xref: /titanic_52/usr/src/uts/common/dtrace/dtrace.c (revision 694c35faa87b858ecdadfe4fc592615f4eefbb07)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
24  * Copyright (c) 2011, Joyent, Inc. All rights reserved.
25  * Copyright (c) 2012 by Delphix. All rights reserved.
26  */
27 
28 /*
29  * DTrace - Dynamic Tracing for Solaris
30  *
31  * This is the implementation of the Solaris Dynamic Tracing framework
32  * (DTrace).  The user-visible interface to DTrace is described at length in
33  * the "Solaris Dynamic Tracing Guide".  The interfaces between the libdtrace
34  * library, the in-kernel DTrace framework, and the DTrace providers are
35  * described in the block comments in the <sys/dtrace.h> header file.  The
36  * internal architecture of DTrace is described in the block comments in the
37  * <sys/dtrace_impl.h> header file.  The comments contained within the DTrace
38  * implementation very much assume mastery of all of these sources; if one has
39  * an unanswered question about the implementation, one should consult them
40  * first.
41  *
42  * The functions here are ordered roughly as follows:
43  *
44  *   - Probe context functions
45  *   - Probe hashing functions
46  *   - Non-probe context utility functions
47  *   - Matching functions
48  *   - Provider-to-Framework API functions
49  *   - Probe management functions
50  *   - DIF object functions
51  *   - Format functions
52  *   - Predicate functions
53  *   - ECB functions
54  *   - Buffer functions
55  *   - Enabling functions
56  *   - DOF functions
57  *   - Anonymous enabling functions
58  *   - Consumer state functions
59  *   - Helper functions
60  *   - Hook functions
61  *   - Driver cookbook functions
62  *
63  * Each group of functions begins with a block comment labelled the "DTrace
64  * [Group] Functions", allowing one to find each block by searching forward
65  * on capital-f functions.
66  */
67 #include <sys/errno.h>
68 #include <sys/stat.h>
69 #include <sys/modctl.h>
70 #include <sys/conf.h>
71 #include <sys/systm.h>
72 #include <sys/ddi.h>
73 #include <sys/sunddi.h>
74 #include <sys/cpuvar.h>
75 #include <sys/kmem.h>
76 #include <sys/strsubr.h>
77 #include <sys/sysmacros.h>
78 #include <sys/dtrace_impl.h>
79 #include <sys/atomic.h>
80 #include <sys/cmn_err.h>
81 #include <sys/mutex_impl.h>
82 #include <sys/rwlock_impl.h>
83 #include <sys/ctf_api.h>
84 #include <sys/panic.h>
85 #include <sys/priv_impl.h>
86 #include <sys/policy.h>
87 #include <sys/cred_impl.h>
88 #include <sys/procfs_isa.h>
89 #include <sys/taskq.h>
90 #include <sys/mkdev.h>
91 #include <sys/kdi.h>
92 #include <sys/zone.h>
93 #include <sys/socket.h>
94 #include <netinet/in.h>
95 
96 /*
97  * DTrace Tunable Variables
98  *
99  * The following variables may be tuned by adding a line to /etc/system that
100  * includes both the name of the DTrace module ("dtrace") and the name of the
101  * variable.  For example:
102  *
103  *   set dtrace:dtrace_destructive_disallow = 1
104  *
105  * In general, the only variables that one should be tuning this way are those
106  * that affect system-wide DTrace behavior, and for which the default behavior
107  * is undesirable.  Most of these variables are tunable on a per-consumer
108  * basis using DTrace options, and need not be tuned on a system-wide basis.
109  * When tuning these variables, avoid pathological values; while some attempt
110  * is made to verify the integrity of these variables, they are not considered
111  * part of the supported interface to DTrace, and they are therefore not
112  * checked comprehensively.  Further, these variables should not be tuned
113  * dynamically via "mdb -kw" or other means; they should only be tuned via
114  * /etc/system.
115  */
116 int		dtrace_destructive_disallow = 0;
117 dtrace_optval_t	dtrace_nonroot_maxsize = (16 * 1024 * 1024);
118 size_t		dtrace_difo_maxsize = (256 * 1024);
119 dtrace_optval_t	dtrace_dof_maxsize = (256 * 1024);
120 size_t		dtrace_global_maxsize = (16 * 1024);
121 size_t		dtrace_actions_max = (16 * 1024);
122 size_t		dtrace_retain_max = 1024;
123 dtrace_optval_t	dtrace_helper_actions_max = 1024;
124 dtrace_optval_t	dtrace_helper_providers_max = 32;
125 dtrace_optval_t	dtrace_dstate_defsize = (1 * 1024 * 1024);
126 size_t		dtrace_strsize_default = 256;
127 dtrace_optval_t	dtrace_cleanrate_default = 9900990;		/* 101 hz */
128 dtrace_optval_t	dtrace_cleanrate_min = 200000;			/* 5000 hz */
129 dtrace_optval_t	dtrace_cleanrate_max = (uint64_t)60 * NANOSEC;	/* 1/minute */
130 dtrace_optval_t	dtrace_aggrate_default = NANOSEC;		/* 1 hz */
131 dtrace_optval_t	dtrace_statusrate_default = NANOSEC;		/* 1 hz */
132 dtrace_optval_t dtrace_statusrate_max = (hrtime_t)10 * NANOSEC;	 /* 6/minute */
133 dtrace_optval_t	dtrace_switchrate_default = NANOSEC;		/* 1 hz */
134 dtrace_optval_t	dtrace_nspec_default = 1;
135 dtrace_optval_t	dtrace_specsize_default = 32 * 1024;
136 dtrace_optval_t dtrace_stackframes_default = 20;
137 dtrace_optval_t dtrace_ustackframes_default = 20;
138 dtrace_optval_t dtrace_jstackframes_default = 50;
139 dtrace_optval_t dtrace_jstackstrsize_default = 512;
140 int		dtrace_msgdsize_max = 128;
141 hrtime_t	dtrace_chill_max = 500 * (NANOSEC / MILLISEC);	/* 500 ms */
142 hrtime_t	dtrace_chill_interval = NANOSEC;		/* 1000 ms */
143 int		dtrace_devdepth_max = 32;
144 int		dtrace_err_verbose;
145 hrtime_t	dtrace_deadman_interval = NANOSEC;
146 hrtime_t	dtrace_deadman_timeout = (hrtime_t)10 * NANOSEC;
147 hrtime_t	dtrace_deadman_user = (hrtime_t)30 * NANOSEC;
148 hrtime_t	dtrace_unregister_defunct_reap = (hrtime_t)60 * NANOSEC;
149 
150 /*
151  * DTrace External Variables
152  *
153  * As dtrace(7D) is a kernel module, any DTrace variables are obviously
154  * available to DTrace consumers via the backtick (`) syntax.  One of these,
155  * dtrace_zero, is made deliberately so:  it is provided as a source of
156  * well-known, zero-filled memory.  While this variable is not documented,
157  * it is used by some translators as an implementation detail.
158  */
159 const char	dtrace_zero[256] = { 0 };	/* zero-filled memory */
160 
161 /*
162  * DTrace Internal Variables
163  */
164 static dev_info_t	*dtrace_devi;		/* device info */
165 static vmem_t		*dtrace_arena;		/* probe ID arena */
166 static vmem_t		*dtrace_minor;		/* minor number arena */
167 static taskq_t		*dtrace_taskq;		/* task queue */
168 static dtrace_probe_t	**dtrace_probes;	/* array of all probes */
169 static int		dtrace_nprobes;		/* number of probes */
170 static dtrace_provider_t *dtrace_provider;	/* provider list */
171 static dtrace_meta_t	*dtrace_meta_pid;	/* user-land meta provider */
172 static int		dtrace_opens;		/* number of opens */
173 static int		dtrace_helpers;		/* number of helpers */
174 static void		*dtrace_softstate;	/* softstate pointer */
175 static dtrace_hash_t	*dtrace_bymod;		/* probes hashed by module */
176 static dtrace_hash_t	*dtrace_byfunc;		/* probes hashed by function */
177 static dtrace_hash_t	*dtrace_byname;		/* probes hashed by name */
178 static dtrace_toxrange_t *dtrace_toxrange;	/* toxic range array */
179 static int		dtrace_toxranges;	/* number of toxic ranges */
180 static int		dtrace_toxranges_max;	/* size of toxic range array */
181 static dtrace_anon_t	dtrace_anon;		/* anonymous enabling */
182 static kmem_cache_t	*dtrace_state_cache;	/* cache for dynamic state */
183 static uint64_t		dtrace_vtime_references; /* number of vtimestamp refs */
184 static kthread_t	*dtrace_panicked;	/* panicking thread */
185 static dtrace_ecb_t	*dtrace_ecb_create_cache; /* cached created ECB */
186 static dtrace_genid_t	dtrace_probegen;	/* current probe generation */
187 static dtrace_helpers_t *dtrace_deferred_pid;	/* deferred helper list */
188 static dtrace_enabling_t *dtrace_retained;	/* list of retained enablings */
189 static dtrace_genid_t	dtrace_retained_gen;	/* current retained enab gen */
190 static dtrace_dynvar_t	dtrace_dynhash_sink;	/* end of dynamic hash chains */
191 static int		dtrace_dynvar_failclean; /* dynvars failed to clean */
192 
193 /*
194  * DTrace Locking
195  * DTrace is protected by three (relatively coarse-grained) locks:
196  *
197  * (1) dtrace_lock is required to manipulate essentially any DTrace state,
198  *     including enabling state, probes, ECBs, consumer state, helper state,
199  *     etc.  Importantly, dtrace_lock is _not_ required when in probe context;
200  *     probe context is lock-free -- synchronization is handled via the
201  *     dtrace_sync() cross call mechanism.
202  *
203  * (2) dtrace_provider_lock is required when manipulating provider state, or
204  *     when provider state must be held constant.
205  *
206  * (3) dtrace_meta_lock is required when manipulating meta provider state, or
207  *     when meta provider state must be held constant.
208  *
209  * The lock ordering between these three locks is dtrace_meta_lock before
210  * dtrace_provider_lock before dtrace_lock.  (In particular, there are
211  * several places where dtrace_provider_lock is held by the framework as it
212  * calls into the providers -- which then call back into the framework,
213  * grabbing dtrace_lock.)
214  *
215  * There are two other locks in the mix:  mod_lock and cpu_lock.  With respect
216  * to dtrace_provider_lock and dtrace_lock, cpu_lock continues its historical
217  * role as a coarse-grained lock; it is acquired before both of these locks.
218  * With respect to dtrace_meta_lock, its behavior is stranger:  cpu_lock must
219  * be acquired _between_ dtrace_meta_lock and any other DTrace locks.
220  * mod_lock is similar with respect to dtrace_provider_lock in that it must be
221  * acquired _between_ dtrace_provider_lock and dtrace_lock.
222  */
223 static kmutex_t		dtrace_lock;		/* probe state lock */
224 static kmutex_t		dtrace_provider_lock;	/* provider state lock */
225 static kmutex_t		dtrace_meta_lock;	/* meta-provider state lock */
226 
227 /*
228  * DTrace Provider Variables
229  *
230  * These are the variables relating to DTrace as a provider (that is, the
231  * provider of the BEGIN, END, and ERROR probes).
232  */
233 static dtrace_pattr_t	dtrace_provider_attr = {
234 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
235 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
236 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
237 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
238 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
239 };
240 
241 static void
242 dtrace_nullop(void)
243 {}
244 
245 static int
246 dtrace_enable_nullop(void)
247 {
248 	return (0);
249 }
250 
251 static dtrace_pops_t	dtrace_provider_ops = {
252 	(void (*)(void *, const dtrace_probedesc_t *))dtrace_nullop,
253 	(void (*)(void *, struct modctl *))dtrace_nullop,
254 	(int (*)(void *, dtrace_id_t, void *))dtrace_enable_nullop,
255 	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
256 	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
257 	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
258 	NULL,
259 	NULL,
260 	NULL,
261 	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop
262 };
263 
264 static dtrace_id_t	dtrace_probeid_begin;	/* special BEGIN probe */
265 static dtrace_id_t	dtrace_probeid_end;	/* special END probe */
266 dtrace_id_t		dtrace_probeid_error;	/* special ERROR probe */
267 
268 /*
269  * DTrace Helper Tracing Variables
270  */
271 uint32_t dtrace_helptrace_next = 0;
272 uint32_t dtrace_helptrace_nlocals;
273 char	*dtrace_helptrace_buffer;
274 int	dtrace_helptrace_bufsize = 512 * 1024;
275 
276 #ifdef DEBUG
277 int	dtrace_helptrace_enabled = 1;
278 #else
279 int	dtrace_helptrace_enabled = 0;
280 #endif
281 
282 /*
283  * DTrace Error Hashing
284  *
285  * On DEBUG kernels, DTrace will track the errors that has seen in a hash
286  * table.  This is very useful for checking coverage of tests that are
287  * expected to induce DIF or DOF processing errors, and may be useful for
288  * debugging problems in the DIF code generator or in DOF generation .  The
289  * error hash may be examined with the ::dtrace_errhash MDB dcmd.
290  */
291 #ifdef DEBUG
292 static dtrace_errhash_t	dtrace_errhash[DTRACE_ERRHASHSZ];
293 static const char *dtrace_errlast;
294 static kthread_t *dtrace_errthread;
295 static kmutex_t dtrace_errlock;
296 #endif
297 
298 /*
299  * DTrace Macros and Constants
300  *
301  * These are various macros that are useful in various spots in the
302  * implementation, along with a few random constants that have no meaning
303  * outside of the implementation.  There is no real structure to this cpp
304  * mishmash -- but is there ever?
305  */
306 #define	DTRACE_HASHSTR(hash, probe)	\
307 	dtrace_hash_str(*((char **)((uintptr_t)(probe) + (hash)->dth_stroffs)))
308 
309 #define	DTRACE_HASHNEXT(hash, probe)	\
310 	(dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_nextoffs)
311 
312 #define	DTRACE_HASHPREV(hash, probe)	\
313 	(dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_prevoffs)
314 
315 #define	DTRACE_HASHEQ(hash, lhs, rhs)	\
316 	(strcmp(*((char **)((uintptr_t)(lhs) + (hash)->dth_stroffs)), \
317 	    *((char **)((uintptr_t)(rhs) + (hash)->dth_stroffs))) == 0)
318 
319 #define	DTRACE_AGGHASHSIZE_SLEW		17
320 
321 #define	DTRACE_V4MAPPED_OFFSET		(sizeof (uint32_t) * 3)
322 
323 /*
324  * The key for a thread-local variable consists of the lower 61 bits of the
325  * t_did, plus the 3 bits of the highest active interrupt above LOCK_LEVEL.
326  * We add DIF_VARIABLE_MAX to t_did to assure that the thread key is never
327  * equal to a variable identifier.  This is necessary (but not sufficient) to
328  * assure that global associative arrays never collide with thread-local
329  * variables.  To guarantee that they cannot collide, we must also define the
330  * order for keying dynamic variables.  That order is:
331  *
332  *   [ key0 ] ... [ keyn ] [ variable-key ] [ tls-key ]
333  *
334  * Because the variable-key and the tls-key are in orthogonal spaces, there is
335  * no way for a global variable key signature to match a thread-local key
336  * signature.
337  */
338 #define	DTRACE_TLS_THRKEY(where) { \
339 	uint_t intr = 0; \
340 	uint_t actv = CPU->cpu_intr_actv >> (LOCK_LEVEL + 1); \
341 	for (; actv; actv >>= 1) \
342 		intr++; \
343 	ASSERT(intr < (1 << 3)); \
344 	(where) = ((curthread->t_did + DIF_VARIABLE_MAX) & \
345 	    (((uint64_t)1 << 61) - 1)) | ((uint64_t)intr << 61); \
346 }
347 
348 #define	DT_BSWAP_8(x)	((x) & 0xff)
349 #define	DT_BSWAP_16(x)	((DT_BSWAP_8(x) << 8) | DT_BSWAP_8((x) >> 8))
350 #define	DT_BSWAP_32(x)	((DT_BSWAP_16(x) << 16) | DT_BSWAP_16((x) >> 16))
351 #define	DT_BSWAP_64(x)	((DT_BSWAP_32(x) << 32) | DT_BSWAP_32((x) >> 32))
352 
353 #define	DT_MASK_LO 0x00000000FFFFFFFFULL
354 
355 #define	DTRACE_STORE(type, tomax, offset, what) \
356 	*((type *)((uintptr_t)(tomax) + (uintptr_t)offset)) = (type)(what);
357 
358 #ifndef __x86
359 #define	DTRACE_ALIGNCHECK(addr, size, flags)				\
360 	if (addr & (size - 1)) {					\
361 		*flags |= CPU_DTRACE_BADALIGN;				\
362 		cpu_core[CPU->cpu_id].cpuc_dtrace_illval = addr;	\
363 		return (0);						\
364 	}
365 #else
366 #define	DTRACE_ALIGNCHECK(addr, size, flags)
367 #endif
368 
369 /*
370  * Test whether a range of memory starting at testaddr of size testsz falls
371  * within the range of memory described by addr, sz.  We take care to avoid
372  * problems with overflow and underflow of the unsigned quantities, and
373  * disallow all negative sizes.  Ranges of size 0 are allowed.
374  */
375 #define	DTRACE_INRANGE(testaddr, testsz, baseaddr, basesz) \
376 	((testaddr) - (baseaddr) < (basesz) && \
377 	(testaddr) + (testsz) - (baseaddr) <= (basesz) && \
378 	(testaddr) + (testsz) >= (testaddr))
379 
380 /*
381  * Test whether alloc_sz bytes will fit in the scratch region.  We isolate
382  * alloc_sz on the righthand side of the comparison in order to avoid overflow
383  * or underflow in the comparison with it.  This is simpler than the INRANGE
384  * check above, because we know that the dtms_scratch_ptr is valid in the
385  * range.  Allocations of size zero are allowed.
386  */
387 #define	DTRACE_INSCRATCH(mstate, alloc_sz) \
388 	((mstate)->dtms_scratch_base + (mstate)->dtms_scratch_size - \
389 	(mstate)->dtms_scratch_ptr >= (alloc_sz))
390 
391 #define	DTRACE_LOADFUNC(bits)						\
392 /*CSTYLED*/								\
393 uint##bits##_t								\
394 dtrace_load##bits(uintptr_t addr)					\
395 {									\
396 	size_t size = bits / NBBY;					\
397 	/*CSTYLED*/							\
398 	uint##bits##_t rval;						\
399 	int i;								\
400 	volatile uint16_t *flags = (volatile uint16_t *)		\
401 	    &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;			\
402 									\
403 	DTRACE_ALIGNCHECK(addr, size, flags);				\
404 									\
405 	for (i = 0; i < dtrace_toxranges; i++) {			\
406 		if (addr >= dtrace_toxrange[i].dtt_limit)		\
407 			continue;					\
408 									\
409 		if (addr + size <= dtrace_toxrange[i].dtt_base)		\
410 			continue;					\
411 									\
412 		/*							\
413 		 * This address falls within a toxic region; return 0.	\
414 		 */							\
415 		*flags |= CPU_DTRACE_BADADDR;				\
416 		cpu_core[CPU->cpu_id].cpuc_dtrace_illval = addr;	\
417 		return (0);						\
418 	}								\
419 									\
420 	*flags |= CPU_DTRACE_NOFAULT;					\
421 	/*CSTYLED*/							\
422 	rval = *((volatile uint##bits##_t *)addr);			\
423 	*flags &= ~CPU_DTRACE_NOFAULT;					\
424 									\
425 	return (!(*flags & CPU_DTRACE_FAULT) ? rval : 0);		\
426 }
427 
428 #ifdef _LP64
429 #define	dtrace_loadptr	dtrace_load64
430 #else
431 #define	dtrace_loadptr	dtrace_load32
432 #endif
433 
434 #define	DTRACE_DYNHASH_FREE	0
435 #define	DTRACE_DYNHASH_SINK	1
436 #define	DTRACE_DYNHASH_VALID	2
437 
438 #define	DTRACE_MATCH_FAIL	-1
439 #define	DTRACE_MATCH_NEXT	0
440 #define	DTRACE_MATCH_DONE	1
441 #define	DTRACE_ANCHORED(probe)	((probe)->dtpr_func[0] != '\0')
442 #define	DTRACE_STATE_ALIGN	64
443 
444 #define	DTRACE_FLAGS2FLT(flags)						\
445 	(((flags) & CPU_DTRACE_BADADDR) ? DTRACEFLT_BADADDR :		\
446 	((flags) & CPU_DTRACE_ILLOP) ? DTRACEFLT_ILLOP :		\
447 	((flags) & CPU_DTRACE_DIVZERO) ? DTRACEFLT_DIVZERO :		\
448 	((flags) & CPU_DTRACE_KPRIV) ? DTRACEFLT_KPRIV :		\
449 	((flags) & CPU_DTRACE_UPRIV) ? DTRACEFLT_UPRIV :		\
450 	((flags) & CPU_DTRACE_TUPOFLOW) ?  DTRACEFLT_TUPOFLOW :		\
451 	((flags) & CPU_DTRACE_BADALIGN) ?  DTRACEFLT_BADALIGN :		\
452 	((flags) & CPU_DTRACE_NOSCRATCH) ?  DTRACEFLT_NOSCRATCH :	\
453 	((flags) & CPU_DTRACE_BADSTACK) ?  DTRACEFLT_BADSTACK :		\
454 	DTRACEFLT_UNKNOWN)
455 
456 #define	DTRACEACT_ISSTRING(act)						\
457 	((act)->dta_kind == DTRACEACT_DIFEXPR &&			\
458 	(act)->dta_difo->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING)
459 
460 static size_t dtrace_strlen(const char *, size_t);
461 static dtrace_probe_t *dtrace_probe_lookup_id(dtrace_id_t id);
462 static void dtrace_enabling_provide(dtrace_provider_t *);
463 static int dtrace_enabling_match(dtrace_enabling_t *, int *);
464 static void dtrace_enabling_matchall(void);
465 static void dtrace_enabling_reap(void);
466 static dtrace_state_t *dtrace_anon_grab(void);
467 static uint64_t dtrace_helper(int, dtrace_mstate_t *,
468     dtrace_state_t *, uint64_t, uint64_t);
469 static dtrace_helpers_t *dtrace_helpers_create(proc_t *);
470 static void dtrace_buffer_drop(dtrace_buffer_t *);
471 static int dtrace_buffer_consumed(dtrace_buffer_t *, hrtime_t when);
472 static intptr_t dtrace_buffer_reserve(dtrace_buffer_t *, size_t, size_t,
473     dtrace_state_t *, dtrace_mstate_t *);
474 static int dtrace_state_option(dtrace_state_t *, dtrace_optid_t,
475     dtrace_optval_t);
476 static int dtrace_ecb_create_enable(dtrace_probe_t *, void *);
477 static void dtrace_helper_provider_destroy(dtrace_helper_provider_t *);
478 
479 /*
480  * DTrace Probe Context Functions
481  *
482  * These functions are called from probe context.  Because probe context is
483  * any context in which C may be called, arbitrarily locks may be held,
484  * interrupts may be disabled, we may be in arbitrary dispatched state, etc.
485  * As a result, functions called from probe context may only call other DTrace
486  * support functions -- they may not interact at all with the system at large.
487  * (Note that the ASSERT macro is made probe-context safe by redefining it in
488  * terms of dtrace_assfail(), a probe-context safe function.) If arbitrary
489  * loads are to be performed from probe context, they _must_ be in terms of
490  * the safe dtrace_load*() variants.
491  *
492  * Some functions in this block are not actually called from probe context;
493  * for these functions, there will be a comment above the function reading
494  * "Note:  not called from probe context."
495  */
496 void
497 dtrace_panic(const char *format, ...)
498 {
499 	va_list alist;
500 
501 	va_start(alist, format);
502 	dtrace_vpanic(format, alist);
503 	va_end(alist);
504 }
505 
506 int
507 dtrace_assfail(const char *a, const char *f, int l)
508 {
509 	dtrace_panic("assertion failed: %s, file: %s, line: %d", a, f, l);
510 
511 	/*
512 	 * We just need something here that even the most clever compiler
513 	 * cannot optimize away.
514 	 */
515 	return (a[(uintptr_t)f]);
516 }
517 
518 /*
519  * Atomically increment a specified error counter from probe context.
520  */
521 static void
522 dtrace_error(uint32_t *counter)
523 {
524 	/*
525 	 * Most counters stored to in probe context are per-CPU counters.
526 	 * However, there are some error conditions that are sufficiently
527 	 * arcane that they don't merit per-CPU storage.  If these counters
528 	 * are incremented concurrently on different CPUs, scalability will be
529 	 * adversely affected -- but we don't expect them to be white-hot in a
530 	 * correctly constructed enabling...
531 	 */
532 	uint32_t oval, nval;
533 
534 	do {
535 		oval = *counter;
536 
537 		if ((nval = oval + 1) == 0) {
538 			/*
539 			 * If the counter would wrap, set it to 1 -- assuring
540 			 * that the counter is never zero when we have seen
541 			 * errors.  (The counter must be 32-bits because we
542 			 * aren't guaranteed a 64-bit compare&swap operation.)
543 			 * To save this code both the infamy of being fingered
544 			 * by a priggish news story and the indignity of being
545 			 * the target of a neo-puritan witch trial, we're
546 			 * carefully avoiding any colorful description of the
547 			 * likelihood of this condition -- but suffice it to
548 			 * say that it is only slightly more likely than the
549 			 * overflow of predicate cache IDs, as discussed in
550 			 * dtrace_predicate_create().
551 			 */
552 			nval = 1;
553 		}
554 	} while (dtrace_cas32(counter, oval, nval) != oval);
555 }
556 
557 /*
558  * Use the DTRACE_LOADFUNC macro to define functions for each of loading a
559  * uint8_t, a uint16_t, a uint32_t and a uint64_t.
560  */
561 DTRACE_LOADFUNC(8)
562 DTRACE_LOADFUNC(16)
563 DTRACE_LOADFUNC(32)
564 DTRACE_LOADFUNC(64)
565 
566 static int
567 dtrace_inscratch(uintptr_t dest, size_t size, dtrace_mstate_t *mstate)
568 {
569 	if (dest < mstate->dtms_scratch_base)
570 		return (0);
571 
572 	if (dest + size < dest)
573 		return (0);
574 
575 	if (dest + size > mstate->dtms_scratch_ptr)
576 		return (0);
577 
578 	return (1);
579 }
580 
581 static int
582 dtrace_canstore_statvar(uint64_t addr, size_t sz,
583     dtrace_statvar_t **svars, int nsvars)
584 {
585 	int i;
586 
587 	for (i = 0; i < nsvars; i++) {
588 		dtrace_statvar_t *svar = svars[i];
589 
590 		if (svar == NULL || svar->dtsv_size == 0)
591 			continue;
592 
593 		if (DTRACE_INRANGE(addr, sz, svar->dtsv_data, svar->dtsv_size))
594 			return (1);
595 	}
596 
597 	return (0);
598 }
599 
600 /*
601  * Check to see if the address is within a memory region to which a store may
602  * be issued.  This includes the DTrace scratch areas, and any DTrace variable
603  * region.  The caller of dtrace_canstore() is responsible for performing any
604  * alignment checks that are needed before stores are actually executed.
605  */
606 static int
607 dtrace_canstore(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
608     dtrace_vstate_t *vstate)
609 {
610 	/*
611 	 * First, check to see if the address is in scratch space...
612 	 */
613 	if (DTRACE_INRANGE(addr, sz, mstate->dtms_scratch_base,
614 	    mstate->dtms_scratch_size))
615 		return (1);
616 
617 	/*
618 	 * Now check to see if it's a dynamic variable.  This check will pick
619 	 * up both thread-local variables and any global dynamically-allocated
620 	 * variables.
621 	 */
622 	if (DTRACE_INRANGE(addr, sz, (uintptr_t)vstate->dtvs_dynvars.dtds_base,
623 	    vstate->dtvs_dynvars.dtds_size)) {
624 		dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
625 		uintptr_t base = (uintptr_t)dstate->dtds_base +
626 		    (dstate->dtds_hashsize * sizeof (dtrace_dynhash_t));
627 		uintptr_t chunkoffs;
628 
629 		/*
630 		 * Before we assume that we can store here, we need to make
631 		 * sure that it isn't in our metadata -- storing to our
632 		 * dynamic variable metadata would corrupt our state.  For
633 		 * the range to not include any dynamic variable metadata,
634 		 * it must:
635 		 *
636 		 *	(1) Start above the hash table that is at the base of
637 		 *	the dynamic variable space
638 		 *
639 		 *	(2) Have a starting chunk offset that is beyond the
640 		 *	dtrace_dynvar_t that is at the base of every chunk
641 		 *
642 		 *	(3) Not span a chunk boundary
643 		 *
644 		 */
645 		if (addr < base)
646 			return (0);
647 
648 		chunkoffs = (addr - base) % dstate->dtds_chunksize;
649 
650 		if (chunkoffs < sizeof (dtrace_dynvar_t))
651 			return (0);
652 
653 		if (chunkoffs + sz > dstate->dtds_chunksize)
654 			return (0);
655 
656 		return (1);
657 	}
658 
659 	/*
660 	 * Finally, check the static local and global variables.  These checks
661 	 * take the longest, so we perform them last.
662 	 */
663 	if (dtrace_canstore_statvar(addr, sz,
664 	    vstate->dtvs_locals, vstate->dtvs_nlocals))
665 		return (1);
666 
667 	if (dtrace_canstore_statvar(addr, sz,
668 	    vstate->dtvs_globals, vstate->dtvs_nglobals))
669 		return (1);
670 
671 	return (0);
672 }
673 
674 
675 /*
676  * Convenience routine to check to see if the address is within a memory
677  * region in which a load may be issued given the user's privilege level;
678  * if not, it sets the appropriate error flags and loads 'addr' into the
679  * illegal value slot.
680  *
681  * DTrace subroutines (DIF_SUBR_*) should use this helper to implement
682  * appropriate memory access protection.
683  */
684 static int
685 dtrace_canload(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
686     dtrace_vstate_t *vstate)
687 {
688 	volatile uintptr_t *illval = &cpu_core[CPU->cpu_id].cpuc_dtrace_illval;
689 
690 	/*
691 	 * If we hold the privilege to read from kernel memory, then
692 	 * everything is readable.
693 	 */
694 	if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
695 		return (1);
696 
697 	/*
698 	 * You can obviously read that which you can store.
699 	 */
700 	if (dtrace_canstore(addr, sz, mstate, vstate))
701 		return (1);
702 
703 	/*
704 	 * We're allowed to read from our own string table.
705 	 */
706 	if (DTRACE_INRANGE(addr, sz, (uintptr_t)mstate->dtms_difo->dtdo_strtab,
707 	    mstate->dtms_difo->dtdo_strlen))
708 		return (1);
709 
710 	DTRACE_CPUFLAG_SET(CPU_DTRACE_KPRIV);
711 	*illval = addr;
712 	return (0);
713 }
714 
715 /*
716  * Convenience routine to check to see if a given string is within a memory
717  * region in which a load may be issued given the user's privilege level;
718  * this exists so that we don't need to issue unnecessary dtrace_strlen()
719  * calls in the event that the user has all privileges.
720  */
721 static int
722 dtrace_strcanload(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
723     dtrace_vstate_t *vstate)
724 {
725 	size_t strsz;
726 
727 	/*
728 	 * If we hold the privilege to read from kernel memory, then
729 	 * everything is readable.
730 	 */
731 	if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
732 		return (1);
733 
734 	strsz = 1 + dtrace_strlen((char *)(uintptr_t)addr, sz);
735 	if (dtrace_canload(addr, strsz, mstate, vstate))
736 		return (1);
737 
738 	return (0);
739 }
740 
741 /*
742  * Convenience routine to check to see if a given variable is within a memory
743  * region in which a load may be issued given the user's privilege level.
744  */
745 static int
746 dtrace_vcanload(void *src, dtrace_diftype_t *type, dtrace_mstate_t *mstate,
747     dtrace_vstate_t *vstate)
748 {
749 	size_t sz;
750 	ASSERT(type->dtdt_flags & DIF_TF_BYREF);
751 
752 	/*
753 	 * If we hold the privilege to read from kernel memory, then
754 	 * everything is readable.
755 	 */
756 	if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
757 		return (1);
758 
759 	if (type->dtdt_kind == DIF_TYPE_STRING)
760 		sz = dtrace_strlen(src,
761 		    vstate->dtvs_state->dts_options[DTRACEOPT_STRSIZE]) + 1;
762 	else
763 		sz = type->dtdt_size;
764 
765 	return (dtrace_canload((uintptr_t)src, sz, mstate, vstate));
766 }
767 
768 /*
769  * Compare two strings using safe loads.
770  */
771 static int
772 dtrace_strncmp(char *s1, char *s2, size_t limit)
773 {
774 	uint8_t c1, c2;
775 	volatile uint16_t *flags;
776 
777 	if (s1 == s2 || limit == 0)
778 		return (0);
779 
780 	flags = (volatile uint16_t *)&cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
781 
782 	do {
783 		if (s1 == NULL) {
784 			c1 = '\0';
785 		} else {
786 			c1 = dtrace_load8((uintptr_t)s1++);
787 		}
788 
789 		if (s2 == NULL) {
790 			c2 = '\0';
791 		} else {
792 			c2 = dtrace_load8((uintptr_t)s2++);
793 		}
794 
795 		if (c1 != c2)
796 			return (c1 - c2);
797 	} while (--limit && c1 != '\0' && !(*flags & CPU_DTRACE_FAULT));
798 
799 	return (0);
800 }
801 
802 /*
803  * Compute strlen(s) for a string using safe memory accesses.  The additional
804  * len parameter is used to specify a maximum length to ensure completion.
805  */
806 static size_t
807 dtrace_strlen(const char *s, size_t lim)
808 {
809 	uint_t len;
810 
811 	for (len = 0; len != lim; len++) {
812 		if (dtrace_load8((uintptr_t)s++) == '\0')
813 			break;
814 	}
815 
816 	return (len);
817 }
818 
819 /*
820  * Check if an address falls within a toxic region.
821  */
822 static int
823 dtrace_istoxic(uintptr_t kaddr, size_t size)
824 {
825 	uintptr_t taddr, tsize;
826 	int i;
827 
828 	for (i = 0; i < dtrace_toxranges; i++) {
829 		taddr = dtrace_toxrange[i].dtt_base;
830 		tsize = dtrace_toxrange[i].dtt_limit - taddr;
831 
832 		if (kaddr - taddr < tsize) {
833 			DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
834 			cpu_core[CPU->cpu_id].cpuc_dtrace_illval = kaddr;
835 			return (1);
836 		}
837 
838 		if (taddr - kaddr < size) {
839 			DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
840 			cpu_core[CPU->cpu_id].cpuc_dtrace_illval = taddr;
841 			return (1);
842 		}
843 	}
844 
845 	return (0);
846 }
847 
848 /*
849  * Copy src to dst using safe memory accesses.  The src is assumed to be unsafe
850  * memory specified by the DIF program.  The dst is assumed to be safe memory
851  * that we can store to directly because it is managed by DTrace.  As with
852  * standard bcopy, overlapping copies are handled properly.
853  */
854 static void
855 dtrace_bcopy(const void *src, void *dst, size_t len)
856 {
857 	if (len != 0) {
858 		uint8_t *s1 = dst;
859 		const uint8_t *s2 = src;
860 
861 		if (s1 <= s2) {
862 			do {
863 				*s1++ = dtrace_load8((uintptr_t)s2++);
864 			} while (--len != 0);
865 		} else {
866 			s2 += len;
867 			s1 += len;
868 
869 			do {
870 				*--s1 = dtrace_load8((uintptr_t)--s2);
871 			} while (--len != 0);
872 		}
873 	}
874 }
875 
876 /*
877  * Copy src to dst using safe memory accesses, up to either the specified
878  * length, or the point that a nul byte is encountered.  The src is assumed to
879  * be unsafe memory specified by the DIF program.  The dst is assumed to be
880  * safe memory that we can store to directly because it is managed by DTrace.
881  * Unlike dtrace_bcopy(), overlapping regions are not handled.
882  */
883 static void
884 dtrace_strcpy(const void *src, void *dst, size_t len)
885 {
886 	if (len != 0) {
887 		uint8_t *s1 = dst, c;
888 		const uint8_t *s2 = src;
889 
890 		do {
891 			*s1++ = c = dtrace_load8((uintptr_t)s2++);
892 		} while (--len != 0 && c != '\0');
893 	}
894 }
895 
896 /*
897  * Copy src to dst, deriving the size and type from the specified (BYREF)
898  * variable type.  The src is assumed to be unsafe memory specified by the DIF
899  * program.  The dst is assumed to be DTrace variable memory that is of the
900  * specified type; we assume that we can store to directly.
901  */
902 static void
903 dtrace_vcopy(void *src, void *dst, dtrace_diftype_t *type)
904 {
905 	ASSERT(type->dtdt_flags & DIF_TF_BYREF);
906 
907 	if (type->dtdt_kind == DIF_TYPE_STRING) {
908 		dtrace_strcpy(src, dst, type->dtdt_size);
909 	} else {
910 		dtrace_bcopy(src, dst, type->dtdt_size);
911 	}
912 }
913 
914 /*
915  * Compare s1 to s2 using safe memory accesses.  The s1 data is assumed to be
916  * unsafe memory specified by the DIF program.  The s2 data is assumed to be
917  * safe memory that we can access directly because it is managed by DTrace.
918  */
919 static int
920 dtrace_bcmp(const void *s1, const void *s2, size_t len)
921 {
922 	volatile uint16_t *flags;
923 
924 	flags = (volatile uint16_t *)&cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
925 
926 	if (s1 == s2)
927 		return (0);
928 
929 	if (s1 == NULL || s2 == NULL)
930 		return (1);
931 
932 	if (s1 != s2 && len != 0) {
933 		const uint8_t *ps1 = s1;
934 		const uint8_t *ps2 = s2;
935 
936 		do {
937 			if (dtrace_load8((uintptr_t)ps1++) != *ps2++)
938 				return (1);
939 		} while (--len != 0 && !(*flags & CPU_DTRACE_FAULT));
940 	}
941 	return (0);
942 }
943 
944 /*
945  * Zero the specified region using a simple byte-by-byte loop.  Note that this
946  * is for safe DTrace-managed memory only.
947  */
948 static void
949 dtrace_bzero(void *dst, size_t len)
950 {
951 	uchar_t *cp;
952 
953 	for (cp = dst; len != 0; len--)
954 		*cp++ = 0;
955 }
956 
957 static void
958 dtrace_add_128(uint64_t *addend1, uint64_t *addend2, uint64_t *sum)
959 {
960 	uint64_t result[2];
961 
962 	result[0] = addend1[0] + addend2[0];
963 	result[1] = addend1[1] + addend2[1] +
964 	    (result[0] < addend1[0] || result[0] < addend2[0] ? 1 : 0);
965 
966 	sum[0] = result[0];
967 	sum[1] = result[1];
968 }
969 
970 /*
971  * Shift the 128-bit value in a by b. If b is positive, shift left.
972  * If b is negative, shift right.
973  */
974 static void
975 dtrace_shift_128(uint64_t *a, int b)
976 {
977 	uint64_t mask;
978 
979 	if (b == 0)
980 		return;
981 
982 	if (b < 0) {
983 		b = -b;
984 		if (b >= 64) {
985 			a[0] = a[1] >> (b - 64);
986 			a[1] = 0;
987 		} else {
988 			a[0] >>= b;
989 			mask = 1LL << (64 - b);
990 			mask -= 1;
991 			a[0] |= ((a[1] & mask) << (64 - b));
992 			a[1] >>= b;
993 		}
994 	} else {
995 		if (b >= 64) {
996 			a[1] = a[0] << (b - 64);
997 			a[0] = 0;
998 		} else {
999 			a[1] <<= b;
1000 			mask = a[0] >> (64 - b);
1001 			a[1] |= mask;
1002 			a[0] <<= b;
1003 		}
1004 	}
1005 }
1006 
1007 /*
1008  * The basic idea is to break the 2 64-bit values into 4 32-bit values,
1009  * use native multiplication on those, and then re-combine into the
1010  * resulting 128-bit value.
1011  *
1012  * (hi1 << 32 + lo1) * (hi2 << 32 + lo2) =
1013  *     hi1 * hi2 << 64 +
1014  *     hi1 * lo2 << 32 +
1015  *     hi2 * lo1 << 32 +
1016  *     lo1 * lo2
1017  */
1018 static void
1019 dtrace_multiply_128(uint64_t factor1, uint64_t factor2, uint64_t *product)
1020 {
1021 	uint64_t hi1, hi2, lo1, lo2;
1022 	uint64_t tmp[2];
1023 
1024 	hi1 = factor1 >> 32;
1025 	hi2 = factor2 >> 32;
1026 
1027 	lo1 = factor1 & DT_MASK_LO;
1028 	lo2 = factor2 & DT_MASK_LO;
1029 
1030 	product[0] = lo1 * lo2;
1031 	product[1] = hi1 * hi2;
1032 
1033 	tmp[0] = hi1 * lo2;
1034 	tmp[1] = 0;
1035 	dtrace_shift_128(tmp, 32);
1036 	dtrace_add_128(product, tmp, product);
1037 
1038 	tmp[0] = hi2 * lo1;
1039 	tmp[1] = 0;
1040 	dtrace_shift_128(tmp, 32);
1041 	dtrace_add_128(product, tmp, product);
1042 }
1043 
1044 /*
1045  * This privilege check should be used by actions and subroutines to
1046  * verify that the user credentials of the process that enabled the
1047  * invoking ECB match the target credentials
1048  */
1049 static int
1050 dtrace_priv_proc_common_user(dtrace_state_t *state)
1051 {
1052 	cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1053 
1054 	/*
1055 	 * We should always have a non-NULL state cred here, since if cred
1056 	 * is null (anonymous tracing), we fast-path bypass this routine.
1057 	 */
1058 	ASSERT(s_cr != NULL);
1059 
1060 	if ((cr = CRED()) != NULL &&
1061 	    s_cr->cr_uid == cr->cr_uid &&
1062 	    s_cr->cr_uid == cr->cr_ruid &&
1063 	    s_cr->cr_uid == cr->cr_suid &&
1064 	    s_cr->cr_gid == cr->cr_gid &&
1065 	    s_cr->cr_gid == cr->cr_rgid &&
1066 	    s_cr->cr_gid == cr->cr_sgid)
1067 		return (1);
1068 
1069 	return (0);
1070 }
1071 
1072 /*
1073  * This privilege check should be used by actions and subroutines to
1074  * verify that the zone of the process that enabled the invoking ECB
1075  * matches the target credentials
1076  */
1077 static int
1078 dtrace_priv_proc_common_zone(dtrace_state_t *state)
1079 {
1080 	cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1081 
1082 	/*
1083 	 * We should always have a non-NULL state cred here, since if cred
1084 	 * is null (anonymous tracing), we fast-path bypass this routine.
1085 	 */
1086 	ASSERT(s_cr != NULL);
1087 
1088 	if ((cr = CRED()) != NULL &&
1089 	    s_cr->cr_zone == cr->cr_zone)
1090 		return (1);
1091 
1092 	return (0);
1093 }
1094 
1095 /*
1096  * This privilege check should be used by actions and subroutines to
1097  * verify that the process has not setuid or changed credentials.
1098  */
1099 static int
1100 dtrace_priv_proc_common_nocd()
1101 {
1102 	proc_t *proc;
1103 
1104 	if ((proc = ttoproc(curthread)) != NULL &&
1105 	    !(proc->p_flag & SNOCD))
1106 		return (1);
1107 
1108 	return (0);
1109 }
1110 
1111 static int
1112 dtrace_priv_proc_destructive(dtrace_state_t *state, dtrace_mstate_t *mstate)
1113 {
1114 	int action = state->dts_cred.dcr_action;
1115 
1116 	if (!(mstate->dtms_access & DTRACE_ACCESS_PROC))
1117 		goto bad;
1118 
1119 	if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE) == 0) &&
1120 	    dtrace_priv_proc_common_zone(state) == 0)
1121 		goto bad;
1122 
1123 	if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER) == 0) &&
1124 	    dtrace_priv_proc_common_user(state) == 0)
1125 		goto bad;
1126 
1127 	if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG) == 0) &&
1128 	    dtrace_priv_proc_common_nocd() == 0)
1129 		goto bad;
1130 
1131 	return (1);
1132 
1133 bad:
1134 	cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1135 
1136 	return (0);
1137 }
1138 
1139 static int
1140 dtrace_priv_proc_control(dtrace_state_t *state, dtrace_mstate_t *mstate)
1141 {
1142 	if (mstate->dtms_access & DTRACE_ACCESS_PROC) {
1143 		if (state->dts_cred.dcr_action & DTRACE_CRA_PROC_CONTROL)
1144 			return (1);
1145 
1146 		if (dtrace_priv_proc_common_zone(state) &&
1147 		    dtrace_priv_proc_common_user(state) &&
1148 		    dtrace_priv_proc_common_nocd())
1149 			return (1);
1150 	}
1151 
1152 	cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1153 
1154 	return (0);
1155 }
1156 
1157 static int
1158 dtrace_priv_proc(dtrace_state_t *state, dtrace_mstate_t *mstate)
1159 {
1160 	if ((mstate->dtms_access & DTRACE_ACCESS_PROC) &&
1161 	    (state->dts_cred.dcr_action & DTRACE_CRA_PROC))
1162 		return (1);
1163 
1164 	cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1165 
1166 	return (0);
1167 }
1168 
1169 static int
1170 dtrace_priv_kernel(dtrace_state_t *state)
1171 {
1172 	if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL)
1173 		return (1);
1174 
1175 	cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1176 
1177 	return (0);
1178 }
1179 
1180 static int
1181 dtrace_priv_kernel_destructive(dtrace_state_t *state)
1182 {
1183 	if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL_DESTRUCTIVE)
1184 		return (1);
1185 
1186 	cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1187 
1188 	return (0);
1189 }
1190 
1191 /*
1192  * Determine if the dte_cond of the specified ECB allows for processing of
1193  * the current probe to continue.  Note that this routine may allow continued
1194  * processing, but with access(es) stripped from the mstate's dtms_access
1195  * field.
1196  */
1197 static int
1198 dtrace_priv_probe(dtrace_state_t *state, dtrace_mstate_t *mstate,
1199     dtrace_ecb_t *ecb)
1200 {
1201 	dtrace_probe_t *probe = ecb->dte_probe;
1202 	dtrace_provider_t *prov = probe->dtpr_provider;
1203 	dtrace_pops_t *pops = &prov->dtpv_pops;
1204 	int mode = DTRACE_MODE_NOPRIV_DROP;
1205 
1206 	ASSERT(ecb->dte_cond);
1207 
1208 	if (pops->dtps_mode != NULL) {
1209 		mode = pops->dtps_mode(prov->dtpv_arg,
1210 		    probe->dtpr_id, probe->dtpr_arg);
1211 
1212 		ASSERT((mode & DTRACE_MODE_USER) ||
1213 		    (mode & DTRACE_MODE_KERNEL));
1214 		ASSERT((mode & DTRACE_MODE_NOPRIV_RESTRICT) ||
1215 		    (mode & DTRACE_MODE_NOPRIV_DROP));
1216 	}
1217 
1218 	/*
1219 	 * If the dte_cond bits indicate that this consumer is only allowed to
1220 	 * see user-mode firings of this probe, call the provider's dtps_mode()
1221 	 * entry point to check that the probe was fired while in a user
1222 	 * context.  If that's not the case, use the policy specified by the
1223 	 * provider to determine if we drop the probe or merely restrict
1224 	 * operation.
1225 	 */
1226 	if (ecb->dte_cond & DTRACE_COND_USERMODE) {
1227 		ASSERT(mode != DTRACE_MODE_NOPRIV_DROP);
1228 
1229 		if (!(mode & DTRACE_MODE_USER)) {
1230 			if (mode & DTRACE_MODE_NOPRIV_DROP)
1231 				return (0);
1232 
1233 			mstate->dtms_access &= ~DTRACE_ACCESS_ARGS;
1234 		}
1235 	}
1236 
1237 	/*
1238 	 * This is more subtle than it looks. We have to be absolutely certain
1239 	 * that CRED() isn't going to change out from under us so it's only
1240 	 * legit to examine that structure if we're in constrained situations.
1241 	 * Currently, the only times we'll this check is if a non-super-user
1242 	 * has enabled the profile or syscall providers -- providers that
1243 	 * allow visibility of all processes. For the profile case, the check
1244 	 * above will ensure that we're examining a user context.
1245 	 */
1246 	if (ecb->dte_cond & DTRACE_COND_OWNER) {
1247 		cred_t *cr;
1248 		cred_t *s_cr = state->dts_cred.dcr_cred;
1249 		proc_t *proc;
1250 
1251 		ASSERT(s_cr != NULL);
1252 
1253 		if ((cr = CRED()) == NULL ||
1254 		    s_cr->cr_uid != cr->cr_uid ||
1255 		    s_cr->cr_uid != cr->cr_ruid ||
1256 		    s_cr->cr_uid != cr->cr_suid ||
1257 		    s_cr->cr_gid != cr->cr_gid ||
1258 		    s_cr->cr_gid != cr->cr_rgid ||
1259 		    s_cr->cr_gid != cr->cr_sgid ||
1260 		    (proc = ttoproc(curthread)) == NULL ||
1261 		    (proc->p_flag & SNOCD)) {
1262 			if (mode & DTRACE_MODE_NOPRIV_DROP)
1263 				return (0);
1264 
1265 			mstate->dtms_access &= ~DTRACE_ACCESS_PROC;
1266 		}
1267 	}
1268 
1269 	/*
1270 	 * If our dte_cond is set to DTRACE_COND_ZONEOWNER and we are not
1271 	 * in our zone, check to see if our mode policy is to restrict rather
1272 	 * than to drop; if to restrict, strip away both DTRACE_ACCESS_PROC
1273 	 * and DTRACE_ACCESS_ARGS
1274 	 */
1275 	if (ecb->dte_cond & DTRACE_COND_ZONEOWNER) {
1276 		cred_t *cr;
1277 		cred_t *s_cr = state->dts_cred.dcr_cred;
1278 
1279 		ASSERT(s_cr != NULL);
1280 
1281 		if ((cr = CRED()) == NULL ||
1282 		    s_cr->cr_zone->zone_id != cr->cr_zone->zone_id) {
1283 			if (mode & DTRACE_MODE_NOPRIV_DROP)
1284 				return (0);
1285 
1286 			mstate->dtms_access &=
1287 			    ~(DTRACE_ACCESS_PROC | DTRACE_ACCESS_ARGS);
1288 		}
1289 	}
1290 
1291 	return (1);
1292 }
1293 
1294 /*
1295  * Note:  not called from probe context.  This function is called
1296  * asynchronously (and at a regular interval) from outside of probe context to
1297  * clean the dirty dynamic variable lists on all CPUs.  Dynamic variable
1298  * cleaning is explained in detail in <sys/dtrace_impl.h>.
1299  */
1300 void
1301 dtrace_dynvar_clean(dtrace_dstate_t *dstate)
1302 {
1303 	dtrace_dynvar_t *dirty;
1304 	dtrace_dstate_percpu_t *dcpu;
1305 	dtrace_dynvar_t **rinsep;
1306 	int i, j, work = 0;
1307 
1308 	for (i = 0; i < NCPU; i++) {
1309 		dcpu = &dstate->dtds_percpu[i];
1310 		rinsep = &dcpu->dtdsc_rinsing;
1311 
1312 		/*
1313 		 * If the dirty list is NULL, there is no dirty work to do.
1314 		 */
1315 		if (dcpu->dtdsc_dirty == NULL)
1316 			continue;
1317 
1318 		if (dcpu->dtdsc_rinsing != NULL) {
1319 			/*
1320 			 * If the rinsing list is non-NULL, then it is because
1321 			 * this CPU was selected to accept another CPU's
1322 			 * dirty list -- and since that time, dirty buffers
1323 			 * have accumulated.  This is a highly unlikely
1324 			 * condition, but we choose to ignore the dirty
1325 			 * buffers -- they'll be picked up a future cleanse.
1326 			 */
1327 			continue;
1328 		}
1329 
1330 		if (dcpu->dtdsc_clean != NULL) {
1331 			/*
1332 			 * If the clean list is non-NULL, then we're in a
1333 			 * situation where a CPU has done deallocations (we
1334 			 * have a non-NULL dirty list) but no allocations (we
1335 			 * also have a non-NULL clean list).  We can't simply
1336 			 * move the dirty list into the clean list on this
1337 			 * CPU, yet we also don't want to allow this condition
1338 			 * to persist, lest a short clean list prevent a
1339 			 * massive dirty list from being cleaned (which in
1340 			 * turn could lead to otherwise avoidable dynamic
1341 			 * drops).  To deal with this, we look for some CPU
1342 			 * with a NULL clean list, NULL dirty list, and NULL
1343 			 * rinsing list -- and then we borrow this CPU to
1344 			 * rinse our dirty list.
1345 			 */
1346 			for (j = 0; j < NCPU; j++) {
1347 				dtrace_dstate_percpu_t *rinser;
1348 
1349 				rinser = &dstate->dtds_percpu[j];
1350 
1351 				if (rinser->dtdsc_rinsing != NULL)
1352 					continue;
1353 
1354 				if (rinser->dtdsc_dirty != NULL)
1355 					continue;
1356 
1357 				if (rinser->dtdsc_clean != NULL)
1358 					continue;
1359 
1360 				rinsep = &rinser->dtdsc_rinsing;
1361 				break;
1362 			}
1363 
1364 			if (j == NCPU) {
1365 				/*
1366 				 * We were unable to find another CPU that
1367 				 * could accept this dirty list -- we are
1368 				 * therefore unable to clean it now.
1369 				 */
1370 				dtrace_dynvar_failclean++;
1371 				continue;
1372 			}
1373 		}
1374 
1375 		work = 1;
1376 
1377 		/*
1378 		 * Atomically move the dirty list aside.
1379 		 */
1380 		do {
1381 			dirty = dcpu->dtdsc_dirty;
1382 
1383 			/*
1384 			 * Before we zap the dirty list, set the rinsing list.
1385 			 * (This allows for a potential assertion in
1386 			 * dtrace_dynvar():  if a free dynamic variable appears
1387 			 * on a hash chain, either the dirty list or the
1388 			 * rinsing list for some CPU must be non-NULL.)
1389 			 */
1390 			*rinsep = dirty;
1391 			dtrace_membar_producer();
1392 		} while (dtrace_casptr(&dcpu->dtdsc_dirty,
1393 		    dirty, NULL) != dirty);
1394 	}
1395 
1396 	if (!work) {
1397 		/*
1398 		 * We have no work to do; we can simply return.
1399 		 */
1400 		return;
1401 	}
1402 
1403 	dtrace_sync();
1404 
1405 	for (i = 0; i < NCPU; i++) {
1406 		dcpu = &dstate->dtds_percpu[i];
1407 
1408 		if (dcpu->dtdsc_rinsing == NULL)
1409 			continue;
1410 
1411 		/*
1412 		 * We are now guaranteed that no hash chain contains a pointer
1413 		 * into this dirty list; we can make it clean.
1414 		 */
1415 		ASSERT(dcpu->dtdsc_clean == NULL);
1416 		dcpu->dtdsc_clean = dcpu->dtdsc_rinsing;
1417 		dcpu->dtdsc_rinsing = NULL;
1418 	}
1419 
1420 	/*
1421 	 * Before we actually set the state to be DTRACE_DSTATE_CLEAN, make
1422 	 * sure that all CPUs have seen all of the dtdsc_clean pointers.
1423 	 * This prevents a race whereby a CPU incorrectly decides that
1424 	 * the state should be something other than DTRACE_DSTATE_CLEAN
1425 	 * after dtrace_dynvar_clean() has completed.
1426 	 */
1427 	dtrace_sync();
1428 
1429 	dstate->dtds_state = DTRACE_DSTATE_CLEAN;
1430 }
1431 
1432 /*
1433  * Depending on the value of the op parameter, this function looks-up,
1434  * allocates or deallocates an arbitrarily-keyed dynamic variable.  If an
1435  * allocation is requested, this function will return a pointer to a
1436  * dtrace_dynvar_t corresponding to the allocated variable -- or NULL if no
1437  * variable can be allocated.  If NULL is returned, the appropriate counter
1438  * will be incremented.
1439  */
1440 dtrace_dynvar_t *
1441 dtrace_dynvar(dtrace_dstate_t *dstate, uint_t nkeys,
1442     dtrace_key_t *key, size_t dsize, dtrace_dynvar_op_t op,
1443     dtrace_mstate_t *mstate, dtrace_vstate_t *vstate)
1444 {
1445 	uint64_t hashval = DTRACE_DYNHASH_VALID;
1446 	dtrace_dynhash_t *hash = dstate->dtds_hash;
1447 	dtrace_dynvar_t *free, *new_free, *next, *dvar, *start, *prev = NULL;
1448 	processorid_t me = CPU->cpu_id, cpu = me;
1449 	dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[me];
1450 	size_t bucket, ksize;
1451 	size_t chunksize = dstate->dtds_chunksize;
1452 	uintptr_t kdata, lock, nstate;
1453 	uint_t i;
1454 
1455 	ASSERT(nkeys != 0);
1456 
1457 	/*
1458 	 * Hash the key.  As with aggregations, we use Jenkins' "One-at-a-time"
1459 	 * algorithm.  For the by-value portions, we perform the algorithm in
1460 	 * 16-bit chunks (as opposed to 8-bit chunks).  This speeds things up a
1461 	 * bit, and seems to have only a minute effect on distribution.  For
1462 	 * the by-reference data, we perform "One-at-a-time" iterating (safely)
1463 	 * over each referenced byte.  It's painful to do this, but it's much
1464 	 * better than pathological hash distribution.  The efficacy of the
1465 	 * hashing algorithm (and a comparison with other algorithms) may be
1466 	 * found by running the ::dtrace_dynstat MDB dcmd.
1467 	 */
1468 	for (i = 0; i < nkeys; i++) {
1469 		if (key[i].dttk_size == 0) {
1470 			uint64_t val = key[i].dttk_value;
1471 
1472 			hashval += (val >> 48) & 0xffff;
1473 			hashval += (hashval << 10);
1474 			hashval ^= (hashval >> 6);
1475 
1476 			hashval += (val >> 32) & 0xffff;
1477 			hashval += (hashval << 10);
1478 			hashval ^= (hashval >> 6);
1479 
1480 			hashval += (val >> 16) & 0xffff;
1481 			hashval += (hashval << 10);
1482 			hashval ^= (hashval >> 6);
1483 
1484 			hashval += val & 0xffff;
1485 			hashval += (hashval << 10);
1486 			hashval ^= (hashval >> 6);
1487 		} else {
1488 			/*
1489 			 * This is incredibly painful, but it beats the hell
1490 			 * out of the alternative.
1491 			 */
1492 			uint64_t j, size = key[i].dttk_size;
1493 			uintptr_t base = (uintptr_t)key[i].dttk_value;
1494 
1495 			if (!dtrace_canload(base, size, mstate, vstate))
1496 				break;
1497 
1498 			for (j = 0; j < size; j++) {
1499 				hashval += dtrace_load8(base + j);
1500 				hashval += (hashval << 10);
1501 				hashval ^= (hashval >> 6);
1502 			}
1503 		}
1504 	}
1505 
1506 	if (DTRACE_CPUFLAG_ISSET(CPU_DTRACE_FAULT))
1507 		return (NULL);
1508 
1509 	hashval += (hashval << 3);
1510 	hashval ^= (hashval >> 11);
1511 	hashval += (hashval << 15);
1512 
1513 	/*
1514 	 * There is a remote chance (ideally, 1 in 2^31) that our hashval
1515 	 * comes out to be one of our two sentinel hash values.  If this
1516 	 * actually happens, we set the hashval to be a value known to be a
1517 	 * non-sentinel value.
1518 	 */
1519 	if (hashval == DTRACE_DYNHASH_FREE || hashval == DTRACE_DYNHASH_SINK)
1520 		hashval = DTRACE_DYNHASH_VALID;
1521 
1522 	/*
1523 	 * Yes, it's painful to do a divide here.  If the cycle count becomes
1524 	 * important here, tricks can be pulled to reduce it.  (However, it's
1525 	 * critical that hash collisions be kept to an absolute minimum;
1526 	 * they're much more painful than a divide.)  It's better to have a
1527 	 * solution that generates few collisions and still keeps things
1528 	 * relatively simple.
1529 	 */
1530 	bucket = hashval % dstate->dtds_hashsize;
1531 
1532 	if (op == DTRACE_DYNVAR_DEALLOC) {
1533 		volatile uintptr_t *lockp = &hash[bucket].dtdh_lock;
1534 
1535 		for (;;) {
1536 			while ((lock = *lockp) & 1)
1537 				continue;
1538 
1539 			if (dtrace_casptr((void *)lockp,
1540 			    (void *)lock, (void *)(lock + 1)) == (void *)lock)
1541 				break;
1542 		}
1543 
1544 		dtrace_membar_producer();
1545 	}
1546 
1547 top:
1548 	prev = NULL;
1549 	lock = hash[bucket].dtdh_lock;
1550 
1551 	dtrace_membar_consumer();
1552 
1553 	start = hash[bucket].dtdh_chain;
1554 	ASSERT(start != NULL && (start->dtdv_hashval == DTRACE_DYNHASH_SINK ||
1555 	    start->dtdv_hashval != DTRACE_DYNHASH_FREE ||
1556 	    op != DTRACE_DYNVAR_DEALLOC));
1557 
1558 	for (dvar = start; dvar != NULL; dvar = dvar->dtdv_next) {
1559 		dtrace_tuple_t *dtuple = &dvar->dtdv_tuple;
1560 		dtrace_key_t *dkey = &dtuple->dtt_key[0];
1561 
1562 		if (dvar->dtdv_hashval != hashval) {
1563 			if (dvar->dtdv_hashval == DTRACE_DYNHASH_SINK) {
1564 				/*
1565 				 * We've reached the sink, and therefore the
1566 				 * end of the hash chain; we can kick out of
1567 				 * the loop knowing that we have seen a valid
1568 				 * snapshot of state.
1569 				 */
1570 				ASSERT(dvar->dtdv_next == NULL);
1571 				ASSERT(dvar == &dtrace_dynhash_sink);
1572 				break;
1573 			}
1574 
1575 			if (dvar->dtdv_hashval == DTRACE_DYNHASH_FREE) {
1576 				/*
1577 				 * We've gone off the rails:  somewhere along
1578 				 * the line, one of the members of this hash
1579 				 * chain was deleted.  Note that we could also
1580 				 * detect this by simply letting this loop run
1581 				 * to completion, as we would eventually hit
1582 				 * the end of the dirty list.  However, we
1583 				 * want to avoid running the length of the
1584 				 * dirty list unnecessarily (it might be quite
1585 				 * long), so we catch this as early as
1586 				 * possible by detecting the hash marker.  In
1587 				 * this case, we simply set dvar to NULL and
1588 				 * break; the conditional after the loop will
1589 				 * send us back to top.
1590 				 */
1591 				dvar = NULL;
1592 				break;
1593 			}
1594 
1595 			goto next;
1596 		}
1597 
1598 		if (dtuple->dtt_nkeys != nkeys)
1599 			goto next;
1600 
1601 		for (i = 0; i < nkeys; i++, dkey++) {
1602 			if (dkey->dttk_size != key[i].dttk_size)
1603 				goto next; /* size or type mismatch */
1604 
1605 			if (dkey->dttk_size != 0) {
1606 				if (dtrace_bcmp(
1607 				    (void *)(uintptr_t)key[i].dttk_value,
1608 				    (void *)(uintptr_t)dkey->dttk_value,
1609 				    dkey->dttk_size))
1610 					goto next;
1611 			} else {
1612 				if (dkey->dttk_value != key[i].dttk_value)
1613 					goto next;
1614 			}
1615 		}
1616 
1617 		if (op != DTRACE_DYNVAR_DEALLOC)
1618 			return (dvar);
1619 
1620 		ASSERT(dvar->dtdv_next == NULL ||
1621 		    dvar->dtdv_next->dtdv_hashval != DTRACE_DYNHASH_FREE);
1622 
1623 		if (prev != NULL) {
1624 			ASSERT(hash[bucket].dtdh_chain != dvar);
1625 			ASSERT(start != dvar);
1626 			ASSERT(prev->dtdv_next == dvar);
1627 			prev->dtdv_next = dvar->dtdv_next;
1628 		} else {
1629 			if (dtrace_casptr(&hash[bucket].dtdh_chain,
1630 			    start, dvar->dtdv_next) != start) {
1631 				/*
1632 				 * We have failed to atomically swing the
1633 				 * hash table head pointer, presumably because
1634 				 * of a conflicting allocation on another CPU.
1635 				 * We need to reread the hash chain and try
1636 				 * again.
1637 				 */
1638 				goto top;
1639 			}
1640 		}
1641 
1642 		dtrace_membar_producer();
1643 
1644 		/*
1645 		 * Now set the hash value to indicate that it's free.
1646 		 */
1647 		ASSERT(hash[bucket].dtdh_chain != dvar);
1648 		dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
1649 
1650 		dtrace_membar_producer();
1651 
1652 		/*
1653 		 * Set the next pointer to point at the dirty list, and
1654 		 * atomically swing the dirty pointer to the newly freed dvar.
1655 		 */
1656 		do {
1657 			next = dcpu->dtdsc_dirty;
1658 			dvar->dtdv_next = next;
1659 		} while (dtrace_casptr(&dcpu->dtdsc_dirty, next, dvar) != next);
1660 
1661 		/*
1662 		 * Finally, unlock this hash bucket.
1663 		 */
1664 		ASSERT(hash[bucket].dtdh_lock == lock);
1665 		ASSERT(lock & 1);
1666 		hash[bucket].dtdh_lock++;
1667 
1668 		return (NULL);
1669 next:
1670 		prev = dvar;
1671 		continue;
1672 	}
1673 
1674 	if (dvar == NULL) {
1675 		/*
1676 		 * If dvar is NULL, it is because we went off the rails:
1677 		 * one of the elements that we traversed in the hash chain
1678 		 * was deleted while we were traversing it.  In this case,
1679 		 * we assert that we aren't doing a dealloc (deallocs lock
1680 		 * the hash bucket to prevent themselves from racing with
1681 		 * one another), and retry the hash chain traversal.
1682 		 */
1683 		ASSERT(op != DTRACE_DYNVAR_DEALLOC);
1684 		goto top;
1685 	}
1686 
1687 	if (op != DTRACE_DYNVAR_ALLOC) {
1688 		/*
1689 		 * If we are not to allocate a new variable, we want to
1690 		 * return NULL now.  Before we return, check that the value
1691 		 * of the lock word hasn't changed.  If it has, we may have
1692 		 * seen an inconsistent snapshot.
1693 		 */
1694 		if (op == DTRACE_DYNVAR_NOALLOC) {
1695 			if (hash[bucket].dtdh_lock != lock)
1696 				goto top;
1697 		} else {
1698 			ASSERT(op == DTRACE_DYNVAR_DEALLOC);
1699 			ASSERT(hash[bucket].dtdh_lock == lock);
1700 			ASSERT(lock & 1);
1701 			hash[bucket].dtdh_lock++;
1702 		}
1703 
1704 		return (NULL);
1705 	}
1706 
1707 	/*
1708 	 * We need to allocate a new dynamic variable.  The size we need is the
1709 	 * size of dtrace_dynvar plus the size of nkeys dtrace_key_t's plus the
1710 	 * size of any auxiliary key data (rounded up to 8-byte alignment) plus
1711 	 * the size of any referred-to data (dsize).  We then round the final
1712 	 * size up to the chunksize for allocation.
1713 	 */
1714 	for (ksize = 0, i = 0; i < nkeys; i++)
1715 		ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
1716 
1717 	/*
1718 	 * This should be pretty much impossible, but could happen if, say,
1719 	 * strange DIF specified the tuple.  Ideally, this should be an
1720 	 * assertion and not an error condition -- but that requires that the
1721 	 * chunksize calculation in dtrace_difo_chunksize() be absolutely
1722 	 * bullet-proof.  (That is, it must not be able to be fooled by
1723 	 * malicious DIF.)  Given the lack of backwards branches in DIF,
1724 	 * solving this would presumably not amount to solving the Halting
1725 	 * Problem -- but it still seems awfully hard.
1726 	 */
1727 	if (sizeof (dtrace_dynvar_t) + sizeof (dtrace_key_t) * (nkeys - 1) +
1728 	    ksize + dsize > chunksize) {
1729 		dcpu->dtdsc_drops++;
1730 		return (NULL);
1731 	}
1732 
1733 	nstate = DTRACE_DSTATE_EMPTY;
1734 
1735 	do {
1736 retry:
1737 		free = dcpu->dtdsc_free;
1738 
1739 		if (free == NULL) {
1740 			dtrace_dynvar_t *clean = dcpu->dtdsc_clean;
1741 			void *rval;
1742 
1743 			if (clean == NULL) {
1744 				/*
1745 				 * We're out of dynamic variable space on
1746 				 * this CPU.  Unless we have tried all CPUs,
1747 				 * we'll try to allocate from a different
1748 				 * CPU.
1749 				 */
1750 				switch (dstate->dtds_state) {
1751 				case DTRACE_DSTATE_CLEAN: {
1752 					void *sp = &dstate->dtds_state;
1753 
1754 					if (++cpu >= NCPU)
1755 						cpu = 0;
1756 
1757 					if (dcpu->dtdsc_dirty != NULL &&
1758 					    nstate == DTRACE_DSTATE_EMPTY)
1759 						nstate = DTRACE_DSTATE_DIRTY;
1760 
1761 					if (dcpu->dtdsc_rinsing != NULL)
1762 						nstate = DTRACE_DSTATE_RINSING;
1763 
1764 					dcpu = &dstate->dtds_percpu[cpu];
1765 
1766 					if (cpu != me)
1767 						goto retry;
1768 
1769 					(void) dtrace_cas32(sp,
1770 					    DTRACE_DSTATE_CLEAN, nstate);
1771 
1772 					/*
1773 					 * To increment the correct bean
1774 					 * counter, take another lap.
1775 					 */
1776 					goto retry;
1777 				}
1778 
1779 				case DTRACE_DSTATE_DIRTY:
1780 					dcpu->dtdsc_dirty_drops++;
1781 					break;
1782 
1783 				case DTRACE_DSTATE_RINSING:
1784 					dcpu->dtdsc_rinsing_drops++;
1785 					break;
1786 
1787 				case DTRACE_DSTATE_EMPTY:
1788 					dcpu->dtdsc_drops++;
1789 					break;
1790 				}
1791 
1792 				DTRACE_CPUFLAG_SET(CPU_DTRACE_DROP);
1793 				return (NULL);
1794 			}
1795 
1796 			/*
1797 			 * The clean list appears to be non-empty.  We want to
1798 			 * move the clean list to the free list; we start by
1799 			 * moving the clean pointer aside.
1800 			 */
1801 			if (dtrace_casptr(&dcpu->dtdsc_clean,
1802 			    clean, NULL) != clean) {
1803 				/*
1804 				 * We are in one of two situations:
1805 				 *
1806 				 *  (a)	The clean list was switched to the
1807 				 *	free list by another CPU.
1808 				 *
1809 				 *  (b)	The clean list was added to by the
1810 				 *	cleansing cyclic.
1811 				 *
1812 				 * In either of these situations, we can
1813 				 * just reattempt the free list allocation.
1814 				 */
1815 				goto retry;
1816 			}
1817 
1818 			ASSERT(clean->dtdv_hashval == DTRACE_DYNHASH_FREE);
1819 
1820 			/*
1821 			 * Now we'll move the clean list to our free list.
1822 			 * It's impossible for this to fail:  the only way
1823 			 * the free list can be updated is through this
1824 			 * code path, and only one CPU can own the clean list.
1825 			 * Thus, it would only be possible for this to fail if
1826 			 * this code were racing with dtrace_dynvar_clean().
1827 			 * (That is, if dtrace_dynvar_clean() updated the clean
1828 			 * list, and we ended up racing to update the free
1829 			 * list.)  This race is prevented by the dtrace_sync()
1830 			 * in dtrace_dynvar_clean() -- which flushes the
1831 			 * owners of the clean lists out before resetting
1832 			 * the clean lists.
1833 			 */
1834 			dcpu = &dstate->dtds_percpu[me];
1835 			rval = dtrace_casptr(&dcpu->dtdsc_free, NULL, clean);
1836 			ASSERT(rval == NULL);
1837 			goto retry;
1838 		}
1839 
1840 		dvar = free;
1841 		new_free = dvar->dtdv_next;
1842 	} while (dtrace_casptr(&dcpu->dtdsc_free, free, new_free) != free);
1843 
1844 	/*
1845 	 * We have now allocated a new chunk.  We copy the tuple keys into the
1846 	 * tuple array and copy any referenced key data into the data space
1847 	 * following the tuple array.  As we do this, we relocate dttk_value
1848 	 * in the final tuple to point to the key data address in the chunk.
1849 	 */
1850 	kdata = (uintptr_t)&dvar->dtdv_tuple.dtt_key[nkeys];
1851 	dvar->dtdv_data = (void *)(kdata + ksize);
1852 	dvar->dtdv_tuple.dtt_nkeys = nkeys;
1853 
1854 	for (i = 0; i < nkeys; i++) {
1855 		dtrace_key_t *dkey = &dvar->dtdv_tuple.dtt_key[i];
1856 		size_t kesize = key[i].dttk_size;
1857 
1858 		if (kesize != 0) {
1859 			dtrace_bcopy(
1860 			    (const void *)(uintptr_t)key[i].dttk_value,
1861 			    (void *)kdata, kesize);
1862 			dkey->dttk_value = kdata;
1863 			kdata += P2ROUNDUP(kesize, sizeof (uint64_t));
1864 		} else {
1865 			dkey->dttk_value = key[i].dttk_value;
1866 		}
1867 
1868 		dkey->dttk_size = kesize;
1869 	}
1870 
1871 	ASSERT(dvar->dtdv_hashval == DTRACE_DYNHASH_FREE);
1872 	dvar->dtdv_hashval = hashval;
1873 	dvar->dtdv_next = start;
1874 
1875 	if (dtrace_casptr(&hash[bucket].dtdh_chain, start, dvar) == start)
1876 		return (dvar);
1877 
1878 	/*
1879 	 * The cas has failed.  Either another CPU is adding an element to
1880 	 * this hash chain, or another CPU is deleting an element from this
1881 	 * hash chain.  The simplest way to deal with both of these cases
1882 	 * (though not necessarily the most efficient) is to free our
1883 	 * allocated block and tail-call ourselves.  Note that the free is
1884 	 * to the dirty list and _not_ to the free list.  This is to prevent
1885 	 * races with allocators, above.
1886 	 */
1887 	dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
1888 
1889 	dtrace_membar_producer();
1890 
1891 	do {
1892 		free = dcpu->dtdsc_dirty;
1893 		dvar->dtdv_next = free;
1894 	} while (dtrace_casptr(&dcpu->dtdsc_dirty, free, dvar) != free);
1895 
1896 	return (dtrace_dynvar(dstate, nkeys, key, dsize, op, mstate, vstate));
1897 }
1898 
1899 /*ARGSUSED*/
1900 static void
1901 dtrace_aggregate_min(uint64_t *oval, uint64_t nval, uint64_t arg)
1902 {
1903 	if ((int64_t)nval < (int64_t)*oval)
1904 		*oval = nval;
1905 }
1906 
1907 /*ARGSUSED*/
1908 static void
1909 dtrace_aggregate_max(uint64_t *oval, uint64_t nval, uint64_t arg)
1910 {
1911 	if ((int64_t)nval > (int64_t)*oval)
1912 		*oval = nval;
1913 }
1914 
1915 static void
1916 dtrace_aggregate_quantize(uint64_t *quanta, uint64_t nval, uint64_t incr)
1917 {
1918 	int i, zero = DTRACE_QUANTIZE_ZEROBUCKET;
1919 	int64_t val = (int64_t)nval;
1920 
1921 	if (val < 0) {
1922 		for (i = 0; i < zero; i++) {
1923 			if (val <= DTRACE_QUANTIZE_BUCKETVAL(i)) {
1924 				quanta[i] += incr;
1925 				return;
1926 			}
1927 		}
1928 	} else {
1929 		for (i = zero + 1; i < DTRACE_QUANTIZE_NBUCKETS; i++) {
1930 			if (val < DTRACE_QUANTIZE_BUCKETVAL(i)) {
1931 				quanta[i - 1] += incr;
1932 				return;
1933 			}
1934 		}
1935 
1936 		quanta[DTRACE_QUANTIZE_NBUCKETS - 1] += incr;
1937 		return;
1938 	}
1939 
1940 	ASSERT(0);
1941 }
1942 
1943 static void
1944 dtrace_aggregate_lquantize(uint64_t *lquanta, uint64_t nval, uint64_t incr)
1945 {
1946 	uint64_t arg = *lquanta++;
1947 	int32_t base = DTRACE_LQUANTIZE_BASE(arg);
1948 	uint16_t step = DTRACE_LQUANTIZE_STEP(arg);
1949 	uint16_t levels = DTRACE_LQUANTIZE_LEVELS(arg);
1950 	int32_t val = (int32_t)nval, level;
1951 
1952 	ASSERT(step != 0);
1953 	ASSERT(levels != 0);
1954 
1955 	if (val < base) {
1956 		/*
1957 		 * This is an underflow.
