xref: /illumos-gate/usr/src/uts/common/dtrace/dtrace.c (revision aedadba64c2e8ed480203bf71975aa68646d7788)
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 2008 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 #pragma ident	"%Z%%M%	%I%	%E% SMI"
28 
29 /*
30  * DTrace - Dynamic Tracing for Solaris
31  *
32  * This is the implementation of the Solaris Dynamic Tracing framework
33  * (DTrace).  The user-visible interface to DTrace is described at length in
34  * the "Solaris Dynamic Tracing Guide".  The interfaces between the libdtrace
35  * library, the in-kernel DTrace framework, and the DTrace providers are
36  * described in the block comments in the <sys/dtrace.h> header file.  The
37  * internal architecture of DTrace is described in the block comments in the
38  * <sys/dtrace_impl.h> header file.  The comments contained within the DTrace
39  * implementation very much assume mastery of all of these sources; if one has
40  * an unanswered question about the implementation, one should consult them
41  * first.
42  *
43  * The functions here are ordered roughly as follows:
44  *
45  *   - Probe context functions
46  *   - Probe hashing functions
47  *   - Non-probe context utility functions
48  *   - Matching functions
49  *   - Provider-to-Framework API functions
50  *   - Probe management functions
51  *   - DIF object functions
52  *   - Format functions
53  *   - Predicate functions
54  *   - ECB functions
55  *   - Buffer functions
56  *   - Enabling functions
57  *   - DOF functions
58  *   - Anonymous enabling functions
59  *   - Consumer state functions
60  *   - Helper functions
61  *   - Hook functions
62  *   - Driver cookbook functions
63  *
64  * Each group of functions begins with a block comment labelled the "DTrace
65  * [Group] Functions", allowing one to find each block by searching forward
66  * on capital-f functions.
67  */
68 #include <sys/errno.h>
69 #include <sys/stat.h>
70 #include <sys/modctl.h>
71 #include <sys/conf.h>
72 #include <sys/systm.h>
73 #include <sys/ddi.h>
74 #include <sys/sunddi.h>
75 #include <sys/cpuvar.h>
76 #include <sys/kmem.h>
77 #include <sys/strsubr.h>
78 #include <sys/sysmacros.h>
79 #include <sys/dtrace_impl.h>
80 #include <sys/atomic.h>
81 #include <sys/cmn_err.h>
82 #include <sys/mutex_impl.h>
83 #include <sys/rwlock_impl.h>
84 #include <sys/ctf_api.h>
85 #include <sys/panic.h>
86 #include <sys/priv_impl.h>
87 #include <sys/policy.h>
88 #include <sys/cred_impl.h>
89 #include <sys/procfs_isa.h>
90 #include <sys/taskq.h>
91 #include <sys/mkdev.h>
92 #include <sys/kdi.h>
93 #include <sys/zone.h>
94 #include <sys/socket.h>
95 #include <netinet/in.h>
96 
97 /*
98  * DTrace Tunable Variables
99  *
100  * The following variables may be tuned by adding a line to /etc/system that
101  * includes both the name of the DTrace module ("dtrace") and the name of the
102  * variable.  For example:
103  *
104  *   set dtrace:dtrace_destructive_disallow = 1
105  *
106  * In general, the only variables that one should be tuning this way are those
107  * that affect system-wide DTrace behavior, and for which the default behavior
108  * is undesirable.  Most of these variables are tunable on a per-consumer
109  * basis using DTrace options, and need not be tuned on a system-wide basis.
110  * When tuning these variables, avoid pathological values; while some attempt
111  * is made to verify the integrity of these variables, they are not considered
112  * part of the supported interface to DTrace, and they are therefore not
113  * checked comprehensively.  Further, these variables should not be tuned
114  * dynamically via "mdb -kw" or other means; they should only be tuned via
115  * /etc/system.
116  */
117 int		dtrace_destructive_disallow = 0;
118 dtrace_optval_t	dtrace_nonroot_maxsize = (16 * 1024 * 1024);
119 size_t		dtrace_difo_maxsize = (256 * 1024);
120 dtrace_optval_t	dtrace_dof_maxsize = (256 * 1024);
121 size_t		dtrace_global_maxsize = (16 * 1024);
122 size_t		dtrace_actions_max = (16 * 1024);
123 size_t		dtrace_retain_max = 1024;
124 dtrace_optval_t	dtrace_helper_actions_max = 32;
125 dtrace_optval_t	dtrace_helper_providers_max = 32;
126 dtrace_optval_t	dtrace_dstate_defsize = (1 * 1024 * 1024);
127 size_t		dtrace_strsize_default = 256;
128 dtrace_optval_t	dtrace_cleanrate_default = 9900990;		/* 101 hz */
129 dtrace_optval_t	dtrace_cleanrate_min = 200000;			/* 5000 hz */
130 dtrace_optval_t	dtrace_cleanrate_max = (uint64_t)60 * NANOSEC;	/* 1/minute */
131 dtrace_optval_t	dtrace_aggrate_default = NANOSEC;		/* 1 hz */
132 dtrace_optval_t	dtrace_statusrate_default = NANOSEC;		/* 1 hz */
133 dtrace_optval_t dtrace_statusrate_max = (hrtime_t)10 * NANOSEC;	 /* 6/minute */
134 dtrace_optval_t	dtrace_switchrate_default = NANOSEC;		/* 1 hz */
135 dtrace_optval_t	dtrace_nspec_default = 1;
136 dtrace_optval_t	dtrace_specsize_default = 32 * 1024;
137 dtrace_optval_t dtrace_stackframes_default = 20;
138 dtrace_optval_t dtrace_ustackframes_default = 20;
139 dtrace_optval_t dtrace_jstackframes_default = 50;
140 dtrace_optval_t dtrace_jstackstrsize_default = 512;
141 int		dtrace_msgdsize_max = 128;
142 hrtime_t	dtrace_chill_max = 500 * (NANOSEC / MILLISEC);	/* 500 ms */
143 hrtime_t	dtrace_chill_interval = NANOSEC;		/* 1000 ms */
144 int		dtrace_devdepth_max = 32;
145 int		dtrace_err_verbose;
146 hrtime_t	dtrace_deadman_interval = NANOSEC;
147 hrtime_t	dtrace_deadman_timeout = (hrtime_t)10 * NANOSEC;
148 hrtime_t	dtrace_deadman_user = (hrtime_t)30 * 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_dynvar_t	dtrace_dynhash_sink;	/* end of dynamic hash chains */
190 
191 /*
192  * DTrace Locking
193  * DTrace is protected by three (relatively coarse-grained) locks:
194  *
195  * (1) dtrace_lock is required to manipulate essentially any DTrace state,
196  *     including enabling state, probes, ECBs, consumer state, helper state,
197  *     etc.  Importantly, dtrace_lock is _not_ required when in probe context;
198  *     probe context is lock-free -- synchronization is handled via the
199  *     dtrace_sync() cross call mechanism.
200  *
201  * (2) dtrace_provider_lock is required when manipulating provider state, or
202  *     when provider state must be held constant.
203  *
204  * (3) dtrace_meta_lock is required when manipulating meta provider state, or
205  *     when meta provider state must be held constant.
206  *
207  * The lock ordering between these three locks is dtrace_meta_lock before
208  * dtrace_provider_lock before dtrace_lock.  (In particular, there are
209  * several places where dtrace_provider_lock is held by the framework as it
210  * calls into the providers -- which then call back into the framework,
211  * grabbing dtrace_lock.)
212  *
213  * There are two other locks in the mix:  mod_lock and cpu_lock.  With respect
214  * to dtrace_provider_lock and dtrace_lock, cpu_lock continues its historical
215  * role as a coarse-grained lock; it is acquired before both of these locks.
216  * With respect to dtrace_meta_lock, its behavior is stranger:  cpu_lock must
217  * be acquired _between_ dtrace_meta_lock and any other DTrace locks.
218  * mod_lock is similar with respect to dtrace_provider_lock in that it must be
219  * acquired _between_ dtrace_provider_lock and dtrace_lock.
220  */
221 static kmutex_t		dtrace_lock;		/* probe state lock */
222 static kmutex_t		dtrace_provider_lock;	/* provider state lock */
223 static kmutex_t		dtrace_meta_lock;	/* meta-provider state lock */
224 
225 /*
226  * DTrace Provider Variables
227  *
228  * These are the variables relating to DTrace as a provider (that is, the
229  * provider of the BEGIN, END, and ERROR probes).
230  */
231 static dtrace_pattr_t	dtrace_provider_attr = {
232 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
233 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
234 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
235 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
236 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
237 };
238 
239 static void
240 dtrace_nullop(void)
241 {}
242 
243 static dtrace_pops_t	dtrace_provider_ops = {
244 	(void (*)(void *, const dtrace_probedesc_t *))dtrace_nullop,
245 	(void (*)(void *, struct modctl *))dtrace_nullop,
246 	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
247 	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
248 	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
249 	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
250 	NULL,
251 	NULL,
252 	NULL,
253 	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop
254 };
255 
256 static dtrace_id_t	dtrace_probeid_begin;	/* special BEGIN probe */
257 static dtrace_id_t	dtrace_probeid_end;	/* special END probe */
258 dtrace_id_t		dtrace_probeid_error;	/* special ERROR probe */
259 
260 /*
261  * DTrace Helper Tracing Variables
262  */
263 uint32_t dtrace_helptrace_next = 0;
264 uint32_t dtrace_helptrace_nlocals;
265 char	*dtrace_helptrace_buffer;
266 int	dtrace_helptrace_bufsize = 512 * 1024;
267 
268 #ifdef DEBUG
269 int	dtrace_helptrace_enabled = 1;
270 #else
271 int	dtrace_helptrace_enabled = 0;
272 #endif
273 
274 /*
275  * DTrace Error Hashing
276  *
277  * On DEBUG kernels, DTrace will track the errors that has seen in a hash
278  * table.  This is very useful for checking coverage of tests that are
279  * expected to induce DIF or DOF processing errors, and may be useful for
280  * debugging problems in the DIF code generator or in DOF generation .  The
281  * error hash may be examined with the ::dtrace_errhash MDB dcmd.
282  */
283 #ifdef DEBUG
284 static dtrace_errhash_t	dtrace_errhash[DTRACE_ERRHASHSZ];
285 static const char *dtrace_errlast;
286 static kthread_t *dtrace_errthread;
287 static kmutex_t dtrace_errlock;
288 #endif
289 
290 /*
291  * DTrace Macros and Constants
292  *
293  * These are various macros that are useful in various spots in the
294  * implementation, along with a few random constants that have no meaning
295  * outside of the implementation.  There is no real structure to this cpp
296  * mishmash -- but is there ever?
297  */
298 #define	DTRACE_HASHSTR(hash, probe)	\
299 	dtrace_hash_str(*((char **)((uintptr_t)(probe) + (hash)->dth_stroffs)))
300 
301 #define	DTRACE_HASHNEXT(hash, probe)	\
302 	(dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_nextoffs)
303 
304 #define	DTRACE_HASHPREV(hash, probe)	\
305 	(dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_prevoffs)
306 
307 #define	DTRACE_HASHEQ(hash, lhs, rhs)	\
308 	(strcmp(*((char **)((uintptr_t)(lhs) + (hash)->dth_stroffs)), \
309 	    *((char **)((uintptr_t)(rhs) + (hash)->dth_stroffs))) == 0)
310 
311 #define	DTRACE_AGGHASHSIZE_SLEW		17
312 
313 #define	DTRACE_V4MAPPED_OFFSET		(sizeof (uint32_t) * 3)
314 
315 /*
316  * The key for a thread-local variable consists of the lower 61 bits of the
317  * t_did, plus the 3 bits of the highest active interrupt above LOCK_LEVEL.
318  * We add DIF_VARIABLE_MAX to t_did to assure that the thread key is never
319  * equal to a variable identifier.  This is necessary (but not sufficient) to
320  * assure that global associative arrays never collide with thread-local
321  * variables.  To guarantee that they cannot collide, we must also define the
322  * order for keying dynamic variables.  That order is:
323  *
324  *   [ key0 ] ... [ keyn ] [ variable-key ] [ tls-key ]
325  *
326  * Because the variable-key and the tls-key are in orthogonal spaces, there is
327  * no way for a global variable key signature to match a thread-local key
328  * signature.
329  */
330 #define	DTRACE_TLS_THRKEY(where) { \
331 	uint_t intr = 0; \
332 	uint_t actv = CPU->cpu_intr_actv >> (LOCK_LEVEL + 1); \
333 	for (; actv; actv >>= 1) \
334 		intr++; \
335 	ASSERT(intr < (1 << 3)); \
336 	(where) = ((curthread->t_did + DIF_VARIABLE_MAX) & \
337 	    (((uint64_t)1 << 61) - 1)) | ((uint64_t)intr << 61); \
338 }
339 
340 #define	DT_BSWAP_8(x)	((x) & 0xff)
341 #define	DT_BSWAP_16(x)	((DT_BSWAP_8(x) << 8) | DT_BSWAP_8((x) >> 8))
342 #define	DT_BSWAP_32(x)	((DT_BSWAP_16(x) << 16) | DT_BSWAP_16((x) >> 16))
343 #define	DT_BSWAP_64(x)	((DT_BSWAP_32(x) << 32) | DT_BSWAP_32((x) >> 32))
344 
345 #define	DT_MASK_LO 0x00000000FFFFFFFFULL
346 
347 #define	DTRACE_STORE(type, tomax, offset, what) \
348 	*((type *)((uintptr_t)(tomax) + (uintptr_t)offset)) = (type)(what);
349 
350 #ifndef __i386
351 #define	DTRACE_ALIGNCHECK(addr, size, flags)				\
352 	if (addr & (size - 1)) {					\
353 		*flags |= CPU_DTRACE_BADALIGN;				\
354 		cpu_core[CPU->cpu_id].cpuc_dtrace_illval = addr;	\
355 		return (0);						\
356 	}
357 #else
358 #define	DTRACE_ALIGNCHECK(addr, size, flags)
359 #endif
360 
361 /*
362  * Test whether a range of memory starting at testaddr of size testsz falls
363  * within the range of memory described by addr, sz.  We take care to avoid
364  * problems with overflow and underflow of the unsigned quantities, and
365  * disallow all negative sizes.  Ranges of size 0 are allowed.
366  */
367 #define	DTRACE_INRANGE(testaddr, testsz, baseaddr, basesz) \
368 	((testaddr) - (baseaddr) < (basesz) && \
369 	(testaddr) + (testsz) - (baseaddr) <= (basesz) && \
370 	(testaddr) + (testsz) >= (testaddr))
371 
372 /*
373  * Test whether alloc_sz bytes will fit in the scratch region.  We isolate
374  * alloc_sz on the righthand side of the comparison in order to avoid overflow
375  * or underflow in the comparison with it.  This is simpler than the INRANGE
376  * check above, because we know that the dtms_scratch_ptr is valid in the
377  * range.  Allocations of size zero are allowed.
378  */
379 #define	DTRACE_INSCRATCH(mstate, alloc_sz) \
380 	((mstate)->dtms_scratch_base + (mstate)->dtms_scratch_size - \
381 	(mstate)->dtms_scratch_ptr >= (alloc_sz))
382 
383 #define	DTRACE_LOADFUNC(bits)						\
384 /*CSTYLED*/								\
385 uint##bits##_t								\
386 dtrace_load##bits(uintptr_t addr)					\
387 {									\
388 	size_t size = bits / NBBY;					\
389 	/*CSTYLED*/							\
390 	uint##bits##_t rval;						\
391 	int i;								\
392 	volatile uint16_t *flags = (volatile uint16_t *)		\
393 	    &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;			\
394 									\
395 	DTRACE_ALIGNCHECK(addr, size, flags);				\
396 									\
397 	for (i = 0; i < dtrace_toxranges; i++) {			\
398 		if (addr >= dtrace_toxrange[i].dtt_limit)		\
399 			continue;					\
400 									\
401 		if (addr + size <= dtrace_toxrange[i].dtt_base)		\
402 			continue;					\
403 									\
404 		/*							\
405 		 * This address falls within a toxic region; return 0.	\
406 		 */							\
407 		*flags |= CPU_DTRACE_BADADDR;				\
408 		cpu_core[CPU->cpu_id].cpuc_dtrace_illval = addr;	\
409 		return (0);						\
410 	}								\
411 									\
412 	*flags |= CPU_DTRACE_NOFAULT;					\
413 	/*CSTYLED*/							\
414 	rval = *((volatile uint##bits##_t *)addr);			\
415 	*flags &= ~CPU_DTRACE_NOFAULT;					\
416 									\
417 	return (!(*flags & CPU_DTRACE_FAULT) ? rval : 0);		\
418 }
419 
420 #ifdef _LP64
421 #define	dtrace_loadptr	dtrace_load64
422 #else
423 #define	dtrace_loadptr	dtrace_load32
424 #endif
425 
426 #define	DTRACE_DYNHASH_FREE	0
427 #define	DTRACE_DYNHASH_SINK	1
428 #define	DTRACE_DYNHASH_VALID	2
429 
430 #define	DTRACE_MATCH_NEXT	0
431 #define	DTRACE_MATCH_DONE	1
432 #define	DTRACE_ANCHORED(probe)	((probe)->dtpr_func[0] != '\0')
433 #define	DTRACE_STATE_ALIGN	64
434 
435 #define	DTRACE_FLAGS2FLT(flags)						\
436 	(((flags) & CPU_DTRACE_BADADDR) ? DTRACEFLT_BADADDR :		\
437 	((flags) & CPU_DTRACE_ILLOP) ? DTRACEFLT_ILLOP :		\
438 	((flags) & CPU_DTRACE_DIVZERO) ? DTRACEFLT_DIVZERO :		\
439 	((flags) & CPU_DTRACE_KPRIV) ? DTRACEFLT_KPRIV :		\
440 	((flags) & CPU_DTRACE_UPRIV) ? DTRACEFLT_UPRIV :		\
441 	((flags) & CPU_DTRACE_TUPOFLOW) ?  DTRACEFLT_TUPOFLOW :		\
442 	((flags) & CPU_DTRACE_BADALIGN) ?  DTRACEFLT_BADALIGN :		\
443 	((flags) & CPU_DTRACE_NOSCRATCH) ?  DTRACEFLT_NOSCRATCH :	\
444 	((flags) & CPU_DTRACE_BADSTACK) ?  DTRACEFLT_BADSTACK :		\
445 	DTRACEFLT_UNKNOWN)
446 
447 #define	DTRACEACT_ISSTRING(act)						\
448 	((act)->dta_kind == DTRACEACT_DIFEXPR &&			\
449 	(act)->dta_difo->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING)
450 
451 static size_t dtrace_strlen(const char *, size_t);
452 static dtrace_probe_t *dtrace_probe_lookup_id(dtrace_id_t id);
453 static void dtrace_enabling_provide(dtrace_provider_t *);
454 static int dtrace_enabling_match(dtrace_enabling_t *, int *);
455 static void dtrace_enabling_matchall(void);
456 static dtrace_state_t *dtrace_anon_grab(void);
457 static uint64_t dtrace_helper(int, dtrace_mstate_t *,
458     dtrace_state_t *, uint64_t, uint64_t);
459 static dtrace_helpers_t *dtrace_helpers_create(proc_t *);
460 static void dtrace_buffer_drop(dtrace_buffer_t *);
461 static intptr_t dtrace_buffer_reserve(dtrace_buffer_t *, size_t, size_t,
462     dtrace_state_t *, dtrace_mstate_t *);
463 static int dtrace_state_option(dtrace_state_t *, dtrace_optid_t,
464     dtrace_optval_t);
465 static int dtrace_ecb_create_enable(dtrace_probe_t *, void *);
466 static void dtrace_helper_provider_destroy(dtrace_helper_provider_t *);
467 
468 /*
469  * DTrace Probe Context Functions
470  *
471  * These functions are called from probe context.  Because probe context is
472  * any context in which C may be called, arbitrarily locks may be held,
473  * interrupts may be disabled, we may be in arbitrary dispatched state, etc.
474  * As a result, functions called from probe context may only call other DTrace
475  * support functions -- they may not interact at all with the system at large.
476  * (Note that the ASSERT macro is made probe-context safe by redefining it in
477  * terms of dtrace_assfail(), a probe-context safe function.) If arbitrary
478  * loads are to be performed from probe context, they _must_ be in terms of
479  * the safe dtrace_load*() variants.
480  *
481  * Some functions in this block are not actually called from probe context;
482  * for these functions, there will be a comment above the function reading
483  * "Note:  not called from probe context."
484  */
485 void
486 dtrace_panic(const char *format, ...)
487 {
488 	va_list alist;
489 
490 	va_start(alist, format);
491 	dtrace_vpanic(format, alist);
492 	va_end(alist);
493 }
494 
495 int
496 dtrace_assfail(const char *a, const char *f, int l)
497 {
498 	dtrace_panic("assertion failed: %s, file: %s, line: %d", a, f, l);
499 
500 	/*
501 	 * We just need something here that even the most clever compiler
502 	 * cannot optimize away.
503 	 */
504 	return (a[(uintptr_t)f]);
505 }
506 
507 /*
508  * Atomically increment a specified error counter from probe context.
509  */
510 static void
511 dtrace_error(uint32_t *counter)
512 {
513 	/*
514 	 * Most counters stored to in probe context are per-CPU counters.
515 	 * However, there are some error conditions that are sufficiently
516 	 * arcane that they don't merit per-CPU storage.  If these counters
517 	 * are incremented concurrently on different CPUs, scalability will be
518 	 * adversely affected -- but we don't expect them to be white-hot in a
519 	 * correctly constructed enabling...
520 	 */
521 	uint32_t oval, nval;
522 
523 	do {
524 		oval = *counter;
525 
526 		if ((nval = oval + 1) == 0) {
527 			/*
528 			 * If the counter would wrap, set it to 1 -- assuring
529 			 * that the counter is never zero when we have seen
530 			 * errors.  (The counter must be 32-bits because we
531 			 * aren't guaranteed a 64-bit compare&swap operation.)
532 			 * To save this code both the infamy of being fingered
533 			 * by a priggish news story and the indignity of being
534 			 * the target of a neo-puritan witch trial, we're
535 			 * carefully avoiding any colorful description of the
536 			 * likelihood of this condition -- but suffice it to
537 			 * say that it is only slightly more likely than the
538 			 * overflow of predicate cache IDs, as discussed in
539 			 * dtrace_predicate_create().
540 			 */
541 			nval = 1;
542 		}
543 	} while (dtrace_cas32(counter, oval, nval) != oval);
544 }
545 
546 /*
547  * Use the DTRACE_LOADFUNC macro to define functions for each of loading a
548  * uint8_t, a uint16_t, a uint32_t and a uint64_t.
549  */
550 DTRACE_LOADFUNC(8)
551 DTRACE_LOADFUNC(16)
552 DTRACE_LOADFUNC(32)
553 DTRACE_LOADFUNC(64)
554 
555 static int
556 dtrace_inscratch(uintptr_t dest, size_t size, dtrace_mstate_t *mstate)
557 {
558 	if (dest < mstate->dtms_scratch_base)
559 		return (0);
560 
561 	if (dest + size < dest)
562 		return (0);
563 
564 	if (dest + size > mstate->dtms_scratch_ptr)
565 		return (0);
566 
567 	return (1);
568 }
569 
570 static int
571 dtrace_canstore_statvar(uint64_t addr, size_t sz,
572     dtrace_statvar_t **svars, int nsvars)
573 {
574 	int i;
575 
576 	for (i = 0; i < nsvars; i++) {
577 		dtrace_statvar_t *svar = svars[i];
578 
579 		if (svar == NULL || svar->dtsv_size == 0)
580 			continue;
581 
582 		if (DTRACE_INRANGE(addr, sz, svar->dtsv_data, svar->dtsv_size))
583 			return (1);
584 	}
585 
586 	return (0);
587 }
588 
589 /*
590  * Check to see if the address is within a memory region to which a store may
591  * be issued.  This includes the DTrace scratch areas, and any DTrace variable
592  * region.  The caller of dtrace_canstore() is responsible for performing any
593  * alignment checks that are needed before stores are actually executed.
594  */
595 static int
596 dtrace_canstore(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
597     dtrace_vstate_t *vstate)
598 {
599 	/*
600 	 * First, check to see if the address is in scratch space...
601 	 */
602 	if (DTRACE_INRANGE(addr, sz, mstate->dtms_scratch_base,
603 	    mstate->dtms_scratch_size))
604 		return (1);
605 
606 	/*
607 	 * Now check to see if it's a dynamic variable.  This check will pick
608 	 * up both thread-local variables and any global dynamically-allocated
609 	 * variables.
610 	 */
611 	if (DTRACE_INRANGE(addr, sz, (uintptr_t)vstate->dtvs_dynvars.dtds_base,
612 	    vstate->dtvs_dynvars.dtds_size)) {
613 		dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
614 		uintptr_t base = (uintptr_t)dstate->dtds_base +
615 		    (dstate->dtds_hashsize * sizeof (dtrace_dynhash_t));
616 		uintptr_t chunkoffs;
617 
618 		/*
619 		 * Before we assume that we can store here, we need to make
620 		 * sure that it isn't in our metadata -- storing to our
621 		 * dynamic variable metadata would corrupt our state.  For
622 		 * the range to not include any dynamic variable metadata,
623 		 * it must:
624 		 *
625 		 *	(1) Start above the hash table that is at the base of
626 		 *	the dynamic variable space
627 		 *
628 		 *	(2) Have a starting chunk offset that is beyond the
629 		 *	dtrace_dynvar_t that is at the base of every chunk
630 		 *
631 		 *	(3) Not span a chunk boundary
632 		 *
633 		 */
634 		if (addr < base)
635 			return (0);
636 
637 		chunkoffs = (addr - base) % dstate->dtds_chunksize;
638 
639 		if (chunkoffs < sizeof (dtrace_dynvar_t))
640 			return (0);
641 
642 		if (chunkoffs + sz > dstate->dtds_chunksize)
643 			return (0);
644 
645 		return (1);
646 	}
647 
648 	/*
649 	 * Finally, check the static local and global variables.  These checks
650 	 * take the longest, so we perform them last.
651 	 */
652 	if (dtrace_canstore_statvar(addr, sz,
653 	    vstate->dtvs_locals, vstate->dtvs_nlocals))
654 		return (1);
655 
656 	if (dtrace_canstore_statvar(addr, sz,
657 	    vstate->dtvs_globals, vstate->dtvs_nglobals))
658 		return (1);
659 
660 	return (0);
661 }
662 
663 
664 /*
665  * Convenience routine to check to see if the address is within a memory
666  * region in which a load may be issued given the user's privilege level;
667  * if not, it sets the appropriate error flags and loads 'addr' into the
668  * illegal value slot.
669  *
670  * DTrace subroutines (DIF_SUBR_*) should use this helper to implement
671  * appropriate memory access protection.
672  */
673 static int
674 dtrace_canload(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
675     dtrace_vstate_t *vstate)
676 {
677 	volatile uintptr_t *illval = &cpu_core[CPU->cpu_id].cpuc_dtrace_illval;
678 
679 	/*
680 	 * If we hold the privilege to read from kernel memory, then
681 	 * everything is readable.
682 	 */
683 	if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
684 		return (1);
685 
686 	/*
687 	 * You can obviously read that which you can store.
688 	 */
689 	if (dtrace_canstore(addr, sz, mstate, vstate))
690 		return (1);
691 
692 	/*
693 	 * We're allowed to read from our own string table.
694 	 */
695 	if (DTRACE_INRANGE(addr, sz, (uintptr_t)mstate->dtms_difo->dtdo_strtab,
696 	    mstate->dtms_difo->dtdo_strlen))
697 		return (1);
698 
699 	DTRACE_CPUFLAG_SET(CPU_DTRACE_KPRIV);
700 	*illval = addr;
701 	return (0);
702 }
703 
704 /*
705  * Convenience routine to check to see if a given string is within a memory
706  * region in which a load may be issued given the user's privilege level;
707  * this exists so that we don't need to issue unnecessary dtrace_strlen()
708  * calls in the event that the user has all privileges.
709  */
710 static int
711 dtrace_strcanload(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
712     dtrace_vstate_t *vstate)
713 {
714 	size_t strsz;
715 
716 	/*
717 	 * If we hold the privilege to read from kernel memory, then
718 	 * everything is readable.
719 	 */
720 	if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
721 		return (1);
722 
723 	strsz = 1 + dtrace_strlen((char *)(uintptr_t)addr, sz);
724 	if (dtrace_canload(addr, strsz, mstate, vstate))
725 		return (1);
726 
727 	return (0);
728 }
729 
730 /*
731  * Convenience routine to check to see if a given variable is within a memory
732  * region in which a load may be issued given the user's privilege level.
733  */
734 static int
735 dtrace_vcanload(void *src, dtrace_diftype_t *type, dtrace_mstate_t *mstate,
736     dtrace_vstate_t *vstate)
737 {
738 	size_t sz;
739 	ASSERT(type->dtdt_flags & DIF_TF_BYREF);
740 
741 	/*
742 	 * If we hold the privilege to read from kernel memory, then
743 	 * everything is readable.
744 	 */
745 	if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
746 		return (1);
747 
748 	if (type->dtdt_kind == DIF_TYPE_STRING)
749 		sz = dtrace_strlen(src,
750 		    vstate->dtvs_state->dts_options[DTRACEOPT_STRSIZE]) + 1;
751 	else
752 		sz = type->dtdt_size;
753 
754 	return (dtrace_canload((uintptr_t)src, sz, mstate, vstate));
755 }
756 
757 /*
758  * Compare two strings using safe loads.
759  */
760 static int
761 dtrace_strncmp(char *s1, char *s2, size_t limit)
762 {
763 	uint8_t c1, c2;
764 	volatile uint16_t *flags;
765 
766 	if (s1 == s2 || limit == 0)
767 		return (0);
768 
769 	flags = (volatile uint16_t *)&cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
770 
771 	do {
772 		if (s1 == NULL) {
773 			c1 = '\0';
774 		} else {
775 			c1 = dtrace_load8((uintptr_t)s1++);
776 		}
777 
778 		if (s2 == NULL) {
779 			c2 = '\0';
780 		} else {
781 			c2 = dtrace_load8((uintptr_t)s2++);
782 		}
783 
784 		if (c1 != c2)
785 			return (c1 - c2);
786 	} while (--limit && c1 != '\0' && !(*flags & CPU_DTRACE_FAULT));
787 
788 	return (0);
789 }
790 
791 /*
792  * Compute strlen(s) for a string using safe memory accesses.  The additional
793  * len parameter is used to specify a maximum length to ensure completion.
794  */
795 static size_t
796 dtrace_strlen(const char *s, size_t lim)
797 {
798 	uint_t len;
799 
800 	for (len = 0; len != lim; len++) {
801 		if (dtrace_load8((uintptr_t)s++) == '\0')
802 			break;
803 	}
804 
805 	return (len);
806 }
807 
808 /*
809  * Check if an address falls within a toxic region.
810  */
811 static int
812 dtrace_istoxic(uintptr_t kaddr, size_t size)
813 {
814 	uintptr_t taddr, tsize;
815 	int i;
816 
817 	for (i = 0; i < dtrace_toxranges; i++) {
818 		taddr = dtrace_toxrange[i].dtt_base;
819 		tsize = dtrace_toxrange[i].dtt_limit - taddr;
820 
821 		if (kaddr - taddr < tsize) {
822 			DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
823 			cpu_core[CPU->cpu_id].cpuc_dtrace_illval = kaddr;
824 			return (1);
825 		}
826 
827 		if (taddr - kaddr < size) {
828 			DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
829 			cpu_core[CPU->cpu_id].cpuc_dtrace_illval = taddr;
830 			return (1);
831 		}
832 	}
833 
834 	return (0);
835 }
836 
837 /*
838  * Copy src to dst using safe memory accesses.  The src is assumed to be unsafe
839  * memory specified by the DIF program.  The dst is assumed to be safe memory
840  * that we can store to directly because it is managed by DTrace.  As with
841  * standard bcopy, overlapping copies are handled properly.
842  */
843 static void
844 dtrace_bcopy(const void *src, void *dst, size_t len)
845 {
846 	if (len != 0) {
847 		uint8_t *s1 = dst;
848 		const uint8_t *s2 = src;
849 
850 		if (s1 <= s2) {
851 			do {
852 				*s1++ = dtrace_load8((uintptr_t)s2++);
853 			} while (--len != 0);
854 		} else {
855 			s2 += len;
856 			s1 += len;
857 
858 			do {
859 				*--s1 = dtrace_load8((uintptr_t)--s2);
860 			} while (--len != 0);
861 		}
862 	}
863 }
864 
865 /*
866  * Copy src to dst using safe memory accesses, up to either the specified
867  * length, or the point that a nul byte is encountered.  The src is assumed to
868  * be unsafe memory specified by the DIF program.  The dst is assumed to be
869  * safe memory that we can store to directly because it is managed by DTrace.
870  * Unlike dtrace_bcopy(), overlapping regions are not handled.
871  */
872 static void
873 dtrace_strcpy(const void *src, void *dst, size_t len)
874 {
875 	if (len != 0) {
876 		uint8_t *s1 = dst, c;
877 		const uint8_t *s2 = src;
878 
879 		do {
880 			*s1++ = c = dtrace_load8((uintptr_t)s2++);
881 		} while (--len != 0 && c != '\0');
882 	}
883 }
884 
885 /*
886  * Copy src to dst, deriving the size and type from the specified (BYREF)
887  * variable type.  The src is assumed to be unsafe memory specified by the DIF
888  * program.  The dst is assumed to be DTrace variable memory that is of the
889  * specified type; we assume that we can store to directly.
890  */
891 static void
892 dtrace_vcopy(void *src, void *dst, dtrace_diftype_t *type)
893 {
894 	ASSERT(type->dtdt_flags & DIF_TF_BYREF);
895 
896 	if (type->dtdt_kind == DIF_TYPE_STRING) {
897 		dtrace_strcpy(src, dst, type->dtdt_size);
898 	} else {
899 		dtrace_bcopy(src, dst, type->dtdt_size);
900 	}
901 }
902 
903 /*
904  * Compare s1 to s2 using safe memory accesses.  The s1 data is assumed to be
905  * unsafe memory specified by the DIF program.  The s2 data is assumed to be
906  * safe memory that we can access directly because it is managed by DTrace.
907  */
908 static int
909 dtrace_bcmp(const void *s1, const void *s2, size_t len)
910 {
911 	volatile uint16_t *flags;
912 
913 	flags = (volatile uint16_t *)&cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
914 
915 	if (s1 == s2)
916 		return (0);
917 
918 	if (s1 == NULL || s2 == NULL)
919 		return (1);
920 
921 	if (s1 != s2 && len != 0) {
922 		const uint8_t *ps1 = s1;
923 		const uint8_t *ps2 = s2;
924 
925 		do {
926 			if (dtrace_load8((uintptr_t)ps1++) != *ps2++)
927 				return (1);
928 		} while (--len != 0 && !(*flags & CPU_DTRACE_FAULT));
929 	}
930 	return (0);
931 }
932 
933 /*
934  * Zero the specified region using a simple byte-by-byte loop.  Note that this
935  * is for safe DTrace-managed memory only.
936  */
937 static void
938 dtrace_bzero(void *dst, size_t len)
939 {
940 	uchar_t *cp;
941 
942 	for (cp = dst; len != 0; len--)
943 		*cp++ = 0;
944 }
945 
946 static void
947 dtrace_add_128(uint64_t *addend1, uint64_t *addend2, uint64_t *sum)
948 {
949 	uint64_t result[2];
950 
951 	result[0] = addend1[0] + addend2[0];
952 	result[1] = addend1[1] + addend2[1] +
953 	    (result[0] < addend1[0] || result[0] < addend2[0] ? 1 : 0);
954 
955 	sum[0] = result[0];
956 	sum[1] = result[1];
957 }
958 
959 /*
960  * Shift the 128-bit value in a by b. If b is positive, shift left.
961  * If b is negative, shift right.
962  */
963 static void
964 dtrace_shift_128(uint64_t *a, int b)
965 {
966 	uint64_t mask;
967 
968 	if (b == 0)
969 		return;
970 
971 	if (b < 0) {
972 		b = -b;
973 		if (b >= 64) {
974 			a[0] = a[1] >> (b - 64);
975 			a[1] = 0;
976 		} else {
977 			a[0] >>= b;
978 			mask = 1LL << (64 - b);
979 			mask -= 1;
980 			a[0] |= ((a[1] & mask) << (64 - b));
981 			a[1] >>= b;
982 		}
983 	} else {
984 		if (b >= 64) {
985 			a[1] = a[0] << (b - 64);
986 			a[0] = 0;
987 		} else {
988 			a[1] <<= b;
989 			mask = a[0] >> (64 - b);
990 			a[1] |= mask;
991 			a[0] <<= b;
992 		}
993 	}
994 }
995 
996 /*
997  * The basic idea is to break the 2 64-bit values into 4 32-bit values,
998  * use native multiplication on those, and then re-combine into the
999  * resulting 128-bit value.
1000  *
1001  * (hi1 << 32 + lo1) * (hi2 << 32 + lo2) =
1002  *     hi1 * hi2 << 64 +
1003  *     hi1 * lo2 << 32 +
1004  *     hi2 * lo1 << 32 +
1005  *     lo1 * lo2
1006  */
1007 static void
1008 dtrace_multiply_128(uint64_t factor1, uint64_t factor2, uint64_t *product)
1009 {
1010 	uint64_t hi1, hi2, lo1, lo2;
1011 	uint64_t tmp[2];
1012 
1013 	hi1 = factor1 >> 32;
1014 	hi2 = factor2 >> 32;
1015 
1016 	lo1 = factor1 & DT_MASK_LO;
1017 	lo2 = factor2 & DT_MASK_LO;
1018 
1019 	product[0] = lo1 * lo2;
1020 	product[1] = hi1 * hi2;
1021 
1022 	tmp[0] = hi1 * lo2;
1023 	tmp[1] = 0;
1024 	dtrace_shift_128(tmp, 32);
1025 	dtrace_add_128(product, tmp, product);
1026 
1027 	tmp[0] = hi2 * lo1;
1028 	tmp[1] = 0;
1029 	dtrace_shift_128(tmp, 32);
1030 	dtrace_add_128(product, tmp, product);
1031 }
1032 
1033 /*
1034  * This privilege check should be used by actions and subroutines to
1035  * verify that the user credentials of the process that enabled the
1036  * invoking ECB match the target credentials
1037  */
1038 static int
1039 dtrace_priv_proc_common_user(dtrace_state_t *state)
1040 {
1041 	cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1042 
1043 	/*
1044 	 * We should always have a non-NULL state cred here, since if cred
1045 	 * is null (anonymous tracing), we fast-path bypass this routine.
1046 	 */
1047 	ASSERT(s_cr != NULL);
1048 
1049 	if ((cr = CRED()) != NULL &&
1050 	    s_cr->cr_uid == cr->cr_uid &&
1051 	    s_cr->cr_uid == cr->cr_ruid &&
1052 	    s_cr->cr_uid == cr->cr_suid &&
1053 	    s_cr->cr_gid == cr->cr_gid &&
1054 	    s_cr->cr_gid == cr->cr_rgid &&
1055 	    s_cr->cr_gid == cr->cr_sgid)
1056 		return (1);
1057 
1058 	return (0);
1059 }
1060 
1061 /*
1062  * This privilege check should be used by actions and subroutines to
1063  * verify that the zone of the process that enabled the invoking ECB
1064  * matches the target credentials
1065  */
1066 static int
1067 dtrace_priv_proc_common_zone(dtrace_state_t *state)
1068 {
1069 	cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1070 
1071 	/*
1072 	 * We should always have a non-NULL state cred here, since if cred
1073 	 * is null (anonymous tracing), we fast-path bypass this routine.
1074 	 */
1075 	ASSERT(s_cr != NULL);
1076 
1077 	if ((cr = CRED()) != NULL &&
1078 	    s_cr->cr_zone == cr->cr_zone)
1079 		return (1);
1080 
1081 	return (0);
1082 }
1083 
1084 /*
1085  * This privilege check should be used by actions and subroutines to
1086  * verify that the process has not setuid or changed credentials.
1087  */
1088 static int
1089 dtrace_priv_proc_common_nocd()
1090 {
1091 	proc_t *proc;
1092 
1093 	if ((proc = ttoproc(curthread)) != NULL &&
1094 	    !(proc->p_flag & SNOCD))
1095 		return (1);
1096 
1097 	return (0);
1098 }
1099 
1100 static int
1101 dtrace_priv_proc_destructive(dtrace_state_t *state)
1102 {
1103 	int action = state->dts_cred.dcr_action;
1104 
1105 	if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE) == 0) &&
1106 	    dtrace_priv_proc_common_zone(state) == 0)
1107 		goto bad;
1108 
1109 	if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER) == 0) &&
1110 	    dtrace_priv_proc_common_user(state) == 0)
1111 		goto bad;
1112 
1113 	if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG) == 0) &&
1114 	    dtrace_priv_proc_common_nocd() == 0)
1115 		goto bad;
1116 
1117 	return (1);
1118 
1119 bad:
1120 	cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1121 
1122 	return (0);
1123 }
1124 
1125 static int
1126 dtrace_priv_proc_control(dtrace_state_t *state)
1127 {
1128 	if (state->dts_cred.dcr_action & DTRACE_CRA_PROC_CONTROL)
1129 		return (1);
1130 
1131 	if (dtrace_priv_proc_common_zone(state) &&
1132 	    dtrace_priv_proc_common_user(state) &&
1133 	    dtrace_priv_proc_common_nocd())
1134 		return (1);
1135 
1136 	cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1137 
1138 	return (0);
1139 }
1140 
1141 static int
1142 dtrace_priv_proc(dtrace_state_t *state)
1143 {
1144 	if (state->dts_cred.dcr_action & DTRACE_CRA_PROC)
1145 		return (1);
1146 
1147 	cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1148 
1149 	return (0);
1150 }
1151 
1152 static int
1153 dtrace_priv_kernel(dtrace_state_t *state)
1154 {
1155 	if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL)
1156 		return (1);
1157 
1158 	cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1159 
1160 	return (0);
1161 }
1162 
1163 static int
1164 dtrace_priv_kernel_destructive(dtrace_state_t *state)
1165 {
1166 	if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL_DESTRUCTIVE)
1167 		return (1);
1168 
1169 	cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1170 
1171 	return (0);
1172 }
1173 
1174 /*
1175  * Note:  not called from probe context.  This function is called
1176  * asynchronously (and at a regular interval) from outside of probe context to
1177  * clean the dirty dynamic variable lists on all CPUs.  Dynamic variable
1178  * cleaning is explained in detail in <sys/dtrace_impl.h>.
1179  */
1180 void
1181 dtrace_dynvar_clean(dtrace_dstate_t *dstate)
1182 {
1183 	dtrace_dynvar_t *dirty;
1184 	dtrace_dstate_percpu_t *dcpu;
1185 	int i, work = 0;
1186 
1187 	for (i = 0; i < NCPU; i++) {
1188 		dcpu = &dstate->dtds_percpu[i];
1189 
1190 		ASSERT(dcpu->dtdsc_rinsing == NULL);
1191 
1192 		/*
1193 		 * If the dirty list is NULL, there is no dirty work to do.
