xref: /titanic_52/usr/src/uts/common/dtrace/dtrace.c (revision a2cedc52cb05701ca77d7e4b4f4aa95448b52ed5)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright 2006 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 #pragma ident	"%Z%%M%	%I%	%E% SMI"
28 
29 /*
30  * DTrace - Dynamic Tracing for Solaris
31  *
32  * This is the implementation of the Solaris Dynamic Tracing framework
33  * (DTrace).  The user-visible interface to DTrace is described at length in
34  * the "Solaris Dynamic Tracing Guide".  The interfaces between the libdtrace
35  * library, the in-kernel DTrace framework, and the DTrace providers are
36  * described in the block comments in the <sys/dtrace.h> header file.  The
37  * internal architecture of DTrace is described in the block comments in the
38  * <sys/dtrace_impl.h> header file.  The comments contained within the DTrace
39  * implementation very much assume mastery of all of these sources; if one has
40  * an unanswered question about the implementation, one should consult them
41  * first.
42  *
43  * The functions here are ordered roughly as follows:
44  *
45  *   - Probe context functions
46  *   - Probe hashing functions
47  *   - Non-probe context utility functions
48  *   - Matching functions
49  *   - Provider-to-Framework API functions
50  *   - Probe management functions
51  *   - DIF object functions
52  *   - Format functions
53  *   - Predicate functions
54  *   - ECB functions
55  *   - Buffer functions
56  *   - Enabling functions
57  *   - DOF functions
58  *   - Anonymous enabling functions
59  *   - Consumer state functions
60  *   - Helper functions
61  *   - Hook functions
62  *   - Driver cookbook functions
63  *
64  * Each group of functions begins with a block comment labelled the "DTrace
65  * [Group] Functions", allowing one to find each block by searching forward
66  * on capital-f functions.
67  */
68 #include <sys/errno.h>
69 #include <sys/stat.h>
70 #include <sys/modctl.h>
71 #include <sys/conf.h>
72 #include <sys/systm.h>
73 #include <sys/ddi.h>
74 #include <sys/sunddi.h>
75 #include <sys/cpuvar.h>
76 #include <sys/kmem.h>
77 #include <sys/strsubr.h>
78 #include <sys/sysmacros.h>
79 #include <sys/dtrace_impl.h>
80 #include <sys/atomic.h>
81 #include <sys/cmn_err.h>
82 #include <sys/mutex_impl.h>
83 #include <sys/rwlock_impl.h>
84 #include <sys/ctf_api.h>
85 #include <sys/panic.h>
86 #include <sys/priv_impl.h>
87 #include <sys/policy.h>
88 #include <sys/cred_impl.h>
89 #include <sys/procfs_isa.h>
90 #include <sys/taskq.h>
91 #include <sys/mkdev.h>
92 #include <sys/kdi.h>
93 #include <sys/zone.h>
94 
95 /*
96  * DTrace Tunable Variables
97  *
98  * The following variables may be tuned by adding a line to /etc/system that
99  * includes both the name of the DTrace module ("dtrace") and the name of the
100  * variable.  For example:
101  *
102  *   set dtrace:dtrace_destructive_disallow = 1
103  *
104  * In general, the only variables that one should be tuning this way are those
105  * that affect system-wide DTrace behavior, and for which the default behavior
106  * is undesirable.  Most of these variables are tunable on a per-consumer
107  * basis using DTrace options, and need not be tuned on a system-wide basis.
108  * When tuning these variables, avoid pathological values; while some attempt
109  * is made to verify the integrity of these variables, they are not considered
110  * part of the supported interface to DTrace, and they are therefore not
111  * checked comprehensively.  Further, these variables should not be tuned
112  * dynamically via "mdb -kw" or other means; they should only be tuned via
113  * /etc/system.
114  */
115 int		dtrace_destructive_disallow = 0;
116 dtrace_optval_t	dtrace_nonroot_maxsize = (16 * 1024 * 1024);
117 size_t		dtrace_difo_maxsize = (256 * 1024);
118 dtrace_optval_t	dtrace_dof_maxsize = (256 * 1024);
119 size_t		dtrace_global_maxsize = (16 * 1024);
120 size_t		dtrace_actions_max = (16 * 1024);
121 size_t		dtrace_retain_max = 1024;
122 dtrace_optval_t	dtrace_helper_actions_max = 32;
123 dtrace_optval_t	dtrace_helper_providers_max = 32;
124 dtrace_optval_t	dtrace_dstate_defsize = (1 * 1024 * 1024);
125 size_t		dtrace_strsize_default = 256;
126 dtrace_optval_t	dtrace_cleanrate_default = 9900990;		/* 101 hz */
127 dtrace_optval_t	dtrace_cleanrate_min = 200000;			/* 5000 hz */
128 dtrace_optval_t	dtrace_cleanrate_max = (uint64_t)60 * NANOSEC;	/* 1/minute */
129 dtrace_optval_t	dtrace_aggrate_default = NANOSEC;		/* 1 hz */
130 dtrace_optval_t	dtrace_statusrate_default = NANOSEC;		/* 1 hz */
131 dtrace_optval_t dtrace_statusrate_max = (hrtime_t)10 * NANOSEC;	 /* 6/minute */
132 dtrace_optval_t	dtrace_switchrate_default = NANOSEC;		/* 1 hz */
133 dtrace_optval_t	dtrace_nspec_default = 1;
134 dtrace_optval_t	dtrace_specsize_default = 32 * 1024;
135 dtrace_optval_t dtrace_stackframes_default = 20;
136 dtrace_optval_t dtrace_ustackframes_default = 20;
137 dtrace_optval_t dtrace_jstackframes_default = 50;
138 dtrace_optval_t dtrace_jstackstrsize_default = 512;
139 int		dtrace_msgdsize_max = 128;
140 hrtime_t	dtrace_chill_max = 500 * (NANOSEC / MILLISEC);	/* 500 ms */
141 hrtime_t	dtrace_chill_interval = NANOSEC;		/* 1000 ms */
142 int		dtrace_devdepth_max = 32;
143 int		dtrace_err_verbose;
144 hrtime_t	dtrace_deadman_interval = NANOSEC;
145 hrtime_t	dtrace_deadman_timeout = (hrtime_t)10 * NANOSEC;
146 hrtime_t	dtrace_deadman_user = (hrtime_t)30 * NANOSEC;
147 
148 /*
149  * DTrace External Variables
150  *
151  * As dtrace(7D) is a kernel module, any DTrace variables are obviously
152  * available to DTrace consumers via the backtick (`) syntax.  One of these,
153  * dtrace_zero, is made deliberately so:  it is provided as a source of
154  * well-known, zero-filled memory.  While this variable is not documented,
155  * it is used by some translators as an implementation detail.
156  */
157 const char	dtrace_zero[256] = { 0 };	/* zero-filled memory */
158 
159 /*
160  * DTrace Internal Variables
161  */
162 static dev_info_t	*dtrace_devi;		/* device info */
163 static vmem_t		*dtrace_arena;		/* probe ID arena */
164 static vmem_t		*dtrace_minor;		/* minor number arena */
165 static taskq_t		*dtrace_taskq;		/* task queue */
166 static dtrace_probe_t	**dtrace_probes;	/* array of all probes */
167 static int		dtrace_nprobes;		/* number of probes */
168 static dtrace_provider_t *dtrace_provider;	/* provider list */
169 static dtrace_meta_t	*dtrace_meta_pid;	/* user-land meta provider */
170 static int		dtrace_opens;		/* number of opens */
171 static int		dtrace_helpers;		/* number of helpers */
172 static void		*dtrace_softstate;	/* softstate pointer */
173 static dtrace_hash_t	*dtrace_bymod;		/* probes hashed by module */
174 static dtrace_hash_t	*dtrace_byfunc;		/* probes hashed by function */
175 static dtrace_hash_t	*dtrace_byname;		/* probes hashed by name */
176 static dtrace_toxrange_t *dtrace_toxrange;	/* toxic range array */
177 static int		dtrace_toxranges;	/* number of toxic ranges */
178 static int		dtrace_toxranges_max;	/* size of toxic range array */
179 static dtrace_anon_t	dtrace_anon;		/* anonymous enabling */
180 static kmem_cache_t	*dtrace_state_cache;	/* cache for dynamic state */
181 static uint64_t		dtrace_vtime_references; /* number of vtimestamp refs */
182 static kthread_t	*dtrace_panicked;	/* panicking thread */
183 static dtrace_ecb_t	*dtrace_ecb_create_cache; /* cached created ECB */
184 static dtrace_genid_t	dtrace_probegen;	/* current probe generation */
185 static dtrace_helpers_t *dtrace_deferred_pid;	/* deferred helper list */
186 static dtrace_enabling_t *dtrace_retained;	/* list of retained enablings */
187 static dtrace_dynvar_t	dtrace_dynhash_sink;	/* end of dynamic hash chains */
188 
189 /*
190  * DTrace Locking
191  * DTrace is protected by three (relatively coarse-grained) locks:
192  *
193  * (1) dtrace_lock is required to manipulate essentially any DTrace state,
194  *     including enabling state, probes, ECBs, consumer state, helper state,
195  *     etc.  Importantly, dtrace_lock is _not_ required when in probe context;
196  *     probe context is lock-free -- synchronization is handled via the
197  *     dtrace_sync() cross call mechanism.
198  *
199  * (2) dtrace_provider_lock is required when manipulating provider state, or
200  *     when provider state must be held constant.
201  *
202  * (3) dtrace_meta_lock is required when manipulating meta provider state, or
203  *     when meta provider state must be held constant.
204  *
205  * The lock ordering between these three locks is dtrace_meta_lock before
206  * dtrace_provider_lock before dtrace_lock.  (In particular, there are
207  * several places where dtrace_provider_lock is held by the framework as it
208  * calls into the providers -- which then call back into the framework,
209  * grabbing dtrace_lock.)
210  *
211  * There are two other locks in the mix:  mod_lock and cpu_lock.  With respect
212  * to dtrace_provider_lock and dtrace_lock, cpu_lock continues its historical
213  * role as a coarse-grained lock; it is acquired before both of these locks.
214  * With respect to dtrace_meta_lock, its behavior is stranger:  cpu_lock must
215  * be acquired _between_ dtrace_meta_lock and any other DTrace locks.
216  * mod_lock is similar with respect to dtrace_provider_lock in that it must be
217  * acquired _between_ dtrace_provider_lock and dtrace_lock.
218  */
219 static kmutex_t		dtrace_lock;		/* probe state lock */
220 static kmutex_t		dtrace_provider_lock;	/* provider state lock */
221 static kmutex_t		dtrace_meta_lock;	/* meta-provider state lock */
222 
223 /*
224  * DTrace Provider Variables
225  *
226  * These are the variables relating to DTrace as a provider (that is, the
227  * provider of the BEGIN, END, and ERROR probes).
228  */
229 static dtrace_pattr_t	dtrace_provider_attr = {
230 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
231 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
232 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
233 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
234 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
235 };
236 
237 static void
238 dtrace_nullop(void)
239 {}
240 
241 static dtrace_pops_t	dtrace_provider_ops = {
242 	(void (*)(void *, const dtrace_probedesc_t *))dtrace_nullop,
243 	(void (*)(void *, struct modctl *))dtrace_nullop,
244 	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
245 	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
246 	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
247 	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
248 	NULL,
249 	NULL,
250 	NULL,
251 	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop
252 };
253 
254 static dtrace_id_t	dtrace_probeid_begin;	/* special BEGIN probe */
255 static dtrace_id_t	dtrace_probeid_end;	/* special END probe */
256 dtrace_id_t		dtrace_probeid_error;	/* special ERROR probe */
257 
258 /*
259  * DTrace Helper Tracing Variables
260  */
261 uint32_t dtrace_helptrace_next = 0;
262 uint32_t dtrace_helptrace_nlocals;
263 char	*dtrace_helptrace_buffer;
264 int	dtrace_helptrace_bufsize = 512 * 1024;
265 
266 #ifdef DEBUG
267 int	dtrace_helptrace_enabled = 1;
268 #else
269 int	dtrace_helptrace_enabled = 0;
270 #endif
271 
272 /*
273  * DTrace Error Hashing
274  *
275  * On DEBUG kernels, DTrace will track the errors that has seen in a hash
276  * table.  This is very useful for checking coverage of tests that are
277  * expected to induce DIF or DOF processing errors, and may be useful for
278  * debugging problems in the DIF code generator or in DOF generation .  The
279  * error hash may be examined with the ::dtrace_errhash MDB dcmd.
280  */
281 #ifdef DEBUG
282 static dtrace_errhash_t	dtrace_errhash[DTRACE_ERRHASHSZ];
283 static const char *dtrace_errlast;
284 static kthread_t *dtrace_errthread;
285 static kmutex_t dtrace_errlock;
286 #endif
287 
288 /*
289  * DTrace Macros and Constants
290  *
291  * These are various macros that are useful in various spots in the
292  * implementation, along with a few random constants that have no meaning
293  * outside of the implementation.  There is no real structure to this cpp
294  * mishmash -- but is there ever?
295  */
296 #define	DTRACE_HASHSTR(hash, probe)	\
297 	dtrace_hash_str(*((char **)((uintptr_t)(probe) + (hash)->dth_stroffs)))
298 
299 #define	DTRACE_HASHNEXT(hash, probe)	\
300 	(dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_nextoffs)
301 
302 #define	DTRACE_HASHPREV(hash, probe)	\
303 	(dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_prevoffs)
304 
305 #define	DTRACE_HASHEQ(hash, lhs, rhs)	\
306 	(strcmp(*((char **)((uintptr_t)(lhs) + (hash)->dth_stroffs)), \
307 	    *((char **)((uintptr_t)(rhs) + (hash)->dth_stroffs))) == 0)
308 
309 #define	DTRACE_AGGHASHSIZE_SLEW		17
310 
311 /*
312  * The key for a thread-local variable consists of the lower 61 bits of the
313  * t_did, plus the 3 bits of the highest active interrupt above LOCK_LEVEL.
314  * We add DIF_VARIABLE_MAX to t_did to assure that the thread key is never
315  * equal to a variable identifier.  This is necessary (but not sufficient) to
316  * assure that global associative arrays never collide with thread-local
317  * variables.  To guarantee that they cannot collide, we must also define the
318  * order for keying dynamic variables.  That order is:
319  *
320  *   [ key0 ] ... [ keyn ] [ variable-key ] [ tls-key ]
321  *
322  * Because the variable-key and the tls-key are in orthogonal spaces, there is
323  * no way for a global variable key signature to match a thread-local key
324  * signature.
325  */
326 #define	DTRACE_TLS_THRKEY(where) { \
327 	uint_t intr = 0; \
328 	uint_t actv = CPU->cpu_intr_actv >> (LOCK_LEVEL + 1); \
329 	for (; actv; actv >>= 1) \
330 		intr++; \
331 	ASSERT(intr < (1 << 3)); \
332 	(where) = ((curthread->t_did + DIF_VARIABLE_MAX) & \
333 	    (((uint64_t)1 << 61) - 1)) | ((uint64_t)intr << 61); \
334 }
335 
336 #define	DTRACE_STORE(type, tomax, offset, what) \
337 	*((type *)((uintptr_t)(tomax) + (uintptr_t)offset)) = (type)(what);
338 
339 #ifndef __i386
340 #define	DTRACE_ALIGNCHECK(addr, size, flags)				\
341 	if (addr & (size - 1)) {					\
342 		*flags |= CPU_DTRACE_BADALIGN;				\
343 		cpu_core[CPU->cpu_id].cpuc_dtrace_illval = addr;	\
344 		return (0);						\
345 	}
346 #else
347 #define	DTRACE_ALIGNCHECK(addr, size, flags)
348 #endif
349 
350 #define	DTRACE_LOADFUNC(bits)						\
351 /*CSTYLED*/								\
352 uint##bits##_t								\
353 dtrace_load##bits(uintptr_t addr)					\
354 {									\
355 	size_t size = bits / NBBY;					\
356 	/*CSTYLED*/							\
357 	uint##bits##_t rval;						\
358 	int i;								\
359 	volatile uint16_t *flags = (volatile uint16_t *)		\
360 	    &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;			\
361 									\
362 	DTRACE_ALIGNCHECK(addr, size, flags);				\
363 									\
364 	for (i = 0; i < dtrace_toxranges; i++) {			\
365 		if (addr >= dtrace_toxrange[i].dtt_limit)		\
366 			continue;					\
367 									\
368 		if (addr + size <= dtrace_toxrange[i].dtt_base)		\
369 			continue;					\
370 									\
371 		/*							\
372 		 * This address falls within a toxic region; return 0.	\
373 		 */							\
374 		*flags |= CPU_DTRACE_BADADDR;				\
375 		cpu_core[CPU->cpu_id].cpuc_dtrace_illval = addr;	\
376 		return (0);						\
377 	}								\
378 									\
379 	*flags |= CPU_DTRACE_NOFAULT;					\
380 	/*CSTYLED*/							\
381 	rval = *((volatile uint##bits##_t *)addr);			\
382 	*flags &= ~CPU_DTRACE_NOFAULT;					\
383 									\
384 	return (rval);							\
385 }
386 
387 #ifdef _LP64
388 #define	dtrace_loadptr	dtrace_load64
389 #else
390 #define	dtrace_loadptr	dtrace_load32
391 #endif
392 
393 #define	DTRACE_DYNHASH_FREE	0
394 #define	DTRACE_DYNHASH_SINK	1
395 #define	DTRACE_DYNHASH_VALID	2
396 
397 #define	DTRACE_MATCH_NEXT	0
398 #define	DTRACE_MATCH_DONE	1
399 #define	DTRACE_ANCHORED(probe)	((probe)->dtpr_func[0] != '\0')
400 #define	DTRACE_STATE_ALIGN	64
401 
402 #define	DTRACE_FLAGS2FLT(flags)						\
403 	(((flags) & CPU_DTRACE_BADADDR) ? DTRACEFLT_BADADDR :		\
404 	((flags) & CPU_DTRACE_ILLOP) ? DTRACEFLT_ILLOP :		\
405 	((flags) & CPU_DTRACE_DIVZERO) ? DTRACEFLT_DIVZERO :		\
406 	((flags) & CPU_DTRACE_KPRIV) ? DTRACEFLT_KPRIV :		\
407 	((flags) & CPU_DTRACE_UPRIV) ? DTRACEFLT_UPRIV :		\
408 	((flags) & CPU_DTRACE_TUPOFLOW) ?  DTRACEFLT_TUPOFLOW :		\
409 	((flags) & CPU_DTRACE_BADALIGN) ?  DTRACEFLT_BADALIGN :		\
410 	((flags) & CPU_DTRACE_NOSCRATCH) ?  DTRACEFLT_NOSCRATCH :	\
411 	DTRACEFLT_UNKNOWN)
412 
413 #define	DTRACEACT_ISSTRING(act)						\
414 	((act)->dta_kind == DTRACEACT_DIFEXPR &&			\
415 	(act)->dta_difo->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING)
416 
417 static dtrace_probe_t *dtrace_probe_lookup_id(dtrace_id_t id);
418 static void dtrace_enabling_provide(dtrace_provider_t *);
419 static int dtrace_enabling_match(dtrace_enabling_t *, int *);
420 static void dtrace_enabling_matchall(void);
421 static dtrace_state_t *dtrace_anon_grab(void);
422 static uint64_t dtrace_helper(int, dtrace_mstate_t *,
423     dtrace_state_t *, uint64_t, uint64_t);
424 static dtrace_helpers_t *dtrace_helpers_create(proc_t *);
425 static void dtrace_buffer_drop(dtrace_buffer_t *);
426 static intptr_t dtrace_buffer_reserve(dtrace_buffer_t *, size_t, size_t,
427     dtrace_state_t *, dtrace_mstate_t *);
428 static int dtrace_state_option(dtrace_state_t *, dtrace_optid_t,
429     dtrace_optval_t);
430 static int dtrace_ecb_create_enable(dtrace_probe_t *, void *);
431 static void dtrace_helper_provider_destroy(dtrace_helper_provider_t *);
432 
433 /*
434  * DTrace Probe Context Functions
435  *
436  * These functions are called from probe context.  Because probe context is
437  * any context in which C may be called, arbitrarily locks may be held,
438  * interrupts may be disabled, we may be in arbitrary dispatched state, etc.
439  * As a result, functions called from probe context may only call other DTrace
440  * support functions -- they may not interact at all with the system at large.
441  * (Note that the ASSERT macro is made probe-context safe by redefining it in
442  * terms of dtrace_assfail(), a probe-context safe function.) If arbitrary
443  * loads are to be performed from probe context, they _must_ be in terms of
444  * the safe dtrace_load*() variants.
445  *
446  * Some functions in this block are not actually called from probe context;
447  * for these functions, there will be a comment above the function reading
448  * "Note:  not called from probe context."
449  */
450 void
451 dtrace_panic(const char *format, ...)
452 {
453 	va_list alist;
454 
455 	va_start(alist, format);
456 	dtrace_vpanic(format, alist);
457 	va_end(alist);
458 }
459 
460 int
461 dtrace_assfail(const char *a, const char *f, int l)
462 {
463 	dtrace_panic("assertion failed: %s, file: %s, line: %d", a, f, l);
464 
465 	/*
466 	 * We just need something here that even the most clever compiler
467 	 * cannot optimize away.
468 	 */
469 	return (a[(uintptr_t)f]);
470 }
471 
472 /*
473  * Atomically increment a specified error counter from probe context.
474  */
475 static void
476 dtrace_error(uint32_t *counter)
477 {
478 	/*
479 	 * Most counters stored to in probe context are per-CPU counters.
480 	 * However, there are some error conditions that are sufficiently
481 	 * arcane that they don't merit per-CPU storage.  If these counters
482 	 * are incremented concurrently on different CPUs, scalability will be
483 	 * adversely affected -- but we don't expect them to be white-hot in a
484 	 * correctly constructed enabling...
485 	 */
486 	uint32_t oval, nval;
487 
488 	do {
489 		oval = *counter;
490 
491 		if ((nval = oval + 1) == 0) {
492 			/*
493 			 * If the counter would wrap, set it to 1 -- assuring
494 			 * that the counter is never zero when we have seen
495 			 * errors.  (The counter must be 32-bits because we
496 			 * aren't guaranteed a 64-bit compare&swap operation.)
497 			 * To save this code both the infamy of being fingered
498 			 * by a priggish news story and the indignity of being
499 			 * the target of a neo-puritan witch trial, we're
500 			 * carefully avoiding any colorful description of the
501 			 * likelihood of this condition -- but suffice it to
502 			 * say that it is only slightly more likely than the
503 			 * overflow of predicate cache IDs, as discussed in
504 			 * dtrace_predicate_create().
505 			 */
506 			nval = 1;
507 		}
508 	} while (dtrace_cas32(counter, oval, nval) != oval);
509 }
510 
511 /*
512  * Use the DTRACE_LOADFUNC macro to define functions for each of loading a
513  * uint8_t, a uint16_t, a uint32_t and a uint64_t.
514  */
515 DTRACE_LOADFUNC(8)
516 DTRACE_LOADFUNC(16)
517 DTRACE_LOADFUNC(32)
518 DTRACE_LOADFUNC(64)
519 
520 static int
521 dtrace_inscratch(uintptr_t dest, size_t size, dtrace_mstate_t *mstate)
522 {
523 	if (dest < mstate->dtms_scratch_base)
524 		return (0);
525 
526 	if (dest + size < dest)
527 		return (0);
528 
529 	if (dest + size > mstate->dtms_scratch_ptr)
530 		return (0);
531 
532 	return (1);
533 }
534 
535 static int
536 dtrace_canstore_statvar(uint64_t addr, size_t sz,
537     dtrace_statvar_t **svars, int nsvars)
538 {
539 	int i;
540 
541 	for (i = 0; i < nsvars; i++) {
542 		dtrace_statvar_t *svar = svars[i];
543 
544 		if (svar == NULL || svar->dtsv_size == 0)
545 			continue;
546 
547 		if (addr - svar->dtsv_data < svar->dtsv_size &&
548 		    addr + sz <= svar->dtsv_data + svar->dtsv_size)
549 			return (1);
550 	}
551 
552 	return (0);
553 }
554 
555 /*
556  * Check to see if the address is within a memory region to which a store may
557  * be issued.  This includes the DTrace scratch areas, and any DTrace variable
558  * region.  The caller of dtrace_canstore() is responsible for performing any
559  * alignment checks that are needed before stores are actually executed.
560  */
561 static int
562 dtrace_canstore(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
563     dtrace_vstate_t *vstate)
564 {
565 	uintptr_t a;
566 	size_t s;
567 
568 	/*
569 	 * First, check to see if the address is in scratch space...
570 	 */
571 	a = mstate->dtms_scratch_base;
572 	s = mstate->dtms_scratch_size;
573 
574 	if (addr - a < s && addr + sz <= a + s)
575 		return (1);
576 
577 	/*
578 	 * Now check to see if it's a dynamic variable.  This check will pick
579 	 * up both thread-local variables and any global dynamically-allocated
580 	 * variables.
581 	 */
582 	a = (uintptr_t)vstate->dtvs_dynvars.dtds_base;
583 	s = vstate->dtvs_dynvars.dtds_size;
584 	if (addr - a < s && addr + sz <= a + s)
585 		return (1);
586 
587 	/*
588 	 * Finally, check the static local and global variables.  These checks
589 	 * take the longest, so we perform them last.
590 	 */
591 	if (dtrace_canstore_statvar(addr, sz,
592 	    vstate->dtvs_locals, vstate->dtvs_nlocals))
593 		return (1);
594 
595 	if (dtrace_canstore_statvar(addr, sz,
596 	    vstate->dtvs_globals, vstate->dtvs_nglobals))
597 		return (1);
598 
599 	return (0);
600 }
601 
602 /*
603  * Compare two strings using safe loads.
604  */
605 static int
606 dtrace_strncmp(char *s1, char *s2, size_t limit)
607 {
608 	uint8_t c1, c2;
609 	volatile uint16_t *flags;
610 
611 	if (s1 == s2 || limit == 0)
612 		return (0);
613 
614 	flags = (volatile uint16_t *)&cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
615 
616 	do {
617 		if (s1 == NULL) {
618 			c1 = '\0';
619 		} else {
620 			c1 = dtrace_load8((uintptr_t)s1++);
621 		}
622 
623 		if (s2 == NULL) {
624 			c2 = '\0';
625 		} else {
626 			c2 = dtrace_load8((uintptr_t)s2++);
627 		}
628 
629 		if (c1 != c2)
630 			return (c1 - c2);
631 	} while (--limit && c1 != '\0' && !(*flags & CPU_DTRACE_FAULT));
632 
633 	return (0);
634 }
635 
636 /*
637  * Compute strlen(s) for a string using safe memory accesses.  The additional
638  * len parameter is used to specify a maximum length to ensure completion.
639  */
640 static size_t
641 dtrace_strlen(const char *s, size_t lim)
642 {
643 	uint_t len;
644 
645 	for (len = 0; len != lim; len++) {
646 		if (dtrace_load8((uintptr_t)s++) == '\0')
647 			break;
648 	}
649 
650 	return (len);
651 }
652 
653 /*
654  * Check if an address falls within a toxic region.
655  */
656 static int
657 dtrace_istoxic(uintptr_t kaddr, size_t size)
658 {
659 	uintptr_t taddr, tsize;
660 	int i;
661 
662 	for (i = 0; i < dtrace_toxranges; i++) {
663 		taddr = dtrace_toxrange[i].dtt_base;
664 		tsize = dtrace_toxrange[i].dtt_limit - taddr;
665 
666 		if (kaddr - taddr < tsize) {
667 			DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
668 			cpu_core[CPU->cpu_id].cpuc_dtrace_illval = kaddr;
669 			return (1);
670 		}
671 
672 		if (taddr - kaddr < size) {
673 			DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
674 			cpu_core[CPU->cpu_id].cpuc_dtrace_illval = taddr;
675 			return (1);
676 		}
677 	}
678 
679 	return (0);
680 }
681 
682 /*
683  * Copy src to dst using safe memory accesses.  The src is assumed to be unsafe
684  * memory specified by the DIF program.  The dst is assumed to be safe memory
685  * that we can store to directly because it is managed by DTrace.  As with
686  * standard bcopy, overlapping copies are handled properly.
687  */
688 static void
689 dtrace_bcopy(const void *src, void *dst, size_t len)
690 {
691 	if (len != 0) {
692 		uint8_t *s1 = dst;
693 		const uint8_t *s2 = src;
694 
695 		if (s1 <= s2) {
696 			do {
697 				*s1++ = dtrace_load8((uintptr_t)s2++);
698 			} while (--len != 0);
699 		} else {
700 			s2 += len;
701 			s1 += len;
702 
703 			do {
704 				*--s1 = dtrace_load8((uintptr_t)--s2);
705 			} while (--len != 0);
706 		}
707 	}
708 }
709 
710 /*
711  * Copy src to dst using safe memory accesses, up to either the specified
712  * length, or the point that a nul byte is encountered.  The src is assumed to
713  * be unsafe memory specified by the DIF program.  The dst is assumed to be
714  * safe memory that we can store to directly because it is managed by DTrace.
715  * Unlike dtrace_bcopy(), overlapping regions are not handled.
716  */
717 static void
718 dtrace_strcpy(const void *src, void *dst, size_t len)
719 {
720 	if (len != 0) {
721 		uint8_t *s1 = dst, c;
722 		const uint8_t *s2 = src;
723 
724 		do {
725 			*s1++ = c = dtrace_load8((uintptr_t)s2++);
726 		} while (--len != 0 && c != '\0');
727 	}
728 }
729 
730 /*
731  * Copy src to dst, deriving the size and type from the specified (BYREF)
732  * variable type.  The src is assumed to be unsafe memory specified by the DIF
733  * program.  The dst is assumed to be DTrace variable memory that is of the
734  * specified type; we assume that we can store to directly.
735  */
736 static void
737 dtrace_vcopy(void *src, void *dst, dtrace_diftype_t *type)
738 {
739 	ASSERT(type->dtdt_flags & DIF_TF_BYREF);
740 
741 	if (type->dtdt_kind == DIF_TYPE_STRING) {
742 		dtrace_strcpy(src, dst, type->dtdt_size);
743 	} else {
744 		dtrace_bcopy(src, dst, type->dtdt_size);
745 	}
746 }
747 
748 /*
749  * Compare s1 to s2 using safe memory accesses.  The s1 data is assumed to be
750  * unsafe memory specified by the DIF program.  The s2 data is assumed to be
751  * safe memory that we can access directly because it is managed by DTrace.
752  */
753 static int
754 dtrace_bcmp(const void *s1, const void *s2, size_t len)
755 {
756 	volatile uint16_t *flags;
757 
758 	flags = (volatile uint16_t *)&cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
759 
760 	if (s1 == s2)
761 		return (0);
762 
763 	if (s1 == NULL || s2 == NULL)
764 		return (1);
765 
766 	if (s1 != s2 && len != 0) {
767 		const uint8_t *ps1 = s1;
768 		const uint8_t *ps2 = s2;
769 
770 		do {
771 			if (dtrace_load8((uintptr_t)ps1++) != *ps2++)
772 				return (1);
773 		} while (--len != 0 && !(*flags & CPU_DTRACE_FAULT));
774 	}
775 	return (0);
776 }
777 
778 /*
779  * Zero the specified region using a simple byte-by-byte loop.  Note that this
780  * is for safe DTrace-managed memory only.
781  */
782 static void
783 dtrace_bzero(void *dst, size_t len)
784 {
785 	uchar_t *cp;
786 
787 	for (cp = dst; len != 0; len--)
788 		*cp++ = 0;
789 }
790 
791 /*
792  * This privilege check should be used by actions and subroutines to
793  * verify that the user credentials of the process that enabled the
794  * invoking ECB match the target credentials
795  */
796 static int
797 dtrace_priv_proc_common_user(dtrace_state_t *state)
798 {
799 	cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
800 
801 	/*
802 	 * We should always have a non-NULL state cred here, since if cred
803 	 * is null (anonymous tracing), we fast-path bypass this routine.
804 	 */
805 	ASSERT(s_cr != NULL);
806 
807 	if ((cr = CRED()) != NULL &&
808 	    s_cr->cr_uid == cr->cr_uid &&
809 	    s_cr->cr_uid == cr->cr_ruid &&
810 	    s_cr->cr_uid == cr->cr_suid &&
811 	    s_cr->cr_gid == cr->cr_gid &&
812 	    s_cr->cr_gid == cr->cr_rgid &&
813 	    s_cr->cr_gid == cr->cr_sgid)
814 		return (1);
815 
816 	return (0);
817 }
818 
819 /*
820  * This privilege check should be used by actions and subroutines to
821  * verify that the zone of the process that enabled the invoking ECB
822  * matches the target credentials
823  */
824 static int
825 dtrace_priv_proc_common_zone(dtrace_state_t *state)
826 {
827 	cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
828 
829 	/*
830 	 * We should always have a non-NULL state cred here, since if cred
831 	 * is null (anonymous tracing), we fast-path bypass this routine.
832 	 */
833 	ASSERT(s_cr != NULL);
834 
835 	if ((cr = CRED()) != NULL &&
836 	    s_cr->cr_zone == cr->cr_zone)
837 		return (1);
838 
839 	return (0);
840 }
841 
842 /*
843  * This privilege check should be used by actions and subroutines to
844  * verify that the process has not setuid or changed credentials.
845  */
846 static int
847 dtrace_priv_proc_common_nocd()
848 {
849 	proc_t *proc;
850 
851 	if ((proc = ttoproc(curthread)) != NULL &&
852 	    !(proc->p_flag & SNOCD))
853 		return (1);
854 
855 	return (0);
856 }
857 
858 static int
859 dtrace_priv_proc_destructive(dtrace_state_t *state)
860 {
861 	int action = state->dts_cred.dcr_action;
862 
863 	if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE) == 0) &&
864 	    dtrace_priv_proc_common_zone(state) == 0)
865 		goto bad;
866 
867 	if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER) == 0) &&
868 	    dtrace_priv_proc_common_user(state) == 0)
869 		goto bad;
870 
871 	if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG) == 0) &&
872 	    dtrace_priv_proc_common_nocd() == 0)
873 		goto bad;
874 
875 	return (1);
876 
877 bad:
878 	cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
879 
880 	return (0);
881 }
882 
883 static int
884 dtrace_priv_proc_control(dtrace_state_t *state)
885 {
886 	if (state->dts_cred.dcr_action & DTRACE_CRA_PROC_CONTROL)
887 		return (1);
888 
889 	if (dtrace_priv_proc_common_zone(state) &&
890 	    dtrace_priv_proc_common_user(state) &&
891 	    dtrace_priv_proc_common_nocd())
892 		return (1);
893 
894 	cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
895 
896 	return (0);
897 }
898 
899 static int
900 dtrace_priv_proc(dtrace_state_t *state)
901 {
902 	if (state->dts_cred.dcr_action & DTRACE_CRA_PROC)
903 		return (1);
904 
905 	cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
906 
907 	return (0);
908 }
909 
910 static int
911 dtrace_priv_kernel(dtrace_state_t *state)
912 {
913 	if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL)
914 		return (1);
915 
916 	cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
917 
918 	return (0);
919 }
920 
921 static int
922 dtrace_priv_kernel_destructive(dtrace_state_t *state)
923 {
924 	if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL_DESTRUCTIVE)
925 		return (1);
926 
927 	cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
928 
929 	return (0);
930 }
931 
932 /*
933  * Note:  not called from probe context.  This function is called
934  * asynchronously (and at a regular interval) from outside of probe context to
935  * clean the dirty dynamic variable lists on all CPUs.  Dynamic variable
936  * cleaning is explained in detail in <sys/dtrace_impl.h>.
937  */
938 void
939 dtrace_dynvar_clean(dtrace_dstate_t *dstate)
940 {
941 	dtrace_dynvar_t *dirty;
942 	dtrace_dstate_percpu_t *dcpu;
943 	int i, work = 0;
944 
945 	for (i = 0; i < NCPU; i++) {
946 		dcpu = &dstate->dtds_percpu[i];
947 
948 		ASSERT(dcpu->dtdsc_rinsing == NULL);
949 
950 		/*
951 		 * If the dirty list is NULL, there is no dirty work to do.
952 		 */
953 		if (dcpu->dtdsc_dirty == NULL)
954 			continue;
955 
956 		/*
957 		 * If the clean list is non-NULL, then we're not going to do
958 		 * any work for this CPU -- it means that there has not been
959 		 * a dtrace_dynvar() allocation on this CPU (or from this CPU)
960 		 * since the last time we cleaned house.
961 		 */
962 		if (dcpu->dtdsc_clean != NULL)
963 			continue;
964 
965 		work = 1;
966 
967 		/*
968 		 * Atomically move the dirty list aside.
969 		 */
970 		do {
971 			dirty = dcpu->dtdsc_dirty;
972 
973 			/*
974 			 * Before we zap the dirty list, set the rinsing list.
975 			 * (This allows for a potential assertion in
976 			 * dtrace_dynvar():  if a free dynamic variable appears
977 			 * on a hash chain, either the dirty list or the
978 			 * rinsing list for some CPU must be non-NULL.)
979 			 */
980 			dcpu->dtdsc_rinsing = dirty;
981 			dtrace_membar_producer();
982 		} while (dtrace_casptr(&dcpu->dtdsc_dirty,
983 		    dirty, NULL) != dirty);
984 	}
985 
986 	if (!work) {
987 		/*
988 		 * We have no work to do; we can simply return.
989 		 */
990 		return;
991 	}
992 
993 	dtrace_sync();
994 
995 	for (i = 0; i < NCPU; i++) {
996 		dcpu = &dstate->dtds_percpu[i];
997 
998 		if (dcpu->dtdsc_rinsing == NULL)
999 			continue;
1000 
1001 		/*
1002 		 * We are now guaranteed that no hash chain contains a pointer
1003 		 * into this dirty list; we can make it clean.
1004 		 */
1005 		ASSERT(dcpu->dtdsc_clean == NULL);
1006 		dcpu->dtdsc_clean = dcpu->dtdsc_rinsing;
1007 		dcpu->dtdsc_rinsing = NULL;
1008 	}
1009 
1010 	/*
1011 	 * Before we actually set the state to be DTRACE_DSTATE_CLEAN, make
1012 	 * sure that all CPUs have seen all of the dtdsc_clean pointers.
1013 	 * This prevents a race whereby a CPU incorrectly decides that
1014 	 * the state should be something other than DTRACE_DSTATE_CLEAN
1015 	 * after dtrace_dynvar_clean() has completed.
1016 	 */
1017 	dtrace_sync();
1018 
1019 	dstate->dtds_state = DTRACE_DSTATE_CLEAN;
1020 }
1021 
1022 /*
1023  * Depending on the value of the op parameter, this function looks-up,
1024  * allocates or deallocates an arbitrarily-keyed dynamic variable.  If an
1025  * allocation is requested, this function will return a pointer to a
1026  * dtrace_dynvar_t corresponding to the allocated variable -- or NULL if no
1027  * variable can be allocated.  If NULL is returned, the appropriate counter
1028  * will be incremented.
1029  */
1030 dtrace_dynvar_t *
1031 dtrace_dynvar(dtrace_dstate_t *dstate, uint_t nkeys,
1032     dtrace_key_t *key, size_t dsize, dtrace_dynvar_op_t op)
1033 {
1034 	uint64_t hashval = DTRACE_DYNHASH_VALID;
1035 	dtrace_dynhash_t *hash = dstate->dtds_hash;
1036 	dtrace_dynvar_t *free, *new_free, *next, *dvar, *start, *prev = NULL;
1037 	processorid_t me = CPU->cpu_id, cpu = me;
1038 	dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[me];
1039 	size_t bucket, ksize;
1040 	size_t chunksize = dstate->dtds_chunksize;
1041 	uintptr_t kdata, lock, nstate;
1042 	uint_t i;
1043 
1044 	ASSERT(nkeys != 0);
1045 
1046 	/*
1047 	 * Hash the key.  As with aggregations, we use Jenkins' "One-at-a-time"
1048 	 * algorithm.  For the by-value portions, we perform the algorithm in
1049 	 * 16-bit chunks (as opposed to 8-bit chunks).  This speeds things up a
1050 	 * bit, and seems to have only a minute effect on distribution.  For
1051 	 * the by-reference data, we perform "One-at-a-time" iterating (safely)
1052 	 * over each referenced byte.  It's painful to do this, but it's much
1053 	 * better than pathological hash distribution.  The efficacy of the
1054 	 * hashing algorithm (and a comparison with other algorithms) may be
1055 	 * found by running the ::dtrace_dynstat MDB dcmd.
1056 	 */
1057 	for (i = 0; i < nkeys; i++) {
1058 		if (key[i].dttk_size == 0) {
1059 			uint64_t val = key[i].dttk_value;
1060 
1061 			hashval += (val >> 48) & 0xffff;
1062 			hashval += (hashval << 10);
1063 			hashval ^= (hashval >> 6);
1064 
1065 			hashval += (val >> 32) & 0xffff;
1066 			hashval += (hashval << 10);
1067 			hashval ^= (hashval >> 6);
1068 
1069 			hashval += (val >> 16) & 0xffff;
1070 			hashval += (hashval << 10);
1071 			hashval ^= (hashval >> 6);
1072 
1073 			hashval += val & 0xffff;
1074 			hashval += (hashval << 10);
1075 			hashval ^= (hashval >> 6);
1076 		} else {
1077 			/*
1078 			 * This is incredibly painful, but it beats the hell
1079 			 * out of the alternative.
1080 			 */
1081 			uint64_t j, size = key[i].dttk_size;
1082 			uintptr_t base = (uintptr_t)key[i].dttk_value;
1083 
1084 			for (j = 0; j < size; j++) {
1085 				hashval += dtrace_load8(base + j);
1086 				hashval += (hashval << 10);
1087 				hashval ^= (hashval >> 6);
1088 			}
1089 		}
1090 	}
1091 
1092 	hashval += (hashval << 3);
1093 	hashval ^= (hashval >> 11);
1094 	hashval += (hashval << 15);
1095 
1096 	/*
1097 	 * There is a remote chance (ideally, 1 in 2^31) that our hashval
1098 	 * comes out to be one of our two sentinel hash values.  If this
1099 	 * actually happens, we set the hashval to be a value known to be a
1100 	 * non-sentinel value.
1101 	 */
1102 	if (hashval == DTRACE_DYNHASH_FREE || hashval == DTRACE_DYNHASH_SINK)
1103 		hashval = DTRACE_DYNHASH_VALID;
1104 
1105 	/*
1106 	 * Yes, it's painful to do a divide here.  If the cycle count becomes
1107 	 * important here, tricks can be pulled to reduce it.  (However, it's
1108 	 * critical that hash collisions be kept to an absolute minimum;
1109 	 * they're much more painful than a divide.)  It's better to have a
1110 	 * solution that generates few collisions and still keeps things
1111 	 * relatively simple.
1112 	 */
1113 	bucket = hashval % dstate->dtds_hashsize;
1114 
1115 	if (op == DTRACE_DYNVAR_DEALLOC) {
1116 		volatile uintptr_t *lockp = &hash[bucket].dtdh_lock;
1117 
1118 		for (;;) {
1119 			while ((lock = *lockp) & 1)
1120 				continue;
1121 
1122 			if (dtrace_casptr((void *)lockp,
1123 			    (void *)lock, (void *)(lock + 1)) == (void *)lock)
1124 				break;
1125 		}
1126 
1127 		dtrace_membar_producer();
1128 	}
1129 
1130 top:
1131 	prev = NULL;
1132 	lock = hash[bucket].dtdh_lock;
1133 
1134 	dtrace_membar_consumer();
1135 
1136 	start = hash[bucket].dtdh_chain;
1137 	ASSERT(start != NULL && (start->dtdv_hashval == DTRACE_DYNHASH_SINK ||
1138 	    start->dtdv_hashval != DTRACE_DYNHASH_FREE ||
1139 	    op != DTRACE_DYNVAR_DEALLOC));
1140 
1141 	for (dvar = start; dvar != NULL; dvar = dvar->dtdv_next) {
1142 		dtrace_tuple_t *dtuple = &dvar->dtdv_tuple;
1143 		dtrace_key_t *dkey = &dtuple->dtt_key[0];
1144 
1145 		if (dvar->dtdv_hashval != hashval) {
1146 			if (dvar->dtdv_hashval == DTRACE_DYNHASH_SINK) {
1147 				/*
1148 				 * We've reached the sink, and therefore the
1149 				 * end of the hash chain; we can kick out of
1150 				 * the loop knowing that we have seen a valid
1151 				 * snapshot of state.
1152 				 */
1153 				ASSERT(dvar->dtdv_next == NULL);
1154 				ASSERT(dvar == &dtrace_dynhash_sink);
1155 				break;
1156 			}
1157 
1158 			if (dvar->dtdv_hashval == DTRACE_DYNHASH_FREE) {
1159 				/*
1160 				 * We've gone off the rails:  somewhere along
1161 				 * the line, one of the members of this hash
1162 				 * chain was deleted.  Note that we could also
1163 				 * detect this by simply letting this loop run
1164 				 * to completion, as we would eventually hit
1165 				 * the end of the dirty list.  However, we
1166 				 * want to avoid running the length of the
1167 				 * dirty list unnecessarily (it might be quite
1168 				 * long), so we catch this as early as
1169 				 * possible by detecting the hash marker.  In
1170 				 * this case, we simply set dvar to NULL and
1171 				 * break; the conditional after the loop will
1172 				 * send us back to top.
1173 				 */
1174 				dvar = NULL;
1175 				break;
1176 			}
1177 
1178 			goto next;
1179 		}
1180 
1181 		if (dtuple->dtt_nkeys != nkeys)
1182 			goto next;
1183 
1184 		for (i = 0; i < nkeys; i++, dkey++) {
1185 			if (dkey->dttk_size != key[i].dttk_size)
1186 				goto next; /* size or type mismatch */
1187 
1188 			if (dkey->dttk_size != 0) {
1189 				if (dtrace_bcmp(
1190 				    (void *)(uintptr_t)key[i].dttk_value,
1191 				    (void *)(uintptr_t)dkey->dttk_value,
1192 				    dkey->dttk_size))
1193 					goto next;
1194 			} else {
1195 				if (dkey->dttk_value != key[i].dttk_value)
1196 					goto next;
1197 			}
1198 		}
1199 
1200 		if (op != DTRACE_DYNVAR_DEALLOC)
1201 			return (dvar);
1202 
1203 		ASSERT(dvar->dtdv_next == NULL ||
1204 		    dvar->dtdv_next->dtdv_hashval != DTRACE_DYNHASH_FREE);
1205 
1206 		if (prev != NULL) {
1207 			ASSERT(hash[bucket].dtdh_chain != dvar);
1208 			ASSERT(start != dvar);
1209 			ASSERT(prev->dtdv_next == dvar);
1210 			prev->dtdv_next = dvar->dtdv_next;
1211 		} else {
1212 			if (dtrace_casptr(&hash[bucket].dtdh_chain,
1213 			    start, dvar->dtdv_next) != start) {
1214 				/*
1215 				 * We have failed to atomically swing the
1216 				 * hash table head pointer, presumably because
1217 				 * of a conflicting allocation on another CPU.
1218 				 * We need to reread the hash chain and try
1219 				 * again.
1220 				 */
1221 				goto top;
1222 			}
1223 		}
1224 
1225 		dtrace_membar_producer();
1226 
1227 		/*
1228 		 * Now set the hash value to indicate that it's free.
1229 		 */
1230 		ASSERT(hash[bucket].dtdh_chain != dvar);
1231 		dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
1232 
1233 		dtrace_membar_producer();
1234 
1235 		/*
1236 		 * Set the next pointer to point at the dirty list, and
1237 		 * atomically swing the dirty pointer to the newly freed dvar.
1238 		 */
1239 		do {
1240 			next = dcpu->dtdsc_dirty;
1241 			dvar->dtdv_next = next;
1242 		} while (dtrace_casptr(&dcpu->dtdsc_dirty, next, dvar) != next);
1243 
1244 		/*
1245 		 * Finally, unlock this hash bucket.
1246 		 */
1247 		ASSERT(hash[bucket].dtdh_lock == lock);
1248 		ASSERT(lock & 1);
1249 		hash[bucket].dtdh_lock++;
1250 
1251 		return (NULL);
1252 next:
1253 		prev = dvar;
1254 		continue;
1255 	}
1256 
1257 	if (dvar == NULL) {
1258 		/*
1259 		 * If dvar is NULL, it is because we went off the rails:
1260 		 * one of the elements that we traversed in the hash chain
1261 		 * was deleted while we were traversing it.  In this case,
1262 		 * we assert that we aren't doing a dealloc (deallocs lock
1263 		 * the hash bucket to prevent themselves from racing with
1264 		 * one another), and retry the hash chain traversal.
1265 		 */
1266 		ASSERT(op != DTRACE_DYNVAR_DEALLOC);
1267 		goto top;
1268 	}
1269 
1270 	if (op != DTRACE_DYNVAR_ALLOC) {
1271 		/*
1272 		 * If we are not to allocate a new variable, we want to
1273 		 * return NULL now.  Before we return, check that the value
1274 		 * of the lock word hasn't changed.  If it has, we may have
1275 		 * seen an inconsistent snapshot.
1276 		 */
1277 		if (op == DTRACE_DYNVAR_NOALLOC) {
1278 			if (hash[bucket].dtdh_lock != lock)
1279 				goto top;
1280 		} else {
1281 			ASSERT(op == DTRACE_DYNVAR_DEALLOC);
1282 			ASSERT(hash[bucket].dtdh_lock == lock);
1283 			ASSERT(lock & 1);
1284 			hash[bucket].dtdh_lock++;
1285 		}
1286 
1287 		return (NULL);
1288 	}
1289 
1290 	/*
1291 	 * We need to allocate a new dynamic variable.  The size we need is the
1292 	 * size of dtrace_dynvar plus the size of nkeys dtrace_key_t's plus the
1293 	 * size of any auxiliary key data (rounded up to 8-byte alignment) plus
1294 	 * the size of any referred-to data (dsize).  We then round the final
1295 	 * size up to the chunksize for allocation.
1296 	 */
1297 	for (ksize = 0, i = 0; i < nkeys; i++)
1298 		ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
1299 
1300 	/*
1301 	 * This should be pretty much impossible, but could happen if, say,
1302 	 * strange DIF specified the tuple.  Ideally, this should be an
1303 	 * assertion and not an error condition -- but that requires that the
1304 	 * chunksize calculation in dtrace_difo_chunksize() be absolutely
1305 	 * bullet-proof.  (That is, it must not be able to be fooled by
1306 	 * malicious DIF.)  Given the lack of backwards branches in DIF,
1307 	 * solving this would presumably not amount to solving the Halting
1308 	 * Problem -- but it still seems awfully hard.
1309 	 */
1310 	if (sizeof (dtrace_dynvar_t) + sizeof (dtrace_key_t) * (nkeys - 1) +
1311 	    ksize + dsize > chunksize) {
1312 		dcpu->dtdsc_drops++;
1313 		return (NULL);
1314 	}
1315 
1316 	nstate = DTRACE_DSTATE_EMPTY;
1317 
1318 	do {
1319 retry:
1320 		free = dcpu->dtdsc_free;
1321 
1322 		if (free == NULL) {
1323 			dtrace_dynvar_t *clean = dcpu->dtdsc_clean;
1324 			void *rval;
1325 
1326 			if (clean == NULL) {
1327 				/*
1328 				 * We're out of dynamic variable space on
1329 				 * this CPU.  Unless we have tried all CPUs,
1330 				 * we'll try to allocate from a different
1331 				 * CPU.
1332 				 */
1333 				switch (dstate->dtds_state) {
1334 				case DTRACE_DSTATE_CLEAN: {
1335 					void *sp = &dstate->dtds_state;
1336 
1337 					if (++cpu >= NCPU)
1338 						cpu = 0;
1339 
1340 					if (dcpu->dtdsc_dirty != NULL &&
1341 					    nstate == DTRACE_DSTATE_EMPTY)
1342 						nstate = DTRACE_DSTATE_DIRTY;
1343 
1344 					if (dcpu->dtdsc_rinsing != NULL)
1345 						nstate = DTRACE_DSTATE_RINSING;
1346 
1347 					dcpu = &dstate->dtds_percpu[cpu];
1348 
1349 					if (cpu != me)
1350 						goto retry;
1351 
1352 					(void) dtrace_cas32(sp,
1353 					    DTRACE_DSTATE_CLEAN, nstate);
1354 
1355 					/*
1356 					 * To increment the correct bean
1357 					 * counter, take another lap.
1358 					 */
1359 					goto retry;
1360 				}
1361 
1362 				case DTRACE_DSTATE_DIRTY:
1363 					dcpu->dtdsc_dirty_drops++;
1364 					break;
1365 
1366 				case DTRACE_DSTATE_RINSING:
1367 					dcpu->dtdsc_rinsing_drops++;
1368 					break;
1369 
1370 				case DTRACE_DSTATE_EMPTY:
1371 					dcpu->dtdsc_drops++;
1372 					break;
1373 				}
1374 
1375 				DTRACE_CPUFLAG_SET(CPU_DTRACE_DROP);
1376 				return (NULL);
1377 			}
1378 
1379 			/*
1380 			 * The clean list appears to be non-empty.  We want to
1381 			 * move the clean list to the free list; we start by
1382 			 * moving the clean pointer aside.
1383 			 */
1384 			if (dtrace_casptr(&dcpu->dtdsc_clean,
1385 			    clean, NULL) != clean) {
1386 				/*
1387 				 * We are in one of two situations:
1388 				 *
1389 				 *  (a)	The clean list was switched to the
1390 				 *	free list by another CPU.
1391 				 *
1392 				 *  (b)	The clean list was added to by the
1393 				 *	cleansing cyclic.
1394 				 *
1395 				 * In either of these situations, we can
1396 				 * just reattempt the free list allocation.
1397 				 */
1398 				goto retry;
1399 			}
1400 
1401 			ASSERT(clean->dtdv_hashval == DTRACE_DYNHASH_FREE);
1402 
1403 			/*
1404 			 * Now we'll move the clean list to the free list.
1405 			 * It's impossible for this to fail:  the only way
1406 			 * the free list can be updated is through this
1407 			 * code path, and only one CPU can own the clean list.
1408 			 * Thus, it would only be possible for this to fail if
1409 			 * this code were racing with dtrace_dynvar_clean().
1410 			 * (That is, if dtrace_dynvar_clean() updated the clean
1411 			 * list, and we ended up racing to update the free
1412 			 * list.)  This race is prevented by the dtrace_sync()
1413 			 * in dtrace_dynvar_clean() -- which flushes the
1414 			 * owners of the clean lists out before resetting
1415 			 * the clean lists.
1416 			 */
1417 			rval = dtrace_casptr(&dcpu->dtdsc_free, NULL, clean);
1418 			ASSERT(rval == NULL);
1419 			goto retry;
1420 		}
1421 
1422 		dvar = free;
1423 		new_free = dvar->dtdv_next;
1424 	} while (dtrace_casptr(&dcpu->dtdsc_free, free, new_free) != free);
1425 
1426 	/*
1427 	 * We have now allocated a new chunk.  We copy the tuple keys into the
1428 	 * tuple array and copy any referenced key data into the data space
1429 	 * following the tuple array.  As we do this, we relocate dttk_value
1430 	 * in the final tuple to point to the key data address in the chunk.
1431 	 */
1432 	kdata = (uintptr_t)&dvar->dtdv_tuple.dtt_key[nkeys];
1433 	dvar->dtdv_data = (void *)(kdata + ksize);
1434 	dvar->dtdv_tuple.dtt_nkeys = nkeys;
1435 
1436 	for (i = 0; i < nkeys; i++) {
1437 		dtrace_key_t *dkey = &dvar->dtdv_tuple.dtt_key[i];
1438 		size_t kesize = key[i].dttk_size;
1439 
1440 		if (kesize != 0) {
1441 			dtrace_bcopy(
1442 			    (const void *)(uintptr_t)key[i].dttk_value,
1443 			    (void *)kdata, kesize);
1444 			dkey->dttk_value = kdata;
1445 			kdata += P2ROUNDUP(kesize, sizeof (uint64_t));
1446 		} else {
1447 			dkey->dttk_value = key[i].dttk_value;
1448 		}
1449 
1450 		dkey->dttk_size = kesize;
1451 	}
1452 
1453 	ASSERT(dvar->dtdv_hashval == DTRACE_DYNHASH_FREE);
1454 	dvar->dtdv_hashval = hashval;
1455 	dvar->dtdv_next = start;
1456 
1457 	if (dtrace_casptr(&hash[bucket].dtdh_chain, start, dvar) == start)
1458 		return (dvar);
1459 
1460 	/*
1461 	 * The cas has failed.  Either another CPU is adding an element to
1462 	 * this hash chain, or another CPU is deleting an element from this
1463 	 * hash chain.  The simplest way to deal with both of these cases
1464 	 * (though not necessarily the most efficient) is to free our
1465 	 * allocated block and tail-call ourselves.  Note that the free is
1466 	 * to the dirty list and _not_ to the free list.  This is to prevent
1467 	 * races with allocators, above.
1468 	 */
1469 	dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
1470 
1471 	dtrace_membar_producer();
1472 
1473 	do {
1474 		free = dcpu->dtdsc_dirty;
1475 		dvar->dtdv_next = free;
1476 	} while (dtrace_casptr(&dcpu->dtdsc_dirty, free, dvar) != free);
1477 
1478 	return (dtrace_dynvar(dstate, nkeys, key, dsize, op));
1479 }
1480 
1481 /*ARGSUSED*/
1482 static void
1483 dtrace_aggregate_min(uint64_t *oval, uint64_t nval, uint64_t arg)
1484 {
1485 	if (nval < *oval)
1486 		*oval = nval;
1487 }
1488 
1489 /*ARGSUSED*/
1490 static void
1491 dtrace_aggregate_max(uint64_t *oval, uint64_t nval, uint64_t arg)
1492 {
1493 	if (nval > *oval)
1494 		*oval = nval;
1495 }
1496 
1497 static void
1498 dtrace_aggregate_quantize(uint64_t *quanta, uint64_t nval, uint64_t incr)
1499 {
1500 	int i, zero = DTRACE_QUANTIZE_ZEROBUCKET;
1501 	int64_t val = (int64_t)nval;
1502 
1503 	if (val < 0) {
1504 		for (i = 0; i < zero; i++) {
1505 			if (val <= DTRACE_QUANTIZE_BUCKETVAL(i)) {
1506 				quanta[i] += incr;
1507 				return;
1508 			}
1509 		}
1510 	} else {
1511 		for (i = zero + 1; i < DTRACE_QUANTIZE_NBUCKETS; i++) {
1512 			if (val < DTRACE_QUANTIZE_BUCKETVAL(i)) {
1513 				quanta[i - 1] += incr;
1514 				return;
1515 			}
1516 		}
1517 
1518 		quanta[DTRACE_QUANTIZE_NBUCKETS - 1] += incr;
1519 		return;
1520 	}
1521 
1522 	ASSERT(0);
1523 }
1524 
1525 static void
1526 dtrace_aggregate_lquantize(uint64_t *lquanta, uint64_t nval, uint64_t incr)
1527 {
1528 	uint64_t arg = *lquanta++;
1529 	int32_t base = DTRACE_LQUANTIZE_BASE(arg);
1530 	uint16_t step = DTRACE_LQUANTIZE_STEP(arg);
1531 	uint16_t levels = DTRACE_LQUANTIZE_LEVELS(arg);
1532 	int32_t val = (int32_t)nval, level;
1533 
1534 	ASSERT(step != 0);
1535 	ASSERT(levels != 0);
1536 
1537 	if (val < base) {
1538 		/*
1539 		 * This is an underflow.
1540 		 */
1541 		lquanta[0] += incr;
1542 		return;
1543 	}
1544 
1545 	level = (val - base) / step;
1546 
1547 	if (level < levels) {
1548 		lquanta[level + 1] += incr;
1549 		return;
1550 	}
1551 
1552 	/*
1553 	 * This is an overflow.
1554 	 */
1555 	lquanta[levels + 1] += incr;
1556 }
1557 
1558 /*ARGSUSED*/
1559 static void
1560 dtrace_aggregate_avg(uint64_t *data, uint64_t nval, uint64_t arg)
1561 {
1562 	data[0]++;
1563 	data[1] += nval;
1564 }
1565 
1566 /*ARGSUSED*/
1567 static void
1568 dtrace_aggregate_count(uint64_t *oval, uint64_t nval, uint64_t arg)
1569 {
1570 	*oval = *oval + 1;
1571 }
1572 
1573 /*ARGSUSED*/
1574 static void
1575 dtrace_aggregate_sum(uint64_t *oval, uint64_t nval, uint64_t arg)
1576 {
1577 	*oval += nval;
1578 }
1579 
1580 /*
1581  * Aggregate given the tuple in the principal data buffer, and the aggregating
1582  * action denoted by the specified dtrace_aggregation_t.  The aggregation
1583  * buffer is specified as the buf parameter.  This routine does not return
1584  * failure; if there is no space in the aggregation buffer, the data will be
1585  * dropped, and a corresponding counter incremented.
1586  */
1587 static void
1588 dtrace_aggregate(dtrace_aggregation_t *agg, dtrace_buffer_t *dbuf,
1589     intptr_t offset, dtrace_buffer_t *buf, uint64_t expr, uint64_t arg)
1590 {
1591 	dtrace_recdesc_t *rec = &agg->dtag_action.dta_rec;
1592 	uint32_t i, ndx, size, fsize;
1593 	uint32_t align = sizeof (uint64_t) - 1;
1594 	dtrace_aggbuffer_t *agb;
1595 	dtrace_aggkey_t *key;
1596 	uint32_t hashval = 0, limit, isstr;
1597 	caddr_t tomax, data, kdata;
1598 	dtrace_actkind_t action;
1599 	dtrace_action_t *act;
1600 	uintptr_t offs;
1601 
1602 	if (buf == NULL)
1603 		return;
1604 
1605 	if (!agg->dtag_hasarg) {
1606 		/*
1607 		 * Currently, only quantize() and lquantize() take additional
1608 		 * arguments, and they have the same semantics:  an increment
1609 		 * value that defaults to 1 when not present.  If additional
1610 		 * aggregating actions take arguments, the setting of the
1611 		 * default argument value will presumably have to become more
1612 		 * sophisticated...
1613 		 */
1614 		arg = 1;
1615 	}
1616 
1617 	action = agg->dtag_action.dta_kind - DTRACEACT_AGGREGATION;
1618 	size = rec->dtrd_offset - agg->dtag_base;
1619 	fsize = size + rec->dtrd_size;
1620 
1621 	ASSERT(dbuf->dtb_tomax != NULL);
1622 	data = dbuf->dtb_tomax + offset + agg->dtag_base;
1623 
1624 	if ((tomax = buf->dtb_tomax) == NULL) {
1625 		dtrace_buffer_drop(buf);
1626 		return;
1627 	}
1628 
1629 	/*
1630 	 * The metastructure is always at the bottom of the buffer.
1631 	 */
1632 	agb = (dtrace_aggbuffer_t *)(tomax + buf->dtb_size -
1633 	    sizeof (dtrace_aggbuffer_t));
1634 
1635 	if (buf->dtb_offset == 0) {
1636 		/*
1637 		 * We just kludge up approximately 1/8th of the size to be
1638 		 * buckets.  If this guess ends up being routinely
1639 		 * off-the-mark, we may need to dynamically readjust this
1640 		 * based on past performance.
1641 		 */
1642 		uintptr_t hashsize = (buf->dtb_size >> 3) / sizeof (uintptr_t);
1643 
1644 		if ((uintptr_t)agb - hashsize * sizeof (dtrace_aggkey_t *) <
1645 		    (uintptr_t)tomax || hashsize == 0) {
1646 			/*
1647 			 * We've been given a ludicrously small buffer;
1648 			 * increment our drop count and leave.
1649 			 */
1650 			dtrace_buffer_drop(buf);
1651 			return;
1652 		}
1653 
1654 		/*
1655 		 * And now, a pathetic attempt to try to get a an odd (or
1656 		 * perchance, a prime) hash size for better hash distribution.
1657 		 */
1658 		if (hashsize > (DTRACE_AGGHASHSIZE_SLEW << 3))
1659 			hashsize -= DTRACE_AGGHASHSIZE_SLEW;
1660 
1661 		agb->dtagb_hashsize = hashsize;
1662 		agb->dtagb_hash = (dtrace_aggkey_t **)((uintptr_t)agb -
1663 		    agb->dtagb_hashsize * sizeof (dtrace_aggkey_t *));
1664 		agb->dtagb_free = (uintptr_t)agb->dtagb_hash;
1665 
1666 		for (i = 0; i < agb->dtagb_hashsize; i++)
1667 			agb->dtagb_hash[i] = NULL;
1668 	}
1669 
1670 	ASSERT(agg->dtag_first != NULL);
1671 	ASSERT(agg->dtag_first->dta_intuple);
1672 
1673 	/*
1674 	 * Calculate the hash value based on the key.  Note that we _don't_
1675 	 * include the aggid in the hashing (but we will store it as part of
1676 	 * the key).  The hashing algorithm is Bob Jenkins' "One-at-a-time"
1677 	 * algorithm: a simple, quick algorithm that has no known funnels, and
1678 	 * gets good distribution in practice.  The efficacy of the hashing
1679 	 * algorithm (and a comparison with other algorithms) may be found by
1680 	 * running the ::dtrace_aggstat MDB dcmd.
1681 	 */
1682 	for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
1683 		i = act->dta_rec.dtrd_offset - agg->dtag_base;
1684 		limit = i + act->dta_rec.dtrd_size;
1685 		ASSERT(limit <= size);
1686 		isstr = DTRACEACT_ISSTRING(act);
1687 
1688 		for (; i < limit; i++) {
1689 			hashval += data[i];
1690 			hashval += (hashval << 10);
1691 			hashval ^= (hashval >> 6);
1692 
1693 			if (isstr && data[i] == '\0')
1694 				break;
1695 		}
1696 	}
1697 
1698 	hashval += (hashval << 3);
1699 	hashval ^= (hashval >> 11);
1700 	hashval += (hashval << 15);
1701 
1702 	/*
1703 	 * Yes, the divide here is expensive -- but it's generally the least
1704 	 * of the performance issues given the amount of data that we iterate
1705 	 * over to compute hash values, compare data, etc.
1706 	 */
1707 	ndx = hashval % agb->dtagb_hashsize;
1708 
1709 	for (key = agb->dtagb_hash[ndx]; key != NULL; key = key->dtak_next) {
1710 		ASSERT((caddr_t)key >= tomax);
1711 		ASSERT((caddr_t)key < tomax + buf->dtb_size);
1712 
1713 		if (hashval != key->dtak_hashval || key->dtak_size != size)
1714 			continue;
1715 
1716 		kdata = key->dtak_data;
1717 		ASSERT(kdata >= tomax && kdata < tomax + buf->dtb_size);
1718 
1719 		for (act = agg->dtag_first; act->dta_intuple;
1720 		    act = act->dta_next) {
1721 			i = act->dta_rec.dtrd_offset - agg->dtag_base;
1722 			limit = i + act->dta_rec.dtrd_size;
1723 			ASSERT(limit <= size);
1724 			isstr = DTRACEACT_ISSTRING(act);
1725 
1726 			for (; i < limit; i++) {
1727 				if (kdata[i] != data[i])
1728 					goto next;
1729 
1730 				if (isstr && data[i] == '\0')
1731 					break;
1732 			}
1733 		}
1734 
1735 		if (action != key->dtak_action) {
1736 			/*
1737 			 * We are aggregating on the same value in the same
1738 			 * aggregation with two different aggregating actions.
1739 			 * (This should have been picked up in the compiler,
1740 			 * so we may be dealing with errant or devious DIF.)
1741 			 * This is an error condition; we indicate as much,
1742 			 * and return.
1743 			 */
1744 			DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
1745 			return;
1746 		}
1747 
1748 		/*
1749 		 * This is a hit:  we need to apply the aggregator to
1750 		 * the value at this key.
1751 		 */
1752 		agg->dtag_aggregate((uint64_t *)(kdata + size), expr, arg);
1753 		return;
1754 next:
1755 		continue;
1756 	}
1757 
1758 	/*
1759 	 * We didn't find it.  We need to allocate some zero-filled space,
1760 	 * link it into the hash table appropriately, and apply the aggregator
1761 	 * to the (zero-filled) value.
1762 	 */
1763 	offs = buf->dtb_offset;
1764 	while (offs & (align - 1))
1765 		offs += sizeof (uint32_t);
1766 
1767 	/*
1768 	 * If we don't have enough room to both allocate a new key _and_
1769 	 * its associated data, increment the drop count and return.
1770 	 */
1771 	if ((uintptr_t)tomax + offs + fsize >
1772 	    agb->dtagb_free - sizeof (dtrace_aggkey_t)) {
1773 		dtrace_buffer_drop(buf);
1774 		return;
1775 	}
1776 
1777 	/*CONSTCOND*/
1778 	ASSERT(!(sizeof (dtrace_aggkey_t) & (sizeof (uintptr_t) - 1)));
1779 	key = (dtrace_aggkey_t *)(agb->dtagb_free - sizeof (dtrace_aggkey_t));
1780 	agb->dtagb_free -= sizeof (dtrace_aggkey_t);
1781 
1782 	key->dtak_data = kdata = tomax + offs;
1783 	buf->dtb_offset = offs + fsize;
1784 
1785 	/*
1786 	 * Now copy the data across.
1787 	 */
1788 	*((dtrace_aggid_t *)kdata) = agg->dtag_id;
1789 
1790 	for (i = sizeof (dtrace_aggid_t); i < size; i++)
1791 		kdata[i] = data[i];
1792 
1793 	/*
1794 	 * Because strings are not zeroed out by default, we need to iterate
1795 	 * looking for actions that store strings, and we need to explicitly
1796 	 * pad these strings out with zeroes.
1797 	 */
1798 	for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
1799 		int nul;
1800 
1801 		if (!DTRACEACT_ISSTRING(act))
1802 			continue;
1803 
1804 		i = act->dta_rec.dtrd_offset - agg->dtag_base;
1805 		limit = i + act->dta_rec.dtrd_size;
1806 		ASSERT(limit <= size);
1807 
1808 		for (nul = 0; i < limit; i++) {
1809 			if (nul) {
1810 				kdata[i] = '\0';
1811 				continue;
1812 			}
1813 
1814 			if (data[i] != '\0')
1815 				continue;
1816 
1817 			nul = 1;
1818 		}
1819 	}
1820 
1821 	for (i = size; i < fsize; i++)
1822 		kdata[i] = 0;
1823 
1824 	key->dtak_hashval = hashval;
1825 	key->dtak_size = size;
1826 	key->dtak_action = action;
1827 	key->dtak_next = agb->dtagb_hash[ndx];
1828 	agb->dtagb_hash[ndx] = key;
1829 
1830 	/*
1831 	 * Finally, apply the aggregator.