1958 		 */
1959 		lquanta[0] += incr;
1960 		return;
1961 	}
1962 
1963 	level = (val - base) / step;
1964 
1965 	if (level < levels) {
1966 		lquanta[level + 1] += incr;
1967 		return;
1968 	}
1969 
1970 	/*
1971 	 * This is an overflow.
1972 	 */
1973 	lquanta[levels + 1] += incr;
1974 }
1975 
1976 static int
1977 dtrace_aggregate_llquantize_bucket(uint16_t factor, uint16_t low,
1978     uint16_t high, uint16_t nsteps, int64_t value)
1979 {
1980 	int64_t this = 1, last, next;
1981 	int base = 1, order;
1982 
1983 	ASSERT(factor <= nsteps);
1984 	ASSERT(nsteps % factor == 0);
1985 
1986 	for (order = 0; order < low; order++)
1987 		this *= factor;
1988 
1989 	/*
1990 	 * If our value is less than our factor taken to the power of the
1991 	 * low order of magnitude, it goes into the zeroth bucket.
1992 	 */
1993 	if (value < (last = this))
1994 		return (0);
1995 
1996 	for (this *= factor; order <= high; order++) {
1997 		int nbuckets = this > nsteps ? nsteps : this;
1998 
1999 		if ((next = this * factor) < this) {
2000 			/*
2001 			 * We should not generally get log/linear quantizations
2002 			 * with a high magnitude that allows 64-bits to
2003 			 * overflow, but we nonetheless protect against this
2004 			 * by explicitly checking for overflow, and clamping
2005 			 * our value accordingly.
2006 			 */
2007 			value = this - 1;
2008 		}
2009 
2010 		if (value < this) {
2011 			/*
2012 			 * If our value lies within this order of magnitude,
2013 			 * determine its position by taking the offset within
2014 			 * the order of magnitude, dividing by the bucket
2015 			 * width, and adding to our (accumulated) base.
2016 			 */
2017 			return (base + (value - last) / (this / nbuckets));
2018 		}
2019 
2020 		base += nbuckets - (nbuckets / factor);
2021 		last = this;
2022 		this = next;
2023 	}
2024 
2025 	/*
2026 	 * Our value is greater than or equal to our factor taken to the
2027 	 * power of one plus the high magnitude -- return the top bucket.
2028 	 */
2029 	return (base);
2030 }
2031 
2032 static void
2033 dtrace_aggregate_llquantize(uint64_t *llquanta, uint64_t nval, uint64_t incr)
2034 {
2035 	uint64_t arg = *llquanta++;
2036 	uint16_t factor = DTRACE_LLQUANTIZE_FACTOR(arg);
2037 	uint16_t low = DTRACE_LLQUANTIZE_LOW(arg);
2038 	uint16_t high = DTRACE_LLQUANTIZE_HIGH(arg);
2039 	uint16_t nsteps = DTRACE_LLQUANTIZE_NSTEP(arg);
2040 
2041 	llquanta[dtrace_aggregate_llquantize_bucket(factor,
2042 	    low, high, nsteps, nval)] += incr;
2043 }
2044 
2045 /*ARGSUSED*/
2046 static void
2047 dtrace_aggregate_avg(uint64_t *data, uint64_t nval, uint64_t arg)
2048 {
2049 	data[0]++;
2050 	data[1] += nval;
2051 }
2052 
2053 /*ARGSUSED*/
2054 static void
2055 dtrace_aggregate_stddev(uint64_t *data, uint64_t nval, uint64_t arg)
2056 {
2057 	int64_t snval = (int64_t)nval;
2058 	uint64_t tmp[2];
2059 
2060 	data[0]++;
2061 	data[1] += nval;
2062 
2063 	/*
2064 	 * What we want to say here is:
2065 	 *
2066 	 * data[2] += nval * nval;
2067 	 *
2068 	 * But given that nval is 64-bit, we could easily overflow, so
2069 	 * we do this as 128-bit arithmetic.
2070 	 */
2071 	if (snval < 0)
2072 		snval = -snval;
2073 
2074 	dtrace_multiply_128((uint64_t)snval, (uint64_t)snval, tmp);
2075 	dtrace_add_128(data + 2, tmp, data + 2);
2076 }
2077 
2078 /*ARGSUSED*/
2079 static void
2080 dtrace_aggregate_count(uint64_t *oval, uint64_t nval, uint64_t arg)
2081 {
2082 	*oval = *oval + 1;
2083 }
2084 
2085 /*ARGSUSED*/
2086 static void
2087 dtrace_aggregate_sum(uint64_t *oval, uint64_t nval, uint64_t arg)
2088 {
2089 	*oval += nval;
2090 }
2091 
2092 /*
2093  * Aggregate given the tuple in the principal data buffer, and the aggregating
2094  * action denoted by the specified dtrace_aggregation_t.  The aggregation
2095  * buffer is specified as the buf parameter.  This routine does not return
2096  * failure; if there is no space in the aggregation buffer, the data will be
2097  * dropped, and a corresponding counter incremented.
2098  */
2099 static void
2100 dtrace_aggregate(dtrace_aggregation_t *agg, dtrace_buffer_t *dbuf,
2101     intptr_t offset, dtrace_buffer_t *buf, uint64_t expr, uint64_t arg)
2102 {
2103 	dtrace_recdesc_t *rec = &agg->dtag_action.dta_rec;
2104 	uint32_t i, ndx, size, fsize;
2105 	uint32_t align = sizeof (uint64_t) - 1;
2106 	dtrace_aggbuffer_t *agb;
2107 	dtrace_aggkey_t *key;
2108 	uint32_t hashval = 0, limit, isstr;
2109 	caddr_t tomax, data, kdata;
2110 	dtrace_actkind_t action;
2111 	dtrace_action_t *act;
2112 	uintptr_t offs;
2113 
2114 	if (buf == NULL)
2115 		return;
2116 
2117 	if (!agg->dtag_hasarg) {
2118 		/*
2119 		 * Currently, only quantize() and lquantize() take additional
2120 		 * arguments, and they have the same semantics:  an increment
2121 		 * value that defaults to 1 when not present.  If additional
2122 		 * aggregating actions take arguments, the setting of the
2123 		 * default argument value will presumably have to become more
2124 		 * sophisticated...
2125 		 */
2126 		arg = 1;
2127 	}
2128 
2129 	action = agg->dtag_action.dta_kind - DTRACEACT_AGGREGATION;
2130 	size = rec->dtrd_offset - agg->dtag_base;
2131 	fsize = size + rec->dtrd_size;
2132 
2133 	ASSERT(dbuf->dtb_tomax != NULL);
2134 	data = dbuf->dtb_tomax + offset + agg->dtag_base;
2135 
2136 	if ((tomax = buf->dtb_tomax) == NULL) {
2137 		dtrace_buffer_drop(buf);
2138 		return;
2139 	}
2140 
2141 	/*
2142 	 * The metastructure is always at the bottom of the buffer.
2143 	 */
2144 	agb = (dtrace_aggbuffer_t *)(tomax + buf->dtb_size -
2145 	    sizeof (dtrace_aggbuffer_t));
2146 
2147 	if (buf->dtb_offset == 0) {
2148 		/*
2149 		 * We just kludge up approximately 1/8th of the size to be
2150 		 * buckets.  If this guess ends up being routinely
2151 		 * off-the-mark, we may need to dynamically readjust this
2152 		 * based on past performance.
2153 		 */
2154 		uintptr_t hashsize = (buf->dtb_size >> 3) / sizeof (uintptr_t);
2155 
2156 		if ((uintptr_t)agb - hashsize * sizeof (dtrace_aggkey_t *) <
2157 		    (uintptr_t)tomax || hashsize == 0) {
2158 			/*
2159 			 * We've been given a ludicrously small buffer;
2160 			 * increment our drop count and leave.
2161 			 */
2162 			dtrace_buffer_drop(buf);
2163 			return;
2164 		}
2165 
2166 		/*
2167 		 * And now, a pathetic attempt to try to get a an odd (or
2168 		 * perchance, a prime) hash size for better hash distribution.
2169 		 */
2170 		if (hashsize > (DTRACE_AGGHASHSIZE_SLEW << 3))
2171 			hashsize -= DTRACE_AGGHASHSIZE_SLEW;
2172 
2173 		agb->dtagb_hashsize = hashsize;
2174 		agb->dtagb_hash = (dtrace_aggkey_t **)((uintptr_t)agb -
2175 		    agb->dtagb_hashsize * sizeof (dtrace_aggkey_t *));
2176 		agb->dtagb_free = (uintptr_t)agb->dtagb_hash;
2177 
2178 		for (i = 0; i < agb->dtagb_hashsize; i++)
2179 			agb->dtagb_hash[i] = NULL;
2180 	}
2181 
2182 	ASSERT(agg->dtag_first != NULL);
2183 	ASSERT(agg->dtag_first->dta_intuple);
2184 
2185 	/*
2186 	 * Calculate the hash value based on the key.  Note that we _don't_
2187 	 * include the aggid in the hashing (but we will store it as part of
2188 	 * the key).  The hashing algorithm is Bob Jenkins' "One-at-a-time"
2189 	 * algorithm: a simple, quick algorithm that has no known funnels, and
2190 	 * gets good distribution in practice.  The efficacy of the hashing
2191 	 * algorithm (and a comparison with other algorithms) may be found by
2192 	 * running the ::dtrace_aggstat MDB dcmd.
2193 	 */
2194 	for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
2195 		i = act->dta_rec.dtrd_offset - agg->dtag_base;
2196 		limit = i + act->dta_rec.dtrd_size;
2197 		ASSERT(limit <= size);
2198 		isstr = DTRACEACT_ISSTRING(act);
2199 
2200 		for (; i < limit; i++) {
2201 			hashval += data[i];
2202 			hashval += (hashval << 10);
2203 			hashval ^= (hashval >> 6);
2204 
2205 			if (isstr && data[i] == '\0')
2206 				break;
2207 		}
2208 	}
2209 
2210 	hashval += (hashval << 3);
2211 	hashval ^= (hashval >> 11);
2212 	hashval += (hashval << 15);
2213 
2214 	/*
2215 	 * Yes, the divide here is expensive -- but it's generally the least
2216 	 * of the performance issues given the amount of data that we iterate
2217 	 * over to compute hash values, compare data, etc.
2218 	 */
2219 	ndx = hashval % agb->dtagb_hashsize;
2220 
2221 	for (key = agb->dtagb_hash[ndx]; key != NULL; key = key->dtak_next) {
2222 		ASSERT((caddr_t)key >= tomax);
2223 		ASSERT((caddr_t)key < tomax + buf->dtb_size);
2224 
2225 		if (hashval != key->dtak_hashval || key->dtak_size != size)
2226 			continue;
2227 
2228 		kdata = key->dtak_data;
2229 		ASSERT(kdata >= tomax && kdata < tomax + buf->dtb_size);
2230 
2231 		for (act = agg->dtag_first; act->dta_intuple;
2232 		    act = act->dta_next) {
2233 			i = act->dta_rec.dtrd_offset - agg->dtag_base;
2234 			limit = i + act->dta_rec.dtrd_size;
2235 			ASSERT(limit <= size);
2236 			isstr = DTRACEACT_ISSTRING(act);
2237 
2238 			for (; i < limit; i++) {
2239 				if (kdata[i] != data[i])
2240 					goto next;
2241 
2242 				if (isstr && data[i] == '\0')
2243 					break;
2244 			}
2245 		}
2246 
2247 		if (action != key->dtak_action) {
2248 			/*
2249 			 * We are aggregating on the same value in the same
2250 			 * aggregation with two different aggregating actions.
2251 			 * (This should have been picked up in the compiler,
2252 			 * so we may be dealing with errant or devious DIF.)
2253 			 * This is an error condition; we indicate as much,
2254 			 * and return.
2255 			 */
2256 			DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
2257 			return;
2258 		}
2259 
2260 		/*
2261 		 * This is a hit:  we need to apply the aggregator to
2262 		 * the value at this key.
2263 		 */
2264 		agg->dtag_aggregate((uint64_t *)(kdata + size), expr, arg);
2265 		return;
2266 next:
2267 		continue;
2268 	}
2269 
2270 	/*
2271 	 * We didn't find it.  We need to allocate some zero-filled space,
2272 	 * link it into the hash table appropriately, and apply the aggregator
2273 	 * to the (zero-filled) value.
2274 	 */
2275 	offs = buf->dtb_offset;
2276 	while (offs & (align - 1))
2277 		offs += sizeof (uint32_t);
2278 
2279 	/*
2280 	 * If we don't have enough room to both allocate a new key _and_
2281 	 * its associated data, increment the drop count and return.
2282 	 */
2283 	if ((uintptr_t)tomax + offs + fsize >
2284 	    agb->dtagb_free - sizeof (dtrace_aggkey_t)) {
2285 		dtrace_buffer_drop(buf);
2286 		return;
2287 	}
2288 
2289 	/*CONSTCOND*/
2290 	ASSERT(!(sizeof (dtrace_aggkey_t) & (sizeof (uintptr_t) - 1)));
2291 	key = (dtrace_aggkey_t *)(agb->dtagb_free - sizeof (dtrace_aggkey_t));
2292 	agb->dtagb_free -= sizeof (dtrace_aggkey_t);
2293 
2294 	key->dtak_data = kdata = tomax + offs;
2295 	buf->dtb_offset = offs + fsize;
2296 
2297 	/*
2298 	 * Now copy the data across.
2299 	 */
2300 	*((dtrace_aggid_t *)kdata) = agg->dtag_id;
2301 
2302 	for (i = sizeof (dtrace_aggid_t); i < size; i++)
2303 		kdata[i] = data[i];
2304 
2305 	/*
2306 	 * Because strings are not zeroed out by default, we need to iterate
2307 	 * looking for actions that store strings, and we need to explicitly
2308 	 * pad these strings out with zeroes.
2309 	 */
2310 	for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
2311 		int nul;
2312 
2313 		if (!DTRACEACT_ISSTRING(act))
2314 			continue;
2315 
2316 		i = act->dta_rec.dtrd_offset - agg->dtag_base;
2317 		limit = i + act->dta_rec.dtrd_size;
2318 		ASSERT(limit <= size);
2319 
2320 		for (nul = 0; i < limit; i++) {
2321 			if (nul) {
2322 				kdata[i] = '\0';
2323 				continue;
2324 			}
2325 
2326 			if (data[i] != '\0')
2327 				continue;
2328 
2329 			nul = 1;
2330 		}
2331 	}
2332 
2333 	for (i = size; i < fsize; i++)
2334 		kdata[i] = 0;
2335 
2336 	key->dtak_hashval = hashval;
2337 	key->dtak_size = size;
2338 	key->dtak_action = action;
2339 	key->dtak_next = agb->dtagb_hash[ndx];
2340 	agb->dtagb_hash[ndx] = key;
2341 
2342 	/*
2343 	 * Finally, apply the aggregator.
2344 	 */
2345 	*((uint64_t *)(key->dtak_data + size)) = agg->dtag_initial;
2346 	agg->dtag_aggregate((uint64_t *)(key->dtak_data + size), expr, arg);
2347 }
2348 
2349 /*
2350  * Given consumer state, this routine finds a speculation in the INACTIVE
2351  * state and transitions it into the ACTIVE state.  If there is no speculation
2352  * in the INACTIVE state, 0 is returned.  In this case, no error counter is
2353  * incremented -- it is up to the caller to take appropriate action.
2354  */
2355 static int
2356 dtrace_speculation(dtrace_state_t *state)
2357 {
2358 	int i = 0;
2359 	dtrace_speculation_state_t current;
2360 	uint32_t *stat = &state->dts_speculations_unavail, count;
2361 
2362 	while (i < state->dts_nspeculations) {
2363 		dtrace_speculation_t *spec = &state->dts_speculations[i];
2364 
2365 		current = spec->dtsp_state;
2366 
2367 		if (current != DTRACESPEC_INACTIVE) {
2368 			if (current == DTRACESPEC_COMMITTINGMANY ||
2369 			    current == DTRACESPEC_COMMITTING ||
2370 			    current == DTRACESPEC_DISCARDING)
2371 				stat = &state->dts_speculations_busy;
2372 			i++;
2373 			continue;
2374 		}
2375 
2376 		if (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2377 		    current, DTRACESPEC_ACTIVE) == current)
2378 			return (i + 1);
2379 	}
2380 
2381 	/*
2382 	 * We couldn't find a speculation.  If we found as much as a single
2383 	 * busy speculation buffer, we'll attribute this failure as "busy"
2384 	 * instead of "unavail".
2385 	 */
2386 	do {
2387 		count = *stat;
2388 	} while (dtrace_cas32(stat, count, count + 1) != count);
2389 
2390 	return (0);
2391 }
2392 
2393 /*
2394  * This routine commits an active speculation.  If the specified speculation
2395  * is not in a valid state to perform a commit(), this routine will silently do
2396  * nothing.  The state of the specified speculation is transitioned according
2397  * to the state transition diagram outlined in <sys/dtrace_impl.h>
2398  */
2399 static void
2400 dtrace_speculation_commit(dtrace_state_t *state, processorid_t cpu,
2401     dtrace_specid_t which)
2402 {
2403 	dtrace_speculation_t *spec;
2404 	dtrace_buffer_t *src, *dest;
2405 	uintptr_t daddr, saddr, dlimit, slimit;
2406 	dtrace_speculation_state_t current, new;
2407 	intptr_t offs;
2408 	uint64_t timestamp;
2409 
2410 	if (which == 0)
2411 		return;
2412 
2413 	if (which > state->dts_nspeculations) {
2414 		cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2415 		return;
2416 	}
2417 
2418 	spec = &state->dts_speculations[which - 1];
2419 	src = &spec->dtsp_buffer[cpu];
2420 	dest = &state->dts_buffer[cpu];
2421 
2422 	do {
2423 		current = spec->dtsp_state;
2424 
2425 		if (current == DTRACESPEC_COMMITTINGMANY)
2426 			break;
2427 
2428 		switch (current) {
2429 		case DTRACESPEC_INACTIVE:
2430 		case DTRACESPEC_DISCARDING:
2431 			return;
2432 
2433 		case DTRACESPEC_COMMITTING:
2434 			/*
2435 			 * This is only possible if we are (a) commit()'ing
2436 			 * without having done a prior speculate() on this CPU
2437 			 * and (b) racing with another commit() on a different
2438 			 * CPU.  There's nothing to do -- we just assert that
2439 			 * our offset is 0.
2440 			 */
2441 			ASSERT(src->dtb_offset == 0);
2442 			return;
2443 
2444 		case DTRACESPEC_ACTIVE:
2445 			new = DTRACESPEC_COMMITTING;
2446 			break;
2447 
2448 		case DTRACESPEC_ACTIVEONE:
2449 			/*
2450 			 * This speculation is active on one CPU.  If our
2451 			 * buffer offset is non-zero, we know that the one CPU
2452 			 * must be us.  Otherwise, we are committing on a
2453 			 * different CPU from the speculate(), and we must
2454 			 * rely on being asynchronously cleaned.
2455 			 */
2456 			if (src->dtb_offset != 0) {
2457 				new = DTRACESPEC_COMMITTING;
2458 				break;
2459 			}
2460 			/*FALLTHROUGH*/
2461 
2462 		case DTRACESPEC_ACTIVEMANY:
2463 			new = DTRACESPEC_COMMITTINGMANY;
2464 			break;
2465 
2466 		default:
2467 			ASSERT(0);
2468 		}
2469 	} while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2470 	    current, new) != current);
2471 
2472 	/*
2473 	 * We have set the state to indicate that we are committing this
2474 	 * speculation.  Now reserve the necessary space in the destination
2475 	 * buffer.
2476 	 */
2477 	if ((offs = dtrace_buffer_reserve(dest, src->dtb_offset,
2478 	    sizeof (uint64_t), state, NULL)) < 0) {
2479 		dtrace_buffer_drop(dest);
2480 		goto out;
2481 	}
2482 
2483 	/*
2484 	 * We have sufficient space to copy the speculative buffer into the
2485 	 * primary buffer.  First, modify the speculative buffer, filling
2486 	 * in the timestamp of all entries with the current time.  The data
2487 	 * must have the commit() time rather than the time it was traced,
2488 	 * so that all entries in the primary buffer are in timestamp order.
2489 	 */
2490 	timestamp = dtrace_gethrtime();
2491 	saddr = (uintptr_t)src->dtb_tomax;
2492 	slimit = saddr + src->dtb_offset;
2493 	while (saddr < slimit) {
2494 		size_t size;
2495 		dtrace_rechdr_t *dtrh = (dtrace_rechdr_t *)saddr;
2496 
2497 		if (dtrh->dtrh_epid == DTRACE_EPIDNONE) {
2498 			saddr += sizeof (dtrace_epid_t);
2499 			continue;
2500 		}
2501 		ASSERT3U(dtrh->dtrh_epid, <=, state->dts_necbs);
2502 		size = state->dts_ecbs[dtrh->dtrh_epid - 1]->dte_size;
2503 
2504 		ASSERT3U(saddr + size, <=, slimit);
2505 		ASSERT3U(size, >=, sizeof (dtrace_rechdr_t));
2506 		ASSERT3U(DTRACE_RECORD_LOAD_TIMESTAMP(dtrh), ==, UINT64_MAX);
2507 
2508 		DTRACE_RECORD_STORE_TIMESTAMP(dtrh, timestamp);
2509 
2510 		saddr += size;
2511 	}
2512 
2513 	/*
2514 	 * Copy the buffer across.  (Note that this is a
2515 	 * highly subobtimal bcopy(); in the unlikely event that this becomes
2516 	 * a serious performance issue, a high-performance DTrace-specific
2517 	 * bcopy() should obviously be invented.)
2518 	 */
2519 	daddr = (uintptr_t)dest->dtb_tomax + offs;
2520 	dlimit = daddr + src->dtb_offset;
2521 	saddr = (uintptr_t)src->dtb_tomax;
2522 
2523 	/*
2524 	 * First, the aligned portion.
2525 	 */
2526 	while (dlimit - daddr >= sizeof (uint64_t)) {
2527 		*((uint64_t *)daddr) = *((uint64_t *)saddr);
2528 
2529 		daddr += sizeof (uint64_t);
2530 		saddr += sizeof (uint64_t);
2531 	}
2532 
2533 	/*
2534 	 * Now any left-over bit...
2535 	 */
2536 	while (dlimit - daddr)
2537 		*((uint8_t *)daddr++) = *((uint8_t *)saddr++);
2538 
2539 	/*
2540 	 * Finally, commit the reserved space in the destination buffer.
2541 	 */
2542 	dest->dtb_offset = offs + src->dtb_offset;
2543 
2544 out:
2545 	/*
2546 	 * If we're lucky enough to be the only active CPU on this speculation
2547 	 * buffer, we can just set the state back to DTRACESPEC_INACTIVE.
2548 	 */
2549 	if (current == DTRACESPEC_ACTIVE ||
2550 	    (current == DTRACESPEC_ACTIVEONE && new == DTRACESPEC_COMMITTING)) {
2551 		uint32_t rval = dtrace_cas32((uint32_t *)&spec->dtsp_state,
2552 		    DTRACESPEC_COMMITTING, DTRACESPEC_INACTIVE);
2553 
2554 		ASSERT(rval == DTRACESPEC_COMMITTING);
2555 	}
2556 
2557 	src->dtb_offset = 0;
2558 	src->dtb_xamot_drops += src->dtb_drops;
2559 	src->dtb_drops = 0;
2560 }
2561 
2562 /*
2563  * This routine discards an active speculation.  If the specified speculation
2564  * is not in a valid state to perform a discard(), this routine will silently
2565  * do nothing.  The state of the specified speculation is transitioned
2566  * according to the state transition diagram outlined in <sys/dtrace_impl.h>
2567  */
2568 static void
2569 dtrace_speculation_discard(dtrace_state_t *state, processorid_t cpu,
2570     dtrace_specid_t which)
2571 {
2572 	dtrace_speculation_t *spec;
2573 	dtrace_speculation_state_t current, new;
2574 	dtrace_buffer_t *buf;
2575 
2576 	if (which == 0)
2577 		return;
2578 
2579 	if (which > state->dts_nspeculations) {
2580 		cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2581 		return;
2582 	}
2583 
2584 	spec = &state->dts_speculations[which - 1];
2585 	buf = &spec->dtsp_buffer[cpu];
2586 
2587 	do {
2588 		current = spec->dtsp_state;
2589 
2590 		switch (current) {
2591 		case DTRACESPEC_INACTIVE:
2592 		case DTRACESPEC_COMMITTINGMANY:
2593 		case DTRACESPEC_COMMITTING:
2594 		case DTRACESPEC_DISCARDING:
2595 			return;
2596 
2597 		case DTRACESPEC_ACTIVE:
2598 		case DTRACESPEC_ACTIVEMANY:
2599 			new = DTRACESPEC_DISCARDING;
2600 			break;
2601 
2602 		case DTRACESPEC_ACTIVEONE:
2603 			if (buf->dtb_offset != 0) {
2604 				new = DTRACESPEC_INACTIVE;
2605 			} else {
2606 				new = DTRACESPEC_DISCARDING;
2607 			}
2608 			break;
2609 
2610 		default:
2611 			ASSERT(0);
2612 		}
2613 	} while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2614 	    current, new) != current);
2615 
2616 	buf->dtb_offset = 0;
2617 	buf->dtb_drops = 0;
2618 }
2619 
2620 /*
2621  * Note:  not called from probe context.  This function is called
2622  * asynchronously from cross call context to clean any speculations that are
2623  * in the COMMITTINGMANY or DISCARDING states.  These speculations may not be
2624  * transitioned back to the INACTIVE state until all CPUs have cleaned the
2625  * speculation.
2626  */
2627 static void
2628 dtrace_speculation_clean_here(dtrace_state_t *state)
2629 {
2630 	dtrace_icookie_t cookie;
2631 	processorid_t cpu = CPU->cpu_id;
2632 	dtrace_buffer_t *dest = &state->dts_buffer[cpu];
2633 	dtrace_specid_t i;
2634 
2635 	cookie = dtrace_interrupt_disable();
2636 
2637 	if (dest->dtb_tomax == NULL) {
2638 		dtrace_interrupt_enable(cookie);
2639 		return;
2640 	}
2641 
2642 	for (i = 0; i < state->dts_nspeculations; i++) {
2643 		dtrace_speculation_t *spec = &state->dts_speculations[i];
2644 		dtrace_buffer_t *src = &spec->dtsp_buffer[cpu];
2645 
2646 		if (src->dtb_tomax == NULL)
2647 			continue;
2648 
2649 		if (spec->dtsp_state == DTRACESPEC_DISCARDING) {
2650 			src->dtb_offset = 0;
2651 			continue;
2652 		}
2653 
2654 		if (spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
2655 			continue;
2656 
2657 		if (src->dtb_offset == 0)
2658 			continue;
2659 
2660 		dtrace_speculation_commit(state, cpu, i + 1);
2661 	}
2662 
2663 	dtrace_interrupt_enable(cookie);
2664 }
2665 
2666 /*
2667  * Note:  not called from probe context.  This function is called
2668  * asynchronously (and at a regular interval) to clean any speculations that
2669  * are in the COMMITTINGMANY or DISCARDING states.  If it discovers that there
2670  * is work to be done, it cross calls all CPUs to perform that work;
2671  * COMMITMANY and DISCARDING speculations may not be transitioned back to the
2672  * INACTIVE state until they have been cleaned by all CPUs.
2673  */
2674 static void
2675 dtrace_speculation_clean(dtrace_state_t *state)
2676 {
2677 	int work = 0, rv;
2678 	dtrace_specid_t i;
2679 
2680 	for (i = 0; i < state->dts_nspeculations; i++) {
2681 		dtrace_speculation_t *spec = &state->dts_speculations[i];
2682 
2683 		ASSERT(!spec->dtsp_cleaning);
2684 
2685 		if (spec->dtsp_state != DTRACESPEC_DISCARDING &&
2686 		    spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
2687 			continue;
2688 
2689 		work++;
2690 		spec->dtsp_cleaning = 1;
2691 	}
2692 
2693 	if (!work)
2694 		return;
2695 
2696 	dtrace_xcall(DTRACE_CPUALL,
2697 	    (dtrace_xcall_t)dtrace_speculation_clean_here, state);
2698 
2699 	/*
2700 	 * We now know that all CPUs have committed or discarded their
2701 	 * speculation buffers, as appropriate.  We can now set the state
2702 	 * to inactive.
2703 	 */
2704 	for (i = 0; i < state->dts_nspeculations; i++) {
2705 		dtrace_speculation_t *spec = &state->dts_speculations[i];
2706 		dtrace_speculation_state_t current, new;
2707 
2708 		if (!spec->dtsp_cleaning)
2709 			continue;
2710 
2711 		current = spec->dtsp_state;
2712 		ASSERT(current == DTRACESPEC_DISCARDING ||
2713 		    current == DTRACESPEC_COMMITTINGMANY);
2714 
2715 		new = DTRACESPEC_INACTIVE;
2716 
2717 		rv = dtrace_cas32((uint32_t *)&spec->dtsp_state, current, new);
2718 		ASSERT(rv == current);
2719 		spec->dtsp_cleaning = 0;
2720 	}
2721 }
2722 
2723 /*
2724  * Called as part of a speculate() to get the speculative buffer associated
2725  * with a given speculation.  Returns NULL if the specified speculation is not
2726  * in an ACTIVE state.  If the speculation is in the ACTIVEONE state -- and
2727  * the active CPU is not the specified CPU -- the speculation will be
2728  * atomically transitioned into the ACTIVEMANY state.
2729  */
2730 static dtrace_buffer_t *
2731 dtrace_speculation_buffer(dtrace_state_t *state, processorid_t cpuid,
2732     dtrace_specid_t which)
2733 {
2734 	dtrace_speculation_t *spec;
2735 	dtrace_speculation_state_t current, new;
2736 	dtrace_buffer_t *buf;
2737 
2738 	if (which == 0)
2739 		return (NULL);
2740 
2741 	if (which > state->dts_nspeculations) {
2742 		cpu_core[cpuid].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2743 		return (NULL);
2744 	}
2745 
2746 	spec = &state->dts_speculations[which - 1];
2747 	buf = &spec->dtsp_buffer[cpuid];
2748 
2749 	do {
2750 		current = spec->dtsp_state;
2751 
2752 		switch (current) {
2753 		case DTRACESPEC_INACTIVE:
2754 		case DTRACESPEC_COMMITTINGMANY:
2755 		case DTRACESPEC_DISCARDING:
2756 			return (NULL);
2757 
2758 		case DTRACESPEC_COMMITTING:
2759 			ASSERT(buf->dtb_offset == 0);
2760 			return (NULL);
2761 
2762 		case DTRACESPEC_ACTIVEONE:
2763 			/*
2764 			 * This speculation is currently active on one CPU.
2765 			 * Check the offset in the buffer; if it's non-zero,
2766 			 * that CPU must be us (and we leave the state alone).
2767 			 * If it's zero, assume that we're starting on a new
2768 			 * CPU -- and change the state to indicate that the
2769 			 * speculation is active on more than one CPU.
2770 			 */
2771 			if (buf->dtb_offset != 0)
2772 				return (buf);
2773 
2774 			new = DTRACESPEC_ACTIVEMANY;
2775 			break;
2776 
2777 		case DTRACESPEC_ACTIVEMANY:
2778 			return (buf);
2779 
2780 		case DTRACESPEC_ACTIVE:
2781 			new = DTRACESPEC_ACTIVEONE;
2782 			break;
2783 
2784 		default:
2785 			ASSERT(0);
2786 		}
2787 	} while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2788 	    current, new) != current);
2789 
2790 	ASSERT(new == DTRACESPEC_ACTIVEONE || new == DTRACESPEC_ACTIVEMANY);
2791 	return (buf);
2792 }
2793 
2794 /*
2795  * Return a string.  In the event that the user lacks the privilege to access
2796  * arbitrary kernel memory, we copy the string out to scratch memory so that we
2797  * don't fail access checking.
2798  *
2799  * dtrace_dif_variable() uses this routine as a helper for various
2800  * builtin values such as 'execname' and 'probefunc.'
2801  */
2802 uintptr_t
2803 dtrace_dif_varstr(uintptr_t addr, dtrace_state_t *state,
2804     dtrace_mstate_t *mstate)
2805 {
2806 	uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
2807 	uintptr_t ret;
2808 	size_t strsz;
2809 
2810 	/*
2811 	 * The easy case: this probe is allowed to read all of memory, so
2812 	 * we can just return this as a vanilla pointer.
2813 	 */
2814 	if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
2815 		return (addr);
2816 
2817 	/*
2818 	 * This is the tougher case: we copy the string in question from
2819 	 * kernel memory into scratch memory and return it that way: this
2820 	 * ensures that we won't trip up when access checking tests the
2821 	 * BYREF return value.
2822 	 */
2823 	strsz = dtrace_strlen((char *)addr, size) + 1;
2824 
2825 	if (mstate->dtms_scratch_ptr + strsz >
2826 	    mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
2827 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
2828 		return (NULL);
2829 	}
2830 
2831 	dtrace_strcpy((const void *)addr, (void *)mstate->dtms_scratch_ptr,
2832 	    strsz);
2833 	ret = mstate->dtms_scratch_ptr;
2834 	mstate->dtms_scratch_ptr += strsz;
2835 	return (ret);
2836 }
2837 
2838 /*
2839  * This function implements the DIF emulator's variable lookups.  The emulator
2840  * passes a reserved variable identifier and optional built-in array index.
2841  */
2842 static uint64_t
2843 dtrace_dif_variable(dtrace_mstate_t *mstate, dtrace_state_t *state, uint64_t v,
2844     uint64_t ndx)
2845 {
2846 	/*
2847 	 * If we're accessing one of the uncached arguments, we'll turn this
2848 	 * into a reference in the args array.
2849 	 */
2850 	if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9) {
2851 		ndx = v - DIF_VAR_ARG0;
2852 		v = DIF_VAR_ARGS;
2853 	}
2854 
2855 	switch (v) {
2856 	case DIF_VAR_ARGS:
2857 		if (!(mstate->dtms_access & DTRACE_ACCESS_ARGS)) {
2858 			cpu_core[CPU->cpu_id].cpuc_dtrace_flags |=
2859 			    CPU_DTRACE_KPRIV;
2860 			return (0);
2861 		}
2862 
2863 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_ARGS);
2864 		if (ndx >= sizeof (mstate->dtms_arg) /
2865 		    sizeof (mstate->dtms_arg[0])) {
2866 			int aframes = mstate->dtms_probe->dtpr_aframes + 2;
2867 			dtrace_provider_t *pv;
2868 			uint64_t val;
2869 
2870 			pv = mstate->dtms_probe->dtpr_provider;
2871 			if (pv->dtpv_pops.dtps_getargval != NULL)
2872 				val = pv->dtpv_pops.dtps_getargval(pv->dtpv_arg,
2873 				    mstate->dtms_probe->dtpr_id,
2874 				    mstate->dtms_probe->dtpr_arg, ndx, aframes);
2875 			else
2876 				val = dtrace_getarg(ndx, aframes);
2877 
2878 			/*
2879 			 * This is regrettably required to keep the compiler
2880 			 * from tail-optimizing the call to dtrace_getarg().
2881 			 * The condition always evaluates to true, but the
2882 			 * compiler has no way of figuring that out a priori.
2883 			 * (None of this would be necessary if the compiler
2884 			 * could be relied upon to _always_ tail-optimize
2885 			 * the call to dtrace_getarg() -- but it can't.)
2886 			 */
2887 			if (mstate->dtms_probe != NULL)
2888 				return (val);
2889 
2890 			ASSERT(0);
2891 		}
2892 
2893 		return (mstate->dtms_arg[ndx]);
2894 
2895 	case DIF_VAR_UREGS: {
2896 		klwp_t *lwp;
2897 
2898 		if (!dtrace_priv_proc(state, mstate))
2899 			return (0);
2900 
2901 		if ((lwp = curthread->t_lwp) == NULL) {
2902 			DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
2903 			cpu_core[CPU->cpu_id].cpuc_dtrace_illval = NULL;
2904 			return (0);
2905 		}
2906 
2907 		return (dtrace_getreg(lwp->lwp_regs, ndx));
2908 	}
2909 
2910 	case DIF_VAR_VMREGS: {
2911 		uint64_t rval;
2912 
2913 		if (!dtrace_priv_kernel(state))
2914 			return (0);
2915 
2916 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
2917 
2918 		rval = dtrace_getvmreg(ndx,
2919 		    &cpu_core[CPU->cpu_id].cpuc_dtrace_flags);
2920 
2921 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
2922 
2923 		return (rval);
2924 	}
2925 
2926 	case DIF_VAR_CURTHREAD:
2927 		if (!dtrace_priv_kernel(state))
2928 			return (0);
2929 		return ((uint64_t)(uintptr_t)curthread);
2930 
2931 	case DIF_VAR_TIMESTAMP:
2932 		if (!(mstate->dtms_present & DTRACE_MSTATE_TIMESTAMP)) {
2933 			mstate->dtms_timestamp = dtrace_gethrtime();
2934 			mstate->dtms_present |= DTRACE_MSTATE_TIMESTAMP;
2935 		}
2936 		return (mstate->dtms_timestamp);
2937 
2938 	case DIF_VAR_VTIMESTAMP:
2939 		ASSERT(dtrace_vtime_references != 0);
2940 		return (curthread->t_dtrace_vtime);
2941 
2942 	case DIF_VAR_WALLTIMESTAMP:
2943 		if (!(mstate->dtms_present & DTRACE_MSTATE_WALLTIMESTAMP)) {
2944 			mstate->dtms_walltimestamp = dtrace_gethrestime();
2945 			mstate->dtms_present |= DTRACE_MSTATE_WALLTIMESTAMP;
2946 		}
2947 		return (mstate->dtms_walltimestamp);
2948 
2949 	case DIF_VAR_IPL:
2950 		if (!dtrace_priv_kernel(state))
2951 			return (0);
2952 		if (!(mstate->dtms_present & DTRACE_MSTATE_IPL)) {
2953 			mstate->dtms_ipl = dtrace_getipl();
2954 			mstate->dtms_present |= DTRACE_MSTATE_IPL;
2955 		}
2956 		return (mstate->dtms_ipl);
2957 
2958 	case DIF_VAR_EPID:
2959 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_EPID);
2960 		return (mstate->dtms_epid);
2961 
2962 	case DIF_VAR_ID:
2963 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
2964 		return (mstate->dtms_probe->dtpr_id);
2965 
2966 	case DIF_VAR_STACKDEPTH:
2967 		if (!dtrace_priv_kernel(state))
2968 			return (0);
2969 		if (!(mstate->dtms_present & DTRACE_MSTATE_STACKDEPTH)) {
2970 			int aframes = mstate->dtms_probe->dtpr_aframes + 2;
2971 
2972 			mstate->dtms_stackdepth = dtrace_getstackdepth(aframes);
2973 			mstate->dtms_present |= DTRACE_MSTATE_STACKDEPTH;
2974 		}
2975 		return (mstate->dtms_stackdepth);
2976 
2977 	case DIF_VAR_USTACKDEPTH:
2978 		if (!dtrace_priv_proc(state, mstate))
2979 			return (0);
2980 		if (!(mstate->dtms_present & DTRACE_MSTATE_USTACKDEPTH)) {
2981 			/*
2982 			 * See comment in DIF_VAR_PID.
2983 			 */
2984 			if (DTRACE_ANCHORED(mstate->dtms_probe) &&
2985 			    CPU_ON_INTR(CPU)) {
2986 				mstate->dtms_ustackdepth = 0;
2987 			} else {
2988 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
2989 				mstate->dtms_ustackdepth =
2990 				    dtrace_getustackdepth();
2991 				DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
2992 			}
2993 			mstate->dtms_present |= DTRACE_MSTATE_USTACKDEPTH;
2994 		}
2995 		return (mstate->dtms_ustackdepth);
2996 
2997 	case DIF_VAR_CALLER:
2998 		if (!dtrace_priv_kernel(state))
2999 			return (0);
3000 		if (!(mstate->dtms_present & DTRACE_MSTATE_CALLER)) {
3001 			int aframes = mstate->dtms_probe->dtpr_aframes + 2;
3002 
3003 			if (!DTRACE_ANCHORED(mstate->dtms_probe)) {
3004 				/*
3005 				 * If this is an unanchored probe, we are
3006 				 * required to go through the slow path:
3007 				 * dtrace_caller() only guarantees correct
3008 				 * results for anchored probes.
3009 				 */
3010 				pc_t caller[2];
3011 
3012 				dtrace_getpcstack(caller, 2, aframes,
3013 				    (uint32_t *)(uintptr_t)mstate->dtms_arg[0]);
3014 				mstate->dtms_caller = caller[1];
3015 			} else if ((mstate->dtms_caller =
3016 			    dtrace_caller(aframes)) == -1) {
3017 				/*
3018 				 * We have failed to do this the quick way;
3019 				 * we must resort to the slower approach of
3020 				 * calling dtrace_getpcstack().
3021 				 */
3022 				pc_t caller;
3023 
3024 				dtrace_getpcstack(&caller, 1, aframes, NULL);
3025 				mstate->dtms_caller = caller;
3026 			}
3027 
3028 			mstate->dtms_present |= DTRACE_MSTATE_CALLER;
3029 		}
3030 		return (mstate->dtms_caller);
3031 
3032 	case DIF_VAR_UCALLER:
3033 		if (!dtrace_priv_proc(state, mstate))
3034 			return (0);
3035 
3036 		if (!(mstate->dtms_present & DTRACE_MSTATE_UCALLER)) {
3037 			uint64_t ustack[3];
3038 
3039 			/*
3040 			 * dtrace_getupcstack() fills in the first uint64_t
3041 			 * with the current PID.  The second uint64_t will
3042 			 * be the program counter at user-level.  The third
3043 			 * uint64_t will contain the caller, which is what
3044 			 * we're after.
3045 			 */
3046 			ustack[2] = NULL;
3047 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3048 			dtrace_getupcstack(ustack, 3);
3049 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3050 			mstate->dtms_ucaller = ustack[2];
3051 			mstate->dtms_present |= DTRACE_MSTATE_UCALLER;
3052 		}
3053 
3054 		return (mstate->dtms_ucaller);
3055 
3056 	case DIF_VAR_PROBEPROV:
3057 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3058 		return (dtrace_dif_varstr(
3059 		    (uintptr_t)mstate->dtms_probe->dtpr_provider->dtpv_name,
3060 		    state, mstate));
3061 
3062 	case DIF_VAR_PROBEMOD:
3063 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3064 		return (dtrace_dif_varstr(
3065 		    (uintptr_t)mstate->dtms_probe->dtpr_mod,
3066 		    state, mstate));
3067 
3068 	case DIF_VAR_PROBEFUNC:
3069 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3070 		return (dtrace_dif_varstr(
3071 		    (uintptr_t)mstate->dtms_probe->dtpr_func,
3072 		    state, mstate));
3073 
3074 	case DIF_VAR_PROBENAME:
3075 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3076 		return (dtrace_dif_varstr(
3077 		    (uintptr_t)mstate->dtms_probe->dtpr_name,
3078 		    state, mstate));
3079 
3080 	case DIF_VAR_PID:
3081 		if (!dtrace_priv_proc(state, mstate))
3082 			return (0);
3083 
3084 		/*
3085 		 * Note that we are assuming that an unanchored probe is
3086 		 * always due to a high-level interrupt.  (And we're assuming
3087 		 * that there is only a single high level interrupt.)
3088 		 */
3089 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3090 			return (pid0.pid_id);
3091 
3092 		/*
3093 		 * It is always safe to dereference one's own t_procp pointer:
3094 		 * it always points to a valid, allocated proc structure.
3095 		 * Further, it is always safe to dereference the p_pidp member
3096 		 * of one's own proc structure.  (These are truisms becuase
3097 		 * threads and processes don't clean up their own state --
3098 		 * they leave that task to whomever reaps them.)
3099 		 */
3100 		return ((uint64_t)curthread->t_procp->p_pidp->pid_id);
3101 
3102 	case DIF_VAR_PPID:
3103 		if (!dtrace_priv_proc(state, mstate))
3104 			return (0);
3105 
3106 		/*
3107 		 * See comment in DIF_VAR_PID.
3108 		 */
3109 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3110 			return (pid0.pid_id);
3111 
3112 		/*
3113 		 * It is always safe to dereference one's own t_procp pointer:
3114 		 * it always points to a valid, allocated proc structure.
3115 		 * (This is true because threads don't clean up their own
3116 		 * state -- they leave that task to whomever reaps them.)
3117 		 */
3118 		return ((uint64_t)curthread->t_procp->p_ppid);
3119 
3120 	case DIF_VAR_TID:
3121 		/*
3122 		 * See comment in DIF_VAR_PID.
3123 		 */
3124 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3125 			return (0);
3126 
3127 		return ((uint64_t)curthread->t_tid);
3128 
3129 	case DIF_VAR_EXECNAME:
3130 		if (!dtrace_priv_proc(state, mstate))
3131 			return (0);
3132 
3133 		/*
3134 		 * See comment in DIF_VAR_PID.
3135 		 */
3136 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3137 			return ((uint64_t)(uintptr_t)p0.p_user.u_comm);
3138 
3139 		/*
3140 		 * It is always safe to dereference one's own t_procp pointer:
3141 		 * it always points to a valid, allocated proc structure.
3142 		 * (This is true because threads don't clean up their own
3143 		 * state -- they leave that task to whomever reaps them.)
3144 		 */
3145 		return (dtrace_dif_varstr(
3146 		    (uintptr_t)curthread->t_procp->p_user.u_comm,
3147 		    state, mstate));
3148 
3149 	case DIF_VAR_ZONENAME:
3150 		if (!dtrace_priv_proc(state, mstate))
3151 			return (0);
3152 
3153 		/*
3154 		 * See comment in DIF_VAR_PID.
3155 		 */
3156 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3157 			return ((uint64_t)(uintptr_t)p0.p_zone->zone_name);
3158 
3159 		/*
3160 		 * It is always safe to dereference one's own t_procp pointer:
3161 		 * it always points to a valid, allocated proc structure.
3162 		 * (This is true because threads don't clean up their own
3163 		 * state -- they leave that task to whomever reaps them.)
3164 		 */
3165 		return (dtrace_dif_varstr(
3166 		    (uintptr_t)curthread->t_procp->p_zone->zone_name,
3167 		    state, mstate));
3168 
3169 	case DIF_VAR_UID:
3170 		if (!dtrace_priv_proc(state, mstate))
3171 			return (0);
3172 
3173 		/*
3174 		 * See comment in DIF_VAR_PID.
3175 		 */
3176 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3177 			return ((uint64_t)p0.p_cred->cr_uid);
3178 
3179 		/*
3180 		 * It is always safe to dereference one's own t_procp pointer:
3181 		 * it always points to a valid, allocated proc structure.
3182 		 * (This is true because threads don't clean up their own
3183 		 * state -- they leave that task to whomever reaps them.)
3184 		 *
3185 		 * Additionally, it is safe to dereference one's own process
3186 		 * credential, since this is never NULL after process birth.
3187 		 */
3188 		return ((uint64_t)curthread->t_procp->p_cred->cr_uid);
3189 
3190 	case DIF_VAR_GID:
3191 		if (!dtrace_priv_proc(state, mstate))
3192 			return (0);
3193 
3194 		/*
3195 		 * See comment in DIF_VAR_PID.
3196 		 */
3197 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3198 			return ((uint64_t)p0.p_cred->cr_gid);
3199 
3200 		/*
3201 		 * It is always safe to dereference one's own t_procp pointer:
3202 		 * it always points to a valid, allocated proc structure.
3203 		 * (This is true because threads don't clean up their own
3204 		 * state -- they leave that task to whomever reaps them.)
3205 		 *
3206 		 * Additionally, it is safe to dereference one's own process
3207 		 * credential, since this is never NULL after process birth.
3208 		 */
3209 		return ((uint64_t)curthread->t_procp->p_cred->cr_gid);
3210 
3211 	case DIF_VAR_ERRNO: {
3212 		klwp_t *lwp;
3213 		if (!dtrace_priv_proc(state, mstate))
3214 			return (0);
3215 
3216 		/*
3217 		 * See comment in DIF_VAR_PID.
3218 		 */
3219 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3220 			return (0);
3221 
3222 		/*
3223 		 * It is always safe to dereference one's own t_lwp pointer in
3224 		 * the event that this pointer is non-NULL.  (This is true
3225 		 * because threads and lwps don't clean up their own state --
3226 		 * they leave that task to whomever reaps them.)
3227 		 */
3228 		if ((lwp = curthread->t_lwp) == NULL)
3229 			return (0);
3230 
3231 		return ((uint64_t)lwp->lwp_errno);
3232 	}
3233 	default:
3234 		DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
3235 		return (0);
3236 	}
3237 }
3238 
3239 /*
3240  * Emulate the execution of DTrace ID subroutines invoked by the call opcode.
3241  * Notice that we don't bother validating the proper number of arguments or
3242  * their types in the tuple stack.  This isn't needed because all argument
3243  * interpretation is safe because of our load safety -- the worst that can
3244  * happen is that a bogus program can obtain bogus results.
3245  */
3246 static void
3247 dtrace_dif_subr(uint_t subr, uint_t rd, uint64_t *regs,
3248     dtrace_key_t *tupregs, int nargs,
3249     dtrace_mstate_t *mstate, dtrace_state_t *state)
3250 {
3251 	volatile uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
3252 	volatile uintptr_t *illval = &cpu_core[CPU->cpu_id].cpuc_dtrace_illval;
3253 	dtrace_vstate_t *vstate = &state->dts_vstate;
3254 
3255 	union {
3256 		mutex_impl_t mi;
3257 		uint64_t mx;
3258 	} m;
3259 
3260 	union {
3261 		krwlock_t ri;
3262 		uintptr_t rw;
3263 	} r;
3264 
3265 	switch (subr) {
3266 	case DIF_SUBR_RAND:
3267 		regs[rd] = (dtrace_gethrtime() * 2416 + 374441) % 1771875;
3268 		break;
3269 
3270 	case DIF_SUBR_MUTEX_OWNED:
3271 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3272 		    mstate, vstate)) {
3273 			regs[rd] = NULL;
3274 			break;
3275 		}
3276 
3277 		m.mx = dtrace_load64(tupregs[0].dttk_value);
3278 		if (MUTEX_TYPE_ADAPTIVE(&m.mi))
3279 			regs[rd] = MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER;
3280 		else
3281 			regs[rd] = LOCK_HELD(&m.mi.m_spin.m_spinlock);
3282 		break;
3283 
3284 	case DIF_SUBR_MUTEX_OWNER:
3285 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3286 		    mstate, vstate)) {
3287 			regs[rd] = NULL;
3288 			break;
3289 		}
3290 
3291 		m.mx = dtrace_load64(tupregs[0].dttk_value);
3292 		if (MUTEX_TYPE_ADAPTIVE(&m.mi) &&
3293 		    MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER)
3294 			regs[rd] = (uintptr_t)MUTEX_OWNER(&m.mi);
3295 		else
3296 			regs[rd] = 0;
3297 		break;
3298 
3299 	case DIF_SUBR_MUTEX_TYPE_ADAPTIVE:
3300 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3301 		    mstate, vstate)) {
3302 			regs[rd] = NULL;
3303 			break;
3304 		}
3305 
3306 		m.mx = dtrace_load64(tupregs[0].dttk_value);
3307 		regs[rd] = MUTEX_TYPE_ADAPTIVE(&m.mi);
3308 		break;
3309 
3310 	case DIF_SUBR_MUTEX_TYPE_SPIN:
3311 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3312 		    mstate, vstate)) {
3313 			regs[rd] = NULL;
3314 			break;
3315 		}
3316 
3317 		m.mx = dtrace_load64(tupregs[0].dttk_value);
3318 		regs[rd] = MUTEX_TYPE_SPIN(&m.mi);
3319 		break;
3320 
3321 	case DIF_SUBR_RW_READ_HELD: {
3322 		uintptr_t tmp;
3323 
3324 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
3325 		    mstate, vstate)) {
3326 			regs[rd] = NULL;
3327 			break;
3328 		}
3329 
3330 		r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3331 		regs[rd] = _RW_READ_HELD(&r.ri, tmp);
3332 		break;
3333 	}
3334 
3335 	case DIF_SUBR_RW_WRITE_HELD:
3336 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
3337 		    mstate, vstate)) {
3338 			regs[rd] = NULL;
3339 			break;
3340 		}
3341 
3342 		r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3343 		regs[rd] = _RW_WRITE_HELD(&r.ri);
3344 		break;
3345 
3346 	case DIF_SUBR_RW_ISWRITER:
3347 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
3348 		    mstate, vstate)) {
3349 			regs[rd] = NULL;
3350 			break;
3351 		}
3352 
3353 		r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3354 		regs[rd] = _RW_ISWRITER(&r.ri);
3355 		break;
3356 
3357 	case DIF_SUBR_BCOPY: {
3358 		/*
3359 		 * We need to be sure that the destination is in the scratch
3360 		 * region -- no other region is allowed.
3361 		 */
3362 		uintptr_t src = tupregs[0].dttk_value;
3363 		uintptr_t dest = tupregs[1].dttk_value;
3364 		size_t size = tupregs[2].dttk_value;
3365 
3366 		if (!dtrace_inscratch(dest, size, mstate)) {
3367 			*flags |= CPU_DTRACE_BADADDR;
3368 			*illval = regs[rd];
3369 			break;
3370 		}
3371 
3372 		if (!dtrace_canload(src, size, mstate, vstate)) {
3373 			regs[rd] = NULL;
3374 			break;
3375 		}
3376 
3377 		dtrace_bcopy((void *)src, (void *)dest, size);
3378 		break;
3379 	}
3380 
3381 	case DIF_SUBR_ALLOCA:
3382 	case DIF_SUBR_COPYIN: {
3383 		uintptr_t dest = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
3384 		uint64_t size =
3385 		    tupregs[subr == DIF_SUBR_ALLOCA ? 0 : 1].dttk_value;
3386 		size_t scratch_size = (dest - mstate->dtms_scratch_ptr) + size;
3387 
3388 		/*
3389 		 * This action doesn't require any credential checks since
3390 		 * probes will not activate in user contexts to which the
3391 		 * enabling user does not have permissions.
3392 		 */
3393 
3394 		/*
3395 		 * Rounding up the user allocation size could have overflowed
3396 		 * a large, bogus allocation (like -1ULL) to 0.
3397 		 */
3398 		if (scratch_size < size ||
3399 		    !DTRACE_INSCRATCH(mstate, scratch_size)) {
3400 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3401 			regs[rd] = NULL;
3402 			break;
3403 		}
3404 
3405 		if (subr == DIF_SUBR_COPYIN) {
3406 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3407 			dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
3408 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3409 		}
3410 
3411 		mstate->dtms_scratch_ptr += scratch_size;
3412 		regs[rd] = dest;
3413 		break;
3414 	}
3415 
3416 	case DIF_SUBR_COPYINTO: {
3417 		uint64_t size = tupregs[1].dttk_value;
3418 		uintptr_t dest = tupregs[2].dttk_value;
3419 
3420 		/*
3421 		 * This action doesn't require any credential checks since
3422 		 * probes will not activate in user contexts to which the
3423 		 * enabling user does not have permissions.
3424 		 */
3425 		if (!dtrace_inscratch(dest, size, mstate)) {
3426 			*flags |= CPU_DTRACE_BADADDR;
3427 			*illval = regs[rd];
3428 			break;
3429 		}
3430 
3431 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3432 		dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
3433 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3434 		break;
3435 	}
3436 
3437 	case DIF_SUBR_COPYINSTR: {
3438 		uintptr_t dest = mstate->dtms_scratch_ptr;
3439 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3440 
3441 		if (nargs > 1 && tupregs[1].dttk_value < size)
3442 			size = tupregs[1].dttk_value + 1;
3443 
3444 		/*
3445 		 * This action doesn't require any credential checks since
3446 		 * probes will not activate in user contexts to which the
3447 		 * enabling user does not have permissions.
3448 		 */
3449 		if (!DTRACE_INSCRATCH(mstate, size)) {
3450 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3451 			regs[rd] = NULL;
3452 			break;
3453 		}
3454 
3455 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3456 		dtrace_copyinstr(tupregs[0].dttk_value, dest, size, flags);
3457 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3458 
3459 		((char *)dest)[size - 1] = '\0';
3460 		mstate->dtms_scratch_ptr += size;
3461 		regs[rd] = dest;
3462 		break;
3463 	}
3464 
3465 	case DIF_SUBR_MSGSIZE:
3466 	case DIF_SUBR_MSGDSIZE: {
3467 		uintptr_t baddr = tupregs[0].dttk_value, daddr;
3468 		uintptr_t wptr, rptr;
3469 		size_t count = 0;
3470 		int cont = 0;
3471 
3472 		while (baddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
3473 
3474 			if (!dtrace_canload(baddr, sizeof (mblk_t), mstate,
3475 			    vstate)) {
3476 				regs[rd] = NULL;
3477 				break;
3478 			}
3479 
3480 			wptr = dtrace_loadptr(baddr +
3481 			    offsetof(mblk_t, b_wptr));
3482 
3483 			rptr = dtrace_loadptr(baddr +
3484 			    offsetof(mblk_t, b_rptr));
3485 
3486 			if (wptr < rptr) {
3487 				*flags |= CPU_DTRACE_BADADDR;
3488 				*illval = tupregs[0].dttk_value;
3489 				break;
3490 			}
3491 
3492 			daddr = dtrace_loadptr(baddr +
3493 			    offsetof(mblk_t, b_datap));
3494 
3495 			baddr = dtrace_loadptr(baddr +
3496 			    offsetof(mblk_t, b_cont));
3497 
3498 			/*
3499 			 * We want to prevent against denial-of-service here,
3500 			 * so we're only going to search the list for
3501 			 * dtrace_msgdsize_max mblks.
3502 			 */
3503 			if (cont++ > dtrace_msgdsize_max) {
3504 				*flags |= CPU_DTRACE_ILLOP;
3505 				break;
3506 			}
3507 
3508 			if (subr == DIF_SUBR_MSGDSIZE) {
3509 				if (dtrace_load8(daddr +
3510 				    offsetof(dblk_t, db_type)) != M_DATA)
3511 					continue;
3512 			}
3513 
3514 			count += wptr - rptr;
3515 		}
3516 
3517 		if (!(*flags & CPU_DTRACE_FAULT))
3518 			regs[rd] = count;
3519 
3520 		break;
3521 	}
3522 
3523 	case DIF_SUBR_PROGENYOF: {
3524 		pid_t pid = tupregs[0].dttk_value;
3525 		proc_t *p;
3526 		int rval = 0;
3527 
3528 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3529 
3530 		for (p = curthread->t_procp; p != NULL; p = p->p_parent) {
3531 			if (p->p_pidp->pid_id == pid) {
3532 				rval = 1;
3533 				break;
3534 			}
3535 		}
3536 
3537 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3538 
3539 		regs[rd] = rval;
3540 		break;
3541 	}
3542 
3543 	case DIF_SUBR_SPECULATION:
3544 		regs[rd] = dtrace_speculation(state);
3545 		break;
3546 
3547 	case DIF_SUBR_COPYOUT: {
3548 		uintptr_t kaddr = tupregs[0].dttk_value;
3549 		uintptr_t uaddr = tupregs[1].dttk_value;
3550 		uint64_t size = tupregs[2].dttk_value;
3551 
3552 		if (!dtrace_destructive_disallow &&
3553 		    dtrace_priv_proc_control(state, mstate) &&
3554 		    !dtrace_istoxic(kaddr, size)) {
3555 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3556 			dtrace_copyout(kaddr, uaddr, size, flags);
3557 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3558 		}
3559 		break;
3560 	}
3561 
3562 	case DIF_SUBR_COPYOUTSTR: {
3563 		uintptr_t kaddr = tupregs[0].dttk_value;
3564 		uintptr_t uaddr = tupregs[1].dttk_value;
3565 		uint64_t size = tupregs[2].dttk_value;
3566 
3567 		if (!dtrace_destructive_disallow &&
3568 		    dtrace_priv_proc_control(state, mstate) &&
3569 		    !dtrace_istoxic(kaddr, size)) {
3570 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3571 			dtrace_copyoutstr(kaddr, uaddr, size, flags);
3572 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3573 		}
3574 		break;
3575 	}
3576 
3577 	case DIF_SUBR_STRLEN: {
3578 		size_t sz;
3579 		uintptr_t addr = (uintptr_t)tupregs[0].dttk_value;
3580 		sz = dtrace_strlen((char *)addr,
3581 		    state->dts_options[DTRACEOPT_STRSIZE]);
3582 
3583 		if (!dtrace_canload(addr, sz + 1, mstate, vstate)) {
3584 			regs[rd] = NULL;
3585 			break;
3586 		}
3587 
3588 		regs[rd] = sz;
3589 
3590 		break;
3591 	}
3592 
3593 	case DIF_SUBR_STRCHR:
3594 	case DIF_SUBR_STRRCHR: {
3595 		/*
3596 		 * We're going to iterate over the string looking for the
3597 		 * specified character.  We will iterate until we have reached
3598 		 * the string length or we have found the character.  If this
3599 		 * is DIF_SUBR_STRRCHR, we will look for the last occurrence
3600 		 * of the specified character instead of the first.
3601 		 */
3602 		uintptr_t saddr = tupregs[0].dttk_value;
3603 		uintptr_t addr = tupregs[0].dttk_value;
3604 		uintptr_t limit = addr + state->dts_options[DTRACEOPT_STRSIZE];
3605 		char c, target = (char)tupregs[1].dttk_value;
3606 
3607 		for (regs[rd] = NULL; addr < limit; addr++) {
3608 			if ((c = dtrace_load8(addr)) == target) {
3609 				regs[rd] = addr;
3610 
3611 				if (subr == DIF_SUBR_STRCHR)
3612 					break;
3613 			}
3614 
3615 			if (c == '\0')
3616 				break;
3617 		}
3618 
3619 		if (!dtrace_canload(saddr, addr - saddr, mstate, vstate)) {
3620 			regs[rd] = NULL;
3621 			break;
3622 		}
3623 
3624 		break;
3625 	}
3626 
3627 	case DIF_SUBR_STRSTR:
3628 	case DIF_SUBR_INDEX:
3629 	case DIF_SUBR_RINDEX: {
3630 		/*
3631 		 * We're going to iterate over the string looking for the
3632 		 * specified string.  We will iterate until we have reached
3633 		 * the string length or we have found the string.  (Yes, this
3634 		 * is done in the most naive way possible -- but considering
3635 		 * that the string we're searching for is likely to be
3636 		 * relatively short, the complexity of Rabin-Karp or similar
3637 		 * hardly seems merited.)
3638 		 */
3639 		char *addr = (char *)(uintptr_t)tupregs[0].dttk_value;
3640 		char *substr = (char *)(uintptr_t)tupregs[1].dttk_value;
3641 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3642 		size_t len = dtrace_strlen(addr, size);
3643 		size_t sublen = dtrace_strlen(substr, size);
3644 		char *limit = addr + len, *orig = addr;
3645 		int notfound = subr == DIF_SUBR_STRSTR ? 0 : -1;
3646 		int inc = 1;
3647 
3648 		regs[rd] = notfound;
3649 
3650 		if (!dtrace_canload((uintptr_t)addr, len + 1, mstate, vstate)) {
3651 			regs[rd] = NULL;
3652 			break;
3653 		}
3654 
3655 		if (!dtrace_canload((uintptr_t)substr, sublen + 1, mstate,
3656 		    vstate)) {
3657 			regs[rd] = NULL;
3658 			break;
3659 		}
3660 
3661 		/*
3662 		 * strstr() and index()/rindex() have similar semantics if
3663 		 * both strings are the empty string: strstr() returns a
3664 		 * pointer to the (empty) string, and index() and rindex()
3665 		 * both return index 0 (regardless of any position argument).
3666 		 */
3667 		if (sublen == 0 && len == 0) {
3668 			if (subr == DIF_SUBR_STRSTR)
3669 				regs[rd] = (uintptr_t)addr;
3670 			else
3671 				regs[rd] = 0;
3672 			break;
3673 		}
3674 
3675 		if (subr != DIF_SUBR_STRSTR) {
3676 			if (subr == DIF_SUBR_RINDEX) {
3677 				limit = orig - 1;
3678 				addr += len;
3679 				inc = -1;
3680 			}
3681 
3682 			/*
3683 			 * Both index() and rindex() take an optional position
3684 			 * argument that denotes the starting position.
3685 			 */
3686 			if (nargs == 3) {
3687 				int64_t pos = (int64_t)tupregs[2].dttk_value;
3688 
3689 				/*
3690 				 * If the position argument to index() is
3691 				 * negative, Perl implicitly clamps it at
3692 				 * zero.  This semantic is a little surprising
3693 				 * given the special meaning of negative
3694 				 * positions to similar Perl functions like
3695 				 * substr(), but it appears to reflect a
3696 				 * notion that index() can start from a
3697 				 * negative index and increment its way up to
3698 				 * the string.  Given this notion, Perl's
3699 				 * rindex() is at least self-consistent in
3700 				 * that it implicitly clamps positions greater
3701 				 * than the string length to be the string
3702 				 * length.  Where Perl completely loses
3703 				 * coherence, however, is when the specified
3704 				 * substring is the empty string ("").  In
3705 				 * this case, even if the position is
3706 				 * negative, rindex() returns 0 -- and even if
3707 				 * the position is greater than the length,
3708 				 * index() returns the string length.  These
3709 				 * semantics violate the notion that index()
3710 				 * should never return a value less than the
3711 				 * specified position and that rindex() should
3712 				 * never return a value greater than the
3713 				 * specified position.  (One assumes that
3714 				 * these semantics are artifacts of Perl's
3715 				 * implementation and not the results of
3716 				 * deliberate design -- it beggars belief that
3717 				 * even Larry Wall could desire such oddness.)
3718 				 * While in the abstract one would wish for
3719 				 * consistent position semantics across
3720 				 * substr(), index() and rindex() -- or at the
3721 				 * very least self-consistent position
3722 				 * semantics for index() and rindex() -- we
3723 				 * instead opt to keep with the extant Perl
3724 				 * semantics, in all their broken glory.  (Do
3725 				 * we have more desire to maintain Perl's
3726 				 * semantics than Perl does?  Probably.)
3727 				 */
3728 				if (subr == DIF_SUBR_RINDEX) {
3729 					if (pos < 0) {
3730 						if (sublen == 0)
3731 							regs[rd] = 0;
3732 						break;
3733 					}
3734 
3735 					if (pos > len)
3736 						pos = len;
3737 				} else {
3738 					if (pos < 0)
3739 						pos = 0;
3740 
3741 					if (pos >= len) {
3742 						if (sublen == 0)
3743 							regs[rd] = len;
3744 						break;
3745 					}
3746 				}
3747 
3748 				addr = orig + pos;
3749 			}
3750 		}
3751 
3752 		for (regs[rd] = notfound; addr != limit; addr += inc) {
3753 			if (dtrace_strncmp(addr, substr, sublen) == 0) {
3754 				if (subr != DIF_SUBR_STRSTR) {
3755 					/*
3756 					 * As D index() and rindex() are
3757 					 * modeled on Perl (and not on awk),
3758 					 * we return a zero-based (and not a
3759 					 * one-based) index.  (For you Perl
3760 					 * weenies: no, we're not going to add
3761 					 * $[ -- and shouldn't you be at a con
3762 					 * or something?)
3763 					 */
3764 					regs[rd] = (uintptr_t)(addr - orig);
3765 					break;
3766 				}
3767 
3768 				ASSERT(subr == DIF_SUBR_STRSTR);
3769 				regs[rd] = (uintptr_t)addr;
3770 				break;
3771 			}
3772 		}
3773 
3774 		break;
3775 	}
3776 
3777 	case DIF_SUBR_STRTOK: {
3778 		uintptr_t addr = tupregs[0].dttk_value;
3779 		uintptr_t tokaddr = tupregs[1].dttk_value;
3780 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3781 		uintptr_t limit, toklimit = tokaddr + size;
3782 		uint8_t c, tokmap[32];	 /* 256 / 8 */
3783 		char *dest = (char *)mstate->dtms_scratch_ptr;
3784 		int i;
3785 
3786 		/*
3787 		 * Check both the token buffer and (later) the input buffer,
3788 		 * since both could be non-scratch addresses.
3789 		 */
3790 		if (!dtrace_strcanload(tokaddr, size, mstate, vstate)) {
3791 			regs[rd] = NULL;
3792 			break;
3793 		}
3794 
3795 		if (!DTRACE_INSCRATCH(mstate, size)) {
3796 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3797 			regs[rd] = NULL;
3798 			break;
3799 		}
3800 
3801 		if (addr == NULL) {
3802 			/*
3803 			 * If the address specified is NULL, we use our saved
3804 			 * strtok pointer from the mstate.  Note that this
3805 			 * means that the saved strtok pointer is _only_
3806 			 * valid within multiple enablings of the same probe --
3807 			 * it behaves like an implicit clause-local variable.
3808 			 */
3809 			addr = mstate->dtms_strtok;
3810 		} else {
3811 			/*
3812 			 * If the user-specified address is non-NULL we must
3813 			 * access check it.  This is the only time we have
3814 			 * a chance to do so, since this address may reside
3815 			 * in the string table of this clause-- future calls
3816 			 * (when we fetch addr from mstate->dtms_strtok)
3817 			 * would fail this access check.
3818 			 */
3819 			if (!dtrace_strcanload(addr, size, mstate, vstate)) {
3820 				regs[rd] = NULL;
3821 				break;
3822 			}
3823 		}
3824 
3825 		/*
3826 		 * First, zero the token map, and then process the token
3827 		 * string -- setting a bit in the map for every character
3828 		 * found in the token string.
3829 		 */
3830 		for (i = 0; i < sizeof (tokmap); i++)
3831 			tokmap[i] = 0;
3832 
3833 		for (; tokaddr < toklimit; tokaddr++) {
3834 			if ((c = dtrace_load8(tokaddr)) == '\0')
3835 				break;
3836 
3837 			ASSERT((c >> 3) < sizeof (tokmap));
3838 			tokmap[c >> 3] |= (1 << (c & 0x7));
3839 		}
3840 
3841 		for (limit = addr + size; addr < limit; addr++) {
3842 			/*
3843 			 * We're looking for a character that is _not_ contained
3844 			 * in the token string.
3845 			 */
3846 			if ((c = dtrace_load8(addr)) == '\0')
3847 				break;
3848 
3849 			if (!(tokmap[c >> 3] & (1 << (c & 0x7))))
3850 				break;
3851 		}
3852 
3853 		if (c == '\0') {
3854 			/*
3855 			 * We reached the end of the string without finding
3856 			 * any character that was not in the token string.
3857 			 * We return NULL in this case, and we set the saved
3858 			 * address to NULL as well.
3859 			 */
3860 			regs[rd] = NULL;
3861 			mstate->dtms_strtok = NULL;
3862 			break;
3863 		}
3864 
3865 		/*
3866 		 * From here on, we're copying into the destination string.
3867 		 */
3868 		for (i = 0; addr < limit && i < size - 1; addr++) {
3869 			if ((c = dtrace_load8(addr)) == '\0')
3870 				break;
3871 
3872 			if (tokmap[c >> 3] & (1 << (c & 0x7)))
3873 				break;
3874 
3875 			ASSERT(i < size);
3876 			dest[i++] = c;
3877 		}
3878 
3879 		ASSERT(i < size);
3880 		dest[i] = '\0';
3881 		regs[rd] = (uintptr_t)dest;
3882 		mstate->dtms_scratch_ptr += size;
3883 		mstate->dtms_strtok = addr;
3884 		break;
3885 	}
3886 
3887 	case DIF_SUBR_SUBSTR: {
3888 		uintptr_t s = tupregs[0].dttk_value;
3889 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3890 		char *d = (char *)mstate->dtms_scratch_ptr;
3891 		int64_t index = (int64_t)tupregs[1].dttk_value;
3892 		int64_t remaining = (int64_t)tupregs[2].dttk_value;
3893 		size_t len = dtrace_strlen((char *)s, size);
3894 		int64_t i;
3895 
3896 		if (!dtrace_canload(s, len + 1, mstate, vstate)) {
3897 			regs[rd] = NULL;
3898 			break;
3899 		}
3900 
3901 		if (!DTRACE_INSCRATCH(mstate, size)) {
3902 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3903 			regs[rd] = NULL;
3904 			break;
3905 		}
3906 
3907 		if (nargs <= 2)
3908 			remaining = (int64_t)size;
3909 
3910 		if (index < 0) {
3911 			index += len;
3912 
3913 			if (index < 0 && index + remaining > 0) {
3914 				remaining += index;
3915 				index = 0;
3916 			}
3917 		}
3918 
3919 		if (index >= len || index < 0) {
3920 			remaining = 0;
3921 		} else if (remaining < 0) {
3922 			remaining += len - index;
3923 		} else if (index + remaining > size) {
3924 			remaining = size - index;
3925 		}
3926 
3927 		for (i = 0; i < remaining; i++) {
3928 			if ((d[i] = dtrace_load8(s + index + i)) == '\0')
3929 				break;
3930 		}
3931 
3932 		d[i] = '\0';
3933 
3934 		mstate->dtms_scratch_ptr += size;
3935 		regs[rd] = (uintptr_t)d;
3936 		break;
3937 	}
3938 
3939 	case DIF_SUBR_TOUPPER:
3940 	case DIF_SUBR_TOLOWER: {
3941 		uintptr_t s = tupregs[0].dttk_value;
3942 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3943 		char *dest = (char *)mstate->dtms_scratch_ptr, c;
3944 		size_t len = dtrace_strlen((char *)s, size);
3945 		char lower, upper, convert;
3946 		int64_t i;
3947 
3948 		if (subr == DIF_SUBR_TOUPPER) {
3949 			lower = 'a';
3950 			upper = 'z';
3951 			convert = 'A';
3952 		} else {
3953 			lower = 'A';
3954 			upper = 'Z';
3955 			convert = 'a';
3956 		}
3957 
3958 		if (!dtrace_canload(s, len + 1, mstate, vstate)) {
3959 			regs[rd] = NULL;
3960 			break;
3961 		}
3962 
3963 		if (!DTRACE_INSCRATCH(mstate, size)) {
3964 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3965 			regs[rd] = NULL;
3966 			break;
3967 		}
3968 
3969 		for (i = 0; i < size - 1; i++) {
3970 			if ((c = dtrace_load8(s + i)) == '\0')
3971 				break;
3972 
3973 			if (c >= lower && c <= upper)
3974 				c = convert + (c - lower);
3975 
3976 			dest[i] = c;
3977 		}
3978 
3979 		ASSERT(i < size);
3980 		dest[i] = '\0';
3981 		regs[rd] = (uintptr_t)dest;
3982 		mstate->dtms_scratch_ptr += size;
3983 		break;
3984 	}
3985 
3986 case DIF_SUBR_GETMAJOR:
3987 #ifdef _LP64
3988 		regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR64) & MAXMAJ64;
3989 #else
3990 		regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR) & MAXMAJ;
3991 #endif
3992 		break;
3993 
3994 	case DIF_SUBR_GETMINOR:
3995 #ifdef _LP64
3996 		regs[rd] = tupregs[0].dttk_value & MAXMIN64;
3997 #else
3998 		regs[rd] = tupregs[0].dttk_value & MAXMIN;
3999 #endif
4000 		break;
4001 
4002 	case DIF_SUBR_DDI_PATHNAME: {
4003 		/*
4004 		 * This one is a galactic mess.  We are going to roughly
4005 		 * emulate ddi_pathname(), but it's made more complicated
4006 		 * by the fact that we (a) want to include the minor name and
4007 		 * (b) must proceed iteratively instead of recursively.
4008 		 */
4009 		uintptr_t dest = mstate->dtms_scratch_ptr;
4010 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4011 		char *start = (char *)dest, *end = start + size - 1;
4012 		uintptr_t daddr = tupregs[0].dttk_value;
4013 		int64_t minor = (int64_t)tupregs[1].dttk_value;
4014 		char *s;
4015 		int i, len, depth = 0;
4016 
4017 		/*
4018 		 * Due to all the pointer jumping we do and context we must
4019 		 * rely upon, we just mandate that the user must have kernel
4020 		 * read privileges to use this routine.