1194 		 */
1195 		if (dcpu->dtdsc_dirty == NULL)
1196 			continue;
1197 
1198 		/*
1199 		 * If the clean list is non-NULL, then we're not going to do
1200 		 * any work for this CPU -- it means that there has not been
1201 		 * a dtrace_dynvar() allocation on this CPU (or from this CPU)
1202 		 * since the last time we cleaned house.
1203 		 */
1204 		if (dcpu->dtdsc_clean != NULL)
1205 			continue;
1206 
1207 		work = 1;
1208 
1209 		/*
1210 		 * Atomically move the dirty list aside.
1211 		 */
1212 		do {
1213 			dirty = dcpu->dtdsc_dirty;
1214 
1215 			/*
1216 			 * Before we zap the dirty list, set the rinsing list.
1217 			 * (This allows for a potential assertion in
1218 			 * dtrace_dynvar():  if a free dynamic variable appears
1219 			 * on a hash chain, either the dirty list or the
1220 			 * rinsing list for some CPU must be non-NULL.)
1221 			 */
1222 			dcpu->dtdsc_rinsing = dirty;
1223 			dtrace_membar_producer();
1224 		} while (dtrace_casptr(&dcpu->dtdsc_dirty,
1225 		    dirty, NULL) != dirty);
1226 	}
1227 
1228 	if (!work) {
1229 		/*
1230 		 * We have no work to do; we can simply return.
1231 		 */
1232 		return;
1233 	}
1234 
1235 	dtrace_sync();
1236 
1237 	for (i = 0; i < NCPU; i++) {
1238 		dcpu = &dstate->dtds_percpu[i];
1239 
1240 		if (dcpu->dtdsc_rinsing == NULL)
1241 			continue;
1242 
1243 		/*
1244 		 * We are now guaranteed that no hash chain contains a pointer
1245 		 * into this dirty list; we can make it clean.
1246 		 */
1247 		ASSERT(dcpu->dtdsc_clean == NULL);
1248 		dcpu->dtdsc_clean = dcpu->dtdsc_rinsing;
1249 		dcpu->dtdsc_rinsing = NULL;
1250 	}
1251 
1252 	/*
1253 	 * Before we actually set the state to be DTRACE_DSTATE_CLEAN, make
1254 	 * sure that all CPUs have seen all of the dtdsc_clean pointers.
1255 	 * This prevents a race whereby a CPU incorrectly decides that
1256 	 * the state should be something other than DTRACE_DSTATE_CLEAN
1257 	 * after dtrace_dynvar_clean() has completed.
1258 	 */
1259 	dtrace_sync();
1260 
1261 	dstate->dtds_state = DTRACE_DSTATE_CLEAN;
1262 }
1263 
1264 /*
1265  * Depending on the value of the op parameter, this function looks-up,
1266  * allocates or deallocates an arbitrarily-keyed dynamic variable.  If an
1267  * allocation is requested, this function will return a pointer to a
1268  * dtrace_dynvar_t corresponding to the allocated variable -- or NULL if no
1269  * variable can be allocated.  If NULL is returned, the appropriate counter
1270  * will be incremented.
1271  */
1272 dtrace_dynvar_t *
1273 dtrace_dynvar(dtrace_dstate_t *dstate, uint_t nkeys,
1274     dtrace_key_t *key, size_t dsize, dtrace_dynvar_op_t op,
1275     dtrace_mstate_t *mstate, dtrace_vstate_t *vstate)
1276 {
1277 	uint64_t hashval = DTRACE_DYNHASH_VALID;
1278 	dtrace_dynhash_t *hash = dstate->dtds_hash;
1279 	dtrace_dynvar_t *free, *new_free, *next, *dvar, *start, *prev = NULL;
1280 	processorid_t me = CPU->cpu_id, cpu = me;
1281 	dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[me];
1282 	size_t bucket, ksize;
1283 	size_t chunksize = dstate->dtds_chunksize;
1284 	uintptr_t kdata, lock, nstate;
1285 	uint_t i;
1286 
1287 	ASSERT(nkeys != 0);
1288 
1289 	/*
1290 	 * Hash the key.  As with aggregations, we use Jenkins' "One-at-a-time"
1291 	 * algorithm.  For the by-value portions, we perform the algorithm in
1292 	 * 16-bit chunks (as opposed to 8-bit chunks).  This speeds things up a
1293 	 * bit, and seems to have only a minute effect on distribution.  For
1294 	 * the by-reference data, we perform "One-at-a-time" iterating (safely)
1295 	 * over each referenced byte.  It's painful to do this, but it's much
1296 	 * better than pathological hash distribution.  The efficacy of the
1297 	 * hashing algorithm (and a comparison with other algorithms) may be
1298 	 * found by running the ::dtrace_dynstat MDB dcmd.
1299 	 */
1300 	for (i = 0; i < nkeys; i++) {
1301 		if (key[i].dttk_size == 0) {
1302 			uint64_t val = key[i].dttk_value;
1303 
1304 			hashval += (val >> 48) & 0xffff;
1305 			hashval += (hashval << 10);
1306 			hashval ^= (hashval >> 6);
1307 
1308 			hashval += (val >> 32) & 0xffff;
1309 			hashval += (hashval << 10);
1310 			hashval ^= (hashval >> 6);
1311 
1312 			hashval += (val >> 16) & 0xffff;
1313 			hashval += (hashval << 10);
1314 			hashval ^= (hashval >> 6);
1315 
1316 			hashval += val & 0xffff;
1317 			hashval += (hashval << 10);
1318 			hashval ^= (hashval >> 6);
1319 		} else {
1320 			/*
1321 			 * This is incredibly painful, but it beats the hell
1322 			 * out of the alternative.
1323 			 */
1324 			uint64_t j, size = key[i].dttk_size;
1325 			uintptr_t base = (uintptr_t)key[i].dttk_value;
1326 
1327 			if (!dtrace_canload(base, size, mstate, vstate))
1328 				break;
1329 
1330 			for (j = 0; j < size; j++) {
1331 				hashval += dtrace_load8(base + j);
1332 				hashval += (hashval << 10);
1333 				hashval ^= (hashval >> 6);
1334 			}
1335 		}
1336 	}
1337 
1338 	if (DTRACE_CPUFLAG_ISSET(CPU_DTRACE_FAULT))
1339 		return (NULL);
1340 
1341 	hashval += (hashval << 3);
1342 	hashval ^= (hashval >> 11);
1343 	hashval += (hashval << 15);
1344 
1345 	/*
1346 	 * There is a remote chance (ideally, 1 in 2^31) that our hashval
1347 	 * comes out to be one of our two sentinel hash values.  If this
1348 	 * actually happens, we set the hashval to be a value known to be a
1349 	 * non-sentinel value.
1350 	 */
1351 	if (hashval == DTRACE_DYNHASH_FREE || hashval == DTRACE_DYNHASH_SINK)
1352 		hashval = DTRACE_DYNHASH_VALID;
1353 
1354 	/*
1355 	 * Yes, it's painful to do a divide here.  If the cycle count becomes
1356 	 * important here, tricks can be pulled to reduce it.  (However, it's
1357 	 * critical that hash collisions be kept to an absolute minimum;
1358 	 * they're much more painful than a divide.)  It's better to have a
1359 	 * solution that generates few collisions and still keeps things
1360 	 * relatively simple.
1361 	 */
1362 	bucket = hashval % dstate->dtds_hashsize;
1363 
1364 	if (op == DTRACE_DYNVAR_DEALLOC) {
1365 		volatile uintptr_t *lockp = &hash[bucket].dtdh_lock;
1366 
1367 		for (;;) {
1368 			while ((lock = *lockp) & 1)
1369 				continue;
1370 
1371 			if (dtrace_casptr((void *)lockp,
1372 			    (void *)lock, (void *)(lock + 1)) == (void *)lock)
1373 				break;
1374 		}
1375 
1376 		dtrace_membar_producer();
1377 	}
1378 
1379 top:
1380 	prev = NULL;
1381 	lock = hash[bucket].dtdh_lock;
1382 
1383 	dtrace_membar_consumer();
1384 
1385 	start = hash[bucket].dtdh_chain;
1386 	ASSERT(start != NULL && (start->dtdv_hashval == DTRACE_DYNHASH_SINK ||
1387 	    start->dtdv_hashval != DTRACE_DYNHASH_FREE ||
1388 	    op != DTRACE_DYNVAR_DEALLOC));
1389 
1390 	for (dvar = start; dvar != NULL; dvar = dvar->dtdv_next) {
1391 		dtrace_tuple_t *dtuple = &dvar->dtdv_tuple;
1392 		dtrace_key_t *dkey = &dtuple->dtt_key[0];
1393 
1394 		if (dvar->dtdv_hashval != hashval) {
1395 			if (dvar->dtdv_hashval == DTRACE_DYNHASH_SINK) {
1396 				/*
1397 				 * We've reached the sink, and therefore the
1398 				 * end of the hash chain; we can kick out of
1399 				 * the loop knowing that we have seen a valid
1400 				 * snapshot of state.
1401 				 */
1402 				ASSERT(dvar->dtdv_next == NULL);
1403 				ASSERT(dvar == &dtrace_dynhash_sink);
1404 				break;
1405 			}
1406 
1407 			if (dvar->dtdv_hashval == DTRACE_DYNHASH_FREE) {
1408 				/*
1409 				 * We've gone off the rails:  somewhere along
1410 				 * the line, one of the members of this hash
1411 				 * chain was deleted.  Note that we could also
1412 				 * detect this by simply letting this loop run
1413 				 * to completion, as we would eventually hit
1414 				 * the end of the dirty list.  However, we
1415 				 * want to avoid running the length of the
1416 				 * dirty list unnecessarily (it might be quite
1417 				 * long), so we catch this as early as
1418 				 * possible by detecting the hash marker.  In
1419 				 * this case, we simply set dvar to NULL and
1420 				 * break; the conditional after the loop will
1421 				 * send us back to top.
1422 				 */
1423 				dvar = NULL;
1424 				break;
1425 			}
1426 
1427 			goto next;
1428 		}
1429 
1430 		if (dtuple->dtt_nkeys != nkeys)
1431 			goto next;
1432 
1433 		for (i = 0; i < nkeys; i++, dkey++) {
1434 			if (dkey->dttk_size != key[i].dttk_size)
1435 				goto next; /* size or type mismatch */
1436 
1437 			if (dkey->dttk_size != 0) {
1438 				if (dtrace_bcmp(
1439 				    (void *)(uintptr_t)key[i].dttk_value,
1440 				    (void *)(uintptr_t)dkey->dttk_value,
1441 				    dkey->dttk_size))
1442 					goto next;
1443 			} else {
1444 				if (dkey->dttk_value != key[i].dttk_value)
1445 					goto next;
1446 			}
1447 		}
1448 
1449 		if (op != DTRACE_DYNVAR_DEALLOC)
1450 			return (dvar);
1451 
1452 		ASSERT(dvar->dtdv_next == NULL ||
1453 		    dvar->dtdv_next->dtdv_hashval != DTRACE_DYNHASH_FREE);
1454 
1455 		if (prev != NULL) {
1456 			ASSERT(hash[bucket].dtdh_chain != dvar);
1457 			ASSERT(start != dvar);
1458 			ASSERT(prev->dtdv_next == dvar);
1459 			prev->dtdv_next = dvar->dtdv_next;
1460 		} else {
1461 			if (dtrace_casptr(&hash[bucket].dtdh_chain,
1462 			    start, dvar->dtdv_next) != start) {
1463 				/*
1464 				 * We have failed to atomically swing the
1465 				 * hash table head pointer, presumably because
1466 				 * of a conflicting allocation on another CPU.
1467 				 * We need to reread the hash chain and try
1468 				 * again.
1469 				 */
1470 				goto top;
1471 			}
1472 		}
1473 
1474 		dtrace_membar_producer();
1475 
1476 		/*
1477 		 * Now set the hash value to indicate that it's free.
1478 		 */
1479 		ASSERT(hash[bucket].dtdh_chain != dvar);
1480 		dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
1481 
1482 		dtrace_membar_producer();
1483 
1484 		/*
1485 		 * Set the next pointer to point at the dirty list, and
1486 		 * atomically swing the dirty pointer to the newly freed dvar.
1487 		 */
1488 		do {
1489 			next = dcpu->dtdsc_dirty;
1490 			dvar->dtdv_next = next;
1491 		} while (dtrace_casptr(&dcpu->dtdsc_dirty, next, dvar) != next);
1492 
1493 		/*
1494 		 * Finally, unlock this hash bucket.
1495 		 */
1496 		ASSERT(hash[bucket].dtdh_lock == lock);
1497 		ASSERT(lock & 1);
1498 		hash[bucket].dtdh_lock++;
1499 
1500 		return (NULL);
1501 next:
1502 		prev = dvar;
1503 		continue;
1504 	}
1505 
1506 	if (dvar == NULL) {
1507 		/*
1508 		 * If dvar is NULL, it is because we went off the rails:
1509 		 * one of the elements that we traversed in the hash chain
1510 		 * was deleted while we were traversing it.  In this case,
1511 		 * we assert that we aren't doing a dealloc (deallocs lock
1512 		 * the hash bucket to prevent themselves from racing with
1513 		 * one another), and retry the hash chain traversal.
1514 		 */
1515 		ASSERT(op != DTRACE_DYNVAR_DEALLOC);
1516 		goto top;
1517 	}
1518 
1519 	if (op != DTRACE_DYNVAR_ALLOC) {
1520 		/*
1521 		 * If we are not to allocate a new variable, we want to
1522 		 * return NULL now.  Before we return, check that the value
1523 		 * of the lock word hasn't changed.  If it has, we may have
1524 		 * seen an inconsistent snapshot.
1525 		 */
1526 		if (op == DTRACE_DYNVAR_NOALLOC) {
1527 			if (hash[bucket].dtdh_lock != lock)
1528 				goto top;
1529 		} else {
1530 			ASSERT(op == DTRACE_DYNVAR_DEALLOC);
1531 			ASSERT(hash[bucket].dtdh_lock == lock);
1532 			ASSERT(lock & 1);
1533 			hash[bucket].dtdh_lock++;
1534 		}
1535 
1536 		return (NULL);
1537 	}
1538 
1539 	/*
1540 	 * We need to allocate a new dynamic variable.  The size we need is the
1541 	 * size of dtrace_dynvar plus the size of nkeys dtrace_key_t's plus the
1542 	 * size of any auxiliary key data (rounded up to 8-byte alignment) plus
1543 	 * the size of any referred-to data (dsize).  We then round the final
1544 	 * size up to the chunksize for allocation.
1545 	 */
1546 	for (ksize = 0, i = 0; i < nkeys; i++)
1547 		ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
1548 
1549 	/*
1550 	 * This should be pretty much impossible, but could happen if, say,
1551 	 * strange DIF specified the tuple.  Ideally, this should be an
1552 	 * assertion and not an error condition -- but that requires that the
1553 	 * chunksize calculation in dtrace_difo_chunksize() be absolutely
1554 	 * bullet-proof.  (That is, it must not be able to be fooled by
1555 	 * malicious DIF.)  Given the lack of backwards branches in DIF,
1556 	 * solving this would presumably not amount to solving the Halting
1557 	 * Problem -- but it still seems awfully hard.
1558 	 */
1559 	if (sizeof (dtrace_dynvar_t) + sizeof (dtrace_key_t) * (nkeys - 1) +
1560 	    ksize + dsize > chunksize) {
1561 		dcpu->dtdsc_drops++;
1562 		return (NULL);
1563 	}
1564 
1565 	nstate = DTRACE_DSTATE_EMPTY;
1566 
1567 	do {
1568 retry:
1569 		free = dcpu->dtdsc_free;
1570 
1571 		if (free == NULL) {
1572 			dtrace_dynvar_t *clean = dcpu->dtdsc_clean;
1573 			void *rval;
1574 
1575 			if (clean == NULL) {
1576 				/*
1577 				 * We're out of dynamic variable space on
1578 				 * this CPU.  Unless we have tried all CPUs,
1579 				 * we'll try to allocate from a different
1580 				 * CPU.
1581 				 */
1582 				switch (dstate->dtds_state) {
1583 				case DTRACE_DSTATE_CLEAN: {
1584 					void *sp = &dstate->dtds_state;
1585 
1586 					if (++cpu >= NCPU)
1587 						cpu = 0;
1588 
1589 					if (dcpu->dtdsc_dirty != NULL &&
1590 					    nstate == DTRACE_DSTATE_EMPTY)
1591 						nstate = DTRACE_DSTATE_DIRTY;
1592 
1593 					if (dcpu->dtdsc_rinsing != NULL)
1594 						nstate = DTRACE_DSTATE_RINSING;
1595 
1596 					dcpu = &dstate->dtds_percpu[cpu];
1597 
1598 					if (cpu != me)
1599 						goto retry;
1600 
1601 					(void) dtrace_cas32(sp,
1602 					    DTRACE_DSTATE_CLEAN, nstate);
1603 
1604 					/*
1605 					 * To increment the correct bean
1606 					 * counter, take another lap.
1607 					 */
1608 					goto retry;
1609 				}
1610 
1611 				case DTRACE_DSTATE_DIRTY:
1612 					dcpu->dtdsc_dirty_drops++;
1613 					break;
1614 
1615 				case DTRACE_DSTATE_RINSING:
1616 					dcpu->dtdsc_rinsing_drops++;
1617 					break;
1618 
1619 				case DTRACE_DSTATE_EMPTY:
1620 					dcpu->dtdsc_drops++;
1621 					break;
1622 				}
1623 
1624 				DTRACE_CPUFLAG_SET(CPU_DTRACE_DROP);
1625 				return (NULL);
1626 			}
1627 
1628 			/*
1629 			 * The clean list appears to be non-empty.  We want to
1630 			 * move the clean list to the free list; we start by
1631 			 * moving the clean pointer aside.
1632 			 */
1633 			if (dtrace_casptr(&dcpu->dtdsc_clean,
1634 			    clean, NULL) != clean) {
1635 				/*
1636 				 * We are in one of two situations:
1637 				 *
1638 				 *  (a)	The clean list was switched to the
1639 				 *	free list by another CPU.
1640 				 *
1641 				 *  (b)	The clean list was added to by the
1642 				 *	cleansing cyclic.
1643 				 *
1644 				 * In either of these situations, we can
1645 				 * just reattempt the free list allocation.
1646 				 */
1647 				goto retry;
1648 			}
1649 
1650 			ASSERT(clean->dtdv_hashval == DTRACE_DYNHASH_FREE);
1651 
1652 			/*
1653 			 * Now we'll move the clean list to the free list.
1654 			 * It's impossible for this to fail:  the only way
1655 			 * the free list can be updated is through this
1656 			 * code path, and only one CPU can own the clean list.
1657 			 * Thus, it would only be possible for this to fail if
1658 			 * this code were racing with dtrace_dynvar_clean().
1659 			 * (That is, if dtrace_dynvar_clean() updated the clean
1660 			 * list, and we ended up racing to update the free
1661 			 * list.)  This race is prevented by the dtrace_sync()
1662 			 * in dtrace_dynvar_clean() -- which flushes the
1663 			 * owners of the clean lists out before resetting
1664 			 * the clean lists.
1665 			 */
1666 			rval = dtrace_casptr(&dcpu->dtdsc_free, NULL, clean);
1667 			ASSERT(rval == NULL);
1668 			goto retry;
1669 		}
1670 
1671 		dvar = free;
1672 		new_free = dvar->dtdv_next;
1673 	} while (dtrace_casptr(&dcpu->dtdsc_free, free, new_free) != free);
1674 
1675 	/*
1676 	 * We have now allocated a new chunk.  We copy the tuple keys into the
1677 	 * tuple array and copy any referenced key data into the data space
1678 	 * following the tuple array.  As we do this, we relocate dttk_value
1679 	 * in the final tuple to point to the key data address in the chunk.
1680 	 */
1681 	kdata = (uintptr_t)&dvar->dtdv_tuple.dtt_key[nkeys];
1682 	dvar->dtdv_data = (void *)(kdata + ksize);
1683 	dvar->dtdv_tuple.dtt_nkeys = nkeys;
1684 
1685 	for (i = 0; i < nkeys; i++) {
1686 		dtrace_key_t *dkey = &dvar->dtdv_tuple.dtt_key[i];
1687 		size_t kesize = key[i].dttk_size;
1688 
1689 		if (kesize != 0) {
1690 			dtrace_bcopy(
1691 			    (const void *)(uintptr_t)key[i].dttk_value,
1692 			    (void *)kdata, kesize);
1693 			dkey->dttk_value = kdata;
1694 			kdata += P2ROUNDUP(kesize, sizeof (uint64_t));
1695 		} else {
1696 			dkey->dttk_value = key[i].dttk_value;
1697 		}
1698 
1699 		dkey->dttk_size = kesize;
1700 	}
1701 
1702 	ASSERT(dvar->dtdv_hashval == DTRACE_DYNHASH_FREE);
1703 	dvar->dtdv_hashval = hashval;
1704 	dvar->dtdv_next = start;
1705 
1706 	if (dtrace_casptr(&hash[bucket].dtdh_chain, start, dvar) == start)
1707 		return (dvar);
1708 
1709 	/*
1710 	 * The cas has failed.  Either another CPU is adding an element to
1711 	 * this hash chain, or another CPU is deleting an element from this
1712 	 * hash chain.  The simplest way to deal with both of these cases
1713 	 * (though not necessarily the most efficient) is to free our
1714 	 * allocated block and tail-call ourselves.  Note that the free is
1715 	 * to the dirty list and _not_ to the free list.  This is to prevent
1716 	 * races with allocators, above.
1717 	 */
1718 	dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
1719 
1720 	dtrace_membar_producer();
1721 
1722 	do {
1723 		free = dcpu->dtdsc_dirty;
1724 		dvar->dtdv_next = free;
1725 	} while (dtrace_casptr(&dcpu->dtdsc_dirty, free, dvar) != free);
1726 
1727 	return (dtrace_dynvar(dstate, nkeys, key, dsize, op, mstate, vstate));
1728 }
1729 
1730 /*ARGSUSED*/
1731 static void
1732 dtrace_aggregate_min(uint64_t *oval, uint64_t nval, uint64_t arg)
1733 {
1734 	if ((int64_t)nval < (int64_t)*oval)
1735 		*oval = nval;
1736 }
1737 
1738 /*ARGSUSED*/
1739 static void
1740 dtrace_aggregate_max(uint64_t *oval, uint64_t nval, uint64_t arg)
1741 {
1742 	if ((int64_t)nval > (int64_t)*oval)
1743 		*oval = nval;
1744 }
1745 
1746 static void
1747 dtrace_aggregate_quantize(uint64_t *quanta, uint64_t nval, uint64_t incr)
1748 {
1749 	int i, zero = DTRACE_QUANTIZE_ZEROBUCKET;
1750 	int64_t val = (int64_t)nval;
1751 
1752 	if (val < 0) {
1753 		for (i = 0; i < zero; i++) {
1754 			if (val <= DTRACE_QUANTIZE_BUCKETVAL(i)) {
1755 				quanta[i] += incr;
1756 				return;
1757 			}
1758 		}
1759 	} else {
1760 		for (i = zero + 1; i < DTRACE_QUANTIZE_NBUCKETS; i++) {
1761 			if (val < DTRACE_QUANTIZE_BUCKETVAL(i)) {
1762 				quanta[i - 1] += incr;
1763 				return;
1764 			}
1765 		}
1766 
1767 		quanta[DTRACE_QUANTIZE_NBUCKETS - 1] += incr;
1768 		return;
1769 	}
1770 
1771 	ASSERT(0);
1772 }
1773 
1774 static void
1775 dtrace_aggregate_lquantize(uint64_t *lquanta, uint64_t nval, uint64_t incr)
1776 {
1777 	uint64_t arg = *lquanta++;
1778 	int32_t base = DTRACE_LQUANTIZE_BASE(arg);
1779 	uint16_t step = DTRACE_LQUANTIZE_STEP(arg);
1780 	uint16_t levels = DTRACE_LQUANTIZE_LEVELS(arg);
1781 	int32_t val = (int32_t)nval, level;
1782 
1783 	ASSERT(step != 0);
1784 	ASSERT(levels != 0);
1785 
1786 	if (val < base) {
1787 		/*
1788 		 * This is an underflow.
1789 		 */
1790 		lquanta[0] += incr;
1791 		return;
1792 	}
1793 
1794 	level = (val - base) / step;
1795 
1796 	if (level < levels) {
1797 		lquanta[level + 1] += incr;
1798 		return;
1799 	}
1800 
1801 	/*
1802 	 * This is an overflow.
1803 	 */
1804 	lquanta[levels + 1] += incr;
1805 }
1806 
1807 /*ARGSUSED*/
1808 static void
1809 dtrace_aggregate_avg(uint64_t *data, uint64_t nval, uint64_t arg)
1810 {
1811 	data[0]++;
1812 	data[1] += nval;
1813 }
1814 
1815 /*ARGSUSED*/
1816 static void
1817 dtrace_aggregate_stddev(uint64_t *data, uint64_t nval, uint64_t arg)
1818 {
1819 	int64_t snval = (int64_t)nval;
1820 	uint64_t tmp[2];
1821 
1822 	data[0]++;
1823 	data[1] += nval;
1824 
1825 	/*
1826 	 * What we want to say here is:
1827 	 *
1828 	 * data[2] += nval * nval;
1829 	 *
1830 	 * But given that nval is 64-bit, we could easily overflow, so
1831 	 * we do this as 128-bit arithmetic.
1832 	 */
1833 	if (snval < 0)
1834 		snval = -snval;
1835 
1836 	dtrace_multiply_128((uint64_t)snval, (uint64_t)snval, tmp);
1837 	dtrace_add_128(data + 2, tmp, data + 2);
1838 }
1839 
1840 /*ARGSUSED*/
1841 static void
1842 dtrace_aggregate_count(uint64_t *oval, uint64_t nval, uint64_t arg)
1843 {
1844 	*oval = *oval + 1;
1845 }
1846 
1847 /*ARGSUSED*/
1848 static void
1849 dtrace_aggregate_sum(uint64_t *oval, uint64_t nval, uint64_t arg)
1850 {
1851 	*oval += nval;
1852 }
1853 
1854 /*
1855  * Aggregate given the tuple in the principal data buffer, and the aggregating
1856  * action denoted by the specified dtrace_aggregation_t.  The aggregation
1857  * buffer is specified as the buf parameter.  This routine does not return
1858  * failure; if there is no space in the aggregation buffer, the data will be
1859  * dropped, and a corresponding counter incremented.
1860  */
1861 static void
1862 dtrace_aggregate(dtrace_aggregation_t *agg, dtrace_buffer_t *dbuf,
1863     intptr_t offset, dtrace_buffer_t *buf, uint64_t expr, uint64_t arg)
1864 {
1865 	dtrace_recdesc_t *rec = &agg->dtag_action.dta_rec;
1866 	uint32_t i, ndx, size, fsize;
1867 	uint32_t align = sizeof (uint64_t) - 1;
1868 	dtrace_aggbuffer_t *agb;
1869 	dtrace_aggkey_t *key;
1870 	uint32_t hashval = 0, limit, isstr;
1871 	caddr_t tomax, data, kdata;
1872 	dtrace_actkind_t action;
1873 	dtrace_action_t *act;
1874 	uintptr_t offs;
1875 
1876 	if (buf == NULL)
1877 		return;
1878 
1879 	if (!agg->dtag_hasarg) {
1880 		/*
1881 		 * Currently, only quantize() and lquantize() take additional
1882 		 * arguments, and they have the same semantics:  an increment
1883 		 * value that defaults to 1 when not present.  If additional
1884 		 * aggregating actions take arguments, the setting of the
1885 		 * default argument value will presumably have to become more
1886 		 * sophisticated...
1887 		 */
1888 		arg = 1;
1889 	}
1890 
1891 	action = agg->dtag_action.dta_kind - DTRACEACT_AGGREGATION;
1892 	size = rec->dtrd_offset - agg->dtag_base;
1893 	fsize = size + rec->dtrd_size;
1894 
1895 	ASSERT(dbuf->dtb_tomax != NULL);
1896 	data = dbuf->dtb_tomax + offset + agg->dtag_base;
1897 
1898 	if ((tomax = buf->dtb_tomax) == NULL) {
1899 		dtrace_buffer_drop(buf);
1900 		return;
1901 	}
1902 
1903 	/*
1904 	 * The metastructure is always at the bottom of the buffer.
1905 	 */
1906 	agb = (dtrace_aggbuffer_t *)(tomax + buf->dtb_size -
1907 	    sizeof (dtrace_aggbuffer_t));
1908 
1909 	if (buf->dtb_offset == 0) {
1910 		/*
1911 		 * We just kludge up approximately 1/8th of the size to be
1912 		 * buckets.  If this guess ends up being routinely
1913 		 * off-the-mark, we may need to dynamically readjust this
1914 		 * based on past performance.
1915 		 */
1916 		uintptr_t hashsize = (buf->dtb_size >> 3) / sizeof (uintptr_t);
1917 
1918 		if ((uintptr_t)agb - hashsize * sizeof (dtrace_aggkey_t *) <
1919 		    (uintptr_t)tomax || hashsize == 0) {
1920 			/*
1921 			 * We've been given a ludicrously small buffer;
1922 			 * increment our drop count and leave.
1923 			 */
1924 			dtrace_buffer_drop(buf);
1925 			return;
1926 		}
1927 
1928 		/*
1929 		 * And now, a pathetic attempt to try to get a an odd (or
1930 		 * perchance, a prime) hash size for better hash distribution.
1931 		 */
1932 		if (hashsize > (DTRACE_AGGHASHSIZE_SLEW << 3))
1933 			hashsize -= DTRACE_AGGHASHSIZE_SLEW;
1934 
1935 		agb->dtagb_hashsize = hashsize;
1936 		agb->dtagb_hash = (dtrace_aggkey_t **)((uintptr_t)agb -
1937 		    agb->dtagb_hashsize * sizeof (dtrace_aggkey_t *));
1938 		agb->dtagb_free = (uintptr_t)agb->dtagb_hash;
1939 
1940 		for (i = 0; i < agb->dtagb_hashsize; i++)
1941 			agb->dtagb_hash[i] = NULL;
1942 	}
1943 
1944 	ASSERT(agg->dtag_first != NULL);
1945 	ASSERT(agg->dtag_first->dta_intuple);
1946 
1947 	/*
1948 	 * Calculate the hash value based on the key.  Note that we _don't_
1949 	 * include the aggid in the hashing (but we will store it as part of
1950 	 * the key).  The hashing algorithm is Bob Jenkins' "One-at-a-time"
1951 	 * algorithm: a simple, quick algorithm that has no known funnels, and
1952 	 * gets good distribution in practice.  The efficacy of the hashing
1953 	 * algorithm (and a comparison with other algorithms) may be found by
1954 	 * running the ::dtrace_aggstat MDB dcmd.
1955 	 */
1956 	for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
1957 		i = act->dta_rec.dtrd_offset - agg->dtag_base;
1958 		limit = i + act->dta_rec.dtrd_size;
1959 		ASSERT(limit <= size);
1960 		isstr = DTRACEACT_ISSTRING(act);
1961 
1962 		for (; i < limit; i++) {
1963 			hashval += data[i];
1964 			hashval += (hashval << 10);
1965 			hashval ^= (hashval >> 6);
1966 
1967 			if (isstr && data[i] == '\0')
1968 				break;
1969 		}
1970 	}
1971 
1972 	hashval += (hashval << 3);
1973 	hashval ^= (hashval >> 11);
1974 	hashval += (hashval << 15);
1975 
1976 	/*
1977 	 * Yes, the divide here is expensive -- but it's generally the least
1978 	 * of the performance issues given the amount of data that we iterate
1979 	 * over to compute hash values, compare data, etc.
1980 	 */
1981 	ndx = hashval % agb->dtagb_hashsize;
1982 
1983 	for (key = agb->dtagb_hash[ndx]; key != NULL; key = key->dtak_next) {
1984 		ASSERT((caddr_t)key >= tomax);
1985 		ASSERT((caddr_t)key < tomax + buf->dtb_size);
1986 
1987 		if (hashval != key->dtak_hashval || key->dtak_size != size)
1988 			continue;
1989 
1990 		kdata = key->dtak_data;
1991 		ASSERT(kdata >= tomax && kdata < tomax + buf->dtb_size);
1992 
1993 		for (act = agg->dtag_first; act->dta_intuple;
1994 		    act = act->dta_next) {
1995 			i = act->dta_rec.dtrd_offset - agg->dtag_base;
1996 			limit = i + act->dta_rec.dtrd_size;
1997 			ASSERT(limit <= size);
1998 			isstr = DTRACEACT_ISSTRING(act);
1999 
2000 			for (; i < limit; i++) {
2001 				if (kdata[i] != data[i])
2002 					goto next;
2003 
2004 				if (isstr && data[i] == '\0')
2005 					break;
2006 			}
2007 		}
2008 
2009 		if (action != key->dtak_action) {
2010 			/*
2011 			 * We are aggregating on the same value in the same
2012 			 * aggregation with two different aggregating actions.
2013 			 * (This should have been picked up in the compiler,
2014 			 * so we may be dealing with errant or devious DIF.)
2015 			 * This is an error condition; we indicate as much,
2016 			 * and return.
2017 			 */
2018 			DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
2019 			return;
2020 		}
2021 
2022 		/*
2023 		 * This is a hit:  we need to apply the aggregator to
2024 		 * the value at this key.
2025 		 */
2026 		agg->dtag_aggregate((uint64_t *)(kdata + size), expr, arg);
2027 		return;
2028 next:
2029 		continue;
2030 	}
2031 
2032 	/*
2033 	 * We didn't find it.  We need to allocate some zero-filled space,
2034 	 * link it into the hash table appropriately, and apply the aggregator
2035 	 * to the (zero-filled) value.
2036 	 */
2037 	offs = buf->dtb_offset;
2038 	while (offs & (align - 1))
2039 		offs += sizeof (uint32_t);
2040 
2041 	/*
2042 	 * If we don't have enough room to both allocate a new key _and_
2043 	 * its associated data, increment the drop count and return.
2044 	 */
2045 	if ((uintptr_t)tomax + offs + fsize >
2046 	    agb->dtagb_free - sizeof (dtrace_aggkey_t)) {
2047 		dtrace_buffer_drop(buf);
2048 		return;
2049 	}
2050 
2051 	/*CONSTCOND*/
2052 	ASSERT(!(sizeof (dtrace_aggkey_t) & (sizeof (uintptr_t) - 1)));
2053 	key = (dtrace_aggkey_t *)(agb->dtagb_free - sizeof (dtrace_aggkey_t));
2054 	agb->dtagb_free -= sizeof (dtrace_aggkey_t);
2055 
2056 	key->dtak_data = kdata = tomax + offs;
2057 	buf->dtb_offset = offs + fsize;
2058 
2059 	/*
2060 	 * Now copy the data across.
2061 	 */
2062 	*((dtrace_aggid_t *)kdata) = agg->dtag_id;
2063 
2064 	for (i = sizeof (dtrace_aggid_t); i < size; i++)
2065 		kdata[i] = data[i];
2066 
2067 	/*
2068 	 * Because strings are not zeroed out by default, we need to iterate
2069 	 * looking for actions that store strings, and we need to explicitly
2070 	 * pad these strings out with zeroes.
2071 	 */
2072 	for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
2073 		int nul;
2074 
2075 		if (!DTRACEACT_ISSTRING(act))
2076 			continue;
2077 
2078 		i = act->dta_rec.dtrd_offset - agg->dtag_base;
2079 		limit = i + act->dta_rec.dtrd_size;
2080 		ASSERT(limit <= size);
2081 
2082 		for (nul = 0; i < limit; i++) {
2083 			if (nul) {
2084 				kdata[i] = '\0';
2085 				continue;
2086 			}
2087 
2088 			if (data[i] != '\0')
2089 				continue;
2090 
2091 			nul = 1;
2092 		}
2093 	}
2094 
2095 	for (i = size; i < fsize; i++)
2096 		kdata[i] = 0;
2097 
2098 	key->dtak_hashval = hashval;
2099 	key->dtak_size = size;
2100 	key->dtak_action = action;
2101 	key->dtak_next = agb->dtagb_hash[ndx];
2102 	agb->dtagb_hash[ndx] = key;
2103 
2104 	/*
2105 	 * Finally, apply the aggregator.
2106 	 */
2107 	*((uint64_t *)(key->dtak_data + size)) = agg->dtag_initial;
2108 	agg->dtag_aggregate((uint64_t *)(key->dtak_data + size), expr, arg);
2109 }
2110 
2111 /*
2112  * Given consumer state, this routine finds a speculation in the INACTIVE
2113  * state and transitions it into the ACTIVE state.  If there is no speculation
2114  * in the INACTIVE state, 0 is returned.  In this case, no error counter is
2115  * incremented -- it is up to the caller to take appropriate action.
2116  */
2117 static int
2118 dtrace_speculation(dtrace_state_t *state)
2119 {
2120 	int i = 0;
2121 	dtrace_speculation_state_t current;
2122 	uint32_t *stat = &state->dts_speculations_unavail, count;
2123 
2124 	while (i < state->dts_nspeculations) {
2125 		dtrace_speculation_t *spec = &state->dts_speculations[i];
2126 
2127 		current = spec->dtsp_state;
2128 
2129 		if (current != DTRACESPEC_INACTIVE) {
2130 			if (current == DTRACESPEC_COMMITTINGMANY ||
2131 			    current == DTRACESPEC_COMMITTING ||
2132 			    current == DTRACESPEC_DISCARDING)
2133 				stat = &state->dts_speculations_busy;
2134 			i++;
2135 			continue;
2136 		}
2137 
2138 		if (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2139 		    current, DTRACESPEC_ACTIVE) == current)
2140 			return (i + 1);
2141 	}
2142 
2143 	/*
2144 	 * We couldn't find a speculation.  If we found as much as a single
2145 	 * busy speculation buffer, we'll attribute this failure as "busy"
2146 	 * instead of "unavail".
2147 	 */
2148 	do {
2149 		count = *stat;
2150 	} while (dtrace_cas32(stat, count, count + 1) != count);
2151 
2152 	return (0);
2153 }
2154 
2155 /*
2156  * This routine commits an active speculation.  If the specified speculation
2157  * is not in a valid state to perform a commit(), this routine will silently do
2158  * nothing.  The state of the specified speculation is transitioned according
2159  * to the state transition diagram outlined in <sys/dtrace_impl.h>
2160  */
2161 static void
2162 dtrace_speculation_commit(dtrace_state_t *state, processorid_t cpu,
2163     dtrace_specid_t which)
2164 {
2165 	dtrace_speculation_t *spec;
2166 	dtrace_buffer_t *src, *dest;
2167 	uintptr_t daddr, saddr, dlimit;
2168 	dtrace_speculation_state_t current, new;
2169 	intptr_t offs;
2170 
2171 	if (which == 0)
2172 		return;
2173 
2174 	if (which > state->dts_nspeculations) {
2175 		cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2176 		return;
2177 	}
2178 
2179 	spec = &state->dts_speculations[which - 1];
2180 	src = &spec->dtsp_buffer[cpu];
2181 	dest = &state->dts_buffer[cpu];
2182 
2183 	do {
2184 		current = spec->dtsp_state;
2185 
2186 		if (current == DTRACESPEC_COMMITTINGMANY)
2187 			break;
2188 
2189 		switch (current) {
2190 		case DTRACESPEC_INACTIVE:
2191 		case DTRACESPEC_DISCARDING:
2192 			return;
2193 
2194 		case DTRACESPEC_COMMITTING:
2195 			/*
2196 			 * This is only possible if we are (a) commit()'ing
2197 			 * without having done a prior speculate() on this CPU
2198 			 * and (b) racing with another commit() on a different
2199 			 * CPU.  There's nothing to do -- we just assert that
2200 			 * our offset is 0.
2201 			 */
2202 			ASSERT(src->dtb_offset == 0);
2203 			return;
2204 
2205 		case DTRACESPEC_ACTIVE:
2206 			new = DTRACESPEC_COMMITTING;
2207 			break;
2208 
2209 		case DTRACESPEC_ACTIVEONE:
2210 			/*
2211 			 * This speculation is active on one CPU.  If our
2212 			 * buffer offset is non-zero, we know that the one CPU
2213 			 * must be us.  Otherwise, we are committing on a
2214 			 * different CPU from the speculate(), and we must
2215 			 * rely on being asynchronously cleaned.
2216 			 */
2217 			if (src->dtb_offset != 0) {
2218 				new = DTRACESPEC_COMMITTING;
2219 				break;
2220 			}
2221 			/*FALLTHROUGH*/
2222 
2223 		case DTRACESPEC_ACTIVEMANY:
2224 			new = DTRACESPEC_COMMITTINGMANY;
2225 			break;
2226 
2227 		default:
2228 			ASSERT(0);
2229 		}
2230 	} while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2231 	    current, new) != current);
2232 
2233 	/*
2234 	 * We have set the state to indicate that we are committing this
2235 	 * speculation.  Now reserve the necessary space in the destination
2236 	 * buffer.
2237 	 */
2238 	if ((offs = dtrace_buffer_reserve(dest, src->dtb_offset,
2239 	    sizeof (uint64_t), state, NULL)) < 0) {
2240 		dtrace_buffer_drop(dest);
2241 		goto out;
2242 	}
2243 
2244 	/*
2245 	 * We have the space; copy the buffer across.  (Note that this is a
2246 	 * highly subobtimal bcopy(); in the unlikely event that this becomes
2247 	 * a serious performance issue, a high-performance DTrace-specific
2248 	 * bcopy() should obviously be invented.)
2249 	 */
2250 	daddr = (uintptr_t)dest->dtb_tomax + offs;
2251 	dlimit = daddr + src->dtb_offset;
2252 	saddr = (uintptr_t)src->dtb_tomax;
2253 
2254 	/*
2255 	 * First, the aligned portion.
2256 	 */
2257 	while (dlimit - daddr >= sizeof (uint64_t)) {
2258 		*((uint64_t *)daddr) = *((uint64_t *)saddr);
2259 
2260 		daddr += sizeof (uint64_t);
2261 		saddr += sizeof (uint64_t);
2262 	}
2263 
2264 	/*
2265 	 * Now any left-over bit...
2266 	 */
2267 	while (dlimit - daddr)
2268 		*((uint8_t *)daddr++) = *((uint8_t *)saddr++);
2269 
2270 	/*
2271 	 * Finally, commit the reserved space in the destination buffer.
2272 	 */
2273 	dest->dtb_offset = offs + src->dtb_offset;
2274 
2275 out:
2276 	/*
2277 	 * If we're lucky enough to be the only active CPU on this speculation
2278 	 * buffer, we can just set the state back to DTRACESPEC_INACTIVE.