1832 	 */
1833 	*((uint64_t *)(key->dtak_data + size)) = agg->dtag_initial;
1834 	agg->dtag_aggregate((uint64_t *)(key->dtak_data + size), expr, arg);
1835 }
1836 
1837 /*
1838  * Given consumer state, this routine finds a speculation in the INACTIVE
1839  * state and transitions it into the ACTIVE state.  If there is no speculation
1840  * in the INACTIVE state, 0 is returned.  In this case, no error counter is
1841  * incremented -- it is up to the caller to take appropriate action.
1842  */
1843 static int
1844 dtrace_speculation(dtrace_state_t *state)
1845 {
1846 	int i = 0;
1847 	dtrace_speculation_state_t current;
1848 	uint32_t *stat = &state->dts_speculations_unavail, count;
1849 
1850 	while (i < state->dts_nspeculations) {
1851 		dtrace_speculation_t *spec = &state->dts_speculations[i];
1852 
1853 		current = spec->dtsp_state;
1854 
1855 		if (current != DTRACESPEC_INACTIVE) {
1856 			if (current == DTRACESPEC_COMMITTINGMANY ||
1857 			    current == DTRACESPEC_COMMITTING ||
1858 			    current == DTRACESPEC_DISCARDING)
1859 				stat = &state->dts_speculations_busy;
1860 			i++;
1861 			continue;
1862 		}
1863 
1864 		if (dtrace_cas32((uint32_t *)&spec->dtsp_state,
1865 		    current, DTRACESPEC_ACTIVE) == current)
1866 			return (i + 1);
1867 	}
1868 
1869 	/*
1870 	 * We couldn't find a speculation.  If we found as much as a single
1871 	 * busy speculation buffer, we'll attribute this failure as "busy"
1872 	 * instead of "unavail".
1873 	 */
1874 	do {
1875 		count = *stat;
1876 	} while (dtrace_cas32(stat, count, count + 1) != count);
1877 
1878 	return (0);
1879 }
1880 
1881 /*
1882  * This routine commits an active speculation.  If the specified speculation
1883  * is not in a valid state to perform a commit(), this routine will silently do
1884  * nothing.  The state of the specified speculation is transitioned according
1885  * to the state transition diagram outlined in <sys/dtrace_impl.h>
1886  */
1887 static void
1888 dtrace_speculation_commit(dtrace_state_t *state, processorid_t cpu,
1889     dtrace_specid_t which)
1890 {
1891 	dtrace_speculation_t *spec;
1892 	dtrace_buffer_t *src, *dest;
1893 	uintptr_t daddr, saddr, dlimit;
1894 	dtrace_speculation_state_t current, new;
1895 	intptr_t offs;
1896 
1897 	if (which == 0)
1898 		return;
1899 
1900 	if (which > state->dts_nspeculations) {
1901 		cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
1902 		return;
1903 	}
1904 
1905 	spec = &state->dts_speculations[which - 1];
1906 	src = &spec->dtsp_buffer[cpu];
1907 	dest = &state->dts_buffer[cpu];
1908 
1909 	do {
1910 		current = spec->dtsp_state;
1911 
1912 		if (current == DTRACESPEC_COMMITTINGMANY)
1913 			break;
1914 
1915 		switch (current) {
1916 		case DTRACESPEC_INACTIVE:
1917 		case DTRACESPEC_DISCARDING:
1918 			return;
1919 
1920 		case DTRACESPEC_COMMITTING:
1921 			/*
1922 			 * This is only possible if we are (a) commit()'ing
1923 			 * without having done a prior speculate() on this CPU
1924 			 * and (b) racing with another commit() on a different
1925 			 * CPU.  There's nothing to do -- we just assert that
1926 			 * our offset is 0.
1927 			 */
1928 			ASSERT(src->dtb_offset == 0);
1929 			return;
1930 
1931 		case DTRACESPEC_ACTIVE:
1932 			new = DTRACESPEC_COMMITTING;
1933 			break;
1934 
1935 		case DTRACESPEC_ACTIVEONE:
1936 			/*
1937 			 * This speculation is active on one CPU.  If our
1938 			 * buffer offset is non-zero, we know that the one CPU
1939 			 * must be us.  Otherwise, we are committing on a
1940 			 * different CPU from the speculate(), and we must
1941 			 * rely on being asynchronously cleaned.
1942 			 */
1943 			if (src->dtb_offset != 0) {
1944 				new = DTRACESPEC_COMMITTING;
1945 				break;
1946 			}
1947 			/*FALLTHROUGH*/
1948 
1949 		case DTRACESPEC_ACTIVEMANY:
1950 			new = DTRACESPEC_COMMITTINGMANY;
1951 			break;
1952 
1953 		default:
1954 			ASSERT(0);
1955 		}
1956 	} while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
1957 	    current, new) != current);
1958 
1959 	/*
1960 	 * We have set the state to indicate that we are committing this
1961 	 * speculation.  Now reserve the necessary space in the destination
1962 	 * buffer.
1963 	 */
1964 	if ((offs = dtrace_buffer_reserve(dest, src->dtb_offset,
1965 	    sizeof (uint64_t), state, NULL)) < 0) {
1966 		dtrace_buffer_drop(dest);
1967 		goto out;
1968 	}
1969 
1970 	/*
1971 	 * We have the space; copy the buffer across.  (Note that this is a
1972 	 * highly subobtimal bcopy(); in the unlikely event that this becomes
1973 	 * a serious performance issue, a high-performance DTrace-specific
1974 	 * bcopy() should obviously be invented.)
1975 	 */
1976 	daddr = (uintptr_t)dest->dtb_tomax + offs;
1977 	dlimit = daddr + src->dtb_offset;
1978 	saddr = (uintptr_t)src->dtb_tomax;
1979 
1980 	/*
1981 	 * First, the aligned portion.
1982 	 */
1983 	while (dlimit - daddr >= sizeof (uint64_t)) {
1984 		*((uint64_t *)daddr) = *((uint64_t *)saddr);
1985 
1986 		daddr += sizeof (uint64_t);
1987 		saddr += sizeof (uint64_t);
1988 	}
1989 
1990 	/*
1991 	 * Now any left-over bit...
1992 	 */
1993 	while (dlimit - daddr)
1994 		*((uint8_t *)daddr++) = *((uint8_t *)saddr++);
1995 
1996 	/*
1997 	 * Finally, commit the reserved space in the destination buffer.
1998 	 */
1999 	dest->dtb_offset = offs + src->dtb_offset;
2000 
2001 out:
2002 	/*
2003 	 * If we're lucky enough to be the only active CPU on this speculation
2004 	 * buffer, we can just set the state back to DTRACESPEC_INACTIVE.
2005 	 */
2006 	if (current == DTRACESPEC_ACTIVE ||
2007 	    (current == DTRACESPEC_ACTIVEONE && new == DTRACESPEC_COMMITTING)) {
2008 		uint32_t rval = dtrace_cas32((uint32_t *)&spec->dtsp_state,
2009 		    DTRACESPEC_COMMITTING, DTRACESPEC_INACTIVE);
2010 
2011 		ASSERT(rval == DTRACESPEC_COMMITTING);
2012 	}
2013 
2014 	src->dtb_offset = 0;
2015 	src->dtb_xamot_drops += src->dtb_drops;
2016 	src->dtb_drops = 0;
2017 }
2018 
2019 /*
2020  * This routine discards an active speculation.  If the specified speculation
2021  * is not in a valid state to perform a discard(), this routine will silently
2022  * do nothing.  The state of the specified speculation is transitioned
2023  * according to the state transition diagram outlined in <sys/dtrace_impl.h>
2024  */
2025 static void
2026 dtrace_speculation_discard(dtrace_state_t *state, processorid_t cpu,
2027     dtrace_specid_t which)
2028 {
2029 	dtrace_speculation_t *spec;
2030 	dtrace_speculation_state_t current, new;
2031 	dtrace_buffer_t *buf;
2032 
2033 	if (which == 0)
2034 		return;
2035 
2036 	if (which > state->dts_nspeculations) {
2037 		cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2038 		return;
2039 	}
2040 
2041 	spec = &state->dts_speculations[which - 1];
2042 	buf = &spec->dtsp_buffer[cpu];
2043 
2044 	do {
2045 		current = spec->dtsp_state;
2046 
2047 		switch (current) {
2048 		case DTRACESPEC_INACTIVE:
2049 		case DTRACESPEC_COMMITTINGMANY:
2050 		case DTRACESPEC_COMMITTING:
2051 		case DTRACESPEC_DISCARDING:
2052 			return;
2053 
2054 		case DTRACESPEC_ACTIVE:
2055 		case DTRACESPEC_ACTIVEMANY:
2056 			new = DTRACESPEC_DISCARDING;
2057 			break;
2058 
2059 		case DTRACESPEC_ACTIVEONE:
2060 			if (buf->dtb_offset != 0) {
2061 				new = DTRACESPEC_INACTIVE;
2062 			} else {
2063 				new = DTRACESPEC_DISCARDING;
2064 			}
2065 			break;
2066 
2067 		default:
2068 			ASSERT(0);
2069 		}
2070 	} while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2071 	    current, new) != current);
2072 
2073 	buf->dtb_offset = 0;
2074 	buf->dtb_drops = 0;
2075 }
2076 
2077 /*
2078  * Note:  not called from probe context.  This function is called
2079  * asynchronously from cross call context to clean any speculations that are
2080  * in the COMMITTINGMANY or DISCARDING states.  These speculations may not be
2081  * transitioned back to the INACTIVE state until all CPUs have cleaned the
2082  * speculation.
2083  */
2084 static void
2085 dtrace_speculation_clean_here(dtrace_state_t *state)
2086 {
2087 	dtrace_icookie_t cookie;
2088 	processorid_t cpu = CPU->cpu_id;
2089 	dtrace_buffer_t *dest = &state->dts_buffer[cpu];
2090 	dtrace_specid_t i;
2091 
2092 	cookie = dtrace_interrupt_disable();
2093 
2094 	if (dest->dtb_tomax == NULL) {
2095 		dtrace_interrupt_enable(cookie);
2096 		return;
2097 	}
2098 
2099 	for (i = 0; i < state->dts_nspeculations; i++) {
2100 		dtrace_speculation_t *spec = &state->dts_speculations[i];
2101 		dtrace_buffer_t *src = &spec->dtsp_buffer[cpu];
2102 
2103 		if (src->dtb_tomax == NULL)
2104 			continue;
2105 
2106 		if (spec->dtsp_state == DTRACESPEC_DISCARDING) {
2107 			src->dtb_offset = 0;
2108 			continue;
2109 		}
2110 
2111 		if (spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
2112 			continue;
2113 
2114 		if (src->dtb_offset == 0)
2115 			continue;
2116 
2117 		dtrace_speculation_commit(state, cpu, i + 1);
2118 	}
2119 
2120 	dtrace_interrupt_enable(cookie);
2121 }
2122 
2123 /*
2124  * Note:  not called from probe context.  This function is called
2125  * asynchronously (and at a regular interval) to clean any speculations that
2126  * are in the COMMITTINGMANY or DISCARDING states.  If it discovers that there
2127  * is work to be done, it cross calls all CPUs to perform that work;
2128  * COMMITMANY and DISCARDING speculations may not be transitioned back to the
2129  * INACTIVE state until they have been cleaned by all CPUs.
2130  */
2131 static void
2132 dtrace_speculation_clean(dtrace_state_t *state)
2133 {
2134 	int work = 0, rv;
2135 	dtrace_specid_t i;
2136 
2137 	for (i = 0; i < state->dts_nspeculations; i++) {
2138 		dtrace_speculation_t *spec = &state->dts_speculations[i];
2139 
2140 		ASSERT(!spec->dtsp_cleaning);
2141 
2142 		if (spec->dtsp_state != DTRACESPEC_DISCARDING &&
2143 		    spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
2144 			continue;
2145 
2146 		work++;
2147 		spec->dtsp_cleaning = 1;
2148 	}
2149 
2150 	if (!work)
2151 		return;
2152 
2153 	dtrace_xcall(DTRACE_CPUALL,
2154 	    (dtrace_xcall_t)dtrace_speculation_clean_here, state);
2155 
2156 	/*
2157 	 * We now know that all CPUs have committed or discarded their
2158 	 * speculation buffers, as appropriate.  We can now set the state
2159 	 * to inactive.
2160 	 */
2161 	for (i = 0; i < state->dts_nspeculations; i++) {
2162 		dtrace_speculation_t *spec = &state->dts_speculations[i];
2163 		dtrace_speculation_state_t current, new;
2164 
2165 		if (!spec->dtsp_cleaning)
2166 			continue;
2167 
2168 		current = spec->dtsp_state;
2169 		ASSERT(current == DTRACESPEC_DISCARDING ||
2170 		    current == DTRACESPEC_COMMITTINGMANY);
2171 
2172 		new = DTRACESPEC_INACTIVE;
2173 
2174 		rv = dtrace_cas32((uint32_t *)&spec->dtsp_state, current, new);
2175 		ASSERT(rv == current);
2176 		spec->dtsp_cleaning = 0;
2177 	}
2178 }
2179 
2180 /*
2181  * Called as part of a speculate() to get the speculative buffer associated
2182  * with a given speculation.  Returns NULL if the specified speculation is not
2183  * in an ACTIVE state.  If the speculation is in the ACTIVEONE state -- and
2184  * the active CPU is not the specified CPU -- the speculation will be
2185  * atomically transitioned into the ACTIVEMANY state.
2186  */
2187 static dtrace_buffer_t *
2188 dtrace_speculation_buffer(dtrace_state_t *state, processorid_t cpuid,
2189     dtrace_specid_t which)
2190 {
2191 	dtrace_speculation_t *spec;
2192 	dtrace_speculation_state_t current, new;
2193 	dtrace_buffer_t *buf;
2194 
2195 	if (which == 0)
2196 		return (NULL);
2197 
2198 	if (which > state->dts_nspeculations) {
2199 		cpu_core[cpuid].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2200 		return (NULL);
2201 	}
2202 
2203 	spec = &state->dts_speculations[which - 1];
2204 	buf = &spec->dtsp_buffer[cpuid];
2205 
2206 	do {
2207 		current = spec->dtsp_state;
2208 
2209 		switch (current) {
2210 		case DTRACESPEC_INACTIVE:
2211 		case DTRACESPEC_COMMITTINGMANY:
2212 		case DTRACESPEC_DISCARDING:
2213 			return (NULL);
2214 
2215 		case DTRACESPEC_COMMITTING:
2216 			ASSERT(buf->dtb_offset == 0);
2217 			return (NULL);
2218 
2219 		case DTRACESPEC_ACTIVEONE:
2220 			/*
2221 			 * This speculation is currently active on one CPU.
2222 			 * Check the offset in the buffer; if it's non-zero,
2223 			 * that CPU must be us (and we leave the state alone).
2224 			 * If it's zero, assume that we're starting on a new
2225 			 * CPU -- and change the state to indicate that the
2226 			 * speculation is active on more than one CPU.
2227 			 */
2228 			if (buf->dtb_offset != 0)
2229 				return (buf);
2230 
2231 			new = DTRACESPEC_ACTIVEMANY;
2232 			break;
2233 
2234 		case DTRACESPEC_ACTIVEMANY:
2235 			return (buf);
2236 
2237 		case DTRACESPEC_ACTIVE:
2238 			new = DTRACESPEC_ACTIVEONE;
2239 			break;
2240 
2241 		default:
2242 			ASSERT(0);
2243 		}
2244 	} while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2245 	    current, new) != current);
2246 
2247 	ASSERT(new == DTRACESPEC_ACTIVEONE || new == DTRACESPEC_ACTIVEMANY);
2248 	return (buf);
2249 }
2250 
2251 /*
2252  * This function implements the DIF emulator's variable lookups.  The emulator
2253  * passes a reserved variable identifier and optional built-in array index.
2254  */
2255 static uint64_t
2256 dtrace_dif_variable(dtrace_mstate_t *mstate, dtrace_state_t *state, uint64_t v,
2257     uint64_t ndx)
2258 {
2259 	/*
2260 	 * If we're accessing one of the uncached arguments, we'll turn this
2261 	 * into a reference in the args array.
2262 	 */
2263 	if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9) {
2264 		ndx = v - DIF_VAR_ARG0;
2265 		v = DIF_VAR_ARGS;
2266 	}
2267 
2268 	switch (v) {
2269 	case DIF_VAR_ARGS:
2270 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_ARGS);
2271 		if (ndx >= sizeof (mstate->dtms_arg) /
2272 		    sizeof (mstate->dtms_arg[0])) {
2273 			int aframes = mstate->dtms_probe->dtpr_aframes + 2;
2274 			dtrace_provider_t *pv;
2275 			uint64_t val;
2276 
2277 			pv = mstate->dtms_probe->dtpr_provider;
2278 			if (pv->dtpv_pops.dtps_getargval != NULL)
2279 				val = pv->dtpv_pops.dtps_getargval(pv->dtpv_arg,
2280 				    mstate->dtms_probe->dtpr_id,
2281 				    mstate->dtms_probe->dtpr_arg, ndx, aframes);
2282 			else
2283 				val = dtrace_getarg(ndx, aframes);
2284 
2285 			/*
2286 			 * This is regrettably required to keep the compiler
2287 			 * from tail-optimizing the call to dtrace_getarg().
2288 			 * The condition always evaluates to true, but the
2289 			 * compiler has no way of figuring that out a priori.
2290 			 * (None of this would be necessary if the compiler
2291 			 * could be relied upon to _always_ tail-optimize
2292 			 * the call to dtrace_getarg() -- but it can't.)
2293 			 */
2294 			if (mstate->dtms_probe != NULL)
2295 				return (val);
2296 
2297 			ASSERT(0);
2298 		}
2299 
2300 		return (mstate->dtms_arg[ndx]);
2301 
2302 	case DIF_VAR_UREGS: {
2303 		klwp_t *lwp;
2304 
2305 		if (!dtrace_priv_proc(state))
2306 			return (0);
2307 
2308 		if ((lwp = curthread->t_lwp) == NULL) {
2309 			DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
2310 			cpu_core[CPU->cpu_id].cpuc_dtrace_illval = NULL;
2311 			return (0);
2312 		}
2313 
2314 		return (dtrace_getreg(lwp->lwp_regs, ndx));
2315 	}
2316 
2317 	case DIF_VAR_CURTHREAD:
2318 		if (!dtrace_priv_kernel(state))
2319 			return (0);
2320 		return ((uint64_t)(uintptr_t)curthread);
2321 
2322 	case DIF_VAR_TIMESTAMP:
2323 		if (!(mstate->dtms_present & DTRACE_MSTATE_TIMESTAMP)) {
2324 			mstate->dtms_timestamp = dtrace_gethrtime();
2325 			mstate->dtms_present |= DTRACE_MSTATE_TIMESTAMP;
2326 		}
2327 		return (mstate->dtms_timestamp);
2328 
2329 	case DIF_VAR_VTIMESTAMP:
2330 		ASSERT(dtrace_vtime_references != 0);
2331 		return (curthread->t_dtrace_vtime);
2332 
2333 	case DIF_VAR_WALLTIMESTAMP:
2334 		if (!(mstate->dtms_present & DTRACE_MSTATE_WALLTIMESTAMP)) {
2335 			mstate->dtms_walltimestamp = dtrace_gethrestime();
2336 			mstate->dtms_present |= DTRACE_MSTATE_WALLTIMESTAMP;
2337 		}
2338 		return (mstate->dtms_walltimestamp);
2339 
2340 	case DIF_VAR_IPL:
2341 		if (!dtrace_priv_kernel(state))
2342 			return (0);
2343 		if (!(mstate->dtms_present & DTRACE_MSTATE_IPL)) {
2344 			mstate->dtms_ipl = dtrace_getipl();
2345 			mstate->dtms_present |= DTRACE_MSTATE_IPL;
2346 		}
2347 		return (mstate->dtms_ipl);
2348 
2349 	case DIF_VAR_EPID:
2350 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_EPID);
2351 		return (mstate->dtms_epid);
2352 
2353 	case DIF_VAR_ID:
2354 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
2355 		return (mstate->dtms_probe->dtpr_id);
2356 
2357 	case DIF_VAR_STACKDEPTH:
2358 		if (!dtrace_priv_kernel(state))
2359 			return (0);
2360 		if (!(mstate->dtms_present & DTRACE_MSTATE_STACKDEPTH)) {
2361 			int aframes = mstate->dtms_probe->dtpr_aframes + 2;
2362 
2363 			mstate->dtms_stackdepth = dtrace_getstackdepth(aframes);
2364 			mstate->dtms_present |= DTRACE_MSTATE_STACKDEPTH;
2365 		}
2366 		return (mstate->dtms_stackdepth);
2367 
2368 	case DIF_VAR_USTACKDEPTH:
2369 		if (!dtrace_priv_proc(state))
2370 			return (0);
2371 		if (!(mstate->dtms_present & DTRACE_MSTATE_USTACKDEPTH)) {
2372 			/*
2373 			 * See comment in DIF_VAR_PID.
2374 			 */
2375 			if (DTRACE_ANCHORED(mstate->dtms_probe) &&
2376 			    CPU_ON_INTR(CPU)) {
2377 				mstate->dtms_ustackdepth = 0;
2378 			} else {
2379 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
2380 				mstate->dtms_ustackdepth =
2381 				    dtrace_getustackdepth();
2382 				DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
2383 			}
2384 			mstate->dtms_present |= DTRACE_MSTATE_USTACKDEPTH;
2385 		}
2386 		return (mstate->dtms_ustackdepth);
2387 
2388 	case DIF_VAR_CALLER:
2389 		if (!dtrace_priv_kernel(state))
2390 			return (0);
2391 		if (!(mstate->dtms_present & DTRACE_MSTATE_CALLER)) {
2392 			int aframes = mstate->dtms_probe->dtpr_aframes + 2;
2393 
2394 			if (!DTRACE_ANCHORED(mstate->dtms_probe)) {
2395 				/*
2396 				 * If this is an unanchored probe, we are
2397 				 * required to go through the slow path:
2398 				 * dtrace_caller() only guarantees correct
2399 				 * results for anchored probes.
2400 				 */
2401 				pc_t caller[2];
2402 
2403 				dtrace_getpcstack(caller, 2, aframes,
2404 				    (uint32_t *)(uintptr_t)mstate->dtms_arg[0]);
2405 				mstate->dtms_caller = caller[1];
2406 			} else if ((mstate->dtms_caller =
2407 			    dtrace_caller(aframes)) == -1) {
2408 				/*
2409 				 * We have failed to do this the quick way;
2410 				 * we must resort to the slower approach of
2411 				 * calling dtrace_getpcstack().
2412 				 */
2413 				pc_t caller;
2414 
2415 				dtrace_getpcstack(&caller, 1, aframes, NULL);
2416 				mstate->dtms_caller = caller;
2417 			}
2418 
2419 			mstate->dtms_present |= DTRACE_MSTATE_CALLER;
2420 		}
2421 		return (mstate->dtms_caller);
2422 
2423 	case DIF_VAR_UCALLER:
2424 		if (!dtrace_priv_proc(state))
2425 			return (0);
2426 
2427 		if (!(mstate->dtms_present & DTRACE_MSTATE_UCALLER)) {
2428 			uint64_t ustack[3];
2429 
2430 			/*
2431 			 * dtrace_getupcstack() fills in the first uint64_t
2432 			 * with the current PID.  The second uint64_t will
2433 			 * be the program counter at user-level.  The third
2434 			 * uint64_t will contain the caller, which is what
2435 			 * we're after.
2436 			 */
2437 			ustack[2] = NULL;
2438 			dtrace_getupcstack(ustack, 3);
2439 			mstate->dtms_ucaller = ustack[2];
2440 			mstate->dtms_present |= DTRACE_MSTATE_UCALLER;
2441 		}
2442 
2443 		return (mstate->dtms_ucaller);
2444 
2445 	case DIF_VAR_PROBEPROV:
2446 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
2447 		return ((uint64_t)(uintptr_t)
2448 		    mstate->dtms_probe->dtpr_provider->dtpv_name);
2449 
2450 	case DIF_VAR_PROBEMOD:
2451 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
2452 		return ((uint64_t)(uintptr_t)
2453 		    mstate->dtms_probe->dtpr_mod);
2454 
2455 	case DIF_VAR_PROBEFUNC:
2456 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
2457 		return ((uint64_t)(uintptr_t)
2458 		    mstate->dtms_probe->dtpr_func);
2459 
2460 	case DIF_VAR_PROBENAME:
2461 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
2462 		return ((uint64_t)(uintptr_t)
2463 		    mstate->dtms_probe->dtpr_name);
2464 
2465 	case DIF_VAR_PID:
2466 		if (!dtrace_priv_proc(state))
2467 			return (0);
2468 
2469 		/*
2470 		 * Note that we are assuming that an unanchored probe is
2471 		 * always due to a high-level interrupt.  (And we're assuming
2472 		 * that there is only a single high level interrupt.)
2473 		 */
2474 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
2475 			return (pid0.pid_id);
2476 
2477 		/*
2478 		 * It is always safe to dereference one's own t_procp pointer:
2479 		 * it always points to a valid, allocated proc structure.
2480 		 * Further, it is always safe to dereference the p_pidp member
2481 		 * of one's own proc structure.  (These are truisms becuase
2482 		 * threads and processes don't clean up their own state --
2483 		 * they leave that task to whomever reaps them.)
2484 		 */
2485 		return ((uint64_t)curthread->t_procp->p_pidp->pid_id);
2486 
2487 	case DIF_VAR_PPID:
2488 		if (!dtrace_priv_proc(state))
2489 			return (0);
2490 
2491 		/*
2492 		 * See comment in DIF_VAR_PID.
2493 		 */
2494 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
2495 			return (pid0.pid_id);
2496 
2497 		return ((uint64_t)curthread->t_procp->p_ppid);
2498 
2499 	case DIF_VAR_TID:
2500 		/*
2501 		 * See comment in DIF_VAR_PID.
2502 		 */
2503 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
2504 			return (0);
2505 
2506 		return ((uint64_t)curthread->t_tid);
2507 
2508 	case DIF_VAR_EXECNAME:
2509 		if (!dtrace_priv_proc(state))
2510 			return (0);
2511 
2512 		/*
2513 		 * See comment in DIF_VAR_PID.
2514 		 */
2515 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
2516 			return ((uint64_t)(uintptr_t)p0.p_user.u_comm);
2517 
2518 		/*
2519 		 * It is always safe to dereference one's own t_procp pointer:
2520 		 * it always points to a valid, allocated proc structure.
2521 		 * (This is true because threads don't clean up their own
2522 		 * state -- they leave that task to whomever reaps them.)
2523 		 */
2524 		return ((uint64_t)(uintptr_t)
2525 		    curthread->t_procp->p_user.u_comm);
2526 
2527 	case DIF_VAR_ZONENAME:
2528 		if (!dtrace_priv_proc(state))
2529 			return (0);
2530 
2531 		/*
2532 		 * See comment in DIF_VAR_PID.
2533 		 */
2534 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
2535 			return ((uint64_t)(uintptr_t)p0.p_zone->zone_name);
2536 
2537 		/*
2538 		 * It is always safe to dereference one's own t_procp pointer:
2539 		 * it always points to a valid, allocated proc structure.
2540 		 * (This is true because threads don't clean up their own
2541 		 * state -- they leave that task to whomever reaps them.)
2542 		 */
2543 		return ((uint64_t)(uintptr_t)
2544 		    curthread->t_procp->p_zone->zone_name);
2545 
2546 	case DIF_VAR_UID:
2547 		if (!dtrace_priv_proc(state))
2548 			return (0);
2549 
2550 		/*
2551 		 * See comment in DIF_VAR_PID.
2552 		 */
2553 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
2554 			return ((uint64_t)p0.p_cred->cr_uid);
2555 
2556 		return ((uint64_t)curthread->t_cred->cr_uid);
2557 
2558 	case DIF_VAR_GID:
2559 		if (!dtrace_priv_proc(state))
2560 			return (0);
2561 
2562 		/*
2563 		 * See comment in DIF_VAR_PID.
2564 		 */
2565 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
2566 			return ((uint64_t)p0.p_cred->cr_gid);
2567 
2568 		return ((uint64_t)curthread->t_cred->cr_gid);
2569 
2570 	case DIF_VAR_ERRNO: {
2571 		klwp_t *lwp;
2572 		if (!dtrace_priv_proc(state))
2573 			return (0);
2574 
2575 		/*
2576 		 * See comment in DIF_VAR_PID.
2577 		 */
2578 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
2579 			return (0);
2580 
2581 		if ((lwp = curthread->t_lwp) == NULL)
2582 			return (0);
2583 
2584 		return ((uint64_t)lwp->lwp_errno);
2585 	}
2586 	default:
2587 		DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
2588 		return (0);
2589 	}
2590 }
2591 
2592 /*
2593  * Emulate the execution of DTrace ID subroutines invoked by the call opcode.
2594  * Notice that we don't bother validating the proper number of arguments or
2595  * their types in the tuple stack.  This isn't needed because all argument
2596  * interpretation is safe because of our load safety -- the worst that can
2597  * happen is that a bogus program can obtain bogus results.
2598  */
2599 static void
2600 dtrace_dif_subr(uint_t subr, uint_t rd, uint64_t *regs,
2601     dtrace_key_t *tupregs, int nargs,
2602     dtrace_mstate_t *mstate, dtrace_state_t *state)
2603 {
2604 	volatile uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
2605 	volatile uintptr_t *illval = &cpu_core[CPU->cpu_id].cpuc_dtrace_illval;
2606 
2607 	union {
2608 		mutex_impl_t mi;
2609 		uint64_t mx;
2610 	} m;
2611 
2612 	union {
2613 		krwlock_t ri;
2614 		uintptr_t rw;
2615 	} r;
2616 
2617 	switch (subr) {
2618 	case DIF_SUBR_RAND:
2619 		regs[rd] = (dtrace_gethrtime() * 2416 + 374441) % 1771875;
2620 		break;
2621 
2622 	case DIF_SUBR_MUTEX_OWNED:
2623 		m.mx = dtrace_load64(tupregs[0].dttk_value);
2624 		if (MUTEX_TYPE_ADAPTIVE(&m.mi))
2625 			regs[rd] = MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER;
2626 		else
2627 			regs[rd] = LOCK_HELD(&m.mi.m_spin.m_spinlock);
2628 		break;
2629 
2630 	case DIF_SUBR_MUTEX_OWNER:
2631 		m.mx = dtrace_load64(tupregs[0].dttk_value);
2632 		if (MUTEX_TYPE_ADAPTIVE(&m.mi) &&
2633 		    MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER)
2634 			regs[rd] = (uintptr_t)MUTEX_OWNER(&m.mi);
2635 		else
2636 			regs[rd] = 0;
2637 		break;
2638 
2639 	case DIF_SUBR_MUTEX_TYPE_ADAPTIVE:
2640 		m.mx = dtrace_load64(tupregs[0].dttk_value);
2641 		regs[rd] = MUTEX_TYPE_ADAPTIVE(&m.mi);
2642 		break;
2643 
2644 	case DIF_SUBR_MUTEX_TYPE_SPIN:
2645 		m.mx = dtrace_load64(tupregs[0].dttk_value);
2646 		regs[rd] = MUTEX_TYPE_SPIN(&m.mi);
2647 		break;
2648 
2649 	case DIF_SUBR_RW_READ_HELD: {
2650 		uintptr_t tmp;
2651 
2652 		r.rw = dtrace_loadptr(tupregs[0].dttk_value);
2653 		regs[rd] = _RW_READ_HELD(&r.ri, tmp);
2654 		break;
2655 	}
2656 
2657 	case DIF_SUBR_RW_WRITE_HELD:
2658 		r.rw = dtrace_loadptr(tupregs[0].dttk_value);
2659 		regs[rd] = _RW_WRITE_HELD(&r.ri);
2660 		break;
2661 
2662 	case DIF_SUBR_RW_ISWRITER:
2663 		r.rw = dtrace_loadptr(tupregs[0].dttk_value);
2664 		regs[rd] = _RW_ISWRITER(&r.ri);
2665 		break;
2666 
2667 	case DIF_SUBR_BCOPY: {
2668 		/*
2669 		 * We need to be sure that the destination is in the scratch
2670 		 * region -- no other region is allowed.
2671 		 */
2672 		uintptr_t src = tupregs[0].dttk_value;
2673 		uintptr_t dest = tupregs[1].dttk_value;
2674 		size_t size = tupregs[2].dttk_value;
2675 
2676 		if (!dtrace_inscratch(dest, size, mstate)) {
2677 			*flags |= CPU_DTRACE_BADADDR;
2678 			*illval = regs[rd];
2679 			break;
2680 		}
2681 
2682 		dtrace_bcopy((void *)src, (void *)dest, size);
2683 		break;
2684 	}
2685 
2686 	case DIF_SUBR_ALLOCA:
2687 	case DIF_SUBR_COPYIN: {
2688 		uintptr_t dest = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
2689 		uint64_t size =
2690 		    tupregs[subr == DIF_SUBR_ALLOCA ? 0 : 1].dttk_value;
2691 		size_t scratch_size = (dest - mstate->dtms_scratch_ptr) + size;
2692 
2693 		/*
2694 		 * This action doesn't require any credential checks since
2695 		 * probes will not activate in user contexts to which the
2696 		 * enabling user does not have permissions.
2697 		 */
2698 		if (mstate->dtms_scratch_ptr + scratch_size >
2699 		    mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
2700 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
2701 			regs[rd] = NULL;
2702 			break;
2703 		}
2704 
2705 		if (subr == DIF_SUBR_COPYIN) {
2706 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
2707 			dtrace_copyin(tupregs[0].dttk_value, dest, size);
2708 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
2709 		}
2710 
2711 		mstate->dtms_scratch_ptr += scratch_size;
2712 		regs[rd] = dest;
2713 		break;
2714 	}
2715 
2716 	case DIF_SUBR_COPYINTO: {
2717 		uint64_t size = tupregs[1].dttk_value;
2718 		uintptr_t dest = tupregs[2].dttk_value;
2719 
2720 		/*
2721 		 * This action doesn't require any credential checks since
2722 		 * probes will not activate in user contexts to which the
2723 		 * enabling user does not have permissions.
2724 		 */
2725 		if (!dtrace_inscratch(dest, size, mstate)) {
2726 			*flags |= CPU_DTRACE_BADADDR;
2727 			*illval = regs[rd];
2728 			break;
2729 		}
2730 
2731 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
2732 		dtrace_copyin(tupregs[0].dttk_value, dest, size);
2733 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
2734 		break;
2735 	}
2736 
2737 	case DIF_SUBR_COPYINSTR: {
2738 		uintptr_t dest = mstate->dtms_scratch_ptr;
2739 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
2740 
2741 		if (nargs > 1 && tupregs[1].dttk_value < size)
2742 			size = tupregs[1].dttk_value + 1;
2743 
2744 		/*
2745 		 * This action doesn't require any credential checks since
2746 		 * probes will not activate in user contexts to which the
2747 		 * enabling user does not have permissions.
2748 		 */
2749 		if (mstate->dtms_scratch_ptr + size >
2750 		    mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
2751 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
2752 			regs[rd] = NULL;
2753 			break;
2754 		}
2755 
2756 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
2757 		dtrace_copyinstr(tupregs[0].dttk_value, dest, size);
2758 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
2759 
2760 		((char *)dest)[size - 1] = '\0';
2761 		mstate->dtms_scratch_ptr += size;
2762 		regs[rd] = dest;
2763 		break;
2764 	}
2765 
2766 	case DIF_SUBR_MSGSIZE:
2767 	case DIF_SUBR_MSGDSIZE: {
2768 		uintptr_t baddr = tupregs[0].dttk_value, daddr;
2769 		uintptr_t wptr, rptr;
2770 		size_t count = 0;
2771 		int cont = 0;
2772 
2773 		while (baddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
2774 			wptr = dtrace_loadptr(baddr +
2775 			    offsetof(mblk_t, b_wptr));
2776 
2777 			rptr = dtrace_loadptr(baddr +
2778 			    offsetof(mblk_t, b_rptr));
2779 
2780 			if (wptr < rptr) {
2781 				*flags |= CPU_DTRACE_BADADDR;
2782 				*illval = tupregs[0].dttk_value;
2783 				break;
2784 			}
2785 
2786 			daddr = dtrace_loadptr(baddr +
2787 			    offsetof(mblk_t, b_datap));
2788 
2789 			baddr = dtrace_loadptr(baddr +
2790 			    offsetof(mblk_t, b_cont));
2791 
2792 			/*
2793 			 * We want to prevent against denial-of-service here,
2794 			 * so we're only going to search the list for
2795 			 * dtrace_msgdsize_max mblks.
2796 			 */
2797 			if (cont++ > dtrace_msgdsize_max) {
2798 				*flags |= CPU_DTRACE_ILLOP;
2799 				break;
2800 			}
2801 
2802 			if (subr == DIF_SUBR_MSGDSIZE) {
2803 				if (dtrace_load8(daddr +
2804 				    offsetof(dblk_t, db_type)) != M_DATA)
2805 					continue;
2806 			}
2807 
2808 			count += wptr - rptr;
2809 		}
2810 
2811 		if (!(*flags & CPU_DTRACE_FAULT))
2812 			regs[rd] = count;
2813 
2814 		break;
2815 	}
2816 
2817 	case DIF_SUBR_PROGENYOF: {
2818 		pid_t pid = tupregs[0].dttk_value;
2819 		proc_t *p;
2820 		int rval = 0;
2821 
2822 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
2823 
2824 		for (p = curthread->t_procp; p != NULL; p = p->p_parent) {
2825 			if (p->p_pidp->pid_id == pid) {
2826 				rval = 1;
2827 				break;
2828 			}
2829 		}
2830 
2831 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
2832 
2833 		regs[rd] = rval;
2834 		break;
2835 	}
2836 
2837 	case DIF_SUBR_SPECULATION:
2838 		regs[rd] = dtrace_speculation(state);
2839 		break;
2840 
2841 	case DIF_SUBR_COPYOUT: {
2842 		uintptr_t kaddr = tupregs[0].dttk_value;
2843 		uintptr_t uaddr = tupregs[1].dttk_value;
2844 		uint64_t size = tupregs[2].dttk_value;
2845 
2846 		if (!dtrace_destructive_disallow &&
2847 		    dtrace_priv_proc_control(state) &&
2848 		    !dtrace_istoxic(kaddr, size)) {
2849 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
2850 			dtrace_copyout(kaddr, uaddr, size);
2851 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
2852 		}
2853 		break;
2854 	}
2855 
2856 	case DIF_SUBR_COPYOUTSTR: {
2857 		uintptr_t kaddr = tupregs[0].dttk_value;
2858 		uintptr_t uaddr = tupregs[1].dttk_value;
2859 		uint64_t size = tupregs[2].dttk_value;
2860 
2861 		if (!dtrace_destructive_disallow &&
2862 		    dtrace_priv_proc_control(state) &&
2863 		    !dtrace_istoxic(kaddr, size)) {
2864 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
2865 			dtrace_copyoutstr(kaddr, uaddr, size);
2866 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
2867 		}
2868 		break;
2869 	}
2870 
2871 	case DIF_SUBR_STRLEN:
2872 		regs[rd] = dtrace_strlen((char *)(uintptr_t)
2873 		    tupregs[0].dttk_value,
2874 		    state->dts_options[DTRACEOPT_STRSIZE]);
2875 		break;
2876 
2877 	case DIF_SUBR_STRCHR:
2878 	case DIF_SUBR_STRRCHR: {
2879 		/*
2880 		 * We're going to iterate over the string looking for the
2881 		 * specified character.  We will iterate until we have reached
2882 		 * the string length or we have found the character.  If this
2883 		 * is DIF_SUBR_STRRCHR, we will look for the last occurrence
2884 		 * of the specified character instead of the first.
2885 		 */
2886 		uintptr_t addr = tupregs[0].dttk_value;
2887 		uintptr_t limit = addr + state->dts_options[DTRACEOPT_STRSIZE];
2888 		char c, target = (char)tupregs[1].dttk_value;
2889 
2890 		for (regs[rd] = NULL; addr < limit; addr++) {
2891 			if ((c = dtrace_load8(addr)) == target) {
2892 				regs[rd] = addr;
2893 
2894 				if (subr == DIF_SUBR_STRCHR)
2895 					break;
2896 			}
2897 
2898 			if (c == '\0')
2899 				break;
2900 		}
2901 
2902 		break;
2903 	}
2904 
2905 	case DIF_SUBR_STRSTR:
2906 	case DIF_SUBR_INDEX:
2907 	case DIF_SUBR_RINDEX: {
2908 		/*
2909 		 * We're going to iterate over the string looking for the
2910 		 * specified string.  We will iterate until we have reached
2911 		 * the string length or we have found the string.  (Yes, this
2912 		 * is done in the most naive way possible -- but considering
2913 		 * that the string we're searching for is likely to be
2914 		 * relatively short, the complexity of Rabin-Karp or similar
2915 		 * hardly seems merited.)
2916 		 */
2917 		char *addr = (char *)(uintptr_t)tupregs[0].dttk_value;
2918 		char *substr = (char *)(uintptr_t)tupregs[1].dttk_value;
2919 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
2920 		size_t len = dtrace_strlen(addr, size);
2921 		size_t sublen = dtrace_strlen(substr, size);
2922 		char *limit = addr + len, *orig = addr;
2923 		int notfound = subr == DIF_SUBR_STRSTR ? 0 : -1;
2924 		int inc = 1;
2925 
2926 		regs[rd] = notfound;
2927 
2928 		/*
2929 		 * strstr() and index()/rindex() have similar semantics if
2930 		 * both strings are the empty string: strstr() returns a
2931 		 * pointer to the (empty) string, and index() and rindex()
2932 		 * both return index 0 (regardless of any position argument).
2933 		 */
2934 		if (sublen == 0 && len == 0) {
2935 			if (subr == DIF_SUBR_STRSTR)
2936 				regs[rd] = (uintptr_t)addr;
2937 			else
2938 				regs[rd] = 0;
2939 			break;
2940 		}
2941 
2942 		if (subr != DIF_SUBR_STRSTR) {
2943 			if (subr == DIF_SUBR_RINDEX) {
2944 				limit = orig - 1;
2945 				addr += len;
2946 				inc = -1;
2947 			}
2948 
2949 			/*
2950 			 * Both index() and rindex() take an optional position
2951 			 * argument that denotes the starting position.
2952 			 */
2953 			if (nargs == 3) {
2954 				int64_t pos = (int64_t)tupregs[2].dttk_value;
2955 
2956 				/*
2957 				 * If the position argument to index() is
2958 				 * negative, Perl implicitly clamps it at
2959 				 * zero.  This semantic is a little surprising
2960 				 * given the special meaning of negative
2961 				 * positions to similar Perl functions like
2962 				 * substr(), but it appears to reflect a
2963 				 * notion that index() can start from a
2964 				 * negative index and increment its way up to
2965 				 * the string.  Given this notion, Perl's
2966 				 * rindex() is at least self-consistent in
2967 				 * that it implicitly clamps positions greater
2968 				 * than the string length to be the string
2969 				 * length.  Where Perl completely loses
2970 				 * coherence, however, is when the specified
2971 				 * substring is the empty string ("").  In
2972 				 * this case, even if the position is
2973 				 * negative, rindex() returns 0 -- and even if
2974 				 * the position is greater than the length,
2975 				 * index() returns the string length.  These
2976 				 * semantics violate the notion that index()
2977 				 * should never return a value less than the
2978 				 * specified position and that rindex() should
2979 				 * never return a value greater than the
2980 				 * specified position.  (One assumes that
2981 				 * these semantics are artifacts of Perl's
2982 				 * implementation and not the results of
2983 				 * deliberate design -- it beggars belief that
2984 				 * even Larry Wall could desire such oddness.)
2985 				 * While in the abstract one would wish for
2986 				 * consistent position semantics across
2987 				 * substr(), index() and rindex() -- or at the
2988 				 * very least self-consistent position
2989 				 * semantics for index() and rindex() -- we
2990 				 * instead opt to keep with the extant Perl
2991 				 * semantics, in all their broken glory.  (Do
2992 				 * we have more desire to maintain Perl's
2993 				 * semantics than Perl does?  Probably.)
2994 				 */
2995 				if (subr == DIF_SUBR_RINDEX) {
2996 					if (pos < 0) {
2997 						if (sublen == 0)
2998 							regs[rd] = 0;
2999 						break;
3000 					}
3001 
3002 					if (pos > len)
3003 						pos = len;
3004 				} else {
3005 					if (pos < 0)
3006 						pos = 0;
3007 
3008 					if (pos >= len) {
3009 						if (sublen == 0)
3010 							regs[rd] = len;
3011 						break;
3012 					}
3013 				}
3014 
3015 				addr = orig + pos;
3016 			}
3017 		}
3018 
3019 		for (regs[rd] = notfound; addr != limit; addr += inc) {
3020 			if (dtrace_strncmp(addr, substr, sublen) == 0) {
3021 				if (subr != DIF_SUBR_STRSTR) {
3022 					/*
3023 					 * As D index() and rindex() are
3024 					 * modeled on Perl (and not on awk),
3025 					 * we return a zero-based (and not a
3026 					 * one-based) index.  (For you Perl
3027 					 * weenies: no, we're not going to add
3028 					 * $[ -- and shouldn't you be at a con
3029 					 * or something?)
3030 					 */
3031 					regs[rd] = (uintptr_t)(addr - orig);
3032 					break;
3033 				}
3034 
3035 				ASSERT(subr == DIF_SUBR_STRSTR);
3036 				regs[rd] = (uintptr_t)addr;
3037 				break;
3038 			}
3039 		}
3040 
3041 		break;
3042 	}
3043 
3044 	case DIF_SUBR_STRTOK: {
3045 		uintptr_t addr = tupregs[0].dttk_value;
3046 		uintptr_t tokaddr = tupregs[1].dttk_value;
3047 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3048 		uintptr_t limit, toklimit = tokaddr + size;
3049 		uint8_t c, tokmap[32];	 /* 256 / 8 */
3050 		char *dest = (char *)mstate->dtms_scratch_ptr;
3051 		int i;
3052 
3053 		if (mstate->dtms_scratch_ptr + size >
3054 		    mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
3055 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3056 			regs[rd] = NULL;
3057 			break;
3058 		}
3059 
3060 		if (addr == NULL) {
3061 			/*
3062 			 * If the address specified is NULL, we use our saved
3063 			 * strtok pointer from the mstate.  Note that this
3064 			 * means that the saved strtok pointer is _only_
3065 			 * valid within multiple enablings of the same probe --
3066 			 * it behaves like an implicit clause-local variable.
3067 			 */
3068 			addr = mstate->dtms_strtok;
3069 		}
3070 
3071 		/*
3072 		 * First, zero the token map, and then process the token
3073 		 * string -- setting a bit in the map for every character
3074 		 * found in the token string.
3075 		 */
3076 		for (i = 0; i < sizeof (tokmap); i++)
3077 			tokmap[i] = 0;
3078 
3079 		for (; tokaddr < toklimit; tokaddr++) {
3080 			if ((c = dtrace_load8(tokaddr)) == '\0')
3081 				break;
3082 
3083 			ASSERT((c >> 3) < sizeof (tokmap));
3084 			tokmap[c >> 3] |= (1 << (c & 0x7));
3085 		}
3086 
3087 		for (limit = addr + size; addr < limit; addr++) {
3088 			/*
3089 			 * We're looking for a character that is _not_ contained
3090 			 * in the token string.
3091 			 */
3092 			if ((c = dtrace_load8(addr)) == '\0')
3093 				break;
3094 
3095 			if (!(tokmap[c >> 3] & (1 << (c & 0x7))))
3096 				break;
3097 		}
3098 
3099 		if (c == '\0') {
3100 			/*
3101 			 * We reached the end of the string without finding
3102 			 * any character that was not in the token string.
3103 			 * We return NULL in this case, and we set the saved
3104 			 * address to NULL as well.
3105 			 */
3106 			regs[rd] = NULL;
3107 			mstate->dtms_strtok = NULL;
3108 			break;
3109 		}
3110 
3111 		/*
3112 		 * From here on, we're copying into the destination string.
3113 		 */
3114 		for (i = 0; addr < limit && i < size - 1; addr++) {
3115 			if ((c = dtrace_load8(addr)) == '\0')
3116 				break;
3117 
3118 			if (tokmap[c >> 3] & (1 << (c & 0x7)))
3119 				break;
3120 
3121 			ASSERT(i < size);
3122 			dest[i++] = c;
3123 		}
3124 
3125 		ASSERT(i < size);
3126 		dest[i] = '\0';
3127 		regs[rd] = (uintptr_t)dest;
3128 		mstate->dtms_scratch_ptr += size;
3129 		mstate->dtms_strtok = addr;
3130 		break;
3131 	}
3132 
3133 	case DIF_SUBR_SUBSTR: {
3134 		uintptr_t s = tupregs[0].dttk_value;
3135 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3136 		char *d = (char *)mstate->dtms_scratch_ptr;
3137 		int64_t index = (int64_t)tupregs[1].dttk_value;
3138 		int64_t remaining = (int64_t)tupregs[2].dttk_value;
3139 		size_t len = dtrace_strlen((char *)s, size);
3140 		int64_t i = 0;
3141 
3142 		if (nargs <= 2)
3143 			remaining = (int64_t)size;
3144 
3145 		if (mstate->dtms_scratch_ptr + size >
3146 		    mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
3147 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3148 			regs[rd] = NULL;
3149 			break;
3150 		}
3151 
3152 		if (index < 0) {
3153 			index += len;
3154 
3155 			if (index < 0 && index + remaining > 0) {
3156 				remaining += index;
3157 				index = 0;
3158 			}
3159 		}
3160 
3161 		if (index >= len || index < 0)
3162 			index = len;
3163 
3164 		for (d[0] = '\0'; remaining > 0; remaining--) {
3165 			if ((d[i++] = dtrace_load8(s++ + index)) == '\0')
3166 				break;
3167 
3168 			if (i == size) {
3169 				d[i - 1] = '\0';
3170 				break;
3171 			}
3172 		}
3173 
3174 		mstate->dtms_scratch_ptr += size;
3175 		regs[rd] = (uintptr_t)d;
3176 		break;
3177 	}
3178 
3179 	case DIF_SUBR_GETMAJOR:
3180 #ifdef _LP64
3181 		regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR64) & MAXMAJ64;
3182 #else
3183 		regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR) & MAXMAJ;
3184 #endif
3185 		break;
3186 
3187 	case DIF_SUBR_GETMINOR:
3188 #ifdef _LP64
3189 		regs[rd] = tupregs[0].dttk_value & MAXMIN64;
3190 #else
3191 		regs[rd] = tupregs[0].dttk_value & MAXMIN;
3192 #endif
3193 		break;
3194 
3195 	case DIF_SUBR_DDI_PATHNAME: {
3196 		/*
3197 		 * This one is a galactic mess.  We are going to roughly
3198 		 * emulate ddi_pathname(), but it's made more complicated
3199 		 * by the fact that we (a) want to include the minor name and
3200 		 * (b) must proceed iteratively instead of recursively.
3201 		 */
3202 		uintptr_t dest = mstate->dtms_scratch_ptr;
3203 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3204 		char *start = (char *)dest, *end = start + size - 1;
3205 		uintptr_t daddr = tupregs[0].dttk_value;
3206 		int64_t minor = (int64_t)tupregs[1].dttk_value;
3207 		char *s;
3208 		int i, len, depth = 0;
3209 
3210 		if (size == 0 || mstate->dtms_scratch_ptr + size >
3211 		    mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
3212 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3213 			regs[rd] = NULL;
3214 			break;
3215 		}
3216 
3217 		*end = '\0';
3218 
3219 		/*
3220 		 * We want to have a name for the minor.  In order to do this,
3221 		 * we need to walk the minor list from the devinfo.  We want
3222 		 * to be sure that we don't infinitely walk a circular list,
3223 		 * so we check for circularity by sending a scout pointer
3224 		 * ahead two elements for every element that we iterate over;
3225 		 * if the list is circular, these will ultimately point to the
3226 		 * same element.  You may recognize this little trick as the
3227 		 * answer to a stupid interview question -- one that always
3228 		 * seems to be asked by those who had to have it laboriously
3229 		 * explained to them, and who can't even concisely describe
3230 		 * the conditions under which one would be forced to resort to
3231 		 * this technique.  Needless to say, those conditions are
3232 		 * found here -- and probably only here.  Is this is the only
3233 		 * use of this infamous trick in shipping, production code?
3234 		 * If it isn't, it probably should be...
3235 		 */
3236 		if (minor != -1) {
3237 			uintptr_t maddr = dtrace_loadptr(daddr +
3238 			    offsetof(struct dev_info, devi_minor));
3239 
3240 			uintptr_t next = offsetof(struct ddi_minor_data, next);
3241 			uintptr_t name = offsetof(struct ddi_minor_data,
3242 			    d_minor) + offsetof(struct ddi_minor, name);
3243 			uintptr_t dev = offsetof(struct ddi_minor_data,
3244 			    d_minor) + offsetof(struct ddi_minor, dev);
3245 			uintptr_t scout;
3246 
3247 			if (maddr != NULL)
3248 				scout = dtrace_loadptr(maddr + next);
3249 
3250 			while (maddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
3251 				uint64_t m;
3252 #ifdef _LP64
3253 				m = dtrace_load64(maddr + dev) & MAXMIN64;
3254 #else
3255 				m = dtrace_load32(maddr + dev) & MAXMIN;
3256 #endif
3257 				if (m != minor) {
3258 					maddr = dtrace_loadptr(maddr + next);
3259 
3260 					if (scout == NULL)
3261 						continue;
3262 
3263 					scout = dtrace_loadptr(scout + next);
3264 
3265 					if (scout == NULL)
3266 						continue;
3267 
3268 					scout = dtrace_loadptr(scout + next);
3269 
3270 					if (scout == NULL)
3271 						continue;
3272 
3273 					if (scout == maddr) {
3274 						*flags |= CPU_DTRACE_ILLOP;
3275 						break;
3276 					}
3277 
3278 					continue;
3279 				}
3280 
3281 				/*
3282 				 * We have the minor data.  Now we need to
3283 				 * copy the minor's name into the end of the
3284 				 * pathname.
3285 				 */
3286 				s = (char *)dtrace_loadptr(maddr + name);
3287 				len = dtrace_strlen(s, size);
3288 
3289 				if (*flags & CPU_DTRACE_FAULT)
3290 					break;
3291 
3292 				if (len != 0) {
3293 					if ((end -= (len + 1)) < start)
3294 						break;
3295 
3296 					*end = ':';
3297 				}
3298 
3299 				for (i = 1; i <= len; i++)
3300 					end[i] = dtrace_load8((uintptr_t)s++);
3301 				break;
3302 			}
3303 		}
3304 
3305 		while (daddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
3306 			ddi_node_state_t devi_state;
3307 
3308 			devi_state = dtrace_load32(daddr +
3309 			    offsetof(struct dev_info, devi_node_state));
3310 
3311 			if (*flags & CPU_DTRACE_FAULT)
3312 				break;
3313 
3314 			if (devi_state >= DS_INITIALIZED) {
3315 				s = (char *)dtrace_loadptr(daddr +
3316 				    offsetof(struct dev_info, devi_addr));
3317 				len = dtrace_strlen(s, size);
3318 
3319 				if (*flags & CPU_DTRACE_FAULT)
3320 					break;
3321 
3322 				if (len != 0) {
3323 					if ((end -= (len + 1)) < start)
3324 						break;
3325 
3326 					*end = '@';
3327 				}
3328 
3329 				for (i = 1; i <= len; i++)
3330 					end[i] = dtrace_load8((uintptr_t)s++);
3331 			}
3332 
3333 			/*
3334 			 * Now for the node name...
3335 			 */
3336 			s = (char *)dtrace_loadptr(daddr +
3337 			    offsetof(struct dev_info, devi_node_name));
3338 
3339 			daddr = dtrace_loadptr(daddr +
3340 			    offsetof(struct dev_info, devi_parent));
3341 
3342 			/*
3343 			 * If our parent is NULL (that is, if we're the root
3344 			 * node), we're going to use the special path
3345 			 * "devices".
3346 			 */
3347 			if (daddr == NULL)
3348 				s = "devices";
3349 
3350 			len = dtrace_strlen(s, size);
3351 			if (*flags & CPU_DTRACE_FAULT)
3352 				break;
3353 
3354 			if ((end -= (len + 1)) < start)
3355 				break;
3356 
3357 			for (i = 1; i <= len; i++)
3358 				end[i] = dtrace_load8((uintptr_t)s++);
3359 			*end = '/';
3360 
3361 			if (depth++ > dtrace_devdepth_max) {
3362 				*flags |= CPU_DTRACE_ILLOP;
3363 				break;
3364 			}
3365 		}
3366 
3367 		if (end < start)
3368 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3369 
3370 		if (daddr == NULL) {
3371 			regs[rd] = (uintptr_t)end;
3372 			mstate->dtms_scratch_ptr += size;
3373 		}
3374 
3375 		break;
3376 	}
3377 
3378 	case DIF_SUBR_STRJOIN: {
3379 		char *d = (char *)mstate->dtms_scratch_ptr;
3380 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3381 		uintptr_t s1 = tupregs[0].dttk_value;
3382 		uintptr_t s2 = tupregs[1].dttk_value;
3383 		int i = 0;
3384 
3385 		if (mstate->dtms_scratch_ptr + size >
3386 		    mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
3387 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3388 			regs[rd] = NULL;
3389 			break;
3390 		}
3391 
3392 		for (;;) {
3393 			if (i >= size) {
3394 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3395 				regs[rd] = NULL;
3396 				break;
3397 			}
3398 
3399 			if ((d[i++] = dtrace_load8(s1++)) == '\0') {
3400 				i--;
3401 				break;
3402 			}
3403 		}
3404 
3405 		for (;;) {
3406 			if (i >= size) {
3407 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3408 				regs[rd] = NULL;
3409 				break;
3410 			}
3411 
3412 			if ((d[i++] = dtrace_load8(s2++)) == '\0')
3413 				break;
3414 		}
3415 
3416 		if (i < size) {
3417 			mstate->dtms_scratch_ptr += i;
3418 			regs[rd] = (uintptr_t)d;
3419 		}
3420 
3421 		break;
3422 	}
3423 
3424 	case DIF_SUBR_LLTOSTR: {
3425 		int64_t i = (int64_t)tupregs[0].dttk_value;
3426 		int64_t val = i < 0 ? i * -1 : i;
3427 		uint64_t size = 22;	/* enough room for 2^64 in decimal */
3428 		char *end = (char *)mstate->dtms_scratch_ptr + size - 1;
3429 
3430 		if (mstate->dtms_scratch_ptr + size >
3431 		    mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
3432 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3433 			regs[rd] = NULL;
3434 			break;
3435 		}
3436 
3437 		for (*end-- = '\0'; val; val /= 10)
3438 			*end-- = '0' + (val % 10);
3439 
3440 		if (i == 0)
3441 			*end-- = '0';
3442 
3443 		if (i < 0)
3444 			*end-- = '-';
3445 
3446 		regs[rd] = (uintptr_t)end + 1;
3447 		mstate->dtms_scratch_ptr += size;
3448 		break;
3449 	}
3450 
3451 	case DIF_SUBR_DIRNAME:
3452 	case DIF_SUBR_BASENAME: {
3453 		char *dest = (char *)mstate->dtms_scratch_ptr;
3454 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3455 		uintptr_t src = tupregs[0].dttk_value;
3456 		int i, j, len = dtrace_strlen((char *)src, size);
3457 		int lastbase = -1, firstbase = -1, lastdir = -1;
3458 		int start, end;
3459 
3460 		if (mstate->dtms_scratch_ptr + size >
3461 		    mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
3462 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3463 			regs[rd] = NULL;
3464 			break;
3465 		}
3466 
3467 		/*
3468 		 * The basename and dirname for a zero-length string is
3469 		 * defined to be "."
3470 		 */
3471 		if (len == 0) {
3472 			len = 1;
3473 			src = (uintptr_t)".";
3474 		}
3475 
3476 		/*
3477 		 * Start from the back of the string, moving back toward the
3478 		 * front until we see a character that isn't a slash.  That
3479 		 * character is the last character in the basename.
3480 		 */
3481 		for (i = len - 1; i >= 0; i--) {
3482 			if (dtrace_load8(src + i) != '/')
3483 				break;
3484 		}
3485 
3486 		if (i >= 0)
3487 			lastbase = i;
3488 
3489 		/*
3490 		 * Starting from the last character in the basename, move
3491 		 * towards the front until we find a slash.  The character
3492 		 * that we processed immediately before that is the first
3493 		 * character in the basename.
3494 		 */
3495 		for (; i >= 0; i--) {
3496 			if (dtrace_load8(src + i) == '/')
3497 				break;
3498 		}
3499 
3500 		if (i >= 0)
3501 			firstbase = i + 1;
3502 
3503 		/*
3504 		 * Now keep going until we find a non-slash character.  That
3505 		 * character is the last character in the dirname.
3506 		 */
3507 		for (; i >= 0; i--) {
3508 			if (dtrace_load8(src + i) != '/')
3509 				break;
3510 		}
3511 
3512 		if (i >= 0)
3513 			lastdir = i;
3514 
3515 		ASSERT(!(lastbase == -1 && firstbase != -1));
3516 		ASSERT(!(firstbase == -1 && lastdir != -1));
3517 
3518 		if (lastbase == -1) {
3519 			/*
3520 			 * We didn't find a non-slash character.  We know that
3521 			 * the length is non-zero, so the whole string must be
3522 			 * slashes.  In either the dirname or the basename
3523 			 * case, we return '/'.
3524 			 */
3525 			ASSERT(firstbase == -1);
3526 			firstbase = lastbase = lastdir = 0;
3527 		}
3528 
3529 		if (firstbase == -1) {
3530 			/*
3531 			 * The entire string consists only of a basename
3532 			 * component.  If we're looking for dirname, we need
3533 			 * to change our string to be just "."; if we're
3534 			 * looking for a basename, we'll just set the first
3535 			 * character of the basename to be 0.
3536 			 */
3537 			if (subr == DIF_SUBR_DIRNAME) {
3538 				ASSERT(lastdir == -1);
3539 				src = (uintptr_t)".";
3540 				lastdir = 0;
3541 			} else {
3542 				firstbase = 0;
3543 			}
3544 		}
3545 
3546 		if (subr == DIF_SUBR_DIRNAME) {
3547 			if (lastdir == -1) {
3548 				/*
3549 				 * We know that we have a slash in the name --
3550 				 * or lastdir would be set to 0, above.  And
3551 				 * because lastdir is -1, we know that this
3552 				 * slash must be the first character.  (That
3553 				 * is, the full string must be of the form
3554 				 * "/basename".)  In this case, the last
3555 				 * character of the directory name is 0.
3556 				 */
3557 				lastdir = 0;
3558 			}
3559 
3560 			start = 0;
3561 			end = lastdir;
3562 		} else {
3563 			ASSERT(subr == DIF_SUBR_BASENAME);
3564 			ASSERT(firstbase != -1 && lastbase != -1);
3565 			start = firstbase;
3566 			end = lastbase;
3567 		}
3568 
3569 		for (i = start, j = 0; i <= end && j < size - 1; i++, j++)
3570 			dest[j] = dtrace_load8(src + i);
3571 
3572 		dest[j] = '\0';
3573 		regs[rd] = (uintptr_t)dest;
3574 		mstate->dtms_scratch_ptr += size;
3575 		break;
3576 	}
3577 
3578 	case DIF_SUBR_CLEANPATH: {
3579 		char *dest = (char *)mstate->dtms_scratch_ptr, c;
3580 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3581 		uintptr_t src = tupregs[0].dttk_value;
3582 		int i = 0, j = 0;
3583 
3584 		if (mstate->dtms_scratch_ptr + size >
3585 		    mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
3586 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3587 			regs[rd] = NULL;
3588 			break;
3589 		}
3590 
3591 		/*
3592 		 * Move forward, loading each character.
3593 		 */
3594 		do {
3595 			c = dtrace_load8(src + i++);
3596 next:
3597 			if (j + 5 >= size)	/* 5 = strlen("/..c\0") */
3598 				break;
3599 
3600 			if (c != '/') {
3601 				dest[j++] = c;
3602 				continue;
3603 			}
3604 
3605 			c = dtrace_load8(src + i++);
3606 
3607 			if (c == '/') {
3608 				/*
3609 				 * We have two slashes -- we can just advance
3610 				 * to the next character.
3611 				 */
3612 				goto next;
3613 			}
3614 
3615 			if (c != '.') {
3616 				/*
3617 				 * This is not "." and it's not ".." -- we can
3618 				 * just store the "/" and this character and
3619 				 * drive on.
3620 				 */
3621 				dest[j++] = '/';
3622 				dest[j++] = c;
3623 				continue;
3624 			}
3625 
3626 			c = dtrace_load8(src + i++);
3627 
3628 			if (c == '/') {
3629 				/*
3630 				 * This is a "/./" component.  We're not going
3631 				 * to store anything in the destination buffer;
3632 				 * we're just going to go to the next component.
3633 				 */
3634 				goto next;
3635 			}
3636 
3637 			if (c != '.') {
3638 				/*
3639 				 * This is not ".." -- we can just store the
3640 				 * "/." and this character and continue
3641 				 * processing.
3642 				 */
3643 				dest[j++] = '/';
3644 				dest[j++] = '.';
3645 				dest[j++] = c;
3646 				continue;
3647 			}
3648 
3649 			c = dtrace_load8(src + i++);
3650 
3651 			if (c != '/' && c != '\0') {
3652 				/*
3653 				 * This is not ".." -- it's "..[mumble]".
3654 				 * We'll store the "/.." and this character
3655 				 * and continue processing.
3656 				 */
3657 				dest[j++] = '/';
3658 				dest[j++] = '.';
3659 				dest[j++] = '.';
3660 				dest[j++] = c;
3661 				continue;
3662 			}
3663 
3664 			/*
3665 			 * This is "/../" or "/..\0".  We need to back up
3666 			 * our destination pointer until we find a "/".
3667 			 */
3668 			i--;
3669 			while (j != 0 && dest[--j] != '/')
3670 				continue;
3671 
3672 			if (c == '\0')
3673 				dest[++j] = '/';
3674 		} while (c != '\0');
3675 
3676 		dest[j] = '\0';
3677 		regs[rd] = (uintptr_t)dest;
3678 		mstate->dtms_scratch_ptr += size;
3679 		break;
3680 	}
3681 	}
3682 }
3683 
3684 /*
3685  * Emulate the execution of DTrace IR instructions specified by the given
3686  * DIF object.  This function is deliberately void of assertions as all of
3687  * the necessary checks are handled by a call to dtrace_difo_validate().