4021 		 */
4022 		if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) == 0) {
4023 			*flags |= CPU_DTRACE_KPRIV;
4024 			*illval = daddr;
4025 			regs[rd] = NULL;
4026 		}
4027 
4028 		if (!DTRACE_INSCRATCH(mstate, size)) {
4029 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4030 			regs[rd] = NULL;
4031 			break;
4032 		}
4033 
4034 		*end = '\0';
4035 
4036 		/*
4037 		 * We want to have a name for the minor.  In order to do this,
4038 		 * we need to walk the minor list from the devinfo.  We want
4039 		 * to be sure that we don't infinitely walk a circular list,
4040 		 * so we check for circularity by sending a scout pointer
4041 		 * ahead two elements for every element that we iterate over;
4042 		 * if the list is circular, these will ultimately point to the
4043 		 * same element.  You may recognize this little trick as the
4044 		 * answer to a stupid interview question -- one that always
4045 		 * seems to be asked by those who had to have it laboriously
4046 		 * explained to them, and who can't even concisely describe
4047 		 * the conditions under which one would be forced to resort to
4048 		 * this technique.  Needless to say, those conditions are
4049 		 * found here -- and probably only here.  Is this the only use
4050 		 * of this infamous trick in shipping, production code?  If it
4051 		 * isn't, it probably should be...
4052 		 */
4053 		if (minor != -1) {
4054 			uintptr_t maddr = dtrace_loadptr(daddr +
4055 			    offsetof(struct dev_info, devi_minor));
4056 
4057 			uintptr_t next = offsetof(struct ddi_minor_data, next);
4058 			uintptr_t name = offsetof(struct ddi_minor_data,
4059 			    d_minor) + offsetof(struct ddi_minor, name);
4060 			uintptr_t dev = offsetof(struct ddi_minor_data,
4061 			    d_minor) + offsetof(struct ddi_minor, dev);
4062 			uintptr_t scout;
4063 
4064 			if (maddr != NULL)
4065 				scout = dtrace_loadptr(maddr + next);
4066 
4067 			while (maddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
4068 				uint64_t m;
4069 #ifdef _LP64
4070 				m = dtrace_load64(maddr + dev) & MAXMIN64;
4071 #else
4072 				m = dtrace_load32(maddr + dev) & MAXMIN;
4073 #endif
4074 				if (m != minor) {
4075 					maddr = dtrace_loadptr(maddr + next);
4076 
4077 					if (scout == NULL)
4078 						continue;
4079 
4080 					scout = dtrace_loadptr(scout + next);
4081 
4082 					if (scout == NULL)
4083 						continue;
4084 
4085 					scout = dtrace_loadptr(scout + next);
4086 
4087 					if (scout == NULL)
4088 						continue;
4089 
4090 					if (scout == maddr) {
4091 						*flags |= CPU_DTRACE_ILLOP;
4092 						break;
4093 					}
4094 
4095 					continue;
4096 				}
4097 
4098 				/*
4099 				 * We have the minor data.  Now we need to
4100 				 * copy the minor's name into the end of the
4101 				 * pathname.
4102 				 */
4103 				s = (char *)dtrace_loadptr(maddr + name);
4104 				len = dtrace_strlen(s, size);
4105 
4106 				if (*flags & CPU_DTRACE_FAULT)
4107 					break;
4108 
4109 				if (len != 0) {
4110 					if ((end -= (len + 1)) < start)
4111 						break;
4112 
4113 					*end = ':';
4114 				}
4115 
4116 				for (i = 1; i <= len; i++)
4117 					end[i] = dtrace_load8((uintptr_t)s++);
4118 				break;
4119 			}
4120 		}
4121 
4122 		while (daddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
4123 			ddi_node_state_t devi_state;
4124 
4125 			devi_state = dtrace_load32(daddr +
4126 			    offsetof(struct dev_info, devi_node_state));
4127 
4128 			if (*flags & CPU_DTRACE_FAULT)
4129 				break;
4130 
4131 			if (devi_state >= DS_INITIALIZED) {
4132 				s = (char *)dtrace_loadptr(daddr +
4133 				    offsetof(struct dev_info, devi_addr));
4134 				len = dtrace_strlen(s, size);
4135 
4136 				if (*flags & CPU_DTRACE_FAULT)
4137 					break;
4138 
4139 				if (len != 0) {
4140 					if ((end -= (len + 1)) < start)
4141 						break;
4142 
4143 					*end = '@';
4144 				}
4145 
4146 				for (i = 1; i <= len; i++)
4147 					end[i] = dtrace_load8((uintptr_t)s++);
4148 			}
4149 
4150 			/*
4151 			 * Now for the node name...
4152 			 */
4153 			s = (char *)dtrace_loadptr(daddr +
4154 			    offsetof(struct dev_info, devi_node_name));
4155 
4156 			daddr = dtrace_loadptr(daddr +
4157 			    offsetof(struct dev_info, devi_parent));
4158 
4159 			/*
4160 			 * If our parent is NULL (that is, if we're the root
4161 			 * node), we're going to use the special path
4162 			 * "devices".
4163 			 */
4164 			if (daddr == NULL)
4165 				s = "devices";
4166 
4167 			len = dtrace_strlen(s, size);
4168 			if (*flags & CPU_DTRACE_FAULT)
4169 				break;
4170 
4171 			if ((end -= (len + 1)) < start)
4172 				break;
4173 
4174 			for (i = 1; i <= len; i++)
4175 				end[i] = dtrace_load8((uintptr_t)s++);
4176 			*end = '/';
4177 
4178 			if (depth++ > dtrace_devdepth_max) {
4179 				*flags |= CPU_DTRACE_ILLOP;
4180 				break;
4181 			}
4182 		}
4183 
4184 		if (end < start)
4185 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4186 
4187 		if (daddr == NULL) {
4188 			regs[rd] = (uintptr_t)end;
4189 			mstate->dtms_scratch_ptr += size;
4190 		}
4191 
4192 		break;
4193 	}
4194 
4195 	case DIF_SUBR_STRJOIN: {
4196 		char *d = (char *)mstate->dtms_scratch_ptr;
4197 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4198 		uintptr_t s1 = tupregs[0].dttk_value;
4199 		uintptr_t s2 = tupregs[1].dttk_value;
4200 		int i = 0;
4201 
4202 		if (!dtrace_strcanload(s1, size, mstate, vstate) ||
4203 		    !dtrace_strcanload(s2, size, mstate, vstate)) {
4204 			regs[rd] = NULL;
4205 			break;
4206 		}
4207 
4208 		if (!DTRACE_INSCRATCH(mstate, size)) {
4209 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4210 			regs[rd] = NULL;
4211 			break;
4212 		}
4213 
4214 		for (;;) {
4215 			if (i >= size) {
4216 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4217 				regs[rd] = NULL;
4218 				break;
4219 			}
4220 
4221 			if ((d[i++] = dtrace_load8(s1++)) == '\0') {
4222 				i--;
4223 				break;
4224 			}
4225 		}
4226 
4227 		for (;;) {
4228 			if (i >= size) {
4229 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4230 				regs[rd] = NULL;
4231 				break;
4232 			}
4233 
4234 			if ((d[i++] = dtrace_load8(s2++)) == '\0')
4235 				break;
4236 		}
4237 
4238 		if (i < size) {
4239 			mstate->dtms_scratch_ptr += i;
4240 			regs[rd] = (uintptr_t)d;
4241 		}
4242 
4243 		break;
4244 	}
4245 
4246 	case DIF_SUBR_LLTOSTR: {
4247 		int64_t i = (int64_t)tupregs[0].dttk_value;
4248 		uint64_t val, digit;
4249 		uint64_t size = 65;	/* enough room for 2^64 in binary */
4250 		char *end = (char *)mstate->dtms_scratch_ptr + size - 1;
4251 		int base = 10;
4252 
4253 		if (nargs > 1) {
4254 			if ((base = tupregs[1].dttk_value) <= 1 ||
4255 			    base > ('z' - 'a' + 1) + ('9' - '0' + 1)) {
4256 				*flags |= CPU_DTRACE_ILLOP;
4257 				break;
4258 			}
4259 		}
4260 
4261 		val = (base == 10 && i < 0) ? i * -1 : i;
4262 
4263 		if (!DTRACE_INSCRATCH(mstate, size)) {
4264 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4265 			regs[rd] = NULL;
4266 			break;
4267 		}
4268 
4269 		for (*end-- = '\0'; val; val /= base) {
4270 			if ((digit = val % base) <= '9' - '0') {
4271 				*end-- = '0' + digit;
4272 			} else {
4273 				*end-- = 'a' + (digit - ('9' - '0') - 1);
4274 			}
4275 		}
4276 
4277 		if (i == 0 && base == 16)
4278 			*end-- = '0';
4279 
4280 		if (base == 16)
4281 			*end-- = 'x';
4282 
4283 		if (i == 0 || base == 8 || base == 16)
4284 			*end-- = '0';
4285 
4286 		if (i < 0 && base == 10)
4287 			*end-- = '-';
4288 
4289 		regs[rd] = (uintptr_t)end + 1;
4290 		mstate->dtms_scratch_ptr += size;
4291 		break;
4292 	}
4293 
4294 	case DIF_SUBR_HTONS:
4295 	case DIF_SUBR_NTOHS:
4296 #ifdef _BIG_ENDIAN
4297 		regs[rd] = (uint16_t)tupregs[0].dttk_value;
4298 #else
4299 		regs[rd] = DT_BSWAP_16((uint16_t)tupregs[0].dttk_value);
4300 #endif
4301 		break;
4302 
4303 
4304 	case DIF_SUBR_HTONL:
4305 	case DIF_SUBR_NTOHL:
4306 #ifdef _BIG_ENDIAN
4307 		regs[rd] = (uint32_t)tupregs[0].dttk_value;
4308 #else
4309 		regs[rd] = DT_BSWAP_32((uint32_t)tupregs[0].dttk_value);
4310 #endif
4311 		break;
4312 
4313 
4314 	case DIF_SUBR_HTONLL:
4315 	case DIF_SUBR_NTOHLL:
4316 #ifdef _BIG_ENDIAN
4317 		regs[rd] = (uint64_t)tupregs[0].dttk_value;
4318 #else
4319 		regs[rd] = DT_BSWAP_64((uint64_t)tupregs[0].dttk_value);
4320 #endif
4321 		break;
4322 
4323 
4324 	case DIF_SUBR_DIRNAME:
4325 	case DIF_SUBR_BASENAME: {
4326 		char *dest = (char *)mstate->dtms_scratch_ptr;
4327 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4328 		uintptr_t src = tupregs[0].dttk_value;
4329 		int i, j, len = dtrace_strlen((char *)src, size);
4330 		int lastbase = -1, firstbase = -1, lastdir = -1;
4331 		int start, end;
4332 
4333 		if (!dtrace_canload(src, len + 1, mstate, vstate)) {
4334 			regs[rd] = NULL;
4335 			break;
4336 		}
4337 
4338 		if (!DTRACE_INSCRATCH(mstate, size)) {
4339 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4340 			regs[rd] = NULL;
4341 			break;
4342 		}
4343 
4344 		/*
4345 		 * The basename and dirname for a zero-length string is
4346 		 * defined to be "."
4347 		 */
4348 		if (len == 0) {
4349 			len = 1;
4350 			src = (uintptr_t)".";
4351 		}
4352 
4353 		/*
4354 		 * Start from the back of the string, moving back toward the
4355 		 * front until we see a character that isn't a slash.  That
4356 		 * character is the last character in the basename.
4357 		 */
4358 		for (i = len - 1; i >= 0; i--) {
4359 			if (dtrace_load8(src + i) != '/')
4360 				break;
4361 		}
4362 
4363 		if (i >= 0)
4364 			lastbase = i;
4365 
4366 		/*
4367 		 * Starting from the last character in the basename, move
4368 		 * towards the front until we find a slash.  The character
4369 		 * that we processed immediately before that is the first
4370 		 * character in the basename.
4371 		 */
4372 		for (; i >= 0; i--) {
4373 			if (dtrace_load8(src + i) == '/')
4374 				break;
4375 		}
4376 
4377 		if (i >= 0)
4378 			firstbase = i + 1;
4379 
4380 		/*
4381 		 * Now keep going until we find a non-slash character.  That
4382 		 * character is the last character in the dirname.
4383 		 */
4384 		for (; i >= 0; i--) {
4385 			if (dtrace_load8(src + i) != '/')
4386 				break;
4387 		}
4388 
4389 		if (i >= 0)
4390 			lastdir = i;
4391 
4392 		ASSERT(!(lastbase == -1 && firstbase != -1));
4393 		ASSERT(!(firstbase == -1 && lastdir != -1));
4394 
4395 		if (lastbase == -1) {
4396 			/*
4397 			 * We didn't find a non-slash character.  We know that
4398 			 * the length is non-zero, so the whole string must be
4399 			 * slashes.  In either the dirname or the basename
4400 			 * case, we return '/'.
4401 			 */
4402 			ASSERT(firstbase == -1);
4403 			firstbase = lastbase = lastdir = 0;
4404 		}
4405 
4406 		if (firstbase == -1) {
4407 			/*
4408 			 * The entire string consists only of a basename
4409 			 * component.  If we're looking for dirname, we need
4410 			 * to change our string to be just "."; if we're
4411 			 * looking for a basename, we'll just set the first
4412 			 * character of the basename to be 0.
4413 			 */
4414 			if (subr == DIF_SUBR_DIRNAME) {
4415 				ASSERT(lastdir == -1);
4416 				src = (uintptr_t)".";
4417 				lastdir = 0;
4418 			} else {
4419 				firstbase = 0;
4420 			}
4421 		}
4422 
4423 		if (subr == DIF_SUBR_DIRNAME) {
4424 			if (lastdir == -1) {
4425 				/*
4426 				 * We know that we have a slash in the name --
4427 				 * or lastdir would be set to 0, above.  And
4428 				 * because lastdir is -1, we know that this
4429 				 * slash must be the first character.  (That
4430 				 * is, the full string must be of the form
4431 				 * "/basename".)  In this case, the last
4432 				 * character of the directory name is 0.
4433 				 */
4434 				lastdir = 0;
4435 			}
4436 
4437 			start = 0;
4438 			end = lastdir;
4439 		} else {
4440 			ASSERT(subr == DIF_SUBR_BASENAME);
4441 			ASSERT(firstbase != -1 && lastbase != -1);
4442 			start = firstbase;
4443 			end = lastbase;
4444 		}
4445 
4446 		for (i = start, j = 0; i <= end && j < size - 1; i++, j++)
4447 			dest[j] = dtrace_load8(src + i);
4448 
4449 		dest[j] = '\0';
4450 		regs[rd] = (uintptr_t)dest;
4451 		mstate->dtms_scratch_ptr += size;
4452 		break;
4453 	}
4454 
4455 	case DIF_SUBR_CLEANPATH: {
4456 		char *dest = (char *)mstate->dtms_scratch_ptr, c;
4457 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4458 		uintptr_t src = tupregs[0].dttk_value;
4459 		int i = 0, j = 0;
4460 
4461 		if (!dtrace_strcanload(src, size, mstate, vstate)) {
4462 			regs[rd] = NULL;
4463 			break;
4464 		}
4465 
4466 		if (!DTRACE_INSCRATCH(mstate, size)) {
4467 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4468 			regs[rd] = NULL;
4469 			break;
4470 		}
4471 
4472 		/*
4473 		 * Move forward, loading each character.
4474 		 */
4475 		do {
4476 			c = dtrace_load8(src + i++);
4477 next:
4478 			if (j + 5 >= size)	/* 5 = strlen("/..c\0") */
4479 				break;
4480 
4481 			if (c != '/') {
4482 				dest[j++] = c;
4483 				continue;
4484 			}
4485 
4486 			c = dtrace_load8(src + i++);
4487 
4488 			if (c == '/') {
4489 				/*
4490 				 * We have two slashes -- we can just advance
4491 				 * to the next character.
4492 				 */
4493 				goto next;
4494 			}
4495 
4496 			if (c != '.') {
4497 				/*
4498 				 * This is not "." and it's not ".." -- we can
4499 				 * just store the "/" and this character and
4500 				 * drive on.
4501 				 */
4502 				dest[j++] = '/';
4503 				dest[j++] = c;
4504 				continue;
4505 			}
4506 
4507 			c = dtrace_load8(src + i++);
4508 
4509 			if (c == '/') {
4510 				/*
4511 				 * This is a "/./" component.  We're not going
4512 				 * to store anything in the destination buffer;
4513 				 * we're just going to go to the next component.
4514 				 */
4515 				goto next;
4516 			}
4517 
4518 			if (c != '.') {
4519 				/*
4520 				 * This is not ".." -- we can just store the
4521 				 * "/." and this character and continue
4522 				 * processing.
4523 				 */
4524 				dest[j++] = '/';
4525 				dest[j++] = '.';
4526 				dest[j++] = c;
4527 				continue;
4528 			}
4529 
4530 			c = dtrace_load8(src + i++);
4531 
4532 			if (c != '/' && c != '\0') {
4533 				/*
4534 				 * This is not ".." -- it's "..[mumble]".
4535 				 * We'll store the "/.." and this character
4536 				 * and continue processing.
4537 				 */
4538 				dest[j++] = '/';
4539 				dest[j++] = '.';
4540 				dest[j++] = '.';
4541 				dest[j++] = c;
4542 				continue;
4543 			}
4544 
4545 			/*
4546 			 * This is "/../" or "/..\0".  We need to back up
4547 			 * our destination pointer until we find a "/".
4548 			 */
4549 			i--;
4550 			while (j != 0 && dest[--j] != '/')
4551 				continue;
4552 
4553 			if (c == '\0')
4554 				dest[++j] = '/';
4555 		} while (c != '\0');
4556 
4557 		dest[j] = '\0';
4558 		regs[rd] = (uintptr_t)dest;
4559 		mstate->dtms_scratch_ptr += size;
4560 		break;
4561 	}
4562 
4563 	case DIF_SUBR_INET_NTOA:
4564 	case DIF_SUBR_INET_NTOA6:
4565 	case DIF_SUBR_INET_NTOP: {
4566 		size_t size;
4567 		int af, argi, i;
4568 		char *base, *end;
4569 
4570 		if (subr == DIF_SUBR_INET_NTOP) {
4571 			af = (int)tupregs[0].dttk_value;
4572 			argi = 1;
4573 		} else {
4574 			af = subr == DIF_SUBR_INET_NTOA ? AF_INET: AF_INET6;
4575 			argi = 0;
4576 		}
4577 
4578 		if (af == AF_INET) {
4579 			ipaddr_t ip4;
4580 			uint8_t *ptr8, val;
4581 
4582 			/*
4583 			 * Safely load the IPv4 address.
4584 			 */
4585 			ip4 = dtrace_load32(tupregs[argi].dttk_value);
4586 
4587 			/*
4588 			 * Check an IPv4 string will fit in scratch.
4589 			 */
4590 			size = INET_ADDRSTRLEN;
4591 			if (!DTRACE_INSCRATCH(mstate, size)) {
4592 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4593 				regs[rd] = NULL;
4594 				break;
4595 			}
4596 			base = (char *)mstate->dtms_scratch_ptr;
4597 			end = (char *)mstate->dtms_scratch_ptr + size - 1;
4598 
4599 			/*
4600 			 * Stringify as a dotted decimal quad.
4601 			 */
4602 			*end-- = '\0';
4603 			ptr8 = (uint8_t *)&ip4;
4604 			for (i = 3; i >= 0; i--) {
4605 				val = ptr8[i];
4606 
4607 				if (val == 0) {
4608 					*end-- = '0';
4609 				} else {
4610 					for (; val; val /= 10) {
4611 						*end-- = '0' + (val % 10);
4612 					}
4613 				}
4614 
4615 				if (i > 0)
4616 					*end-- = '.';
4617 			}
4618 			ASSERT(end + 1 >= base);
4619 
4620 		} else if (af == AF_INET6) {
4621 			struct in6_addr ip6;
4622 			int firstzero, tryzero, numzero, v6end;
4623 			uint16_t val;
4624 			const char digits[] = "0123456789abcdef";
4625 
4626 			/*
4627 			 * Stringify using RFC 1884 convention 2 - 16 bit
4628 			 * hexadecimal values with a zero-run compression.
4629 			 * Lower case hexadecimal digits are used.
4630 			 * 	eg, fe80::214:4fff:fe0b:76c8.
4631 			 * The IPv4 embedded form is returned for inet_ntop,
4632 			 * just the IPv4 string is returned for inet_ntoa6.
4633 			 */
4634 
4635 			/*
4636 			 * Safely load the IPv6 address.
4637 			 */
4638 			dtrace_bcopy(
4639 			    (void *)(uintptr_t)tupregs[argi].dttk_value,
4640 			    (void *)(uintptr_t)&ip6, sizeof (struct in6_addr));
4641 
4642 			/*
4643 			 * Check an IPv6 string will fit in scratch.
4644 			 */
4645 			size = INET6_ADDRSTRLEN;
4646 			if (!DTRACE_INSCRATCH(mstate, size)) {
4647 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4648 				regs[rd] = NULL;
4649 				break;
4650 			}
4651 			base = (char *)mstate->dtms_scratch_ptr;
4652 			end = (char *)mstate->dtms_scratch_ptr + size - 1;
4653 			*end-- = '\0';
4654 
4655 			/*
4656 			 * Find the longest run of 16 bit zero values
4657 			 * for the single allowed zero compression - "::".
4658 			 */
4659 			firstzero = -1;
4660 			tryzero = -1;
4661 			numzero = 1;
4662 			for (i = 0; i < sizeof (struct in6_addr); i++) {
4663 				if (ip6._S6_un._S6_u8[i] == 0 &&
4664 				    tryzero == -1 && i % 2 == 0) {
4665 					tryzero = i;
4666 					continue;
4667 				}
4668 
4669 				if (tryzero != -1 &&
4670 				    (ip6._S6_un._S6_u8[i] != 0 ||
4671 				    i == sizeof (struct in6_addr) - 1)) {
4672 
4673 					if (i - tryzero <= numzero) {
4674 						tryzero = -1;
4675 						continue;
4676 					}
4677 
4678 					firstzero = tryzero;
4679 					numzero = i - i % 2 - tryzero;
4680 					tryzero = -1;
4681 
4682 					if (ip6._S6_un._S6_u8[i] == 0 &&
4683 					    i == sizeof (struct in6_addr) - 1)
4684 						numzero += 2;
4685 				}
4686 			}
4687 			ASSERT(firstzero + numzero <= sizeof (struct in6_addr));
4688 
4689 			/*
4690 			 * Check for an IPv4 embedded address.
4691 			 */
4692 			v6end = sizeof (struct in6_addr) - 2;
4693 			if (IN6_IS_ADDR_V4MAPPED(&ip6) ||
4694 			    IN6_IS_ADDR_V4COMPAT(&ip6)) {
4695 				for (i = sizeof (struct in6_addr) - 1;
4696 				    i >= DTRACE_V4MAPPED_OFFSET; i--) {
4697 					ASSERT(end >= base);
4698 
4699 					val = ip6._S6_un._S6_u8[i];
4700 
4701 					if (val == 0) {
4702 						*end-- = '0';
4703 					} else {
4704 						for (; val; val /= 10) {
4705 							*end-- = '0' + val % 10;
4706 						}
4707 					}
4708 
4709 					if (i > DTRACE_V4MAPPED_OFFSET)
4710 						*end-- = '.';
4711 				}
4712 
4713 				if (subr == DIF_SUBR_INET_NTOA6)
4714 					goto inetout;
4715 
4716 				/*
4717 				 * Set v6end to skip the IPv4 address that
4718 				 * we have already stringified.
4719 				 */
4720 				v6end = 10;
4721 			}
4722 
4723 			/*
4724 			 * Build the IPv6 string by working through the
4725 			 * address in reverse.
4726 			 */
4727 			for (i = v6end; i >= 0; i -= 2) {
4728 				ASSERT(end >= base);
4729 
4730 				if (i == firstzero + numzero - 2) {
4731 					*end-- = ':';
4732 					*end-- = ':';
4733 					i -= numzero - 2;
4734 					continue;
4735 				}
4736 
4737 				if (i < 14 && i != firstzero - 2)
4738 					*end-- = ':';
4739 
4740 				val = (ip6._S6_un._S6_u8[i] << 8) +
4741 				    ip6._S6_un._S6_u8[i + 1];
4742 
4743 				if (val == 0) {
4744 					*end-- = '0';
4745 				} else {
4746 					for (; val; val /= 16) {
4747 						*end-- = digits[val % 16];
4748 					}
4749 				}
4750 			}
4751 			ASSERT(end + 1 >= base);
4752 
4753 		} else {
4754 			/*
4755 			 * The user didn't use AH_INET or AH_INET6.
4756 			 */
4757 			DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
4758 			regs[rd] = NULL;
4759 			break;
4760 		}
4761 
4762 inetout:	regs[rd] = (uintptr_t)end + 1;
4763 		mstate->dtms_scratch_ptr += size;
4764 		break;
4765 	}
4766 
4767 	}
4768 }
4769 
4770 /*
4771  * Emulate the execution of DTrace IR instructions specified by the given
4772  * DIF object.  This function is deliberately void of assertions as all of
4773  * the necessary checks are handled by a call to dtrace_difo_validate().
4774  */
4775 static uint64_t
4776 dtrace_dif_emulate(dtrace_difo_t *difo, dtrace_mstate_t *mstate,
4777     dtrace_vstate_t *vstate, dtrace_state_t *state)
4778 {
4779 	const dif_instr_t *text = difo->dtdo_buf;
4780 	const uint_t textlen = difo->dtdo_len;
4781 	const char *strtab = difo->dtdo_strtab;
4782 	const uint64_t *inttab = difo->dtdo_inttab;
4783 
4784 	uint64_t rval = 0;
4785 	dtrace_statvar_t *svar;
4786 	dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
4787 	dtrace_difv_t *v;
4788 	volatile uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
4789 	volatile uintptr_t *illval = &cpu_core[CPU->cpu_id].cpuc_dtrace_illval;
4790 
4791 	dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
4792 	uint64_t regs[DIF_DIR_NREGS];
4793 	uint64_t *tmp;
4794 
4795 	uint8_t cc_n = 0, cc_z = 0, cc_v = 0, cc_c = 0;
4796 	int64_t cc_r;
4797 	uint_t pc = 0, id, opc;
4798 	uint8_t ttop = 0;
4799 	dif_instr_t instr;
4800 	uint_t r1, r2, rd;
4801 
4802 	/*
4803 	 * We stash the current DIF object into the machine state: we need it
4804 	 * for subsequent access checking.
4805 	 */
4806 	mstate->dtms_difo = difo;
4807 
4808 	regs[DIF_REG_R0] = 0; 		/* %r0 is fixed at zero */
4809 
4810 	while (pc < textlen && !(*flags & CPU_DTRACE_FAULT)) {
4811 		opc = pc;
4812 
4813 		instr = text[pc++];
4814 		r1 = DIF_INSTR_R1(instr);
4815 		r2 = DIF_INSTR_R2(instr);
4816 		rd = DIF_INSTR_RD(instr);
4817 
4818 		switch (DIF_INSTR_OP(instr)) {
4819 		case DIF_OP_OR:
4820 			regs[rd] = regs[r1] | regs[r2];
4821 			break;
4822 		case DIF_OP_XOR:
4823 			regs[rd] = regs[r1] ^ regs[r2];
4824 			break;
4825 		case DIF_OP_AND:
4826 			regs[rd] = regs[r1] & regs[r2];
4827 			break;
4828 		case DIF_OP_SLL:
4829 			regs[rd] = regs[r1] << regs[r2];
4830 			break;
4831 		case DIF_OP_SRL:
4832 			regs[rd] = regs[r1] >> regs[r2];
4833 			break;
4834 		case DIF_OP_SUB:
4835 			regs[rd] = regs[r1] - regs[r2];
4836 			break;
4837 		case DIF_OP_ADD:
4838 			regs[rd] = regs[r1] + regs[r2];
4839 			break;
4840 		case DIF_OP_MUL:
4841 			regs[rd] = regs[r1] * regs[r2];
4842 			break;
4843 		case DIF_OP_SDIV:
4844 			if (regs[r2] == 0) {
4845 				regs[rd] = 0;
4846 				*flags |= CPU_DTRACE_DIVZERO;
4847 			} else {
4848 				regs[rd] = (int64_t)regs[r1] /
4849 				    (int64_t)regs[r2];
4850 			}
4851 			break;
4852 
4853 		case DIF_OP_UDIV:
4854 			if (regs[r2] == 0) {
4855 				regs[rd] = 0;
4856 				*flags |= CPU_DTRACE_DIVZERO;
4857 			} else {
4858 				regs[rd] = regs[r1] / regs[r2];
4859 			}
4860 			break;
4861 
4862 		case DIF_OP_SREM:
4863 			if (regs[r2] == 0) {
4864 				regs[rd] = 0;
4865 				*flags |= CPU_DTRACE_DIVZERO;
4866 			} else {
4867 				regs[rd] = (int64_t)regs[r1] %
4868 				    (int64_t)regs[r2];
4869 			}
4870 			break;
4871 
4872 		case DIF_OP_UREM:
4873 			if (regs[r2] == 0) {
4874 				regs[rd] = 0;
4875 				*flags |= CPU_DTRACE_DIVZERO;
4876 			} else {
4877 				regs[rd] = regs[r1] % regs[r2];
4878 			}
4879 			break;
4880 
4881 		case DIF_OP_NOT:
4882 			regs[rd] = ~regs[r1];
4883 			break;
4884 		case DIF_OP_MOV:
4885 			regs[rd] = regs[r1];
4886 			break;
4887 		case DIF_OP_CMP:
4888 			cc_r = regs[r1] - regs[r2];
4889 			cc_n = cc_r < 0;
4890 			cc_z = cc_r == 0;
4891 			cc_v = 0;
4892 			cc_c = regs[r1] < regs[r2];
4893 			break;
4894 		case DIF_OP_TST:
4895 			cc_n = cc_v = cc_c = 0;
4896 			cc_z = regs[r1] == 0;
4897 			break;
4898 		case DIF_OP_BA:
4899 			pc = DIF_INSTR_LABEL(instr);
4900 			break;
4901 		case DIF_OP_BE:
4902 			if (cc_z)
4903 				pc = DIF_INSTR_LABEL(instr);
4904 			break;
4905 		case DIF_OP_BNE:
4906 			if (cc_z == 0)
4907 				pc = DIF_INSTR_LABEL(instr);
4908 			break;
4909 		case DIF_OP_BG:
4910 			if ((cc_z | (cc_n ^ cc_v)) == 0)
4911 				pc = DIF_INSTR_LABEL(instr);
4912 			break;
4913 		case DIF_OP_BGU:
4914 			if ((cc_c | cc_z) == 0)
4915 				pc = DIF_INSTR_LABEL(instr);
4916 			break;
4917 		case DIF_OP_BGE:
4918 			if ((cc_n ^ cc_v) == 0)
4919 				pc = DIF_INSTR_LABEL(instr);
4920 			break;
4921 		case DIF_OP_BGEU:
4922 			if (cc_c == 0)
4923 				pc = DIF_INSTR_LABEL(instr);
4924 			break;
4925 		case DIF_OP_BL:
4926 			if (cc_n ^ cc_v)
4927 				pc = DIF_INSTR_LABEL(instr);
4928 			break;
4929 		case DIF_OP_BLU:
4930 			if (cc_c)
4931 				pc = DIF_INSTR_LABEL(instr);
4932 			break;
4933 		case DIF_OP_BLE:
4934 			if (cc_z | (cc_n ^ cc_v))
4935 				pc = DIF_INSTR_LABEL(instr);
4936 			break;
4937 		case DIF_OP_BLEU:
4938 			if (cc_c | cc_z)
4939 				pc = DIF_INSTR_LABEL(instr);
4940 			break;
4941 		case DIF_OP_RLDSB:
4942 			if (!dtrace_canstore(regs[r1], 1, mstate, vstate)) {
4943 				*flags |= CPU_DTRACE_KPRIV;
4944 				*illval = regs[r1];
4945 				break;
4946 			}
4947 			/*FALLTHROUGH*/
4948 		case DIF_OP_LDSB:
4949 			regs[rd] = (int8_t)dtrace_load8(regs[r1]);
4950 			break;
4951 		case DIF_OP_RLDSH:
4952 			if (!dtrace_canstore(regs[r1], 2, mstate, vstate)) {
4953 				*flags |= CPU_DTRACE_KPRIV;
4954 				*illval = regs[r1];
4955 				break;
4956 			}
4957 			/*FALLTHROUGH*/
4958 		case DIF_OP_LDSH:
4959 			regs[rd] = (int16_t)dtrace_load16(regs[r1]);
4960 			break;
4961 		case DIF_OP_RLDSW:
4962 			if (!dtrace_canstore(regs[r1], 4, mstate, vstate)) {
4963 				*flags |= CPU_DTRACE_KPRIV;
4964 				*illval = regs[r1];
4965 				break;
4966 			}
4967 			/*FALLTHROUGH*/
4968 		case DIF_OP_LDSW:
4969 			regs[rd] = (int32_t)dtrace_load32(regs[r1]);
4970 			break;
4971 		case DIF_OP_RLDUB:
4972 			if (!dtrace_canstore(regs[r1], 1, mstate, vstate)) {
4973 				*flags |= CPU_DTRACE_KPRIV;
4974 				*illval = regs[r1];
4975 				break;
4976 			}
4977 			/*FALLTHROUGH*/
4978 		case DIF_OP_LDUB:
4979 			regs[rd] = dtrace_load8(regs[r1]);
4980 			break;
4981 		case DIF_OP_RLDUH:
4982 			if (!dtrace_canstore(regs[r1], 2, mstate, vstate)) {
4983 				*flags |= CPU_DTRACE_KPRIV;
4984 				*illval = regs[r1];
4985 				break;
4986 			}
4987 			/*FALLTHROUGH*/
4988 		case DIF_OP_LDUH:
4989 			regs[rd] = dtrace_load16(regs[r1]);
4990 			break;
4991 		case DIF_OP_RLDUW:
4992 			if (!dtrace_canstore(regs[r1], 4, mstate, vstate)) {
4993 				*flags |= CPU_DTRACE_KPRIV;
4994 				*illval = regs[r1];
4995 				break;
4996 			}
4997 			/*FALLTHROUGH*/
4998 		case DIF_OP_LDUW:
4999 			regs[rd] = dtrace_load32(regs[r1]);
5000 			break;
5001 		case DIF_OP_RLDX:
5002 			if (!dtrace_canstore(regs[r1], 8, mstate, vstate)) {
5003 				*flags |= CPU_DTRACE_KPRIV;
5004 				*illval = regs[r1];
5005 				break;
5006 			}
5007 			/*FALLTHROUGH*/
5008 		case DIF_OP_LDX:
5009 			regs[rd] = dtrace_load64(regs[r1]);
5010 			break;
5011 		case DIF_OP_ULDSB:
5012 			regs[rd] = (int8_t)
5013 			    dtrace_fuword8((void *)(uintptr_t)regs[r1]);
5014 			break;
5015 		case DIF_OP_ULDSH:
5016 			regs[rd] = (int16_t)
5017 			    dtrace_fuword16((void *)(uintptr_t)regs[r1]);
5018 			break;
5019 		case DIF_OP_ULDSW:
5020 			regs[rd] = (int32_t)
5021 			    dtrace_fuword32((void *)(uintptr_t)regs[r1]);
5022 			break;
5023 		case DIF_OP_ULDUB:
5024 			regs[rd] =
5025 			    dtrace_fuword8((void *)(uintptr_t)regs[r1]);
5026 			break;
5027 		case DIF_OP_ULDUH:
5028 			regs[rd] =
5029 			    dtrace_fuword16((void *)(uintptr_t)regs[r1]);
5030 			break;
5031 		case DIF_OP_ULDUW:
5032 			regs[rd] =
5033 			    dtrace_fuword32((void *)(uintptr_t)regs[r1]);
5034 			break;
5035 		case DIF_OP_ULDX:
5036 			regs[rd] =
5037 			    dtrace_fuword64((void *)(uintptr_t)regs[r1]);
5038 			break;
5039 		case DIF_OP_RET:
5040 			rval = regs[rd];
5041 			pc = textlen;
5042 			break;
5043 		case DIF_OP_NOP:
5044 			break;
5045 		case DIF_OP_SETX:
5046 			regs[rd] = inttab[DIF_INSTR_INTEGER(instr)];
5047 			break;
5048 		case DIF_OP_SETS:
5049 			regs[rd] = (uint64_t)(uintptr_t)
5050 			    (strtab + DIF_INSTR_STRING(instr));
5051 			break;
5052 		case DIF_OP_SCMP: {
5053 			size_t sz = state->dts_options[DTRACEOPT_STRSIZE];
5054 			uintptr_t s1 = regs[r1];
5055 			uintptr_t s2 = regs[r2];
5056 
5057 			if (s1 != NULL &&
5058 			    !dtrace_strcanload(s1, sz, mstate, vstate))
5059 				break;
5060 			if (s2 != NULL &&
5061 			    !dtrace_strcanload(s2, sz, mstate, vstate))
5062 				break;
5063 
5064 			cc_r = dtrace_strncmp((char *)s1, (char *)s2, sz);
5065 
5066 			cc_n = cc_r < 0;
5067 			cc_z = cc_r == 0;
5068 			cc_v = cc_c = 0;
5069 			break;
5070 		}
5071 		case DIF_OP_LDGA:
5072 			regs[rd] = dtrace_dif_variable(mstate, state,
5073 			    r1, regs[r2]);
5074 			break;
5075 		case DIF_OP_LDGS:
5076 			id = DIF_INSTR_VAR(instr);
5077 
5078 			if (id >= DIF_VAR_OTHER_UBASE) {
5079 				uintptr_t a;
5080 
5081 				id -= DIF_VAR_OTHER_UBASE;
5082 				svar = vstate->dtvs_globals[id];
5083 				ASSERT(svar != NULL);
5084 				v = &svar->dtsv_var;
5085 
5086 				if (!(v->dtdv_type.dtdt_flags & DIF_TF_BYREF)) {
5087 					regs[rd] = svar->dtsv_data;
5088 					break;
5089 				}
5090 
5091 				a = (uintptr_t)svar->dtsv_data;
5092 
5093 				if (*(uint8_t *)a == UINT8_MAX) {
5094 					/*
5095 					 * If the 0th byte is set to UINT8_MAX
5096 					 * then this is to be treated as a
5097 					 * reference to a NULL variable.
5098 					 */
5099 					regs[rd] = NULL;
5100 				} else {
5101 					regs[rd] = a + sizeof (uint64_t);
5102 				}
5103 
5104 				break;
5105 			}
5106 
5107 			regs[rd] = dtrace_dif_variable(mstate, state, id, 0);
5108 			break;
5109 
5110 		case DIF_OP_STGS:
5111 			id = DIF_INSTR_VAR(instr);
5112 
5113 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
5114 			id -= DIF_VAR_OTHER_UBASE;
5115 
5116 			svar = vstate->dtvs_globals[id];
5117 			ASSERT(svar != NULL);
5118 			v = &svar->dtsv_var;
5119 
5120 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5121 				uintptr_t a = (uintptr_t)svar->dtsv_data;
5122 
5123 				ASSERT(a != NULL);
5124 				ASSERT(svar->dtsv_size != 0);
5125 
5126 				if (regs[rd] == NULL) {
5127 					*(uint8_t *)a = UINT8_MAX;
5128 					break;
5129 				} else {
5130 					*(uint8_t *)a = 0;
5131 					a += sizeof (uint64_t);
5132 				}
5133 				if (!dtrace_vcanload(
5134 				    (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5135 				    mstate, vstate))
5136 					break;
5137 
5138 				dtrace_vcopy((void *)(uintptr_t)regs[rd],
5139 				    (void *)a, &v->dtdv_type);
5140 				break;
5141 			}
5142 
5143 			svar->dtsv_data = regs[rd];
5144 			break;
5145 
5146 		case DIF_OP_LDTA:
5147 			/*
5148 			 * There are no DTrace built-in thread-local arrays at
5149 			 * present.  This opcode is saved for future work.
5150 			 */
5151 			*flags |= CPU_DTRACE_ILLOP;
5152 			regs[rd] = 0;
5153 			break;
5154 
5155 		case DIF_OP_LDLS:
5156 			id = DIF_INSTR_VAR(instr);
5157 
5158 			if (id < DIF_VAR_OTHER_UBASE) {
5159 				/*
5160 				 * For now, this has no meaning.
5161 				 */
5162 				regs[rd] = 0;
5163 				break;
5164 			}
5165 
5166 			id -= DIF_VAR_OTHER_UBASE;
5167 
5168 			ASSERT(id < vstate->dtvs_nlocals);
5169 			ASSERT(vstate->dtvs_locals != NULL);
5170 
5171 			svar = vstate->dtvs_locals[id];
5172 			ASSERT(svar != NULL);
5173 			v = &svar->dtsv_var;
5174 
5175 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5176 				uintptr_t a = (uintptr_t)svar->dtsv_data;
5177 				size_t sz = v->dtdv_type.dtdt_size;
5178 
5179 				sz += sizeof (uint64_t);
5180 				ASSERT(svar->dtsv_size == NCPU * sz);
5181 				a += CPU->cpu_id * sz;
5182 
5183 				if (*(uint8_t *)a == UINT8_MAX) {
5184 					/*
5185 					 * If the 0th byte is set to UINT8_MAX
5186 					 * then this is to be treated as a
5187 					 * reference to a NULL variable.
5188 					 */
5189 					regs[rd] = NULL;
5190 				} else {
5191 					regs[rd] = a + sizeof (uint64_t);
5192 				}
5193 
5194 				break;
5195 			}
5196 
5197 			ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
5198 			tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
5199 			regs[rd] = tmp[CPU->cpu_id];
5200 			break;
5201 
5202 		case DIF_OP_STLS:
5203 			id = DIF_INSTR_VAR(instr);
5204 
5205 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
5206 			id -= DIF_VAR_OTHER_UBASE;
5207 			ASSERT(id < vstate->dtvs_nlocals);
5208 
5209 			ASSERT(vstate->dtvs_locals != NULL);
5210 			svar = vstate->dtvs_locals[id];
5211 			ASSERT(svar != NULL);
5212 			v = &svar->dtsv_var;
5213 
5214 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5215 				uintptr_t a = (uintptr_t)svar->dtsv_data;
5216 				size_t sz = v->dtdv_type.dtdt_size;
5217 
5218 				sz += sizeof (uint64_t);
5219 				ASSERT(svar->dtsv_size == NCPU * sz);
5220 				a += CPU->cpu_id * sz;
5221 
5222 				if (regs[rd] == NULL) {
5223 					*(uint8_t *)a = UINT8_MAX;
5224 					break;
5225 				} else {
5226 					*(uint8_t *)a = 0;
5227 					a += sizeof (uint64_t);
5228 				}
5229 
5230 				if (!dtrace_vcanload(
5231 				    (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5232 				    mstate, vstate))
5233 					break;
5234 
5235 				dtrace_vcopy((void *)(uintptr_t)regs[rd],
5236 				    (void *)a, &v->dtdv_type);
5237 				break;
5238 			}
5239 
5240 			ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
5241 			tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
5242 			tmp[CPU->cpu_id] = regs[rd];
5243 			break;
5244 
5245 		case DIF_OP_LDTS: {
5246 			dtrace_dynvar_t *dvar;
5247 			dtrace_key_t *key;
5248 
5249 			id = DIF_INSTR_VAR(instr);
5250 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
5251 			id -= DIF_VAR_OTHER_UBASE;
5252 			v = &vstate->dtvs_tlocals[id];
5253 
5254 			key = &tupregs[DIF_DTR_NREGS];
5255 			key[0].dttk_value = (uint64_t)id;
5256 			key[0].dttk_size = 0;
5257 			DTRACE_TLS_THRKEY(key[1].dttk_value);
5258 			key[1].dttk_size = 0;
5259 
5260 			dvar = dtrace_dynvar(dstate, 2, key,
5261 			    sizeof (uint64_t), DTRACE_DYNVAR_NOALLOC,
5262 			    mstate, vstate);
5263 
5264 			if (dvar == NULL) {
5265 				regs[rd] = 0;
5266 				break;
5267 			}
5268 
5269 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5270 				regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
5271 			} else {
5272 				regs[rd] = *((uint64_t *)dvar->dtdv_data);
5273 			}
5274 
5275 			break;
5276 		}
5277 
5278 		case DIF_OP_STTS: {
5279 			dtrace_dynvar_t *dvar;
5280 			dtrace_key_t *key;
5281 
5282 			id = DIF_INSTR_VAR(instr);
5283 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
5284 			id -= DIF_VAR_OTHER_UBASE;
5285 
5286 			key = &tupregs[DIF_DTR_NREGS];
5287 			key[0].dttk_value = (uint64_t)id;
5288 			key[0].dttk_size = 0;
5289 			DTRACE_TLS_THRKEY(key[1].dttk_value);
5290 			key[1].dttk_size = 0;
5291 			v = &vstate->dtvs_tlocals[id];
5292 
5293 			dvar = dtrace_dynvar(dstate, 2, key,
5294 			    v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
5295 			    v->dtdv_type.dtdt_size : sizeof (uint64_t),
5296 			    regs[rd] ? DTRACE_DYNVAR_ALLOC :
5297 			    DTRACE_DYNVAR_DEALLOC, mstate, vstate);
5298 
5299 			/*
5300 			 * Given that we're storing to thread-local data,
5301 			 * we need to flush our predicate cache.
5302 			 */
5303 			curthread->t_predcache = NULL;
5304 
5305 			if (dvar == NULL)
5306 				break;
5307 
5308 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5309 				if (!dtrace_vcanload(
5310 				    (void *)(uintptr_t)regs[rd],
5311 				    &v->dtdv_type, mstate, vstate))
5312 					break;
5313 
5314 				dtrace_vcopy((void *)(uintptr_t)regs[rd],
5315 				    dvar->dtdv_data, &v->dtdv_type);
5316 			} else {
5317 				*((uint64_t *)dvar->dtdv_data) = regs[rd];
5318 			}
5319 
5320 			break;
5321 		}
5322 
5323 		case DIF_OP_SRA:
5324 			regs[rd] = (int64_t)regs[r1] >> regs[r2];
5325 			break;
5326 
5327 		case DIF_OP_CALL:
5328 			dtrace_dif_subr(DIF_INSTR_SUBR(instr), rd,
5329 			    regs, tupregs, ttop, mstate, state);
5330 			break;
5331 
5332 		case DIF_OP_PUSHTR:
5333 			if (ttop == DIF_DTR_NREGS) {
5334 				*flags |= CPU_DTRACE_TUPOFLOW;
5335 				break;
5336 			}
5337 
5338 			if (r1 == DIF_TYPE_STRING) {
5339 				/*
5340 				 * If this is a string type and the size is 0,
5341 				 * we'll use the system-wide default string
5342 				 * size.  Note that we are _not_ looking at
5343 				 * the value of the DTRACEOPT_STRSIZE option;
5344 				 * had this been set, we would expect to have
5345 				 * a non-zero size value in the "pushtr".
5346 				 */
5347 				tupregs[ttop].dttk_size =
5348 				    dtrace_strlen((char *)(uintptr_t)regs[rd],
5349 				    regs[r2] ? regs[r2] :
5350 				    dtrace_strsize_default) + 1;
5351 			} else {
5352 				tupregs[ttop].dttk_size = regs[r2];
5353 			}
5354 
5355 			tupregs[ttop++].dttk_value = regs[rd];
5356 			break;
5357 
5358 		case DIF_OP_PUSHTV:
5359 			if (ttop == DIF_DTR_NREGS) {
5360 				*flags |= CPU_DTRACE_TUPOFLOW;
5361 				break;
5362 			}
5363 
5364 			tupregs[ttop].dttk_value = regs[rd];
5365 			tupregs[ttop++].dttk_size = 0;
5366 			break;
5367 
5368 		case DIF_OP_POPTS:
5369 			if (ttop != 0)
5370 				ttop--;
5371 			break;
5372 
5373 		case DIF_OP_FLUSHTS:
5374 			ttop = 0;
5375 			break;
5376 
5377 		case DIF_OP_LDGAA:
5378 		case DIF_OP_LDTAA: {
5379 			dtrace_dynvar_t *dvar;
5380 			dtrace_key_t *key = tupregs;
5381 			uint_t nkeys = ttop;
5382 
5383 			id = DIF_INSTR_VAR(instr);
5384 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
5385 			id -= DIF_VAR_OTHER_UBASE;
5386 
5387 			key[nkeys].dttk_value = (uint64_t)id;
5388 			key[nkeys++].dttk_size = 0;
5389 
5390 			if (DIF_INSTR_OP(instr) == DIF_OP_LDTAA) {
5391 				DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
5392 				key[nkeys++].dttk_size = 0;
5393 				v = &vstate->dtvs_tlocals[id];
5394 			} else {
5395 				v = &vstate->dtvs_globals[id]->dtsv_var;
5396 			}
5397 
5398 			dvar = dtrace_dynvar(dstate, nkeys, key,
5399 			    v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
5400 			    v->dtdv_type.dtdt_size : sizeof (uint64_t),
5401 			    DTRACE_DYNVAR_NOALLOC, mstate, vstate);
5402 
5403 			if (dvar == NULL) {
5404 				regs[rd] = 0;
5405 				break;
5406 			}
5407 
5408 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5409 				regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
5410 			} else {
5411 				regs[rd] = *((uint64_t *)dvar->dtdv_data);
5412 			}
5413 
5414 			break;
5415 		}
5416 
5417 		case DIF_OP_STGAA:
5418 		case DIF_OP_STTAA: {
5419 			dtrace_dynvar_t *dvar;
5420 			dtrace_key_t *key = tupregs;
5421 			uint_t nkeys = ttop;
5422 
5423 			id = DIF_INSTR_VAR(instr);
5424 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
5425 			id -= DIF_VAR_OTHER_UBASE;
5426 
5427 			key[nkeys].dttk_value = (uint64_t)id;
5428 			key[nkeys++].dttk_size = 0;
5429 
5430 			if (DIF_INSTR_OP(instr) == DIF_OP_STTAA) {
5431 				DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
5432 				key[nkeys++].dttk_size = 0;
5433 				v = &vstate->dtvs_tlocals[id];
5434 			} else {
5435 				v = &vstate->dtvs_globals[id]->dtsv_var;
5436 			}
5437 
5438 			dvar = dtrace_dynvar(dstate, nkeys, key,
5439 			    v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
5440 			    v->dtdv_type.dtdt_size : sizeof (uint64_t),
5441 			    regs[rd] ? DTRACE_DYNVAR_ALLOC :
5442 			    DTRACE_DYNVAR_DEALLOC, mstate, vstate);
5443 
5444 			if (dvar == NULL)
5445 				break;
5446 
5447 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5448 				if (!dtrace_vcanload(
5449 				    (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5450 				    mstate, vstate))
5451 					break;
5452 
5453 				dtrace_vcopy((void *)(uintptr_t)regs[rd],
5454 				    dvar->dtdv_data, &v->dtdv_type);
5455 			} else {
5456 				*((uint64_t *)dvar->dtdv_data) = regs[rd];
5457 			}
5458 
5459 			break;
5460 		}
5461 
5462 		case DIF_OP_ALLOCS: {
5463 			uintptr_t ptr = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
5464 			size_t size = ptr - mstate->dtms_scratch_ptr + regs[r1];
5465 
5466 			/*
5467 			 * Rounding up the user allocation size could have
5468 			 * overflowed large, bogus allocations (like -1ULL) to
5469 			 * 0.
5470 			 */
5471 			if (size < regs[r1] ||
5472 			    !DTRACE_INSCRATCH(mstate, size)) {
5473 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5474 				regs[rd] = NULL;
5475 				break;
5476 			}
5477 
5478 			dtrace_bzero((void *) mstate->dtms_scratch_ptr, size);
5479 			mstate->dtms_scratch_ptr += size;
5480 			regs[rd] = ptr;
5481 			break;
5482 		}
5483 
5484 		case DIF_OP_COPYS:
5485 			if (!dtrace_canstore(regs[rd], regs[r2],
5486 			    mstate, vstate)) {
5487 				*flags |= CPU_DTRACE_BADADDR;
5488 				*illval = regs[rd];
5489 				break;
5490 			}
5491 
5492 			if (!dtrace_canload(regs[r1], regs[r2], mstate, vstate))
5493 				break;
5494 
5495 			dtrace_bcopy((void *)(uintptr_t)regs[r1],
5496 			    (void *)(uintptr_t)regs[rd], (size_t)regs[r2]);
5497 			break;
5498 
5499 		case DIF_OP_STB:
5500 			if (!dtrace_canstore(regs[rd], 1, mstate, vstate)) {
5501 				*flags |= CPU_DTRACE_BADADDR;
5502 				*illval = regs[rd];
5503 				break;
5504 			}
5505 			*((uint8_t *)(uintptr_t)regs[rd]) = (uint8_t)regs[r1];
5506 			break;
5507 
5508 		case DIF_OP_STH:
5509 			if (!dtrace_canstore(regs[rd], 2, mstate, vstate)) {
5510 				*flags |= CPU_DTRACE_BADADDR;
5511 				*illval = regs[rd];
5512 				break;
5513 			}
5514 			if (regs[rd] & 1) {
5515 				*flags |= CPU_DTRACE_BADALIGN;
5516 				*illval = regs[rd];
5517 				break;
5518 			}
5519 			*((uint16_t *)(uintptr_t)regs[rd]) = (uint16_t)regs[r1];
5520 			break;
5521 
5522 		case DIF_OP_STW:
5523 			if (!dtrace_canstore(regs[rd], 4, mstate, vstate)) {
5524 				*flags |= CPU_DTRACE_BADADDR;
5525 				*illval = regs[rd];
5526 				break;
5527 			}
5528 			if (regs[rd] & 3) {
5529 				*flags |= CPU_DTRACE_BADALIGN;
5530 				*illval = regs[rd];
5531 				break;
5532 			}
5533 			*((uint32_t *)(uintptr_t)regs[rd]) = (uint32_t)regs[r1];
5534 			break;
5535 
5536 		case DIF_OP_STX:
5537 			if (!dtrace_canstore(regs[rd], 8, mstate, vstate)) {
5538 				*flags |= CPU_DTRACE_BADADDR;
5539 				*illval = regs[rd];
5540 				break;
5541 			}
5542 			if (regs[rd] & 7) {
5543 				*flags |= CPU_DTRACE_BADALIGN;
5544 				*illval = regs[rd];
5545 				break;
5546 			}
5547 			*((uint64_t *)(uintptr_t)regs[rd]) = regs[r1];
5548 			break;
5549 		}
5550 	}
5551 
5552 	if (!(*flags & CPU_DTRACE_FAULT))
5553 		return (rval);
5554 
5555 	mstate->dtms_fltoffs = opc * sizeof (dif_instr_t);
5556 	mstate->dtms_present |= DTRACE_MSTATE_FLTOFFS;
5557 
5558 	return (0);
5559 }
5560 
5561 static void
5562 dtrace_action_breakpoint(dtrace_ecb_t *ecb)
5563 {
5564 	dtrace_probe_t *probe = ecb->dte_probe;
5565 	dtrace_provider_t *prov = probe->dtpr_provider;
5566 	char c[DTRACE_FULLNAMELEN + 80], *str;
5567 	char *msg = "dtrace: breakpoint action at probe ";
5568 	char *ecbmsg = " (ecb ";
5569 	uintptr_t mask = (0xf << (sizeof (uintptr_t) * NBBY / 4));
5570 	uintptr_t val = (uintptr_t)ecb;
5571 	int shift = (sizeof (uintptr_t) * NBBY) - 4, i = 0;
5572 
5573 	if (dtrace_destructive_disallow)
5574 		return;
5575 
5576 	/*
5577 	 * It's impossible to be taking action on the NULL probe.
5578 	 */
5579 	ASSERT(probe != NULL);
5580 
5581 	/*
5582 	 * This is a poor man's (destitute man's?) sprintf():  we want to
5583 	 * print the provider name, module name, function name and name of
5584 	 * the probe, along with the hex address of the ECB with the breakpoint
5585 	 * action -- all of which we must place in the character buffer by
5586 	 * hand.
5587 	 */
5588 	while (*msg != '\0')
5589 		c[i++] = *msg++;
5590 
5591 	for (str = prov->dtpv_name; *str != '\0'; str++)
5592 		c[i++] = *str;
5593 	c[i++] = ':';
5594 
5595 	for (str = probe->dtpr_mod; *str != '\0'; str++)
5596 		c[i++] = *str;
5597 	c[i++] = ':';
5598 
5599 	for (str = probe->dtpr_func; *str != '\0'; str++)
5600 		c[i++] = *str;
5601 	c[i++] = ':';
5602 
5603 	for (str = probe->dtpr_name; *str != '\0'; str++)
5604 		c[i++] = *str;
5605 
5606 	while (*ecbmsg != '\0')
5607 		c[i++] = *ecbmsg++;
5608 
5609 	while (shift >= 0) {
5610 		mask = (uintptr_t)0xf << shift;
5611 
5612 		if (val >= ((uintptr_t)1 << shift))
5613 			c[i++] = "0123456789abcdef"[(val & mask) >> shift];
5614 		shift -= 4;
5615 	}
5616 
5617 	c[i++] = ')';
5618 	c[i] = '\0';
5619 
5620 	debug_enter(c);
5621 }
5622 
5623 static void
5624 dtrace_action_panic(dtrace_ecb_t *ecb)
5625 {
5626 	dtrace_probe_t *probe = ecb->dte_probe;
5627 
5628 	/*
5629 	 * It's impossible to be taking action on the NULL probe.
5630 	 */
5631 	ASSERT(probe != NULL);
5632 
5633 	if (dtrace_destructive_disallow)
5634 		return;
5635 
5636 	if (dtrace_panicked != NULL)
5637 		return;
5638 
5639 	if (dtrace_casptr(&dtrace_panicked, NULL, curthread) != NULL)
5640 		return;
5641 
5642 	/*
5643 	 * We won the right to panic.  (We want to be sure that only one
5644 	 * thread calls panic() from dtrace_probe(), and that panic() is
5645 	 * called exactly once.)
5646 	 */
5647 	dtrace_panic("dtrace: panic action at probe %s:%s:%s:%s (ecb %p)",
5648 	    probe->dtpr_provider->dtpv_name, probe->dtpr_mod,
5649 	    probe->dtpr_func, probe->dtpr_name, (void *)ecb);
5650 }
5651 
5652 static void
5653 dtrace_action_raise(uint64_t sig)
5654 {
5655 	if (dtrace_destructive_disallow)
5656 		return;
5657 
5658 	if (sig >= NSIG) {
5659 		DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
5660 		return;
5661 	}
5662 
5663 	/*
5664 	 * raise() has a queue depth of 1 -- we ignore all subsequent
5665 	 * invocations of the raise() action.
5666 	 */
5667 	if (curthread->t_dtrace_sig == 0)
5668 		curthread->t_dtrace_sig = (uint8_t)sig;
5669 
5670 	curthread->t_sig_check = 1;
5671 	aston(curthread);
5672 }
5673 
5674 static void
5675 dtrace_action_stop(void)
5676 {
5677 	if (dtrace_destructive_disallow)
5678 		return;
5679 
5680 	if (!curthread->t_dtrace_stop) {
5681 		curthread->t_dtrace_stop = 1;
5682 		curthread->t_sig_check = 1;
5683 		aston(curthread);
5684 	}
5685 }
5686 
5687 static void
5688 dtrace_action_chill(dtrace_mstate_t *mstate, hrtime_t val)
5689 {
5690 	hrtime_t now;
5691 	volatile uint16_t *flags;
5692 	cpu_t *cpu = CPU;
5693 
5694 	if (dtrace_destructive_disallow)
5695 		return;
5696 
5697 	flags = (volatile uint16_t *)&cpu_core[cpu->cpu_id].cpuc_dtrace_flags;
5698 
5699 	now = dtrace_gethrtime();
5700 
5701 	if (now - cpu->cpu_dtrace_chillmark > dtrace_chill_interval) {
5702 		/*
5703 		 * We need to advance the mark to the current time.
5704 		 */
5705 		cpu->cpu_dtrace_chillmark = now;
5706 		cpu->cpu_dtrace_chilled = 0;
5707 	}
5708 
5709 	/*
5710 	 * Now check to see if the requested chill time would take us over
5711 	 * the maximum amount of time allowed in the chill interval.  (Or
5712 	 * worse, if the calculation itself induces overflow.)
5713 	 */
5714 	if (cpu->cpu_dtrace_chilled + val > dtrace_chill_max ||
5715 	    cpu->cpu_dtrace_chilled + val < cpu->cpu_dtrace_chilled) {
5716 		*flags |= CPU_DTRACE_ILLOP;
5717 		return;
5718 	}
5719 
5720 	while (dtrace_gethrtime() - now < val)
5721 		continue;
5722 
5723 	/*
5724 	 * Normally, we assure that the value of the variable "timestamp" does
5725 	 * not change within an ECB.  The presence of chill() represents an
5726 	 * exception to this rule, however.
5727 	 */
5728 	mstate->dtms_present &= ~DTRACE_MSTATE_TIMESTAMP;
5729 	cpu->cpu_dtrace_chilled += val;
5730 }
5731 
5732 static void
5733 dtrace_action_ustack(dtrace_mstate_t *mstate, dtrace_state_t *state,
5734     uint64_t *buf, uint64_t arg)
5735 {
5736 	int nframes = DTRACE_USTACK_NFRAMES(arg);
5737 	int strsize = DTRACE_USTACK_STRSIZE(arg);
5738 	uint64_t *pcs = &buf[1], *fps;
5739 	char *str = (char *)&pcs[nframes];
5740 	int size, offs = 0, i, j;
5741 	uintptr_t old = mstate->dtms_scratch_ptr, saved;
5742 	uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
5743 	char *sym;
5744 
5745 	/*
5746 	 * Should be taking a faster path if string space has not been
5747 	 * allocated.
5748 	 */
5749 	ASSERT(strsize != 0);
5750 
5751 	/*
5752 	 * We will first allocate some temporary space for the frame pointers.
5753 	 */
5754 	fps = (uint64_t *)P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
5755 	size = (uintptr_t)fps - mstate->dtms_scratch_ptr +
5756 	    (nframes * sizeof (uint64_t));
5757 
5758 	if (!DTRACE_INSCRATCH(mstate, size)) {
5759 		/*
5760 		 * Not enough room for our frame pointers -- need to indicate
5761 		 * that we ran out of scratch space.
5762 		 */
5763 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5764 		return;
5765 	}
5766 
5767 	mstate->dtms_scratch_ptr += size;
5768 	saved = mstate->dtms_scratch_ptr;
5769 
5770 	/*
5771 	 * Now get a stack with both program counters and frame pointers.
5772 	 */
5773 	DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5774 	dtrace_getufpstack(buf, fps, nframes + 1);
5775 	DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5776 
5777 	/*
5778 	 * If that faulted, we're cooked.
5779 	 */
5780 	if (*flags & CPU_DTRACE_FAULT)
5781 		goto out;
5782 
5783 	/*
5784 	 * Now we want to walk up the stack, calling the USTACK helper.  For
5785 	 * each iteration, we restore the scratch pointer.
5786 	 */
5787 	for (i = 0; i < nframes; i++) {
5788 		mstate->dtms_scratch_ptr = saved;
5789 
5790 		if (offs >= strsize)
5791 			break;
5792 
5793 		sym = (char *)(uintptr_t)dtrace_helper(
5794 		    DTRACE_HELPER_ACTION_USTACK,
5795 		    mstate, state, pcs[i], fps[i]);
5796 
5797 		/*
5798 		 * If we faulted while running the helper, we're going to
5799 		 * clear the fault and null out the corresponding string.
5800 		 */
5801 		if (*flags & CPU_DTRACE_FAULT) {
5802 			*flags &= ~CPU_DTRACE_FAULT;
5803 			str[offs++] = '\0';
5804 			continue;
5805 		}
5806 
5807 		if (sym == NULL) {
5808 			str[offs++] = '\0';
5809 			continue;
5810 		}
5811 
5812 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5813 
5814 		/*
5815 		 * Now copy in the string that the helper returned to us.
5816 		 */
5817 		for (j = 0; offs + j < strsize; j++) {
5818 			if ((str[offs + j] = sym[j]) == '\0')
5819 				break;
5820 		}
5821 
5822 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5823 
5824 		offs += j + 1;
5825 	}
5826 
5827 	if (offs >= strsize) {
5828 		/*
5829 		 * If we didn't have room for all of the strings, we don't
5830 		 * abort processing -- this needn't be a fatal error -- but we
5831 		 * still want to increment a counter (dts_stkstroverflows) to
5832 		 * allow this condition to be warned about.  (If this is from
5833 		 * a jstack() action, it is easily tuned via jstackstrsize.)
5834 		 */
5835 		dtrace_error(&state->dts_stkstroverflows);
5836 	}
5837 
5838 	while (offs < strsize)
5839 		str[offs++] = '\0';
5840 
5841 out:
5842 	mstate->dtms_scratch_ptr = old;
5843 }
5844 
5845 /*
5846  * If you're looking for the epicenter of DTrace, you just found it.  This
5847  * is the function called by the provider to fire a probe -- from which all
5848  * subsequent probe-context DTrace activity emanates.
5849  */
5850 void
5851 dtrace_probe(dtrace_id_t id, uintptr_t arg0, uintptr_t arg1,
5852     uintptr_t arg2, uintptr_t arg3, uintptr_t arg4)
5853 {
5854 	processorid_t cpuid;
5855 	dtrace_icookie_t cookie;
5856 	dtrace_probe_t *probe;
5857 	dtrace_mstate_t mstate;
5858 	dtrace_ecb_t *ecb;
5859 	dtrace_action_t *act;
5860 	intptr_t offs;
5861 	size_t size;
5862 	int vtime, onintr;
5863 	volatile uint16_t *flags;
5864 	hrtime_t now, end;
5865 
5866 	/*
5867 	 * Kick out immediately if this CPU is still being born (in which case
5868 	 * curthread will be set to -1) or the current thread can't allow
5869 	 * probes in its current context.
5870 	 */
5871 	if (((uintptr_t)curthread & 1) || (curthread->t_flag & T_DONTDTRACE))
5872 		return;
5873 
5874 	cookie = dtrace_interrupt_disable();
5875 	probe = dtrace_probes[id - 1];
5876 	cpuid = CPU->cpu_id;
5877 	onintr = CPU_ON_INTR(CPU);
5878 
5879 	CPU->cpu_dtrace_probes++;
5880 
5881 	if (!onintr && probe->dtpr_predcache != DTRACE_CACHEIDNONE &&
5882 	    probe->dtpr_predcache == curthread->t_predcache) {
5883 		/*
5884 		 * We have hit in the predicate cache; we know that
5885 		 * this predicate would evaluate to be false.
5886 		 */
5887 		dtrace_interrupt_enable(cookie);
5888 		return;
5889 	}
5890 
5891 	if (panic_quiesce) {
5892 		/*
5893 		 * We don't trace anything if we're panicking.
5894 		 */
5895 		dtrace_interrupt_enable(cookie);
5896 		return;
5897 	}
5898 
5899 	now = dtrace_gethrtime();
5900 	vtime = dtrace_vtime_references != 0;
5901 
5902 	if (vtime && curthread->t_dtrace_start)
5903 		curthread->t_dtrace_vtime += now - curthread->t_dtrace_start;
5904 
5905 	mstate.dtms_difo = NULL;
5906 	mstate.dtms_probe = probe;
5907 	mstate.dtms_strtok = NULL;
5908 	mstate.dtms_arg[0] = arg0;
5909 	mstate.dtms_arg[1] = arg1;
5910 	mstate.dtms_arg[2] = arg2;
5911 	mstate.dtms_arg[3] = arg3;
5912 	mstate.dtms_arg[4] = arg4;
5913 
5914 	flags = (volatile uint16_t *)&cpu_core[cpuid].cpuc_dtrace_flags;
5915 
5916 	for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
5917 		dtrace_predicate_t *pred = ecb->dte_predicate;
5918 		dtrace_state_t *state = ecb->dte_state;
5919 		dtrace_buffer_t *buf = &state->dts_buffer[cpuid];
5920 		dtrace_buffer_t *aggbuf = &state->dts_aggbuffer[cpuid];
5921 		dtrace_vstate_t *vstate = &state->dts_vstate;
5922 		dtrace_provider_t *prov = probe->dtpr_provider;
5923 		uint64_t tracememsize = 0;
5924 		int committed = 0;
5925 		caddr_t tomax;
5926 
5927 		/*
5928 		 * A little subtlety with the following (seemingly innocuous)
5929 		 * declaration of the automatic 'val':  by looking at the
5930 		 * code, you might think that it could be declared in the
5931 		 * action processing loop, below.  (That is, it's only used in
5932 		 * the action processing loop.)  However, it must be declared
5933 		 * out of that scope because in the case of DIF expression
5934 		 * arguments to aggregating actions, one iteration of the
5935 		 * action loop will use the last iteration's value.
5936 		 */
5937 #ifdef lint
5938 		uint64_t val = 0;
5939 #else
5940 		uint64_t val;
5941 #endif
5942 
5943 		mstate.dtms_present = DTRACE_MSTATE_ARGS | DTRACE_MSTATE_PROBE;
5944 		mstate.dtms_access = DTRACE_ACCESS_ARGS | DTRACE_ACCESS_PROC;
5945 		*flags &= ~CPU_DTRACE_ERROR;
5946 
5947 		if (prov == dtrace_provider) {
5948 			/*
5949 			 * If dtrace itself is the provider of this probe,
5950 			 * we're only going to continue processing the ECB if
5951 			 * arg0 (the dtrace_state_t) is equal to the ECB's
5952 			 * creating state.  (This prevents disjoint consumers
5953 			 * from seeing one another's metaprobes.)
5954 			 */
5955 			if (arg0 != (uint64_t)(uintptr_t)state)
5956 				continue;
5957 		}
5958 
5959 		if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE) {
5960 			/*
5961 			 * We're not currently active.  If our provider isn't
5962 			 * the dtrace pseudo provider, we're not interested.
5963 			 */
5964 			if (prov != dtrace_provider)
5965 				continue;
5966 
5967 			/*
5968 			 * Now we must further check if we are in the BEGIN
5969 			 * probe.  If we are, we will only continue processing
5970 			 * if we're still in WARMUP -- if one BEGIN enabling
5971 			 * has invoked the exit() action, we don't want to
5972 			 * evaluate subsequent BEGIN enablings.
5973 			 */
5974 			if (probe->dtpr_id == dtrace_probeid_begin &&
5975 			    state->dts_activity != DTRACE_ACTIVITY_WARMUP) {
5976 				ASSERT(state->dts_activity ==
5977 				    DTRACE_ACTIVITY_DRAINING);
5978 				continue;
5979 			}
5980 		}
5981 
5982 		if (ecb->dte_cond && !dtrace_priv_probe(state, &mstate, ecb))
5983 			continue;
5984 
5985 		if (now - state->dts_alive > dtrace_deadman_timeout) {
5986 			/*
5987 			 * We seem to be dead.  Unless we (a) have kernel
5988 			 * destructive permissions (b) have explicitly enabled
5989 			 * destructive actions and (c) destructive actions have
5990 			 * not been disabled, we're going to transition into
5991 			 * the KILLED state, from which no further processing
5992 			 * on this state will be performed.
5993 			 */
5994 			if (!dtrace_priv_kernel_destructive(state) ||
5995 			    !state->dts_cred.dcr_destructive ||
5996 			    dtrace_destructive_disallow) {
5997 				void *activity = &state->dts_activity;
5998 				dtrace_activity_t current;
5999 
6000 				do {
6001 					current = state->dts_activity;
6002 				} while (dtrace_cas32(activity, current,
6003 				    DTRACE_ACTIVITY_KILLED) != current);
6004 
6005 				continue;
6006 			}
6007 		}
6008 
6009 		if ((offs = dtrace_buffer_reserve(buf, ecb->dte_needed,
6010 		    ecb->dte_alignment, state, &mstate)) < 0)
6011 			continue;
6012 
6013 		tomax = buf->dtb_tomax;
6014 		ASSERT(tomax != NULL);
6015 
6016 		if (ecb->dte_size != 0) {
6017 			dtrace_rechdr_t dtrh;
6018 			if (!(mstate.dtms_present & DTRACE_MSTATE_TIMESTAMP)) {
6019 				mstate.dtms_timestamp = dtrace_gethrtime();
6020 				mstate.dtms_present |= DTRACE_MSTATE_TIMESTAMP;
6021 			}
6022 			ASSERT3U(ecb->dte_size, >=, sizeof (dtrace_rechdr_t));
6023 			dtrh.dtrh_epid = ecb->dte_epid;
6024 			DTRACE_RECORD_STORE_TIMESTAMP(&dtrh,
6025 			    mstate.dtms_timestamp);
6026 			*((dtrace_rechdr_t *)(tomax + offs)) = dtrh;
6027 		}
6028 
6029 		mstate.dtms_epid = ecb->dte_epid;
6030 		mstate.dtms_present |= DTRACE_MSTATE_EPID;
6031 
6032 		if (state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)
6033 			mstate.dtms_access |= DTRACE_ACCESS_KERNEL;
6034 
6035 		if (pred != NULL) {
6036 			dtrace_difo_t *dp = pred->dtp_difo;
6037 			int rval;
6038 
6039 			rval = dtrace_dif_emulate(dp, &mstate, vstate, state);
6040 
6041 			if (!(*flags & CPU_DTRACE_ERROR) && !rval) {
6042 				dtrace_cacheid_t cid = probe->dtpr_predcache;
6043 
6044 				if (cid != DTRACE_CACHEIDNONE && !onintr) {
6045 					/*
6046 					 * Update the predicate cache...
6047 					 */
6048 					ASSERT(cid == pred->dtp_cacheid);
6049 					curthread->t_predcache = cid;
6050 				}
6051 
6052 				continue;
6053 			}
6054 		}
6055 
6056 		for (act = ecb->dte_action; !(*flags & CPU_DTRACE_ERROR) &&
6057 		    act != NULL; act = act->dta_next) {
6058 			size_t valoffs;
6059 			dtrace_difo_t *dp;
6060 			dtrace_recdesc_t *rec = &act->dta_rec;
6061 
6062 			size = rec->dtrd_size;
6063 			valoffs = offs + rec->dtrd_offset;
6064 
6065 			if (DTRACEACT_ISAGG(act->dta_kind)) {
6066 				uint64_t v = 0xbad;
6067 				dtrace_aggregation_t *agg;
6068 
6069 				agg = (dtrace_aggregation_t *)act;
6070 
6071 				if ((dp = act->dta_difo) != NULL)
6072 					v = dtrace_dif_emulate(dp,
6073 					    &mstate, vstate, state);
6074 
6075 				if (*flags & CPU_DTRACE_ERROR)
6076 					continue;
6077 
6078 				/*
6079 				 * Note that we always pass the expression
6080 				 * value from the previous iteration of the
6081 				 * action loop.  This value will only be used
6082 				 * if there is an expression argument to the
6083 				 * aggregating action, denoted by the
6084 				 * dtag_hasarg field.
6085 				 */
6086 				dtrace_aggregate(agg, buf,
6087 				    offs, aggbuf, v, val);
6088 				continue;
6089 			}
6090 
6091 			switch (act->dta_kind) {
6092 			case DTRACEACT_STOP:
6093 				if (dtrace_priv_proc_destructive(state,
6094 				    &mstate))
6095 					dtrace_action_stop();
6096 				continue;
6097 
6098 			case DTRACEACT_BREAKPOINT:
6099 				if (dtrace_priv_kernel_destructive(state))
6100 					dtrace_action_breakpoint(ecb);
6101 				continue;
6102 
6103 			case DTRACEACT_PANIC:
6104 				if (dtrace_priv_kernel_destructive(state))
6105 					dtrace_action_panic(ecb);
6106 				continue;
6107 
6108 			case DTRACEACT_STACK:
6109 				if (!dtrace_priv_kernel(state))
6110 					continue;
6111 
6112 				dtrace_getpcstack((pc_t *)(tomax + valoffs),
6113 				    size / sizeof (pc_t), probe->dtpr_aframes,
6114 				    DTRACE_ANCHORED(probe) ? NULL :
6115 				    (uint32_t *)arg0);
6116 
6117 				continue;
6118 
6119 			case DTRACEACT_JSTACK:
6120 			case DTRACEACT_USTACK:
6121 				if (!dtrace_priv_proc(state, &mstate))
6122 					continue;
6123 
6124 				/*
6125 				 * See comment in DIF_VAR_PID.
6126 				 */
6127 				if (DTRACE_ANCHORED(mstate.dtms_probe) &&
6128 				    CPU_ON_INTR(CPU)) {
6129 					int depth = DTRACE_USTACK_NFRAMES(
6130 					    rec->dtrd_arg) + 1;
6131 
6132 					dtrace_bzero((void *)(tomax + valoffs),
6133 					    DTRACE_USTACK_STRSIZE(rec->dtrd_arg)
6134 					    + depth * sizeof (uint64_t));
6135 
6136 					continue;
6137 				}
6138 
6139 				if (DTRACE_USTACK_STRSIZE(rec->dtrd_arg) != 0 &&
6140 				    curproc->p_dtrace_helpers != NULL) {
6141 					/*
6142 					 * This is the slow path -- we have
6143 					 * allocated string space, and we're
6144 					 * getting the stack of a process that
6145 					 * has helpers.  Call into a separate
6146 					 * routine to perform this processing.