2279 	 */
2280 	if (current == DTRACESPEC_ACTIVE ||
2281 	    (current == DTRACESPEC_ACTIVEONE && new == DTRACESPEC_COMMITTING)) {
2282 		uint32_t rval = dtrace_cas32((uint32_t *)&spec->dtsp_state,
2283 		    DTRACESPEC_COMMITTING, DTRACESPEC_INACTIVE);
2284 
2285 		ASSERT(rval == DTRACESPEC_COMMITTING);
2286 	}
2287 
2288 	src->dtb_offset = 0;
2289 	src->dtb_xamot_drops += src->dtb_drops;
2290 	src->dtb_drops = 0;
2291 }
2292 
2293 /*
2294  * This routine discards an active speculation.  If the specified speculation
2295  * is not in a valid state to perform a discard(), this routine will silently
2296  * do nothing.  The state of the specified speculation is transitioned
2297  * according to the state transition diagram outlined in <sys/dtrace_impl.h>
2298  */
2299 static void
2300 dtrace_speculation_discard(dtrace_state_t *state, processorid_t cpu,
2301     dtrace_specid_t which)
2302 {
2303 	dtrace_speculation_t *spec;
2304 	dtrace_speculation_state_t current, new;
2305 	dtrace_buffer_t *buf;
2306 
2307 	if (which == 0)
2308 		return;
2309 
2310 	if (which > state->dts_nspeculations) {
2311 		cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2312 		return;
2313 	}
2314 
2315 	spec = &state->dts_speculations[which - 1];
2316 	buf = &spec->dtsp_buffer[cpu];
2317 
2318 	do {
2319 		current = spec->dtsp_state;
2320 
2321 		switch (current) {
2322 		case DTRACESPEC_INACTIVE:
2323 		case DTRACESPEC_COMMITTINGMANY:
2324 		case DTRACESPEC_COMMITTING:
2325 		case DTRACESPEC_DISCARDING:
2326 			return;
2327 
2328 		case DTRACESPEC_ACTIVE:
2329 		case DTRACESPEC_ACTIVEMANY:
2330 			new = DTRACESPEC_DISCARDING;
2331 			break;
2332 
2333 		case DTRACESPEC_ACTIVEONE:
2334 			if (buf->dtb_offset != 0) {
2335 				new = DTRACESPEC_INACTIVE;
2336 			} else {
2337 				new = DTRACESPEC_DISCARDING;
2338 			}
2339 			break;
2340 
2341 		default:
2342 			ASSERT(0);
2343 		}
2344 	} while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2345 	    current, new) != current);
2346 
2347 	buf->dtb_offset = 0;
2348 	buf->dtb_drops = 0;
2349 }
2350 
2351 /*
2352  * Note:  not called from probe context.  This function is called
2353  * asynchronously from cross call context to clean any speculations that are
2354  * in the COMMITTINGMANY or DISCARDING states.  These speculations may not be
2355  * transitioned back to the INACTIVE state until all CPUs have cleaned the
2356  * speculation.
2357  */
2358 static void
2359 dtrace_speculation_clean_here(dtrace_state_t *state)
2360 {
2361 	dtrace_icookie_t cookie;
2362 	processorid_t cpu = CPU->cpu_id;
2363 	dtrace_buffer_t *dest = &state->dts_buffer[cpu];
2364 	dtrace_specid_t i;
2365 
2366 	cookie = dtrace_interrupt_disable();
2367 
2368 	if (dest->dtb_tomax == NULL) {
2369 		dtrace_interrupt_enable(cookie);
2370 		return;
2371 	}
2372 
2373 	for (i = 0; i < state->dts_nspeculations; i++) {
2374 		dtrace_speculation_t *spec = &state->dts_speculations[i];
2375 		dtrace_buffer_t *src = &spec->dtsp_buffer[cpu];
2376 
2377 		if (src->dtb_tomax == NULL)
2378 			continue;
2379 
2380 		if (spec->dtsp_state == DTRACESPEC_DISCARDING) {
2381 			src->dtb_offset = 0;
2382 			continue;
2383 		}
2384 
2385 		if (spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
2386 			continue;
2387 
2388 		if (src->dtb_offset == 0)
2389 			continue;
2390 
2391 		dtrace_speculation_commit(state, cpu, i + 1);
2392 	}
2393 
2394 	dtrace_interrupt_enable(cookie);
2395 }
2396 
2397 /*
2398  * Note:  not called from probe context.  This function is called
2399  * asynchronously (and at a regular interval) to clean any speculations that
2400  * are in the COMMITTINGMANY or DISCARDING states.  If it discovers that there
2401  * is work to be done, it cross calls all CPUs to perform that work;
2402  * COMMITMANY and DISCARDING speculations may not be transitioned back to the
2403  * INACTIVE state until they have been cleaned by all CPUs.
2404  */
2405 static void
2406 dtrace_speculation_clean(dtrace_state_t *state)
2407 {
2408 	int work = 0, rv;
2409 	dtrace_specid_t i;
2410 
2411 	for (i = 0; i < state->dts_nspeculations; i++) {
2412 		dtrace_speculation_t *spec = &state->dts_speculations[i];
2413 
2414 		ASSERT(!spec->dtsp_cleaning);
2415 
2416 		if (spec->dtsp_state != DTRACESPEC_DISCARDING &&
2417 		    spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
2418 			continue;
2419 
2420 		work++;
2421 		spec->dtsp_cleaning = 1;
2422 	}
2423 
2424 	if (!work)
2425 		return;
2426 
2427 	dtrace_xcall(DTRACE_CPUALL,
2428 	    (dtrace_xcall_t)dtrace_speculation_clean_here, state);
2429 
2430 	/*
2431 	 * We now know that all CPUs have committed or discarded their
2432 	 * speculation buffers, as appropriate.  We can now set the state
2433 	 * to inactive.
2434 	 */
2435 	for (i = 0; i < state->dts_nspeculations; i++) {
2436 		dtrace_speculation_t *spec = &state->dts_speculations[i];
2437 		dtrace_speculation_state_t current, new;
2438 
2439 		if (!spec->dtsp_cleaning)
2440 			continue;
2441 
2442 		current = spec->dtsp_state;
2443 		ASSERT(current == DTRACESPEC_DISCARDING ||
2444 		    current == DTRACESPEC_COMMITTINGMANY);
2445 
2446 		new = DTRACESPEC_INACTIVE;
2447 
2448 		rv = dtrace_cas32((uint32_t *)&spec->dtsp_state, current, new);
2449 		ASSERT(rv == current);
2450 		spec->dtsp_cleaning = 0;
2451 	}
2452 }
2453 
2454 /*
2455  * Called as part of a speculate() to get the speculative buffer associated
2456  * with a given speculation.  Returns NULL if the specified speculation is not
2457  * in an ACTIVE state.  If the speculation is in the ACTIVEONE state -- and
2458  * the active CPU is not the specified CPU -- the speculation will be
2459  * atomically transitioned into the ACTIVEMANY state.
2460  */
2461 static dtrace_buffer_t *
2462 dtrace_speculation_buffer(dtrace_state_t *state, processorid_t cpuid,
2463     dtrace_specid_t which)
2464 {
2465 	dtrace_speculation_t *spec;
2466 	dtrace_speculation_state_t current, new;
2467 	dtrace_buffer_t *buf;
2468 
2469 	if (which == 0)
2470 		return (NULL);
2471 
2472 	if (which > state->dts_nspeculations) {
2473 		cpu_core[cpuid].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2474 		return (NULL);
2475 	}
2476 
2477 	spec = &state->dts_speculations[which - 1];
2478 	buf = &spec->dtsp_buffer[cpuid];
2479 
2480 	do {
2481 		current = spec->dtsp_state;
2482 
2483 		switch (current) {
2484 		case DTRACESPEC_INACTIVE:
2485 		case DTRACESPEC_COMMITTINGMANY:
2486 		case DTRACESPEC_DISCARDING:
2487 			return (NULL);
2488 
2489 		case DTRACESPEC_COMMITTING:
2490 			ASSERT(buf->dtb_offset == 0);
2491 			return (NULL);
2492 
2493 		case DTRACESPEC_ACTIVEONE:
2494 			/*
2495 			 * This speculation is currently active on one CPU.
2496 			 * Check the offset in the buffer; if it's non-zero,
2497 			 * that CPU must be us (and we leave the state alone).
2498 			 * If it's zero, assume that we're starting on a new
2499 			 * CPU -- and change the state to indicate that the
2500 			 * speculation is active on more than one CPU.
2501 			 */
2502 			if (buf->dtb_offset != 0)
2503 				return (buf);
2504 
2505 			new = DTRACESPEC_ACTIVEMANY;
2506 			break;
2507 
2508 		case DTRACESPEC_ACTIVEMANY:
2509 			return (buf);
2510 
2511 		case DTRACESPEC_ACTIVE:
2512 			new = DTRACESPEC_ACTIVEONE;
2513 			break;
2514 
2515 		default:
2516 			ASSERT(0);
2517 		}
2518 	} while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2519 	    current, new) != current);
2520 
2521 	ASSERT(new == DTRACESPEC_ACTIVEONE || new == DTRACESPEC_ACTIVEMANY);
2522 	return (buf);
2523 }
2524 
2525 /*
2526  * Return a string.  In the event that the user lacks the privilege to access
2527  * arbitrary kernel memory, we copy the string out to scratch memory so that we
2528  * don't fail access checking.
2529  *
2530  * dtrace_dif_variable() uses this routine as a helper for various
2531  * builtin values such as 'execname' and 'probefunc.'
2532  */
2533 uintptr_t
2534 dtrace_dif_varstr(uintptr_t addr, dtrace_state_t *state,
2535     dtrace_mstate_t *mstate)
2536 {
2537 	uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
2538 	uintptr_t ret;
2539 	size_t strsz;
2540 
2541 	/*
2542 	 * The easy case: this probe is allowed to read all of memory, so
2543 	 * we can just return this as a vanilla pointer.
2544 	 */
2545 	if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
2546 		return (addr);
2547 
2548 	/*
2549 	 * This is the tougher case: we copy the string in question from
2550 	 * kernel memory into scratch memory and return it that way: this
2551 	 * ensures that we won't trip up when access checking tests the
2552 	 * BYREF return value.
2553 	 */
2554 	strsz = dtrace_strlen((char *)addr, size) + 1;
2555 
2556 	if (mstate->dtms_scratch_ptr + strsz >
2557 	    mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
2558 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
2559 		return (NULL);
2560 	}
2561 
2562 	dtrace_strcpy((const void *)addr, (void *)mstate->dtms_scratch_ptr,
2563 	    strsz);
2564 	ret = mstate->dtms_scratch_ptr;
2565 	mstate->dtms_scratch_ptr += strsz;
2566 	return (ret);
2567 }
2568 
2569 /*
2570  * This function implements the DIF emulator's variable lookups.  The emulator
2571  * passes a reserved variable identifier and optional built-in array index.
2572  */
2573 static uint64_t
2574 dtrace_dif_variable(dtrace_mstate_t *mstate, dtrace_state_t *state, uint64_t v,
2575     uint64_t ndx)
2576 {
2577 	/*
2578 	 * If we're accessing one of the uncached arguments, we'll turn this
2579 	 * into a reference in the args array.
2580 	 */
2581 	if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9) {
2582 		ndx = v - DIF_VAR_ARG0;
2583 		v = DIF_VAR_ARGS;
2584 	}
2585 
2586 	switch (v) {
2587 	case DIF_VAR_ARGS:
2588 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_ARGS);
2589 		if (ndx >= sizeof (mstate->dtms_arg) /
2590 		    sizeof (mstate->dtms_arg[0])) {
2591 			int aframes = mstate->dtms_probe->dtpr_aframes + 2;
2592 			dtrace_provider_t *pv;
2593 			uint64_t val;
2594 
2595 			pv = mstate->dtms_probe->dtpr_provider;
2596 			if (pv->dtpv_pops.dtps_getargval != NULL)
2597 				val = pv->dtpv_pops.dtps_getargval(pv->dtpv_arg,
2598 				    mstate->dtms_probe->dtpr_id,
2599 				    mstate->dtms_probe->dtpr_arg, ndx, aframes);
2600 			else
2601 				val = dtrace_getarg(ndx, aframes);
2602 
2603 			/*
2604 			 * This is regrettably required to keep the compiler
2605 			 * from tail-optimizing the call to dtrace_getarg().
2606 			 * The condition always evaluates to true, but the
2607 			 * compiler has no way of figuring that out a priori.
2608 			 * (None of this would be necessary if the compiler
2609 			 * could be relied upon to _always_ tail-optimize
2610 			 * the call to dtrace_getarg() -- but it can't.)
2611 			 */
2612 			if (mstate->dtms_probe != NULL)
2613 				return (val);
2614 
2615 			ASSERT(0);
2616 		}
2617 
2618 		return (mstate->dtms_arg[ndx]);
2619 
2620 	case DIF_VAR_UREGS: {
2621 		klwp_t *lwp;
2622 
2623 		if (!dtrace_priv_proc(state))
2624 			return (0);
2625 
2626 		if ((lwp = curthread->t_lwp) == NULL) {
2627 			DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
2628 			cpu_core[CPU->cpu_id].cpuc_dtrace_illval = NULL;
2629 			return (0);
2630 		}
2631 
2632 		return (dtrace_getreg(lwp->lwp_regs, ndx));
2633 	}
2634 
2635 	case DIF_VAR_CURTHREAD:
2636 		if (!dtrace_priv_kernel(state))
2637 			return (0);
2638 		return ((uint64_t)(uintptr_t)curthread);
2639 
2640 	case DIF_VAR_TIMESTAMP:
2641 		if (!(mstate->dtms_present & DTRACE_MSTATE_TIMESTAMP)) {
2642 			mstate->dtms_timestamp = dtrace_gethrtime();
2643 			mstate->dtms_present |= DTRACE_MSTATE_TIMESTAMP;
2644 		}
2645 		return (mstate->dtms_timestamp);
2646 
2647 	case DIF_VAR_VTIMESTAMP:
2648 		ASSERT(dtrace_vtime_references != 0);
2649 		return (curthread->t_dtrace_vtime);
2650 
2651 	case DIF_VAR_WALLTIMESTAMP:
2652 		if (!(mstate->dtms_present & DTRACE_MSTATE_WALLTIMESTAMP)) {
2653 			mstate->dtms_walltimestamp = dtrace_gethrestime();
2654 			mstate->dtms_present |= DTRACE_MSTATE_WALLTIMESTAMP;
2655 		}
2656 		return (mstate->dtms_walltimestamp);
2657 
2658 	case DIF_VAR_IPL:
2659 		if (!dtrace_priv_kernel(state))
2660 			return (0);
2661 		if (!(mstate->dtms_present & DTRACE_MSTATE_IPL)) {
2662 			mstate->dtms_ipl = dtrace_getipl();
2663 			mstate->dtms_present |= DTRACE_MSTATE_IPL;
2664 		}
2665 		return (mstate->dtms_ipl);
2666 
2667 	case DIF_VAR_EPID:
2668 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_EPID);
2669 		return (mstate->dtms_epid);
2670 
2671 	case DIF_VAR_ID:
2672 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
2673 		return (mstate->dtms_probe->dtpr_id);
2674 
2675 	case DIF_VAR_STACKDEPTH:
2676 		if (!dtrace_priv_kernel(state))
2677 			return (0);
2678 		if (!(mstate->dtms_present & DTRACE_MSTATE_STACKDEPTH)) {
2679 			int aframes = mstate->dtms_probe->dtpr_aframes + 2;
2680 
2681 			mstate->dtms_stackdepth = dtrace_getstackdepth(aframes);
2682 			mstate->dtms_present |= DTRACE_MSTATE_STACKDEPTH;
2683 		}
2684 		return (mstate->dtms_stackdepth);
2685 
2686 	case DIF_VAR_USTACKDEPTH:
2687 		if (!dtrace_priv_proc(state))
2688 			return (0);
2689 		if (!(mstate->dtms_present & DTRACE_MSTATE_USTACKDEPTH)) {
2690 			/*
2691 			 * See comment in DIF_VAR_PID.
2692 			 */
2693 			if (DTRACE_ANCHORED(mstate->dtms_probe) &&
2694 			    CPU_ON_INTR(CPU)) {
2695 				mstate->dtms_ustackdepth = 0;
2696 			} else {
2697 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
2698 				mstate->dtms_ustackdepth =
2699 				    dtrace_getustackdepth();
2700 				DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
2701 			}
2702 			mstate->dtms_present |= DTRACE_MSTATE_USTACKDEPTH;
2703 		}
2704 		return (mstate->dtms_ustackdepth);
2705 
2706 	case DIF_VAR_CALLER:
2707 		if (!dtrace_priv_kernel(state))
2708 			return (0);
2709 		if (!(mstate->dtms_present & DTRACE_MSTATE_CALLER)) {
2710 			int aframes = mstate->dtms_probe->dtpr_aframes + 2;
2711 
2712 			if (!DTRACE_ANCHORED(mstate->dtms_probe)) {
2713 				/*
2714 				 * If this is an unanchored probe, we are
2715 				 * required to go through the slow path:
2716 				 * dtrace_caller() only guarantees correct
2717 				 * results for anchored probes.
2718 				 */
2719 				pc_t caller[2];
2720 
2721 				dtrace_getpcstack(caller, 2, aframes,
2722 				    (uint32_t *)(uintptr_t)mstate->dtms_arg[0]);
2723 				mstate->dtms_caller = caller[1];
2724 			} else if ((mstate->dtms_caller =
2725 			    dtrace_caller(aframes)) == -1) {
2726 				/*
2727 				 * We have failed to do this the quick way;
2728 				 * we must resort to the slower approach of
2729 				 * calling dtrace_getpcstack().
2730 				 */
2731 				pc_t caller;
2732 
2733 				dtrace_getpcstack(&caller, 1, aframes, NULL);
2734 				mstate->dtms_caller = caller;
2735 			}
2736 
2737 			mstate->dtms_present |= DTRACE_MSTATE_CALLER;
2738 		}
2739 		return (mstate->dtms_caller);
2740 
2741 	case DIF_VAR_UCALLER:
2742 		if (!dtrace_priv_proc(state))
2743 			return (0);
2744 
2745 		if (!(mstate->dtms_present & DTRACE_MSTATE_UCALLER)) {
2746 			uint64_t ustack[3];
2747 
2748 			/*
2749 			 * dtrace_getupcstack() fills in the first uint64_t
2750 			 * with the current PID.  The second uint64_t will
2751 			 * be the program counter at user-level.  The third
2752 			 * uint64_t will contain the caller, which is what
2753 			 * we're after.
2754 			 */
2755 			ustack[2] = NULL;
2756 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
2757 			dtrace_getupcstack(ustack, 3);
2758 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
2759 			mstate->dtms_ucaller = ustack[2];
2760 			mstate->dtms_present |= DTRACE_MSTATE_UCALLER;
2761 		}
2762 
2763 		return (mstate->dtms_ucaller);
2764 
2765 	case DIF_VAR_PROBEPROV:
2766 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
2767 		return (dtrace_dif_varstr(
2768 		    (uintptr_t)mstate->dtms_probe->dtpr_provider->dtpv_name,
2769 		    state, mstate));
2770 
2771 	case DIF_VAR_PROBEMOD:
2772 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
2773 		return (dtrace_dif_varstr(
2774 		    (uintptr_t)mstate->dtms_probe->dtpr_mod,
2775 		    state, mstate));
2776 
2777 	case DIF_VAR_PROBEFUNC:
2778 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
2779 		return (dtrace_dif_varstr(
2780 		    (uintptr_t)mstate->dtms_probe->dtpr_func,
2781 		    state, mstate));
2782 
2783 	case DIF_VAR_PROBENAME:
2784 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
2785 		return (dtrace_dif_varstr(
2786 		    (uintptr_t)mstate->dtms_probe->dtpr_name,
2787 		    state, mstate));
2788 
2789 	case DIF_VAR_PID:
2790 		if (!dtrace_priv_proc(state))
2791 			return (0);
2792 
2793 		/*
2794 		 * Note that we are assuming that an unanchored probe is
2795 		 * always due to a high-level interrupt.  (And we're assuming
2796 		 * that there is only a single high level interrupt.)
2797 		 */
2798 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
2799 			return (pid0.pid_id);
2800 
2801 		/*
2802 		 * It is always safe to dereference one's own t_procp pointer:
2803 		 * it always points to a valid, allocated proc structure.
2804 		 * Further, it is always safe to dereference the p_pidp member
2805 		 * of one's own proc structure.  (These are truisms becuase
2806 		 * threads and processes don't clean up their own state --
2807 		 * they leave that task to whomever reaps them.)
2808 		 */
2809 		return ((uint64_t)curthread->t_procp->p_pidp->pid_id);
2810 
2811 	case DIF_VAR_PPID:
2812 		if (!dtrace_priv_proc(state))
2813 			return (0);
2814 
2815 		/*
2816 		 * See comment in DIF_VAR_PID.
2817 		 */
2818 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
2819 			return (pid0.pid_id);
2820 
2821 		/*
2822 		 * It is always safe to dereference one's own t_procp pointer:
2823 		 * it always points to a valid, allocated proc structure.
2824 		 * (This is true because threads don't clean up their own
2825 		 * state -- they leave that task to whomever reaps them.)
2826 		 */
2827 		return ((uint64_t)curthread->t_procp->p_ppid);
2828 
2829 	case DIF_VAR_TID:
2830 		/*
2831 		 * See comment in DIF_VAR_PID.
2832 		 */
2833 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
2834 			return (0);
2835 
2836 		return ((uint64_t)curthread->t_tid);
2837 
2838 	case DIF_VAR_EXECNAME:
2839 		if (!dtrace_priv_proc(state))
2840 			return (0);
2841 
2842 		/*
2843 		 * See comment in DIF_VAR_PID.
2844 		 */
2845 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
2846 			return ((uint64_t)(uintptr_t)p0.p_user.u_comm);
2847 
2848 		/*
2849 		 * It is always safe to dereference one's own t_procp pointer:
2850 		 * it always points to a valid, allocated proc structure.
2851 		 * (This is true because threads don't clean up their own
2852 		 * state -- they leave that task to whomever reaps them.)
2853 		 */
2854 		return (dtrace_dif_varstr(
2855 		    (uintptr_t)curthread->t_procp->p_user.u_comm,
2856 		    state, mstate));
2857 
2858 	case DIF_VAR_ZONENAME:
2859 		if (!dtrace_priv_proc(state))
2860 			return (0);
2861 
2862 		/*
2863 		 * See comment in DIF_VAR_PID.
2864 		 */
2865 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
2866 			return ((uint64_t)(uintptr_t)p0.p_zone->zone_name);
2867 
2868 		/*
2869 		 * It is always safe to dereference one's own t_procp pointer:
2870 		 * it always points to a valid, allocated proc structure.
2871 		 * (This is true because threads don't clean up their own
2872 		 * state -- they leave that task to whomever reaps them.)
2873 		 */
2874 		return (dtrace_dif_varstr(
2875 		    (uintptr_t)curthread->t_procp->p_zone->zone_name,
2876 		    state, mstate));
2877 
2878 	case DIF_VAR_UID:
2879 		if (!dtrace_priv_proc(state))
2880 			return (0);
2881 
2882 		/*
2883 		 * See comment in DIF_VAR_PID.
2884 		 */
2885 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
2886 			return ((uint64_t)p0.p_cred->cr_uid);
2887 
2888 		/*
2889 		 * It is always safe to dereference one's own t_procp pointer:
2890 		 * it always points to a valid, allocated proc structure.
2891 		 * (This is true because threads don't clean up their own
2892 		 * state -- they leave that task to whomever reaps them.)
2893 		 *
2894 		 * Additionally, it is safe to dereference one's own process
2895 		 * credential, since this is never NULL after process birth.
2896 		 */
2897 		return ((uint64_t)curthread->t_procp->p_cred->cr_uid);
2898 
2899 	case DIF_VAR_GID:
2900 		if (!dtrace_priv_proc(state))
2901 			return (0);
2902 
2903 		/*
2904 		 * See comment in DIF_VAR_PID.
2905 		 */
2906 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
2907 			return ((uint64_t)p0.p_cred->cr_gid);
2908 
2909 		/*
2910 		 * It is always safe to dereference one's own t_procp pointer:
2911 		 * it always points to a valid, allocated proc structure.
2912 		 * (This is true because threads don't clean up their own
2913 		 * state -- they leave that task to whomever reaps them.)
2914 		 *
2915 		 * Additionally, it is safe to dereference one's own process
2916 		 * credential, since this is never NULL after process birth.
2917 		 */
2918 		return ((uint64_t)curthread->t_procp->p_cred->cr_gid);
2919 
2920 	case DIF_VAR_ERRNO: {
2921 		klwp_t *lwp;
2922 		if (!dtrace_priv_proc(state))
2923 			return (0);
2924 
2925 		/*
2926 		 * See comment in DIF_VAR_PID.
2927 		 */
2928 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
2929 			return (0);
2930 
2931 		/*
2932 		 * It is always safe to dereference one's own t_lwp pointer in
2933 		 * the event that this pointer is non-NULL.  (This is true
2934 		 * because threads and lwps don't clean up their own state --
2935 		 * they leave that task to whomever reaps them.)
2936 		 */
2937 		if ((lwp = curthread->t_lwp) == NULL)
2938 			return (0);
2939 
2940 		return ((uint64_t)lwp->lwp_errno);
2941 	}
2942 	default:
2943 		DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
2944 		return (0);
2945 	}
2946 }
2947 
2948 /*
2949  * Emulate the execution of DTrace ID subroutines invoked by the call opcode.
2950  * Notice that we don't bother validating the proper number of arguments or
2951  * their types in the tuple stack.  This isn't needed because all argument
2952  * interpretation is safe because of our load safety -- the worst that can
2953  * happen is that a bogus program can obtain bogus results.
2954  */
2955 static void
2956 dtrace_dif_subr(uint_t subr, uint_t rd, uint64_t *regs,
2957     dtrace_key_t *tupregs, int nargs,
2958     dtrace_mstate_t *mstate, dtrace_state_t *state)
2959 {
2960 	volatile uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
2961 	volatile uintptr_t *illval = &cpu_core[CPU->cpu_id].cpuc_dtrace_illval;
2962 	dtrace_vstate_t *vstate = &state->dts_vstate;
2963 
2964 	union {
2965 		mutex_impl_t mi;
2966 		uint64_t mx;
2967 	} m;
2968 
2969 	union {
2970 		krwlock_t ri;
2971 		uintptr_t rw;
2972 	} r;
2973 
2974 	switch (subr) {
2975 	case DIF_SUBR_RAND:
2976 		regs[rd] = (dtrace_gethrtime() * 2416 + 374441) % 1771875;
2977 		break;
2978 
2979 	case DIF_SUBR_MUTEX_OWNED:
2980 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
2981 		    mstate, vstate)) {
2982 			regs[rd] = NULL;
2983 			break;
2984 		}
2985 
2986 		m.mx = dtrace_load64(tupregs[0].dttk_value);
2987 		if (MUTEX_TYPE_ADAPTIVE(&m.mi))
2988 			regs[rd] = MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER;
2989 		else
2990 			regs[rd] = LOCK_HELD(&m.mi.m_spin.m_spinlock);
2991 		break;
2992 
2993 	case DIF_SUBR_MUTEX_OWNER:
2994 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
2995 		    mstate, vstate)) {
2996 			regs[rd] = NULL;
2997 			break;
2998 		}
2999 
3000 		m.mx = dtrace_load64(tupregs[0].dttk_value);
3001 		if (MUTEX_TYPE_ADAPTIVE(&m.mi) &&
3002 		    MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER)
3003 			regs[rd] = (uintptr_t)MUTEX_OWNER(&m.mi);
3004 		else
3005 			regs[rd] = 0;
3006 		break;
3007 
3008 	case DIF_SUBR_MUTEX_TYPE_ADAPTIVE:
3009 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3010 		    mstate, vstate)) {
3011 			regs[rd] = NULL;
3012 			break;
3013 		}
3014 
3015 		m.mx = dtrace_load64(tupregs[0].dttk_value);
3016 		regs[rd] = MUTEX_TYPE_ADAPTIVE(&m.mi);
3017 		break;
3018 
3019 	case DIF_SUBR_MUTEX_TYPE_SPIN:
3020 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3021 		    mstate, vstate)) {
3022 			regs[rd] = NULL;
3023 			break;
3024 		}
3025 
3026 		m.mx = dtrace_load64(tupregs[0].dttk_value);
3027 		regs[rd] = MUTEX_TYPE_SPIN(&m.mi);
3028 		break;
3029 
3030 	case DIF_SUBR_RW_READ_HELD: {
3031 		uintptr_t tmp;
3032 
3033 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
3034 		    mstate, vstate)) {
3035 			regs[rd] = NULL;
3036 			break;
3037 		}
3038 
3039 		r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3040 		regs[rd] = _RW_READ_HELD(&r.ri, tmp);
3041 		break;
3042 	}
3043 
3044 	case DIF_SUBR_RW_WRITE_HELD:
3045 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
3046 		    mstate, vstate)) {
3047 			regs[rd] = NULL;
3048 			break;
3049 		}
3050 
3051 		r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3052 		regs[rd] = _RW_WRITE_HELD(&r.ri);
3053 		break;
3054 
3055 	case DIF_SUBR_RW_ISWRITER:
3056 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
3057 		    mstate, vstate)) {
3058 			regs[rd] = NULL;
3059 			break;
3060 		}
3061 
3062 		r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3063 		regs[rd] = _RW_ISWRITER(&r.ri);
3064 		break;
3065 
3066 	case DIF_SUBR_BCOPY: {
3067 		/*
3068 		 * We need to be sure that the destination is in the scratch
3069 		 * region -- no other region is allowed.
3070 		 */
3071 		uintptr_t src = tupregs[0].dttk_value;
3072 		uintptr_t dest = tupregs[1].dttk_value;
3073 		size_t size = tupregs[2].dttk_value;
3074 
3075 		if (!dtrace_inscratch(dest, size, mstate)) {
3076 			*flags |= CPU_DTRACE_BADADDR;
3077 			*illval = regs[rd];
3078 			break;
3079 		}
3080 
3081 		if (!dtrace_canload(src, size, mstate, vstate)) {
3082 			regs[rd] = NULL;
3083 			break;
3084 		}
3085 
3086 		dtrace_bcopy((void *)src, (void *)dest, size);
3087 		break;
3088 	}
3089 
3090 	case DIF_SUBR_ALLOCA:
3091 	case DIF_SUBR_COPYIN: {
3092 		uintptr_t dest = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
3093 		uint64_t size =
3094 		    tupregs[subr == DIF_SUBR_ALLOCA ? 0 : 1].dttk_value;
3095 		size_t scratch_size = (dest - mstate->dtms_scratch_ptr) + size;
3096 
3097 		/*
3098 		 * This action doesn't require any credential checks since
3099 		 * probes will not activate in user contexts to which the
3100 		 * enabling user does not have permissions.
3101 		 */
3102 
3103 		/*
3104 		 * Rounding up the user allocation size could have overflowed
3105 		 * a large, bogus allocation (like -1ULL) to 0.
3106 		 */
3107 		if (scratch_size < size ||
3108 		    !DTRACE_INSCRATCH(mstate, scratch_size)) {
3109 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3110 			regs[rd] = NULL;
3111 			break;
3112 		}
3113 
3114 		if (subr == DIF_SUBR_COPYIN) {
3115 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3116 			dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
3117 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3118 		}
3119 
3120 		mstate->dtms_scratch_ptr += scratch_size;
3121 		regs[rd] = dest;
3122 		break;
3123 	}
3124 
3125 	case DIF_SUBR_COPYINTO: {
3126 		uint64_t size = tupregs[1].dttk_value;
3127 		uintptr_t dest = tupregs[2].dttk_value;
3128 
3129 		/*
3130 		 * This action doesn't require any credential checks since
3131 		 * probes will not activate in user contexts to which the
3132 		 * enabling user does not have permissions.
3133 		 */
3134 		if (!dtrace_inscratch(dest, size, mstate)) {
3135 			*flags |= CPU_DTRACE_BADADDR;
3136 			*illval = regs[rd];
3137 			break;
3138 		}
3139 
3140 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3141 		dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
3142 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3143 		break;
3144 	}
3145 
3146 	case DIF_SUBR_COPYINSTR: {
3147 		uintptr_t dest = mstate->dtms_scratch_ptr;
3148 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3149 
3150 		if (nargs > 1 && tupregs[1].dttk_value < size)
3151 			size = tupregs[1].dttk_value + 1;
3152 
3153 		/*
3154 		 * This action doesn't require any credential checks since
3155 		 * probes will not activate in user contexts to which the
3156 		 * enabling user does not have permissions.
3157 		 */
3158 		if (!DTRACE_INSCRATCH(mstate, size)) {
3159 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3160 			regs[rd] = NULL;
3161 			break;
3162 		}
3163 
3164 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3165 		dtrace_copyinstr(tupregs[0].dttk_value, dest, size, flags);
3166 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3167 
3168 		((char *)dest)[size - 1] = '\0';
3169 		mstate->dtms_scratch_ptr += size;
3170 		regs[rd] = dest;
3171 		break;
3172 	}
3173 
3174 	case DIF_SUBR_MSGSIZE:
3175 	case DIF_SUBR_MSGDSIZE: {
3176 		uintptr_t baddr = tupregs[0].dttk_value, daddr;
3177 		uintptr_t wptr, rptr;
3178 		size_t count = 0;
3179 		int cont = 0;
3180 
3181 		while (baddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
3182 
3183 			if (!dtrace_canload(baddr, sizeof (mblk_t), mstate,
3184 			    vstate)) {
3185 				regs[rd] = NULL;
3186 				break;
3187 			}
3188 
3189 			wptr = dtrace_loadptr(baddr +
3190 			    offsetof(mblk_t, b_wptr));
3191 
3192 			rptr = dtrace_loadptr(baddr +
3193 			    offsetof(mblk_t, b_rptr));
3194 
3195 			if (wptr < rptr) {
3196 				*flags |= CPU_DTRACE_BADADDR;
3197 				*illval = tupregs[0].dttk_value;
3198 				break;
3199 			}
3200 
3201 			daddr = dtrace_loadptr(baddr +
3202 			    offsetof(mblk_t, b_datap));
3203 
3204 			baddr = dtrace_loadptr(baddr +
3205 			    offsetof(mblk_t, b_cont));
3206 
3207 			/*
3208 			 * We want to prevent against denial-of-service here,
3209 			 * so we're only going to search the list for
3210 			 * dtrace_msgdsize_max mblks.
3211 			 */
3212 			if (cont++ > dtrace_msgdsize_max) {
3213 				*flags |= CPU_DTRACE_ILLOP;
3214 				break;
3215 			}
3216 
3217 			if (subr == DIF_SUBR_MSGDSIZE) {
3218 				if (dtrace_load8(daddr +
3219 				    offsetof(dblk_t, db_type)) != M_DATA)
3220 					continue;
3221 			}
3222 
3223 			count += wptr - rptr;
3224 		}
3225 
3226 		if (!(*flags & CPU_DTRACE_FAULT))
3227 			regs[rd] = count;
3228 
3229 		break;
3230 	}
3231 
3232 	case DIF_SUBR_PROGENYOF: {
3233 		pid_t pid = tupregs[0].dttk_value;
3234 		proc_t *p;
3235 		int rval = 0;
3236 
3237 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3238 
3239 		for (p = curthread->t_procp; p != NULL; p = p->p_parent) {
3240 			if (p->p_pidp->pid_id == pid) {
3241 				rval = 1;
3242 				break;
3243 			}
3244 		}
3245 
3246 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3247 
3248 		regs[rd] = rval;
3249 		break;
3250 	}
3251 
3252 	case DIF_SUBR_SPECULATION:
3253 		regs[rd] = dtrace_speculation(state);
3254 		break;
3255 
3256 	case DIF_SUBR_COPYOUT: {
3257 		uintptr_t kaddr = tupregs[0].dttk_value;
3258 		uintptr_t uaddr = tupregs[1].dttk_value;
3259 		uint64_t size = tupregs[2].dttk_value;
3260 
3261 		if (!dtrace_destructive_disallow &&
3262 		    dtrace_priv_proc_control(state) &&
3263 		    !dtrace_istoxic(kaddr, size)) {
3264 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3265 			dtrace_copyout(kaddr, uaddr, size, flags);
3266 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3267 		}
3268 		break;
3269 	}
3270 
3271 	case DIF_SUBR_COPYOUTSTR: {
3272 		uintptr_t kaddr = tupregs[0].dttk_value;
3273 		uintptr_t uaddr = tupregs[1].dttk_value;
3274 		uint64_t size = tupregs[2].dttk_value;
3275 
3276 		if (!dtrace_destructive_disallow &&
3277 		    dtrace_priv_proc_control(state) &&
3278 		    !dtrace_istoxic(kaddr, size)) {
3279 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3280 			dtrace_copyoutstr(kaddr, uaddr, size, flags);
3281 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3282 		}
3283 		break;
3284 	}
3285 
3286 	case DIF_SUBR_STRLEN: {
3287 		size_t sz;
3288 		uintptr_t addr = (uintptr_t)tupregs[0].dttk_value;
3289 		sz = dtrace_strlen((char *)addr,
3290 		    state->dts_options[DTRACEOPT_STRSIZE]);
3291 
3292 		if (!dtrace_canload(addr, sz + 1, mstate, vstate)) {
3293 			regs[rd] = NULL;
3294 			break;
3295 		}
3296 
3297 		regs[rd] = sz;
3298 
3299 		break;
3300 	}
3301 
3302 	case DIF_SUBR_STRCHR:
3303 	case DIF_SUBR_STRRCHR: {
3304 		/*
3305 		 * We're going to iterate over the string looking for the
3306 		 * specified character.  We will iterate until we have reached
3307 		 * the string length or we have found the character.  If this
3308 		 * is DIF_SUBR_STRRCHR, we will look for the last occurrence
3309 		 * of the specified character instead of the first.
3310 		 */
3311 		uintptr_t saddr = tupregs[0].dttk_value;
3312 		uintptr_t addr = tupregs[0].dttk_value;
3313 		uintptr_t limit = addr + state->dts_options[DTRACEOPT_STRSIZE];
3314 		char c, target = (char)tupregs[1].dttk_value;
3315 
3316 		for (regs[rd] = NULL; addr < limit; addr++) {
3317 			if ((c = dtrace_load8(addr)) == target) {
3318 				regs[rd] = addr;
3319 
3320 				if (subr == DIF_SUBR_STRCHR)
3321 					break;
3322 			}
3323 
3324 			if (c == '\0')
3325 				break;
3326 		}
3327 
3328 		if (!dtrace_canload(saddr, addr - saddr, mstate, vstate)) {
3329 			regs[rd] = NULL;
3330 			break;
3331 		}
3332 
3333 		break;
3334 	}
3335 
3336 	case DIF_SUBR_STRSTR:
3337 	case DIF_SUBR_INDEX:
3338 	case DIF_SUBR_RINDEX: {
3339 		/*
3340 		 * We're going to iterate over the string looking for the
3341 		 * specified string.  We will iterate until we have reached
3342 		 * the string length or we have found the string.  (Yes, this
3343 		 * is done in the most naive way possible -- but considering
3344 		 * that the string we're searching for is likely to be
3345 		 * relatively short, the complexity of Rabin-Karp or similar
3346 		 * hardly seems merited.)
3347 		 */
3348 		char *addr = (char *)(uintptr_t)tupregs[0].dttk_value;
3349 		char *substr = (char *)(uintptr_t)tupregs[1].dttk_value;
3350 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3351 		size_t len = dtrace_strlen(addr, size);
3352 		size_t sublen = dtrace_strlen(substr, size);
3353 		char *limit = addr + len, *orig = addr;
3354 		int notfound = subr == DIF_SUBR_STRSTR ? 0 : -1;
3355 		int inc = 1;
3356 
3357 		regs[rd] = notfound;
3358 
3359 		if (!dtrace_canload((uintptr_t)addr, len + 1, mstate, vstate)) {
3360 			regs[rd] = NULL;
3361 			break;
3362 		}
3363 
3364 		if (!dtrace_canload((uintptr_t)substr, sublen + 1, mstate,
3365 		    vstate)) {
3366 			regs[rd] = NULL;
3367 			break;
3368 		}
3369 
3370 		/*
3371 		 * strstr() and index()/rindex() have similar semantics if
3372 		 * both strings are the empty string: strstr() returns a
3373 		 * pointer to the (empty) string, and index() and rindex()
3374 		 * both return index 0 (regardless of any position argument).
3375 		 */
3376 		if (sublen == 0 && len == 0) {
3377 			if (subr == DIF_SUBR_STRSTR)
3378 				regs[rd] = (uintptr_t)addr;
3379 			else
3380 				regs[rd] = 0;
3381 			break;
3382 		}
3383 
3384 		if (subr != DIF_SUBR_STRSTR) {
3385 			if (subr == DIF_SUBR_RINDEX) {
3386 				limit = orig - 1;
3387 				addr += len;
3388 				inc = -1;
3389 			}
3390 
3391 			/*
3392 			 * Both index() and rindex() take an optional position
3393 			 * argument that denotes the starting position.
3394 			 */
3395 			if (nargs == 3) {
3396 				int64_t pos = (int64_t)tupregs[2].dttk_value;
3397 
3398 				/*
3399 				 * If the position argument to index() is
3400 				 * negative, Perl implicitly clamps it at
3401 				 * zero.  This semantic is a little surprising
3402 				 * given the special meaning of negative
3403 				 * positions to similar Perl functions like
3404 				 * substr(), but it appears to reflect a
3405 				 * notion that index() can start from a
3406 				 * negative index and increment its way up to
3407 				 * the string.  Given this notion, Perl's
3408 				 * rindex() is at least self-consistent in
3409 				 * that it implicitly clamps positions greater
3410 				 * than the string length to be the string
3411 				 * length.  Where Perl completely loses
3412 				 * coherence, however, is when the specified
3413 				 * substring is the empty string ("").  In
3414 				 * this case, even if the position is
3415 				 * negative, rindex() returns 0 -- and even if
3416 				 * the position is greater than the length,
3417 				 * index() returns the string length.  These
3418 				 * semantics violate the notion that index()
3419 				 * should never return a value less than the
3420 				 * specified position and that rindex() should
3421 				 * never return a value greater than the
3422 				 * specified position.  (One assumes that
3423 				 * these semantics are artifacts of Perl's
3424 				 * implementation and not the results of
3425 				 * deliberate design -- it beggars belief that
3426 				 * even Larry Wall could desire such oddness.)
3427 				 * While in the abstract one would wish for
3428 				 * consistent position semantics across
3429 				 * substr(), index() and rindex() -- or at the
3430 				 * very least self-consistent position
3431 				 * semantics for index() and rindex() -- we
3432 				 * instead opt to keep with the extant Perl
3433 				 * semantics, in all their broken glory.  (Do
3434 				 * we have more desire to maintain Perl's
3435 				 * semantics than Perl does?  Probably.)