3688  */
3689 static uint64_t
3690 dtrace_dif_emulate(dtrace_difo_t *difo, dtrace_mstate_t *mstate,
3691     dtrace_vstate_t *vstate, dtrace_state_t *state)
3692 {
3693 	const dif_instr_t *text = difo->dtdo_buf;
3694 	const uint_t textlen = difo->dtdo_len;
3695 	const char *strtab = difo->dtdo_strtab;
3696 	const uint64_t *inttab = difo->dtdo_inttab;
3697 
3698 	uint64_t rval = 0;
3699 	dtrace_statvar_t *svar;
3700 	dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
3701 	dtrace_difv_t *v;
3702 	volatile uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
3703 	volatile uintptr_t *illval = &cpu_core[CPU->cpu_id].cpuc_dtrace_illval;
3704 
3705 	dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
3706 	uint64_t regs[DIF_DIR_NREGS];
3707 	uint64_t *tmp;
3708 
3709 	uint8_t cc_n = 0, cc_z = 0, cc_v = 0, cc_c = 0;
3710 	int64_t cc_r;
3711 	uint_t pc = 0, id, opc;
3712 	uint8_t ttop = 0;
3713 	dif_instr_t instr;
3714 	uint_t r1, r2, rd;
3715 
3716 	regs[DIF_REG_R0] = 0; 		/* %r0 is fixed at zero */
3717 
3718 	while (pc < textlen && !(*flags & CPU_DTRACE_FAULT)) {
3719 		opc = pc;
3720 
3721 		instr = text[pc++];
3722 		r1 = DIF_INSTR_R1(instr);
3723 		r2 = DIF_INSTR_R2(instr);
3724 		rd = DIF_INSTR_RD(instr);
3725 
3726 		switch (DIF_INSTR_OP(instr)) {
3727 		case DIF_OP_OR:
3728 			regs[rd] = regs[r1] | regs[r2];
3729 			break;
3730 		case DIF_OP_XOR:
3731 			regs[rd] = regs[r1] ^ regs[r2];
3732 			break;
3733 		case DIF_OP_AND:
3734 			regs[rd] = regs[r1] & regs[r2];
3735 			break;
3736 		case DIF_OP_SLL:
3737 			regs[rd] = regs[r1] << regs[r2];
3738 			break;
3739 		case DIF_OP_SRL:
3740 			regs[rd] = regs[r1] >> regs[r2];
3741 			break;
3742 		case DIF_OP_SUB:
3743 			regs[rd] = regs[r1] - regs[r2];
3744 			break;
3745 		case DIF_OP_ADD:
3746 			regs[rd] = regs[r1] + regs[r2];
3747 			break;
3748 		case DIF_OP_MUL:
3749 			regs[rd] = regs[r1] * regs[r2];
3750 			break;
3751 		case DIF_OP_SDIV:
3752 			if (regs[r2] == 0) {
3753 				regs[rd] = 0;
3754 				*flags |= CPU_DTRACE_DIVZERO;
3755 			} else {
3756 				regs[rd] = (int64_t)regs[r1] /
3757 				    (int64_t)regs[r2];
3758 			}
3759 			break;
3760 
3761 		case DIF_OP_UDIV:
3762 			if (regs[r2] == 0) {
3763 				regs[rd] = 0;
3764 				*flags |= CPU_DTRACE_DIVZERO;
3765 			} else {
3766 				regs[rd] = regs[r1] / regs[r2];
3767 			}
3768 			break;
3769 
3770 		case DIF_OP_SREM:
3771 			if (regs[r2] == 0) {
3772 				regs[rd] = 0;
3773 				*flags |= CPU_DTRACE_DIVZERO;
3774 			} else {
3775 				regs[rd] = (int64_t)regs[r1] %
3776 				    (int64_t)regs[r2];
3777 			}
3778 			break;
3779 
3780 		case DIF_OP_UREM:
3781 			if (regs[r2] == 0) {
3782 				regs[rd] = 0;
3783 				*flags |= CPU_DTRACE_DIVZERO;
3784 			} else {
3785 				regs[rd] = regs[r1] % regs[r2];
3786 			}
3787 			break;
3788 
3789 		case DIF_OP_NOT:
3790 			regs[rd] = ~regs[r1];
3791 			break;
3792 		case DIF_OP_MOV:
3793 			regs[rd] = regs[r1];
3794 			break;
3795 		case DIF_OP_CMP:
3796 			cc_r = regs[r1] - regs[r2];
3797 			cc_n = cc_r < 0;
3798 			cc_z = cc_r == 0;
3799 			cc_v = 0;
3800 			cc_c = regs[r1] < regs[r2];
3801 			break;
3802 		case DIF_OP_TST:
3803 			cc_n = cc_v = cc_c = 0;
3804 			cc_z = regs[r1] == 0;
3805 			break;
3806 		case DIF_OP_BA:
3807 			pc = DIF_INSTR_LABEL(instr);
3808 			break;
3809 		case DIF_OP_BE:
3810 			if (cc_z)
3811 				pc = DIF_INSTR_LABEL(instr);
3812 			break;
3813 		case DIF_OP_BNE:
3814 			if (cc_z == 0)
3815 				pc = DIF_INSTR_LABEL(instr);
3816 			break;
3817 		case DIF_OP_BG:
3818 			if ((cc_z | (cc_n ^ cc_v)) == 0)
3819 				pc = DIF_INSTR_LABEL(instr);
3820 			break;
3821 		case DIF_OP_BGU:
3822 			if ((cc_c | cc_z) == 0)
3823 				pc = DIF_INSTR_LABEL(instr);
3824 			break;
3825 		case DIF_OP_BGE:
3826 			if ((cc_n ^ cc_v) == 0)
3827 				pc = DIF_INSTR_LABEL(instr);
3828 			break;
3829 		case DIF_OP_BGEU:
3830 			if (cc_c == 0)
3831 				pc = DIF_INSTR_LABEL(instr);
3832 			break;
3833 		case DIF_OP_BL:
3834 			if (cc_n ^ cc_v)
3835 				pc = DIF_INSTR_LABEL(instr);
3836 			break;
3837 		case DIF_OP_BLU:
3838 			if (cc_c)
3839 				pc = DIF_INSTR_LABEL(instr);
3840 			break;
3841 		case DIF_OP_BLE:
3842 			if (cc_z | (cc_n ^ cc_v))
3843 				pc = DIF_INSTR_LABEL(instr);
3844 			break;
3845 		case DIF_OP_BLEU:
3846 			if (cc_c | cc_z)
3847 				pc = DIF_INSTR_LABEL(instr);
3848 			break;
3849 		case DIF_OP_RLDSB:
3850 			if (!dtrace_canstore(regs[r1], 1, mstate, vstate)) {
3851 				*flags |= CPU_DTRACE_KPRIV;
3852 				*illval = regs[r1];
3853 				break;
3854 			}
3855 			/*FALLTHROUGH*/
3856 		case DIF_OP_LDSB:
3857 			regs[rd] = (int8_t)dtrace_load8(regs[r1]);
3858 			break;
3859 		case DIF_OP_RLDSH:
3860 			if (!dtrace_canstore(regs[r1], 2, mstate, vstate)) {
3861 				*flags |= CPU_DTRACE_KPRIV;
3862 				*illval = regs[r1];
3863 				break;
3864 			}
3865 			/*FALLTHROUGH*/
3866 		case DIF_OP_LDSH:
3867 			regs[rd] = (int16_t)dtrace_load16(regs[r1]);
3868 			break;
3869 		case DIF_OP_RLDSW:
3870 			if (!dtrace_canstore(regs[r1], 4, mstate, vstate)) {
3871 				*flags |= CPU_DTRACE_KPRIV;
3872 				*illval = regs[r1];
3873 				break;
3874 			}
3875 			/*FALLTHROUGH*/
3876 		case DIF_OP_LDSW:
3877 			regs[rd] = (int32_t)dtrace_load32(regs[r1]);
3878 			break;
3879 		case DIF_OP_RLDUB:
3880 			if (!dtrace_canstore(regs[r1], 1, mstate, vstate)) {
3881 				*flags |= CPU_DTRACE_KPRIV;
3882 				*illval = regs[r1];
3883 				break;
3884 			}
3885 			/*FALLTHROUGH*/
3886 		case DIF_OP_LDUB:
3887 			regs[rd] = dtrace_load8(regs[r1]);
3888 			break;
3889 		case DIF_OP_RLDUH:
3890 			if (!dtrace_canstore(regs[r1], 2, mstate, vstate)) {
3891 				*flags |= CPU_DTRACE_KPRIV;
3892 				*illval = regs[r1];
3893 				break;
3894 			}
3895 			/*FALLTHROUGH*/
3896 		case DIF_OP_LDUH:
3897 			regs[rd] = dtrace_load16(regs[r1]);
3898 			break;
3899 		case DIF_OP_RLDUW:
3900 			if (!dtrace_canstore(regs[r1], 4, mstate, vstate)) {
3901 				*flags |= CPU_DTRACE_KPRIV;
3902 				*illval = regs[r1];
3903 				break;
3904 			}
3905 			/*FALLTHROUGH*/
3906 		case DIF_OP_LDUW:
3907 			regs[rd] = dtrace_load32(regs[r1]);
3908 			break;
3909 		case DIF_OP_RLDX:
3910 			if (!dtrace_canstore(regs[r1], 8, mstate, vstate)) {
3911 				*flags |= CPU_DTRACE_KPRIV;
3912 				*illval = regs[r1];
3913 				break;
3914 			}
3915 			/*FALLTHROUGH*/
3916 		case DIF_OP_LDX:
3917 			regs[rd] = dtrace_load64(regs[r1]);
3918 			break;
3919 		case DIF_OP_ULDSB:
3920 			regs[rd] = (int8_t)
3921 			    dtrace_fuword8((void *)(uintptr_t)regs[r1]);
3922 			break;
3923 		case DIF_OP_ULDSH:
3924 			regs[rd] = (int16_t)
3925 			    dtrace_fuword16((void *)(uintptr_t)regs[r1]);
3926 			break;
3927 		case DIF_OP_ULDSW:
3928 			regs[rd] = (int32_t)
3929 			    dtrace_fuword32((void *)(uintptr_t)regs[r1]);
3930 			break;
3931 		case DIF_OP_ULDUB:
3932 			regs[rd] =
3933 			    dtrace_fuword8((void *)(uintptr_t)regs[r1]);
3934 			break;
3935 		case DIF_OP_ULDUH:
3936 			regs[rd] =
3937 			    dtrace_fuword16((void *)(uintptr_t)regs[r1]);
3938 			break;
3939 		case DIF_OP_ULDUW:
3940 			regs[rd] =
3941 			    dtrace_fuword32((void *)(uintptr_t)regs[r1]);
3942 			break;
3943 		case DIF_OP_ULDX:
3944 			regs[rd] =
3945 			    dtrace_fuword64((void *)(uintptr_t)regs[r1]);
3946 			break;
3947 		case DIF_OP_RET:
3948 			rval = regs[rd];
3949 			break;
3950 		case DIF_OP_NOP:
3951 			break;
3952 		case DIF_OP_SETX:
3953 			regs[rd] = inttab[DIF_INSTR_INTEGER(instr)];
3954 			break;
3955 		case DIF_OP_SETS:
3956 			regs[rd] = (uint64_t)(uintptr_t)
3957 			    (strtab + DIF_INSTR_STRING(instr));
3958 			break;
3959 		case DIF_OP_SCMP:
3960 			cc_r = dtrace_strncmp((char *)(uintptr_t)regs[r1],
3961 			    (char *)(uintptr_t)regs[r2],
3962 			    state->dts_options[DTRACEOPT_STRSIZE]);
3963 
3964 			cc_n = cc_r < 0;
3965 			cc_z = cc_r == 0;
3966 			cc_v = cc_c = 0;
3967 			break;
3968 		case DIF_OP_LDGA:
3969 			regs[rd] = dtrace_dif_variable(mstate, state,
3970 			    r1, regs[r2]);
3971 			break;
3972 		case DIF_OP_LDGS:
3973 			id = DIF_INSTR_VAR(instr);
3974 
3975 			if (id >= DIF_VAR_OTHER_UBASE) {
3976 				uintptr_t a;
3977 
3978 				id -= DIF_VAR_OTHER_UBASE;
3979 				svar = vstate->dtvs_globals[id];
3980 				ASSERT(svar != NULL);
3981 				v = &svar->dtsv_var;
3982 
3983 				if (!(v->dtdv_type.dtdt_flags & DIF_TF_BYREF)) {
3984 					regs[rd] = svar->dtsv_data;
3985 					break;
3986 				}
3987 
3988 				a = (uintptr_t)svar->dtsv_data;
3989 
3990 				if (*(uint8_t *)a == UINT8_MAX) {
3991 					/*
3992 					 * If the 0th byte is set to UINT8_MAX
3993 					 * then this is to be treated as a
3994 					 * reference to a NULL variable.
3995 					 */
3996 					regs[rd] = NULL;
3997 				} else {
3998 					regs[rd] = a + sizeof (uint64_t);
3999 				}
4000 
4001 				break;
4002 			}
4003 
4004 			regs[rd] = dtrace_dif_variable(mstate, state, id, 0);
4005 			break;
4006 
4007 		case DIF_OP_STGS:
4008 			id = DIF_INSTR_VAR(instr);
4009 
4010 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
4011 			id -= DIF_VAR_OTHER_UBASE;
4012 
4013 			svar = vstate->dtvs_globals[id];
4014 			ASSERT(svar != NULL);
4015 			v = &svar->dtsv_var;
4016 
4017 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
4018 				uintptr_t a = (uintptr_t)svar->dtsv_data;
4019 
4020 				ASSERT(a != NULL);
4021 				ASSERT(svar->dtsv_size != 0);
4022 
4023 				if (regs[rd] == NULL) {
4024 					*(uint8_t *)a = UINT8_MAX;
4025 					break;
4026 				} else {
4027 					*(uint8_t *)a = 0;
4028 					a += sizeof (uint64_t);
4029 				}
4030 
4031 				dtrace_vcopy((void *)(uintptr_t)regs[rd],
4032 				    (void *)a, &v->dtdv_type);
4033 				break;
4034 			}
4035 
4036 			svar->dtsv_data = regs[rd];
4037 			break;
4038 
4039 		case DIF_OP_LDTA:
4040 			/*
4041 			 * There are no DTrace built-in thread-local arrays at
4042 			 * present.  This opcode is saved for future work.
4043 			 */
4044 			*flags |= CPU_DTRACE_ILLOP;
4045 			regs[rd] = 0;
4046 			break;
4047 
4048 		case DIF_OP_LDLS:
4049 			id = DIF_INSTR_VAR(instr);
4050 
4051 			if (id < DIF_VAR_OTHER_UBASE) {
4052 				/*
4053 				 * For now, this has no meaning.
4054 				 */
4055 				regs[rd] = 0;
4056 				break;
4057 			}
4058 
4059 			id -= DIF_VAR_OTHER_UBASE;
4060 
4061 			ASSERT(id < vstate->dtvs_nlocals);
4062 			ASSERT(vstate->dtvs_locals != NULL);
4063 
4064 			svar = vstate->dtvs_locals[id];
4065 			ASSERT(svar != NULL);
4066 			v = &svar->dtsv_var;
4067 
4068 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
4069 				uintptr_t a = (uintptr_t)svar->dtsv_data;
4070 				size_t sz = v->dtdv_type.dtdt_size;
4071 
4072 				sz += sizeof (uint64_t);
4073 				ASSERT(svar->dtsv_size == NCPU * sz);
4074 				a += CPU->cpu_id * sz;
4075 
4076 				if (*(uint8_t *)a == UINT8_MAX) {
4077 					/*
4078 					 * If the 0th byte is set to UINT8_MAX
4079 					 * then this is to be treated as a
4080 					 * reference to a NULL variable.
4081 					 */
4082 					regs[rd] = NULL;
4083 				} else {
4084 					regs[rd] = a + sizeof (uint64_t);
4085 				}
4086 
4087 				break;
4088 			}
4089 
4090 			ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
4091 			tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
4092 			regs[rd] = tmp[CPU->cpu_id];
4093 			break;
4094 
4095 		case DIF_OP_STLS:
4096 			id = DIF_INSTR_VAR(instr);
4097 
4098 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
4099 			id -= DIF_VAR_OTHER_UBASE;
4100 			ASSERT(id < vstate->dtvs_nlocals);
4101 
4102 			ASSERT(vstate->dtvs_locals != NULL);
4103 			svar = vstate->dtvs_locals[id];
4104 			ASSERT(svar != NULL);
4105 			v = &svar->dtsv_var;
4106 
4107 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
4108 				uintptr_t a = (uintptr_t)svar->dtsv_data;
4109 				size_t sz = v->dtdv_type.dtdt_size;
4110 
4111 				sz += sizeof (uint64_t);
4112 				ASSERT(svar->dtsv_size == NCPU * sz);
4113 				a += CPU->cpu_id * sz;
4114 
4115 				if (regs[rd] == NULL) {
4116 					*(uint8_t *)a = UINT8_MAX;
4117 					break;
4118 				} else {
4119 					*(uint8_t *)a = 0;
4120 					a += sizeof (uint64_t);
4121 				}
4122 
4123 				dtrace_vcopy((void *)(uintptr_t)regs[rd],
4124 				    (void *)a, &v->dtdv_type);
4125 				break;
4126 			}
4127 
4128 			ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
4129 			tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
4130 			tmp[CPU->cpu_id] = regs[rd];
4131 			break;
4132 
4133 		case DIF_OP_LDTS: {
4134 			dtrace_dynvar_t *dvar;
4135 			dtrace_key_t *key;
4136 
4137 			id = DIF_INSTR_VAR(instr);
4138 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
4139 			id -= DIF_VAR_OTHER_UBASE;
4140 			v = &vstate->dtvs_tlocals[id];
4141 
4142 			key = &tupregs[DIF_DTR_NREGS];
4143 			key[0].dttk_value = (uint64_t)id;
4144 			key[0].dttk_size = 0;
4145 			DTRACE_TLS_THRKEY(key[1].dttk_value);
4146 			key[1].dttk_size = 0;
4147 
4148 			dvar = dtrace_dynvar(dstate, 2, key,
4149 			    sizeof (uint64_t), DTRACE_DYNVAR_NOALLOC);
4150 
4151 			if (dvar == NULL) {
4152 				regs[rd] = 0;
4153 				break;
4154 			}
4155 
4156 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
4157 				regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
4158 			} else {
4159 				regs[rd] = *((uint64_t *)dvar->dtdv_data);
4160 			}
4161 
4162 			break;
4163 		}
4164 
4165 		case DIF_OP_STTS: {
4166 			dtrace_dynvar_t *dvar;
4167 			dtrace_key_t *key;
4168 
4169 			id = DIF_INSTR_VAR(instr);
4170 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
4171 			id -= DIF_VAR_OTHER_UBASE;
4172 
4173 			key = &tupregs[DIF_DTR_NREGS];
4174 			key[0].dttk_value = (uint64_t)id;
4175 			key[0].dttk_size = 0;
4176 			DTRACE_TLS_THRKEY(key[1].dttk_value);
4177 			key[1].dttk_size = 0;
4178 			v = &vstate->dtvs_tlocals[id];
4179 
4180 			dvar = dtrace_dynvar(dstate, 2, key,
4181 			    v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
4182 			    v->dtdv_type.dtdt_size : sizeof (uint64_t),
4183 			    regs[rd] ? DTRACE_DYNVAR_ALLOC :
4184 			    DTRACE_DYNVAR_DEALLOC);
4185 
4186 			/*
4187 			 * Given that we're storing to thread-local data,
4188 			 * we need to flush our predicate cache.
4189 			 */
4190 			curthread->t_predcache = NULL;
4191 
4192 			if (dvar == NULL)
4193 				break;
4194 
4195 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
4196 				dtrace_vcopy((void *)(uintptr_t)regs[rd],
4197 				    dvar->dtdv_data, &v->dtdv_type);
4198 			} else {
4199 				*((uint64_t *)dvar->dtdv_data) = regs[rd];
4200 			}
4201 
4202 			break;
4203 		}
4204 
4205 		case DIF_OP_SRA:
4206 			regs[rd] = (int64_t)regs[r1] >> regs[r2];
4207 			break;
4208 
4209 		case DIF_OP_CALL:
4210 			dtrace_dif_subr(DIF_INSTR_SUBR(instr), rd,
4211 			    regs, tupregs, ttop, mstate, state);
4212 			break;
4213 
4214 		case DIF_OP_PUSHTR:
4215 			if (ttop == DIF_DTR_NREGS) {
4216 				*flags |= CPU_DTRACE_TUPOFLOW;
4217 				break;
4218 			}
4219 
4220 			if (r1 == DIF_TYPE_STRING) {
4221 				/*
4222 				 * If this is a string type and the size is 0,
4223 				 * we'll use the system-wide default string
4224 				 * size.  Note that we are _not_ looking at
4225 				 * the value of the DTRACEOPT_STRSIZE option;
4226 				 * had this been set, we would expect to have
4227 				 * a non-zero size value in the "pushtr".
4228 				 */
4229 				tupregs[ttop].dttk_size =
4230 				    dtrace_strlen((char *)(uintptr_t)regs[rd],
4231 				    regs[r2] ? regs[r2] :
4232 				    dtrace_strsize_default) + 1;
4233 			} else {
4234 				tupregs[ttop].dttk_size = regs[r2];
4235 			}
4236 
4237 			tupregs[ttop++].dttk_value = regs[rd];
4238 			break;
4239 
4240 		case DIF_OP_PUSHTV:
4241 			if (ttop == DIF_DTR_NREGS) {
4242 				*flags |= CPU_DTRACE_TUPOFLOW;
4243 				break;
4244 			}
4245 
4246 			tupregs[ttop].dttk_value = regs[rd];
4247 			tupregs[ttop++].dttk_size = 0;
4248 			break;
4249 
4250 		case DIF_OP_POPTS:
4251 			if (ttop != 0)
4252 				ttop--;
4253 			break;
4254 
4255 		case DIF_OP_FLUSHTS:
4256 			ttop = 0;
4257 			break;
4258 
4259 		case DIF_OP_LDGAA:
4260 		case DIF_OP_LDTAA: {
4261 			dtrace_dynvar_t *dvar;
4262 			dtrace_key_t *key = tupregs;
4263 			uint_t nkeys = ttop;
4264 
4265 			id = DIF_INSTR_VAR(instr);
4266 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
4267 			id -= DIF_VAR_OTHER_UBASE;
4268 
4269 			key[nkeys].dttk_value = (uint64_t)id;
4270 			key[nkeys++].dttk_size = 0;
4271 
4272 			if (DIF_INSTR_OP(instr) == DIF_OP_LDTAA) {
4273 				DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
4274 				key[nkeys++].dttk_size = 0;
4275 				v = &vstate->dtvs_tlocals[id];
4276 			} else {
4277 				v = &vstate->dtvs_globals[id]->dtsv_var;
4278 			}
4279 
4280 			dvar = dtrace_dynvar(dstate, nkeys, key,
4281 			    v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
4282 			    v->dtdv_type.dtdt_size : sizeof (uint64_t),
4283 			    DTRACE_DYNVAR_NOALLOC);
4284 
4285 			if (dvar == NULL) {
4286 				regs[rd] = 0;
4287 				break;
4288 			}
4289 
4290 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
4291 				regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
4292 			} else {
4293 				regs[rd] = *((uint64_t *)dvar->dtdv_data);
4294 			}
4295 
4296 			break;
4297 		}
4298 
4299 		case DIF_OP_STGAA:
4300 		case DIF_OP_STTAA: {
4301 			dtrace_dynvar_t *dvar;
4302 			dtrace_key_t *key = tupregs;
4303 			uint_t nkeys = ttop;
4304 
4305 			id = DIF_INSTR_VAR(instr);
4306 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
4307 			id -= DIF_VAR_OTHER_UBASE;
4308 
4309 			key[nkeys].dttk_value = (uint64_t)id;
4310 			key[nkeys++].dttk_size = 0;
4311 
4312 			if (DIF_INSTR_OP(instr) == DIF_OP_STTAA) {
4313 				DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
4314 				key[nkeys++].dttk_size = 0;
4315 				v = &vstate->dtvs_tlocals[id];
4316 			} else {
4317 				v = &vstate->dtvs_globals[id]->dtsv_var;
4318 			}
4319 
4320 			dvar = dtrace_dynvar(dstate, nkeys, key,
4321 			    v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
4322 			    v->dtdv_type.dtdt_size : sizeof (uint64_t),
4323 			    regs[rd] ? DTRACE_DYNVAR_ALLOC :
4324 			    DTRACE_DYNVAR_DEALLOC);
4325 
4326 			if (dvar == NULL)
4327 				break;
4328 
4329 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
4330 				dtrace_vcopy((void *)(uintptr_t)regs[rd],
4331 				    dvar->dtdv_data, &v->dtdv_type);
4332 			} else {
4333 				*((uint64_t *)dvar->dtdv_data) = regs[rd];
4334 			}
4335 
4336 			break;
4337 		}
4338 
4339 		case DIF_OP_ALLOCS: {
4340 			uintptr_t ptr = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
4341 			size_t size = ptr - mstate->dtms_scratch_ptr + regs[r1];
4342 
4343 			if (mstate->dtms_scratch_ptr + size >
4344 			    mstate->dtms_scratch_base +
4345 			    mstate->dtms_scratch_size) {
4346 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4347 				regs[rd] = NULL;
4348 			} else {
4349 				dtrace_bzero((void *)
4350 				    mstate->dtms_scratch_ptr, size);
4351 				mstate->dtms_scratch_ptr += size;
4352 				regs[rd] = ptr;
4353 			}
4354 			break;
4355 		}
4356 
4357 		case DIF_OP_COPYS:
4358 			if (!dtrace_canstore(regs[rd], regs[r2],
4359 			    mstate, vstate)) {
4360 				*flags |= CPU_DTRACE_BADADDR;
4361 				*illval = regs[rd];
4362 				break;
4363 			}
4364 
4365 			dtrace_bcopy((void *)(uintptr_t)regs[r1],
4366 			    (void *)(uintptr_t)regs[rd], (size_t)regs[r2]);
4367 			break;
4368 
4369 		case DIF_OP_STB:
4370 			if (!dtrace_canstore(regs[rd], 1, mstate, vstate)) {
4371 				*flags |= CPU_DTRACE_BADADDR;
4372 				*illval = regs[rd];
4373 				break;
4374 			}
4375 			*((uint8_t *)(uintptr_t)regs[rd]) = (uint8_t)regs[r1];
4376 			break;
4377 
4378 		case DIF_OP_STH:
4379 			if (!dtrace_canstore(regs[rd], 2, mstate, vstate)) {
4380 				*flags |= CPU_DTRACE_BADADDR;
4381 				*illval = regs[rd];
4382 				break;
4383 			}
4384 			if (regs[rd] & 1) {
4385 				*flags |= CPU_DTRACE_BADALIGN;
4386 				*illval = regs[rd];
4387 				break;
4388 			}
4389 			*((uint16_t *)(uintptr_t)regs[rd]) = (uint16_t)regs[r1];
4390 			break;
4391 
4392 		case DIF_OP_STW:
4393 			if (!dtrace_canstore(regs[rd], 4, mstate, vstate)) {
4394 				*flags |= CPU_DTRACE_BADADDR;
4395 				*illval = regs[rd];
4396 				break;
4397 			}
4398 			if (regs[rd] & 3) {
4399 				*flags |= CPU_DTRACE_BADALIGN;
4400 				*illval = regs[rd];
4401 				break;
4402 			}
4403 			*((uint32_t *)(uintptr_t)regs[rd]) = (uint32_t)regs[r1];
4404 			break;
4405 
4406 		case DIF_OP_STX:
4407 			if (!dtrace_canstore(regs[rd], 8, mstate, vstate)) {
4408 				*flags |= CPU_DTRACE_BADADDR;
4409 				*illval = regs[rd];
4410 				break;
4411 			}
4412 			if (regs[rd] & 7) {
4413 				*flags |= CPU_DTRACE_BADALIGN;
4414 				*illval = regs[rd];
4415 				break;
4416 			}
4417 			*((uint64_t *)(uintptr_t)regs[rd]) = regs[r1];
4418 			break;
4419 		}
4420 	}
4421 
4422 	if (!(*flags & CPU_DTRACE_FAULT))
4423 		return (rval);
4424 
4425 	mstate->dtms_fltoffs = opc * sizeof (dif_instr_t);
4426 	mstate->dtms_present |= DTRACE_MSTATE_FLTOFFS;
4427 
4428 	return (0);
4429 }
4430 
4431 static void
4432 dtrace_action_breakpoint(dtrace_ecb_t *ecb)
4433 {
4434 	dtrace_probe_t *probe = ecb->dte_probe;
4435 	dtrace_provider_t *prov = probe->dtpr_provider;
4436 	char c[DTRACE_FULLNAMELEN + 80], *str;
4437 	char *msg = "dtrace: breakpoint action at probe ";
4438 	char *ecbmsg = " (ecb ";
4439 	uintptr_t mask = (0xf << (sizeof (uintptr_t) * NBBY / 4));
4440 	uintptr_t val = (uintptr_t)ecb;
4441 	int shift = (sizeof (uintptr_t) * NBBY) - 4, i = 0;
4442 
4443 	if (dtrace_destructive_disallow)
4444 		return;
4445 
4446 	/*
4447 	 * It's impossible to be taking action on the NULL probe.
4448 	 */
4449 	ASSERT(probe != NULL);
4450 
4451 	/*
4452 	 * This is a poor man's (destitute man's?) sprintf():  we want to
4453 	 * print the provider name, module name, function name and name of
4454 	 * the probe, along with the hex address of the ECB with the breakpoint
4455 	 * action -- all of which we must place in the character buffer by
4456 	 * hand.
4457 	 */
4458 	while (*msg != '\0')
4459 		c[i++] = *msg++;
4460 
4461 	for (str = prov->dtpv_name; *str != '\0'; str++)
4462 		c[i++] = *str;
4463 	c[i++] = ':';
4464 
4465 	for (str = probe->dtpr_mod; *str != '\0'; str++)
4466 		c[i++] = *str;
4467 	c[i++] = ':';
4468 
4469 	for (str = probe->dtpr_func; *str != '\0'; str++)
4470 		c[i++] = *str;
4471 	c[i++] = ':';
4472 
4473 	for (str = probe->dtpr_name; *str != '\0'; str++)
4474 		c[i++] = *str;
4475 
4476 	while (*ecbmsg != '\0')
4477 		c[i++] = *ecbmsg++;
4478 
4479 	while (shift >= 0) {
4480 		mask = (uintptr_t)0xf << shift;
4481 
4482 		if (val >= ((uintptr_t)1 << shift))
4483 			c[i++] = "0123456789abcdef"[(val & mask) >> shift];
4484 		shift -= 4;
4485 	}
4486 
4487 	c[i++] = ')';
4488 	c[i] = '\0';
4489 
4490 	debug_enter(c);
4491 }
4492 
4493 static void
4494 dtrace_action_panic(dtrace_ecb_t *ecb)
4495 {
4496 	dtrace_probe_t *probe = ecb->dte_probe;
4497 
4498 	/*
4499 	 * It's impossible to be taking action on the NULL probe.
4500 	 */
4501 	ASSERT(probe != NULL);
4502 
4503 	if (dtrace_destructive_disallow)
4504 		return;
4505 
4506 	if (dtrace_panicked != NULL)
4507 		return;
4508 
4509 	if (dtrace_casptr(&dtrace_panicked, NULL, curthread) != NULL)
4510 		return;
4511 
4512 	/*
4513 	 * We won the right to panic.  (We want to be sure that only one
4514 	 * thread calls panic() from dtrace_probe(), and that panic() is
4515 	 * called exactly once.)
4516 	 */
4517 	dtrace_panic("dtrace: panic action at probe %s:%s:%s:%s (ecb %p)",
4518 	    probe->dtpr_provider->dtpv_name, probe->dtpr_mod,
4519 	    probe->dtpr_func, probe->dtpr_name, (void *)ecb);
4520 }
4521 
4522 static void
4523 dtrace_action_raise(uint64_t sig)
4524 {
4525 	if (dtrace_destructive_disallow)
4526 		return;
4527 
4528 	if (sig >= NSIG) {
4529 		DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
4530 		return;
4531 	}
4532 
4533 	/*
4534 	 * raise() has a queue depth of 1 -- we ignore all subsequent
4535 	 * invocations of the raise() action.
4536 	 */
4537 	if (curthread->t_dtrace_sig == 0)
4538 		curthread->t_dtrace_sig = (uint8_t)sig;
4539 
4540 	curthread->t_sig_check = 1;
4541 	aston(curthread);
4542 }
4543 
4544 static void
4545 dtrace_action_stop(void)
4546 {
4547 	if (dtrace_destructive_disallow)
4548 		return;
4549 
4550 	if (!curthread->t_dtrace_stop) {
4551 		curthread->t_dtrace_stop = 1;
4552 		curthread->t_sig_check = 1;
4553 		aston(curthread);
4554 	}
4555 }
4556 
4557 static void
4558 dtrace_action_chill(dtrace_mstate_t *mstate, hrtime_t val)
4559 {
4560 	hrtime_t now;
4561 	volatile uint16_t *flags;
4562 	cpu_t *cpu = CPU;
4563 
4564 	if (dtrace_destructive_disallow)
4565 		return;
4566 
4567 	flags = (volatile uint16_t *)&cpu_core[cpu->cpu_id].cpuc_dtrace_flags;
4568 
4569 	now = dtrace_gethrtime();
4570 
4571 	if (now - cpu->cpu_dtrace_chillmark > dtrace_chill_interval) {
4572 		/*
4573 		 * We need to advance the mark to the current time.
4574 		 */
4575 		cpu->cpu_dtrace_chillmark = now;
4576 		cpu->cpu_dtrace_chilled = 0;
4577 	}
4578 
4579 	/*
4580 	 * Now check to see if the requested chill time would take us over
4581 	 * the maximum amount of time allowed in the chill interval.  (Or
4582 	 * worse, if the calculation itself induces overflow.)
4583 	 */
4584 	if (cpu->cpu_dtrace_chilled + val > dtrace_chill_max ||
4585 	    cpu->cpu_dtrace_chilled + val < cpu->cpu_dtrace_chilled) {
4586 		*flags |= CPU_DTRACE_ILLOP;
4587 		return;
4588 	}
4589 
4590 	while (dtrace_gethrtime() - now < val)
4591 		continue;
4592 
4593 	/*
4594 	 * Normally, we assure that the value of the variable "timestamp" does
4595 	 * not change within an ECB.  The presence of chill() represents an
4596 	 * exception to this rule, however.
4597 	 */
4598 	mstate->dtms_present &= ~DTRACE_MSTATE_TIMESTAMP;
4599 	cpu->cpu_dtrace_chilled += val;
4600 }
4601 
4602 static void
4603 dtrace_action_ustack(dtrace_mstate_t *mstate, dtrace_state_t *state,
4604     uint64_t *buf, uint64_t arg)
4605 {
4606 	int nframes = DTRACE_USTACK_NFRAMES(arg);
4607 	int strsize = DTRACE_USTACK_STRSIZE(arg);
4608 	uint64_t *pcs = &buf[1], *fps;
4609 	char *str = (char *)&pcs[nframes];
4610 	int size, offs = 0, i, j;
4611 	uintptr_t old = mstate->dtms_scratch_ptr, saved;
4612 	uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
4613 	char *sym;
4614 
4615 	/*
4616 	 * Should be taking a faster path if string space has not been
4617 	 * allocated.
4618 	 */
4619 	ASSERT(strsize != 0);
4620 
4621 	/*
4622 	 * We will first allocate some temporary space for the frame pointers.
4623 	 */
4624 	fps = (uint64_t *)P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
4625 	size = (uintptr_t)fps - mstate->dtms_scratch_ptr +
4626 	    (nframes * sizeof (uint64_t));
4627 
4628 	if (mstate->dtms_scratch_ptr + size >
4629 	    mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
4630 		/*
4631 		 * Not enough room for our frame pointers -- need to indicate
4632 		 * that we ran out of scratch space.
4633 		 */
4634 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4635 		return;
4636 	}
4637 
4638 	mstate->dtms_scratch_ptr += size;
4639 	saved = mstate->dtms_scratch_ptr;
4640 
4641 	/*
4642 	 * Now get a stack with both program counters and frame pointers.
4643 	 */
4644 	DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4645 	dtrace_getufpstack(buf, fps, nframes + 1);
4646 	DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4647 
4648 	/*
4649 	 * If that faulted, we're cooked.
4650 	 */
4651 	if (*flags & CPU_DTRACE_FAULT)
4652 		goto out;
4653 
4654 	/*
4655 	 * Now we want to walk up the stack, calling the USTACK helper.  For
4656 	 * each iteration, we restore the scratch pointer.
4657 	 */
4658 	for (i = 0; i < nframes; i++) {
4659 		mstate->dtms_scratch_ptr = saved;
4660 
4661 		if (offs >= strsize)
4662 			break;
4663 
4664 		sym = (char *)(uintptr_t)dtrace_helper(
4665 		    DTRACE_HELPER_ACTION_USTACK,
4666 		    mstate, state, pcs[i], fps[i]);
4667 
4668 		/*
4669 		 * If we faulted while running the helper, we're going to
4670 		 * clear the fault and null out the corresponding string.
4671 		 */
4672 		if (*flags & CPU_DTRACE_FAULT) {
4673 			*flags &= ~CPU_DTRACE_FAULT;
4674 			str[offs++] = '\0';
4675 			continue;
4676 		}
4677 
4678 		if (sym == NULL) {
4679 			str[offs++] = '\0';
4680 			continue;
4681 		}
4682 
4683 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4684 
4685 		/*
4686 		 * Now copy in the string that the helper returned to us.
4687 		 */
4688 		for (j = 0; offs + j < strsize; j++) {
4689 			if ((str[offs + j] = sym[j]) == '\0')
4690 				break;
4691 		}
4692 
4693 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4694 
4695 		offs += j + 1;
4696 	}
4697 
4698 	if (offs >= strsize) {
4699 		/*
4700 		 * If we didn't have room for all of the strings, we don't
4701 		 * abort processing -- this needn't be a fatal error -- but we
4702 		 * still want to increment a counter (dts_stkstroverflows) to
4703 		 * allow this condition to be warned about.  (If this is from
4704 		 * a jstack() action, it is easily tuned via jstackstrsize.)
4705 		 */
4706 		dtrace_error(&state->dts_stkstroverflows);
4707 	}
4708 
4709 	while (offs < strsize)
4710 		str[offs++] = '\0';
4711 
4712 out:
4713 	mstate->dtms_scratch_ptr = old;
4714 }
4715 
4716 /*
4717  * If you're looking for the epicenter of DTrace, you just found it.  This
4718  * is the function called by the provider to fire a probe -- from which all
4719  * subsequent probe-context DTrace activity emanates.
4720  */
4721 void
4722 dtrace_probe(dtrace_id_t id, uintptr_t arg0, uintptr_t arg1,
4723     uintptr_t arg2, uintptr_t arg3, uintptr_t arg4)
4724 {
4725 	processorid_t cpuid;
4726 	dtrace_icookie_t cookie;
4727 	dtrace_probe_t *probe;
4728 	dtrace_mstate_t mstate;
4729 	dtrace_ecb_t *ecb;
4730 	dtrace_action_t *act;
4731 	intptr_t offs;
4732 	size_t size;
4733 	int vtime, onintr;
4734 	volatile uint16_t *flags;
4735 	hrtime_t now;
4736 
4737 	/*
4738 	 * Kick out immediately if this CPU is still being born (in which case
4739 	 * curthread will be set to -1)
4740 	 */
4741 	if ((uintptr_t)curthread & 1)
4742 		return;
4743 
4744 	cookie = dtrace_interrupt_disable();
4745 	probe = dtrace_probes[id - 1];
4746 	cpuid = CPU->cpu_id;
4747 	onintr = CPU_ON_INTR(CPU);
4748 
4749 	if (!onintr && probe->dtpr_predcache != DTRACE_CACHEIDNONE &&
4750 	    probe->dtpr_predcache == curthread->t_predcache) {
4751 		/*
4752 		 * We have hit in the predicate cache; we know that
4753 		 * this predicate would evaluate to be false.
4754 		 */
4755 		dtrace_interrupt_enable(cookie);
4756 		return;
4757 	}
4758 
4759 	if (panic_quiesce) {
4760 		/*
4761 		 * We don't trace anything if we're panicking.
4762 		 */
4763 		dtrace_interrupt_enable(cookie);
4764 		return;
4765 	}
4766 
4767 	now = dtrace_gethrtime();
4768 	vtime = dtrace_vtime_references != 0;
4769 
4770 	if (vtime && curthread->t_dtrace_start)
4771 		curthread->t_dtrace_vtime += now - curthread->t_dtrace_start;
4772 
4773 	mstate.dtms_probe = probe;
4774 	mstate.dtms_arg[0] = arg0;
4775 	mstate.dtms_arg[1] = arg1;
4776 	mstate.dtms_arg[2] = arg2;
4777 	mstate.dtms_arg[3] = arg3;
4778 	mstate.dtms_arg[4] = arg4;
4779 
4780 	flags = (volatile uint16_t *)&cpu_core[cpuid].cpuc_dtrace_flags;
4781 
4782 	for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
4783 		dtrace_predicate_t *pred = ecb->dte_predicate;
4784 		dtrace_state_t *state = ecb->dte_state;
4785 		dtrace_buffer_t *buf = &state->dts_buffer[cpuid];
4786 		dtrace_buffer_t *aggbuf = &state->dts_aggbuffer[cpuid];
4787 		dtrace_vstate_t *vstate = &state->dts_vstate;
4788 		dtrace_provider_t *prov = probe->dtpr_provider;
4789 		int committed = 0;
4790 		caddr_t tomax;
4791 
4792 		/*
4793 		 * A little subtlety with the following (seemingly innocuous)
4794 		 * declaration of the automatic 'val':  by looking at the
4795 		 * code, you might think that it could be declared in the
4796 		 * action processing loop, below.  (That is, it's only used in
4797 		 * the action processing loop.)  However, it must be declared
4798 		 * out of that scope because in the case of DIF expression
4799 		 * arguments to aggregating actions, one iteration of the
4800 		 * action loop will use the last iteration's value.
4801 		 */
4802 #ifdef lint
4803 		uint64_t val = 0;
4804 #else
4805 		uint64_t val;
4806 #endif
4807 
4808 		mstate.dtms_present = DTRACE_MSTATE_ARGS | DTRACE_MSTATE_PROBE;
4809 		*flags &= ~CPU_DTRACE_ERROR;
4810 
4811 		if (prov == dtrace_provider) {
4812 			/*
4813 			 * If dtrace itself is the provider of this probe,
4814 			 * we're only going to continue processing the ECB if
4815 			 * arg0 (the dtrace_state_t) is equal to the ECB's
4816 			 * creating state.  (This prevents disjoint consumers
4817 			 * from seeing one another's metaprobes.)
4818 			 */
4819 			if (arg0 != (uint64_t)(uintptr_t)state)
4820 				continue;
4821 		}
4822 
4823 		if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE) {
4824 			/*
4825 			 * We're not currently active.  If our provider isn't
4826 			 * the dtrace pseudo provider, we're not interested.
4827 			 */
4828 			if (prov != dtrace_provider)
4829 				continue;
4830 
4831 			/*
4832 			 * Now we must further check if we are in the BEGIN
4833 			 * probe.  If we are, we will only continue processing
4834 			 * if we're still in WARMUP -- if one BEGIN enabling
4835 			 * has invoked the exit() action, we don't want to
4836 			 * evaluate subsequent BEGIN enablings.
4837 			 */
4838 			if (probe->dtpr_id == dtrace_probeid_begin &&
4839 			    state->dts_activity != DTRACE_ACTIVITY_WARMUP) {
4840 				ASSERT(state->dts_activity ==
4841 				    DTRACE_ACTIVITY_DRAINING);
4842 				continue;
4843 			}
4844 		}
4845 
4846 		if (ecb->dte_cond) {
4847 			/*
4848 			 * If the dte_cond bits indicate that this
4849 			 * consumer is only allowed to see user-mode firings
4850 			 * of this probe, call the provider's dtps_usermode()
4851 			 * entry point to check that the probe was fired
4852 			 * while in a user context. Skip this ECB if that's
4853 			 * not the case.
4854 			 */
4855 			if ((ecb->dte_cond & DTRACE_COND_USERMODE) &&
4856 			    prov->dtpv_pops.dtps_usermode(prov->dtpv_arg,
4857 			    probe->dtpr_id, probe->dtpr_arg) == 0)
4858 				continue;
4859 
4860 			/*
4861 			 * This is more subtle than it looks. We have to be
4862 			 * absolutely certain that CRED() isn't going to
4863 			 * change out from under us so it's only legit to
4864 			 * examine that structure if we're in constrained
4865 			 * situations. Currently, the only times we'll this
4866 			 * check is if a non-super-user has enabled the
4867 			 * profile or syscall providers -- providers that
4868 			 * allow visibility of all processes. For the
4869 			 * profile case, the check above will ensure that
4870 			 * we're examining a user context.
4871 			 */
4872 			if (ecb->dte_cond & DTRACE_COND_OWNER) {
4873 				cred_t *cr;
4874 				cred_t *s_cr =
4875 				    ecb->dte_state->dts_cred.dcr_cred;
4876 				proc_t *proc;
4877 
4878 				ASSERT(s_cr != NULL);
4879 
4880 				if ((cr = CRED()) == NULL ||
4881 				    s_cr->cr_uid != cr->cr_uid ||
4882 				    s_cr->cr_uid != cr->cr_ruid ||
4883 				    s_cr->cr_uid != cr->cr_suid ||
4884 				    s_cr->cr_gid != cr->cr_gid ||
4885 				    s_cr->cr_gid != cr->cr_rgid ||
4886 				    s_cr->cr_gid != cr->cr_sgid ||
4887 				    (proc = ttoproc(curthread)) == NULL ||
4888 				    (proc->p_flag & SNOCD))
4889 					continue;
4890 			}
4891 
4892 			if (ecb->dte_cond & DTRACE_COND_ZONEOWNER) {
4893 				cred_t *cr;
4894 				cred_t *s_cr =
4895 				    ecb->dte_state->dts_cred.dcr_cred;
4896 
4897 				ASSERT(s_cr != NULL);
4898 
4899 				if ((cr = CRED()) == NULL ||
4900 				    s_cr->cr_zone->zone_id !=
4901 				    cr->cr_zone->zone_id)
4902 					continue;
4903 			}
4904 		}
4905 
4906 		if (now - state->dts_alive > dtrace_deadman_timeout) {
4907 			/*
4908 			 * We seem to be dead.  Unless we (a) have kernel
4909 			 * destructive permissions (b) have expicitly enabled
4910 			 * destructive actions and (c) destructive actions have
4911 			 * not been disabled, we're going to transition into
4912 			 * the KILLED state, from which no further processing
4913 			 * on this state will be performed.
4914 			 */
4915 			if (!dtrace_priv_kernel_destructive(state) ||
4916 			    !state->dts_cred.dcr_destructive ||
4917 			    dtrace_destructive_disallow) {
4918 				void *activity = &state->dts_activity;
4919 				dtrace_activity_t current;
4920 
4921 				do {
4922 					current = state->dts_activity;
4923 				} while (dtrace_cas32(activity, current,
4924 				    DTRACE_ACTIVITY_KILLED) != current);
4925 
4926 				continue;
4927 			}
4928 		}
4929 
4930 		if ((offs = dtrace_buffer_reserve(buf, ecb->dte_needed,
4931 		    ecb->dte_alignment, state, &mstate)) < 0)
4932 			continue;
4933 
4934 		tomax = buf->dtb_tomax;
4935 		ASSERT(tomax != NULL);
4936 
4937 		if (ecb->dte_size != 0)
4938 			DTRACE_STORE(uint32_t, tomax, offs, ecb->dte_epid);
4939 
4940 		mstate.dtms_epid = ecb->dte_epid;
4941 		mstate.dtms_present |= DTRACE_MSTATE_EPID;
4942 
4943 		if (pred != NULL) {
4944 			dtrace_difo_t *dp = pred->dtp_difo;
4945 			int rval;
4946 
4947 			rval = dtrace_dif_emulate(dp, &mstate, vstate, state);
4948 
4949 			if (!(*flags & CPU_DTRACE_ERROR) && !rval) {
4950 				dtrace_cacheid_t cid = probe->dtpr_predcache;
4951 
4952 				if (cid != DTRACE_CACHEIDNONE && !onintr) {
4953 					/*
4954 					 * Update the predicate cache...
4955 					 */
4956 					ASSERT(cid == pred->dtp_cacheid);
4957 					curthread->t_predcache = cid;
4958 				}
4959 
4960 				continue;
4961 			}
4962 		}
4963 
4964 		for (act = ecb->dte_action; !(*flags & CPU_DTRACE_ERROR) &&
4965 		    act != NULL; act = act->dta_next) {
4966 			size_t valoffs;
4967 			dtrace_difo_t *dp;
4968 			dtrace_recdesc_t *rec = &act->dta_rec;
4969 
4970 			size = rec->dtrd_size;
4971 			valoffs = offs + rec->dtrd_offset;
4972 
4973 			if (DTRACEACT_ISAGG(act->dta_kind)) {
4974 				uint64_t v = 0xbad;
4975 				dtrace_aggregation_t *agg;
4976 
4977 				agg = (dtrace_aggregation_t *)act;
4978 
4979 				if ((dp = act->dta_difo) != NULL)
4980 					v = dtrace_dif_emulate(dp,
4981 					    &mstate, vstate, state);
4982 
4983 				if (*flags & CPU_DTRACE_ERROR)
4984 					continue;
4985 
4986 				/*
4987 				 * Note that we always pass the expression
4988 				 * value from the previous iteration of the
4989 				 * action loop.  This value will only be used
4990 				 * if there is an expression argument to the
4991 				 * aggregating action, denoted by the
4992 				 * dtag_hasarg field.
4993 				 */
4994 				dtrace_aggregate(agg, buf,
4995 				    offs, aggbuf, v, val);
4996 				continue;
4997 			}
4998 
4999 			switch (act->dta_kind) {
5000 			case DTRACEACT_STOP:
5001 				if (dtrace_priv_proc_destructive(state))
5002 					dtrace_action_stop();
5003 				continue;
5004 
5005 			case DTRACEACT_BREAKPOINT:
5006 				if (dtrace_priv_kernel_destructive(state))
5007 					dtrace_action_breakpoint(ecb);
5008 				continue;
5009 
5010 			case DTRACEACT_PANIC:
5011 				if (dtrace_priv_kernel_destructive(state))
5012 					dtrace_action_panic(ecb);
5013 				continue;
5014 
5015 			case DTRACEACT_STACK:
5016 				if (!dtrace_priv_kernel(state))
5017 					continue;
5018 
5019 				dtrace_getpcstack((pc_t *)(tomax + valoffs),
5020 				    size / sizeof (pc_t), probe->dtpr_aframes,
5021 				    DTRACE_ANCHORED(probe) ? NULL :
5022 				    (uint32_t *)arg0);
5023 
5024 				continue;
5025 
5026 			case DTRACEACT_JSTACK:
5027 			case DTRACEACT_USTACK:
5028 				if (!dtrace_priv_proc(state))
5029 					continue;
5030 
5031 				/*
5032 				 * See comment in DIF_VAR_PID.
5033 				 */
5034 				if (DTRACE_ANCHORED(mstate.dtms_probe) &&
5035 				    CPU_ON_INTR(CPU)) {
5036 					int depth = DTRACE_USTACK_NFRAMES(
5037 					    rec->dtrd_arg) + 1;
5038 
5039 					dtrace_bzero((void *)(tomax + valoffs),
5040 					    DTRACE_USTACK_STRSIZE(rec->dtrd_arg)
5041 					    + depth * sizeof (uint64_t));
5042 
5043 					continue;
5044 				}
5045 
5046 				if (DTRACE_USTACK_STRSIZE(rec->dtrd_arg) != 0 &&
5047 				    curproc->p_dtrace_helpers != NULL) {
5048 					/*
5049 					 * This is the slow path -- we have
5050 					 * allocated string space, and we're
5051 					 * getting the stack of a process that
5052 					 * has helpers.  Call into a separate
5053 					 * routine to perform this processing.
5054 					 */
5055 					dtrace_action_ustack(&mstate, state,
5056 					    (uint64_t *)(tomax + valoffs),
5057 					    rec->dtrd_arg);
5058 					continue;
5059 				}
5060 
5061 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5062 				dtrace_getupcstack((uint64_t *)
5063 				    (tomax + valoffs),
5064 				    DTRACE_USTACK_NFRAMES(rec->dtrd_arg) + 1);
5065 				DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5066 				continue;
5067 
5068 			default:
5069 				break;
5070 			}
5071 
5072 			dp = act->dta_difo;
5073 			ASSERT(dp != NULL);
5074 
5075 			val = dtrace_dif_emulate(dp, &mstate, vstate, state);
5076 
5077 			if (*flags & CPU_DTRACE_ERROR)
5078 				continue;
5079 
5080 			switch (act->dta_kind) {
5081 			case DTRACEACT_SPECULATE:
5082 				ASSERT(buf == &state->dts_buffer[cpuid]);
5083 				buf = dtrace_speculation_buffer(state,
5084 				    cpuid, val);
5085 
5086 				if (buf == NULL) {
5087 					*flags |= CPU_DTRACE_DROP;
5088 					continue;
5089 				}
5090 
5091 				offs = dtrace_buffer_reserve(buf,
5092 				    ecb->dte_needed, ecb->dte_alignment,
5093 				    state, NULL);
5094 
5095 				if (offs < 0) {
5096 					*flags |= CPU_DTRACE_DROP;
5097 					continue;
5098 				}
5099 
5100 				tomax = buf->dtb_tomax;
5101 				ASSERT(tomax != NULL);
5102 
5103 				if (ecb->dte_size != 0)
5104 					DTRACE_STORE(uint32_t, tomax, offs,
5105 					    ecb->dte_epid);
5106 				continue;
5107 
5108 			case DTRACEACT_CHILL:
5109 				if (dtrace_priv_kernel_destructive(state))
5110 					dtrace_action_chill(&mstate, val);
5111 				continue;
5112 
5113 			case DTRACEACT_RAISE:
5114 				if (dtrace_priv_proc_destructive(state))
5115 					dtrace_action_raise(val);
5116 				continue;
5117 
5118 			case DTRACEACT_COMMIT:
5119 				ASSERT(!committed);
5120 
5121 				/*
5122 				 * We need to commit our buffer state.
5123 				 */
5124 				if (ecb->dte_size)
5125 					buf->dtb_offset = offs + ecb->dte_size;
5126 				buf = &state->dts_buffer[cpuid];
5127 				dtrace_speculation_commit(state, cpuid, val);
5128 				committed = 1;
5129 				continue;
5130 
5131 			case DTRACEACT_DISCARD:
5132 				dtrace_speculation_discard(state, cpuid, val);
5133 				continue;
5134 
5135 			case DTRACEACT_DIFEXPR:
5136 			case DTRACEACT_LIBACT:
5137 			case DTRACEACT_PRINTF:
5138 			case DTRACEACT_PRINTA:
5139 			case DTRACEACT_SYSTEM:
5140 			case DTRACEACT_FREOPEN:
5141 				break;
5142 
5143 			case DTRACEACT_SYM:
5144 			case DTRACEACT_MOD:
5145 				if (!dtrace_priv_kernel(state))
5146 					continue;
5147 				break;
5148 
5149 			case DTRACEACT_USYM:
5150 			case DTRACEACT_UMOD:
5151 			case DTRACEACT_UADDR: {
5152 				struct pid *pid = curthread->t_procp->p_pidp;
5153 
5154 				if (!dtrace_priv_proc(state))
5155 					continue;
5156 
5157 				DTRACE_STORE(uint64_t, tomax,
5158 				    valoffs, (uint64_t)pid->pid_id);
5159 				DTRACE_STORE(uint64_t, tomax,
5160 				    valoffs + sizeof (uint64_t), val);
5161 
5162 				continue;
5163 			}
5164 
5165 			case DTRACEACT_EXIT: {
5166 				/*
5167 				 * For the exit action, we are going to attempt
5168 				 * to atomically set our activity to be
5169 				 * draining.  If this fails (either because
5170 				 * another CPU has beat us to the exit action,
5171 				 * or because our current activity is something
5172 				 * other than ACTIVE or WARMUP), we will
5173 				 * continue.  This assures that the exit action
5174 				 * can be successfully recorded at most once
5175 				 * when we're in the ACTIVE state.  If we're
5176 				 * encountering the exit() action while in
5177 				 * COOLDOWN, however, we want to honor the new
5178 				 * status code.  (We know that we're the only
5179 				 * thread in COOLDOWN, so there is no race.)
5180 				 */
5181 				void *activity = &state->dts_activity;
5182 				dtrace_activity_t current = state->dts_activity;
5183 
5184 				if (current == DTRACE_ACTIVITY_COOLDOWN)
5185 					break;
5186 
5187 				if (current != DTRACE_ACTIVITY_WARMUP)
5188 					current = DTRACE_ACTIVITY_ACTIVE;
5189 
5190 				if (dtrace_cas32(activity, current,
5191 				    DTRACE_ACTIVITY_DRAINING) != current) {
5192 					*flags |= CPU_DTRACE_DROP;
5193 					continue;
5194 				}
5195 
5196 				break;
5197 			}
5198 
5199 			default:
5200 				ASSERT(0);
5201 			}
5202 
5203 			if (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF) {
5204 				uintptr_t end = valoffs + size;
5205 
5206 				/*
5207 				 * If this is a string, we're going to only
5208 				 * load until we find the zero byte -- after
5209 				 * which we'll store zero bytes.
5210 				 */
5211 				if (dp->dtdo_rtype.dtdt_kind ==
5212 				    DIF_TYPE_STRING) {
5213 					char c = '\0' + 1;
5214 					int intuple = act->dta_intuple;
5215 					size_t s;
5216 
5217 					for (s = 0; s < size; s++) {
5218 						if (c != '\0')
5219 							c = dtrace_load8(val++);
5220 
5221 						DTRACE_STORE(uint8_t, tomax,
5222 						    valoffs++, c);
5223 
5224 						if (c == '\0' && intuple)
5225 							break;
5226 					}
5227 
5228 					continue;
5229 				}
5230 
5231 				while (valoffs < end) {
5232 					DTRACE_STORE(uint8_t, tomax, valoffs++,
5233 					    dtrace_load8(val++));
5234 				}
5235 
5236 				continue;
5237 			}
5238 
5239 			switch (size) {
5240 			case 0:
5241 				break;
5242 
5243 			case sizeof (uint8_t):
5244 				DTRACE_STORE(uint8_t, tomax, valoffs, val);
5245 				break;
5246 			case sizeof (uint16_t):
5247 				DTRACE_STORE(uint16_t, tomax, valoffs, val);
5248 				break;
5249 			case sizeof (uint32_t):
5250 				DTRACE_STORE(uint32_t, tomax, valoffs, val);
5251 				break;
5252 			case sizeof (uint64_t):
5253 				DTRACE_STORE(uint64_t, tomax, valoffs, val);
5254 				break;
5255 			default:
5256 				/*
5257 				 * Any other size should have been returned by
5258 				 * reference, not by value.
5259 				 */
5260 				ASSERT(0);
5261 				break;
5262 			}
5263 		}
5264 
5265 		if (*flags & CPU_DTRACE_DROP)
5266 			continue;
5267 
5268 		if (*flags & CPU_DTRACE_FAULT) {
5269 			int ndx;
5270 			dtrace_action_t *err;
5271 
5272 			buf->dtb_errors++;
5273 
5274 			if (probe->dtpr_id == dtrace_probeid_error) {
5275 				/*
5276 				 * There's nothing we can do -- we had an
5277 				 * error on the error probe.  We bump an
5278 				 * error counter to at least indicate that
5279 				 * this condition happened.
5280 				 */
5281 				dtrace_error(&state->dts_dblerrors);
5282 				continue;
5283 			}
5284 
5285 			if (vtime) {
5286 				/*
5287 				 * Before recursing on dtrace_probe(), we
5288 				 * need to explicitly clear out our start
5289 				 * time to prevent it from being accumulated
5290 				 * into t_dtrace_vtime.
5291 				 */
5292 				curthread->t_dtrace_start = 0;
5293 			}
5294 
5295 			/*
5296 			 * Iterate over the actions to figure out which action
5297 			 * we were processing when we experienced the error.
5298 			 * Note that act points _past_ the faulting action; if
5299 			 * act is ecb->dte_action, the fault was in the
5300 			 * predicate, if it's ecb->dte_action->dta_next it's
5301 			 * in action #1, and so on.
5302 			 */
5303 			for (err = ecb->dte_action, ndx = 0;
5304 			    err != act; err = err->dta_next, ndx++)
5305 				continue;
5306 
5307 			dtrace_probe_error(state, ecb->dte_epid, ndx,
5308 			    (mstate.dtms_present & DTRACE_MSTATE_FLTOFFS) ?
5309 			    mstate.dtms_fltoffs : -1, DTRACE_FLAGS2FLT(*flags),
5310 			    cpu_core[cpuid].cpuc_dtrace_illval);
5311 
5312 			continue;
5313 		}
5314 
5315 		if (!committed)
5316 			buf->dtb_offset = offs + ecb->dte_size;
5317 	}
5318 
5319 	if (vtime)
5320 		curthread->t_dtrace_start = dtrace_gethrtime();
5321 
5322 	dtrace_interrupt_enable(cookie);
5323 }
5324 
5325 /*
5326  * DTrace Probe Hashing Functions
5327  *
5328  * The functions in this section (and indeed, the functions in remaining
5329  * sections) are not _called_ from probe context.  (Any exceptions to this are
5330  * marked with a "Note:".)  Rather, they are called from elsewhere in the
5331  * DTrace framework to look-up probes in, add probes to and remove probes from
5332  * the DTrace probe hashes.  (Each probe is hashed by each element of the
5333  * probe tuple -- allowing for fast lookups, regardless of what was
5334  * specified.)
5335  */
5336 static uint_t
5337 dtrace_hash_str(char *p)
5338 {
5339 	unsigned int g;
5340 	uint_t hval = 0;
5341 
5342 	while (*p) {
5343 		hval = (hval << 4) + *p++;
5344 		if ((g = (hval & 0xf0000000)) != 0)
5345 			hval ^= g >> 24;
5346 		hval &= ~g;
5347 	}
5348 	return (hval);
5349 }
5350 
5351 static dtrace_hash_t *
5352 dtrace_hash_create(uintptr_t stroffs, uintptr_t nextoffs, uintptr_t prevoffs)
5353 {
5354 	dtrace_hash_t *hash = kmem_zalloc(sizeof (dtrace_hash_t), KM_SLEEP);
5355 
5356 	hash->dth_stroffs = stroffs;
5357 	hash->dth_nextoffs = nextoffs;
5358 	hash->dth_prevoffs = prevoffs;
5359 
5360 	hash->dth_size = 1;
5361 	hash->dth_mask = hash->dth_size - 1;
5362 
5363 	hash->dth_tab = kmem_zalloc(hash->dth_size *
5364 	    sizeof (dtrace_hashbucket_t *), KM_SLEEP);
5365 
5366 	return (hash);
5367 }
5368 
5369 static void
5370 dtrace_hash_destroy(dtrace_hash_t *hash)
5371 {
5372 #ifdef DEBUG
5373 	int i;
5374 
5375 	for (i = 0; i < hash->dth_size; i++)
5376 		ASSERT(hash->dth_tab[i] == NULL);
5377 #endif
5378 
5379 	kmem_free(hash->dth_tab,
5380 	    hash->dth_size * sizeof (dtrace_hashbucket_t *));
5381 	kmem_free(hash, sizeof (dtrace_hash_t));
5382 }
5383 
5384 static void
5385 dtrace_hash_resize(dtrace_hash_t *hash)
5386 {
5387 	int size = hash->dth_size, i, ndx;
5388 	int new_size = hash->dth_size << 1;
5389 	int new_mask = new_size - 1;
5390 	dtrace_hashbucket_t **new_tab, *bucket, *next;
5391 
5392 	ASSERT((new_size & new_mask) == 0);
5393 
5394 	new_tab = kmem_zalloc(new_size * sizeof (void *), KM_SLEEP);
5395 
5396 	for (i = 0; i < size; i++) {
5397 		for (bucket = hash->dth_tab[i]; bucket != NULL; bucket = next) {
5398 			dtrace_probe_t *probe = bucket->dthb_chain;
5399 
5400 			ASSERT(probe != NULL);
5401 			ndx = DTRACE_HASHSTR(hash, probe) & new_mask;
5402 
5403 			next = bucket->dthb_next;
5404 			bucket->dthb_next = new_tab[ndx];
5405 			new_tab[ndx] = bucket;
5406 		}
5407 	}
5408 
5409 	kmem_free(hash->dth_tab, hash->dth_size * sizeof (void *));
5410 	hash->dth_tab = new_tab;
5411 	hash->dth_size = new_size;
5412 	hash->dth_mask = new_mask;
5413 }
5414 
5415 static void
5416 dtrace_hash_add(dtrace_hash_t *hash, dtrace_probe_t *new)
5417 {
5418 	int hashval = DTRACE_HASHSTR(hash, new);
5419 	int ndx = hashval & hash->dth_mask;
5420 	dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
5421 	dtrace_probe_t **nextp, **prevp;
5422 
5423 	for (; bucket != NULL; bucket = bucket->dthb_next) {
5424 		if (DTRACE_HASHEQ(hash, bucket->dthb_chain, new))
5425 			goto add;
5426 	}
5427 
5428 	if ((hash->dth_nbuckets >> 1) > hash->dth_size) {
5429 		dtrace_hash_resize(hash);
5430 		dtrace_hash_add(hash, new);
5431 		return;
5432 	}
5433 
5434 	bucket = kmem_zalloc(sizeof (dtrace_hashbucket_t), KM_SLEEP);
5435 	bucket->dthb_next = hash->dth_tab[ndx];
5436 	hash->dth_tab[ndx] = bucket;
5437 	hash->dth_nbuckets++;
5438 
5439 add:
5440 	nextp = DTRACE_HASHNEXT(hash, new);
5441 	ASSERT(*nextp == NULL && *(DTRACE_HASHPREV(hash, new)) == NULL);
5442 	*nextp = bucket->dthb_chain;
5443 
5444 	if (bucket->dthb_chain != NULL) {
5445 		prevp = DTRACE_HASHPREV(hash, bucket->dthb_chain);
5446 		ASSERT(*prevp == NULL);
5447 		*prevp = new;
5448 	}
5449 
5450 	bucket->dthb_chain = new;
5451 	bucket->dthb_len++;
5452 }
5453 
5454 static dtrace_probe_t *
5455 dtrace_hash_lookup(dtrace_hash_t *hash, dtrace_probe_t *template)
5456 {
5457 	int hashval = DTRACE_HASHSTR(hash, template);
5458 	int ndx = hashval & hash->dth_mask;
5459 	dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
5460 
5461 	for (; bucket != NULL; bucket = bucket->dthb_next) {
5462 		if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
5463 			return (bucket->dthb_chain);
5464 	}
5465 
5466 	return (NULL);
5467 }
5468 
5469 static int
5470 dtrace_hash_collisions(dtrace_hash_t *hash, dtrace_probe_t *template)
5471 {
5472 	int hashval = DTRACE_HASHSTR(hash, template);
5473 	int ndx = hashval & hash->dth_mask;
5474 	dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
5475 
5476 	for (; bucket != NULL; bucket = bucket->dthb_next) {
5477 		if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
5478 			return (bucket->dthb_len);
5479 	}
5480 
5481 	return (NULL);
5482 }
5483 
5484 static void
5485 dtrace_hash_remove(dtrace_hash_t *hash, dtrace_probe_t *probe)
5486 {
5487 	int ndx = DTRACE_HASHSTR(hash, probe) & hash->dth_mask;
5488 	dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
5489 
5490 	dtrace_probe_t **prevp = DTRACE_HASHPREV(hash, probe);
5491 	dtrace_probe_t **nextp = DTRACE_HASHNEXT(hash, probe);
5492 
5493 	/*
5494 	 * Find the bucket that we're removing this probe from.
5495 	 */
5496 	for (; bucket != NULL; bucket = bucket->dthb_next) {
5497 		if (DTRACE_HASHEQ(hash, bucket->dthb_chain, probe))
5498 			break;
5499 	}
5500 
5501 	ASSERT(bucket != NULL);
5502 
5503 	if (*prevp == NULL) {
5504 		if (*nextp == NULL) {
5505 			/*
5506 			 * The removed probe was the only probe on this
5507 			 * bucket; we need to remove the bucket.
5508 			 */
5509 			dtrace_hashbucket_t *b = hash->dth_tab[ndx];
5510 
5511 			ASSERT(bucket->dthb_chain == probe);
5512 			ASSERT(b != NULL);
5513 
5514 			if (b == bucket) {
5515 				hash->dth_tab[ndx] = bucket->dthb_next;
5516 			} else {
5517 				while (b->dthb_next != bucket)
5518 					b = b->dthb_next;
5519 				b->dthb_next = bucket->dthb_next;
5520 			}
5521 
5522 			ASSERT(hash->dth_nbuckets > 0);
5523 			hash->dth_nbuckets--;
5524 			kmem_free(bucket, sizeof (dtrace_hashbucket_t));
5525 			return;
5526 		}
5527 
5528 		bucket->dthb_chain = *nextp;
5529 	} else {
5530 		*(DTRACE_HASHNEXT(hash, *prevp)) = *nextp;
5531 	}
5532 
5533 	if (*nextp != NULL)
5534 		*(DTRACE_HASHPREV(hash, *nextp)) = *prevp;
5535 }
5536 
5537 /*
5538  * DTrace Utility Functions
5539  *
5540  * These are random utility functions that are _not_ called from probe context.
5541  */
5542 static int
5543 dtrace_badattr(const dtrace_attribute_t *a)
5544 {
5545 	return (a->dtat_name > DTRACE_STABILITY_MAX ||
5546 	    a->dtat_data > DTRACE_STABILITY_MAX ||
5547 	    a->dtat_class > DTRACE_CLASS_MAX);
5548 }
5549 
5550 /*
5551  * Return a duplicate copy of a string.  If the specified string is NULL,
5552  * this function returns a zero-length string.
5553  */
5554 static char *
5555 dtrace_strdup(const char *str)
5556 {
5557 	char *new = kmem_zalloc((str != NULL ? strlen(str) : 0) + 1, KM_SLEEP);
5558 
5559 	if (str != NULL)
5560 		(void) strcpy(new, str);
5561 
5562 	return (new);
5563 }
5564 
5565 #define	DTRACE_ISALPHA(c)	\
5566 	(((c) >= 'a' && (c) <= 'z') || ((c) >= 'A' && (c) <= 'Z'))
5567 
5568 static int
5569 dtrace_badname(const char *s)
5570 {
5571 	char c;
5572 
5573 	if (s == NULL || (c = *s++) == '\0')
5574 		return (0);
5575 
5576 	if (!DTRACE_ISALPHA(c) && c != '-' && c != '_' && c != '.')
5577 		return (1);
5578 
5579 	while ((c = *s++) != '\0') {
5580 		if (!DTRACE_ISALPHA(c) && (c < '0' || c > '9') &&
5581 		    c != '-' && c != '_' && c != '.' && c != '`')
5582 			return (1);
5583 	}
5584 
5585 	return (0);
5586 }
5587 
5588 static void
5589 dtrace_cred2priv(cred_t *cr, uint32_t *privp, uid_t *uidp, zoneid_t *zoneidp)
5590 {
5591 	uint32_t priv;
5592 
5593 	if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
5594 		/*
5595 		 * For DTRACE_PRIV_ALL, the uid and zoneid don't matter.
5596 		 */
5597 		priv = DTRACE_PRIV_ALL;
5598 	} else {
5599 		*uidp = crgetuid(cr);
5600 		*zoneidp = crgetzoneid(cr);
5601 
5602 		priv = 0;
5603 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE))
5604 			priv |= DTRACE_PRIV_KERNEL | DTRACE_PRIV_USER;
5605 		else if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE))
5606 			priv |= DTRACE_PRIV_USER;
5607 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE))
5608 			priv |= DTRACE_PRIV_PROC;
5609 		if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
5610 			priv |= DTRACE_PRIV_OWNER;
5611 		if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
5612 			priv |= DTRACE_PRIV_ZONEOWNER;
5613 	}
5614 
5615 	*privp = priv;
5616 }
5617 
5618 #ifdef DTRACE_ERRDEBUG
5619 static void
5620 dtrace_errdebug(const char *str)
5621 {
5622 	int hval = dtrace_hash_str((char *)str) % DTRACE_ERRHASHSZ;
5623 	int occupied = 0;
5624 
5625 	mutex_enter(&dtrace_errlock);
5626 	dtrace_errlast = str;
5627 	dtrace_errthread = curthread;
5628 
5629 	while (occupied++ < DTRACE_ERRHASHSZ) {
5630 		if (dtrace_errhash[hval].dter_msg == str) {
5631 			dtrace_errhash[hval].dter_count++;
5632 			goto out;
5633 		}
5634 
5635 		if (dtrace_errhash[hval].dter_msg != NULL) {
5636 			hval = (hval + 1) % DTRACE_ERRHASHSZ;
5637 			continue;
5638 		}
5639 
5640 		dtrace_errhash[hval].dter_msg = str;
5641 		dtrace_errhash[hval].dter_count = 1;
5642 		goto out;
5643 	}
5644 
5645 	panic("dtrace: undersized error hash");
5646 out:
5647 	mutex_exit(&dtrace_errlock);
5648 }
5649 #endif
5650 
5651 /*
5652  * DTrace Matching Functions
5653  *
5654  * These functions are used to match groups of probes, given some elements of
5655  * a probe tuple, or some globbed expressions for elements of a probe tuple.