6147 					 */
6148 					dtrace_action_ustack(&mstate, state,
6149 					    (uint64_t *)(tomax + valoffs),
6150 					    rec->dtrd_arg);
6151 					continue;
6152 				}
6153 
6154 				/*
6155 				 * Clear the string space, since there's no
6156 				 * helper to do it for us.
6157 				 */
6158 				if (DTRACE_USTACK_STRSIZE(rec->dtrd_arg) != 0) {
6159 					int depth = DTRACE_USTACK_NFRAMES(
6160 					    rec->dtrd_arg);
6161 					size_t strsize = DTRACE_USTACK_STRSIZE(
6162 					    rec->dtrd_arg);
6163 					uint64_t *buf = (uint64_t *)(tomax +
6164 					    valoffs);
6165 					void *strspace = &buf[depth + 1];
6166 
6167 					dtrace_bzero(strspace,
6168 					    MIN(depth, strsize));
6169 				}
6170 
6171 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6172 				dtrace_getupcstack((uint64_t *)
6173 				    (tomax + valoffs),
6174 				    DTRACE_USTACK_NFRAMES(rec->dtrd_arg) + 1);
6175 				DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6176 				continue;
6177 
6178 			default:
6179 				break;
6180 			}
6181 
6182 			dp = act->dta_difo;
6183 			ASSERT(dp != NULL);
6184 
6185 			val = dtrace_dif_emulate(dp, &mstate, vstate, state);
6186 
6187 			if (*flags & CPU_DTRACE_ERROR)
6188 				continue;
6189 
6190 			switch (act->dta_kind) {
6191 			case DTRACEACT_SPECULATE: {
6192 				dtrace_rechdr_t *dtrh;
6193 
6194 				ASSERT(buf == &state->dts_buffer[cpuid]);
6195 				buf = dtrace_speculation_buffer(state,
6196 				    cpuid, val);
6197 
6198 				if (buf == NULL) {
6199 					*flags |= CPU_DTRACE_DROP;
6200 					continue;
6201 				}
6202 
6203 				offs = dtrace_buffer_reserve(buf,
6204 				    ecb->dte_needed, ecb->dte_alignment,
6205 				    state, NULL);
6206 
6207 				if (offs < 0) {
6208 					*flags |= CPU_DTRACE_DROP;
6209 					continue;
6210 				}
6211 
6212 				tomax = buf->dtb_tomax;
6213 				ASSERT(tomax != NULL);
6214 
6215 				if (ecb->dte_size == 0)
6216 					continue;
6217 
6218 				ASSERT3U(ecb->dte_size, >=,
6219 				    sizeof (dtrace_rechdr_t));
6220 				dtrh = ((void *)(tomax + offs));
6221 				dtrh->dtrh_epid = ecb->dte_epid;
6222 				/*
6223 				 * When the speculation is committed, all of
6224 				 * the records in the speculative buffer will
6225 				 * have their timestamps set to the commit
6226 				 * time.  Until then, it is set to a sentinel
6227 				 * value, for debugability.
6228 				 */
6229 				DTRACE_RECORD_STORE_TIMESTAMP(dtrh, UINT64_MAX);
6230 				continue;
6231 			}
6232 
6233 			case DTRACEACT_CHILL:
6234 				if (dtrace_priv_kernel_destructive(state))
6235 					dtrace_action_chill(&mstate, val);
6236 				continue;
6237 
6238 			case DTRACEACT_RAISE:
6239 				if (dtrace_priv_proc_destructive(state,
6240 				    &mstate))
6241 					dtrace_action_raise(val);
6242 				continue;
6243 
6244 			case DTRACEACT_COMMIT:
6245 				ASSERT(!committed);
6246 
6247 				/*
6248 				 * We need to commit our buffer state.
6249 				 */
6250 				if (ecb->dte_size)
6251 					buf->dtb_offset = offs + ecb->dte_size;
6252 				buf = &state->dts_buffer[cpuid];
6253 				dtrace_speculation_commit(state, cpuid, val);
6254 				committed = 1;
6255 				continue;
6256 
6257 			case DTRACEACT_DISCARD:
6258 				dtrace_speculation_discard(state, cpuid, val);
6259 				continue;
6260 
6261 			case DTRACEACT_DIFEXPR:
6262 			case DTRACEACT_LIBACT:
6263 			case DTRACEACT_PRINTF:
6264 			case DTRACEACT_PRINTA:
6265 			case DTRACEACT_SYSTEM:
6266 			case DTRACEACT_FREOPEN:
6267 			case DTRACEACT_TRACEMEM:
6268 				break;
6269 
6270 			case DTRACEACT_TRACEMEM_DYNSIZE:
6271 				tracememsize = val;
6272 				break;
6273 
6274 			case DTRACEACT_SYM:
6275 			case DTRACEACT_MOD:
6276 				if (!dtrace_priv_kernel(state))
6277 					continue;
6278 				break;
6279 
6280 			case DTRACEACT_USYM:
6281 			case DTRACEACT_UMOD:
6282 			case DTRACEACT_UADDR: {
6283 				struct pid *pid = curthread->t_procp->p_pidp;
6284 
6285 				if (!dtrace_priv_proc(state, &mstate))
6286 					continue;
6287 
6288 				DTRACE_STORE(uint64_t, tomax,
6289 				    valoffs, (uint64_t)pid->pid_id);
6290 				DTRACE_STORE(uint64_t, tomax,
6291 				    valoffs + sizeof (uint64_t), val);
6292 
6293 				continue;
6294 			}
6295 
6296 			case DTRACEACT_EXIT: {
6297 				/*
6298 				 * For the exit action, we are going to attempt
6299 				 * to atomically set our activity to be
6300 				 * draining.  If this fails (either because
6301 				 * another CPU has beat us to the exit action,
6302 				 * or because our current activity is something
6303 				 * other than ACTIVE or WARMUP), we will
6304 				 * continue.  This assures that the exit action
6305 				 * can be successfully recorded at most once
6306 				 * when we're in the ACTIVE state.  If we're
6307 				 * encountering the exit() action while in
6308 				 * COOLDOWN, however, we want to honor the new
6309 				 * status code.  (We know that we're the only
6310 				 * thread in COOLDOWN, so there is no race.)
6311 				 */
6312 				void *activity = &state->dts_activity;
6313 				dtrace_activity_t current = state->dts_activity;
6314 
6315 				if (current == DTRACE_ACTIVITY_COOLDOWN)
6316 					break;
6317 
6318 				if (current != DTRACE_ACTIVITY_WARMUP)
6319 					current = DTRACE_ACTIVITY_ACTIVE;
6320 
6321 				if (dtrace_cas32(activity, current,
6322 				    DTRACE_ACTIVITY_DRAINING) != current) {
6323 					*flags |= CPU_DTRACE_DROP;
6324 					continue;
6325 				}
6326 
6327 				break;
6328 			}
6329 
6330 			default:
6331 				ASSERT(0);
6332 			}
6333 
6334 			if (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF) {
6335 				uintptr_t end = valoffs + size;
6336 
6337 				if (tracememsize != 0 &&
6338 				    valoffs + tracememsize < end) {
6339 					end = valoffs + tracememsize;
6340 					tracememsize = 0;
6341 				}
6342 
6343 				if (!dtrace_vcanload((void *)(uintptr_t)val,
6344 				    &dp->dtdo_rtype, &mstate, vstate))
6345 					continue;
6346 
6347 				/*
6348 				 * If this is a string, we're going to only
6349 				 * load until we find the zero byte -- after
6350 				 * which we'll store zero bytes.
6351 				 */
6352 				if (dp->dtdo_rtype.dtdt_kind ==
6353 				    DIF_TYPE_STRING) {
6354 					char c = '\0' + 1;
6355 					int intuple = act->dta_intuple;
6356 					size_t s;
6357 
6358 					for (s = 0; s < size; s++) {
6359 						if (c != '\0')
6360 							c = dtrace_load8(val++);
6361 
6362 						DTRACE_STORE(uint8_t, tomax,
6363 						    valoffs++, c);
6364 
6365 						if (c == '\0' && intuple)
6366 							break;
6367 					}
6368 
6369 					continue;
6370 				}
6371 
6372 				while (valoffs < end) {
6373 					DTRACE_STORE(uint8_t, tomax, valoffs++,
6374 					    dtrace_load8(val++));
6375 				}
6376 
6377 				continue;
6378 			}
6379 
6380 			switch (size) {
6381 			case 0:
6382 				break;
6383 
6384 			case sizeof (uint8_t):
6385 				DTRACE_STORE(uint8_t, tomax, valoffs, val);
6386 				break;
6387 			case sizeof (uint16_t):
6388 				DTRACE_STORE(uint16_t, tomax, valoffs, val);
6389 				break;
6390 			case sizeof (uint32_t):
6391 				DTRACE_STORE(uint32_t, tomax, valoffs, val);
6392 				break;
6393 			case sizeof (uint64_t):
6394 				DTRACE_STORE(uint64_t, tomax, valoffs, val);
6395 				break;
6396 			default:
6397 				/*
6398 				 * Any other size should have been returned by
6399 				 * reference, not by value.
6400 				 */
6401 				ASSERT(0);
6402 				break;
6403 			}
6404 		}
6405 
6406 		if (*flags & CPU_DTRACE_DROP)
6407 			continue;
6408 
6409 		if (*flags & CPU_DTRACE_FAULT) {
6410 			int ndx;
6411 			dtrace_action_t *err;
6412 
6413 			buf->dtb_errors++;
6414 
6415 			if (probe->dtpr_id == dtrace_probeid_error) {
6416 				/*
6417 				 * There's nothing we can do -- we had an
6418 				 * error on the error probe.  We bump an
6419 				 * error counter to at least indicate that
6420 				 * this condition happened.
6421 				 */
6422 				dtrace_error(&state->dts_dblerrors);
6423 				continue;
6424 			}
6425 
6426 			if (vtime) {
6427 				/*
6428 				 * Before recursing on dtrace_probe(), we
6429 				 * need to explicitly clear out our start
6430 				 * time to prevent it from being accumulated
6431 				 * into t_dtrace_vtime.
6432 				 */
6433 				curthread->t_dtrace_start = 0;
6434 			}
6435 
6436 			/*
6437 			 * Iterate over the actions to figure out which action
6438 			 * we were processing when we experienced the error.
6439 			 * Note that act points _past_ the faulting action; if
6440 			 * act is ecb->dte_action, the fault was in the
6441 			 * predicate, if it's ecb->dte_action->dta_next it's
6442 			 * in action #1, and so on.
6443 			 */
6444 			for (err = ecb->dte_action, ndx = 0;
6445 			    err != act; err = err->dta_next, ndx++)
6446 				continue;
6447 
6448 			dtrace_probe_error(state, ecb->dte_epid, ndx,
6449 			    (mstate.dtms_present & DTRACE_MSTATE_FLTOFFS) ?
6450 			    mstate.dtms_fltoffs : -1, DTRACE_FLAGS2FLT(*flags),
6451 			    cpu_core[cpuid].cpuc_dtrace_illval);
6452 
6453 			continue;
6454 		}
6455 
6456 		if (!committed)
6457 			buf->dtb_offset = offs + ecb->dte_size;
6458 	}
6459 
6460 	end = dtrace_gethrtime();
6461 	if (vtime)
6462 		curthread->t_dtrace_start = end;
6463 
6464 	CPU->cpu_dtrace_nsec += end - now;
6465 
6466 	dtrace_interrupt_enable(cookie);
6467 }
6468 
6469 /*
6470  * DTrace Probe Hashing Functions
6471  *
6472  * The functions in this section (and indeed, the functions in remaining
6473  * sections) are not _called_ from probe context.  (Any exceptions to this are
6474  * marked with a "Note:".)  Rather, they are called from elsewhere in the
6475  * DTrace framework to look-up probes in, add probes to and remove probes from
6476  * the DTrace probe hashes.  (Each probe is hashed by each element of the
6477  * probe tuple -- allowing for fast lookups, regardless of what was
6478  * specified.)
6479  */
6480 static uint_t
6481 dtrace_hash_str(char *p)
6482 {
6483 	unsigned int g;
6484 	uint_t hval = 0;
6485 
6486 	while (*p) {
6487 		hval = (hval << 4) + *p++;
6488 		if ((g = (hval & 0xf0000000)) != 0)
6489 			hval ^= g >> 24;
6490 		hval &= ~g;
6491 	}
6492 	return (hval);
6493 }
6494 
6495 static dtrace_hash_t *
6496 dtrace_hash_create(uintptr_t stroffs, uintptr_t nextoffs, uintptr_t prevoffs)
6497 {
6498 	dtrace_hash_t *hash = kmem_zalloc(sizeof (dtrace_hash_t), KM_SLEEP);
6499 
6500 	hash->dth_stroffs = stroffs;
6501 	hash->dth_nextoffs = nextoffs;
6502 	hash->dth_prevoffs = prevoffs;
6503 
6504 	hash->dth_size = 1;
6505 	hash->dth_mask = hash->dth_size - 1;
6506 
6507 	hash->dth_tab = kmem_zalloc(hash->dth_size *
6508 	    sizeof (dtrace_hashbucket_t *), KM_SLEEP);
6509 
6510 	return (hash);
6511 }
6512 
6513 static void
6514 dtrace_hash_destroy(dtrace_hash_t *hash)
6515 {
6516 #ifdef DEBUG
6517 	int i;
6518 
6519 	for (i = 0; i < hash->dth_size; i++)
6520 		ASSERT(hash->dth_tab[i] == NULL);
6521 #endif
6522 
6523 	kmem_free(hash->dth_tab,
6524 	    hash->dth_size * sizeof (dtrace_hashbucket_t *));
6525 	kmem_free(hash, sizeof (dtrace_hash_t));
6526 }
6527 
6528 static void
6529 dtrace_hash_resize(dtrace_hash_t *hash)
6530 {
6531 	int size = hash->dth_size, i, ndx;
6532 	int new_size = hash->dth_size << 1;
6533 	int new_mask = new_size - 1;
6534 	dtrace_hashbucket_t **new_tab, *bucket, *next;
6535 
6536 	ASSERT((new_size & new_mask) == 0);
6537 
6538 	new_tab = kmem_zalloc(new_size * sizeof (void *), KM_SLEEP);
6539 
6540 	for (i = 0; i < size; i++) {
6541 		for (bucket = hash->dth_tab[i]; bucket != NULL; bucket = next) {
6542 			dtrace_probe_t *probe = bucket->dthb_chain;
6543 
6544 			ASSERT(probe != NULL);
6545 			ndx = DTRACE_HASHSTR(hash, probe) & new_mask;
6546 
6547 			next = bucket->dthb_next;
6548 			bucket->dthb_next = new_tab[ndx];
6549 			new_tab[ndx] = bucket;
6550 		}
6551 	}
6552 
6553 	kmem_free(hash->dth_tab, hash->dth_size * sizeof (void *));
6554 	hash->dth_tab = new_tab;
6555 	hash->dth_size = new_size;
6556 	hash->dth_mask = new_mask;
6557 }
6558 
6559 static void
6560 dtrace_hash_add(dtrace_hash_t *hash, dtrace_probe_t *new)
6561 {
6562 	int hashval = DTRACE_HASHSTR(hash, new);
6563 	int ndx = hashval & hash->dth_mask;
6564 	dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6565 	dtrace_probe_t **nextp, **prevp;
6566 
6567 	for (; bucket != NULL; bucket = bucket->dthb_next) {
6568 		if (DTRACE_HASHEQ(hash, bucket->dthb_chain, new))
6569 			goto add;
6570 	}
6571 
6572 	if ((hash->dth_nbuckets >> 1) > hash->dth_size) {
6573 		dtrace_hash_resize(hash);
6574 		dtrace_hash_add(hash, new);
6575 		return;
6576 	}
6577 
6578 	bucket = kmem_zalloc(sizeof (dtrace_hashbucket_t), KM_SLEEP);
6579 	bucket->dthb_next = hash->dth_tab[ndx];
6580 	hash->dth_tab[ndx] = bucket;
6581 	hash->dth_nbuckets++;
6582 
6583 add:
6584 	nextp = DTRACE_HASHNEXT(hash, new);
6585 	ASSERT(*nextp == NULL && *(DTRACE_HASHPREV(hash, new)) == NULL);
6586 	*nextp = bucket->dthb_chain;
6587 
6588 	if (bucket->dthb_chain != NULL) {
6589 		prevp = DTRACE_HASHPREV(hash, bucket->dthb_chain);
6590 		ASSERT(*prevp == NULL);
6591 		*prevp = new;
6592 	}
6593 
6594 	bucket->dthb_chain = new;
6595 	bucket->dthb_len++;
6596 }
6597 
6598 static dtrace_probe_t *
6599 dtrace_hash_lookup(dtrace_hash_t *hash, dtrace_probe_t *template)
6600 {
6601 	int hashval = DTRACE_HASHSTR(hash, template);
6602 	int ndx = hashval & hash->dth_mask;
6603 	dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6604 
6605 	for (; bucket != NULL; bucket = bucket->dthb_next) {
6606 		if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
6607 			return (bucket->dthb_chain);
6608 	}
6609 
6610 	return (NULL);
6611 }
6612 
6613 static int
6614 dtrace_hash_collisions(dtrace_hash_t *hash, dtrace_probe_t *template)
6615 {
6616 	int hashval = DTRACE_HASHSTR(hash, template);
6617 	int ndx = hashval & hash->dth_mask;
6618 	dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6619 
6620 	for (; bucket != NULL; bucket = bucket->dthb_next) {
6621 		if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
6622 			return (bucket->dthb_len);
6623 	}
6624 
6625 	return (NULL);
6626 }
6627 
6628 static void
6629 dtrace_hash_remove(dtrace_hash_t *hash, dtrace_probe_t *probe)
6630 {
6631 	int ndx = DTRACE_HASHSTR(hash, probe) & hash->dth_mask;
6632 	dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6633 
6634 	dtrace_probe_t **prevp = DTRACE_HASHPREV(hash, probe);
6635 	dtrace_probe_t **nextp = DTRACE_HASHNEXT(hash, probe);
6636 
6637 	/*
6638 	 * Find the bucket that we're removing this probe from.
6639 	 */
6640 	for (; bucket != NULL; bucket = bucket->dthb_next) {
6641 		if (DTRACE_HASHEQ(hash, bucket->dthb_chain, probe))
6642 			break;
6643 	}
6644 
6645 	ASSERT(bucket != NULL);
6646 
6647 	if (*prevp == NULL) {
6648 		if (*nextp == NULL) {
6649 			/*
6650 			 * The removed probe was the only probe on this
6651 			 * bucket; we need to remove the bucket.
6652 			 */
6653 			dtrace_hashbucket_t *b = hash->dth_tab[ndx];
6654 
6655 			ASSERT(bucket->dthb_chain == probe);
6656 			ASSERT(b != NULL);
6657 
6658 			if (b == bucket) {
6659 				hash->dth_tab[ndx] = bucket->dthb_next;
6660 			} else {
6661 				while (b->dthb_next != bucket)
6662 					b = b->dthb_next;
6663 				b->dthb_next = bucket->dthb_next;
6664 			}
6665 
6666 			ASSERT(hash->dth_nbuckets > 0);
6667 			hash->dth_nbuckets--;
6668 			kmem_free(bucket, sizeof (dtrace_hashbucket_t));
6669 			return;
6670 		}
6671 
6672 		bucket->dthb_chain = *nextp;
6673 	} else {
6674 		*(DTRACE_HASHNEXT(hash, *prevp)) = *nextp;
6675 	}
6676 
6677 	if (*nextp != NULL)
6678 		*(DTRACE_HASHPREV(hash, *nextp)) = *prevp;
6679 }
6680 
6681 /*
6682  * DTrace Utility Functions
6683  *
6684  * These are random utility functions that are _not_ called from probe context.
6685  */
6686 static int
6687 dtrace_badattr(const dtrace_attribute_t *a)
6688 {
6689 	return (a->dtat_name > DTRACE_STABILITY_MAX ||
6690 	    a->dtat_data > DTRACE_STABILITY_MAX ||
6691 	    a->dtat_class > DTRACE_CLASS_MAX);
6692 }
6693 
6694 /*
6695  * Return a duplicate copy of a string.  If the specified string is NULL,
6696  * this function returns a zero-length string.
6697  */
6698 static char *
6699 dtrace_strdup(const char *str)
6700 {
6701 	char *new = kmem_zalloc((str != NULL ? strlen(str) : 0) + 1, KM_SLEEP);
6702 
6703 	if (str != NULL)
6704 		(void) strcpy(new, str);
6705 
6706 	return (new);
6707 }
6708 
6709 #define	DTRACE_ISALPHA(c)	\
6710 	(((c) >= 'a' && (c) <= 'z') || ((c) >= 'A' && (c) <= 'Z'))
6711 
6712 static int
6713 dtrace_badname(const char *s)
6714 {
6715 	char c;
6716 
6717 	if (s == NULL || (c = *s++) == '\0')
6718 		return (0);
6719 
6720 	if (!DTRACE_ISALPHA(c) && c != '-' && c != '_' && c != '.')
6721 		return (1);
6722 
6723 	while ((c = *s++) != '\0') {
6724 		if (!DTRACE_ISALPHA(c) && (c < '0' || c > '9') &&
6725 		    c != '-' && c != '_' && c != '.' && c != '`')
6726 			return (1);
6727 	}
6728 
6729 	return (0);
6730 }
6731 
6732 static void
6733 dtrace_cred2priv(cred_t *cr, uint32_t *privp, uid_t *uidp, zoneid_t *zoneidp)
6734 {
6735 	uint32_t priv;
6736 
6737 	if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
6738 		/*
6739 		 * For DTRACE_PRIV_ALL, the uid and zoneid don't matter.
6740 		 */
6741 		priv = DTRACE_PRIV_ALL;
6742 	} else {
6743 		*uidp = crgetuid(cr);
6744 		*zoneidp = crgetzoneid(cr);
6745 
6746 		priv = 0;
6747 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE))
6748 			priv |= DTRACE_PRIV_KERNEL | DTRACE_PRIV_USER;
6749 		else if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE))
6750 			priv |= DTRACE_PRIV_USER;
6751 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE))
6752 			priv |= DTRACE_PRIV_PROC;
6753 		if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
6754 			priv |= DTRACE_PRIV_OWNER;
6755 		if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
6756 			priv |= DTRACE_PRIV_ZONEOWNER;
6757 	}
6758 
6759 	*privp = priv;
6760 }
6761 
6762 #ifdef DTRACE_ERRDEBUG
6763 static void
6764 dtrace_errdebug(const char *str)
6765 {
6766 	int hval = dtrace_hash_str((char *)str) % DTRACE_ERRHASHSZ;
6767 	int occupied = 0;
6768 
6769 	mutex_enter(&dtrace_errlock);
6770 	dtrace_errlast = str;
6771 	dtrace_errthread = curthread;
6772 
6773 	while (occupied++ < DTRACE_ERRHASHSZ) {
6774 		if (dtrace_errhash[hval].dter_msg == str) {
6775 			dtrace_errhash[hval].dter_count++;
6776 			goto out;
6777 		}
6778 
6779 		if (dtrace_errhash[hval].dter_msg != NULL) {
6780 			hval = (hval + 1) % DTRACE_ERRHASHSZ;
6781 			continue;
6782 		}
6783 
6784 		dtrace_errhash[hval].dter_msg = str;
6785 		dtrace_errhash[hval].dter_count = 1;
6786 		goto out;
6787 	}
6788 
6789 	panic("dtrace: undersized error hash");
6790 out:
6791 	mutex_exit(&dtrace_errlock);
6792 }
6793 #endif
6794 
6795 /*
6796  * DTrace Matching Functions
6797  *
6798  * These functions are used to match groups of probes, given some elements of
6799  * a probe tuple, or some globbed expressions for elements of a probe tuple.
6800  */
6801 static int
6802 dtrace_match_priv(const dtrace_probe_t *prp, uint32_t priv, uid_t uid,
6803     zoneid_t zoneid)
6804 {
6805 	if (priv != DTRACE_PRIV_ALL) {
6806 		uint32_t ppriv = prp->dtpr_provider->dtpv_priv.dtpp_flags;
6807 		uint32_t match = priv & ppriv;
6808 
6809 		/*
6810 		 * No PRIV_DTRACE_* privileges...
6811 		 */
6812 		if ((priv & (DTRACE_PRIV_PROC | DTRACE_PRIV_USER |
6813 		    DTRACE_PRIV_KERNEL)) == 0)
6814 			return (0);
6815 
6816 		/*
6817 		 * No matching bits, but there were bits to match...
6818 		 */
6819 		if (match == 0 && ppriv != 0)
6820 			return (0);
6821 
6822 		/*
6823 		 * Need to have permissions to the process, but don't...
6824 		 */
6825 		if (((ppriv & ~match) & DTRACE_PRIV_OWNER) != 0 &&
6826 		    uid != prp->dtpr_provider->dtpv_priv.dtpp_uid) {
6827 			return (0);
6828 		}
6829 
6830 		/*
6831 		 * Need to be in the same zone unless we possess the
6832 		 * privilege to examine all zones.
6833 		 */
6834 		if (((ppriv & ~match) & DTRACE_PRIV_ZONEOWNER) != 0 &&
6835 		    zoneid != prp->dtpr_provider->dtpv_priv.dtpp_zoneid) {
6836 			return (0);
6837 		}
6838 	}
6839 
6840 	return (1);
6841 }
6842 
6843 /*
6844  * dtrace_match_probe compares a dtrace_probe_t to a pre-compiled key, which
6845  * consists of input pattern strings and an ops-vector to evaluate them.
6846  * This function returns >0 for match, 0 for no match, and <0 for error.
6847  */
6848 static int
6849 dtrace_match_probe(const dtrace_probe_t *prp, const dtrace_probekey_t *pkp,
6850     uint32_t priv, uid_t uid, zoneid_t zoneid)
6851 {
6852 	dtrace_provider_t *pvp = prp->dtpr_provider;
6853 	int rv;
6854 
6855 	if (pvp->dtpv_defunct)
6856 		return (0);
6857 
6858 	if ((rv = pkp->dtpk_pmatch(pvp->dtpv_name, pkp->dtpk_prov, 0)) <= 0)
6859 		return (rv);
6860 
6861 	if ((rv = pkp->dtpk_mmatch(prp->dtpr_mod, pkp->dtpk_mod, 0)) <= 0)
6862 		return (rv);
6863 
6864 	if ((rv = pkp->dtpk_fmatch(prp->dtpr_func, pkp->dtpk_func, 0)) <= 0)
6865 		return (rv);
6866 
6867 	if ((rv = pkp->dtpk_nmatch(prp->dtpr_name, pkp->dtpk_name, 0)) <= 0)
6868 		return (rv);
6869 
6870 	if (dtrace_match_priv(prp, priv, uid, zoneid) == 0)
6871 		return (0);
6872 
6873 	return (rv);
6874 }
6875 
6876 /*
6877  * dtrace_match_glob() is a safe kernel implementation of the gmatch(3GEN)
6878  * interface for matching a glob pattern 'p' to an input string 's'.  Unlike
6879  * libc's version, the kernel version only applies to 8-bit ASCII strings.
6880  * In addition, all of the recursion cases except for '*' matching have been
6881  * unwound.  For '*', we still implement recursive evaluation, but a depth
6882  * counter is maintained and matching is aborted if we recurse too deep.
6883  * The function returns 0 if no match, >0 if match, and <0 if recursion error.
6884  */
6885 static int
6886 dtrace_match_glob(const char *s, const char *p, int depth)
6887 {
6888 	const char *olds;
6889 	char s1, c;
6890 	int gs;
6891 
6892 	if (depth > DTRACE_PROBEKEY_MAXDEPTH)
6893 		return (-1);
6894 
6895 	if (s == NULL)
6896 		s = ""; /* treat NULL as empty string */
6897 
6898 top:
6899 	olds = s;
6900 	s1 = *s++;
6901 
6902 	if (p == NULL)
6903 		return (0);
6904 
6905 	if ((c = *p++) == '\0')
6906 		return (s1 == '\0');
6907 
6908 	switch (c) {
6909 	case '[': {
6910 		int ok = 0, notflag = 0;
6911 		char lc = '\0';
6912 
6913 		if (s1 == '\0')
6914 			return (0);
6915 
6916 		if (*p == '!') {
6917 			notflag = 1;
6918 			p++;
6919 		}
6920 
6921 		if ((c = *p++) == '\0')
6922 			return (0);
6923 
6924 		do {
6925 			if (c == '-' && lc != '\0' && *p != ']') {
6926 				if ((c = *p++) == '\0')
6927 					return (0);
6928 				if (c == '\\' && (c = *p++) == '\0')
6929 					return (0);
6930 
6931 				if (notflag) {
6932 					if (s1 < lc || s1 > c)
6933 						ok++;
6934 					else
6935 						return (0);
6936 				} else if (lc <= s1 && s1 <= c)
6937 					ok++;
6938 
6939 			} else if (c == '\\' && (c = *p++) == '\0')
6940 				return (0);
6941 
6942 			lc = c; /* save left-hand 'c' for next iteration */
6943 
6944 			if (notflag) {
6945 				if (s1 != c)
6946 					ok++;
6947 				else
6948 					return (0);
6949 			} else if (s1 == c)
6950 				ok++;
6951 
6952 			if ((c = *p++) == '\0')
6953 				return (0);
6954 
6955 		} while (c != ']');
6956 
6957 		if (ok)
6958 			goto top;
6959 
6960 		return (0);
6961 	}
6962 
6963 	case '\\':
6964 		if ((c = *p++) == '\0')
6965 			return (0);
6966 		/*FALLTHRU*/
6967 
6968 	default:
6969 		if (c != s1)
6970 			return (0);
6971 		/*FALLTHRU*/
6972 
6973 	case '?':
6974 		if (s1 != '\0')
6975 			goto top;
6976 		return (0);
6977 
6978 	case '*':
6979 		while (*p == '*')
6980 			p++; /* consecutive *'s are identical to a single one */
6981 
6982 		if (*p == '\0')
6983 			return (1);
6984 
6985 		for (s = olds; *s != '\0'; s++) {
6986 			if ((gs = dtrace_match_glob(s, p, depth + 1)) != 0)
6987 				return (gs);
6988 		}
6989 
6990 		return (0);
6991 	}
6992 }
6993 
6994 /*ARGSUSED*/
6995 static int
6996 dtrace_match_string(const char *s, const char *p, int depth)
6997 {
6998 	return (s != NULL && strcmp(s, p) == 0);
6999 }
7000 
7001 /*ARGSUSED*/
7002 static int
7003 dtrace_match_nul(const char *s, const char *p, int depth)
7004 {
7005 	return (1); /* always match the empty pattern */
7006 }
7007 
7008 /*ARGSUSED*/
7009 static int
7010 dtrace_match_nonzero(const char *s, const char *p, int depth)
7011 {
7012 	return (s != NULL && s[0] != '\0');
7013 }
7014 
7015 static int
7016 dtrace_match(const dtrace_probekey_t *pkp, uint32_t priv, uid_t uid,
7017     zoneid_t zoneid, int (*matched)(dtrace_probe_t *, void *), void *arg)
7018 {
7019 	dtrace_probe_t template, *probe;
7020 	dtrace_hash_t *hash = NULL;
7021 	int len, rc, best = INT_MAX, nmatched = 0;
7022 	dtrace_id_t i;
7023 
7024 	ASSERT(MUTEX_HELD(&dtrace_lock));
7025 
7026 	/*
7027 	 * If the probe ID is specified in the key, just lookup by ID and
7028 	 * invoke the match callback once if a matching probe is found.
7029 	 */
7030 	if (pkp->dtpk_id != DTRACE_IDNONE) {
7031 		if ((probe = dtrace_probe_lookup_id(pkp->dtpk_id)) != NULL &&
7032 		    dtrace_match_probe(probe, pkp, priv, uid, zoneid) > 0) {
7033 			if ((*matched)(probe, arg) == DTRACE_MATCH_FAIL)
7034 				return (DTRACE_MATCH_FAIL);
7035 			nmatched++;
7036 		}
7037 		return (nmatched);
7038 	}
7039 
7040 	template.dtpr_mod = (char *)pkp->dtpk_mod;
7041 	template.dtpr_func = (char *)pkp->dtpk_func;
7042 	template.dtpr_name = (char *)pkp->dtpk_name;
7043 
7044 	/*
7045 	 * We want to find the most distinct of the module name, function
7046 	 * name, and name.  So for each one that is not a glob pattern or
7047 	 * empty string, we perform a lookup in the corresponding hash and
7048 	 * use the hash table with the fewest collisions to do our search.
7049 	 */
7050 	if (pkp->dtpk_mmatch == &dtrace_match_string &&
7051 	    (len = dtrace_hash_collisions(dtrace_bymod, &template)) < best) {
7052 		best = len;
7053 		hash = dtrace_bymod;
7054 	}
7055 
7056 	if (pkp->dtpk_fmatch == &dtrace_match_string &&
7057 	    (len = dtrace_hash_collisions(dtrace_byfunc, &template)) < best) {
7058 		best = len;
7059 		hash = dtrace_byfunc;
7060 	}
7061 
7062 	if (pkp->dtpk_nmatch == &dtrace_match_string &&
7063 	    (len = dtrace_hash_collisions(dtrace_byname, &template)) < best) {
7064 		best = len;
7065 		hash = dtrace_byname;
7066 	}
7067 
7068 	/*
7069 	 * If we did not select a hash table, iterate over every probe and
7070 	 * invoke our callback for each one that matches our input probe key.
7071 	 */
7072 	if (hash == NULL) {
7073 		for (i = 0; i < dtrace_nprobes; i++) {
7074 			if ((probe = dtrace_probes[i]) == NULL ||
7075 			    dtrace_match_probe(probe, pkp, priv, uid,
7076 			    zoneid) <= 0)
7077 				continue;
7078 
7079 			nmatched++;
7080 
7081 			if ((rc = (*matched)(probe, arg)) !=
7082 			    DTRACE_MATCH_NEXT) {
7083 				if (rc == DTRACE_MATCH_FAIL)
7084 					return (DTRACE_MATCH_FAIL);
7085 				break;
7086 			}
7087 		}
7088 
7089 		return (nmatched);
7090 	}
7091 
7092 	/*
7093 	 * If we selected a hash table, iterate over each probe of the same key
7094 	 * name and invoke the callback for every probe that matches the other
7095 	 * attributes of our input probe key.
7096 	 */
7097 	for (probe = dtrace_hash_lookup(hash, &template); probe != NULL;
7098 	    probe = *(DTRACE_HASHNEXT(hash, probe))) {
7099 
7100 		if (dtrace_match_probe(probe, pkp, priv, uid, zoneid) <= 0)
7101 			continue;
7102 
7103 		nmatched++;
7104 
7105 		if ((rc = (*matched)(probe, arg)) != DTRACE_MATCH_NEXT) {
7106 			if (rc == DTRACE_MATCH_FAIL)
7107 				return (DTRACE_MATCH_FAIL);
7108 			break;
7109 		}
7110 	}
7111 
7112 	return (nmatched);
7113 }
7114 
7115 /*
7116  * Return the function pointer dtrace_probecmp() should use to compare the
7117  * specified pattern with a string.  For NULL or empty patterns, we select
7118  * dtrace_match_nul().  For glob pattern strings, we use dtrace_match_glob().
7119  * For non-empty non-glob strings, we use dtrace_match_string().
7120  */
7121 static dtrace_probekey_f *
7122 dtrace_probekey_func(const char *p)
7123 {
7124 	char c;
7125 
7126 	if (p == NULL || *p == '\0')
7127 		return (&dtrace_match_nul);
7128 
7129 	while ((c = *p++) != '\0') {
7130 		if (c == '[' || c == '?' || c == '*' || c == '\\')
7131 			return (&dtrace_match_glob);
7132 	}
7133 
7134 	return (&dtrace_match_string);
7135 }
7136 
7137 /*
7138  * Build a probe comparison key for use with dtrace_match_probe() from the
7139  * given probe description.  By convention, a null key only matches anchored
7140  * probes: if each field is the empty string, reset dtpk_fmatch to
7141  * dtrace_match_nonzero().
7142  */
7143 static void
7144 dtrace_probekey(const dtrace_probedesc_t *pdp, dtrace_probekey_t *pkp)
7145 {
7146 	pkp->dtpk_prov = pdp->dtpd_provider;
7147 	pkp->dtpk_pmatch = dtrace_probekey_func(pdp->dtpd_provider);
7148 
7149 	pkp->dtpk_mod = pdp->dtpd_mod;
7150 	pkp->dtpk_mmatch = dtrace_probekey_func(pdp->dtpd_mod);
7151 
7152 	pkp->dtpk_func = pdp->dtpd_func;
7153 	pkp->dtpk_fmatch = dtrace_probekey_func(pdp->dtpd_func);
7154 
7155 	pkp->dtpk_name = pdp->dtpd_name;
7156 	pkp->dtpk_nmatch = dtrace_probekey_func(pdp->dtpd_name);
7157 
7158 	pkp->dtpk_id = pdp->dtpd_id;
7159 
7160 	if (pkp->dtpk_id == DTRACE_IDNONE &&
7161 	    pkp->dtpk_pmatch == &dtrace_match_nul &&
7162 	    pkp->dtpk_mmatch == &dtrace_match_nul &&
7163 	    pkp->dtpk_fmatch == &dtrace_match_nul &&
7164 	    pkp->dtpk_nmatch == &dtrace_match_nul)
7165 		pkp->dtpk_fmatch = &dtrace_match_nonzero;
7166 }
7167 
7168 /*
7169  * DTrace Provider-to-Framework API Functions
7170  *
7171  * These functions implement much of the Provider-to-Framework API, as
7172  * described in <sys/dtrace.h>.  The parts of the API not in this section are
7173  * the functions in the API for probe management (found below), and
7174  * dtrace_probe() itself (found above).
7175  */
7176 
7177 /*
7178  * Register the calling provider with the DTrace framework.  This should
7179  * generally be called by DTrace providers in their attach(9E) entry point.
7180  */
7181 int
7182 dtrace_register(const char *name, const dtrace_pattr_t *pap, uint32_t priv,
7183     cred_t *cr, const dtrace_pops_t *pops, void *arg, dtrace_provider_id_t *idp)
7184 {
7185 	dtrace_provider_t *provider;
7186 
7187 	if (name == NULL || pap == NULL || pops == NULL || idp == NULL) {
7188 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7189 		    "arguments", name ? name : "<NULL>");
7190 		return (EINVAL);
7191 	}
7192 
7193 	if (name[0] == '\0' || dtrace_badname(name)) {
7194 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7195 		    "provider name", name);
7196 		return (EINVAL);
7197 	}
7198 
7199 	if ((pops->dtps_provide == NULL && pops->dtps_provide_module == NULL) ||
7200 	    pops->dtps_enable == NULL || pops->dtps_disable == NULL ||
7201 	    pops->dtps_destroy == NULL ||
7202 	    ((pops->dtps_resume == NULL) != (pops->dtps_suspend == NULL))) {
7203 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7204 		    "provider ops", name);
7205 		return (EINVAL);
7206 	}
7207 
7208 	if (dtrace_badattr(&pap->dtpa_provider) ||
7209 	    dtrace_badattr(&pap->dtpa_mod) ||
7210 	    dtrace_badattr(&pap->dtpa_func) ||
7211 	    dtrace_badattr(&pap->dtpa_name) ||
7212 	    dtrace_badattr(&pap->dtpa_args)) {
7213 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7214 		    "provider attributes", name);
7215 		return (EINVAL);
7216 	}
7217 
7218 	if (priv & ~DTRACE_PRIV_ALL) {
7219 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7220 		    "privilege attributes", name);
7221 		return (EINVAL);
7222 	}
7223 
7224 	if ((priv & DTRACE_PRIV_KERNEL) &&
7225 	    (priv & (DTRACE_PRIV_USER | DTRACE_PRIV_OWNER)) &&
7226 	    pops->dtps_mode == NULL) {
7227 		cmn_err(CE_WARN, "failed to register provider '%s': need "
7228 		    "dtps_mode() op for given privilege attributes", name);
7229 		return (EINVAL);
7230 	}
7231 
7232 	provider = kmem_zalloc(sizeof (dtrace_provider_t), KM_SLEEP);
7233 	provider->dtpv_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
7234 	(void) strcpy(provider->dtpv_name, name);
7235 
7236 	provider->dtpv_attr = *pap;
7237 	provider->dtpv_priv.dtpp_flags = priv;
7238 	if (cr != NULL) {
7239 		provider->dtpv_priv.dtpp_uid = crgetuid(cr);
7240 		provider->dtpv_priv.dtpp_zoneid = crgetzoneid(cr);
7241 	}
7242 	provider->dtpv_pops = *pops;
7243 
7244 	if (pops->dtps_provide == NULL) {
7245 		ASSERT(pops->dtps_provide_module != NULL);
7246 		provider->dtpv_pops.dtps_provide =
7247 		    (void (*)(void *, const dtrace_probedesc_t *))dtrace_nullop;
7248 	}
7249 
7250 	if (pops->dtps_provide_module == NULL) {
7251 		ASSERT(pops->dtps_provide != NULL);
7252 		provider->dtpv_pops.dtps_provide_module =
7253 		    (void (*)(void *, struct modctl *))dtrace_nullop;
7254 	}
7255 
7256 	if (pops->dtps_suspend == NULL) {
7257 		ASSERT(pops->dtps_resume == NULL);
7258 		provider->dtpv_pops.dtps_suspend =
7259 		    (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
7260 		provider->dtpv_pops.dtps_resume =
7261 		    (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
7262 	}
7263 
7264 	provider->dtpv_arg = arg;
7265 	*idp = (dtrace_provider_id_t)provider;
7266 
7267 	if (pops == &dtrace_provider_ops) {
7268 		ASSERT(MUTEX_HELD(&dtrace_provider_lock));
7269 		ASSERT(MUTEX_HELD(&dtrace_lock));
7270 		ASSERT(dtrace_anon.dta_enabling == NULL);
7271 
7272 		/*
7273 		 * We make sure that the DTrace provider is at the head of
7274 		 * the provider chain.
7275 		 */
7276 		provider->dtpv_next = dtrace_provider;
7277 		dtrace_provider = provider;
7278 		return (0);
7279 	}
7280 
7281 	mutex_enter(&dtrace_provider_lock);
7282 	mutex_enter(&dtrace_lock);
7283 
7284 	/*
7285 	 * If there is at least one provider registered, we'll add this
7286 	 * provider after the first provider.
7287 	 */
7288 	if (dtrace_provider != NULL) {
7289 		provider->dtpv_next = dtrace_provider->dtpv_next;
7290 		dtrace_provider->dtpv_next = provider;
7291 	} else {
7292 		dtrace_provider = provider;
7293 	}
7294 
7295 	if (dtrace_retained != NULL) {
7296 		dtrace_enabling_provide(provider);
7297 
7298 		/*
7299 		 * Now we need to call dtrace_enabling_matchall() -- which
7300 		 * will acquire cpu_lock and dtrace_lock.  We therefore need
7301 		 * to drop all of our locks before calling into it...
7302 		 */
7303 		mutex_exit(&dtrace_lock);
7304 		mutex_exit(&dtrace_provider_lock);
7305 		dtrace_enabling_matchall();
7306 
7307 		return (0);
7308 	}
7309 
7310 	mutex_exit(&dtrace_lock);
7311 	mutex_exit(&dtrace_provider_lock);
7312 
7313 	return (0);
7314 }
7315 
7316 /*
7317  * Unregister the specified provider from the DTrace framework.  This should
7318  * generally be called by DTrace providers in their detach(9E) entry point.
7319  */
7320 int
7321 dtrace_unregister(dtrace_provider_id_t id)
7322 {
7323 	dtrace_provider_t *old = (dtrace_provider_t *)id;
7324 	dtrace_provider_t *prev = NULL;
7325 	int i, self = 0, noreap = 0;
7326 	dtrace_probe_t *probe, *first = NULL;
7327 
7328 	if (old->dtpv_pops.dtps_enable ==
7329 	    (int (*)(void *, dtrace_id_t, void *))dtrace_enable_nullop) {
7330 		/*
7331 		 * If DTrace itself is the provider, we're called with locks
7332 		 * already held.
7333 		 */
7334 		ASSERT(old == dtrace_provider);
7335 		ASSERT(dtrace_devi != NULL);
7336 		ASSERT(MUTEX_HELD(&dtrace_provider_lock));
7337 		ASSERT(MUTEX_HELD(&dtrace_lock));
7338 		self = 1;
7339 
7340 		if (dtrace_provider->dtpv_next != NULL) {
7341 			/*
7342 			 * There's another provider here; return failure.
7343 			 */
7344 			return (EBUSY);
7345 		}
7346 	} else {
7347 		mutex_enter(&dtrace_provider_lock);
7348 		mutex_enter(&mod_lock);
7349 		mutex_enter(&dtrace_lock);
7350 	}
7351 
7352 	/*
7353 	 * If anyone has /dev/dtrace open, or if there are anonymous enabled
7354 	 * probes, we refuse to let providers slither away, unless this
7355 	 * provider has already been explicitly invalidated.
7356 	 */
7357 	if (!old->dtpv_defunct &&
7358 	    (dtrace_opens || (dtrace_anon.dta_state != NULL &&
7359 	    dtrace_anon.dta_state->dts_necbs > 0))) {
7360 		if (!self) {
7361 			mutex_exit(&dtrace_lock);
7362 			mutex_exit(&mod_lock);
7363 			mutex_exit(&dtrace_provider_lock);
7364 		}
7365 		return (EBUSY);
7366 	}
7367 
7368 	/*
7369 	 * Attempt to destroy the probes associated with this provider.
7370 	 */
7371 	for (i = 0; i < dtrace_nprobes; i++) {
7372 		if ((probe = dtrace_probes[i]) == NULL)
7373 			continue;
7374 
7375 		if (probe->dtpr_provider != old)
7376 			continue;
7377 
7378 		if (probe->dtpr_ecb == NULL)
7379 			continue;
7380 
7381 		/*
7382 		 * If we are trying to unregister a defunct provider, and the
7383 		 * provider was made defunct within the interval dictated by
7384 		 * dtrace_unregister_defunct_reap, we'll (asynchronously)
7385 		 * attempt to reap our enablings.  To denote that the provider
7386 		 * should reattempt to unregister itself at some point in the
7387 		 * future, we will return a differentiable error code (EAGAIN
7388 		 * instead of EBUSY) in this case.
7389 		 */
7390 		if (dtrace_gethrtime() - old->dtpv_defunct >
7391 		    dtrace_unregister_defunct_reap)
7392 			noreap = 1;
7393 
7394 		if (!self) {
7395 			mutex_exit(&dtrace_lock);
7396 			mutex_exit(&mod_lock);
7397 			mutex_exit(&dtrace_provider_lock);
7398 		}
7399 
7400 		if (noreap)
7401 			return (EBUSY);
7402 
7403 		(void) taskq_dispatch(dtrace_taskq,
7404 		    (task_func_t *)dtrace_enabling_reap, NULL, TQ_SLEEP);
7405 
7406 		return (EAGAIN);
7407 	}
7408 
7409 	/*
7410 	 * All of the probes for this provider are disabled; we can safely
7411 	 * remove all of them from their hash chains and from the probe array.
7412 	 */
7413 	for (i = 0; i < dtrace_nprobes; i++) {
7414 		if ((probe = dtrace_probes[i]) == NULL)
7415 			continue;
7416 
7417 		if (probe->dtpr_provider != old)
7418 			continue;
7419 
7420 		dtrace_probes[i] = NULL;
7421 
7422 		dtrace_hash_remove(dtrace_bymod, probe);
7423 		dtrace_hash_remove(dtrace_byfunc, probe);
7424 		dtrace_hash_remove(dtrace_byname, probe);
7425 
7426 		if (first == NULL) {
7427 			first = probe;
7428 			probe->dtpr_nextmod = NULL;
7429 		} else {
7430 			probe->dtpr_nextmod = first;
7431 			first = probe;
7432 		}
7433 	}
7434 
7435 	/*
7436 	 * The provider's probes have been removed from the hash chains and
7437 	 * from the probe array.  Now issue a dtrace_sync() to be sure that
7438 	 * everyone has cleared out from any probe array processing.
7439 	 */
7440 	dtrace_sync();
7441 
7442 	for (probe = first; probe != NULL; probe = first) {
7443 		first = probe->dtpr_nextmod;
7444 
7445 		old->dtpv_pops.dtps_destroy(old->dtpv_arg, probe->dtpr_id,
7446 		    probe->dtpr_arg);
7447 		kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
7448 		kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
7449 		kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
7450 		vmem_free(dtrace_arena, (void *)(uintptr_t)(probe->dtpr_id), 1);
7451 		kmem_free(probe, sizeof (dtrace_probe_t));
7452 	}
7453 
7454 	if ((prev = dtrace_provider) == old) {
7455 		ASSERT(self || dtrace_devi == NULL);
7456 		ASSERT(old->dtpv_next == NULL || dtrace_devi == NULL);
7457 		dtrace_provider = old->dtpv_next;
7458 	} else {
7459 		while (prev != NULL && prev->dtpv_next != old)
7460 			prev = prev->dtpv_next;
7461 
7462 		if (prev == NULL) {
7463 			panic("attempt to unregister non-existent "
7464 			    "dtrace provider %p\n", (void *)id);
7465 		}
7466 
7467 		prev->dtpv_next = old->dtpv_next;
7468 	}
7469 
7470 	if (!self) {
7471 		mutex_exit(&dtrace_lock);
7472 		mutex_exit(&mod_lock);
7473 		mutex_exit(&dtrace_provider_lock);
7474 	}
7475 
7476 	kmem_free(old->dtpv_name, strlen(old->dtpv_name) + 1);
7477 	kmem_free(old, sizeof (dtrace_provider_t));
7478 
7479 	return (0);
7480 }
7481 
7482 /*
7483  * Invalidate the specified provider.  All subsequent probe lookups for the
7484  * specified provider will fail, but its probes will not be removed.
7485  */
7486 void
7487 dtrace_invalidate(dtrace_provider_id_t id)
7488 {
7489 	dtrace_provider_t *pvp = (dtrace_provider_t *)id;
7490 
7491 	ASSERT(pvp->dtpv_pops.dtps_enable !=
7492 	    (int (*)(void *, dtrace_id_t, void *))dtrace_enable_nullop);
7493 
7494 	mutex_enter(&dtrace_provider_lock);
7495 	mutex_enter(&dtrace_lock);
7496 
7497 	pvp->dtpv_defunct = dtrace_gethrtime();
7498 
7499 	mutex_exit(&dtrace_lock);
7500 	mutex_exit(&dtrace_provider_lock);
7501 }
7502 
7503 /*
7504  * Indicate whether or not DTrace has attached.
7505  */
7506 int
7507 dtrace_attached(void)
7508 {
7509 	/*
7510 	 * dtrace_provider will be non-NULL iff the DTrace driver has
7511 	 * attached.  (It's non-NULL because DTrace is always itself a
7512 	 * provider.)
7513 	 */
7514 	return (dtrace_provider != NULL);
7515 }
7516 
7517 /*
7518  * Remove all the unenabled probes for the given provider.  This function is
7519  * not unlike dtrace_unregister(), except that it doesn't remove the provider
7520  * -- just as many of its associated probes as it can.
7521  */
7522 int
7523 dtrace_condense(dtrace_provider_id_t id)
7524 {
7525 	dtrace_provider_t *prov = (dtrace_provider_t *)id;
7526 	int i;
7527 	dtrace_probe_t *probe;
7528 
7529 	/*
7530 	 * Make sure this isn't the dtrace provider itself.
7531 	 */
7532 	ASSERT(prov->dtpv_pops.dtps_enable !=
7533 	    (int (*)(void *, dtrace_id_t, void *))dtrace_enable_nullop);
7534 
7535 	mutex_enter(&dtrace_provider_lock);
7536 	mutex_enter(&dtrace_lock);
7537 
7538 	/*
7539 	 * Attempt to destroy the probes associated with this provider.
7540 	 */
7541 	for (i = 0; i < dtrace_nprobes; i++) {
7542 		if ((probe = dtrace_probes[i]) == NULL)
7543 			continue;
7544 
7545 		if (probe->dtpr_provider != prov)
7546 			continue;
7547 
7548 		if (probe->dtpr_ecb != NULL)
7549 			continue;
7550 
7551 		dtrace_probes[i] = NULL;
7552 
7553 		dtrace_hash_remove(dtrace_bymod, probe);
7554 		dtrace_hash_remove(dtrace_byfunc, probe);
7555 		dtrace_hash_remove(dtrace_byname, probe);
7556 
7557 		prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, i + 1,
7558 		    probe->dtpr_arg);
7559 		kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
7560 		kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
7561 		kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
7562 		kmem_free(probe, sizeof (dtrace_probe_t));
7563 		vmem_free(dtrace_arena, (void *)((uintptr_t)i + 1), 1);
7564 	}
7565 
7566 	mutex_exit(&dtrace_lock);
7567 	mutex_exit(&dtrace_provider_lock);
7568 
7569 	return (0);
7570 }
7571 
7572 /*
7573  * DTrace Probe Management Functions
7574  *
7575  * The functions in this section perform the DTrace probe management,
7576  * including functions to create probes, look-up probes, and call into the
7577  * providers to request that probes be provided.  Some of these functions are
7578  * in the Provider-to-Framework API; these functions can be identified by the
7579  * fact that they are not declared "static".
7580  */
7581 
7582 /*
7583  * Create a probe with the specified module name, function name, and name.
7584  */
7585 dtrace_id_t
7586 dtrace_probe_create(dtrace_provider_id_t prov, const char *mod,
7587     const char *func, const char *name, int aframes, void *arg)
7588 {
7589 	dtrace_probe_t *probe, **probes;
7590 	dtrace_provider_t *provider = (dtrace_provider_t *)prov;
7591 	dtrace_id_t id;
7592 
7593 	if (provider == dtrace_provider) {
7594 		ASSERT(MUTEX_HELD(&dtrace_lock));
7595 	} else {
7596 		mutex_enter(&dtrace_lock);
7597 	}
7598 
7599 	id = (dtrace_id_t)(uintptr_t)vmem_alloc(dtrace_arena, 1,
7600 	    VM_BESTFIT | VM_SLEEP);
7601 	probe = kmem_zalloc(sizeof (dtrace_probe_t), KM_SLEEP);
7602 
7603 	probe->dtpr_id = id;
7604 	probe->dtpr_gen = dtrace_probegen++;
7605 	probe->dtpr_mod = dtrace_strdup(mod);
7606 	probe->dtpr_func = dtrace_strdup(func);
7607 	probe->dtpr_name = dtrace_strdup(name);
7608 	probe->dtpr_arg = arg;
7609 	probe->dtpr_aframes = aframes;
7610 	probe->dtpr_provider = provider;
7611 
7612 	dtrace_hash_add(dtrace_bymod, probe);
7613 	dtrace_hash_add(dtrace_byfunc, probe);
7614 	dtrace_hash_add(dtrace_byname, probe);
7615 
7616 	if (id - 1 >= dtrace_nprobes) {
7617 		size_t osize = dtrace_nprobes * sizeof (dtrace_probe_t *);
7618 		size_t nsize = osize << 1;
7619 
7620 		if (nsize == 0) {
7621 			ASSERT(osize == 0);
7622 			ASSERT(dtrace_probes == NULL);
7623 			nsize = sizeof (dtrace_probe_t *);
7624 		}
7625 
7626 		probes = kmem_zalloc(nsize, KM_SLEEP);
7627 
7628 		if (dtrace_probes == NULL) {
7629 			ASSERT(osize == 0);
7630 			dtrace_probes = probes;
7631 			dtrace_nprobes = 1;
7632 		} else {
7633 			dtrace_probe_t **oprobes = dtrace_probes;
7634 
7635 			bcopy(oprobes, probes, osize);
7636 			dtrace_membar_producer();
7637 			dtrace_probes = probes;
7638 
7639 			dtrace_sync();
7640 
7641 			/*
7642 			 * All CPUs are now seeing the new probes array; we can
7643 			 * safely free the old array.
7644 			 */
7645 			kmem_free(oprobes, osize);
7646 			dtrace_nprobes <<= 1;
7647 		}
7648 
7649 		ASSERT(id - 1 < dtrace_nprobes);
7650 	}
7651 
7652 	ASSERT(dtrace_probes[id - 1] == NULL);
7653 	dtrace_probes[id - 1] = probe;
7654 
7655 	if (provider != dtrace_provider)
7656 		mutex_exit(&dtrace_lock);
7657 
7658 	return (id);
7659 }
7660 
7661 static dtrace_probe_t *
7662 dtrace_probe_lookup_id(dtrace_id_t id)
7663 {
7664 	ASSERT(MUTEX_HELD(&dtrace_lock));
7665 
7666 	if (id == 0 || id > dtrace_nprobes)
7667 		return (NULL);
7668 
7669 	return (dtrace_probes[id - 1]);
7670 }
7671 
7672 static int
7673 dtrace_probe_lookup_match(dtrace_probe_t *probe, void *arg)
7674 {
7675 	*((dtrace_id_t *)arg) = probe->dtpr_id;
7676 
7677 	return (DTRACE_MATCH_DONE);
7678 }
7679 
7680 /*
7681  * Look up a probe based on provider and one or more of module name, function
7682  * name and probe name.
7683  */
7684 dtrace_id_t
7685 dtrace_probe_lookup(dtrace_provider_id_t prid, const char *mod,
7686     const char *func, const char *name)
7687 {
7688 	dtrace_probekey_t pkey;
7689 	dtrace_id_t id;
7690 	int match;
7691 
7692 	pkey.dtpk_prov = ((dtrace_provider_t *)prid)->dtpv_name;
7693 	pkey.dtpk_pmatch = &dtrace_match_string;
7694 	pkey.dtpk_mod = mod;
7695 	pkey.dtpk_mmatch = mod ? &dtrace_match_string : &dtrace_match_nul;
7696 	pkey.dtpk_func = func;
7697 	pkey.dtpk_fmatch = func ? &dtrace_match_string : &dtrace_match_nul;
7698 	pkey.dtpk_name = name;
7699 	pkey.dtpk_nmatch = name ? &dtrace_match_string : &dtrace_match_nul;
7700 	pkey.dtpk_id = DTRACE_IDNONE;
7701 
7702 	mutex_enter(&dtrace_lock);
7703 	match = dtrace_match(&pkey, DTRACE_PRIV_ALL, 0, 0,
7704 	    dtrace_probe_lookup_match, &id);
7705 	mutex_exit(&dtrace_lock);
7706 
7707 	ASSERT(match == 1 || match == 0);
7708 	return (match ? id : 0);
7709 }
7710 
7711 /*
7712  * Returns the probe argument associated with the specified probe.
7713  */
7714 void *
7715 dtrace_probe_arg(dtrace_provider_id_t id, dtrace_id_t pid)
7716 {
7717 	dtrace_probe_t *probe;
7718 	void *rval = NULL;
7719 
7720 	mutex_enter(&dtrace_lock);
7721 
7722 	if ((probe = dtrace_probe_lookup_id(pid)) != NULL &&
7723 	    probe->dtpr_provider == (dtrace_provider_t *)id)
7724 		rval = probe->dtpr_arg;
7725 
7726 	mutex_exit(&dtrace_lock);
7727 
7728 	return (rval);
7729 }
7730 
7731 /*
7732  * Copy a probe into a probe description.
7733  */
7734 static void
7735 dtrace_probe_description(const dtrace_probe_t *prp, dtrace_probedesc_t *pdp)
7736 {
7737 	bzero(pdp, sizeof (dtrace_probedesc_t));
7738 	pdp->dtpd_id = prp->dtpr_id;
7739 
7740 	(void) strncpy(pdp->dtpd_provider,
7741 	    prp->dtpr_provider->dtpv_name, DTRACE_PROVNAMELEN - 1);
7742 
7743 	(void) strncpy(pdp->dtpd_mod, prp->dtpr_mod, DTRACE_MODNAMELEN - 1);
7744 	(void) strncpy(pdp->dtpd_func, prp->dtpr_func, DTRACE_FUNCNAMELEN - 1);
7745 	(void) strncpy(pdp->dtpd_name, prp->dtpr_name, DTRACE_NAMELEN - 1);
7746 }
7747 
7748 /*
7749  * Called to indicate that a probe -- or probes -- should be provided by a
7750  * specfied provider.  If the specified description is NULL, the provider will
7751  * be told to provide all of its probes.  (This is done whenever a new
7752  * consumer comes along, or whenever a retained enabling is to be matched.) If
7753  * the specified description is non-NULL, the provider is given the
7754  * opportunity to dynamically provide the specified probe, allowing providers
7755  * to support the creation of probes on-the-fly.  (So-called _autocreated_
7756  * probes.)  If the provider is NULL, the operations will be applied to all
7757  * providers; if the provider is non-NULL the operations will only be applied
7758  * to the specified provider.  The dtrace_provider_lock must be held, and the
7759  * dtrace_lock must _not_ be held -- the provider's dtps_provide() operation
7760  * will need to grab the dtrace_lock when it reenters the framework through
7761  * dtrace_probe_lookup(), dtrace_probe_create(), etc.
7762  */
7763 static void
7764 dtrace_probe_provide(dtrace_probedesc_t *desc, dtrace_provider_t *prv)
7765 {
7766 	struct modctl *ctl;
7767 	int all = 0;
7768 
7769 	ASSERT(MUTEX_HELD(&dtrace_provider_lock));
7770 
7771 	if (prv == NULL) {
7772 		all = 1;
7773 		prv = dtrace_provider;
7774 	}
7775 
7776 	do {
7777 		/*
7778 		 * First, call the blanket provide operation.
7779 		 */
7780 		prv->dtpv_pops.dtps_provide(prv->dtpv_arg, desc);
7781 
7782 		/*
7783 		 * Now call the per-module provide operation.  We will grab
7784 		 * mod_lock to prevent the list from being modified.  Note
7785 		 * that this also prevents the mod_busy bits from changing.
7786 		 * (mod_busy can only be changed with mod_lock held.)
7787 		 */
7788 		mutex_enter(&mod_lock);
7789 
7790 		ctl = &modules;
7791 		do {
7792 			if (ctl->mod_busy || ctl->mod_mp == NULL)
7793 				continue;
7794 
7795 			prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
7796 
7797 		} while ((ctl = ctl->mod_next) != &modules);
7798 
7799 		mutex_exit(&mod_lock);
7800 	} while (all && (prv = prv->dtpv_next) != NULL);
7801 }
7802 
7803 /*
7804  * Iterate over each probe, and call the Framework-to-Provider API function
7805  * denoted by offs.
7806  */
7807 static void
7808 dtrace_probe_foreach(uintptr_t offs)
7809 {
7810 	dtrace_provider_t *prov;
7811 	void (*func)(void *, dtrace_id_t, void *);
7812 	dtrace_probe_t *probe;
7813 	dtrace_icookie_t cookie;
7814 	int i;
7815 
7816 	/*
7817 	 * We disable interrupts to walk through the probe array.  This is
7818 	 * safe -- the dtrace_sync() in dtrace_unregister() assures that we
7819 	 * won't see stale data.
7820 	 */
7821 	cookie = dtrace_interrupt_disable();
7822 
7823 	for (i = 0; i < dtrace_nprobes; i++) {
7824 		if ((probe = dtrace_probes[i]) == NULL)
7825 			continue;
7826 
7827 		if (probe->dtpr_ecb == NULL) {
7828 			/*
7829 			 * This probe isn't enabled -- don't call the function.
7830 			 */
7831 			continue;
7832 		}
7833 
7834 		prov = probe->dtpr_provider;
7835 		func = *((void(**)(void *, dtrace_id_t, void *))
7836 		    ((uintptr_t)&prov->dtpv_pops + offs));
7837 
7838 		func(prov->dtpv_arg, i + 1, probe->dtpr_arg);
7839 	}
7840 
7841 	dtrace_interrupt_enable(cookie);
7842 }
7843 
7844 static int
7845 dtrace_probe_enable(const dtrace_probedesc_t *desc, dtrace_enabling_t *enab)
7846 {
7847 	dtrace_probekey_t pkey;
7848 	uint32_t priv;
7849 	uid_t uid;
7850 	zoneid_t zoneid;
7851 
7852 	ASSERT(MUTEX_HELD(&dtrace_lock));
7853 	dtrace_ecb_create_cache = NULL;
7854 
7855 	if (desc == NULL) {
7856 		/*
7857 		 * If we're passed a NULL description, we're being asked to
7858 		 * create an ECB with a NULL probe.
7859 		 */
7860 		(void) dtrace_ecb_create_enable(NULL, enab);
7861 		return (0);
7862 	}
7863 
7864 	dtrace_probekey(desc, &pkey);
7865 	dtrace_cred2priv(enab->dten_vstate->dtvs_state->dts_cred.dcr_cred,
7866 	    &priv, &uid, &zoneid);
7867 
7868 	return (dtrace_match(&pkey, priv, uid, zoneid, dtrace_ecb_create_enable,
7869 	    enab));
7870 }
7871 
7872 /*
7873  * DTrace Helper Provider Functions
7874  */
7875 static void
7876 dtrace_dofattr2attr(dtrace_attribute_t *attr, const dof_attr_t dofattr)
7877 {
7878 	attr->dtat_name = DOF_ATTR_NAME(dofattr);
7879 	attr->dtat_data = DOF_ATTR_DATA(dofattr);
7880 	attr->dtat_class = DOF_ATTR_CLASS(dofattr);
7881 }
7882 
7883 static void
7884 dtrace_dofprov2hprov(dtrace_helper_provdesc_t *hprov,
7885     const dof_provider_t *dofprov, char *strtab)
7886 {
7887 	hprov->dthpv_provname = strtab + dofprov->dofpv_name;
7888 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_provider,
7889 	    dofprov->dofpv_provattr);
7890 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_mod,
7891 	    dofprov->dofpv_modattr);
7892 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_func,
7893 	    dofprov->dofpv_funcattr);
7894 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_name,
7895 	    dofprov->dofpv_nameattr);
7896 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_args,
7897 	    dofprov->dofpv_argsattr);
7898 }
7899 
7900 static void
7901 dtrace_helper_provide_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
7902 {
7903 	uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
7904 	dof_hdr_t *dof = (dof_hdr_t *)daddr;
7905 	dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
7906 	dof_provider_t *provider;
7907 	dof_probe_t *probe;
7908 	uint32_t *off, *enoff;
7909 	uint8_t *arg;
7910 	char *strtab;
7911 	uint_t i, nprobes;
7912 	dtrace_helper_provdesc_t dhpv;
7913 	dtrace_helper_probedesc_t dhpb;
7914 	dtrace_meta_t *meta = dtrace_meta_pid;
7915 	dtrace_mops_t *mops = &meta->dtm_mops;
7916 	void *parg;
7917 
7918 	provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
7919 	str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
7920 	    provider->dofpv_strtab * dof->dofh_secsize);
7921 	prb_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
7922 	    provider->dofpv_probes * dof->dofh_secsize);
7923 	arg_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
7924 	    provider->dofpv_prargs * dof->dofh_secsize);
7925 	off_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
7926 	    provider->dofpv_proffs * dof->dofh_secsize);
7927 
7928 	strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
7929 	off = (uint32_t *)(uintptr_t)(daddr + off_sec->dofs_offset);
7930 	arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
7931 	enoff = NULL;
7932 
7933 	/*
7934 	 * See dtrace_helper_provider_validate().
7935 	 */
7936 	if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
7937 	    provider->dofpv_prenoffs != DOF_SECT_NONE) {
7938 		enoff_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
7939 		    provider->dofpv_prenoffs * dof->dofh_secsize);
7940 		enoff = (uint32_t *)(uintptr_t)(daddr + enoff_sec->dofs_offset);
7941 	}
7942 
7943 	nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
7944 
7945 	/*
7946 	 * Create the provider.
7947 	 */
7948 	dtrace_dofprov2hprov(&dhpv, provider, strtab);
7949 
7950 	if ((parg = mops->dtms_provide_pid(meta->dtm_arg, &dhpv, pid)) == NULL)
7951 		return;
7952 
7953 	meta->dtm_count++;
7954 
7955 	/*
7956 	 * Create the probes.
7957 	 */
7958 	for (i = 0; i < nprobes; i++) {
7959 		probe = (dof_probe_t *)(uintptr_t)(daddr +
7960 		    prb_sec->dofs_offset + i * prb_sec->dofs_entsize);
7961 
7962 		dhpb.dthpb_mod = dhp->dofhp_mod;
7963 		dhpb.dthpb_func = strtab + probe->dofpr_func;
7964 		dhpb.dthpb_name = strtab + probe->dofpr_name;
7965 		dhpb.dthpb_base = probe->dofpr_addr;
7966 		dhpb.dthpb_offs = off + probe->dofpr_offidx;
7967 		dhpb.dthpb_noffs = probe->dofpr_noffs;
7968 		if (enoff != NULL) {
7969 			dhpb.dthpb_enoffs = enoff + probe->dofpr_enoffidx;
7970 			dhpb.dthpb_nenoffs = probe->dofpr_nenoffs;
7971 		} else {
7972 			dhpb.dthpb_enoffs = NULL;
7973 			dhpb.dthpb_nenoffs = 0;
7974 		}
7975 		dhpb.dthpb_args = arg + probe->dofpr_argidx;
7976 		dhpb.dthpb_nargc = probe->dofpr_nargc;
7977 		dhpb.dthpb_xargc = probe->dofpr_xargc;
7978 		dhpb.dthpb_ntypes = strtab + probe->dofpr_nargv;
7979 		dhpb.dthpb_xtypes = strtab + probe->dofpr_xargv;
7980 
7981 		mops->dtms_create_probe(meta->dtm_arg, parg, &dhpb);
7982 	}
7983 }
7984 
7985 static void
7986 dtrace_helper_provide(dof_helper_t *dhp, pid_t pid)
7987 {
7988 	uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
7989 	dof_hdr_t *dof = (dof_hdr_t *)daddr;
7990 	int i;
7991 
7992 	ASSERT(MUTEX_HELD(&dtrace_meta_lock));
7993 
7994 	for (i = 0; i < dof->dofh_secnum; i++) {
7995 		dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
7996 		    dof->dofh_secoff + i * dof->dofh_secsize);
7997 
7998 		if (sec->dofs_type != DOF_SECT_PROVIDER)
7999 			continue;
8000 
8001 		dtrace_helper_provide_one(dhp, sec, pid);
8002 	}
8003 
8004 	/*
8005 	 * We may have just created probes, so we must now rematch against
8006 	 * any retained enablings.  Note that this call will acquire both
8007 	 * cpu_lock and dtrace_lock; the fact that we are holding
8008 	 * dtrace_meta_lock now is what defines the ordering with respect to
8009 	 * these three locks.
8010 	 */
8011 	dtrace_enabling_matchall();
8012 }
8013 
8014 static void
8015 dtrace_helper_provider_remove_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
8016 {
8017 	uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8018 	dof_hdr_t *dof = (dof_hdr_t *)daddr;
8019 	dof_sec_t *str_sec;
8020 	dof_provider_t *provider;
8021 	char *strtab;
8022 	dtrace_helper_provdesc_t dhpv;
8023 	dtrace_meta_t *meta = dtrace_meta_pid;
8024 	dtrace_mops_t *mops = &meta->dtm_mops;
8025 
8026 	provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
8027 	str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8028 	    provider->dofpv_strtab * dof->dofh_secsize);
8029 
8030 	strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
8031 
8032 	/*
8033 	 * Create the provider.
8034 	 */
8035 	dtrace_dofprov2hprov(&dhpv, provider, strtab);
8036 
8037 	mops->dtms_remove_pid(meta->dtm_arg, &dhpv, pid);
8038 
8039 	meta->dtm_count--;
8040 }
8041 
8042 static void
8043 dtrace_helper_provider_remove(dof_helper_t *dhp, pid_t pid)
8044 {
8045 	uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8046 	dof_hdr_t *dof = (dof_hdr_t *)daddr;
8047 	int i;
8048 
8049 	ASSERT(MUTEX_HELD(&dtrace_meta_lock));
8050 
8051 	for (i = 0; i < dof->dofh_secnum; i++) {
8052 		dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
8053 		    dof->dofh_secoff + i * dof->dofh_secsize);
8054 
8055 		if (sec->dofs_type != DOF_SECT_PROVIDER)
8056 			continue;
8057 
8058 		dtrace_helper_provider_remove_one(dhp, sec, pid);
8059 	}
8060 }
8061 
8062 /*
8063  * DTrace Meta Provider-to-Framework API Functions
8064  *
8065  * These functions implement the Meta Provider-to-Framework API, as described
8066  * in <sys/dtrace.h>.
8067  */
8068 int
8069 dtrace_meta_register(const char *name, const dtrace_mops_t *mops, void *arg,
8070     dtrace_meta_provider_id_t *idp)
8071 {
8072 	dtrace_meta_t *meta;
8073 	dtrace_helpers_t *help, *next;
8074 	int i;
8075 
8076 	*idp = DTRACE_METAPROVNONE;
8077 
8078 	/*
8079 	 * We strictly don't need the name, but we hold onto it for
8080 	 * debuggability. All hail error queues!
8081 	 */
8082 	if (name == NULL) {
8083 		cmn_err(CE_WARN, "failed to register meta-provider: "
8084 		    "invalid name");
8085 		return (EINVAL);
8086 	}
8087 
8088 	if (mops == NULL ||
8089 	    mops->dtms_create_probe == NULL ||
8090 	    mops->dtms_provide_pid == NULL ||
8091 	    mops->dtms_remove_pid == NULL) {
8092 		cmn_err(CE_WARN, "failed to register meta-register %s: "
8093 		    "invalid ops", name);
8094 		return (EINVAL);
8095 	}
8096 
8097 	meta = kmem_zalloc(sizeof (dtrace_meta_t), KM_SLEEP);
8098 	meta->dtm_mops = *mops;
8099 	meta->dtm_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
8100 	(void) strcpy(meta->dtm_name, name);
8101 	meta->dtm_arg = arg;
8102 
8103 	mutex_enter(&dtrace_meta_lock);
8104 	mutex_enter(&dtrace_lock);
8105 
8106 	if (dtrace_meta_pid != NULL) {
8107 		mutex_exit(&dtrace_lock);
8108 		mutex_exit(&dtrace_meta_lock);
8109 		cmn_err(CE_WARN, "failed to register meta-register %s: "
8110 		    "user-land meta-provider exists", name);
8111 		kmem_free(meta->dtm_name, strlen(meta->dtm_name) + 1);
8112 		kmem_free(meta, sizeof (dtrace_meta_t));
8113 		return (EINVAL);
8114 	}
8115 
8116 	dtrace_meta_pid = meta;
8117 	*idp = (dtrace_meta_provider_id_t)meta;
8118 
8119 	/*
8120 	 * If there are providers and probes ready to go, pass them
8121 	 * off to the new meta provider now.
8122 	 */
8123 
8124 	help = dtrace_deferred_pid;
8125 	dtrace_deferred_pid = NULL;
8126 
8127 	mutex_exit(&dtrace_lock);
8128 
8129 	while (help != NULL) {
8130 		for (i = 0; i < help->dthps_nprovs; i++) {
8131 			dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
8132 			    help->dthps_pid);
8133 		}
8134 
8135 		next = help->dthps_next;
8136 		help->dthps_next = NULL;
8137 		help->dthps_prev = NULL;
8138 		help->dthps_deferred = 0;
8139 		help = next;
8140 	}
8141 
8142 	mutex_exit(&dtrace_meta_lock);
8143 
8144 	return (0);
8145 }
8146 
8147 int
8148 dtrace_meta_unregister(dtrace_meta_provider_id_t id)
8149 {
8150 	dtrace_meta_t **pp, *old = (dtrace_meta_t *)id;
8151 
8152 	mutex_enter(&dtrace_meta_lock);
8153 	mutex_enter(&dtrace_lock);
8154 
8155 	if (old == dtrace_meta_pid) {
8156 		pp = &dtrace_meta_pid;
8157 	} else {
8158 		panic("attempt to unregister non-existent "
8159 		    "dtrace meta-provider %p\n", (void *)old);
8160 	}
8161 
8162 	if (old->dtm_count != 0) {
8163 		mutex_exit(&dtrace_lock);
8164 		mutex_exit(&dtrace_meta_lock);
8165 		return (EBUSY);
8166 	}
8167 
8168 	*pp = NULL;
8169 
8170 	mutex_exit(&dtrace_lock);
8171 	mutex_exit(&dtrace_meta_lock);
8172 
8173 	kmem_free(old->dtm_name, strlen(old->dtm_name) + 1);
8174 	kmem_free(old, sizeof (dtrace_meta_t));
8175 
8176 	return (0);
8177 }
8178 
8179 
8180 /*
8181  * DTrace DIF Object Functions
8182  */
8183 static int
8184 dtrace_difo_err(uint_t pc, const char *format, ...)