3436 				 */
3437 				if (subr == DIF_SUBR_RINDEX) {
3438 					if (pos < 0) {
3439 						if (sublen == 0)
3440 							regs[rd] = 0;
3441 						break;
3442 					}
3443 
3444 					if (pos > len)
3445 						pos = len;
3446 				} else {
3447 					if (pos < 0)
3448 						pos = 0;
3449 
3450 					if (pos >= len) {
3451 						if (sublen == 0)
3452 							regs[rd] = len;
3453 						break;
3454 					}
3455 				}
3456 
3457 				addr = orig + pos;
3458 			}
3459 		}
3460 
3461 		for (regs[rd] = notfound; addr != limit; addr += inc) {
3462 			if (dtrace_strncmp(addr, substr, sublen) == 0) {
3463 				if (subr != DIF_SUBR_STRSTR) {
3464 					/*
3465 					 * As D index() and rindex() are
3466 					 * modeled on Perl (and not on awk),
3467 					 * we return a zero-based (and not a
3468 					 * one-based) index.  (For you Perl
3469 					 * weenies: no, we're not going to add
3470 					 * $[ -- and shouldn't you be at a con
3471 					 * or something?)
3472 					 */
3473 					regs[rd] = (uintptr_t)(addr - orig);
3474 					break;
3475 				}
3476 
3477 				ASSERT(subr == DIF_SUBR_STRSTR);
3478 				regs[rd] = (uintptr_t)addr;
3479 				break;
3480 			}
3481 		}
3482 
3483 		break;
3484 	}
3485 
3486 	case DIF_SUBR_STRTOK: {
3487 		uintptr_t addr = tupregs[0].dttk_value;
3488 		uintptr_t tokaddr = tupregs[1].dttk_value;
3489 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3490 		uintptr_t limit, toklimit = tokaddr + size;
3491 		uint8_t c, tokmap[32];	 /* 256 / 8 */
3492 		char *dest = (char *)mstate->dtms_scratch_ptr;
3493 		int i;
3494 
3495 		/*
3496 		 * Check both the token buffer and (later) the input buffer,
3497 		 * since both could be non-scratch addresses.
3498 		 */
3499 		if (!dtrace_strcanload(tokaddr, size, mstate, vstate)) {
3500 			regs[rd] = NULL;
3501 			break;
3502 		}
3503 
3504 		if (!DTRACE_INSCRATCH(mstate, size)) {
3505 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3506 			regs[rd] = NULL;
3507 			break;
3508 		}
3509 
3510 		if (addr == NULL) {
3511 			/*
3512 			 * If the address specified is NULL, we use our saved
3513 			 * strtok pointer from the mstate.  Note that this
3514 			 * means that the saved strtok pointer is _only_
3515 			 * valid within multiple enablings of the same probe --
3516 			 * it behaves like an implicit clause-local variable.
3517 			 */
3518 			addr = mstate->dtms_strtok;
3519 		} else {
3520 			/*
3521 			 * If the user-specified address is non-NULL we must
3522 			 * access check it.  This is the only time we have
3523 			 * a chance to do so, since this address may reside
3524 			 * in the string table of this clause-- future calls
3525 			 * (when we fetch addr from mstate->dtms_strtok)
3526 			 * would fail this access check.
3527 			 */
3528 			if (!dtrace_strcanload(addr, size, mstate, vstate)) {
3529 				regs[rd] = NULL;
3530 				break;
3531 			}
3532 		}
3533 
3534 		/*
3535 		 * First, zero the token map, and then process the token
3536 		 * string -- setting a bit in the map for every character
3537 		 * found in the token string.
3538 		 */
3539 		for (i = 0; i < sizeof (tokmap); i++)
3540 			tokmap[i] = 0;
3541 
3542 		for (; tokaddr < toklimit; tokaddr++) {
3543 			if ((c = dtrace_load8(tokaddr)) == '\0')
3544 				break;
3545 
3546 			ASSERT((c >> 3) < sizeof (tokmap));
3547 			tokmap[c >> 3] |= (1 << (c & 0x7));
3548 		}
3549 
3550 		for (limit = addr + size; addr < limit; addr++) {
3551 			/*
3552 			 * We're looking for a character that is _not_ contained
3553 			 * in the token string.
3554 			 */
3555 			if ((c = dtrace_load8(addr)) == '\0')
3556 				break;
3557 
3558 			if (!(tokmap[c >> 3] & (1 << (c & 0x7))))
3559 				break;
3560 		}
3561 
3562 		if (c == '\0') {
3563 			/*
3564 			 * We reached the end of the string without finding
3565 			 * any character that was not in the token string.
3566 			 * We return NULL in this case, and we set the saved
3567 			 * address to NULL as well.
3568 			 */
3569 			regs[rd] = NULL;
3570 			mstate->dtms_strtok = NULL;
3571 			break;
3572 		}
3573 
3574 		/*
3575 		 * From here on, we're copying into the destination string.
3576 		 */
3577 		for (i = 0; addr < limit && i < size - 1; addr++) {
3578 			if ((c = dtrace_load8(addr)) == '\0')
3579 				break;
3580 
3581 			if (tokmap[c >> 3] & (1 << (c & 0x7)))
3582 				break;
3583 
3584 			ASSERT(i < size);
3585 			dest[i++] = c;
3586 		}
3587 
3588 		ASSERT(i < size);
3589 		dest[i] = '\0';
3590 		regs[rd] = (uintptr_t)dest;
3591 		mstate->dtms_scratch_ptr += size;
3592 		mstate->dtms_strtok = addr;
3593 		break;
3594 	}
3595 
3596 	case DIF_SUBR_SUBSTR: {
3597 		uintptr_t s = tupregs[0].dttk_value;
3598 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3599 		char *d = (char *)mstate->dtms_scratch_ptr;
3600 		int64_t index = (int64_t)tupregs[1].dttk_value;
3601 		int64_t remaining = (int64_t)tupregs[2].dttk_value;
3602 		size_t len = dtrace_strlen((char *)s, size);
3603 		int64_t i = 0;
3604 
3605 		if (!dtrace_canload(s, len + 1, mstate, vstate)) {
3606 			regs[rd] = NULL;
3607 			break;
3608 		}
3609 
3610 		if (!DTRACE_INSCRATCH(mstate, size)) {
3611 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3612 			regs[rd] = NULL;
3613 			break;
3614 		}
3615 
3616 		if (nargs <= 2)
3617 			remaining = (int64_t)size;
3618 
3619 		if (index < 0) {
3620 			index += len;
3621 
3622 			if (index < 0 && index + remaining > 0) {
3623 				remaining += index;
3624 				index = 0;
3625 			}
3626 		}
3627 
3628 		if (index >= len || index < 0) {
3629 			remaining = 0;
3630 		} else if (remaining < 0) {
3631 			remaining += len - index;
3632 		} else if (index + remaining > size) {
3633 			remaining = size - index;
3634 		}
3635 
3636 		for (i = 0; i < remaining; i++) {
3637 			if ((d[i] = dtrace_load8(s + index + i)) == '\0')
3638 				break;
3639 		}
3640 
3641 		d[i] = '\0';
3642 
3643 		mstate->dtms_scratch_ptr += size;
3644 		regs[rd] = (uintptr_t)d;
3645 		break;
3646 	}
3647 
3648 	case DIF_SUBR_GETMAJOR:
3649 #ifdef _LP64
3650 		regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR64) & MAXMAJ64;
3651 #else
3652 		regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR) & MAXMAJ;
3653 #endif
3654 		break;
3655 
3656 	case DIF_SUBR_GETMINOR:
3657 #ifdef _LP64
3658 		regs[rd] = tupregs[0].dttk_value & MAXMIN64;
3659 #else
3660 		regs[rd] = tupregs[0].dttk_value & MAXMIN;
3661 #endif
3662 		break;
3663 
3664 	case DIF_SUBR_DDI_PATHNAME: {
3665 		/*
3666 		 * This one is a galactic mess.  We are going to roughly
3667 		 * emulate ddi_pathname(), but it's made more complicated
3668 		 * by the fact that we (a) want to include the minor name and
3669 		 * (b) must proceed iteratively instead of recursively.
3670 		 */
3671 		uintptr_t dest = mstate->dtms_scratch_ptr;
3672 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3673 		char *start = (char *)dest, *end = start + size - 1;
3674 		uintptr_t daddr = tupregs[0].dttk_value;
3675 		int64_t minor = (int64_t)tupregs[1].dttk_value;
3676 		char *s;
3677 		int i, len, depth = 0;
3678 
3679 		/*
3680 		 * Due to all the pointer jumping we do and context we must
3681 		 * rely upon, we just mandate that the user must have kernel
3682 		 * read privileges to use this routine.
3683 		 */
3684 		if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) == 0) {
3685 			*flags |= CPU_DTRACE_KPRIV;
3686 			*illval = daddr;
3687 			regs[rd] = NULL;
3688 		}
3689 
3690 		if (!DTRACE_INSCRATCH(mstate, size)) {
3691 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3692 			regs[rd] = NULL;
3693 			break;
3694 		}
3695 
3696 		*end = '\0';
3697 
3698 		/*
3699 		 * We want to have a name for the minor.  In order to do this,
3700 		 * we need to walk the minor list from the devinfo.  We want
3701 		 * to be sure that we don't infinitely walk a circular list,
3702 		 * so we check for circularity by sending a scout pointer
3703 		 * ahead two elements for every element that we iterate over;
3704 		 * if the list is circular, these will ultimately point to the
3705 		 * same element.  You may recognize this little trick as the
3706 		 * answer to a stupid interview question -- one that always
3707 		 * seems to be asked by those who had to have it laboriously
3708 		 * explained to them, and who can't even concisely describe
3709 		 * the conditions under which one would be forced to resort to
3710 		 * this technique.  Needless to say, those conditions are
3711 		 * found here -- and probably only here.  Is this the only use
3712 		 * of this infamous trick in shipping, production code?  If it
3713 		 * isn't, it probably should be...
3714 		 */
3715 		if (minor != -1) {
3716 			uintptr_t maddr = dtrace_loadptr(daddr +
3717 			    offsetof(struct dev_info, devi_minor));
3718 
3719 			uintptr_t next = offsetof(struct ddi_minor_data, next);
3720 			uintptr_t name = offsetof(struct ddi_minor_data,
3721 			    d_minor) + offsetof(struct ddi_minor, name);
3722 			uintptr_t dev = offsetof(struct ddi_minor_data,
3723 			    d_minor) + offsetof(struct ddi_minor, dev);
3724 			uintptr_t scout;
3725 
3726 			if (maddr != NULL)
3727 				scout = dtrace_loadptr(maddr + next);
3728 
3729 			while (maddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
3730 				uint64_t m;
3731 #ifdef _LP64
3732 				m = dtrace_load64(maddr + dev) & MAXMIN64;
3733 #else
3734 				m = dtrace_load32(maddr + dev) & MAXMIN;
3735 #endif
3736 				if (m != minor) {
3737 					maddr = dtrace_loadptr(maddr + next);
3738 
3739 					if (scout == NULL)
3740 						continue;
3741 
3742 					scout = dtrace_loadptr(scout + next);
3743 
3744 					if (scout == NULL)
3745 						continue;
3746 
3747 					scout = dtrace_loadptr(scout + next);
3748 
3749 					if (scout == NULL)
3750 						continue;
3751 
3752 					if (scout == maddr) {
3753 						*flags |= CPU_DTRACE_ILLOP;
3754 						break;
3755 					}
3756 
3757 					continue;
3758 				}
3759 
3760 				/*
3761 				 * We have the minor data.  Now we need to
3762 				 * copy the minor's name into the end of the
3763 				 * pathname.
3764 				 */
3765 				s = (char *)dtrace_loadptr(maddr + name);
3766 				len = dtrace_strlen(s, size);
3767 
3768 				if (*flags & CPU_DTRACE_FAULT)
3769 					break;
3770 
3771 				if (len != 0) {
3772 					if ((end -= (len + 1)) < start)
3773 						break;
3774 
3775 					*end = ':';
3776 				}
3777 
3778 				for (i = 1; i <= len; i++)
3779 					end[i] = dtrace_load8((uintptr_t)s++);
3780 				break;
3781 			}
3782 		}
3783 
3784 		while (daddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
3785 			ddi_node_state_t devi_state;
3786 
3787 			devi_state = dtrace_load32(daddr +
3788 			    offsetof(struct dev_info, devi_node_state));
3789 
3790 			if (*flags & CPU_DTRACE_FAULT)
3791 				break;
3792 
3793 			if (devi_state >= DS_INITIALIZED) {
3794 				s = (char *)dtrace_loadptr(daddr +
3795 				    offsetof(struct dev_info, devi_addr));
3796 				len = dtrace_strlen(s, size);
3797 
3798 				if (*flags & CPU_DTRACE_FAULT)
3799 					break;
3800 
3801 				if (len != 0) {
3802 					if ((end -= (len + 1)) < start)
3803 						break;
3804 
3805 					*end = '@';
3806 				}
3807 
3808 				for (i = 1; i <= len; i++)
3809 					end[i] = dtrace_load8((uintptr_t)s++);
3810 			}
3811 
3812 			/*
3813 			 * Now for the node name...
3814 			 */
3815 			s = (char *)dtrace_loadptr(daddr +
3816 			    offsetof(struct dev_info, devi_node_name));
3817 
3818 			daddr = dtrace_loadptr(daddr +
3819 			    offsetof(struct dev_info, devi_parent));
3820 
3821 			/*
3822 			 * If our parent is NULL (that is, if we're the root
3823 			 * node), we're going to use the special path
3824 			 * "devices".
3825 			 */
3826 			if (daddr == NULL)
3827 				s = "devices";
3828 
3829 			len = dtrace_strlen(s, size);
3830 			if (*flags & CPU_DTRACE_FAULT)
3831 				break;
3832 
3833 			if ((end -= (len + 1)) < start)
3834 				break;
3835 
3836 			for (i = 1; i <= len; i++)
3837 				end[i] = dtrace_load8((uintptr_t)s++);
3838 			*end = '/';
3839 
3840 			if (depth++ > dtrace_devdepth_max) {
3841 				*flags |= CPU_DTRACE_ILLOP;
3842 				break;
3843 			}
3844 		}
3845 
3846 		if (end < start)
3847 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3848 
3849 		if (daddr == NULL) {
3850 			regs[rd] = (uintptr_t)end;
3851 			mstate->dtms_scratch_ptr += size;
3852 		}
3853 
3854 		break;
3855 	}
3856 
3857 	case DIF_SUBR_STRJOIN: {
3858 		char *d = (char *)mstate->dtms_scratch_ptr;
3859 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3860 		uintptr_t s1 = tupregs[0].dttk_value;
3861 		uintptr_t s2 = tupregs[1].dttk_value;
3862 		int i = 0;
3863 
3864 		if (!dtrace_strcanload(s1, size, mstate, vstate) ||
3865 		    !dtrace_strcanload(s2, size, mstate, vstate)) {
3866 			regs[rd] = NULL;
3867 			break;
3868 		}
3869 
3870 		if (!DTRACE_INSCRATCH(mstate, size)) {
3871 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3872 			regs[rd] = NULL;
3873 			break;
3874 		}
3875 
3876 		for (;;) {
3877 			if (i >= size) {
3878 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3879 				regs[rd] = NULL;
3880 				break;
3881 			}
3882 
3883 			if ((d[i++] = dtrace_load8(s1++)) == '\0') {
3884 				i--;
3885 				break;
3886 			}
3887 		}
3888 
3889 		for (;;) {
3890 			if (i >= size) {
3891 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3892 				regs[rd] = NULL;
3893 				break;
3894 			}
3895 
3896 			if ((d[i++] = dtrace_load8(s2++)) == '\0')
3897 				break;
3898 		}
3899 
3900 		if (i < size) {
3901 			mstate->dtms_scratch_ptr += i;
3902 			regs[rd] = (uintptr_t)d;
3903 		}
3904 
3905 		break;
3906 	}
3907 
3908 	case DIF_SUBR_LLTOSTR: {
3909 		int64_t i = (int64_t)tupregs[0].dttk_value;
3910 		int64_t val = i < 0 ? i * -1 : i;
3911 		uint64_t size = 22;	/* enough room for 2^64 in decimal */
3912 		char *end = (char *)mstate->dtms_scratch_ptr + size - 1;
3913 
3914 		if (!DTRACE_INSCRATCH(mstate, size)) {
3915 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3916 			regs[rd] = NULL;
3917 			break;
3918 		}
3919 
3920 		for (*end-- = '\0'; val; val /= 10)
3921 			*end-- = '0' + (val % 10);
3922 
3923 		if (i == 0)
3924 			*end-- = '0';
3925 
3926 		if (i < 0)
3927 			*end-- = '-';
3928 
3929 		regs[rd] = (uintptr_t)end + 1;
3930 		mstate->dtms_scratch_ptr += size;
3931 		break;
3932 	}
3933 
3934 	case DIF_SUBR_HTONS:
3935 	case DIF_SUBR_NTOHS:
3936 #ifdef _BIG_ENDIAN
3937 		regs[rd] = (uint16_t)tupregs[0].dttk_value;
3938 #else
3939 		regs[rd] = DT_BSWAP_16((uint16_t)tupregs[0].dttk_value);
3940 #endif
3941 		break;
3942 
3943 
3944 	case DIF_SUBR_HTONL:
3945 	case DIF_SUBR_NTOHL:
3946 #ifdef _BIG_ENDIAN
3947 		regs[rd] = (uint32_t)tupregs[0].dttk_value;
3948 #else
3949 		regs[rd] = DT_BSWAP_32((uint32_t)tupregs[0].dttk_value);
3950 #endif
3951 		break;
3952 
3953 
3954 	case DIF_SUBR_HTONLL:
3955 	case DIF_SUBR_NTOHLL:
3956 #ifdef _BIG_ENDIAN
3957 		regs[rd] = (uint64_t)tupregs[0].dttk_value;
3958 #else
3959 		regs[rd] = DT_BSWAP_64((uint64_t)tupregs[0].dttk_value);
3960 #endif
3961 		break;
3962 
3963 
3964 	case DIF_SUBR_DIRNAME:
3965 	case DIF_SUBR_BASENAME: {
3966 		char *dest = (char *)mstate->dtms_scratch_ptr;
3967 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3968 		uintptr_t src = tupregs[0].dttk_value;
3969 		int i, j, len = dtrace_strlen((char *)src, size);
3970 		int lastbase = -1, firstbase = -1, lastdir = -1;
3971 		int start, end;
3972 
3973 		if (!dtrace_canload(src, len + 1, mstate, vstate)) {
3974 			regs[rd] = NULL;
3975 			break;
3976 		}
3977 
3978 		if (!DTRACE_INSCRATCH(mstate, size)) {
3979 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3980 			regs[rd] = NULL;
3981 			break;
3982 		}
3983 
3984 		/*
3985 		 * The basename and dirname for a zero-length string is
3986 		 * defined to be "."
3987 		 */
3988 		if (len == 0) {
3989 			len = 1;
3990 			src = (uintptr_t)".";
3991 		}
3992 
3993 		/*
3994 		 * Start from the back of the string, moving back toward the
3995 		 * front until we see a character that isn't a slash.  That
3996 		 * character is the last character in the basename.
3997 		 */
3998 		for (i = len - 1; i >= 0; i--) {
3999 			if (dtrace_load8(src + i) != '/')
4000 				break;
4001 		}
4002 
4003 		if (i >= 0)
4004 			lastbase = i;
4005 
4006 		/*
4007 		 * Starting from the last character in the basename, move
4008 		 * towards the front until we find a slash.  The character
4009 		 * that we processed immediately before that is the first
4010 		 * character in the basename.
4011 		 */
4012 		for (; i >= 0; i--) {
4013 			if (dtrace_load8(src + i) == '/')
4014 				break;
4015 		}
4016 
4017 		if (i >= 0)
4018 			firstbase = i + 1;
4019 
4020 		/*
4021 		 * Now keep going until we find a non-slash character.  That
4022 		 * character is the last character in the dirname.
4023 		 */
4024 		for (; i >= 0; i--) {
4025 			if (dtrace_load8(src + i) != '/')
4026 				break;
4027 		}
4028 
4029 		if (i >= 0)
4030 			lastdir = i;
4031 
4032 		ASSERT(!(lastbase == -1 && firstbase != -1));
4033 		ASSERT(!(firstbase == -1 && lastdir != -1));
4034 
4035 		if (lastbase == -1) {
4036 			/*
4037 			 * We didn't find a non-slash character.  We know that
4038 			 * the length is non-zero, so the whole string must be
4039 			 * slashes.  In either the dirname or the basename
4040 			 * case, we return '/'.
4041 			 */
4042 			ASSERT(firstbase == -1);
4043 			firstbase = lastbase = lastdir = 0;
4044 		}
4045 
4046 		if (firstbase == -1) {
4047 			/*
4048 			 * The entire string consists only of a basename
4049 			 * component.  If we're looking for dirname, we need
4050 			 * to change our string to be just "."; if we're
4051 			 * looking for a basename, we'll just set the first
4052 			 * character of the basename to be 0.
4053 			 */
4054 			if (subr == DIF_SUBR_DIRNAME) {
4055 				ASSERT(lastdir == -1);
4056 				src = (uintptr_t)".";
4057 				lastdir = 0;
4058 			} else {
4059 				firstbase = 0;
4060 			}
4061 		}
4062 
4063 		if (subr == DIF_SUBR_DIRNAME) {
4064 			if (lastdir == -1) {
4065 				/*
4066 				 * We know that we have a slash in the name --
4067 				 * or lastdir would be set to 0, above.  And
4068 				 * because lastdir is -1, we know that this
4069 				 * slash must be the first character.  (That
4070 				 * is, the full string must be of the form
4071 				 * "/basename".)  In this case, the last
4072 				 * character of the directory name is 0.
4073 				 */
4074 				lastdir = 0;
4075 			}
4076 
4077 			start = 0;
4078 			end = lastdir;
4079 		} else {
4080 			ASSERT(subr == DIF_SUBR_BASENAME);
4081 			ASSERT(firstbase != -1 && lastbase != -1);
4082 			start = firstbase;
4083 			end = lastbase;
4084 		}
4085 
4086 		for (i = start, j = 0; i <= end && j < size - 1; i++, j++)
4087 			dest[j] = dtrace_load8(src + i);
4088 
4089 		dest[j] = '\0';
4090 		regs[rd] = (uintptr_t)dest;
4091 		mstate->dtms_scratch_ptr += size;
4092 		break;
4093 	}
4094 
4095 	case DIF_SUBR_CLEANPATH: {
4096 		char *dest = (char *)mstate->dtms_scratch_ptr, c;
4097 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4098 		uintptr_t src = tupregs[0].dttk_value;
4099 		int i = 0, j = 0;
4100 
4101 		if (!dtrace_strcanload(src, size, mstate, vstate)) {
4102 			regs[rd] = NULL;
4103 			break;
4104 		}
4105 
4106 		if (!DTRACE_INSCRATCH(mstate, size)) {
4107 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4108 			regs[rd] = NULL;
4109 			break;
4110 		}
4111 
4112 		/*
4113 		 * Move forward, loading each character.
4114 		 */
4115 		do {
4116 			c = dtrace_load8(src + i++);
4117 next:
4118 			if (j + 5 >= size)	/* 5 = strlen("/..c\0") */
4119 				break;
4120 
4121 			if (c != '/') {
4122 				dest[j++] = c;
4123 				continue;
4124 			}
4125 
4126 			c = dtrace_load8(src + i++);
4127 
4128 			if (c == '/') {
4129 				/*
4130 				 * We have two slashes -- we can just advance
4131 				 * to the next character.
4132 				 */
4133 				goto next;
4134 			}
4135 
4136 			if (c != '.') {
4137 				/*
4138 				 * This is not "." and it's not ".." -- we can
4139 				 * just store the "/" and this character and
4140 				 * drive on.
4141 				 */
4142 				dest[j++] = '/';
4143 				dest[j++] = c;
4144 				continue;
4145 			}
4146 
4147 			c = dtrace_load8(src + i++);
4148 
4149 			if (c == '/') {
4150 				/*
4151 				 * This is a "/./" component.  We're not going
4152 				 * to store anything in the destination buffer;
4153 				 * we're just going to go to the next component.
4154 				 */
4155 				goto next;
4156 			}
4157 
4158 			if (c != '.') {
4159 				/*
4160 				 * This is not ".." -- we can just store the
4161 				 * "/." and this character and continue
4162 				 * processing.
4163 				 */
4164 				dest[j++] = '/';
4165 				dest[j++] = '.';
4166 				dest[j++] = c;
4167 				continue;
4168 			}
4169 
4170 			c = dtrace_load8(src + i++);
4171 
4172 			if (c != '/' && c != '\0') {
4173 				/*
4174 				 * This is not ".." -- it's "..[mumble]".
4175 				 * We'll store the "/.." and this character
4176 				 * and continue processing.
4177 				 */
4178 				dest[j++] = '/';
4179 				dest[j++] = '.';
4180 				dest[j++] = '.';
4181 				dest[j++] = c;
4182 				continue;
4183 			}
4184 
4185 			/*
4186 			 * This is "/../" or "/..\0".  We need to back up
4187 			 * our destination pointer until we find a "/".
4188 			 */
4189 			i--;
4190 			while (j != 0 && dest[--j] != '/')
4191 				continue;
4192 
4193 			if (c == '\0')
4194 				dest[++j] = '/';
4195 		} while (c != '\0');
4196 
4197 		dest[j] = '\0';
4198 		regs[rd] = (uintptr_t)dest;
4199 		mstate->dtms_scratch_ptr += size;
4200 		break;
4201 	}
4202 
4203 	case DIF_SUBR_INET_NTOA:
4204 	case DIF_SUBR_INET_NTOA6:
4205 	case DIF_SUBR_INET_NTOP: {
4206 		size_t size;
4207 		int af, argi, i;
4208 		char *base, *end;
4209 
4210 		if (subr == DIF_SUBR_INET_NTOP) {
4211 			af = (int)tupregs[0].dttk_value;
4212 			argi = 1;
4213 		} else {
4214 			af = subr == DIF_SUBR_INET_NTOA ? AF_INET: AF_INET6;
4215 			argi = 0;
4216 		}
4217 
4218 		if (af == AF_INET) {
4219 			ipaddr_t ip4;
4220 			uint8_t *ptr8, val;
4221 
4222 			/*
4223 			 * Safely load the IPv4 address.
4224 			 */
4225 			ip4 = dtrace_load32(tupregs[argi].dttk_value);
4226 
4227 			/*
4228 			 * Check an IPv4 string will fit in scratch.
4229 			 */
4230 			size = INET_ADDRSTRLEN;
4231 			if (!DTRACE_INSCRATCH(mstate, size)) {
4232 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4233 				regs[rd] = NULL;
4234 				break;
4235 			}
4236 			base = (char *)mstate->dtms_scratch_ptr;
4237 			end = (char *)mstate->dtms_scratch_ptr + size - 1;
4238 
4239 			/*
4240 			 * Stringify as a dotted decimal quad.
4241 			 */
4242 			*end-- = '\0';
4243 			ptr8 = (uint8_t *)&ip4;
4244 			for (i = 3; i >= 0; i--) {
4245 				val = ptr8[i];
4246 
4247 				if (val == 0) {
4248 					*end-- = '0';
4249 				} else {
4250 					for (; val; val /= 10) {
4251 						*end-- = '0' + (val % 10);
4252 					}
4253 				}
4254 
4255 				if (i > 0)
4256 					*end-- = '.';
4257 			}
4258 			ASSERT(end + 1 >= base);
4259 
4260 		} else if (af == AF_INET6) {
4261 			struct in6_addr ip6;
4262 			int firstzero, tryzero, numzero, v6end;
4263 			uint16_t val;
4264 			const char digits[] = "0123456789abcdef";
4265 
4266 			/*
4267 			 * Stringify using RFC 1884 convention 2 - 16 bit
4268 			 * hexadecimal values with a zero-run compression.
4269 			 * Lower case hexadecimal digits are used.
4270 			 * 	eg, fe80::214:4fff:fe0b:76c8.
4271 			 * The IPv4 embedded form is returned for inet_ntop,
4272 			 * just the IPv4 string is returned for inet_ntoa6.
4273 			 */
4274 
4275 			/*
4276 			 * Safely load the IPv6 address.
4277 			 */
4278 			dtrace_bcopy(
4279 			    (void *)(uintptr_t)tupregs[argi].dttk_value,
4280 			    (void *)(uintptr_t)&ip6, sizeof (struct in6_addr));
4281 
4282 			/*
4283 			 * Check an IPv6 string will fit in scratch.
4284 			 */
4285 			size = INET6_ADDRSTRLEN;
4286 			if (!DTRACE_INSCRATCH(mstate, size)) {
4287 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4288 				regs[rd] = NULL;
4289 				break;
4290 			}
4291 			base = (char *)mstate->dtms_scratch_ptr;
4292 			end = (char *)mstate->dtms_scratch_ptr + size - 1;
4293 			*end-- = '\0';
4294 
4295 			/*
4296 			 * Find the longest run of 16 bit zero values
4297 			 * for the single allowed zero compression - "::".
4298 			 */
4299 			firstzero = -1;
4300 			tryzero = -1;
4301 			numzero = 1;
4302 			for (i = 0; i < sizeof (struct in6_addr); i++) {
4303 				if (ip6._S6_un._S6_u8[i] == 0 &&
4304 				    tryzero == -1 && i % 2 == 0) {
4305 					tryzero = i;
4306 					continue;
4307 				}
4308 
4309 				if (tryzero != -1 &&
4310 				    (ip6._S6_un._S6_u8[i] != 0 ||
4311 				    i == sizeof (struct in6_addr) - 1)) {
4312 
4313 					if (i - tryzero <= numzero) {
4314 						tryzero = -1;
4315 						continue;
4316 					}
4317 
4318 					firstzero = tryzero;
4319 					numzero = i - i % 2 - tryzero;
4320 					tryzero = -1;
4321 
4322 					if (ip6._S6_un._S6_u8[i] == 0 &&
4323 					    i == sizeof (struct in6_addr) - 1)
4324 						numzero += 2;
4325 				}
4326 			}
4327 			ASSERT(firstzero + numzero <= sizeof (struct in6_addr));
4328 
4329 			/*
4330 			 * Check for an IPv4 embedded address.
4331 			 */
4332 			v6end = sizeof (struct in6_addr) - 2;
4333 			if (IN6_IS_ADDR_V4MAPPED(&ip6) ||
4334 			    IN6_IS_ADDR_V4COMPAT(&ip6)) {
4335 				for (i = sizeof (struct in6_addr) - 1;
4336 				    i >= DTRACE_V4MAPPED_OFFSET; i--) {
4337 					ASSERT(end >= base);
4338 
4339 					val = ip6._S6_un._S6_u8[i];
4340 
4341 					if (val == 0) {
4342 						*end-- = '0';
4343 					} else {
4344 						for (; val; val /= 10) {
4345 							*end-- = '0' + val % 10;
4346 						}
4347 					}
4348 
4349 					if (i > DTRACE_V4MAPPED_OFFSET)
4350 						*end-- = '.';
4351 				}
4352 
4353 				if (subr == DIF_SUBR_INET_NTOA6)
4354 					goto inetout;
4355 
4356 				/*
4357 				 * Set v6end to skip the IPv4 address that
4358 				 * we have already stringified.
4359 				 */
4360 				v6end = 10;
4361 			}
4362 
4363 			/*
4364 			 * Build the IPv6 string by working through the
4365 			 * address in reverse.
4366 			 */
4367 			for (i = v6end; i >= 0; i -= 2) {
4368 				ASSERT(end >= base);
4369 
4370 				if (i == firstzero + numzero - 2) {
4371 					*end-- = ':';
4372 					*end-- = ':';
4373 					i -= numzero - 2;
4374 					continue;
4375 				}
4376 
4377 				if (i < 14 && i != firstzero - 2)
4378 					*end-- = ':';
4379 
4380 				val = (ip6._S6_un._S6_u8[i] << 8) +
4381 				    ip6._S6_un._S6_u8[i + 1];
4382 
4383 				if (val == 0) {
4384 					*end-- = '0';
4385 				} else {
4386 					for (; val; val /= 16) {
4387 						*end-- = digits[val % 16];
4388 					}
4389 				}
4390 			}
4391 			ASSERT(end + 1 >= base);
4392 
4393 		} else {
4394 			/*
4395 			 * The user didn't use AH_INET or AH_INET6.
4396 			 */
4397 			DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
4398 			regs[rd] = NULL;
4399 			break;
4400 		}
4401 
4402 inetout:	regs[rd] = (uintptr_t)end + 1;
4403 		mstate->dtms_scratch_ptr += size;
4404 		break;
4405 	}
4406 
4407 	}
4408 }
4409 
4410 /*
4411  * Emulate the execution of DTrace IR instructions specified by the given
4412  * DIF object.  This function is deliberately void of assertions as all of
4413  * the necessary checks are handled by a call to dtrace_difo_validate().
4414  */
4415 static uint64_t
4416 dtrace_dif_emulate(dtrace_difo_t *difo, dtrace_mstate_t *mstate,
4417     dtrace_vstate_t *vstate, dtrace_state_t *state)
4418 {
4419 	const dif_instr_t *text = difo->dtdo_buf;
4420 	const uint_t textlen = difo->dtdo_len;
4421 	const char *strtab = difo->dtdo_strtab;
4422 	const uint64_t *inttab = difo->dtdo_inttab;
4423 
4424 	uint64_t rval = 0;
4425 	dtrace_statvar_t *svar;
4426 	dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
4427 	dtrace_difv_t *v;
4428 	volatile uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
4429 	volatile uintptr_t *illval = &cpu_core[CPU->cpu_id].cpuc_dtrace_illval;
4430 
4431 	dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
4432 	uint64_t regs[DIF_DIR_NREGS];
4433 	uint64_t *tmp;
4434 
4435 	uint8_t cc_n = 0, cc_z = 0, cc_v = 0, cc_c = 0;
4436 	int64_t cc_r;
4437 	uint_t pc = 0, id, opc;
4438 	uint8_t ttop = 0;
4439 	dif_instr_t instr;
4440 	uint_t r1, r2, rd;
4441 
4442 	/*
4443 	 * We stash the current DIF object into the machine state: we need it
4444 	 * for subsequent access checking.
4445 	 */
4446 	mstate->dtms_difo = difo;
4447 
4448 	regs[DIF_REG_R0] = 0; 		/* %r0 is fixed at zero */
4449 
4450 	while (pc < textlen && !(*flags & CPU_DTRACE_FAULT)) {
4451 		opc = pc;
4452 
4453 		instr = text[pc++];
4454 		r1 = DIF_INSTR_R1(instr);
4455 		r2 = DIF_INSTR_R2(instr);
4456 		rd = DIF_INSTR_RD(instr);
4457 
4458 		switch (DIF_INSTR_OP(instr)) {
4459 		case DIF_OP_OR:
4460 			regs[rd] = regs[r1] | regs[r2];
4461 			break;
4462 		case DIF_OP_XOR:
4463 			regs[rd] = regs[r1] ^ regs[r2];
4464 			break;
4465 		case DIF_OP_AND:
4466 			regs[rd] = regs[r1] & regs[r2];
4467 			break;
4468 		case DIF_OP_SLL:
4469 			regs[rd] = regs[r1] << regs[r2];
4470 			break;
4471 		case DIF_OP_SRL:
4472 			regs[rd] = regs[r1] >> regs[r2];
4473 			break;
4474 		case DIF_OP_SUB:
4475 			regs[rd] = regs[r1] - regs[r2];
4476 			break;
4477 		case DIF_OP_ADD:
4478 			regs[rd] = regs[r1] + regs[r2];
4479 			break;
4480 		case DIF_OP_MUL:
4481 			regs[rd] = regs[r1] * regs[r2];
4482 			break;
4483 		case DIF_OP_SDIV:
4484 			if (regs[r2] == 0) {
4485 				regs[rd] = 0;
4486 				*flags |= CPU_DTRACE_DIVZERO;
4487 			} else {
4488 				regs[rd] = (int64_t)regs[r1] /
4489 				    (int64_t)regs[r2];
4490 			}
4491 			break;
4492 
4493 		case DIF_OP_UDIV:
4494 			if (regs[r2] == 0) {
4495 				regs[rd] = 0;
4496 				*flags |= CPU_DTRACE_DIVZERO;
4497 			} else {
4498 				regs[rd] = regs[r1] / regs[r2];
4499 			}
4500 			break;
4501 
4502 		case DIF_OP_SREM:
4503 			if (regs[r2] == 0) {
4504 				regs[rd] = 0;
4505 				*flags |= CPU_DTRACE_DIVZERO;
4506 			} else {
4507 				regs[rd] = (int64_t)regs[r1] %
4508 				    (int64_t)regs[r2];
4509 			}
4510 			break;
4511 
4512 		case DIF_OP_UREM:
4513 			if (regs[r2] == 0) {
4514 				regs[rd] = 0;
4515 				*flags |= CPU_DTRACE_DIVZERO;
4516 			} else {
4517 				regs[rd] = regs[r1] % regs[r2];
4518 			}
4519 			break;
4520 
4521 		case DIF_OP_NOT:
4522 			regs[rd] = ~regs[r1];
4523 			break;
4524 		case DIF_OP_MOV:
4525 			regs[rd] = regs[r1];
4526 			break;
4527 		case DIF_OP_CMP:
4528 			cc_r = regs[r1] - regs[r2];
4529 			cc_n = cc_r < 0;
4530 			cc_z = cc_r == 0;
4531 			cc_v = 0;
4532 			cc_c = regs[r1] < regs[r2];
4533 			break;
4534 		case DIF_OP_TST:
4535 			cc_n = cc_v = cc_c = 0;
4536 			cc_z = regs[r1] == 0;
4537 			break;
4538 		case DIF_OP_BA:
4539 			pc = DIF_INSTR_LABEL(instr);
4540 			break;
4541 		case DIF_OP_BE:
4542 			if (cc_z)
4543 				pc = DIF_INSTR_LABEL(instr);
4544 			break;
4545 		case DIF_OP_BNE:
4546 			if (cc_z == 0)
4547 				pc = DIF_INSTR_LABEL(instr);
4548 			break;
4549 		case DIF_OP_BG:
4550 			if ((cc_z | (cc_n ^ cc_v)) == 0)
4551 				pc = DIF_INSTR_LABEL(instr);
4552 			break;
4553 		case DIF_OP_BGU:
4554 			if ((cc_c | cc_z) == 0)
4555 				pc = DIF_INSTR_LABEL(instr);
4556 			break;
4557 		case DIF_OP_BGE:
4558 			if ((cc_n ^ cc_v) == 0)
4559 				pc = DIF_INSTR_LABEL(instr);
4560 			break;
4561 		case DIF_OP_BGEU:
4562 			if (cc_c == 0)
4563 				pc = DIF_INSTR_LABEL(instr);
4564 			break;
4565 		case DIF_OP_BL:
4566 			if (cc_n ^ cc_v)
4567 				pc = DIF_INSTR_LABEL(instr);
4568 			break;
4569 		case DIF_OP_BLU:
4570 			if (cc_c)
4571 				pc = DIF_INSTR_LABEL(instr);
4572 			break;
4573 		case DIF_OP_BLE:
4574 			if (cc_z | (cc_n ^ cc_v))
4575 				pc = DIF_INSTR_LABEL(instr);
4576 			break;
4577 		case DIF_OP_BLEU:
4578 			if (cc_c | cc_z)
4579 				pc = DIF_INSTR_LABEL(instr);
4580 			break;
4581 		case DIF_OP_RLDSB:
4582 			if (!dtrace_canstore(regs[r1], 1, mstate, vstate)) {
4583 				*flags |= CPU_DTRACE_KPRIV;
4584 				*illval = regs[r1];
4585 				break;
4586 			}
4587 			/*FALLTHROUGH*/
4588 		case DIF_OP_LDSB:
4589 			regs[rd] = (int8_t)dtrace_load8(regs[r1]);
4590 			break;
4591 		case DIF_OP_RLDSH:
4592 			if (!dtrace_canstore(regs[r1], 2, mstate, vstate)) {
4593 				*flags |= CPU_DTRACE_KPRIV;
4594 				*illval = regs[r1];
4595 				break;
4596 			}
4597 			/*FALLTHROUGH*/
4598 		case DIF_OP_LDSH:
4599 			regs[rd] = (int16_t)dtrace_load16(regs[r1]);
4600 			break;
4601 		case DIF_OP_RLDSW:
4602 			if (!dtrace_canstore(regs[r1], 4, mstate, vstate)) {
4603 				*flags |= CPU_DTRACE_KPRIV;
4604 				*illval = regs[r1];
4605 				break;
4606 			}
4607 			/*FALLTHROUGH*/
4608 		case DIF_OP_LDSW:
4609 			regs[rd] = (int32_t)dtrace_load32(regs[r1]);
4610 			break;
4611 		case DIF_OP_RLDUB:
4612 			if (!dtrace_canstore(regs[r1], 1, mstate, vstate)) {
4613 				*flags |= CPU_DTRACE_KPRIV;
4614 				*illval = regs[r1];
4615 				break;
4616 			}
4617 			/*FALLTHROUGH*/
4618 		case DIF_OP_LDUB:
4619 			regs[rd] = dtrace_load8(regs[r1]);
4620 			break;
4621 		case DIF_OP_RLDUH:
4622 			if (!dtrace_canstore(regs[r1], 2, mstate, vstate)) {
4623 				*flags |= CPU_DTRACE_KPRIV;
4624 				*illval = regs[r1];
4625 				break;
4626 			}
4627 			/*FALLTHROUGH*/
4628 		case DIF_OP_LDUH:
4629 			regs[rd] = dtrace_load16(regs[r1]);
4630 			break;
4631 		case DIF_OP_RLDUW:
4632 			if (!dtrace_canstore(regs[r1], 4, mstate, vstate)) {
4633 				*flags |= CPU_DTRACE_KPRIV;
4634 				*illval = regs[r1];
4635 				break;
4636 			}
4637 			/*FALLTHROUGH*/
4638 		case DIF_OP_LDUW:
4639 			regs[rd] = dtrace_load32(regs[r1]);
4640 			break;
4641 		case DIF_OP_RLDX:
4642 			if (!dtrace_canstore(regs[r1], 8, mstate, vstate)) {
4643 				*flags |= CPU_DTRACE_KPRIV;
4644 				*illval = regs[r1];
4645 				break;
4646 			}
4647 			/*FALLTHROUGH*/
4648 		case DIF_OP_LDX:
4649 			regs[rd] = dtrace_load64(regs[r1]);
4650 			break;
4651 		case DIF_OP_ULDSB:
4652 			regs[rd] = (int8_t)
4653 			    dtrace_fuword8((void *)(uintptr_t)regs[r1]);
4654 			break;
4655 		case DIF_OP_ULDSH:
4656 			regs[rd] = (int16_t)
4657 			    dtrace_fuword16((void *)(uintptr_t)regs[r1]);
4658 			break;
4659 		case DIF_OP_ULDSW:
4660 			regs[rd] = (int32_t)
4661 			    dtrace_fuword32((void *)(uintptr_t)regs[r1]);
4662 			break;
4663 		case DIF_OP_ULDUB:
4664 			regs[rd] =
4665 			    dtrace_fuword8((void *)(uintptr_t)regs[r1]);
4666 			break;
4667 		case DIF_OP_ULDUH:
4668 			regs[rd] =
4669 			    dtrace_fuword16((void *)(uintptr_t)regs[r1]);
4670 			break;
4671 		case DIF_OP_ULDUW:
4672 			regs[rd] =
4673 			    dtrace_fuword32((void *)(uintptr_t)regs[r1]);
4674 			break;
4675 		case DIF_OP_ULDX:
4676 			regs[rd] =
4677 			    dtrace_fuword64((void *)(uintptr_t)regs[r1]);
4678 			break;
4679 		case DIF_OP_RET:
4680 			rval = regs[rd];
4681 			pc = textlen;
4682 			break;
4683 		case DIF_OP_NOP:
4684 			break;
4685 		case DIF_OP_SETX:
4686 			regs[rd] = inttab[DIF_INSTR_INTEGER(instr)];
4687 			break;
4688 		case DIF_OP_SETS:
4689 			regs[rd] = (uint64_t)(uintptr_t)
4690 			    (strtab + DIF_INSTR_STRING(instr));
4691 			break;
4692 		case DIF_OP_SCMP: {
4693 			size_t sz = state->dts_options[DTRACEOPT_STRSIZE];
4694 			uintptr_t s1 = regs[r1];
4695 			uintptr_t s2 = regs[r2];
4696 
4697 			if (s1 != NULL &&
4698 			    !dtrace_strcanload(s1, sz, mstate, vstate))
4699 				break;
4700 			if (s2 != NULL &&
4701 			    !dtrace_strcanload(s2, sz, mstate, vstate))
4702 				break;
4703 
4704 			cc_r = dtrace_strncmp((char *)s1, (char *)s2, sz);
4705 
4706 			cc_n = cc_r < 0;
4707 			cc_z = cc_r == 0;
4708 			cc_v = cc_c = 0;
4709 			break;
4710 		}
4711 		case DIF_OP_LDGA:
4712 			regs[rd] = dtrace_dif_variable(mstate, state,
4713 			    r1, regs[r2]);
4714 			break;
4715 		case DIF_OP_LDGS:
4716 			id = DIF_INSTR_VAR(instr);
4717 
4718 			if (id >= DIF_VAR_OTHER_UBASE) {
4719 				uintptr_t a;
4720 
4721 				id -= DIF_VAR_OTHER_UBASE;
4722 				svar = vstate->dtvs_globals[id];
4723 				ASSERT(svar != NULL);
4724 				v = &svar->dtsv_var;
4725 
4726 				if (!(v->dtdv_type.dtdt_flags & DIF_TF_BYREF)) {
4727 					regs[rd] = svar->dtsv_data;
4728 					break;
4729 				}
4730 
4731 				a = (uintptr_t)svar->dtsv_data;
4732 
4733 				if (*(uint8_t *)a == UINT8_MAX) {
4734 					/*
4735 					 * If the 0th byte is set to UINT8_MAX
4736 					 * then this is to be treated as a
4737 					 * reference to a NULL variable.
