor_node(dtrace_devi, NULL);
18098 	dtrace_devi = NULL;
18099 
18100 	ddi_soft_state_fini(&dtrace_softstate);
18101 
18102 	ASSERT(dtrace_vtime_references == 0);
18103 	ASSERT(dtrace_opens == 0);
18104 	ASSERT(dtrace_retained == NULL);
18105 
18106 	mutex_exit(&dtrace_lock);
18107 	mutex_exit(&dtrace_provider_lock);
18108 
18109 	/*
18110 	 * We don't destroy the task queue until after we have dropped our
18111 	 * locks (taskq_destroy() may block on running tasks).  To prevent
18112 	 * attempting to do work after we have effectively detached but before
18113 	 * the task queue has been destroyed, all tasks dispatched via the
18114 	 * task queue must check that DTrace is still attached before
18115 	 * performing any operation.
18116 	 */
18117 	taskq_destroy(dtrace_taskq);
18118 	dtrace_taskq = NULL;
18119 
18120 	return (DDI_SUCCESS);
18121 }
18122 #endif
18123 
18124 #ifdef illumos
18125 /*ARGSUSED*/
18126 static int
18127 dtrace_info(dev_info_t *dip, ddi_info_cmd_t infocmd, void *arg, void **result)
18128 {
18129 	int error;
18130 
18131 	switch (infocmd) {
18132 	case DDI_INFO_DEVT2DEVINFO:
18133 		*result = (void *)dtrace_devi;
18134 		error = DDI_SUCCESS;
18135 		break;
18136 	case DDI_INFO_DEVT2INSTANCE:
18137 		*result = (void *)0;
18138 		error = DDI_SUCCESS;
18139 		break;
18140 	default:
18141 		error = DDI_FAILURE;
18142 	}
18143 	return (error);
18144 }
18145 #endif
18146 
18147 #ifdef illumos
18148 static struct cb_ops dtrace_cb_ops = {
18149 	dtrace_open,		/* open */
18150 	dtrace_close,		/* close */
18151 	nulldev,		/* strategy */
18152 	nulldev,		/* print */
18153 	nodev,			/* dump */
18154 	nodev,			/* read */
18155 	nodev,			/* write */
18156 	dtrace_ioctl,		/* ioctl */
18157 	nodev,			/* devmap */
18158 	nodev,			/* mmap */
18159 	nodev,			/* segmap */
18160 	nochpoll,		/* poll */
18161 	ddi_prop_op,		/* cb_prop_op */
18162 	0,			/* streamtab  */
18163 	D_NEW | D_MP		/* Driver compatibility flag */
18164 };
18165 
18166 static struct dev_ops dtrace_ops = {
18167 	DEVO_REV,		/* devo_rev */
18168 	0,			/* refcnt */
18169 	dtrace_info,		/* get_dev_info */
18170 	nulldev,		/* identify */
18171 	nulldev,		/* probe */
18172 	dtrace_attach,		/* attach */
18173 	dtrace_detach,		/* detach */
18174 	nodev,			/* reset */
18175 	&dtrace_cb_ops,		/* driver operations */
18176 	NULL,			/* bus operations */
18177 	nodev			/* dev power */
18178 };
18179 
18180 static struct modldrv modldrv = {
18181 	&mod_driverops,		/* module type (this is a pseudo driver) */
18182 	"Dynamic Tracing",	/* name of module */
18183 	&dtrace_ops,		/* driver ops */
18184 };
18185 
18186 static struct modlinkage modlinkage = {
18187 	MODREV_1,
18188 	(void *)&modldrv,
18189 	NULL
18190 };
18191 
18192 int
18193 _init(void)
18194 {
18195 	return (mod_install(&modlinkage));
18196 }
18197 
18198 int
18199 _info(struct modinfo *modinfop)
18200 {
18201 	return (mod_info(&modlinkage, modinfop));
18202 }
18203 
18204 int
18205 _fini(void)
18206 {
18207 	return (mod_remove(&modlinkage));
18208 }
18209 #else
18210 
18211 static d_ioctl_t	dtrace_ioctl;
18212 static d_ioctl_t	dtrace_ioctl_helper;
18213 static void		dtrace_load(void *);
18214 static int		dtrace_unload(void);
18215 static struct cdev	*dtrace_dev;
18216 static struct cdev	*helper_dev;
18217 
18218 void dtrace_invop_init(void);
18219 void dtrace_invop_uninit(void);
18220 
18221 static struct cdevsw dtrace_cdevsw = {
18222 	.d_version	= D_VERSION,
18223 	.d_ioctl	= dtrace_ioctl,
18224 	.d_open		= dtrace_open,
18225 	.d_name		= "dtrace",
18226 };
18227 
18228 static struct cdevsw helper_cdevsw = {
18229 	.d_version	= D_VERSION,
18230 	.d_ioctl	= dtrace_ioctl_helper,
18231 	.d_name		= "helper",
18232 };
18233 
18234 #include <dtrace_anon.c>
18235 #include <dtrace_ioctl.c>
18236 #include <dtrace_load.c>
18237 #include <dtrace_modevent.c>
18238 #include <dtrace_sysctl.c>
18239 #include <dtrace_unload.c>
18240 #include <dtrace_vtime.c>
18241 #include <dtrace_hacks.c>
18242 #include <dtrace_isa.c>
18243 
18244 SYSINIT(dtrace_load, SI_SUB_DTRACE, SI_ORDER_FIRST, dtrace_load, NULL);
18245 SYSUNINIT(dtrace_unload, SI_SUB_DTRACE, SI_ORDER_FIRST, dtrace_unload, NULL);
18246 SYSINIT(dtrace_anon_init, SI_SUB_DTRACE_ANON, SI_ORDER_FIRST, dtrace_anon_init, NULL);
18247 
18248 DEV_MODULE(dtrace, dtrace_modevent, NULL);
18249 MODULE_VERSION(dtrace, 1);
18250 MODULE_DEPEND(dtrace, opensolaris, 1, 1, 1);
18251 #endif
18252