8185 {
8186 	if (dtrace_err_verbose) {
8187 		va_list alist;
8188 
8189 		(void) uprintf("dtrace DIF object error: [%u]: ", pc);
8190 		va_start(alist, format);
8191 		(void) vuprintf(format, alist);
8192 		va_end(alist);
8193 	}
8194 
8195 #ifdef DTRACE_ERRDEBUG
8196 	dtrace_errdebug(format);
8197 #endif
8198 	return (1);
8199 }
8200 
8201 /*
8202  * Validate a DTrace DIF object by checking the IR instructions.  The following
8203  * rules are currently enforced by dtrace_difo_validate():
8204  *
8205  * 1. Each instruction must have a valid opcode
8206  * 2. Each register, string, variable, or subroutine reference must be valid
8207  * 3. No instruction can modify register %r0 (must be zero)
8208  * 4. All instruction reserved bits must be set to zero
8209  * 5. The last instruction must be a "ret" instruction
8210  * 6. All branch targets must reference a valid instruction _after_ the branch
8211  */
8212 static int
8213 dtrace_difo_validate(dtrace_difo_t *dp, dtrace_vstate_t *vstate, uint_t nregs,
8214     cred_t *cr)
8215 {
8216 	int err = 0, i;
8217 	int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
8218 	int kcheckload;
8219 	uint_t pc;
8220 
8221 	kcheckload = cr == NULL ||
8222 	    (vstate->dtvs_state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) == 0;
8223 
8224 	dp->dtdo_destructive = 0;
8225 
8226 	for (pc = 0; pc < dp->dtdo_len && err == 0; pc++) {
8227 		dif_instr_t instr = dp->dtdo_buf[pc];
8228 
8229 		uint_t r1 = DIF_INSTR_R1(instr);
8230 		uint_t r2 = DIF_INSTR_R2(instr);
8231 		uint_t rd = DIF_INSTR_RD(instr);
8232 		uint_t rs = DIF_INSTR_RS(instr);
8233 		uint_t label = DIF_INSTR_LABEL(instr);
8234 		uint_t v = DIF_INSTR_VAR(instr);
8235 		uint_t subr = DIF_INSTR_SUBR(instr);
8236 		uint_t type = DIF_INSTR_TYPE(instr);
8237 		uint_t op = DIF_INSTR_OP(instr);
8238 
8239 		switch (op) {
8240 		case DIF_OP_OR:
8241 		case DIF_OP_XOR:
8242 		case DIF_OP_AND:
8243 		case DIF_OP_SLL:
8244 		case DIF_OP_SRL:
8245 		case DIF_OP_SRA:
8246 		case DIF_OP_SUB:
8247 		case DIF_OP_ADD:
8248 		case DIF_OP_MUL:
8249 		case DIF_OP_SDIV:
8250 		case DIF_OP_UDIV:
8251 		case DIF_OP_SREM:
8252 		case DIF_OP_UREM:
8253 		case DIF_OP_COPYS:
8254 			if (r1 >= nregs)
8255 				err += efunc(pc, "invalid register %u\n", r1);
8256 			if (r2 >= nregs)
8257 				err += efunc(pc, "invalid register %u\n", r2);
8258 			if (rd >= nregs)
8259 				err += efunc(pc, "invalid register %u\n", rd);
8260 			if (rd == 0)
8261 				err += efunc(pc, "cannot write to %r0\n");
8262 			break;
8263 		case DIF_OP_NOT:
8264 		case DIF_OP_MOV:
8265 		case DIF_OP_ALLOCS:
8266 			if (r1 >= nregs)
8267 				err += efunc(pc, "invalid register %u\n", r1);
8268 			if (r2 != 0)
8269 				err += efunc(pc, "non-zero reserved bits\n");
8270 			if (rd >= nregs)
8271 				err += efunc(pc, "invalid register %u\n", rd);
8272 			if (rd == 0)
8273 				err += efunc(pc, "cannot write to %r0\n");
8274 			break;
8275 		case DIF_OP_LDSB:
8276 		case DIF_OP_LDSH:
8277 		case DIF_OP_LDSW:
8278 		case DIF_OP_LDUB:
8279 		case DIF_OP_LDUH:
8280 		case DIF_OP_LDUW:
8281 		case DIF_OP_LDX:
8282 			if (r1 >= nregs)
8283 				err += efunc(pc, "invalid register %u\n", r1);
8284 			if (r2 != 0)
8285 				err += efunc(pc, "non-zero reserved bits\n");
8286 			if (rd >= nregs)
8287 				err += efunc(pc, "invalid register %u\n", rd);
8288 			if (rd == 0)
8289 				err += efunc(pc, "cannot write to %r0\n");
8290 			if (kcheckload)
8291 				dp->dtdo_buf[pc] = DIF_INSTR_LOAD(op +
8292 				    DIF_OP_RLDSB - DIF_OP_LDSB, r1, rd);
8293 			break;
8294 		case DIF_OP_RLDSB:
8295 		case DIF_OP_RLDSH:
8296 		case DIF_OP_RLDSW:
8297 		case DIF_OP_RLDUB:
8298 		case DIF_OP_RLDUH:
8299 		case DIF_OP_RLDUW:
8300 		case DIF_OP_RLDX:
8301 			if (r1 >= nregs)
8302 				err += efunc(pc, "invalid register %u\n", r1);
8303 			if (r2 != 0)
8304 				err += efunc(pc, "non-zero reserved bits\n");
8305 			if (rd >= nregs)
8306 				err += efunc(pc, "invalid register %u\n", rd);
8307 			if (rd == 0)
8308 				err += efunc(pc, "cannot write to %r0\n");
8309 			break;
8310 		case DIF_OP_ULDSB:
8311 		case DIF_OP_ULDSH:
8312 		case DIF_OP_ULDSW:
8313 		case DIF_OP_ULDUB:
8314 		case DIF_OP_ULDUH:
8315 		case DIF_OP_ULDUW:
8316 		case DIF_OP_ULDX:
8317 			if (r1 >= nregs)
8318 				err += efunc(pc, "invalid register %u\n", r1);
8319 			if (r2 != 0)
8320 				err += efunc(pc, "non-zero reserved bits\n");
8321 			if (rd >= nregs)
8322 				err += efunc(pc, "invalid register %u\n", rd);
8323 			if (rd == 0)
8324 				err += efunc(pc, "cannot write to %r0\n");
8325 			break;
8326 		case DIF_OP_STB:
8327 		case DIF_OP_STH:
8328 		case DIF_OP_STW:
8329 		case DIF_OP_STX:
8330 			if (r1 >= nregs)
8331 				err += efunc(pc, "invalid register %u\n", r1);
8332 			if (r2 != 0)
8333 				err += efunc(pc, "non-zero reserved bits\n");
8334 			if (rd >= nregs)
8335 				err += efunc(pc, "invalid register %u\n", rd);
8336 			if (rd == 0)
8337 				err += efunc(pc, "cannot write to 0 address\n");
8338 			break;
8339 		case DIF_OP_CMP:
8340 		case DIF_OP_SCMP:
8341 			if (r1 >= nregs)
8342 				err += efunc(pc, "invalid register %u\n", r1);
8343 			if (r2 >= nregs)
8344 				err += efunc(pc, "invalid register %u\n", r2);
8345 			if (rd != 0)
8346 				err += efunc(pc, "non-zero reserved bits\n");
8347 			break;
8348 		case DIF_OP_TST:
8349 			if (r1 >= nregs)
8350 				err += efunc(pc, "invalid register %u\n", r1);
8351 			if (r2 != 0 || rd != 0)
8352 				err += efunc(pc, "non-zero reserved bits\n");
8353 			break;
8354 		case DIF_OP_BA:
8355 		case DIF_OP_BE:
8356 		case DIF_OP_BNE:
8357 		case DIF_OP_BG:
8358 		case DIF_OP_BGU:
8359 		case DIF_OP_BGE:
8360 		case DIF_OP_BGEU:
8361 		case DIF_OP_BL:
8362 		case DIF_OP_BLU:
8363 		case DIF_OP_BLE:
8364 		case DIF_OP_BLEU:
8365 			if (label >= dp->dtdo_len) {
8366 				err += efunc(pc, "invalid branch target %u\n",
8367 				    label);
8368 			}
8369 			if (label <= pc) {
8370 				err += efunc(pc, "backward branch to %u\n",
8371 				    label);
8372 			}
8373 			break;
8374 		case DIF_OP_RET:
8375 			if (r1 != 0 || r2 != 0)
8376 				err += efunc(pc, "non-zero reserved bits\n");
8377 			if (rd >= nregs)
8378 				err += efunc(pc, "invalid register %u\n", rd);
8379 			break;
8380 		case DIF_OP_NOP:
8381 		case DIF_OP_POPTS:
8382 		case DIF_OP_FLUSHTS:
8383 			if (r1 != 0 || r2 != 0 || rd != 0)
8384 				err += efunc(pc, "non-zero reserved bits\n");
8385 			break;
8386 		case DIF_OP_SETX:
8387 			if (DIF_INSTR_INTEGER(instr) >= dp->dtdo_intlen) {
8388 				err += efunc(pc, "invalid integer ref %u\n",
8389 				    DIF_INSTR_INTEGER(instr));
8390 			}
8391 			if (rd >= nregs)
8392 				err += efunc(pc, "invalid register %u\n", rd);
8393 			if (rd == 0)
8394 				err += efunc(pc, "cannot write to %r0\n");
8395 			break;
8396 		case DIF_OP_SETS:
8397 			if (DIF_INSTR_STRING(instr) >= dp->dtdo_strlen) {
8398 				err += efunc(pc, "invalid string ref %u\n",
8399 				    DIF_INSTR_STRING(instr));
8400 			}
8401 			if (rd >= nregs)
8402 				err += efunc(pc, "invalid register %u\n", rd);
8403 			if (rd == 0)
8404 				err += efunc(pc, "cannot write to %r0\n");
8405 			break;
8406 		case DIF_OP_LDGA:
8407 		case DIF_OP_LDTA:
8408 			if (r1 > DIF_VAR_ARRAY_MAX)
8409 				err += efunc(pc, "invalid array %u\n", r1);
8410 			if (r2 >= nregs)
8411 				err += efunc(pc, "invalid register %u\n", r2);
8412 			if (rd >= nregs)
8413 				err += efunc(pc, "invalid register %u\n", rd);
8414 			if (rd == 0)
8415 				err += efunc(pc, "cannot write to %r0\n");
8416 			break;
8417 		case DIF_OP_LDGS:
8418 		case DIF_OP_LDTS:
8419 		case DIF_OP_LDLS:
8420 		case DIF_OP_LDGAA:
8421 		case DIF_OP_LDTAA:
8422 			if (v < DIF_VAR_OTHER_MIN || v > DIF_VAR_OTHER_MAX)
8423 				err += efunc(pc, "invalid variable %u\n", v);
8424 			if (rd >= nregs)
8425 				err += efunc(pc, "invalid register %u\n", rd);
8426 			if (rd == 0)
8427 				err += efunc(pc, "cannot write to %r0\n");
8428 			break;
8429 		case DIF_OP_STGS:
8430 		case DIF_OP_STTS:
8431 		case DIF_OP_STLS:
8432 		case DIF_OP_STGAA:
8433 		case DIF_OP_STTAA:
8434 			if (v < DIF_VAR_OTHER_UBASE || v > DIF_VAR_OTHER_MAX)
8435 				err += efunc(pc, "invalid variable %u\n", v);
8436 			if (rs >= nregs)
8437 				err += efunc(pc, "invalid register %u\n", rd);
8438 			break;
8439 		case DIF_OP_CALL:
8440 			if (subr > DIF_SUBR_MAX)
8441 				err += efunc(pc, "invalid subr %u\n", subr);
8442 			if (rd >= nregs)
8443 				err += efunc(pc, "invalid register %u\n", rd);
8444 			if (rd == 0)
8445 				err += efunc(pc, "cannot write to %r0\n");
8446 
8447 			if (subr == DIF_SUBR_COPYOUT ||
8448 			    subr == DIF_SUBR_COPYOUTSTR) {
8449 				dp->dtdo_destructive = 1;
8450 			}
8451 			break;
8452 		case DIF_OP_PUSHTR:
8453 			if (type != DIF_TYPE_STRING && type != DIF_TYPE_CTF)
8454 				err += efunc(pc, "invalid ref type %u\n", type);
8455 			if (r2 >= nregs)
8456 				err += efunc(pc, "invalid register %u\n", r2);
8457 			if (rs >= nregs)
8458 				err += efunc(pc, "invalid register %u\n", rs);
8459 			break;
8460 		case DIF_OP_PUSHTV:
8461 			if (type != DIF_TYPE_CTF)
8462 				err += efunc(pc, "invalid val type %u\n", type);
8463 			if (r2 >= nregs)
8464 				err += efunc(pc, "invalid register %u\n", r2);
8465 			if (rs >= nregs)
8466 				err += efunc(pc, "invalid register %u\n", rs);
8467 			break;
8468 		default:
8469 			err += efunc(pc, "invalid opcode %u\n",
8470 			    DIF_INSTR_OP(instr));
8471 		}
8472 	}
8473 
8474 	if (dp->dtdo_len != 0 &&
8475 	    DIF_INSTR_OP(dp->dtdo_buf[dp->dtdo_len - 1]) != DIF_OP_RET) {
8476 		err += efunc(dp->dtdo_len - 1,
8477 		    "expected 'ret' as last DIF instruction\n");
8478 	}
8479 
8480 	if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF)) {
8481 		/*
8482 		 * If we're not returning by reference, the size must be either
8483 		 * 0 or the size of one of the base types.
8484 		 */
8485 		switch (dp->dtdo_rtype.dtdt_size) {
8486 		case 0:
8487 		case sizeof (uint8_t):
8488 		case sizeof (uint16_t):
8489 		case sizeof (uint32_t):
8490 		case sizeof (uint64_t):
8491 			break;
8492 
8493 		default:
8494 			err += efunc(dp->dtdo_len - 1, "bad return size\n");
8495 		}
8496 	}
8497 
8498 	for (i = 0; i < dp->dtdo_varlen && err == 0; i++) {
8499 		dtrace_difv_t *v = &dp->dtdo_vartab[i], *existing = NULL;
8500 		dtrace_diftype_t *vt, *et;
8501 		uint_t id, ndx;
8502 
8503 		if (v->dtdv_scope != DIFV_SCOPE_GLOBAL &&
8504 		    v->dtdv_scope != DIFV_SCOPE_THREAD &&
8505 		    v->dtdv_scope != DIFV_SCOPE_LOCAL) {
8506 			err += efunc(i, "unrecognized variable scope %d\n",
8507 			    v->dtdv_scope);
8508 			break;
8509 		}
8510 
8511 		if (v->dtdv_kind != DIFV_KIND_ARRAY &&
8512 		    v->dtdv_kind != DIFV_KIND_SCALAR) {
8513 			err += efunc(i, "unrecognized variable type %d\n",
8514 			    v->dtdv_kind);
8515 			break;
8516 		}
8517 
8518 		if ((id = v->dtdv_id) > DIF_VARIABLE_MAX) {
8519 			err += efunc(i, "%d exceeds variable id limit\n", id);
8520 			break;
8521 		}
8522 
8523 		if (id < DIF_VAR_OTHER_UBASE)
8524 			continue;
8525 
8526 		/*
8527 		 * For user-defined variables, we need to check that this
8528 		 * definition is identical to any previous definition that we
8529 		 * encountered.
8530 		 */
8531 		ndx = id - DIF_VAR_OTHER_UBASE;
8532 
8533 		switch (v->dtdv_scope) {
8534 		case DIFV_SCOPE_GLOBAL:
8535 			if (ndx < vstate->dtvs_nglobals) {
8536 				dtrace_statvar_t *svar;
8537 
8538 				if ((svar = vstate->dtvs_globals[ndx]) != NULL)
8539 					existing = &svar->dtsv_var;
8540 			}
8541 
8542 			break;
8543 
8544 		case DIFV_SCOPE_THREAD:
8545 			if (ndx < vstate->dtvs_ntlocals)
8546 				existing = &vstate->dtvs_tlocals[ndx];
8547 			break;
8548 
8549 		case DIFV_SCOPE_LOCAL:
8550 			if (ndx < vstate->dtvs_nlocals) {
8551 				dtrace_statvar_t *svar;
8552 
8553 				if ((svar = vstate->dtvs_locals[ndx]) != NULL)
8554 					existing = &svar->dtsv_var;
8555 			}
8556 
8557 			break;
8558 		}
8559 
8560 		vt = &v->dtdv_type;
8561 
8562 		if (vt->dtdt_flags & DIF_TF_BYREF) {
8563 			if (vt->dtdt_size == 0) {
8564 				err += efunc(i, "zero-sized variable\n");
8565 				break;
8566 			}
8567 
8568 			if (v->dtdv_scope == DIFV_SCOPE_GLOBAL &&
8569 			    vt->dtdt_size > dtrace_global_maxsize) {
8570 				err += efunc(i, "oversized by-ref global\n");
8571 				break;
8572 			}
8573 		}
8574 
8575 		if (existing == NULL || existing->dtdv_id == 0)
8576 			continue;
8577 
8578 		ASSERT(existing->dtdv_id == v->dtdv_id);
8579 		ASSERT(existing->dtdv_scope == v->dtdv_scope);
8580 
8581 		if (existing->dtdv_kind != v->dtdv_kind)
8582 			err += efunc(i, "%d changed variable kind\n", id);
8583 
8584 		et = &existing->dtdv_type;
8585 
8586 		if (vt->dtdt_flags != et->dtdt_flags) {
8587 			err += efunc(i, "%d changed variable type flags\n", id);
8588 			break;
8589 		}
8590 
8591 		if (vt->dtdt_size != 0 && vt->dtdt_size != et->dtdt_size) {
8592 			err += efunc(i, "%d changed variable type size\n", id);
8593 			break;
8594 		}
8595 	}
8596 
8597 	return (err);
8598 }
8599 
8600 /*
8601  * Validate a DTrace DIF object that it is to be used as a helper.  Helpers
8602  * are much more constrained than normal DIFOs.  Specifically, they may
8603  * not:
8604  *
8605  * 1. Make calls to subroutines other than copyin(), copyinstr() or
8606  *    miscellaneous string routines
8607  * 2. Access DTrace variables other than the args[] array, and the
8608  *    curthread, pid, ppid, tid, execname, zonename, uid and gid variables.
8609  * 3. Have thread-local variables.
8610  * 4. Have dynamic variables.
8611  */
8612 static int
8613 dtrace_difo_validate_helper(dtrace_difo_t *dp)
8614 {
8615 	int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
8616 	int err = 0;
8617 	uint_t pc;
8618 
8619 	for (pc = 0; pc < dp->dtdo_len; pc++) {
8620 		dif_instr_t instr = dp->dtdo_buf[pc];
8621 
8622 		uint_t v = DIF_INSTR_VAR(instr);
8623 		uint_t subr = DIF_INSTR_SUBR(instr);
8624 		uint_t op = DIF_INSTR_OP(instr);
8625 
8626 		switch (op) {
8627 		case DIF_OP_OR:
8628 		case DIF_OP_XOR:
8629 		case DIF_OP_AND:
8630 		case DIF_OP_SLL:
8631 		case DIF_OP_SRL:
8632 		case DIF_OP_SRA:
8633 		case DIF_OP_SUB:
8634 		case DIF_OP_ADD:
8635 		case DIF_OP_MUL:
8636 		case DIF_OP_SDIV:
8637 		case DIF_OP_UDIV:
8638 		case DIF_OP_SREM:
8639 		case DIF_OP_UREM:
8640 		case DIF_OP_COPYS:
8641 		case DIF_OP_NOT:
8642 		case DIF_OP_MOV:
8643 		case DIF_OP_RLDSB:
8644 		case DIF_OP_RLDSH:
8645 		case DIF_OP_RLDSW:
8646 		case DIF_OP_RLDUB:
8647 		case DIF_OP_RLDUH:
8648 		case DIF_OP_RLDUW:
8649 		case DIF_OP_RLDX:
8650 		case DIF_OP_ULDSB:
8651 		case DIF_OP_ULDSH:
8652 		case DIF_OP_ULDSW:
8653 		case DIF_OP_ULDUB:
8654 		case DIF_OP_ULDUH:
8655 		case DIF_OP_ULDUW:
8656 		case DIF_OP_ULDX:
8657 		case DIF_OP_STB:
8658 		case DIF_OP_STH:
8659 		case DIF_OP_STW:
8660 		case DIF_OP_STX:
8661 		case DIF_OP_ALLOCS:
8662 		case DIF_OP_CMP:
8663 		case DIF_OP_SCMP:
8664 		case DIF_OP_TST:
8665 		case DIF_OP_BA:
8666 		case DIF_OP_BE:
8667 		case DIF_OP_BNE:
8668 		case DIF_OP_BG:
8669 		case DIF_OP_BGU:
8670 		case DIF_OP_BGE:
8671 		case DIF_OP_BGEU:
8672 		case DIF_OP_BL:
8673 		case DIF_OP_BLU:
8674 		case DIF_OP_BLE:
8675 		case DIF_OP_BLEU:
8676 		case DIF_OP_RET:
8677 		case DIF_OP_NOP:
8678 		case DIF_OP_POPTS:
8679 		case DIF_OP_FLUSHTS:
8680 		case DIF_OP_SETX:
8681 		case DIF_OP_SETS:
8682 		case DIF_OP_LDGA:
8683 		case DIF_OP_LDLS:
8684 		case DIF_OP_STGS:
8685 		case DIF_OP_STLS:
8686 		case DIF_OP_PUSHTR:
8687 		case DIF_OP_PUSHTV:
8688 			break;
8689 
8690 		case DIF_OP_LDGS:
8691 			if (v >= DIF_VAR_OTHER_UBASE)
8692 				break;
8693 
8694 			if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9)
8695 				break;
8696 
8697 			if (v == DIF_VAR_CURTHREAD || v == DIF_VAR_PID ||
8698 			    v == DIF_VAR_PPID || v == DIF_VAR_TID ||
8699 			    v == DIF_VAR_EXECNAME || v == DIF_VAR_ZONENAME ||
8700 			    v == DIF_VAR_UID || v == DIF_VAR_GID)
8701 				break;
8702 
8703 			err += efunc(pc, "illegal variable %u\n", v);
8704 			break;
8705 
8706 		case DIF_OP_LDTA:
8707 		case DIF_OP_LDTS:
8708 		case DIF_OP_LDGAA:
8709 		case DIF_OP_LDTAA:
8710 			err += efunc(pc, "illegal dynamic variable load\n");
8711 			break;
8712 
8713 		case DIF_OP_STTS:
8714 		case DIF_OP_STGAA:
8715 		case DIF_OP_STTAA:
8716 			err += efunc(pc, "illegal dynamic variable store\n");
8717 			break;
8718 
8719 		case DIF_OP_CALL:
8720 			if (subr == DIF_SUBR_ALLOCA ||
8721 			    subr == DIF_SUBR_BCOPY ||
8722 			    subr == DIF_SUBR_COPYIN ||
8723 			    subr == DIF_SUBR_COPYINTO ||
8724 			    subr == DIF_SUBR_COPYINSTR ||
8725 			    subr == DIF_SUBR_INDEX ||
8726 			    subr == DIF_SUBR_INET_NTOA ||
8727 			    subr == DIF_SUBR_INET_NTOA6 ||
8728 			    subr == DIF_SUBR_INET_NTOP ||
8729 			    subr == DIF_SUBR_LLTOSTR ||
8730 			    subr == DIF_SUBR_RINDEX ||
8731 			    subr == DIF_SUBR_STRCHR ||
8732 			    subr == DIF_SUBR_STRJOIN ||
8733 			    subr == DIF_SUBR_STRRCHR ||
8734 			    subr == DIF_SUBR_STRSTR ||
8735 			    subr == DIF_SUBR_HTONS ||
8736 			    subr == DIF_SUBR_HTONL ||
8737 			    subr == DIF_SUBR_HTONLL ||
8738 			    subr == DIF_SUBR_NTOHS ||
8739 			    subr == DIF_SUBR_NTOHL ||
8740 			    subr == DIF_SUBR_NTOHLL)
8741 				break;
8742 
8743 			err += efunc(pc, "invalid subr %u\n", subr);
8744 			break;
8745 
8746 		default:
8747 			err += efunc(pc, "invalid opcode %u\n",
8748 			    DIF_INSTR_OP(instr));
8749 		}
8750 	}
8751 
8752 	return (err);
8753 }
8754 
8755 /*
8756  * Returns 1 if the expression in the DIF object can be cached on a per-thread
8757  * basis; 0 if not.
8758  */
8759 static int
8760 dtrace_difo_cacheable(dtrace_difo_t *dp)
8761 {
8762 	int i;
8763 
8764 	if (dp == NULL)
8765 		return (0);
8766 
8767 	for (i = 0; i < dp->dtdo_varlen; i++) {
8768 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
8769 
8770 		if (v->dtdv_scope != DIFV_SCOPE_GLOBAL)
8771 			continue;
8772 
8773 		switch (v->dtdv_id) {
8774 		case DIF_VAR_CURTHREAD:
8775 		case DIF_VAR_PID:
8776 		case DIF_VAR_TID:
8777 		case DIF_VAR_EXECNAME:
8778 		case DIF_VAR_ZONENAME:
8779 			break;
8780 
8781 		default:
8782 			return (0);
8783 		}
8784 	}
8785 
8786 	/*
8787 	 * This DIF object may be cacheable.  Now we need to look for any
8788 	 * array loading instructions, any memory loading instructions, or
8789 	 * any stores to thread-local variables.
8790 	 */
8791 	for (i = 0; i < dp->dtdo_len; i++) {
8792 		uint_t op = DIF_INSTR_OP(dp->dtdo_buf[i]);
8793 
8794 		if ((op >= DIF_OP_LDSB && op <= DIF_OP_LDX) ||
8795 		    (op >= DIF_OP_ULDSB && op <= DIF_OP_ULDX) ||
8796 		    (op >= DIF_OP_RLDSB && op <= DIF_OP_RLDX) ||
8797 		    op == DIF_OP_LDGA || op == DIF_OP_STTS)
8798 			return (0);
8799 	}
8800 
8801 	return (1);
8802 }
8803 
8804 static void
8805 dtrace_difo_hold(dtrace_difo_t *dp)
8806 {
8807 	int i;
8808 
8809 	ASSERT(MUTEX_HELD(&dtrace_lock));
8810 
8811 	dp->dtdo_refcnt++;
8812 	ASSERT(dp->dtdo_refcnt != 0);
8813 
8814 	/*
8815 	 * We need to check this DIF object for references to the variable
8816 	 * DIF_VAR_VTIMESTAMP.
8817 	 */
8818 	for (i = 0; i < dp->dtdo_varlen; i++) {
8819 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
8820 
8821 		if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
8822 			continue;
8823 
8824 		if (dtrace_vtime_references++ == 0)
8825 			dtrace_vtime_enable();
8826 	}
8827 }
8828 
8829 /*
8830  * This routine calculates the dynamic variable chunksize for a given DIF
8831  * object.  The calculation is not fool-proof, and can probably be tricked by
8832  * malicious DIF -- but it works for all compiler-generated DIF.  Because this
8833  * calculation is likely imperfect, dtrace_dynvar() is able to gracefully fail
8834  * if a dynamic variable size exceeds the chunksize.
8835  */
8836 static void
8837 dtrace_difo_chunksize(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
8838 {
8839 	uint64_t sval;
8840 	dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
8841 	const dif_instr_t *text = dp->dtdo_buf;
8842 	uint_t pc, srd = 0;
8843 	uint_t ttop = 0;
8844 	size_t size, ksize;
8845 	uint_t id, i;
8846 
8847 	for (pc = 0; pc < dp->dtdo_len; pc++) {
8848 		dif_instr_t instr = text[pc];
8849 		uint_t op = DIF_INSTR_OP(instr);
8850 		uint_t rd = DIF_INSTR_RD(instr);
8851 		uint_t r1 = DIF_INSTR_R1(instr);
8852 		uint_t nkeys = 0;
8853 		uchar_t scope;
8854 
8855 		dtrace_key_t *key = tupregs;
8856 
8857 		switch (op) {
8858 		case DIF_OP_SETX:
8859 			sval = dp->dtdo_inttab[DIF_INSTR_INTEGER(instr)];
8860 			srd = rd;
8861 			continue;
8862 
8863 		case DIF_OP_STTS:
8864 			key = &tupregs[DIF_DTR_NREGS];
8865 			key[0].dttk_size = 0;
8866 			key[1].dttk_size = 0;
8867 			nkeys = 2;
8868 			scope = DIFV_SCOPE_THREAD;
8869 			break;
8870 
8871 		case DIF_OP_STGAA:
8872 		case DIF_OP_STTAA:
8873 			nkeys = ttop;
8874 
8875 			if (DIF_INSTR_OP(instr) == DIF_OP_STTAA)
8876 				key[nkeys++].dttk_size = 0;
8877 
8878 			key[nkeys++].dttk_size = 0;
8879 
8880 			if (op == DIF_OP_STTAA) {
8881 				scope = DIFV_SCOPE_THREAD;
8882 			} else {
8883 				scope = DIFV_SCOPE_GLOBAL;
8884 			}
8885 
8886 			break;
8887 
8888 		case DIF_OP_PUSHTR:
8889 			if (ttop == DIF_DTR_NREGS)
8890 				return;
8891 
8892 			if ((srd == 0 || sval == 0) && r1 == DIF_TYPE_STRING) {
8893 				/*
8894 				 * If the register for the size of the "pushtr"
8895 				 * is %r0 (or the value is 0) and the type is
8896 				 * a string, we'll use the system-wide default
8897 				 * string size.
8898 				 */
8899 				tupregs[ttop++].dttk_size =
8900 				    dtrace_strsize_default;
8901 			} else {
8902 				if (srd == 0)
8903 					return;
8904 
8905 				tupregs[ttop++].dttk_size = sval;
8906 			}
8907 
8908 			break;
8909 
8910 		case DIF_OP_PUSHTV:
8911 			if (ttop == DIF_DTR_NREGS)
8912 				return;
8913 
8914 			tupregs[ttop++].dttk_size = 0;
8915 			break;
8916 
8917 		case DIF_OP_FLUSHTS:
8918 			ttop = 0;
8919 			break;
8920 
8921 		case DIF_OP_POPTS:
8922 			if (ttop != 0)
8923 				ttop--;
8924 			break;
8925 		}
8926 
8927 		sval = 0;
8928 		srd = 0;
8929 
8930 		if (nkeys == 0)
8931 			continue;
8932 
8933 		/*
8934 		 * We have a dynamic variable allocation; calculate its size.
8935 		 */
8936 		for (ksize = 0, i = 0; i < nkeys; i++)
8937 			ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
8938 
8939 		size = sizeof (dtrace_dynvar_t);
8940 		size += sizeof (dtrace_key_t) * (nkeys - 1);
8941 		size += ksize;
8942 
8943 		/*
8944 		 * Now we need to determine the size of the stored data.
8945 		 */
8946 		id = DIF_INSTR_VAR(instr);
8947 
8948 		for (i = 0; i < dp->dtdo_varlen; i++) {
8949 			dtrace_difv_t *v = &dp->dtdo_vartab[i];
8950 
8951 			if (v->dtdv_id == id && v->dtdv_scope == scope) {
8952 				size += v->dtdv_type.dtdt_size;
8953 				break;
8954 			}
8955 		}
8956 
8957 		if (i == dp->dtdo_varlen)
8958 			return;
8959 
8960 		/*
8961 		 * We have the size.  If this is larger than the chunk size
8962 		 * for our dynamic variable state, reset the chunk size.
8963 		 */
8964 		size = P2ROUNDUP(size, sizeof (uint64_t));
8965 
8966 		if (size > vstate->dtvs_dynvars.dtds_chunksize)
8967 			vstate->dtvs_dynvars.dtds_chunksize = size;
8968 	}
8969 }
8970 
8971 static void
8972 dtrace_difo_init(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
8973 {
8974 	int i, oldsvars, osz, nsz, otlocals, ntlocals;
8975 	uint_t id;
8976 
8977 	ASSERT(MUTEX_HELD(&dtrace_lock));
8978 	ASSERT(dp->dtdo_buf != NULL && dp->dtdo_len != 0);
8979 
8980 	for (i = 0; i < dp->dtdo_varlen; i++) {
8981 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
8982 		dtrace_statvar_t *svar, ***svarp;
8983 		size_t dsize = 0;
8984 		uint8_t scope = v->dtdv_scope;
8985 		int *np;
8986 
8987 		if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
8988 			continue;
8989 
8990 		id -= DIF_VAR_OTHER_UBASE;
8991 
8992 		switch (scope) {
8993 		case DIFV_SCOPE_THREAD:
8994 			while (id >= (otlocals = vstate->dtvs_ntlocals)) {
8995 				dtrace_difv_t *tlocals;
8996 
8997 				if ((ntlocals = (otlocals << 1)) == 0)
8998 					ntlocals = 1;
8999 
9000 				osz = otlocals * sizeof (dtrace_difv_t);
9001 				nsz = ntlocals * sizeof (dtrace_difv_t);
9002 
9003 				tlocals = kmem_zalloc(nsz, KM_SLEEP);
9004 
9005 				if (osz != 0) {
9006 					bcopy(vstate->dtvs_tlocals,
9007 					    tlocals, osz);
9008 					kmem_free(vstate->dtvs_tlocals, osz);
9009 				}
9010 
9011 				vstate->dtvs_tlocals = tlocals;
9012 				vstate->dtvs_ntlocals = ntlocals;
9013 			}
9014 
9015 			vstate->dtvs_tlocals[id] = *v;
9016 			continue;
9017 
9018 		case DIFV_SCOPE_LOCAL:
9019 			np = &vstate->dtvs_nlocals;
9020 			svarp = &vstate->dtvs_locals;
9021 
9022 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
9023 				dsize = NCPU * (v->dtdv_type.dtdt_size +
9024 				    sizeof (uint64_t));
9025 			else
9026 				dsize = NCPU * sizeof (uint64_t);
9027 
9028 			break;
9029 
9030 		case DIFV_SCOPE_GLOBAL:
9031 			np = &vstate->dtvs_nglobals;
9032 			svarp = &vstate->dtvs_globals;
9033 
9034 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
9035 				dsize = v->dtdv_type.dtdt_size +
9036 				    sizeof (uint64_t);
9037 
9038 			break;
9039 
9040 		default:
9041 			ASSERT(0);
9042 		}
9043 
9044 		while (id >= (oldsvars = *np)) {
9045 			dtrace_statvar_t **statics;
9046 			int newsvars, oldsize, newsize;
9047 
9048 			if ((newsvars = (oldsvars << 1)) == 0)
9049 				newsvars = 1;
9050 
9051 			oldsize = oldsvars * sizeof (dtrace_statvar_t *);
9052 			newsize = newsvars * sizeof (dtrace_statvar_t *);
9053 
9054 			statics = kmem_zalloc(newsize, KM_SLEEP);
9055 
9056 			if (oldsize != 0) {
9057 				bcopy(*svarp, statics, oldsize);
9058 				kmem_free(*svarp, oldsize);
9059 			}
9060 
9061 			*svarp = statics;
9062 			*np = newsvars;
9063 		}
9064 
9065 		if ((svar = (*svarp)[id]) == NULL) {
9066 			svar = kmem_zalloc(sizeof (dtrace_statvar_t), KM_SLEEP);
9067 			svar->dtsv_var = *v;
9068 
9069 			if ((svar->dtsv_size = dsize) != 0) {
9070 				svar->dtsv_data = (uint64_t)(uintptr_t)
9071 				    kmem_zalloc(dsize, KM_SLEEP);
9072 			}
9073 
9074 			(*svarp)[id] = svar;
9075 		}
9076 
9077 		svar->dtsv_refcnt++;
9078 	}
9079 
9080 	dtrace_difo_chunksize(dp, vstate);
9081 	dtrace_difo_hold(dp);
9082 }
9083 
9084 static dtrace_difo_t *
9085 dtrace_difo_duplicate(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9086 {
9087 	dtrace_difo_t *new;
9088 	size_t sz;
9089 
9090 	ASSERT(dp->dtdo_buf != NULL);
9091 	ASSERT(dp->dtdo_refcnt != 0);
9092 
9093 	new = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
9094 
9095 	ASSERT(dp->dtdo_buf != NULL);
9096 	sz = dp->dtdo_len * sizeof (dif_instr_t);
9097 	new->dtdo_buf = kmem_alloc(sz, KM_SLEEP);
9098 	bcopy(dp->dtdo_buf, new->dtdo_buf, sz);
9099 	new->dtdo_len = dp->dtdo_len;
9100 
9101 	if (dp->dtdo_strtab != NULL) {
9102 		ASSERT(dp->dtdo_strlen != 0);
9103 		new->dtdo_strtab = kmem_alloc(dp->dtdo_strlen, KM_SLEEP);
9104 		bcopy(dp->dtdo_strtab, new->dtdo_strtab, dp->dtdo_strlen);
9105 		new->dtdo_strlen = dp->dtdo_strlen;
9106 	}
9107 
9108 	if (dp->dtdo_inttab != NULL) {
9109 		ASSERT(dp->dtdo_intlen != 0);
9110 		sz = dp->dtdo_intlen * sizeof (uint64_t);
9111 		new->dtdo_inttab = kmem_alloc(sz, KM_SLEEP);
9112 		bcopy(dp->dtdo_inttab, new->dtdo_inttab, sz);
9113 		new->dtdo_intlen = dp->dtdo_intlen;
9114 	}
9115 
9116 	if (dp->dtdo_vartab != NULL) {
9117 		ASSERT(dp->dtdo_varlen != 0);
9118 		sz = dp->dtdo_varlen * sizeof (dtrace_difv_t);
9119 		new->dtdo_vartab = kmem_alloc(sz, KM_SLEEP);
9120 		bcopy(dp->dtdo_vartab, new->dtdo_vartab, sz);
9121 		new->dtdo_varlen = dp->dtdo_varlen;
9122 	}
9123 
9124 	dtrace_difo_init(new, vstate);
9125 	return (new);
9126 }
9127 
9128 static void
9129 dtrace_difo_destroy(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9130 {
9131 	int i;
9132 
9133 	ASSERT(dp->dtdo_refcnt == 0);
9134 
9135 	for (i = 0; i < dp->dtdo_varlen; i++) {
9136 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
9137 		dtrace_statvar_t *svar, **svarp;
9138 		uint_t id;
9139 		uint8_t scope = v->dtdv_scope;
9140 		int *np;
9141 
9142 		switch (scope) {
9143 		case DIFV_SCOPE_THREAD:
9144 			continue;
9145 
9146 		case DIFV_SCOPE_LOCAL:
9147 			np = &vstate->dtvs_nlocals;
9148 			svarp = vstate->dtvs_locals;
9149 			break;
9150 
9151 		case DIFV_SCOPE_GLOBAL:
9152 			np = &vstate->dtvs_nglobals;
9153 			svarp = vstate->dtvs_globals;
9154 			break;
9155 
9156 		default:
9157 			ASSERT(0);
9158 		}
9159 
9160 		if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
9161 			continue;
9162 
9163 		id -= DIF_VAR_OTHER_UBASE;
9164 		ASSERT(id < *np);
9165 
9166 		svar = svarp[id];
9167 		ASSERT(svar != NULL);
9168 		ASSERT(svar->dtsv_refcnt > 0);
9169 
9170 		if (--svar->dtsv_refcnt > 0)
9171 			continue;
9172 
9173 		if (svar->dtsv_size != 0) {
9174 			ASSERT(svar->dtsv_data != NULL);
9175 			kmem_free((void *)(uintptr_t)svar->dtsv_data,
9176 			    svar->dtsv_size);
9177 		}
9178 
9179 		kmem_free(svar, sizeof (dtrace_statvar_t));
9180 		svarp[id] = NULL;
9181 	}
9182 
9183 	kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
9184 	kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
9185 	kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
9186 	kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
9187 
9188 	kmem_free(dp, sizeof (dtrace_difo_t));
9189 }
9190 
9191 static void
9192 dtrace_difo_release(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9193 {
9194 	int i;
9195 
9196 	ASSERT(MUTEX_HELD(&dtrace_lock));
9197 	ASSERT(dp->dtdo_refcnt != 0);
9198 
9199 	for (i = 0; i < dp->dtdo_varlen; i++) {
9200 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
9201 
9202 		if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
9203 			continue;
9204 
9205 		ASSERT(dtrace_vtime_references > 0);
9206 		if (--dtrace_vtime_references == 0)
9207 			dtrace_vtime_disable();
9208 	}
9209 
9210 	if (--dp->dtdo_refcnt == 0)
9211 		dtrace_difo_destroy(dp, vstate);
9212 }
9213 
9214 /*
9215  * DTrace Format Functions
9216  */
9217 static uint16_t
9218 dtrace_format_add(dtrace_state_t *state, char *str)
9219 {
9220 	char *fmt, **new;
9221 	uint16_t ndx, len = strlen(str) + 1;
9222 
9223 	fmt = kmem_zalloc(len, KM_SLEEP);
9224 	bcopy(str, fmt, len);
9225 
9226 	for (ndx = 0; ndx < state->dts_nformats; ndx++) {
9227 		if (state->dts_formats[ndx] == NULL) {
9228 			state->dts_formats[ndx] = fmt;
9229 			return (ndx + 1);
9230 		}
9231 	}
9232 
9233 	if (state->dts_nformats == USHRT_MAX) {
9234 		/*
9235 		 * This is only likely if a denial-of-service attack is being
9236 		 * attempted.  As such, it's okay to fail silently here.
9237 		 */
9238 		kmem_free(fmt, len);
9239 		return (0);
9240 	}
9241 
9242 	/*
9243 	 * For simplicity, we always resize the formats array to be exactly the
9244 	 * number of formats.
9245 	 */
9246 	ndx = state->dts_nformats++;
9247 	new = kmem_alloc((ndx + 1) * sizeof (char *), KM_SLEEP);
9248 
9249 	if (state->dts_formats != NULL) {
9250 		ASSERT(ndx != 0);
9251 		bcopy(state->dts_formats, new, ndx * sizeof (char *));
9252 		kmem_free(state->dts_formats, ndx * sizeof (char *));
9253 	}
9254 
9255 	state->dts_formats = new;
9256 	state->dts_formats[ndx] = fmt;
9257 
9258 	return (ndx + 1);
9259 }
9260 
9261 static void
9262 dtrace_format_remove(dtrace_state_t *state, uint16_t format)
9263 {
9264 	char *fmt;
9265 
9266 	ASSERT(state->dts_formats != NULL);
9267 	ASSERT(format <= state->dts_nformats);
9268 	ASSERT(state->dts_formats[format - 1] != NULL);
9269 
9270 	fmt = state->dts_formats[format - 1];
9271 	kmem_free(fmt, strlen(fmt) + 1);
9272 	state->dts_formats[format - 1] = NULL;
9273 }
9274 
9275 static void
9276 dtrace_format_destroy(dtrace_state_t *state)
9277 {
9278 	int i;
9279 
9280 	if (state->dts_nformats == 0) {
9281 		ASSERT(state->dts_formats == NULL);
9282 		return;
9283 	}
9284 
9285 	ASSERT(state->dts_formats != NULL);
9286 
9287 	for (i = 0; i < state->dts_nformats; i++) {
9288 		char *fmt = state->dts_formats[i];
9289 
9290 		if (fmt == NULL)
9291 			continue;
9292 
9293 		kmem_free(fmt, strlen(fmt) + 1);
9294 	}
9295 
9296 	kmem_free(state->dts_formats, state->dts_nformats * sizeof (char *));
9297 	state->dts_nformats = 0;
9298 	state->dts_formats = NULL;
9299 }
9300 
9301 /*
9302  * DTrace Predicate Functions
9303  */
9304 static dtrace_predicate_t *
9305 dtrace_predicate_create(dtrace_difo_t *dp)
9306 {
9307 	dtrace_predicate_t *pred;
9308 
9309 	ASSERT(MUTEX_HELD(&dtrace_lock));
9310 	ASSERT(dp->dtdo_refcnt != 0);
9311 
9312 	pred = kmem_zalloc(sizeof (dtrace_predicate_t), KM_SLEEP);
9313 	pred->dtp_difo = dp;
9314 	pred->dtp_refcnt = 1;
9315 
9316 	if (!dtrace_difo_cacheable(dp))
9317 		return (pred);
9318 
9319 	if (dtrace_predcache_id == DTRACE_CACHEIDNONE) {
9320 		/*
9321 		 * This is only theoretically possible -- we have had 2^32
9322 		 * cacheable predicates on this machine.  We cannot allow any
9323 		 * more predicates to become cacheable:  as unlikely as it is,
9324 		 * there may be a thread caching a (now stale) predicate cache
9325 		 * ID. (N.B.: the temptation is being successfully resisted to
9326 		 * have this cmn_err() "Holy shit -- we executed this code!")
9327 		 */
9328 		return (pred);
9329 	}
9330 
9331 	pred->dtp_cacheid = dtrace_predcache_id++;
9332 
9333 	return (pred);
9334 }
9335 
9336 static void
9337 dtrace_predicate_hold(dtrace_predicate_t *pred)
9338 {
9339 	ASSERT(MUTEX_HELD(&dtrace_lock));
9340 	ASSERT(pred->dtp_difo != NULL && pred->dtp_difo->dtdo_refcnt != 0);
9341 	ASSERT(pred->dtp_refcnt > 0);
9342 
9343 	pred->dtp_refcnt++;
9344 }
9345 
9346 static void
9347 dtrace_predicate_release(dtrace_predicate_t *pred, dtrace_vstate_t *vstate)
9348 {
9349 	dtrace_difo_t *dp = pred->dtp_difo;
9350 
9351 	ASSERT(MUTEX_HELD(&dtrace_lock));
9352 	ASSERT(dp != NULL && dp->dtdo_refcnt != 0);
9353 	ASSERT(pred->dtp_refcnt > 0);
9354 
9355 	if (--pred->dtp_refcnt == 0) {
9356 		dtrace_difo_release(pred->dtp_difo, vstate);
9357 		kmem_free(pred, sizeof (dtrace_predicate_t));
9358 	}
9359 }
9360 
9361 /*
9362  * DTrace Action Description Functions
9363  */
9364 static dtrace_actdesc_t *
9365 dtrace_actdesc_create(dtrace_actkind_t kind, uint32_t ntuple,
9366     uint64_t uarg, uint64_t arg)
9367 {
9368 	dtrace_actdesc_t *act;
9369 
9370 	ASSERT(!DTRACEACT_ISPRINTFLIKE(kind) || (arg != NULL &&
9371 	    arg >= KERNELBASE) || (arg == NULL && kind == DTRACEACT_PRINTA));
9372 
9373 	act = kmem_zalloc(sizeof (dtrace_actdesc_t), KM_SLEEP);
9374 	act->dtad_kind = kind;
9375 	act->dtad_ntuple = ntuple;
9376 	act->dtad_uarg = uarg;
9377 	act->dtad_arg = arg;
9378 	act->dtad_refcnt = 1;
9379 
9380 	return (act);
9381 }
9382 
9383 static void
9384 dtrace_actdesc_hold(dtrace_actdesc_t *act)
9385 {
9386 	ASSERT(act->dtad_refcnt >= 1);
9387 	act->dtad_refcnt++;
9388 }
9389 
9390 static void
9391 dtrace_actdesc_release(dtrace_actdesc_t *act, dtrace_vstate_t *vstate)
9392 {
9393 	dtrace_actkind_t kind = act->dtad_kind;
9394 	dtrace_difo_t *dp;
9395 
9396 	ASSERT(act->dtad_refcnt >= 1);
9397 
9398 	if (--act->dtad_refcnt != 0)
9399 		return;
9400 
9401 	if ((dp = act->dtad_difo) != NULL)
9402 		dtrace_difo_release(dp, vstate);
9403 
9404 	if (DTRACEACT_ISPRINTFLIKE(kind)) {
9405 		char *str = (char *)(uintptr_t)act->dtad_arg;
9406 
9407 		ASSERT((str != NULL && (uintptr_t)str >= KERNELBASE) ||
9408 		    (str == NULL && act->dtad_kind == DTRACEACT_PRINTA));
9409 
9410 		if (str != NULL)
9411 			kmem_free(str, strlen(str) + 1);
9412 	}
9413 
9414 	kmem_free(act, sizeof (dtrace_actdesc_t));
9415 }
9416 
9417 /*
9418  * DTrace ECB Functions
9419  */
9420 static dtrace_ecb_t *
9421 dtrace_ecb_add(dtrace_state_t *state, dtrace_probe_t *probe)
9422 {
9423 	dtrace_ecb_t *ecb;
9424 	dtrace_epid_t epid;
9425 
9426 	ASSERT(MUTEX_HELD(&dtrace_lock));
9427 
9428 	ecb = kmem_zalloc(sizeof (dtrace_ecb_t), KM_SLEEP);
9429 	ecb->dte_predicate = NULL;
9430 	ecb->dte_probe = probe;
9431 
9432 	/*
9433 	 * The default size is the size of the default action: recording
9434 	 * the header.
9435 	 */
9436 	ecb->dte_size = ecb->dte_needed = sizeof (dtrace_rechdr_t);
9437 	ecb->dte_alignment = sizeof (dtrace_epid_t);
9438 
9439 	epid = state->dts_epid++;
9440 
9441 	if (epid - 1 >= state->dts_necbs) {
9442 		dtrace_ecb_t **oecbs = state->dts_ecbs, **ecbs;
9443 		int necbs = state->dts_necbs << 1;
9444 
9445 		ASSERT(epid == state->dts_necbs + 1);
9446 
9447 		if (necbs == 0) {
9448 			ASSERT(oecbs == NULL);
9449 			necbs = 1;
9450 		}
9451 
9452 		ecbs = kmem_zalloc(necbs * sizeof (*ecbs), KM_SLEEP);
9453 
9454 		if (oecbs != NULL)
9455 			bcopy(oecbs, ecbs, state->dts_necbs * sizeof (*ecbs));
9456 
9457 		dtrace_membar_producer();
9458 		state->dts_ecbs = ecbs;
9459 
9460 		if (oecbs != NULL) {
9461 			/*
9462 			 * If this state is active, we must dtrace_sync()
9463 			 * before we can free the old dts_ecbs array:  we're
9464 			 * coming in hot, and there may be active ring
9465 			 * buffer processing (which indexes into the dts_ecbs
9466 			 * array) on another CPU.
9467 			 */
9468 			if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
9469 				dtrace_sync();
9470 
9471 			kmem_free(oecbs, state->dts_necbs * sizeof (*ecbs));
9472 		}
9473 
9474 		dtrace_membar_producer();
9475 		state->dts_necbs = necbs;
9476 	}
9477 
9478 	ecb->dte_state = state;
9479 
9480 	ASSERT(state->dts_ecbs[epid - 1] == NULL);
9481 	dtrace_membar_producer();
9482 	state->dts_ecbs[(ecb->dte_epid = epid) - 1] = ecb;
9483 
9484 	return (ecb);
9485 }
9486 
9487 static int
9488 dtrace_ecb_enable(dtrace_ecb_t *ecb)
9489 {
9490 	dtrace_probe_t *probe = ecb->dte_probe;
9491 
9492 	ASSERT(MUTEX_HELD(&cpu_lock));
9493 	ASSERT(MUTEX_HELD(&dtrace_lock));
9494 	ASSERT(ecb->dte_next == NULL);
9495 
9496 	if (probe == NULL) {
9497 		/*
9498 		 * This is the NULL probe -- there's nothing to do.
9499 		 */
9500 		return (0);
9501 	}
9502 
9503 	if (probe->dtpr_ecb == NULL) {
9504 		dtrace_provider_t *prov = probe->dtpr_provider;
9505 
9506 		/*
9507 		 * We're the first ECB on this probe.
9508 		 */
9509 		probe->dtpr_ecb = probe->dtpr_ecb_last = ecb;
9510 
9511 		if (ecb->dte_predicate != NULL)
9512 			probe->dtpr_predcache = ecb->dte_predicate->dtp_cacheid;
9513 
9514 		return (prov->dtpv_pops.dtps_enable(prov->dtpv_arg,
9515 		    probe->dtpr_id, probe->dtpr_arg));
9516 	} else {
9517 		/*
9518 		 * This probe is already active.  Swing the last pointer to
9519 		 * point to the new ECB, and issue a dtrace_sync() to assure
9520 		 * that all CPUs have seen the change.
9521 		 */
9522 		ASSERT(probe->dtpr_ecb_last != NULL);
9523 		probe->dtpr_ecb_last->dte_next = ecb;
9524 		probe->dtpr_ecb_last = ecb;
9525 		probe->dtpr_predcache = 0;
9526 
9527 		dtrace_sync();
9528 		return (0);
9529 	}
9530 }
9531 
9532 static void
9533 dtrace_ecb_resize(dtrace_ecb_t *ecb)
9534 {
9535 	dtrace_action_t *act;
9536 	uint32_t curneeded = UINT32_MAX;
9537 	uint32_t aggbase = UINT32_MAX;
9538 
9539 	/*
9540 	 * If we record anything, we always record the dtrace_rechdr_t.  (And
9541 	 * we always record it first.)
9542 	 */
9543 	ecb->dte_size = sizeof (dtrace_rechdr_t);
9544 	ecb->dte_alignment = sizeof (dtrace_epid_t);
9545 
9546 	for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
9547 		dtrace_recdesc_t *rec = &act->dta_rec;
9548 		ASSERT(rec->dtrd_size > 0 || rec->dtrd_alignment == 1);
9549 
9550 		ecb->dte_alignment = MAX(ecb->dte_alignment,
9551 		    rec->dtrd_alignment);
9552 
9553 		if (DTRACEACT_ISAGG(act->dta_kind)) {
9554 			dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
9555 
9556 			ASSERT(rec->dtrd_size != 0);
9557 			ASSERT(agg->dtag_first != NULL);
9558 			ASSERT(act->dta_prev->dta_intuple);
9559 			ASSERT(aggbase != UINT32_MAX);
9560 			ASSERT(curneeded != UINT32_MAX);
9561 
9562 			agg->dtag_base = aggbase;
9563 
9564 			curneeded = P2ROUNDUP(curneeded, rec->dtrd_alignment);
9565 			rec->dtrd_offset = curneeded;
9566 			curneeded += rec->dtrd_size;
9567 			ecb->dte_needed = MAX(ecb->dte_needed, curneeded);
9568 
9569 			aggbase = UINT32_MAX;
9570 			curneeded = UINT32_MAX;
9571 		} else if (act->dta_intuple) {
9572 			if (curneeded == UINT32_MAX) {
9573 				/*
9574 				 * This is the first record in a tuple.  Align
9575 				 * curneeded to be at offset 4 in an 8-byte
9576 				 * aligned block.
9577 				 */
9578 				ASSERT(act->dta_prev == NULL ||
9579 				    !act->dta_prev->dta_intuple);
9580 				ASSERT3U(aggbase, ==, UINT32_MAX);
9581 				curneeded = P2PHASEUP(ecb->dte_size,
9582 				    sizeof (uint64_t), sizeof (dtrace_aggid_t));
9583 
9584 				aggbase = curneeded - sizeof (dtrace_aggid_t);
9585 				ASSERT(IS_P2ALIGNED(aggbase,
9586 				    sizeof (uint64_t)));
9587 			}
9588 			curneeded = P2ROUNDUP(curneeded, rec->dtrd_alignment);
9589 			rec->dtrd_offset = curneeded;
9590 			curneeded += rec->dtrd_size;
9591 		} else {
9592 			/* tuples must be followed by an aggregation */
9593 			ASSERT(act->dta_prev == NULL ||
9594 			    !act->dta_prev->dta_intuple);
9595 
9596 			ecb->dte_size = P2ROUNDUP(ecb->dte_size,
9597 			    rec->dtrd_alignment);
9598 			rec->dtrd_offset = ecb->dte_size;
9599 			ecb->dte_size += rec->dtrd_size;
9600 			ecb->dte_needed = MAX(ecb->dte_needed, ecb->dte_size);
9601 		}
9602 	}
9603 
9604 	if ((act = ecb->dte_action) != NULL &&
9605 	    !(act->dta_kind == DTRACEACT_SPECULATE && act->dta_next == NULL) &&
9606 	    ecb->dte_size == sizeof (dtrace_rechdr_t)) {
9607 		/*
9608 		 * If the size is still sizeof (dtrace_rechdr_t), then all
9609 		 * actions store no data; set the size to 0.
9610 		 */
9611 		ecb->dte_size = 0;
9612 	}
9613 
9614 	ecb->dte_size = P2ROUNDUP(ecb->dte_size, sizeof (dtrace_epid_t));
9615 	ecb->dte_needed = P2ROUNDUP(ecb->dte_needed, (sizeof (dtrace_epid_t)));
9616 	ecb->dte_state->dts_needed = MAX(ecb->dte_state->dts_needed,
9617 	    ecb->dte_needed);
9618 }
9619 
9620 static dtrace_action_t *
9621 dtrace_ecb_aggregation_create(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
9622 {
9623 	dtrace_aggregation_t *agg;
9624 	size_t size = sizeof (uint64_t);
9625 	int ntuple = desc->dtad_ntuple;
9626 	dtrace_action_t *act;
9627 	dtrace_recdesc_t *frec;
9628 	dtrace_aggid_t aggid;
9629 	dtrace_state_t *state = ecb->dte_state;
9630 
9631 	agg = kmem_zalloc(sizeof (dtrace_aggregation_t), KM_SLEEP);
9632 	agg->dtag_ecb = ecb;
9633 
9634 	ASSERT(DTRACEACT_ISAGG(desc->dtad_kind));
9635 
9636 	switch (desc->dtad_kind) {
9637 	case DTRACEAGG_MIN:
9638 		agg->dtag_initial = INT64_MAX;
9639 		agg->dtag_aggregate = dtrace_aggregate_min;
9640 		break;
9641 
9642 	case DTRACEAGG_MAX:
9643 		agg->dtag_initial = INT64_MIN;
9644 		agg->dtag_aggregate = dtrace_aggregate_max;
9645 		break;
9646 
9647 	case DTRACEAGG_COUNT:
9648 		agg->dtag_aggregate = dtrace_aggregate_count;
9649 		break;
9650 
9651 	case DTRACEAGG_QUANTIZE:
9652 		agg->dtag_aggregate = dtrace_aggregate_quantize;
9653 		size = (((sizeof (uint64_t) * NBBY) - 1) * 2 + 1) *
9654 		    sizeof (uint64_t);
9655 		break;
9656 
9657 	case DTRACEAGG_LQUANTIZE: {
9658 		uint16_t step = DTRACE_LQUANTIZE_STEP(desc->dtad_arg);
9659 		uint16_t levels = DTRACE_LQUANTIZE_LEVELS(desc->dtad_arg);
9660 
9661 		agg->dtag_initial = desc->dtad_arg;
9662 		agg->dtag_aggregate = dtrace_aggregate_lquantize;
9663 
9664 		if (step == 0 || levels == 0)
9665 			goto err;
9666 
9667 		size = levels * sizeof (uint64_t) + 3 * sizeof (uint64_t);
9668 		break;
9669 	}
9670 
9671 	case DTRACEAGG_LLQUANTIZE: {
9672 		uint16_t factor = DTRACE_LLQUANTIZE_FACTOR(desc->dtad_arg);
9673 		uint16_t low = DTRACE_LLQUANTIZE_LOW(desc->dtad_arg);
9674 		uint16_t high = DTRACE_LLQUANTIZE_HIGH(desc->dtad_arg);
9675 		uint16_t nsteps = DTRACE_LLQUANTIZE_NSTEP(desc->dtad_arg);
9676 		int64_t v;
9677 
9678 		agg->dtag_initial = desc->dtad_arg;
9679 		agg->dtag_aggregate = dtrace_aggregate_llquantize;
9680 
9681 		if (factor < 2 || low >= high || nsteps < factor)
9682 			goto err;
9683 
9684 		/*
9685 		 * Now check that the number of steps evenly divides a power
9686 		 * of the factor.  (This assures both integer bucket size and
9687 		 * linearity within each magnitude.)
9688 		 */
9689 		for (v = factor; v < nsteps; v *= factor)
9690 			continue;
9691 
9692 		if ((v % nsteps) || (nsteps % factor))
9693 			goto err;
9694 
9695 		size = (dtrace_aggregate_llquantize_bucket(factor,
9696 		    low, high, nsteps, INT64_MAX) + 2) * sizeof (uint64_t);
9697 		break;
9698 	}
9699 
9700 	case DTRACEAGG_AVG:
9701 		agg->dtag_aggregate = dtrace_aggregate_avg;
9702 		size = sizeof (uint64_t) * 2;
9703 		break;
9704 
9705 	case DTRACEAGG_STDDEV:
9706 		agg->dtag_aggregate = dtrace_aggregate_stddev;
9707 		size = sizeof (uint64_t) * 4;
9708 		break;
9709 
9710 	case DTRACEAGG_SUM:
9711 		agg->dtag_aggregate = dtrace_aggregate_sum;
9712 		break;
9713 
9714 	default:
9715 		goto err;
9716 	}
9717 
9718 	agg->dtag_action.dta_rec.dtrd_size = size;
9719 
9720 	if (ntuple == 0)
9721 		goto err;
9722 
9723 	/*
9724 	 * We must make sure that we have enough actions for the n-tuple.
9725 	 */
9726 	for (act = ecb->dte_action_last; act != NULL; act = act->dta_prev) {
9727 		if (DTRACEACT_ISAGG(act->dta_kind))
9728 			break;
9729 
9730 		if (--ntuple == 0) {
9731 			/*
9732 			 * This is the action with which our n-tuple begins.
9733 			 */
9734 			agg->dtag_first = act;
9735 			goto success;
9736 		}
9737 	}
9738 
9739 	/*
9740 	 * This n-tuple is short by ntuple elements.  Return failure.
9741 	 */
9742 	ASSERT(ntuple != 0);
9743 err:
9744 	kmem_free(agg, sizeof (dtrace_aggregation_t));
9745 	return (NULL);
9746 
9747 success:
9748 	/*
9749 	 * If the last action in the tuple has a size of zero, it's actually
9750 	 * an expression argument for the aggregating action.
9751 	 */
9752 	ASSERT(ecb->dte_action_last != NULL);
9753 	act = ecb->dte_action_last;
9754 
9755 	if (act->dta_kind == DTRACEACT_DIFEXPR) {
9756 		ASSERT(act->dta_difo != NULL);
9757 
9758 		if (act->dta_difo->dtdo_rtype.dtdt_size == 0)
9759 			agg->dtag_hasarg = 1;
9760 	}
9761 
9762 	/*
9763 	 * We need to allocate an id for this aggregation.
9764 	 */
9765 	aggid = (dtrace_aggid_t)(uintptr_t)vmem_alloc(state->dts_aggid_arena, 1,
9766 	    VM_BESTFIT | VM_SLEEP);
9767 
9768 	if (aggid - 1 >= state->dts_naggregations) {
9769 		dtrace_aggregation_t **oaggs = state->dts_aggregations;
9770 		dtrace_aggregation_t **aggs;
9771 		int naggs = state->dts_naggregations << 1;
9772 		int onaggs = state->dts_naggregations;
9773 
9774 		ASSERT(aggid == state->dts_naggregations + 1);
9775 
9776 		if (naggs == 0) {
9777 			ASSERT(oaggs == NULL);
9778 			naggs = 1;
9779 		}
9780 
9781 		aggs = kmem_zalloc(naggs * sizeof (*aggs), KM_SLEEP);
9782 
9783 		if (oaggs != NULL) {
9784 			bcopy(oaggs, aggs, onaggs * sizeof (*aggs));
9785 			kmem_free(oaggs, onaggs * sizeof (*aggs));
9786 		}
9787 
9788 		state->dts_aggregations = aggs;
9789 		state->dts_naggregations = naggs;
9790 	}
9791 
9792 	ASSERT(state->dts_aggregations[aggid - 1] == NULL);
9793 	state->dts_aggregations[(agg->dtag_id = aggid) - 1] = agg;
9794 
9795 	frec = &agg->dtag_first->dta_rec;
9796 	if (frec->dtrd_alignment < sizeof (dtrace_aggid_t))
9797 		frec->dtrd_alignment = sizeof (dtrace_aggid_t);
9798 
9799 	for (act = agg->dtag_first; act != NULL; act = act->dta_next) {
9800 		ASSERT(!act->dta_intuple);
9801 		act->dta_intuple = 1;
9802 	}
9803 
9804 	return (&agg->dtag_action);
9805 }
9806 
9807 static void
9808 dtrace_ecb_aggregation_destroy(dtrace_ecb_t *ecb, dtrace_action_t *act)
9809 {
9810 	dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
9811 	dtrace_state_t *state = ecb->dte_state;
9812 	dtrace_aggid_t aggid = agg->dtag_id;
9813 
9814 	ASSERT(DTRACEACT_ISAGG(act->dta_kind));
9815 	vmem_free(state->dts_aggid_arena, (void *)(uintptr_t)aggid, 1);
9816 
9817 	ASSERT(state->dts_aggregations[aggid - 1] == agg);
9818 	state->dts_aggregations[aggid - 1] = NULL;
9819 
9820 	kmem_free(agg, sizeof (dtrace_aggregation_t));
9821 }
9822 
9823 static int
9824 dtrace_ecb_action_add(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
9825 {
9826 	dtrace_action_t *action, *last;
9827 	dtrace_difo_t *dp = desc->dtad_difo;
9828 	uint32_t size = 0, align = sizeof (uint8_t), mask;
9829 	uint16_t format = 0;
9830 	dtrace_recdesc_t *rec;
9831 	dtrace_state_t *state = ecb->dte_state;
9832 	dtrace_optval_t *opt = state->dts_options, nframes, strsize;
9833 	uint64_t arg = desc->dtad_arg;
9834 
9835 	ASSERT(MUTEX_HELD(&dtrace_lock));
9836 	ASSERT(ecb->dte_action == NULL || ecb->dte_action->dta_refcnt == 1);
9837 
9838 	if (DTRACEACT_ISAGG(desc->dtad_kind)) {
9839 		/*
9840 		 * If this is an aggregating action, there must be neither
9841 		 * a speculate nor a commit on the action chain.
9842 		 */
9843 		dtrace_action_t *act;
9844 
9845 		for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
9846 			if (act->dta_kind == DTRACEACT_COMMIT)
9847 				return (EINVAL);
9848 
9849 			if (act->dta_kind == DTRACEACT_SPECULATE)
9850 				return (EINVAL);
9851 		}
9852 
9853 		action = dtrace_ecb_aggregation_create(ecb, desc);
9854 
9855 		if (action == NULL)
9856 			return (EINVAL);
9857 	} else {
9858 		if (DTRACEACT_ISDESTRUCTIVE(desc->dtad_kind) ||
9859 		    (desc->dtad_kind == DTRACEACT_DIFEXPR &&
9860 		    dp != NULL && dp->dtdo_destructive)) {
9861 			state->dts_destructive = 1;
9862 		}
9863 
9864 		switch (desc->dtad_kind) {
9865 		case DTRACEACT_PRINTF:
9866 		case DTRACEACT_PRINTA:
9867 		case DTRACEACT_SYSTEM:
9868 		case DTRACEACT_FREOPEN:
9869 		case DTRACEACT_DIFEXPR:
9870 			/*
9871 			 * We know that our arg is a string -- turn it into a
9872 			 * format.
9873 			 */
9874 			if (arg == NULL) {
9875 				ASSERT(desc->dtad_kind == DTRACEACT_PRINTA ||
9876 				    desc->dtad_kind == DTRACEACT_DIFEXPR);
9877 				format = 0;
9878 			} else {
9879 				ASSERT(arg != NULL);
9880 				ASSERT(arg > KERNELBASE);
9881 				format = dtrace_format_add(state,
9882 				    (char *)(uintptr_t)arg);
9883 			}
9884 
9885 			/*FALLTHROUGH*/
9886 		case DTRACEACT_LIBACT:
9887 		case DTRACEACT_TRACEMEM:
9888 		case DTRACEACT_TRACEMEM_DYNSIZE:
9889 			if (dp == NULL)
9890 				return (EINVAL);
9891 
9892 			if ((size = dp->dtdo_rtype.dtdt_size) != 0)
9893 				break;
9894 
9895 			if (dp->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING) {
9896 				if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
9897 					return (EINVAL);
9898 
9899 				size = opt[DTRACEOPT_STRSIZE];
9900 			}
9901 
9902 			break;
9903 
9904 		case DTRACEACT_STACK:
9905 			if ((nframes = arg) == 0) {
9906 				nframes = opt[DTRACEOPT_STACKFRAMES];
9907 				ASSERT(nframes > 0);
9908 				arg = nframes;
9909 			}
9910 
9911 			size = nframes * sizeof (pc_t);
9912 			break;
9913 
9914 		case DTRACEACT_JSTACK:
9915 			if ((strsize = DTRACE_USTACK_STRSIZE(arg)) == 0)
9916 				strsize = opt[DTRACEOPT_JSTACKSTRSIZE];
9917 
9918 			if ((nframes = DTRACE_USTACK_NFRAMES(arg)) == 0)
9919 				nframes = opt[DTRACEOPT_JSTACKFRAMES];
9920 
9921 			arg = DTRACE_USTACK_ARG(nframes, strsize);
9922 
9923 			/*FALLTHROUGH*/
9924 		case DTRACEACT_USTACK:
9925 			if (desc->dtad_kind != DTRACEACT_JSTACK &&
9926 			    (nframes = DTRACE_USTACK_NFRAMES(arg)) == 0) {
9927 				strsize = DTRACE_USTACK_STRSIZE(arg);
9928 				nframes = opt[DTRACEOPT_USTACKFRAMES];
9929 				ASSERT(nframes > 0);
9930 				arg = DTRACE_USTACK_ARG(nframes, strsize);
9931 			}
9932 
9933 			/*
9934 			 * Save a slot for the pid.
9935 			 */
9936 			size = (nframes + 1) * sizeof (uint64_t);
9937 			size += DTRACE_USTACK_STRSIZE(arg);
9938 			size = P2ROUNDUP(size, (uint32_t)(sizeof (uintptr_t)));
9939 
9940 			break;
9941 
9942 		case DTRACEACT_SYM:
9943 		case DTRACEACT_MOD:
9944 			if (dp == NULL || ((size = dp->dtdo_rtype.dtdt_size) !=
9945 			    sizeof (uint64_t)) ||
9946 			    (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
9947 				return (EINVAL);
9948 			break;
9949 
9950 		case DTRACEACT_USYM:
9951 		case DTRACEACT_UMOD:
9952 		case DTRACEACT_UADDR:
9953 			if (dp == NULL ||
9954 			    (dp->dtdo_rtype.dtdt_size != sizeof (uint64_t)) ||
9955 			    (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
9956 				return (EINVAL);
9957 
9958 			/*
9959 			 * We have a slot for the pid, plus a slot for the
9960 			 * argument.  To keep things simple (aligned with
9961 			 * bitness-neutral sizing), we store each as a 64-bit
9962 			 * quantity.
9963 			 */
9964 			size = 2 * sizeof (uint64_t);
9965 			break;
9966 
9967 		case DTRACEACT_STOP:
9968 		case DTRACEACT_BREAKPOINT:
9969 		case DTRACEACT_PANIC:
9970 			break;
9971 
9972 		case DTRACEACT_CHILL:
9973 		case DTRACEACT_DISCARD:
9974 		case DTRACEACT_RAISE:
9975 			if (dp == NULL)
9976 				return (EINVAL);
9977 			break;
9978 
9979 		case DTRACEACT_EXIT:
9980 			if (dp == NULL ||
9981 			    (size = dp->dtdo_rtype.dtdt_size) != sizeof (int) ||
9982 			    (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
9983 				return (EINVAL);
9984 			break;
9985 
9986 		case DTRACEACT_SPECULATE:
9987 			if (ecb->dte_size > sizeof (dtrace_rechdr_t))
9988 				return (EINVAL);
9989 
9990 			if (dp == NULL)
9991 				return (EINVAL);
9992 
9993 			state->dts_speculates = 1;
9994 			break;
9995 
9996 		case DTRACEACT_COMMIT: {
9997 			dtrace_action_t *act = ecb->dte_action;
9998 
9999 			for (; act != NULL; act = act->dta_next) {
10000 				if (act->dta_kind == DTRACEACT_COMMIT)
10001 					return (EINVAL);
10002 			}
10003 
10004 			if (dp == NULL)
10005 				return (EINVAL);
10006 			break;
10007 		}
10008 
10009 		default:
10010 			return (EINVAL);
10011 		}
10012 
10013 		if (size != 0 || desc->dtad_kind == DTRACEACT_SPECULATE) {
10014 			/*
10015 			 * If this is a data-storing action or a speculate,
10016 			 * we must be sure that there isn't a commit on the
10017 			 * action chain.
10018 			 */
10019 			dtrace_action_t *act = ecb->dte_action;
10020 
10021 			for (; act != NULL; act = act->dta_next) {
10022 				if (act->dta_kind == DTRACEACT_COMMIT)
10023 					return (EINVAL);
10024 			}
10025 		}
10026 
10027 		action = kmem_zalloc(sizeof (dtrace_action_t), KM_SLEEP);
10028 		action->dta_rec.dtrd_size = size;
10029 	}
10030 
10031 	action->dta_refcnt = 1;
10032 	rec = &action->dta_rec;
10033 	size = rec->dtrd_size;
10034 
10035 	for (mask = sizeof (uint64_t) - 1; size != 0 && mask > 0; mask >>= 1) {
10036 		if (!(size & mask)) {
10037 			align = mask + 1;
10038 			break;
10039 		}
10040 	}
10041 
10042 	action->dta_kind = desc->dtad_kind;
10043 
10044 	if ((action->dta_difo = dp) != NULL)
10045 		dtrace_difo_hold(dp);
10046 
10047 	rec->dtrd_action = action->dta_kind;
10048 	rec->dtrd_arg = arg;
10049 	rec->dtrd_uarg = desc->dtad_uarg;
10050 	rec->dtrd_alignment = (uint16_t)align;
10051 	rec->dtrd_format = format;
10052 
10053 	if ((last = ecb->dte_action_last) != NULL) {
10054 		ASSERT(ecb->dte_action != NULL);
10055 		action->dta_prev = last;
10056 		last->dta_next = action;
10057 	} else {
10058 		ASSERT(ecb->dte_action == NULL);
10059 		ecb->dte_action = action;
10060 	}
10061 
10062 	ecb->dte_action_last = action;
10063 
10064 	return (0);
10065 }
10066 
10067 static void
10068 dtrace_ecb_action_remove(dtrace_ecb_t *ecb)
10069 {
10070 	dtrace_action_t *act = ecb->dte_action, *next;
10071 	dtrace_vstate_t *vstate = &ecb->dte_state->dts_vstate;
10072 	dtrace_difo_t *dp;
10073 	uint16_t format;
10074 
10075 	if (act != NULL && act->dta_refcnt > 1) {
10076 		ASSERT(act->dta_next == NULL || act->dta_next->dta_refcnt == 1);
10077 		act->dta_refcnt--;
10078 	} else {
10079 		for (; act != NULL; act = next) {
10080 			next = act->dta_next;
10081 			ASSERT(next != NULL || act == ecb->dte_action_last);
10082 			ASSERT(act->dta_refcnt == 1);
10083 
10084 			if ((format = act->dta_rec.dtrd_format) != 0)
10085 				dtrace_format_remove(ecb->dte_state, format);
10086 
10087 			if ((dp = act->dta_difo) != NULL)
10088 				dtrace_difo_release(dp, vstate);
10089 
10090 			if (DTRACEACT_ISAGG(act->dta_kind)) {
10091 				dtrace_ecb_aggregation_destroy(ecb, act);
10092 			} else {
10093 				kmem_free(act, sizeof (dtrace_action_t));
10094 			}
10095 		}
10096 	}
10097 
10098 	ecb->dte_action = NULL;
10099 	ecb->dte_action_last = NULL;
10100 	ecb->dte_size = 0;
10101 }
10102 
10103 static void
10104 dtrace_ecb_disable(dtrace_ecb_t *ecb)
10105 {
10106 	/*
10107 	 * We disable the ECB by removing it from its probe.
10108 	 */
10109 	dtrace_ecb_t *pecb, *prev = NULL;
10110 	dtrace_probe_t *probe = ecb->dte_probe;
10111 
10112 	ASSERT(MUTEX_HELD(&dtrace_lock));
10113 
10114 	if (probe == NULL) {
10115 		/*
10116 		 * This is the NULL probe; there is nothing to disable.
10117 		 */
10118 		return;
10119 	}
10120 
10121 	for (pecb = probe->dtpr_ecb; pecb != NULL; pecb = pecb->dte_next) {
10122 		if (pecb == ecb)
10123 			break;
10124 		prev = pecb;
10125 	}
10126 
10127 	ASSERT(pecb != NULL);
10128 
10129 	if (prev == NULL) {
10130 		probe->dtpr_ecb = ecb->dte_next;
10131 	} else {
10132 		prev->dte_next = ecb->dte_next;
10133 	}
10134 
10135 	if (ecb == probe->dtpr_ecb_last) {
10136 		ASSERT(ecb->dte_next == NULL);
10137 		probe->dtpr_ecb_last = prev;
10138 	}
10139 
10140 	/*
10141 	 * The ECB has been disconnected from the probe; now sync to assure
10142 	 * that all CPUs have seen the change before returning.