4738 					 */
4739 					regs[rd] = NULL;
4740 				} else {
4741 					regs[rd] = a + sizeof (uint64_t);
4742 				}
4743 
4744 				break;
4745 			}
4746 
4747 			regs[rd] = dtrace_dif_variable(mstate, state, id, 0);
4748 			break;
4749 
4750 		case DIF_OP_STGS:
4751 			id = DIF_INSTR_VAR(instr);
4752 
4753 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
4754 			id -= DIF_VAR_OTHER_UBASE;
4755 
4756 			svar = vstate->dtvs_globals[id];
4757 			ASSERT(svar != NULL);
4758 			v = &svar->dtsv_var;
4759 
4760 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
4761 				uintptr_t a = (uintptr_t)svar->dtsv_data;
4762 
4763 				ASSERT(a != NULL);
4764 				ASSERT(svar->dtsv_size != 0);
4765 
4766 				if (regs[rd] == NULL) {
4767 					*(uint8_t *)a = UINT8_MAX;
4768 					break;
4769 				} else {
4770 					*(uint8_t *)a = 0;
4771 					a += sizeof (uint64_t);
4772 				}
4773 				if (!dtrace_vcanload(
4774 				    (void *)(uintptr_t)regs[rd], &v->dtdv_type,
4775 				    mstate, vstate))
4776 					break;
4777 
4778 				dtrace_vcopy((void *)(uintptr_t)regs[rd],
4779 				    (void *)a, &v->dtdv_type);
4780 				break;
4781 			}
4782 
4783 			svar->dtsv_data = regs[rd];
4784 			break;
4785 
4786 		case DIF_OP_LDTA:
4787 			/*
4788 			 * There are no DTrace built-in thread-local arrays at
4789 			 * present.  This opcode is saved for future work.
4790 			 */
4791 			*flags |= CPU_DTRACE_ILLOP;
4792 			regs[rd] = 0;
4793 			break;
4794 
4795 		case DIF_OP_LDLS:
4796 			id = DIF_INSTR_VAR(instr);
4797 
4798 			if (id < DIF_VAR_OTHER_UBASE) {
4799 				/*
4800 				 * For now, this has no meaning.
4801 				 */
4802 				regs[rd] = 0;
4803 				break;
4804 			}
4805 
4806 			id -= DIF_VAR_OTHER_UBASE;
4807 
4808 			ASSERT(id < vstate->dtvs_nlocals);
4809 			ASSERT(vstate->dtvs_locals != NULL);
4810 
4811 			svar = vstate->dtvs_locals[id];
4812 			ASSERT(svar != NULL);
4813 			v = &svar->dtsv_var;
4814 
4815 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
4816 				uintptr_t a = (uintptr_t)svar->dtsv_data;
4817 				size_t sz = v->dtdv_type.dtdt_size;
4818 
4819 				sz += sizeof (uint64_t);
4820 				ASSERT(svar->dtsv_size == NCPU * sz);
4821 				a += CPU->cpu_id * sz;
4822 
4823 				if (*(uint8_t *)a == UINT8_MAX) {
4824 					/*
4825 					 * If the 0th byte is set to UINT8_MAX
4826 					 * then this is to be treated as a
4827 					 * reference to a NULL variable.
4828 					 */
4829 					regs[rd] = NULL;
4830 				} else {
4831 					regs[rd] = a + sizeof (uint64_t);
4832 				}
4833 
4834 				break;
4835 			}
4836 
4837 			ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
4838 			tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
4839 			regs[rd] = tmp[CPU->cpu_id];
4840 			break;
4841 
4842 		case DIF_OP_STLS:
4843 			id = DIF_INSTR_VAR(instr);
4844 
4845 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
4846 			id -= DIF_VAR_OTHER_UBASE;
4847 			ASSERT(id < vstate->dtvs_nlocals);
4848 
4849 			ASSERT(vstate->dtvs_locals != NULL);
4850 			svar = vstate->dtvs_locals[id];
4851 			ASSERT(svar != NULL);
4852 			v = &svar->dtsv_var;
4853 
4854 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
4855 				uintptr_t a = (uintptr_t)svar->dtsv_data;
4856 				size_t sz = v->dtdv_type.dtdt_size;
4857 
4858 				sz += sizeof (uint64_t);
4859 				ASSERT(svar->dtsv_size == NCPU * sz);
4860 				a += CPU->cpu_id * sz;
4861 
4862 				if (regs[rd] == NULL) {
4863 					*(uint8_t *)a = UINT8_MAX;
4864 					break;
4865 				} else {
4866 					*(uint8_t *)a = 0;
4867 					a += sizeof (uint64_t);
4868 				}
4869 
4870 				if (!dtrace_vcanload(
4871 				    (void *)(uintptr_t)regs[rd], &v->dtdv_type,
4872 				    mstate, vstate))
4873 					break;
4874 
4875 				dtrace_vcopy((void *)(uintptr_t)regs[rd],
4876 				    (void *)a, &v->dtdv_type);
4877 				break;
4878 			}
4879 
4880 			ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
4881 			tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
4882 			tmp[CPU->cpu_id] = regs[rd];
4883 			break;
4884 
4885 		case DIF_OP_LDTS: {
4886 			dtrace_dynvar_t *dvar;
4887 			dtrace_key_t *key;
4888 
4889 			id = DIF_INSTR_VAR(instr);
4890 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
4891 			id -= DIF_VAR_OTHER_UBASE;
4892 			v = &vstate->dtvs_tlocals[id];
4893 
4894 			key = &tupregs[DIF_DTR_NREGS];
4895 			key[0].dttk_value = (uint64_t)id;
4896 			key[0].dttk_size = 0;
4897 			DTRACE_TLS_THRKEY(key[1].dttk_value);
4898 			key[1].dttk_size = 0;
4899 
4900 			dvar = dtrace_dynvar(dstate, 2, key,
4901 			    sizeof (uint64_t), DTRACE_DYNVAR_NOALLOC,
4902 			    mstate, vstate);
4903 
4904 			if (dvar == NULL) {
4905 				regs[rd] = 0;
4906 				break;
4907 			}
4908 
4909 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
4910 				regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
4911 			} else {
4912 				regs[rd] = *((uint64_t *)dvar->dtdv_data);
4913 			}
4914 
4915 			break;
4916 		}
4917 
4918 		case DIF_OP_STTS: {
4919 			dtrace_dynvar_t *dvar;
4920 			dtrace_key_t *key;
4921 
4922 			id = DIF_INSTR_VAR(instr);
4923 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
4924 			id -= DIF_VAR_OTHER_UBASE;
4925 
4926 			key = &tupregs[DIF_DTR_NREGS];
4927 			key[0].dttk_value = (uint64_t)id;
4928 			key[0].dttk_size = 0;
4929 			DTRACE_TLS_THRKEY(key[1].dttk_value);
4930 			key[1].dttk_size = 0;
4931 			v = &vstate->dtvs_tlocals[id];
4932 
4933 			dvar = dtrace_dynvar(dstate, 2, key,
4934 			    v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
4935 			    v->dtdv_type.dtdt_size : sizeof (uint64_t),
4936 			    regs[rd] ? DTRACE_DYNVAR_ALLOC :
4937 			    DTRACE_DYNVAR_DEALLOC, mstate, vstate);
4938 
4939 			/*
4940 			 * Given that we're storing to thread-local data,
4941 			 * we need to flush our predicate cache.
4942 			 */
4943 			curthread->t_predcache = NULL;
4944 
4945 			if (dvar == NULL)
4946 				break;
4947 
4948 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
4949 				if (!dtrace_vcanload(
4950 				    (void *)(uintptr_t)regs[rd],
4951 				    &v->dtdv_type, mstate, vstate))
4952 					break;
4953 
4954 				dtrace_vcopy((void *)(uintptr_t)regs[rd],
4955 				    dvar->dtdv_data, &v->dtdv_type);
4956 			} else {
4957 				*((uint64_t *)dvar->dtdv_data) = regs[rd];
4958 			}
4959 
4960 			break;
4961 		}
4962 
4963 		case DIF_OP_SRA:
4964 			regs[rd] = (int64_t)regs[r1] >> regs[r2];
4965 			break;
4966 
4967 		case DIF_OP_CALL:
4968 			dtrace_dif_subr(DIF_INSTR_SUBR(instr), rd,
4969 			    regs, tupregs, ttop, mstate, state);
4970 			break;
4971 
4972 		case DIF_OP_PUSHTR:
4973 			if (ttop == DIF_DTR_NREGS) {
4974 				*flags |= CPU_DTRACE_TUPOFLOW;
4975 				break;
4976 			}
4977 
4978 			if (r1 == DIF_TYPE_STRING) {
4979 				/*
4980 				 * If this is a string type and the size is 0,
4981 				 * we'll use the system-wide default string
4982 				 * size.  Note that we are _not_ looking at
4983 				 * the value of the DTRACEOPT_STRSIZE option;
4984 				 * had this been set, we would expect to have
4985 				 * a non-zero size value in the "pushtr".
4986 				 */
4987 				tupregs[ttop].dttk_size =
4988 				    dtrace_strlen((char *)(uintptr_t)regs[rd],
4989 				    regs[r2] ? regs[r2] :
4990 				    dtrace_strsize_default) + 1;
4991 			} else {
4992 				tupregs[ttop].dttk_size = regs[r2];
4993 			}
4994 
4995 			tupregs[ttop++].dttk_value = regs[rd];
4996 			break;
4997 
4998 		case DIF_OP_PUSHTV:
4999 			if (ttop == DIF_DTR_NREGS) {
5000 				*flags |= CPU_DTRACE_TUPOFLOW;
5001 				break;
5002 			}
5003 
5004 			tupregs[ttop].dttk_value = regs[rd];
5005 			tupregs[ttop++].dttk_size = 0;
5006 			break;
5007 
5008 		case DIF_OP_POPTS:
5009 			if (ttop != 0)
5010 				ttop--;
5011 			break;
5012 
5013 		case DIF_OP_FLUSHTS:
5014 			ttop = 0;
5015 			break;
5016 
5017 		case DIF_OP_LDGAA:
5018 		case DIF_OP_LDTAA: {
5019 			dtrace_dynvar_t *dvar;
5020 			dtrace_key_t *key = tupregs;
5021 			uint_t nkeys = ttop;
5022 
5023 			id = DIF_INSTR_VAR(instr);
5024 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
5025 			id -= DIF_VAR_OTHER_UBASE;
5026 
5027 			key[nkeys].dttk_value = (uint64_t)id;
5028 			key[nkeys++].dttk_size = 0;
5029 
5030 			if (DIF_INSTR_OP(instr) == DIF_OP_LDTAA) {
5031 				DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
5032 				key[nkeys++].dttk_size = 0;
5033 				v = &vstate->dtvs_tlocals[id];
5034 			} else {
5035 				v = &vstate->dtvs_globals[id]->dtsv_var;
5036 			}
5037 
5038 			dvar = dtrace_dynvar(dstate, nkeys, key,
5039 			    v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
5040 			    v->dtdv_type.dtdt_size : sizeof (uint64_t),
5041 			    DTRACE_DYNVAR_NOALLOC, mstate, vstate);
5042 
5043 			if (dvar == NULL) {
5044 				regs[rd] = 0;
5045 				break;
5046 			}
5047 
5048 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5049 				regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
5050 			} else {
5051 				regs[rd] = *((uint64_t *)dvar->dtdv_data);
5052 			}
5053 
5054 			break;
5055 		}
5056 
5057 		case DIF_OP_STGAA:
5058 		case DIF_OP_STTAA: {
5059 			dtrace_dynvar_t *dvar;
5060 			dtrace_key_t *key = tupregs;
5061 			uint_t nkeys = ttop;
5062 
5063 			id = DIF_INSTR_VAR(instr);
5064 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
5065 			id -= DIF_VAR_OTHER_UBASE;
5066 
5067 			key[nkeys].dttk_value = (uint64_t)id;
5068 			key[nkeys++].dttk_size = 0;
5069 
5070 			if (DIF_INSTR_OP(instr) == DIF_OP_STTAA) {
5071 				DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
5072 				key[nkeys++].dttk_size = 0;
5073 				v = &vstate->dtvs_tlocals[id];
5074 			} else {
5075 				v = &vstate->dtvs_globals[id]->dtsv_var;
5076 			}
5077 
5078 			dvar = dtrace_dynvar(dstate, nkeys, key,
5079 			    v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
5080 			    v->dtdv_type.dtdt_size : sizeof (uint64_t),
5081 			    regs[rd] ? DTRACE_DYNVAR_ALLOC :
5082 			    DTRACE_DYNVAR_DEALLOC, mstate, vstate);
5083 
5084 			if (dvar == NULL)
5085 				break;
5086 
5087 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5088 				if (!dtrace_vcanload(
5089 				    (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5090 				    mstate, vstate))
5091 					break;
5092 
5093 				dtrace_vcopy((void *)(uintptr_t)regs[rd],
5094 				    dvar->dtdv_data, &v->dtdv_type);
5095 			} else {
5096 				*((uint64_t *)dvar->dtdv_data) = regs[rd];
5097 			}
5098 
5099 			break;
5100 		}
5101 
5102 		case DIF_OP_ALLOCS: {
5103 			uintptr_t ptr = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
5104 			size_t size = ptr - mstate->dtms_scratch_ptr + regs[r1];
5105 
5106 			/*
5107 			 * Rounding up the user allocation size could have
5108 			 * overflowed large, bogus allocations (like -1ULL) to
5109 			 * 0.
5110 			 */
5111 			if (size < regs[r1] ||
5112 			    !DTRACE_INSCRATCH(mstate, size)) {
5113 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5114 				regs[rd] = NULL;
5115 				break;
5116 			}
5117 
5118 			dtrace_bzero((void *) mstate->dtms_scratch_ptr, size);
5119 			mstate->dtms_scratch_ptr += size;
5120 			regs[rd] = ptr;
5121 			break;
5122 		}
5123 
5124 		case DIF_OP_COPYS:
5125 			if (!dtrace_canstore(regs[rd], regs[r2],
5126 			    mstate, vstate)) {
5127 				*flags |= CPU_DTRACE_BADADDR;
5128 				*illval = regs[rd];
5129 				break;
5130 			}
5131 
5132 			if (!dtrace_canload(regs[r1], regs[r2], mstate, vstate))
5133 				break;
5134 
5135 			dtrace_bcopy((void *)(uintptr_t)regs[r1],
5136 			    (void *)(uintptr_t)regs[rd], (size_t)regs[r2]);
5137 			break;
5138 
5139 		case DIF_OP_STB:
5140 			if (!dtrace_canstore(regs[rd], 1, mstate, vstate)) {
5141 				*flags |= CPU_DTRACE_BADADDR;
5142 				*illval = regs[rd];
5143 				break;
5144 			}
5145 			*((uint8_t *)(uintptr_t)regs[rd]) = (uint8_t)regs[r1];
5146 			break;
5147 
5148 		case DIF_OP_STH:
5149 			if (!dtrace_canstore(regs[rd], 2, mstate, vstate)) {
5150 				*flags |= CPU_DTRACE_BADADDR;
5151 				*illval = regs[rd];
5152 				break;
5153 			}
5154 			if (regs[rd] & 1) {
5155 				*flags |= CPU_DTRACE_BADALIGN;
5156 				*illval = regs[rd];
5157 				break;
5158 			}
5159 			*((uint16_t *)(uintptr_t)regs[rd]) = (uint16_t)regs[r1];
5160 			break;
5161 
5162 		case DIF_OP_STW:
5163 			if (!dtrace_canstore(regs[rd], 4, mstate, vstate)) {
5164 				*flags |= CPU_DTRACE_BADADDR;
5165 				*illval = regs[rd];
5166 				break;
5167 			}
5168 			if (regs[rd] & 3) {
5169 				*flags |= CPU_DTRACE_BADALIGN;
5170 				*illval = regs[rd];
5171 				break;
5172 			}
5173 			*((uint32_t *)(uintptr_t)regs[rd]) = (uint32_t)regs[r1];
5174 			break;
5175 
5176 		case DIF_OP_STX:
5177 			if (!dtrace_canstore(regs[rd], 8, mstate, vstate)) {
5178 				*flags |= CPU_DTRACE_BADADDR;
5179 				*illval = regs[rd];
5180 				break;
5181 			}
5182 			if (regs[rd] & 7) {
5183 				*flags |= CPU_DTRACE_BADALIGN;
5184 				*illval = regs[rd];
5185 				break;
5186 			}
5187 			*((uint64_t *)(uintptr_t)regs[rd]) = regs[r1];
5188 			break;
5189 		}
5190 	}
5191 
5192 	if (!(*flags & CPU_DTRACE_FAULT))
5193 		return (rval);
5194 
5195 	mstate->dtms_fltoffs = opc * sizeof (dif_instr_t);
5196 	mstate->dtms_present |= DTRACE_MSTATE_FLTOFFS;
5197 
5198 	return (0);
5199 }
5200 
5201 static void
5202 dtrace_action_breakpoint(dtrace_ecb_t *ecb)
5203 {
5204 	dtrace_probe_t *probe = ecb->dte_probe;
5205 	dtrace_provider_t *prov = probe->dtpr_provider;
5206 	char c[DTRACE_FULLNAMELEN + 80], *str;
5207 	char *msg = "dtrace: breakpoint action at probe ";
5208 	char *ecbmsg = " (ecb ";
5209 	uintptr_t mask = (0xf << (sizeof (uintptr_t) * NBBY / 4));
5210 	uintptr_t val = (uintptr_t)ecb;
5211 	int shift = (sizeof (uintptr_t) * NBBY) - 4, i = 0;
5212 
5213 	if (dtrace_destructive_disallow)
5214 		return;
5215 
5216 	/*
5217 	 * It's impossible to be taking action on the NULL probe.
5218 	 */
5219 	ASSERT(probe != NULL);
5220 
5221 	/*
5222 	 * This is a poor man's (destitute man's?) sprintf():  we want to
5223 	 * print the provider name, module name, function name and name of
5224 	 * the probe, along with the hex address of the ECB with the breakpoint
5225 	 * action -- all of which we must place in the character buffer by
5226 	 * hand.
5227 	 */
5228 	while (*msg != '\0')
5229 		c[i++] = *msg++;
5230 
5231 	for (str = prov->dtpv_name; *str != '\0'; str++)
5232 		c[i++] = *str;
5233 	c[i++] = ':';
5234 
5235 	for (str = probe->dtpr_mod; *str != '\0'; str++)
5236 		c[i++] = *str;
5237 	c[i++] = ':';
5238 
5239 	for (str = probe->dtpr_func; *str != '\0'; str++)
5240 		c[i++] = *str;
5241 	c[i++] = ':';
5242 
5243 	for (str = probe->dtpr_name; *str != '\0'; str++)
5244 		c[i++] = *str;
5245 
5246 	while (*ecbmsg != '\0')
5247 		c[i++] = *ecbmsg++;
5248 
5249 	while (shift >= 0) {
5250 		mask = (uintptr_t)0xf << shift;
5251 
5252 		if (val >= ((uintptr_t)1 << shift))
5253 			c[i++] = "0123456789abcdef"[(val & mask) >> shift];
5254 		shift -= 4;
5255 	}
5256 
5257 	c[i++] = ')';
5258 	c[i] = '\0';
5259 
5260 	debug_enter(c);
5261 }
5262 
5263 static void
5264 dtrace_action_panic(dtrace_ecb_t *ecb)
5265 {
5266 	dtrace_probe_t *probe = ecb->dte_probe;
5267 
5268 	/*
5269 	 * It's impossible to be taking action on the NULL probe.
5270 	 */
5271 	ASSERT(probe != NULL);
5272 
5273 	if (dtrace_destructive_disallow)
5274 		return;
5275 
5276 	if (dtrace_panicked != NULL)
5277 		return;
5278 
5279 	if (dtrace_casptr(&dtrace_panicked, NULL, curthread) != NULL)
5280 		return;
5281 
5282 	/*
5283 	 * We won the right to panic.  (We want to be sure that only one
5284 	 * thread calls panic() from dtrace_probe(), and that panic() is
5285 	 * called exactly once.)
5286 	 */
5287 	dtrace_panic("dtrace: panic action at probe %s:%s:%s:%s (ecb %p)",
5288 	    probe->dtpr_provider->dtpv_name, probe->dtpr_mod,
5289 	    probe->dtpr_func, probe->dtpr_name, (void *)ecb);
5290 }
5291 
5292 static void
5293 dtrace_action_raise(uint64_t sig)
5294 {
5295 	if (dtrace_destructive_disallow)
5296 		return;
5297 
5298 	if (sig >= NSIG) {
5299 		DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
5300 		return;
5301 	}
5302 
5303 	/*
5304 	 * raise() has a queue depth of 1 -- we ignore all subsequent
5305 	 * invocations of the raise() action.
5306 	 */
5307 	if (curthread->t_dtrace_sig == 0)
5308 		curthread->t_dtrace_sig = (uint8_t)sig;
5309 
5310 	curthread->t_sig_check = 1;
5311 	aston(curthread);
5312 }
5313 
5314 static void
5315 dtrace_action_stop(void)
5316 {
5317 	if (dtrace_destructive_disallow)
5318 		return;
5319 
5320 	if (!curthread->t_dtrace_stop) {
5321 		curthread->t_dtrace_stop = 1;
5322 		curthread->t_sig_check = 1;
5323 		aston(curthread);
5324 	}
5325 }
5326 
5327 static void
5328 dtrace_action_chill(dtrace_mstate_t *mstate, hrtime_t val)
5329 {
5330 	hrtime_t now;
5331 	volatile uint16_t *flags;
5332 	cpu_t *cpu = CPU;
5333 
5334 	if (dtrace_destructive_disallow)
5335 		return;
5336 
5337 	flags = (volatile uint16_t *)&cpu_core[cpu->cpu_id].cpuc_dtrace_flags;
5338 
5339 	now = dtrace_gethrtime();
5340 
5341 	if (now - cpu->cpu_dtrace_chillmark > dtrace_chill_interval) {
5342 		/*
5343 		 * We need to advance the mark to the current time.
5344 		 */
5345 		cpu->cpu_dtrace_chillmark = now;
5346 		cpu->cpu_dtrace_chilled = 0;
5347 	}
5348 
5349 	/*
5350 	 * Now check to see if the requested chill time would take us over
5351 	 * the maximum amount of time allowed in the chill interval.  (Or
5352 	 * worse, if the calculation itself induces overflow.)
5353 	 */
5354 	if (cpu->cpu_dtrace_chilled + val > dtrace_chill_max ||
5355 	    cpu->cpu_dtrace_chilled + val < cpu->cpu_dtrace_chilled) {
5356 		*flags |= CPU_DTRACE_ILLOP;
5357 		return;
5358 	}
5359 
5360 	while (dtrace_gethrtime() - now < val)
5361 		continue;
5362 
5363 	/*
5364 	 * Normally, we assure that the value of the variable "timestamp" does
5365 	 * not change within an ECB.  The presence of chill() represents an
5366 	 * exception to this rule, however.
5367 	 */
5368 	mstate->dtms_present &= ~DTRACE_MSTATE_TIMESTAMP;
5369 	cpu->cpu_dtrace_chilled += val;
5370 }
5371 
5372 static void
5373 dtrace_action_ustack(dtrace_mstate_t *mstate, dtrace_state_t *state,
5374     uint64_t *buf, uint64_t arg)
5375 {
5376 	int nframes = DTRACE_USTACK_NFRAMES(arg);
5377 	int strsize = DTRACE_USTACK_STRSIZE(arg);
5378 	uint64_t *pcs = &buf[1], *fps;
5379 	char *str = (char *)&pcs[nframes];
5380 	int size, offs = 0, i, j;
5381 	uintptr_t old = mstate->dtms_scratch_ptr, saved;
5382 	uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
5383 	char *sym;
5384 
5385 	/*
5386 	 * Should be taking a faster path if string space has not been
5387 	 * allocated.
5388 	 */
5389 	ASSERT(strsize != 0);
5390 
5391 	/*
5392 	 * We will first allocate some temporary space for the frame pointers.
5393 	 */
5394 	fps = (uint64_t *)P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
5395 	size = (uintptr_t)fps - mstate->dtms_scratch_ptr +
5396 	    (nframes * sizeof (uint64_t));
5397 
5398 	if (!DTRACE_INSCRATCH(mstate, size)) {
5399 		/*
5400 		 * Not enough room for our frame pointers -- need to indicate
5401 		 * that we ran out of scratch space.
5402 		 */
5403 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5404 		return;
5405 	}
5406 
5407 	mstate->dtms_scratch_ptr += size;
5408 	saved = mstate->dtms_scratch_ptr;
5409 
5410 	/*
5411 	 * Now get a stack with both program counters and frame pointers.
5412 	 */
5413 	DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5414 	dtrace_getufpstack(buf, fps, nframes + 1);
5415 	DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5416 
5417 	/*
5418 	 * If that faulted, we're cooked.
5419 	 */
5420 	if (*flags & CPU_DTRACE_FAULT)
5421 		goto out;
5422 
5423 	/*
5424 	 * Now we want to walk up the stack, calling the USTACK helper.  For
5425 	 * each iteration, we restore the scratch pointer.
5426 	 */
5427 	for (i = 0; i < nframes; i++) {
5428 		mstate->dtms_scratch_ptr = saved;
5429 
5430 		if (offs >= strsize)
5431 			break;
5432 
5433 		sym = (char *)(uintptr_t)dtrace_helper(
5434 		    DTRACE_HELPER_ACTION_USTACK,
5435 		    mstate, state, pcs[i], fps[i]);
5436 
5437 		/*
5438 		 * If we faulted while running the helper, we're going to
5439 		 * clear the fault and null out the corresponding string.
5440 		 */
5441 		if (*flags & CPU_DTRACE_FAULT) {
5442 			*flags &= ~CPU_DTRACE_FAULT;
5443 			str[offs++] = '\0';
5444 			continue;
5445 		}
5446 
5447 		if (sym == NULL) {
5448 			str[offs++] = '\0';
5449 			continue;
5450 		}
5451 
5452 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5453 
5454 		/*
5455 		 * Now copy in the string that the helper returned to us.
5456 		 */
5457 		for (j = 0; offs + j < strsize; j++) {
5458 			if ((str[offs + j] = sym[j]) == '\0')
5459 				break;
5460 		}
5461 
5462 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5463 
5464 		offs += j + 1;
5465 	}
5466 
5467 	if (offs >= strsize) {
5468 		/*
5469 		 * If we didn't have room for all of the strings, we don't
5470 		 * abort processing -- this needn't be a fatal error -- but we
5471 		 * still want to increment a counter (dts_stkstroverflows) to
5472 		 * allow this condition to be warned about.  (If this is from
5473 		 * a jstack() action, it is easily tuned via jstackstrsize.)
5474 		 */
5475 		dtrace_error(&state->dts_stkstroverflows);
5476 	}
5477 
5478 	while (offs < strsize)
5479 		str[offs++] = '\0';
5480 
5481 out:
5482 	mstate->dtms_scratch_ptr = old;
5483 }
5484 
5485 /*
5486  * If you're looking for the epicenter of DTrace, you just found it.  This
5487  * is the function called by the provider to fire a probe -- from which all
5488  * subsequent probe-context DTrace activity emanates.
5489  */
5490 void
5491 dtrace_probe(dtrace_id_t id, uintptr_t arg0, uintptr_t arg1,
5492     uintptr_t arg2, uintptr_t arg3, uintptr_t arg4)
5493 {
5494 	processorid_t cpuid;
5495 	dtrace_icookie_t cookie;
5496 	dtrace_probe_t *probe;
5497 	dtrace_mstate_t mstate;
5498 	dtrace_ecb_t *ecb;
5499 	dtrace_action_t *act;
5500 	intptr_t offs;
5501 	size_t size;
5502 	int vtime, onintr;
5503 	volatile uint16_t *flags;
5504 	hrtime_t now;
5505 
5506 	/*
5507 	 * Kick out immediately if this CPU is still being born (in which case
5508 	 * curthread will be set to -1) or the current thread can't allow
5509 	 * probes in its current context.
5510 	 */
5511 	if (((uintptr_t)curthread & 1) || (curthread->t_flag & T_DONTDTRACE))
5512 		return;
5513 
5514 	cookie = dtrace_interrupt_disable();
5515 	probe = dtrace_probes[id - 1];
5516 	cpuid = CPU->cpu_id;
5517 	onintr = CPU_ON_INTR(CPU);
5518 
5519 	if (!onintr && probe->dtpr_predcache != DTRACE_CACHEIDNONE &&
5520 	    probe->dtpr_predcache == curthread->t_predcache) {
5521 		/*
5522 		 * We have hit in the predicate cache; we know that
5523 		 * this predicate would evaluate to be false.
5524 		 */
5525 		dtrace_interrupt_enable(cookie);
5526 		return;
5527 	}
5528 
5529 	if (panic_quiesce) {
5530 		/*
5531 		 * We don't trace anything if we're panicking.
5532 		 */
5533 		dtrace_interrupt_enable(cookie);
5534 		return;
5535 	}
5536 
5537 	now = dtrace_gethrtime();
5538 	vtime = dtrace_vtime_references != 0;
5539 
5540 	if (vtime && curthread->t_dtrace_start)
5541 		curthread->t_dtrace_vtime += now - curthread->t_dtrace_start;
5542 
5543 	mstate.dtms_difo = NULL;
5544 	mstate.dtms_probe = probe;
5545 	mstate.dtms_strtok = NULL;
5546 	mstate.dtms_arg[0] = arg0;
5547 	mstate.dtms_arg[1] = arg1;
5548 	mstate.dtms_arg[2] = arg2;
5549 	mstate.dtms_arg[3] = arg3;
5550 	mstate.dtms_arg[4] = arg4;
5551 
5552 	flags = (volatile uint16_t *)&cpu_core[cpuid].cpuc_dtrace_flags;
5553 
5554 	for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
5555 		dtrace_predicate_t *pred = ecb->dte_predicate;
5556 		dtrace_state_t *state = ecb->dte_state;
5557 		dtrace_buffer_t *buf = &state->dts_buffer[cpuid];
5558 		dtrace_buffer_t *aggbuf = &state->dts_aggbuffer[cpuid];
5559 		dtrace_vstate_t *vstate = &state->dts_vstate;
5560 		dtrace_provider_t *prov = probe->dtpr_provider;
5561 		int committed = 0;
5562 		caddr_t tomax;
5563 
5564 		/*
5565 		 * A little subtlety with the following (seemingly innocuous)
5566 		 * declaration of the automatic 'val':  by looking at the
5567 		 * code, you might think that it could be declared in the
5568 		 * action processing loop, below.  (That is, it's only used in
5569 		 * the action processing loop.)  However, it must be declared
5570 		 * out of that scope because in the case of DIF expression
5571 		 * arguments to aggregating actions, one iteration of the
5572 		 * action loop will use the last iteration's value.
5573 		 */
5574 #ifdef lint
5575 		uint64_t val = 0;
5576 #else
5577 		uint64_t val;
5578 #endif
5579 
5580 		mstate.dtms_present = DTRACE_MSTATE_ARGS | DTRACE_MSTATE_PROBE;
5581 		*flags &= ~CPU_DTRACE_ERROR;
5582 
5583 		if (prov == dtrace_provider) {
5584 			/*
5585 			 * If dtrace itself is the provider of this probe,
5586 			 * we're only going to continue processing the ECB if
5587 			 * arg0 (the dtrace_state_t) is equal to the ECB's
5588 			 * creating state.  (This prevents disjoint consumers
5589 			 * from seeing one another's metaprobes.)
5590 			 */
5591 			if (arg0 != (uint64_t)(uintptr_t)state)
5592 				continue;
5593 		}
5594 
5595 		if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE) {
5596 			/*
5597 			 * We're not currently active.  If our provider isn't
5598 			 * the dtrace pseudo provider, we're not interested.
5599 			 */
5600 			if (prov != dtrace_provider)
5601 				continue;
5602 
5603 			/*
5604 			 * Now we must further check if we are in the BEGIN
5605 			 * probe.  If we are, we will only continue processing
5606 			 * if we're still in WARMUP -- if one BEGIN enabling
5607 			 * has invoked the exit() action, we don't want to
5608 			 * evaluate subsequent BEGIN enablings.
5609 			 */
5610 			if (probe->dtpr_id == dtrace_probeid_begin &&
5611 			    state->dts_activity != DTRACE_ACTIVITY_WARMUP) {
5612 				ASSERT(state->dts_activity ==
5613 				    DTRACE_ACTIVITY_DRAINING);
5614 				continue;
5615 			}
5616 		}
5617 
5618 		if (ecb->dte_cond) {
5619 			/*
5620 			 * If the dte_cond bits indicate that this
5621 			 * consumer is only allowed to see user-mode firings
5622 			 * of this probe, call the provider's dtps_usermode()
5623 			 * entry point to check that the probe was fired
5624 			 * while in a user context. Skip this ECB if that's
5625 			 * not the case.
5626 			 */
5627 			if ((ecb->dte_cond & DTRACE_COND_USERMODE) &&
5628 			    prov->dtpv_pops.dtps_usermode(prov->dtpv_arg,
5629 			    probe->dtpr_id, probe->dtpr_arg) == 0)
5630 				continue;
5631 
5632 			/*
5633 			 * This is more subtle than it looks. We have to be
5634 			 * absolutely certain that CRED() isn't going to
5635 			 * change out from under us so it's only legit to
5636 			 * examine that structure if we're in constrained
5637 			 * situations. Currently, the only times we'll this
5638 			 * check is if a non-super-user has enabled the
5639 			 * profile or syscall providers -- providers that
5640 			 * allow visibility of all processes. For the
5641 			 * profile case, the check above will ensure that
5642 			 * we're examining a user context.
5643 			 */
5644 			if (ecb->dte_cond & DTRACE_COND_OWNER) {
5645 				cred_t *cr;
5646 				cred_t *s_cr =
5647 				    ecb->dte_state->dts_cred.dcr_cred;
5648 				proc_t *proc;
5649 
5650 				ASSERT(s_cr != NULL);
5651 
5652 				if ((cr = CRED()) == NULL ||
5653 				    s_cr->cr_uid != cr->cr_uid ||
5654 				    s_cr->cr_uid != cr->cr_ruid ||
5655 				    s_cr->cr_uid != cr->cr_suid ||
5656 				    s_cr->cr_gid != cr->cr_gid ||
5657 				    s_cr->cr_gid != cr->cr_rgid ||
5658 				    s_cr->cr_gid != cr->cr_sgid ||
5659 				    (proc = ttoproc(curthread)) == NULL ||
5660 				    (proc->p_flag & SNOCD))
5661 					continue;
5662 			}
5663 
5664 			if (ecb->dte_cond & DTRACE_COND_ZONEOWNER) {
5665 				cred_t *cr;
5666 				cred_t *s_cr =
5667 				    ecb->dte_state->dts_cred.dcr_cred;
5668 
5669 				ASSERT(s_cr != NULL);
5670 
5671 				if ((cr = CRED()) == NULL ||
5672 				    s_cr->cr_zone->zone_id !=
5673 				    cr->cr_zone->zone_id)
5674 					continue;
5675 			}
5676 		}
5677 
5678 		if (now - state->dts_alive > dtrace_deadman_timeout) {
5679 			/*
5680 			 * We seem to be dead.  Unless we (a) have kernel
5681 			 * destructive permissions (b) have expicitly enabled
5682 			 * destructive actions and (c) destructive actions have
5683 			 * not been disabled, we're going to transition into
5684 			 * the KILLED state, from which no further processing
5685 			 * on this state will be performed.
5686 			 */
5687 			if (!dtrace_priv_kernel_destructive(state) ||
5688 			    !state->dts_cred.dcr_destructive ||
5689 			    dtrace_destructive_disallow) {
5690 				void *activity = &state->dts_activity;
5691 				dtrace_activity_t current;
5692 
5693 				do {
5694 					current = state->dts_activity;
5695 				} while (dtrace_cas32(activity, current,
5696 				    DTRACE_ACTIVITY_KILLED) != current);
5697 
5698 				continue;
5699 			}
5700 		}
5701 
5702 		if ((offs = dtrace_buffer_reserve(buf, ecb->dte_needed,
5703 		    ecb->dte_alignment, state, &mstate)) < 0)
5704 			continue;
5705 
5706 		tomax = buf->dtb_tomax;
5707 		ASSERT(tomax != NULL);
5708 
5709 		if (ecb->dte_size != 0)
5710 			DTRACE_STORE(uint32_t, tomax, offs, ecb->dte_epid);
5711 
5712 		mstate.dtms_epid = ecb->dte_epid;
5713 		mstate.dtms_present |= DTRACE_MSTATE_EPID;
5714 
5715 		if (state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)
5716 			mstate.dtms_access = DTRACE_ACCESS_KERNEL;
5717 		else
5718 			mstate.dtms_access = 0;
5719 
5720 		if (pred != NULL) {
5721 			dtrace_difo_t *dp = pred->dtp_difo;
5722 			int rval;
5723 
5724 			rval = dtrace_dif_emulate(dp, &mstate, vstate, state);
5725 
5726 			if (!(*flags & CPU_DTRACE_ERROR) && !rval) {
5727 				dtrace_cacheid_t cid = probe->dtpr_predcache;
5728 
5729 				if (cid != DTRACE_CACHEIDNONE && !onintr) {
5730 					/*
5731 					 * Update the predicate cache...
5732 					 */
5733 					ASSERT(cid == pred->dtp_cacheid);
5734 					curthread->t_predcache = cid;
5735 				}
5736 
5737 				continue;
5738 			}
5739 		}
5740 
5741 		for (act = ecb->dte_action; !(*flags & CPU_DTRACE_ERROR) &&
5742 		    act != NULL; act = act->dta_next) {
5743 			size_t valoffs;
5744 			dtrace_difo_t *dp;
5745 			dtrace_recdesc_t *rec = &act->dta_rec;
5746 
5747 			size = rec->dtrd_size;
5748 			valoffs = offs + rec->dtrd_offset;
5749 
5750 			if (DTRACEACT_ISAGG(act->dta_kind)) {
5751 				uint64_t v = 0xbad;
5752 				dtrace_aggregation_t *agg;
5753 
5754 				agg = (dtrace_aggregation_t *)act;
5755 
5756 				if ((dp = act->dta_difo) != NULL)
5757 					v = dtrace_dif_emulate(dp,
5758 					    &mstate, vstate, state);
5759 
5760 				if (*flags & CPU_DTRACE_ERROR)
5761 					continue;
5762 
5763 				/*
5764 				 * Note that we always pass the expression
5765 				 * value from the previous iteration of the
5766 				 * action loop.  This value will only be used
5767 				 * if there is an expression argument to the
5768 				 * aggregating action, denoted by the
5769 				 * dtag_hasarg field.
5770 				 */
5771 				dtrace_aggregate(agg, buf,
5772 				    offs, aggbuf, v, val);
5773 				continue;
5774 			}
5775 
5776 			switch (act->dta_kind) {
5777 			case DTRACEACT_STOP:
5778 				if (dtrace_priv_proc_destructive(state))
5779 					dtrace_action_stop();
5780 				continue;
5781 
5782 			case DTRACEACT_BREAKPOINT:
5783 				if (dtrace_priv_kernel_destructive(state))
5784 					dtrace_action_breakpoint(ecb);
5785 				continue;
5786 
5787 			case DTRACEACT_PANIC:
5788 				if (dtrace_priv_kernel_destructive(state))
5789 					dtrace_action_panic(ecb);
5790 				continue;
5791 
5792 			case DTRACEACT_STACK:
5793 				if (!dtrace_priv_kernel(state))
5794 					continue;
5795 
5796 				dtrace_getpcstack((pc_t *)(tomax + valoffs),
5797 				    size / sizeof (pc_t), probe->dtpr_aframes,
5798 				    DTRACE_ANCHORED(probe) ? NULL :
5799 				    (uint32_t *)arg0);
5800 
5801 				continue;
5802 
5803 			case DTRACEACT_JSTACK:
5804 			case DTRACEACT_USTACK:
5805 				if (!dtrace_priv_proc(state))
5806 					continue;
5807 
5808 				/*
5809 				 * See comment in DIF_VAR_PID.