5656  */
5657 static int
5658 dtrace_match_priv(const dtrace_probe_t *prp, uint32_t priv, uid_t uid,
5659     zoneid_t zoneid)
5660 {
5661 	if (priv != DTRACE_PRIV_ALL) {
5662 		uint32_t ppriv = prp->dtpr_provider->dtpv_priv.dtpp_flags;
5663 		uint32_t match = priv & ppriv;
5664 
5665 		/*
5666 		 * No PRIV_DTRACE_* privileges...
5667 		 */
5668 		if ((priv & (DTRACE_PRIV_PROC | DTRACE_PRIV_USER |
5669 		    DTRACE_PRIV_KERNEL)) == 0)
5670 			return (0);
5671 
5672 		/*
5673 		 * No matching bits, but there were bits to match...
5674 		 */
5675 		if (match == 0 && ppriv != 0)
5676 			return (0);
5677 
5678 		/*
5679 		 * Need to have permissions to the process, but don't...
5680 		 */
5681 		if (((ppriv & ~match) & DTRACE_PRIV_OWNER) != 0 &&
5682 		    uid != prp->dtpr_provider->dtpv_priv.dtpp_uid) {
5683 			return (0);
5684 		}
5685 
5686 		/*
5687 		 * Need to be in the same zone unless we possess the
5688 		 * privilege to examine all zones.
5689 		 */
5690 		if (((ppriv & ~match) & DTRACE_PRIV_ZONEOWNER) != 0 &&
5691 		    zoneid != prp->dtpr_provider->dtpv_priv.dtpp_zoneid) {
5692 			return (0);
5693 		}
5694 	}
5695 
5696 	return (1);
5697 }
5698 
5699 /*
5700  * dtrace_match_probe compares a dtrace_probe_t to a pre-compiled key, which
5701  * consists of input pattern strings and an ops-vector to evaluate them.
5702  * This function returns >0 for match, 0 for no match, and <0 for error.
5703  */
5704 static int
5705 dtrace_match_probe(const dtrace_probe_t *prp, const dtrace_probekey_t *pkp,
5706     uint32_t priv, uid_t uid, zoneid_t zoneid)
5707 {
5708 	dtrace_provider_t *pvp = prp->dtpr_provider;
5709 	int rv;
5710 
5711 	if (pvp->dtpv_defunct)
5712 		return (0);
5713 
5714 	if ((rv = pkp->dtpk_pmatch(pvp->dtpv_name, pkp->dtpk_prov, 0)) <= 0)
5715 		return (rv);
5716 
5717 	if ((rv = pkp->dtpk_mmatch(prp->dtpr_mod, pkp->dtpk_mod, 0)) <= 0)
5718 		return (rv);
5719 
5720 	if ((rv = pkp->dtpk_fmatch(prp->dtpr_func, pkp->dtpk_func, 0)) <= 0)
5721 		return (rv);
5722 
5723 	if ((rv = pkp->dtpk_nmatch(prp->dtpr_name, pkp->dtpk_name, 0)) <= 0)
5724 		return (rv);
5725 
5726 	if (dtrace_match_priv(prp, priv, uid, zoneid) == 0)
5727 		return (0);
5728 
5729 	return (rv);
5730 }
5731 
5732 /*
5733  * dtrace_match_glob() is a safe kernel implementation of the gmatch(3GEN)
5734  * interface for matching a glob pattern 'p' to an input string 's'.  Unlike
5735  * libc's version, the kernel version only applies to 8-bit ASCII strings.
5736  * In addition, all of the recursion cases except for '*' matching have been
5737  * unwound.  For '*', we still implement recursive evaluation, but a depth
5738  * counter is maintained and matching is aborted if we recurse too deep.
5739  * The function returns 0 if no match, >0 if match, and <0 if recursion error.
5740  */
5741 static int
5742 dtrace_match_glob(const char *s, const char *p, int depth)
5743 {
5744 	const char *olds;
5745 	char s1, c;
5746 	int gs;
5747 
5748 	if (depth > DTRACE_PROBEKEY_MAXDEPTH)
5749 		return (-1);
5750 
5751 	if (s == NULL)
5752 		s = ""; /* treat NULL as empty string */
5753 
5754 top:
5755 	olds = s;
5756 	s1 = *s++;
5757 
5758 	if (p == NULL)
5759 		return (0);
5760 
5761 	if ((c = *p++) == '\0')
5762 		return (s1 == '\0');
5763 
5764 	switch (c) {
5765 	case '[': {
5766 		int ok = 0, notflag = 0;
5767 		char lc = '\0';
5768 
5769 		if (s1 == '\0')
5770 			return (0);
5771 
5772 		if (*p == '!') {
5773 			notflag = 1;
5774 			p++;
5775 		}
5776 
5777 		if ((c = *p++) == '\0')
5778 			return (0);
5779 
5780 		do {
5781 			if (c == '-' && lc != '\0' && *p != ']') {
5782 				if ((c = *p++) == '\0')
5783 					return (0);
5784 				if (c == '\\' && (c = *p++) == '\0')
5785 					return (0);
5786 
5787 				if (notflag) {
5788 					if (s1 < lc || s1 > c)
5789 						ok++;
5790 					else
5791 						return (0);
5792 				} else if (lc <= s1 && s1 <= c)
5793 					ok++;
5794 
5795 			} else if (c == '\\' && (c = *p++) == '\0')
5796 				return (0);
5797 
5798 			lc = c; /* save left-hand 'c' for next iteration */
5799 
5800 			if (notflag) {
5801 				if (s1 != c)
5802 					ok++;
5803 				else
5804 					return (0);
5805 			} else if (s1 == c)
5806 				ok++;
5807 
5808 			if ((c = *p++) == '\0')
5809 				return (0);
5810 
5811 		} while (c != ']');
5812 
5813 		if (ok)
5814 			goto top;
5815 
5816 		return (0);
5817 	}
5818 
5819 	case '\\':
5820 		if ((c = *p++) == '\0')
5821 			return (0);
5822 		/*FALLTHRU*/
5823 
5824 	default:
5825 		if (c != s1)
5826 			return (0);
5827 		/*FALLTHRU*/
5828 
5829 	case '?':
5830 		if (s1 != '\0')
5831 			goto top;
5832 		return (0);
5833 
5834 	case '*':
5835 		while (*p == '*')
5836 			p++; /* consecutive *'s are identical to a single one */
5837 
5838 		if (*p == '\0')
5839 			return (1);
5840 
5841 		for (s = olds; *s != '\0'; s++) {
5842 			if ((gs = dtrace_match_glob(s, p, depth + 1)) != 0)
5843 				return (gs);
5844 		}
5845 
5846 		return (0);
5847 	}
5848 }
5849 
5850 /*ARGSUSED*/
5851 static int
5852 dtrace_match_string(const char *s, const char *p, int depth)
5853 {
5854 	return (s != NULL && strcmp(s, p) == 0);
5855 }
5856 
5857 /*ARGSUSED*/
5858 static int
5859 dtrace_match_nul(const char *s, const char *p, int depth)
5860 {
5861 	return (1); /* always match the empty pattern */
5862 }
5863 
5864 /*ARGSUSED*/
5865 static int
5866 dtrace_match_nonzero(const char *s, const char *p, int depth)
5867 {
5868 	return (s != NULL && s[0] != '\0');
5869 }
5870 
5871 static int
5872 dtrace_match(const dtrace_probekey_t *pkp, uint32_t priv, uid_t uid,
5873     zoneid_t zoneid, int (*matched)(dtrace_probe_t *, void *), void *arg)
5874 {
5875 	dtrace_probe_t template, *probe;
5876 	dtrace_hash_t *hash = NULL;
5877 	int len, best = INT_MAX, nmatched = 0;
5878 	dtrace_id_t i;
5879 
5880 	ASSERT(MUTEX_HELD(&dtrace_lock));
5881 
5882 	/*
5883 	 * If the probe ID is specified in the key, just lookup by ID and
5884 	 * invoke the match callback once if a matching probe is found.
5885 	 */
5886 	if (pkp->dtpk_id != DTRACE_IDNONE) {
5887 		if ((probe = dtrace_probe_lookup_id(pkp->dtpk_id)) != NULL &&
5888 		    dtrace_match_probe(probe, pkp, priv, uid, zoneid) > 0) {
5889 			(void) (*matched)(probe, arg);
5890 			nmatched++;
5891 		}
5892 		return (nmatched);
5893 	}
5894 
5895 	template.dtpr_mod = (char *)pkp->dtpk_mod;
5896 	template.dtpr_func = (char *)pkp->dtpk_func;
5897 	template.dtpr_name = (char *)pkp->dtpk_name;
5898 
5899 	/*
5900 	 * We want to find the most distinct of the module name, function
5901 	 * name, and name.  So for each one that is not a glob pattern or
5902 	 * empty string, we perform a lookup in the corresponding hash and
5903 	 * use the hash table with the fewest collisions to do our search.
5904 	 */
5905 	if (pkp->dtpk_mmatch == &dtrace_match_string &&
5906 	    (len = dtrace_hash_collisions(dtrace_bymod, &template)) < best) {
5907 		best = len;
5908 		hash = dtrace_bymod;
5909 	}
5910 
5911 	if (pkp->dtpk_fmatch == &dtrace_match_string &&
5912 	    (len = dtrace_hash_collisions(dtrace_byfunc, &template)) < best) {
5913 		best = len;
5914 		hash = dtrace_byfunc;
5915 	}
5916 
5917 	if (pkp->dtpk_nmatch == &dtrace_match_string &&
5918 	    (len = dtrace_hash_collisions(dtrace_byname, &template)) < best) {
5919 		best = len;
5920 		hash = dtrace_byname;
5921 	}
5922 
5923 	/*
5924 	 * If we did not select a hash table, iterate over every probe and
5925 	 * invoke our callback for each one that matches our input probe key.
5926 	 */
5927 	if (hash == NULL) {
5928 		for (i = 0; i < dtrace_nprobes; i++) {
5929 			if ((probe = dtrace_probes[i]) == NULL ||
5930 			    dtrace_match_probe(probe, pkp, priv, uid,
5931 			    zoneid) <= 0)
5932 				continue;
5933 
5934 			nmatched++;
5935 
5936 			if ((*matched)(probe, arg) != DTRACE_MATCH_NEXT)
5937 				break;
5938 		}
5939 
5940 		return (nmatched);
5941 	}
5942 
5943 	/*
5944 	 * If we selected a hash table, iterate over each probe of the same key
5945 	 * name and invoke the callback for every probe that matches the other
5946 	 * attributes of our input probe key.
5947 	 */
5948 	for (probe = dtrace_hash_lookup(hash, &template); probe != NULL;
5949 	    probe = *(DTRACE_HASHNEXT(hash, probe))) {
5950 
5951 		if (dtrace_match_probe(probe, pkp, priv, uid, zoneid) <= 0)
5952 			continue;
5953 
5954 		nmatched++;
5955 
5956 		if ((*matched)(probe, arg) != DTRACE_MATCH_NEXT)
5957 			break;
5958 	}
5959 
5960 	return (nmatched);
5961 }
5962 
5963 /*
5964  * Return the function pointer dtrace_probecmp() should use to compare the
5965  * specified pattern with a string.  For NULL or empty patterns, we select
5966  * dtrace_match_nul().  For glob pattern strings, we use dtrace_match_glob().
5967  * For non-empty non-glob strings, we use dtrace_match_string().
5968  */
5969 static dtrace_probekey_f *
5970 dtrace_probekey_func(const char *p)
5971 {
5972 	char c;
5973 
5974 	if (p == NULL || *p == '\0')
5975 		return (&dtrace_match_nul);
5976 
5977 	while ((c = *p++) != '\0') {
5978 		if (c == '[' || c == '?' || c == '*' || c == '\\')
5979 			return (&dtrace_match_glob);
5980 	}
5981 
5982 	return (&dtrace_match_string);
5983 }
5984 
5985 /*
5986  * Build a probe comparison key for use with dtrace_match_probe() from the
5987  * given probe description.  By convention, a null key only matches anchored
5988  * probes: if each field is the empty string, reset dtpk_fmatch to
5989  * dtrace_match_nonzero().
5990  */
5991 static void
5992 dtrace_probekey(const dtrace_probedesc_t *pdp, dtrace_probekey_t *pkp)
5993 {
5994 	pkp->dtpk_prov = pdp->dtpd_provider;
5995 	pkp->dtpk_pmatch = dtrace_probekey_func(pdp->dtpd_provider);
5996 
5997 	pkp->dtpk_mod = pdp->dtpd_mod;
5998 	pkp->dtpk_mmatch = dtrace_probekey_func(pdp->dtpd_mod);
5999 
6000 	pkp->dtpk_func = pdp->dtpd_func;
6001 	pkp->dtpk_fmatch = dtrace_probekey_func(pdp->dtpd_func);
6002 
6003 	pkp->dtpk_name = pdp->dtpd_name;
6004 	pkp->dtpk_nmatch = dtrace_probekey_func(pdp->dtpd_name);
6005 
6006 	pkp->dtpk_id = pdp->dtpd_id;
6007 
6008 	if (pkp->dtpk_id == DTRACE_IDNONE &&
6009 	    pkp->dtpk_pmatch == &dtrace_match_nul &&
6010 	    pkp->dtpk_mmatch == &dtrace_match_nul &&
6011 	    pkp->dtpk_fmatch == &dtrace_match_nul &&
6012 	    pkp->dtpk_nmatch == &dtrace_match_nul)
6013 		pkp->dtpk_fmatch = &dtrace_match_nonzero;
6014 }
6015 
6016 /*
6017  * DTrace Provider-to-Framework API Functions
6018  *
6019  * These functions implement much of the Provider-to-Framework API, as
6020  * described in <sys/dtrace.h>.  The parts of the API not in this section are
6021  * the functions in the API for probe management (found below), and
6022  * dtrace_probe() itself (found above).
6023  */
6024 
6025 /*
6026  * Register the calling provider with the DTrace framework.  This should
6027  * generally be called by DTrace providers in their attach(9E) entry point.
6028  */
6029 int
6030 dtrace_register(const char *name, const dtrace_pattr_t *pap, uint32_t priv,
6031     cred_t *cr, const dtrace_pops_t *pops, void *arg, dtrace_provider_id_t *idp)
6032 {
6033 	dtrace_provider_t *provider;
6034 
6035 	if (name == NULL || pap == NULL || pops == NULL || idp == NULL) {
6036 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
6037 		    "arguments", name ? name : "<NULL>");
6038 		return (EINVAL);
6039 	}
6040 
6041 	if (name[0] == '\0' || dtrace_badname(name)) {
6042 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
6043 		    "provider name", name);
6044 		return (EINVAL);
6045 	}
6046 
6047 	if ((pops->dtps_provide == NULL && pops->dtps_provide_module == NULL) ||
6048 	    pops->dtps_enable == NULL || pops->dtps_disable == NULL ||
6049 	    pops->dtps_destroy == NULL ||
6050 	    ((pops->dtps_resume == NULL) != (pops->dtps_suspend == NULL))) {
6051 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
6052 		    "provider ops", name);
6053 		return (EINVAL);
6054 	}
6055 
6056 	if (dtrace_badattr(&pap->dtpa_provider) ||
6057 	    dtrace_badattr(&pap->dtpa_mod) ||
6058 	    dtrace_badattr(&pap->dtpa_func) ||
6059 	    dtrace_badattr(&pap->dtpa_name) ||
6060 	    dtrace_badattr(&pap->dtpa_args)) {
6061 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
6062 		    "provider attributes", name);
6063 		return (EINVAL);
6064 	}
6065 
6066 	if (priv & ~DTRACE_PRIV_ALL) {
6067 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
6068 		    "privilege attributes", name);
6069 		return (EINVAL);
6070 	}
6071 
6072 	if ((priv & DTRACE_PRIV_KERNEL) &&
6073 	    (priv & (DTRACE_PRIV_USER | DTRACE_PRIV_OWNER)) &&
6074 	    pops->dtps_usermode == NULL) {
6075 		cmn_err(CE_WARN, "failed to register provider '%s': need "
6076 		    "dtps_usermode() op for given privilege attributes", name);
6077 		return (EINVAL);
6078 	}
6079 
6080 	provider = kmem_zalloc(sizeof (dtrace_provider_t), KM_SLEEP);
6081 	provider->dtpv_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
6082 	(void) strcpy(provider->dtpv_name, name);
6083 
6084 	provider->dtpv_attr = *pap;
6085 	provider->dtpv_priv.dtpp_flags = priv;
6086 	if (cr != NULL) {
6087 		provider->dtpv_priv.dtpp_uid = crgetuid(cr);
6088 		provider->dtpv_priv.dtpp_zoneid = crgetzoneid(cr);
6089 	}
6090 	provider->dtpv_pops = *pops;
6091 
6092 	if (pops->dtps_provide == NULL) {
6093 		ASSERT(pops->dtps_provide_module != NULL);
6094 		provider->dtpv_pops.dtps_provide =
6095 		    (void (*)(void *, const dtrace_probedesc_t *))dtrace_nullop;
6096 	}
6097 
6098 	if (pops->dtps_provide_module == NULL) {
6099 		ASSERT(pops->dtps_provide != NULL);
6100 		provider->dtpv_pops.dtps_provide_module =
6101 		    (void (*)(void *, struct modctl *))dtrace_nullop;
6102 	}
6103 
6104 	if (pops->dtps_suspend == NULL) {
6105 		ASSERT(pops->dtps_resume == NULL);
6106 		provider->dtpv_pops.dtps_suspend =
6107 		    (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
6108 		provider->dtpv_pops.dtps_resume =
6109 		    (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
6110 	}
6111 
6112 	provider->dtpv_arg = arg;
6113 	*idp = (dtrace_provider_id_t)provider;
6114 
6115 	if (pops == &dtrace_provider_ops) {
6116 		ASSERT(MUTEX_HELD(&dtrace_provider_lock));
6117 		ASSERT(MUTEX_HELD(&dtrace_lock));
6118 		ASSERT(dtrace_anon.dta_enabling == NULL);
6119 
6120 		/*
6121 		 * We make sure that the DTrace provider is at the head of
6122 		 * the provider chain.
6123 		 */
6124 		provider->dtpv_next = dtrace_provider;
6125 		dtrace_provider = provider;
6126 		return (0);
6127 	}
6128 
6129 	mutex_enter(&dtrace_provider_lock);
6130 	mutex_enter(&dtrace_lock);
6131 
6132 	/*
6133 	 * If there is at least one provider registered, we'll add this
6134 	 * provider after the first provider.
6135 	 */
6136 	if (dtrace_provider != NULL) {
6137 		provider->dtpv_next = dtrace_provider->dtpv_next;
6138 		dtrace_provider->dtpv_next = provider;
6139 	} else {
6140 		dtrace_provider = provider;
6141 	}
6142 
6143 	if (dtrace_retained != NULL) {
6144 		dtrace_enabling_provide(provider);
6145 
6146 		/*
6147 		 * Now we need to call dtrace_enabling_matchall() -- which
6148 		 * will acquire cpu_lock and dtrace_lock.  We therefore need
6149 		 * to drop all of our locks before calling into it...
6150 		 */
6151 		mutex_exit(&dtrace_lock);
6152 		mutex_exit(&dtrace_provider_lock);
6153 		dtrace_enabling_matchall();
6154 
6155 		return (0);
6156 	}
6157 
6158 	mutex_exit(&dtrace_lock);
6159 	mutex_exit(&dtrace_provider_lock);
6160 
6161 	return (0);
6162 }
6163 
6164 /*
6165  * Unregister the specified provider from the DTrace framework.  This should
6166  * generally be called by DTrace providers in their detach(9E) entry point.
6167  */
6168 int
6169 dtrace_unregister(dtrace_provider_id_t id)
6170 {
6171 	dtrace_provider_t *old = (dtrace_provider_t *)id;
6172 	dtrace_provider_t *prev = NULL;
6173 	int i, self = 0;
6174 	dtrace_probe_t *probe, *first = NULL;
6175 
6176 	if (old->dtpv_pops.dtps_enable ==
6177 	    (void (*)(void *, dtrace_id_t, void *))dtrace_nullop) {
6178 		/*
6179 		 * If DTrace itself is the provider, we're called with locks
6180 		 * already held.
6181 		 */
6182 		ASSERT(old == dtrace_provider);
6183 		ASSERT(dtrace_devi != NULL);
6184 		ASSERT(MUTEX_HELD(&dtrace_provider_lock));
6185 		ASSERT(MUTEX_HELD(&dtrace_lock));
6186 		self = 1;
6187 
6188 		if (dtrace_provider->dtpv_next != NULL) {
6189 			/*
6190 			 * There's another provider here; return failure.
6191 			 */
6192 			return (EBUSY);
6193 		}
6194 	} else {
6195 		mutex_enter(&dtrace_provider_lock);
6196 		mutex_enter(&mod_lock);
6197 		mutex_enter(&dtrace_lock);
6198 	}
6199 
6200 	/*
6201 	 * If anyone has /dev/dtrace open, or if there are anonymous enabled
6202 	 * probes, we refuse to let providers slither away, unless this
6203 	 * provider has already been explicitly invalidated.
6204 	 */
6205 	if (!old->dtpv_defunct &&
6206 	    (dtrace_opens || (dtrace_anon.dta_state != NULL &&
6207 	    dtrace_anon.dta_state->dts_necbs > 0))) {
6208 		if (!self) {
6209 			mutex_exit(&dtrace_lock);
6210 			mutex_exit(&mod_lock);
6211 			mutex_exit(&dtrace_provider_lock);
6212 		}
6213 		return (EBUSY);
6214 	}
6215 
6216 	/*
6217 	 * Attempt to destroy the probes associated with this provider.
6218 	 */
6219 	for (i = 0; i < dtrace_nprobes; i++) {
6220 		if ((probe = dtrace_probes[i]) == NULL)
6221 			continue;
6222 
6223 		if (probe->dtpr_provider != old)
6224 			continue;
6225 
6226 		if (probe->dtpr_ecb == NULL)
6227 			continue;
6228 
6229 		/*
6230 		 * We have at least one ECB; we can't remove this provider.
6231 		 */
6232 		if (!self) {
6233 			mutex_exit(&dtrace_lock);
6234 			mutex_exit(&mod_lock);
6235 			mutex_exit(&dtrace_provider_lock);
6236 		}
6237 		return (EBUSY);
6238 	}
6239 
6240 	/*
6241 	 * All of the probes for this provider are disabled; we can safely
6242 	 * remove all of them from their hash chains and from the probe array.
6243 	 */
6244 	for (i = 0; i < dtrace_nprobes; i++) {
6245 		if ((probe = dtrace_probes[i]) == NULL)
6246 			continue;
6247 
6248 		if (probe->dtpr_provider != old)
6249 			continue;
6250 
6251 		dtrace_probes[i] = NULL;
6252 
6253 		dtrace_hash_remove(dtrace_bymod, probe);
6254 		dtrace_hash_remove(dtrace_byfunc, probe);
6255 		dtrace_hash_remove(dtrace_byname, probe);
6256 
6257 		if (first == NULL) {
6258 			first = probe;
6259 			probe->dtpr_nextmod = NULL;
6260 		} else {
6261 			probe->dtpr_nextmod = first;
6262 			first = probe;
6263 		}
6264 	}
6265 
6266 	/*
6267 	 * The provider's probes have been removed from the hash chains and
6268 	 * from the probe array.  Now issue a dtrace_sync() to be sure that
6269 	 * everyone has cleared out from any probe array processing.
6270 	 */
6271 	dtrace_sync();
6272 
6273 	for (probe = first; probe != NULL; probe = first) {
6274 		first = probe->dtpr_nextmod;
6275 
6276 		old->dtpv_pops.dtps_destroy(old->dtpv_arg, probe->dtpr_id,
6277 		    probe->dtpr_arg);
6278 		kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
6279 		kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
6280 		kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
6281 		vmem_free(dtrace_arena, (void *)(uintptr_t)(probe->dtpr_id), 1);
6282 		kmem_free(probe, sizeof (dtrace_probe_t));
6283 	}
6284 
6285 	if ((prev = dtrace_provider) == old) {
6286 		ASSERT(self || dtrace_devi == NULL);
6287 		ASSERT(old->dtpv_next == NULL || dtrace_devi == NULL);
6288 		dtrace_provider = old->dtpv_next;
6289 	} else {
6290 		while (prev != NULL && prev->dtpv_next != old)
6291 			prev = prev->dtpv_next;
6292 
6293 		if (prev == NULL) {
6294 			panic("attempt to unregister non-existent "
6295 			    "dtrace provider %p\n", (void *)id);
6296 		}
6297 
6298 		prev->dtpv_next = old->dtpv_next;
6299 	}
6300 
6301 	if (!self) {
6302 		mutex_exit(&dtrace_lock);
6303 		mutex_exit(&mod_lock);
6304 		mutex_exit(&dtrace_provider_lock);
6305 	}
6306 
6307 	kmem_free(old->dtpv_name, strlen(old->dtpv_name) + 1);
6308 	kmem_free(old, sizeof (dtrace_provider_t));
6309 
6310 	return (0);
6311 }
6312 
6313 /*
6314  * Invalidate the specified provider.  All subsequent probe lookups for the
6315  * specified provider will fail, but its probes will not be removed.
6316  */
6317 void
6318 dtrace_invalidate(dtrace_provider_id_t id)
6319 {
6320 	dtrace_provider_t *pvp = (dtrace_provider_t *)id;
6321 
6322 	ASSERT(pvp->dtpv_pops.dtps_enable !=
6323 	    (void (*)(void *, dtrace_id_t, void *))dtrace_nullop);
6324 
6325 	mutex_enter(&dtrace_provider_lock);
6326 	mutex_enter(&dtrace_lock);
6327 
6328 	pvp->dtpv_defunct = 1;
6329 
6330 	mutex_exit(&dtrace_lock);
6331 	mutex_exit(&dtrace_provider_lock);
6332 }
6333 
6334 /*
6335  * Indicate whether or not DTrace has attached.
6336  */
6337 int
6338 dtrace_attached(void)
6339 {
6340 	/*
6341 	 * dtrace_provider will be non-NULL iff the DTrace driver has
6342 	 * attached.  (It's non-NULL because DTrace is always itself a
6343 	 * provider.)
6344 	 */
6345 	return (dtrace_provider != NULL);
6346 }
6347 
6348 /*
6349  * Remove all the unenabled probes for the given provider.  This function is
6350  * not unlike dtrace_unregister(), except that it doesn't remove the provider
6351  * -- just as many of its associated probes as it can.
6352  */
6353 int
6354 dtrace_condense(dtrace_provider_id_t id)
6355 {
6356 	dtrace_provider_t *prov = (dtrace_provider_t *)id;
6357 	int i;
6358 	dtrace_probe_t *probe;
6359 
6360 	/*
6361 	 * Make sure this isn't the dtrace provider itself.
6362 	 */
6363 	ASSERT(prov->dtpv_pops.dtps_enable !=
6364 	    (void (*)(void *, dtrace_id_t, void *))dtrace_nullop);
6365 
6366 	mutex_enter(&dtrace_provider_lock);
6367 	mutex_enter(&dtrace_lock);
6368 
6369 	/*
6370 	 * Attempt to destroy the probes associated with this provider.
6371 	 */
6372 	for (i = 0; i < dtrace_nprobes; i++) {
6373 		if ((probe = dtrace_probes[i]) == NULL)
6374 			continue;
6375 
6376 		if (probe->dtpr_provider != prov)
6377 			continue;
6378 
6379 		if (probe->dtpr_ecb != NULL)
6380 			continue;
6381 
6382 		dtrace_probes[i] = NULL;
6383 
6384 		dtrace_hash_remove(dtrace_bymod, probe);
6385 		dtrace_hash_remove(dtrace_byfunc, probe);
6386 		dtrace_hash_remove(dtrace_byname, probe);
6387 
6388 		prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, i + 1,
6389 		    probe->dtpr_arg);
6390 		kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
6391 		kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
6392 		kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
6393 		kmem_free(probe, sizeof (dtrace_probe_t));
6394 		vmem_free(dtrace_arena, (void *)((uintptr_t)i + 1), 1);
6395 	}
6396 
6397 	mutex_exit(&dtrace_lock);
6398 	mutex_exit(&dtrace_provider_lock);
6399 
6400 	return (0);
6401 }
6402 
6403 /*
6404  * DTrace Probe Management Functions
6405  *
6406  * The functions in this section perform the DTrace probe management,
6407  * including functions to create probes, look-up probes, and call into the
6408  * providers to request that probes be provided.  Some of these functions are
6409  * in the Provider-to-Framework API; these functions can be identified by the
6410  * fact that they are not declared "static".
6411  */
6412 
6413 /*
6414  * Create a probe with the specified module name, function name, and name.
6415  */
6416 dtrace_id_t
6417 dtrace_probe_create(dtrace_provider_id_t prov, const char *mod,
6418     const char *func, const char *name, int aframes, void *arg)
6419 {
6420 	dtrace_probe_t *probe, **probes;
6421 	dtrace_provider_t *provider = (dtrace_provider_t *)prov;
6422 	dtrace_id_t id;
6423 
6424 	if (provider == dtrace_provider) {
6425 		ASSERT(MUTEX_HELD(&dtrace_lock));
6426 	} else {
6427 		mutex_enter(&dtrace_lock);
6428 	}
6429 
6430 	id = (dtrace_id_t)(uintptr_t)vmem_alloc(dtrace_arena, 1,
6431 	    VM_BESTFIT | VM_SLEEP);
6432 	probe = kmem_zalloc(sizeof (dtrace_probe_t), KM_SLEEP);
6433 
6434 	probe->dtpr_id = id;
6435 	probe->dtpr_gen = dtrace_probegen++;
6436 	probe->dtpr_mod = dtrace_strdup(mod);
6437 	probe->dtpr_func = dtrace_strdup(func);
6438 	probe->dtpr_name = dtrace_strdup(name);
6439 	probe->dtpr_arg = arg;
6440 	probe->dtpr_aframes = aframes;
6441 	probe->dtpr_provider = provider;
6442 
6443 	dtrace_hash_add(dtrace_bymod, probe);
6444 	dtrace_hash_add(dtrace_byfunc, probe);
6445 	dtrace_hash_add(dtrace_byname, probe);
6446 
6447 	if (id - 1 >= dtrace_nprobes) {
6448 		size_t osize = dtrace_nprobes * sizeof (dtrace_probe_t *);
6449 		size_t nsize = osize << 1;
6450 
6451 		if (nsize == 0) {
6452 			ASSERT(osize == 0);
6453 			ASSERT(dtrace_probes == NULL);
6454 			nsize = sizeof (dtrace_probe_t *);
6455 		}
6456 
6457 		probes = kmem_zalloc(nsize, KM_SLEEP);
6458 
6459 		if (dtrace_probes == NULL) {
6460 			ASSERT(osize == 0);
6461 			dtrace_probes = probes;
6462 			dtrace_nprobes = 1;
6463 		} else {
6464 			dtrace_probe_t **oprobes = dtrace_probes;
6465 
6466 			bcopy(oprobes, probes, osize);
6467 			dtrace_membar_producer();
6468 			dtrace_probes = probes;
6469 
6470 			dtrace_sync();
6471 
6472 			/*
6473 			 * All CPUs are now seeing the new probes array; we can
6474 			 * safely free the old array.
6475 			 */
6476 			kmem_free(oprobes, osize);
6477 			dtrace_nprobes <<= 1;
6478 		}
6479 
6480 		ASSERT(id - 1 < dtrace_nprobes);
6481 	}
6482 
6483 	ASSERT(dtrace_probes[id - 1] == NULL);
6484 	dtrace_probes[id - 1] = probe;
6485 
6486 	if (provider != dtrace_provider)
6487 		mutex_exit(&dtrace_lock);
6488 
6489 	return (id);
6490 }
6491 
6492 static dtrace_probe_t *
6493 dtrace_probe_lookup_id(dtrace_id_t id)
6494 {
6495 	ASSERT(MUTEX_HELD(&dtrace_lock));
6496 
6497 	if (id == 0 || id > dtrace_nprobes)
6498 		return (NULL);
6499 
6500 	return (dtrace_probes[id - 1]);
6501 }
6502 
6503 static int
6504 dtrace_probe_lookup_match(dtrace_probe_t *probe, void *arg)
6505 {
6506 	*((dtrace_id_t *)arg) = probe->dtpr_id;
6507 
6508 	return (DTRACE_MATCH_DONE);
6509 }
6510 
6511 /*
6512  * Look up a probe based on provider and one or more of module name, function
6513  * name and probe name.
6514  */
6515 dtrace_id_t
6516 dtrace_probe_lookup(dtrace_provider_id_t prid, const char *mod,
6517     const char *func, const char *name)
6518 {
6519 	dtrace_probekey_t pkey;
6520 	dtrace_id_t id;
6521 	int match;
6522 
6523 	pkey.dtpk_prov = ((dtrace_provider_t *)prid)->dtpv_name;
6524 	pkey.dtpk_pmatch = &dtrace_match_string;
6525 	pkey.dtpk_mod = mod;
6526 	pkey.dtpk_mmatch = mod ? &dtrace_match_string : &dtrace_match_nul;
6527 	pkey.dtpk_func = func;
6528 	pkey.dtpk_fmatch = func ? &dtrace_match_string : &dtrace_match_nul;
6529 	pkey.dtpk_name = name;
6530 	pkey.dtpk_nmatch = name ? &dtrace_match_string : &dtrace_match_nul;
6531 	pkey.dtpk_id = DTRACE_IDNONE;
6532 
6533 	mutex_enter(&dtrace_lock);
6534 	match = dtrace_match(&pkey, DTRACE_PRIV_ALL, 0, 0,
6535 	    dtrace_probe_lookup_match, &id);
6536 	mutex_exit(&dtrace_lock);
6537 
6538 	ASSERT(match == 1 || match == 0);
6539 	return (match ? id : 0);
6540 }
6541 
6542 /*
6543  * Returns the probe argument associated with the specified probe.
6544  */
6545 void *
6546 dtrace_probe_arg(dtrace_provider_id_t id, dtrace_id_t pid)
6547 {
6548 	dtrace_probe_t *probe;
6549 	void *rval = NULL;
6550 
6551 	mutex_enter(&dtrace_lock);
6552 
6553 	if ((probe = dtrace_probe_lookup_id(pid)) != NULL &&
6554 	    probe->dtpr_provider == (dtrace_provider_t *)id)
6555 		rval = probe->dtpr_arg;
6556 
6557 	mutex_exit(&dtrace_lock);
6558 
6559 	return (rval);
6560 }
6561 
6562 /*
6563  * Copy a probe into a probe description.
6564  */
6565 static void
6566 dtrace_probe_description(const dtrace_probe_t *prp, dtrace_probedesc_t *pdp)
6567 {
6568 	bzero(pdp, sizeof (dtrace_probedesc_t));
6569 	pdp->dtpd_id = prp->dtpr_id;
6570 
6571 	(void) strncpy(pdp->dtpd_provider,
6572 	    prp->dtpr_provider->dtpv_name, DTRACE_PROVNAMELEN - 1);
6573 
6574 	(void) strncpy(pdp->dtpd_mod, prp->dtpr_mod, DTRACE_MODNAMELEN - 1);
6575 	(void) strncpy(pdp->dtpd_func, prp->dtpr_func, DTRACE_FUNCNAMELEN - 1);
6576 	(void) strncpy(pdp->dtpd_name, prp->dtpr_name, DTRACE_NAMELEN - 1);
6577 }
6578 
6579 /*
6580  * Called to indicate that a probe -- or probes -- should be provided by a
6581  * specfied provider.  If the specified description is NULL, the provider will
6582  * be told to provide all of its probes.  (This is done whenever a new
6583  * consumer comes along, or whenever a retained enabling is to be matched.) If
6584  * the specified description is non-NULL, the provider is given the
6585  * opportunity to dynamically provide the specified probe, allowing providers
6586  * to support the creation of probes on-the-fly.  (So-called _autocreated_
6587  * probes.)  If the provider is NULL, the operations will be applied to all
6588  * providers; if the provider is non-NULL the operations will only be applied
6589  * to the specified provider.  The dtrace_provider_lock must be held, and the
6590  * dtrace_lock must _not_ be held -- the provider's dtps_provide() operation
6591  * will need to grab the dtrace_lock when it reenters the framework through
6592  * dtrace_probe_lookup(), dtrace_probe_create(), etc.
6593  */
6594 static void
6595 dtrace_probe_provide(dtrace_probedesc_t *desc, dtrace_provider_t *prv)
6596 {
6597 	struct modctl *ctl;
6598 	int all = 0;
6599 
6600 	ASSERT(MUTEX_HELD(&dtrace_provider_lock));
6601 
6602 	if (prv == NULL) {
6603 		all = 1;
6604 		prv = dtrace_provider;
6605 	}
6606 
6607 	do {
6608 		/*
6609 		 * First, call the blanket provide operation.
6610 		 */
6611 		prv->dtpv_pops.dtps_provide(prv->dtpv_arg, desc);
6612 
6613 		/*
6614 		 * Now call the per-module provide operation.  We will grab
6615 		 * mod_lock to prevent the list from being modified.  Note
6616 		 * that this also prevents the mod_busy bits from changing.
6617 		 * (mod_busy can only be changed with mod_lock held.)
6618 		 */
6619 		mutex_enter(&mod_lock);
6620 
6621 		ctl = &modules;
6622 		do {
6623 			if (ctl->mod_busy || ctl->mod_mp == NULL)
6624 				continue;
6625 
6626 			prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
6627 
6628 		} while ((ctl = ctl->mod_next) != &modules);
6629 
6630 		mutex_exit(&mod_lock);
6631 	} while (all && (prv = prv->dtpv_next) != NULL);
6632 }
6633 
6634 /*
6635  * Iterate over each probe, and call the Framework-to-Provider API function
6636  * denoted by offs.
6637  */
6638 static void
6639 dtrace_probe_foreach(uintptr_t offs)
6640 {
6641 	dtrace_provider_t *prov;
6642 	void (*func)(void *, dtrace_id_t, void *);
6643 	dtrace_probe_t *probe;
6644 	dtrace_icookie_t cookie;
6645 	int i;
6646 
6647 	/*
6648 	 * We disable interrupts to walk through the probe array.  This is
6649 	 * safe -- the dtrace_sync() in dtrace_unregister() assures that we
6650 	 * won't see stale data.
6651 	 */
6652 	cookie = dtrace_interrupt_disable();
6653 
6654 	for (i = 0; i < dtrace_nprobes; i++) {
6655 		if ((probe = dtrace_probes[i]) == NULL)
6656 			continue;
6657 
6658 		if (probe->dtpr_ecb == NULL) {
6659 			/*
6660 			 * This probe isn't enabled -- don't call the function.
6661 			 */
6662 			continue;
6663 		}
6664 
6665 		prov = probe->dtpr_provider;
6666 		func = *((void(**)(void *, dtrace_id_t, void *))
6667 		    ((uintptr_t)&prov->dtpv_pops + offs));
6668 
6669 		func(prov->dtpv_arg, i + 1, probe->dtpr_arg);
6670 	}
6671 
6672 	dtrace_interrupt_enable(cookie);
6673 }
6674 
6675 static int
6676 dtrace_probe_enable(const dtrace_probedesc_t *desc, dtrace_enabling_t *enab)
6677 {
6678 	dtrace_probekey_t pkey;
6679 	uint32_t priv;
6680 	uid_t uid;
6681 	zoneid_t zoneid;
6682 
6683 	ASSERT(MUTEX_HELD(&dtrace_lock));
6684 	dtrace_ecb_create_cache = NULL;
6685 
6686 	if (desc == NULL) {
6687 		/*
6688 		 * If we're passed a NULL description, we're being asked to
6689 		 * create an ECB with a NULL probe.
6690 		 */
6691 		(void) dtrace_ecb_create_enable(NULL, enab);
6692 		return (0);
6693 	}
6694 
6695 	dtrace_probekey(desc, &pkey);
6696 	dtrace_cred2priv(enab->dten_vstate->dtvs_state->dts_cred.dcr_cred,
6697 	    &priv, &uid, &zoneid);
6698 
6699 	return (dtrace_match(&pkey, priv, uid, zoneid, dtrace_ecb_create_enable,
6700 	    enab));
6701 }
6702 
6703 /*
6704  * DTrace Helper Provider Functions
6705  */
6706 static void
6707 dtrace_dofattr2attr(dtrace_attribute_t *attr, const dof_attr_t dofattr)
6708 {
6709 	attr->dtat_name = DOF_ATTR_NAME(dofattr);
6710 	attr->dtat_data = DOF_ATTR_DATA(dofattr);
6711 	attr->dtat_class = DOF_ATTR_CLASS(dofattr);
6712 }
6713 
6714 static void
6715 dtrace_dofprov2hprov(dtrace_helper_provdesc_t *hprov,
6716     const dof_provider_t *dofprov, char *strtab)
6717 {
6718 	hprov->dthpv_provname = strtab + dofprov->dofpv_name;
6719 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_provider,
6720 	    dofprov->dofpv_provattr);
6721 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_mod,
6722 	    dofprov->dofpv_modattr);
6723 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_func,
6724 	    dofprov->dofpv_funcattr);
6725 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_name,
6726 	    dofprov->dofpv_nameattr);
6727 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_args,
6728 	    dofprov->dofpv_argsattr);
6729 }
6730 
6731 static void
6732 dtrace_helper_provide_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
6733 {
6734 	uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
6735 	dof_hdr_t *dof = (dof_hdr_t *)daddr;
6736 	dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
6737 	dof_provider_t *provider;
6738 	dof_probe_t *probe;
6739 	uint32_t *off, *enoff;
6740 	uint8_t *arg;
6741 	char *strtab;
6742 	uint_t i, nprobes;
6743 	dtrace_helper_provdesc_t dhpv;
6744 	dtrace_helper_probedesc_t dhpb;
6745 	dtrace_meta_t *meta = dtrace_meta_pid;
6746 	dtrace_mops_t *mops = &meta->dtm_mops;
6747 	void *parg;
6748 
6749 	provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
6750 	str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
6751 	    provider->dofpv_strtab * dof->dofh_secsize);
6752 	prb_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
6753 	    provider->dofpv_probes * dof->dofh_secsize);
6754 	arg_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
6755 	    provider->dofpv_prargs * dof->dofh_secsize);
6756 	off_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
6757 	    provider->dofpv_proffs * dof->dofh_secsize);
6758 
6759 	strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
6760 	off = (uint32_t *)(uintptr_t)(daddr + off_sec->dofs_offset);
6761 	arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
6762 	enoff = NULL;
6763 
6764 	/*
6765 	 * See dtrace_helper_provider_validate().
6766 	 */
6767 	if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
6768 	    provider->dofpv_prenoffs != DOF_SECT_NONE) {
6769 		enoff_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
6770 		    provider->dofpv_prenoffs * dof->dofh_secsize);
6771 		enoff = (uint32_t *)(uintptr_t)(daddr + enoff_sec->dofs_offset);
6772 	}
6773 
6774 	nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
6775 
6776 	/*
6777 	 * Create the provider.
6778 	 */
6779 	dtrace_dofprov2hprov(&dhpv, provider, strtab);
6780 
6781 	if ((parg = mops->dtms_provide_pid(meta->dtm_arg, &dhpv, pid)) == NULL)
6782 		return;
6783 
6784 	meta->dtm_count++;
6785 
6786 	/*
6787 	 * Create the probes.
6788 	 */
6789 	for (i = 0; i < nprobes; i++) {
6790 		probe = (dof_probe_t *)(uintptr_t)(daddr +
6791 		    prb_sec->dofs_offset + i * prb_sec->dofs_entsize);
6792 
6793 		dhpb.dthpb_mod = dhp->dofhp_mod;
6794 		dhpb.dthpb_func = strtab + probe->dofpr_func;
6795 		dhpb.dthpb_name = strtab + probe->dofpr_name;
6796 		dhpb.dthpb_base = probe->dofpr_addr;
6797 		dhpb.dthpb_offs = off + probe->dofpr_offidx;
6798 		dhpb.dthpb_noffs = probe->dofpr_noffs;
6799 		if (enoff != NULL) {
6800 			dhpb.dthpb_enoffs = enoff + probe->dofpr_enoffidx;
6801 			dhpb.dthpb_nenoffs = probe->dofpr_nenoffs;
6802 		} else {
6803 			dhpb.dthpb_enoffs = NULL;
6804 			dhpb.dthpb_nenoffs = 0;
6805 		}
6806 		dhpb.dthpb_args = arg + probe->dofpr_argidx;
6807 		dhpb.dthpb_nargc = probe->dofpr_nargc;
6808 		dhpb.dthpb_xargc = probe->dofpr_xargc;
6809 		dhpb.dthpb_ntypes = strtab + probe->dofpr_nargv;
6810 		dhpb.dthpb_xtypes = strtab + probe->dofpr_xargv;
6811 
6812 		mops->dtms_create_probe(meta->dtm_arg, parg, &dhpb);
6813 	}
6814 }
6815 
6816 static void
6817 dtrace_helper_provide(dof_helper_t *dhp, pid_t pid)
6818 {
6819 	uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
6820 	dof_hdr_t *dof = (dof_hdr_t *)daddr;
6821 	int i;
6822 
6823 	ASSERT(MUTEX_HELD(&dtrace_meta_lock));
6824 
6825 	for (i = 0; i < dof->dofh_secnum; i++) {
6826 		dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
6827 		    dof->dofh_secoff + i * dof->dofh_secsize);
6828 
6829 		if (sec->dofs_type != DOF_SECT_PROVIDER)
6830 			continue;
6831 
6832 		dtrace_helper_provide_one(dhp, sec, pid);
6833 	}
6834 
6835 	/*
6836 	 * We may have just created probes, so we must now rematch against
6837 	 * any retained enablings.  Note that this call will acquire both
6838 	 * cpu_lock and dtrace_lock; the fact that we are holding
6839 	 * dtrace_meta_lock now is what defines the ordering with respect to
6840 	 * these three locks.
6841 	 */
6842 	dtrace_enabling_matchall();
6843 }
6844 
6845 static void
6846 dtrace_helper_provider_remove_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
6847 {
6848 	uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
6849 	dof_hdr_t *dof = (dof_hdr_t *)daddr;
6850 	dof_sec_t *str_sec;
6851 	dof_provider_t *provider;
6852 	char *strtab;
6853 	dtrace_helper_provdesc_t dhpv;
6854 	dtrace_meta_t *meta = dtrace_meta_pid;
6855 	dtrace_mops_t *mops = &meta->dtm_mops;
6856 
6857 	provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
6858 	str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
6859 	    provider->dofpv_strtab * dof->dofh_secsize);
6860 
6861 	strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
6862 
6863 	/*
6864 	 * Create the provider.
6865 	 */
6866 	dtrace_dofprov2hprov(&dhpv, provider, strtab);
6867 
6868 	mops->dtms_remove_pid(meta->dtm_arg, &dhpv, pid);
6869 
6870 	meta->dtm_count--;
6871 }
6872 
6873 static void
6874 dtrace_helper_provider_remove(dof_helper_t *dhp, pid_t pid)
6875 {
6876 	uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
6877 	dof_hdr_t *dof = (dof_hdr_t *)daddr;
6878 	int i;
6879 
6880 	ASSERT(MUTEX_HELD(&dtrace_meta_lock));
6881 
6882 	for (i = 0; i < dof->dofh_secnum; i++) {
6883 		dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
6884 		    dof->dofh_secoff + i * dof->dofh_secsize);
6885 
6886 		if (sec->dofs_type != DOF_SECT_PROVIDER)
6887 			continue;
6888 
6889 		dtrace_helper_provider_remove_one(dhp, sec, pid);
6890 	}
6891 }
6892 
6893 /*
6894  * DTrace Meta Provider-to-Framework API Functions
6895  *
6896  * These functions implement the Meta Provider-to-Framework API, as described
6897  * in <sys/dtrace.h>.
6898  */
6899 int
6900 dtrace_meta_register(const char *name, const dtrace_mops_t *mops, void *arg,
6901     dtrace_meta_provider_id_t *idp)
6902 {
6903 	dtrace_meta_t *meta;
6904 	dtrace_helpers_t *help, *next;
6905 	int i;
6906 
6907 	*idp = DTRACE_METAPROVNONE;
6908 
6909 	/*
6910 	 * We strictly don't need the name, but we hold onto it for
6911 	 * debuggability. All hail error queues!
6912 	 */
6913 	if (name == NULL) {
6914 		cmn_err(CE_WARN, "failed to register meta-provider: "
6915 		    "invalid name");
6916 		return (EINVAL);
6917 	}
6918 
6919 	if (mops == NULL ||
6920 	    mops->dtms_create_probe == NULL ||
6921 	    mops->dtms_provide_pid == NULL ||
6922 	    mops->dtms_remove_pid == NULL) {
6923 		cmn_err(CE_WARN, "failed to register meta-register %s: "
6924 		    "invalid ops", name);
6925 		return (EINVAL);
6926 	}
6927 
6928 	meta = kmem_zalloc(sizeof (dtrace_meta_t), KM_SLEEP);
6929 	meta->dtm_mops = *mops;
6930 	meta->dtm_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
6931 	(void) strcpy(meta->dtm_name, name);
6932 	meta->dtm_arg = arg;
6933 
6934 	mutex_enter(&dtrace_meta_lock);
6935 	mutex_enter(&dtrace_lock);
6936 
6937 	if (dtrace_meta_pid != NULL) {
6938 		mutex_exit(&dtrace_lock);
6939 		mutex_exit(&dtrace_meta_lock);
6940 		cmn_err(CE_WARN, "failed to register meta-register %s: "
6941 		    "user-land meta-provider exists", name);
6942 		kmem_free(meta->dtm_name, strlen(meta->dtm_name) + 1);
6943 		kmem_free(meta, sizeof (dtrace_meta_t));
6944 		return (EINVAL);
6945 	}
6946 
6947 	dtrace_meta_pid = meta;
6948 	*idp = (dtrace_meta_provider_id_t)meta;
6949 
6950 	/*
6951 	 * If there are providers and probes ready to go, pass them
6952 	 * off to the new meta provider now.
6953 	 */
6954 
6955 	help = dtrace_deferred_pid;
6956 	dtrace_deferred_pid = NULL;
6957 
6958 	mutex_exit(&dtrace_lock);
6959 
6960 	while (help != NULL) {
6961 		for (i = 0; i < help->dthps_nprovs; i++) {
6962 			dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
6963 			    help->dthps_pid);
6964 		}
6965 
6966 		next = help->dthps_next;
6967 		help->dthps_next = NULL;
6968 		help->dthps_prev = NULL;
6969 		help->dthps_deferred = 0;
6970 		help = next;
6971 	}
6972 
6973 	mutex_exit(&dtrace_meta_lock);
6974 
6975 	return (0);
6976 }
6977 
6978 int
6979 dtrace_meta_unregister(dtrace_meta_provider_id_t id)
6980 {
6981 	dtrace_meta_t **pp, *old = (dtrace_meta_t *)id;
6982 
6983 	mutex_enter(&dtrace_meta_lock);
6984 	mutex_enter(&dtrace_lock);
6985 
6986 	if (old == dtrace_meta_pid) {
6987 		pp = &dtrace_meta_pid;
6988 	} else {
6989 		panic("attempt to unregister non-existent "
6990 		    "dtrace meta-provider %p\n", (void *)old);
6991 	}
6992 
6993 	if (old->dtm_count != 0) {
6994 		mutex_exit(&dtrace_lock);
6995 		mutex_exit(&dtrace_meta_lock);
6996 		return (EBUSY);
6997 	}
6998 
6999 	*pp = NULL;
7000 
7001 	mutex_exit(&dtrace_lock);
7002 	mutex_exit(&dtrace_meta_lock);
7003 
7004 	kmem_free(old->dtm_name, strlen(old->dtm_name) + 1);
7005 	kmem_free(old, sizeof (dtrace_meta_t));
7006 
7007 	return (0);
7008 }
7009 
7010 
7011 /*
7012  * DTrace DIF Object Functions
7013  */
7014 static int
7015 dtrace_difo_err(uint_t pc, const char *format, ...)
7016 {
7017 	if (dtrace_err_verbose) {
7018 		va_list alist;
7019 
7020 		(void) uprintf("dtrace DIF object error: [%u]: ", pc);
7021 		va_start(alist, format);
7022 		(void) vuprintf(format, alist);
7023 		va_end(alist);
7024 	}
7025 
7026 #ifdef DTRACE_ERRDEBUG
7027 	dtrace_errdebug(format);
7028 #endif
7029 	return (1);
7030 }
7031 
7032 /*
7033  * Validate a DTrace DIF object by checking the IR instructions.  The following
7034  * rules are currently enforced by dtrace_difo_validate():
7035  *
7036  * 1. Each instruction must have a valid opcode
7037  * 2. Each register, string, variable, or subroutine reference must be valid
7038  * 3. No instruction can modify register %r0 (must be zero)
7039  * 4. All instruction reserved bits must be set to zero
7040  * 5. The last instruction must be a "ret" instruction
7041  * 6. All branch targets must reference a valid instruction _after_ the branch
7042  */
7043 static int
7044 dtrace_difo_validate(dtrace_difo_t *dp, dtrace_vstate_t *vstate, uint_t nregs,
7045     cred_t *cr)
7046 {
7047 	int err = 0, i;
7048 	int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
7049 	int kcheck;
7050 	uint_t pc;
7051 
7052 	kcheck = cr == NULL ||
7053 	    PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE) == 0;
7054 
7055 	dp->dtdo_destructive = 0;
7056 
7057 	for (pc = 0; pc < dp->dtdo_len && err == 0; pc++) {
7058 		dif_instr_t instr = dp->dtdo_buf[pc];
7059 
7060 		uint_t r1 = DIF_INSTR_R1(instr);
7061 		uint_t r2 = DIF_INSTR_R2(instr);
7062 		uint_t rd = DIF_INSTR_RD(instr);
7063 		uint_t rs = DIF_INSTR_RS(instr);
7064 		uint_t label = DIF_INSTR_LABEL(instr);
7065 		uint_t v = DIF_INSTR_VAR(instr);
7066 		uint_t subr = DIF_INSTR_SUBR(instr);
7067 		uint_t type = DIF_INSTR_TYPE(instr);
7068 		uint_t op = DIF_INSTR_OP(instr);
7069 
7070 		switch (op) {
7071 		case DIF_OP_OR:
7072 		case DIF_OP_XOR:
7073 		case DIF_OP_AND:
7074 		case DIF_OP_SLL:
7075 		case DIF_OP_SRL:
7076 		case DIF_OP_SRA:
7077 		case DIF_OP_SUB:
7078 		case DIF_OP_ADD:
7079 		case DIF_OP_MUL:
7080 		case DIF_OP_SDIV:
7081 		case DIF_OP_UDIV:
7082 		case DIF_OP_SREM:
7083 		case DIF_OP_UREM:
7084 		case DIF_OP_COPYS:
7085 			if (r1 >= nregs)
7086 				err += efunc(pc, "invalid register %u\n", r1);
7087 			if (r2 >= nregs)
7088 				err += efunc(pc, "invalid register %u\n", r2);
7089 			if (rd >= nregs)
7090 				err += efunc(pc, "invalid register %u\n", rd);
7091 			if (rd == 0)
7092 				err += efunc(pc, "cannot write to %r0\n");
7093 			break;
7094 		case DIF_OP_NOT:
7095 		case DIF_OP_MOV:
7096 		case DIF_OP_ALLOCS:
7097 			if (r1 >= nregs)
7098 				err += efunc(pc, "invalid register %u\n", r1);
7099 			if (r2 != 0)
7100 				err += efunc(pc, "non-zero reserved bits\n");
7101 			if (rd >= nregs)
7102 				err += efunc(pc, "invalid register %u\n", rd);
7103 			if (rd == 0)
7104 				err += efunc(pc, "cannot write to %r0\n");
7105 			break;
7106 		case DIF_OP_LDSB:
7107 		case DIF_OP_LDSH:
7108 		case DIF_OP_LDSW:
7109 		case DIF_OP_LDUB:
7110 		case DIF_OP_LDUH:
7111 		case DIF_OP_LDUW:
7112 		case DIF_OP_LDX:
7113 			if (r1 >= nregs)
7114 				err += efunc(pc, "invalid register %u\n", r1);
7115 			if (r2 != 0)
7116 				err += efunc(pc, "non-zero reserved bits\n");
7117 			if (rd >= nregs)
7118 				err += efunc(pc, "invalid register %u\n", rd);
7119 			if (rd == 0)
7120 				err += efunc(pc, "cannot write to %r0\n");
7121 			if (kcheck)
7122 				dp->dtdo_buf[pc] = DIF_INSTR_LOAD(op +
7123 				    DIF_OP_RLDSB - DIF_OP_LDSB, r1, rd);
7124 			break;
7125 		case DIF_OP_RLDSB:
7126 		case DIF_OP_RLDSH:
7127 		case DIF_OP_RLDSW:
7128 		case DIF_OP_RLDUB:
7129 		case DIF_OP_RLDUH:
7130 		case DIF_OP_RLDUW:
7131 		case DIF_OP_RLDX:
7132 			if (r1 >= nregs)
7133 				err += efunc(pc, "invalid register %u\n", r1);
7134 			if (r2 != 0)
7135 				err += efunc(pc, "non-zero reserved bits\n");
7136 			if (rd >= nregs)
7137 				err += efunc(pc, "invalid register %u\n", rd);
7138 			if (rd == 0)
7139 				err += efunc(pc, "cannot write to %r0\n");
7140 			break;
7141 		case DIF_OP_ULDSB:
7142 		case DIF_OP_ULDSH:
7143 		case DIF_OP_ULDSW:
7144 		case DIF_OP_ULDUB:
7145 		case DIF_OP_ULDUH:
7146 		case DIF_OP_ULDUW:
7147 		case DIF_OP_ULDX:
7148 			if (r1 >= nregs)
7149 				err += efunc(pc, "invalid register %u\n", r1);
7150 			if (r2 != 0)
7151 				err += efunc(pc, "non-zero reserved bits\n");
7152 			if (rd >= nregs)
7153 				err += efunc(pc, "invalid register %u\n", rd);
7154 			if (rd == 0)
7155 				err += efunc(pc, "cannot write to %r0\n");
7156 			break;
7157 		case DIF_OP_STB:
7158 		case DIF_OP_STH:
7159 		case DIF_OP_STW:
7160 		case DIF_OP_STX:
7161 			if (r1 >= nregs)
7162 				err += efunc(pc, "invalid register %u\n", r1);
7163 			if (r2 != 0)
7164 				err += efunc(pc, "non-zero reserved bits\n");
7165 			if (rd >= nregs)
7166 				err += efunc(pc, "invalid register %u\n", rd);
7167 			if (rd == 0)
7168 				err += efunc(pc, "cannot write to 0 address\n");
7169 			break;
7170 		case DIF_OP_CMP:
7171 		case DIF_OP_SCMP:
7172 			if (r1 >= nregs)
7173 				err += efunc(pc, "invalid register %u\n", r1);
7174 			if (r2 >= nregs)
7175 				err += efunc(pc, "invalid register %u\n", r2);
7176 			if (rd != 0)
7177 				err += efunc(pc, "non-zero reserved bits\n");
7178 			break;
7179 		case DIF_OP_TST:
7180 			if (r1 >= nregs)
7181 				err += efunc(pc, "invalid register %u\n", r1);
7182 			if (r2 != 0 || rd != 0)
7183 				err += efunc(pc, "non-zero reserved bits\n");
7184 			break;
7185 		case DIF_OP_BA:
7186 		case DIF_OP_BE:
7187 		case DIF_OP_BNE:
7188 		case DIF_OP_BG:
7189 		case DIF_OP_BGU:
7190 		case DIF_OP_BGE:
7191 		case DIF_OP_BGEU:
7192 		case DIF_OP_BL:
7193 		case DIF_OP_BLU:
7194 		case DIF_OP_BLE:
7195 		case DIF_OP_BLEU:
7196 			if (label >= dp->dtdo_len) {
7197 				err += efunc(pc, "invalid branch target %u\n",
7198 				    label);
7199 			}
7200 			if (label <= pc) {
7201 				err += efunc(pc, "backward branch to %u\n",
7202 				    label);
7203 			}
7204 			break;
7205 		case DIF_OP_RET:
7206 			if (r1 != 0 || r2 != 0)
7207 				err += efunc(pc, "non-zero reserved bits\n");
7208 			if (rd >= nregs)
7209 				err += efunc(pc, "invalid register %u\n", rd);
7210 			break;
7211 		case DIF_OP_NOP:
7212 		case DIF_OP_POPTS:
7213 		case DIF_OP_FLUSHTS:
7214 			if (r1 != 0 || r2 != 0 || rd != 0)
7215 				err += efunc(pc, "non-zero reserved bits\n");
7216 			break;
7217 		case DIF_OP_SETX:
7218 			if (DIF_INSTR_INTEGER(instr) >= dp->dtdo_intlen) {
7219 				err += efunc(pc, "invalid integer ref %u\n",
7220 				    DIF_INSTR_INTEGER(instr));
7221 			}
7222 			if (rd >= nregs)
7223 				err += efunc(pc, "invalid register %u\n", rd);
7224 			if (rd == 0)
7225 				err += efunc(pc, "cannot write to %r0\n");
7226 			break;
7227 		case DIF_OP_SETS:
7228 			if (DIF_INSTR_STRING(instr) >= dp->dtdo_strlen) {
7229 				err += efunc(pc, "invalid string ref %u\n",
7230 				    DIF_INSTR_STRING(instr));
7231 			}
7232 			if (rd >= nregs)
7233 				err += efunc(pc, "invalid register %u\n", rd);
7234 			if (rd == 0)
7235 				err += efunc(pc, "cannot write to %r0\n");
7236 			break;
7237 		case DIF_OP_LDGA:
7238 		case DIF_OP_LDTA:
7239 			if (r1 > DIF_VAR_ARRAY_MAX)
7240 				err += efunc(pc, "invalid array %u\n", r1);
7241 			if (r2 >= nregs)
7242 				err += efunc(pc, "invalid register %u\n", r2);
7243 			if (rd >= nregs)
7244 				err += efunc(pc, "invalid register %u\n", rd);
7245 			if (rd == 0)
7246 				err += efunc(pc, "cannot write to %r0\n");
7247 			break;
7248 		case DIF_OP_LDGS:
7249 		case DIF_OP_LDTS:
7250 		case DIF_OP_LDLS:
7251 		case DIF_OP_LDGAA:
7252 		case DIF_OP_LDTAA:
7253 			if (v < DIF_VAR_OTHER_MIN || v > DIF_VAR_OTHER_MAX)
7254 				err += efunc(pc, "invalid variable %u\n", v);
7255 			if (rd >= nregs)
7256 				err += efunc(pc, "invalid register %u\n", rd);
7257 			if (rd == 0)
7258 				err += efunc(pc, "cannot write to %r0\n");
7259 			break;
7260 		case DIF_OP_STGS:
7261 		case DIF_OP_STTS:
7262 		case DIF_OP_STLS:
7263 		case DIF_OP_STGAA:
7264 		case DIF_OP_STTAA:
7265 			if (v < DIF_VAR_OTHER_UBASE || v > DIF_VAR_OTHER_MAX)
7266 				err += efunc(pc, "invalid variable %u\n", v);
7267 			if (rs >= nregs)
7268 				err += efunc(pc, "invalid register %u\n", rd);
7269 			break;
7270 		case DIF_OP_CALL:
7271 			if (subr > DIF_SUBR_MAX)
7272 				err += efunc(pc, "invalid subr %u\n", subr);
7273 			if (rd >= nregs)
7274 				err += efunc(pc, "invalid register %u\n", rd);
7275 			if (rd == 0)
7276 				err += efunc(pc, "cannot write to %r0\n");
7277 
7278 			if (subr == DIF_SUBR_COPYOUT ||
7279 			    subr == DIF_SUBR_COPYOUTSTR) {
7280 				dp->dtdo_destructive = 1;
7281 			}
7282 			break;
7283 		case DIF_OP_PUSHTR:
7284 			if (type != DIF_TYPE_STRING && type != DIF_TYPE_CTF)
7285 				err += efunc(pc, "invalid ref type %u\n", type);
7286 			if (r2 >= nregs)
7287 				err += efunc(pc, "invalid register %u\n", r2);
7288 			if (rs >= nregs)
7289 				err += efunc(pc, "invalid register %u\n", rs);
7290 			break;
7291 		case DIF_OP_PUSHTV:
7292 			if (type != DIF_TYPE_CTF)
7293 				err += efunc(pc, "invalid val type %u\n", type);
7294 			if (r2 >= nregs)
7295 				err += efunc(pc, "invalid register %u\n", r2);
7296 			if (rs >= nregs)
7297 				err += efunc(pc, "invalid register %u\n", rs);
7298 			break;
7299 		default:
7300 			err += efunc(pc, "invalid opcode %u\n",
7301 			    DIF_INSTR_OP(instr));
7302 		}
7303 	}
7304 
7305 	if (dp->dtdo_len != 0 &&
7306 	    DIF_INSTR_OP(dp->dtdo_buf[dp->dtdo_len - 1]) != DIF_OP_RET) {
7307 		err += efunc(dp->dtdo_len - 1,
7308 		    "expected 'ret' as last DIF instruction\n");
7309 	}
7310 
7311 	if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF)) {
7312 		/*
7313 		 * If we're not returning by reference, the size must be either
7314 		 * 0 or the size of one of the base types.
7315 		 */
7316 		switch (dp->dtdo_rtype.dtdt_size) {
7317 		case 0:
7318 		case sizeof (uint8_t):
7319 		case sizeof (uint16_t):
7320 		case sizeof (uint32_t):
7321 		case sizeof (uint64_t):
7322 			break;
7323 
7324 		default:
7325 			err += efunc(dp->dtdo_len - 1, "bad return size");
7326 		}
7327 	}
7328 
7329 	for (i = 0; i < dp->dtdo_varlen && err == 0; i++) {
7330 		dtrace_difv_t *v = &dp->dtdo_vartab[i], *existing = NULL;
7331 		dtrace_diftype_t *vt, *et;
7332 		uint_t id, ndx;
7333 
7334 		if (v->dtdv_scope != DIFV_SCOPE_GLOBAL &&
7335 		    v->dtdv_scope != DIFV_SCOPE_THREAD &&
7336 		    v->dtdv_scope != DIFV_SCOPE_LOCAL) {
7337 			err += efunc(i, "unrecognized variable scope %d\n",
7338 			    v->dtdv_scope);
7339 			break;
7340 		}
7341 
7342 		if (v->dtdv_kind != DIFV_KIND_ARRAY &&
7343 		    v->dtdv_kind != DIFV_KIND_SCALAR) {
7344 			err += efunc(i, "unrecognized variable type %d\n",
7345 			    v->dtdv_kind);
7346 			break;
7347 		}
7348 
7349 		if ((id = v->dtdv_id) > DIF_VARIABLE_MAX) {
7350 			err += efunc(i, "%d exceeds variable id limit\n", id);
7351 			break;
7352 		}
7353 
7354 		if (id < DIF_VAR_OTHER_UBASE)
7355 			continue;
7356 
7357 		/*
7358 		 * For user-defined variables, we need to check that this
7359 		 * definition is identical to any previous definition that we
7360 		 * encountered.
7361 		 */
7362 		ndx = id - DIF_VAR_OTHER_UBASE;
7363 
7364 		switch (v->dtdv_scope) {
7365 		case DIFV_SCOPE_GLOBAL:
7366 			if (ndx < vstate->dtvs_nglobals) {
7367 				dtrace_statvar_t *svar;
7368 
7369 				if ((svar = vstate->dtvs_globals[ndx]) != NULL)
7370 					existing = &svar->dtsv_var;
7371 			}
7372 
7373 			break;
7374 
7375 		case DIFV_SCOPE_THREAD:
7376 			if (ndx < vstate->dtvs_ntlocals)
7377 				existing = &vstate->dtvs_tlocals[ndx];
7378 			break;
7379 
7380 		case DIFV_SCOPE_LOCAL:
7381 			if (ndx < vstate->dtvs_nlocals) {
7382 				dtrace_statvar_t *svar;
7383 
7384 				if ((svar = vstate->dtvs_locals[ndx]) != NULL)
7385 					existing = &svar->dtsv_var;
7386 			}
7387 
7388 			break;
7389 		}
7390 
7391 		vt = &v->dtdv_type;
7392 
7393 		if (vt->dtdt_flags & DIF_TF_BYREF) {
7394 			if (vt->dtdt_size == 0) {
7395 				err += efunc(i, "zero-sized variable\n");
7396 				break;
7397 			}
7398 
7399 			if (v->dtdv_scope == DIFV_SCOPE_GLOBAL &&
7400 			    vt->dtdt_size > dtrace_global_maxsize) {
7401 				err += efunc(i, "oversized by-ref global\n");
7402 				break;
7403 			}
7404 		}
7405 
7406 		if (existing == NULL || existing->dtdv_id == 0)
7407 			continue;
7408 
7409 		ASSERT(existing->dtdv_id == v->dtdv_id);
7410 		ASSERT(existing->dtdv_scope == v->dtdv_scope);
7411 
7412 		if (existing->dtdv_kind != v->dtdv_kind)
7413 			err += efunc(i, "%d changed variable kind\n", id);
7414 
7415 		et = &existing->dtdv_type;
7416 
7417 		if (vt->dtdt_flags != et->dtdt_flags) {
7418 			err += efunc(i, "%d changed variable type flags\n", id);
7419 			break;
7420 		}
7421 
7422 		if (vt->dtdt_size != 0 && vt->dtdt_size != et->dtdt_size) {
7423 			err += efunc(i, "%d changed variable type size\n", id);
7424 			break;
7425 		}
7426 	}
7427 
7428 	return (err);
7429 }
7430 
7431 /*
7432  * Validate a DTrace DIF object that it is to be used as a helper.  Helpers
7433  * are much more constrained than normal DIFOs.  Specifically, they may
7434  * not:
7435  *
7436  * 1. Make calls to subroutines other than copyin(), copyinstr() or
7437  *    miscellaneous string routines
7438  * 2. Access DTrace variables other than the args[] array, and the
7439  *    curthread, pid, ppid, tid, execname, zonename, uid and gid variables.
7440  * 3. Have thread-local variables.
7441  * 4. Have dynamic variables.