10143 	 */
10144 	dtrace_sync();
10145 
10146 	if (probe->dtpr_ecb == NULL) {
10147 		/*
10148 		 * That was the last ECB on the probe; clear the predicate
10149 		 * cache ID for the probe, disable it and sync one more time
10150 		 * to assure that we'll never hit it again.
10151 		 */
10152 		dtrace_provider_t *prov = probe->dtpr_provider;
10153 
10154 		ASSERT(ecb->dte_next == NULL);
10155 		ASSERT(probe->dtpr_ecb_last == NULL);
10156 		probe->dtpr_predcache = DTRACE_CACHEIDNONE;
10157 		prov->dtpv_pops.dtps_disable(prov->dtpv_arg,
10158 		    probe->dtpr_id, probe->dtpr_arg);
10159 		dtrace_sync();
10160 	} else {
10161 		/*
10162 		 * There is at least one ECB remaining on the probe.  If there
10163 		 * is _exactly_ one, set the probe's predicate cache ID to be
10164 		 * the predicate cache ID of the remaining ECB.
10165 		 */
10166 		ASSERT(probe->dtpr_ecb_last != NULL);
10167 		ASSERT(probe->dtpr_predcache == DTRACE_CACHEIDNONE);
10168 
10169 		if (probe->dtpr_ecb == probe->dtpr_ecb_last) {
10170 			dtrace_predicate_t *p = probe->dtpr_ecb->dte_predicate;
10171 
10172 			ASSERT(probe->dtpr_ecb->dte_next == NULL);
10173 
10174 			if (p != NULL)
10175 				probe->dtpr_predcache = p->dtp_cacheid;
10176 		}
10177 
10178 		ecb->dte_next = NULL;
10179 	}
10180 }
10181 
10182 static void
10183 dtrace_ecb_destroy(dtrace_ecb_t *ecb)
10184 {
10185 	dtrace_state_t *state = ecb->dte_state;
10186 	dtrace_vstate_t *vstate = &state->dts_vstate;
10187 	dtrace_predicate_t *pred;
10188 	dtrace_epid_t epid = ecb->dte_epid;
10189 
10190 	ASSERT(MUTEX_HELD(&dtrace_lock));
10191 	ASSERT(ecb->dte_next == NULL);
10192 	ASSERT(ecb->dte_probe == NULL || ecb->dte_probe->dtpr_ecb != ecb);
10193 
10194 	if ((pred = ecb->dte_predicate) != NULL)
10195 		dtrace_predicate_release(pred, vstate);
10196 
10197 	dtrace_ecb_action_remove(ecb);
10198 
10199 	ASSERT(state->dts_ecbs[epid - 1] == ecb);
10200 	state->dts_ecbs[epid - 1] = NULL;
10201 
10202 	kmem_free(ecb, sizeof (dtrace_ecb_t));
10203 }
10204 
10205 static dtrace_ecb_t *
10206 dtrace_ecb_create(dtrace_state_t *state, dtrace_probe_t *probe,
10207     dtrace_enabling_t *enab)
10208 {
10209 	dtrace_ecb_t *ecb;
10210 	dtrace_predicate_t *pred;
10211 	dtrace_actdesc_t *act;
10212 	dtrace_provider_t *prov;
10213 	dtrace_ecbdesc_t *desc = enab->dten_current;
10214 
10215 	ASSERT(MUTEX_HELD(&dtrace_lock));
10216 	ASSERT(state != NULL);
10217 
10218 	ecb = dtrace_ecb_add(state, probe);
10219 	ecb->dte_uarg = desc->dted_uarg;
10220 
10221 	if ((pred = desc->dted_pred.dtpdd_predicate) != NULL) {
10222 		dtrace_predicate_hold(pred);
10223 		ecb->dte_predicate = pred;
10224 	}
10225 
10226 	if (probe != NULL) {
10227 		/*
10228 		 * If the provider shows more leg than the consumer is old
10229 		 * enough to see, we need to enable the appropriate implicit
10230 		 * predicate bits to prevent the ecb from activating at
10231 		 * revealing times.
10232 		 *
10233 		 * Providers specifying DTRACE_PRIV_USER at register time
10234 		 * are stating that they need the /proc-style privilege
10235 		 * model to be enforced, and this is what DTRACE_COND_OWNER
10236 		 * and DTRACE_COND_ZONEOWNER will then do at probe time.
10237 		 */
10238 		prov = probe->dtpr_provider;
10239 		if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLPROC) &&
10240 		    (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
10241 			ecb->dte_cond |= DTRACE_COND_OWNER;
10242 
10243 		if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLZONE) &&
10244 		    (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
10245 			ecb->dte_cond |= DTRACE_COND_ZONEOWNER;
10246 
10247 		/*
10248 		 * If the provider shows us kernel innards and the user
10249 		 * is lacking sufficient privilege, enable the
10250 		 * DTRACE_COND_USERMODE implicit predicate.
10251 		 */
10252 		if (!(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) &&
10253 		    (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_KERNEL))
10254 			ecb->dte_cond |= DTRACE_COND_USERMODE;
10255 	}
10256 
10257 	if (dtrace_ecb_create_cache != NULL) {
10258 		/*
10259 		 * If we have a cached ecb, we'll use its action list instead
10260 		 * of creating our own (saving both time and space).
10261 		 */
10262 		dtrace_ecb_t *cached = dtrace_ecb_create_cache;
10263 		dtrace_action_t *act = cached->dte_action;
10264 
10265 		if (act != NULL) {
10266 			ASSERT(act->dta_refcnt > 0);
10267 			act->dta_refcnt++;
10268 			ecb->dte_action = act;
10269 			ecb->dte_action_last = cached->dte_action_last;
10270 			ecb->dte_needed = cached->dte_needed;
10271 			ecb->dte_size = cached->dte_size;
10272 			ecb->dte_alignment = cached->dte_alignment;
10273 		}
10274 
10275 		return (ecb);
10276 	}
10277 
10278 	for (act = desc->dted_action; act != NULL; act = act->dtad_next) {
10279 		if ((enab->dten_error = dtrace_ecb_action_add(ecb, act)) != 0) {
10280 			dtrace_ecb_destroy(ecb);
10281 			return (NULL);
10282 		}
10283 	}
10284 
10285 	dtrace_ecb_resize(ecb);
10286 
10287 	return (dtrace_ecb_create_cache = ecb);
10288 }
10289 
10290 static int
10291 dtrace_ecb_create_enable(dtrace_probe_t *probe, void *arg)
10292 {
10293 	dtrace_ecb_t *ecb;
10294 	dtrace_enabling_t *enab = arg;
10295 	dtrace_state_t *state = enab->dten_vstate->dtvs_state;
10296 
10297 	ASSERT(state != NULL);
10298 
10299 	if (probe != NULL && probe->dtpr_gen < enab->dten_probegen) {
10300 		/*
10301 		 * This probe was created in a generation for which this
10302 		 * enabling has previously created ECBs; we don't want to
10303 		 * enable it again, so just kick out.
10304 		 */
10305 		return (DTRACE_MATCH_NEXT);
10306 	}
10307 
10308 	if ((ecb = dtrace_ecb_create(state, probe, enab)) == NULL)
10309 		return (DTRACE_MATCH_DONE);
10310 
10311 	if (dtrace_ecb_enable(ecb) < 0)
10312 		return (DTRACE_MATCH_FAIL);
10313 
10314 	return (DTRACE_MATCH_NEXT);
10315 }
10316 
10317 static dtrace_ecb_t *
10318 dtrace_epid2ecb(dtrace_state_t *state, dtrace_epid_t id)
10319 {
10320 	dtrace_ecb_t *ecb;
10321 
10322 	ASSERT(MUTEX_HELD(&dtrace_lock));
10323 
10324 	if (id == 0 || id > state->dts_necbs)
10325 		return (NULL);
10326 
10327 	ASSERT(state->dts_necbs > 0 && state->dts_ecbs != NULL);
10328 	ASSERT((ecb = state->dts_ecbs[id - 1]) == NULL || ecb->dte_epid == id);
10329 
10330 	return (state->dts_ecbs[id - 1]);
10331 }
10332 
10333 static dtrace_aggregation_t *
10334 dtrace_aggid2agg(dtrace_state_t *state, dtrace_aggid_t id)
10335 {
10336 	dtrace_aggregation_t *agg;
10337 
10338 	ASSERT(MUTEX_HELD(&dtrace_lock));
10339 
10340 	if (id == 0 || id > state->dts_naggregations)
10341 		return (NULL);
10342 
10343 	ASSERT(state->dts_naggregations > 0 && state->dts_aggregations != NULL);
10344 	ASSERT((agg = state->dts_aggregations[id - 1]) == NULL ||
10345 	    agg->dtag_id == id);
10346 
10347 	return (state->dts_aggregations[id - 1]);
10348 }
10349 
10350 /*
10351  * DTrace Buffer Functions
10352  *
10353  * The following functions manipulate DTrace buffers.  Most of these functions
10354  * are called in the context of establishing or processing consumer state;
10355  * exceptions are explicitly noted.
10356  */
10357 
10358 /*
10359  * Note:  called from cross call context.  This function switches the two
10360  * buffers on a given CPU.  The atomicity of this operation is assured by
10361  * disabling interrupts while the actual switch takes place; the disabling of
10362  * interrupts serializes the execution with any execution of dtrace_probe() on
10363  * the same CPU.
10364  */
10365 static void
10366 dtrace_buffer_switch(dtrace_buffer_t *buf)
10367 {
10368 	caddr_t tomax = buf->dtb_tomax;
10369 	caddr_t xamot = buf->dtb_xamot;
10370 	dtrace_icookie_t cookie;
10371 	hrtime_t now;
10372 
10373 	ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
10374 	ASSERT(!(buf->dtb_flags & DTRACEBUF_RING));
10375 
10376 	cookie = dtrace_interrupt_disable();
10377 	now = dtrace_gethrtime();
10378 	buf->dtb_tomax = xamot;
10379 	buf->dtb_xamot = tomax;
10380 	buf->dtb_xamot_drops = buf->dtb_drops;
10381 	buf->dtb_xamot_offset = buf->dtb_offset;
10382 	buf->dtb_xamot_errors = buf->dtb_errors;
10383 	buf->dtb_xamot_flags = buf->dtb_flags;
10384 	buf->dtb_offset = 0;
10385 	buf->dtb_drops = 0;
10386 	buf->dtb_errors = 0;
10387 	buf->dtb_flags &= ~(DTRACEBUF_ERROR | DTRACEBUF_DROPPED);
10388 	buf->dtb_interval = now - buf->dtb_switched;
10389 	buf->dtb_switched = now;
10390 	dtrace_interrupt_enable(cookie);
10391 }
10392 
10393 /*
10394  * Note:  called from cross call context.  This function activates a buffer
10395  * on a CPU.  As with dtrace_buffer_switch(), the atomicity of the operation
10396  * is guaranteed by the disabling of interrupts.
10397  */
10398 static void
10399 dtrace_buffer_activate(dtrace_state_t *state)
10400 {
10401 	dtrace_buffer_t *buf;
10402 	dtrace_icookie_t cookie = dtrace_interrupt_disable();
10403 
10404 	buf = &state->dts_buffer[CPU->cpu_id];
10405 
10406 	if (buf->dtb_tomax != NULL) {
10407 		/*
10408 		 * We might like to assert that the buffer is marked inactive,
10409 		 * but this isn't necessarily true:  the buffer for the CPU
10410 		 * that processes the BEGIN probe has its buffer activated
10411 		 * manually.  In this case, we take the (harmless) action
10412 		 * re-clearing the bit INACTIVE bit.
10413 		 */
10414 		buf->dtb_flags &= ~DTRACEBUF_INACTIVE;
10415 	}
10416 
10417 	dtrace_interrupt_enable(cookie);
10418 }
10419 
10420 static int
10421 dtrace_buffer_alloc(dtrace_buffer_t *bufs, size_t size, int flags,
10422     processorid_t cpu, int *factor)
10423 {
10424 	cpu_t *cp;
10425 	dtrace_buffer_t *buf;
10426 	int allocated = 0, desired = 0;
10427 
10428 	ASSERT(MUTEX_HELD(&cpu_lock));
10429 	ASSERT(MUTEX_HELD(&dtrace_lock));
10430 
10431 	*factor = 1;
10432 
10433 	if (size > dtrace_nonroot_maxsize &&
10434 	    !PRIV_POLICY_CHOICE(CRED(), PRIV_ALL, B_FALSE))
10435 		return (EFBIG);
10436 
10437 	cp = cpu_list;
10438 
10439 	do {
10440 		if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
10441 			continue;
10442 
10443 		buf = &bufs[cp->cpu_id];
10444 
10445 		/*
10446 		 * If there is already a buffer allocated for this CPU, it
10447 		 * is only possible that this is a DR event.  In this case,
10448 		 * the buffer size must match our specified size.
10449 		 */
10450 		if (buf->dtb_tomax != NULL) {
10451 			ASSERT(buf->dtb_size == size);
10452 			continue;
10453 		}
10454 
10455 		ASSERT(buf->dtb_xamot == NULL);
10456 
10457 		if ((buf->dtb_tomax = kmem_zalloc(size,
10458 		    KM_NOSLEEP | KM_NORMALPRI)) == NULL)
10459 			goto err;
10460 
10461 		buf->dtb_size = size;
10462 		buf->dtb_flags = flags;
10463 		buf->dtb_offset = 0;
10464 		buf->dtb_drops = 0;
10465 
10466 		if (flags & DTRACEBUF_NOSWITCH)
10467 			continue;
10468 
10469 		if ((buf->dtb_xamot = kmem_zalloc(size,
10470 		    KM_NOSLEEP | KM_NORMALPRI)) == NULL)
10471 			goto err;
10472 	} while ((cp = cp->cpu_next) != cpu_list);
10473 
10474 	return (0);
10475 
10476 err:
10477 	cp = cpu_list;
10478 
10479 	do {
10480 		if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
10481 			continue;
10482 
10483 		buf = &bufs[cp->cpu_id];
10484 		desired += 2;
10485 
10486 		if (buf->dtb_xamot != NULL) {
10487 			ASSERT(buf->dtb_tomax != NULL);
10488 			ASSERT(buf->dtb_size == size);
10489 			kmem_free(buf->dtb_xamot, size);
10490 			allocated++;
10491 		}
10492 
10493 		if (buf->dtb_tomax != NULL) {
10494 			ASSERT(buf->dtb_size == size);
10495 			kmem_free(buf->dtb_tomax, size);
10496 			allocated++;
10497 		}
10498 
10499 		buf->dtb_tomax = NULL;
10500 		buf->dtb_xamot = NULL;
10501 		buf->dtb_size = 0;
10502 	} while ((cp = cp->cpu_next) != cpu_list);
10503 
10504 	*factor = desired / (allocated > 0 ? allocated : 1);
10505 
10506 	return (ENOMEM);
10507 }
10508 
10509 /*
10510  * Note:  called from probe context.  This function just increments the drop
10511  * count on a buffer.  It has been made a function to allow for the
10512  * possibility of understanding the source of mysterious drop counts.  (A
10513  * problem for which one may be particularly disappointed that DTrace cannot
10514  * be used to understand DTrace.)
10515  */
10516 static void
10517 dtrace_buffer_drop(dtrace_buffer_t *buf)
10518 {
10519 	buf->dtb_drops++;
10520 }
10521 
10522 /*
10523  * Note:  called from probe context.  This function is called to reserve space
10524  * in a buffer.  If mstate is non-NULL, sets the scratch base and size in the
10525  * mstate.  Returns the new offset in the buffer, or a negative value if an
10526  * error has occurred.
10527  */
10528 static intptr_t
10529 dtrace_buffer_reserve(dtrace_buffer_t *buf, size_t needed, size_t align,
10530     dtrace_state_t *state, dtrace_mstate_t *mstate)
10531 {
10532 	intptr_t offs = buf->dtb_offset, soffs;
10533 	intptr_t woffs;
10534 	caddr_t tomax;
10535 	size_t total;
10536 
10537 	if (buf->dtb_flags & DTRACEBUF_INACTIVE)
10538 		return (-1);
10539 
10540 	if ((tomax = buf->dtb_tomax) == NULL) {
10541 		dtrace_buffer_drop(buf);
10542 		return (-1);
10543 	}
10544 
10545 	if (!(buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL))) {
10546 		while (offs & (align - 1)) {
10547 			/*
10548 			 * Assert that our alignment is off by a number which
10549 			 * is itself sizeof (uint32_t) aligned.
10550 			 */
10551 			ASSERT(!((align - (offs & (align - 1))) &
10552 			    (sizeof (uint32_t) - 1)));
10553 			DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
10554 			offs += sizeof (uint32_t);
10555 		}
10556 
10557 		if ((soffs = offs + needed) > buf->dtb_size) {
10558 			dtrace_buffer_drop(buf);
10559 			return (-1);
10560 		}
10561 
10562 		if (mstate == NULL)
10563 			return (offs);
10564 
10565 		mstate->dtms_scratch_base = (uintptr_t)tomax + soffs;
10566 		mstate->dtms_scratch_size = buf->dtb_size - soffs;
10567 		mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
10568 
10569 		return (offs);
10570 	}
10571 
10572 	if (buf->dtb_flags & DTRACEBUF_FILL) {
10573 		if (state->dts_activity != DTRACE_ACTIVITY_COOLDOWN &&
10574 		    (buf->dtb_flags & DTRACEBUF_FULL))
10575 			return (-1);
10576 		goto out;
10577 	}
10578 
10579 	total = needed + (offs & (align - 1));
10580 
10581 	/*
10582 	 * For a ring buffer, life is quite a bit more complicated.  Before
10583 	 * we can store any padding, we need to adjust our wrapping offset.
10584 	 * (If we've never before wrapped or we're not about to, no adjustment
10585 	 * is required.)
10586 	 */
10587 	if ((buf->dtb_flags & DTRACEBUF_WRAPPED) ||
10588 	    offs + total > buf->dtb_size) {
10589 		woffs = buf->dtb_xamot_offset;
10590 
10591 		if (offs + total > buf->dtb_size) {
10592 			/*
10593 			 * We can't fit in the end of the buffer.  First, a
10594 			 * sanity check that we can fit in the buffer at all.
10595 			 */
10596 			if (total > buf->dtb_size) {
10597 				dtrace_buffer_drop(buf);
10598 				return (-1);
10599 			}
10600 
10601 			/*
10602 			 * We're going to be storing at the top of the buffer,
10603 			 * so now we need to deal with the wrapped offset.  We
10604 			 * only reset our wrapped offset to 0 if it is
10605 			 * currently greater than the current offset.  If it
10606 			 * is less than the current offset, it is because a
10607 			 * previous allocation induced a wrap -- but the
10608 			 * allocation didn't subsequently take the space due
10609 			 * to an error or false predicate evaluation.  In this
10610 			 * case, we'll just leave the wrapped offset alone: if
10611 			 * the wrapped offset hasn't been advanced far enough
10612 			 * for this allocation, it will be adjusted in the
10613 			 * lower loop.
10614 			 */
10615 			if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
10616 				if (woffs >= offs)
10617 					woffs = 0;
10618 			} else {
10619 				woffs = 0;
10620 			}
10621 
10622 			/*
10623 			 * Now we know that we're going to be storing to the
10624 			 * top of the buffer and that there is room for us
10625 			 * there.  We need to clear the buffer from the current
10626 			 * offset to the end (there may be old gunk there).
10627 			 */
10628 			while (offs < buf->dtb_size)
10629 				tomax[offs++] = 0;
10630 
10631 			/*
10632 			 * We need to set our offset to zero.  And because we
10633 			 * are wrapping, we need to set the bit indicating as
10634 			 * much.  We can also adjust our needed space back
10635 			 * down to the space required by the ECB -- we know
10636 			 * that the top of the buffer is aligned.
10637 			 */
10638 			offs = 0;
10639 			total = needed;
10640 			buf->dtb_flags |= DTRACEBUF_WRAPPED;
10641 		} else {
10642 			/*
10643 			 * There is room for us in the buffer, so we simply
10644 			 * need to check the wrapped offset.
10645 			 */
10646 			if (woffs < offs) {
10647 				/*
10648 				 * The wrapped offset is less than the offset.
10649 				 * This can happen if we allocated buffer space
10650 				 * that induced a wrap, but then we didn't
10651 				 * subsequently take the space due to an error
10652 				 * or false predicate evaluation.  This is
10653 				 * okay; we know that _this_ allocation isn't
10654 				 * going to induce a wrap.  We still can't
10655 				 * reset the wrapped offset to be zero,
10656 				 * however: the space may have been trashed in
10657 				 * the previous failed probe attempt.  But at
10658 				 * least the wrapped offset doesn't need to
10659 				 * be adjusted at all...
10660 				 */
10661 				goto out;
10662 			}
10663 		}
10664 
10665 		while (offs + total > woffs) {
10666 			dtrace_epid_t epid = *(uint32_t *)(tomax + woffs);
10667 			size_t size;
10668 
10669 			if (epid == DTRACE_EPIDNONE) {
10670 				size = sizeof (uint32_t);
10671 			} else {
10672 				ASSERT3U(epid, <=, state->dts_necbs);
10673 				ASSERT(state->dts_ecbs[epid - 1] != NULL);
10674 
10675 				size = state->dts_ecbs[epid - 1]->dte_size;
10676 			}
10677 
10678 			ASSERT(woffs + size <= buf->dtb_size);
10679 			ASSERT(size != 0);
10680 
10681 			if (woffs + size == buf->dtb_size) {
10682 				/*
10683 				 * We've reached the end of the buffer; we want
10684 				 * to set the wrapped offset to 0 and break
10685 				 * out.  However, if the offs is 0, then we're
10686 				 * in a strange edge-condition:  the amount of
10687 				 * space that we want to reserve plus the size
10688 				 * of the record that we're overwriting is
10689 				 * greater than the size of the buffer.  This
10690 				 * is problematic because if we reserve the
10691 				 * space but subsequently don't consume it (due
10692 				 * to a failed predicate or error) the wrapped
10693 				 * offset will be 0 -- yet the EPID at offset 0
10694 				 * will not be committed.  This situation is
10695 				 * relatively easy to deal with:  if we're in
10696 				 * this case, the buffer is indistinguishable
10697 				 * from one that hasn't wrapped; we need only
10698 				 * finish the job by clearing the wrapped bit,
10699 				 * explicitly setting the offset to be 0, and
10700 				 * zero'ing out the old data in the buffer.
10701 				 */
10702 				if (offs == 0) {
10703 					buf->dtb_flags &= ~DTRACEBUF_WRAPPED;
10704 					buf->dtb_offset = 0;
10705 					woffs = total;
10706 
10707 					while (woffs < buf->dtb_size)
10708 						tomax[woffs++] = 0;
10709 				}
10710 
10711 				woffs = 0;
10712 				break;
10713 			}
10714 
10715 			woffs += size;
10716 		}
10717 
10718 		/*
10719 		 * We have a wrapped offset.  It may be that the wrapped offset
10720 		 * has become zero -- that's okay.
10721 		 */
10722 		buf->dtb_xamot_offset = woffs;
10723 	}
10724 
10725 out:
10726 	/*
10727 	 * Now we can plow the buffer with any necessary padding.
10728 	 */
10729 	while (offs & (align - 1)) {
10730 		/*
10731 		 * Assert that our alignment is off by a number which
10732 		 * is itself sizeof (uint32_t) aligned.
10733 		 */
10734 		ASSERT(!((align - (offs & (align - 1))) &
10735 		    (sizeof (uint32_t) - 1)));
10736 		DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
10737 		offs += sizeof (uint32_t);
10738 	}
10739 
10740 	if (buf->dtb_flags & DTRACEBUF_FILL) {
10741 		if (offs + needed > buf->dtb_size - state->dts_reserve) {
10742 			buf->dtb_flags |= DTRACEBUF_FULL;
10743 			return (-1);
10744 		}
10745 	}
10746 
10747 	if (mstate == NULL)
10748 		return (offs);
10749 
10750 	/*
10751 	 * For ring buffers and fill buffers, the scratch space is always
10752 	 * the inactive buffer.
10753 	 */
10754 	mstate->dtms_scratch_base = (uintptr_t)buf->dtb_xamot;
10755 	mstate->dtms_scratch_size = buf->dtb_size;
10756 	mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
10757 
10758 	return (offs);
10759 }
10760 
10761 static void
10762 dtrace_buffer_polish(dtrace_buffer_t *buf)
10763 {
10764 	ASSERT(buf->dtb_flags & DTRACEBUF_RING);
10765 	ASSERT(MUTEX_HELD(&dtrace_lock));
10766 
10767 	if (!(buf->dtb_flags & DTRACEBUF_WRAPPED))
10768 		return;
10769 
10770 	/*
10771 	 * We need to polish the ring buffer.  There are three cases:
10772 	 *
10773 	 * - The first (and presumably most common) is that there is no gap
10774 	 *   between the buffer offset and the wrapped offset.  In this case,
10775 	 *   there is nothing in the buffer that isn't valid data; we can
10776 	 *   mark the buffer as polished and return.
10777 	 *
10778 	 * - The second (less common than the first but still more common
10779 	 *   than the third) is that there is a gap between the buffer offset
10780 	 *   and the wrapped offset, and the wrapped offset is larger than the
10781 	 *   buffer offset.  This can happen because of an alignment issue, or
10782 	 *   can happen because of a call to dtrace_buffer_reserve() that
10783 	 *   didn't subsequently consume the buffer space.  In this case,
10784 	 *   we need to zero the data from the buffer offset to the wrapped
10785 	 *   offset.
10786 	 *
10787 	 * - The third (and least common) is that there is a gap between the
10788 	 *   buffer offset and the wrapped offset, but the wrapped offset is
10789 	 *   _less_ than the buffer offset.  This can only happen because a
10790 	 *   call to dtrace_buffer_reserve() induced a wrap, but the space
10791 	 *   was not subsequently consumed.  In this case, we need to zero the
10792 	 *   space from the offset to the end of the buffer _and_ from the
10793 	 *   top of the buffer to the wrapped offset.
10794 	 */
10795 	if (buf->dtb_offset < buf->dtb_xamot_offset) {
10796 		bzero(buf->dtb_tomax + buf->dtb_offset,
10797 		    buf->dtb_xamot_offset - buf->dtb_offset);
10798 	}
10799 
10800 	if (buf->dtb_offset > buf->dtb_xamot_offset) {
10801 		bzero(buf->dtb_tomax + buf->dtb_offset,
10802 		    buf->dtb_size - buf->dtb_offset);
10803 		bzero(buf->dtb_tomax, buf->dtb_xamot_offset);
10804 	}
10805 }
10806 
10807 /*
10808  * This routine determines if data generated at the specified time has likely
10809  * been entirely consumed at user-level.  This routine is called to determine
10810  * if an ECB on a defunct probe (but for an active enabling) can be safely
10811  * disabled and destroyed.
10812  */
10813 static int
10814 dtrace_buffer_consumed(dtrace_buffer_t *bufs, hrtime_t when)
10815 {
10816 	int i;
10817 
10818 	for (i = 0; i < NCPU; i++) {
10819 		dtrace_buffer_t *buf = &bufs[i];
10820 
10821 		if (buf->dtb_size == 0)
10822 			continue;
10823 
10824 		if (buf->dtb_flags & DTRACEBUF_RING)
10825 			return (0);
10826 
10827 		if (!buf->dtb_switched && buf->dtb_offset != 0)
10828 			return (0);
10829 
10830 		if (buf->dtb_switched - buf->dtb_interval < when)
10831 			return (0);
10832 	}
10833 
10834 	return (1);
10835 }
10836 
10837 static void
10838 dtrace_buffer_free(dtrace_buffer_t *bufs)
10839 {
10840 	int i;
10841 
10842 	for (i = 0; i < NCPU; i++) {
10843 		dtrace_buffer_t *buf = &bufs[i];
10844 
10845 		if (buf->dtb_tomax == NULL) {
10846 			ASSERT(buf->dtb_xamot == NULL);
10847 			ASSERT(buf->dtb_size == 0);
10848 			continue;
10849 		}
10850 
10851 		if (buf->dtb_xamot != NULL) {
10852 			ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
10853 			kmem_free(buf->dtb_xamot, buf->dtb_size);
10854 		}
10855 
10856 		kmem_free(buf->dtb_tomax, buf->dtb_size);
10857 		buf->dtb_size = 0;
10858 		buf->dtb_tomax = NULL;
10859 		buf->dtb_xamot = NULL;
10860 	}
10861 }
10862 
10863 /*
10864  * DTrace Enabling Functions
10865  */
10866 static dtrace_enabling_t *
10867 dtrace_enabling_create(dtrace_vstate_t *vstate)
10868 {
10869 	dtrace_enabling_t *enab;
10870 
10871 	enab = kmem_zalloc(sizeof (dtrace_enabling_t), KM_SLEEP);
10872 	enab->dten_vstate = vstate;
10873 
10874 	return (enab);
10875 }
10876 
10877 static void
10878 dtrace_enabling_add(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb)
10879 {
10880 	dtrace_ecbdesc_t **ndesc;
10881 	size_t osize, nsize;
10882 
10883 	/*
10884 	 * We can't add to enablings after we've enabled them, or after we've
10885 	 * retained them.
10886 	 */
10887 	ASSERT(enab->dten_probegen == 0);
10888 	ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
10889 
10890 	if (enab->dten_ndesc < enab->dten_maxdesc) {
10891 		enab->dten_desc[enab->dten_ndesc++] = ecb;
10892 		return;
10893 	}
10894 
10895 	osize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
10896 
10897 	if (enab->dten_maxdesc == 0) {
10898 		enab->dten_maxdesc = 1;
10899 	} else {
10900 		enab->dten_maxdesc <<= 1;
10901 	}
10902 
10903 	ASSERT(enab->dten_ndesc < enab->dten_maxdesc);
10904 
10905 	nsize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
10906 	ndesc = kmem_zalloc(nsize, KM_SLEEP);
10907 	bcopy(enab->dten_desc, ndesc, osize);
10908 	kmem_free(enab->dten_desc, osize);
10909 
10910 	enab->dten_desc = ndesc;
10911 	enab->dten_desc[enab->dten_ndesc++] = ecb;
10912 }
10913 
10914 static void
10915 dtrace_enabling_addlike(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb,
10916     dtrace_probedesc_t *pd)
10917 {
10918 	dtrace_ecbdesc_t *new;
10919 	dtrace_predicate_t *pred;
10920 	dtrace_actdesc_t *act;
10921 
10922 	/*
10923 	 * We're going to create a new ECB description that matches the
10924 	 * specified ECB in every way, but has the specified probe description.
10925 	 */
10926 	new = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
10927 
10928 	if ((pred = ecb->dted_pred.dtpdd_predicate) != NULL)
10929 		dtrace_predicate_hold(pred);
10930 
10931 	for (act = ecb->dted_action; act != NULL; act = act->dtad_next)
10932 		dtrace_actdesc_hold(act);
10933 
10934 	new->dted_action = ecb->dted_action;
10935 	new->dted_pred = ecb->dted_pred;
10936 	new->dted_probe = *pd;
10937 	new->dted_uarg = ecb->dted_uarg;
10938 
10939 	dtrace_enabling_add(enab, new);
10940 }
10941 
10942 static void
10943 dtrace_enabling_dump(dtrace_enabling_t *enab)
10944 {
10945 	int i;
10946 
10947 	for (i = 0; i < enab->dten_ndesc; i++) {
10948 		dtrace_probedesc_t *desc = &enab->dten_desc[i]->dted_probe;
10949 
10950 		cmn_err(CE_NOTE, "enabling probe %d (%s:%s:%s:%s)", i,
10951 		    desc->dtpd_provider, desc->dtpd_mod,
10952 		    desc->dtpd_func, desc->dtpd_name);
10953 	}
10954 }
10955 
10956 static void
10957 dtrace_enabling_destroy(dtrace_enabling_t *enab)
10958 {
10959 	int i;
10960 	dtrace_ecbdesc_t *ep;
10961 	dtrace_vstate_t *vstate = enab->dten_vstate;
10962 
10963 	ASSERT(MUTEX_HELD(&dtrace_lock));
10964 
10965 	for (i = 0; i < enab->dten_ndesc; i++) {
10966 		dtrace_actdesc_t *act, *next;
10967 		dtrace_predicate_t *pred;
10968 
10969 		ep = enab->dten_desc[i];
10970 
10971 		if ((pred = ep->dted_pred.dtpdd_predicate) != NULL)
10972 			dtrace_predicate_release(pred, vstate);
10973 
10974 		for (act = ep->dted_action; act != NULL; act = next) {
10975 			next = act->dtad_next;
10976 			dtrace_actdesc_release(act, vstate);
10977 		}
10978 
10979 		kmem_free(ep, sizeof (dtrace_ecbdesc_t));
10980 	}
10981 
10982 	kmem_free(enab->dten_desc,
10983 	    enab->dten_maxdesc * sizeof (dtrace_enabling_t *));
10984 
10985 	/*
10986 	 * If this was a retained enabling, decrement the dts_nretained count
10987 	 * and take it off of the dtrace_retained list.
10988 	 */
10989 	if (enab->dten_prev != NULL || enab->dten_next != NULL ||
10990 	    dtrace_retained == enab) {
10991 		ASSERT(enab->dten_vstate->dtvs_state != NULL);
10992 		ASSERT(enab->dten_vstate->dtvs_state->dts_nretained > 0);
10993 		enab->dten_vstate->dtvs_state->dts_nretained--;
10994 		dtrace_retained_gen++;
10995 	}
10996 
10997 	if (enab->dten_prev == NULL) {
10998 		if (dtrace_retained == enab) {
10999 			dtrace_retained = enab->dten_next;
11000 
11001 			if (dtrace_retained != NULL)
11002 				dtrace_retained->dten_prev = NULL;
11003 		}
11004 	} else {
11005 		ASSERT(enab != dtrace_retained);
11006 		ASSERT(dtrace_retained != NULL);
11007 		enab->dten_prev->dten_next = enab->dten_next;
11008 	}
11009 
11010 	if (enab->dten_next != NULL) {
11011 		ASSERT(dtrace_retained != NULL);
11012 		enab->dten_next->dten_prev = enab->dten_prev;
11013 	}
11014 
11015 	kmem_free(enab, sizeof (dtrace_enabling_t));
11016 }
11017 
11018 static int
11019 dtrace_enabling_retain(dtrace_enabling_t *enab)
11020 {
11021 	dtrace_state_t *state;
11022 
11023 	ASSERT(MUTEX_HELD(&dtrace_lock));
11024 	ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
11025 	ASSERT(enab->dten_vstate != NULL);
11026 
11027 	state = enab->dten_vstate->dtvs_state;
11028 	ASSERT(state != NULL);
11029 
11030 	/*
11031 	 * We only allow each state to retain dtrace_retain_max enablings.
11032 	 */
11033 	if (state->dts_nretained >= dtrace_retain_max)
11034 		return (ENOSPC);
11035 
11036 	state->dts_nretained++;
11037 	dtrace_retained_gen++;
11038 
11039 	if (dtrace_retained == NULL) {
11040 		dtrace_retained = enab;
11041 		return (0);
11042 	}
11043 
11044 	enab->dten_next = dtrace_retained;
11045 	dtrace_retained->dten_prev = enab;
11046 	dtrace_retained = enab;
11047 
11048 	return (0);
11049 }
11050 
11051 static int
11052 dtrace_enabling_replicate(dtrace_state_t *state, dtrace_probedesc_t *match,
11053     dtrace_probedesc_t *create)
11054 {
11055 	dtrace_enabling_t *new, *enab;
11056 	int found = 0, err = ENOENT;
11057 
11058 	ASSERT(MUTEX_HELD(&dtrace_lock));
11059 	ASSERT(strlen(match->dtpd_provider) < DTRACE_PROVNAMELEN);
11060 	ASSERT(strlen(match->dtpd_mod) < DTRACE_MODNAMELEN);
11061 	ASSERT(strlen(match->dtpd_func) < DTRACE_FUNCNAMELEN);
11062 	ASSERT(strlen(match->dtpd_name) < DTRACE_NAMELEN);
11063 
11064 	new = dtrace_enabling_create(&state->dts_vstate);
11065 
11066 	/*
11067 	 * Iterate over all retained enablings, looking for enablings that
11068 	 * match the specified state.
11069 	 */
11070 	for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11071 		int i;
11072 
11073 		/*
11074 		 * dtvs_state can only be NULL for helper enablings -- and
11075 		 * helper enablings can't be retained.
11076 		 */
11077 		ASSERT(enab->dten_vstate->dtvs_state != NULL);
11078 
11079 		if (enab->dten_vstate->dtvs_state != state)
11080 			continue;
11081 
11082 		/*
11083 		 * Now iterate over each probe description; we're looking for
11084 		 * an exact match to the specified probe description.
11085 		 */
11086 		for (i = 0; i < enab->dten_ndesc; i++) {
11087 			dtrace_ecbdesc_t *ep = enab->dten_desc[i];
11088 			dtrace_probedesc_t *pd = &ep->dted_probe;
11089 
11090 			if (strcmp(pd->dtpd_provider, match->dtpd_provider))
11091 				continue;
11092 
11093 			if (strcmp(pd->dtpd_mod, match->dtpd_mod))
11094 				continue;
11095 
11096 			if (strcmp(pd->dtpd_func, match->dtpd_func))
11097 				continue;
11098 
11099 			if (strcmp(pd->dtpd_name, match->dtpd_name))
11100 				continue;
11101 
11102 			/*
11103 			 * We have a winning probe!  Add it to our growing
11104 			 * enabling.
11105 			 */
11106 			found = 1;
11107 			dtrace_enabling_addlike(new, ep, create);
11108 		}
11109 	}
11110 
11111 	if (!found || (err = dtrace_enabling_retain(new)) != 0) {
11112 		dtrace_enabling_destroy(new);
11113 		return (err);
11114 	}
11115 
11116 	return (0);
11117 }
11118 
11119 static void
11120 dtrace_enabling_retract(dtrace_state_t *state)
11121 {
11122 	dtrace_enabling_t *enab, *next;
11123 
11124 	ASSERT(MUTEX_HELD(&dtrace_lock));
11125 
11126 	/*
11127 	 * Iterate over all retained enablings, destroy the enablings retained
11128 	 * for the specified state.
11129 	 */
11130 	for (enab = dtrace_retained; enab != NULL; enab = next) {
11131 		next = enab->dten_next;
11132 
11133 		/*
11134 		 * dtvs_state can only be NULL for helper enablings -- and
11135 		 * helper enablings can't be retained.
11136 		 */
11137 		ASSERT(enab->dten_vstate->dtvs_state != NULL);
11138 
11139 		if (enab->dten_vstate->dtvs_state == state) {
11140 			ASSERT(state->dts_nretained > 0);
11141 			dtrace_enabling_destroy(enab);
11142 		}
11143 	}
11144 
11145 	ASSERT(state->dts_nretained == 0);
11146 }
11147 
11148 static int
11149 dtrace_enabling_match(dtrace_enabling_t *enab, int *nmatched)
11150 {
11151 	int i = 0;
11152 	int total_matched = 0, matched = 0;
11153 
11154 	ASSERT(MUTEX_HELD(&cpu_lock));
11155 	ASSERT(MUTEX_HELD(&dtrace_lock));
11156 
11157 	for (i = 0; i < enab->dten_ndesc; i++) {
11158 		dtrace_ecbdesc_t *ep = enab->dten_desc[i];
11159 
11160 		enab->dten_current = ep;
11161 		enab->dten_error = 0;
11162 
11163 		/*
11164 		 * If a provider failed to enable a probe then get out and
11165 		 * let the consumer know we failed.
11166 		 */
11167 		if ((matched = dtrace_probe_enable(&ep->dted_probe, enab)) < 0)
11168 			return (EBUSY);
11169 
11170 		total_matched += matched;
11171 
11172 		if (enab->dten_error != 0) {
11173 			/*
11174 			 * If we get an error half-way through enabling the
11175 			 * probes, we kick out -- perhaps with some number of
11176 			 * them enabled.  Leaving enabled probes enabled may
11177 			 * be slightly confusing for user-level, but we expect
11178 			 * that no one will attempt to actually drive on in
11179 			 * the face of such errors.  If this is an anonymous
11180 			 * enabling (indicated with a NULL nmatched pointer),
11181 			 * we cmn_err() a message.  We aren't expecting to
11182 			 * get such an error -- such as it can exist at all,
11183 			 * it would be a result of corrupted DOF in the driver
11184 			 * properties.
11185 			 */
11186 			if (nmatched == NULL) {
11187 				cmn_err(CE_WARN, "dtrace_enabling_match() "
11188 				    "error on %p: %d", (void *)ep,
11189 				    enab->dten_error);
11190 			}
11191 
11192 			return (enab->dten_error);
11193 		}
11194 	}
11195 
11196 	enab->dten_probegen = dtrace_probegen;
11197 	if (nmatched != NULL)
11198 		*nmatched = total_matched;
11199 
11200 	return (0);
11201 }
11202 
11203 static void
11204 dtrace_enabling_matchall(void)
11205 {
11206 	dtrace_enabling_t *enab;
11207 
11208 	mutex_enter(&cpu_lock);
11209 	mutex_enter(&dtrace_lock);
11210 
11211 	/*
11212 	 * Iterate over all retained enablings to see if any probes match
11213 	 * against them.  We only perform this operation on enablings for which
11214 	 * we have sufficient permissions by virtue of being in the global zone
11215 	 * or in the same zone as the DTrace client.  Because we can be called
11216 	 * after dtrace_detach() has been called, we cannot assert that there
11217 	 * are retained enablings.  We can safely load from dtrace_retained,
11218 	 * however:  the taskq_destroy() at the end of dtrace_detach() will
11219 	 * block pending our completion.
11220 	 */
11221 	for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11222 		dtrace_cred_t *dcr = &enab->dten_vstate->dtvs_state->dts_cred;
11223 		cred_t *cr = dcr->dcr_cred;
11224 		zoneid_t zone = cr != NULL ? crgetzoneid(cr) : 0;
11225 
11226 		if ((dcr->dcr_visible & DTRACE_CRV_ALLZONE) || (cr != NULL &&
11227 		    (zone == GLOBAL_ZONEID || getzoneid() == zone)))
11228 			(void) dtrace_enabling_match(enab, NULL);
11229 	}
11230 
11231 	mutex_exit(&dtrace_lock);
11232 	mutex_exit(&cpu_lock);
11233 }
11234 
11235 /*
11236  * If an enabling is to be enabled without having matched probes (that is, if
11237  * dtrace_state_go() is to be called on the underlying dtrace_state_t), the
11238  * enabling must be _primed_ by creating an ECB for every ECB description.
11239  * This must be done to assure that we know the number of speculations, the
11240  * number of aggregations, the minimum buffer size needed, etc. before we
11241  * transition out of DTRACE_ACTIVITY_INACTIVE.  To do this without actually
11242  * enabling any probes, we create ECBs for every ECB decription, but with a
11243  * NULL probe -- which is exactly what this function does.
11244  */
11245 static void
11246 dtrace_enabling_prime(dtrace_state_t *state)
11247 {
11248 	dtrace_enabling_t *enab;
11249 	int i;
11250 
11251 	for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11252 		ASSERT(enab->dten_vstate->dtvs_state != NULL);
11253 
11254 		if (enab->dten_vstate->dtvs_state != state)
11255 			continue;
11256 
11257 		/*
11258 		 * We don't want to prime an enabling more than once, lest
11259 		 * we allow a malicious user to induce resource exhaustion.
11260 		 * (The ECBs that result from priming an enabling aren't
11261 		 * leaked -- but they also aren't deallocated until the
11262 		 * consumer state is destroyed.)
11263 		 */
11264 		if (enab->dten_primed)
11265 			continue;
11266 
11267 		for (i = 0; i < enab->dten_ndesc; i++) {
11268 			enab->dten_current = enab->dten_desc[i];
11269 			(void) dtrace_probe_enable(NULL, enab);
11270 		}
11271 
11272 		enab->dten_primed = 1;
11273 	}
11274 }
11275 
11276 /*
11277  * Called to indicate that probes should be provided due to retained
11278  * enablings.  This is implemented in terms of dtrace_probe_provide(), but it
11279  * must take an initial lap through the enabling calling the dtps_provide()
11280  * entry point explicitly to allow for autocreated probes.
11281  */
11282 static void
11283 dtrace_enabling_provide(dtrace_provider_t *prv)
11284 {
11285 	int i, all = 0;
11286 	dtrace_probedesc_t desc;
11287 	dtrace_genid_t gen;
11288 
11289 	ASSERT(MUTEX_HELD(&dtrace_lock));
11290 	ASSERT(MUTEX_HELD(&dtrace_provider_lock));
11291 
11292 	if (prv == NULL) {
11293 		all = 1;
11294 		prv = dtrace_provider;
11295 	}
11296 
11297 	do {
11298 		dtrace_enabling_t *enab;
11299 		void *parg = prv->dtpv_arg;
11300 
11301 retry:
11302 		gen = dtrace_retained_gen;
11303 		for (enab = dtrace_retained; enab != NULL;
11304 		    enab = enab->dten_next) {
11305 			for (i = 0; i < enab->dten_ndesc; i++) {
11306 				desc = enab->dten_desc[i]->dted_probe;
11307 				mutex_exit(&dtrace_lock);
11308 				prv->dtpv_pops.dtps_provide(parg, &desc);
11309 				mutex_enter(&dtrace_lock);
11310 				/*
11311 				 * Process the retained enablings again if
11312 				 * they have changed while we weren't holding
11313 				 * dtrace_lock.
11314 				 */
11315 				if (gen != dtrace_retained_gen)
11316 					goto retry;
11317 			}
11318 		}
11319 	} while (all && (prv = prv->dtpv_next) != NULL);
11320 
11321 	mutex_exit(&dtrace_lock);
11322 	dtrace_probe_provide(NULL, all ? NULL : prv);
11323 	mutex_enter(&dtrace_lock);
11324 }
11325 
11326 /*
11327  * Called to reap ECBs that are attached to probes from defunct providers.
11328  */
11329 static void
11330 dtrace_enabling_reap(void)
11331 {
11332 	dtrace_provider_t *prov;
11333 	dtrace_probe_t *probe;
11334 	dtrace_ecb_t *ecb;
11335 	hrtime_t when;
11336 	int i;
11337 
11338 	mutex_enter(&cpu_lock);
11339 	mutex_enter(&dtrace_lock);
11340 
11341 	for (i = 0; i < dtrace_nprobes; i++) {
11342 		if ((probe = dtrace_probes[i]) == NULL)
11343 			continue;
11344 
11345 		if (probe->dtpr_ecb == NULL)
11346 			continue;
11347 
11348 		prov = probe->dtpr_provider;
11349 
11350 		if ((when = prov->dtpv_defunct) == 0)
11351 			continue;
11352 
11353 		/*
11354 		 * We have ECBs on a defunct provider:  we want to reap these
11355 		 * ECBs to allow the provider to unregister.  The destruction
11356 		 * of these ECBs must be done carefully:  if we destroy the ECB
11357 		 * and the consumer later wishes to consume an EPID that
11358 		 * corresponds to the destroyed ECB (and if the EPID metadata
11359 		 * has not been previously consumed), the consumer will abort
11360 		 * processing on the unknown EPID.  To reduce (but not, sadly,
11361 		 * eliminate) the possibility of this, we will only destroy an
11362 		 * ECB for a defunct provider if, for the state that
11363 		 * corresponds to the ECB:
11364 		 *
11365 		 *  (a)	There is no speculative tracing (which can effectively
11366 		 *	cache an EPID for an arbitrary amount of time).
11367 		 *
11368 		 *  (b)	The principal buffers have been switched twice since the
11369 		 *	provider became defunct.
11370 		 *
11371 		 *  (c)	The aggregation buffers are of zero size or have been
11372 		 *	switched twice since the provider became defunct.
11373 		 *
11374 		 * We use dts_speculates to determine (a) and call a function
11375 		 * (dtrace_buffer_consumed()) to determine (b) and (c).  Note
11376 		 * that as soon as we've been unable to destroy one of the ECBs
11377 		 * associated with the probe, we quit trying -- reaping is only
11378 		 * fruitful in as much as we can destroy all ECBs associated
11379 		 * with the defunct provider's probes.
11380 		 */
11381 		while ((ecb = probe->dtpr_ecb) != NULL) {
11382 			dtrace_state_t *state = ecb->dte_state;
11383 			dtrace_buffer_t *buf = state->dts_buffer;
11384 			dtrace_buffer_t *aggbuf = state->dts_aggbuffer;
11385 
11386 			if (state->dts_speculates)
11387 				break;
11388 
11389 			if (!dtrace_buffer_consumed(buf, when))
11390 				break;
11391 
11392 			if (!dtrace_buffer_consumed(aggbuf, when))
11393 				break;
11394 
11395 			dtrace_ecb_disable(ecb);
11396 			ASSERT(probe->dtpr_ecb != ecb);
11397 			dtrace_ecb_destroy(ecb);
11398 		}
11399 	}
11400 
11401 	mutex_exit(&dtrace_lock);
11402 	mutex_exit(&cpu_lock);
11403 }
11404 
11405 /*
11406  * DTrace DOF Functions
11407  */
11408 /*ARGSUSED*/
11409 static void
11410 dtrace_dof_error(dof_hdr_t *dof, const char *str)
11411 {
11412 	if (dtrace_err_verbose)
11413 		cmn_err(CE_WARN, "failed to process DOF: %s", str);
11414 
11415 #ifdef DTRACE_ERRDEBUG
11416 	dtrace_errdebug(str);
11417 #endif
11418 }
11419 
11420 /*
11421  * Create DOF out of a currently enabled state.  Right now, we only create
11422  * DOF containing the run-time options -- but this could be expanded to create
11423  * complete DOF representing the enabled state.
11424  */
11425 static dof_hdr_t *
11426 dtrace_dof_create(dtrace_state_t *state)
11427 {
11428 	dof_hdr_t *dof;
11429 	dof_sec_t *sec;
11430 	dof_optdesc_t *opt;
11431 	int i, len = sizeof (dof_hdr_t) +
11432 	    roundup(sizeof (dof_sec_t), sizeof (uint64_t)) +
11433 	    sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
11434 
11435 	ASSERT(MUTEX_HELD(&dtrace_lock));
11436 
11437 	dof = kmem_zalloc(len, KM_SLEEP);
11438 	dof->dofh_ident[DOF_ID_MAG0] = DOF_MAG_MAG0;
11439 	dof->dofh_ident[DOF_ID_MAG1] = DOF_MAG_MAG1;
11440 	dof->dofh_ident[DOF_ID_MAG2] = DOF_MAG_MAG2;
11441 	dof->dofh_ident[DOF_ID_MAG3] = DOF_MAG_MAG3;
11442 
11443 	dof->dofh_ident[DOF_ID_MODEL] = DOF_MODEL_NATIVE;
11444 	dof->dofh_ident[DOF_ID_ENCODING] = DOF_ENCODE_NATIVE;
11445 	dof->dofh_ident[DOF_ID_VERSION] = DOF_VERSION;
11446 	dof->dofh_ident[DOF_ID_DIFVERS] = DIF_VERSION;
11447 	dof->dofh_ident[DOF_ID_DIFIREG] = DIF_DIR_NREGS;
11448 	dof->dofh_ident[DOF_ID_DIFTREG] = DIF_DTR_NREGS;
11449 
11450 	dof->dofh_flags = 0;
11451 	dof->dofh_hdrsize = sizeof (dof_hdr_t);
11452 	dof->dofh_secsize = sizeof (dof_sec_t);
11453 	dof->dofh_secnum = 1;	/* only DOF_SECT_OPTDESC */
11454 	dof->dofh_secoff = sizeof (dof_hdr_t);
11455 	dof->dofh_loadsz = len;
11456 	dof->dofh_filesz = len;
11457 	dof->dofh_pad = 0;
11458 
11459 	/*
11460 	 * Fill in the option section header...
11461 	 */
11462 	sec = (dof_sec_t *)((uintptr_t)dof + sizeof (dof_hdr_t));
11463 	sec->dofs_type = DOF_SECT_OPTDESC;
11464 	sec->dofs_align = sizeof (uint64_t);
11465 	sec->dofs_flags = DOF_SECF_LOAD;
11466 	sec->dofs_entsize = sizeof (dof_optdesc_t);
11467 
11468 	opt = (dof_optdesc_t *)((uintptr_t)sec +
11469 	    roundup(sizeof (dof_sec_t), sizeof (uint64_t)));
11470 
11471 	sec->dofs_offset = (uintptr_t)opt - (uintptr_t)dof;
11472 	sec->dofs_size = sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
11473 
11474 	for (i = 0; i < DTRACEOPT_MAX; i++) {
11475 		opt[i].dofo_option = i;
11476 		opt[i].dofo_strtab = DOF_SECIDX_NONE;
11477 		opt[i].dofo_value = state->dts_options[i];
11478 	}
11479 
11480 	return (dof);
11481 }
11482 
11483 static dof_hdr_t *
11484 dtrace_dof_copyin(uintptr_t uarg, int *errp)
11485 {
11486 	dof_hdr_t hdr, *dof;
11487 
11488 	ASSERT(!MUTEX_HELD(&dtrace_lock));
11489 
11490 	/*
11491 	 * First, we're going to copyin() the sizeof (dof_hdr_t).
11492 	 */
11493 	if (copyin((void *)uarg, &hdr, sizeof (hdr)) != 0) {
11494 		dtrace_dof_error(NULL, "failed to copyin DOF header");
11495 		*errp = EFAULT;
11496 		return (NULL);
11497 	}
11498 
11499 	/*
11500 	 * Now we'll allocate the entire DOF and copy it in -- provided
11501 	 * that the length isn't outrageous.
11502 	 */
11503 	if (hdr.dofh_loadsz >= dtrace_dof_maxsize) {
11504 		dtrace_dof_error(&hdr, "load size exceeds maximum");
11505 		*errp = E2BIG;
11506 		return (NULL);
11507 	}
11508 
11509 	if (hdr.dofh_loadsz < sizeof (hdr)) {
11510 		dtrace_dof_error(&hdr, "invalid load size");
11511 		*errp = EINVAL;
11512 		return (NULL);
11513 	}
11514 
11515 	dof = kmem_alloc(hdr.dofh_loadsz, KM_SLEEP);
11516 
11517 	if (copyin((void *)uarg, dof, hdr.dofh_loadsz) != 0 ||
11518 	    dof->dofh_loadsz != hdr.dofh_loadsz) {
11519 		kmem_free(dof, hdr.dofh_loadsz);
11520 		*errp = EFAULT;
11521 		return (NULL);
11522 	}
11523 
11524 	return (dof);
11525 }
11526 
11527 static dof_hdr_t *
11528 dtrace_dof_property(const char *name)
11529 {
11530 	uchar_t *buf;
11531 	uint64_t loadsz;
11532 	unsigned int len, i;
11533 	dof_hdr_t *dof;
11534 
11535 	/*
11536 	 * Unfortunately, array of values in .conf files are always (and
11537 	 * only) interpreted to be integer arrays.  We must read our DOF
11538 	 * as an integer array, and then squeeze it into a byte array.
11539 	 */
11540 	if (ddi_prop_lookup_int_array(DDI_DEV_T_ANY, dtrace_devi, 0,
11541 	    (char *)name, (int **)&buf, &len) != DDI_PROP_SUCCESS)
11542 		return (NULL);
11543 
11544 	for (i = 0; i < len; i++)
11545 		buf[i] = (uchar_t)(((int *)buf)[i]);
11546 
11547 	if (len < sizeof (dof_hdr_t)) {
11548 		ddi_prop_free(buf);
11549 		dtrace_dof_error(NULL, "truncated header");
11550 		return (NULL);
11551 	}
11552 
11553 	if (len < (loadsz = ((dof_hdr_t *)buf)->dofh_loadsz)) {
11554 		ddi_prop_free(buf);
11555 		dtrace_dof_error(NULL, "truncated DOF");
11556 		return (NULL);
11557 	}
11558 
11559 	if (loadsz >= dtrace_dof_maxsize) {
11560 		ddi_prop_free(buf);
11561 		dtrace_dof_error(NULL, "oversized DOF");
11562 		return (NULL);
11563 	}
11564 
11565 	dof = kmem_alloc(loadsz, KM_SLEEP);
11566 	bcopy(buf, dof, loadsz);
11567 	ddi_prop_free(buf);
11568 
11569 	return (dof);
11570 }
11571 
11572 static void
11573 dtrace_dof_destroy(dof_hdr_t *dof)
11574 {
11575 	kmem_free(dof, dof->dofh_loadsz);
11576 }
11577 
11578 /*
11579  * Return the dof_sec_t pointer corresponding to a given section index.  If the
11580  * index is not valid, dtrace_dof_error() is called and NULL is returned.  If
11581  * a type other than DOF_SECT_NONE is specified, the header is checked against
11582  * this type and NULL is returned if the types do not match.
11583  */
11584 static dof_sec_t *
11585 dtrace_dof_sect(dof_hdr_t *dof, uint32_t type, dof_secidx_t i)
11586 {
11587 	dof_sec_t *sec = (dof_sec_t *)(uintptr_t)
11588 	    ((uintptr_t)dof + dof->dofh_secoff + i * dof->dofh_secsize);
11589 
11590 	if (i >= dof->dofh_secnum) {
11591 		dtrace_dof_error(dof, "referenced section index is invalid");
11592 		return (NULL);
11593 	}
11594 
11595 	if (!(sec->dofs_flags & DOF_SECF_LOAD)) {
11596 		dtrace_dof_error(dof, "referenced section is not loadable");
11597 		return (NULL);
11598 	}
11599 
11600 	if (type != DOF_SECT_NONE && type != sec->dofs_type) {
11601 		dtrace_dof_error(dof, "referenced section is the wrong type");
11602 		return (NULL);
11603 	}
11604 
11605 	return (sec);
11606 }
11607 
11608 static dtrace_probedesc_t *
11609 dtrace_dof_probedesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_probedesc_t *desc)
11610 {
11611 	dof_probedesc_t *probe;
11612 	dof_sec_t *strtab;
11613 	uintptr_t daddr = (uintptr_t)dof;
11614 	uintptr_t str;
11615 	size_t size;
11616 
11617 	if (sec->dofs_type != DOF_SECT_PROBEDESC) {
11618 		dtrace_dof_error(dof, "invalid probe section");
11619 		return (NULL);
11620 	}
11621 
11622 	if (sec->dofs_align != sizeof (dof_secidx_t)) {
11623 		dtrace_dof_error(dof, "bad alignment in probe description");
11624 		return (NULL);
11625 	}
11626 
11627 	if (sec->dofs_offset + sizeof (dof_probedesc_t) > dof->dofh_loadsz) {
11628 		dtrace_dof_error(dof, "truncated probe description");
11629 		return (NULL);
11630 	}
11631 
11632 	probe = (dof_probedesc_t *)(uintptr_t)(daddr + sec->dofs_offset);
11633 	strtab = dtrace_dof_sect(dof, DOF_SECT_STRTAB, probe->dofp_strtab);
11634 
11635 	if (strtab == NULL)
11636 		return (NULL);
11637 
11638 	str = daddr + strtab->dofs_offset;
11639 	size = strtab->dofs_size;
11640 
11641 	if (probe->dofp_provider >= strtab->dofs_size) {
11642 		dtrace_dof_error(dof, "corrupt probe provider");
11643 		return (NULL);
11644 	}
11645 
11646 	(void) strncpy(desc->dtpd_provider,
11647 	    (char *)(str + probe->dofp_provider),
11648 	    MIN(DTRACE_PROVNAMELEN - 1, size - probe->dofp_provider));
11649 
11650 	if (probe->dofp_mod >= strtab->dofs_size) {
11651 		dtrace_dof_error(dof, "corrupt probe module");
11652 		return (NULL);
11653 	}
11654 
11655 	(void) strncpy(desc->dtpd_mod, (char *)(str + probe->dofp_mod),
11656 	    MIN(DTRACE_MODNAMELEN - 1, size - probe->dofp_mod));
11657 
11658 	if (probe->dofp_func >= strtab->dofs_size) {
11659 		dtrace_dof_error(dof, "corrupt probe function");
11660 		return (NULL);
11661 	}
11662 
11663 	(void) strncpy(desc->dtpd_func, (char *)(str + probe->dofp_func),
11664 	    MIN(DTRACE_FUNCNAMELEN - 1, size - probe->dofp_func));
11665 
11666 	if (probe->dofp_name >= strtab->dofs_size) {
11667 		dtrace_dof_error(dof, "corrupt probe name");
11668 		return (NULL);
11669 	}
11670 
11671 	(void) strncpy(desc->dtpd_name, (char *)(str + probe->dofp_name),
11672 	    MIN(DTRACE_NAMELEN - 1, size - probe->dofp_name));
11673 
11674 	return (desc);
11675 }
11676 
11677 static dtrace_difo_t *
11678 dtrace_dof_difo(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
11679     cred_t *cr)
11680 {
11681 	dtrace_difo_t *dp;
11682 	size_t ttl = 0;
11683 	dof_difohdr_t *dofd;
11684 	uintptr_t daddr = (uintptr_t)dof;
11685 	size_t max = dtrace_difo_maxsize;
11686 	int i, l, n;
11687 
11688 	static const struct {
11689 		int section;
11690 		int bufoffs;
11691 		int lenoffs;
11692 		int entsize;
11693 		int align;
11694 		const char *msg;
11695 	} difo[] = {
11696 		{ DOF_SECT_DIF, offsetof(dtrace_difo_t, dtdo_buf),
11697 		offsetof(dtrace_difo_t, dtdo_len), sizeof (dif_instr_t),
11698 		sizeof (dif_instr_t), "multiple DIF sections" },
11699 
11700 		{ DOF_SECT_INTTAB, offsetof(dtrace_difo_t, dtdo_inttab),
11701 		offsetof(dtrace_difo_t, dtdo_intlen), sizeof (uint64_t),
11702 		sizeof (uint64_t), "multiple integer tables" },
11703 
11704 		{ DOF_SECT_STRTAB, offsetof(dtrace_difo_t, dtdo_strtab),
11705 		offsetof(dtrace_difo_t, dtdo_strlen), 0,
11706 		sizeof (char), "multiple string tables" },
11707 
11708 		{ DOF_SECT_VARTAB, offsetof(dtrace_difo_t, dtdo_vartab),
11709 		offsetof(dtrace_difo_t, dtdo_varlen), sizeof (dtrace_difv_t),
11710 		sizeof (uint_t), "multiple variable tables" },
11711 
11712 		{ DOF_SECT_NONE, 0, 0, 0, NULL }
11713 	};
11714 
11715 	if (sec->dofs_type != DOF_SECT_DIFOHDR) {
11716 		dtrace_dof_error(dof, "invalid DIFO header section");
11717 		return (NULL);
11718 	}
11719 
11720 	if (sec->dofs_align != sizeof (dof_secidx_t)) {
11721 		dtrace_dof_error(dof, "bad alignment in DIFO header");
11722 		return (NULL);
11723 	}
11724 
11725 	if (sec->dofs_size < sizeof (dof_difohdr_t) ||
11726 	    sec->dofs_size % sizeof (dof_secidx_t)) {
11727 		dtrace_dof_error(dof, "bad size in DIFO header");
11728 		return (NULL);
11729 	}
11730 
11731 	dofd = (dof_difohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
11732 	n = (sec->dofs_size - sizeof (*dofd)) / sizeof (dof_secidx_t) + 1;
11733 
11734 	dp = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
11735 	dp->dtdo_rtype = dofd->dofd_rtype;
11736 
11737 	for (l = 0; l < n; l++) {
11738 		dof_sec_t *subsec;
11739 		void **bufp;
11740 		uint32_t *lenp;
11741 
11742 		if ((subsec = dtrace_dof_sect(dof, DOF_SECT_NONE,
11743 		    dofd->dofd_links[l])) == NULL)
11744 			goto err; /* invalid section link */
11745 
11746 		if (ttl + subsec->dofs_size > max) {
11747 			dtrace_dof_error(dof, "exceeds maximum size");
11748 			goto err;
11749 		}
11750 
11751 		ttl += subsec->dofs_size;
11752 
11753 		for (i = 0; difo[i].section != DOF_SECT_NONE; i++) {
11754 			if (subsec->dofs_type != difo[i].section)
11755 				continue;
11756 
11757 			if (!(subsec->dofs_flags & DOF_SECF_LOAD)) {
11758 				dtrace_dof_error(dof, "section not loaded");
11759 				goto err;
11760 			}
11761 
11762 			if (subsec->dofs_align != difo[i].align) {
11763 				dtrace_dof_error(dof, "bad alignment");
11764 				goto err;
11765 			}
11766 
11767 			bufp = (void **)((uintptr_t)dp + difo[i].bufoffs);
11768 			lenp = (uint32_t *)((uintptr_t)dp + difo[i].lenoffs);
11769 
11770 			if (*bufp != NULL) {
11771 				dtrace_dof_error(dof, difo[i].msg);
11772 				goto err;
11773 			}
11774 
11775 			if (difo[i].entsize != subsec->dofs_entsize) {
11776 				dtrace_dof_error(dof, "entry size mismatch");
11777 				goto err;
11778 			}
11779 
11780 			if (subsec->dofs_entsize != 0 &&
11781 			    (subsec->dofs_size % subsec->dofs_entsize) != 0) {
11782 				dtrace_dof_error(dof, "corrupt entry size");
11783 				goto err;
11784 			}
11785 
11786 			*lenp = subsec->dofs_size;
11787 			*bufp = kmem_alloc(subsec->dofs_size, KM_SLEEP);
11788 			bcopy((char *)(uintptr_t)(daddr + subsec->dofs_offset),
11789 			    *bufp, subsec->dofs_size);
11790 
11791 			if (subsec->dofs_entsize != 0)
11792 				*lenp /= subsec->dofs_entsize;
11793 
11794 			break;
11795 		}
11796 
11797 		/*
11798 		 * If we encounter a loadable DIFO sub-section that is not
11799 		 * known to us, assume this is a broken program and fail.
11800 		 */
11801 		if (difo[i].section == DOF_SECT_NONE &&
11802 		    (subsec->dofs_flags & DOF_SECF_LOAD)) {
11803 			dtrace_dof_error(dof, "unrecognized DIFO subsection");
11804 			goto err;
11805 		}
11806 	}
11807 
11808 	if (dp->dtdo_buf == NULL) {
11809 		/*
11810 		 * We can't have a DIF object without DIF text.
11811 		 */
11812 		dtrace_dof_error(dof, "missing DIF text");
11813 		goto err;
11814 	}
11815 
11816 	/*
11817 	 * Before we validate the DIF object, run through the variable table
11818 	 * looking for the strings -- if any of their size are under, we'll set
11819 	 * their size to be the system-wide default string size.  Note that
11820 	 * this should _not_ happen if the "strsize" option has been set --
11821 	 * in this case, the compiler should have set the size to reflect the
11822 	 * setting of the option.
11823 	 */
11824 	for (i = 0; i < dp->dtdo_varlen; i++) {
11825 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
11826 		dtrace_diftype_t *t = &v->dtdv_type;
11827 
11828 		if (v->dtdv_id < DIF_VAR_OTHER_UBASE)
11829 			continue;
11830 
11831 		if (t->dtdt_kind == DIF_TYPE_STRING && t->dtdt_size == 0)
11832 			t->dtdt_size = dtrace_strsize_default;
11833 	}
11834 
11835 	if (dtrace_difo_validate(dp, vstate, DIF_DIR_NREGS, cr) != 0)
11836 		goto err;
11837 
11838 	dtrace_difo_init(dp, vstate);
11839 	return (dp);
11840 
11841 err:
11842 	kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
11843 	kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
11844 	kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
11845 	kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
11846 
11847 	kmem_free(dp, sizeof (dtrace_difo_t));
11848 	return (NULL);
11849 }
11850 
11851 static dtrace_predicate_t *
11852 dtrace_dof_predicate(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
11853     cred_t *cr)
11854 {
11855 	dtrace_difo_t *dp;
11856 
11857 	if ((dp = dtrace_dof_difo(dof, sec, vstate, cr)) == NULL)
11858 		return (NULL);
11859 
11860 	return (dtrace_predicate_create(dp));
11861 }
11862 
11863 static dtrace_actdesc_t *
11864 dtrace_dof_actdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
11865     cred_t *cr)
11866 {
11867 	dtrace_actdesc_t *act, *first = NULL, *last = NULL, *next;
11868 	dof_actdesc_t *desc;
11869 	dof_sec_t *difosec;
11870 	size_t offs;
11871 	uintptr_t daddr = (uintptr_t)dof;
11872 	uint64_t arg;
11873 	dtrace_actkind_t kind;
11874 
11875 	if (sec->dofs_type != DOF_SECT_ACTDESC) {
11876 		dtrace_dof_error(dof, "invalid action section");
11877 		return (NULL);
11878 	}
11879 
11880 	if (sec->dofs_offset + sizeof (dof_actdesc_t) > dof->dofh_loadsz) {
11881 		dtrace_dof_error(dof, "truncated action description");
11882 		return (NULL);
11883 	}
11884 
11885 	if (sec->dofs_align != sizeof (uint64_t)) {
11886 		dtrace_dof_error(dof, "bad alignment in action description");
11887 		return (NULL);
11888 	}
11889 
11890 	if (sec->dofs_size < sec->dofs_entsize) {
11891 		dtrace_dof_error(dof, "section entry size exceeds total size");
11892 		return (NULL);
11893 	}
11894 
11895 	if (sec->dofs_entsize != sizeof (dof_actdesc_t)) {
11896 		dtrace_dof_error(dof, "bad entry size in action description");
11897 		return (NULL);
11898 	}
11899 
11900 	if (sec->dofs_size / sec->dofs_entsize > dtrace_actions_max) {
11901 		dtrace_dof_error(dof, "actions exceed dtrace_actions_max");
11902 		return (NULL);
11903 	}
11904 
11905 	for (offs = 0; offs < sec->dofs_size; offs += sec->dofs_entsize) {
11906 		desc = (dof_actdesc_t *)(daddr +
11907 		    (uintptr_t)sec->dofs_offset + offs);
11908 		kind = (dtrace_actkind_t)desc->dofa_kind;
11909 
11910 		if ((DTRACEACT_ISPRINTFLIKE(kind) &&
11911 		    (kind != DTRACEACT_PRINTA ||
11912 		    desc->dofa_strtab != DOF_SECIDX_NONE)) ||
11913 		    (kind == DTRACEACT_DIFEXPR &&
11914 		    desc->dofa_strtab != DOF_SECIDX_NONE)) {
11915 			dof_sec_t *strtab;
11916 			char *str, *fmt;
11917 			uint64_t i;
11918 
11919 			/*
11920 			 * The argument to these actions is an index into the
11921 			 * DOF string table.  For printf()-like actions, this
11922 			 * is the format string.  For print(), this is the
11923 			 * CTF type of the expression result.
11924 			 */
11925 			if ((strtab = dtrace_dof_sect(dof,
11926 			    DOF_SECT_STRTAB, desc->dofa_strtab)) == NULL)
11927 				goto err;
11928 
11929 			str = (char *)((uintptr_t)dof +
11930 			    (uintptr_t)strtab->dofs_offset);
11931 
11932 			for (i = desc->dofa_arg; i < strtab->dofs_size; i++) {
11933 				if (str[i] == '\0')
11934 					break;
11935 			}
11936 
11937 			if (i >= strtab->dofs_size) {
11938 				dtrace_dof_error(dof, "bogus format string");
11939 				goto err;
11940 			}
11941 
11942 			if (i == desc->dofa_arg) {
11943 				dtrace_dof_error(dof, "empty format string");
11944 				goto err;
11945 			}
11946 
11947 			i -= desc->dofa_arg;
11948 			fmt = kmem_alloc(i + 1, KM_SLEEP);
11949 			bcopy(&str[desc->dofa_arg], fmt, i + 1);
11950 			arg = (uint64_t)(uintptr_t)fmt;
11951 		} else {
11952 			if (kind == DTRACEACT_PRINTA) {
11953 				ASSERT(desc->dofa_strtab == DOF_SECIDX_NONE);
11954 				arg = 0;
11955 			} else {
11956 				arg = desc->dofa_arg;
11957 			}
11958 		}
11959 
11960 		act = dtrace_actdesc_create(kind, desc->dofa_ntuple,
11961 		    desc->dofa_uarg, arg);
11962 
11963 		if (last != NULL) {
11964 			last->dtad_next = act;
11965 		} else {
11966 			first = act;
11967 		}
11968 
11969 		last = act;
11970 
11971 		if (desc->dofa_difo == DOF_SECIDX_NONE)
11972 			continue;
11973 
11974 		if ((difosec = dtrace_dof_sect(dof,
11975 		    DOF_SECT_DIFOHDR, desc->dofa_difo)) == NULL)
11976 			goto err;
11977 
11978 		act->dtad_difo = dtrace_dof_difo(dof, difosec, vstate, cr);
11979 
11980 		if (act->dtad_difo == NULL)
11981 			goto err;
11982 	}
11983 
11984 	ASSERT(first != NULL);
11985 	return (first);
11986 
11987 err:
11988 	for (act = first; act != NULL; act = next) {
11989 		next = act->dtad_next;
11990 		dtrace_actdesc_release(act, vstate);
11991 	}
11992 
11993 	return (NULL);
11994 }
11995 
11996 static dtrace_ecbdesc_t *
11997 dtrace_dof_ecbdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
11998     cred_t *cr)
11999 {
12000 	dtrace_ecbdesc_t *ep;
12001 	dof_ecbdesc_t *ecb;
12002 	dtrace_probedesc_t *desc;
12003 	dtrace_predicate_t *pred = NULL;
12004 
12005 	if (sec->dofs_size < sizeof (dof_ecbdesc_t)) {
12006 		dtrace_dof_error(dof, "truncated ECB description");
12007 		return (NULL);
12008 	}
12009 
12010 	if (sec->dofs_align != sizeof (uint64_t)) {
12011 		dtrace_dof_error(dof, "bad alignment in ECB description");
12012 		return (NULL);
12013 	}
12014 
12015 	ecb = (dof_ecbdesc_t *)((uintptr_t)dof + (uintptr_t)sec->dofs_offset);
12016 	sec = dtrace_dof_sect(dof, DOF_SECT_PROBEDESC, ecb->dofe_probes);
12017 
12018 	if (sec == NULL)
12019 		return (NULL);
12020 
12021 	ep = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
12022 	ep->dted_uarg = ecb->dofe_uarg;
12023 	desc = &ep->dted_probe;
12024 
12025 	if (dtrace_dof_probedesc(dof, sec, desc) == NULL)
12026 		goto err;
12027 
12028 	if (ecb->dofe_pred != DOF_SECIDX_NONE) {
12029 		if ((sec = dtrace_dof_sect(dof,
12030 		    DOF_SECT_DIFOHDR, ecb->dofe_pred)) == NULL)
12031 			goto err;
12032 
12033 		if ((pred = dtrace_dof_predicate(dof, sec, vstate, cr)) == NULL)
12034 			goto err;
12035 
12036 		ep->dted_pred.dtpdd_predicate = pred;
12037 	}
12038 
12039 	if (ecb->dofe_actions != DOF_SECIDX_NONE) {
12040 		if ((sec = dtrace_dof_sect(dof,
12041 		    DOF_SECT_ACTDESC, ecb->dofe_actions)) == NULL)
12042 			goto err;
12043 
12044 		ep->dted_action = dtrace_dof_actdesc(dof, sec, vstate, cr);
12045 
12046 		if (ep->dted_action == NULL)
12047 			goto err;
12048 	}
12049 
12050 	return (ep);
12051 
12052 err:
12053 	if (pred != NULL)
12054 		dtrace_predicate_release(pred, vstate);
12055 	kmem_free(ep, sizeof (dtrace_ecbdesc_t));
12056 	return (NULL);
12057 }
12058 
12059 /*
12060  * Apply the relocations from the specified 'sec' (a DOF_SECT_URELHDR) to the
12061  * specified DOF.  At present, this amounts to simply adding 'ubase' to the
12062  * site of any user SETX relocations to account for load object base address.
12063  * In the future, if we need other relocations, this function can be extended.