5810 				 */
5811 				if (DTRACE_ANCHORED(mstate.dtms_probe) &&
5812 				    CPU_ON_INTR(CPU)) {
5813 					int depth = DTRACE_USTACK_NFRAMES(
5814 					    rec->dtrd_arg) + 1;
5815 
5816 					dtrace_bzero((void *)(tomax + valoffs),
5817 					    DTRACE_USTACK_STRSIZE(rec->dtrd_arg)
5818 					    + depth * sizeof (uint64_t));
5819 
5820 					continue;
5821 				}
5822 
5823 				if (DTRACE_USTACK_STRSIZE(rec->dtrd_arg) != 0 &&
5824 				    curproc->p_dtrace_helpers != NULL) {
5825 					/*
5826 					 * This is the slow path -- we have
5827 					 * allocated string space, and we're
5828 					 * getting the stack of a process that
5829 					 * has helpers.  Call into a separate
5830 					 * routine to perform this processing.
5831 					 */
5832 					dtrace_action_ustack(&mstate, state,
5833 					    (uint64_t *)(tomax + valoffs),
5834 					    rec->dtrd_arg);
5835 					continue;
5836 				}
5837 
5838 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5839 				dtrace_getupcstack((uint64_t *)
5840 				    (tomax + valoffs),
5841 				    DTRACE_USTACK_NFRAMES(rec->dtrd_arg) + 1);
5842 				DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5843 				continue;
5844 
5845 			default:
5846 				break;
5847 			}
5848 
5849 			dp = act->dta_difo;
5850 			ASSERT(dp != NULL);
5851 
5852 			val = dtrace_dif_emulate(dp, &mstate, vstate, state);
5853 
5854 			if (*flags & CPU_DTRACE_ERROR)
5855 				continue;
5856 
5857 			switch (act->dta_kind) {
5858 			case DTRACEACT_SPECULATE:
5859 				ASSERT(buf == &state->dts_buffer[cpuid]);
5860 				buf = dtrace_speculation_buffer(state,
5861 				    cpuid, val);
5862 
5863 				if (buf == NULL) {
5864 					*flags |= CPU_DTRACE_DROP;
5865 					continue;
5866 				}
5867 
5868 				offs = dtrace_buffer_reserve(buf,
5869 				    ecb->dte_needed, ecb->dte_alignment,
5870 				    state, NULL);
5871 
5872 				if (offs < 0) {
5873 					*flags |= CPU_DTRACE_DROP;
5874 					continue;
5875 				}
5876 
5877 				tomax = buf->dtb_tomax;
5878 				ASSERT(tomax != NULL);
5879 
5880 				if (ecb->dte_size != 0)
5881 					DTRACE_STORE(uint32_t, tomax, offs,
5882 					    ecb->dte_epid);
5883 				continue;
5884 
5885 			case DTRACEACT_CHILL:
5886 				if (dtrace_priv_kernel_destructive(state))
5887 					dtrace_action_chill(&mstate, val);
5888 				continue;
5889 
5890 			case DTRACEACT_RAISE:
5891 				if (dtrace_priv_proc_destructive(state))
5892 					dtrace_action_raise(val);
5893 				continue;
5894 
5895 			case DTRACEACT_COMMIT:
5896 				ASSERT(!committed);
5897 
5898 				/*
5899 				 * We need to commit our buffer state.
5900 				 */
5901 				if (ecb->dte_size)
5902 					buf->dtb_offset = offs + ecb->dte_size;
5903 				buf = &state->dts_buffer[cpuid];
5904 				dtrace_speculation_commit(state, cpuid, val);
5905 				committed = 1;
5906 				continue;
5907 
5908 			case DTRACEACT_DISCARD:
5909 				dtrace_speculation_discard(state, cpuid, val);
5910 				continue;
5911 
5912 			case DTRACEACT_DIFEXPR:
5913 			case DTRACEACT_LIBACT:
5914 			case DTRACEACT_PRINTF:
5915 			case DTRACEACT_PRINTA:
5916 			case DTRACEACT_SYSTEM:
5917 			case DTRACEACT_FREOPEN:
5918 				break;
5919 
5920 			case DTRACEACT_SYM:
5921 			case DTRACEACT_MOD:
5922 				if (!dtrace_priv_kernel(state))
5923 					continue;
5924 				break;
5925 
5926 			case DTRACEACT_USYM:
5927 			case DTRACEACT_UMOD:
5928 			case DTRACEACT_UADDR: {
5929 				struct pid *pid = curthread->t_procp->p_pidp;
5930 
5931 				if (!dtrace_priv_proc(state))
5932 					continue;
5933 
5934 				DTRACE_STORE(uint64_t, tomax,
5935 				    valoffs, (uint64_t)pid->pid_id);
5936 				DTRACE_STORE(uint64_t, tomax,
5937 				    valoffs + sizeof (uint64_t), val);
5938 
5939 				continue;
5940 			}
5941 
5942 			case DTRACEACT_EXIT: {
5943 				/*
5944 				 * For the exit action, we are going to attempt
5945 				 * to atomically set our activity to be
5946 				 * draining.  If this fails (either because
5947 				 * another CPU has beat us to the exit action,
5948 				 * or because our current activity is something
5949 				 * other than ACTIVE or WARMUP), we will
5950 				 * continue.  This assures that the exit action
5951 				 * can be successfully recorded at most once
5952 				 * when we're in the ACTIVE state.  If we're
5953 				 * encountering the exit() action while in
5954 				 * COOLDOWN, however, we want to honor the new
5955 				 * status code.  (We know that we're the only
5956 				 * thread in COOLDOWN, so there is no race.)
5957 				 */
5958 				void *activity = &state->dts_activity;
5959 				dtrace_activity_t current = state->dts_activity;
5960 
5961 				if (current == DTRACE_ACTIVITY_COOLDOWN)
5962 					break;
5963 
5964 				if (current != DTRACE_ACTIVITY_WARMUP)
5965 					current = DTRACE_ACTIVITY_ACTIVE;
5966 
5967 				if (dtrace_cas32(activity, current,
5968 				    DTRACE_ACTIVITY_DRAINING) != current) {
5969 					*flags |= CPU_DTRACE_DROP;
5970 					continue;
5971 				}
5972 
5973 				break;
5974 			}
5975 
5976 			default:
5977 				ASSERT(0);
5978 			}
5979 
5980 			if (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF) {
5981 				uintptr_t end = valoffs + size;
5982 
5983 				if (!dtrace_vcanload((void *)(uintptr_t)val,
5984 				    &dp->dtdo_rtype, &mstate, vstate))
5985 					continue;
5986 
5987 				/*
5988 				 * If this is a string, we're going to only
5989 				 * load until we find the zero byte -- after
5990 				 * which we'll store zero bytes.
5991 				 */
5992 				if (dp->dtdo_rtype.dtdt_kind ==
5993 				    DIF_TYPE_STRING) {
5994 					char c = '\0' + 1;
5995 					int intuple = act->dta_intuple;
5996 					size_t s;
5997 
5998 					for (s = 0; s < size; s++) {
5999 						if (c != '\0')
6000 							c = dtrace_load8(val++);
6001 
6002 						DTRACE_STORE(uint8_t, tomax,
6003 						    valoffs++, c);
6004 
6005 						if (c == '\0' && intuple)
6006 							break;
6007 					}
6008 
6009 					continue;
6010 				}
6011 
6012 				while (valoffs < end) {
6013 					DTRACE_STORE(uint8_t, tomax, valoffs++,
6014 					    dtrace_load8(val++));
6015 				}
6016 
6017 				continue;
6018 			}
6019 
6020 			switch (size) {
6021 			case 0:
6022 				break;
6023 
6024 			case sizeof (uint8_t):
6025 				DTRACE_STORE(uint8_t, tomax, valoffs, val);
6026 				break;
6027 			case sizeof (uint16_t):
6028 				DTRACE_STORE(uint16_t, tomax, valoffs, val);
6029 				break;
6030 			case sizeof (uint32_t):
6031 				DTRACE_STORE(uint32_t, tomax, valoffs, val);
6032 				break;
6033 			case sizeof (uint64_t):
6034 				DTRACE_STORE(uint64_t, tomax, valoffs, val);
6035 				break;
6036 			default:
6037 				/*
6038 				 * Any other size should have been returned by
6039 				 * reference, not by value.
6040 				 */
6041 				ASSERT(0);
6042 				break;
6043 			}
6044 		}
6045 
6046 		if (*flags & CPU_DTRACE_DROP)
6047 			continue;
6048 
6049 		if (*flags & CPU_DTRACE_FAULT) {
6050 			int ndx;
6051 			dtrace_action_t *err;
6052 
6053 			buf->dtb_errors++;
6054 
6055 			if (probe->dtpr_id == dtrace_probeid_error) {
6056 				/*
6057 				 * There's nothing we can do -- we had an
6058 				 * error on the error probe.  We bump an
6059 				 * error counter to at least indicate that
6060 				 * this condition happened.
6061 				 */
6062 				dtrace_error(&state->dts_dblerrors);
6063 				continue;
6064 			}
6065 
6066 			if (vtime) {
6067 				/*
6068 				 * Before recursing on dtrace_probe(), we
6069 				 * need to explicitly clear out our start
6070 				 * time to prevent it from being accumulated
6071 				 * into t_dtrace_vtime.
6072 				 */
6073 				curthread->t_dtrace_start = 0;
6074 			}
6075 
6076 			/*
6077 			 * Iterate over the actions to figure out which action
6078 			 * we were processing when we experienced the error.
6079 			 * Note that act points _past_ the faulting action; if
6080 			 * act is ecb->dte_action, the fault was in the
6081 			 * predicate, if it's ecb->dte_action->dta_next it's
6082 			 * in action #1, and so on.
6083 			 */
6084 			for (err = ecb->dte_action, ndx = 0;
6085 			    err != act; err = err->dta_next, ndx++)
6086 				continue;
6087 
6088 			dtrace_probe_error(state, ecb->dte_epid, ndx,
6089 			    (mstate.dtms_present & DTRACE_MSTATE_FLTOFFS) ?
6090 			    mstate.dtms_fltoffs : -1, DTRACE_FLAGS2FLT(*flags),
6091 			    cpu_core[cpuid].cpuc_dtrace_illval);
6092 
6093 			continue;
6094 		}
6095 
6096 		if (!committed)
6097 			buf->dtb_offset = offs + ecb->dte_size;
6098 	}
6099 
6100 	if (vtime)
6101 		curthread->t_dtrace_start = dtrace_gethrtime();
6102 
6103 	dtrace_interrupt_enable(cookie);
6104 }
6105 
6106 /*
6107  * DTrace Probe Hashing Functions
6108  *
6109  * The functions in this section (and indeed, the functions in remaining
6110  * sections) are not _called_ from probe context.  (Any exceptions to this are
6111  * marked with a "Note:".)  Rather, they are called from elsewhere in the
6112  * DTrace framework to look-up probes in, add probes to and remove probes from
6113  * the DTrace probe hashes.  (Each probe is hashed by each element of the
6114  * probe tuple -- allowing for fast lookups, regardless of what was
6115  * specified.)
6116  */
6117 static uint_t
6118 dtrace_hash_str(char *p)
6119 {
6120 	unsigned int g;
6121 	uint_t hval = 0;
6122 
6123 	while (*p) {
6124 		hval = (hval << 4) + *p++;
6125 		if ((g = (hval & 0xf0000000)) != 0)
6126 			hval ^= g >> 24;
6127 		hval &= ~g;
6128 	}
6129 	return (hval);
6130 }
6131 
6132 static dtrace_hash_t *
6133 dtrace_hash_create(uintptr_t stroffs, uintptr_t nextoffs, uintptr_t prevoffs)
6134 {
6135 	dtrace_hash_t *hash = kmem_zalloc(sizeof (dtrace_hash_t), KM_SLEEP);
6136 
6137 	hash->dth_stroffs = stroffs;
6138 	hash->dth_nextoffs = nextoffs;
6139 	hash->dth_prevoffs = prevoffs;
6140 
6141 	hash->dth_size = 1;
6142 	hash->dth_mask = hash->dth_size - 1;
6143 
6144 	hash->dth_tab = kmem_zalloc(hash->dth_size *
6145 	    sizeof (dtrace_hashbucket_t *), KM_SLEEP);
6146 
6147 	return (hash);
6148 }
6149 
6150 static void
6151 dtrace_hash_destroy(dtrace_hash_t *hash)
6152 {
6153 #ifdef DEBUG
6154 	int i;
6155 
6156 	for (i = 0; i < hash->dth_size; i++)
6157 		ASSERT(hash->dth_tab[i] == NULL);
6158 #endif
6159 
6160 	kmem_free(hash->dth_tab,
6161 	    hash->dth_size * sizeof (dtrace_hashbucket_t *));
6162 	kmem_free(hash, sizeof (dtrace_hash_t));
6163 }
6164 
6165 static void
6166 dtrace_hash_resize(dtrace_hash_t *hash)
6167 {
6168 	int size = hash->dth_size, i, ndx;
6169 	int new_size = hash->dth_size << 1;
6170 	int new_mask = new_size - 1;
6171 	dtrace_hashbucket_t **new_tab, *bucket, *next;
6172 
6173 	ASSERT((new_size & new_mask) == 0);
6174 
6175 	new_tab = kmem_zalloc(new_size * sizeof (void *), KM_SLEEP);
6176 
6177 	for (i = 0; i < size; i++) {
6178 		for (bucket = hash->dth_tab[i]; bucket != NULL; bucket = next) {
6179 			dtrace_probe_t *probe = bucket->dthb_chain;
6180 
6181 			ASSERT(probe != NULL);
6182 			ndx = DTRACE_HASHSTR(hash, probe) & new_mask;
6183 
6184 			next = bucket->dthb_next;
6185 			bucket->dthb_next = new_tab[ndx];
6186 			new_tab[ndx] = bucket;
6187 		}
6188 	}
6189 
6190 	kmem_free(hash->dth_tab, hash->dth_size * sizeof (void *));
6191 	hash->dth_tab = new_tab;
6192 	hash->dth_size = new_size;
6193 	hash->dth_mask = new_mask;
6194 }
6195 
6196 static void
6197 dtrace_hash_add(dtrace_hash_t *hash, dtrace_probe_t *new)
6198 {
6199 	int hashval = DTRACE_HASHSTR(hash, new);
6200 	int ndx = hashval & hash->dth_mask;
6201 	dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6202 	dtrace_probe_t **nextp, **prevp;
6203 
6204 	for (; bucket != NULL; bucket = bucket->dthb_next) {
6205 		if (DTRACE_HASHEQ(hash, bucket->dthb_chain, new))
6206 			goto add;
6207 	}
6208 
6209 	if ((hash->dth_nbuckets >> 1) > hash->dth_size) {
6210 		dtrace_hash_resize(hash);
6211 		dtrace_hash_add(hash, new);
6212 		return;
6213 	}
6214 
6215 	bucket = kmem_zalloc(sizeof (dtrace_hashbucket_t), KM_SLEEP);
6216 	bucket->dthb_next = hash->dth_tab[ndx];
6217 	hash->dth_tab[ndx] = bucket;
6218 	hash->dth_nbuckets++;
6219 
6220 add:
6221 	nextp = DTRACE_HASHNEXT(hash, new);
6222 	ASSERT(*nextp == NULL && *(DTRACE_HASHPREV(hash, new)) == NULL);
6223 	*nextp = bucket->dthb_chain;
6224 
6225 	if (bucket->dthb_chain != NULL) {
6226 		prevp = DTRACE_HASHPREV(hash, bucket->dthb_chain);
6227 		ASSERT(*prevp == NULL);
6228 		*prevp = new;
6229 	}
6230 
6231 	bucket->dthb_chain = new;
6232 	bucket->dthb_len++;
6233 }
6234 
6235 static dtrace_probe_t *
6236 dtrace_hash_lookup(dtrace_hash_t *hash, dtrace_probe_t *template)
6237 {
6238 	int hashval = DTRACE_HASHSTR(hash, template);
6239 	int ndx = hashval & hash->dth_mask;
6240 	dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6241 
6242 	for (; bucket != NULL; bucket = bucket->dthb_next) {
6243 		if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
6244 			return (bucket->dthb_chain);
6245 	}
6246 
6247 	return (NULL);
6248 }
6249 
6250 static int
6251 dtrace_hash_collisions(dtrace_hash_t *hash, dtrace_probe_t *template)
6252 {
6253 	int hashval = DTRACE_HASHSTR(hash, template);
6254 	int ndx = hashval & hash->dth_mask;
6255 	dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6256 
6257 	for (; bucket != NULL; bucket = bucket->dthb_next) {
6258 		if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
6259 			return (bucket->dthb_len);
6260 	}
6261 
6262 	return (NULL);
6263 }
6264 
6265 static void
6266 dtrace_hash_remove(dtrace_hash_t *hash, dtrace_probe_t *probe)
6267 {
6268 	int ndx = DTRACE_HASHSTR(hash, probe) & hash->dth_mask;
6269 	dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6270 
6271 	dtrace_probe_t **prevp = DTRACE_HASHPREV(hash, probe);
6272 	dtrace_probe_t **nextp = DTRACE_HASHNEXT(hash, probe);
6273 
6274 	/*
6275 	 * Find the bucket that we're removing this probe from.
6276 	 */
6277 	for (; bucket != NULL; bucket = bucket->dthb_next) {
6278 		if (DTRACE_HASHEQ(hash, bucket->dthb_chain, probe))
6279 			break;
6280 	}
6281 
6282 	ASSERT(bucket != NULL);
6283 
6284 	if (*prevp == NULL) {
6285 		if (*nextp == NULL) {
6286 			/*
6287 			 * The removed probe was the only probe on this
6288 			 * bucket; we need to remove the bucket.
6289 			 */
6290 			dtrace_hashbucket_t *b = hash->dth_tab[ndx];
6291 
6292 			ASSERT(bucket->dthb_chain == probe);
6293 			ASSERT(b != NULL);
6294 
6295 			if (b == bucket) {
6296 				hash->dth_tab[ndx] = bucket->dthb_next;
6297 			} else {
6298 				while (b->dthb_next != bucket)
6299 					b = b->dthb_next;
6300 				b->dthb_next = bucket->dthb_next;
6301 			}
6302 
6303 			ASSERT(hash->dth_nbuckets > 0);
6304 			hash->dth_nbuckets--;
6305 			kmem_free(bucket, sizeof (dtrace_hashbucket_t));
6306 			return;
6307 		}
6308 
6309 		bucket->dthb_chain = *nextp;
6310 	} else {
6311 		*(DTRACE_HASHNEXT(hash, *prevp)) = *nextp;
6312 	}
6313 
6314 	if (*nextp != NULL)
6315 		*(DTRACE_HASHPREV(hash, *nextp)) = *prevp;
6316 }
6317 
6318 /*
6319  * DTrace Utility Functions
6320  *
6321  * These are random utility functions that are _not_ called from probe context.
6322  */
6323 static int
6324 dtrace_badattr(const dtrace_attribute_t *a)
6325 {
6326 	return (a->dtat_name > DTRACE_STABILITY_MAX ||
6327 	    a->dtat_data > DTRACE_STABILITY_MAX ||
6328 	    a->dtat_class > DTRACE_CLASS_MAX);
6329 }
6330 
6331 /*
6332  * Return a duplicate copy of a string.  If the specified string is NULL,
6333  * this function returns a zero-length string.
6334  */
6335 static char *
6336 dtrace_strdup(const char *str)
6337 {
6338 	char *new = kmem_zalloc((str != NULL ? strlen(str) : 0) + 1, KM_SLEEP);
6339 
6340 	if (str != NULL)
6341 		(void) strcpy(new, str);
6342 
6343 	return (new);
6344 }
6345 
6346 #define	DTRACE_ISALPHA(c)	\
6347 	(((c) >= 'a' && (c) <= 'z') || ((c) >= 'A' && (c) <= 'Z'))
6348 
6349 static int
6350 dtrace_badname(const char *s)
6351 {
6352 	char c;
6353 
6354 	if (s == NULL || (c = *s++) == '\0')
6355 		return (0);
6356 
6357 	if (!DTRACE_ISALPHA(c) && c != '-' && c != '_' && c != '.')
6358 		return (1);
6359 
6360 	while ((c = *s++) != '\0') {
6361 		if (!DTRACE_ISALPHA(c) && (c < '0' || c > '9') &&
6362 		    c != '-' && c != '_' && c != '.' && c != '`')
6363 			return (1);
6364 	}
6365 
6366 	return (0);
6367 }
6368 
6369 static void
6370 dtrace_cred2priv(cred_t *cr, uint32_t *privp, uid_t *uidp, zoneid_t *zoneidp)
6371 {
6372 	uint32_t priv;
6373 
6374 	if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
6375 		/*
6376 		 * For DTRACE_PRIV_ALL, the uid and zoneid don't matter.
6377 		 */
6378 		priv = DTRACE_PRIV_ALL;
6379 	} else {
6380 		*uidp = crgetuid(cr);
6381 		*zoneidp = crgetzoneid(cr);
6382 
6383 		priv = 0;
6384 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE))
6385 			priv |= DTRACE_PRIV_KERNEL | DTRACE_PRIV_USER;
6386 		else if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE))
6387 			priv |= DTRACE_PRIV_USER;
6388 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE))
6389 			priv |= DTRACE_PRIV_PROC;
6390 		if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
6391 			priv |= DTRACE_PRIV_OWNER;
6392 		if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
6393 			priv |= DTRACE_PRIV_ZONEOWNER;
6394 	}
6395 
6396 	*privp = priv;
6397 }
6398 
6399 #ifdef DTRACE_ERRDEBUG
6400 static void
6401 dtrace_errdebug(const char *str)
6402 {
6403 	int hval = dtrace_hash_str((char *)str) % DTRACE_ERRHASHSZ;
6404 	int occupied = 0;
6405 
6406 	mutex_enter(&dtrace_errlock);
6407 	dtrace_errlast = str;
6408 	dtrace_errthread = curthread;
6409 
6410 	while (occupied++ < DTRACE_ERRHASHSZ) {
6411 		if (dtrace_errhash[hval].dter_msg == str) {
6412 			dtrace_errhash[hval].dter_count++;
6413 			goto out;
6414 		}
6415 
6416 		if (dtrace_errhash[hval].dter_msg != NULL) {
6417 			hval = (hval + 1) % DTRACE_ERRHASHSZ;
6418 			continue;
6419 		}
6420 
6421 		dtrace_errhash[hval].dter_msg = str;
6422 		dtrace_errhash[hval].dter_count = 1;
6423 		goto out;
6424 	}
6425 
6426 	panic("dtrace: undersized error hash");
6427 out:
6428 	mutex_exit(&dtrace_errlock);
6429 }
6430 #endif
6431 
6432 /*
6433  * DTrace Matching Functions
6434  *
6435  * These functions are used to match groups of probes, given some elements of
6436  * a probe tuple, or some globbed expressions for elements of a probe tuple.
6437  */
6438 static int
6439 dtrace_match_priv(const dtrace_probe_t *prp, uint32_t priv, uid_t uid,
6440     zoneid_t zoneid)
6441 {
6442 	if (priv != DTRACE_PRIV_ALL) {
6443 		uint32_t ppriv = prp->dtpr_provider->dtpv_priv.dtpp_flags;
6444 		uint32_t match = priv & ppriv;
6445 
6446 		/*
6447 		 * No PRIV_DTRACE_* privileges...
6448 		 */
6449 		if ((priv & (DTRACE_PRIV_PROC | DTRACE_PRIV_USER |
6450 		    DTRACE_PRIV_KERNEL)) == 0)
6451 			return (0);
6452 
6453 		/*
6454 		 * No matching bits, but there were bits to match...
6455 		 */
6456 		if (match == 0 && ppriv != 0)
6457 			return (0);
6458 
6459 		/*
6460 		 * Need to have permissions to the process, but don't...
6461 		 */
6462 		if (((ppriv & ~match) & DTRACE_PRIV_OWNER) != 0 &&
6463 		    uid != prp->dtpr_provider->dtpv_priv.dtpp_uid) {
6464 			return (0);
6465 		}
6466 
6467 		/*
6468 		 * Need to be in the same zone unless we possess the
6469 		 * privilege to examine all zones.
6470 		 */
6471 		if (((ppriv & ~match) & DTRACE_PRIV_ZONEOWNER) != 0 &&
6472 		    zoneid != prp->dtpr_provider->dtpv_priv.dtpp_zoneid) {
6473 			return (0);
6474 		}
6475 	}
6476 
6477 	return (1);
6478 }
6479 
6480 /*
6481  * dtrace_match_probe compares a dtrace_probe_t to a pre-compiled key, which
6482  * consists of input pattern strings and an ops-vector to evaluate them.
6483  * This function returns >0 for match, 0 for no match, and <0 for error.
6484  */
6485 static int
6486 dtrace_match_probe(const dtrace_probe_t *prp, const dtrace_probekey_t *pkp,
6487     uint32_t priv, uid_t uid, zoneid_t zoneid)
6488 {
6489 	dtrace_provider_t *pvp = prp->dtpr_provider;
6490 	int rv;
6491 
6492 	if (pvp->dtpv_defunct)
6493 		return (0);
6494 
6495 	if ((rv = pkp->dtpk_pmatch(pvp->dtpv_name, pkp->dtpk_prov, 0)) <= 0)
6496 		return (rv);
6497 
6498 	if ((rv = pkp->dtpk_mmatch(prp->dtpr_mod, pkp->dtpk_mod, 0)) <= 0)
6499 		return (rv);
6500 
6501 	if ((rv = pkp->dtpk_fmatch(prp->dtpr_func, pkp->dtpk_func, 0)) <= 0)
6502 		return (rv);
6503 
6504 	if ((rv = pkp->dtpk_nmatch(prp->dtpr_name, pkp->dtpk_name, 0)) <= 0)
6505 		return (rv);
6506 
6507 	if (dtrace_match_priv(prp, priv, uid, zoneid) == 0)
6508 		return (0);
6509 
6510 	return (rv);
6511 }
6512 
6513 /*
6514  * dtrace_match_glob() is a safe kernel implementation of the gmatch(3GEN)
6515  * interface for matching a glob pattern 'p' to an input string 's'.  Unlike
6516  * libc's version, the kernel version only applies to 8-bit ASCII strings.
6517  * In addition, all of the recursion cases except for '*' matching have been
6518  * unwound.  For '*', we still implement recursive evaluation, but a depth
6519  * counter is maintained and matching is aborted if we recurse too deep.
6520  * The function returns 0 if no match, >0 if match, and <0 if recursion error.
6521  */
6522 static int
6523 dtrace_match_glob(const char *s, const char *p, int depth)
6524 {
6525 	const char *olds;
6526 	char s1, c;
6527 	int gs;
6528 
6529 	if (depth > DTRACE_PROBEKEY_MAXDEPTH)
6530 		return (-1);
6531 
6532 	if (s == NULL)
6533 		s = ""; /* treat NULL as empty string */
6534 
6535 top:
6536 	olds = s;
6537 	s1 = *s++;
6538 
6539 	if (p == NULL)
6540 		return (0);
6541 
6542 	if ((c = *p++) == '\0')
6543 		return (s1 == '\0');
6544 
6545 	switch (c) {
6546 	case '[': {
6547 		int ok = 0, notflag = 0;
6548 		char lc = '\0';
6549 
6550 		if (s1 == '\0')
6551 			return (0);
6552 
6553 		if (*p == '!') {
6554 			notflag = 1;
6555 			p++;
6556 		}
6557 
6558 		if ((c = *p++) == '\0')
6559 			return (0);
6560 
6561 		do {
6562 			if (c == '-' && lc != '\0' && *p != ']') {
6563 				if ((c = *p++) == '\0')
6564 					return (0);
6565 				if (c == '\\' && (c = *p++) == '\0')
6566 					return (0);
6567 
6568 				if (notflag) {
6569 					if (s1 < lc || s1 > c)
6570 						ok++;
6571 					else
6572 						return (0);
6573 				} else if (lc <= s1 && s1 <= c)
6574 					ok++;
6575 
6576 			} else if (c == '\\' && (c = *p++) == '\0')
6577 				return (0);
6578 
6579 			lc = c; /* save left-hand 'c' for next iteration */
6580 
6581 			if (notflag) {
6582 				if (s1 != c)
6583 					ok++;
6584 				else
6585 					return (0);
6586 			} else if (s1 == c)
6587 				ok++;
6588 
6589 			if ((c = *p++) == '\0')
6590 				return (0);
6591 
6592 		} while (c != ']');
6593 
6594 		if (ok)
6595 			goto top;
6596 
6597 		return (0);
6598 	}
6599 
6600 	case '\\':
6601 		if ((c = *p++) == '\0')
6602 			return (0);
6603 		/*FALLTHRU*/
6604 
6605 	default:
6606 		if (c != s1)
6607 			return (0);
6608 		/*FALLTHRU*/
6609 
6610 	case '?':
6611 		if (s1 != '\0')
6612 			goto top;
6613 		return (0);
6614 
6615 	case '*':
6616 		while (*p == '*')
6617 			p++; /* consecutive *'s are identical to a single one */
6618 
6619 		if (*p == '\0')
6620 			return (1);
6621 
6622 		for (s = olds; *s != '\0'; s++) {
6623 			if ((gs = dtrace_match_glob(s, p, depth + 1)) != 0)
6624 				return (gs);
6625 		}
6626 
6627 		return (0);
6628 	}
6629 }
6630 
6631 /*ARGSUSED*/
6632 static int
6633 dtrace_match_string(const char *s, const char *p, int depth)
6634 {
6635 	return (s != NULL && strcmp(s, p) == 0);
6636 }
6637 
6638 /*ARGSUSED*/
6639 static int
6640 dtrace_match_nul(const char *s, const char *p, int depth)
6641 {
6642 	return (1); /* always match the empty pattern */
6643 }
6644 
6645 /*ARGSUSED*/
6646 static int
6647 dtrace_match_nonzero(const char *s, const char *p, int depth)
6648 {
6649 	return (s != NULL && s[0] != '\0');
6650 }
6651 
6652 static int
6653 dtrace_match(const dtrace_probekey_t *pkp, uint32_t priv, uid_t uid,
6654     zoneid_t zoneid, int (*matched)(dtrace_probe_t *, void *), void *arg)
6655 {
6656 	dtrace_probe_t template, *probe;
6657 	dtrace_hash_t *hash = NULL;
6658 	int len, best = INT_MAX, nmatched = 0;
6659 	dtrace_id_t i;
6660 
6661 	ASSERT(MUTEX_HELD(&dtrace_lock));
6662 
6663 	/*
6664 	 * If the probe ID is specified in the key, just lookup by ID and
6665 	 * invoke the match callback once if a matching probe is found.
6666 	 */
6667 	if (pkp->dtpk_id != DTRACE_IDNONE) {
6668 		if ((probe = dtrace_probe_lookup_id(pkp->dtpk_id)) != NULL &&
6669 		    dtrace_match_probe(probe, pkp, priv, uid, zoneid) > 0) {
6670 			(void) (*matched)(probe, arg);
6671 			nmatched++;
6672 		}
6673 		return (nmatched);
6674 	}
6675 
6676 	template.dtpr_mod = (char *)pkp->dtpk_mod;
6677 	template.dtpr_func = (char *)pkp->dtpk_func;
6678 	template.dtpr_name = (char *)pkp->dtpk_name;
6679 
6680 	/*
6681 	 * We want to find the most distinct of the module name, function
6682 	 * name, and name.  So for each one that is not a glob pattern or
6683 	 * empty string, we perform a lookup in the corresponding hash and
6684 	 * use the hash table with the fewest collisions to do our search.
6685 	 */
6686 	if (pkp->dtpk_mmatch == &dtrace_match_string &&
6687 	    (len = dtrace_hash_collisions(dtrace_bymod, &template)) < best) {
6688 		best = len;
6689 		hash = dtrace_bymod;
6690 	}
6691 
6692 	if (pkp->dtpk_fmatch == &dtrace_match_string &&
6693 	    (len = dtrace_hash_collisions(dtrace_byfunc, &template)) < best) {
6694 		best = len;
6695 		hash = dtrace_byfunc;
6696 	}
6697 
6698 	if (pkp->dtpk_nmatch == &dtrace_match_string &&
6699 	    (len = dtrace_hash_collisions(dtrace_byname, &template)) < best) {
6700 		best = len;
6701 		hash = dtrace_byname;
6702 	}
6703 
6704 	/*
6705 	 * If we did not select a hash table, iterate over every probe and
6706 	 * invoke our callback for each one that matches our input probe key.
6707 	 */
6708 	if (hash == NULL) {
6709 		for (i = 0; i < dtrace_nprobes; i++) {
6710 			if ((probe = dtrace_probes[i]) == NULL ||
6711 			    dtrace_match_probe(probe, pkp, priv, uid,
6712 			    zoneid) <= 0)
6713 				continue;
6714 
6715 			nmatched++;
6716 
6717 			if ((*matched)(probe, arg) != DTRACE_MATCH_NEXT)
6718 				break;
6719 		}
6720 
6721 		return (nmatched);
6722 	}
6723 
6724 	/*
6725 	 * If we selected a hash table, iterate over each probe of the same key
6726 	 * name and invoke the callback for every probe that matches the other
6727 	 * attributes of our input probe key.
6728 	 */
6729 	for (probe = dtrace_hash_lookup(hash, &template); probe != NULL;
6730 	    probe = *(DTRACE_HASHNEXT(hash, probe))) {
6731 
6732 		if (dtrace_match_probe(probe, pkp, priv, uid, zoneid) <= 0)
6733 			continue;
6734 
6735 		nmatched++;
6736 
6737 		if ((*matched)(probe, arg) != DTRACE_MATCH_NEXT)
6738 			break;
6739 	}
6740 
6741 	return (nmatched);
6742 }
6743 
6744 /*
6745  * Return the function pointer dtrace_probecmp() should use to compare the
6746  * specified pattern with a string.  For NULL or empty patterns, we select
6747  * dtrace_match_nul().  For glob pattern strings, we use dtrace_match_glob().
6748  * For non-empty non-glob strings, we use dtrace_match_string().
6749  */
6750 static dtrace_probekey_f *
6751 dtrace_probekey_func(const char *p)
6752 {
6753 	char c;
6754 
6755 	if (p == NULL || *p == '\0')
6756 		return (&dtrace_match_nul);
6757 
6758 	while ((c = *p++) != '\0') {
6759 		if (c == '[' || c == '?' || c == '*' || c == '\\')
6760 			return (&dtrace_match_glob);
6761 	}
6762 
6763 	return (&dtrace_match_string);
6764 }
6765 
6766 /*
6767  * Build a probe comparison key for use with dtrace_match_probe() from the
6768  * given probe description.  By convention, a null key only matches anchored
6769  * probes: if each field is the empty string, reset dtpk_fmatch to
6770  * dtrace_match_nonzero().
6771  */
6772 static void
6773 dtrace_probekey(const dtrace_probedesc_t *pdp, dtrace_probekey_t *pkp)
6774 {
6775 	pkp->dtpk_prov = pdp->dtpd_provider;
6776 	pkp->dtpk_pmatch = dtrace_probekey_func(pdp->dtpd_provider);
6777 
6778 	pkp->dtpk_mod = pdp->dtpd_mod;
6779 	pkp->dtpk_mmatch = dtrace_probekey_func(pdp->dtpd_mod);
6780 
6781 	pkp->dtpk_func = pdp->dtpd_func;
6782 	pkp->dtpk_fmatch = dtrace_probekey_func(pdp->dtpd_func);
6783 
6784 	pkp->dtpk_name = pdp->dtpd_name;
6785 	pkp->dtpk_nmatch = dtrace_probekey_func(pdp->dtpd_name);
6786 
6787 	pkp->dtpk_id = pdp->dtpd_id;
6788 
6789 	if (pkp->dtpk_id == DTRACE_IDNONE &&
6790 	    pkp->dtpk_pmatch == &dtrace_match_nul &&
6791 	    pkp->dtpk_mmatch == &dtrace_match_nul &&
6792 	    pkp->dtpk_fmatch == &dtrace_match_nul &&
6793 	    pkp->dtpk_nmatch == &dtrace_match_nul)
6794 		pkp->dtpk_fmatch = &dtrace_match_nonzero;
6795 }
6796 
6797 /*
6798  * DTrace Provider-to-Framework API Functions
6799  *
6800  * These functions implement much of the Provider-to-Framework API, as
6801  * described in <sys/dtrace.h>.  The parts of the API not in this section are
6802  * the functions in the API for probe management (found below), and
6803  * dtrace_probe() itself (found above).
6804  */
6805 
6806 /*
6807  * Register the calling provider with the DTrace framework.  This should
6808  * generally be called by DTrace providers in their attach(9E) entry point.
6809  */
6810 int
6811 dtrace_register(const char *name, const dtrace_pattr_t *pap, uint32_t priv,
6812     cred_t *cr, const dtrace_pops_t *pops, void *arg, dtrace_provider_id_t *idp)
6813 {
6814 	dtrace_provider_t *provider;
6815 
6816 	if (name == NULL || pap == NULL || pops == NULL || idp == NULL) {
6817 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
6818 		    "arguments", name ? name : "<NULL>");
6819 		return (EINVAL);
6820 	}
6821 
6822 	if (name[0] == '\0' || dtrace_badname(name)) {
6823 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
6824 		    "provider name", name);
6825 		return (EINVAL);
6826 	}
6827 
6828 	if ((pops->dtps_provide == NULL && pops->dtps_provide_module == NULL) ||
6829 	    pops->dtps_enable == NULL || pops->dtps_disable == NULL ||
6830 	    pops->dtps_destroy == NULL ||
6831 	    ((pops->dtps_resume == NULL) != (pops->dtps_suspend == NULL))) {
6832 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
6833 		    "provider ops", name);
6834 		return (EINVAL);
6835 	}
6836 
6837 	if (dtrace_badattr(&pap->dtpa_provider) ||
6838 	    dtrace_badattr(&pap->dtpa_mod) ||
6839 	    dtrace_badattr(&pap->dtpa_func) ||
6840 	    dtrace_badattr(&pap->dtpa_name) ||
6841 	    dtrace_badattr(&pap->dtpa_args)) {
6842 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
6843 		    "provider attributes", name);
6844 		return (EINVAL);
6845 	}
6846 
6847 	if (priv & ~DTRACE_PRIV_ALL) {
6848 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
6849 		    "privilege attributes", name);
6850 		return (EINVAL);
6851 	}
6852 
6853 	if ((priv & DTRACE_PRIV_KERNEL) &&
6854 	    (priv & (DTRACE_PRIV_USER | DTRACE_PRIV_OWNER)) &&
6855 	    pops->dtps_usermode == NULL) {
6856 		cmn_err(CE_WARN, "failed to register provider '%s': need "
6857 		    "dtps_usermode() op for given privilege attributes", name);
6858 		return (EINVAL);
6859 	}
6860 
6861 	provider = kmem_zalloc(sizeof (dtrace_provider_t), KM_SLEEP);
6862 	provider->dtpv_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
6863 	(void) strcpy(provider->dtpv_name, name);
6864 
6865 	provider->dtpv_attr = *pap;
6866 	provider->dtpv_priv.dtpp_flags = priv;
6867 	if (cr != NULL) {
6868 		provider->dtpv_priv.dtpp_uid = crgetuid(cr);
6869 		provider->dtpv_priv.dtpp_zoneid = crgetzoneid(cr);
6870 	}
6871 	provider->dtpv_pops = *pops;
6872 
6873 	if (pops->dtps_provide == NULL) {
6874 		ASSERT(pops->dtps_provide_module != NULL);
6875 		provider->dtpv_pops.dtps_provide =
6876 		    (void (*)(void *, const dtrace_probedesc_t *))dtrace_nullop;
6877 	}
6878 
6879 	if (pops->dtps_provide_module == NULL) {
6880 		ASSERT(pops->dtps_provide != NULL);
6881 		provider->dtpv_pops.dtps_provide_module =
6882 		    (void (*)(void *, struct modctl *))dtrace_nullop;
6883 	}
6884 
6885 	if (pops->dtps_suspend == NULL) {
6886 		ASSERT(pops->dtps_resume == NULL);
6887 		provider->dtpv_pops.dtps_suspend =
6888 		    (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
6889 		provider->dtpv_pops.dtps_resume =
6890 		    (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
6891 	}
6892 
6893 	provider->dtpv_arg = arg;
6894 	*idp = (dtrace_provider_id_t)provider;
6895 
6896 	if (pops == &dtrace_provider_ops) {
6897 		ASSERT(MUTEX_HELD(&dtrace_provider_lock));
6898 		ASSERT(MUTEX_HELD(&dtrace_lock));
6899 		ASSERT(dtrace_anon.dta_enabling == NULL);
6900 
6901 		/*
6902 		 * We make sure that the DTrace provider is at the head of
6903 		 * the provider chain.
6904 		 */
6905 		provider->dtpv_next = dtrace_provider;
6906 		dtrace_provider = provider;
6907 		return (0);
6908 	}
6909 
6910 	mutex_enter(&dtrace_provider_lock);
6911 	mutex_enter(&dtrace_lock);
6912 
6913 	/*
6914 	 * If there is at least one provider registered, we'll add this
6915 	 * provider after the first provider.
6916 	 */
6917 	if (dtrace_provider != NULL) {
6918 		provider->dtpv_next = dtrace_provider->dtpv_next;
6919 		dtrace_provider->dtpv_next = provider;
6920 	} else {
6921 		dtrace_provider = provider;
6922 	}
6923 
6924 	if (dtrace_retained != NULL) {
6925 		dtrace_enabling_provide(provider);
6926 
6927 		/*
6928 		 * Now we need to call dtrace_enabling_matchall() -- which
6929 		 * will acquire cpu_lock and dtrace_lock.  We therefore need
6930 		 * to drop all of our locks before calling into it...
6931 		 */
6932 		mutex_exit(&dtrace_lock);
6933 		mutex_exit(&dtrace_provider_lock);
6934 		dtrace_enabling_matchall();
6935 
6936 		return (0);
6937 	}
6938 
6939 	mutex_exit(&dtrace_lock);
6940 	mutex_exit(&dtrace_provider_lock);
6941 
6942 	return (0);
6943 }
6944 
6945 /*
6946  * Unregister the specified provider from the DTrace framework.  This should
6947  * generally be called by DTrace providers in their detach(9E) entry point.
6948  */
6949 int
6950 dtrace_unregister(dtrace_provider_id_t id)
6951 {
6952 	dtrace_provider_t *old = (dtrace_provider_t *)id;
6953 	dtrace_provider_t *prev = NULL;
6954 	int i, self = 0;
6955 	dtrace_probe_t *probe, *first = NULL;
6956 
6957 	if (old->dtpv_pops.dtps_enable ==
6958 	    (void (*)(void *, dtrace_id_t, void *))dtrace_nullop) {
6959 		/*
6960 		 * If DTrace itself is the provider, we're called with locks
6961 		 * already held.