7442  */
7443 static int
7444 dtrace_difo_validate_helper(dtrace_difo_t *dp)
7445 {
7446 	int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
7447 	int err = 0;
7448 	uint_t pc;
7449 
7450 	for (pc = 0; pc < dp->dtdo_len; pc++) {
7451 		dif_instr_t instr = dp->dtdo_buf[pc];
7452 
7453 		uint_t v = DIF_INSTR_VAR(instr);
7454 		uint_t subr = DIF_INSTR_SUBR(instr);
7455 		uint_t op = DIF_INSTR_OP(instr);
7456 
7457 		switch (op) {
7458 		case DIF_OP_OR:
7459 		case DIF_OP_XOR:
7460 		case DIF_OP_AND:
7461 		case DIF_OP_SLL:
7462 		case DIF_OP_SRL:
7463 		case DIF_OP_SRA:
7464 		case DIF_OP_SUB:
7465 		case DIF_OP_ADD:
7466 		case DIF_OP_MUL:
7467 		case DIF_OP_SDIV:
7468 		case DIF_OP_UDIV:
7469 		case DIF_OP_SREM:
7470 		case DIF_OP_UREM:
7471 		case DIF_OP_COPYS:
7472 		case DIF_OP_NOT:
7473 		case DIF_OP_MOV:
7474 		case DIF_OP_RLDSB:
7475 		case DIF_OP_RLDSH:
7476 		case DIF_OP_RLDSW:
7477 		case DIF_OP_RLDUB:
7478 		case DIF_OP_RLDUH:
7479 		case DIF_OP_RLDUW:
7480 		case DIF_OP_RLDX:
7481 		case DIF_OP_ULDSB:
7482 		case DIF_OP_ULDSH:
7483 		case DIF_OP_ULDSW:
7484 		case DIF_OP_ULDUB:
7485 		case DIF_OP_ULDUH:
7486 		case DIF_OP_ULDUW:
7487 		case DIF_OP_ULDX:
7488 		case DIF_OP_STB:
7489 		case DIF_OP_STH:
7490 		case DIF_OP_STW:
7491 		case DIF_OP_STX:
7492 		case DIF_OP_ALLOCS:
7493 		case DIF_OP_CMP:
7494 		case DIF_OP_SCMP:
7495 		case DIF_OP_TST:
7496 		case DIF_OP_BA:
7497 		case DIF_OP_BE:
7498 		case DIF_OP_BNE:
7499 		case DIF_OP_BG:
7500 		case DIF_OP_BGU:
7501 		case DIF_OP_BGE:
7502 		case DIF_OP_BGEU:
7503 		case DIF_OP_BL:
7504 		case DIF_OP_BLU:
7505 		case DIF_OP_BLE:
7506 		case DIF_OP_BLEU:
7507 		case DIF_OP_RET:
7508 		case DIF_OP_NOP:
7509 		case DIF_OP_POPTS:
7510 		case DIF_OP_FLUSHTS:
7511 		case DIF_OP_SETX:
7512 		case DIF_OP_SETS:
7513 		case DIF_OP_LDGA:
7514 		case DIF_OP_LDLS:
7515 		case DIF_OP_STGS:
7516 		case DIF_OP_STLS:
7517 		case DIF_OP_PUSHTR:
7518 		case DIF_OP_PUSHTV:
7519 			break;
7520 
7521 		case DIF_OP_LDGS:
7522 			if (v >= DIF_VAR_OTHER_UBASE)
7523 				break;
7524 
7525 			if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9)
7526 				break;
7527 
7528 			if (v == DIF_VAR_CURTHREAD || v == DIF_VAR_PID ||
7529 			    v == DIF_VAR_PPID || v == DIF_VAR_TID ||
7530 			    v == DIF_VAR_EXECNAME || v == DIF_VAR_ZONENAME ||
7531 			    v == DIF_VAR_UID || v == DIF_VAR_GID)
7532 				break;
7533 
7534 			err += efunc(pc, "illegal variable %u\n", v);
7535 			break;
7536 
7537 		case DIF_OP_LDTA:
7538 		case DIF_OP_LDTS:
7539 		case DIF_OP_LDGAA:
7540 		case DIF_OP_LDTAA:
7541 			err += efunc(pc, "illegal dynamic variable load\n");
7542 			break;
7543 
7544 		case DIF_OP_STTS:
7545 		case DIF_OP_STGAA:
7546 		case DIF_OP_STTAA:
7547 			err += efunc(pc, "illegal dynamic variable store\n");
7548 			break;
7549 
7550 		case DIF_OP_CALL:
7551 			if (subr == DIF_SUBR_ALLOCA ||
7552 			    subr == DIF_SUBR_BCOPY ||
7553 			    subr == DIF_SUBR_COPYIN ||
7554 			    subr == DIF_SUBR_COPYINTO ||
7555 			    subr == DIF_SUBR_COPYINSTR ||
7556 			    subr == DIF_SUBR_INDEX ||
7557 			    subr == DIF_SUBR_LLTOSTR ||
7558 			    subr == DIF_SUBR_RINDEX ||
7559 			    subr == DIF_SUBR_STRCHR ||
7560 			    subr == DIF_SUBR_STRJOIN ||
7561 			    subr == DIF_SUBR_STRRCHR ||
7562 			    subr == DIF_SUBR_STRSTR)
7563 				break;
7564 
7565 			err += efunc(pc, "invalid subr %u\n", subr);
7566 			break;
7567 
7568 		default:
7569 			err += efunc(pc, "invalid opcode %u\n",
7570 			    DIF_INSTR_OP(instr));
7571 		}
7572 	}
7573 
7574 	return (err);
7575 }
7576 
7577 /*
7578  * Returns 1 if the expression in the DIF object can be cached on a per-thread
7579  * basis; 0 if not.
7580  */
7581 static int
7582 dtrace_difo_cacheable(dtrace_difo_t *dp)
7583 {
7584 	int i;
7585 
7586 	if (dp == NULL)
7587 		return (0);
7588 
7589 	for (i = 0; i < dp->dtdo_varlen; i++) {
7590 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
7591 
7592 		if (v->dtdv_scope != DIFV_SCOPE_GLOBAL)
7593 			continue;
7594 
7595 		switch (v->dtdv_id) {
7596 		case DIF_VAR_CURTHREAD:
7597 		case DIF_VAR_PID:
7598 		case DIF_VAR_TID:
7599 		case DIF_VAR_EXECNAME:
7600 		case DIF_VAR_ZONENAME:
7601 			break;
7602 
7603 		default:
7604 			return (0);
7605 		}
7606 	}
7607 
7608 	/*
7609 	 * This DIF object may be cacheable.  Now we need to look for any
7610 	 * array loading instructions, any memory loading instructions, or
7611 	 * any stores to thread-local variables.
7612 	 */
7613 	for (i = 0; i < dp->dtdo_len; i++) {
7614 		uint_t op = DIF_INSTR_OP(dp->dtdo_buf[i]);
7615 
7616 		if ((op >= DIF_OP_LDSB && op <= DIF_OP_LDX) ||
7617 		    (op >= DIF_OP_ULDSB && op <= DIF_OP_ULDX) ||
7618 		    (op >= DIF_OP_RLDSB && op <= DIF_OP_RLDX) ||
7619 		    op == DIF_OP_LDGA || op == DIF_OP_STTS)
7620 			return (0);
7621 	}
7622 
7623 	return (1);
7624 }
7625 
7626 static void
7627 dtrace_difo_hold(dtrace_difo_t *dp)
7628 {
7629 	int i;
7630 
7631 	ASSERT(MUTEX_HELD(&dtrace_lock));
7632 
7633 	dp->dtdo_refcnt++;
7634 	ASSERT(dp->dtdo_refcnt != 0);
7635 
7636 	/*
7637 	 * We need to check this DIF object for references to the variable
7638 	 * DIF_VAR_VTIMESTAMP.
7639 	 */
7640 	for (i = 0; i < dp->dtdo_varlen; i++) {
7641 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
7642 
7643 		if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
7644 			continue;
7645 
7646 		if (dtrace_vtime_references++ == 0)
7647 			dtrace_vtime_enable();
7648 	}
7649 }
7650 
7651 /*
7652  * This routine calculates the dynamic variable chunksize for a given DIF
7653  * object.  The calculation is not fool-proof, and can probably be tricked by
7654  * malicious DIF -- but it works for all compiler-generated DIF.  Because this
7655  * calculation is likely imperfect, dtrace_dynvar() is able to gracefully fail
7656  * if a dynamic variable size exceeds the chunksize.
7657  */
7658 static void
7659 dtrace_difo_chunksize(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
7660 {
7661 	uint64_t sval;
7662 	dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
7663 	const dif_instr_t *text = dp->dtdo_buf;
7664 	uint_t pc, srd = 0;
7665 	uint_t ttop = 0;
7666 	size_t size, ksize;
7667 	uint_t id, i;
7668 
7669 	for (pc = 0; pc < dp->dtdo_len; pc++) {
7670 		dif_instr_t instr = text[pc];
7671 		uint_t op = DIF_INSTR_OP(instr);
7672 		uint_t rd = DIF_INSTR_RD(instr);
7673 		uint_t r1 = DIF_INSTR_R1(instr);
7674 		uint_t nkeys = 0;
7675 		uchar_t scope;
7676 
7677 		dtrace_key_t *key = tupregs;
7678 
7679 		switch (op) {
7680 		case DIF_OP_SETX:
7681 			sval = dp->dtdo_inttab[DIF_INSTR_INTEGER(instr)];
7682 			srd = rd;
7683 			continue;
7684 
7685 		case DIF_OP_STTS:
7686 			key = &tupregs[DIF_DTR_NREGS];
7687 			key[0].dttk_size = 0;
7688 			key[1].dttk_size = 0;
7689 			nkeys = 2;
7690 			scope = DIFV_SCOPE_THREAD;
7691 			break;
7692 
7693 		case DIF_OP_STGAA:
7694 		case DIF_OP_STTAA:
7695 			nkeys = ttop;
7696 
7697 			if (DIF_INSTR_OP(instr) == DIF_OP_STTAA)
7698 				key[nkeys++].dttk_size = 0;
7699 
7700 			key[nkeys++].dttk_size = 0;
7701 
7702 			if (op == DIF_OP_STTAA) {
7703 				scope = DIFV_SCOPE_THREAD;
7704 			} else {
7705 				scope = DIFV_SCOPE_GLOBAL;
7706 			}
7707 
7708 			break;
7709 
7710 		case DIF_OP_PUSHTR:
7711 			if (ttop == DIF_DTR_NREGS)
7712 				return;
7713 
7714 			if ((srd == 0 || sval == 0) && r1 == DIF_TYPE_STRING) {
7715 				/*
7716 				 * If the register for the size of the "pushtr"
7717 				 * is %r0 (or the value is 0) and the type is
7718 				 * a string, we'll use the system-wide default
7719 				 * string size.
7720 				 */
7721 				tupregs[ttop++].dttk_size =
7722 				    dtrace_strsize_default;
7723 			} else {
7724 				if (srd == 0)
7725 					return;
7726 
7727 				tupregs[ttop++].dttk_size = sval;
7728 			}
7729 
7730 			break;
7731 
7732 		case DIF_OP_PUSHTV:
7733 			if (ttop == DIF_DTR_NREGS)
7734 				return;
7735 
7736 			tupregs[ttop++].dttk_size = 0;
7737 			break;
7738 
7739 		case DIF_OP_FLUSHTS:
7740 			ttop = 0;
7741 			break;
7742 
7743 		case DIF_OP_POPTS:
7744 			if (ttop != 0)
7745 				ttop--;
7746 			break;
7747 		}
7748 
7749 		sval = 0;
7750 		srd = 0;
7751 
7752 		if (nkeys == 0)
7753 			continue;
7754 
7755 		/*
7756 		 * We have a dynamic variable allocation; calculate its size.
7757 		 */
7758 		for (ksize = 0, i = 0; i < nkeys; i++)
7759 			ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
7760 
7761 		size = sizeof (dtrace_dynvar_t);
7762 		size += sizeof (dtrace_key_t) * (nkeys - 1);
7763 		size += ksize;
7764 
7765 		/*
7766 		 * Now we need to determine the size of the stored data.
7767 		 */
7768 		id = DIF_INSTR_VAR(instr);
7769 
7770 		for (i = 0; i < dp->dtdo_varlen; i++) {
7771 			dtrace_difv_t *v = &dp->dtdo_vartab[i];
7772 
7773 			if (v->dtdv_id == id && v->dtdv_scope == scope) {
7774 				size += v->dtdv_type.dtdt_size;
7775 				break;
7776 			}
7777 		}
7778 
7779 		if (i == dp->dtdo_varlen)
7780 			return;
7781 
7782 		/*
7783 		 * We have the size.  If this is larger than the chunk size
7784 		 * for our dynamic variable state, reset the chunk size.
7785 		 */
7786 		size = P2ROUNDUP(size, sizeof (uint64_t));
7787 
7788 		if (size > vstate->dtvs_dynvars.dtds_chunksize)
7789 			vstate->dtvs_dynvars.dtds_chunksize = size;
7790 	}
7791 }
7792 
7793 static void
7794 dtrace_difo_init(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
7795 {
7796 	int i, oldsvars, osz, nsz, otlocals, ntlocals;
7797 	uint_t id;
7798 
7799 	ASSERT(MUTEX_HELD(&dtrace_lock));
7800 	ASSERT(dp->dtdo_buf != NULL && dp->dtdo_len != 0);
7801 
7802 	for (i = 0; i < dp->dtdo_varlen; i++) {
7803 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
7804 		dtrace_statvar_t *svar, ***svarp;
7805 		size_t dsize = 0;
7806 		uint8_t scope = v->dtdv_scope;
7807 		int *np;
7808 
7809 		if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
7810 			continue;
7811 
7812 		id -= DIF_VAR_OTHER_UBASE;
7813 
7814 		switch (scope) {
7815 		case DIFV_SCOPE_THREAD:
7816 			while (id >= (otlocals = vstate->dtvs_ntlocals)) {
7817 				dtrace_difv_t *tlocals;
7818 
7819 				if ((ntlocals = (otlocals << 1)) == 0)
7820 					ntlocals = 1;
7821 
7822 				osz = otlocals * sizeof (dtrace_difv_t);
7823 				nsz = ntlocals * sizeof (dtrace_difv_t);
7824 
7825 				tlocals = kmem_zalloc(nsz, KM_SLEEP);
7826 
7827 				if (osz != 0) {
7828 					bcopy(vstate->dtvs_tlocals,
7829 					    tlocals, osz);
7830 					kmem_free(vstate->dtvs_tlocals, osz);
7831 				}
7832 
7833 				vstate->dtvs_tlocals = tlocals;
7834 				vstate->dtvs_ntlocals = ntlocals;
7835 			}
7836 
7837 			vstate->dtvs_tlocals[id] = *v;
7838 			continue;
7839 
7840 		case DIFV_SCOPE_LOCAL:
7841 			np = &vstate->dtvs_nlocals;
7842 			svarp = &vstate->dtvs_locals;
7843 
7844 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
7845 				dsize = NCPU * (v->dtdv_type.dtdt_size +
7846 				    sizeof (uint64_t));
7847 			else
7848 				dsize = NCPU * sizeof (uint64_t);
7849 
7850 			break;
7851 
7852 		case DIFV_SCOPE_GLOBAL:
7853 			np = &vstate->dtvs_nglobals;
7854 			svarp = &vstate->dtvs_globals;
7855 
7856 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
7857 				dsize = v->dtdv_type.dtdt_size +
7858 				    sizeof (uint64_t);
7859 
7860 			break;
7861 
7862 		default:
7863 			ASSERT(0);
7864 		}
7865 
7866 		while (id >= (oldsvars = *np)) {
7867 			dtrace_statvar_t **statics;
7868 			int newsvars, oldsize, newsize;
7869 
7870 			if ((newsvars = (oldsvars << 1)) == 0)
7871 				newsvars = 1;
7872 
7873 			oldsize = oldsvars * sizeof (dtrace_statvar_t *);
7874 			newsize = newsvars * sizeof (dtrace_statvar_t *);
7875 
7876 			statics = kmem_zalloc(newsize, KM_SLEEP);
7877 
7878 			if (oldsize != 0) {
7879 				bcopy(*svarp, statics, oldsize);
7880 				kmem_free(*svarp, oldsize);
7881 			}
7882 
7883 			*svarp = statics;
7884 			*np = newsvars;
7885 		}
7886 
7887 		if ((svar = (*svarp)[id]) == NULL) {
7888 			svar = kmem_zalloc(sizeof (dtrace_statvar_t), KM_SLEEP);
7889 			svar->dtsv_var = *v;
7890 
7891 			if ((svar->dtsv_size = dsize) != 0) {
7892 				svar->dtsv_data = (uint64_t)(uintptr_t)
7893 				    kmem_zalloc(dsize, KM_SLEEP);
7894 			}
7895 
7896 			(*svarp)[id] = svar;
7897 		}
7898 
7899 		svar->dtsv_refcnt++;
7900 	}
7901 
7902 	dtrace_difo_chunksize(dp, vstate);
7903 	dtrace_difo_hold(dp);
7904 }
7905 
7906 static dtrace_difo_t *
7907 dtrace_difo_duplicate(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
7908 {
7909 	dtrace_difo_t *new;
7910 	size_t sz;
7911 
7912 	ASSERT(dp->dtdo_buf != NULL);
7913 	ASSERT(dp->dtdo_refcnt != 0);
7914 
7915 	new = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
7916 
7917 	ASSERT(dp->dtdo_buf != NULL);
7918 	sz = dp->dtdo_len * sizeof (dif_instr_t);
7919 	new->dtdo_buf = kmem_alloc(sz, KM_SLEEP);
7920 	bcopy(dp->dtdo_buf, new->dtdo_buf, sz);
7921 	new->dtdo_len = dp->dtdo_len;
7922 
7923 	if (dp->dtdo_strtab != NULL) {
7924 		ASSERT(dp->dtdo_strlen != 0);
7925 		new->dtdo_strtab = kmem_alloc(dp->dtdo_strlen, KM_SLEEP);
7926 		bcopy(dp->dtdo_strtab, new->dtdo_strtab, dp->dtdo_strlen);
7927 		new->dtdo_strlen = dp->dtdo_strlen;
7928 	}
7929 
7930 	if (dp->dtdo_inttab != NULL) {
7931 		ASSERT(dp->dtdo_intlen != 0);
7932 		sz = dp->dtdo_intlen * sizeof (uint64_t);
7933 		new->dtdo_inttab = kmem_alloc(sz, KM_SLEEP);
7934 		bcopy(dp->dtdo_inttab, new->dtdo_inttab, sz);
7935 		new->dtdo_intlen = dp->dtdo_intlen;
7936 	}
7937 
7938 	if (dp->dtdo_vartab != NULL) {
7939 		ASSERT(dp->dtdo_varlen != 0);
7940 		sz = dp->dtdo_varlen * sizeof (dtrace_difv_t);
7941 		new->dtdo_vartab = kmem_alloc(sz, KM_SLEEP);
7942 		bcopy(dp->dtdo_vartab, new->dtdo_vartab, sz);
7943 		new->dtdo_varlen = dp->dtdo_varlen;
7944 	}
7945 
7946 	dtrace_difo_init(new, vstate);
7947 	return (new);
7948 }
7949 
7950 static void
7951 dtrace_difo_destroy(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
7952 {
7953 	int i;
7954 
7955 	ASSERT(dp->dtdo_refcnt == 0);
7956 
7957 	for (i = 0; i < dp->dtdo_varlen; i++) {
7958 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
7959 		dtrace_statvar_t *svar, **svarp;
7960 		uint_t id;
7961 		uint8_t scope = v->dtdv_scope;
7962 		int *np;
7963 
7964 		switch (scope) {
7965 		case DIFV_SCOPE_THREAD:
7966 			continue;
7967 
7968 		case DIFV_SCOPE_LOCAL:
7969 			np = &vstate->dtvs_nlocals;
7970 			svarp = vstate->dtvs_locals;
7971 			break;
7972 
7973 		case DIFV_SCOPE_GLOBAL:
7974 			np = &vstate->dtvs_nglobals;
7975 			svarp = vstate->dtvs_globals;
7976 			break;
7977 
7978 		default:
7979 			ASSERT(0);
7980 		}
7981 
7982 		if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
7983 			continue;
7984 
7985 		id -= DIF_VAR_OTHER_UBASE;
7986 		ASSERT(id < *np);
7987 
7988 		svar = svarp[id];
7989 		ASSERT(svar != NULL);
7990 		ASSERT(svar->dtsv_refcnt > 0);
7991 
7992 		if (--svar->dtsv_refcnt > 0)
7993 			continue;
7994 
7995 		if (svar->dtsv_size != 0) {
7996 			ASSERT(svar->dtsv_data != NULL);
7997 			kmem_free((void *)(uintptr_t)svar->dtsv_data,
7998 			    svar->dtsv_size);
7999 		}
8000 
8001 		kmem_free(svar, sizeof (dtrace_statvar_t));
8002 		svarp[id] = NULL;
8003 	}
8004 
8005 	kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
8006 	kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
8007 	kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
8008 	kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
8009 
8010 	kmem_free(dp, sizeof (dtrace_difo_t));
8011 }
8012 
8013 static void
8014 dtrace_difo_release(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
8015 {
8016 	int i;
8017 
8018 	ASSERT(MUTEX_HELD(&dtrace_lock));
8019 	ASSERT(dp->dtdo_refcnt != 0);
8020 
8021 	for (i = 0; i < dp->dtdo_varlen; i++) {
8022 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
8023 
8024 		if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
8025 			continue;
8026 
8027 		ASSERT(dtrace_vtime_references > 0);
8028 		if (--dtrace_vtime_references == 0)
8029 			dtrace_vtime_disable();
8030 	}
8031 
8032 	if (--dp->dtdo_refcnt == 0)
8033 		dtrace_difo_destroy(dp, vstate);
8034 }
8035 
8036 /*
8037  * DTrace Format Functions
8038  */
8039 static uint16_t
8040 dtrace_format_add(dtrace_state_t *state, char *str)
8041 {
8042 	char *fmt, **new;
8043 	uint16_t ndx, len = strlen(str) + 1;
8044 
8045 	fmt = kmem_zalloc(len, KM_SLEEP);
8046 	bcopy(str, fmt, len);
8047 
8048 	for (ndx = 0; ndx < state->dts_nformats; ndx++) {
8049 		if (state->dts_formats[ndx] == NULL) {
8050 			state->dts_formats[ndx] = fmt;
8051 			return (ndx + 1);
8052 		}
8053 	}
8054 
8055 	if (state->dts_nformats == USHRT_MAX) {
8056 		/*
8057 		 * This is only likely if a denial-of-service attack is being
8058 		 * attempted.  As such, it's okay to fail silently here.
8059 		 */
8060 		kmem_free(fmt, len);
8061 		return (0);
8062 	}
8063 
8064 	/*
8065 	 * For simplicity, we always resize the formats array to be exactly the
8066 	 * number of formats.
8067 	 */
8068 	ndx = state->dts_nformats++;
8069 	new = kmem_alloc((ndx + 1) * sizeof (char *), KM_SLEEP);
8070 
8071 	if (state->dts_formats != NULL) {
8072 		ASSERT(ndx != 0);
8073 		bcopy(state->dts_formats, new, ndx * sizeof (char *));
8074 		kmem_free(state->dts_formats, ndx * sizeof (char *));
8075 	}
8076 
8077 	state->dts_formats = new;
8078 	state->dts_formats[ndx] = fmt;
8079 
8080 	return (ndx + 1);
8081 }
8082 
8083 static void
8084 dtrace_format_remove(dtrace_state_t *state, uint16_t format)
8085 {
8086 	char *fmt;
8087 
8088 	ASSERT(state->dts_formats != NULL);
8089 	ASSERT(format <= state->dts_nformats);
8090 	ASSERT(state->dts_formats[format - 1] != NULL);
8091 
8092 	fmt = state->dts_formats[format - 1];
8093 	kmem_free(fmt, strlen(fmt) + 1);
8094 	state->dts_formats[format - 1] = NULL;
8095 }
8096 
8097 static void
8098 dtrace_format_destroy(dtrace_state_t *state)
8099 {
8100 	int i;
8101 
8102 	if (state->dts_nformats == 0) {
8103 		ASSERT(state->dts_formats == NULL);
8104 		return;
8105 	}
8106 
8107 	ASSERT(state->dts_formats != NULL);
8108 
8109 	for (i = 0; i < state->dts_nformats; i++) {
8110 		char *fmt = state->dts_formats[i];
8111 
8112 		if (fmt == NULL)
8113 			continue;
8114 
8115 		kmem_free(fmt, strlen(fmt) + 1);
8116 	}
8117 
8118 	kmem_free(state->dts_formats, state->dts_nformats * sizeof (char *));
8119 	state->dts_nformats = 0;
8120 	state->dts_formats = NULL;
8121 }
8122 
8123 /*
8124  * DTrace Predicate Functions
8125  */
8126 static dtrace_predicate_t *
8127 dtrace_predicate_create(dtrace_difo_t *dp)
8128 {
8129 	dtrace_predicate_t *pred;
8130 
8131 	ASSERT(MUTEX_HELD(&dtrace_lock));
8132 	ASSERT(dp->dtdo_refcnt != 0);
8133 
8134 	pred = kmem_zalloc(sizeof (dtrace_predicate_t), KM_SLEEP);
8135 	pred->dtp_difo = dp;
8136 	pred->dtp_refcnt = 1;
8137 
8138 	if (!dtrace_difo_cacheable(dp))
8139 		return (pred);
8140 
8141 	if (dtrace_predcache_id == DTRACE_CACHEIDNONE) {
8142 		/*
8143 		 * This is only theoretically possible -- we have had 2^32
8144 		 * cacheable predicates on this machine.  We cannot allow any
8145 		 * more predicates to become cacheable:  as unlikely as it is,
8146 		 * there may be a thread caching a (now stale) predicate cache
8147 		 * ID. (N.B.: the temptation is being successfully resisted to
8148 		 * have this cmn_err() "Holy shit -- we executed this code!")
8149 		 */
8150 		return (pred);
8151 	}
8152 
8153 	pred->dtp_cacheid = dtrace_predcache_id++;
8154 
8155 	return (pred);
8156 }
8157 
8158 static void
8159 dtrace_predicate_hold(dtrace_predicate_t *pred)
8160 {
8161 	ASSERT(MUTEX_HELD(&dtrace_lock));
8162 	ASSERT(pred->dtp_difo != NULL && pred->dtp_difo->dtdo_refcnt != 0);
8163 	ASSERT(pred->dtp_refcnt > 0);
8164 
8165 	pred->dtp_refcnt++;
8166 }
8167 
8168 static void
8169 dtrace_predicate_release(dtrace_predicate_t *pred, dtrace_vstate_t *vstate)
8170 {
8171 	dtrace_difo_t *dp = pred->dtp_difo;
8172 
8173 	ASSERT(MUTEX_HELD(&dtrace_lock));
8174 	ASSERT(dp != NULL && dp->dtdo_refcnt != 0);
8175 	ASSERT(pred->dtp_refcnt > 0);
8176 
8177 	if (--pred->dtp_refcnt == 0) {
8178 		dtrace_difo_release(pred->dtp_difo, vstate);
8179 		kmem_free(pred, sizeof (dtrace_predicate_t));
8180 	}
8181 }
8182 
8183 /*
8184  * DTrace Action Description Functions
8185  */
8186 static dtrace_actdesc_t *
8187 dtrace_actdesc_create(dtrace_actkind_t kind, uint32_t ntuple,
8188     uint64_t uarg, uint64_t arg)
8189 {
8190 	dtrace_actdesc_t *act;
8191 
8192 	ASSERT(!DTRACEACT_ISPRINTFLIKE(kind) || (arg != NULL &&
8193 	    arg >= KERNELBASE) || (arg == NULL && kind == DTRACEACT_PRINTA));
8194 
8195 	act = kmem_zalloc(sizeof (dtrace_actdesc_t), KM_SLEEP);
8196 	act->dtad_kind = kind;
8197 	act->dtad_ntuple = ntuple;
8198 	act->dtad_uarg = uarg;
8199 	act->dtad_arg = arg;
8200 	act->dtad_refcnt = 1;
8201 
8202 	return (act);
8203 }
8204 
8205 static void
8206 dtrace_actdesc_hold(dtrace_actdesc_t *act)
8207 {
8208 	ASSERT(act->dtad_refcnt >= 1);
8209 	act->dtad_refcnt++;
8210 }
8211 
8212 static void
8213 dtrace_actdesc_release(dtrace_actdesc_t *act, dtrace_vstate_t *vstate)
8214 {
8215 	dtrace_actkind_t kind = act->dtad_kind;
8216 	dtrace_difo_t *dp;
8217 
8218 	ASSERT(act->dtad_refcnt >= 1);
8219 
8220 	if (--act->dtad_refcnt != 0)
8221 		return;
8222 
8223 	if ((dp = act->dtad_difo) != NULL)
8224 		dtrace_difo_release(dp, vstate);
8225 
8226 	if (DTRACEACT_ISPRINTFLIKE(kind)) {
8227 		char *str = (char *)(uintptr_t)act->dtad_arg;
8228 
8229 		ASSERT((str != NULL && (uintptr_t)str >= KERNELBASE) ||
8230 		    (str == NULL && act->dtad_kind == DTRACEACT_PRINTA));
8231 
8232 		if (str != NULL)
8233 			kmem_free(str, strlen(str) + 1);
8234 	}
8235 
8236 	kmem_free(act, sizeof (dtrace_actdesc_t));
8237 }
8238 
8239 /*
8240  * DTrace ECB Functions
8241  */
8242 static dtrace_ecb_t *
8243 dtrace_ecb_add(dtrace_state_t *state, dtrace_probe_t *probe)
8244 {
8245 	dtrace_ecb_t *ecb;
8246 	dtrace_epid_t epid;
8247 
8248 	ASSERT(MUTEX_HELD(&dtrace_lock));
8249 
8250 	ecb = kmem_zalloc(sizeof (dtrace_ecb_t), KM_SLEEP);
8251 	ecb->dte_predicate = NULL;
8252 	ecb->dte_probe = probe;
8253 
8254 	/*
8255 	 * The default size is the size of the default action: recording
8256 	 * the epid.
8257 	 */
8258 	ecb->dte_size = ecb->dte_needed = sizeof (dtrace_epid_t);
8259 	ecb->dte_alignment = sizeof (dtrace_epid_t);
8260 
8261 	epid = state->dts_epid++;
8262 
8263 	if (epid - 1 >= state->dts_necbs) {
8264 		dtrace_ecb_t **oecbs = state->dts_ecbs, **ecbs;
8265 		int necbs = state->dts_necbs << 1;
8266 
8267 		ASSERT(epid == state->dts_necbs + 1);
8268 
8269 		if (necbs == 0) {
8270 			ASSERT(oecbs == NULL);
8271 			necbs = 1;
8272 		}
8273 
8274 		ecbs = kmem_zalloc(necbs * sizeof (*ecbs), KM_SLEEP);
8275 
8276 		if (oecbs != NULL)
8277 			bcopy(oecbs, ecbs, state->dts_necbs * sizeof (*ecbs));
8278 
8279 		dtrace_membar_producer();
8280 		state->dts_ecbs = ecbs;
8281 
8282 		if (oecbs != NULL) {
8283 			/*
8284 			 * If this state is active, we must dtrace_sync()
8285 			 * before we can free the old dts_ecbs array:  we're
8286 			 * coming in hot, and there may be active ring
8287 			 * buffer processing (which indexes into the dts_ecbs
8288 			 * array) on another CPU.
8289 			 */
8290 			if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
8291 				dtrace_sync();
8292 
8293 			kmem_free(oecbs, state->dts_necbs * sizeof (*ecbs));
8294 		}
8295 
8296 		dtrace_membar_producer();
8297 		state->dts_necbs = necbs;
8298 	}
8299 
8300 	ecb->dte_state = state;
8301 
8302 	ASSERT(state->dts_ecbs[epid - 1] == NULL);
8303 	dtrace_membar_producer();
8304 	state->dts_ecbs[(ecb->dte_epid = epid) - 1] = ecb;
8305 
8306 	return (ecb);
8307 }
8308 
8309 static void
8310 dtrace_ecb_enable(dtrace_ecb_t *ecb)
8311 {
8312 	dtrace_probe_t *probe = ecb->dte_probe;
8313 
8314 	ASSERT(MUTEX_HELD(&cpu_lock));
8315 	ASSERT(MUTEX_HELD(&dtrace_lock));
8316 	ASSERT(ecb->dte_next == NULL);
8317 
8318 	if (probe == NULL) {
8319 		/*
8320 		 * This is the NULL probe -- there's nothing to do.
8321 		 */
8322 		return;
8323 	}
8324 
8325 	if (probe->dtpr_ecb == NULL) {
8326 		dtrace_provider_t *prov = probe->dtpr_provider;
8327 
8328 		/*
8329 		 * We're the first ECB on this probe.
8330 		 */
8331 		probe->dtpr_ecb = probe->dtpr_ecb_last = ecb;
8332 
8333 		if (ecb->dte_predicate != NULL)
8334 			probe->dtpr_predcache = ecb->dte_predicate->dtp_cacheid;
8335 
8336 		prov->dtpv_pops.dtps_enable(prov->dtpv_arg,
8337 		    probe->dtpr_id, probe->dtpr_arg);
8338 	} else {
8339 		/*
8340 		 * This probe is already active.  Swing the last pointer to
8341 		 * point to the new ECB, and issue a dtrace_sync() to assure
8342 		 * that all CPUs have seen the change.
8343 		 */
8344 		ASSERT(probe->dtpr_ecb_last != NULL);
8345 		probe->dtpr_ecb_last->dte_next = ecb;
8346 		probe->dtpr_ecb_last = ecb;
8347 		probe->dtpr_predcache = 0;
8348 
8349 		dtrace_sync();
8350 	}
8351 }
8352 
8353 static void
8354 dtrace_ecb_resize(dtrace_ecb_t *ecb)
8355 {
8356 	uint32_t maxalign = sizeof (dtrace_epid_t);
8357 	uint32_t align = sizeof (uint8_t), offs, diff;
8358 	dtrace_action_t *act;
8359 	int wastuple = 0;
8360 	uint32_t aggbase = UINT32_MAX;
8361 	dtrace_state_t *state = ecb->dte_state;
8362 
8363 	/*
8364 	 * If we record anything, we always record the epid.  (And we always
8365 	 * record it first.)
8366 	 */
8367 	offs = sizeof (dtrace_epid_t);
8368 	ecb->dte_size = ecb->dte_needed = sizeof (dtrace_epid_t);
8369 
8370 	for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
8371 		dtrace_recdesc_t *rec = &act->dta_rec;
8372 
8373 		if ((align = rec->dtrd_alignment) > maxalign)
8374 			maxalign = align;
8375 
8376 		if (!wastuple && act->dta_intuple) {
8377 			/*
8378 			 * This is the first record in a tuple.  Align the
8379 			 * offset to be at offset 4 in an 8-byte aligned
8380 			 * block.
8381 			 */
8382 			diff = offs + sizeof (dtrace_aggid_t);
8383 
8384 			if (diff = (diff & (sizeof (uint64_t) - 1)))
8385 				offs += sizeof (uint64_t) - diff;
8386 
8387 			aggbase = offs - sizeof (dtrace_aggid_t);
8388 			ASSERT(!(aggbase & (sizeof (uint64_t) - 1)));
8389 		}
8390 
8391 		/*LINTED*/
8392 		if (rec->dtrd_size != 0 && (diff = (offs & (align - 1)))) {
8393 			/*
8394 			 * The current offset is not properly aligned; align it.
8395 			 */
8396 			offs += align - diff;
8397 		}
8398 
8399 		rec->dtrd_offset = offs;
8400 
8401 		if (offs + rec->dtrd_size > ecb->dte_needed) {
8402 			ecb->dte_needed = offs + rec->dtrd_size;
8403 
8404 			if (ecb->dte_needed > state->dts_needed)
8405 				state->dts_needed = ecb->dte_needed;
8406 		}
8407 
8408 		if (DTRACEACT_ISAGG(act->dta_kind)) {
8409 			dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
8410 			dtrace_action_t *first = agg->dtag_first, *prev;
8411 
8412 			ASSERT(rec->dtrd_size != 0 && first != NULL);
8413 			ASSERT(wastuple);
8414 			ASSERT(aggbase != UINT32_MAX);
8415 
8416 			agg->dtag_base = aggbase;
8417 
8418 			while ((prev = first->dta_prev) != NULL &&
8419 			    DTRACEACT_ISAGG(prev->dta_kind)) {
8420 				agg = (dtrace_aggregation_t *)prev;
8421 				first = agg->dtag_first;
8422 			}
8423 
8424 			if (prev != NULL) {
8425 				offs = prev->dta_rec.dtrd_offset +
8426 				    prev->dta_rec.dtrd_size;
8427 			} else {
8428 				offs = sizeof (dtrace_epid_t);
8429 			}
8430 			wastuple = 0;
8431 		} else {
8432 			if (!act->dta_intuple)
8433 				ecb->dte_size = offs + rec->dtrd_size;
8434 
8435 			offs += rec->dtrd_size;
8436 		}
8437 
8438 		wastuple = act->dta_intuple;
8439 	}
8440 
8441 	if ((act = ecb->dte_action) != NULL &&
8442 	    !(act->dta_kind == DTRACEACT_SPECULATE && act->dta_next == NULL) &&
8443 	    ecb->dte_size == sizeof (dtrace_epid_t)) {
8444 		/*
8445 		 * If the size is still sizeof (dtrace_epid_t), then all
8446 		 * actions store no data; set the size to 0.
8447 		 */
8448 		ecb->dte_alignment = maxalign;
8449 		ecb->dte_size = 0;
8450 
8451 		/*
8452 		 * If the needed space is still sizeof (dtrace_epid_t), then
8453 		 * all actions need no additional space; set the needed
8454 		 * size to 0.
8455 		 */
8456 		if (ecb->dte_needed == sizeof (dtrace_epid_t))
8457 			ecb->dte_needed = 0;
8458 
8459 		return;
8460 	}
8461 
8462 	/*
8463 	 * Set our alignment, and make sure that the dte_size and dte_needed
8464 	 * are aligned to the size of an EPID.
8465 	 */
8466 	ecb->dte_alignment = maxalign;
8467 	ecb->dte_size = (ecb->dte_size + (sizeof (dtrace_epid_t) - 1)) &
8468 	    ~(sizeof (dtrace_epid_t) - 1);
8469 	ecb->dte_needed = (ecb->dte_needed + (sizeof (dtrace_epid_t) - 1)) &
8470 	    ~(sizeof (dtrace_epid_t) - 1);
8471 	ASSERT(ecb->dte_size <= ecb->dte_needed);
8472 }
8473 
8474 static dtrace_action_t *
8475 dtrace_ecb_aggregation_create(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
8476 {
8477 	dtrace_aggregation_t *agg;
8478 	size_t size = sizeof (uint64_t);
8479 	int ntuple = desc->dtad_ntuple;
8480 	dtrace_action_t *act;
8481 	dtrace_recdesc_t *frec;
8482 	dtrace_aggid_t aggid;
8483 	dtrace_state_t *state = ecb->dte_state;
8484 
8485 	agg = kmem_zalloc(sizeof (dtrace_aggregation_t), KM_SLEEP);
8486 	agg->dtag_ecb = ecb;
8487 
8488 	ASSERT(DTRACEACT_ISAGG(desc->dtad_kind));
8489 
8490 	switch (desc->dtad_kind) {
8491 	case DTRACEAGG_MIN:
8492 		agg->dtag_initial = UINT64_MAX;
8493 		agg->dtag_aggregate = dtrace_aggregate_min;
8494 		break;
8495 
8496 	case DTRACEAGG_MAX:
8497 		agg->dtag_aggregate = dtrace_aggregate_max;
8498 		break;
8499 
8500 	case DTRACEAGG_COUNT:
8501 		agg->dtag_aggregate = dtrace_aggregate_count;
8502 		break;
8503 
8504 	case DTRACEAGG_QUANTIZE:
8505 		agg->dtag_aggregate = dtrace_aggregate_quantize;
8506 		size = (((sizeof (uint64_t) * NBBY) - 1) * 2 + 1) *
8507 		    sizeof (uint64_t);
8508 		break;
8509 
8510 	case DTRACEAGG_LQUANTIZE: {
8511 		uint16_t step = DTRACE_LQUANTIZE_STEP(desc->dtad_arg);
8512 		uint16_t levels = DTRACE_LQUANTIZE_LEVELS(desc->dtad_arg);
8513 
8514 		agg->dtag_initial = desc->dtad_arg;
8515 		agg->dtag_aggregate = dtrace_aggregate_lquantize;
8516 
8517 		if (step == 0 || levels == 0)
8518 			goto err;
8519 
8520 		size = levels * sizeof (uint64_t) + 3 * sizeof (uint64_t);
8521 		break;
8522 	}
8523 
8524 	case DTRACEAGG_AVG:
8525 		agg->dtag_aggregate = dtrace_aggregate_avg;
8526 		size = sizeof (uint64_t) * 2;
8527 		break;
8528 
8529 	case DTRACEAGG_SUM:
8530 		agg->dtag_aggregate = dtrace_aggregate_sum;
8531 		break;
8532 
8533 	default:
8534 		goto err;
8535 	}
8536 
8537 	agg->dtag_action.dta_rec.dtrd_size = size;
8538 
8539 	if (ntuple == 0)
8540 		goto err;
8541 
8542 	/*
8543 	 * We must make sure that we have enough actions for the n-tuple.
8544 	 */
8545 	for (act = ecb->dte_action_last; act != NULL; act = act->dta_prev) {
8546 		if (DTRACEACT_ISAGG(act->dta_kind))
8547 			break;
8548 
8549 		if (--ntuple == 0) {
8550 			/*
8551 			 * This is the action with which our n-tuple begins.
8552 			 */
8553 			agg->dtag_first = act;
8554 			goto success;
8555 		}
8556 	}
8557 
8558 	/*
8559 	 * This n-tuple is short by ntuple elements.  Return failure.
8560 	 */
8561 	ASSERT(ntuple != 0);
8562 err:
8563 	kmem_free(agg, sizeof (dtrace_aggregation_t));
8564 	return (NULL);
8565 
8566 success:
8567 	/*
8568 	 * If the last action in the tuple has a size of zero, it's actually
8569 	 * an expression argument for the aggregating action.
8570 	 */
8571 	ASSERT(ecb->dte_action_last != NULL);
8572 	act = ecb->dte_action_last;
8573 
8574 	if (act->dta_kind == DTRACEACT_DIFEXPR) {
8575 		ASSERT(act->dta_difo != NULL);
8576 
8577 		if (act->dta_difo->dtdo_rtype.dtdt_size == 0)
8578 			agg->dtag_hasarg = 1;
8579 	}
8580 
8581 	/*
8582 	 * We need to allocate an id for this aggregation.
8583 	 */
8584 	aggid = (dtrace_aggid_t)(uintptr_t)vmem_alloc(state->dts_aggid_arena, 1,
8585 	    VM_BESTFIT | VM_SLEEP);
8586 
8587 	if (aggid - 1 >= state->dts_naggregations) {
8588 		dtrace_aggregation_t **oaggs = state->dts_aggregations;
8589 		dtrace_aggregation_t **aggs;
8590 		int naggs = state->dts_naggregations << 1;
8591 		int onaggs = state->dts_naggregations;
8592 
8593 		ASSERT(aggid == state->dts_naggregations + 1);
8594 
8595 		if (naggs == 0) {
8596 			ASSERT(oaggs == NULL);
8597 			naggs = 1;
8598 		}
8599 
8600 		aggs = kmem_zalloc(naggs * sizeof (*aggs), KM_SLEEP);
8601 
8602 		if (oaggs != NULL) {
8603 			bcopy(oaggs, aggs, onaggs * sizeof (*aggs));
8604 			kmem_free(oaggs, onaggs * sizeof (*aggs));
8605 		}
8606 
8607 		state->dts_aggregations = aggs;
8608 		state->dts_naggregations = naggs;
8609 	}
8610 
8611 	ASSERT(state->dts_aggregations[aggid - 1] == NULL);
8612 	state->dts_aggregations[(agg->dtag_id = aggid) - 1] = agg;
8613 
8614 	frec = &agg->dtag_first->dta_rec;
8615 	if (frec->dtrd_alignment < sizeof (dtrace_aggid_t))
8616 		frec->dtrd_alignment = sizeof (dtrace_aggid_t);
8617 
8618 	for (act = agg->dtag_first; act != NULL; act = act->dta_next) {
8619 		ASSERT(!act->dta_intuple);
8620 		act->dta_intuple = 1;
8621 	}
8622 
8623 	return (&agg->dtag_action);
8624 }
8625 
8626 static void
8627 dtrace_ecb_aggregation_destroy(dtrace_ecb_t *ecb, dtrace_action_t *act)
8628 {
8629 	dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
8630 	dtrace_state_t *state = ecb->dte_state;
8631 	dtrace_aggid_t aggid = agg->dtag_id;
8632 
8633 	ASSERT(DTRACEACT_ISAGG(act->dta_kind));
8634 	vmem_free(state->dts_aggid_arena, (void *)(uintptr_t)aggid, 1);
8635 
8636 	ASSERT(state->dts_aggregations[aggid - 1] == agg);
8637 	state->dts_aggregations[aggid - 1] = NULL;
8638 
8639 	kmem_free(agg, sizeof (dtrace_aggregation_t));
8640 }
8641 
8642 static int
8643 dtrace_ecb_action_add(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
8644 {
8645 	dtrace_action_t *action, *last;
8646 	dtrace_difo_t *dp = desc->dtad_difo;
8647 	uint32_t size = 0, align = sizeof (uint8_t), mask;
8648 	uint16_t format = 0;
8649 	dtrace_recdesc_t *rec;
8650 	dtrace_state_t *state = ecb->dte_state;
8651 	dtrace_optval_t *opt = state->dts_options, nframes, strsize;
8652 	uint64_t arg = desc->dtad_arg;
8653 
8654 	ASSERT(MUTEX_HELD(&dtrace_lock));
8655 	ASSERT(ecb->dte_action == NULL || ecb->dte_action->dta_refcnt == 1);
8656 
8657 	if (DTRACEACT_ISAGG(desc->dtad_kind)) {
8658 		/*
8659 		 * If this is an aggregating action, there must be neither
8660 		 * a speculate nor a commit on the action chain.
8661 		 */
8662 		dtrace_action_t *act;
8663 
8664 		for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
8665 			if (act->dta_kind == DTRACEACT_COMMIT)
8666 				return (EINVAL);
8667 
8668 			if (act->dta_kind == DTRACEACT_SPECULATE)
8669 				return (EINVAL);
8670 		}
8671 
8672 		action = dtrace_ecb_aggregation_create(ecb, desc);
8673 
8674 		if (action == NULL)
8675 			return (EINVAL);
8676 	} else {
8677 		if (DTRACEACT_ISDESTRUCTIVE(desc->dtad_kind) ||
8678 		    (desc->dtad_kind == DTRACEACT_DIFEXPR &&
8679 		    dp != NULL && dp->dtdo_destructive)) {
8680 			state->dts_destructive = 1;
8681 		}
8682 
8683 		switch (desc->dtad_kind) {
8684 		case DTRACEACT_PRINTF:
8685 		case DTRACEACT_PRINTA:
8686 		case DTRACEACT_SYSTEM:
8687 		case DTRACEACT_FREOPEN:
8688 			/*
8689 			 * We know that our arg is a string -- turn it into a
8690 			 * format.
8691 			 */
8692 			if (arg == NULL) {
8693 				ASSERT(desc->dtad_kind == DTRACEACT_PRINTA);
8694 				format = 0;
8695 			} else {
8696 				ASSERT(arg != NULL);
8697 				ASSERT(arg > KERNELBASE);
8698 				format = dtrace_format_add(state,
8699 				    (char *)(uintptr_t)arg);
8700 			}
8701 
8702 			/*FALLTHROUGH*/
8703 		case DTRACEACT_LIBACT:
8704 		case DTRACEACT_DIFEXPR:
8705 			if (dp == NULL)
8706 				return (EINVAL);
8707 
8708 			if ((size = dp->dtdo_rtype.dtdt_size) != 0)
8709 				break;
8710 
8711 			if (dp->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING) {
8712 				if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
8713 					return (EINVAL);
8714 
8715 				size = opt[DTRACEOPT_STRSIZE];
8716 			}
8717 
8718 			break;
8719 
8720 		case DTRACEACT_STACK:
8721 			if ((nframes = arg) == 0) {
8722 				nframes = opt[DTRACEOPT_STACKFRAMES];
8723 				ASSERT(nframes > 0);
8724 				arg = nframes;
8725 			}
8726 
8727 			size = nframes * sizeof (pc_t);
8728 			break;
8729 
8730 		case DTRACEACT_JSTACK:
8731 			if ((strsize = DTRACE_USTACK_STRSIZE(arg)) == 0)
8732 				strsize = opt[DTRACEOPT_JSTACKSTRSIZE];
8733 
8734 			if ((nframes = DTRACE_USTACK_NFRAMES(arg)) == 0)
8735 				nframes = opt[DTRACEOPT_JSTACKFRAMES];
8736 
8737 			arg = DTRACE_USTACK_ARG(nframes, strsize);
8738 
8739 			/*FALLTHROUGH*/
8740 		case DTRACEACT_USTACK:
8741 			if (desc->dtad_kind != DTRACEACT_JSTACK &&
8742 			    (nframes = DTRACE_USTACK_NFRAMES(arg)) == 0) {
8743 				strsize = DTRACE_USTACK_STRSIZE(arg);
8744 				nframes = opt[DTRACEOPT_USTACKFRAMES];
8745 				ASSERT(nframes > 0);
8746 				arg = DTRACE_USTACK_ARG(nframes, strsize);
8747 			}
8748 
8749 			/*
8750 			 * Save a slot for the pid.
8751 			 */
8752 			size = (nframes + 1) * sizeof (uint64_t);
8753 			size += DTRACE_USTACK_STRSIZE(arg);
8754 			size = P2ROUNDUP(size, (uint32_t)(sizeof (uintptr_t)));
8755 
8756 			break;
8757 
8758 		case DTRACEACT_SYM:
8759 		case DTRACEACT_MOD:
8760 			if (dp == NULL || ((size = dp->dtdo_rtype.dtdt_size) !=
8761 			    sizeof (uint64_t)) ||
8762 			    (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
8763 				return (EINVAL);
8764 			break;
8765 
8766 		case DTRACEACT_USYM:
8767 		case DTRACEACT_UMOD:
8768 		case DTRACEACT_UADDR:
8769 			if (dp == NULL ||
8770 			    (dp->dtdo_rtype.dtdt_size != sizeof (uint64_t)) ||
8771 			    (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
8772 				return (EINVAL);
8773 
8774 			/*
8775 			 * We have a slot for the pid, plus a slot for the
8776 			 * argument.  To keep things simple (aligned with
8777 			 * bitness-neutral sizing), we store each as a 64-bit
8778 			 * quantity.
8779 			 */
8780 			size = 2 * sizeof (uint64_t);
8781 			break;
8782 
8783 		case DTRACEACT_STOP:
8784 		case DTRACEACT_BREAKPOINT:
8785 		case DTRACEACT_PANIC:
8786 			break;
8787 
8788 		case DTRACEACT_CHILL:
8789 		case DTRACEACT_DISCARD:
8790 		case DTRACEACT_RAISE:
8791 			if (dp == NULL)
8792 				return (EINVAL);
8793 			break;
8794 
8795 		case DTRACEACT_EXIT:
8796 			if (dp == NULL ||
8797 			    (size = dp->dtdo_rtype.dtdt_size) != sizeof (int) ||
8798 			    (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
8799 				return (EINVAL);
8800 			break;
8801 
8802 		case DTRACEACT_SPECULATE:
8803 			if (ecb->dte_size > sizeof (dtrace_epid_t))
8804 				return (EINVAL);
8805 
8806 			if (dp == NULL)
8807 				return (EINVAL);
8808 
8809 			state->dts_speculates = 1;
8810 			break;
8811 
8812 		case DTRACEACT_COMMIT: {
8813 			dtrace_action_t *act = ecb->dte_action;
8814 
8815 			for (; act != NULL; act = act->dta_next) {
8816 				if (act->dta_kind == DTRACEACT_COMMIT)
8817 					return (EINVAL);
8818 			}
8819 
8820 			if (dp == NULL)
8821 				return (EINVAL);
8822 			break;
8823 		}
8824 
8825 		default:
8826 			return (EINVAL);
8827 		}
8828 
8829 		if (size != 0 || desc->dtad_kind == DTRACEACT_SPECULATE) {
8830 			/*
8831 			 * If this is a data-storing action or a speculate,
8832 			 * we must be sure that there isn't a commit on the
8833 			 * action chain.
8834 			 */
8835 			dtrace_action_t *act = ecb->dte_action;
8836 
8837 			for (; act != NULL; act = act->dta_next) {
8838 				if (act->dta_kind == DTRACEACT_COMMIT)
8839 					return (EINVAL);
8840 			}
8841 		}
8842 
8843 		action = kmem_zalloc(sizeof (dtrace_action_t), KM_SLEEP);
8844 		action->dta_rec.dtrd_size = size;
8845 	}
8846 
8847 	action->dta_refcnt = 1;
8848 	rec = &action->dta_rec;
8849 	size = rec->dtrd_size;
8850 
8851 	for (mask = sizeof (uint64_t) - 1; size != 0 && mask > 0; mask >>= 1) {
8852 		if (!(size & mask)) {
8853 			align = mask + 1;
8854 			break;
8855 		}
8856 	}
8857 
8858 	action->dta_kind = desc->dtad_kind;
8859 
8860 	if ((action->dta_difo = dp) != NULL)
8861 		dtrace_difo_hold(dp);
8862 
8863 	rec->dtrd_action = action->dta_kind;
8864 	rec->dtrd_arg = arg;
8865 	rec->dtrd_uarg = desc->dtad_uarg;
8866 	rec->dtrd_alignment = (uint16_t)align;
8867 	rec->dtrd_format = format;
8868 
8869 	if ((last = ecb->dte_action_last) != NULL) {
8870 		ASSERT(ecb->dte_action != NULL);
8871 		action->dta_prev = last;
8872 		last->dta_next = action;
8873 	} else {
8874 		ASSERT(ecb->dte_action == NULL);
8875 		ecb->dte_action = action;
8876 	}
8877 
8878 	ecb->dte_action_last = action;
8879 
8880 	return (0);
8881 }
8882 
8883 static void
8884 dtrace_ecb_action_remove(dtrace_ecb_t *ecb)
8885 {
8886 	dtrace_action_t *act = ecb->dte_action, *next;
8887 	dtrace_vstate_t *vstate = &ecb->dte_state->dts_vstate;
8888 	dtrace_difo_t *dp;
8889 	uint16_t format;
8890 
8891 	if (act != NULL && act->dta_refcnt > 1) {
8892 		ASSERT(act->dta_next == NULL || act->dta_next->dta_refcnt == 1);
8893 		act->dta_refcnt--;
8894 	} else {
8895 		for (; act != NULL; act = next) {
8896 			next = act->dta_next;
8897 			ASSERT(next != NULL || act == ecb->dte_action_last);
8898 			ASSERT(act->dta_refcnt == 1);
8899 
8900 			if ((format = act->dta_rec.dtrd_format) != 0)
8901 				dtrace_format_remove(ecb->dte_state, format);
8902 
8903 			if ((dp = act->dta_difo) != NULL)
8904 				dtrace_difo_release(dp, vstate);
8905 
8906 			if (DTRACEACT_ISAGG(act->dta_kind)) {
8907 				dtrace_ecb_aggregation_destroy(ecb, act);
8908 			} else {
8909 				kmem_free(act, sizeof (dtrace_action_t));
8910 			}
8911 		}
8912 	}
8913 
8914 	ecb->dte_action = NULL;
8915 	ecb->dte_action_last = NULL;
8916 	ecb->dte_size = sizeof (dtrace_epid_t);
8917 }
8918 
8919 static void
8920 dtrace_ecb_disable(dtrace_ecb_t *ecb)
8921 {
8922 	/*
8923 	 * We disable the ECB by removing it from its probe.
8924 	 */
8925 	dtrace_ecb_t *pecb, *prev = NULL;
8926 	dtrace_probe_t *probe = ecb->dte_probe;
8927 
8928 	ASSERT(MUTEX_HELD(&dtrace_lock));
8929 
8930 	if (probe == NULL) {
8931 		/*
8932 		 * This is the NULL probe; there is nothing to disable.
8933 		 */
8934 		return;
8935 	}
8936 
8937 	for (pecb = probe->dtpr_ecb; pecb != NULL; pecb = pecb->dte_next) {
8938 		if (pecb == ecb)
8939 			break;
8940 		prev = pecb;
8941 	}
8942 
8943 	ASSERT(pecb != NULL);
8944 
8945 	if (prev == NULL) {
8946 		probe->dtpr_ecb = ecb->dte_next;
8947 	} else {
8948 		prev->dte_next = ecb->dte_next;
8949 	}
8950 
8951 	if (ecb == probe->dtpr_ecb_last) {
8952 		ASSERT(ecb->dte_next == NULL);
8953 		probe->dtpr_ecb_last = prev;
8954 	}
8955 
8956 	/*
8957 	 * The ECB has been disconnected from the probe; now sync to assure
8958 	 * that all CPUs have seen the change before returning.
8959 	 */
8960 	dtrace_sync();
8961 
8962 	if (probe->dtpr_ecb == NULL) {
8963 		/*
8964 		 * That was the last ECB on the probe; clear the predicate
8965 		 * cache ID for the probe, disable it and sync one more time
8966 		 * to assure that we'll never hit it again.
8967 		 */
8968 		dtrace_provider_t *prov = probe->dtpr_provider;
8969 
8970 		ASSERT(ecb->dte_next == NULL);
8971 		ASSERT(probe->dtpr_ecb_last == NULL);
8972 		probe->dtpr_predcache = DTRACE_CACHEIDNONE;
8973 		prov->dtpv_pops.dtps_disable(prov->dtpv_arg,
8974 		    probe->dtpr_id, probe->dtpr_arg);
8975 		dtrace_sync();
8976 	} else {
8977 		/*
8978 		 * There is at least one ECB remaining on the probe.  If there
8979 		 * is _exactly_ one, set the probe's predicate cache ID to be
8980 		 * the predicate cache ID of the remaining ECB.
8981 		 */
8982 		ASSERT(probe->dtpr_ecb_last != NULL);
8983 		ASSERT(probe->dtpr_predcache == DTRACE_CACHEIDNONE);
8984 
8985 		if (probe->dtpr_ecb == probe->dtpr_ecb_last) {
8986 			dtrace_predicate_t *p = probe->dtpr_ecb->dte_predicate;
8987 
8988 			ASSERT(probe->dtpr_ecb->dte_next == NULL);
8989 
8990 			if (p != NULL)
8991 				probe->dtpr_predcache = p->dtp_cacheid;
8992 		}
8993 
8994 		ecb->dte_next = NULL;
8995 	}
8996 }
8997 
8998 static void
8999 dtrace_ecb_destroy(dtrace_ecb_t *ecb)
9000 {
9001 	dtrace_state_t *state = ecb->dte_state;
9002 	dtrace_vstate_t *vstate = &state->dts_vstate;
9003 	dtrace_predicate_t *pred;
9004 	dtrace_epid_t epid = ecb->dte_epid;
9005 
9006 	ASSERT(MUTEX_HELD(&dtrace_lock));
9007 	ASSERT(ecb->dte_next == NULL);
9008 	ASSERT(ecb->dte_probe == NULL || ecb->dte_probe->dtpr_ecb != ecb);
9009 
9010 	if ((pred = ecb->dte_predicate) != NULL)
9011 		dtrace_predicate_release(pred, vstate);
9012 
9013 	dtrace_ecb_action_remove(ecb);
9014 
9015 	ASSERT(state->dts_ecbs[epid - 1] == ecb);
9016 	state->dts_ecbs[epid - 1] = NULL;
9017 
9018 	kmem_free(ecb, sizeof (dtrace_ecb_t));
9019 }
9020 
9021 static dtrace_ecb_t *
9022 dtrace_ecb_create(dtrace_state_t *state, dtrace_probe_t *probe,
9023     dtrace_enabling_t *enab)
9024 {
9025 	dtrace_ecb_t *ecb;
9026 	dtrace_predicate_t *pred;
9027 	dtrace_actdesc_t *act;
9028 	dtrace_provider_t *prov;
9029 	dtrace_ecbdesc_t *desc = enab->dten_current;
9030 
9031 	ASSERT(MUTEX_HELD(&dtrace_lock));
9032 	ASSERT(state != NULL);
9033 
9034 	ecb = dtrace_ecb_add(state, probe);
9035 	ecb->dte_uarg = desc->dted_uarg;
9036 
9037 	if ((pred = desc->dted_pred.dtpdd_predicate) != NULL) {
9038 		dtrace_predicate_hold(pred);
9039 		ecb->dte_predicate = pred;
9040 	}
9041 
9042 	if (probe != NULL) {
9043 		/*
9044 		 * If the provider shows more leg than the consumer is old
9045 		 * enough to see, we need to enable the appropriate implicit
9046 		 * predicate bits to prevent the ecb from activating at
9047 		 * revealing times.
9048 		 *
9049 		 * Providers specifying DTRACE_PRIV_USER at register time
9050 		 * are stating that they need the /proc-style privilege
9051 		 * model to be enforced, and this is what DTRACE_COND_OWNER
9052 		 * and DTRACE_COND_ZONEOWNER will then do at probe time.
9053 		 */
9054 		prov = probe->dtpr_provider;
9055 		if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLPROC) &&
9056 		    (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
9057 			ecb->dte_cond |= DTRACE_COND_OWNER;
9058 
9059 		if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLZONE) &&
9060 		    (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
9061 			ecb->dte_cond |= DTRACE_COND_ZONEOWNER;
9062 
9063 		/*
9064 		 * If the provider shows us kernel innards and the user
9065 		 * is lacking sufficient privilege, enable the
9066 		 * DTRACE_COND_USERMODE implicit predicate.
9067 		 */
9068 		if (!(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) &&
9069 		    (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_KERNEL))
9070 			ecb->dte_cond |= DTRACE_COND_USERMODE;
9071 	}
9072 
9073 	if (dtrace_ecb_create_cache != NULL) {
9074 		/*
9075 		 * If we have a cached ecb, we'll use its action list instead
9076 		 * of creating our own (saving both time and space).
9077 		 */
9078 		dtrace_ecb_t *cached = dtrace_ecb_create_cache;
9079 		dtrace_action_t *act = cached->dte_action;
9080 
9081 		if (act != NULL) {
9082 			ASSERT(act->dta_refcnt > 0);
9083 			act->dta_refcnt++;
9084 			ecb->dte_action = act;
9085 			ecb->dte_action_last = cached->dte_action_last;
9086 			ecb->dte_needed = cached->dte_needed;
9087 			ecb->dte_size = cached->dte_size;
9088 			ecb->dte_alignment = cached->dte_alignment;
9089 		}
9090 
9091 		return (ecb);
9092 	}
9093 
9094 	for (act = desc->dted_action; act != NULL; act = act->dtad_next) {
9095 		if ((enab->dten_error = dtrace_ecb_action_add(ecb, act)) != 0) {
9096 			dtrace_ecb_destroy(ecb);
9097 			return (NULL);
9098 		}
9099 	}
9100 
9101 	dtrace_ecb_resize(ecb);
9102 
9103 	return (dtrace_ecb_create_cache = ecb);
9104 }
9105 
9106 static int
9107 dtrace_ecb_create_enable(dtrace_probe_t *probe, void *arg)
9108 {
9109 	dtrace_ecb_t *ecb;
9110 	dtrace_enabling_t *enab = arg;
9111 	dtrace_state_t *state = enab->dten_vstate->dtvs_state;
9112 
9113 	ASSERT(state != NULL);
9114 
9115 	if (probe != NULL && probe->dtpr_gen < enab->dten_probegen) {
9116 		/*
9117 		 * This probe was created in a generation for which this
9118 		 * enabling has previously created ECBs; we don't want to
9119 		 * enable it again, so just kick out.
9120 		 */
9121 		return (DTRACE_MATCH_NEXT);
9122 	}
9123 
9124 	if ((ecb = dtrace_ecb_create(state, probe, enab)) == NULL)
9125 		return (DTRACE_MATCH_DONE);
9126 
9127 	dtrace_ecb_enable(ecb);
9128 	return (DTRACE_MATCH_NEXT);
9129 }
9130 
9131 static dtrace_ecb_t *
9132 dtrace_epid2ecb(dtrace_state_t *state, dtrace_epid_t id)
9133 {
9134 	dtrace_ecb_t *ecb;
9135 
9136 	ASSERT(MUTEX_HELD(&dtrace_lock));
9137 
9138 	if (id == 0 || id > state->dts_necbs)
9139 		return (NULL);
9140 
9141 	ASSERT(state->dts_necbs > 0 && state->dts_ecbs != NULL);
9142 	ASSERT((ecb = state->dts_ecbs[id - 1]) == NULL || ecb->dte_epid == id);
9143 
9144 	return (state->dts_ecbs[id - 1]);
9145 }
9146 
9147 static dtrace_aggregation_t *
9148 dtrace_aggid2agg(dtrace_state_t *state, dtrace_aggid_t id)
9149 {
9150 	dtrace_aggregation_t *agg;
9151 
9152 	ASSERT(MUTEX_HELD(&dtrace_lock));
9153 
9154 	if (id == 0 || id > state->dts_naggregations)
9155 		return (NULL);
9156 
9157 	ASSERT(state->dts_naggregations > 0 && state->dts_aggregations != NULL);
9158 	ASSERT((agg = state->dts_aggregations[id - 1]) == NULL ||
9159 	    agg->dtag_id == id);
9160 
9161 	return (state->dts_aggregations[id - 1]);
9162 }
9163 
9164 /*
9165  * DTrace Buffer Functions
9166  *
9167  * The following functions manipulate DTrace buffers.  Most of these functions
9168  * are called in the context of establishing or processing consumer state;
9169  * exceptions are explicitly noted.
9170  */
9171 
9172 /*
9173  * Note:  called from cross call context.  This function switches the two
9174  * buffers on a given CPU.  The atomicity of this operation is assured by
9175  * disabling interrupts while the actual switch takes place; the disabling of
9176  * interrupts serializes the execution with any execution of dtrace_probe() on
9177  * the same CPU.
9178  */
9179 static void
9180 dtrace_buffer_switch(dtrace_buffer_t *buf)
9181 {
9182 	caddr_t tomax = buf->dtb_tomax;
9183 	caddr_t xamot = buf->dtb_xamot;
9184 	dtrace_icookie_t cookie;
9185 
9186 	ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
9187 	ASSERT(!(buf->dtb_flags & DTRACEBUF_RING));
9188 
9189 	cookie = dtrace_interrupt_disable();
9190 	buf->dtb_tomax = xamot;
9191 	buf->dtb_xamot = tomax;
9192 	buf->dtb_xamot_drops = buf->dtb_drops;
9193 	buf->dtb_xamot_offset = buf->dtb_offset;
9194 	buf->dtb_xamot_errors = buf->dtb_errors;
9195 	buf->dtb_xamot_flags = buf->dtb_flags;
9196 	buf->dtb_offset = 0;
9197 	buf->dtb_drops = 0;
9198 	buf->dtb_errors = 0;
9199 	buf->dtb_flags &= ~(DTRACEBUF_ERROR | DTRACEBUF_DROPPED);
9200 	dtrace_interrupt_enable(cookie);
9201 }
9202 
9203 /*
9204  * Note:  called from cross call context.  This function activates a buffer
9205  * on a CPU.  As with dtrace_buffer_switch(), the atomicity of the operation
9206  * is guaranteed by the disabling of interrupts.
9207  */
9208 static void
9209 dtrace_buffer_activate(dtrace_state_t *state)
9210 {
9211 	dtrace_buffer_t *buf;
9212 	dtrace_icookie_t cookie = dtrace_interrupt_disable();
9213 
9214 	buf = &state->dts_buffer[CPU->cpu_id];
9215 
9216 	if (buf->dtb_tomax != NULL) {
9217 		/*
9218 		 * We might like to assert that the buffer is marked inactive,
9219 		 * but this isn't necessarily true:  the buffer for the CPU
9220 		 * that processes the BEGIN probe has its buffer activated
9221 		 * manually.  In this case, we take the (harmless) action
9222 		 * re-clearing the bit INACTIVE bit.
9223 		 */
9224 		buf->dtb_flags &= ~DTRACEBUF_INACTIVE;
9225 	}
9226 
9227 	dtrace_interrupt_enable(cookie);
9228 }
9229 
9230 static int
9231 dtrace_buffer_alloc(dtrace_buffer_t *bufs, size_t size, int flags,
9232     processorid_t cpu)
9233 {
9234 	cpu_t *cp;
9235 	dtrace_buffer_t *buf;
9236 
9237 	ASSERT(MUTEX_HELD(&cpu_lock));
9238 	ASSERT(MUTEX_HELD(&dtrace_lock));
9239 
9240 	if (size > dtrace_nonroot_maxsize &&
9241 	    !PRIV_POLICY_CHOICE(CRED(), PRIV_ALL, B_FALSE))
9242 		return (EFBIG);
9243 
9244 	cp = cpu_list;
9245 
9246 	do {
9247 		if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
9248 			continue;
9249 
9250 		buf = &bufs[cp->cpu_id];
9251 
9252 		/*
9253 		 * If there is already a buffer allocated for this CPU, it
9254 		 * is only possible that this is a DR event.  In this case,
9255 		 * the buffer size must match our specified size.
9256 		 */
9257 		if (buf->dtb_tomax != NULL) {
9258 			ASSERT(buf->dtb_size == size);
9259 			continue;
9260 		}
9261 
9262 		ASSERT(buf->dtb_xamot == NULL);
9263 
9264 		if ((buf->dtb_tomax = kmem_zalloc(size, KM_NOSLEEP)) == NULL)
9265 			goto err;
9266 
9267 		buf->dtb_size = size;
9268 		buf->dtb_flags = flags;
9269 		buf->dtb_offset = 0;
9270 		buf->dtb_drops = 0;
9271 
9272 		if (flags & DTRACEBUF_NOSWITCH)
9273 			continue;
9274 
9275 		if ((buf->dtb_xamot = kmem_zalloc(size, KM_NOSLEEP)) == NULL)
9276 			goto err;
9277 	} while ((cp = cp->cpu_next) != cpu_list);
9278 
9279 	return (0);
9280 
9281 err:
9282 	cp = cpu_list;
9283 
9284 	do {
9285 		if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
9286 			continue;
9287 
9288 		buf = &bufs[cp->cpu_id];
9289 
9290 		if (buf->dtb_xamot != NULL) {
9291 			ASSERT(buf->dtb_tomax != NULL);
9292 			ASSERT(buf->dtb_size == size);
9293 			kmem_free(buf->dtb_xamot, size);
9294 		}
9295 
9296 		if (buf->dtb_tomax != NULL) {
9297 			ASSERT(buf->dtb_size == size);
9298 			kmem_free(buf->dtb_tomax, size);
9299 		}
9300 
9301 		buf->dtb_tomax = NULL;
9302 		buf->dtb_xamot = NULL;
9303 		buf->dtb_size = 0;
9304 	} while ((cp = cp->cpu_next) != cpu_list);
9305 
9306 	return (ENOMEM);
9307 }
9308 
9309 /*
9310  * Note:  called from probe context.  This function just increments the drop
9311  * count on a buffer.  It has been made a function to allow for the
9312  * possibility of understanding the source of mysterious drop counts.  (A
9313  * problem for which one may be particularly disappointed that DTrace cannot
9314  * be used to understand DTrace.)