12064  */
12065 static int
12066 dtrace_dof_relocate(dof_hdr_t *dof, dof_sec_t *sec, uint64_t ubase)
12067 {
12068 	uintptr_t daddr = (uintptr_t)dof;
12069 	dof_relohdr_t *dofr =
12070 	    (dof_relohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
12071 	dof_sec_t *ss, *rs, *ts;
12072 	dof_relodesc_t *r;
12073 	uint_t i, n;
12074 
12075 	if (sec->dofs_size < sizeof (dof_relohdr_t) ||
12076 	    sec->dofs_align != sizeof (dof_secidx_t)) {
12077 		dtrace_dof_error(dof, "invalid relocation header");
12078 		return (-1);
12079 	}
12080 
12081 	ss = dtrace_dof_sect(dof, DOF_SECT_STRTAB, dofr->dofr_strtab);
12082 	rs = dtrace_dof_sect(dof, DOF_SECT_RELTAB, dofr->dofr_relsec);
12083 	ts = dtrace_dof_sect(dof, DOF_SECT_NONE, dofr->dofr_tgtsec);
12084 
12085 	if (ss == NULL || rs == NULL || ts == NULL)
12086 		return (-1); /* dtrace_dof_error() has been called already */
12087 
12088 	if (rs->dofs_entsize < sizeof (dof_relodesc_t) ||
12089 	    rs->dofs_align != sizeof (uint64_t)) {
12090 		dtrace_dof_error(dof, "invalid relocation section");
12091 		return (-1);
12092 	}
12093 
12094 	r = (dof_relodesc_t *)(uintptr_t)(daddr + rs->dofs_offset);
12095 	n = rs->dofs_size / rs->dofs_entsize;
12096 
12097 	for (i = 0; i < n; i++) {
12098 		uintptr_t taddr = daddr + ts->dofs_offset + r->dofr_offset;
12099 
12100 		switch (r->dofr_type) {
12101 		case DOF_RELO_NONE:
12102 			break;
12103 		case DOF_RELO_SETX:
12104 			if (r->dofr_offset >= ts->dofs_size || r->dofr_offset +
12105 			    sizeof (uint64_t) > ts->dofs_size) {
12106 				dtrace_dof_error(dof, "bad relocation offset");
12107 				return (-1);
12108 			}
12109 
12110 			if (!IS_P2ALIGNED(taddr, sizeof (uint64_t))) {
12111 				dtrace_dof_error(dof, "misaligned setx relo");
12112 				return (-1);
12113 			}
12114 
12115 			*(uint64_t *)taddr += ubase;
12116 			break;
12117 		default:
12118 			dtrace_dof_error(dof, "invalid relocation type");
12119 			return (-1);
12120 		}
12121 
12122 		r = (dof_relodesc_t *)((uintptr_t)r + rs->dofs_entsize);
12123 	}
12124 
12125 	return (0);
12126 }
12127 
12128 /*
12129  * The dof_hdr_t passed to dtrace_dof_slurp() should be a partially validated
12130  * header:  it should be at the front of a memory region that is at least
12131  * sizeof (dof_hdr_t) in size -- and then at least dof_hdr.dofh_loadsz in
12132  * size.  It need not be validated in any other way.
12133  */
12134 static int
12135 dtrace_dof_slurp(dof_hdr_t *dof, dtrace_vstate_t *vstate, cred_t *cr,
12136     dtrace_enabling_t **enabp, uint64_t ubase, int noprobes)
12137 {
12138 	uint64_t len = dof->dofh_loadsz, seclen;
12139 	uintptr_t daddr = (uintptr_t)dof;
12140 	dtrace_ecbdesc_t *ep;
12141 	dtrace_enabling_t *enab;
12142 	uint_t i;
12143 
12144 	ASSERT(MUTEX_HELD(&dtrace_lock));
12145 	ASSERT(dof->dofh_loadsz >= sizeof (dof_hdr_t));
12146 
12147 	/*
12148 	 * Check the DOF header identification bytes.  In addition to checking
12149 	 * valid settings, we also verify that unused bits/bytes are zeroed so
12150 	 * we can use them later without fear of regressing existing binaries.
12151 	 */
12152 	if (bcmp(&dof->dofh_ident[DOF_ID_MAG0],
12153 	    DOF_MAG_STRING, DOF_MAG_STRLEN) != 0) {
12154 		dtrace_dof_error(dof, "DOF magic string mismatch");
12155 		return (-1);
12156 	}
12157 
12158 	if (dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_ILP32 &&
12159 	    dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_LP64) {
12160 		dtrace_dof_error(dof, "DOF has invalid data model");
12161 		return (-1);
12162 	}
12163 
12164 	if (dof->dofh_ident[DOF_ID_ENCODING] != DOF_ENCODE_NATIVE) {
12165 		dtrace_dof_error(dof, "DOF encoding mismatch");
12166 		return (-1);
12167 	}
12168 
12169 	if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
12170 	    dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_2) {
12171 		dtrace_dof_error(dof, "DOF version mismatch");
12172 		return (-1);
12173 	}
12174 
12175 	if (dof->dofh_ident[DOF_ID_DIFVERS] != DIF_VERSION_2) {
12176 		dtrace_dof_error(dof, "DOF uses unsupported instruction set");
12177 		return (-1);
12178 	}
12179 
12180 	if (dof->dofh_ident[DOF_ID_DIFIREG] > DIF_DIR_NREGS) {
12181 		dtrace_dof_error(dof, "DOF uses too many integer registers");
12182 		return (-1);
12183 	}
12184 
12185 	if (dof->dofh_ident[DOF_ID_DIFTREG] > DIF_DTR_NREGS) {
12186 		dtrace_dof_error(dof, "DOF uses too many tuple registers");
12187 		return (-1);
12188 	}
12189 
12190 	for (i = DOF_ID_PAD; i < DOF_ID_SIZE; i++) {
12191 		if (dof->dofh_ident[i] != 0) {
12192 			dtrace_dof_error(dof, "DOF has invalid ident byte set");
12193 			return (-1);
12194 		}
12195 	}
12196 
12197 	if (dof->dofh_flags & ~DOF_FL_VALID) {
12198 		dtrace_dof_error(dof, "DOF has invalid flag bits set");
12199 		return (-1);
12200 	}
12201 
12202 	if (dof->dofh_secsize == 0) {
12203 		dtrace_dof_error(dof, "zero section header size");
12204 		return (-1);
12205 	}
12206 
12207 	/*
12208 	 * Check that the section headers don't exceed the amount of DOF
12209 	 * data.  Note that we cast the section size and number of sections
12210 	 * to uint64_t's to prevent possible overflow in the multiplication.
12211 	 */
12212 	seclen = (uint64_t)dof->dofh_secnum * (uint64_t)dof->dofh_secsize;
12213 
12214 	if (dof->dofh_secoff > len || seclen > len ||
12215 	    dof->dofh_secoff + seclen > len) {
12216 		dtrace_dof_error(dof, "truncated section headers");
12217 		return (-1);
12218 	}
12219 
12220 	if (!IS_P2ALIGNED(dof->dofh_secoff, sizeof (uint64_t))) {
12221 		dtrace_dof_error(dof, "misaligned section headers");
12222 		return (-1);
12223 	}
12224 
12225 	if (!IS_P2ALIGNED(dof->dofh_secsize, sizeof (uint64_t))) {
12226 		dtrace_dof_error(dof, "misaligned section size");
12227 		return (-1);
12228 	}
12229 
12230 	/*
12231 	 * Take an initial pass through the section headers to be sure that
12232 	 * the headers don't have stray offsets.  If the 'noprobes' flag is
12233 	 * set, do not permit sections relating to providers, probes, or args.
12234 	 */
12235 	for (i = 0; i < dof->dofh_secnum; i++) {
12236 		dof_sec_t *sec = (dof_sec_t *)(daddr +
12237 		    (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12238 
12239 		if (noprobes) {
12240 			switch (sec->dofs_type) {
12241 			case DOF_SECT_PROVIDER:
12242 			case DOF_SECT_PROBES:
12243 			case DOF_SECT_PRARGS:
12244 			case DOF_SECT_PROFFS:
12245 				dtrace_dof_error(dof, "illegal sections "
12246 				    "for enabling");
12247 				return (-1);
12248 			}
12249 		}
12250 
12251 		if (DOF_SEC_ISLOADABLE(sec->dofs_type) &&
12252 		    !(sec->dofs_flags & DOF_SECF_LOAD)) {
12253 			dtrace_dof_error(dof, "loadable section with load "
12254 			    "flag unset");
12255 			return (-1);
12256 		}
12257 
12258 		if (!(sec->dofs_flags & DOF_SECF_LOAD))
12259 			continue; /* just ignore non-loadable sections */
12260 
12261 		if (sec->dofs_align & (sec->dofs_align - 1)) {
12262 			dtrace_dof_error(dof, "bad section alignment");
12263 			return (-1);
12264 		}
12265 
12266 		if (sec->dofs_offset & (sec->dofs_align - 1)) {
12267 			dtrace_dof_error(dof, "misaligned section");
12268 			return (-1);
12269 		}
12270 
12271 		if (sec->dofs_offset > len || sec->dofs_size > len ||
12272 		    sec->dofs_offset + sec->dofs_size > len) {
12273 			dtrace_dof_error(dof, "corrupt section header");
12274 			return (-1);
12275 		}
12276 
12277 		if (sec->dofs_type == DOF_SECT_STRTAB && *((char *)daddr +
12278 		    sec->dofs_offset + sec->dofs_size - 1) != '\0') {
12279 			dtrace_dof_error(dof, "non-terminating string table");
12280 			return (-1);
12281 		}
12282 	}
12283 
12284 	/*
12285 	 * Take a second pass through the sections and locate and perform any
12286 	 * relocations that are present.  We do this after the first pass to
12287 	 * be sure that all sections have had their headers validated.
12288 	 */
12289 	for (i = 0; i < dof->dofh_secnum; i++) {
12290 		dof_sec_t *sec = (dof_sec_t *)(daddr +
12291 		    (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12292 
12293 		if (!(sec->dofs_flags & DOF_SECF_LOAD))
12294 			continue; /* skip sections that are not loadable */
12295 
12296 		switch (sec->dofs_type) {
12297 		case DOF_SECT_URELHDR:
12298 			if (dtrace_dof_relocate(dof, sec, ubase) != 0)
12299 				return (-1);
12300 			break;
12301 		}
12302 	}
12303 
12304 	if ((enab = *enabp) == NULL)
12305 		enab = *enabp = dtrace_enabling_create(vstate);
12306 
12307 	for (i = 0; i < dof->dofh_secnum; i++) {
12308 		dof_sec_t *sec = (dof_sec_t *)(daddr +
12309 		    (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12310 
12311 		if (sec->dofs_type != DOF_SECT_ECBDESC)
12312 			continue;
12313 
12314 		if ((ep = dtrace_dof_ecbdesc(dof, sec, vstate, cr)) == NULL) {
12315 			dtrace_enabling_destroy(enab);
12316 			*enabp = NULL;
12317 			return (-1);
12318 		}
12319 
12320 		dtrace_enabling_add(enab, ep);
12321 	}
12322 
12323 	return (0);
12324 }
12325 
12326 /*
12327  * Process DOF for any options.  This routine assumes that the DOF has been
12328  * at least processed by dtrace_dof_slurp().
12329  */
12330 static int
12331 dtrace_dof_options(dof_hdr_t *dof, dtrace_state_t *state)
12332 {
12333 	int i, rval;
12334 	uint32_t entsize;
12335 	size_t offs;
12336 	dof_optdesc_t *desc;
12337 
12338 	for (i = 0; i < dof->dofh_secnum; i++) {
12339 		dof_sec_t *sec = (dof_sec_t *)((uintptr_t)dof +
12340 		    (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12341 
12342 		if (sec->dofs_type != DOF_SECT_OPTDESC)
12343 			continue;
12344 
12345 		if (sec->dofs_align != sizeof (uint64_t)) {
12346 			dtrace_dof_error(dof, "bad alignment in "
12347 			    "option description");
12348 			return (EINVAL);
12349 		}
12350 
12351 		if ((entsize = sec->dofs_entsize) == 0) {
12352 			dtrace_dof_error(dof, "zeroed option entry size");
12353 			return (EINVAL);
12354 		}
12355 
12356 		if (entsize < sizeof (dof_optdesc_t)) {
12357 			dtrace_dof_error(dof, "bad option entry size");
12358 			return (EINVAL);
12359 		}
12360 
12361 		for (offs = 0; offs < sec->dofs_size; offs += entsize) {
12362 			desc = (dof_optdesc_t *)((uintptr_t)dof +
12363 			    (uintptr_t)sec->dofs_offset + offs);
12364 
12365 			if (desc->dofo_strtab != DOF_SECIDX_NONE) {
12366 				dtrace_dof_error(dof, "non-zero option string");
12367 				return (EINVAL);
12368 			}
12369 
12370 			if (desc->dofo_value == DTRACEOPT_UNSET) {
12371 				dtrace_dof_error(dof, "unset option");
12372 				return (EINVAL);
12373 			}
12374 
12375 			if ((rval = dtrace_state_option(state,
12376 			    desc->dofo_option, desc->dofo_value)) != 0) {
12377 				dtrace_dof_error(dof, "rejected option");
12378 				return (rval);
12379 			}
12380 		}
12381 	}
12382 
12383 	return (0);
12384 }
12385 
12386 /*
12387  * DTrace Consumer State Functions
12388  */
12389 int
12390 dtrace_dstate_init(dtrace_dstate_t *dstate, size_t size)
12391 {
12392 	size_t hashsize, maxper, min, chunksize = dstate->dtds_chunksize;
12393 	void *base;
12394 	uintptr_t limit;
12395 	dtrace_dynvar_t *dvar, *next, *start;
12396 	int i;
12397 
12398 	ASSERT(MUTEX_HELD(&dtrace_lock));
12399 	ASSERT(dstate->dtds_base == NULL && dstate->dtds_percpu == NULL);
12400 
12401 	bzero(dstate, sizeof (dtrace_dstate_t));
12402 
12403 	if ((dstate->dtds_chunksize = chunksize) == 0)
12404 		dstate->dtds_chunksize = DTRACE_DYNVAR_CHUNKSIZE;
12405 
12406 	if (size < (min = dstate->dtds_chunksize + sizeof (dtrace_dynhash_t)))
12407 		size = min;
12408 
12409 	if ((base = kmem_zalloc(size, KM_NOSLEEP | KM_NORMALPRI)) == NULL)
12410 		return (ENOMEM);
12411 
12412 	dstate->dtds_size = size;
12413 	dstate->dtds_base = base;
12414 	dstate->dtds_percpu = kmem_cache_alloc(dtrace_state_cache, KM_SLEEP);
12415 	bzero(dstate->dtds_percpu, NCPU * sizeof (dtrace_dstate_percpu_t));
12416 
12417 	hashsize = size / (dstate->dtds_chunksize + sizeof (dtrace_dynhash_t));
12418 
12419 	if (hashsize != 1 && (hashsize & 1))
12420 		hashsize--;
12421 
12422 	dstate->dtds_hashsize = hashsize;
12423 	dstate->dtds_hash = dstate->dtds_base;
12424 
12425 	/*
12426 	 * Set all of our hash buckets to point to the single sink, and (if
12427 	 * it hasn't already been set), set the sink's hash value to be the
12428 	 * sink sentinel value.  The sink is needed for dynamic variable
12429 	 * lookups to know that they have iterated over an entire, valid hash
12430 	 * chain.
12431 	 */
12432 	for (i = 0; i < hashsize; i++)
12433 		dstate->dtds_hash[i].dtdh_chain = &dtrace_dynhash_sink;
12434 
12435 	if (dtrace_dynhash_sink.dtdv_hashval != DTRACE_DYNHASH_SINK)
12436 		dtrace_dynhash_sink.dtdv_hashval = DTRACE_DYNHASH_SINK;
12437 
12438 	/*
12439 	 * Determine number of active CPUs.  Divide free list evenly among
12440 	 * active CPUs.
12441 	 */
12442 	start = (dtrace_dynvar_t *)
12443 	    ((uintptr_t)base + hashsize * sizeof (dtrace_dynhash_t));
12444 	limit = (uintptr_t)base + size;
12445 
12446 	maxper = (limit - (uintptr_t)start) / NCPU;
12447 	maxper = (maxper / dstate->dtds_chunksize) * dstate->dtds_chunksize;
12448 
12449 	for (i = 0; i < NCPU; i++) {
12450 		dstate->dtds_percpu[i].dtdsc_free = dvar = start;
12451 
12452 		/*
12453 		 * If we don't even have enough chunks to make it once through
12454 		 * NCPUs, we're just going to allocate everything to the first
12455 		 * CPU.  And if we're on the last CPU, we're going to allocate
12456 		 * whatever is left over.  In either case, we set the limit to
12457 		 * be the limit of the dynamic variable space.
12458 		 */
12459 		if (maxper == 0 || i == NCPU - 1) {
12460 			limit = (uintptr_t)base + size;
12461 			start = NULL;
12462 		} else {
12463 			limit = (uintptr_t)start + maxper;
12464 			start = (dtrace_dynvar_t *)limit;
12465 		}
12466 
12467 		ASSERT(limit <= (uintptr_t)base + size);
12468 
12469 		for (;;) {
12470 			next = (dtrace_dynvar_t *)((uintptr_t)dvar +
12471 			    dstate->dtds_chunksize);
12472 
12473 			if ((uintptr_t)next + dstate->dtds_chunksize >= limit)
12474 				break;
12475 
12476 			dvar->dtdv_next = next;
12477 			dvar = next;
12478 		}
12479 
12480 		if (maxper == 0)
12481 			break;
12482 	}
12483 
12484 	return (0);
12485 }
12486 
12487 void
12488 dtrace_dstate_fini(dtrace_dstate_t *dstate)
12489 {
12490 	ASSERT(MUTEX_HELD(&cpu_lock));
12491 
12492 	if (dstate->dtds_base == NULL)
12493 		return;
12494 
12495 	kmem_free(dstate->dtds_base, dstate->dtds_size);
12496 	kmem_cache_free(dtrace_state_cache, dstate->dtds_percpu);
12497 }
12498 
12499 static void
12500 dtrace_vstate_fini(dtrace_vstate_t *vstate)
12501 {
12502 	/*
12503 	 * Logical XOR, where are you?
12504 	 */
12505 	ASSERT((vstate->dtvs_nglobals == 0) ^ (vstate->dtvs_globals != NULL));
12506 
12507 	if (vstate->dtvs_nglobals > 0) {
12508 		kmem_free(vstate->dtvs_globals, vstate->dtvs_nglobals *
12509 		    sizeof (dtrace_statvar_t *));
12510 	}
12511 
12512 	if (vstate->dtvs_ntlocals > 0) {
12513 		kmem_free(vstate->dtvs_tlocals, vstate->dtvs_ntlocals *
12514 		    sizeof (dtrace_difv_t));
12515 	}
12516 
12517 	ASSERT((vstate->dtvs_nlocals == 0) ^ (vstate->dtvs_locals != NULL));
12518 
12519 	if (vstate->dtvs_nlocals > 0) {
12520 		kmem_free(vstate->dtvs_locals, vstate->dtvs_nlocals *
12521 		    sizeof (dtrace_statvar_t *));
12522 	}
12523 }
12524 
12525 static void
12526 dtrace_state_clean(dtrace_state_t *state)
12527 {
12528 	if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE)
12529 		return;
12530 
12531 	dtrace_dynvar_clean(&state->dts_vstate.dtvs_dynvars);
12532 	dtrace_speculation_clean(state);
12533 }
12534 
12535 static void
12536 dtrace_state_deadman(dtrace_state_t *state)
12537 {
12538 	hrtime_t now;
12539 
12540 	dtrace_sync();
12541 
12542 	now = dtrace_gethrtime();
12543 
12544 	if (state != dtrace_anon.dta_state &&
12545 	    now - state->dts_laststatus >= dtrace_deadman_user)
12546 		return;
12547 
12548 	/*
12549 	 * We must be sure that dts_alive never appears to be less than the
12550 	 * value upon entry to dtrace_state_deadman(), and because we lack a
12551 	 * dtrace_cas64(), we cannot store to it atomically.  We thus instead
12552 	 * store INT64_MAX to it, followed by a memory barrier, followed by
12553 	 * the new value.  This assures that dts_alive never appears to be
12554 	 * less than its true value, regardless of the order in which the
12555 	 * stores to the underlying storage are issued.
12556 	 */
12557 	state->dts_alive = INT64_MAX;
12558 	dtrace_membar_producer();
12559 	state->dts_alive = now;
12560 }
12561 
12562 dtrace_state_t *
12563 dtrace_state_create(dev_t *devp, cred_t *cr)
12564 {
12565 	minor_t minor;
12566 	major_t major;
12567 	char c[30];
12568 	dtrace_state_t *state;
12569 	dtrace_optval_t *opt;
12570 	int bufsize = NCPU * sizeof (dtrace_buffer_t), i;
12571 
12572 	ASSERT(MUTEX_HELD(&dtrace_lock));
12573 	ASSERT(MUTEX_HELD(&cpu_lock));
12574 
12575 	minor = (minor_t)(uintptr_t)vmem_alloc(dtrace_minor, 1,
12576 	    VM_BESTFIT | VM_SLEEP);
12577 
12578 	if (ddi_soft_state_zalloc(dtrace_softstate, minor) != DDI_SUCCESS) {
12579 		vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
12580 		return (NULL);
12581 	}
12582 
12583 	state = ddi_get_soft_state(dtrace_softstate, minor);
12584 	state->dts_epid = DTRACE_EPIDNONE + 1;
12585 
12586 	(void) snprintf(c, sizeof (c), "dtrace_aggid_%d", minor);
12587 	state->dts_aggid_arena = vmem_create(c, (void *)1, UINT32_MAX, 1,
12588 	    NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
12589 
12590 	if (devp != NULL) {
12591 		major = getemajor(*devp);
12592 	} else {
12593 		major = ddi_driver_major(dtrace_devi);
12594 	}
12595 
12596 	state->dts_dev = makedevice(major, minor);
12597 
12598 	if (devp != NULL)
12599 		*devp = state->dts_dev;
12600 
12601 	/*
12602 	 * We allocate NCPU buffers.  On the one hand, this can be quite
12603 	 * a bit of memory per instance (nearly 36K on a Starcat).  On the
12604 	 * other hand, it saves an additional memory reference in the probe
12605 	 * path.
12606 	 */
12607 	state->dts_buffer = kmem_zalloc(bufsize, KM_SLEEP);
12608 	state->dts_aggbuffer = kmem_zalloc(bufsize, KM_SLEEP);
12609 	state->dts_cleaner = CYCLIC_NONE;
12610 	state->dts_deadman = CYCLIC_NONE;
12611 	state->dts_vstate.dtvs_state = state;
12612 
12613 	for (i = 0; i < DTRACEOPT_MAX; i++)
12614 		state->dts_options[i] = DTRACEOPT_UNSET;
12615 
12616 	/*
12617 	 * Set the default options.
12618 	 */
12619 	opt = state->dts_options;
12620 	opt[DTRACEOPT_BUFPOLICY] = DTRACEOPT_BUFPOLICY_SWITCH;
12621 	opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_AUTO;
12622 	opt[DTRACEOPT_NSPEC] = dtrace_nspec_default;
12623 	opt[DTRACEOPT_SPECSIZE] = dtrace_specsize_default;
12624 	opt[DTRACEOPT_CPU] = (dtrace_optval_t)DTRACE_CPUALL;
12625 	opt[DTRACEOPT_STRSIZE] = dtrace_strsize_default;
12626 	opt[DTRACEOPT_STACKFRAMES] = dtrace_stackframes_default;
12627 	opt[DTRACEOPT_USTACKFRAMES] = dtrace_ustackframes_default;
12628 	opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_default;
12629 	opt[DTRACEOPT_AGGRATE] = dtrace_aggrate_default;
12630 	opt[DTRACEOPT_SWITCHRATE] = dtrace_switchrate_default;
12631 	opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_default;
12632 	opt[DTRACEOPT_JSTACKFRAMES] = dtrace_jstackframes_default;
12633 	opt[DTRACEOPT_JSTACKSTRSIZE] = dtrace_jstackstrsize_default;
12634 
12635 	state->dts_activity = DTRACE_ACTIVITY_INACTIVE;
12636 
12637 	/*
12638 	 * Depending on the user credentials, we set flag bits which alter probe
12639 	 * visibility or the amount of destructiveness allowed.  In the case of
12640 	 * actual anonymous tracing, or the possession of all privileges, all of
12641 	 * the normal checks are bypassed.
12642 	 */
12643 	if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
12644 		state->dts_cred.dcr_visible = DTRACE_CRV_ALL;
12645 		state->dts_cred.dcr_action = DTRACE_CRA_ALL;
12646 	} else {
12647 		/*
12648 		 * Set up the credentials for this instantiation.  We take a
12649 		 * hold on the credential to prevent it from disappearing on
12650 		 * us; this in turn prevents the zone_t referenced by this
12651 		 * credential from disappearing.  This means that we can
12652 		 * examine the credential and the zone from probe context.
12653 		 */
12654 		crhold(cr);
12655 		state->dts_cred.dcr_cred = cr;
12656 
12657 		/*
12658 		 * CRA_PROC means "we have *some* privilege for dtrace" and
12659 		 * unlocks the use of variables like pid, zonename, etc.
12660 		 */
12661 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE) ||
12662 		    PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
12663 			state->dts_cred.dcr_action |= DTRACE_CRA_PROC;
12664 		}
12665 
12666 		/*
12667 		 * dtrace_user allows use of syscall and profile providers.
12668 		 * If the user also has proc_owner and/or proc_zone, we
12669 		 * extend the scope to include additional visibility and
12670 		 * destructive power.
12671 		 */
12672 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE)) {
12673 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE)) {
12674 				state->dts_cred.dcr_visible |=
12675 				    DTRACE_CRV_ALLPROC;
12676 
12677 				state->dts_cred.dcr_action |=
12678 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
12679 			}
12680 
12681 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE)) {
12682 				state->dts_cred.dcr_visible |=
12683 				    DTRACE_CRV_ALLZONE;
12684 
12685 				state->dts_cred.dcr_action |=
12686 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
12687 			}
12688 
12689 			/*
12690 			 * If we have all privs in whatever zone this is,
12691 			 * we can do destructive things to processes which
12692 			 * have altered credentials.
12693 			 */
12694 			if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
12695 			    cr->cr_zone->zone_privset)) {
12696 				state->dts_cred.dcr_action |=
12697 				    DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
12698 			}
12699 		}
12700 
12701 		/*
12702 		 * Holding the dtrace_kernel privilege also implies that
12703 		 * the user has the dtrace_user privilege from a visibility
12704 		 * perspective.  But without further privileges, some
12705 		 * destructive actions are not available.
12706 		 */
12707 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE)) {
12708 			/*
12709 			 * Make all probes in all zones visible.  However,
12710 			 * this doesn't mean that all actions become available
12711 			 * to all zones.
12712 			 */
12713 			state->dts_cred.dcr_visible |= DTRACE_CRV_KERNEL |
12714 			    DTRACE_CRV_ALLPROC | DTRACE_CRV_ALLZONE;
12715 
12716 			state->dts_cred.dcr_action |= DTRACE_CRA_KERNEL |
12717 			    DTRACE_CRA_PROC;
12718 			/*
12719 			 * Holding proc_owner means that destructive actions
12720 			 * for *this* zone are allowed.
12721 			 */
12722 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
12723 				state->dts_cred.dcr_action |=
12724 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
12725 
12726 			/*
12727 			 * Holding proc_zone means that destructive actions
12728 			 * for this user/group ID in all zones is allowed.
12729 			 */
12730 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
12731 				state->dts_cred.dcr_action |=
12732 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
12733 
12734 			/*
12735 			 * If we have all privs in whatever zone this is,
12736 			 * we can do destructive things to processes which
12737 			 * have altered credentials.
12738 			 */
12739 			if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
12740 			    cr->cr_zone->zone_privset)) {
12741 				state->dts_cred.dcr_action |=
12742 				    DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
12743 			}
12744 		}
12745 
12746 		/*
12747 		 * Holding the dtrace_proc privilege gives control over fasttrap
12748 		 * and pid providers.  We need to grant wider destructive
12749 		 * privileges in the event that the user has proc_owner and/or
12750 		 * proc_zone.
12751 		 */
12752 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
12753 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
12754 				state->dts_cred.dcr_action |=
12755 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
12756 
12757 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
12758 				state->dts_cred.dcr_action |=
12759 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
12760 		}
12761 	}
12762 
12763 	return (state);
12764 }
12765 
12766 static int
12767 dtrace_state_buffer(dtrace_state_t *state, dtrace_buffer_t *buf, int which)
12768 {
12769 	dtrace_optval_t *opt = state->dts_options, size;
12770 	processorid_t cpu;
12771 	int flags = 0, rval, factor, divisor = 1;
12772 
12773 	ASSERT(MUTEX_HELD(&dtrace_lock));
12774 	ASSERT(MUTEX_HELD(&cpu_lock));
12775 	ASSERT(which < DTRACEOPT_MAX);
12776 	ASSERT(state->dts_activity == DTRACE_ACTIVITY_INACTIVE ||
12777 	    (state == dtrace_anon.dta_state &&
12778 	    state->dts_activity == DTRACE_ACTIVITY_ACTIVE));
12779 
12780 	if (opt[which] == DTRACEOPT_UNSET || opt[which] == 0)
12781 		return (0);
12782 
12783 	if (opt[DTRACEOPT_CPU] != DTRACEOPT_UNSET)
12784 		cpu = opt[DTRACEOPT_CPU];
12785 
12786 	if (which == DTRACEOPT_SPECSIZE)
12787 		flags |= DTRACEBUF_NOSWITCH;
12788 
12789 	if (which == DTRACEOPT_BUFSIZE) {
12790 		if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_RING)
12791 			flags |= DTRACEBUF_RING;
12792 
12793 		if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_FILL)
12794 			flags |= DTRACEBUF_FILL;
12795 
12796 		if (state != dtrace_anon.dta_state ||
12797 		    state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
12798 			flags |= DTRACEBUF_INACTIVE;
12799 	}
12800 
12801 	for (size = opt[which]; size >= sizeof (uint64_t); size /= divisor) {
12802 		/*
12803 		 * The size must be 8-byte aligned.  If the size is not 8-byte
12804 		 * aligned, drop it down by the difference.
12805 		 */
12806 		if (size & (sizeof (uint64_t) - 1))
12807 			size -= size & (sizeof (uint64_t) - 1);
12808 
12809 		if (size < state->dts_reserve) {
12810 			/*
12811 			 * Buffers always must be large enough to accommodate
12812 			 * their prereserved space.  We return E2BIG instead
12813 			 * of ENOMEM in this case to allow for user-level
12814 			 * software to differentiate the cases.
12815 			 */
12816 			return (E2BIG);
12817 		}
12818 
12819 		rval = dtrace_buffer_alloc(buf, size, flags, cpu, &factor);
12820 
12821 		if (rval != ENOMEM) {
12822 			opt[which] = size;
12823 			return (rval);
12824 		}
12825 
12826 		if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
12827 			return (rval);
12828 
12829 		for (divisor = 2; divisor < factor; divisor <<= 1)
12830 			continue;
12831 	}
12832 
12833 	return (ENOMEM);
12834 }
12835 
12836 static int
12837 dtrace_state_buffers(dtrace_state_t *state)
12838 {
12839 	dtrace_speculation_t *spec = state->dts_speculations;
12840 	int rval, i;
12841 
12842 	if ((rval = dtrace_state_buffer(state, state->dts_buffer,
12843 	    DTRACEOPT_BUFSIZE)) != 0)
12844 		return (rval);
12845 
12846 	if ((rval = dtrace_state_buffer(state, state->dts_aggbuffer,
12847 	    DTRACEOPT_AGGSIZE)) != 0)
12848 		return (rval);
12849 
12850 	for (i = 0; i < state->dts_nspeculations; i++) {
12851 		if ((rval = dtrace_state_buffer(state,
12852 		    spec[i].dtsp_buffer, DTRACEOPT_SPECSIZE)) != 0)
12853 			return (rval);
12854 	}
12855 
12856 	return (0);
12857 }
12858 
12859 static void
12860 dtrace_state_prereserve(dtrace_state_t *state)
12861 {
12862 	dtrace_ecb_t *ecb;
12863 	dtrace_probe_t *probe;
12864 
12865 	state->dts_reserve = 0;
12866 
12867 	if (state->dts_options[DTRACEOPT_BUFPOLICY] != DTRACEOPT_BUFPOLICY_FILL)
12868 		return;
12869 
12870 	/*
12871 	 * If our buffer policy is a "fill" buffer policy, we need to set the
12872 	 * prereserved space to be the space required by the END probes.
12873 	 */
12874 	probe = dtrace_probes[dtrace_probeid_end - 1];
12875 	ASSERT(probe != NULL);
12876 
12877 	for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
12878 		if (ecb->dte_state != state)
12879 			continue;
12880 
12881 		state->dts_reserve += ecb->dte_needed + ecb->dte_alignment;
12882 	}
12883 }
12884 
12885 static int
12886 dtrace_state_go(dtrace_state_t *state, processorid_t *cpu)
12887 {
12888 	dtrace_optval_t *opt = state->dts_options, sz, nspec;
12889 	dtrace_speculation_t *spec;
12890 	dtrace_buffer_t *buf;
12891 	cyc_handler_t hdlr;
12892 	cyc_time_t when;
12893 	int rval = 0, i, bufsize = NCPU * sizeof (dtrace_buffer_t);
12894 	dtrace_icookie_t cookie;
12895 
12896 	mutex_enter(&cpu_lock);
12897 	mutex_enter(&dtrace_lock);
12898 
12899 	if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
12900 		rval = EBUSY;
12901 		goto out;
12902 	}
12903 
12904 	/*
12905 	 * Before we can perform any checks, we must prime all of the
12906 	 * retained enablings that correspond to this state.
12907 	 */
12908 	dtrace_enabling_prime(state);
12909 
12910 	if (state->dts_destructive && !state->dts_cred.dcr_destructive) {
12911 		rval = EACCES;
12912 		goto out;
12913 	}
12914 
12915 	dtrace_state_prereserve(state);
12916 
12917 	/*
12918 	 * Now we want to do is try to allocate our speculations.
12919 	 * We do not automatically resize the number of speculations; if
12920 	 * this fails, we will fail the operation.
12921 	 */
12922 	nspec = opt[DTRACEOPT_NSPEC];
12923 	ASSERT(nspec != DTRACEOPT_UNSET);
12924 
12925 	if (nspec > INT_MAX) {
12926 		rval = ENOMEM;
12927 		goto out;
12928 	}
12929 
12930 	spec = kmem_zalloc(nspec * sizeof (dtrace_speculation_t),
12931 	    KM_NOSLEEP | KM_NORMALPRI);
12932 
12933 	if (spec == NULL) {
12934 		rval = ENOMEM;
12935 		goto out;
12936 	}
12937 
12938 	state->dts_speculations = spec;
12939 	state->dts_nspeculations = (int)nspec;
12940 
12941 	for (i = 0; i < nspec; i++) {
12942 		if ((buf = kmem_zalloc(bufsize,
12943 		    KM_NOSLEEP | KM_NORMALPRI)) == NULL) {
12944 			rval = ENOMEM;
12945 			goto err;
12946 		}
12947 
12948 		spec[i].dtsp_buffer = buf;
12949 	}
12950 
12951 	if (opt[DTRACEOPT_GRABANON] != DTRACEOPT_UNSET) {
12952 		if (dtrace_anon.dta_state == NULL) {
12953 			rval = ENOENT;
12954 			goto out;
12955 		}
12956 
12957 		if (state->dts_necbs != 0) {
12958 			rval = EALREADY;
12959 			goto out;
12960 		}
12961 
12962 		state->dts_anon = dtrace_anon_grab();
12963 		ASSERT(state->dts_anon != NULL);
12964 		state = state->dts_anon;
12965 
12966 		/*
12967 		 * We want "grabanon" to be set in the grabbed state, so we'll
12968 		 * copy that option value from the grabbing state into the
12969 		 * grabbed state.
12970 		 */
12971 		state->dts_options[DTRACEOPT_GRABANON] =
12972 		    opt[DTRACEOPT_GRABANON];
12973 
12974 		*cpu = dtrace_anon.dta_beganon;
12975 
12976 		/*
12977 		 * If the anonymous state is active (as it almost certainly
12978 		 * is if the anonymous enabling ultimately matched anything),
12979 		 * we don't allow any further option processing -- but we
12980 		 * don't return failure.
12981 		 */
12982 		if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
12983 			goto out;
12984 	}
12985 
12986 	if (opt[DTRACEOPT_AGGSIZE] != DTRACEOPT_UNSET &&
12987 	    opt[DTRACEOPT_AGGSIZE] != 0) {
12988 		if (state->dts_aggregations == NULL) {
12989 			/*
12990 			 * We're not going to create an aggregation buffer
12991 			 * because we don't have any ECBs that contain
12992 			 * aggregations -- set this option to 0.
12993 			 */
12994 			opt[DTRACEOPT_AGGSIZE] = 0;
12995 		} else {
12996 			/*
12997 			 * If we have an aggregation buffer, we must also have
12998 			 * a buffer to use as scratch.
12999 			 */
13000 			if (opt[DTRACEOPT_BUFSIZE] == DTRACEOPT_UNSET ||
13001 			    opt[DTRACEOPT_BUFSIZE] < state->dts_needed) {
13002 				opt[DTRACEOPT_BUFSIZE] = state->dts_needed;
13003 			}
13004 		}
13005 	}
13006 
13007 	if (opt[DTRACEOPT_SPECSIZE] != DTRACEOPT_UNSET &&
13008 	    opt[DTRACEOPT_SPECSIZE] != 0) {
13009 		if (!state->dts_speculates) {
13010 			/*
13011 			 * We're not going to create speculation buffers
13012 			 * because we don't have any ECBs that actually
13013 			 * speculate -- set the speculation size to 0.
13014 			 */
13015 			opt[DTRACEOPT_SPECSIZE] = 0;
13016 		}
13017 	}
13018 
13019 	/*
13020 	 * The bare minimum size for any buffer that we're actually going to
13021 	 * do anything to is sizeof (uint64_t).
13022 	 */
13023 	sz = sizeof (uint64_t);
13024 
13025 	if ((state->dts_needed != 0 && opt[DTRACEOPT_BUFSIZE] < sz) ||
13026 	    (state->dts_speculates && opt[DTRACEOPT_SPECSIZE] < sz) ||
13027 	    (state->dts_aggregations != NULL && opt[DTRACEOPT_AGGSIZE] < sz)) {
13028 		/*
13029 		 * A buffer size has been explicitly set to 0 (or to a size
13030 		 * that will be adjusted to 0) and we need the space -- we
13031 		 * need to return failure.  We return ENOSPC to differentiate
13032 		 * it from failing to allocate a buffer due to failure to meet
13033 		 * the reserve (for which we return E2BIG).
13034 		 */
13035 		rval = ENOSPC;
13036 		goto out;
13037 	}
13038 
13039 	if ((rval = dtrace_state_buffers(state)) != 0)
13040 		goto err;
13041 
13042 	if ((sz = opt[DTRACEOPT_DYNVARSIZE]) == DTRACEOPT_UNSET)
13043 		sz = dtrace_dstate_defsize;
13044 
13045 	do {
13046 		rval = dtrace_dstate_init(&state->dts_vstate.dtvs_dynvars, sz);
13047 
13048 		if (rval == 0)
13049 			break;
13050 
13051 		if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
13052 			goto err;
13053 	} while (sz >>= 1);
13054 
13055 	opt[DTRACEOPT_DYNVARSIZE] = sz;
13056 
13057 	if (rval != 0)
13058 		goto err;
13059 
13060 	if (opt[DTRACEOPT_STATUSRATE] > dtrace_statusrate_max)
13061 		opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_max;
13062 
13063 	if (opt[DTRACEOPT_CLEANRATE] == 0)
13064 		opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
13065 
13066 	if (opt[DTRACEOPT_CLEANRATE] < dtrace_cleanrate_min)
13067 		opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_min;
13068 
13069 	if (opt[DTRACEOPT_CLEANRATE] > dtrace_cleanrate_max)
13070 		opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
13071 
13072 	hdlr.cyh_func = (cyc_func_t)dtrace_state_clean;
13073 	hdlr.cyh_arg = state;
13074 	hdlr.cyh_level = CY_LOW_LEVEL;
13075 
13076 	when.cyt_when = 0;
13077 	when.cyt_interval = opt[DTRACEOPT_CLEANRATE];
13078 
13079 	state->dts_cleaner = cyclic_add(&hdlr, &when);
13080 
13081 	hdlr.cyh_func = (cyc_func_t)dtrace_state_deadman;
13082 	hdlr.cyh_arg = state;
13083 	hdlr.cyh_level = CY_LOW_LEVEL;
13084 
13085 	when.cyt_when = 0;
13086 	when.cyt_interval = dtrace_deadman_interval;
13087 
13088 	state->dts_alive = state->dts_laststatus = dtrace_gethrtime();
13089 	state->dts_deadman = cyclic_add(&hdlr, &when);
13090 
13091 	state->dts_activity = DTRACE_ACTIVITY_WARMUP;
13092 
13093 	/*
13094 	 * Now it's time to actually fire the BEGIN probe.  We need to disable
13095 	 * interrupts here both to record the CPU on which we fired the BEGIN
13096 	 * probe (the data from this CPU will be processed first at user
13097 	 * level) and to manually activate the buffer for this CPU.
13098 	 */
13099 	cookie = dtrace_interrupt_disable();
13100 	*cpu = CPU->cpu_id;
13101 	ASSERT(state->dts_buffer[*cpu].dtb_flags & DTRACEBUF_INACTIVE);
13102 	state->dts_buffer[*cpu].dtb_flags &= ~DTRACEBUF_INACTIVE;
13103 
13104 	dtrace_probe(dtrace_probeid_begin,
13105 	    (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
13106 	dtrace_interrupt_enable(cookie);
13107 	/*
13108 	 * We may have had an exit action from a BEGIN probe; only change our
13109 	 * state to ACTIVE if we're still in WARMUP.
13110 	 */
13111 	ASSERT(state->dts_activity == DTRACE_ACTIVITY_WARMUP ||
13112 	    state->dts_activity == DTRACE_ACTIVITY_DRAINING);
13113 
13114 	if (state->dts_activity == DTRACE_ACTIVITY_WARMUP)
13115 		state->dts_activity = DTRACE_ACTIVITY_ACTIVE;
13116 
13117 	/*
13118 	 * Regardless of whether or not now we're in ACTIVE or DRAINING, we
13119 	 * want each CPU to transition its principal buffer out of the
13120 	 * INACTIVE state.  Doing this assures that no CPU will suddenly begin
13121 	 * processing an ECB halfway down a probe's ECB chain; all CPUs will
13122 	 * atomically transition from processing none of a state's ECBs to
13123 	 * processing all of them.
13124 	 */
13125 	dtrace_xcall(DTRACE_CPUALL,
13126 	    (dtrace_xcall_t)dtrace_buffer_activate, state);
13127 	goto out;
13128 
13129 err:
13130 	dtrace_buffer_free(state->dts_buffer);
13131 	dtrace_buffer_free(state->dts_aggbuffer);
13132 
13133 	if ((nspec = state->dts_nspeculations) == 0) {
13134 		ASSERT(state->dts_speculations == NULL);
13135 		goto out;
13136 	}
13137 
13138 	spec = state->dts_speculations;
13139 	ASSERT(spec != NULL);
13140 
13141 	for (i = 0; i < state->dts_nspeculations; i++) {
13142 		if ((buf = spec[i].dtsp_buffer) == NULL)
13143 			break;
13144 
13145 		dtrace_buffer_free(buf);
13146 		kmem_free(buf, bufsize);
13147 	}
13148 
13149 	kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
13150 	state->dts_nspeculations = 0;
13151 	state->dts_speculations = NULL;
13152 
13153 out:
13154 	mutex_exit(&dtrace_lock);
13155 	mutex_exit(&cpu_lock);
13156 
13157 	return (rval);
13158 }
13159 
13160 static int
13161 dtrace_state_stop(dtrace_state_t *state, processorid_t *cpu)
13162 {
13163 	dtrace_icookie_t cookie;
13164 
13165 	ASSERT(MUTEX_HELD(&dtrace_lock));
13166 
13167 	if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE &&
13168 	    state->dts_activity != DTRACE_ACTIVITY_DRAINING)
13169 		return (EINVAL);
13170 
13171 	/*
13172 	 * We'll set the activity to DTRACE_ACTIVITY_DRAINING, and issue a sync
13173 	 * to be sure that every CPU has seen it.  See below for the details
13174 	 * on why this is done.
13175 	 */
13176 	state->dts_activity = DTRACE_ACTIVITY_DRAINING;
13177 	dtrace_sync();
13178 
13179 	/*
13180 	 * By this point, it is impossible for any CPU to be still processing
13181 	 * with DTRACE_ACTIVITY_ACTIVE.  We can thus set our activity to
13182 	 * DTRACE_ACTIVITY_COOLDOWN and know that we're not racing with any
13183 	 * other CPU in dtrace_buffer_reserve().  This allows dtrace_probe()
13184 	 * and callees to know that the activity is DTRACE_ACTIVITY_COOLDOWN
13185 	 * iff we're in the END probe.
13186 	 */
13187 	state->dts_activity = DTRACE_ACTIVITY_COOLDOWN;
13188 	dtrace_sync();
13189 	ASSERT(state->dts_activity == DTRACE_ACTIVITY_COOLDOWN);
13190 
13191 	/*
13192 	 * Finally, we can release the reserve and call the END probe.  We
13193 	 * disable interrupts across calling the END probe to allow us to
13194 	 * return the CPU on which we actually called the END probe.  This
13195 	 * allows user-land to be sure that this CPU's principal buffer is
13196 	 * processed last.
13197 	 */
13198 	state->dts_reserve = 0;
13199 
13200 	cookie = dtrace_interrupt_disable();
13201 	*cpu = CPU->cpu_id;
13202 	dtrace_probe(dtrace_probeid_end,
13203 	    (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
13204 	dtrace_interrupt_enable(cookie);
13205 
13206 	state->dts_activity = DTRACE_ACTIVITY_STOPPED;
13207 	dtrace_sync();
13208 
13209 	return (0);
13210 }
13211 
13212 static int
13213 dtrace_state_option(dtrace_state_t *state, dtrace_optid_t option,
13214     dtrace_optval_t val)
13215 {
13216 	ASSERT(MUTEX_HELD(&dtrace_lock));
13217 
13218 	if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
13219 		return (EBUSY);
13220 
13221 	if (option >= DTRACEOPT_MAX)
13222 		return (EINVAL);
13223 
13224 	if (option != DTRACEOPT_CPU && val < 0)
13225 		return (EINVAL);
13226 
13227 	switch (option) {
13228 	case DTRACEOPT_DESTRUCTIVE:
13229 		if (dtrace_destructive_disallow)
13230 			return (EACCES);
13231 
13232 		state->dts_cred.dcr_destructive = 1;
13233 		break;
13234 
13235 	case DTRACEOPT_BUFSIZE:
13236 	case DTRACEOPT_DYNVARSIZE:
13237 	case DTRACEOPT_AGGSIZE:
13238 	case DTRACEOPT_SPECSIZE:
13239 	case DTRACEOPT_STRSIZE:
13240 		if (val < 0)
13241 			return (EINVAL);
13242 
13243 		if (val >= LONG_MAX) {
13244 			/*
13245 			 * If this is an otherwise negative value, set it to
13246 			 * the highest multiple of 128m less than LONG_MAX.
13247 			 * Technically, we're adjusting the size without
13248 			 * regard to the buffer resizing policy, but in fact,
13249 			 * this has no effect -- if we set the buffer size to
13250 			 * ~LONG_MAX and the buffer policy is ultimately set to
13251 			 * be "manual", the buffer allocation is guaranteed to
13252 			 * fail, if only because the allocation requires two
13253 			 * buffers.  (We set the the size to the highest
13254 			 * multiple of 128m because it ensures that the size
13255 			 * will remain a multiple of a megabyte when
13256 			 * repeatedly halved -- all the way down to 15m.)
13257 			 */
13258 			val = LONG_MAX - (1 << 27) + 1;
13259 		}
13260 	}
13261 
13262 	state->dts_options[option] = val;
13263 
13264 	return (0);
13265 }
13266 
13267 static void
13268 dtrace_state_destroy(dtrace_state_t *state)
13269 {
13270 	dtrace_ecb_t *ecb;
13271 	dtrace_vstate_t *vstate = &state->dts_vstate;
13272 	minor_t minor = getminor(state->dts_dev);
13273 	int i, bufsize = NCPU * sizeof (dtrace_buffer_t);
13274 	dtrace_speculation_t *spec = state->dts_speculations;
13275 	int nspec = state->dts_nspeculations;
13276 	uint32_t match;
13277 
13278 	ASSERT(MUTEX_HELD(&dtrace_lock));
13279 	ASSERT(MUTEX_HELD(&cpu_lock));
13280 
13281 	/*
13282 	 * First, retract any retained enablings for this state.
13283 	 */
13284 	dtrace_enabling_retract(state);
13285 	ASSERT(state->dts_nretained == 0);
13286 
13287 	if (state->dts_activity == DTRACE_ACTIVITY_ACTIVE ||
13288 	    state->dts_activity == DTRACE_ACTIVITY_DRAINING) {
13289 		/*
13290 		 * We have managed to come into dtrace_state_destroy() on a
13291 		 * hot enabling -- almost certainly because of a disorderly
13292 		 * shutdown of a consumer.  (That is, a consumer that is
13293 		 * exiting without having called dtrace_stop().) In this case,
13294 		 * we're going to set our activity to be KILLED, and then
13295 		 * issue a sync to be sure that everyone is out of probe
13296 		 * context before we start blowing away ECBs.
13297 		 */
13298 		state->dts_activity = DTRACE_ACTIVITY_KILLED;
13299 		dtrace_sync();
13300 	}
13301 
13302 	/*
13303 	 * Release the credential hold we took in dtrace_state_create().
13304 	 */
13305 	if (state->dts_cred.dcr_cred != NULL)
13306 		crfree(state->dts_cred.dcr_cred);
13307 
13308 	/*
13309 	 * Now we can safely disable and destroy any enabled probes.  Because
13310 	 * any DTRACE_PRIV_KERNEL probes may actually be slowing our progress
13311 	 * (especially if they're all enabled), we take two passes through the
13312 	 * ECBs:  in the first, we disable just DTRACE_PRIV_KERNEL probes, and
13313 	 * in the second we disable whatever is left over.
13314 	 */
13315 	for (match = DTRACE_PRIV_KERNEL; ; match = 0) {
13316 		for (i = 0; i < state->dts_necbs; i++) {
13317 			if ((ecb = state->dts_ecbs[i]) == NULL)
13318 				continue;
13319 
13320 			if (match && ecb->dte_probe != NULL) {
13321 				dtrace_probe_t *probe = ecb->dte_probe;
13322 				dtrace_provider_t *prov = probe->dtpr_provider;
13323 
13324 				if (!(prov->dtpv_priv.dtpp_flags & match))
13325 					continue;
13326 			}
13327 
13328 			dtrace_ecb_disable(ecb);
13329 			dtrace_ecb_destroy(ecb);
13330 		}
13331 
13332 		if (!match)
13333 			break;
13334 	}
13335 
13336 	/*
13337 	 * Before we free the buffers, perform one more sync to assure that
13338 	 * every CPU is out of probe context.
13339 	 */
13340 	dtrace_sync();
13341 
13342 	dtrace_buffer_free(state->dts_buffer);
13343 	dtrace_buffer_free(state->dts_aggbuffer);
13344 
13345 	for (i = 0; i < nspec; i++)
13346 		dtrace_buffer_free(spec[i].dtsp_buffer);
13347 
13348 	if (state->dts_cleaner != CYCLIC_NONE)
13349 		cyclic_remove(state->dts_cleaner);
13350 
13351 	if (state->dts_deadman != CYCLIC_NONE)
13352 		cyclic_remove(state->dts_deadman);
13353 
13354 	dtrace_dstate_fini(&vstate->dtvs_dynvars);
13355 	dtrace_vstate_fini(vstate);
13356 	kmem_free(state->dts_ecbs, state->dts_necbs * sizeof (dtrace_ecb_t *));
13357 
13358 	if (state->dts_aggregations != NULL) {
13359 #ifdef DEBUG
13360 		for (i = 0; i < state->dts_naggregations; i++)
13361 			ASSERT(state->dts_aggregations[i] == NULL);
13362 #endif
13363 		ASSERT(state->dts_naggregations > 0);
13364 		kmem_free(state->dts_aggregations,
13365 		    state->dts_naggregations * sizeof (dtrace_aggregation_t *));
13366 	}
13367 
13368 	kmem_free(state->dts_buffer, bufsize);
13369 	kmem_free(state->dts_aggbuffer, bufsize);
13370 
13371 	for (i = 0; i < nspec; i++)
13372 		kmem_free(spec[i].dtsp_buffer, bufsize);
13373 
13374 	kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
13375 
13376 	dtrace_format_destroy(state);
13377 
13378 	vmem_destroy(state->dts_aggid_arena);
13379 	ddi_soft_state_free(dtrace_softstate, minor);
13380 	vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
13381 }
13382 
13383 /*
13384  * DTrace Anonymous Enabling Functions
13385  */
13386 static dtrace_state_t *
13387 dtrace_anon_grab(void)
13388 {
13389 	dtrace_state_t *state;
13390 
13391 	ASSERT(MUTEX_HELD(&dtrace_lock));
13392 
13393 	if ((state = dtrace_anon.dta_state) == NULL) {
13394 		ASSERT(dtrace_anon.dta_enabling == NULL);
13395 		return (NULL);
13396 	}
13397 
13398 	ASSERT(dtrace_anon.dta_enabling != NULL);
13399 	ASSERT(dtrace_retained != NULL);
13400 
13401 	dtrace_enabling_destroy(dtrace_anon.dta_enabling);
13402 	dtrace_anon.dta_enabling = NULL;
13403 	dtrace_anon.dta_state = NULL;
13404 
13405 	return (state);
13406 }
13407 
13408 static void
13409 dtrace_anon_property(void)
13410 {
13411 	int i, rv;
13412 	dtrace_state_t *state;
13413 	dof_hdr_t *dof;
13414 	char c[32];		/* enough for "dof-data-" + digits */
13415 
13416 	ASSERT(MUTEX_HELD(&dtrace_lock));
13417 	ASSERT(MUTEX_HELD(&cpu_lock));
13418 
13419 	for (i = 0; ; i++) {
13420 		(void) snprintf(c, sizeof (c), "dof-data-%d", i);
13421 
13422 		dtrace_err_verbose = 1;
13423 
13424 		if ((dof = dtrace_dof_property(c)) == NULL) {
13425 			dtrace_err_verbose = 0;
13426 			break;
13427 		}
13428 
13429 		/*
13430 		 * We want to create anonymous state, so we need to transition
13431 		 * the kernel debugger to indicate that DTrace is active.  If
13432 		 * this fails (e.g. because the debugger has modified text in
13433 		 * some way), we won't continue with the processing.
13434 		 */
13435 		if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
13436 			cmn_err(CE_NOTE, "kernel debugger active; anonymous "
13437 			    "enabling ignored.");
13438 			dtrace_dof_destroy(dof);
13439 			break;
13440 		}
13441 
13442 		/*
13443 		 * If we haven't allocated an anonymous state, we'll do so now.
13444 		 */
13445 		if ((state = dtrace_anon.dta_state) == NULL) {
13446 			state = dtrace_state_create(NULL, NULL);
13447 			dtrace_anon.dta_state = state;
13448 
13449 			if (state == NULL) {
13450 				/*
13451 				 * This basically shouldn't happen:  the only
13452 				 * failure mode from dtrace_state_create() is a
13453 				 * failure of ddi_soft_state_zalloc() that
13454 				 * itself should never happen.  Still, the
13455 				 * interface allows for a failure mode, and
13456 				 * we want to fail as gracefully as possible:
13457 				 * we'll emit an error message and cease
13458 				 * processing anonymous state in this case.
13459 				 */
13460 				cmn_err(CE_WARN, "failed to create "
13461 				    "anonymous state");
13462 				dtrace_dof_destroy(dof);
13463 				break;
13464 			}
13465 		}
13466 
13467 		rv = dtrace_dof_slurp(dof, &state->dts_vstate, CRED(),
13468 		    &dtrace_anon.dta_enabling, 0, B_TRUE);
13469 
13470 		if (rv == 0)
13471 			rv = dtrace_dof_options(dof, state);
13472 
13473 		dtrace_err_verbose = 0;
13474 		dtrace_dof_destroy(dof);
13475 
13476 		if (rv != 0) {
13477 			/*
13478 			 * This is malformed DOF; chuck any anonymous state
13479 			 * that we created.
13480 			 */
13481 			ASSERT(dtrace_anon.dta_enabling == NULL);
13482 			dtrace_state_destroy(state);
13483 			dtrace_anon.dta_state = NULL;
13484 			break;
13485 		}
13486 
13487 		ASSERT(dtrace_anon.dta_enabling != NULL);
13488 	}
13489 
13490 	if (dtrace_anon.dta_enabling != NULL) {
13491 		int rval;
13492 
13493 		/*
13494 		 * dtrace_enabling_retain() can only fail because we are
13495 		 * trying to retain more enablings than are allowed -- but
13496 		 * we only have one anonymous enabling, and we are guaranteed
13497 		 * to be allowed at least one retained enabling; we assert
13498 		 * that dtrace_enabling_retain() returns success.
13499 		 */
13500 		rval = dtrace_enabling_retain(dtrace_anon.dta_enabling);
13501 		ASSERT(rval == 0);
13502 
13503 		dtrace_enabling_dump(dtrace_anon.dta_enabling);
13504 	}
13505 }
13506 
13507 /*
13508  * DTrace Helper Functions
13509  */
13510 static void
13511 dtrace_helper_trace(dtrace_helper_action_t *helper,
13512     dtrace_mstate_t *mstate, dtrace_vstate_t *vstate, int where)
13513 {
13514 	uint32_t size, next, nnext, i;
13515 	dtrace_helptrace_t *ent;
13516 	uint16_t flags = cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
13517 
13518 	if (!dtrace_helptrace_enabled)
13519 		return;
13520 
13521 	ASSERT(vstate->dtvs_nlocals <= dtrace_helptrace_nlocals);
13522 
13523 	/*
13524 	 * What would a tracing framework be without its own tracing
13525 	 * framework?  (Well, a hell of a lot simpler, for starters...)
13526 	 */
13527 	size = sizeof (dtrace_helptrace_t) + dtrace_helptrace_nlocals *
13528 	    sizeof (uint64_t) - sizeof (uint64_t);
13529 
13530 	/*
13531 	 * Iterate until we can allocate a slot in the trace buffer.
13532 	 */
13533 	do {
13534 		next = dtrace_helptrace_next;
13535 
13536 		if (next + size < dtrace_helptrace_bufsize) {
13537 			nnext = next + size;
13538 		} else {
13539 			nnext = size;
13540 		}
13541 	} while (dtrace_cas32(&dtrace_helptrace_next, next, nnext) != next);
13542 
13543 	/*
13544 	 * We have our slot; fill it in.
13545 	 */
13546 	if (nnext == size)
13547 		next = 0;
13548 
13549 	ent = (dtrace_helptrace_t *)&dtrace_helptrace_buffer[next];
13550 	ent->dtht_helper = helper;
13551 	ent->dtht_where = where;
13552 	ent->dtht_nlocals = vstate->dtvs_nlocals;
13553 
13554 	ent->dtht_fltoffs = (mstate->dtms_present & DTRACE_MSTATE_FLTOFFS) ?
13555 	    mstate->dtms_fltoffs : -1;
13556 	ent->dtht_fault = DTRACE_FLAGS2FLT(flags);
13557 	ent->dtht_illval = cpu_core[CPU->cpu_id].cpuc_dtrace_illval;
13558 
13559 	for (i = 0; i < vstate->dtvs_nlocals; i++) {
13560 		dtrace_statvar_t *svar;
13561 
13562 		if ((svar = vstate->dtvs_locals[i]) == NULL)
13563 			continue;
13564 
13565 		ASSERT(svar->dtsv_size >= NCPU * sizeof (uint64_t));
13566 		ent->dtht_locals[i] =
13567 		    ((uint64_t *)(uintptr_t)svar->dtsv_data)[CPU->cpu_id];
13568 	}
13569 }
13570 
13571 static uint64_t
13572 dtrace_helper(int which, dtrace_mstate_t *mstate,
13573     dtrace_state_t *state, uint64_t arg0, uint64_t arg1)
13574 {
13575 	uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
13576 	uint64_t sarg0 = mstate->dtms_arg[0];
13577 	uint64_t sarg1 = mstate->dtms_arg[1];
13578 	uint64_t rval;
13579 	dtrace_helpers_t *helpers = curproc->p_dtrace_helpers;
13580 	dtrace_helper_action_t *helper;
13581 	dtrace_vstate_t *vstate;
13582 	dtrace_difo_t *pred;
13583 	int i, trace = dtrace_helptrace_enabled;
13584 
13585 	ASSERT(which >= 0 && which < DTRACE_NHELPER_ACTIONS);
13586 
13587 	if (helpers == NULL)
13588 		return (0);
13589 
13590 	if ((helper = helpers->dthps_actions[which]) == NULL)
13591 		return (0);
13592 
13593 	vstate = &helpers->dthps_vstate;
13594 	mstate->dtms_arg[0] = arg0;
13595 	mstate->dtms_arg[1] = arg1;
13596 
13597 	/*
13598 	 * Now iterate over each helper.  If its predicate evaluates to 'true',
13599 	 * we'll call the corresponding actions.  Note that the below calls
13600 	 * to dtrace_dif_emulate() may set faults in machine state.  This is
13601 	 * okay:  our caller (the outer dtrace_dif_emulate()) will simply plow
13602 	 * the stored DIF offset with its own (which is the desired behavior).
13603 	 * Also, note the calls to dtrace_dif_emulate() may allocate scratch
13604 	 * from machine state; this is okay, too.
13605 	 */
13606 	for (; helper != NULL; helper = helper->dtha_next) {
13607 		if ((pred = helper->dtha_predicate) != NULL) {
13608 			if (trace)
13609 				dtrace_helper_trace(helper, mstate, vstate, 0);
13610 
13611 			if (!dtrace_dif_emulate(pred, mstate, vstate, state))
13612 				goto next;
13613 
13614 			if (*flags & CPU_DTRACE_FAULT)
13615 				goto err;
13616 		}
13617 
13618 		for (i = 0; i < helper->dtha_nactions; i++) {
13619 			if (trace)
13620 				dtrace_helper_trace(helper,
13621 				    mstate, vstate, i + 1);
13622 
13623 			rval = dtrace_dif_emulate(helper->dtha_actions[i],
13624 			    mstate, vstate, state);
13625 
13626 			if (*flags & CPU_DTRACE_FAULT)
13627 				goto err;
13628 		}
13629 
13630 next:
13631 		if (trace)
13632 			dtrace_helper_trace(helper, mstate, vstate,
13633 			    DTRACE_HELPTRACE_NEXT);
13634 	}
13635 
13636 	if (trace)
13637 		dtrace_helper_trace(helper, mstate, vstate,
13638 		    DTRACE_HELPTRACE_DONE);
13639 
13640 	/*
13641 	 * Restore the arg0 that we saved upon entry.
13642 	 */
13643 	mstate->dtms_arg[0] = sarg0;
13644 	mstate->dtms_arg[1] = sarg1;
13645 
13646 	return (rval);
13647 
13648 err:
13649 	if (trace)
13650 		dtrace_helper_trace(helper, mstate, vstate,
13651 		    DTRACE_HELPTRACE_ERR);
13652 
13653 	/*
13654 	 * Restore the arg0 that we saved upon entry.
13655 	 */
13656 	mstate->dtms_arg[0] = sarg0;
13657 	mstate->dtms_arg[1] = sarg1;
13658 
13659 	return (NULL);
13660 }
13661 
13662 static void
13663 dtrace_helper_action_destroy(dtrace_helper_action_t *helper,
13664     dtrace_vstate_t *vstate)
13665 {
13666 	int i;
13667 
13668 	if (helper->dtha_predicate != NULL)
13669 		dtrace_difo_release(helper->dtha_predicate, vstate);
13670 
13671 	for (i = 0; i < helper->dtha_nactions; i++) {
13672 		ASSERT(helper->dtha_actions[i] != NULL);
13673 		dtrace_difo_release(helper->dtha_actions[i], vstate);
13674 	}
13675 
13676 	kmem_free(helper->dtha_actions,
13677 	    helper->dtha_nactions * sizeof (dtrace_difo_t *));
13678 	kmem_free(helper, sizeof (dtrace_helper_action_t));
13679 }
13680 
13681 static int
13682 dtrace_helper_destroygen(int gen)
13683 {
13684 	proc_t *p = curproc;
13685 	dtrace_helpers_t *help = p->p_dtrace_helpers;
13686 	dtrace_vstate_t *vstate;
13687 	int i;
13688 
13689 	ASSERT(MUTEX_HELD(&dtrace_lock));
13690 
13691 	if (help == NULL || gen > help->dthps_generation)
13692 		return (EINVAL);
13693 
13694 	vstate = &help->dthps_vstate;
13695 
13696 	for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
13697 		dtrace_helper_action_t *last = NULL, *h, *next;
13698 
13699 		for (h = help->dthps_actions[i]; h != NULL; h = next) {
13700 			next = h->dtha_next;
13701 
13702 			if (h->dtha_generation == gen) {
13703 				if (last != NULL) {
13704 					last->dtha_next = next;
13705 				} else {
13706 					help->dthps_actions[i] = next;
13707 				}
13708 
13709 				dtrace_helper_action_destroy(h, vstate);
13710 			} else {
13711 				last = h;
13712 			}
13713 		}
13714 	}
13715 
13716 	/*
13717 	 * Interate until we've cleared out all helper providers with the
13718 	 * given generation number.
13719 	 */
13720 	for (;;) {
13721 		dtrace_helper_provider_t *prov;
13722 
13723 		/*
13724 		 * Look for a helper provider with the right generation. We
13725 		 * have to start back at the beginning of the list each time
13726 		 * because we drop dtrace_lock. It's unlikely that we'll make
13727 		 * more than two passes.
13728 		 */
13729 		for (i = 0; i < help->dthps_nprovs; i++) {
13730 			prov = help->dthps_provs[i];
13731 
13732 			if (prov->dthp_generation == gen)
13733 				break;
13734 		}
13735 
13736 		/*
13737 		 * If there were no matches, we're done.
13738 		 */
13739 		if (i == help->dthps_nprovs)
13740 			break;
13741 
13742 		/*
13743 		 * Move the last helper provider into this slot.
13744 		 */
13745 		help->dthps_nprovs--;
13746 		help->dthps_provs[i] = help->dthps_provs[help->dthps_nprovs];
13747 		help->dthps_provs[help->dthps_nprovs] = NULL;
13748 
13749 		mutex_exit(&dtrace_lock);
13750 
13751 		/*
13752 		 * If we have a meta provider, remove this helper provider.
13753 		 */
13754 		mutex_enter(&dtrace_meta_lock);
13755 		if (dtrace_meta_pid != NULL) {
13756 			ASSERT(dtrace_deferred_pid == NULL);
13757 			dtrace_helper_provider_remove(&prov->dthp_prov,
13758 			    p->p_pid);
13759 		}
13760 		mutex_exit(&dtrace_meta_lock);
13761 
13762 		dtrace_helper_provider_destroy(prov);
13763 
13764 		mutex_enter(&dtrace_lock);
13765 	}
13766 
13767 	return (0);
13768 }
13769 
13770 static int
13771 dtrace_helper_validate(dtrace_helper_action_t *helper)
13772 {
13773 	int err = 0, i;
13774 	dtrace_difo_t *dp;
13775 
13776 	if ((dp = helper->dtha_predicate) != NULL)
13777 		err += dtrace_difo_validate_helper(dp);
13778 
13779 	for (i = 0; i < helper->dtha_nactions; i++)
13780 		err += dtrace_difo_validate_helper(helper->dtha_actions[i]);
13781 
13782 	return (err == 0);
13783 }
13784 
13785 static int
13786 dtrace_helper_action_add(int which, dtrace_ecbdesc_t *ep)
13787 {
13788 	dtrace_helpers_t *help;
13789 	dtrace_helper_action_t *helper, *last;
13790 	dtrace_actdesc_t *act;
13791 	dtrace_vstate_t *vstate;
13792 	dtrace_predicate_t *pred;
13793 	int count = 0, nactions = 0, i;
13794 
13795 	if (which < 0 || which >= DTRACE_NHELPER_ACTIONS)
13796 		return (EINVAL);
13797 
13798 	help = curproc->p_dtrace_helpers;
13799 	last = help->dthps_actions[which];
13800 	vstate = &help->dthps_vstate;
13801 
13802 	for (count = 0; last != NULL; last = last->dtha_next) {
13803 		count++;
13804 		if (last->dtha_next == NULL)
13805 			break;
13806 	}
13807 
13808 	/*
13809 	 * If we already have dtrace_helper_actions_max helper actions for this
13810 	 * helper action type, we'll refuse to add a new one.
13811 	 */
13812 	if (count >= dtrace_helper_actions_max)
13813 		return (ENOSPC);
13814 
13815 	helper = kmem_zalloc(sizeof (dtrace_helper_action_t), KM_SLEEP);
13816 	helper->dtha_generation = help->dthps_generation;
13817 
13818 	if ((pred = ep->dted_pred.dtpdd_predicate) != NULL) {
13819 		ASSERT(pred->dtp_difo != NULL);
13820 		dtrace_difo_hold(pred->dtp_difo);
13821 		helper->dtha_predicate = pred->dtp_difo;
13822 	}
13823 
13824 	for (act = ep->dted_action; act != NULL; act = act->dtad_next) {
13825 		if (act->dtad_kind != DTRACEACT_DIFEXPR)
13826 			goto err;
13827 
13828 		if (act->dtad_difo == NULL)
13829 			goto err;
13830 
13831 		nactions++;
13832 	}
13833 
13834 	helper->dtha_actions = kmem_zalloc(sizeof (dtrace_difo_t *) *
13835 	    (helper->dtha_nactions = nactions), KM_SLEEP);
13836 
13837 	for (act = ep->dted_action, i = 0; act != NULL; act = act->dtad_next) {
13838 		dtrace_difo_hold(act->dtad_difo);
13839 		helper->dtha_actions[i++] = act->dtad_difo;
13840 	}
13841 
13842 	if (!dtrace_helper_validate(helper))
13843 		goto err;
13844 
13845 	if (last == NULL) {
13846 		help->dthps_actions[which] = helper;
13847 	} else {
13848 		last->dtha_next = helper;
13849 	}
13850 
13851 	if (vstate->dtvs_nlocals > dtrace_helptrace_nlocals) {
13852 		dtrace_helptrace_nlocals = vstate->dtvs_nlocals;
13853 		dtrace_helptrace_next = 0;
13854 	}
13855 
13856 	return (0);
13857 err:
13858 	dtrace_helper_action_destroy(helper, vstate);
13859 	return (EINVAL);
13860 }
13861 
13862 static void
13863 dtrace_helper_provider_register(proc_t *p, dtrace_helpers_t *help,
13864     dof_helper_t *dofhp)
13865 {
13866 	ASSERT(MUTEX_NOT_HELD(&dtrace_lock));
13867 
13868 	mutex_enter(&dtrace_meta_lock);
13869 	mutex_enter(&dtrace_lock);
13870 
13871 	if (!dtrace_attached() || dtrace_meta_pid == NULL) {
13872 		/*
13873 		 * If the dtrace module is loaded but not attached, or if
13874 		 * there aren't isn't a meta provider registered to deal with
13875 		 * these provider descriptions, we need to postpone creating
13876 		 * the actual providers until later.
13877 		 */
13878 
13879 		if (help->dthps_next == NULL && help->dthps_prev == NULL &&
13880 		    dtrace_deferred_pid != help) {
13881 			help->dthps_deferred = 1;
13882 			help->dthps_pid = p->p_pid;
13883 			help->dthps_next = dtrace_deferred_pid;
13884 			help->dthps_prev = NULL;
13885 			if (dtrace_deferred_pid != NULL)
13886 				dtrace_deferred_pid->dthps_prev = help;
13887 			dtrace_deferred_pid = help;
13888 		}
13889 
13890 		mutex_exit(&dtrace_lock);
13891 
13892 	} else if (dofhp != NULL) {
13893 		/*
13894 		 * If the dtrace module is loaded and we have a particular
13895 		 * helper provider description, pass that off to the
13896 		 * meta provider.
13897 		 */
13898 
13899 		mutex_exit(&dtrace_lock);
13900 
13901 		dtrace_helper_provide(dofhp, p->p_pid);
13902 
13903 	} else {
13904 		/*
13905 		 * Otherwise, just pass all the helper provider descriptions
13906 		 * off to the meta provider.
13907 		 */
13908 
13909 		int i;
13910 		mutex_exit(&dtrace_lock);
13911 
13912 		for (i = 0; i < help->dthps_nprovs; i++) {
13913 			dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
13914 			    p->p_pid);
13915 		}
13916 	}
13917 
13918 	mutex_exit(&dtrace_meta_lock);
13919 }
13920 
13921 static int
13922 dtrace_helper_provider_add(dof_helper_t *dofhp, int gen)
13923 {
13924 	dtrace_helpers_t *help;
13925 	dtrace_helper_provider_t *hprov, **tmp_provs;
13926 	uint_t tmp_maxprovs, i;
13927 
13928 	ASSERT(MUTEX_HELD(&dtrace_lock));
13929 
13930 	help = curproc->p_dtrace_helpers;
13931 	ASSERT(help != NULL);
13932 
13933 	/*
13934 	 * If we already have dtrace_helper_providers_max helper providers,
13935 	 * we're refuse to add a new one.
13936 	 */
13937 	if (help->dthps_nprovs >= dtrace_helper_providers_max)
13938 		return (ENOSPC);
13939 
13940 	/*
13941 	 * Check to make sure this isn't a duplicate.
13942 	 */
13943 	for (i = 0; i < help->dthps_nprovs; i++) {
13944 		if (dofhp->dofhp_addr ==
13945 		    help->dthps_provs[i]->dthp_prov.dofhp_addr)
13946 			return (EALREADY);
13947 	}
13948 
13949 	hprov = kmem_zalloc(sizeof (dtrace_helper_provider_t), KM_SLEEP);
13950 	hprov->dthp_prov = *dofhp;
13951 	hprov->dthp_ref = 1;
13952 	hprov->dthp_generation = gen;
13953 
13954 	/*
13955 	 * Allocate a bigger table for helper providers if it's already full.
13956 	 */
13957 	if (help->dthps_maxprovs == help->dthps_nprovs) {
13958 		tmp_maxprovs = help->dthps_maxprovs;
13959 		tmp_provs = help->dthps_provs;
13960 
13961 		if (help->dthps_maxprovs == 0)
13962 			help->dthps_maxprovs = 2;
13963 		else
13964 			help->dthps_maxprovs *= 2;
13965 		if (help->dthps_maxprovs > dtrace_helper_providers_max)
13966 			help->dthps_maxprovs = dtrace_helper_providers_max;
13967 
13968 		ASSERT(tmp_maxprovs < help->dthps_maxprovs);
13969 
13970 		help->dthps_provs = kmem_zalloc(help->dthps_maxprovs *
13971 		    sizeof (dtrace_helper_provider_t *), KM_SLEEP);
13972 
13973 		if (tmp_provs != NULL) {
13974 			bcopy(tmp_provs, help->dthps_provs, tmp_maxprovs *
13975 			    sizeof (dtrace_helper_provider_t *));
13976 			kmem_free(tmp_provs, tmp_maxprovs *
13977 			    sizeof (dtrace_helper_provider_t *));
13978 		}
13979 	}
13980 
13981 	help->dthps_provs[help->dthps_nprovs] = hprov;
13982 	help->dthps_nprovs++;
13983 
13984 	return (0);
13985 }
13986 
13987 static void
13988 dtrace_helper_provider_destroy(dtrace_helper_provider_t *hprov)
13989 {
13990 	mutex_enter(&dtrace_lock);
13991 
13992 	if (--hprov->dthp_ref == 0) {
13993 		dof_hdr_t *dof;
13994 		mutex_exit(&dtrace_lock);
13995 		dof = (dof_hdr_t *)(uintptr_t)hprov->dthp_prov.dofhp_dof;
13996 		dtrace_dof_destroy(dof);
13997 		kmem_free(hprov, sizeof (dtrace_helper_provider_t));
13998 	} else {
13999 		mutex_exit(&dtrace_lock);
14000 	}
14001 }
14002 
14003 static int
14004 dtrace_helper_provider_validate(dof_hdr_t *dof, dof_sec_t *sec)
14005 {
14006 	uintptr_t daddr = (uintptr_t)dof;
14007 	dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
14008 	dof_provider_t *provider;
14009 	dof_probe_t *probe;
14010 	uint8_t *arg;
14011 	char *strtab, *typestr;
14012 	dof_stridx_t typeidx;
14013 	size_t typesz;
14014 	uint_t nprobes, j, k;
14015 
14016 	ASSERT(sec->dofs_type == DOF_SECT_PROVIDER);
14017 
14018 	if (sec->dofs_offset & (sizeof (uint_t) - 1)) {
14019 		dtrace_dof_error(dof, "misaligned section offset");
14020 		return (-1);
14021 	}
14022 
14023 	/*
14024 	 * The section needs to be large enough to contain the DOF provider
14025 	 * structure appropriate for the given version.
14026 	 */
14027 	if (sec->dofs_size <
14028 	    ((dof->dofh_ident[DOF_ID_VERSION] == DOF_VERSION_1) ?