6962 		 */
6963 		ASSERT(old == dtrace_provider);
6964 		ASSERT(dtrace_devi != NULL);
6965 		ASSERT(MUTEX_HELD(&dtrace_provider_lock));
6966 		ASSERT(MUTEX_HELD(&dtrace_lock));
6967 		self = 1;
6968 
6969 		if (dtrace_provider->dtpv_next != NULL) {
6970 			/*
6971 			 * There's another provider here; return failure.
6972 			 */
6973 			return (EBUSY);
6974 		}
6975 	} else {
6976 		mutex_enter(&dtrace_provider_lock);
6977 		mutex_enter(&mod_lock);
6978 		mutex_enter(&dtrace_lock);
6979 	}
6980 
6981 	/*
6982 	 * If anyone has /dev/dtrace open, or if there are anonymous enabled
6983 	 * probes, we refuse to let providers slither away, unless this
6984 	 * provider has already been explicitly invalidated.
6985 	 */
6986 	if (!old->dtpv_defunct &&
6987 	    (dtrace_opens || (dtrace_anon.dta_state != NULL &&
6988 	    dtrace_anon.dta_state->dts_necbs > 0))) {
6989 		if (!self) {
6990 			mutex_exit(&dtrace_lock);
6991 			mutex_exit(&mod_lock);
6992 			mutex_exit(&dtrace_provider_lock);
6993 		}
6994 		return (EBUSY);
6995 	}
6996 
6997 	/*
6998 	 * Attempt to destroy the probes associated with this provider.
6999 	 */
7000 	for (i = 0; i < dtrace_nprobes; i++) {
7001 		if ((probe = dtrace_probes[i]) == NULL)
7002 			continue;
7003 
7004 		if (probe->dtpr_provider != old)
7005 			continue;
7006 
7007 		if (probe->dtpr_ecb == NULL)
7008 			continue;
7009 
7010 		/*
7011 		 * We have at least one ECB; we can't remove this provider.
7012 		 */
7013 		if (!self) {
7014 			mutex_exit(&dtrace_lock);
7015 			mutex_exit(&mod_lock);
7016 			mutex_exit(&dtrace_provider_lock);
7017 		}
7018 		return (EBUSY);
7019 	}
7020 
7021 	/*
7022 	 * All of the probes for this provider are disabled; we can safely
7023 	 * remove all of them from their hash chains and from the probe array.
7024 	 */
7025 	for (i = 0; i < dtrace_nprobes; i++) {
7026 		if ((probe = dtrace_probes[i]) == NULL)
7027 			continue;
7028 
7029 		if (probe->dtpr_provider != old)
7030 			continue;
7031 
7032 		dtrace_probes[i] = NULL;
7033 
7034 		dtrace_hash_remove(dtrace_bymod, probe);
7035 		dtrace_hash_remove(dtrace_byfunc, probe);
7036 		dtrace_hash_remove(dtrace_byname, probe);
7037 
7038 		if (first == NULL) {
7039 			first = probe;
7040 			probe->dtpr_nextmod = NULL;
7041 		} else {
7042 			probe->dtpr_nextmod = first;
7043 			first = probe;
7044 		}
7045 	}
7046 
7047 	/*
7048 	 * The provider's probes have been removed from the hash chains and
7049 	 * from the probe array.  Now issue a dtrace_sync() to be sure that
7050 	 * everyone has cleared out from any probe array processing.
7051 	 */
7052 	dtrace_sync();
7053 
7054 	for (probe = first; probe != NULL; probe = first) {
7055 		first = probe->dtpr_nextmod;
7056 
7057 		old->dtpv_pops.dtps_destroy(old->dtpv_arg, probe->dtpr_id,
7058 		    probe->dtpr_arg);
7059 		kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
7060 		kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
7061 		kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
7062 		vmem_free(dtrace_arena, (void *)(uintptr_t)(probe->dtpr_id), 1);
7063 		kmem_free(probe, sizeof (dtrace_probe_t));
7064 	}
7065 
7066 	if ((prev = dtrace_provider) == old) {
7067 		ASSERT(self || dtrace_devi == NULL);
7068 		ASSERT(old->dtpv_next == NULL || dtrace_devi == NULL);
7069 		dtrace_provider = old->dtpv_next;
7070 	} else {
7071 		while (prev != NULL && prev->dtpv_next != old)
7072 			prev = prev->dtpv_next;
7073 
7074 		if (prev == NULL) {
7075 			panic("attempt to unregister non-existent "
7076 			    "dtrace provider %p\n", (void *)id);
7077 		}
7078 
7079 		prev->dtpv_next = old->dtpv_next;
7080 	}
7081 
7082 	if (!self) {
7083 		mutex_exit(&dtrace_lock);
7084 		mutex_exit(&mod_lock);
7085 		mutex_exit(&dtrace_provider_lock);
7086 	}
7087 
7088 	kmem_free(old->dtpv_name, strlen(old->dtpv_name) + 1);
7089 	kmem_free(old, sizeof (dtrace_provider_t));
7090 
7091 	return (0);
7092 }
7093 
7094 /*
7095  * Invalidate the specified provider.  All subsequent probe lookups for the
7096  * specified provider will fail, but its probes will not be removed.
7097  */
7098 void
7099 dtrace_invalidate(dtrace_provider_id_t id)
7100 {
7101 	dtrace_provider_t *pvp = (dtrace_provider_t *)id;
7102 
7103 	ASSERT(pvp->dtpv_pops.dtps_enable !=
7104 	    (void (*)(void *, dtrace_id_t, void *))dtrace_nullop);
7105 
7106 	mutex_enter(&dtrace_provider_lock);
7107 	mutex_enter(&dtrace_lock);
7108 
7109 	pvp->dtpv_defunct = 1;
7110 
7111 	mutex_exit(&dtrace_lock);
7112 	mutex_exit(&dtrace_provider_lock);
7113 }
7114 
7115 /*
7116  * Indicate whether or not DTrace has attached.
7117  */
7118 int
7119 dtrace_attached(void)
7120 {
7121 	/*
7122 	 * dtrace_provider will be non-NULL iff the DTrace driver has
7123 	 * attached.  (It's non-NULL because DTrace is always itself a
7124 	 * provider.)
7125 	 */
7126 	return (dtrace_provider != NULL);
7127 }
7128 
7129 /*
7130  * Remove all the unenabled probes for the given provider.  This function is
7131  * not unlike dtrace_unregister(), except that it doesn't remove the provider
7132  * -- just as many of its associated probes as it can.
7133  */
7134 int
7135 dtrace_condense(dtrace_provider_id_t id)
7136 {
7137 	dtrace_provider_t *prov = (dtrace_provider_t *)id;
7138 	int i;
7139 	dtrace_probe_t *probe;
7140 
7141 	/*
7142 	 * Make sure this isn't the dtrace provider itself.
7143 	 */
7144 	ASSERT(prov->dtpv_pops.dtps_enable !=
7145 	    (void (*)(void *, dtrace_id_t, void *))dtrace_nullop);
7146 
7147 	mutex_enter(&dtrace_provider_lock);
7148 	mutex_enter(&dtrace_lock);
7149 
7150 	/*
7151 	 * Attempt to destroy the probes associated with this provider.
7152 	 */
7153 	for (i = 0; i < dtrace_nprobes; i++) {
7154 		if ((probe = dtrace_probes[i]) == NULL)
7155 			continue;
7156 
7157 		if (probe->dtpr_provider != prov)
7158 			continue;
7159 
7160 		if (probe->dtpr_ecb != NULL)
7161 			continue;
7162 
7163 		dtrace_probes[i] = NULL;
7164 
7165 		dtrace_hash_remove(dtrace_bymod, probe);
7166 		dtrace_hash_remove(dtrace_byfunc, probe);
7167 		dtrace_hash_remove(dtrace_byname, probe);
7168 
7169 		prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, i + 1,
7170 		    probe->dtpr_arg);
7171 		kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
7172 		kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
7173 		kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
7174 		kmem_free(probe, sizeof (dtrace_probe_t));
7175 		vmem_free(dtrace_arena, (void *)((uintptr_t)i + 1), 1);
7176 	}
7177 
7178 	mutex_exit(&dtrace_lock);
7179 	mutex_exit(&dtrace_provider_lock);
7180 
7181 	return (0);
7182 }
7183 
7184 /*
7185  * DTrace Probe Management Functions
7186  *
7187  * The functions in this section perform the DTrace probe management,
7188  * including functions to create probes, look-up probes, and call into the
7189  * providers to request that probes be provided.  Some of these functions are
7190  * in the Provider-to-Framework API; these functions can be identified by the
7191  * fact that they are not declared "static".
7192  */
7193 
7194 /*
7195  * Create a probe with the specified module name, function name, and name.
7196  */
7197 dtrace_id_t
7198 dtrace_probe_create(dtrace_provider_id_t prov, const char *mod,
7199     const char *func, const char *name, int aframes, void *arg)
7200 {
7201 	dtrace_probe_t *probe, **probes;
7202 	dtrace_provider_t *provider = (dtrace_provider_t *)prov;
7203 	dtrace_id_t id;
7204 
7205 	if (provider == dtrace_provider) {
7206 		ASSERT(MUTEX_HELD(&dtrace_lock));
7207 	} else {
7208 		mutex_enter(&dtrace_lock);
7209 	}
7210 
7211 	id = (dtrace_id_t)(uintptr_t)vmem_alloc(dtrace_arena, 1,
7212 	    VM_BESTFIT | VM_SLEEP);
7213 	probe = kmem_zalloc(sizeof (dtrace_probe_t), KM_SLEEP);
7214 
7215 	probe->dtpr_id = id;
7216 	probe->dtpr_gen = dtrace_probegen++;
7217 	probe->dtpr_mod = dtrace_strdup(mod);
7218 	probe->dtpr_func = dtrace_strdup(func);
7219 	probe->dtpr_name = dtrace_strdup(name);
7220 	probe->dtpr_arg = arg;
7221 	probe->dtpr_aframes = aframes;
7222 	probe->dtpr_provider = provider;
7223 
7224 	dtrace_hash_add(dtrace_bymod, probe);
7225 	dtrace_hash_add(dtrace_byfunc, probe);
7226 	dtrace_hash_add(dtrace_byname, probe);
7227 
7228 	if (id - 1 >= dtrace_nprobes) {
7229 		size_t osize = dtrace_nprobes * sizeof (dtrace_probe_t *);
7230 		size_t nsize = osize << 1;
7231 
7232 		if (nsize == 0) {
7233 			ASSERT(osize == 0);
7234 			ASSERT(dtrace_probes == NULL);
7235 			nsize = sizeof (dtrace_probe_t *);
7236 		}
7237 
7238 		probes = kmem_zalloc(nsize, KM_SLEEP);
7239 
7240 		if (dtrace_probes == NULL) {
7241 			ASSERT(osize == 0);
7242 			dtrace_probes = probes;
7243 			dtrace_nprobes = 1;
7244 		} else {
7245 			dtrace_probe_t **oprobes = dtrace_probes;
7246 
7247 			bcopy(oprobes, probes, osize);
7248 			dtrace_membar_producer();
7249 			dtrace_probes = probes;
7250 
7251 			dtrace_sync();
7252 
7253 			/*
7254 			 * All CPUs are now seeing the new probes array; we can
7255 			 * safely free the old array.
7256 			 */
7257 			kmem_free(oprobes, osize);
7258 			dtrace_nprobes <<= 1;
7259 		}
7260 
7261 		ASSERT(id - 1 < dtrace_nprobes);
7262 	}
7263 
7264 	ASSERT(dtrace_probes[id - 1] == NULL);
7265 	dtrace_probes[id - 1] = probe;
7266 
7267 	if (provider != dtrace_provider)
7268 		mutex_exit(&dtrace_lock);
7269 
7270 	return (id);
7271 }
7272 
7273 static dtrace_probe_t *
7274 dtrace_probe_lookup_id(dtrace_id_t id)
7275 {
7276 	ASSERT(MUTEX_HELD(&dtrace_lock));
7277 
7278 	if (id == 0 || id > dtrace_nprobes)
7279 		return (NULL);
7280 
7281 	return (dtrace_probes[id - 1]);
7282 }
7283 
7284 static int
7285 dtrace_probe_lookup_match(dtrace_probe_t *probe, void *arg)
7286 {
7287 	*((dtrace_id_t *)arg) = probe->dtpr_id;
7288 
7289 	return (DTRACE_MATCH_DONE);
7290 }
7291 
7292 /*
7293  * Look up a probe based on provider and one or more of module name, function
7294  * name and probe name.
7295  */
7296 dtrace_id_t
7297 dtrace_probe_lookup(dtrace_provider_id_t prid, const char *mod,
7298     const char *func, const char *name)
7299 {
7300 	dtrace_probekey_t pkey;
7301 	dtrace_id_t id;
7302 	int match;
7303 
7304 	pkey.dtpk_prov = ((dtrace_provider_t *)prid)->dtpv_name;
7305 	pkey.dtpk_pmatch = &dtrace_match_string;
7306 	pkey.dtpk_mod = mod;
7307 	pkey.dtpk_mmatch = mod ? &dtrace_match_string : &dtrace_match_nul;
7308 	pkey.dtpk_func = func;
7309 	pkey.dtpk_fmatch = func ? &dtrace_match_string : &dtrace_match_nul;
7310 	pkey.dtpk_name = name;
7311 	pkey.dtpk_nmatch = name ? &dtrace_match_string : &dtrace_match_nul;
7312 	pkey.dtpk_id = DTRACE_IDNONE;
7313 
7314 	mutex_enter(&dtrace_lock);
7315 	match = dtrace_match(&pkey, DTRACE_PRIV_ALL, 0, 0,
7316 	    dtrace_probe_lookup_match, &id);
7317 	mutex_exit(&dtrace_lock);
7318 
7319 	ASSERT(match == 1 || match == 0);
7320 	return (match ? id : 0);
7321 }
7322 
7323 /*
7324  * Returns the probe argument associated with the specified probe.
7325  */
7326 void *
7327 dtrace_probe_arg(dtrace_provider_id_t id, dtrace_id_t pid)
7328 {
7329 	dtrace_probe_t *probe;
7330 	void *rval = NULL;
7331 
7332 	mutex_enter(&dtrace_lock);
7333 
7334 	if ((probe = dtrace_probe_lookup_id(pid)) != NULL &&
7335 	    probe->dtpr_provider == (dtrace_provider_t *)id)
7336 		rval = probe->dtpr_arg;
7337 
7338 	mutex_exit(&dtrace_lock);
7339 
7340 	return (rval);
7341 }
7342 
7343 /*
7344  * Copy a probe into a probe description.
7345  */
7346 static void
7347 dtrace_probe_description(const dtrace_probe_t *prp, dtrace_probedesc_t *pdp)
7348 {
7349 	bzero(pdp, sizeof (dtrace_probedesc_t));
7350 	pdp->dtpd_id = prp->dtpr_id;
7351 
7352 	(void) strncpy(pdp->dtpd_provider,
7353 	    prp->dtpr_provider->dtpv_name, DTRACE_PROVNAMELEN - 1);
7354 
7355 	(void) strncpy(pdp->dtpd_mod, prp->dtpr_mod, DTRACE_MODNAMELEN - 1);
7356 	(void) strncpy(pdp->dtpd_func, prp->dtpr_func, DTRACE_FUNCNAMELEN - 1);
7357 	(void) strncpy(pdp->dtpd_name, prp->dtpr_name, DTRACE_NAMELEN - 1);
7358 }
7359 
7360 /*
7361  * Called to indicate that a probe -- or probes -- should be provided by a
7362  * specfied provider.  If the specified description is NULL, the provider will
7363  * be told to provide all of its probes.  (This is done whenever a new
7364  * consumer comes along, or whenever a retained enabling is to be matched.) If
7365  * the specified description is non-NULL, the provider is given the
7366  * opportunity to dynamically provide the specified probe, allowing providers
7367  * to support the creation of probes on-the-fly.  (So-called _autocreated_
7368  * probes.)  If the provider is NULL, the operations will be applied to all
7369  * providers; if the provider is non-NULL the operations will only be applied
7370  * to the specified provider.  The dtrace_provider_lock must be held, and the
7371  * dtrace_lock must _not_ be held -- the provider's dtps_provide() operation
7372  * will need to grab the dtrace_lock when it reenters the framework through
7373  * dtrace_probe_lookup(), dtrace_probe_create(), etc.
7374  */
7375 static void
7376 dtrace_probe_provide(dtrace_probedesc_t *desc, dtrace_provider_t *prv)
7377 {
7378 	struct modctl *ctl;
7379 	int all = 0;
7380 
7381 	ASSERT(MUTEX_HELD(&dtrace_provider_lock));
7382 
7383 	if (prv == NULL) {
7384 		all = 1;
7385 		prv = dtrace_provider;
7386 	}
7387 
7388 	do {
7389 		/*
7390 		 * First, call the blanket provide operation.
7391 		 */
7392 		prv->dtpv_pops.dtps_provide(prv->dtpv_arg, desc);
7393 
7394 		/*
7395 		 * Now call the per-module provide operation.  We will grab
7396 		 * mod_lock to prevent the list from being modified.  Note
7397 		 * that this also prevents the mod_busy bits from changing.
7398 		 * (mod_busy can only be changed with mod_lock held.)
7399 		 */
7400 		mutex_enter(&mod_lock);
7401 
7402 		ctl = &modules;
7403 		do {
7404 			if (ctl->mod_busy || ctl->mod_mp == NULL)
7405 				continue;
7406 
7407 			prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
7408 
7409 		} while ((ctl = ctl->mod_next) != &modules);
7410 
7411 		mutex_exit(&mod_lock);
7412 	} while (all && (prv = prv->dtpv_next) != NULL);
7413 }
7414 
7415 /*
7416  * Iterate over each probe, and call the Framework-to-Provider API function
7417  * denoted by offs.
7418  */
7419 static void
7420 dtrace_probe_foreach(uintptr_t offs)
7421 {
7422 	dtrace_provider_t *prov;
7423 	void (*func)(void *, dtrace_id_t, void *);
7424 	dtrace_probe_t *probe;
7425 	dtrace_icookie_t cookie;
7426 	int i;
7427 
7428 	/*
7429 	 * We disable interrupts to walk through the probe array.  This is
7430 	 * safe -- the dtrace_sync() in dtrace_unregister() assures that we
7431 	 * won't see stale data.
7432 	 */
7433 	cookie = dtrace_interrupt_disable();
7434 
7435 	for (i = 0; i < dtrace_nprobes; i++) {
7436 		if ((probe = dtrace_probes[i]) == NULL)
7437 			continue;
7438 
7439 		if (probe->dtpr_ecb == NULL) {
7440 			/*
7441 			 * This probe isn't enabled -- don't call the function.
7442 			 */
7443 			continue;
7444 		}
7445 
7446 		prov = probe->dtpr_provider;
7447 		func = *((void(**)(void *, dtrace_id_t, void *))
7448 		    ((uintptr_t)&prov->dtpv_pops + offs));
7449 
7450 		func(prov->dtpv_arg, i + 1, probe->dtpr_arg);
7451 	}
7452 
7453 	dtrace_interrupt_enable(cookie);
7454 }
7455 
7456 static int
7457 dtrace_probe_enable(const dtrace_probedesc_t *desc, dtrace_enabling_t *enab)
7458 {
7459 	dtrace_probekey_t pkey;
7460 	uint32_t priv;
7461 	uid_t uid;
7462 	zoneid_t zoneid;
7463 
7464 	ASSERT(MUTEX_HELD(&dtrace_lock));
7465 	dtrace_ecb_create_cache = NULL;
7466 
7467 	if (desc == NULL) {
7468 		/*
7469 		 * If we're passed a NULL description, we're being asked to
7470 		 * create an ECB with a NULL probe.
7471 		 */
7472 		(void) dtrace_ecb_create_enable(NULL, enab);
7473 		return (0);
7474 	}
7475 
7476 	dtrace_probekey(desc, &pkey);
7477 	dtrace_cred2priv(enab->dten_vstate->dtvs_state->dts_cred.dcr_cred,
7478 	    &priv, &uid, &zoneid);
7479 
7480 	return (dtrace_match(&pkey, priv, uid, zoneid, dtrace_ecb_create_enable,
7481 	    enab));
7482 }
7483 
7484 /*
7485  * DTrace Helper Provider Functions
7486  */
7487 static void
7488 dtrace_dofattr2attr(dtrace_attribute_t *attr, const dof_attr_t dofattr)
7489 {
7490 	attr->dtat_name = DOF_ATTR_NAME(dofattr);
7491 	attr->dtat_data = DOF_ATTR_DATA(dofattr);
7492 	attr->dtat_class = DOF_ATTR_CLASS(dofattr);
7493 }
7494 
7495 static void
7496 dtrace_dofprov2hprov(dtrace_helper_provdesc_t *hprov,
7497     const dof_provider_t *dofprov, char *strtab)
7498 {
7499 	hprov->dthpv_provname = strtab + dofprov->dofpv_name;
7500 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_provider,
7501 	    dofprov->dofpv_provattr);
7502 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_mod,
7503 	    dofprov->dofpv_modattr);
7504 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_func,
7505 	    dofprov->dofpv_funcattr);
7506 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_name,
7507 	    dofprov->dofpv_nameattr);
7508 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_args,
7509 	    dofprov->dofpv_argsattr);
7510 }
7511 
7512 static void
7513 dtrace_helper_provide_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
7514 {
7515 	uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
7516 	dof_hdr_t *dof = (dof_hdr_t *)daddr;
7517 	dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
7518 	dof_provider_t *provider;
7519 	dof_probe_t *probe;
7520 	uint32_t *off, *enoff;
7521 	uint8_t *arg;
7522 	char *strtab;
7523 	uint_t i, nprobes;
7524 	dtrace_helper_provdesc_t dhpv;
7525 	dtrace_helper_probedesc_t dhpb;
7526 	dtrace_meta_t *meta = dtrace_meta_pid;
7527 	dtrace_mops_t *mops = &meta->dtm_mops;
7528 	void *parg;
7529 
7530 	provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
7531 	str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
7532 	    provider->dofpv_strtab * dof->dofh_secsize);
7533 	prb_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
7534 	    provider->dofpv_probes * dof->dofh_secsize);
7535 	arg_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
7536 	    provider->dofpv_prargs * dof->dofh_secsize);
7537 	off_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
7538 	    provider->dofpv_proffs * dof->dofh_secsize);
7539 
7540 	strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
7541 	off = (uint32_t *)(uintptr_t)(daddr + off_sec->dofs_offset);
7542 	arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
7543 	enoff = NULL;
7544 
7545 	/*
7546 	 * See dtrace_helper_provider_validate().
7547 	 */
7548 	if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
7549 	    provider->dofpv_prenoffs != DOF_SECT_NONE) {
7550 		enoff_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
7551 		    provider->dofpv_prenoffs * dof->dofh_secsize);
7552 		enoff = (uint32_t *)(uintptr_t)(daddr + enoff_sec->dofs_offset);
7553 	}
7554 
7555 	nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
7556 
7557 	/*
7558 	 * Create the provider.
7559 	 */
7560 	dtrace_dofprov2hprov(&dhpv, provider, strtab);
7561 
7562 	if ((parg = mops->dtms_provide_pid(meta->dtm_arg, &dhpv, pid)) == NULL)
7563 		return;
7564 
7565 	meta->dtm_count++;
7566 
7567 	/*
7568 	 * Create the probes.
7569 	 */
7570 	for (i = 0; i < nprobes; i++) {
7571 		probe = (dof_probe_t *)(uintptr_t)(daddr +
7572 		    prb_sec->dofs_offset + i * prb_sec->dofs_entsize);
7573 
7574 		dhpb.dthpb_mod = dhp->dofhp_mod;
7575 		dhpb.dthpb_func = strtab + probe->dofpr_func;
7576 		dhpb.dthpb_name = strtab + probe->dofpr_name;
7577 		dhpb.dthpb_base = probe->dofpr_addr;
7578 		dhpb.dthpb_offs = off + probe->dofpr_offidx;
7579 		dhpb.dthpb_noffs = probe->dofpr_noffs;
7580 		if (enoff != NULL) {
7581 			dhpb.dthpb_enoffs = enoff + probe->dofpr_enoffidx;
7582 			dhpb.dthpb_nenoffs = probe->dofpr_nenoffs;
7583 		} else {
7584 			dhpb.dthpb_enoffs = NULL;
7585 			dhpb.dthpb_nenoffs = 0;
7586 		}
7587 		dhpb.dthpb_args = arg + probe->dofpr_argidx;
7588 		dhpb.dthpb_nargc = probe->dofpr_nargc;
7589 		dhpb.dthpb_xargc = probe->dofpr_xargc;
7590 		dhpb.dthpb_ntypes = strtab + probe->dofpr_nargv;
7591 		dhpb.dthpb_xtypes = strtab + probe->dofpr_xargv;
7592 
7593 		mops->dtms_create_probe(meta->dtm_arg, parg, &dhpb);
7594 	}
7595 }
7596 
7597 static void
7598 dtrace_helper_provide(dof_helper_t *dhp, pid_t pid)
7599 {
7600 	uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
7601 	dof_hdr_t *dof = (dof_hdr_t *)daddr;
7602 	int i;
7603 
7604 	ASSERT(MUTEX_HELD(&dtrace_meta_lock));
7605 
7606 	for (i = 0; i < dof->dofh_secnum; i++) {
7607 		dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
7608 		    dof->dofh_secoff + i * dof->dofh_secsize);
7609 
7610 		if (sec->dofs_type != DOF_SECT_PROVIDER)
7611 			continue;
7612 
7613 		dtrace_helper_provide_one(dhp, sec, pid);
7614 	}
7615 
7616 	/*
7617 	 * We may have just created probes, so we must now rematch against
7618 	 * any retained enablings.  Note that this call will acquire both
7619 	 * cpu_lock and dtrace_lock; the fact that we are holding
7620 	 * dtrace_meta_lock now is what defines the ordering with respect to
7621 	 * these three locks.
7622 	 */
7623 	dtrace_enabling_matchall();
7624 }
7625 
7626 static void
7627 dtrace_helper_provider_remove_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
7628 {
7629 	uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
7630 	dof_hdr_t *dof = (dof_hdr_t *)daddr;
7631 	dof_sec_t *str_sec;
7632 	dof_provider_t *provider;
7633 	char *strtab;
7634 	dtrace_helper_provdesc_t dhpv;
7635 	dtrace_meta_t *meta = dtrace_meta_pid;
7636 	dtrace_mops_t *mops = &meta->dtm_mops;
7637 
7638 	provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
7639 	str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
7640 	    provider->dofpv_strtab * dof->dofh_secsize);
7641 
7642 	strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
7643 
7644 	/*
7645 	 * Create the provider.
7646 	 */
7647 	dtrace_dofprov2hprov(&dhpv, provider, strtab);
7648 
7649 	mops->dtms_remove_pid(meta->dtm_arg, &dhpv, pid);
7650 
7651 	meta->dtm_count--;
7652 }
7653 
7654 static void
7655 dtrace_helper_provider_remove(dof_helper_t *dhp, pid_t pid)
7656 {
7657 	uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
7658 	dof_hdr_t *dof = (dof_hdr_t *)daddr;
7659 	int i;
7660 
7661 	ASSERT(MUTEX_HELD(&dtrace_meta_lock));
7662 
7663 	for (i = 0; i < dof->dofh_secnum; i++) {
7664 		dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
7665 		    dof->dofh_secoff + i * dof->dofh_secsize);
7666 
7667 		if (sec->dofs_type != DOF_SECT_PROVIDER)
7668 			continue;
7669 
7670 		dtrace_helper_provider_remove_one(dhp, sec, pid);
7671 	}
7672 }
7673 
7674 /*
7675  * DTrace Meta Provider-to-Framework API Functions
7676  *
7677  * These functions implement the Meta Provider-to-Framework API, as described
7678  * in <sys/dtrace.h>.
7679  */
7680 int
7681 dtrace_meta_register(const char *name, const dtrace_mops_t *mops, void *arg,
7682     dtrace_meta_provider_id_t *idp)
7683 {
7684 	dtrace_meta_t *meta;
7685 	dtrace_helpers_t *help, *next;
7686 	int i;
7687 
7688 	*idp = DTRACE_METAPROVNONE;
7689 
7690 	/*
7691 	 * We strictly don't need the name, but we hold onto it for
7692 	 * debuggability. All hail error queues!
7693 	 */
7694 	if (name == NULL) {
7695 		cmn_err(CE_WARN, "failed to register meta-provider: "
7696 		    "invalid name");
7697 		return (EINVAL);
7698 	}
7699 
7700 	if (mops == NULL ||
7701 	    mops->dtms_create_probe == NULL ||
7702 	    mops->dtms_provide_pid == NULL ||
7703 	    mops->dtms_remove_pid == NULL) {
7704 		cmn_err(CE_WARN, "failed to register meta-register %s: "
7705 		    "invalid ops", name);
7706 		return (EINVAL);
7707 	}
7708 
7709 	meta = kmem_zalloc(sizeof (dtrace_meta_t), KM_SLEEP);
7710 	meta->dtm_mops = *mops;
7711 	meta->dtm_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
7712 	(void) strcpy(meta->dtm_name, name);
7713 	meta->dtm_arg = arg;
7714 
7715 	mutex_enter(&dtrace_meta_lock);
7716 	mutex_enter(&dtrace_lock);
7717 
7718 	if (dtrace_meta_pid != NULL) {
7719 		mutex_exit(&dtrace_lock);
7720 		mutex_exit(&dtrace_meta_lock);
7721 		cmn_err(CE_WARN, "failed to register meta-register %s: "
7722 		    "user-land meta-provider exists", name);
7723 		kmem_free(meta->dtm_name, strlen(meta->dtm_name) + 1);
7724 		kmem_free(meta, sizeof (dtrace_meta_t));
7725 		return (EINVAL);
7726 	}
7727 
7728 	dtrace_meta_pid = meta;
7729 	*idp = (dtrace_meta_provider_id_t)meta;
7730 
7731 	/*
7732 	 * If there are providers and probes ready to go, pass them
7733 	 * off to the new meta provider now.
7734 	 */
7735 
7736 	help = dtrace_deferred_pid;
7737 	dtrace_deferred_pid = NULL;
7738 
7739 	mutex_exit(&dtrace_lock);
7740 
7741 	while (help != NULL) {
7742 		for (i = 0; i < help->dthps_nprovs; i++) {
7743 			dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
7744 			    help->dthps_pid);
7745 		}
7746 
7747 		next = help->dthps_next;
7748 		help->dthps_next = NULL;
7749 		help->dthps_prev = NULL;
7750 		help->dthps_deferred = 0;
7751 		help = next;
7752 	}
7753 
7754 	mutex_exit(&dtrace_meta_lock);
7755 
7756 	return (0);
7757 }
7758 
7759 int
7760 dtrace_meta_unregister(dtrace_meta_provider_id_t id)
7761 {
7762 	dtrace_meta_t **pp, *old = (dtrace_meta_t *)id;
7763 
7764 	mutex_enter(&dtrace_meta_lock);
7765 	mutex_enter(&dtrace_lock);
7766 
7767 	if (old == dtrace_meta_pid) {
7768 		pp = &dtrace_meta_pid;
7769 	} else {
7770 		panic("attempt to unregister non-existent "
7771 		    "dtrace meta-provider %p\n", (void *)old);
7772 	}
7773 
7774 	if (old->dtm_count != 0) {
7775 		mutex_exit(&dtrace_lock);
7776 		mutex_exit(&dtrace_meta_lock);
7777 		return (EBUSY);
7778 	}
7779 
7780 	*pp = NULL;
7781 
7782 	mutex_exit(&dtrace_lock);
7783 	mutex_exit(&dtrace_meta_lock);
7784 
7785 	kmem_free(old->dtm_name, strlen(old->dtm_name) + 1);
7786 	kmem_free(old, sizeof (dtrace_meta_t));
7787 
7788 	return (0);
7789 }
7790 
7791 
7792 /*
7793  * DTrace DIF Object Functions
7794  */
7795 static int
7796 dtrace_difo_err(uint_t pc, const char *format, ...)
7797 {
7798 	if (dtrace_err_verbose) {
7799 		va_list alist;
7800 
7801 		(void) uprintf("dtrace DIF object error: [%u]: ", pc);
7802 		va_start(alist, format);
7803 		(void) vuprintf(format, alist);
7804 		va_end(alist);
7805 	}
7806 
7807 #ifdef DTRACE_ERRDEBUG
7808 	dtrace_errdebug(format);
7809 #endif
7810 	return (1);
7811 }
7812 
7813 /*
7814  * Validate a DTrace DIF object by checking the IR instructions.  The following
7815  * rules are currently enforced by dtrace_difo_validate():
7816  *
7817  * 1. Each instruction must have a valid opcode
7818  * 2. Each register, string, variable, or subroutine reference must be valid
7819  * 3. No instruction can modify register %r0 (must be zero)
7820  * 4. All instruction reserved bits must be set to zero
7821  * 5. The last instruction must be a "ret" instruction
7822  * 6. All branch targets must reference a valid instruction _after_ the branch
7823  */
7824 static int
7825 dtrace_difo_validate(dtrace_difo_t *dp, dtrace_vstate_t *vstate, uint_t nregs,
7826     cred_t *cr)
7827 {
7828 	int err = 0, i;
7829 	int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
7830 	int kcheckload;
7831 	uint_t pc;
7832 
7833 	kcheckload = cr == NULL ||
7834 	    (vstate->dtvs_state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) == 0;
7835 
7836 	dp->dtdo_destructive = 0;
7837 
7838 	for (pc = 0; pc < dp->dtdo_len && err == 0; pc++) {
7839 		dif_instr_t instr = dp->dtdo_buf[pc];
7840 
7841 		uint_t r1 = DIF_INSTR_R1(instr);
7842 		uint_t r2 = DIF_INSTR_R2(instr);
7843 		uint_t rd = DIF_INSTR_RD(instr);
7844 		uint_t rs = DIF_INSTR_RS(instr);
7845 		uint_t label = DIF_INSTR_LABEL(instr);
7846 		uint_t v = DIF_INSTR_VAR(instr);
7847 		uint_t subr = DIF_INSTR_SUBR(instr);
7848 		uint_t type = DIF_INSTR_TYPE(instr);
7849 		uint_t op = DIF_INSTR_OP(instr);
7850 
7851 		switch (op) {
7852 		case DIF_OP_OR:
7853 		case DIF_OP_XOR:
7854 		case DIF_OP_AND:
7855 		case DIF_OP_SLL:
7856 		case DIF_OP_SRL:
7857 		case DIF_OP_SRA:
7858 		case DIF_OP_SUB:
7859 		case DIF_OP_ADD:
7860 		case DIF_OP_MUL:
7861 		case DIF_OP_SDIV:
7862 		case DIF_OP_UDIV:
7863 		case DIF_OP_SREM:
7864 		case DIF_OP_UREM:
7865 		case DIF_OP_COPYS:
7866 			if (r1 >= nregs)
7867 				err += efunc(pc, "invalid register %u\n", r1);
7868 			if (r2 >= nregs)
7869 				err += efunc(pc, "invalid register %u\n", r2);
7870 			if (rd >= nregs)
7871 				err += efunc(pc, "invalid register %u\n", rd);
7872 			if (rd == 0)
7873 				err += efunc(pc, "cannot write to %r0\n");
7874 			break;
7875 		case DIF_OP_NOT:
7876 		case DIF_OP_MOV:
7877 		case DIF_OP_ALLOCS:
7878 			if (r1 >= nregs)
7879 				err += efunc(pc, "invalid register %u\n", r1);
7880 			if (r2 != 0)
7881 				err += efunc(pc, "non-zero reserved bits\n");
7882 			if (rd >= nregs)
7883 				err += efunc(pc, "invalid register %u\n", rd);
7884 			if (rd == 0)
7885 				err += efunc(pc, "cannot write to %r0\n");
7886 			break;
7887 		case DIF_OP_LDSB:
7888 		case DIF_OP_LDSH:
7889 		case DIF_OP_LDSW:
7890 		case DIF_OP_LDUB:
7891 		case DIF_OP_LDUH:
7892 		case DIF_OP_LDUW:
7893 		case DIF_OP_LDX:
7894 			if (r1 >= nregs)
7895 				err += efunc(pc, "invalid register %u\n", r1);
7896 			if (r2 != 0)
7897 				err += efunc(pc, "non-zero reserved bits\n");
7898 			if (rd >= nregs)
7899 				err += efunc(pc, "invalid register %u\n", rd);
7900 			if (rd == 0)
7901 				err += efunc(pc, "cannot write to %r0\n");
7902 			if (kcheckload)
7903 				dp->dtdo_buf[pc] = DIF_INSTR_LOAD(op +
7904 				    DIF_OP_RLDSB - DIF_OP_LDSB, r1, rd);
7905 			break;
7906 		case DIF_OP_RLDSB:
7907 		case DIF_OP_RLDSH:
7908 		case DIF_OP_RLDSW:
7909 		case DIF_OP_RLDUB:
7910 		case DIF_OP_RLDUH:
7911 		case DIF_OP_RLDUW:
7912 		case DIF_OP_RLDX:
7913 			if (r1 >= nregs)
7914 				err += efunc(pc, "invalid register %u\n", r1);
7915 			if (r2 != 0)
7916 				err += efunc(pc, "non-zero reserved bits\n");
7917 			if (rd >= nregs)
7918 				err += efunc(pc, "invalid register %u\n", rd);
7919 			if (rd == 0)
7920 				err += efunc(pc, "cannot write to %r0\n");
7921 			break;
7922 		case DIF_OP_ULDSB:
7923 		case DIF_OP_ULDSH:
7924 		case DIF_OP_ULDSW:
7925 		case DIF_OP_ULDUB:
7926 		case DIF_OP_ULDUH:
7927 		case DIF_OP_ULDUW:
7928 		case DIF_OP_ULDX:
7929 			if (r1 >= nregs)
7930 				err += efunc(pc, "invalid register %u\n", r1);
7931 			if (r2 != 0)
7932 				err += efunc(pc, "non-zero reserved bits\n");
7933 			if (rd >= nregs)
7934 				err += efunc(pc, "invalid register %u\n", rd);
7935 			if (rd == 0)
7936 				err += efunc(pc, "cannot write to %r0\n");
7937 			break;
7938 		case DIF_OP_STB:
7939 		case DIF_OP_STH:
7940 		case DIF_OP_STW:
7941 		case DIF_OP_STX:
7942 			if (r1 >= nregs)
7943 				err += efunc(pc, "invalid register %u\n", r1);
7944 			if (r2 != 0)
7945 				err += efunc(pc, "non-zero reserved bits\n");
7946 			if (rd >= nregs)
7947 				err += efunc(pc, "invalid register %u\n", rd);
7948 			if (rd == 0)
7949 				err += efunc(pc, "cannot write to 0 address\n");
7950 			break;
7951 		case DIF_OP_CMP:
7952 		case DIF_OP_SCMP:
7953 			if (r1 >= nregs)
7954 				err += efunc(pc, "invalid register %u\n", r1);
7955 			if (r2 >= nregs)
7956 				err += efunc(pc, "invalid register %u\n", r2);
7957 			if (rd != 0)
7958 				err += efunc(pc, "non-zero reserved bits\n");
7959 			break;
7960 		case DIF_OP_TST:
7961 			if (r1 >= nregs)
7962 				err += efunc(pc, "invalid register %u\n", r1);
7963 			if (r2 != 0 || rd != 0)
7964 				err += efunc(pc, "non-zero reserved bits\n");
7965 			break;
7966 		case DIF_OP_BA:
7967 		case DIF_OP_BE:
7968 		case DIF_OP_BNE:
7969 		case DIF_OP_BG:
7970 		case DIF_OP_BGU:
7971 		case DIF_OP_BGE:
7972 		case DIF_OP_BGEU:
7973 		case DIF_OP_BL:
7974 		case DIF_OP_BLU:
7975 		case DIF_OP_BLE:
7976 		case DIF_OP_BLEU:
7977 			if (label >= dp->dtdo_len) {
7978 				err += efunc(pc, "invalid branch target %u\n",
7979 				    label);
7980 			}
7981 			if (label <= pc) {
7982 				err += efunc(pc, "backward branch to %u\n",
7983 				    label);
7984 			}
7985 			break;
7986 		case DIF_OP_RET:
7987 			if (r1 != 0 || r2 != 0)
7988 				err += efunc(pc, "non-zero reserved bits\n");
7989 			if (rd >= nregs)
7990 				err += efunc(pc, "invalid register %u\n", rd);
7991 			break;
7992 		case DIF_OP_NOP:
7993 		case DIF_OP_POPTS:
7994 		case DIF_OP_FLUSHTS:
7995 			if (r1 != 0 || r2 != 0 || rd != 0)
7996 				err += efunc(pc, "non-zero reserved bits\n");
7997 			break;
7998 		case DIF_OP_SETX:
7999 			if (DIF_INSTR_INTEGER(instr) >= dp->dtdo_intlen) {
8000 				err += efunc(pc, "invalid integer ref %u\n",
8001 				    DIF_INSTR_INTEGER(instr));
8002 			}
8003 			if (rd >= nregs)
8004 				err += efunc(pc, "invalid register %u\n", rd);
8005 			if (rd == 0)
8006 				err += efunc(pc, "cannot write to %r0\n");
8007 			break;
8008 		case DIF_OP_SETS:
8009 			if (DIF_INSTR_STRING(instr) >= dp->dtdo_strlen) {
8010 				err += efunc(pc, "invalid string ref %u\n",
8011 				    DIF_INSTR_STRING(instr));
8012 			}
8013 			if (rd >= nregs)
8014 				err += efunc(pc, "invalid register %u\n", rd);
8015 			if (rd == 0)
8016 				err += efunc(pc, "cannot write to %r0\n");
8017 			break;
8018 		case DIF_OP_LDGA:
8019 		case DIF_OP_LDTA:
8020 			if (r1 > DIF_VAR_ARRAY_MAX)
8021 				err += efunc(pc, "invalid array %u\n", r1);
8022 			if (r2 >= nregs)
8023 				err += efunc(pc, "invalid register %u\n", r2);
8024 			if (rd >= nregs)
8025 				err += efunc(pc, "invalid register %u\n", rd);
8026 			if (rd == 0)
8027 				err += efunc(pc, "cannot write to %r0\n");
8028 			break;
8029 		case DIF_OP_LDGS:
8030 		case DIF_OP_LDTS:
8031 		case DIF_OP_LDLS:
8032 		case DIF_OP_LDGAA:
8033 		case DIF_OP_LDTAA:
8034 			if (v < DIF_VAR_OTHER_MIN || v > DIF_VAR_OTHER_MAX)
8035 				err += efunc(pc, "invalid variable %u\n", v);
8036 			if (rd >= nregs)
8037 				err += efunc(pc, "invalid register %u\n", rd);
8038 			if (rd == 0)
8039 				err += efunc(pc, "cannot write to %r0\n");
8040 			break;
8041 		case DIF_OP_STGS:
8042 		case DIF_OP_STTS:
8043 		case DIF_OP_STLS:
8044 		case DIF_OP_STGAA:
8045 		case DIF_OP_STTAA:
8046 			if (v < DIF_VAR_OTHER_UBASE || v > DIF_VAR_OTHER_MAX)
8047 				err += efunc(pc, "invalid variable %u\n", v);
8048 			if (rs >= nregs)
8049 				err += efunc(pc, "invalid register %u\n", rd);
8050 			break;
8051 		case DIF_OP_CALL:
8052 			if (subr > DIF_SUBR_MAX)
8053 				err += efunc(pc, "invalid subr %u\n", subr);
8054 			if (rd >= nregs)
8055 				err += efunc(pc, "invalid register %u\n", rd);
8056 			if (rd == 0)
8057 				err += efunc(pc, "cannot write to %r0\n");
8058 
8059 			if (subr == DIF_SUBR_COPYOUT ||
8060 			    subr == DIF_SUBR_COPYOUTSTR) {
8061 				dp->dtdo_destructive = 1;
8062 			}
8063 			break;
8064 		case DIF_OP_PUSHTR:
8065 			if (type != DIF_TYPE_STRING && type != DIF_TYPE_CTF)
8066 				err += efunc(pc, "invalid ref type %u\n", type);
8067 			if (r2 >= nregs)
8068 				err += efunc(pc, "invalid register %u\n", r2);
8069 			if (rs >= nregs)
8070 				err += efunc(pc, "invalid register %u\n", rs);
8071 			break;
8072 		case DIF_OP_PUSHTV:
8073 			if (type != DIF_TYPE_CTF)
8074 				err += efunc(pc, "invalid val type %u\n", type);
8075 			if (r2 >= nregs)
8076 				err += efunc(pc, "invalid register %u\n", r2);
8077 			if (rs >= nregs)
8078 				err += efunc(pc, "invalid register %u\n", rs);
8079 			break;
8080 		default:
8081 			err += efunc(pc, "invalid opcode %u\n",
8082 			    DIF_INSTR_OP(instr));
8083 		}
8084 	}
8085 
8086 	if (dp->dtdo_len != 0 &&
8087 	    DIF_INSTR_OP(dp->dtdo_buf[dp->dtdo_len - 1]) != DIF_OP_RET) {
8088 		err += efunc(dp->dtdo_len - 1,
8089 		    "expected 'ret' as last DIF instruction\n");
8090 	}
8091 
8092 	if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF)) {
8093 		/*
8094 		 * If we're not returning by reference, the size must be either
8095 		 * 0 or the size of one of the base types.