9315  */
9316 static void
9317 dtrace_buffer_drop(dtrace_buffer_t *buf)
9318 {
9319 	buf->dtb_drops++;
9320 }
9321 
9322 /*
9323  * Note:  called from probe context.  This function is called to reserve space
9324  * in a buffer.  If mstate is non-NULL, sets the scratch base and size in the
9325  * mstate.  Returns the new offset in the buffer, or a negative value if an
9326  * error has occurred.
9327  */
9328 static intptr_t
9329 dtrace_buffer_reserve(dtrace_buffer_t *buf, size_t needed, size_t align,
9330     dtrace_state_t *state, dtrace_mstate_t *mstate)
9331 {
9332 	intptr_t offs = buf->dtb_offset, soffs;
9333 	intptr_t woffs;
9334 	caddr_t tomax;
9335 	size_t total;
9336 
9337 	if (buf->dtb_flags & DTRACEBUF_INACTIVE)
9338 		return (-1);
9339 
9340 	if ((tomax = buf->dtb_tomax) == NULL) {
9341 		dtrace_buffer_drop(buf);
9342 		return (-1);
9343 	}
9344 
9345 	if (!(buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL))) {
9346 		while (offs & (align - 1)) {
9347 			/*
9348 			 * Assert that our alignment is off by a number which
9349 			 * is itself sizeof (uint32_t) aligned.
9350 			 */
9351 			ASSERT(!((align - (offs & (align - 1))) &
9352 			    (sizeof (uint32_t) - 1)));
9353 			DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
9354 			offs += sizeof (uint32_t);
9355 		}
9356 
9357 		if ((soffs = offs + needed) > buf->dtb_size) {
9358 			dtrace_buffer_drop(buf);
9359 			return (-1);
9360 		}
9361 
9362 		if (mstate == NULL)
9363 			return (offs);
9364 
9365 		mstate->dtms_scratch_base = (uintptr_t)tomax + soffs;
9366 		mstate->dtms_scratch_size = buf->dtb_size - soffs;
9367 		mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
9368 
9369 		return (offs);
9370 	}
9371 
9372 	if (buf->dtb_flags & DTRACEBUF_FILL) {
9373 		if (state->dts_activity != DTRACE_ACTIVITY_COOLDOWN &&
9374 		    (buf->dtb_flags & DTRACEBUF_FULL))
9375 			return (-1);
9376 		goto out;
9377 	}
9378 
9379 	total = needed + (offs & (align - 1));
9380 
9381 	/*
9382 	 * For a ring buffer, life is quite a bit more complicated.  Before
9383 	 * we can store any padding, we need to adjust our wrapping offset.
9384 	 * (If we've never before wrapped or we're not about to, no adjustment
9385 	 * is required.)
9386 	 */
9387 	if ((buf->dtb_flags & DTRACEBUF_WRAPPED) ||
9388 	    offs + total > buf->dtb_size) {
9389 		woffs = buf->dtb_xamot_offset;
9390 
9391 		if (offs + total > buf->dtb_size) {
9392 			/*
9393 			 * We can't fit in the end of the buffer.  First, a
9394 			 * sanity check that we can fit in the buffer at all.
9395 			 */
9396 			if (total > buf->dtb_size) {
9397 				dtrace_buffer_drop(buf);
9398 				return (-1);
9399 			}
9400 
9401 			/*
9402 			 * We're going to be storing at the top of the buffer,
9403 			 * so now we need to deal with the wrapped offset.  We
9404 			 * only reset our wrapped offset to 0 if it is
9405 			 * currently greater than the current offset.  If it
9406 			 * is less than the current offset, it is because a
9407 			 * previous allocation induced a wrap -- but the
9408 			 * allocation didn't subsequently take the space due
9409 			 * to an error or false predicate evaluation.  In this
9410 			 * case, we'll just leave the wrapped offset alone: if
9411 			 * the wrapped offset hasn't been advanced far enough
9412 			 * for this allocation, it will be adjusted in the
9413 			 * lower loop.
9414 			 */
9415 			if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
9416 				if (woffs >= offs)
9417 					woffs = 0;
9418 			} else {
9419 				woffs = 0;
9420 			}
9421 
9422 			/*
9423 			 * Now we know that we're going to be storing to the
9424 			 * top of the buffer and that there is room for us
9425 			 * there.  We need to clear the buffer from the current
9426 			 * offset to the end (there may be old gunk there).
9427 			 */
9428 			while (offs < buf->dtb_size)
9429 				tomax[offs++] = 0;
9430 
9431 			/*
9432 			 * We need to set our offset to zero.  And because we
9433 			 * are wrapping, we need to set the bit indicating as
9434 			 * much.  We can also adjust our needed space back
9435 			 * down to the space required by the ECB -- we know
9436 			 * that the top of the buffer is aligned.
9437 			 */
9438 			offs = 0;
9439 			total = needed;
9440 			buf->dtb_flags |= DTRACEBUF_WRAPPED;
9441 		} else {
9442 			/*
9443 			 * There is room for us in the buffer, so we simply
9444 			 * need to check the wrapped offset.
9445 			 */
9446 			if (woffs < offs) {
9447 				/*
9448 				 * The wrapped offset is less than the offset.
9449 				 * This can happen if we allocated buffer space
9450 				 * that induced a wrap, but then we didn't
9451 				 * subsequently take the space due to an error
9452 				 * or false predicate evaluation.  This is
9453 				 * okay; we know that _this_ allocation isn't
9454 				 * going to induce a wrap.  We still can't
9455 				 * reset the wrapped offset to be zero,
9456 				 * however: the space may have been trashed in
9457 				 * the previous failed probe attempt.  But at
9458 				 * least the wrapped offset doesn't need to
9459 				 * be adjusted at all...
9460 				 */
9461 				goto out;
9462 			}
9463 		}
9464 
9465 		while (offs + total > woffs) {
9466 			dtrace_epid_t epid = *(uint32_t *)(tomax + woffs);
9467 			size_t size;
9468 
9469 			if (epid == DTRACE_EPIDNONE) {
9470 				size = sizeof (uint32_t);
9471 			} else {
9472 				ASSERT(epid <= state->dts_necbs);
9473 				ASSERT(state->dts_ecbs[epid - 1] != NULL);
9474 
9475 				size = state->dts_ecbs[epid - 1]->dte_size;
9476 			}
9477 
9478 			ASSERT(woffs + size <= buf->dtb_size);
9479 			ASSERT(size != 0);
9480 
9481 			if (woffs + size == buf->dtb_size) {
9482 				/*
9483 				 * We've reached the end of the buffer; we want
9484 				 * to set the wrapped offset to 0 and break
9485 				 * out.  However, if the offs is 0, then we're
9486 				 * in a strange edge-condition:  the amount of
9487 				 * space that we want to reserve plus the size
9488 				 * of the record that we're overwriting is
9489 				 * greater than the size of the buffer.  This
9490 				 * is problematic because if we reserve the
9491 				 * space but subsequently don't consume it (due
9492 				 * to a failed predicate or error) the wrapped
9493 				 * offset will be 0 -- yet the EPID at offset 0
9494 				 * will not be committed.  This situation is
9495 				 * relatively easy to deal with:  if we're in
9496 				 * this case, the buffer is indistinguishable
9497 				 * from one that hasn't wrapped; we need only
9498 				 * finish the job by clearing the wrapped bit,
9499 				 * explicitly setting the offset to be 0, and
9500 				 * zero'ing out the old data in the buffer.
9501 				 */
9502 				if (offs == 0) {
9503 					buf->dtb_flags &= ~DTRACEBUF_WRAPPED;
9504 					buf->dtb_offset = 0;
9505 					woffs = total;
9506 
9507 					while (woffs < buf->dtb_size)
9508 						tomax[woffs++] = 0;
9509 				}
9510 
9511 				woffs = 0;
9512 				break;
9513 			}
9514 
9515 			woffs += size;
9516 		}
9517 
9518 		/*
9519 		 * We have a wrapped offset.  It may be that the wrapped offset
9520 		 * has become zero -- that's okay.
9521 		 */
9522 		buf->dtb_xamot_offset = woffs;
9523 	}
9524 
9525 out:
9526 	/*
9527 	 * Now we can plow the buffer with any necessary padding.
9528 	 */
9529 	while (offs & (align - 1)) {
9530 		/*
9531 		 * Assert that our alignment is off by a number which
9532 		 * is itself sizeof (uint32_t) aligned.
9533 		 */
9534 		ASSERT(!((align - (offs & (align - 1))) &
9535 		    (sizeof (uint32_t) - 1)));
9536 		DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
9537 		offs += sizeof (uint32_t);
9538 	}
9539 
9540 	if (buf->dtb_flags & DTRACEBUF_FILL) {
9541 		if (offs + needed > buf->dtb_size - state->dts_reserve) {
9542 			buf->dtb_flags |= DTRACEBUF_FULL;
9543 			return (-1);
9544 		}
9545 	}
9546 
9547 	if (mstate == NULL)
9548 		return (offs);
9549 
9550 	/*
9551 	 * For ring buffers and fill buffers, the scratch space is always
9552 	 * the inactive buffer.
9553 	 */
9554 	mstate->dtms_scratch_base = (uintptr_t)buf->dtb_xamot;
9555 	mstate->dtms_scratch_size = buf->dtb_size;
9556 	mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
9557 
9558 	return (offs);
9559 }
9560 
9561 static void
9562 dtrace_buffer_polish(dtrace_buffer_t *buf)
9563 {
9564 	ASSERT(buf->dtb_flags & DTRACEBUF_RING);
9565 	ASSERT(MUTEX_HELD(&dtrace_lock));
9566 
9567 	if (!(buf->dtb_flags & DTRACEBUF_WRAPPED))
9568 		return;
9569 
9570 	/*
9571 	 * We need to polish the ring buffer.  There are three cases:
9572 	 *
9573 	 * - The first (and presumably most common) is that there is no gap
9574 	 *   between the buffer offset and the wrapped offset.  In this case,
9575 	 *   there is nothing in the buffer that isn't valid data; we can
9576 	 *   mark the buffer as polished and return.
9577 	 *
9578 	 * - The second (less common than the first but still more common
9579 	 *   than the third) is that there is a gap between the buffer offset
9580 	 *   and the wrapped offset, and the wrapped offset is larger than the
9581 	 *   buffer offset.  This can happen because of an alignment issue, or
9582 	 *   can happen because of a call to dtrace_buffer_reserve() that
9583 	 *   didn't subsequently consume the buffer space.  In this case,
9584 	 *   we need to zero the data from the buffer offset to the wrapped
9585 	 *   offset.
9586 	 *
9587 	 * - The third (and least common) is that there is a gap between the
9588 	 *   buffer offset and the wrapped offset, but the wrapped offset is
9589 	 *   _less_ than the buffer offset.  This can only happen because a
9590 	 *   call to dtrace_buffer_reserve() induced a wrap, but the space
9591 	 *   was not subsequently consumed.  In this case, we need to zero the
9592 	 *   space from the offset to the end of the buffer _and_ from the
9593 	 *   top of the buffer to the wrapped offset.
9594 	 */
9595 	if (buf->dtb_offset < buf->dtb_xamot_offset) {
9596 		bzero(buf->dtb_tomax + buf->dtb_offset,
9597 		    buf->dtb_xamot_offset - buf->dtb_offset);
9598 	}
9599 
9600 	if (buf->dtb_offset > buf->dtb_xamot_offset) {
9601 		bzero(buf->dtb_tomax + buf->dtb_offset,
9602 		    buf->dtb_size - buf->dtb_offset);
9603 		bzero(buf->dtb_tomax, buf->dtb_xamot_offset);
9604 	}
9605 }
9606 
9607 static void
9608 dtrace_buffer_free(dtrace_buffer_t *bufs)
9609 {
9610 	int i;
9611 
9612 	for (i = 0; i < NCPU; i++) {
9613 		dtrace_buffer_t *buf = &bufs[i];
9614 
9615 		if (buf->dtb_tomax == NULL) {
9616 			ASSERT(buf->dtb_xamot == NULL);
9617 			ASSERT(buf->dtb_size == 0);
9618 			continue;
9619 		}
9620 
9621 		if (buf->dtb_xamot != NULL) {
9622 			ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
9623 			kmem_free(buf->dtb_xamot, buf->dtb_size);
9624 		}
9625 
9626 		kmem_free(buf->dtb_tomax, buf->dtb_size);
9627 		buf->dtb_size = 0;
9628 		buf->dtb_tomax = NULL;
9629 		buf->dtb_xamot = NULL;
9630 	}
9631 }
9632 
9633 /*
9634  * DTrace Enabling Functions
9635  */
9636 static dtrace_enabling_t *
9637 dtrace_enabling_create(dtrace_vstate_t *vstate)
9638 {
9639 	dtrace_enabling_t *enab;
9640 
9641 	enab = kmem_zalloc(sizeof (dtrace_enabling_t), KM_SLEEP);
9642 	enab->dten_vstate = vstate;
9643 
9644 	return (enab);
9645 }
9646 
9647 static void
9648 dtrace_enabling_add(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb)
9649 {
9650 	dtrace_ecbdesc_t **ndesc;
9651 	size_t osize, nsize;
9652 
9653 	/*
9654 	 * We can't add to enablings after we've enabled them, or after we've
9655 	 * retained them.
9656 	 */
9657 	ASSERT(enab->dten_probegen == 0);
9658 	ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
9659 
9660 	if (enab->dten_ndesc < enab->dten_maxdesc) {
9661 		enab->dten_desc[enab->dten_ndesc++] = ecb;
9662 		return;
9663 	}
9664 
9665 	osize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
9666 
9667 	if (enab->dten_maxdesc == 0) {
9668 		enab->dten_maxdesc = 1;
9669 	} else {
9670 		enab->dten_maxdesc <<= 1;
9671 	}
9672 
9673 	ASSERT(enab->dten_ndesc < enab->dten_maxdesc);
9674 
9675 	nsize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
9676 	ndesc = kmem_zalloc(nsize, KM_SLEEP);
9677 	bcopy(enab->dten_desc, ndesc, osize);
9678 	kmem_free(enab->dten_desc, osize);
9679 
9680 	enab->dten_desc = ndesc;
9681 	enab->dten_desc[enab->dten_ndesc++] = ecb;
9682 }
9683 
9684 static void
9685 dtrace_enabling_addlike(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb,
9686     dtrace_probedesc_t *pd)
9687 {
9688 	dtrace_ecbdesc_t *new;
9689 	dtrace_predicate_t *pred;
9690 	dtrace_actdesc_t *act;
9691 
9692 	/*
9693 	 * We're going to create a new ECB description that matches the
9694 	 * specified ECB in every way, but has the specified probe description.
9695 	 */
9696 	new = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
9697 
9698 	if ((pred = ecb->dted_pred.dtpdd_predicate) != NULL)
9699 		dtrace_predicate_hold(pred);
9700 
9701 	for (act = ecb->dted_action; act != NULL; act = act->dtad_next)
9702 		dtrace_actdesc_hold(act);
9703 
9704 	new->dted_action = ecb->dted_action;
9705 	new->dted_pred = ecb->dted_pred;
9706 	new->dted_probe = *pd;
9707 	new->dted_uarg = ecb->dted_uarg;
9708 
9709 	dtrace_enabling_add(enab, new);
9710 }
9711 
9712 static void
9713 dtrace_enabling_dump(dtrace_enabling_t *enab)
9714 {
9715 	int i;
9716 
9717 	for (i = 0; i < enab->dten_ndesc; i++) {
9718 		dtrace_probedesc_t *desc = &enab->dten_desc[i]->dted_probe;
9719 
9720 		cmn_err(CE_NOTE, "enabling probe %d (%s:%s:%s:%s)", i,
9721 		    desc->dtpd_provider, desc->dtpd_mod,
9722 		    desc->dtpd_func, desc->dtpd_name);
9723 	}
9724 }
9725 
9726 static void
9727 dtrace_enabling_destroy(dtrace_enabling_t *enab)
9728 {
9729 	int i;
9730 	dtrace_ecbdesc_t *ep;
9731 	dtrace_vstate_t *vstate = enab->dten_vstate;
9732 
9733 	ASSERT(MUTEX_HELD(&dtrace_lock));
9734 
9735 	for (i = 0; i < enab->dten_ndesc; i++) {
9736 		dtrace_actdesc_t *act, *next;
9737 		dtrace_predicate_t *pred;
9738 
9739 		ep = enab->dten_desc[i];
9740 
9741 		if ((pred = ep->dted_pred.dtpdd_predicate) != NULL)
9742 			dtrace_predicate_release(pred, vstate);
9743 
9744 		for (act = ep->dted_action; act != NULL; act = next) {
9745 			next = act->dtad_next;
9746 			dtrace_actdesc_release(act, vstate);
9747 		}
9748 
9749 		kmem_free(ep, sizeof (dtrace_ecbdesc_t));
9750 	}
9751 
9752 	kmem_free(enab->dten_desc,
9753 	    enab->dten_maxdesc * sizeof (dtrace_enabling_t *));
9754 
9755 	/*
9756 	 * If this was a retained enabling, decrement the dts_nretained count
9757 	 * and take it off of the dtrace_retained list.
9758 	 */
9759 	if (enab->dten_prev != NULL || enab->dten_next != NULL ||
9760 	    dtrace_retained == enab) {
9761 		ASSERT(enab->dten_vstate->dtvs_state != NULL);
9762 		ASSERT(enab->dten_vstate->dtvs_state->dts_nretained > 0);
9763 		enab->dten_vstate->dtvs_state->dts_nretained--;
9764 	}
9765 
9766 	if (enab->dten_prev == NULL) {
9767 		if (dtrace_retained == enab) {
9768 			dtrace_retained = enab->dten_next;
9769 
9770 			if (dtrace_retained != NULL)
9771 				dtrace_retained->dten_prev = NULL;
9772 		}
9773 	} else {
9774 		ASSERT(enab != dtrace_retained);
9775 		ASSERT(dtrace_retained != NULL);
9776 		enab->dten_prev->dten_next = enab->dten_next;
9777 	}
9778 
9779 	if (enab->dten_next != NULL) {
9780 		ASSERT(dtrace_retained != NULL);
9781 		enab->dten_next->dten_prev = enab->dten_prev;
9782 	}
9783 
9784 	kmem_free(enab, sizeof (dtrace_enabling_t));
9785 }
9786 
9787 static int
9788 dtrace_enabling_retain(dtrace_enabling_t *enab)
9789 {
9790 	dtrace_state_t *state;
9791 
9792 	ASSERT(MUTEX_HELD(&dtrace_lock));
9793 	ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
9794 	ASSERT(enab->dten_vstate != NULL);
9795 
9796 	state = enab->dten_vstate->dtvs_state;
9797 	ASSERT(state != NULL);
9798 
9799 	/*
9800 	 * We only allow each state to retain dtrace_retain_max enablings.
9801 	 */
9802 	if (state->dts_nretained >= dtrace_retain_max)
9803 		return (ENOSPC);
9804 
9805 	state->dts_nretained++;
9806 
9807 	if (dtrace_retained == NULL) {
9808 		dtrace_retained = enab;
9809 		return (0);
9810 	}
9811 
9812 	enab->dten_next = dtrace_retained;
9813 	dtrace_retained->dten_prev = enab;
9814 	dtrace_retained = enab;
9815 
9816 	return (0);
9817 }
9818 
9819 static int
9820 dtrace_enabling_replicate(dtrace_state_t *state, dtrace_probedesc_t *match,
9821     dtrace_probedesc_t *create)
9822 {
9823 	dtrace_enabling_t *new, *enab;
9824 	int found = 0, err = ENOENT;
9825 
9826 	ASSERT(MUTEX_HELD(&dtrace_lock));
9827 	ASSERT(strlen(match->dtpd_provider) < DTRACE_PROVNAMELEN);
9828 	ASSERT(strlen(match->dtpd_mod) < DTRACE_MODNAMELEN);
9829 	ASSERT(strlen(match->dtpd_func) < DTRACE_FUNCNAMELEN);
9830 	ASSERT(strlen(match->dtpd_name) < DTRACE_NAMELEN);
9831 
9832 	new = dtrace_enabling_create(&state->dts_vstate);
9833 
9834 	/*
9835 	 * Iterate over all retained enablings, looking for enablings that
9836 	 * match the specified state.
9837 	 */
9838 	for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
9839 		int i;
9840 
9841 		/*
9842 		 * dtvs_state can only be NULL for helper enablings -- and
9843 		 * helper enablings can't be retained.
9844 		 */
9845 		ASSERT(enab->dten_vstate->dtvs_state != NULL);
9846 
9847 		if (enab->dten_vstate->dtvs_state != state)
9848 			continue;
9849 
9850 		/*
9851 		 * Now iterate over each probe description; we're looking for
9852 		 * an exact match to the specified probe description.
9853 		 */
9854 		for (i = 0; i < enab->dten_ndesc; i++) {
9855 			dtrace_ecbdesc_t *ep = enab->dten_desc[i];
9856 			dtrace_probedesc_t *pd = &ep->dted_probe;
9857 
9858 			if (strcmp(pd->dtpd_provider, match->dtpd_provider))
9859 				continue;
9860 
9861 			if (strcmp(pd->dtpd_mod, match->dtpd_mod))
9862 				continue;
9863 
9864 			if (strcmp(pd->dtpd_func, match->dtpd_func))
9865 				continue;
9866 
9867 			if (strcmp(pd->dtpd_name, match->dtpd_name))
9868 				continue;
9869 
9870 			/*
9871 			 * We have a winning probe!  Add it to our growing
9872 			 * enabling.
9873 			 */
9874 			found = 1;
9875 			dtrace_enabling_addlike(new, ep, create);
9876 		}
9877 	}
9878 
9879 	if (!found || (err = dtrace_enabling_retain(new)) != 0) {
9880 		dtrace_enabling_destroy(new);
9881 		return (err);
9882 	}
9883 
9884 	return (0);
9885 }
9886 
9887 static void
9888 dtrace_enabling_retract(dtrace_state_t *state)
9889 {
9890 	dtrace_enabling_t *enab, *next;
9891 
9892 	ASSERT(MUTEX_HELD(&dtrace_lock));
9893 
9894 	/*
9895 	 * Iterate over all retained enablings, destroy the enablings retained
9896 	 * for the specified state.
9897 	 */
9898 	for (enab = dtrace_retained; enab != NULL; enab = next) {
9899 		next = enab->dten_next;
9900 
9901 		/*
9902 		 * dtvs_state can only be NULL for helper enablings -- and
9903 		 * helper enablings can't be retained.
9904 		 */
9905 		ASSERT(enab->dten_vstate->dtvs_state != NULL);
9906 
9907 		if (enab->dten_vstate->dtvs_state == state) {
9908 			ASSERT(state->dts_nretained > 0);
9909 			dtrace_enabling_destroy(enab);
9910 		}
9911 	}
9912 
9913 	ASSERT(state->dts_nretained == 0);
9914 }
9915 
9916 static int
9917 dtrace_enabling_match(dtrace_enabling_t *enab, int *nmatched)
9918 {
9919 	int i = 0;
9920 	int matched = 0;
9921 
9922 	ASSERT(MUTEX_HELD(&cpu_lock));
9923 	ASSERT(MUTEX_HELD(&dtrace_lock));
9924 
9925 	for (i = 0; i < enab->dten_ndesc; i++) {
9926 		dtrace_ecbdesc_t *ep = enab->dten_desc[i];
9927 
9928 		enab->dten_current = ep;
9929 		enab->dten_error = 0;
9930 
9931 		matched += dtrace_probe_enable(&ep->dted_probe, enab);
9932 
9933 		if (enab->dten_error != 0) {
9934 			/*
9935 			 * If we get an error half-way through enabling the
9936 			 * probes, we kick out -- perhaps with some number of
9937 			 * them enabled.  Leaving enabled probes enabled may
9938 			 * be slightly confusing for user-level, but we expect
9939 			 * that no one will attempt to actually drive on in
9940 			 * the face of such errors.  If this is an anonymous
9941 			 * enabling (indicated with a NULL nmatched pointer),
9942 			 * we cmn_err() a message.  We aren't expecting to
9943 			 * get such an error -- such as it can exist at all,
9944 			 * it would be a result of corrupted DOF in the driver
9945 			 * properties.
9946 			 */
9947 			if (nmatched == NULL) {
9948 				cmn_err(CE_WARN, "dtrace_enabling_match() "
9949 				    "error on %p: %d", (void *)ep,
9950 				    enab->dten_error);
9951 			}
9952 
9953 			return (enab->dten_error);
9954 		}
9955 	}
9956 
9957 	enab->dten_probegen = dtrace_probegen;
9958 	if (nmatched != NULL)
9959 		*nmatched = matched;
9960 
9961 	return (0);
9962 }
9963 
9964 static void
9965 dtrace_enabling_matchall(void)
9966 {
9967 	dtrace_enabling_t *enab;
9968 
9969 	mutex_enter(&cpu_lock);
9970 	mutex_enter(&dtrace_lock);
9971 
9972 	/*
9973 	 * Because we can be called after dtrace_detach() has been called, we
9974 	 * cannot assert that there are retained enablings.  We can safely
9975 	 * load from dtrace_retained, however:  the taskq_destroy() at the
9976 	 * end of dtrace_detach() will block pending our completion.
9977 	 */
9978 	for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next)
9979 		(void) dtrace_enabling_match(enab, NULL);
9980 
9981 	mutex_exit(&dtrace_lock);
9982 	mutex_exit(&cpu_lock);
9983 }
9984 
9985 static int
9986 dtrace_enabling_matchstate(dtrace_state_t *state, int *nmatched)
9987 {
9988 	dtrace_enabling_t *enab;
9989 	int matched, total = 0, err;
9990 
9991 	ASSERT(MUTEX_HELD(&cpu_lock));
9992 	ASSERT(MUTEX_HELD(&dtrace_lock));
9993 
9994 	for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
9995 		ASSERT(enab->dten_vstate->dtvs_state != NULL);
9996 
9997 		if (enab->dten_vstate->dtvs_state != state)
9998 			continue;
9999 
10000 		if ((err = dtrace_enabling_match(enab, &matched)) != 0)
10001 			return (err);
10002 
10003 		total += matched;
10004 	}
10005 
10006 	if (nmatched != NULL)
10007 		*nmatched = total;
10008 
10009 	return (0);
10010 }
10011 
10012 /*
10013  * If an enabling is to be enabled without having matched probes (that is, if
10014  * dtrace_state_go() is to be called on the underlying dtrace_state_t), the
10015  * enabling must be _primed_ by creating an ECB for every ECB description.
10016  * This must be done to assure that we know the number of speculations, the
10017  * number of aggregations, the minimum buffer size needed, etc. before we
10018  * transition out of DTRACE_ACTIVITY_INACTIVE.  To do this without actually
10019  * enabling any probes, we create ECBs for every ECB decription, but with a
10020  * NULL probe -- which is exactly what this function does.
10021  */
10022 static void
10023 dtrace_enabling_prime(dtrace_state_t *state)
10024 {
10025 	dtrace_enabling_t *enab;
10026 	int i;
10027 
10028 	for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
10029 		ASSERT(enab->dten_vstate->dtvs_state != NULL);
10030 
10031 		if (enab->dten_vstate->dtvs_state != state)
10032 			continue;
10033 
10034 		/*
10035 		 * We don't want to prime an enabling more than once, lest
10036 		 * we allow a malicious user to induce resource exhaustion.
10037 		 * (The ECBs that result from priming an enabling aren't
10038 		 * leaked -- but they also aren't deallocated until the
10039 		 * consumer state is destroyed.)
10040 		 */
10041 		if (enab->dten_primed)
10042 			continue;
10043 
10044 		for (i = 0; i < enab->dten_ndesc; i++) {
10045 			enab->dten_current = enab->dten_desc[i];
10046 			(void) dtrace_probe_enable(NULL, enab);
10047 		}
10048 
10049 		enab->dten_primed = 1;
10050 	}
10051 }
10052 
10053 /*
10054  * Called to indicate that probes should be provided due to retained
10055  * enablings.  This is implemented in terms of dtrace_probe_provide(), but it
10056  * must take an initial lap through the enabling calling the dtps_provide()
10057  * entry point explicitly to allow for autocreated probes.
10058  */
10059 static void
10060 dtrace_enabling_provide(dtrace_provider_t *prv)
10061 {
10062 	int i, all = 0;
10063 	dtrace_probedesc_t desc;
10064 
10065 	ASSERT(MUTEX_HELD(&dtrace_lock));
10066 	ASSERT(MUTEX_HELD(&dtrace_provider_lock));
10067 
10068 	if (prv == NULL) {
10069 		all = 1;
10070 		prv = dtrace_provider;
10071 	}
10072 
10073 	do {
10074 		dtrace_enabling_t *enab = dtrace_retained;
10075 		void *parg = prv->dtpv_arg;
10076 
10077 		for (; enab != NULL; enab = enab->dten_next) {
10078 			for (i = 0; i < enab->dten_ndesc; i++) {
10079 				desc = enab->dten_desc[i]->dted_probe;
10080 				mutex_exit(&dtrace_lock);
10081 				prv->dtpv_pops.dtps_provide(parg, &desc);
10082 				mutex_enter(&dtrace_lock);
10083 			}
10084 		}
10085 	} while (all && (prv = prv->dtpv_next) != NULL);
10086 
10087 	mutex_exit(&dtrace_lock);
10088 	dtrace_probe_provide(NULL, all ? NULL : prv);
10089 	mutex_enter(&dtrace_lock);
10090 }
10091 
10092 /*
10093  * DTrace DOF Functions
10094  */
10095 /*ARGSUSED*/
10096 static void
10097 dtrace_dof_error(dof_hdr_t *dof, const char *str)
10098 {
10099 	if (dtrace_err_verbose)
10100 		cmn_err(CE_WARN, "failed to process DOF: %s", str);
10101 
10102 #ifdef DTRACE_ERRDEBUG
10103 	dtrace_errdebug(str);
10104 #endif
10105 }
10106 
10107 /*
10108  * Create DOF out of a currently enabled state.  Right now, we only create
10109  * DOF containing the run-time options -- but this could be expanded to create
10110  * complete DOF representing the enabled state.
10111  */
10112 static dof_hdr_t *
10113 dtrace_dof_create(dtrace_state_t *state)
10114 {
10115 	dof_hdr_t *dof;
10116 	dof_sec_t *sec;
10117 	dof_optdesc_t *opt;
10118 	int i, len = sizeof (dof_hdr_t) +
10119 	    roundup(sizeof (dof_sec_t), sizeof (uint64_t)) +
10120 	    sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
10121 
10122 	ASSERT(MUTEX_HELD(&dtrace_lock));
10123 
10124 	dof = kmem_zalloc(len, KM_SLEEP);
10125 	dof->dofh_ident[DOF_ID_MAG0] = DOF_MAG_MAG0;
10126 	dof->dofh_ident[DOF_ID_MAG1] = DOF_MAG_MAG1;
10127 	dof->dofh_ident[DOF_ID_MAG2] = DOF_MAG_MAG2;
10128 	dof->dofh_ident[DOF_ID_MAG3] = DOF_MAG_MAG3;
10129 
10130 	dof->dofh_ident[DOF_ID_MODEL] = DOF_MODEL_NATIVE;
10131 	dof->dofh_ident[DOF_ID_ENCODING] = DOF_ENCODE_NATIVE;
10132 	dof->dofh_ident[DOF_ID_VERSION] = DOF_VERSION;
10133 	dof->dofh_ident[DOF_ID_DIFVERS] = DIF_VERSION;
10134 	dof->dofh_ident[DOF_ID_DIFIREG] = DIF_DIR_NREGS;
10135 	dof->dofh_ident[DOF_ID_DIFTREG] = DIF_DTR_NREGS;
10136 
10137 	dof->dofh_flags = 0;
10138 	dof->dofh_hdrsize = sizeof (dof_hdr_t);
10139 	dof->dofh_secsize = sizeof (dof_sec_t);
10140 	dof->dofh_secnum = 1;	/* only DOF_SECT_OPTDESC */
10141 	dof->dofh_secoff = sizeof (dof_hdr_t);
10142 	dof->dofh_loadsz = len;
10143 	dof->dofh_filesz = len;
10144 	dof->dofh_pad = 0;
10145 
10146 	/*
10147 	 * Fill in the option section header...
10148 	 */
10149 	sec = (dof_sec_t *)((uintptr_t)dof + sizeof (dof_hdr_t));
10150 	sec->dofs_type = DOF_SECT_OPTDESC;
10151 	sec->dofs_align = sizeof (uint64_t);
10152 	sec->dofs_flags = DOF_SECF_LOAD;
10153 	sec->dofs_entsize = sizeof (dof_optdesc_t);
10154 
10155 	opt = (dof_optdesc_t *)((uintptr_t)sec +
10156 	    roundup(sizeof (dof_sec_t), sizeof (uint64_t)));
10157 
10158 	sec->dofs_offset = (uintptr_t)opt - (uintptr_t)dof;
10159 	sec->dofs_size = sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
10160 
10161 	for (i = 0; i < DTRACEOPT_MAX; i++) {
10162 		opt[i].dofo_option = i;
10163 		opt[i].dofo_strtab = DOF_SECIDX_NONE;
10164 		opt[i].dofo_value = state->dts_options[i];
10165 	}
10166 
10167 	return (dof);
10168 }
10169 
10170 static dof_hdr_t *
10171 dtrace_dof_copyin(uintptr_t uarg, int *errp)
10172 {
10173 	dof_hdr_t hdr, *dof;
10174 
10175 	ASSERT(!MUTEX_HELD(&dtrace_lock));
10176 
10177 	/*
10178 	 * First, we're going to copyin() the sizeof (dof_hdr_t).
10179 	 */
10180 	if (copyin((void *)uarg, &hdr, sizeof (hdr)) != 0) {
10181 		dtrace_dof_error(NULL, "failed to copyin DOF header");
10182 		*errp = EFAULT;
10183 		return (NULL);
10184 	}
10185 
10186 	/*
10187 	 * Now we'll allocate the entire DOF and copy it in -- provided
10188 	 * that the length isn't outrageous.
10189 	 */
10190 	if (hdr.dofh_loadsz >= dtrace_dof_maxsize) {
10191 		dtrace_dof_error(&hdr, "load size exceeds maximum");
10192 		*errp = E2BIG;
10193 		return (NULL);
10194 	}
10195 
10196 	if (hdr.dofh_loadsz < sizeof (hdr)) {
10197 		dtrace_dof_error(&hdr, "invalid load size");
10198 		*errp = EINVAL;
10199 		return (NULL);
10200 	}
10201 
10202 	dof = kmem_alloc(hdr.dofh_loadsz, KM_SLEEP);
10203 
10204 	if (copyin((void *)uarg, dof, hdr.dofh_loadsz) != 0) {
10205 		kmem_free(dof, hdr.dofh_loadsz);
10206 		*errp = EFAULT;
10207 		return (NULL);
10208 	}
10209 
10210 	return (dof);
10211 }
10212 
10213 static dof_hdr_t *
10214 dtrace_dof_property(const char *name)
10215 {
10216 	uchar_t *buf;
10217 	uint64_t loadsz;
10218 	unsigned int len, i;
10219 	dof_hdr_t *dof;
10220 
10221 	/*
10222 	 * Unfortunately, array of values in .conf files are always (and
10223 	 * only) interpreted to be integer arrays.  We must read our DOF
10224 	 * as an integer array, and then squeeze it into a byte array.
10225 	 */
10226 	if (ddi_prop_lookup_int_array(DDI_DEV_T_ANY, dtrace_devi, 0,
10227 	    (char *)name, (int **)&buf, &len) != DDI_PROP_SUCCESS)
10228 		return (NULL);
10229 
10230 	for (i = 0; i < len; i++)
10231 		buf[i] = (uchar_t)(((int *)buf)[i]);
10232 
10233 	if (len < sizeof (dof_hdr_t)) {
10234 		ddi_prop_free(buf);
10235 		dtrace_dof_error(NULL, "truncated header");
10236 		return (NULL);
10237 	}
10238 
10239 	if (len < (loadsz = ((dof_hdr_t *)buf)->dofh_loadsz)) {
10240 		ddi_prop_free(buf);
10241 		dtrace_dof_error(NULL, "truncated DOF");
10242 		return (NULL);
10243 	}
10244 
10245 	if (loadsz >= dtrace_dof_maxsize) {
10246 		ddi_prop_free(buf);
10247 		dtrace_dof_error(NULL, "oversized DOF");
10248 		return (NULL);
10249 	}
10250 
10251 	dof = kmem_alloc(loadsz, KM_SLEEP);
10252 	bcopy(buf, dof, loadsz);
10253 	ddi_prop_free(buf);
10254 
10255 	return (dof);
10256 }
10257 
10258 static void
10259 dtrace_dof_destroy(dof_hdr_t *dof)
10260 {
10261 	kmem_free(dof, dof->dofh_loadsz);
10262 }
10263 
10264 /*
10265  * Return the dof_sec_t pointer corresponding to a given section index.  If the
10266  * index is not valid, dtrace_dof_error() is called and NULL is returned.  If
10267  * a type other than DOF_SECT_NONE is specified, the header is checked against
10268  * this type and NULL is returned if the types do not match.
10269  */
10270 static dof_sec_t *
10271 dtrace_dof_sect(dof_hdr_t *dof, uint32_t type, dof_secidx_t i)
10272 {
10273 	dof_sec_t *sec = (dof_sec_t *)(uintptr_t)
10274 	    ((uintptr_t)dof + dof->dofh_secoff + i * dof->dofh_secsize);
10275 
10276 	if (i >= dof->dofh_secnum) {
10277 		dtrace_dof_error(dof, "referenced section index is invalid");
10278 		return (NULL);
10279 	}
10280 
10281 	if (!(sec->dofs_flags & DOF_SECF_LOAD)) {
10282 		dtrace_dof_error(dof, "referenced section is not loadable");
10283 		return (NULL);
10284 	}
10285 
10286 	if (type != DOF_SECT_NONE && type != sec->dofs_type) {
10287 		dtrace_dof_error(dof, "referenced section is the wrong type");
10288 		return (NULL);
10289 	}
10290 
10291 	return (sec);
10292 }
10293 
10294 static dtrace_probedesc_t *
10295 dtrace_dof_probedesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_probedesc_t *desc)
10296 {
10297 	dof_probedesc_t *probe;
10298 	dof_sec_t *strtab;
10299 	uintptr_t daddr = (uintptr_t)dof;
10300 	uintptr_t str;
10301 	size_t size;
10302 
10303 	if (sec->dofs_type != DOF_SECT_PROBEDESC) {
10304 		dtrace_dof_error(dof, "invalid probe section");
10305 		return (NULL);
10306 	}
10307 
10308 	if (sec->dofs_align != sizeof (dof_secidx_t)) {
10309 		dtrace_dof_error(dof, "bad alignment in probe description");
10310 		return (NULL);
10311 	}
10312 
10313 	if (sec->dofs_offset + sizeof (dof_probedesc_t) > dof->dofh_loadsz) {
10314 		dtrace_dof_error(dof, "truncated probe description");
10315 		return (NULL);
10316 	}
10317 
10318 	probe = (dof_probedesc_t *)(uintptr_t)(daddr + sec->dofs_offset);
10319 	strtab = dtrace_dof_sect(dof, DOF_SECT_STRTAB, probe->dofp_strtab);
10320 
10321 	if (strtab == NULL)
10322 		return (NULL);
10323 
10324 	str = daddr + strtab->dofs_offset;
10325 	size = strtab->dofs_size;
10326 
10327 	if (probe->dofp_provider >= strtab->dofs_size) {
10328 		dtrace_dof_error(dof, "corrupt probe provider");
10329 		return (NULL);
10330 	}
10331 
10332 	(void) strncpy(desc->dtpd_provider,
10333 	    (char *)(str + probe->dofp_provider),
10334 	    MIN(DTRACE_PROVNAMELEN - 1, size - probe->dofp_provider));
10335 
10336 	if (probe->dofp_mod >= strtab->dofs_size) {
10337 		dtrace_dof_error(dof, "corrupt probe module");
10338 		return (NULL);
10339 	}
10340 
10341 	(void) strncpy(desc->dtpd_mod, (char *)(str + probe->dofp_mod),
10342 	    MIN(DTRACE_MODNAMELEN - 1, size - probe->dofp_mod));
10343 
10344 	if (probe->dofp_func >= strtab->dofs_size) {
10345 		dtrace_dof_error(dof, "corrupt probe function");
10346 		return (NULL);
10347 	}
10348 
10349 	(void) strncpy(desc->dtpd_func, (char *)(str + probe->dofp_func),
10350 	    MIN(DTRACE_FUNCNAMELEN - 1, size - probe->dofp_func));
10351 
10352 	if (probe->dofp_name >= strtab->dofs_size) {
10353 		dtrace_dof_error(dof, "corrupt probe name");
10354 		return (NULL);
10355 	}
10356 
10357 	(void) strncpy(desc->dtpd_name, (char *)(str + probe->dofp_name),
10358 	    MIN(DTRACE_NAMELEN - 1, size - probe->dofp_name));
10359 
10360 	return (desc);
10361 }
10362 
10363 static dtrace_difo_t *
10364 dtrace_dof_difo(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
10365     cred_t *cr)
10366 {
10367 	dtrace_difo_t *dp;
10368 	size_t ttl = 0;
10369 	dof_difohdr_t *dofd;
10370 	uintptr_t daddr = (uintptr_t)dof;
10371 	size_t max = dtrace_difo_maxsize;
10372 	int i, l, n;
10373 
10374 	static const struct {
10375 		int section;
10376 		int bufoffs;
10377 		int lenoffs;
10378 		int entsize;
10379 		int align;
10380 		const char *msg;
10381 	} difo[] = {
10382 		{ DOF_SECT_DIF, offsetof(dtrace_difo_t, dtdo_buf),
10383 		offsetof(dtrace_difo_t, dtdo_len), sizeof (dif_instr_t),
10384 		sizeof (dif_instr_t), "multiple DIF sections" },
10385 
10386 		{ DOF_SECT_INTTAB, offsetof(dtrace_difo_t, dtdo_inttab),
10387 		offsetof(dtrace_difo_t, dtdo_intlen), sizeof (uint64_t),
10388 		sizeof (uint64_t), "multiple integer tables" },
10389 
10390 		{ DOF_SECT_STRTAB, offsetof(dtrace_difo_t, dtdo_strtab),
10391 		offsetof(dtrace_difo_t, dtdo_strlen), 0,
10392 		sizeof (char), "multiple string tables" },
10393 
10394 		{ DOF_SECT_VARTAB, offsetof(dtrace_difo_t, dtdo_vartab),
10395 		offsetof(dtrace_difo_t, dtdo_varlen), sizeof (dtrace_difv_t),
10396 		sizeof (uint_t), "multiple variable tables" },
10397 
10398 		{ DOF_SECT_NONE, 0, 0, 0, NULL }
10399 	};
10400 
10401 	if (sec->dofs_type != DOF_SECT_DIFOHDR) {
10402 		dtrace_dof_error(dof, "invalid DIFO header section");
10403 		return (NULL);
10404 	}
10405 
10406 	if (sec->dofs_align != sizeof (dof_secidx_t)) {
10407 		dtrace_dof_error(dof, "bad alignment in DIFO header");
10408 		return (NULL);
10409 	}
10410 
10411 	if (sec->dofs_size < sizeof (dof_difohdr_t) ||
10412 	    sec->dofs_size % sizeof (dof_secidx_t)) {
10413 		dtrace_dof_error(dof, "bad size in DIFO header");
10414 		return (NULL);
10415 	}
10416 
10417 	dofd = (dof_difohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
10418 	n = (sec->dofs_size - sizeof (*dofd)) / sizeof (dof_secidx_t) + 1;
10419 
10420 	dp = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
10421 	dp->dtdo_rtype = dofd->dofd_rtype;
10422 
10423 	for (l = 0; l < n; l++) {
10424 		dof_sec_t *subsec;
10425 		void **bufp;
10426 		uint32_t *lenp;
10427 
10428 		if ((subsec = dtrace_dof_sect(dof, DOF_SECT_NONE,
10429 		    dofd->dofd_links[l])) == NULL)
10430 			goto err; /* invalid section link */
10431 
10432 		if (ttl + subsec->dofs_size > max) {
10433 			dtrace_dof_error(dof, "exceeds maximum size");
10434 			goto err;
10435 		}
10436 
10437 		ttl += subsec->dofs_size;
10438 
10439 		for (i = 0; difo[i].section != DOF_SECT_NONE; i++) {
10440 			if (subsec->dofs_type != difo[i].section)
10441 				continue;
10442 
10443 			if (!(subsec->dofs_flags & DOF_SECF_LOAD)) {
10444 				dtrace_dof_error(dof, "section not loaded");
10445 				goto err;
10446 			}
10447 
10448 			if (subsec->dofs_align != difo[i].align) {
10449 				dtrace_dof_error(dof, "bad alignment");
10450 				goto err;
10451 			}
10452 
10453 			bufp = (void **)((uintptr_t)dp + difo[i].bufoffs);
10454 			lenp = (uint32_t *)((uintptr_t)dp + difo[i].lenoffs);
10455 
10456 			if (*bufp != NULL) {
10457 				dtrace_dof_error(dof, difo[i].msg);
10458 				goto err;
10459 			}
10460 
10461 			if (difo[i].entsize != subsec->dofs_entsize) {
10462 				dtrace_dof_error(dof, "entry size mismatch");
10463 				goto err;
10464 			}
10465 
10466 			if (subsec->dofs_entsize != 0 &&
10467 			    (subsec->dofs_size % subsec->dofs_entsize) != 0) {
10468 				dtrace_dof_error(dof, "corrupt entry size");
10469 				goto err;
10470 			}
10471 
10472 			*lenp = subsec->dofs_size;
10473 			*bufp = kmem_alloc(subsec->dofs_size, KM_SLEEP);
10474 			bcopy((char *)(uintptr_t)(daddr + subsec->dofs_offset),
10475 			    *bufp, subsec->dofs_size);
10476 
10477 			if (subsec->dofs_entsize != 0)
10478 				*lenp /= subsec->dofs_entsize;
10479 
10480 			break;
10481 		}
10482 
10483 		/*
10484 		 * If we encounter a loadable DIFO sub-section that is not
10485 		 * known to us, assume this is a broken program and fail.
10486 		 */
10487 		if (difo[i].section == DOF_SECT_NONE &&
10488 		    (subsec->dofs_flags & DOF_SECF_LOAD)) {
10489 			dtrace_dof_error(dof, "unrecognized DIFO subsection");
10490 			goto err;
10491 		}
10492 	}
10493 
10494 	if (dp->dtdo_buf == NULL) {
10495 		/*
10496 		 * We can't have a DIF object without DIF text.
10497 		 */
10498 		dtrace_dof_error(dof, "missing DIF text");
10499 		goto err;
10500 	}
10501 
10502 	/*
10503 	 * Before we validate the DIF object, run through the variable table
10504 	 * looking for the strings -- if any of their size are under, we'll set
10505 	 * their size to be the system-wide default string size.  Note that
10506 	 * this should _not_ happen if the "strsize" option has been set --
10507 	 * in this case, the compiler should have set the size to reflect the
10508 	 * setting of the option.
10509 	 */
10510 	for (i = 0; i < dp->dtdo_varlen; i++) {
10511 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
10512 		dtrace_diftype_t *t = &v->dtdv_type;
10513 
10514 		if (v->dtdv_id < DIF_VAR_OTHER_UBASE)
10515 			continue;
10516 
10517 		if (t->dtdt_kind == DIF_TYPE_STRING && t->dtdt_size == 0)
10518 			t->dtdt_size = dtrace_strsize_default;
10519 	}
10520 
10521 	if (dtrace_difo_validate(dp, vstate, DIF_DIR_NREGS, cr) != 0)
10522 		goto err;
10523 
10524 	dtrace_difo_init(dp, vstate);
10525 	return (dp);
10526 
10527 err:
10528 	kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
10529 	kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
10530 	kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
10531 	kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
10532 
10533 	kmem_free(dp, sizeof (dtrace_difo_t));
10534 	return (NULL);
10535 }
10536 
10537 static dtrace_predicate_t *
10538 dtrace_dof_predicate(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
10539     cred_t *cr)
10540 {
10541 	dtrace_difo_t *dp;
10542 
10543 	if ((dp = dtrace_dof_difo(dof, sec, vstate, cr)) == NULL)
10544 		return (NULL);
10545 
10546 	return (dtrace_predicate_create(dp));
10547 }
10548 
10549 static dtrace_actdesc_t *
10550 dtrace_dof_actdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
10551     cred_t *cr)
10552 {
10553 	dtrace_actdesc_t *act, *first = NULL, *last = NULL, *next;
10554 	dof_actdesc_t *desc;
10555 	dof_sec_t *difosec;
10556 	size_t offs;
10557 	uintptr_t daddr = (uintptr_t)dof;
10558 	uint64_t arg;
10559 	dtrace_actkind_t kind;
10560 
10561 	if (sec->dofs_type != DOF_SECT_ACTDESC) {
10562 		dtrace_dof_error(dof, "invalid action section");
10563 		return (NULL);
10564 	}
10565 
10566 	if (sec->dofs_offset + sizeof (dof_actdesc_t) > dof->dofh_loadsz) {
10567 		dtrace_dof_error(dof, "truncated action description");
10568 		return (NULL);
10569 	}
10570 
10571 	if (sec->dofs_align != sizeof (uint64_t)) {
10572 		dtrace_dof_error(dof, "bad alignment in action description");
10573 		return (NULL);
10574 	}
10575 
10576 	if (sec->dofs_size < sec->dofs_entsize) {
10577 		dtrace_dof_error(dof, "section entry size exceeds total size");
10578 		return (NULL);
10579 	}
10580 
10581 	if (sec->dofs_entsize != sizeof (dof_actdesc_t)) {
10582 		dtrace_dof_error(dof, "bad entry size in action description");
10583 		return (NULL);
10584 	}
10585 
10586 	if (sec->dofs_size / sec->dofs_entsize > dtrace_actions_max) {
10587 		dtrace_dof_error(dof, "actions exceed dtrace_actions_max");
10588 		return (NULL);
10589 	}
10590 
10591 	for (offs = 0; offs < sec->dofs_size; offs += sec->dofs_entsize) {
10592 		desc = (dof_actdesc_t *)(daddr +
10593 		    (uintptr_t)sec->dofs_offset + offs);
10594 		kind = (dtrace_actkind_t)desc->dofa_kind;
10595 
10596 		if (DTRACEACT_ISPRINTFLIKE(kind) &&
10597 		    (kind != DTRACEACT_PRINTA ||
10598 		    desc->dofa_strtab != DOF_SECIDX_NONE)) {
10599 			dof_sec_t *strtab;
10600 			char *str, *fmt;
10601 			uint64_t i;
10602 
10603 			/*
10604 			 * printf()-like actions must have a format string.
10605 			 */
10606 			if ((strtab = dtrace_dof_sect(dof,
10607 			    DOF_SECT_STRTAB, desc->dofa_strtab)) == NULL)
10608 				goto err;
10609 
10610 			str = (char *)((uintptr_t)dof +
10611 			    (uintptr_t)strtab->dofs_offset);
10612 
10613 			for (i = desc->dofa_arg; i < strtab->dofs_size; i++) {
10614 				if (str[i] == '\0')
10615 					break;
10616 			}
10617 
10618 			if (i >= strtab->dofs_size) {
10619 				dtrace_dof_error(dof, "bogus format string");
10620 				goto err;
10621 			}
10622 
10623 			if (i == desc->dofa_arg) {
10624 				dtrace_dof_error(dof, "empty format string");
10625 				goto err;
10626 			}
10627 
10628 			i -= desc->dofa_arg;
10629 			fmt = kmem_alloc(i + 1, KM_SLEEP);
10630 			bcopy(&str[desc->dofa_arg], fmt, i + 1);
10631 			arg = (uint64_t)(uintptr_t)fmt;
10632 		} else {
10633 			if (kind == DTRACEACT_PRINTA) {
10634 				ASSERT(desc->dofa_strtab == DOF_SECIDX_NONE);
10635 				arg = 0;
10636 			} else {
10637 				arg = desc->dofa_arg;
10638 			}
10639 		}
10640 
10641 		act = dtrace_actdesc_create(kind, desc->dofa_ntuple,
10642 		    desc->dofa_uarg, arg);
10643 
10644 		if (last != NULL) {
10645 			last->dtad_next = act;
10646 		} else {
10647 			first = act;
10648 		}
10649 
10650 		last = act;
10651 
10652 		if (desc->dofa_difo == DOF_SECIDX_NONE)
10653 			continue;
10654 
10655 		if ((difosec = dtrace_dof_sect(dof,
10656 		    DOF_SECT_DIFOHDR, desc->dofa_difo)) == NULL)
10657 			goto err;
10658 
10659 		act->dtad_difo = dtrace_dof_difo(dof, difosec, vstate, cr);
10660 
10661 		if (act->dtad_difo == NULL)
10662 			goto err;
10663 	}
10664 
10665 	ASSERT(first != NULL);
10666 	return (first);
10667 
10668 err:
10669 	for (act = first; act != NULL; act = next) {
10670 		next = act->dtad_next;
10671 		dtrace_actdesc_release(act, vstate);
10672 	}
10673 
10674 	return (NULL);
10675 }
10676 
10677 static dtrace_ecbdesc_t *
10678 dtrace_dof_ecbdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
10679     cred_t *cr)
10680 {
10681 	dtrace_ecbdesc_t *ep;
10682 	dof_ecbdesc_t *ecb;
10683 	dtrace_probedesc_t *desc;
10684 	dtrace_predicate_t *pred = NULL;
10685 
10686 	if (sec->dofs_size < sizeof (dof_ecbdesc_t)) {
10687 		dtrace_dof_error(dof, "truncated ECB description");
10688 		return (NULL);
10689 	}
10690 
10691 	if (sec->dofs_align != sizeof (uint64_t)) {
10692 		dtrace_dof_error(dof, "bad alignment in ECB description");
10693 		return (NULL);
10694 	}
10695 
10696 	ecb = (dof_ecbdesc_t *)((uintptr_t)dof + (uintptr_t)sec->dofs_offset);
10697 	sec = dtrace_dof_sect(dof, DOF_SECT_PROBEDESC, ecb->dofe_probes);
10698 
10699 	if (sec == NULL)
10700 		return (NULL);
10701 
10702 	ep = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
10703 	ep->dted_uarg = ecb->dofe_uarg;
10704 	desc = &ep->dted_probe;
10705 
10706 	if (dtrace_dof_probedesc(dof, sec, desc) == NULL)
10707 		goto err;
10708 
10709 	if (ecb->dofe_pred != DOF_SECIDX_NONE) {
10710 		if ((sec = dtrace_dof_sect(dof,
10711 		    DOF_SECT_DIFOHDR, ecb->dofe_pred)) == NULL)
10712 			goto err;
10713 
10714 		if ((pred = dtrace_dof_predicate(dof, sec, vstate, cr)) == NULL)
10715 			goto err;
10716 
10717 		ep->dted_pred.dtpdd_predicate = pred;
10718 	}
10719 
10720 	if (ecb->dofe_actions != DOF_SECIDX_NONE) {
10721 		if ((sec = dtrace_dof_sect(dof,
10722 		    DOF_SECT_ACTDESC, ecb->dofe_actions)) == NULL)
10723 			goto err;
10724 
10725 		ep->dted_action = dtrace_dof_actdesc(dof, sec, vstate, cr);
10726 
10727 		if (ep->dted_action == NULL)
10728 			goto err;
10729 	}
10730 
10731 	return (ep);
10732 
10733 err:
10734 	if (pred != NULL)
10735 		dtrace_predicate_release(pred, vstate);
10736 	kmem_free(ep, sizeof (dtrace_ecbdesc_t));
10737 	return (NULL);
10738 }
10739 
10740 /*
10741  * Apply the relocations from the specified 'sec' (a DOF_SECT_URELHDR) to the
10742  * specified DOF.  At present, this amounts to simply adding 'ubase' to the
10743  * site of any user SETX relocations to account for load object base address.
10744  * In the future, if we need other relocations, this function can be extended.
10745  */
10746 static int
10747 dtrace_dof_relocate(dof_hdr_t *dof, dof_sec_t *sec, uint64_t ubase)
10748 {
10749 	uintptr_t daddr = (uintptr_t)dof;
10750 	dof_relohdr_t *dofr =
10751 	    (dof_relohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
10752 	dof_sec_t *ss, *rs, *ts;
10753 	dof_relodesc_t *r;
10754 	uint_t i, n;
10755 
10756 	if (sec->dofs_size < sizeof (dof_relohdr_t) ||
10757 	    sec->dofs_align != sizeof (dof_secidx_t)) {
10758 		dtrace_dof_error(dof, "invalid relocation header");
10759 		return (-1);
10760 	}
10761 
10762 	ss = dtrace_dof_sect(dof, DOF_SECT_STRTAB, dofr->dofr_strtab);
10763 	rs = dtrace_dof_sect(dof, DOF_SECT_RELTAB, dofr->dofr_relsec);
10764 	ts = dtrace_dof_sect(dof, DOF_SECT_NONE, dofr->dofr_tgtsec);
10765 
10766 	if (ss == NULL || rs == NULL || ts == NULL)
10767 		return (-1); /* dtrace_dof_error() has been called already */
10768 
10769 	if (rs->dofs_entsize < sizeof (dof_relodesc_t) ||
10770 	    rs->dofs_align != sizeof (uint64_t)) {
10771 		dtrace_dof_error(dof, "invalid relocation section");
10772 		return (-1);
10773 	}
10774 
10775 	r = (dof_relodesc_t *)(uintptr_t)(daddr + rs->dofs_offset);
10776 	n = rs->dofs_size / rs->dofs_entsize;
10777 
10778 	for (i = 0; i < n; i++) {
10779 		uintptr_t taddr = daddr + ts->dofs_offset + r->dofr_offset;
10780 
10781 		switch (r->dofr_type) {
10782 		case DOF_RELO_NONE:
10783 			break;
10784 		case DOF_RELO_SETX:
10785 			if (r->dofr_offset >= ts->dofs_size || r->dofr_offset +
10786 			    sizeof (uint64_t) > ts->dofs_size) {
10787 				dtrace_dof_error(dof, "bad relocation offset");
10788 				return (-1);
10789 			}
10790 
10791 			if (!IS_P2ALIGNED(taddr, sizeof (uint64_t))) {
10792 				dtrace_dof_error(dof, "misaligned setx relo");
10793 				return (-1);
10794 			}
10795 
10796 			*(uint64_t *)taddr += ubase;
10797 			break;
10798 		default:
10799 			dtrace_dof_error(dof, "invalid relocation type");
10800 			return (-1);
10801 		}
10802 
10803 		r = (dof_relodesc_t *)((uintptr_t)r + rs->dofs_entsize);
10804 	}
10805 
10806 	return (0);
10807 }
10808 
10809 /*
10810  * The dof_hdr_t passed to dtrace_dof_slurp() should be a partially validated
10811  * header:  it should be at the front of a memory region that is at least
10812  * sizeof (dof_hdr_t) in size -- and then at least dof_hdr.dofh_loadsz in
10813  * size.  It need not be validated in any other way.
10814  */
10815 static int
10816 dtrace_dof_slurp(dof_hdr_t *dof, dtrace_vstate_t *vstate, cred_t *cr,
10817     dtrace_enabling_t **enabp, uint64_t ubase, int noprobes)
10818 {
10819 	uint64_t len = dof->dofh_loadsz, seclen;
10820 	uintptr_t daddr = (uintptr_t)dof;
10821 	dtrace_ecbdesc_t *ep;
10822 	dtrace_enabling_t *enab;
10823 	uint_t i;
10824 
10825 	ASSERT(MUTEX_HELD(&dtrace_lock));
10826 	ASSERT(dof->dofh_loadsz >= sizeof (dof_hdr_t));
10827 
10828 	/*
10829 	 * Check the DOF header identification bytes.  In addition to checking
10830 	 * valid settings, we also verify that unused bits/bytes are zeroed so
10831 	 * we can use them later without fear of regressing existing binaries.
10832 	 */
10833 	if (bcmp(&dof->dofh_ident[DOF_ID_MAG0],
10834 	    DOF_MAG_STRING, DOF_MAG_STRLEN) != 0) {
10835 		dtrace_dof_error(dof, "DOF magic string mismatch");
10836 		return (-1);
10837 	}
10838 
10839 	if (dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_ILP32 &&
10840 	    dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_LP64) {
10841 		dtrace_dof_error(dof, "DOF has invalid data model");
10842 		return (-1);
10843 	}
10844 
10845 	if (dof->dofh_ident[DOF_ID_ENCODING] != DOF_ENCODE_NATIVE) {
10846 		dtrace_dof_error(dof, "DOF encoding mismatch");
10847 		return (-1);
10848 	}
10849 
10850 	if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
10851 	    dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_2) {
10852 		dtrace_dof_error(dof, "DOF version mismatch");
10853 		return (-1);
10854 	}
10855 
10856 	if (dof->dofh_ident[DOF_ID_DIFVERS] != DIF_VERSION_2) {
10857 		dtrace_dof_error(dof, "DOF uses unsupported instruction set");
10858 		return (-1);
10859 	}
10860 
10861 	if (dof->dofh_ident[DOF_ID_DIFIREG] > DIF_DIR_NREGS) {
10862 		dtrace_dof_error(dof, "DOF uses too many integer registers");
10863 		return (-1);
10864 	}
10865 
10866 	if (dof->dofh_ident[DOF_ID_DIFTREG] > DIF_DTR_NREGS) {
10867 		dtrace_dof_error(dof, "DOF uses too many tuple registers");
10868 		return (-1);
10869 	}
10870 
10871 	for (i = DOF_ID_PAD; i < DOF_ID_SIZE; i++) {
10872 		if (dof->dofh_ident[i] != 0) {
10873 			dtrace_dof_error(dof, "DOF has invalid ident byte set");
10874 			return (-1);
10875 		}
10876 	}
10877 
10878 	if (dof->dofh_flags & ~DOF_FL_VALID) {
10879 		dtrace_dof_error(dof, "DOF has invalid flag bits set");
10880 		return (-1);
10881 	}
10882 
10883 	if (dof->dofh_secsize == 0) {
10884 		dtrace_dof_error(dof, "zero section header size");
10885 		return (-1);
10886 	}
10887 
10888 	/*
10889 	 * Check that the section headers don't exceed the amount of DOF
10890 	 * data.  Note that we cast the section size and number of sections
10891 	 * to uint64_t's to prevent possible overflow in the multiplication.
10892 	 */
10893 	seclen = (uint64_t)dof->dofh_secnum * (uint64_t)dof->dofh_secsize;
10894 
10895 	if (dof->dofh_secoff > len || seclen > len ||
10896 	    dof->dofh_secoff + seclen > len) {
10897 		dtrace_dof_error(dof, "truncated section headers");
10898 		return (-1);
10899 	}
10900 
10901 	if (!IS_P2ALIGNED(dof->dofh_secoff, sizeof (uint64_t))) {
10902 		dtrace_dof_error(dof, "misaligned section headers");
10903 		return (-1);
10904 	}
10905 
10906 	if (!IS_P2ALIGNED(dof->dofh_secsize, sizeof (uint64_t))) {
10907 		dtrace_dof_error(dof, "misaligned section size");
10908 		return (-1);
10909 	}
10910 
10911 	/*
10912 	 * Take an initial pass through the section headers to be sure that
10913 	 * the headers don't have stray offsets.  If the 'noprobes' flag is
10914 	 * set, do not permit sections relating to providers, probes, or args.
10915 	 */
10916 	for (i = 0; i < dof->dofh_secnum; i++) {
10917 		dof_sec_t *sec = (dof_sec_t *)(daddr +
10918 		    (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
10919 
10920 		if (noprobes) {
10921 			switch (sec->dofs_type) {
10922 			case DOF_SECT_PROVIDER:
10923 			case DOF_SECT_PROBES:
10924 			case DOF_SECT_PRARGS:
10925 			case DOF_SECT_PROFFS:
10926 				dtrace_dof_error(dof, "illegal sections "
10927 				    "for enabling");
10928 				return (-1);
10929 			}
10930 		}
10931 
10932 		if (!(sec->dofs_flags & DOF_SECF_LOAD))
10933 			continue; /* just ignore non-loadable sections */
10934 
10935 		if (sec->dofs_align & (sec->dofs_align - 1)) {
10936 			dtrace_dof_error(dof, "bad section alignment");
10937 			return (-1);
10938 		}
10939 
10940 		if (sec->dofs_offset & (sec->dofs_align - 1)) {
10941 			dtrace_dof_error(dof, "misaligned section");
10942 			return (-1);
10943 		}
10944 
10945 		if (sec->dofs_offset > len || sec->dofs_size > len ||
10946 		    sec->dofs_offset + sec->dofs_size > len) {
10947 			dtrace_dof_error(dof, "corrupt section header");
10948 			return (-1);
10949 		}
10950 
10951 		if (sec->dofs_type == DOF_SECT_STRTAB && *((char *)daddr +
10952 		    sec->dofs_offset + sec->dofs_size - 1) != '\0') {
10953 			dtrace_dof_error(dof, "non-terminating string table");
10954 			return (-1);
10955 		}
10956 	}
10957 
10958 	/*
10959 	 * Take a second pass through the sections and locate and perform any
10960 	 * relocations that are present.  We do this after the first pass to
10961 	 * be sure that all sections have had their headers validated.
10962 	 */
10963 	for (i = 0; i < dof->dofh_secnum; i++) {
10964 		dof_sec_t *sec = (dof_sec_t *)(daddr +
10965 		    (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
10966 
10967 		if (!(sec->dofs_flags & DOF_SECF_LOAD))
10968 			continue; /* skip sections that are not loadable */
10969 
10970 		switch (sec->dofs_type) {
10971 		case DOF_SECT_URELHDR:
10972 			if (dtrace_dof_relocate(dof, sec, ubase) != 0)
10973 				return (-1);
10974 			break;
10975 		}
10976 	}
10977 
10978 	if ((enab = *enabp) == NULL)
10979 		enab = *enabp = dtrace_enabling_create(vstate);
10980 
10981 	for (i = 0; i < dof->dofh_secnum; i++) {
10982 		dof_sec_t *sec = (dof_sec_t *)(daddr +
10983 		    (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
10984 
10985 		if (sec->dofs_type != DOF_SECT_ECBDESC)
10986 			continue;
10987 
10988 		if ((ep = dtrace_dof_ecbdesc(dof, sec, vstate, cr)) == NULL) {
10989 			dtrace_enabling_destroy(enab);
10990 			*enabp = NULL;
10991 			return (-1);
10992 		}
10993 
10994 		dtrace_enabling_add(enab, ep);
10995 	}
10996 
10997 	return (0);
10998 }
10999 
11000 /*
11001  * Process DOF for any options.  This routine assumes that the DOF has been
11002  * at least processed by dtrace_dof_slurp().
11003  */
11004 static int
11005 dtrace_dof_options(dof_hdr_t *dof, dtrace_state_t *state)
11006 {
11007 	int i, rval;
11008 	uint32_t entsize;
11009 	size_t offs;
11010 	dof_optdesc_t *desc;
11011 
11012 	for (i = 0; i < dof->dofh_secnum; i++) {
11013 		dof_sec_t *sec = (dof_sec_t *)((uintptr_t)dof +
11014 		    (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
11015 
11016 		if (sec->dofs_type != DOF_SECT_OPTDESC)
11017 			continue;
11018 
11019 		if (sec->dofs_align != sizeof (uint64_t)) {
11020 			dtrace_dof_error(dof, "bad alignment in "
11021 			    "option description");
11022 			return (EINVAL);
11023 		}
11024 
11025 		if ((entsize = sec->dofs_entsize) == 0) {
11026 			dtrace_dof_error(dof, "zeroed option entry size");
11027 			return (EINVAL);
11028 		}
11029 
11030 		if (entsize < sizeof (dof_optdesc_t)) {
11031 			dtrace_dof_error(dof, "bad option entry size");
11032 			return (EINVAL);
11033 		}
11034 
11035 		for (offs = 0; offs < sec->dofs_size; offs += entsize) {
11036 			desc = (dof_optdesc_t *)((uintptr_t)dof +
11037 			    (uintptr_t)sec->dofs_offset + offs);
11038 
11039 			if (desc->dofo_strtab != DOF_SECIDX_NONE) {
11040 				dtrace_dof_error(dof, "non-zero option string");
11041 				return (EINVAL);
11042 			}
11043 
11044 			if (desc->dofo_value == DTRACEOPT_UNSET) {
11045 				dtrace_dof_error(dof, "unset option");
11046 				return (EINVAL);
11047 			}
11048 
11049 			if ((rval = dtrace_state_option(state,
11050 			    desc->dofo_option, desc->dofo_value)) != 0) {
11051 				dtrace_dof_error(dof, "rejected option");
11052 				return (rval);
11053 			}
11054 		}
11055 	}
11056 
11057 	return (0);
11058 }
11059 
11060 /*
11061  * DTrace Consumer State Functions
11062  */
11063 int
11064 dtrace_dstate_init(dtrace_dstate_t *dstate, size_t size)
11065 {
11066 	size_t hashsize, maxper, min, chunksize = dstate->dtds_chunksize;
11067 	void *base;
11068 	uintptr_t limit;
11069 	dtrace_dynvar_t *dvar, *next, *start;
11070 	int i;
11071 
11072 	ASSERT(MUTEX_HELD(&dtrace_lock));
11073 	ASSERT(dstate->dtds_base == NULL && dstate->dtds_percpu == NULL);
11074 
11075 	bzero(dstate, sizeof (dtrace_dstate_t));
11076 
11077 	if ((dstate->dtds_chunksize = chunksize) == 0)
11078 		dstate->dtds_chunksize = DTRACE_DYNVAR_CHUNKSIZE;
11079 
11080 	if (size < (min = dstate->dtds_chunksize + sizeof (dtrace_dynhash_t)))
11081 		size = min;
11082 
11083 	if ((base = kmem_zalloc(size, KM_NOSLEEP)) == NULL)
11084 		return (ENOMEM);
11085 
11086 	dstate->dtds_size = size;
11087 	dstate->dtds_base = base;
11088 	dstate->dtds_percpu = kmem_cache_alloc(dtrace_state_cache, KM_SLEEP);
11089 	bzero(dstate->dtds_percpu, NCPU * sizeof (dtrace_dstate_percpu_t));
11090 
11091 	hashsize = size / (dstate->dtds_chunksize + sizeof (dtrace_dynhash_t));
11092 
11093 	if (hashsize != 1 && (hashsize & 1))
11094 		hashsize--;
11095 
11096 	dstate->dtds_hashsize = hashsize;
11097 	dstate->dtds_hash = dstate->dtds_base;
11098 
11099 	/*
11100 	 * Set all of our hash buckets to point to the single sink, and (if
11101 	 * it hasn't already been set), set the sink's hash value to be the
11102 	 * sink sentinel value.  The sink is needed for dynamic variable
11103 	 * lookups to know that they have iterated over an entire, valid hash
11104 	 * chain.