14029 	    offsetof(dof_provider_t, dofpv_prenoffs) :
14030 	    sizeof (dof_provider_t))) {
14031 		dtrace_dof_error(dof, "provider section too small");
14032 		return (-1);
14033 	}
14034 
14035 	provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
14036 	str_sec = dtrace_dof_sect(dof, DOF_SECT_STRTAB, provider->dofpv_strtab);
14037 	prb_sec = dtrace_dof_sect(dof, DOF_SECT_PROBES, provider->dofpv_probes);
14038 	arg_sec = dtrace_dof_sect(dof, DOF_SECT_PRARGS, provider->dofpv_prargs);
14039 	off_sec = dtrace_dof_sect(dof, DOF_SECT_PROFFS, provider->dofpv_proffs);
14040 
14041 	if (str_sec == NULL || prb_sec == NULL ||
14042 	    arg_sec == NULL || off_sec == NULL)
14043 		return (-1);
14044 
14045 	enoff_sec = NULL;
14046 
14047 	if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
14048 	    provider->dofpv_prenoffs != DOF_SECT_NONE &&
14049 	    (enoff_sec = dtrace_dof_sect(dof, DOF_SECT_PRENOFFS,
14050 	    provider->dofpv_prenoffs)) == NULL)
14051 		return (-1);
14052 
14053 	strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
14054 
14055 	if (provider->dofpv_name >= str_sec->dofs_size ||
14056 	    strlen(strtab + provider->dofpv_name) >= DTRACE_PROVNAMELEN) {
14057 		dtrace_dof_error(dof, "invalid provider name");
14058 		return (-1);
14059 	}
14060 
14061 	if (prb_sec->dofs_entsize == 0 ||
14062 	    prb_sec->dofs_entsize > prb_sec->dofs_size) {
14063 		dtrace_dof_error(dof, "invalid entry size");
14064 		return (-1);
14065 	}
14066 
14067 	if (prb_sec->dofs_entsize & (sizeof (uintptr_t) - 1)) {
14068 		dtrace_dof_error(dof, "misaligned entry size");
14069 		return (-1);
14070 	}
14071 
14072 	if (off_sec->dofs_entsize != sizeof (uint32_t)) {
14073 		dtrace_dof_error(dof, "invalid entry size");
14074 		return (-1);
14075 	}
14076 
14077 	if (off_sec->dofs_offset & (sizeof (uint32_t) - 1)) {
14078 		dtrace_dof_error(dof, "misaligned section offset");
14079 		return (-1);
14080 	}
14081 
14082 	if (arg_sec->dofs_entsize != sizeof (uint8_t)) {
14083 		dtrace_dof_error(dof, "invalid entry size");
14084 		return (-1);
14085 	}
14086 
14087 	arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
14088 
14089 	nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
14090 
14091 	/*
14092 	 * Take a pass through the probes to check for errors.
14093 	 */
14094 	for (j = 0; j < nprobes; j++) {
14095 		probe = (dof_probe_t *)(uintptr_t)(daddr +
14096 		    prb_sec->dofs_offset + j * prb_sec->dofs_entsize);
14097 
14098 		if (probe->dofpr_func >= str_sec->dofs_size) {
14099 			dtrace_dof_error(dof, "invalid function name");
14100 			return (-1);
14101 		}
14102 
14103 		if (strlen(strtab + probe->dofpr_func) >= DTRACE_FUNCNAMELEN) {
14104 			dtrace_dof_error(dof, "function name too long");
14105 			return (-1);
14106 		}
14107 
14108 		if (probe->dofpr_name >= str_sec->dofs_size ||
14109 		    strlen(strtab + probe->dofpr_name) >= DTRACE_NAMELEN) {
14110 			dtrace_dof_error(dof, "invalid probe name");
14111 			return (-1);
14112 		}
14113 
14114 		/*
14115 		 * The offset count must not wrap the index, and the offsets
14116 		 * must also not overflow the section's data.
14117 		 */
14118 		if (probe->dofpr_offidx + probe->dofpr_noffs <
14119 		    probe->dofpr_offidx ||
14120 		    (probe->dofpr_offidx + probe->dofpr_noffs) *
14121 		    off_sec->dofs_entsize > off_sec->dofs_size) {
14122 			dtrace_dof_error(dof, "invalid probe offset");
14123 			return (-1);
14124 		}
14125 
14126 		if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1) {
14127 			/*
14128 			 * If there's no is-enabled offset section, make sure
14129 			 * there aren't any is-enabled offsets. Otherwise
14130 			 * perform the same checks as for probe offsets
14131 			 * (immediately above).
14132 			 */
14133 			if (enoff_sec == NULL) {
14134 				if (probe->dofpr_enoffidx != 0 ||
14135 				    probe->dofpr_nenoffs != 0) {
14136 					dtrace_dof_error(dof, "is-enabled "
14137 					    "offsets with null section");
14138 					return (-1);
14139 				}
14140 			} else if (probe->dofpr_enoffidx +
14141 			    probe->dofpr_nenoffs < probe->dofpr_enoffidx ||
14142 			    (probe->dofpr_enoffidx + probe->dofpr_nenoffs) *
14143 			    enoff_sec->dofs_entsize > enoff_sec->dofs_size) {
14144 				dtrace_dof_error(dof, "invalid is-enabled "
14145 				    "offset");
14146 				return (-1);
14147 			}
14148 
14149 			if (probe->dofpr_noffs + probe->dofpr_nenoffs == 0) {
14150 				dtrace_dof_error(dof, "zero probe and "
14151 				    "is-enabled offsets");
14152 				return (-1);
14153 			}
14154 		} else if (probe->dofpr_noffs == 0) {
14155 			dtrace_dof_error(dof, "zero probe offsets");
14156 			return (-1);
14157 		}
14158 
14159 		if (probe->dofpr_argidx + probe->dofpr_xargc <
14160 		    probe->dofpr_argidx ||
14161 		    (probe->dofpr_argidx + probe->dofpr_xargc) *
14162 		    arg_sec->dofs_entsize > arg_sec->dofs_size) {
14163 			dtrace_dof_error(dof, "invalid args");
14164 			return (-1);
14165 		}
14166 
14167 		typeidx = probe->dofpr_nargv;
14168 		typestr = strtab + probe->dofpr_nargv;
14169 		for (k = 0; k < probe->dofpr_nargc; k++) {
14170 			if (typeidx >= str_sec->dofs_size) {
14171 				dtrace_dof_error(dof, "bad "
14172 				    "native argument type");
14173 				return (-1);
14174 			}
14175 
14176 			typesz = strlen(typestr) + 1;
14177 			if (typesz > DTRACE_ARGTYPELEN) {
14178 				dtrace_dof_error(dof, "native "
14179 				    "argument type too long");
14180 				return (-1);
14181 			}
14182 			typeidx += typesz;
14183 			typestr += typesz;
14184 		}
14185 
14186 		typeidx = probe->dofpr_xargv;
14187 		typestr = strtab + probe->dofpr_xargv;
14188 		for (k = 0; k < probe->dofpr_xargc; k++) {
14189 			if (arg[probe->dofpr_argidx + k] > probe->dofpr_nargc) {
14190 				dtrace_dof_error(dof, "bad "
14191 				    "native argument index");
14192 				return (-1);
14193 			}
14194 
14195 			if (typeidx >= str_sec->dofs_size) {
14196 				dtrace_dof_error(dof, "bad "
14197 				    "translated argument type");
14198 				return (-1);
14199 			}
14200 
14201 			typesz = strlen(typestr) + 1;
14202 			if (typesz > DTRACE_ARGTYPELEN) {
14203 				dtrace_dof_error(dof, "translated argument "
14204 				    "type too long");
14205 				return (-1);
14206 			}
14207 
14208 			typeidx += typesz;
14209 			typestr += typesz;
14210 		}
14211 	}
14212 
14213 	return (0);
14214 }
14215 
14216 static int
14217 dtrace_helper_slurp(dof_hdr_t *dof, dof_helper_t *dhp)
14218 {
14219 	dtrace_helpers_t *help;
14220 	dtrace_vstate_t *vstate;
14221 	dtrace_enabling_t *enab = NULL;
14222 	int i, gen, rv, nhelpers = 0, nprovs = 0, destroy = 1;
14223 	uintptr_t daddr = (uintptr_t)dof;
14224 
14225 	ASSERT(MUTEX_HELD(&dtrace_lock));
14226 
14227 	if ((help = curproc->p_dtrace_helpers) == NULL)
14228 		help = dtrace_helpers_create(curproc);
14229 
14230 	vstate = &help->dthps_vstate;
14231 
14232 	if ((rv = dtrace_dof_slurp(dof, vstate, NULL, &enab,
14233 	    dhp != NULL ? dhp->dofhp_addr : 0, B_FALSE)) != 0) {
14234 		dtrace_dof_destroy(dof);
14235 		return (rv);
14236 	}
14237 
14238 	/*
14239 	 * Look for helper providers and validate their descriptions.
14240 	 */
14241 	if (dhp != NULL) {
14242 		for (i = 0; i < dof->dofh_secnum; i++) {
14243 			dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
14244 			    dof->dofh_secoff + i * dof->dofh_secsize);
14245 
14246 			if (sec->dofs_type != DOF_SECT_PROVIDER)
14247 				continue;
14248 
14249 			if (dtrace_helper_provider_validate(dof, sec) != 0) {
14250 				dtrace_enabling_destroy(enab);
14251 				dtrace_dof_destroy(dof);
14252 				return (-1);
14253 			}
14254 
14255 			nprovs++;
14256 		}
14257 	}
14258 
14259 	/*
14260 	 * Now we need to walk through the ECB descriptions in the enabling.
14261 	 */
14262 	for (i = 0; i < enab->dten_ndesc; i++) {
14263 		dtrace_ecbdesc_t *ep = enab->dten_desc[i];
14264 		dtrace_probedesc_t *desc = &ep->dted_probe;
14265 
14266 		if (strcmp(desc->dtpd_provider, "dtrace") != 0)
14267 			continue;
14268 
14269 		if (strcmp(desc->dtpd_mod, "helper") != 0)
14270 			continue;
14271 
14272 		if (strcmp(desc->dtpd_func, "ustack") != 0)
14273 			continue;
14274 
14275 		if ((rv = dtrace_helper_action_add(DTRACE_HELPER_ACTION_USTACK,
14276 		    ep)) != 0) {
14277 			/*
14278 			 * Adding this helper action failed -- we are now going
14279 			 * to rip out the entire generation and return failure.
14280 			 */
14281 			(void) dtrace_helper_destroygen(help->dthps_generation);
14282 			dtrace_enabling_destroy(enab);
14283 			dtrace_dof_destroy(dof);
14284 			return (-1);
14285 		}
14286 
14287 		nhelpers++;
14288 	}
14289 
14290 	if (nhelpers < enab->dten_ndesc)
14291 		dtrace_dof_error(dof, "unmatched helpers");
14292 
14293 	gen = help->dthps_generation++;
14294 	dtrace_enabling_destroy(enab);
14295 
14296 	if (dhp != NULL && nprovs > 0) {
14297 		dhp->dofhp_dof = (uint64_t)(uintptr_t)dof;
14298 		if (dtrace_helper_provider_add(dhp, gen) == 0) {
14299 			mutex_exit(&dtrace_lock);
14300 			dtrace_helper_provider_register(curproc, help, dhp);
14301 			mutex_enter(&dtrace_lock);
14302 
14303 			destroy = 0;
14304 		}
14305 	}
14306 
14307 	if (destroy)
14308 		dtrace_dof_destroy(dof);
14309 
14310 	return (gen);
14311 }
14312 
14313 static dtrace_helpers_t *
14314 dtrace_helpers_create(proc_t *p)
14315 {
14316 	dtrace_helpers_t *help;
14317 
14318 	ASSERT(MUTEX_HELD(&dtrace_lock));
14319 	ASSERT(p->p_dtrace_helpers == NULL);
14320 
14321 	help = kmem_zalloc(sizeof (dtrace_helpers_t), KM_SLEEP);
14322 	help->dthps_actions = kmem_zalloc(sizeof (dtrace_helper_action_t *) *
14323 	    DTRACE_NHELPER_ACTIONS, KM_SLEEP);
14324 
14325 	p->p_dtrace_helpers = help;
14326 	dtrace_helpers++;
14327 
14328 	return (help);
14329 }
14330 
14331 static void
14332 dtrace_helpers_destroy(void)
14333 {
14334 	dtrace_helpers_t *help;
14335 	dtrace_vstate_t *vstate;
14336 	proc_t *p = curproc;
14337 	int i;
14338 
14339 	mutex_enter(&dtrace_lock);
14340 
14341 	ASSERT(p->p_dtrace_helpers != NULL);
14342 	ASSERT(dtrace_helpers > 0);
14343 
14344 	help = p->p_dtrace_helpers;
14345 	vstate = &help->dthps_vstate;
14346 
14347 	/*
14348 	 * We're now going to lose the help from this process.
14349 	 */
14350 	p->p_dtrace_helpers = NULL;
14351 	dtrace_sync();
14352 
14353 	/*
14354 	 * Destory the helper actions.
14355 	 */
14356 	for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
14357 		dtrace_helper_action_t *h, *next;
14358 
14359 		for (h = help->dthps_actions[i]; h != NULL; h = next) {
14360 			next = h->dtha_next;
14361 			dtrace_helper_action_destroy(h, vstate);
14362 			h = next;
14363 		}
14364 	}
14365 
14366 	mutex_exit(&dtrace_lock);
14367 
14368 	/*
14369 	 * Destroy the helper providers.
14370 	 */
14371 	if (help->dthps_maxprovs > 0) {
14372 		mutex_enter(&dtrace_meta_lock);
14373 		if (dtrace_meta_pid != NULL) {
14374 			ASSERT(dtrace_deferred_pid == NULL);
14375 
14376 			for (i = 0; i < help->dthps_nprovs; i++) {
14377 				dtrace_helper_provider_remove(
14378 				    &help->dthps_provs[i]->dthp_prov, p->p_pid);
14379 			}
14380 		} else {
14381 			mutex_enter(&dtrace_lock);
14382 			ASSERT(help->dthps_deferred == 0 ||
14383 			    help->dthps_next != NULL ||
14384 			    help->dthps_prev != NULL ||
14385 			    help == dtrace_deferred_pid);
14386 
14387 			/*
14388 			 * Remove the helper from the deferred list.
14389 			 */
14390 			if (help->dthps_next != NULL)
14391 				help->dthps_next->dthps_prev = help->dthps_prev;
14392 			if (help->dthps_prev != NULL)
14393 				help->dthps_prev->dthps_next = help->dthps_next;
14394 			if (dtrace_deferred_pid == help) {
14395 				dtrace_deferred_pid = help->dthps_next;
14396 				ASSERT(help->dthps_prev == NULL);
14397 			}
14398 
14399 			mutex_exit(&dtrace_lock);
14400 		}
14401 
14402 		mutex_exit(&dtrace_meta_lock);
14403 
14404 		for (i = 0; i < help->dthps_nprovs; i++) {
14405 			dtrace_helper_provider_destroy(help->dthps_provs[i]);
14406 		}
14407 
14408 		kmem_free(help->dthps_provs, help->dthps_maxprovs *
14409 		    sizeof (dtrace_helper_provider_t *));
14410 	}
14411 
14412 	mutex_enter(&dtrace_lock);
14413 
14414 	dtrace_vstate_fini(&help->dthps_vstate);
14415 	kmem_free(help->dthps_actions,
14416 	    sizeof (dtrace_helper_action_t *) * DTRACE_NHELPER_ACTIONS);
14417 	kmem_free(help, sizeof (dtrace_helpers_t));
14418 
14419 	--dtrace_helpers;
14420 	mutex_exit(&dtrace_lock);
14421 }
14422 
14423 static void
14424 dtrace_helpers_duplicate(proc_t *from, proc_t *to)
14425 {
14426 	dtrace_helpers_t *help, *newhelp;
14427 	dtrace_helper_action_t *helper, *new, *last;
14428 	dtrace_difo_t *dp;
14429 	dtrace_vstate_t *vstate;
14430 	int i, j, sz, hasprovs = 0;
14431 
14432 	mutex_enter(&dtrace_lock);
14433 	ASSERT(from->p_dtrace_helpers != NULL);
14434 	ASSERT(dtrace_helpers > 0);
14435 
14436 	help = from->p_dtrace_helpers;
14437 	newhelp = dtrace_helpers_create(to);
14438 	ASSERT(to->p_dtrace_helpers != NULL);
14439 
14440 	newhelp->dthps_generation = help->dthps_generation;
14441 	vstate = &newhelp->dthps_vstate;
14442 
14443 	/*
14444 	 * Duplicate the helper actions.
14445 	 */
14446 	for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
14447 		if ((helper = help->dthps_actions[i]) == NULL)
14448 			continue;
14449 
14450 		for (last = NULL; helper != NULL; helper = helper->dtha_next) {
14451 			new = kmem_zalloc(sizeof (dtrace_helper_action_t),
14452 			    KM_SLEEP);
14453 			new->dtha_generation = helper->dtha_generation;
14454 
14455 			if ((dp = helper->dtha_predicate) != NULL) {
14456 				dp = dtrace_difo_duplicate(dp, vstate);
14457 				new->dtha_predicate = dp;
14458 			}
14459 
14460 			new->dtha_nactions = helper->dtha_nactions;
14461 			sz = sizeof (dtrace_difo_t *) * new->dtha_nactions;
14462 			new->dtha_actions = kmem_alloc(sz, KM_SLEEP);
14463 
14464 			for (j = 0; j < new->dtha_nactions; j++) {
14465 				dtrace_difo_t *dp = helper->dtha_actions[j];
14466 
14467 				ASSERT(dp != NULL);
14468 				dp = dtrace_difo_duplicate(dp, vstate);
14469 				new->dtha_actions[j] = dp;
14470 			}
14471 
14472 			if (last != NULL) {
14473 				last->dtha_next = new;
14474 			} else {
14475 				newhelp->dthps_actions[i] = new;
14476 			}
14477 
14478 			last = new;
14479 		}
14480 	}
14481 
14482 	/*
14483 	 * Duplicate the helper providers and register them with the
14484 	 * DTrace framework.
14485 	 */
14486 	if (help->dthps_nprovs > 0) {
14487 		newhelp->dthps_nprovs = help->dthps_nprovs;
14488 		newhelp->dthps_maxprovs = help->dthps_nprovs;
14489 		newhelp->dthps_provs = kmem_alloc(newhelp->dthps_nprovs *
14490 		    sizeof (dtrace_helper_provider_t *), KM_SLEEP);
14491 		for (i = 0; i < newhelp->dthps_nprovs; i++) {
14492 			newhelp->dthps_provs[i] = help->dthps_provs[i];
14493 			newhelp->dthps_provs[i]->dthp_ref++;
14494 		}
14495 
14496 		hasprovs = 1;
14497 	}
14498 
14499 	mutex_exit(&dtrace_lock);
14500 
14501 	if (hasprovs)
14502 		dtrace_helper_provider_register(to, newhelp, NULL);
14503 }
14504 
14505 /*
14506  * DTrace Hook Functions
14507  */
14508 static void
14509 dtrace_module_loaded(struct modctl *ctl)
14510 {
14511 	dtrace_provider_t *prv;
14512 
14513 	mutex_enter(&dtrace_provider_lock);
14514 	mutex_enter(&mod_lock);
14515 
14516 	ASSERT(ctl->mod_busy);
14517 
14518 	/*
14519 	 * We're going to call each providers per-module provide operation
14520 	 * specifying only this module.
14521 	 */
14522 	for (prv = dtrace_provider; prv != NULL; prv = prv->dtpv_next)
14523 		prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
14524 
14525 	mutex_exit(&mod_lock);
14526 	mutex_exit(&dtrace_provider_lock);
14527 
14528 	/*
14529 	 * If we have any retained enablings, we need to match against them.
14530 	 * Enabling probes requires that cpu_lock be held, and we cannot hold
14531 	 * cpu_lock here -- it is legal for cpu_lock to be held when loading a
14532 	 * module.  (In particular, this happens when loading scheduling
14533 	 * classes.)  So if we have any retained enablings, we need to dispatch
14534 	 * our task queue to do the match for us.
14535 	 */
14536 	mutex_enter(&dtrace_lock);
14537 
14538 	if (dtrace_retained == NULL) {
14539 		mutex_exit(&dtrace_lock);
14540 		return;
14541 	}
14542 
14543 	(void) taskq_dispatch(dtrace_taskq,
14544 	    (task_func_t *)dtrace_enabling_matchall, NULL, TQ_SLEEP);
14545 
14546 	mutex_exit(&dtrace_lock);
14547 
14548 	/*
14549 	 * And now, for a little heuristic sleaze:  in general, we want to
14550 	 * match modules as soon as they load.  However, we cannot guarantee
14551 	 * this, because it would lead us to the lock ordering violation
14552 	 * outlined above.  The common case, of course, is that cpu_lock is
14553 	 * _not_ held -- so we delay here for a clock tick, hoping that that's
14554 	 * long enough for the task queue to do its work.  If it's not, it's
14555 	 * not a serious problem -- it just means that the module that we
14556 	 * just loaded may not be immediately instrumentable.
14557 	 */
14558 	delay(1);
14559 }
14560 
14561 static void
14562 dtrace_module_unloaded(struct modctl *ctl)
14563 {
14564 	dtrace_probe_t template, *probe, *first, *next;
14565 	dtrace_provider_t *prov;
14566 
14567 	template.dtpr_mod = ctl->mod_modname;
14568 
14569 	mutex_enter(&dtrace_provider_lock);
14570 	mutex_enter(&mod_lock);
14571 	mutex_enter(&dtrace_lock);
14572 
14573 	if (dtrace_bymod == NULL) {
14574 		/*
14575 		 * The DTrace module is loaded (obviously) but not attached;
14576 		 * we don't have any work to do.
14577 		 */
14578 		mutex_exit(&dtrace_provider_lock);
14579 		mutex_exit(&mod_lock);
14580 		mutex_exit(&dtrace_lock);
14581 		return;
14582 	}
14583 
14584 	for (probe = first = dtrace_hash_lookup(dtrace_bymod, &template);
14585 	    probe != NULL; probe = probe->dtpr_nextmod) {
14586 		if (probe->dtpr_ecb != NULL) {
14587 			mutex_exit(&dtrace_provider_lock);
14588 			mutex_exit(&mod_lock);
14589 			mutex_exit(&dtrace_lock);
14590 
14591 			/*
14592 			 * This shouldn't _actually_ be possible -- we're
14593 			 * unloading a module that has an enabled probe in it.
14594 			 * (It's normally up to the provider to make sure that
14595 			 * this can't happen.)  However, because dtps_enable()
14596 			 * doesn't have a failure mode, there can be an
14597 			 * enable/unload race.  Upshot:  we don't want to
14598 			 * assert, but we're not going to disable the
14599 			 * probe, either.
14600 			 */
14601 			if (dtrace_err_verbose) {
14602 				cmn_err(CE_WARN, "unloaded module '%s' had "
14603 				    "enabled probes", ctl->mod_modname);
14604 			}
14605 
14606 			return;
14607 		}
14608 	}
14609 
14610 	probe = first;
14611 
14612 	for (first = NULL; probe != NULL; probe = next) {
14613 		ASSERT(dtrace_probes[probe->dtpr_id - 1] == probe);
14614 
14615 		dtrace_probes[probe->dtpr_id - 1] = NULL;
14616 
14617 		next = probe->dtpr_nextmod;
14618 		dtrace_hash_remove(dtrace_bymod, probe);
14619 		dtrace_hash_remove(dtrace_byfunc, probe);
14620 		dtrace_hash_remove(dtrace_byname, probe);
14621 
14622 		if (first == NULL) {
14623 			first = probe;
14624 			probe->dtpr_nextmod = NULL;
14625 		} else {
14626 			probe->dtpr_nextmod = first;
14627 			first = probe;
14628 		}
14629 	}
14630 
14631 	/*
14632 	 * We've removed all of the module's probes from the hash chains and
14633 	 * from the probe array.  Now issue a dtrace_sync() to be sure that
14634 	 * everyone has cleared out from any probe array processing.
14635 	 */
14636 	dtrace_sync();
14637 
14638 	for (probe = first; probe != NULL; probe = first) {
14639 		first = probe->dtpr_nextmod;
14640 		prov = probe->dtpr_provider;
14641 		prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, probe->dtpr_id,
14642 		    probe->dtpr_arg);
14643 		kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
14644 		kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
14645 		kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
14646 		vmem_free(dtrace_arena, (void *)(uintptr_t)probe->dtpr_id, 1);
14647 		kmem_free(probe, sizeof (dtrace_probe_t));
14648 	}
14649 
14650 	mutex_exit(&dtrace_lock);
14651 	mutex_exit(&mod_lock);
14652 	mutex_exit(&dtrace_provider_lock);
14653 }
14654 
14655 void
14656 dtrace_suspend(void)
14657 {
14658 	dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_suspend));
14659 }
14660 
14661 void
14662 dtrace_resume(void)
14663 {
14664 	dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_resume));
14665 }
14666 
14667 static int
14668 dtrace_cpu_setup(cpu_setup_t what, processorid_t cpu)
14669 {
14670 	ASSERT(MUTEX_HELD(&cpu_lock));
14671 	mutex_enter(&dtrace_lock);
14672 
14673 	switch (what) {
14674 	case CPU_CONFIG: {
14675 		dtrace_state_t *state;
14676 		dtrace_optval_t *opt, rs, c;
14677 
14678 		/*
14679 		 * For now, we only allocate a new buffer for anonymous state.
14680 		 */
14681 		if ((state = dtrace_anon.dta_state) == NULL)
14682 			break;
14683 
14684 		if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
14685 			break;
14686 
14687 		opt = state->dts_options;
14688 		c = opt[DTRACEOPT_CPU];
14689 
14690 		if (c != DTRACE_CPUALL && c != DTRACEOPT_UNSET && c != cpu)
14691 			break;
14692 
14693 		/*
14694 		 * Regardless of what the actual policy is, we're going to
14695 		 * temporarily set our resize policy to be manual.  We're
14696 		 * also going to temporarily set our CPU option to denote
14697 		 * the newly configured CPU.
14698 		 */
14699 		rs = opt[DTRACEOPT_BUFRESIZE];
14700 		opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_MANUAL;
14701 		opt[DTRACEOPT_CPU] = (dtrace_optval_t)cpu;
14702 
14703 		(void) dtrace_state_buffers(state);
14704 
14705 		opt[DTRACEOPT_BUFRESIZE] = rs;
14706 		opt[DTRACEOPT_CPU] = c;
14707 
14708 		break;
14709 	}
14710 
14711 	case CPU_UNCONFIG:
14712 		/*
14713 		 * We don't free the buffer in the CPU_UNCONFIG case.  (The
14714 		 * buffer will be freed when the consumer exits.)
14715 		 */
14716 		break;
14717 
14718 	default:
14719 		break;
14720 	}
14721 
14722 	mutex_exit(&dtrace_lock);
14723 	return (0);
14724 }
14725 
14726 static void
14727 dtrace_cpu_setup_initial(processorid_t cpu)
14728 {
14729 	(void) dtrace_cpu_setup(CPU_CONFIG, cpu);
14730 }
14731 
14732 static void
14733 dtrace_toxrange_add(uintptr_t base, uintptr_t limit)
14734 {
14735 	if (dtrace_toxranges >= dtrace_toxranges_max) {
14736 		int osize, nsize;
14737 		dtrace_toxrange_t *range;
14738 
14739 		osize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
14740 
14741 		if (osize == 0) {
14742 			ASSERT(dtrace_toxrange == NULL);
14743 			ASSERT(dtrace_toxranges_max == 0);
14744 			dtrace_toxranges_max = 1;
14745 		} else {
14746 			dtrace_toxranges_max <<= 1;
14747 		}
14748 
14749 		nsize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
14750 		range = kmem_zalloc(nsize, KM_SLEEP);
14751 
14752 		if (dtrace_toxrange != NULL) {
14753 			ASSERT(osize != 0);
14754 			bcopy(dtrace_toxrange, range, osize);
14755 			kmem_free(dtrace_toxrange, osize);
14756 		}
14757 
14758 		dtrace_toxrange = range;
14759 	}
14760 
14761 	ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_base == NULL);
14762 	ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_limit == NULL);
14763 
14764 	dtrace_toxrange[dtrace_toxranges].dtt_base = base;
14765 	dtrace_toxrange[dtrace_toxranges].dtt_limit = limit;
14766 	dtrace_toxranges++;
14767 }
14768 
14769 /*
14770  * DTrace Driver Cookbook Functions
14771  */
14772 /*ARGSUSED*/
14773 static int
14774 dtrace_attach(dev_info_t *devi, ddi_attach_cmd_t cmd)
14775 {
14776 	dtrace_provider_id_t id;
14777 	dtrace_state_t *state = NULL;
14778 	dtrace_enabling_t *enab;
14779 
14780 	mutex_enter(&cpu_lock);
14781 	mutex_enter(&dtrace_provider_lock);
14782 	mutex_enter(&dtrace_lock);
14783 
14784 	if (ddi_soft_state_init(&dtrace_softstate,
14785 	    sizeof (dtrace_state_t), 0) != 0) {
14786 		cmn_err(CE_NOTE, "/dev/dtrace failed to initialize soft state");
14787 		mutex_exit(&cpu_lock);
14788 		mutex_exit(&dtrace_provider_lock);
14789 		mutex_exit(&dtrace_lock);
14790 		return (DDI_FAILURE);
14791 	}
14792 
14793 	if (ddi_create_minor_node(devi, DTRACEMNR_DTRACE, S_IFCHR,
14794 	    DTRACEMNRN_DTRACE, DDI_PSEUDO, NULL) == DDI_FAILURE ||
14795 	    ddi_create_minor_node(devi, DTRACEMNR_HELPER, S_IFCHR,
14796 	    DTRACEMNRN_HELPER, DDI_PSEUDO, NULL) == DDI_FAILURE) {
14797 		cmn_err(CE_NOTE, "/dev/dtrace couldn't create minor nodes");
14798 		ddi_remove_minor_node(devi, NULL);
14799 		ddi_soft_state_fini(&dtrace_softstate);
14800 		mutex_exit(&cpu_lock);
14801 		mutex_exit(&dtrace_provider_lock);
14802 		mutex_exit(&dtrace_lock);
14803 		return (DDI_FAILURE);
14804 	}
14805 
14806 	ddi_report_dev(devi);
14807 	dtrace_devi = devi;
14808 
14809 	dtrace_modload = dtrace_module_loaded;
14810 	dtrace_modunload = dtrace_module_unloaded;
14811 	dtrace_cpu_init = dtrace_cpu_setup_initial;
14812 	dtrace_helpers_cleanup = dtrace_helpers_destroy;
14813 	dtrace_helpers_fork = dtrace_helpers_duplicate;
14814 	dtrace_cpustart_init = dtrace_suspend;
14815 	dtrace_cpustart_fini = dtrace_resume;
14816 	dtrace_debugger_init = dtrace_suspend;
14817 	dtrace_debugger_fini = dtrace_resume;
14818 
14819 	register_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
14820 
14821 	ASSERT(MUTEX_HELD(&cpu_lock));
14822 
14823 	dtrace_arena = vmem_create("dtrace", (void *)1, UINT32_MAX, 1,
14824 	    NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
14825 	dtrace_minor = vmem_create("dtrace_minor", (void *)DTRACEMNRN_CLONE,
14826 	    UINT32_MAX - DTRACEMNRN_CLONE, 1, NULL, NULL, NULL, 0,
14827 	    VM_SLEEP | VMC_IDENTIFIER);
14828 	dtrace_taskq = taskq_create("dtrace_taskq", 1, maxclsyspri,
14829 	    1, INT_MAX, 0);
14830 
14831 	dtrace_state_cache = kmem_cache_create("dtrace_state_cache",
14832 	    sizeof (dtrace_dstate_percpu_t) * NCPU, DTRACE_STATE_ALIGN,
14833 	    NULL, NULL, NULL, NULL, NULL, 0);
14834 
14835 	ASSERT(MUTEX_HELD(&cpu_lock));
14836 	dtrace_bymod = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_mod),
14837 	    offsetof(dtrace_probe_t, dtpr_nextmod),
14838 	    offsetof(dtrace_probe_t, dtpr_prevmod));
14839 
14840 	dtrace_byfunc = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_func),
14841 	    offsetof(dtrace_probe_t, dtpr_nextfunc),
14842 	    offsetof(dtrace_probe_t, dtpr_prevfunc));
14843 
14844 	dtrace_byname = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_name),
14845 	    offsetof(dtrace_probe_t, dtpr_nextname),
14846 	    offsetof(dtrace_probe_t, dtpr_prevname));
14847 
14848 	if (dtrace_retain_max < 1) {
14849 		cmn_err(CE_WARN, "illegal value (%lu) for dtrace_retain_max; "
14850 		    "setting to 1", dtrace_retain_max);
14851 		dtrace_retain_max = 1;
14852 	}
14853 
14854 	/*
14855 	 * Now discover our toxic ranges.
14856 	 */
14857 	dtrace_toxic_ranges(dtrace_toxrange_add);
14858 
14859 	/*
14860 	 * Before we register ourselves as a provider to our own framework,
14861 	 * we would like to assert that dtrace_provider is NULL -- but that's
14862 	 * not true if we were loaded as a dependency of a DTrace provider.
14863 	 * Once we've registered, we can assert that dtrace_provider is our
14864 	 * pseudo provider.
14865 	 */
14866 	(void) dtrace_register("dtrace", &dtrace_provider_attr,
14867 	    DTRACE_PRIV_NONE, 0, &dtrace_provider_ops, NULL, &id);
14868 
14869 	ASSERT(dtrace_provider != NULL);
14870 	ASSERT((dtrace_provider_id_t)dtrace_provider == id);
14871 
14872 	dtrace_probeid_begin = dtrace_probe_create((dtrace_provider_id_t)
14873 	    dtrace_provider, NULL, NULL, "BEGIN", 0, NULL);
14874 	dtrace_probeid_end = dtrace_probe_create((dtrace_provider_id_t)
14875 	    dtrace_provider, NULL, NULL, "END", 0, NULL);
14876 	dtrace_probeid_error = dtrace_probe_create((dtrace_provider_id_t)
14877 	    dtrace_provider, NULL, NULL, "ERROR", 1, NULL);
14878 
14879 	dtrace_anon_property();
14880 	mutex_exit(&cpu_lock);
14881 
14882 	/*
14883 	 * If DTrace helper tracing is enabled, we need to allocate the
14884 	 * trace buffer and initialize the values.
14885 	 */
14886 	if (dtrace_helptrace_enabled) {
14887 		ASSERT(dtrace_helptrace_buffer == NULL);
14888 		dtrace_helptrace_buffer =
14889 		    kmem_zalloc(dtrace_helptrace_bufsize, KM_SLEEP);
14890 		dtrace_helptrace_next = 0;
14891 	}
14892 
14893 	/*
14894 	 * If there are already providers, we must ask them to provide their
14895 	 * probes, and then match any anonymous enabling against them.  Note
14896 	 * that there should be no other retained enablings at this time:
14897 	 * the only retained enablings at this time should be the anonymous
14898 	 * enabling.
14899 	 */
14900 	if (dtrace_anon.dta_enabling != NULL) {
14901 		ASSERT(dtrace_retained == dtrace_anon.dta_enabling);
14902 
14903 		dtrace_enabling_provide(NULL);
14904 		state = dtrace_anon.dta_state;
14905 
14906 		/*
14907 		 * We couldn't hold cpu_lock across the above call to
14908 		 * dtrace_enabling_provide(), but we must hold it to actually
14909 		 * enable the probes.  We have to drop all of our locks, pick
14910 		 * up cpu_lock, and regain our locks before matching the
14911 		 * retained anonymous enabling.
14912 		 */
14913 		mutex_exit(&dtrace_lock);
14914 		mutex_exit(&dtrace_provider_lock);
14915 
14916 		mutex_enter(&cpu_lock);
14917 		mutex_enter(&dtrace_provider_lock);
14918 		mutex_enter(&dtrace_lock);
14919 
14920 		if ((enab = dtrace_anon.dta_enabling) != NULL)
14921 			(void) dtrace_enabling_match(enab, NULL);
14922 
14923 		mutex_exit(&cpu_lock);
14924 	}
14925 
14926 	mutex_exit(&dtrace_lock);
14927 	mutex_exit(&dtrace_provider_lock);
14928 
14929 	if (state != NULL) {
14930 		/*
14931 		 * If we created any anonymous state, set it going now.
14932 		 */
14933 		(void) dtrace_state_go(state, &dtrace_anon.dta_beganon);
14934 	}
14935 
14936 	return (DDI_SUCCESS);
14937 }
14938 
14939 /*ARGSUSED*/
14940 static int
14941 dtrace_open(dev_t *devp, int flag, int otyp, cred_t *cred_p)
14942 {
14943 	dtrace_state_t *state;
14944 	uint32_t priv;
14945 	uid_t uid;
14946 	zoneid_t zoneid;
14947 
14948 	if (getminor(*devp) == DTRACEMNRN_HELPER)
14949 		return (0);
14950 
14951 	/*
14952 	 * If this wasn't an open with the "helper" minor, then it must be
14953 	 * the "dtrace" minor.
14954 	 */
14955 	if (getminor(*devp) != DTRACEMNRN_DTRACE)
14956 		return (ENXIO);
14957 
14958 	/*
14959 	 * If no DTRACE_PRIV_* bits are set in the credential, then the
14960 	 * caller lacks sufficient permission to do anything with DTrace.
14961 	 */
14962 	dtrace_cred2priv(cred_p, &priv, &uid, &zoneid);
14963 	if (priv == DTRACE_PRIV_NONE)
14964 		return (EACCES);
14965 
14966 	/*
14967 	 * Ask all providers to provide all their probes.
14968 	 */
14969 	mutex_enter(&dtrace_provider_lock);
14970 	dtrace_probe_provide(NULL, NULL);
14971 	mutex_exit(&dtrace_provider_lock);
14972 
14973 	mutex_enter(&cpu_lock);
14974 	mutex_enter(&dtrace_lock);
14975 	dtrace_opens++;
14976 	dtrace_membar_producer();
14977 
14978 	/*
14979 	 * If the kernel debugger is active (that is, if the kernel debugger
14980 	 * modified text in some way), we won't allow the open.
14981 	 */
14982 	if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
14983 		dtrace_opens--;
14984 		mutex_exit(&cpu_lock);
14985 		mutex_exit(&dtrace_lock);
14986 		return (EBUSY);
14987 	}
14988 
14989 	state = dtrace_state_create(devp, cred_p);
14990 	mutex_exit(&cpu_lock);
14991 
14992 	if (state == NULL) {
14993 		if (--dtrace_opens == 0 && dtrace_anon.dta_enabling == NULL)
14994 			(void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
14995 		mutex_exit(&dtrace_lock);
14996 		return (EAGAIN);
14997 	}
14998 
14999 	mutex_exit(&dtrace_lock);
15000 
15001 	return (0);
15002 }
15003 
15004 /*ARGSUSED*/
15005 static int
15006 dtrace_close(dev_t dev, int flag, int otyp, cred_t *cred_p)
15007 {
15008 	minor_t minor = getminor(dev);
15009 	dtrace_state_t *state;
15010 
15011 	if (minor == DTRACEMNRN_HELPER)
15012 		return (0);
15013 
15014 	state = ddi_get_soft_state(dtrace_softstate, minor);
15015 
15016 	mutex_enter(&cpu_lock);
15017 	mutex_enter(&dtrace_lock);
15018 
15019 	if (state->dts_anon) {
15020 		/*
15021 		 * There is anonymous state. Destroy that first.
15022 		 */
15023 		ASSERT(dtrace_anon.dta_state == NULL);
15024 		dtrace_state_destroy(state->dts_anon);
15025 	}
15026 
15027 	dtrace_state_destroy(state);
15028 	ASSERT(dtrace_opens > 0);
15029 
15030 	/*
15031 	 * Only relinquish control of the kernel debugger interface when there
15032 	 * are no consumers and no anonymous enablings.
15033 	 */
15034 	if (--dtrace_opens == 0 && dtrace_anon.dta_enabling == NULL)
15035 		(void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
15036 
15037 	mutex_exit(&dtrace_lock);
15038 	mutex_exit(&cpu_lock);
15039 
15040 	return (0);
15041 }
15042 
15043 /*ARGSUSED*/
15044 static int
15045 dtrace_ioctl_helper(int cmd, intptr_t arg, int *rv)
15046 {
15047 	int rval;
15048 	dof_helper_t help, *dhp = NULL;
15049 
15050 	switch (cmd) {
15051 	case DTRACEHIOC_ADDDOF:
15052 		if (copyin((void *)arg, &help, sizeof (help)) != 0) {
15053 			dtrace_dof_error(NULL, "failed to copyin DOF helper");
15054 			return (EFAULT);
15055 		}
15056 
15057 		dhp = &help;
15058 		arg = (intptr_t)help.dofhp_dof;
15059 		/*FALLTHROUGH*/
15060 
15061 	case DTRACEHIOC_ADD: {
15062 		dof_hdr_t *dof = dtrace_dof_copyin(arg, &rval);
15063 
15064 		if (dof == NULL)
15065 			return (rval);
15066 
15067 		mutex_enter(&dtrace_lock);
15068 
15069 		/*
15070 		 * dtrace_helper_slurp() takes responsibility for the dof --
15071 		 * it may free it now or it may save it and free it later.
15072 		 */
15073 		if ((rval = dtrace_helper_slurp(dof, dhp)) != -1) {
15074 			*rv = rval;
15075 			rval = 0;
15076 		} else {
15077 			rval = EINVAL;
15078 		}
15079 
15080 		mutex_exit(&dtrace_lock);
15081 		return (rval);
15082 	}
15083 
15084 	case DTRACEHIOC_REMOVE: {
15085 		mutex_enter(&dtrace_lock);
15086 		rval = dtrace_helper_destroygen(arg);
15087 		mutex_exit(&dtrace_lock);
15088 
15089 		return (rval);
15090 	}
15091 
15092 	default:
15093 		break;
15094 	}
15095 
15096 	return (ENOTTY);
15097 }
15098 
15099 /*ARGSUSED*/
15100 static int
15101 dtrace_ioctl(dev_t dev, int cmd, intptr_t arg, int md, cred_t *cr, int *rv)
15102 {
15103 	minor_t minor = getminor(dev);
15104 	dtrace_state_t *state;
15105 	int rval;
15106 
15107 	if (minor == DTRACEMNRN_HELPER)
15108 		return (dtrace_ioctl_helper(cmd, arg, rv));
15109 
15110 	state = ddi_get_soft_state(dtrace_softstate, minor);
15111 
15112 	if (state->dts_anon) {
15113 		ASSERT(dtrace_anon.dta_state == NULL);
15114 		state = state->dts_anon;
15115 	}
15116 
15117 	switch (cmd) {
15118 	case DTRACEIOC_PROVIDER: {
15119 		dtrace_providerdesc_t pvd;
15120 		dtrace_provider_t *pvp;
15121 
15122 		if (copyin((void *)arg, &pvd, sizeof (pvd)) != 0)
15123 			return (EFAULT);
15124 
15125 		pvd.dtvd_name[DTRACE_PROVNAMELEN - 1] = '\0';
15126 		mutex_enter(&dtrace_provider_lock);
15127 
15128 		for (pvp = dtrace_provider; pvp != NULL; pvp = pvp->dtpv_next) {
15129 			if (strcmp(pvp->dtpv_name, pvd.dtvd_name) == 0)
15130 				break;
15131 		}
15132 
15133 		mutex_exit(&dtrace_provider_lock);
15134 
15135 		if (pvp == NULL)
15136 			return (ESRCH);
15137 
15138 		bcopy(&pvp->dtpv_priv, &pvd.dtvd_priv, sizeof (dtrace_ppriv_t));
15139 		bcopy(&pvp->dtpv_attr, &pvd.dtvd_attr, sizeof (dtrace_pattr_t));
15140 		if (copyout(&pvd, (void *)arg, sizeof (pvd)) != 0)
15141 			return (EFAULT);
15142 
15143 		return (0);
15144 	}
15145 
15146 	case DTRACEIOC_EPROBE: {
15147 		dtrace_eprobedesc_t epdesc;
15148 		dtrace_ecb_t *ecb;
15149 		dtrace_action_t *act;
15150 		void *buf;
15151 		size_t size;
15152 		uintptr_t dest;
15153 		int nrecs;
15154 
15155 		if (copyin((void *)arg, &epdesc, sizeof (epdesc)) != 0)
15156 			return (EFAULT);
15157 
15158 		mutex_enter(&dtrace_lock);
15159 
15160 		if ((ecb = dtrace_epid2ecb(state, epdesc.dtepd_epid)) == NULL) {
15161 			mutex_exit(&dtrace_lock);
15162 			return (EINVAL);
15163 		}
15164 
15165 		if (ecb->dte_probe == NULL) {
15166 			mutex_exit(&dtrace_lock);
15167 			return (EINVAL);
15168 		}
15169 
15170 		epdesc.dtepd_probeid = ecb->dte_probe->dtpr_id;
15171 		epdesc.dtepd_uarg = ecb->dte_uarg;
15172 		epdesc.dtepd_size = ecb->dte_size;
15173 
15174 		nrecs = epdesc.dtepd_nrecs;
15175 		epdesc.dtepd_nrecs = 0;
15176 		for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
15177 			if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
15178 				continue;
15179 
15180 			epdesc.dtepd_nrecs++;
15181 		}
15182 
15183 		/*
15184 		 * Now that we have the size, we need to allocate a temporary
15185 		 * buffer in which to store the complete description.  We need
15186 		 * the temporary buffer to be able to drop dtrace_lock()
15187 		 * across the copyout(), below.
15188 		 */
15189 		size = sizeof (dtrace_eprobedesc_t) +
15190 		    (epdesc.dtepd_nrecs * sizeof (dtrace_recdesc_t));
15191 
15192 		buf = kmem_alloc(size, KM_SLEEP);
15193 		dest = (uintptr_t)buf;
15194 
15195 		bcopy(&epdesc, (void *)dest, sizeof (epdesc));
15196 		dest += offsetof(dtrace_eprobedesc_t, dtepd_rec[0]);
15197 
15198 		for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
15199 			if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
15200 				continue;
15201 
15202 			if (nrecs-- == 0)
15203 				break;
15204 
15205 			bcopy(&act->dta_rec, (void *)dest,
15206 			    sizeof (dtrace_recdesc_t));
15207 			dest += sizeof (dtrace_recdesc_t);
15208 		}
15209 
15210 		mutex_exit(&dtrace_lock);
15211 
15212 		if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
15213 			kmem_free(buf, size);
15214 			return (EFAULT);
15215 		}
15216 
15217 		kmem_free(buf, size);
15218 		return (0);
15219 	}
15220 
15221 	case DTRACEIOC_AGGDESC: {
15222 		dtrace_aggdesc_t aggdesc;
15223 		dtrace_action_t *act;
15224 		dtrace_aggregation_t *agg;
15225 		int nrecs;
15226 		uint32_t offs;
15227 		dtrace_recdesc_t *lrec;
15228 		void *buf;
15229 		size_t size;
15230 		uintptr_t dest;
15231 
15232 		if (copyin((void *)arg, &aggdesc, sizeof (aggdesc)) != 0)
15233 			return (EFAULT);
15234 
15235 		mutex_enter(&dtrace_lock);
15236 
15237 		if ((agg = dtrace_aggid2agg(state, aggdesc.dtagd_id)) == NULL) {
15238 			mutex_exit(&dtrace_lock);
15239 			return (EINVAL);
15240 		}
15241 
15242 		aggdesc.dtagd_epid = agg->dtag_ecb->dte_epid;
15243 
15244 		nrecs = aggdesc.dtagd_nrecs;
15245 		aggdesc.dtagd_nrecs = 0;
15246 
15247 		offs = agg->dtag_base;
15248 		lrec = &agg->dtag_action.dta_rec;
15249 		aggdesc.dtagd_size = lrec->dtrd_offset + lrec->dtrd_size - offs;
15250 
15251 		for (act = agg->dtag_first; ; act = act->dta_next) {
15252 			ASSERT(act->dta_intuple ||
15253 			    DTRACEACT_ISAGG(act->dta_kind));
15254 
15255 			/*
15256 			 * If this action has a record size of zero, it
15257 			 * denotes an argument to the aggregating action.
15258 			 * Because the presence of this record doesn't (or
15259 			 * shouldn't) affect the way the data is interpreted,
15260 			 * we don't copy it out to save user-level the
15261 			 * confusion of dealing with a zero-length record.
15262 			 */
15263 			if (act->dta_rec.dtrd_size == 0) {
15264 				ASSERT(agg->dtag_hasarg);
15265 				continue;
15266 			}
15267 
15268 			aggdesc.dtagd_nrecs++;
15269 
15270 			if (act == &agg->dtag_action)
15271 				break;
15272 		}
15273 
15274 		/*
15275 		 * Now that we have the size, we need to allocate a temporary
15276 		 * buffer in which to store the complete description.  We need
15277 		 * the temporary buffer to be able to drop dtrace_lock()
15278 		 * across the copyout(), below.
15279 		 */
15280 		size = sizeof (dtrace_aggdesc_t) +
15281 		    (aggdesc.dtagd_nrecs * sizeof (dtrace_recdesc_t));
15282 
15283 		buf = kmem_alloc(size, KM_SLEEP);
15284 		dest = (uintptr_t)buf;
15285 
15286 		bcopy(&aggdesc, (void *)dest, sizeof (aggdesc));
15287 		dest += offsetof(dtrace_aggdesc_t, dtagd_rec[0]);
15288 
15289 		for (act = agg->dtag_first; ; act = act->dta_next) {
15290 			dtrace_recdesc_t rec = act->dta_rec;
15291 
15292 			/*
15293 			 * See the comment in the above loop for why we pass
15294 			 * over zero-length records.
15295 			 */
15296 			if (rec.dtrd_size == 0) {
15297 				ASSERT(agg->dtag_hasarg);
15298 				continue;
15299 			}
15300 
15301 			if (nrecs-- == 0)
15302 				break;
15303 
15304 			rec.dtrd_offset -= offs;
15305 			bcopy(&rec, (void *)dest, sizeof (rec));
15306 			dest += sizeof (dtrace_recdesc_t);
15307 
15308 			if (act == &agg->dtag_action)
15309 				break;
15310 		}
15311 
15312 		mutex_exit(&dtrace_lock);
15313 
15314 		if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
15315 			kmem_free(buf, size);
15316 			return (EFAULT);
15317 		}
15318 
15319 		kmem_free(buf, size);
15320 		return (0);
15321 	}
15322 
15323 	case DTRACEIOC_ENABLE: {
15324 		dof_hdr_t *dof;
15325 		dtrace_enabling_t *enab = NULL;
15326 		dtrace_vstate_t *vstate;
15327 		int err = 0;
15328 
15329 		*rv = 0;
15330 
15331 		/*
15332 		 * If a NULL argument has been passed, we take this as our
15333 		 * cue to reevaluate our enablings.
15334 		 */
15335 		if (arg == NULL) {
15336 			dtrace_enabling_matchall();
15337 
15338 			return (0);
15339 		}
15340 
15341 		if ((dof = dtrace_dof_copyin(arg, &rval)) == NULL)
15342 			return (rval);
15343 
15344 		mutex_enter(&cpu_lock);
15345 		mutex_enter(&dtrace_lock);
15346 		vstate = &state->dts_vstate;
15347 
15348 		if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
15349 			mutex_exit(&dtrace_lock);
15350 			mutex_exit(&cpu_lock);
15351 			dtrace_dof_destroy(dof);
15352 			return (EBUSY);
15353 		}
15354 
15355 		if (dtrace_dof_slurp(dof, vstate, cr, &enab, 0, B_TRUE) != 0) {
15356 			mutex_exit(&dtrace_lock);
15357 			mutex_exit(&cpu_lock);
15358 			dtrace_dof_destroy(dof);
15359 			return (EINVAL);
15360 		}
15361 
15362 		if ((rval = dtrace_dof_options(dof, state)) != 0) {
15363 			dtrace_enabling_destroy(enab);
15364 			mutex_exit(&dtrace_lock);
15365 			mutex_exit(&cpu_lock);
15366 			dtrace_dof_destroy(dof);
15367 			return (rval);
15368 		}
15369 
15370 		if ((err = dtrace_enabling_match(enab, rv)) == 0) {
15371 			err = dtrace_enabling_retain(enab);
15372 		} else {
15373 			dtrace_enabling_destroy(enab);
15374 		}
15375 
15376 		mutex_exit(&cpu_lock);
15377 		mutex_exit(&dtrace_lock);
15378 		dtrace_dof_destroy(dof);
15379 
15380 		return (err);
15381 	}
15382 
15383 	case DTRACEIOC_REPLICATE: {
15384 		dtrace_repldesc_t desc;
15385 		dtrace_probedesc_t *match = &desc.dtrpd_match;
15386 		dtrace_probedesc_t *create = &desc.dtrpd_create;
15387 		int err;
15388 
15389 		if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
15390 			return (EFAULT);
15391 
15392 		match->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
15393 		match->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
15394 		match->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
15395 		match->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
15396 
15397 		create->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
15398 		create->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
15399 		create->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
15400 		create->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
15401 
15402 		mutex_enter(&dtrace_lock);
15403 		err = dtrace_enabling_replicate(state, match, create);
15404 		mutex_exit(&dtrace_lock);
15405 
15406 		return (err);
15407 	}
15408 
15409 	case DTRACEIOC_PROBEMATCH:
15410 	case DTRACEIOC_PROBES: {
15411 		dtrace_probe_t *probe = NULL;
15412 		dtrace_probedesc_t desc;
15413 		dtrace_probekey_t pkey;
15414 		dtrace_id_t i;
15415 		int m = 0;
15416 		uint32_t priv;
15417 		uid_t uid;
15418 		zoneid_t zoneid;
15419 
15420 		if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
15421 			return (EFAULT);
15422 
15423 		desc.dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
15424 		desc.dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
15425 		desc.dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
15426 		desc.dtpd_name[DTRACE_NAMELEN - 1] = '\0';
15427 
15428 		/*
15429 		 * Before we attempt to match this probe, we want to give
15430 		 * all providers the opportunity to provide it.
15431 		 */
15432 		if (desc.dtpd_id == DTRACE_IDNONE) {
15433 			mutex_enter(&dtrace_provider_lock);
15434 			dtrace_probe_provide(&desc, NULL);
15435 			mutex_exit(&dtrace_provider_lock);
15436 			desc.dtpd_id++;
15437 		}
15438 
15439 		if (cmd == DTRACEIOC_PROBEMATCH)  {
15440 			dtrace_probekey(&desc, &pkey);
15441 			pkey.dtpk_id = DTRACE_IDNONE;
15442 		}
15443 
15444 		dtrace_cred2priv(cr, &priv, &uid, &zoneid);
15445 
15446 		mutex_enter(&dtrace_lock);
15447 
15448 		if (cmd == DTRACEIOC_PROBEMATCH) {
15449 			for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
15450 				if ((probe = dtrace_probes[i - 1]) != NULL &&
15451 				    (m = dtrace_match_probe(probe, &pkey,
15452 				    priv, uid, zoneid)) != 0)
15453 					break;
15454 			}
15455 
15456 			if (m < 0) {
15457 				mutex_exit(&dtrace_lock);
15458 				return (EINVAL);
15459 			}
15460 
15461 		} else {
15462 			for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
15463 				if ((probe = dtrace_probes[i - 1]) != NULL &&
15464 				    dtrace_match_priv(probe, priv, uid, zoneid))
15465 					break;
15466 			}
15467 		}
15468 
15469 		if (probe == NULL) {
15470 			mutex_exit(&dtrace_lock);
15471 			return (ESRCH);
15472 		}
15473 
15474 		dtrace_probe_description(probe, &desc);
15475 		mutex_exit(&dtrace_lock);
15476 
15477 		if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
15478 			return (EFAULT);
15479 
15480 		return (0);
15481 	}
15482 
15483 	case DTRACEIOC_PROBEARG: {
15484 		dtrace_argdesc_t desc;
15485 		dtrace_probe_t *probe;
15486 		dtrace_provider_t *prov;
15487 
15488 		if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
15489 			return (EFAULT);
15490 
15491 		if (desc.dtargd_id == DTRACE_IDNONE)
15492 			return (EINVAL);
15493 
15494 		if (desc.dtargd_ndx == DTRACE_ARGNONE)
15495 			return (EINVAL);
15496 
15497 		mutex_enter(&dtrace_provider_lock);
15498 		mutex_enter(&mod_lock);
15499 		mutex_enter(&dtrace_lock);
15500 
15501 		if (desc.dtargd_id > dtrace_nprobes) {
15502 			mutex_exit(&dtrace_lock);
15503 			mutex_exit(&mod_lock);
15504 			mutex_exit(&dtrace_provider_lock);
15505 			return (EINVAL);
15506 		}
15507 
15508 		if ((probe = dtrace_probes[desc.dtargd_id - 1]) == NULL) {
15509 			mutex_exit(&dtrace_lock);
15510 			mutex_exit(&mod_lock);
15511 			mutex_exit(&dtrace_provider_lock);
15512 			return (EINVAL);
15513 		}
15514 
15515 		mutex_exit(&dtrace_lock);
15516 
15517 		prov = probe->dtpr_provider;
15518 
15519 		if (prov->dtpv_pops.dtps_getargdesc == NULL) {
15520 			/*
15521 			 * There isn't any typed information for this probe.
15522 			 * Set the argument number to DTRACE_ARGNONE.
15523 			 */
15524 			desc.dtargd_ndx = DTRACE_ARGNONE;
15525 		} else {
15526 			desc.dtargd_native[0] = '\0';
15527 			desc.dtargd_xlate[0] = '\0';
15528 			desc.dtargd_mapping = desc.dtargd_ndx;
15529 
15530 			prov->dtpv_pops.dtps_getargdesc(prov->dtpv_arg,
15531 			    probe->dtpr_id, probe->dtpr_arg, &desc);
15532 		}
15533 
15534 		mutex_exit(&mod_lock);
15535 		mutex_exit(&dtrace_provider_lock);
15536 
15537 		if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
15538 			return (EFAULT);
15539 
15540 		return (0);
15541 	}
15542 
15543 	case DTRACEIOC_GO: {
15544 		processorid_t cpuid;
15545 		rval = dtrace_state_go(state, &cpuid);
15546 
15547 		if (rval != 0)
15548 			return (rval);
15549 
15550 		if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
15551 			return (EFAULT);
15552 
15553 		return (0);
15554 	}
15555 
15556 	case DTRACEIOC_STOP: {
15557 		processorid_t cpuid;
15558 
15559 		mutex_enter(&dtrace_lock);
15560 		rval = dtrace_state_stop(state, &cpuid);
15561 		mutex_exit(&dtrace_lock);
15562 
15563 		if (rval != 0)
15564 			return (rval);
15565 
15566 		if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
15567 			return (EFAULT);
15568 
15569 		return (0);
15570 	}
15571 
15572 	case DTRACEIOC_DOFGET: {
15573 		dof_hdr_t hdr, *dof;
15574 		uint64_t len;
15575 
15576 		if (copyin((void *)arg, &hdr, sizeof (hdr)) != 0)
15577 			return (EFAULT);
15578 
15579 		mutex_enter(&dtrace_lock);
15580 		dof = dtrace_dof_create(state);
15581 		mutex_exit(&dtrace_lock);
15582 
15583 		len = MIN(hdr.dofh_loadsz, dof->dofh_loadsz);
15584 		rval = copyout(dof, (void *)arg, len);
15585 		dtrace_dof_destroy(dof);
15586 
15587 		return (rval == 0 ? 0 : EFAULT);
15588 	}
15589 
15590 	case DTRACEIOC_AGGSNAP:
15591 	case DTRACEIOC_BUFSNAP: {
15592 		dtrace_bufdesc_t desc;
15593 		caddr_t cached;
15594 		dtrace_buffer_t *buf;
15595 
15596 		if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
15597 			return (EFAULT);
15598 
15599 		if (desc.dtbd_cpu < 0 || desc.dtbd_cpu >= NCPU)
15600 			return (EINVAL);
15601 
15602 		mutex_enter(&dtrace_lock);
15603 
15604 		if (cmd == DTRACEIOC_BUFSNAP) {
15605 			buf = &state->dts_buffer[desc.dtbd_cpu];
15606 		} else {
15607 			buf = &state->dts_aggbuffer[desc.dtbd_cpu];
15608 		}
15609 
15610 		if (buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL)) {
15611 			size_t sz = buf->dtb_offset;
15612 
15613 			if (state->dts_activity != DTRACE_ACTIVITY_STOPPED) {
15614 				mutex_exit(&dtrace_lock);
15615 				return (EBUSY);
15616 			}
15617 
15618 			/*
15619 			 * If this buffer has already been consumed, we're
15620 			 * going to indicate that there's nothing left here
15621 			 * to consume.
15622 			 */
15623 			if (buf->dtb_flags & DTRACEBUF_CONSUMED) {
15624 				mutex_exit(&dtrace_lock);
15625 
15626 				desc.dtbd_size = 0;
15627 				desc.dtbd_drops = 0;
15628 				desc.dtbd_errors = 0;
15629 				desc.dtbd_oldest = 0;
15630 				sz = sizeof (desc);
15631 
15632 				if (copyout(&desc, (void *)arg, sz) != 0)
15633 					return (EFAULT);
15634 
15635 				return (0);
15636 			}
15637 
15638 			/*
15639 			 * If this is a ring buffer that has wrapped, we want
15640 			 * to copy the whole thing out.
15641 			 */
15642 			if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
15643 				dtrace_buffer_polish(buf);
15644 				sz = buf->dtb_size;
15645 			}
15646 
15647 			if (copyout(buf->dtb_tomax, desc.dtbd_data, sz) != 0) {
15648 				mutex_exit(&dtrace_lock);
15649 				return (EFAULT);
15650 			}
15651 
15652 			desc.dtbd_size = sz;
15653 			desc.dtbd_drops = buf->dtb_drops;
15654 			desc.dtbd_errors = buf->dtb_errors;
15655 			desc.dtbd_oldest = buf->dtb_xamot_offset;
15656 			desc.dtbd_timestamp = dtrace_gethrtime();
15657 
15658 			mutex_exit(&dtrace_lock);
15659 
15660 			if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
15661 				return (EFAULT);
15662 
15663 			buf->dtb_flags |= DTRACEBUF_CONSUMED;
15664 
15665 			return (0);
15666 		}
15667 
15668 		if (buf->dtb_tomax == NULL) {
15669 			ASSERT(buf->dtb_xamot == NULL);
15670 			mutex_exit(&dtrace_lock);
15671 			return (ENOENT);
15672 		}
15673 
15674 		cached = buf->dtb_tomax;
15675 		ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
15676 
15677 		dtrace_xcall(desc.dtbd_cpu,
15678 		    (dtrace_xcall_t)dtrace_buffer_switch, buf);
15679 
15680 		state->dts_errors += buf->dtb_xamot_errors;
15681 
15682 		/*
15683 		 * If the buffers did not actually switch, then the cross call
15684 		 * did not take place -- presumably because the given CPU is
15685 		 * not in the ready set.  If this is the case, we'll return
15686 		 * ENOENT.
15687 		 */
15688 		if (buf->dtb_tomax == cached) {
15689 			ASSERT(buf->dtb_xamot != cached);
15690 			mutex_exit(&dtrace_lock);
15691 			return (ENOENT);
15692 		}
15693 
15694 		ASSERT(cached == buf->dtb_xamot);
15695 
15696 		/*
15697 		 * We have our snapshot; now copy it out.
15698 		 */
15699 		if (copyout(buf->dtb_xamot, desc.dtbd_data,
15700 		    buf->dtb_xamot_offset) != 0) {
15701 			mutex_exit(&dtrace_lock);
15702 			return (EFAULT);
15703 		}
15704 
15705 		desc.dtbd_size = buf->dtb_xamot_offset;
15706 		desc.dtbd_drops = buf->dtb_xamot_drops;
15707 		desc.dtbd_errors = buf->dtb_xamot_errors;
15708 		desc.dtbd_oldest = 0;
15709 		desc.dtbd_timestamp = buf->dtb_switched;
15710 
15711 		mutex_exit(&dtrace_lock);
15712 
15713 		/*
15714 		 * Finally, copy out the buffer description.
15715 		 */
15716 		if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
15717 			return (EFAULT);
15718 
15719 		return (0);
15720 	}
15721 
15722 	case DTRACEIOC_CONF: {
15723 		dtrace_conf_t conf;
15724 
15725 		bzero(&conf, sizeof (conf));
15726 		conf.dtc_difversion = DIF_VERSION;
15727 		conf.dtc_difintregs = DIF_DIR_NREGS;
15728 		conf.dtc_diftupregs = DIF_DTR_NREGS;
15729 		conf.dtc_ctfmodel = CTF_MODEL_NATIVE;
15730 
15731 		if (copyout(&conf, (void *)arg, sizeof (conf)) != 0)
15732 			return (EFAULT);
15733 
15734 		return (0);
15735 	}
15736 
15737 	case DTRACEIOC_STATUS: {
15738 		dtrace_status_t stat;
15739 		dtrace_dstate_t *dstate;
15740 		int i, j;
15741 		uint64_t nerrs;
15742 
15743 		/*
15744 		 * See the comment in dtrace_state_deadman() for the reason
15745 		 * for setting dts_laststatus to INT64_MAX before setting
15746 		 * it to the correct value.
15747 		 */
15748 		state->dts_laststatus = INT64_MAX;
15749 		dtrace_membar_producer();
15750 		state->dts_laststatus = dtrace_gethrtime();
15751 
15752 		bzero(&stat, sizeof (stat));
15753 
15754 		mutex_enter(&dtrace_lock);
15755 
15756 		if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE) {
15757 			mutex_exit(&dtrace_lock);
15758 			return (ENOENT);
15759 		}
15760 
15761 		if (state->dts_activity == DTRACE_ACTIVITY_DRAINING)
15762 			stat.dtst_exiting = 1;
15763 
15764 		nerrs = state->dts_errors;
15765 		dstate = &state->dts_vstate.dtvs_dynvars;
15766 
15767 		for (i = 0; i < NCPU; i++) {
15768 			dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[i];
15769 
15770 			stat.dtst_dyndrops += dcpu->dtdsc_drops;
15771 			stat.dtst_dyndrops_dirty += dcpu->dtdsc_dirty_drops;
15772 			stat.dtst_dyndrops_rinsing += dcpu->dtdsc_rinsing_drops;
15773 
15774 			if (state->dts_buffer[i].dtb_flags & DTRACEBUF_FULL)
15775 				stat.dtst_filled++;
15776 
15777 			nerrs += state->dts_buffer[i].dtb_errors;
15778 
15779 			for (j = 0; j < state->dts_nspeculations; j++) {
15780 				dtrace_speculation_t *spec;
15781 				dtrace_buffer_t *buf;
15782 
15783 				spec = &state->dts_speculations[j];
15784 				buf = &spec->dtsp_buffer[i];
15785 				stat.dtst_specdrops += buf->dtb_xamot_drops;
15786 			}
15787 		}
15788 
15789 		stat.dtst_specdrops_busy = state->dts_speculations_busy;
15790 		stat.dtst_specdrops_unavail = state->dts_speculations_unavail;
15791 		stat.dtst_stkstroverflows = state->dts_stkstroverflows;
15792 		stat.dtst_dblerrors = state->dts_dblerrors;
15793 		stat.dtst_killed =
15794 		    (state->dts_activity == DTRACE_ACTIVITY_KILLED);
15795 		stat.dtst_errors = nerrs;
15796 
15797 		mutex_exit(&dtrace_lock);
15798 
15799 		if (copyout(&stat, (void *)arg, sizeof (stat)) != 0)
15800 			return (EFAULT);
15801 
15802 		return (0);
15803 	}
15804 
15805 	case DTRACEIOC_FORMAT: {
15806 		dtrace_fmtdesc_t fmt;
15807 		char *str;
15808 		int len;
15809 
15810 		if (copyin((void *)arg, &fmt, sizeof (fmt)) != 0)
15811 			return (EFAULT);
15812 
15813 		mutex_enter(&dtrace_lock);
15814 
15815 		if (fmt.dtfd_format == 0 ||
15816 		    fmt.dtfd_format > state->dts_nformats) {
15817 			mutex_exit(&dtrace_lock);
15818 			return (EINVAL);
15819 		}
15820 
15821 		/*
15822 		 * Format strings are allocated contiguously and they are
15823 		 * never freed; if a format index is less than the number
15824 		 * of formats, we can assert that the format map is non-NULL
15825 		 * and that the format for the specified index is non-NULL.
15826 		 */
15827 		ASSERT(state->dts_formats != NULL);
15828 		str = state->dts_formats[fmt.dtfd_format - 1];
15829 		ASSERT(str != NULL);
15830 
15831 		len = strlen(str) + 1;
15832 
15833 		if (len > fmt.dtfd_length) {
15834 			fmt.dtfd_length = len;
15835 
15836 			if (copyout(&fmt, (void *)arg, sizeof (fmt)) != 0) {
15837 				mutex_exit(&dtrace_lock);
15838 				return (EINVAL);
15839 			}
15840 		} else {
15841 			if (copyout(str, fmt.dtfd_string, len) != 0) {
15842 				mutex_exit(&dtrace_lock);
15843 				return (EINVAL);
15844 			}
15845 		}
15846 
15847 		mutex_exit(&dtrace_lock);
15848 		return (0);
15849 	}
15850 
15851 	default:
15852 		break;
15853 	}
15854 
15855 	return (ENOTTY);
15856 }
15857 
15858 /*ARGSUSED*/
15859 static int
15860 dtrace_detach(dev_info_t *dip, ddi_detach_cmd_t cmd)
15861 {
15862 	dtrace_state_t *state;
15863 
15864 	switch (cmd) {
15865 	case DDI_DETACH:
15866 		break;
15867 
15868 	case DDI_SUSPEND:
15869 		return (DDI_SUCCESS);
15870 
15871 	default:
15872 		return (DDI_FAILURE);
15873 	}
15874 
15875 	mutex_enter(&cpu_lock);
15876 	mutex_enter(&dtrace_provider_lock);
15877 	mutex_enter(&dtrace_lock);
15878 
15879 	ASSERT(dtrace_opens == 0);
15880 
15881 	if (dtrace_helpers > 0) {
15882 		mutex_exit(&dtrace_provider_lock);
15883 		mutex_exit(&dtrace_lock);
15884 		mutex_exit(&cpu_lock);
15885 		return (DDI_FAILURE);
15886 	}
15887 
15888 	if (dtrace_unregister((dtrace_provider_id_t)dtrace_provider) != 0) {
15889 		mutex_exit(&dtrace_provider_lock);
15890 		mutex_exit(&dtrace_lock);
15891 		mutex_exit(&cpu_lock);
15892 		return (DDI_FAILURE);
15893 	}
15894 
15895 	dtrace_provider = NULL;
15896 
15897 	if ((state = dtrace_anon_grab()) != NULL) {
15898 		/*
15899 		 * If there were ECBs on this state, the provider should
15900 		 * have not been allowed to detach; assert that there is
15901 		 * none.
15902 		 */
15903 		ASSERT(state->dts_necbs == 0);
15904 		dtrace_state_destroy(state);
15905 
15906 		/*
15907 		 * If we're being detached with anonymous state, we need to
15908 		 * indicate to the kernel debugger that DTrace is now inactive.
15909 		 */
15910 		(void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
15911 	}
15912 
15913 	bzero(&dtrace_anon, sizeof (dtrace_anon_t));
15914 	unregister_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
15915 	dtrace_cpu_init = NULL;
15916 	dtrace_helpers_cleanup = NULL;
15917 	dtrace_helpers_fork = NULL;
15918 	dtrace_cpustart_init = NULL;
15919 	dtrace_cpustart_fini = NULL;
15920 	dtrace_debugger_init = NULL;
15921 	dtrace_debugger_fini = NULL;
15922 	dtrace_modload = NULL;
15923 	dtrace_modunload = NULL;
15924 
15925 	mutex_exit(&cpu_lock);
15926 
15927 	if (dtrace_helptrace_enabled) {
15928 		kmem_free(dtrace_helptrace_buffer, dtrace_helptrace_bufsize);
15929 		dtrace_helptrace_buffer = NULL;
15930 	}
15931 
15932 	kmem_free(dtrace_probes, dtrace_nprobes * sizeof (dtrace_probe_t *));
15933 	dtrace_probes = NULL;
15934 	dtrace_nprobes = 0;
15935 
15936 	dtrace_hash_destroy(dtrace_bymod);
15937 	dtrace_hash_destroy(dtrace_byfunc);
15938 	dtrace_hash_destroy(dtrace_byname);
15939 	dtrace_bymod = NULL;
15940 	dtrace_byfunc = NULL;
15941 	dtrace_byname = NULL;
15942 
15943 	kmem_cache_destroy(dtrace_state_cache);
15944 	vmem_destroy(dtrace_minor);
15945 	vmem_destroy(dtrace_arena);
15946 
15947 	if (dtrace_toxrange != NULL) {
15948 		kmem_free(dtrace_toxrange,
15949 		    dtrace_toxranges_max * sizeof (dtrace_toxrange_t));
15950 		dtrace_toxrange = NULL;
15951 		dtrace_toxranges = 0;
15952 		dtrace_toxranges_max = 0;
15953 	}
15954 
15955 	ddi_remove_minor_node(dtrace_devi, NULL);
15956 	dtrace_devi = NULL;
15957 
15958 	ddi_soft_state_fini(&dtrace_softstate);
15959 
15960 	ASSERT(dtrace_vtime_references == 0);
15961 	ASSERT(dtrace_opens == 0);
15962 	ASSERT(dtrace_retained == NULL);
15963 
15964 	mutex_exit(&dtrace_lock);
15965 	mutex_exit(&dtrace_provider_lock);
15966 
15967 	/*
15968 	 * We don't destroy the task queue until after we have dropped our
15969 	 * locks (taskq_destroy() may block on running tasks).  To prevent
15970 	 * attempting to do work after we have effectively detached but before
15971 	 * the task queue has been destroyed, all tasks dispatched via the
15972 	 * task queue must check that DTrace is still attached before
15973 	 * performing any operation.
15974 	 */
15975 	taskq_destroy(dtrace_taskq);
15976 	dtrace_taskq = NULL;
15977 
15978 	return (DDI_SUCCESS);
15979 }
15980 
15981 /*ARGSUSED*/
15982 static int
15983 dtrace_info(dev_info_t *dip, ddi_info_cmd_t infocmd, void *arg, void **result)
15984 {
15985 	int error;
15986 
15987 	switch (infocmd) {
15988 	case DDI_INFO_DEVT2DEVINFO:
15989 		*result = (void *)dtrace_devi;
15990 		error = DDI_SUCCESS;
15991 		break;
15992 	case DDI_INFO_DEVT2INSTANCE:
15993 		*result = (void *)0;
15994 		error = DDI_SUCCESS;
15995 		break;
15996 	default:
15997 		error = DDI_FAILURE;
15998 	}
15999 	return (error);
16000 }
16001 
16002 static struct cb_ops dtrace_cb_ops = {
16003 	dtrace_open,		/* open */
16004 	dtrace_close,		/* close */
16005 	nulldev,		/* strategy */
16006 	nulldev,		/* print */
16007 	nodev,			/* dump */
16008 	nodev,			/* read */
16009 	nodev,			/* write */
16010 	dtrace_ioctl,		/* ioctl */
16011 	nodev,			/* devmap */
16012 	nodev,			/* mmap */
16013 	nodev,			/* segmap */
16014 	nochpoll,		/* poll */
16015 	ddi_prop_op,		/* cb_prop_op */
16016 	0,			/* streamtab  */
16017 	D_NEW | D_MP		/* Driver compatibility flag */
16018 };
16019 
16020 static struct dev_ops dtrace_ops = {
16021 	DEVO_REV,		/* devo_rev */
16022 	0,			/* refcnt */
16023 	dtrace_info,		/* get_dev_info */
16024 	nulldev,		/* identify */
16025 	nulldev,		/* probe */
16026 	dtrace_attach,		/* attach */
16027 	dtrace_detach,		/* detach */
16028 	nodev,			/* reset */
16029 	&dtrace_cb_ops,		/* driver operations */
16030 	NULL,			/* bus operations */
16031 	nodev,			/* dev power */
16032 	ddi_quiesce_not_needed,		/* quiesce */
16033 };
16034 
16035 static struct modldrv modldrv = {
16036 	&mod_driverops,		/* module type (this is a pseudo driver) */
16037 	"Dynamic Tracing",	/* name of module */
16038 	&dtrace_ops,		/* driver ops */
16039 };
16040 
16041 static struct modlinkage modlinkage = {
16042 	MODREV_1,
16043 	(void *)&modldrv,
16044 	NULL
16045 };
16046 
16047 int
16048 _init(void)
16049 {
16050 	return (mod_install(&modlinkage));
16051 }
16052 
16053 int
16054 _info(struct modinfo *modinfop)
16055 {
16056 	return (mod_info(&modlinkage, modinfop));
16057 }
16058 
16059 int
16060 _fini(void)
16061 {
16062 	return (mod_remove(&modlinkage));
16063 }
16064