11105 	 */
11106 	for (i = 0; i < hashsize; i++)
11107 		dstate->dtds_hash[i].dtdh_chain = &dtrace_dynhash_sink;
11108 
11109 	if (dtrace_dynhash_sink.dtdv_hashval != DTRACE_DYNHASH_SINK)
11110 		dtrace_dynhash_sink.dtdv_hashval = DTRACE_DYNHASH_SINK;
11111 
11112 	/*
11113 	 * Determine number of active CPUs.  Divide free list evenly among
11114 	 * active CPUs.
11115 	 */
11116 	start = (dtrace_dynvar_t *)
11117 	    ((uintptr_t)base + hashsize * sizeof (dtrace_dynhash_t));
11118 	limit = (uintptr_t)base + size;
11119 
11120 	maxper = (limit - (uintptr_t)start) / NCPU;
11121 	maxper = (maxper / dstate->dtds_chunksize) * dstate->dtds_chunksize;
11122 
11123 	for (i = 0; i < NCPU; i++) {
11124 		dstate->dtds_percpu[i].dtdsc_free = dvar = start;
11125 
11126 		/*
11127 		 * If we don't even have enough chunks to make it once through
11128 		 * NCPUs, we're just going to allocate everything to the first
11129 		 * CPU.  And if we're on the last CPU, we're going to allocate
11130 		 * whatever is left over.  In either case, we set the limit to
11131 		 * be the limit of the dynamic variable space.
11132 		 */
11133 		if (maxper == 0 || i == NCPU - 1) {
11134 			limit = (uintptr_t)base + size;
11135 			start = NULL;
11136 		} else {
11137 			limit = (uintptr_t)start + maxper;
11138 			start = (dtrace_dynvar_t *)limit;
11139 		}
11140 
11141 		ASSERT(limit <= (uintptr_t)base + size);
11142 
11143 		for (;;) {
11144 			next = (dtrace_dynvar_t *)((uintptr_t)dvar +
11145 			    dstate->dtds_chunksize);
11146 
11147 			if ((uintptr_t)next + dstate->dtds_chunksize >= limit)
11148 				break;
11149 
11150 			dvar->dtdv_next = next;
11151 			dvar = next;
11152 		}
11153 
11154 		if (maxper == 0)
11155 			break;
11156 	}
11157 
11158 	return (0);
11159 }
11160 
11161 void
11162 dtrace_dstate_fini(dtrace_dstate_t *dstate)
11163 {
11164 	ASSERT(MUTEX_HELD(&cpu_lock));
11165 
11166 	if (dstate->dtds_base == NULL)
11167 		return;
11168 
11169 	kmem_free(dstate->dtds_base, dstate->dtds_size);
11170 	kmem_cache_free(dtrace_state_cache, dstate->dtds_percpu);
11171 }
11172 
11173 static void
11174 dtrace_vstate_fini(dtrace_vstate_t *vstate)
11175 {
11176 	/*
11177 	 * Logical XOR, where are you?
11178 	 */
11179 	ASSERT((vstate->dtvs_nglobals == 0) ^ (vstate->dtvs_globals != NULL));
11180 
11181 	if (vstate->dtvs_nglobals > 0) {
11182 		kmem_free(vstate->dtvs_globals, vstate->dtvs_nglobals *
11183 		    sizeof (dtrace_statvar_t *));
11184 	}
11185 
11186 	if (vstate->dtvs_ntlocals > 0) {
11187 		kmem_free(vstate->dtvs_tlocals, vstate->dtvs_ntlocals *
11188 		    sizeof (dtrace_difv_t));
11189 	}
11190 
11191 	ASSERT((vstate->dtvs_nlocals == 0) ^ (vstate->dtvs_locals != NULL));
11192 
11193 	if (vstate->dtvs_nlocals > 0) {
11194 		kmem_free(vstate->dtvs_locals, vstate->dtvs_nlocals *
11195 		    sizeof (dtrace_statvar_t *));
11196 	}
11197 }
11198 
11199 static void
11200 dtrace_state_clean(dtrace_state_t *state)
11201 {
11202 	if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE)
11203 		return;
11204 
11205 	dtrace_dynvar_clean(&state->dts_vstate.dtvs_dynvars);
11206 	dtrace_speculation_clean(state);
11207 }
11208 
11209 static void
11210 dtrace_state_deadman(dtrace_state_t *state)
11211 {
11212 	hrtime_t now;
11213 
11214 	dtrace_sync();
11215 
11216 	now = dtrace_gethrtime();
11217 
11218 	if (state != dtrace_anon.dta_state &&
11219 	    now - state->dts_laststatus >= dtrace_deadman_user)
11220 		return;
11221 
11222 	/*
11223 	 * We must be sure that dts_alive never appears to be less than the
11224 	 * value upon entry to dtrace_state_deadman(), and because we lack a
11225 	 * dtrace_cas64(), we cannot store to it atomically.  We thus instead
11226 	 * store INT64_MAX to it, followed by a memory barrier, followed by
11227 	 * the new value.  This assures that dts_alive never appears to be
11228 	 * less than its true value, regardless of the order in which the
11229 	 * stores to the underlying storage are issued.
11230 	 */
11231 	state->dts_alive = INT64_MAX;
11232 	dtrace_membar_producer();
11233 	state->dts_alive = now;
11234 }
11235 
11236 dtrace_state_t *
11237 dtrace_state_create(dev_t *devp, cred_t *cr)
11238 {
11239 	minor_t minor;
11240 	major_t major;
11241 	char c[30];
11242 	dtrace_state_t *state;
11243 	dtrace_optval_t *opt;
11244 	int bufsize = NCPU * sizeof (dtrace_buffer_t), i;
11245 
11246 	ASSERT(MUTEX_HELD(&dtrace_lock));
11247 	ASSERT(MUTEX_HELD(&cpu_lock));
11248 
11249 	minor = (minor_t)(uintptr_t)vmem_alloc(dtrace_minor, 1,
11250 	    VM_BESTFIT | VM_SLEEP);
11251 
11252 	if (ddi_soft_state_zalloc(dtrace_softstate, minor) != DDI_SUCCESS) {
11253 		vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
11254 		return (NULL);
11255 	}
11256 
11257 	state = ddi_get_soft_state(dtrace_softstate, minor);
11258 	state->dts_epid = DTRACE_EPIDNONE + 1;
11259 
11260 	(void) snprintf(c, sizeof (c), "dtrace_aggid_%d", minor);
11261 	state->dts_aggid_arena = vmem_create(c, (void *)1, UINT32_MAX, 1,
11262 	    NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
11263 
11264 	if (devp != NULL) {
11265 		major = getemajor(*devp);
11266 	} else {
11267 		major = ddi_driver_major(dtrace_devi);
11268 	}
11269 
11270 	state->dts_dev = makedevice(major, minor);
11271 
11272 	if (devp != NULL)
11273 		*devp = state->dts_dev;
11274 
11275 	/*
11276 	 * We allocate NCPU buffers.  On the one hand, this can be quite
11277 	 * a bit of memory per instance (nearly 36K on a Starcat).  On the
11278 	 * other hand, it saves an additional memory reference in the probe
11279 	 * path.
11280 	 */
11281 	state->dts_buffer = kmem_zalloc(bufsize, KM_SLEEP);
11282 	state->dts_aggbuffer = kmem_zalloc(bufsize, KM_SLEEP);
11283 	state->dts_cleaner = CYCLIC_NONE;
11284 	state->dts_deadman = CYCLIC_NONE;
11285 	state->dts_vstate.dtvs_state = state;
11286 
11287 	for (i = 0; i < DTRACEOPT_MAX; i++)
11288 		state->dts_options[i] = DTRACEOPT_UNSET;
11289 
11290 	/*
11291 	 * Set the default options.
11292 	 */
11293 	opt = state->dts_options;
11294 	opt[DTRACEOPT_BUFPOLICY] = DTRACEOPT_BUFPOLICY_SWITCH;
11295 	opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_AUTO;
11296 	opt[DTRACEOPT_NSPEC] = dtrace_nspec_default;
11297 	opt[DTRACEOPT_SPECSIZE] = dtrace_specsize_default;
11298 	opt[DTRACEOPT_CPU] = (dtrace_optval_t)DTRACE_CPUALL;
11299 	opt[DTRACEOPT_STRSIZE] = dtrace_strsize_default;
11300 	opt[DTRACEOPT_STACKFRAMES] = dtrace_stackframes_default;
11301 	opt[DTRACEOPT_USTACKFRAMES] = dtrace_ustackframes_default;
11302 	opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_default;
11303 	opt[DTRACEOPT_AGGRATE] = dtrace_aggrate_default;
11304 	opt[DTRACEOPT_SWITCHRATE] = dtrace_switchrate_default;
11305 	opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_default;
11306 	opt[DTRACEOPT_JSTACKFRAMES] = dtrace_jstackframes_default;
11307 	opt[DTRACEOPT_JSTACKSTRSIZE] = dtrace_jstackstrsize_default;
11308 
11309 	state->dts_activity = DTRACE_ACTIVITY_INACTIVE;
11310 
11311 	/*
11312 	 * Depending on the user credentials, we set flag bits which alter probe
11313 	 * visibility or the amount of destructiveness allowed.  In the case of
11314 	 * actual anonymous tracing, or the possession of all privileges, all of
11315 	 * the normal checks are bypassed.
11316 	 */
11317 	if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
11318 		state->dts_cred.dcr_visible = DTRACE_CRV_ALL;
11319 		state->dts_cred.dcr_action = DTRACE_CRA_ALL;
11320 	} else {
11321 		/*
11322 		 * Set up the credentials for this instantiation.  We take a
11323 		 * hold on the credential to prevent it from disappearing on
11324 		 * us; this in turn prevents the zone_t referenced by this
11325 		 * credential from disappearing.  This means that we can
11326 		 * examine the credential and the zone from probe context.
11327 		 */
11328 		crhold(cr);
11329 		state->dts_cred.dcr_cred = cr;
11330 
11331 		/*
11332 		 * CRA_PROC means "we have *some* privilege for dtrace" and
11333 		 * unlocks the use of variables like pid, zonename, etc.
11334 		 */
11335 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE) ||
11336 		    PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
11337 			state->dts_cred.dcr_action |= DTRACE_CRA_PROC;
11338 		}
11339 
11340 		/*
11341 		 * dtrace_user allows use of syscall and profile providers.
11342 		 * If the user also has proc_owner and/or proc_zone, we
11343 		 * extend the scope to include additional visibility and
11344 		 * destructive power.
11345 		 */
11346 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE)) {
11347 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE)) {
11348 				state->dts_cred.dcr_visible |=
11349 				    DTRACE_CRV_ALLPROC;
11350 
11351 				state->dts_cred.dcr_action |=
11352 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
11353 			}
11354 
11355 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE)) {
11356 				state->dts_cred.dcr_visible |=
11357 				    DTRACE_CRV_ALLZONE;
11358 
11359 				state->dts_cred.dcr_action |=
11360 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
11361 			}
11362 
11363 			/*
11364 			 * If we have all privs in whatever zone this is,
11365 			 * we can do destructive things to processes which
11366 			 * have altered credentials.
11367 			 */
11368 			if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
11369 			    cr->cr_zone->zone_privset)) {
11370 				state->dts_cred.dcr_action |=
11371 				    DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
11372 			}
11373 		}
11374 
11375 		/*
11376 		 * Holding the dtrace_kernel privilege also implies that
11377 		 * the user has the dtrace_user privilege from a visibility
11378 		 * perspective.  But without further privileges, some
11379 		 * destructive actions are not available.
11380 		 */
11381 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE)) {
11382 			/*
11383 			 * Make all probes in all zones visible.  However,
11384 			 * this doesn't mean that all actions become available
11385 			 * to all zones.
11386 			 */
11387 			state->dts_cred.dcr_visible |= DTRACE_CRV_KERNEL |
11388 			    DTRACE_CRV_ALLPROC | DTRACE_CRV_ALLZONE;
11389 
11390 			state->dts_cred.dcr_action |= DTRACE_CRA_KERNEL |
11391 			    DTRACE_CRA_PROC;
11392 			/*
11393 			 * Holding proc_owner means that destructive actions
11394 			 * for *this* zone are allowed.
11395 			 */
11396 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
11397 				state->dts_cred.dcr_action |=
11398 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
11399 
11400 			/*
11401 			 * Holding proc_zone means that destructive actions
11402 			 * for this user/group ID in all zones is allowed.
11403 			 */
11404 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
11405 				state->dts_cred.dcr_action |=
11406 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
11407 
11408 			/*
11409 			 * If we have all privs in whatever zone this is,
11410 			 * we can do destructive things to processes which
11411 			 * have altered credentials.
11412 			 */
11413 			if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
11414 			    cr->cr_zone->zone_privset)) {
11415 				state->dts_cred.dcr_action |=
11416 				    DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
11417 			}
11418 		}
11419 
11420 		/*
11421 		 * Holding the dtrace_proc privilege gives control over fasttrap
11422 		 * and pid providers.  We need to grant wider destructive
11423 		 * privileges in the event that the user has proc_owner and/or
11424 		 * proc_zone.
11425 		 */
11426 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
11427 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
11428 				state->dts_cred.dcr_action |=
11429 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
11430 
11431 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
11432 				state->dts_cred.dcr_action |=
11433 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
11434 		}
11435 	}
11436 
11437 	return (state);
11438 }
11439 
11440 static int
11441 dtrace_state_buffer(dtrace_state_t *state, dtrace_buffer_t *buf, int which)
11442 {
11443 	dtrace_optval_t *opt = state->dts_options, size;
11444 	processorid_t cpu;
11445 	int flags = 0, rval;
11446 
11447 	ASSERT(MUTEX_HELD(&dtrace_lock));
11448 	ASSERT(MUTEX_HELD(&cpu_lock));
11449 	ASSERT(which < DTRACEOPT_MAX);
11450 	ASSERT(state->dts_activity == DTRACE_ACTIVITY_INACTIVE ||
11451 	    (state == dtrace_anon.dta_state &&
11452 	    state->dts_activity == DTRACE_ACTIVITY_ACTIVE));
11453 
11454 	if (opt[which] == DTRACEOPT_UNSET || opt[which] == 0)
11455 		return (0);
11456 
11457 	if (opt[DTRACEOPT_CPU] != DTRACEOPT_UNSET)
11458 		cpu = opt[DTRACEOPT_CPU];
11459 
11460 	if (which == DTRACEOPT_SPECSIZE)
11461 		flags |= DTRACEBUF_NOSWITCH;
11462 
11463 	if (which == DTRACEOPT_BUFSIZE) {
11464 		if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_RING)
11465 			flags |= DTRACEBUF_RING;
11466 
11467 		if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_FILL)
11468 			flags |= DTRACEBUF_FILL;
11469 
11470 		if (state != dtrace_anon.dta_state ||
11471 		    state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
11472 			flags |= DTRACEBUF_INACTIVE;
11473 	}
11474 
11475 	for (size = opt[which]; size >= sizeof (uint64_t); size >>= 1) {
11476 		/*
11477 		 * The size must be 8-byte aligned.  If the size is not 8-byte
11478 		 * aligned, drop it down by the difference.
11479 		 */
11480 		if (size & (sizeof (uint64_t) - 1))
11481 			size -= size & (sizeof (uint64_t) - 1);
11482 
11483 		if (size < state->dts_reserve) {
11484 			/*
11485 			 * Buffers always must be large enough to accommodate
11486 			 * their prereserved space.  We return E2BIG instead
11487 			 * of ENOMEM in this case to allow for user-level
11488 			 * software to differentiate the cases.
11489 			 */
11490 			return (E2BIG);
11491 		}
11492 
11493 		rval = dtrace_buffer_alloc(buf, size, flags, cpu);
11494 
11495 		if (rval != ENOMEM) {
11496 			opt[which] = size;
11497 			return (rval);
11498 		}
11499 
11500 		if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
11501 			return (rval);
11502 	}
11503 
11504 	return (ENOMEM);
11505 }
11506 
11507 static int
11508 dtrace_state_buffers(dtrace_state_t *state)
11509 {
11510 	dtrace_speculation_t *spec = state->dts_speculations;
11511 	int rval, i;
11512 
11513 	if ((rval = dtrace_state_buffer(state, state->dts_buffer,
11514 	    DTRACEOPT_BUFSIZE)) != 0)
11515 		return (rval);
11516 
11517 	if ((rval = dtrace_state_buffer(state, state->dts_aggbuffer,
11518 	    DTRACEOPT_AGGSIZE)) != 0)
11519 		return (rval);
11520 
11521 	for (i = 0; i < state->dts_nspeculations; i++) {
11522 		if ((rval = dtrace_state_buffer(state,
11523 		    spec[i].dtsp_buffer, DTRACEOPT_SPECSIZE)) != 0)
11524 			return (rval);
11525 	}
11526 
11527 	return (0);
11528 }
11529 
11530 static void
11531 dtrace_state_prereserve(dtrace_state_t *state)
11532 {
11533 	dtrace_ecb_t *ecb;
11534 	dtrace_probe_t *probe;
11535 
11536 	state->dts_reserve = 0;
11537 
11538 	if (state->dts_options[DTRACEOPT_BUFPOLICY] != DTRACEOPT_BUFPOLICY_FILL)
11539 		return;
11540 
11541 	/*
11542 	 * If our buffer policy is a "fill" buffer policy, we need to set the
11543 	 * prereserved space to be the space required by the END probes.
11544 	 */
11545 	probe = dtrace_probes[dtrace_probeid_end - 1];
11546 	ASSERT(probe != NULL);
11547 
11548 	for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
11549 		if (ecb->dte_state != state)
11550 			continue;
11551 
11552 		state->dts_reserve += ecb->dte_needed + ecb->dte_alignment;
11553 	}
11554 }
11555 
11556 static int
11557 dtrace_state_go(dtrace_state_t *state, processorid_t *cpu)
11558 {
11559 	dtrace_optval_t *opt = state->dts_options, sz, nspec;
11560 	dtrace_speculation_t *spec;
11561 	dtrace_buffer_t *buf;
11562 	cyc_handler_t hdlr;
11563 	cyc_time_t when;
11564 	int rval = 0, i, bufsize = NCPU * sizeof (dtrace_buffer_t);
11565 	dtrace_icookie_t cookie;
11566 
11567 	mutex_enter(&cpu_lock);
11568 	mutex_enter(&dtrace_lock);
11569 
11570 	if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
11571 		rval = EBUSY;
11572 		goto out;
11573 	}
11574 
11575 	/*
11576 	 * Before we can perform any checks, we must prime all of the
11577 	 * retained enablings that correspond to this state.
11578 	 */
11579 	dtrace_enabling_prime(state);
11580 
11581 	if (state->dts_destructive && !state->dts_cred.dcr_destructive) {
11582 		rval = EACCES;
11583 		goto out;
11584 	}
11585 
11586 	dtrace_state_prereserve(state);
11587 
11588 	/*
11589 	 * Now we want to do is try to allocate our speculations.
11590 	 * We do not automatically resize the number of speculations; if
11591 	 * this fails, we will fail the operation.
11592 	 */
11593 	nspec = opt[DTRACEOPT_NSPEC];
11594 	ASSERT(nspec != DTRACEOPT_UNSET);
11595 
11596 	if (nspec > INT_MAX) {
11597 		rval = ENOMEM;
11598 		goto out;
11599 	}
11600 
11601 	spec = kmem_zalloc(nspec * sizeof (dtrace_speculation_t), KM_NOSLEEP);
11602 
11603 	if (spec == NULL) {
11604 		rval = ENOMEM;
11605 		goto out;
11606 	}
11607 
11608 	state->dts_speculations = spec;
11609 	state->dts_nspeculations = (int)nspec;
11610 
11611 	for (i = 0; i < nspec; i++) {
11612 		if ((buf = kmem_zalloc(bufsize, KM_NOSLEEP)) == NULL) {
11613 			rval = ENOMEM;
11614 			goto err;
11615 		}
11616 
11617 		spec[i].dtsp_buffer = buf;
11618 	}
11619 
11620 	if (opt[DTRACEOPT_GRABANON] != DTRACEOPT_UNSET) {
11621 		if (dtrace_anon.dta_state == NULL) {
11622 			rval = ENOENT;
11623 			goto out;
11624 		}
11625 
11626 		if (state->dts_necbs != 0) {
11627 			rval = EALREADY;
11628 			goto out;
11629 		}
11630 
11631 		state->dts_anon = dtrace_anon_grab();
11632 		ASSERT(state->dts_anon != NULL);
11633 		state = state->dts_anon;
11634 
11635 		/*
11636 		 * We want "grabanon" to be set in the grabbed state, so we'll
11637 		 * copy that option value from the grabbing state into the
11638 		 * grabbed state.
11639 		 */
11640 		state->dts_options[DTRACEOPT_GRABANON] =
11641 		    opt[DTRACEOPT_GRABANON];
11642 
11643 		*cpu = dtrace_anon.dta_beganon;
11644 
11645 		/*
11646 		 * If the anonymous state is active (as it almost certainly
11647 		 * is if the anonymous enabling ultimately matched anything),
11648 		 * we don't allow any further option processing -- but we
11649 		 * don't return failure.
11650 		 */
11651 		if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
11652 			goto out;
11653 	}
11654 
11655 	if (opt[DTRACEOPT_AGGSIZE] != DTRACEOPT_UNSET &&
11656 	    opt[DTRACEOPT_AGGSIZE] != 0) {
11657 		if (state->dts_aggregations == NULL) {
11658 			/*
11659 			 * We're not going to create an aggregation buffer
11660 			 * because we don't have any ECBs that contain
11661 			 * aggregations -- set this option to 0.
11662 			 */
11663 			opt[DTRACEOPT_AGGSIZE] = 0;
11664 		} else {
11665 			/*
11666 			 * If we have an aggregation buffer, we must also have
11667 			 * a buffer to use as scratch.
11668 			 */
11669 			if (opt[DTRACEOPT_BUFSIZE] == DTRACEOPT_UNSET ||
11670 			    opt[DTRACEOPT_BUFSIZE] < state->dts_needed) {
11671 				opt[DTRACEOPT_BUFSIZE] = state->dts_needed;
11672 			}
11673 		}
11674 	}
11675 
11676 	if (opt[DTRACEOPT_SPECSIZE] != DTRACEOPT_UNSET &&
11677 	    opt[DTRACEOPT_SPECSIZE] != 0) {
11678 		if (!state->dts_speculates) {
11679 			/*
11680 			 * We're not going to create speculation buffers
11681 			 * because we don't have any ECBs that actually
11682 			 * speculate -- set the speculation size to 0.
11683 			 */
11684 			opt[DTRACEOPT_SPECSIZE] = 0;
11685 		}
11686 	}
11687 
11688 	/*
11689 	 * The bare minimum size for any buffer that we're actually going to
11690 	 * do anything to is sizeof (uint64_t).
11691 	 */
11692 	sz = sizeof (uint64_t);
11693 
11694 	if ((state->dts_needed != 0 && opt[DTRACEOPT_BUFSIZE] < sz) ||
11695 	    (state->dts_speculates && opt[DTRACEOPT_SPECSIZE] < sz) ||
11696 	    (state->dts_aggregations != NULL && opt[DTRACEOPT_AGGSIZE] < sz)) {
11697 		/*
11698 		 * A buffer size has been explicitly set to 0 (or to a size
11699 		 * that will be adjusted to 0) and we need the space -- we
11700 		 * need to return failure.  We return ENOSPC to differentiate
11701 		 * it from failing to allocate a buffer due to failure to meet
11702 		 * the reserve (for which we return E2BIG).
11703 		 */
11704 		rval = ENOSPC;
11705 		goto out;
11706 	}
11707 
11708 	if ((rval = dtrace_state_buffers(state)) != 0)
11709 		goto err;
11710 
11711 	if ((sz = opt[DTRACEOPT_DYNVARSIZE]) == DTRACEOPT_UNSET)
11712 		sz = dtrace_dstate_defsize;
11713 
11714 	do {
11715 		rval = dtrace_dstate_init(&state->dts_vstate.dtvs_dynvars, sz);
11716 
11717 		if (rval == 0)
11718 			break;
11719 
11720 		if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
11721 			goto err;
11722 	} while (sz >>= 1);
11723 
11724 	opt[DTRACEOPT_DYNVARSIZE] = sz;
11725 
11726 	if (rval != 0)
11727 		goto err;
11728 
11729 	if (opt[DTRACEOPT_STATUSRATE] > dtrace_statusrate_max)
11730 		opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_max;
11731 
11732 	if (opt[DTRACEOPT_CLEANRATE] == 0)
11733 		opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
11734 
11735 	if (opt[DTRACEOPT_CLEANRATE] < dtrace_cleanrate_min)
11736 		opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_min;
11737 
11738 	if (opt[DTRACEOPT_CLEANRATE] > dtrace_cleanrate_max)
11739 		opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
11740 
11741 	hdlr.cyh_func = (cyc_func_t)dtrace_state_clean;
11742 	hdlr.cyh_arg = state;
11743 	hdlr.cyh_level = CY_LOW_LEVEL;
11744 
11745 	when.cyt_when = 0;
11746 	when.cyt_interval = opt[DTRACEOPT_CLEANRATE];
11747 
11748 	state->dts_cleaner = cyclic_add(&hdlr, &when);
11749 
11750 	hdlr.cyh_func = (cyc_func_t)dtrace_state_deadman;
11751 	hdlr.cyh_arg = state;
11752 	hdlr.cyh_level = CY_LOW_LEVEL;
11753 
11754 	when.cyt_when = 0;
11755 	when.cyt_interval = dtrace_deadman_interval;
11756 
11757 	state->dts_alive = state->dts_laststatus = dtrace_gethrtime();
11758 	state->dts_deadman = cyclic_add(&hdlr, &when);
11759 
11760 	state->dts_activity = DTRACE_ACTIVITY_WARMUP;
11761 
11762 	/*
11763 	 * Now it's time to actually fire the BEGIN probe.  We need to disable
11764 	 * interrupts here both to record the CPU on which we fired the BEGIN
11765 	 * probe (the data from this CPU will be processed first at user
11766 	 * level) and to manually activate the buffer for this CPU.
11767 	 */
11768 	cookie = dtrace_interrupt_disable();
11769 	*cpu = CPU->cpu_id;
11770 	ASSERT(state->dts_buffer[*cpu].dtb_flags & DTRACEBUF_INACTIVE);
11771 	state->dts_buffer[*cpu].dtb_flags &= ~DTRACEBUF_INACTIVE;
11772 
11773 	dtrace_probe(dtrace_probeid_begin,
11774 	    (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
11775 	dtrace_interrupt_enable(cookie);
11776 	/*
11777 	 * We may have had an exit action from a BEGIN probe; only change our
11778 	 * state to ACTIVE if we're still in WARMUP.
11779 	 */
11780 	ASSERT(state->dts_activity == DTRACE_ACTIVITY_WARMUP ||
11781 	    state->dts_activity == DTRACE_ACTIVITY_DRAINING);
11782 
11783 	if (state->dts_activity == DTRACE_ACTIVITY_WARMUP)
11784 		state->dts_activity = DTRACE_ACTIVITY_ACTIVE;
11785 
11786 	/*
11787 	 * Regardless of whether or not now we're in ACTIVE or DRAINING, we
11788 	 * want each CPU to transition its principal buffer out of the
11789 	 * INACTIVE state.  Doing this assures that no CPU will suddenly begin
11790 	 * processing an ECB halfway down a probe's ECB chain; all CPUs will
11791 	 * atomically transition from processing none of a state's ECBs to
11792 	 * processing all of them.
11793 	 */
11794 	dtrace_xcall(DTRACE_CPUALL,
11795 	    (dtrace_xcall_t)dtrace_buffer_activate, state);
11796 	goto out;
11797 
11798 err:
11799 	dtrace_buffer_free(state->dts_buffer);
11800 	dtrace_buffer_free(state->dts_aggbuffer);
11801 
11802 	if ((nspec = state->dts_nspeculations) == 0) {
11803 		ASSERT(state->dts_speculations == NULL);
11804 		goto out;
11805 	}
11806 
11807 	spec = state->dts_speculations;
11808 	ASSERT(spec != NULL);
11809 
11810 	for (i = 0; i < state->dts_nspeculations; i++) {
11811 		if ((buf = spec[i].dtsp_buffer) == NULL)
11812 			break;
11813 
11814 		dtrace_buffer_free(buf);
11815 		kmem_free(buf, bufsize);
11816 	}
11817 
11818 	kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
11819 	state->dts_nspeculations = 0;
11820 	state->dts_speculations = NULL;
11821 
11822 out:
11823 	mutex_exit(&dtrace_lock);
11824 	mutex_exit(&cpu_lock);
11825 
11826 	return (rval);
11827 }
11828 
11829 static int
11830 dtrace_state_stop(dtrace_state_t *state, processorid_t *cpu)
11831 {
11832 	dtrace_icookie_t cookie;
11833 
11834 	ASSERT(MUTEX_HELD(&dtrace_lock));
11835 
11836 	if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE &&
11837 	    state->dts_activity != DTRACE_ACTIVITY_DRAINING)
11838 		return (EINVAL);
11839 
11840 	/*
11841 	 * We'll set the activity to DTRACE_ACTIVITY_DRAINING, and issue a sync
11842 	 * to be sure that every CPU has seen it.  See below for the details
11843 	 * on why this is done.
11844 	 */
11845 	state->dts_activity = DTRACE_ACTIVITY_DRAINING;
11846 	dtrace_sync();
11847 
11848 	/*
11849 	 * By this point, it is impossible for any CPU to be still processing
11850 	 * with DTRACE_ACTIVITY_ACTIVE.  We can thus set our activity to
11851 	 * DTRACE_ACTIVITY_COOLDOWN and know that we're not racing with any
11852 	 * other CPU in dtrace_buffer_reserve().  This allows dtrace_probe()
11853 	 * and callees to know that the activity is DTRACE_ACTIVITY_COOLDOWN
11854 	 * iff we're in the END probe.
11855 	 */
11856 	state->dts_activity = DTRACE_ACTIVITY_COOLDOWN;
11857 	dtrace_sync();
11858 	ASSERT(state->dts_activity == DTRACE_ACTIVITY_COOLDOWN);
11859 
11860 	/*
11861 	 * Finally, we can release the reserve and call the END probe.  We
11862 	 * disable interrupts across calling the END probe to allow us to
11863 	 * return the CPU on which we actually called the END probe.  This
11864 	 * allows user-land to be sure that this CPU's principal buffer is
11865 	 * processed last.
11866 	 */
11867 	state->dts_reserve = 0;
11868 
11869 	cookie = dtrace_interrupt_disable();
11870 	*cpu = CPU->cpu_id;
11871 	dtrace_probe(dtrace_probeid_end,
11872 	    (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
11873 	dtrace_interrupt_enable(cookie);
11874 
11875 	state->dts_activity = DTRACE_ACTIVITY_STOPPED;
11876 	dtrace_sync();
11877 
11878 	return (0);
11879 }
11880 
11881 static int
11882 dtrace_state_option(dtrace_state_t *state, dtrace_optid_t option,
11883     dtrace_optval_t val)
11884 {
11885 	ASSERT(MUTEX_HELD(&dtrace_lock));
11886 
11887 	if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
11888 		return (EBUSY);
11889 
11890 	if (option >= DTRACEOPT_MAX)
11891 		return (EINVAL);
11892 
11893 	if (option != DTRACEOPT_CPU && val < 0)
11894 		return (EINVAL);
11895 
11896 	switch (option) {
11897 	case DTRACEOPT_DESTRUCTIVE:
11898 		if (dtrace_destructive_disallow)
11899 			return (EACCES);
11900 
11901 		state->dts_cred.dcr_destructive = 1;
11902 		break;
11903 
11904 	case DTRACEOPT_BUFSIZE:
11905 	case DTRACEOPT_DYNVARSIZE:
11906 	case DTRACEOPT_AGGSIZE:
11907 	case DTRACEOPT_SPECSIZE:
11908 	case DTRACEOPT_STRSIZE:
11909 		if (val < 0)
11910 			return (EINVAL);
11911 
11912 		if (val >= LONG_MAX) {
11913 			/*
11914 			 * If this is an otherwise negative value, set it to
11915 			 * the highest multiple of 128m less than LONG_MAX.
11916 			 * Technically, we're adjusting the size without
11917 			 * regard to the buffer resizing policy, but in fact,
11918 			 * this has no effect -- if we set the buffer size to
11919 			 * ~LONG_MAX and the buffer policy is ultimately set to
11920 			 * be "manual", the buffer allocation is guaranteed to
11921 			 * fail, if only because the allocation requires two
11922 			 * buffers.  (We set the the size to the highest
11923 			 * multiple of 128m because it ensures that the size
11924 			 * will remain a multiple of a megabyte when
11925 			 * repeatedly halved -- all the way down to 15m.)
11926 			 */
11927 			val = LONG_MAX - (1 << 27) + 1;
11928 		}
11929 	}
11930 
11931 	state->dts_options[option] = val;
11932 
11933 	return (0);
11934 }
11935 
11936 static void
11937 dtrace_state_destroy(dtrace_state_t *state)
11938 {
11939 	dtrace_ecb_t *ecb;
11940 	dtrace_vstate_t *vstate = &state->dts_vstate;
11941 	minor_t minor = getminor(state->dts_dev);
11942 	int i, bufsize = NCPU * sizeof (dtrace_buffer_t);
11943 	dtrace_speculation_t *spec = state->dts_speculations;
11944 	int nspec = state->dts_nspeculations;
11945 	uint32_t match;
11946 
11947 	ASSERT(MUTEX_HELD(&dtrace_lock));
11948 	ASSERT(MUTEX_HELD(&cpu_lock));
11949 
11950 	/*
11951 	 * First, retract any retained enablings for this state.
11952 	 */
11953 	dtrace_enabling_retract(state);
11954 	ASSERT(state->dts_nretained == 0);
11955 
11956 	if (state->dts_activity == DTRACE_ACTIVITY_ACTIVE ||
11957 	    state->dts_activity == DTRACE_ACTIVITY_DRAINING) {
11958 		/*
11959 		 * We have managed to come into dtrace_state_destroy() on a
11960 		 * hot enabling -- almost certainly because of a disorderly
11961 		 * shutdown of a consumer.  (That is, a consumer that is
11962 		 * exiting without having called dtrace_stop().) In this case,
11963 		 * we're going to set our activity to be KILLED, and then
11964 		 * issue a sync to be sure that everyone is out of probe
11965 		 * context before we start blowing away ECBs.
11966 		 */
11967 		state->dts_activity = DTRACE_ACTIVITY_KILLED;
11968 		dtrace_sync();
11969 	}
11970 
11971 	/*
11972 	 * Release the credential hold we took in dtrace_state_create().
11973 	 */
11974 	if (state->dts_cred.dcr_cred != NULL)
11975 		crfree(state->dts_cred.dcr_cred);
11976 
11977 	/*
11978 	 * Now we can safely disable and destroy any enabled probes.  Because
11979 	 * any DTRACE_PRIV_KERNEL probes may actually be slowing our progress
11980 	 * (especially if they're all enabled), we take two passes through the
11981 	 * ECBs:  in the first, we disable just DTRACE_PRIV_KERNEL probes, and
11982 	 * in the second we disable whatever is left over.
11983 	 */
11984 	for (match = DTRACE_PRIV_KERNEL; ; match = 0) {
11985 		for (i = 0; i < state->dts_necbs; i++) {
11986 			if ((ecb = state->dts_ecbs[i]) == NULL)
11987 				continue;
11988 
11989 			if (match && ecb->dte_probe != NULL) {
11990 				dtrace_probe_t *probe = ecb->dte_probe;
11991 				dtrace_provider_t *prov = probe->dtpr_provider;
11992 
11993 				if (!(prov->dtpv_priv.dtpp_flags & match))
11994 					continue;
11995 			}
11996 
11997 			dtrace_ecb_disable(ecb);
11998 			dtrace_ecb_destroy(ecb);
11999 		}
12000 
12001 		if (!match)
12002 			break;
12003 	}
12004 
12005 	/*
12006 	 * Before we free the buffers, perform one more sync to assure that
12007 	 * every CPU is out of probe context.
12008 	 */
12009 	dtrace_sync();
12010 
12011 	dtrace_buffer_free(state->dts_buffer);
12012 	dtrace_buffer_free(state->dts_aggbuffer);
12013 
12014 	for (i = 0; i < nspec; i++)
12015 		dtrace_buffer_free(spec[i].dtsp_buffer);
12016 
12017 	if (state->dts_cleaner != CYCLIC_NONE)
12018 		cyclic_remove(state->dts_cleaner);
12019 
12020 	if (state->dts_deadman != CYCLIC_NONE)
12021 		cyclic_remove(state->dts_deadman);
12022 
12023 	dtrace_dstate_fini(&vstate->dtvs_dynvars);
12024 	dtrace_vstate_fini(vstate);
12025 	kmem_free(state->dts_ecbs, state->dts_necbs * sizeof (dtrace_ecb_t *));
12026 
12027 	if (state->dts_aggregations != NULL) {
12028 #ifdef DEBUG
12029 		for (i = 0; i < state->dts_naggregations; i++)
12030 			ASSERT(state->dts_aggregations[i] == NULL);
12031 #endif
12032 		ASSERT(state->dts_naggregations > 0);
12033 		kmem_free(state->dts_aggregations,
12034 		    state->dts_naggregations * sizeof (dtrace_aggregation_t *));
12035 	}
12036 
12037 	kmem_free(state->dts_buffer, bufsize);
12038 	kmem_free(state->dts_aggbuffer, bufsize);
12039 
12040 	for (i = 0; i < nspec; i++)
12041 		kmem_free(spec[i].dtsp_buffer, bufsize);
12042 
12043 	kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
12044 
12045 	dtrace_format_destroy(state);
12046 
12047 	vmem_destroy(state->dts_aggid_arena);
12048 	ddi_soft_state_free(dtrace_softstate, minor);
12049 	vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
12050 }
12051 
12052 /*
12053  * DTrace Anonymous Enabling Functions
12054  */
12055 static dtrace_state_t *
12056 dtrace_anon_grab(void)
12057 {
12058 	dtrace_state_t *state;
12059 
12060 	ASSERT(MUTEX_HELD(&dtrace_lock));
12061 
12062 	if ((state = dtrace_anon.dta_state) == NULL) {
12063 		ASSERT(dtrace_anon.dta_enabling == NULL);
12064 		return (NULL);
12065 	}
12066 
12067 	ASSERT(dtrace_anon.dta_enabling != NULL);
12068 	ASSERT(dtrace_retained != NULL);
12069 
12070 	dtrace_enabling_destroy(dtrace_anon.dta_enabling);
12071 	dtrace_anon.dta_enabling = NULL;
12072 	dtrace_anon.dta_state = NULL;
12073 
12074 	return (state);
12075 }
12076 
12077 static void
12078 dtrace_anon_property(void)
12079 {
12080 	int i, rv;
12081 	dtrace_state_t *state;
12082 	dof_hdr_t *dof;
12083 	char c[32];		/* enough for "dof-data-" + digits */
12084 
12085 	ASSERT(MUTEX_HELD(&dtrace_lock));
12086 	ASSERT(MUTEX_HELD(&cpu_lock));
12087 
12088 	for (i = 0; ; i++) {
12089 		(void) snprintf(c, sizeof (c), "dof-data-%d", i);
12090 
12091 		dtrace_err_verbose = 1;
12092 
12093 		if ((dof = dtrace_dof_property(c)) == NULL) {
12094 			dtrace_err_verbose = 0;
12095 			break;
12096 		}
12097 
12098 		/*
12099 		 * We want to create anonymous state, so we need to transition
12100 		 * the kernel debugger to indicate that DTrace is active.  If
12101 		 * this fails (e.g. because the debugger has modified text in
12102 		 * some way), we won't continue with the processing.
12103 		 */
12104 		if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
12105 			cmn_err(CE_NOTE, "kernel debugger active; anonymous "
12106 			    "enabling ignored.");
12107 			dtrace_dof_destroy(dof);
12108 			break;
12109 		}
12110 
12111 		/*
12112 		 * If we haven't allocated an anonymous state, we'll do so now.
12113 		 */
12114 		if ((state = dtrace_anon.dta_state) == NULL) {
12115 			state = dtrace_state_create(NULL, NULL);
12116 			dtrace_anon.dta_state = state;
12117 
12118 			if (state == NULL) {
12119 				/*
12120 				 * This basically shouldn't happen:  the only
12121 				 * failure mode from dtrace_state_create() is a
12122 				 * failure of ddi_soft_state_zalloc() that
12123 				 * itself should never happen.  Still, the
12124 				 * interface allows for a failure mode, and
12125 				 * we want to fail as gracefully as possible:
12126 				 * we'll emit an error message and cease
12127 				 * processing anonymous state in this case.
12128 				 */
12129 				cmn_err(CE_WARN, "failed to create "
12130 				    "anonymous state");
12131 				dtrace_dof_destroy(dof);
12132 				break;
12133 			}
12134 		}
12135 
12136 		rv = dtrace_dof_slurp(dof, &state->dts_vstate, CRED(),
12137 		    &dtrace_anon.dta_enabling, 0, B_TRUE);
12138 
12139 		if (rv == 0)
12140 			rv = dtrace_dof_options(dof, state);
12141 
12142 		dtrace_err_verbose = 0;
12143 		dtrace_dof_destroy(dof);
12144 
12145 		if (rv != 0) {
12146 			/*
12147 			 * This is malformed DOF; chuck any anonymous state
12148 			 * that we created.
12149 			 */
12150 			ASSERT(dtrace_anon.dta_enabling == NULL);
12151 			dtrace_state_destroy(state);
12152 			dtrace_anon.dta_state = NULL;
12153 			break;
12154 		}
12155 
12156 		ASSERT(dtrace_anon.dta_enabling != NULL);
12157 	}
12158 
12159 	if (dtrace_anon.dta_enabling != NULL) {
12160 		int rval;
12161 
12162 		/*
12163 		 * dtrace_enabling_retain() can only fail because we are
12164 		 * trying to retain more enablings than are allowed -- but
12165 		 * we only have one anonymous enabling, and we are guaranteed
12166 		 * to be allowed at least one retained enabling; we assert
12167 		 * that dtrace_enabling_retain() returns success.
12168 		 */
12169 		rval = dtrace_enabling_retain(dtrace_anon.dta_enabling);
12170 		ASSERT(rval == 0);
12171 
12172 		dtrace_enabling_dump(dtrace_anon.dta_enabling);
12173 	}
12174 }
12175 
12176 /*
12177  * DTrace Helper Functions
12178  */
12179 static void
12180 dtrace_helper_trace(dtrace_helper_action_t *helper,
12181     dtrace_mstate_t *mstate, dtrace_vstate_t *vstate, int where)
12182 {
12183 	uint32_t size, next, nnext, i;
12184 	dtrace_helptrace_t *ent;
12185 	uint16_t flags = cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
12186 
12187 	if (!dtrace_helptrace_enabled)
12188 		return;
12189 
12190 	ASSERT(vstate->dtvs_nlocals <= dtrace_helptrace_nlocals);
12191 
12192 	/*
12193 	 * What would a tracing framework be without its own tracing
12194 	 * framework?  (Well, a hell of a lot simpler, for starters...)
12195 	 */
12196 	size = sizeof (dtrace_helptrace_t) + dtrace_helptrace_nlocals *
12197 	    sizeof (uint64_t) - sizeof (uint64_t);
12198 
12199 	/*
12200 	 * Iterate until we can allocate a slot in the trace buffer.
12201 	 */
12202 	do {
12203 		next = dtrace_helptrace_next;
12204 
12205 		if (next + size < dtrace_helptrace_bufsize) {
12206 			nnext = next + size;
12207 		} else {
12208 			nnext = size;
12209 		}
12210 	} while (dtrace_cas32(&dtrace_helptrace_next, next, nnext) != next);
12211 
12212 	/*
12213 	 * We have our slot; fill it in.
12214 	 */
12215 	if (nnext == size)
12216 		next = 0;
12217 
12218 	ent = (dtrace_helptrace_t *)&dtrace_helptrace_buffer[next];
12219 	ent->dtht_helper = helper;
12220 	ent->dtht_where = where;
12221 	ent->dtht_nlocals = vstate->dtvs_nlocals;
12222 
12223 	ent->dtht_fltoffs = (mstate->dtms_present & DTRACE_MSTATE_FLTOFFS) ?
12224 	    mstate->dtms_fltoffs : -1;
12225 	ent->dtht_fault = DTRACE_FLAGS2FLT(flags);
12226 	ent->dtht_illval = cpu_core[CPU->cpu_id].cpuc_dtrace_illval;
12227 
12228 	for (i = 0; i < vstate->dtvs_nlocals; i++) {
12229 		dtrace_statvar_t *svar;
12230 
12231 		if ((svar = vstate->dtvs_locals[i]) == NULL)
12232 			continue;
12233 
12234 		ASSERT(svar->dtsv_size >= NCPU * sizeof (uint64_t));
12235 		ent->dtht_locals[i] =
12236 		    ((uint64_t *)(uintptr_t)svar->dtsv_data)[CPU->cpu_id];
12237 	}
12238 }
12239 
12240 static uint64_t
12241 dtrace_helper(int which, dtrace_mstate_t *mstate,
12242     dtrace_state_t *state, uint64_t arg0, uint64_t arg1)
12243 {
12244 	uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
12245 	uint64_t sarg0 = mstate->dtms_arg[0];
12246 	uint64_t sarg1 = mstate->dtms_arg[1];
12247 	uint64_t rval;
12248 	dtrace_helpers_t *helpers = curproc->p_dtrace_helpers;
12249 	dtrace_helper_action_t *helper;
12250 	dtrace_vstate_t *vstate;
12251 	dtrace_difo_t *pred;
12252 	int i, trace = dtrace_helptrace_enabled;
12253 
12254 	ASSERT(which >= 0 && which < DTRACE_NHELPER_ACTIONS);
12255 
12256 	if (helpers == NULL)
12257 		return (0);
12258 
12259 	if ((helper = helpers->dthps_actions[which]) == NULL)
12260 		return (0);
12261 
12262 	vstate = &helpers->dthps_vstate;
12263 	mstate->dtms_arg[0] = arg0;
12264 	mstate->dtms_arg[1] = arg1;
12265 
12266 	/*
12267 	 * Now iterate over each helper.  If its predicate evaluates to 'true',
12268 	 * we'll call the corresponding actions.  Note that the below calls
12269 	 * to dtrace_dif_emulate() may set faults in machine state.  This is
12270 	 * okay:  our caller (the outer dtrace_dif_emulate()) will simply plow
12271 	 * the stored DIF offset with its own (which is the desired behavior).
12272 	 * Also, note the calls to dtrace_dif_emulate() may allocate scratch
12273 	 * from machine state; this is okay, too.
12274 	 */
12275 	for (; helper != NULL; helper = helper->dtha_next) {
12276 		if ((pred = helper->dtha_predicate) != NULL) {
12277 			if (trace)
12278 				dtrace_helper_trace(helper, mstate, vstate, 0);
12279 
12280 			if (!dtrace_dif_emulate(pred, mstate, vstate, state))
12281 				goto next;
12282 
12283 			if (*flags & CPU_DTRACE_FAULT)
12284 				goto err;
12285 		}
12286 
12287 		for (i = 0; i < helper->dtha_nactions; i++) {
12288 			if (trace)
12289 				dtrace_helper_trace(helper,
12290 				    mstate, vstate, i + 1);
12291 
12292 			rval = dtrace_dif_emulate(helper->dtha_actions[i],
12293 			    mstate, vstate, state);
12294 
12295 			if (*flags & CPU_DTRACE_FAULT)
12296 				goto err;
12297 		}
12298 
12299 next:
12300 		if (trace)
12301 			dtrace_helper_trace(helper, mstate, vstate,
12302 			    DTRACE_HELPTRACE_NEXT);
12303 	}
12304 
12305 	if (trace)
12306 		dtrace_helper_trace(helper, mstate, vstate,
12307 		    DTRACE_HELPTRACE_DONE);
12308 
12309 	/*
12310 	 * Restore the arg0 that we saved upon entry.
12311 	 */
12312 	mstate->dtms_arg[0] = sarg0;
12313 	mstate->dtms_arg[1] = sarg1;
12314 
12315 	return (rval);
12316 
12317 err:
12318 	if (trace)
12319 		dtrace_helper_trace(helper, mstate, vstate,
12320 		    DTRACE_HELPTRACE_ERR);
12321 
12322 	/*
12323 	 * Restore the arg0 that we saved upon entry.
12324 	 */
12325 	mstate->dtms_arg[0] = sarg0;
12326 	mstate->dtms_arg[1] = sarg1;
12327 
12328 	return (NULL);
12329 }
12330 
12331 static void
12332 dtrace_helper_action_destroy(dtrace_helper_action_t *helper,
12333     dtrace_vstate_t *vstate)
12334 {
12335 	int i;
12336 
12337 	if (helper->dtha_predicate != NULL)
12338 		dtrace_difo_release(helper->dtha_predicate, vstate);
12339 
12340 	for (i = 0; i < helper->dtha_nactions; i++) {
12341 		ASSERT(helper->dtha_actions[i] != NULL);
12342 		dtrace_difo_release(helper->dtha_actions[i], vstate);
12343 	}
12344 
12345 	kmem_free(helper->dtha_actions,
12346 	    helper->dtha_nactions * sizeof (dtrace_difo_t *));
12347 	kmem_free(helper, sizeof (dtrace_helper_action_t));
12348 }
12349 
12350 static int
12351 dtrace_helper_destroygen(int gen)
12352 {
12353 	proc_t *p = curproc;
12354 	dtrace_helpers_t *help = p->p_dtrace_helpers;
12355 	dtrace_vstate_t *vstate;
12356 	int i;
12357 
12358 	ASSERT(MUTEX_HELD(&dtrace_lock));
12359 
12360 	if (help == NULL || gen > help->dthps_generation)
12361 		return (EINVAL);
12362 
12363 	vstate = &help->dthps_vstate;
12364 
12365 	for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
12366 		dtrace_helper_action_t *last = NULL, *h, *next;
12367 
12368 		for (h = help->dthps_actions[i]; h != NULL; h = next) {
12369 			next = h->dtha_next;
12370 
12371 			if (h->dtha_generation == gen) {
12372 				if (last != NULL) {
12373 					last->dtha_next = next;
12374 				} else {
12375 					help->dthps_actions[i] = next;
12376 				}
12377 
12378 				dtrace_helper_action_destroy(h, vstate);
12379 			} else {
12380 				last = h;
12381 			}
12382 		}
12383 	}
12384 
12385 	/*
12386 	 * Interate until we've cleared out all helper providers with the
12387 	 * given generation number.
12388 	 */
12389 	for (;;) {
12390 		dtrace_helper_provider_t *prov;
12391 
12392 		/*
12393 		 * Look for a helper provider with the right generation. We
12394 		 * have to start back at the beginning of the list each time
12395 		 * because we drop dtrace_lock. It's unlikely that we'll make
12396 		 * more than two passes.
12397 		 */
12398 		for (i = 0; i < help->dthps_nprovs; i++) {
12399 			prov = help->dthps_provs[i];
12400 
12401 			if (prov->dthp_generation == gen)
12402 				break;
12403 		}
12404 
12405 		/*
12406 		 * If there were no matches, we're done.
12407 		 */
12408 		if (i == help->dthps_nprovs)
12409 			break;
12410 
12411 		/*
12412 		 * Move the last helper provider into this slot.
12413 		 */
12414 		help->dthps_nprovs--;
12415 		help->dthps_provs[i] = help->dthps_provs[help->dthps_nprovs];
12416 		help->dthps_provs[help->dthps_nprovs] = NULL;
12417 
12418 		mutex_exit(&dtrace_lock);
12419 
12420 		/*
12421 		 * If we have a meta provider, remove this helper provider.
12422 		 */
12423 		mutex_enter(&dtrace_meta_lock);
12424 		if (dtrace_meta_pid != NULL) {
12425 			ASSERT(dtrace_deferred_pid == NULL);
12426 			dtrace_helper_provider_remove(&prov->dthp_prov,
12427 			    p->p_pid);
12428 		}
12429 		mutex_exit(&dtrace_meta_lock);
12430 
12431 		dtrace_helper_provider_destroy(prov);
12432 
12433 		mutex_enter(&dtrace_lock);
12434 	}
12435 
12436 	return (0);
12437 }
12438 
12439 static int
12440 dtrace_helper_validate(dtrace_helper_action_t *helper)
12441 {
12442 	int err = 0, i;
12443 	dtrace_difo_t *dp;
12444 
12445 	if ((dp = helper->dtha_predicate) != NULL)
12446 		err += dtrace_difo_validate_helper(dp);
12447 
12448 	for (i = 0; i < helper->dtha_nactions; i++)
12449 		err += dtrace_difo_validate_helper(helper->dtha_actions[i]);
12450 
12451 	return (err == 0);
12452 }
12453 
12454 static int
12455 dtrace_helper_action_add(int which, dtrace_ecbdesc_t *ep)
12456 {
12457 	dtrace_helpers_t *help;
12458 	dtrace_helper_action_t *helper, *last;
12459 	dtrace_actdesc_t *act;
12460 	dtrace_vstate_t *vstate;
12461 	dtrace_predicate_t *pred;
12462 	int count = 0, nactions = 0, i;
12463 
12464 	if (which < 0 || which >= DTRACE_NHELPER_ACTIONS)
12465 		return (EINVAL);
12466 
12467 	help = curproc->p_dtrace_helpers;
12468 	last = help->dthps_actions[which];
12469 	vstate = &help->dthps_vstate;
12470 
12471 	for (count = 0; last != NULL; last = last->dtha_next) {
12472 		count++;
12473 		if (last->dtha_next == NULL)
12474 			break;
12475 	}
12476 
12477 	/*
12478 	 * If we already have dtrace_helper_actions_max helper actions for this
12479 	 * helper action type, we'll refuse to add a new one.
12480 	 */
12481 	if (count >= dtrace_helper_actions_max)
12482 		return (ENOSPC);
12483 
12484 	helper = kmem_zalloc(sizeof (dtrace_helper_action_t), KM_SLEEP);
12485 	helper->dtha_generation = help->dthps_generation;
12486 
12487 	if ((pred = ep->dted_pred.dtpdd_predicate) != NULL) {
12488 		ASSERT(pred->dtp_difo != NULL);
12489 		dtrace_difo_hold(pred->dtp_difo);
12490 		helper->dtha_predicate = pred->dtp_difo;
12491 	}
12492 
12493 	for (act = ep->dted_action; act != NULL; act = act->dtad_next) {
12494 		if (act->dtad_kind != DTRACEACT_DIFEXPR)
12495 			goto err;
12496 
12497 		if (act->dtad_difo == NULL)
12498 			goto err;
12499 
12500 		nactions++;
12501 	}
12502 
12503 	helper->dtha_actions = kmem_zalloc(sizeof (dtrace_difo_t *) *
12504 	    (helper->dtha_nactions = nactions), KM_SLEEP);
12505 
12506 	for (act = ep->dted_action, i = 0; act != NULL; act = act->dtad_next) {
12507 		dtrace_difo_hold(act->dtad_difo);
12508 		helper->dtha_actions[i++] = act->dtad_difo;
12509 	}
12510 
12511 	if (!dtrace_helper_validate(helper))
12512 		goto err;
12513 
12514 	if (last == NULL) {
12515 		help->dthps_actions[which] = helper;
12516 	} else {
12517 		last->dtha_next = helper;
12518 	}
12519 
12520 	if (vstate->dtvs_nlocals > dtrace_helptrace_nlocals) {
12521 		dtrace_helptrace_nlocals = vstate->dtvs_nlocals;
12522 		dtrace_helptrace_next = 0;
12523 	}
12524 
12525 	return (0);
12526 err:
12527 	dtrace_helper_action_destroy(helper, vstate);
12528 	return (EINVAL);
12529 }
12530 
12531 static void
12532 dtrace_helper_provider_register(proc_t *p, dtrace_helpers_t *help,
12533     dof_helper_t *dofhp)
12534 {
12535 	ASSERT(MUTEX_NOT_HELD(&dtrace_lock));
12536 
12537 	mutex_enter(&dtrace_meta_lock);
12538 	mutex_enter(&dtrace_lock);
12539 
12540 	if (!dtrace_attached() || dtrace_meta_pid == NULL) {
12541 		/*
12542 		 * If the dtrace module is loaded but not attached, or if
12543 		 * there aren't isn't a meta provider registered to deal with
12544 		 * these provider descriptions, we need to postpone creating
12545 		 * the actual providers until later.
12546 		 */
12547 
12548 		if (help->dthps_next == NULL && help->dthps_prev == NULL &&
12549 		    dtrace_deferred_pid != help) {
12550 			help->dthps_deferred = 1;
12551 			help->dthps_pid = p->p_pid;
12552 			help->dthps_next = dtrace_deferred_pid;
12553 			help->dthps_prev = NULL;
12554 			if (dtrace_deferred_pid != NULL)
12555 				dtrace_deferred_pid->dthps_prev = help;
12556 			dtrace_deferred_pid = help;
12557 		}
12558 
12559 		mutex_exit(&dtrace_lock);
12560 
12561 	} else if (dofhp != NULL) {
12562 		/*
12563 		 * If the dtrace module is loaded and we have a particular
12564 		 * helper provider description, pass that off to the
12565 		 * meta provider.
12566 		 */
12567 
12568 		mutex_exit(&dtrace_lock);
12569 
12570 		dtrace_helper_provide(dofhp, p->p_pid);
12571 
12572 	} else {
12573 		/*
12574 		 * Otherwise, just pass all the helper provider descriptions
12575 		 * off to the meta provider.
12576 		 */
12577 
12578 		int i;
12579 		mutex_exit(&dtrace_lock);
12580 
12581 		for (i = 0; i < help->dthps_nprovs; i++) {
12582 			dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
12583 			    p->p_pid);
12584 		}
12585 	}
12586 
12587 	mutex_exit(&dtrace_meta_lock);
12588 }
12589 
12590 static int
12591 dtrace_helper_provider_add(dof_helper_t *dofhp, int gen)
12592 {
12593 	dtrace_helpers_t *help;
12594 	dtrace_helper_provider_t *hprov, **tmp_provs;
12595 	uint_t tmp_maxprovs, i;
12596 
12597 	ASSERT(MUTEX_HELD(&dtrace_lock));
12598 
12599 	help = curproc->p_dtrace_helpers;
12600 	ASSERT(help != NULL);
12601 
12602 	/*
12603 	 * If we already have dtrace_helper_providers_max helper providers,
12604 	 * we're refuse to add a new one.
12605 	 */
12606 	if (help->dthps_nprovs >= dtrace_helper_providers_max)
12607 		return (ENOSPC);
12608 
12609 	/*
12610 	 * Check to make sure this isn't a duplicate.
12611 	 */
12612 	for (i = 0; i < help->dthps_nprovs; i++) {
12613 		if (dofhp->dofhp_addr ==
12614 		    help->dthps_provs[i]->dthp_prov.dofhp_addr)
12615 			return (EALREADY);
12616 	}
12617 
12618 	hprov = kmem_zalloc(sizeof (dtrace_helper_provider_t), KM_SLEEP);
12619 	hprov->dthp_prov = *dofhp;
12620 	hprov->dthp_ref = 1;
12621 	hprov->dthp_generation = gen;
12622 
12623 	/*
12624 	 * Allocate a bigger table for helper providers if it's already full.
12625 	 */
12626 	if (help->dthps_maxprovs == help->dthps_nprovs) {
12627 		tmp_maxprovs = help->dthps_maxprovs;
12628 		tmp_provs = help->dthps_provs;
12629 
12630 		if (help->dthps_maxprovs == 0)
12631 			help->dthps_maxprovs = 2;
12632 		else
12633 			help->dthps_maxprovs *= 2;
12634 		if (help->dthps_maxprovs > dtrace_helper_providers_max)
12635 			help->dthps_maxprovs = dtrace_helper_providers_max;
12636 
12637 		ASSERT(tmp_maxprovs < help->dthps_maxprovs);
12638 
12639 		help->dthps_provs = kmem_zalloc(help->dthps_maxprovs *
12640 		    sizeof (dtrace_helper_provider_t *), KM_SLEEP);
12641 
12642 		if (tmp_provs != NULL) {
12643 			bcopy(tmp_provs, help->dthps_provs, tmp_maxprovs *
12644 			    sizeof (dtrace_helper_provider_t *));
12645 			kmem_free(tmp_provs, tmp_maxprovs *
12646 			    sizeof (dtrace_helper_provider_t *));
12647 		}
12648 	}
12649 
12650 	help->dthps_provs[help->dthps_nprovs] = hprov;
12651 	help->dthps_nprovs++;
12652 
12653 	return (0);
12654 }
12655 
12656 static void
12657 dtrace_helper_provider_destroy(dtrace_helper_provider_t *hprov)
12658 {
12659 	mutex_enter(&dtrace_lock);
12660 
12661 	if (--hprov->dthp_ref == 0) {
12662 		dof_hdr_t *dof;
12663 		mutex_exit(&dtrace_lock);
12664 		dof = (dof_hdr_t *)(uintptr_t)hprov->dthp_prov.dofhp_dof;
12665 		dtrace_dof_destroy(dof);
12666 		kmem_free(hprov, sizeof (dtrace_helper_provider_t));
12667 	} else {
12668 		mutex_exit(&dtrace_lock);
12669 	}
12670 }
12671 
12672 static int
12673 dtrace_helper_provider_validate(dof_hdr_t *dof, dof_sec_t *sec)
12674 {
12675 	uintptr_t daddr = (uintptr_t)dof;
12676 	dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
12677 	dof_provider_t *provider;
12678 	dof_probe_t *probe;
12679 	uint8_t *arg;
12680 	char *strtab, *typestr;
12681 	dof_stridx_t typeidx;
12682 	size_t typesz;
12683 	uint_t nprobes, j, k;
12684 
12685 	ASSERT(sec->dofs_type == DOF_SECT_PROVIDER);
12686 
12687 	if (sec->dofs_offset & (sizeof (uint_t) - 1)) {
12688 		dtrace_dof_error(dof, "misaligned section offset");
12689 		return (-1);
12690 	}
12691 
12692 	/*
12693 	 * The section needs to be large enough to contain the DOF provider
12694 	 * structure appropriate for the given version.
12695 	 */
12696 	if (sec->dofs_size <
12697 	    ((dof->dofh_ident[DOF_ID_VERSION] == DOF_VERSION_1) ?
12698 	    offsetof(dof_provider_t, dofpv_prenoffs) :
12699 	    sizeof (dof_provider_t))) {
12700 		dtrace_dof_error(dof, "provider section too small");
12701 		return (-1);
12702 	}
12703 
12704 	provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
12705 	str_sec = dtrace_dof_sect(dof, DOF_SECT_STRTAB, provider->dofpv_strtab);
12706 	prb_sec = dtrace_dof_sect(dof, DOF_SECT_PROBES, provider->dofpv_probes);
12707 	arg_sec = dtrace_dof_sect(dof, DOF_SECT_PRARGS, provider->dofpv_prargs);
12708 	off_sec = dtrace_dof_sect(dof, DOF_SECT_PROFFS, provider->dofpv_proffs);
12709 
12710 	if (str_sec == NULL || prb_sec == NULL ||
12711 	    arg_sec == NULL || off_sec == NULL)
12712 		return (-1);
12713 
12714 	enoff_sec = NULL;
12715 
12716 	if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
12717 	    provider->dofpv_prenoffs != DOF_SECT_NONE &&
12718 	    (enoff_sec = dtrace_dof_sect(dof, DOF_SECT_PRENOFFS,
12719 	    provider->dofpv_prenoffs)) == NULL)
12720 		return (-1);
12721 
12722 	strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
12723 
12724 	if (provider->dofpv_name >= str_sec->dofs_size ||
12725 	    strlen(strtab + provider->dofpv_name) >= DTRACE_PROVNAMELEN) {
12726 		dtrace_dof_error(dof, "invalid provider name");
12727 		return (-1);
12728 	}
12729 
12730 	if (prb_sec->dofs_entsize == 0 ||
12731 	    prb_sec->dofs_entsize > prb_sec->dofs_size) {
12732 		dtrace_dof_error(dof, "invalid entry size");
12733 		return (-1);
12734 	}
12735 
12736 	if (prb_sec->dofs_entsize & (sizeof (uintptr_t) - 1)) {
12737 		dtrace_dof_error(dof, "misaligned entry size");
12738 		return (-1);
12739 	}
12740 
12741 	if (off_sec->dofs_entsize != sizeof (uint32_t)) {
12742 		dtrace_dof_error(dof, "invalid entry size");
12743 		return (-1);
12744 	}
12745 
12746 	if (off_sec->dofs_offset & (sizeof (uint32_t) - 1)) {
12747 		dtrace_dof_error(dof, "misaligned section offset");
12748 		return (-1);
12749 	}
12750 
12751 	if (arg_sec->dofs_entsize != sizeof (uint8_t)) {
12752 		dtrace_dof_error(dof, "invalid entry size");
12753 		return (-1);
12754 	}
12755 
12756 	arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
12757 
12758 	nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
12759 
12760 	/*
12761 	 * Take a pass through the probes to check for errors.
12762 	 */
12763 	for (j = 0; j < nprobes; j++) {
12764 		probe = (dof_probe_t *)(uintptr_t)(daddr +
12765 		    prb_sec->dofs_offset + j * prb_sec->dofs_entsize);
12766 
12767 		if (probe->dofpr_func >= str_sec->dofs_size) {
12768 			dtrace_dof_error(dof, "invalid function name");
12769 			return (-1);
12770 		}
12771 
12772 		if (strlen(strtab + probe->dofpr_func) >= DTRACE_FUNCNAMELEN) {
12773 			dtrace_dof_error(dof, "function name too long");
12774 			return (-1);
12775 		}
12776 
12777 		if (probe->dofpr_name >= str_sec->dofs_size ||
12778 		    strlen(strtab + probe->dofpr_name) >= DTRACE_NAMELEN) {
12779 			dtrace_dof_error(dof, "invalid probe name");
12780 			return (-1);
12781 		}
12782 
12783 		/*
12784 		 * The offset count must not wrap the index, and the offsets
12785 		 * must also not overflow the section's data.
12786 		 */
12787 		if (probe->dofpr_offidx + probe->dofpr_noffs <
12788 		    probe->dofpr_offidx ||
12789 		    (probe->dofpr_offidx + probe->dofpr_noffs) *
12790 		    off_sec->dofs_entsize > off_sec->dofs_size) {
12791 			dtrace_dof_error(dof, "invalid probe offset");
12792 			return (-1);
12793 		}
12794 
12795 		if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1) {
12796 			/*
12797 			 * If there's no is-enabled offset section, make sure
12798 			 * there aren't any is-enabled offsets. Otherwise
12799 			 * perform the same checks as for probe offsets
12800 			 * (immediately above).
12801 			 */
12802 			if (enoff_sec == NULL) {
12803 				if (probe->dofpr_enoffidx != 0 ||
12804 				    probe->dofpr_nenoffs != 0) {
12805 					dtrace_dof_error(dof, "is-enabled "
12806 					    "offsets with null section");
12807 					return (-1);
12808 				}
12809 			} else if (probe->dofpr_enoffidx +
12810 			    probe->dofpr_nenoffs < probe->dofpr_enoffidx ||
12811 			    (probe->dofpr_enoffidx + probe->dofpr_nenoffs) *
12812 			    enoff_sec->dofs_entsize > enoff_sec->dofs_size) {
12813 				dtrace_dof_error(dof, "invalid is-enabled "
12814 				    "offset");
12815 				return (-1);
12816 			}
12817 
12818 			if (probe->dofpr_noffs + probe->dofpr_nenoffs == 0) {
12819 				dtrace_dof_error(dof, "zero probe and "
12820 				    "is-enabled offsets");
12821 				return (-1);
12822 			}
12823 		} else if (probe->dofpr_noffs == 0) {
12824 			dtrace_dof_error(dof, "zero probe offsets");
12825 			return (-1);
12826 		}
12827 
12828 		if (probe->dofpr_argidx + probe->dofpr_xargc <
12829 		    probe->dofpr_argidx ||
12830 		    (probe->dofpr_argidx + probe->dofpr_xargc) *
12831 		    arg_sec->dofs_entsize > arg_sec->dofs_size) {
12832 			dtrace_dof_error(dof, "invalid args");
12833 			return (-1);
12834 		}
12835 
12836 		typeidx = probe->dofpr_nargv;
12837 		typestr = strtab + probe->dofpr_nargv;
12838 		for (k = 0; k < probe->dofpr_nargc; k++) {
12839 			if (typeidx >= str_sec->dofs_size) {
12840 				dtrace_dof_error(dof, "bad "
12841 				    "native argument type");
12842 				return (-1);
12843 			}
12844 
12845 			typesz = strlen(typestr) + 1;
12846 			if (typesz > DTRACE_ARGTYPELEN) {
12847 				dtrace_dof_error(dof, "native "
12848 				    "argument type too long");
12849 				return (-1);
12850 			}
12851 			typeidx += typesz;
12852 			typestr += typesz;
12853 		}
12854 
12855 		typeidx = probe->dofpr_xargv;
12856 		typestr = strtab + probe->dofpr_xargv;
12857 		for (k = 0; k < probe->dofpr_xargc; k++) {
12858 			if (arg[probe->dofpr_argidx + k] > probe->dofpr_nargc) {
12859 				dtrace_dof_error(dof, "bad "
12860 				    "native argument index");
12861 				return (-1);
12862 			}
12863 
12864 			if (typeidx >= str_sec->dofs_size) {
12865 				dtrace_dof_error(dof, "bad "
12866 				    "translated argument type");
12867 				return (-1);
12868 			}
12869 
12870 			typesz = strlen(typestr) + 1;
12871 			if (typesz > DTRACE_ARGTYPELEN) {
12872 				dtrace_dof_error(dof, "translated argument "
12873 				    "type too long");
12874 				return (-1);
12875 			}
12876 
12877 			typeidx += typesz;
12878 			typestr += typesz;
12879 		}
12880 	}
12881 
12882 	return (0);
12883 }
12884 
12885 static int
12886 dtrace_helper_slurp(dof_hdr_t *dof, dof_helper_t *dhp)
12887 {
12888 	dtrace_helpers_t *help;
12889 	dtrace_vstate_t *vstate;
12890 	dtrace_enabling_t *enab = NULL;
12891 	int i, gen, rv, nhelpers = 0, nprovs = 0, destroy = 1;
12892 	uintptr_t daddr = (uintptr_t)dof;
12893 
12894 	ASSERT(MUTEX_HELD(&dtrace_lock));
12895 
12896 	if ((help = curproc->p_dtrace_helpers) == NULL)
12897 		help = dtrace_helpers_create(curproc);
12898 
12899 	vstate = &help->dthps_vstate;
12900 
12901 	if ((rv = dtrace_dof_slurp(dof, vstate, NULL, &enab,
12902 	    dhp != NULL ? dhp->dofhp_addr : 0, B_FALSE)) != 0) {
12903 		dtrace_dof_destroy(dof);
12904 		return (rv);
12905 	}
12906 
12907 	/*
12908 	 * Look for helper providers and validate their descriptions.
12909 	 */
12910 	if (dhp != NULL) {
12911 		for (i = 0; i < dof->dofh_secnum; i++) {
12912 			dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
12913 			    dof->dofh_secoff + i * dof->dofh_secsize);
12914 
12915 			if (sec->dofs_type != DOF_SECT_PROVIDER)
12916 				continue;
12917 
12918 			if (dtrace_helper_provider_validate(dof, sec) != 0) {
12919 				dtrace_enabling_destroy(enab);
12920 				dtrace_dof_destroy(dof);
12921 				return (-1);
12922 			}
12923 
12924 			nprovs++;
12925 		}
12926 	}
12927 
12928 	/*
12929 	 * Now we need to walk through the ECB descriptions in the enabling.
12930 	 */
12931 	for (i = 0; i < enab->dten_ndesc; i++) {
12932 		dtrace_ecbdesc_t *ep = enab->dten_desc[i];
12933 		dtrace_probedesc_t *desc = &ep->dted_probe;
12934 
12935 		if (strcmp(desc->dtpd_provider, "dtrace") != 0)
12936 			continue;
12937 
12938 		if (strcmp(desc->dtpd_mod, "helper") != 0)
12939 			continue;
12940 
12941 		if (strcmp(desc->dtpd_func, "ustack") != 0)
12942 			continue;
12943 
12944 		if ((rv = dtrace_helper_action_add(DTRACE_HELPER_ACTION_USTACK,
12945 		    ep)) != 0) {
12946 			/*
12947 			 * Adding this helper action failed -- we are now going
12948 			 * to rip out the entire generation and return failure.
12949 			 */
12950 			(void) dtrace_helper_destroygen(help->dthps_generation);
12951 			dtrace_enabling_destroy(enab);
12952 			dtrace_dof_destroy(dof);
12953 			return (-1);
12954 		}
12955 
12956 		nhelpers++;
12957 	}
12958 
12959 	if (nhelpers < enab->dten_ndesc)
12960 		dtrace_dof_error(dof, "unmatched helpers");
12961 
12962 	gen = help->dthps_generation++;
12963 	dtrace_enabling_destroy(enab);
12964 
12965 	if (dhp != NULL && nprovs > 0) {
12966 		dhp->dofhp_dof = (uint64_t)(uintptr_t)dof;
12967 		if (dtrace_helper_provider_add(dhp, gen) == 0) {
12968 			mutex_exit(&dtrace_lock);
12969 			dtrace_helper_provider_register(curproc, help, dhp);
12970 			mutex_enter(&dtrace_lock);
12971 
12972 			destroy = 0;
12973 		}
12974 	}
12975 
12976 	if (destroy)
12977 		dtrace_dof_destroy(dof);
12978 
12979 	return (gen);
12980 }
12981 
12982 static dtrace_helpers_t *
12983 dtrace_helpers_create(proc_t *p)
12984 {
12985 	dtrace_helpers_t *help;
12986 
12987 	ASSERT(MUTEX_HELD(&dtrace_lock));
12988 	ASSERT(p->p_dtrace_helpers == NULL);
12989 
12990 	help = kmem_zalloc(sizeof (dtrace_helpers_t), KM_SLEEP);
12991 	help->dthps_actions = kmem_zalloc(sizeof (dtrace_helper_action_t *) *
12992 	    DTRACE_NHELPER_ACTIONS, KM_SLEEP);
12993 
12994 	p->p_dtrace_helpers = help;
12995 	dtrace_helpers++;
12996 
12997 	return (help);
12998 }
12999 
13000 static void
13001 dtrace_helpers_destroy(void)
13002 {
13003 	dtrace_helpers_t *help;
13004 	dtrace_vstate_t *vstate;
13005 	proc_t *p = curproc;
13006 	int i;
13007 
13008 	mutex_enter(&dtrace_lock);
13009 
13010 	ASSERT(p->p_dtrace_helpers != NULL);
13011 	ASSERT(dtrace_helpers > 0);
13012 
13013 	help = p->p_dtrace_helpers;
13014 	vstate = &help->dthps_vstate;
13015 
13016 	/*
13017 	 * We're now going to lose the help from this process.
13018 	 */
13019 	p->p_dtrace_helpers = NULL;
13020 	dtrace_sync();
13021 
13022 	/*
13023 	 * Destory the helper actions.
13024 	 */
13025 	for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
13026 		dtrace_helper_action_t *h, *next;
13027 
13028 		for (h = help->dthps_actions[i]; h != NULL; h = next) {
13029 			next = h->dtha_next;
13030 			dtrace_helper_action_destroy(h, vstate);
13031 			h = next;
13032 		}
13033 	}
13034 
13035 	mutex_exit(&dtrace_lock);
13036 
13037 	/*
13038 	 * Destroy the helper providers.
13039 	 */
13040 	if (help->dthps_maxprovs > 0) {
13041 		mutex_enter(&dtrace_meta_lock);
13042 		if (dtrace_meta_pid != NULL) {
13043 			ASSERT(dtrace_deferred_pid == NULL);
13044 
13045 			for (i = 0; i < help->dthps_nprovs; i++) {
13046 				dtrace_helper_provider_remove(
13047 				    &help->dthps_provs[i]->dthp_prov, p->p_pid);
13048 			}
13049 		} else {
13050 			mutex_enter(&dtrace_lock);
13051 			ASSERT(help->dthps_deferred == 0 ||
13052 			    help->dthps_next != NULL ||
13053 			    help->dthps_prev != NULL ||
13054 			    help == dtrace_deferred_pid);
13055 
13056 			/*
13057 			 * Remove the helper from the deferred list.
13058 			 */
13059 			if (help->dthps_next != NULL)
13060 				help->dthps_next->dthps_prev = help->dthps_prev;
13061 			if (help->dthps_prev != NULL)
13062 				help->dthps_prev->dthps_next = help->dthps_next;
13063 			if (dtrace_deferred_pid == help) {
13064 				dtrace_deferred_pid = help->dthps_next;
13065 				ASSERT(help->dthps_prev == NULL);
13066 			}
13067 
13068 			mutex_exit(&dtrace_lock);
13069 		}
13070 
13071 		mutex_exit(&dtrace_meta_lock);
13072 
13073 		for (i = 0; i < help->dthps_nprovs; i++) {
13074 			dtrace_helper_provider_destroy(help->dthps_provs[i]);
13075 		}
13076 
13077 		kmem_free(help->dthps_provs, help->dthps_maxprovs *
13078 		    sizeof (dtrace_helper_provider_t *));
13079 	}
13080 
13081 	mutex_enter(&dtrace_lock);
13082 
13083 	dtrace_vstate_fini(&help->dthps_vstate);
13084 	kmem_free(help->dthps_actions,
13085 	    sizeof (dtrace_helper_action_t *) * DTRACE_NHELPER_ACTIONS);
13086 	kmem_free(help, sizeof (dtrace_helpers_t));
13087 
13088 	--dtrace_helpers;
13089 	mutex_exit(&dtrace_lock);
13090 }
13091 
13092 static void
13093 dtrace_helpers_duplicate(proc_t *from, proc_t *to)
13094 {
13095 	dtrace_helpers_t *help, *newhelp;
13096 	dtrace_helper_action_t *helper, *new, *last;
13097 	dtrace_difo_t *dp;
13098 	dtrace_vstate_t *vstate;
13099 	int i, j, sz, hasprovs = 0;
13100 
13101 	mutex_enter(&dtrace_lock);
13102 	ASSERT(from->p_dtrace_helpers != NULL);
13103 	ASSERT(dtrace_helpers > 0);
13104 
13105 	help = from->p_dtrace_helpers;
13106 	newhelp = dtrace_helpers_create(to);
13107 	ASSERT(to->p_dtrace_helpers != NULL);
13108 
13109 	newhelp->dthps_generation = help->dthps_generation;
13110 	vstate = &newhelp->dthps_vstate;
13111 
13112 	/*
13113 	 * Duplicate the helper actions.
13114 	 */
13115 	for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
13116 		if ((helper = help->dthps_actions[i]) == NULL)
13117 			continue;
13118 
13119 		for (last = NULL; helper != NULL; helper = helper->dtha_next) {
13120 			new = kmem_zalloc(sizeof (dtrace_helper_action_t),
13121 			    KM_SLEEP);
13122 			new->dtha_generation = helper->dtha_generation;
13123 
13124 			if ((dp = helper->dtha_predicate) != NULL) {
13125 				dp = dtrace_difo_duplicate(dp, vstate);
13126 				new->dtha_predicate = dp;
13127 			}
13128 
13129 			new->dtha_nactions = helper->dtha_nactions;
13130 			sz = sizeof (dtrace_difo_t *) * new->dtha_nactions;
13131 			new->dtha_actions = kmem_alloc(sz, KM_SLEEP);
13132 
13133 			for (j = 0; j < new->dtha_nactions; j++) {
13134 				dtrace_difo_t *dp = helper->dtha_actions[j];
13135 
13136 				ASSERT(dp != NULL);
13137 				dp = dtrace_difo_duplicate(dp, vstate);
13138 				new->dtha_actions[j] = dp;
13139 			}
13140 
13141 			if (last != NULL) {
13142 				last->dtha_next = new;
13143 			} else {
13144 				newhelp->dthps_actions[i] = new;
13145 			}
13146 
13147 			last = new;
13148 		}
13149 	}
13150 
13151 	/*
13152 	 * Duplicate the helper providers and register them with the
13153 	 * DTrace framework.
13154 	 */
13155 	if (help->dthps_nprovs > 0) {
13156 		newhelp->dthps_nprovs = help->dthps_nprovs;
13157 		newhelp->dthps_maxprovs = help->dthps_nprovs;
13158 		newhelp->dthps_provs = kmem_alloc(newhelp->dthps_nprovs *
13159 		    sizeof (dtrace_helper_provider_t *), KM_SLEEP);
13160 		for (i = 0; i < newhelp->dthps_nprovs; i++) {
13161 			newhelp->dthps_provs[i] = help->dthps_provs[i];
13162 			newhelp->dthps_provs[i]->dthp_ref++;
13163 		}
13164 
13165 		hasprovs = 1;
13166 	}
13167 
13168 	mutex_exit(&dtrace_lock);
13169 
13170 	if (hasprovs)
13171 		dtrace_helper_provider_register(to, newhelp, NULL);
13172 }
13173 
13174 /*
13175  * DTrace Hook Functions
13176  */
13177 static void
13178 dtrace_module_loaded(struct modctl *ctl)
13179 {
13180 	dtrace_provider_t *prv;
13181 
13182 	mutex_enter(&dtrace_provider_lock);
13183 	mutex_enter(&mod_lock);
13184 
13185 	ASSERT(ctl->mod_busy);
13186 
13187 	/*
13188 	 * We're going to call each providers per-module provide operation
13189 	 * specifying only this module.
13190 	 */
13191 	for (prv = dtrace_provider; prv != NULL; prv = prv->dtpv_next)
13192 		prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
13193 
13194 	mutex_exit(&mod_lock);
13195 	mutex_exit(&dtrace_provider_lock);
13196 
13197 	/*
13198 	 * If we have any retained enablings, we need to match against them.
13199 	 * Enabling probes requires that cpu_lock be held, and we cannot hold
13200 	 * cpu_lock here -- it is legal for cpu_lock to be held when loading a
13201 	 * module.  (In particular, this happens when loading scheduling
13202 	 * classes.)  So if we have any retained enablings, we need to dispatch
13203 	 * our task queue to do the match for us.
13204 	 */
13205 	mutex_enter(&dtrace_lock);
13206 
13207 	if (dtrace_retained == NULL) {
13208 		mutex_exit(&dtrace_lock);
13209 		return;
13210 	}
13211 
13212 	(void) taskq_dispatch(dtrace_taskq,
13213 	    (task_func_t *)dtrace_enabling_matchall, NULL, TQ_SLEEP);
13214 
13215 	mutex_exit(&dtrace_lock);
13216 
13217 	/*
13218 	 * And now, for a little heuristic sleaze:  in general, we want to
13219 	 * match modules as soon as they load.  However, we cannot guarantee
13220 	 * this, because it would lead us to the lock ordering violation
13221 	 * outlined above.  The common case, of course, is that cpu_lock is
13222 	 * _not_ held -- so we delay here for a clock tick, hoping that that's
13223 	 * long enough for the task queue to do its work.  If it's not, it's
13224 	 * not a serious problem -- it just means that the module that we
13225 	 * just loaded may not be immediately instrumentable.
13226 	 */
13227 	delay(1);
13228 }
13229 
13230 static void
13231 dtrace_module_unloaded(struct modctl *ctl)
13232 {
13233 	dtrace_probe_t template, *probe, *first, *next;
13234 	dtrace_provider_t *prov;
13235 
13236 	template.dtpr_mod = ctl->mod_modname;
13237 
13238 	mutex_enter(&dtrace_provider_lock);
13239 	mutex_enter(&mod_lock);
13240 	mutex_enter(&dtrace_lock);
13241 
13242 	if (dtrace_bymod == NULL) {
13243 		/*
13244 		 * The DTrace module is loaded (obviously) but not attached;
13245 		 * we don't have any work to do.
13246 		 */
13247 		mutex_exit(&dtrace_provider_lock);
13248 		mutex_exit(&mod_lock);
13249 		mutex_exit(&dtrace_lock);
13250 		return;
13251 	}
13252 
13253 	for (probe = first = dtrace_hash_lookup(dtrace_bymod, &template);
13254 	    probe != NULL; probe = probe->dtpr_nextmod) {
13255 		if (probe->dtpr_ecb != NULL) {
13256 			mutex_exit(&dtrace_provider_lock);
13257 			mutex_exit(&mod_lock);
13258 			mutex_exit(&dtrace_lock);
13259 
13260 			/*
13261 			 * This shouldn't _actually_ be possible -- we're
13262 			 * unloading a module that has an enabled probe in it.
13263 			 * (It's normally up to the provider to make sure that
13264 			 * this can't happen.)  However, because dtps_enable()
13265 			 * doesn't have a failure mode, there can be an
13266 			 * enable/unload race.  Upshot:  we don't want to
13267 			 * assert, but we're not going to disable the
13268 			 * probe, either.
13269 			 */
13270 			if (dtrace_err_verbose) {
13271 				cmn_err(CE_WARN, "unloaded module '%s' had "
13272 				    "enabled probes", ctl->mod_modname);
13273 			}
13274 
13275 			return;
13276 		}
13277 	}
13278 
13279 	probe = first;
13280 
13281 	for (first = NULL; probe != NULL; probe = next) {
13282 		ASSERT(dtrace_probes[probe->dtpr_id - 1] == probe);
13283 
13284 		dtrace_probes[probe->dtpr_id - 1] = NULL;
13285 
13286 		next = probe->dtpr_nextmod;
13287 		dtrace_hash_remove(dtrace_bymod, probe);
13288 		dtrace_hash_remove(dtrace_byfunc, probe);
13289 		dtrace_hash_remove(dtrace_byname, probe);
13290 
13291 		if (first == NULL) {
13292 			first = probe;
13293 			probe->dtpr_nextmod = NULL;
13294 		} else {
13295 			probe->dtpr_nextmod = first;
13296 			first = probe;
13297 		}
13298 	}
13299 
13300 	/*
13301 	 * We've removed all of the module's probes from the hash chains and
13302 	 * from the probe array.  Now issue a dtrace_sync() to be sure that
13303 	 * everyone has cleared out from any probe array processing.
13304 	 */
13305 	dtrace_sync();
13306 
13307 	for (probe = first; probe != NULL; probe = first) {
13308 		first = probe->dtpr_nextmod;
13309 		prov = probe->dtpr_provider;
13310 		prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, probe->dtpr_id,
13311 		    probe->dtpr_arg);
13312 		kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
13313 		kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
13314 		kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
13315 		vmem_free(dtrace_arena, (void *)(uintptr_t)probe->dtpr_id, 1);
13316 		kmem_free(probe, sizeof (dtrace_probe_t));
13317 	}
13318 
13319 	mutex_exit(&dtrace_lock);
13320 	mutex_exit(&mod_lock);
13321 	mutex_exit(&dtrace_provider_lock);
13322 }
13323 
13324 void
13325 dtrace_suspend(void)
13326 {
13327 	dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_suspend));
13328 }
13329 
13330 void
13331 dtrace_resume(void)
13332 {
13333 	dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_resume));
13334 }
13335 
13336 static int
13337 dtrace_cpu_setup(cpu_setup_t what, processorid_t cpu)
13338 {
13339 	ASSERT(MUTEX_HELD(&cpu_lock));
13340 	mutex_enter(&dtrace_lock);
13341 
13342 	switch (what) {
13343 	case CPU_CONFIG: {
13344 		dtrace_state_t *state;
13345 		dtrace_optval_t *opt, rs, c;
13346 
13347 		/*
13348 		 * For now, we only allocate a new buffer for anonymous state.
13349 		 */
13350 		if ((state = dtrace_anon.dta_state) == NULL)
13351 			break;
13352 
13353 		if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
13354 			break;
13355 
13356 		opt = state->dts_options;
13357 		c = opt[DTRACEOPT_CPU];
13358 
13359 		if (c != DTRACE_CPUALL && c != DTRACEOPT_UNSET && c != cpu)
13360 			break;
13361 
13362 		/*
13363 		 * Regardless of what the actual policy is, we're going to
13364 		 * temporarily set our resize policy to be manual.  We're
13365 		 * also going to temporarily set our CPU option to denote
13366 		 * the newly configured CPU.
13367 		 */
13368 		rs = opt[DTRACEOPT_BUFRESIZE];
13369 		opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_MANUAL;
13370 		opt[DTRACEOPT_CPU] = (dtrace_optval_t)cpu;
13371 
13372 		(void) dtrace_state_buffers(state);
13373 
13374 		opt[DTRACEOPT_BUFRESIZE] = rs;
13375 		opt[DTRACEOPT_CPU] = c;
13376 
13377 		break;
13378 	}
13379 
13380 	case CPU_UNCONFIG:
13381 		/*
13382 		 * We don't free the buffer in the CPU_UNCONFIG case.  (The
13383 		 * buffer will be freed when the consumer exits.)
13384 		 */
13385 		break;
13386 
13387 	default:
13388 		break;
13389 	}
13390 
13391 	mutex_exit(&dtrace_lock);
13392 	return (0);
13393 }
13394 
13395 static void
13396 dtrace_cpu_setup_initial(processorid_t cpu)
13397 {
13398 	(void) dtrace_cpu_setup(CPU_CONFIG, cpu);
13399 }
13400 
13401 static void
13402 dtrace_toxrange_add(uintptr_t base, uintptr_t limit)
13403 {
13404 	if (dtrace_toxranges >= dtrace_toxranges_max) {
13405 		int osize, nsize;
13406 		dtrace_toxrange_t *range;
13407 
13408 		osize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
13409 
13410 		if (osize == 0) {
13411 			ASSERT(dtrace_toxrange == NULL);
13412 			ASSERT(dtrace_toxranges_max == 0);
13413 			dtrace_toxranges_max = 1;
13414 		} else {
13415 			dtrace_toxranges_max <<= 1;
13416 		}
13417 
13418 		nsize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
13419 		range = kmem_zalloc(nsize, KM_SLEEP);
13420 
13421 		if (dtrace_toxrange != NULL) {
13422 			ASSERT(osize != 0);
13423 			bcopy(dtrace_toxrange, range, osize);
13424 			kmem_free(dtrace_toxrange, osize);
13425 		}
13426 
13427 		dtrace_toxrange = range;
13428 	}
13429 
13430 	ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_base == NULL);
13431 	ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_limit == NULL);
13432 
13433 	dtrace_toxrange[dtrace_toxranges].dtt_base = base;
13434 	dtrace_toxrange[dtrace_toxranges].dtt_limit = limit;
13435 	dtrace_toxranges++;
13436 }
13437 
13438 /*
13439  * DTrace Driver Cookbook Functions
13440  */
13441 /*ARGSUSED*/
13442 static int
13443 dtrace_attach(dev_info_t *devi, ddi_attach_cmd_t cmd)
13444 {
13445 	dtrace_provider_id_t id;
13446 	dtrace_state_t *state = NULL;
13447 	dtrace_enabling_t *enab;
13448 
13449 	mutex_enter(&cpu_lock);
13450 	mutex_enter(&dtrace_provider_lock);
13451 	mutex_enter(&dtrace_lock);
13452 
13453 	if (ddi_soft_state_init(&dtrace_softstate,
13454 	    sizeof (dtrace_state_t), 0) != 0) {
13455 		cmn_err(CE_NOTE, "/dev/dtrace failed to initialize soft state");
13456 		mutex_exit(&cpu_lock);
13457 		mutex_exit(&dtrace_provider_lock);
13458 		mutex_exit(&dtrace_lock);
13459 		return (DDI_FAILURE);
13460 	}
13461 
13462 	if (ddi_create_minor_node(devi, DTRACEMNR_DTRACE, S_IFCHR,
13463 	    DTRACEMNRN_DTRACE, DDI_PSEUDO, NULL) == DDI_FAILURE ||
13464 	    ddi_create_minor_node(devi, DTRACEMNR_HELPER, S_IFCHR,
13465 	    DTRACEMNRN_HELPER, DDI_PSEUDO, NULL) == DDI_FAILURE) {
13466 		cmn_err(CE_NOTE, "/dev/dtrace couldn't create minor nodes");
13467 		ddi_remove_minor_node(devi, NULL);
13468 		ddi_soft_state_fini(&dtrace_softstate);
13469 		mutex_exit(&cpu_lock);
13470 		mutex_exit(&dtrace_provider_lock);
13471 		mutex_exit(&dtrace_lock);
13472 		return (DDI_FAILURE);
13473 	}
13474 
13475 	ddi_report_dev(devi);
13476 	dtrace_devi = devi;
13477 
13478 	dtrace_modload = dtrace_module_loaded;
13479 	dtrace_modunload = dtrace_module_unloaded;
13480 	dtrace_cpu_init = dtrace_cpu_setup_initial;
13481 	dtrace_helpers_cleanup = dtrace_helpers_destroy;
13482 	dtrace_helpers_fork = dtrace_helpers_duplicate;
13483 	dtrace_cpustart_init = dtrace_suspend;
13484 	dtrace_cpustart_fini = dtrace_resume;
13485 	dtrace_debugger_init = dtrace_suspend;
13486 	dtrace_debugger_fini = dtrace_resume;
13487 	dtrace_kreloc_init = dtrace_suspend;
13488 	dtrace_kreloc_fini = dtrace_resume;
13489 
13490 	register_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
13491 
13492 	ASSERT(MUTEX_HELD(&cpu_lock));
13493 
13494 	dtrace_arena = vmem_create("dtrace", (void *)1, UINT32_MAX, 1,
13495 	    NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
13496 	dtrace_minor = vmem_create("dtrace_minor", (void *)DTRACEMNRN_CLONE,
13497 	    UINT32_MAX - DTRACEMNRN_CLONE, 1, NULL, NULL, NULL, 0,
13498 	    VM_SLEEP | VMC_IDENTIFIER);
13499 	dtrace_taskq = taskq_create("dtrace_taskq", 1, maxclsyspri,
13500 	    1, INT_MAX, 0);
13501 
13502 	dtrace_state_cache = kmem_cache_create("dtrace_state_cache",
13503 	    sizeof (dtrace_dstate_percpu_t) * NCPU, DTRACE_STATE_ALIGN,
13504 	    NULL, NULL, NULL, NULL, NULL, 0);
13505 
13506 	ASSERT(MUTEX_HELD(&cpu_lock));
13507 	dtrace_bymod = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_mod),
13508 	    offsetof(dtrace_probe_t, dtpr_nextmod),
13509 	    offsetof(dtrace_probe_t, dtpr_prevmod));
13510 
13511 	dtrace_byfunc = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_func),
13512 	    offsetof(dtrace_probe_t, dtpr_nextfunc),
13513 	    offsetof(dtrace_probe_t, dtpr_prevfunc));
13514 
13515 	dtrace_byname = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_name),
13516 	    offsetof(dtrace_probe_t, dtpr_nextname),
13517 	    offsetof(dtrace_probe_t, dtpr_prevname));
13518 
13519 	if (dtrace_retain_max < 1) {
13520 		cmn_err(CE_WARN, "illegal value (%lu) for dtrace_retain_max; "
13521 		    "setting to 1", dtrace_retain_max);
13522 		dtrace_retain_max = 1;
13523 	}
13524 
13525 	/*
13526 	 * Now discover our toxic ranges.
13527 	 */
13528 	dtrace_toxic_ranges(dtrace_toxrange_add);
13529 
13530 	/*
13531 	 * Before we register ourselves as a provider to our own framework,
13532 	 * we would like to assert that dtrace_provider is NULL -- but that's
13533 	 * not true if we were loaded as a dependency of a DTrace provider.
13534 	 * Once we've registered, we can assert that dtrace_provider is our
13535 	 * pseudo provider.
13536 	 */
13537 	(void) dtrace_register("dtrace", &dtrace_provider_attr,
13538 	    DTRACE_PRIV_NONE, 0, &dtrace_provider_ops, NULL, &id);
13539 
13540 	ASSERT(dtrace_provider != NULL);
13541 	ASSERT((dtrace_provider_id_t)dtrace_provider == id);
13542 
13543 	dtrace_probeid_begin = dtrace_probe_create((dtrace_provider_id_t)
13544 	    dtrace_provider, NULL, NULL, "BEGIN", 0, NULL);
13545 	dtrace_probeid_end = dtrace_probe_create((dtrace_provider_id_t)
13546 	    dtrace_provider, NULL, NULL, "END", 0, NULL);
13547 	dtrace_probeid_error = dtrace_probe_create((dtrace_provider_id_t)
13548 	    dtrace_provider, NULL, NULL, "ERROR", 1, NULL);
13549 
13550 	dtrace_anon_property();
13551 	mutex_exit(&cpu_lock);
13552 
13553 	/*
13554 	 * If DTrace helper tracing is enabled, we need to allocate the
13555 	 * trace buffer and initialize the values.
13556 	 */
13557 	if (dtrace_helptrace_enabled) {
13558 		ASSERT(dtrace_helptrace_buffer == NULL);
13559 		dtrace_helptrace_buffer =
13560 		    kmem_zalloc(dtrace_helptrace_bufsize, KM_SLEEP);
13561 		dtrace_helptrace_next = 0;
13562 	}
13563 
13564 	/*
13565 	 * If there are already providers, we must ask them to provide their
13566 	 * probes, and then match any anonymous enabling against them.  Note
13567 	 * that there should be no other retained enablings at this time:
13568 	 * the only retained enablings at this time should be the anonymous
13569 	 * enabling.
13570 	 */
13571 	if (dtrace_anon.dta_enabling != NULL) {
13572 		ASSERT(dtrace_retained == dtrace_anon.dta_enabling);
13573 
13574 		dtrace_enabling_provide(NULL);
13575 		state = dtrace_anon.dta_state;
13576 
13577 		/*
13578 		 * We couldn't hold cpu_lock across the above call to
13579 		 * dtrace_enabling_provide(), but we must hold it to actually
13580 		 * enable the probes.  We have to drop all of our locks, pick
13581 		 * up cpu_lock, and regain our locks before matching the
13582 		 * retained anonymous enabling.
13583 		 */
13584 		mutex_exit(&dtrace_lock);
13585 		mutex_exit(&dtrace_provider_lock);
13586 
13587 		mutex_enter(&cpu_lock);
13588 		mutex_enter(&dtrace_provider_lock);
13589 		mutex_enter(&dtrace_lock);
13590 
13591 		if ((enab = dtrace_anon.dta_enabling) != NULL)
13592 			(void) dtrace_enabling_match(enab, NULL);
13593 
13594 		mutex_exit(&cpu_lock);
13595 	}
13596 
13597 	mutex_exit(&dtrace_lock);
13598 	mutex_exit(&dtrace_provider_lock);
13599 
13600 	if (state != NULL) {
13601 		/*
13602 		 * If we created any anonymous state, set it going now.
13603 		 */
13604 		(void) dtrace_state_go(state, &dtrace_anon.dta_beganon);
13605 	}
13606 
13607 	return (DDI_SUCCESS);
13608 }
13609 
13610 /*ARGSUSED*/
13611 static int
13612 dtrace_open(dev_t *devp, int flag, int otyp, cred_t *cred_p)
13613 {
13614 	dtrace_state_t *state;
13615 	uint32_t priv;
13616 	uid_t uid;
13617 	zoneid_t zoneid;
13618 
13619 	if (getminor(*devp) == DTRACEMNRN_HELPER)
13620 		return (0);
13621 
13622 	/*
13623 	 * If this wasn't an open with the "helper" minor, then it must be
13624 	 * the "dtrace" minor.
13625 	 */
13626 	ASSERT(getminor(*devp) == DTRACEMNRN_DTRACE);
13627 
13628 	/*
13629 	 * If no DTRACE_PRIV_* bits are set in the credential, then the
13630 	 * caller lacks sufficient permission to do anything with DTrace.
13631 	 */
13632 	dtrace_cred2priv(cred_p, &priv, &uid, &zoneid);
13633 	if (priv == DTRACE_PRIV_NONE)
13634 		return (EACCES);
13635 
13636 	/*
13637 	 * Ask all providers to provide all their probes.
13638 	 */
13639 	mutex_enter(&dtrace_provider_lock);
13640 	dtrace_probe_provide(NULL, NULL);
13641 	mutex_exit(&dtrace_provider_lock);
13642 
13643 	mutex_enter(&cpu_lock);
13644 	mutex_enter(&dtrace_lock);
13645 	dtrace_opens++;
13646 	dtrace_membar_producer();
13647 
13648 	/*
13649 	 * If the kernel debugger is active (that is, if the kernel debugger
13650 	 * modified text in some way), we won't allow the open.
13651 	 */
13652 	if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
13653 		dtrace_opens--;
13654 		mutex_exit(&cpu_lock);
13655 		mutex_exit(&dtrace_lock);
13656 		return (EBUSY);
13657 	}
13658 
13659 	state = dtrace_state_create(devp, cred_p);
13660 	mutex_exit(&cpu_lock);
13661 
13662 	if (state == NULL) {
13663 		if (--dtrace_opens == 0)
13664 			(void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
13665 		mutex_exit(&dtrace_lock);
13666 		return (EAGAIN);
13667 	}
13668 
13669 	mutex_exit(&dtrace_lock);
13670 
13671 	return (0);
13672 }
13673 
13674 /*ARGSUSED*/
13675 static int
13676 dtrace_close(dev_t dev, int flag, int otyp, cred_t *cred_p)
13677 {
13678 	minor_t minor = getminor(dev);
13679 	dtrace_state_t *state;
13680 
13681 	if (minor == DTRACEMNRN_HELPER)
13682 		return (0);
13683 
13684 	state = ddi_get_soft_state(dtrace_softstate, minor);
13685 
13686 	mutex_enter(&cpu_lock);
13687 	mutex_enter(&dtrace_lock);
13688 
13689 	if (state->dts_anon) {
13690 		/*
13691 		 * There is anonymous state. Destroy that first.
13692 		 */
13693 		ASSERT(dtrace_anon.dta_state == NULL);
13694 		dtrace_state_destroy(state->dts_anon);
13695 	}
13696 
13697 	dtrace_state_destroy(state);
13698 	ASSERT(dtrace_opens > 0);
13699 	if (--dtrace_opens == 0)
13700 		(void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
13701 
13702 	mutex_exit(&dtrace_lock);
13703 	mutex_exit(&cpu_lock);
13704 
13705 	return (0);
13706 }
13707 
13708 /*ARGSUSED*/
13709 static int
13710 dtrace_ioctl_helper(int cmd, intptr_t arg, int *rv)
13711 {
13712 	int rval;
13713 	dof_helper_t help, *dhp = NULL;
13714 
13715 	switch (cmd) {
13716 	case DTRACEHIOC_ADDDOF:
13717 		if (copyin((void *)arg, &help, sizeof (help)) != 0) {
13718 			dtrace_dof_error(NULL, "failed to copyin DOF helper");
13719 			return (EFAULT);
13720 		}
13721 
13722 		dhp = &help;
13723 		arg = (intptr_t)help.dofhp_dof;
13724 		/*FALLTHROUGH*/
13725 
13726 	case DTRACEHIOC_ADD: {
13727 		dof_hdr_t *dof = dtrace_dof_copyin(arg, &rval);
13728 
13729 		if (dof == NULL)
13730 			return (rval);
13731 
13732 		mutex_enter(&dtrace_lock);
13733 
13734 		/*
13735 		 * dtrace_helper_slurp() takes responsibility for the dof --
13736 		 * it may free it now or it may save it and free it later.
13737 		 */
13738 		if ((rval = dtrace_helper_slurp(dof, dhp)) != -1) {
13739 			*rv = rval;
13740 			rval = 0;
13741 		} else {
13742 			rval = EINVAL;
13743 		}
13744 
13745 		mutex_exit(&dtrace_lock);
13746 		return (rval);
13747 	}
13748 
13749 	case DTRACEHIOC_REMOVE: {
13750 		mutex_enter(&dtrace_lock);
13751 		rval = dtrace_helper_destroygen(arg);
13752 		mutex_exit(&dtrace_lock);
13753 
13754 		return (rval);
13755 	}
13756 
13757 	default:
13758 		break;
13759 	}
13760 
13761 	return (ENOTTY);
13762 }
13763 
13764 /*ARGSUSED*/
13765 static int
13766 dtrace_ioctl(dev_t dev, int cmd, intptr_t arg, int md, cred_t *cr, int *rv)
13767 {
13768 	minor_t minor = getminor(dev);
13769 	dtrace_state_t *state;
13770 	int rval;
13771 
13772 	if (minor == DTRACEMNRN_HELPER)
13773 		return (dtrace_ioctl_helper(cmd, arg, rv));
13774 
13775 	state = ddi_get_soft_state(dtrace_softstate, minor);
13776 
13777 	if (state->dts_anon) {
13778 		ASSERT(dtrace_anon.dta_state == NULL);
13779 		state = state->dts_anon;
13780 	}
13781 
13782 	switch (cmd) {
13783 	case DTRACEIOC_PROVIDER: {
13784 		dtrace_providerdesc_t pvd;
13785 		dtrace_provider_t *pvp;
13786 
13787 		if (copyin((void *)arg, &pvd, sizeof (pvd)) != 0)
13788 			return (EFAULT);
13789 
13790 		pvd.dtvd_name[DTRACE_PROVNAMELEN - 1] = '\0';
13791 		mutex_enter(&dtrace_provider_lock);
13792 
13793 		for (pvp = dtrace_provider; pvp != NULL; pvp = pvp->dtpv_next) {
13794 			if (strcmp(pvp->dtpv_name, pvd.dtvd_name) == 0)
13795 				break;
13796 		}
13797 
13798 		mutex_exit(&dtrace_provider_lock);
13799 
13800 		if (pvp == NULL)
13801 			return (ESRCH);
13802 
13803 		bcopy(&pvp->dtpv_priv, &pvd.dtvd_priv, sizeof (dtrace_ppriv_t));
13804 		bcopy(&pvp->dtpv_attr, &pvd.dtvd_attr, sizeof (dtrace_pattr_t));
13805 		if (copyout(&pvd, (void *)arg, sizeof (pvd)) != 0)
13806 			return (EFAULT);
13807 
13808 		return (0);
13809 	}
13810 
13811 	case DTRACEIOC_EPROBE: {
13812 		dtrace_eprobedesc_t epdesc;
13813 		dtrace_ecb_t *ecb;
13814 		dtrace_action_t *act;
13815 		void *buf;
13816 		size_t size;
13817 		uintptr_t dest;
13818 		int nrecs;
13819 
13820 		if (copyin((void *)arg, &epdesc, sizeof (epdesc)) != 0)
13821 			return (EFAULT);
13822 
13823 		mutex_enter(&dtrace_lock);
13824 
13825 		if ((ecb = dtrace_epid2ecb(state, epdesc.dtepd_epid)) == NULL) {
13826 			mutex_exit(&dtrace_lock);
13827 			return (EINVAL);
13828 		}
13829 
13830 		if (ecb->dte_probe == NULL) {
13831 			mutex_exit(&dtrace_lock);
13832 			return (EINVAL);
13833 		}
13834 
13835 		epdesc.dtepd_probeid = ecb->dte_probe->dtpr_id;
13836 		epdesc.dtepd_uarg = ecb->dte_uarg;
13837 		epdesc.dtepd_size = ecb->dte_size;
13838 
13839 		nrecs = epdesc.dtepd_nrecs;
13840 		epdesc.dtepd_nrecs = 0;
13841 		for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
13842 			if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
13843 				continue;
13844 
13845 			epdesc.dtepd_nrecs++;
13846 		}
13847 
13848 		/*
13849 		 * Now that we have the size, we need to allocate a temporary
13850 		 * buffer in which to store the complete description.  We need
13851 		 * the temporary buffer to be able to drop dtrace_lock()
13852 		 * across the copyout(), below.
13853 		 */
13854 		size = sizeof (dtrace_eprobedesc_t) +
13855 		    (epdesc.dtepd_nrecs * sizeof (dtrace_recdesc_t));
13856 
13857 		buf = kmem_alloc(size, KM_SLEEP);
13858 		dest = (uintptr_t)buf;
13859 
13860 		bcopy(&epdesc, (void *)dest, sizeof (epdesc));
13861 		dest += offsetof(dtrace_eprobedesc_t, dtepd_rec[0]);
13862 
13863 		for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
13864 			if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
13865 				continue;
13866 
13867 			if (nrecs-- == 0)
13868 				break;
13869 
13870 			bcopy(&act->dta_rec, (void *)dest,
13871 			    sizeof (dtrace_recdesc_t));
13872 			dest += sizeof (dtrace_recdesc_t);
13873 		}
13874 
13875 		mutex_exit(&dtrace_lock);
13876 
13877 		if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
13878 			kmem_free(buf, size);
13879 			return (EFAULT);
13880 		}
13881 
13882 		kmem_free(buf, size);
13883 		return (0);
13884 	}
13885 
13886 	case DTRACEIOC_AGGDESC: {
13887 		dtrace_aggdesc_t aggdesc;
13888 		dtrace_action_t *act;
13889 		dtrace_aggregation_t *agg;
13890 		int nrecs;
13891 		uint32_t offs;
13892 		dtrace_recdesc_t *lrec;
13893 		void *buf;
13894 		size_t size;
13895 		uintptr_t dest;
13896 
13897 		if (copyin((void *)arg, &aggdesc, sizeof (aggdesc)) != 0)
13898 			return (EFAULT);
13899 
13900 		mutex_enter(&dtrace_lock);
13901 
13902 		if ((agg = dtrace_aggid2agg(state, aggdesc.dtagd_id)) == NULL) {
13903 			mutex_exit(&dtrace_lock);
13904 			return (EINVAL);
13905 		}
13906 
13907 		aggdesc.dtagd_epid = agg->dtag_ecb->dte_epid;
13908 
13909 		nrecs = aggdesc.dtagd_nrecs;
13910 		aggdesc.dtagd_nrecs = 0;
13911 
13912 		offs = agg->dtag_base;
13913 		lrec = &agg->dtag_action.dta_rec;
13914 		aggdesc.dtagd_size = lrec->dtrd_offset + lrec->dtrd_size - offs;
13915 
13916 		for (act = agg->dtag_first; ; act = act->dta_next) {
13917 			ASSERT(act->dta_intuple ||
13918 			    DTRACEACT_ISAGG(act->dta_kind));
13919 
13920 			/*
13921 			 * If this action has a record size of zero, it
13922 			 * denotes an argument to the aggregating action.
13923 			 * Because the presence of this record doesn't (or
13924 			 * shouldn't) affect the way the data is interpreted,
13925 			 * we don't copy it out to save user-level the
13926 			 * confusion of dealing with a zero-length record.
13927 			 */
13928 			if (act->dta_rec.dtrd_size == 0) {
13929 				ASSERT(agg->dtag_hasarg);
13930 				continue;
13931 			}
13932 
13933 			aggdesc.dtagd_nrecs++;
13934 
13935 			if (act == &agg->dtag_action)
13936 				break;
13937 		}
13938 
13939 		/*
13940 		 * Now that we have the size, we need to allocate a temporary
13941 		 * buffer in which to store the complete description.  We need
13942 		 * the temporary buffer to be able to drop dtrace_lock()
13943 		 * across the copyout(), below.
13944 		 */
13945 		size = sizeof (dtrace_aggdesc_t) +
13946 		    (aggdesc.dtagd_nrecs * sizeof (dtrace_recdesc_t));
13947 
13948 		buf = kmem_alloc(size, KM_SLEEP);
13949 		dest = (uintptr_t)buf;
13950 
13951 		bcopy(&aggdesc, (void *)dest, sizeof (aggdesc));
13952 		dest += offsetof(dtrace_aggdesc_t, dtagd_rec[0]);
13953 
13954 		for (act = agg->dtag_first; ; act = act->dta_next) {
13955 			dtrace_recdesc_t rec = act->dta_rec;
13956 
13957 			/*
13958 			 * See the comment in the above loop for why we pass
13959 			 * over zero-length records.
13960 			 */
13961 			if (rec.dtrd_size == 0) {
13962 				ASSERT(agg->dtag_hasarg);
13963 				continue;
13964 			}
13965 
13966 			if (nrecs-- == 0)
13967 				break;
13968 
13969 			rec.dtrd_offset -= offs;
13970 			bcopy(&rec, (void *)dest, sizeof (rec));
13971 			dest += sizeof (dtrace_recdesc_t);
13972 
13973 			if (act == &agg->dtag_action)
13974 				break;
13975 		}
13976 
13977 		mutex_exit(&dtrace_lock);
13978 
13979 		if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
13980 			kmem_free(buf, size);
13981 			return (EFAULT);
13982 		}
13983 
13984 		kmem_free(buf, size);
13985 		return (0);
13986 	}
13987 
13988 	case DTRACEIOC_ENABLE: {
13989 		dof_hdr_t *dof;
13990 		dtrace_enabling_t *enab = NULL;
13991 		dtrace_vstate_t *vstate;
13992 		int err = 0;
13993 
13994 		*rv = 0;
13995 
13996 		/*
13997 		 * If a NULL argument has been passed, we take this as our
13998 		 * cue to reevaluate our enablings.
13999 		 */
14000 		if (arg == NULL) {
14001 			mutex_enter(&cpu_lock);
14002 			mutex_enter(&dtrace_lock);
14003 			err = dtrace_enabling_matchstate(state, rv);
14004 			mutex_exit(&dtrace_lock);
14005 			mutex_exit(&cpu_lock);
14006 
14007 			return (err);
14008 		}
14009 
14010 		if ((dof = dtrace_dof_copyin(arg, &rval)) == NULL)
14011 			return (rval);
14012 
14013 		mutex_enter(&cpu_lock);
14014 		mutex_enter(&dtrace_lock);
14015 		vstate = &state->dts_vstate;
14016 
14017 		if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
14018 			mutex_exit(&dtrace_lock);
14019 			mutex_exit(&cpu_lock);
14020 			dtrace_dof_destroy(dof);
14021 			return (EBUSY);
14022 		}
14023 
14024 		if (dtrace_dof_slurp(dof, vstate, cr, &enab, 0, B_TRUE) != 0) {
14025 			mutex_exit(&dtrace_lock);
14026 			mutex_exit(&cpu_lock);
14027 			dtrace_dof_destroy(dof);
14028 			return (EINVAL);
14029 		}
14030 
14031 		if ((rval = dtrace_dof_options(dof, state)) != 0) {
14032 			dtrace_enabling_destroy(enab);
14033 			mutex_exit(&dtrace_lock);
14034 			mutex_exit(&cpu_lock);
14035 			dtrace_dof_destroy(dof);
14036 			return (rval);
14037 		}
14038 
14039 		if ((err = dtrace_enabling_match(enab, rv)) == 0) {
14040 			err = dtrace_enabling_retain(enab);
14041 		} else {
14042 			dtrace_enabling_destroy(enab);
14043 		}
14044 
14045 		mutex_exit(&cpu_lock);
14046 		mutex_exit(&dtrace_lock);
14047 		dtrace_dof_destroy(dof);
14048 
14049 		return (err);
14050 	}
14051 
14052 	case DTRACEIOC_REPLICATE: {
14053 		dtrace_repldesc_t desc;
14054 		dtrace_probedesc_t *match = &desc.dtrpd_match;
14055 		dtrace_probedesc_t *create = &desc.dtrpd_create;
14056 		int err;
14057 
14058 		if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
14059 			return (EFAULT);
14060 
14061 		match->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
14062 		match->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
14063 		match->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
14064 		match->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
14065 
14066 		create->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
14067 		create->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
14068 		create->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
14069 		create->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
14070 
14071 		mutex_enter(&dtrace_lock);
14072 		err = dtrace_enabling_replicate(state, match, create);
14073 		mutex_exit(&dtrace_lock);
14074 
14075 		return (err);
14076 	}
14077 
14078 	case DTRACEIOC_PROBEMATCH:
14079 	case DTRACEIOC_PROBES: {
14080 		dtrace_probe_t *probe = NULL;
14081 		dtrace_probedesc_t desc;
14082 		dtrace_probekey_t pkey;
14083 		dtrace_id_t i;
14084 		int m = 0;
14085 		uint32_t priv;
14086 		uid_t uid;
14087 		zoneid_t zoneid;
14088 
14089 		if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
14090 			return (EFAULT);
14091 
14092 		desc.dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
14093 		desc.dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
14094 		desc.dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
14095 		desc.dtpd_name[DTRACE_NAMELEN - 1] = '\0';
14096 
14097 		/*
14098 		 * Before we attempt to match this probe, we want to give
14099 		 * all providers the opportunity to provide it.
14100 		 */
14101 		if (desc.dtpd_id == DTRACE_IDNONE) {
14102 			mutex_enter(&dtrace_provider_lock);
14103 			dtrace_probe_provide(&desc, NULL);
14104 			mutex_exit(&dtrace_provider_lock);
14105 			desc.dtpd_id++;
14106 		}
14107 
14108 		if (cmd == DTRACEIOC_PROBEMATCH)  {
14109 			dtrace_probekey(&desc, &pkey);
14110 			pkey.dtpk_id = DTRACE_IDNONE;
14111 		}
14112 
14113 		dtrace_cred2priv(cr, &priv, &uid, &zoneid);
14114 
14115 		mutex_enter(&dtrace_lock);
14116 
14117 		if (cmd == DTRACEIOC_PROBEMATCH) {
14118 			for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
14119 				if ((probe = dtrace_probes[i - 1]) != NULL &&
14120 				    (m = dtrace_match_probe(probe, &pkey,
14121 				    priv, uid, zoneid)) != 0)
14122 					break;
14123 			}
14124 
14125 			if (m < 0) {
14126 				mutex_exit(&dtrace_lock);
14127 				return (EINVAL);
14128 			}
14129 
14130 		} else {
14131 			for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
14132 				if ((probe = dtrace_probes[i - 1]) != NULL &&
14133 				    dtrace_match_priv(probe, priv, uid, zoneid))
14134 					break;
14135 			}
14136 		}
14137 
14138 		if (probe == NULL) {
14139 			mutex_exit(&dtrace_lock);
14140 			return (ESRCH);
14141 		}
14142 
14143 		dtrace_probe_description(probe, &desc);
14144 		mutex_exit(&dtrace_lock);
14145 
14146 		if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
14147 			return (EFAULT);
14148 
14149 		return (0);
14150 	}
14151 
14152 	case DTRACEIOC_PROBEARG: {
14153 		dtrace_argdesc_t desc;
14154 		dtrace_probe_t *probe;
14155 		dtrace_provider_t *prov;
14156 
14157 		if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
14158 			return (EFAULT);
14159 
14160 		if (desc.dtargd_id == DTRACE_IDNONE)
14161 			return (EINVAL);
14162 
14163 		if (desc.dtargd_ndx == DTRACE_ARGNONE)
14164 			return (EINVAL);
14165 
14166 		mutex_enter(&dtrace_provider_lock);
14167 		mutex_enter(&mod_lock);
14168 		mutex_enter(&dtrace_lock);
14169 
14170 		if (desc.dtargd_id > dtrace_nprobes) {
14171 			mutex_exit(&dtrace_lock);
14172 			mutex_exit(&mod_lock);
14173 			mutex_exit(&dtrace_provider_lock);
14174 			return (EINVAL);
14175 		}
14176 
14177 		if ((probe = dtrace_probes[desc.dtargd_id - 1]) == NULL) {
14178 			mutex_exit(&dtrace_lock);
14179 			mutex_exit(&mod_lock);
14180 			mutex_exit(&dtrace_provider_lock);
14181 			return (EINVAL);
14182 		}
14183 
14184 		mutex_exit(&dtrace_lock);
14185 
14186 		prov = probe->dtpr_provider;
14187 
14188 		if (prov->dtpv_pops.dtps_getargdesc == NULL) {
14189 			/*
14190 			 * There isn't any typed information for this probe.
14191 			 * Set the argument number to DTRACE_ARGNONE.
14192 			 */
14193 			desc.dtargd_ndx = DTRACE_ARGNONE;
14194 		} else {
14195 			desc.dtargd_native[0] = '\0';
14196 			desc.dtargd_xlate[0] = '\0';
14197 			desc.dtargd_mapping = desc.dtargd_ndx;
14198 
14199 			prov->dtpv_pops.dtps_getargdesc(prov->dtpv_arg,
14200 			    probe->dtpr_id, probe->dtpr_arg, &desc);
14201 		}
14202 
14203 		mutex_exit(&mod_lock);
14204 		mutex_exit(&dtrace_provider_lock);
14205 
14206 		if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
14207 			return (EFAULT);
14208 
14209 		return (0);
14210 	}
14211 
14212 	case DTRACEIOC_GO: {
14213 		processorid_t cpuid;
14214 		rval = dtrace_state_go(state, &cpuid);
14215 
14216 		if (rval != 0)
14217 			return (rval);
14218 
14219 		if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
14220 			return (EFAULT);
14221 
14222 		return (0);
14223 	}
14224 
14225 	case DTRACEIOC_STOP: {
14226 		processorid_t cpuid;
14227 
14228 		mutex_enter(&dtrace_lock);
14229 		rval = dtrace_state_stop(state, &cpuid);
14230 		mutex_exit(&dtrace_lock);
14231 
14232 		if (rval != 0)
14233 			return (rval);
14234 
14235 		if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
14236 			return (EFAULT);
14237 
14238 		return (0);
14239 	}
14240 
14241 	case DTRACEIOC_DOFGET: {
14242 		dof_hdr_t hdr, *dof;
14243 		uint64_t len;
14244 
14245 		if (copyin((void *)arg, &hdr, sizeof (hdr)) != 0)
14246 			return (EFAULT);
14247 
14248 		mutex_enter(&dtrace_lock);
14249 		dof = dtrace_dof_create(state);
14250 		mutex_exit(&dtrace_lock);
14251 
14252 		len = MIN(hdr.dofh_loadsz, dof->dofh_loadsz);
14253 		rval = copyout(dof, (void *)arg, len);
14254 		dtrace_dof_destroy(dof);
14255 
14256 		return (rval == 0 ? 0 : EFAULT);
14257 	}
14258 
14259 	case DTRACEIOC_AGGSNAP:
14260 	case DTRACEIOC_BUFSNAP: {
14261 		dtrace_bufdesc_t desc;
14262 		caddr_t cached;
14263 		dtrace_buffer_t *buf;
14264 
14265 		if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
14266 			return (EFAULT);
14267 
14268 		if (desc.dtbd_cpu < 0 || desc.dtbd_cpu >= NCPU)
14269 			return (EINVAL);
14270 
14271 		mutex_enter(&dtrace_lock);
14272 
14273 		if (cmd == DTRACEIOC_BUFSNAP) {
14274 			buf = &state->dts_buffer[desc.dtbd_cpu];
14275 		} else {
14276 			buf = &state->dts_aggbuffer[desc.dtbd_cpu];
14277 		}
14278 
14279 		if (buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL)) {
14280 			size_t sz = buf->dtb_offset;
14281 
14282 			if (state->dts_activity != DTRACE_ACTIVITY_STOPPED) {
14283 				mutex_exit(&dtrace_lock);
14284 				return (EBUSY);
14285 			}
14286 
14287 			/*
14288 			 * If this buffer has already been consumed, we're
14289 			 * going to indicate that there's nothing left here
14290 			 * to consume.
14291 			 */
14292 			if (buf->dtb_flags & DTRACEBUF_CONSUMED) {
14293 				mutex_exit(&dtrace_lock);
14294 
14295 				desc.dtbd_size = 0;
14296 				desc.dtbd_drops = 0;
14297 				desc.dtbd_errors = 0;
14298 				desc.dtbd_oldest = 0;
14299 				sz = sizeof (desc);
14300 
14301 				if (copyout(&desc, (void *)arg, sz) != 0)
14302 					return (EFAULT);
14303 
14304 				return (0);
14305 			}
14306 
14307 			/*
14308 			 * If this is a ring buffer that has wrapped, we want
14309 			 * to copy the whole thing out.
14310 			 */
14311 			if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
14312 				dtrace_buffer_polish(buf);
14313 				sz = buf->dtb_size;
14314 			}
14315 
14316 			if (copyout(buf->dtb_tomax, desc.dtbd_data, sz) != 0) {
14317 				mutex_exit(&dtrace_lock);
14318 				return (EFAULT);
14319 			}
14320 
14321 			desc.dtbd_size = sz;
14322 			desc.dtbd_drops = buf->dtb_drops;
14323 			desc.dtbd_errors = buf->dtb_errors;
14324 			desc.dtbd_oldest = buf->dtb_xamot_offset;
14325 
14326 			mutex_exit(&dtrace_lock);
14327 
14328 			if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
14329 				return (EFAULT);
14330 
14331 			buf->dtb_flags |= DTRACEBUF_CONSUMED;
14332 
14333 			return (0);
14334 		}
14335 
14336 		if (buf->dtb_tomax == NULL) {
14337 			ASSERT(buf->dtb_xamot == NULL);
14338 			mutex_exit(&dtrace_lock);
14339 			return (ENOENT);
14340 		}
14341 
14342 		cached = buf->dtb_tomax;
14343 		ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
14344 
14345 		dtrace_xcall(desc.dtbd_cpu,
14346 		    (dtrace_xcall_t)dtrace_buffer_switch, buf);
14347 
14348 		state->dts_errors += buf->dtb_xamot_errors;
14349 
14350 		/*
14351 		 * If the buffers did not actually switch, then the cross call
14352 		 * did not take place -- presumably because the given CPU is
14353 		 * not in the ready set.  If this is the case, we'll return
14354 		 * ENOENT.
14355 		 */
14356 		if (buf->dtb_tomax == cached) {
14357 			ASSERT(buf->dtb_xamot != cached);
14358 			mutex_exit(&dtrace_lock);
14359 			return (ENOENT);
14360 		}
14361 
14362 		ASSERT(cached == buf->dtb_xamot);
14363 
14364 		/*
14365 		 * We have our snapshot; now copy it out.
14366 		 */
14367 		if (copyout(buf->dtb_xamot, desc.dtbd_data,
14368 		    buf->dtb_xamot_offset) != 0) {
14369 			mutex_exit(&dtrace_lock);
14370 			return (EFAULT);
14371 		}
14372 
14373 		desc.dtbd_size = buf->dtb_xamot_offset;
14374 		desc.dtbd_drops = buf->dtb_xamot_drops;
14375 		desc.dtbd_errors = buf->dtb_xamot_errors;
14376 		desc.dtbd_oldest = 0;
14377 
14378 		mutex_exit(&dtrace_lock);
14379 
14380 		/*
14381 		 * Finally, copy out the buffer description.
14382 		 */
14383 		if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
14384 			return (EFAULT);
14385 
14386 		return (0);
14387 	}
14388 
14389 	case DTRACEIOC_CONF: {
14390 		dtrace_conf_t conf;
14391 
14392 		bzero(&conf, sizeof (conf));
14393 		conf.dtc_difversion = DIF_VERSION;
14394 		conf.dtc_difintregs = DIF_DIR_NREGS;
14395 		conf.dtc_diftupregs = DIF_DTR_NREGS;
14396 		conf.dtc_ctfmodel = CTF_MODEL_NATIVE;
14397 
14398 		if (copyout(&conf, (void *)arg, sizeof (conf)) != 0)
14399 			return (EFAULT);
14400 
14401 		return (0);
14402 	}
14403 
14404 	case DTRACEIOC_STATUS: {
14405 		dtrace_status_t stat;
14406 		dtrace_dstate_t *dstate;
14407 		int i, j;
14408 		uint64_t nerrs;
14409 
14410 		/*
14411 		 * See the comment in dtrace_state_deadman() for the reason
14412 		 * for setting dts_laststatus to INT64_MAX before setting
14413 		 * it to the correct value.
14414 		 */
14415 		state->dts_laststatus = INT64_MAX;
14416 		dtrace_membar_producer();
14417 		state->dts_laststatus = dtrace_gethrtime();
14418 
14419 		bzero(&stat, sizeof (stat));
14420 
14421 		mutex_enter(&dtrace_lock);
14422 
14423 		if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE) {
14424 			mutex_exit(&dtrace_lock);
14425 			return (ENOENT);
14426 		}
14427 
14428 		if (state->dts_activity == DTRACE_ACTIVITY_DRAINING)
14429 			stat.dtst_exiting = 1;
14430 
14431 		nerrs = state->dts_errors;
14432 		dstate = &state->dts_vstate.dtvs_dynvars;
14433 
14434 		for (i = 0; i < NCPU; i++) {
14435 			dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[i];
14436 
14437 			stat.dtst_dyndrops += dcpu->dtdsc_drops;
14438 			stat.dtst_dyndrops_dirty += dcpu->dtdsc_dirty_drops;
14439 			stat.dtst_dyndrops_rinsing += dcpu->dtdsc_rinsing_drops;
14440 
14441 			if (state->dts_buffer[i].dtb_flags & DTRACEBUF_FULL)
14442 				stat.dtst_filled++;
14443 
14444 			nerrs += state->dts_buffer[i].dtb_errors;
14445 
14446 			for (j = 0; j < state->dts_nspeculations; j++) {
14447 				dtrace_speculation_t *spec;
14448 				dtrace_buffer_t *buf;
14449 
14450 				spec = &state->dts_speculations[j];
14451 				buf = &spec->dtsp_buffer[i];
14452 				stat.dtst_specdrops += buf->dtb_xamot_drops;
14453 			}
14454 		}
14455 
14456 		stat.dtst_specdrops_busy = state->dts_speculations_busy;
14457 		stat.dtst_specdrops_unavail = state->dts_speculations_unavail;
14458 		stat.dtst_stkstroverflows = state->dts_stkstroverflows;
14459 		stat.dtst_dblerrors = state->dts_dblerrors;
14460 		stat.dtst_killed =
14461 		    (state->dts_activity == DTRACE_ACTIVITY_KILLED);
14462 		stat.dtst_errors = nerrs;
14463 
14464 		mutex_exit(&dtrace_lock);
14465 
14466 		if (copyout(&stat, (void *)arg, sizeof (stat)) != 0)
14467 			return (EFAULT);
14468 
14469 		return (0);
14470 	}
14471 
14472 	case DTRACEIOC_FORMAT: {
14473 		dtrace_fmtdesc_t fmt;
14474 		char *str;
14475 		int len;
14476 
14477 		if (copyin((void *)arg, &fmt, sizeof (fmt)) != 0)
14478 			return (EFAULT);
14479 
14480 		mutex_enter(&dtrace_lock);
14481 
14482 		if (fmt.dtfd_format == 0 ||
14483 		    fmt.dtfd_format > state->dts_nformats) {
14484 			mutex_exit(&dtrace_lock);
14485 			return (EINVAL);
14486 		}
14487 
14488 		/*
14489 		 * Format strings are allocated contiguously and they are
14490 		 * never freed; if a format index is less than the number
14491 		 * of formats, we can assert that the format map is non-NULL
14492 		 * and that the format for the specified index is non-NULL.
14493 		 */
14494 		ASSERT(state->dts_formats != NULL);
14495 		str = state->dts_formats[fmt.dtfd_format - 1];
14496 		ASSERT(str != NULL);
14497 
14498 		len = strlen(str) + 1;
14499 
14500 		if (len > fmt.dtfd_length) {
14501 			fmt.dtfd_length = len;
14502 
14503 			if (copyout(&fmt, (void *)arg, sizeof (fmt)) != 0) {
14504 				mutex_exit(&dtrace_lock);
14505 				return (EINVAL);
14506 			}
14507 		} else {
14508 			if (copyout(str, fmt.dtfd_string, len) != 0) {
14509 				mutex_exit(&dtrace_lock);
14510 				return (EINVAL);
14511 			}
14512 		}
14513 
14514 		mutex_exit(&dtrace_lock);
14515 		return (0);
14516 	}
14517 
14518 	default:
14519 		break;
14520 	}
14521 
14522 	return (ENOTTY);
14523 }
14524 
14525 /*ARGSUSED*/
14526 static int
14527 dtrace_detach(dev_info_t *dip, ddi_detach_cmd_t cmd)
14528 {
14529 	dtrace_state_t *state;
14530 
14531 	switch (cmd) {
14532 	case DDI_DETACH:
14533 		break;
14534 
14535 	case DDI_SUSPEND:
14536 		return (DDI_SUCCESS);
14537 
14538 	default:
14539 		return (DDI_FAILURE);
14540 	}
14541 
14542 	mutex_enter(&cpu_lock);
14543 	mutex_enter(&dtrace_provider_lock);
14544 	mutex_enter(&dtrace_lock);
14545 
14546 	ASSERT(dtrace_opens == 0);
14547 
14548 	if (dtrace_helpers > 0) {
14549 		mutex_exit(&dtrace_provider_lock);
14550 		mutex_exit(&dtrace_lock);
14551 		mutex_exit(&cpu_lock);
14552 		return (DDI_FAILURE);
14553 	}
14554 
14555 	if (dtrace_unregister((dtrace_provider_id_t)dtrace_provider) != 0) {
14556 		mutex_exit(&dtrace_provider_lock);
14557 		mutex_exit(&dtrace_lock);
14558 		mutex_exit(&cpu_lock);
14559 		return (DDI_FAILURE);
14560 	}
14561 
14562 	dtrace_provider = NULL;
14563 
14564 	if ((state = dtrace_anon_grab()) != NULL) {
14565 		/*
14566 		 * If there were ECBs on this state, the provider should
14567 		 * have not been allowed to detach; assert that there is
14568 		 * none.
14569 		 */
14570 		ASSERT(state->dts_necbs == 0);
14571 		dtrace_state_destroy(state);
14572 
14573 		/*
14574 		 * If we're being detached with anonymous state, we need to
14575 		 * indicate to the kernel debugger that DTrace is now inactive.
14576 		 */
14577 		(void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
14578 	}
14579 
14580 	bzero(&dtrace_anon, sizeof (dtrace_anon_t));
14581 	unregister_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
14582 	dtrace_cpu_init = NULL;
14583 	dtrace_helpers_cleanup = NULL;
14584 	dtrace_helpers_fork = NULL;
14585 	dtrace_cpustart_init = NULL;
14586 	dtrace_cpustart_fini = NULL;
14587 	dtrace_debugger_init = NULL;
14588 	dtrace_debugger_fini = NULL;
14589 	dtrace_kreloc_init = NULL;
14590 	dtrace_kreloc_fini = NULL;
14591 	dtrace_modload = NULL;
14592 	dtrace_modunload = NULL;
14593 
14594 	mutex_exit(&cpu_lock);
14595 
14596 	if (dtrace_helptrace_enabled) {
14597 		kmem_free(dtrace_helptrace_buffer, dtrace_helptrace_bufsize);
14598 		dtrace_helptrace_buffer = NULL;
14599 	}
14600 
14601 	kmem_free(dtrace_probes, dtrace_nprobes * sizeof (dtrace_probe_t *));
14602 	dtrace_probes = NULL;
14603 	dtrace_nprobes = 0;
14604 
14605 	dtrace_hash_destroy(dtrace_bymod);
14606 	dtrace_hash_destroy(dtrace_byfunc);
14607 	dtrace_hash_destroy(dtrace_byname);
14608 	dtrace_bymod = NULL;
14609 	dtrace_byfunc = NULL;
14610 	dtrace_byname = NULL;
14611 
14612 	kmem_cache_destroy(dtrace_state_cache);
14613 	vmem_destroy(dtrace_minor);
14614 	vmem_destroy(dtrace_arena);
14615 
14616 	if (dtrace_toxrange != NULL) {
14617 		kmem_free(dtrace_toxrange,
14618 		    dtrace_toxranges_max * sizeof (dtrace_toxrange_t));
14619 		dtrace_toxrange = NULL;
14620 		dtrace_toxranges = 0;
14621 		dtrace_toxranges_max = 0;
14622 	}
14623 
14624 	ddi_remove_minor_node(dtrace_devi, NULL);
14625 	dtrace_devi = NULL;
14626 
14627 	ddi_soft_state_fini(&dtrace_softstate);
14628 
14629 	ASSERT(dtrace_vtime_references == 0);
14630 	ASSERT(dtrace_opens == 0);
14631 	ASSERT(dtrace_retained == NULL);
14632 
14633 	mutex_exit(&dtrace_lock);
14634 	mutex_exit(&dtrace_provider_lock);
14635 
14636 	/*
14637 	 * We don't destroy the task queue until after we have dropped our
14638 	 * locks (taskq_destroy() may block on running tasks).  To prevent
14639 	 * attempting to do work after we have effectively detached but before
14640 	 * the task queue has been destroyed, all tasks dispatched via the
14641 	 * task queue must check that DTrace is still attached before
14642 	 * performing any operation.
14643 	 */
14644 	taskq_destroy(dtrace_taskq);
14645 	dtrace_taskq = NULL;
14646 
14647 	return (DDI_SUCCESS);
14648 }
14649 
14650 /*ARGSUSED*/
14651 static int
14652 dtrace_info(dev_info_t *dip, ddi_info_cmd_t infocmd, void *arg, void **result)
14653 {
14654 	int error;
14655 
14656 	switch (infocmd) {
14657 	case DDI_INFO_DEVT2DEVINFO:
14658 		*result = (void *)dtrace_devi;
14659 		error = DDI_SUCCESS;
14660 		break;
14661 	case DDI_INFO_DEVT2INSTANCE:
14662 		*result = (void *)0;
14663 		error = DDI_SUCCESS;
14664 		break;
14665 	default:
14666 		error = DDI_FAILURE;
14667 	}
14668 	return (error);
14669 }
14670 
14671 static struct cb_ops dtrace_cb_ops = {
14672 	dtrace_open,		/* open */
14673 	dtrace_close,		/* close */
14674 	nulldev,		/* strategy */
14675 	nulldev,		/* print */
14676 	nodev,			/* dump */
14677 	nodev,			/* read */
14678 	nodev,			/* write */
14679 	dtrace_ioctl,		/* ioctl */
14680 	nodev,			/* devmap */
14681 	nodev,			/* mmap */
14682 	nodev,			/* segmap */
14683 	nochpoll,		/* poll */
14684 	ddi_prop_op,		/* cb_prop_op */
14685 	0,			/* streamtab  */
14686 	D_NEW | D_MP		/* Driver compatibility flag */
14687 };
14688 
14689 static struct dev_ops dtrace_ops = {
14690 	DEVO_REV,		/* devo_rev */
14691 	0,			/* refcnt */
14692 	dtrace_info,		/* get_dev_info */
14693 	nulldev,		/* identify */
14694 	nulldev,		/* probe */
14695 	dtrace_attach,		/* attach */
14696 	dtrace_detach,		/* detach */
14697 	nodev,			/* reset */
14698 	&dtrace_cb_ops,		/* driver operations */
14699 	NULL,			/* bus operations */
14700 	nodev			/* dev power */
14701 };
14702 
14703 static struct modldrv modldrv = {
14704 	&mod_driverops,		/* module type (this is a pseudo driver) */
14705 	"Dynamic Tracing",	/* name of module */
14706 	&dtrace_ops,		/* driver ops */
14707 };
14708 
14709 static struct modlinkage modlinkage = {
14710 	MODREV_1,
14711 	(void *)&modldrv,
14712 	NULL
14713 };
14714 
14715 int
14716 _init(void)
14717 {
14718 	return (mod_install(&modlinkage));
14719 }
14720 
14721 int
14722 _info(struct modinfo *modinfop)
14723 {
14724 	return (mod_info(&modlinkage, modinfop));
14725 }
14726 
14727 int
14728 _fini(void)
14729 {
14730 	return (mod_remove(&modlinkage));
14731 }
14732