8096 		 */
8097 		switch (dp->dtdo_rtype.dtdt_size) {
8098 		case 0:
8099 		case sizeof (uint8_t):
8100 		case sizeof (uint16_t):
8101 		case sizeof (uint32_t):
8102 		case sizeof (uint64_t):
8103 			break;
8104 
8105 		default:
8106 			err += efunc(dp->dtdo_len - 1, "bad return size");
8107 		}
8108 	}
8109 
8110 	for (i = 0; i < dp->dtdo_varlen && err == 0; i++) {
8111 		dtrace_difv_t *v = &dp->dtdo_vartab[i], *existing = NULL;
8112 		dtrace_diftype_t *vt, *et;
8113 		uint_t id, ndx;
8114 
8115 		if (v->dtdv_scope != DIFV_SCOPE_GLOBAL &&
8116 		    v->dtdv_scope != DIFV_SCOPE_THREAD &&
8117 		    v->dtdv_scope != DIFV_SCOPE_LOCAL) {
8118 			err += efunc(i, "unrecognized variable scope %d\n",
8119 			    v->dtdv_scope);
8120 			break;
8121 		}
8122 
8123 		if (v->dtdv_kind != DIFV_KIND_ARRAY &&
8124 		    v->dtdv_kind != DIFV_KIND_SCALAR) {
8125 			err += efunc(i, "unrecognized variable type %d\n",
8126 			    v->dtdv_kind);
8127 			break;
8128 		}
8129 
8130 		if ((id = v->dtdv_id) > DIF_VARIABLE_MAX) {
8131 			err += efunc(i, "%d exceeds variable id limit\n", id);
8132 			break;
8133 		}
8134 
8135 		if (id < DIF_VAR_OTHER_UBASE)
8136 			continue;
8137 
8138 		/*
8139 		 * For user-defined variables, we need to check that this
8140 		 * definition is identical to any previous definition that we
8141 		 * encountered.
8142 		 */
8143 		ndx = id - DIF_VAR_OTHER_UBASE;
8144 
8145 		switch (v->dtdv_scope) {
8146 		case DIFV_SCOPE_GLOBAL:
8147 			if (ndx < vstate->dtvs_nglobals) {
8148 				dtrace_statvar_t *svar;
8149 
8150 				if ((svar = vstate->dtvs_globals[ndx]) != NULL)
8151 					existing = &svar->dtsv_var;
8152 			}
8153 
8154 			break;
8155 
8156 		case DIFV_SCOPE_THREAD:
8157 			if (ndx < vstate->dtvs_ntlocals)
8158 				existing = &vstate->dtvs_tlocals[ndx];
8159 			break;
8160 
8161 		case DIFV_SCOPE_LOCAL:
8162 			if (ndx < vstate->dtvs_nlocals) {
8163 				dtrace_statvar_t *svar;
8164 
8165 				if ((svar = vstate->dtvs_locals[ndx]) != NULL)
8166 					existing = &svar->dtsv_var;
8167 			}
8168 
8169 			break;
8170 		}
8171 
8172 		vt = &v->dtdv_type;
8173 
8174 		if (vt->dtdt_flags & DIF_TF_BYREF) {
8175 			if (vt->dtdt_size == 0) {
8176 				err += efunc(i, "zero-sized variable\n");
8177 				break;
8178 			}
8179 
8180 			if (v->dtdv_scope == DIFV_SCOPE_GLOBAL &&
8181 			    vt->dtdt_size > dtrace_global_maxsize) {
8182 				err += efunc(i, "oversized by-ref global\n");
8183 				break;
8184 			}
8185 		}
8186 
8187 		if (existing == NULL || existing->dtdv_id == 0)
8188 			continue;
8189 
8190 		ASSERT(existing->dtdv_id == v->dtdv_id);
8191 		ASSERT(existing->dtdv_scope == v->dtdv_scope);
8192 
8193 		if (existing->dtdv_kind != v->dtdv_kind)
8194 			err += efunc(i, "%d changed variable kind\n", id);
8195 
8196 		et = &existing->dtdv_type;
8197 
8198 		if (vt->dtdt_flags != et->dtdt_flags) {
8199 			err += efunc(i, "%d changed variable type flags\n", id);
8200 			break;
8201 		}
8202 
8203 		if (vt->dtdt_size != 0 && vt->dtdt_size != et->dtdt_size) {
8204 			err += efunc(i, "%d changed variable type size\n", id);
8205 			break;
8206 		}
8207 	}
8208 
8209 	return (err);
8210 }
8211 
8212 /*
8213  * Validate a DTrace DIF object that it is to be used as a helper.  Helpers
8214  * are much more constrained than normal DIFOs.  Specifically, they may
8215  * not:
8216  *
8217  * 1. Make calls to subroutines other than copyin(), copyinstr() or
8218  *    miscellaneous string routines
8219  * 2. Access DTrace variables other than the args[] array, and the
8220  *    curthread, pid, ppid, tid, execname, zonename, uid and gid variables.
8221  * 3. Have thread-local variables.
8222  * 4. Have dynamic variables.
8223  */
8224 static int
8225 dtrace_difo_validate_helper(dtrace_difo_t *dp)
8226 {
8227 	int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
8228 	int err = 0;
8229 	uint_t pc;
8230 
8231 	for (pc = 0; pc < dp->dtdo_len; pc++) {
8232 		dif_instr_t instr = dp->dtdo_buf[pc];
8233 
8234 		uint_t v = DIF_INSTR_VAR(instr);
8235 		uint_t subr = DIF_INSTR_SUBR(instr);
8236 		uint_t op = DIF_INSTR_OP(instr);
8237 
8238 		switch (op) {
8239 		case DIF_OP_OR:
8240 		case DIF_OP_XOR:
8241 		case DIF_OP_AND:
8242 		case DIF_OP_SLL:
8243 		case DIF_OP_SRL:
8244 		case DIF_OP_SRA:
8245 		case DIF_OP_SUB:
8246 		case DIF_OP_ADD:
8247 		case DIF_OP_MUL:
8248 		case DIF_OP_SDIV:
8249 		case DIF_OP_UDIV:
8250 		case DIF_OP_SREM:
8251 		case DIF_OP_UREM:
8252 		case DIF_OP_COPYS:
8253 		case DIF_OP_NOT:
8254 		case DIF_OP_MOV:
8255 		case DIF_OP_RLDSB:
8256 		case DIF_OP_RLDSH:
8257 		case DIF_OP_RLDSW:
8258 		case DIF_OP_RLDUB:
8259 		case DIF_OP_RLDUH:
8260 		case DIF_OP_RLDUW:
8261 		case DIF_OP_RLDX:
8262 		case DIF_OP_ULDSB:
8263 		case DIF_OP_ULDSH:
8264 		case DIF_OP_ULDSW:
8265 		case DIF_OP_ULDUB:
8266 		case DIF_OP_ULDUH:
8267 		case DIF_OP_ULDUW:
8268 		case DIF_OP_ULDX:
8269 		case DIF_OP_STB:
8270 		case DIF_OP_STH:
8271 		case DIF_OP_STW:
8272 		case DIF_OP_STX:
8273 		case DIF_OP_ALLOCS:
8274 		case DIF_OP_CMP:
8275 		case DIF_OP_SCMP:
8276 		case DIF_OP_TST:
8277 		case DIF_OP_BA:
8278 		case DIF_OP_BE:
8279 		case DIF_OP_BNE:
8280 		case DIF_OP_BG:
8281 		case DIF_OP_BGU:
8282 		case DIF_OP_BGE:
8283 		case DIF_OP_BGEU:
8284 		case DIF_OP_BL:
8285 		case DIF_OP_BLU:
8286 		case DIF_OP_BLE:
8287 		case DIF_OP_BLEU:
8288 		case DIF_OP_RET:
8289 		case DIF_OP_NOP:
8290 		case DIF_OP_POPTS:
8291 		case DIF_OP_FLUSHTS:
8292 		case DIF_OP_SETX:
8293 		case DIF_OP_SETS:
8294 		case DIF_OP_LDGA:
8295 		case DIF_OP_LDLS:
8296 		case DIF_OP_STGS:
8297 		case DIF_OP_STLS:
8298 		case DIF_OP_PUSHTR:
8299 		case DIF_OP_PUSHTV:
8300 			break;
8301 
8302 		case DIF_OP_LDGS:
8303 			if (v >= DIF_VAR_OTHER_UBASE)
8304 				break;
8305 
8306 			if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9)
8307 				break;
8308 
8309 			if (v == DIF_VAR_CURTHREAD || v == DIF_VAR_PID ||
8310 			    v == DIF_VAR_PPID || v == DIF_VAR_TID ||
8311 			    v == DIF_VAR_EXECNAME || v == DIF_VAR_ZONENAME ||
8312 			    v == DIF_VAR_UID || v == DIF_VAR_GID)
8313 				break;
8314 
8315 			err += efunc(pc, "illegal variable %u\n", v);
8316 			break;
8317 
8318 		case DIF_OP_LDTA:
8319 		case DIF_OP_LDTS:
8320 		case DIF_OP_LDGAA:
8321 		case DIF_OP_LDTAA:
8322 			err += efunc(pc, "illegal dynamic variable load\n");
8323 			break;
8324 
8325 		case DIF_OP_STTS:
8326 		case DIF_OP_STGAA:
8327 		case DIF_OP_STTAA:
8328 			err += efunc(pc, "illegal dynamic variable store\n");
8329 			break;
8330 
8331 		case DIF_OP_CALL:
8332 			if (subr == DIF_SUBR_ALLOCA ||
8333 			    subr == DIF_SUBR_BCOPY ||
8334 			    subr == DIF_SUBR_COPYIN ||
8335 			    subr == DIF_SUBR_COPYINTO ||
8336 			    subr == DIF_SUBR_COPYINSTR ||
8337 			    subr == DIF_SUBR_INDEX ||
8338 			    subr == DIF_SUBR_INET_NTOA ||
8339 			    subr == DIF_SUBR_INET_NTOA6 ||
8340 			    subr == DIF_SUBR_INET_NTOP ||
8341 			    subr == DIF_SUBR_LLTOSTR ||
8342 			    subr == DIF_SUBR_RINDEX ||
8343 			    subr == DIF_SUBR_STRCHR ||
8344 			    subr == DIF_SUBR_STRJOIN ||
8345 			    subr == DIF_SUBR_STRRCHR ||
8346 			    subr == DIF_SUBR_STRSTR ||
8347 			    subr == DIF_SUBR_HTONS ||
8348 			    subr == DIF_SUBR_HTONL ||
8349 			    subr == DIF_SUBR_HTONLL ||
8350 			    subr == DIF_SUBR_NTOHS ||
8351 			    subr == DIF_SUBR_NTOHL ||
8352 			    subr == DIF_SUBR_NTOHLL)
8353 				break;
8354 
8355 			err += efunc(pc, "invalid subr %u\n", subr);
8356 			break;
8357 
8358 		default:
8359 			err += efunc(pc, "invalid opcode %u\n",
8360 			    DIF_INSTR_OP(instr));
8361 		}
8362 	}
8363 
8364 	return (err);
8365 }
8366 
8367 /*
8368  * Returns 1 if the expression in the DIF object can be cached on a per-thread
8369  * basis; 0 if not.
8370  */
8371 static int
8372 dtrace_difo_cacheable(dtrace_difo_t *dp)
8373 {
8374 	int i;
8375 
8376 	if (dp == NULL)
8377 		return (0);
8378 
8379 	for (i = 0; i < dp->dtdo_varlen; i++) {
8380 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
8381 
8382 		if (v->dtdv_scope != DIFV_SCOPE_GLOBAL)
8383 			continue;
8384 
8385 		switch (v->dtdv_id) {
8386 		case DIF_VAR_CURTHREAD:
8387 		case DIF_VAR_PID:
8388 		case DIF_VAR_TID:
8389 		case DIF_VAR_EXECNAME:
8390 		case DIF_VAR_ZONENAME:
8391 			break;
8392 
8393 		default:
8394 			return (0);
8395 		}
8396 	}
8397 
8398 	/*
8399 	 * This DIF object may be cacheable.  Now we need to look for any
8400 	 * array loading instructions, any memory loading instructions, or
8401 	 * any stores to thread-local variables.
8402 	 */
8403 	for (i = 0; i < dp->dtdo_len; i++) {
8404 		uint_t op = DIF_INSTR_OP(dp->dtdo_buf[i]);
8405 
8406 		if ((op >= DIF_OP_LDSB && op <= DIF_OP_LDX) ||
8407 		    (op >= DIF_OP_ULDSB && op <= DIF_OP_ULDX) ||
8408 		    (op >= DIF_OP_RLDSB && op <= DIF_OP_RLDX) ||
8409 		    op == DIF_OP_LDGA || op == DIF_OP_STTS)
8410 			return (0);
8411 	}
8412 
8413 	return (1);
8414 }
8415 
8416 static void
8417 dtrace_difo_hold(dtrace_difo_t *dp)
8418 {
8419 	int i;
8420 
8421 	ASSERT(MUTEX_HELD(&dtrace_lock));
8422 
8423 	dp->dtdo_refcnt++;
8424 	ASSERT(dp->dtdo_refcnt != 0);
8425 
8426 	/*
8427 	 * We need to check this DIF object for references to the variable
8428 	 * DIF_VAR_VTIMESTAMP.
8429 	 */
8430 	for (i = 0; i < dp->dtdo_varlen; i++) {
8431 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
8432 
8433 		if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
8434 			continue;
8435 
8436 		if (dtrace_vtime_references++ == 0)
8437 			dtrace_vtime_enable();
8438 	}
8439 }
8440 
8441 /*
8442  * This routine calculates the dynamic variable chunksize for a given DIF
8443  * object.  The calculation is not fool-proof, and can probably be tricked by
8444  * malicious DIF -- but it works for all compiler-generated DIF.  Because this
8445  * calculation is likely imperfect, dtrace_dynvar() is able to gracefully fail
8446  * if a dynamic variable size exceeds the chunksize.
8447  */
8448 static void
8449 dtrace_difo_chunksize(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
8450 {
8451 	uint64_t sval;
8452 	dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
8453 	const dif_instr_t *text = dp->dtdo_buf;
8454 	uint_t pc, srd = 0;
8455 	uint_t ttop = 0;
8456 	size_t size, ksize;
8457 	uint_t id, i;
8458 
8459 	for (pc = 0; pc < dp->dtdo_len; pc++) {
8460 		dif_instr_t instr = text[pc];
8461 		uint_t op = DIF_INSTR_OP(instr);
8462 		uint_t rd = DIF_INSTR_RD(instr);
8463 		uint_t r1 = DIF_INSTR_R1(instr);
8464 		uint_t nkeys = 0;
8465 		uchar_t scope;
8466 
8467 		dtrace_key_t *key = tupregs;
8468 
8469 		switch (op) {
8470 		case DIF_OP_SETX:
8471 			sval = dp->dtdo_inttab[DIF_INSTR_INTEGER(instr)];
8472 			srd = rd;
8473 			continue;
8474 
8475 		case DIF_OP_STTS:
8476 			key = &tupregs[DIF_DTR_NREGS];
8477 			key[0].dttk_size = 0;
8478 			key[1].dttk_size = 0;
8479 			nkeys = 2;
8480 			scope = DIFV_SCOPE_THREAD;
8481 			break;
8482 
8483 		case DIF_OP_STGAA:
8484 		case DIF_OP_STTAA:
8485 			nkeys = ttop;
8486 
8487 			if (DIF_INSTR_OP(instr) == DIF_OP_STTAA)
8488 				key[nkeys++].dttk_size = 0;
8489 
8490 			key[nkeys++].dttk_size = 0;
8491 
8492 			if (op == DIF_OP_STTAA) {
8493 				scope = DIFV_SCOPE_THREAD;
8494 			} else {
8495 				scope = DIFV_SCOPE_GLOBAL;
8496 			}
8497 
8498 			break;
8499 
8500 		case DIF_OP_PUSHTR:
8501 			if (ttop == DIF_DTR_NREGS)
8502 				return;
8503 
8504 			if ((srd == 0 || sval == 0) && r1 == DIF_TYPE_STRING) {
8505 				/*
8506 				 * If the register for the size of the "pushtr"
8507 				 * is %r0 (or the value is 0) and the type is
8508 				 * a string, we'll use the system-wide default
8509 				 * string size.
8510 				 */
8511 				tupregs[ttop++].dttk_size =
8512 				    dtrace_strsize_default;
8513 			} else {
8514 				if (srd == 0)
8515 					return;
8516 
8517 				tupregs[ttop++].dttk_size = sval;
8518 			}
8519 
8520 			break;
8521 
8522 		case DIF_OP_PUSHTV:
8523 			if (ttop == DIF_DTR_NREGS)
8524 				return;
8525 
8526 			tupregs[ttop++].dttk_size = 0;
8527 			break;
8528 
8529 		case DIF_OP_FLUSHTS:
8530 			ttop = 0;
8531 			break;
8532 
8533 		case DIF_OP_POPTS:
8534 			if (ttop != 0)
8535 				ttop--;
8536 			break;
8537 		}
8538 
8539 		sval = 0;
8540 		srd = 0;
8541 
8542 		if (nkeys == 0)
8543 			continue;
8544 
8545 		/*
8546 		 * We have a dynamic variable allocation; calculate its size.
8547 		 */
8548 		for (ksize = 0, i = 0; i < nkeys; i++)
8549 			ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
8550 
8551 		size = sizeof (dtrace_dynvar_t);
8552 		size += sizeof (dtrace_key_t) * (nkeys - 1);
8553 		size += ksize;
8554 
8555 		/*
8556 		 * Now we need to determine the size of the stored data.
8557 		 */
8558 		id = DIF_INSTR_VAR(instr);
8559 
8560 		for (i = 0; i < dp->dtdo_varlen; i++) {
8561 			dtrace_difv_t *v = &dp->dtdo_vartab[i];
8562 
8563 			if (v->dtdv_id == id && v->dtdv_scope == scope) {
8564 				size += v->dtdv_type.dtdt_size;
8565 				break;
8566 			}
8567 		}
8568 
8569 		if (i == dp->dtdo_varlen)
8570 			return;
8571 
8572 		/*
8573 		 * We have the size.  If this is larger than the chunk size
8574 		 * for our dynamic variable state, reset the chunk size.
8575 		 */
8576 		size = P2ROUNDUP(size, sizeof (uint64_t));
8577 
8578 		if (size > vstate->dtvs_dynvars.dtds_chunksize)
8579 			vstate->dtvs_dynvars.dtds_chunksize = size;
8580 	}
8581 }
8582 
8583 static void
8584 dtrace_difo_init(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
8585 {
8586 	int i, oldsvars, osz, nsz, otlocals, ntlocals;
8587 	uint_t id;
8588 
8589 	ASSERT(MUTEX_HELD(&dtrace_lock));
8590 	ASSERT(dp->dtdo_buf != NULL && dp->dtdo_len != 0);
8591 
8592 	for (i = 0; i < dp->dtdo_varlen; i++) {
8593 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
8594 		dtrace_statvar_t *svar, ***svarp;
8595 		size_t dsize = 0;
8596 		uint8_t scope = v->dtdv_scope;
8597 		int *np;
8598 
8599 		if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
8600 			continue;
8601 
8602 		id -= DIF_VAR_OTHER_UBASE;
8603 
8604 		switch (scope) {
8605 		case DIFV_SCOPE_THREAD:
8606 			while (id >= (otlocals = vstate->dtvs_ntlocals)) {
8607 				dtrace_difv_t *tlocals;
8608 
8609 				if ((ntlocals = (otlocals << 1)) == 0)
8610 					ntlocals = 1;
8611 
8612 				osz = otlocals * sizeof (dtrace_difv_t);
8613 				nsz = ntlocals * sizeof (dtrace_difv_t);
8614 
8615 				tlocals = kmem_zalloc(nsz, KM_SLEEP);
8616 
8617 				if (osz != 0) {
8618 					bcopy(vstate->dtvs_tlocals,
8619 					    tlocals, osz);
8620 					kmem_free(vstate->dtvs_tlocals, osz);
8621 				}
8622 
8623 				vstate->dtvs_tlocals = tlocals;
8624 				vstate->dtvs_ntlocals = ntlocals;
8625 			}
8626 
8627 			vstate->dtvs_tlocals[id] = *v;
8628 			continue;
8629 
8630 		case DIFV_SCOPE_LOCAL:
8631 			np = &vstate->dtvs_nlocals;
8632 			svarp = &vstate->dtvs_locals;
8633 
8634 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
8635 				dsize = NCPU * (v->dtdv_type.dtdt_size +
8636 				    sizeof (uint64_t));
8637 			else
8638 				dsize = NCPU * sizeof (uint64_t);
8639 
8640 			break;
8641 
8642 		case DIFV_SCOPE_GLOBAL:
8643 			np = &vstate->dtvs_nglobals;
8644 			svarp = &vstate->dtvs_globals;
8645 
8646 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
8647 				dsize = v->dtdv_type.dtdt_size +
8648 				    sizeof (uint64_t);
8649 
8650 			break;
8651 
8652 		default:
8653 			ASSERT(0);
8654 		}
8655 
8656 		while (id >= (oldsvars = *np)) {
8657 			dtrace_statvar_t **statics;
8658 			int newsvars, oldsize, newsize;
8659 
8660 			if ((newsvars = (oldsvars << 1)) == 0)
8661 				newsvars = 1;
8662 
8663 			oldsize = oldsvars * sizeof (dtrace_statvar_t *);
8664 			newsize = newsvars * sizeof (dtrace_statvar_t *);
8665 
8666 			statics = kmem_zalloc(newsize, KM_SLEEP);
8667 
8668 			if (oldsize != 0) {
8669 				bcopy(*svarp, statics, oldsize);
8670 				kmem_free(*svarp, oldsize);
8671 			}
8672 
8673 			*svarp = statics;
8674 			*np = newsvars;
8675 		}
8676 
8677 		if ((svar = (*svarp)[id]) == NULL) {
8678 			svar = kmem_zalloc(sizeof (dtrace_statvar_t), KM_SLEEP);
8679 			svar->dtsv_var = *v;
8680 
8681 			if ((svar->dtsv_size = dsize) != 0) {
8682 				svar->dtsv_data = (uint64_t)(uintptr_t)
8683 				    kmem_zalloc(dsize, KM_SLEEP);
8684 			}
8685 
8686 			(*svarp)[id] = svar;
8687 		}
8688 
8689 		svar->dtsv_refcnt++;
8690 	}
8691 
8692 	dtrace_difo_chunksize(dp, vstate);
8693 	dtrace_difo_hold(dp);
8694 }
8695 
8696 static dtrace_difo_t *
8697 dtrace_difo_duplicate(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
8698 {
8699 	dtrace_difo_t *new;
8700 	size_t sz;
8701 
8702 	ASSERT(dp->dtdo_buf != NULL);
8703 	ASSERT(dp->dtdo_refcnt != 0);
8704 
8705 	new = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
8706 
8707 	ASSERT(dp->dtdo_buf != NULL);
8708 	sz = dp->dtdo_len * sizeof (dif_instr_t);
8709 	new->dtdo_buf = kmem_alloc(sz, KM_SLEEP);
8710 	bcopy(dp->dtdo_buf, new->dtdo_buf, sz);
8711 	new->dtdo_len = dp->dtdo_len;
8712 
8713 	if (dp->dtdo_strtab != NULL) {
8714 		ASSERT(dp->dtdo_strlen != 0);
8715 		new->dtdo_strtab = kmem_alloc(dp->dtdo_strlen, KM_SLEEP);
8716 		bcopy(dp->dtdo_strtab, new->dtdo_strtab, dp->dtdo_strlen);
8717 		new->dtdo_strlen = dp->dtdo_strlen;
8718 	}
8719 
8720 	if (dp->dtdo_inttab != NULL) {
8721 		ASSERT(dp->dtdo_intlen != 0);
8722 		sz = dp->dtdo_intlen * sizeof (uint64_t);
8723 		new->dtdo_inttab = kmem_alloc(sz, KM_SLEEP);
8724 		bcopy(dp->dtdo_inttab, new->dtdo_inttab, sz);
8725 		new->dtdo_intlen = dp->dtdo_intlen;
8726 	}
8727 
8728 	if (dp->dtdo_vartab != NULL) {
8729 		ASSERT(dp->dtdo_varlen != 0);
8730 		sz = dp->dtdo_varlen * sizeof (dtrace_difv_t);
8731 		new->dtdo_vartab = kmem_alloc(sz, KM_SLEEP);
8732 		bcopy(dp->dtdo_vartab, new->dtdo_vartab, sz);
8733 		new->dtdo_varlen = dp->dtdo_varlen;
8734 	}
8735 
8736 	dtrace_difo_init(new, vstate);
8737 	return (new);
8738 }
8739 
8740 static void
8741 dtrace_difo_destroy(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
8742 {
8743 	int i;
8744 
8745 	ASSERT(dp->dtdo_refcnt == 0);
8746 
8747 	for (i = 0; i < dp->dtdo_varlen; i++) {
8748 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
8749 		dtrace_statvar_t *svar, **svarp;
8750 		uint_t id;
8751 		uint8_t scope = v->dtdv_scope;
8752 		int *np;
8753 
8754 		switch (scope) {
8755 		case DIFV_SCOPE_THREAD:
8756 			continue;
8757 
8758 		case DIFV_SCOPE_LOCAL:
8759 			np = &vstate->dtvs_nlocals;
8760 			svarp = vstate->dtvs_locals;
8761 			break;
8762 
8763 		case DIFV_SCOPE_GLOBAL:
8764 			np = &vstate->dtvs_nglobals;
8765 			svarp = vstate->dtvs_globals;
8766 			break;
8767 
8768 		default:
8769 			ASSERT(0);
8770 		}
8771 
8772 		if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
8773 			continue;
8774 
8775 		id -= DIF_VAR_OTHER_UBASE;
8776 		ASSERT(id < *np);
8777 
8778 		svar = svarp[id];
8779 		ASSERT(svar != NULL);
8780 		ASSERT(svar->dtsv_refcnt > 0);
8781 
8782 		if (--svar->dtsv_refcnt > 0)
8783 			continue;
8784 
8785 		if (svar->dtsv_size != 0) {
8786 			ASSERT(svar->dtsv_data != NULL);
8787 			kmem_free((void *)(uintptr_t)svar->dtsv_data,
8788 			    svar->dtsv_size);
8789 		}
8790 
8791 		kmem_free(svar, sizeof (dtrace_statvar_t));
8792 		svarp[id] = NULL;
8793 	}
8794 
8795 	kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
8796 	kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
8797 	kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
8798 	kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
8799 
8800 	kmem_free(dp, sizeof (dtrace_difo_t));
8801 }
8802 
8803 static void
8804 dtrace_difo_release(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
8805 {
8806 	int i;
8807 
8808 	ASSERT(MUTEX_HELD(&dtrace_lock));
8809 	ASSERT(dp->dtdo_refcnt != 0);
8810 
8811 	for (i = 0; i < dp->dtdo_varlen; i++) {
8812 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
8813 
8814 		if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
8815 			continue;
8816 
8817 		ASSERT(dtrace_vtime_references > 0);
8818 		if (--dtrace_vtime_references == 0)
8819 			dtrace_vtime_disable();
8820 	}
8821 
8822 	if (--dp->dtdo_refcnt == 0)
8823 		dtrace_difo_destroy(dp, vstate);
8824 }
8825 
8826 /*
8827  * DTrace Format Functions
8828  */
8829 static uint16_t
8830 dtrace_format_add(dtrace_state_t *state, char *str)
8831 {
8832 	char *fmt, **new;
8833 	uint16_t ndx, len = strlen(str) + 1;
8834 
8835 	fmt = kmem_zalloc(len, KM_SLEEP);
8836 	bcopy(str, fmt, len);
8837 
8838 	for (ndx = 0; ndx < state->dts_nformats; ndx++) {
8839 		if (state->dts_formats[ndx] == NULL) {
8840 			state->dts_formats[ndx] = fmt;
8841 			return (ndx + 1);
8842 		}
8843 	}
8844 
8845 	if (state->dts_nformats == USHRT_MAX) {
8846 		/*
8847 		 * This is only likely if a denial-of-service attack is being
8848 		 * attempted.  As such, it's okay to fail silently here.
8849 		 */
8850 		kmem_free(fmt, len);
8851 		return (0);
8852 	}
8853 
8854 	/*
8855 	 * For simplicity, we always resize the formats array to be exactly the
8856 	 * number of formats.
8857 	 */
8858 	ndx = state->dts_nformats++;
8859 	new = kmem_alloc((ndx + 1) * sizeof (char *), KM_SLEEP);
8860 
8861 	if (state->dts_formats != NULL) {
8862 		ASSERT(ndx != 0);
8863 		bcopy(state->dts_formats, new, ndx * sizeof (char *));
8864 		kmem_free(state->dts_formats, ndx * sizeof (char *));
8865 	}
8866 
8867 	state->dts_formats = new;
8868 	state->dts_formats[ndx] = fmt;
8869 
8870 	return (ndx + 1);
8871 }
8872 
8873 static void
8874 dtrace_format_remove(dtrace_state_t *state, uint16_t format)
8875 {
8876 	char *fmt;
8877 
8878 	ASSERT(state->dts_formats != NULL);
8879 	ASSERT(format <= state->dts_nformats);
8880 	ASSERT(state->dts_formats[format - 1] != NULL);
8881 
8882 	fmt = state->dts_formats[format - 1];
8883 	kmem_free(fmt, strlen(fmt) + 1);
8884 	state->dts_formats[format - 1] = NULL;
8885 }
8886 
8887 static void
8888 dtrace_format_destroy(dtrace_state_t *state)
8889 {
8890 	int i;
8891 
8892 	if (state->dts_nformats == 0) {
8893 		ASSERT(state->dts_formats == NULL);
8894 		return;
8895 	}
8896 
8897 	ASSERT(state->dts_formats != NULL);
8898 
8899 	for (i = 0; i < state->dts_nformats; i++) {
8900 		char *fmt = state->dts_formats[i];
8901 
8902 		if (fmt == NULL)
8903 			continue;
8904 
8905 		kmem_free(fmt, strlen(fmt) + 1);
8906 	}
8907 
8908 	kmem_free(state->dts_formats, state->dts_nformats * sizeof (char *));
8909 	state->dts_nformats = 0;
8910 	state->dts_formats = NULL;
8911 }
8912 
8913 /*
8914  * DTrace Predicate Functions
8915  */
8916 static dtrace_predicate_t *
8917 dtrace_predicate_create(dtrace_difo_t *dp)
8918 {
8919 	dtrace_predicate_t *pred;
8920 
8921 	ASSERT(MUTEX_HELD(&dtrace_lock));
8922 	ASSERT(dp->dtdo_refcnt != 0);
8923 
8924 	pred = kmem_zalloc(sizeof (dtrace_predicate_t), KM_SLEEP);
8925 	pred->dtp_difo = dp;
8926 	pred->dtp_refcnt = 1;
8927 
8928 	if (!dtrace_difo_cacheable(dp))
8929 		return (pred);
8930 
8931 	if (dtrace_predcache_id == DTRACE_CACHEIDNONE) {
8932 		/*
8933 		 * This is only theoretically possible -- we have had 2^32
8934 		 * cacheable predicates on this machine.  We cannot allow any
8935 		 * more predicates to become cacheable:  as unlikely as it is,
8936 		 * there may be a thread caching a (now stale) predicate cache
8937 		 * ID. (N.B.: the temptation is being successfully resisted to
8938 		 * have this cmn_err() "Holy shit -- we executed this code!")
8939 		 */
8940 		return (pred);
8941 	}
8942 
8943 	pred->dtp_cacheid = dtrace_predcache_id++;
8944 
8945 	return (pred);
8946 }
8947 
8948 static void
8949 dtrace_predicate_hold(dtrace_predicate_t *pred)
8950 {
8951 	ASSERT(MUTEX_HELD(&dtrace_lock));
8952 	ASSERT(pred->dtp_difo != NULL && pred->dtp_difo->dtdo_refcnt != 0);
8953 	ASSERT(pred->dtp_refcnt > 0);
8954 
8955 	pred->dtp_refcnt++;
8956 }
8957 
8958 static void
8959 dtrace_predicate_release(dtrace_predicate_t *pred, dtrace_vstate_t *vstate)
8960 {
8961 	dtrace_difo_t *dp = pred->dtp_difo;
8962 
8963 	ASSERT(MUTEX_HELD(&dtrace_lock));
8964 	ASSERT(dp != NULL && dp->dtdo_refcnt != 0);
8965 	ASSERT(pred->dtp_refcnt > 0);
8966 
8967 	if (--pred->dtp_refcnt == 0) {
8968 		dtrace_difo_release(pred->dtp_difo, vstate);
8969 		kmem_free(pred, sizeof (dtrace_predicate_t));
8970 	}
8971 }
8972 
8973 /*
8974  * DTrace Action Description Functions
8975  */
8976 static dtrace_actdesc_t *
8977 dtrace_actdesc_create(dtrace_actkind_t kind, uint32_t ntuple,
8978     uint64_t uarg, uint64_t arg)
8979 {
8980 	dtrace_actdesc_t *act;
8981 
8982 	ASSERT(!DTRACEACT_ISPRINTFLIKE(kind) || (arg != NULL &&
8983 	    arg >= KERNELBASE) || (arg == NULL && kind == DTRACEACT_PRINTA));
8984 
8985 	act = kmem_zalloc(sizeof (dtrace_actdesc_t), KM_SLEEP);
8986 	act->dtad_kind = kind;
8987 	act->dtad_ntuple = ntuple;
8988 	act->dtad_uarg = uarg;
8989 	act->dtad_arg = arg;
8990 	act->dtad_refcnt = 1;
8991 
8992 	return (act);
8993 }
8994 
8995 static void
8996 dtrace_actdesc_hold(dtrace_actdesc_t *act)
8997 {
8998 	ASSERT(act->dtad_refcnt >= 1);
8999 	act->dtad_refcnt++;
9000 }
9001 
9002 static void
9003 dtrace_actdesc_release(dtrace_actdesc_t *act, dtrace_vstate_t *vstate)
9004 {
9005 	dtrace_actkind_t kind = act->dtad_kind;
9006 	dtrace_difo_t *dp;
9007 
9008 	ASSERT(act->dtad_refcnt >= 1);
9009 
9010 	if (--act->dtad_refcnt != 0)
9011 		return;
9012 
9013 	if ((dp = act->dtad_difo) != NULL)
9014 		dtrace_difo_release(dp, vstate);
9015 
9016 	if (DTRACEACT_ISPRINTFLIKE(kind)) {
9017 		char *str = (char *)(uintptr_t)act->dtad_arg;
9018 
9019 		ASSERT((str != NULL && (uintptr_t)str >= KERNELBASE) ||
9020 		    (str == NULL && act->dtad_kind == DTRACEACT_PRINTA));
9021 
9022 		if (str != NULL)
9023 			kmem_free(str, strlen(str) + 1);
9024 	}
9025 
9026 	kmem_free(act, sizeof (dtrace_actdesc_t));
9027 }
9028 
9029 /*
9030  * DTrace ECB Functions
9031  */
9032 static dtrace_ecb_t *
9033 dtrace_ecb_add(dtrace_state_t *state, dtrace_probe_t *probe)
9034 {
9035 	dtrace_ecb_t *ecb;
9036 	dtrace_epid_t epid;
9037 
9038 	ASSERT(MUTEX_HELD(&dtrace_lock));
9039 
9040 	ecb = kmem_zalloc(sizeof (dtrace_ecb_t), KM_SLEEP);
9041 	ecb->dte_predicate = NULL;
9042 	ecb->dte_probe = probe;
9043 
9044 	/*
9045 	 * The default size is the size of the default action: recording
9046 	 * the epid.
9047 	 */
9048 	ecb->dte_size = ecb->dte_needed = sizeof (dtrace_epid_t);
9049 	ecb->dte_alignment = sizeof (dtrace_epid_t);
9050 
9051 	epid = state->dts_epid++;
9052 
9053 	if (epid - 1 >= state->dts_necbs) {
9054 		dtrace_ecb_t **oecbs = state->dts_ecbs, **ecbs;
9055 		int necbs = state->dts_necbs << 1;
9056 
9057 		ASSERT(epid == state->dts_necbs + 1);
9058 
9059 		if (necbs == 0) {
9060 			ASSERT(oecbs == NULL);
9061 			necbs = 1;
9062 		}
9063 
9064 		ecbs = kmem_zalloc(necbs * sizeof (*ecbs), KM_SLEEP);
9065 
9066 		if (oecbs != NULL)
9067 			bcopy(oecbs, ecbs, state->dts_necbs * sizeof (*ecbs));
9068 
9069 		dtrace_membar_producer();
9070 		state->dts_ecbs = ecbs;
9071 
9072 		if (oecbs != NULL) {
9073 			/*
9074 			 * If this state is active, we must dtrace_sync()
9075 			 * before we can free the old dts_ecbs array:  we're
9076 			 * coming in hot, and there may be active ring
9077 			 * buffer processing (which indexes into the dts_ecbs
9078 			 * array) on another CPU.
9079 			 */
9080 			if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
9081 				dtrace_sync();
9082 
9083 			kmem_free(oecbs, state->dts_necbs * sizeof (*ecbs));
9084 		}
9085 
9086 		dtrace_membar_producer();
9087 		state->dts_necbs = necbs;
9088 	}
9089 
9090 	ecb->dte_state = state;
9091 
9092 	ASSERT(state->dts_ecbs[epid - 1] == NULL);
9093 	dtrace_membar_producer();
9094 	state->dts_ecbs[(ecb->dte_epid = epid) - 1] = ecb;
9095 
9096 	return (ecb);
9097 }
9098 
9099 static void
9100 dtrace_ecb_enable(dtrace_ecb_t *ecb)
9101 {
9102 	dtrace_probe_t *probe = ecb->dte_probe;
9103 
9104 	ASSERT(MUTEX_HELD(&cpu_lock));
9105 	ASSERT(MUTEX_HELD(&dtrace_lock));
9106 	ASSERT(ecb->dte_next == NULL);
9107 
9108 	if (probe == NULL) {
9109 		/*
9110 		 * This is the NULL probe -- there's nothing to do.
9111 		 */
9112 		return;
9113 	}
9114 
9115 	if (probe->dtpr_ecb == NULL) {
9116 		dtrace_provider_t *prov = probe->dtpr_provider;
9117 
9118 		/*
9119 		 * We're the first ECB on this probe.
9120 		 */
9121 		probe->dtpr_ecb = probe->dtpr_ecb_last = ecb;
9122 
9123 		if (ecb->dte_predicate != NULL)
9124 			probe->dtpr_predcache = ecb->dte_predicate->dtp_cacheid;
9125 
9126 		prov->dtpv_pops.dtps_enable(prov->dtpv_arg,
9127 		    probe->dtpr_id, probe->dtpr_arg);
9128 	} else {
9129 		/*
9130 		 * This probe is already active.  Swing the last pointer to
9131 		 * point to the new ECB, and issue a dtrace_sync() to assure
9132 		 * that all CPUs have seen the change.
9133 		 */
9134 		ASSERT(probe->dtpr_ecb_last != NULL);
9135 		probe->dtpr_ecb_last->dte_next = ecb;
9136 		probe->dtpr_ecb_last = ecb;
9137 		probe->dtpr_predcache = 0;
9138 
9139 		dtrace_sync();
9140 	}
9141 }
9142 
9143 static void
9144 dtrace_ecb_resize(dtrace_ecb_t *ecb)
9145 {
9146 	uint32_t maxalign = sizeof (dtrace_epid_t);
9147 	uint32_t align = sizeof (uint8_t), offs, diff;
9148 	dtrace_action_t *act;
9149 	int wastuple = 0;
9150 	uint32_t aggbase = UINT32_MAX;
9151 	dtrace_state_t *state = ecb->dte_state;
9152 
9153 	/*
9154 	 * If we record anything, we always record the epid.  (And we always
9155 	 * record it first.)
9156 	 */
9157 	offs = sizeof (dtrace_epid_t);
9158 	ecb->dte_size = ecb->dte_needed = sizeof (dtrace_epid_t);
9159 
9160 	for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
9161 		dtrace_recdesc_t *rec = &act->dta_rec;
9162 
9163 		if ((align = rec->dtrd_alignment) > maxalign)
9164 			maxalign = align;
9165 
9166 		if (!wastuple && act->dta_intuple) {
9167 			/*
9168 			 * This is the first record in a tuple.  Align the
9169 			 * offset to be at offset 4 in an 8-byte aligned
9170 			 * block.
9171 			 */
9172 			diff = offs + sizeof (dtrace_aggid_t);
9173 
9174 			if (diff = (diff & (sizeof (uint64_t) - 1)))
9175 				offs += sizeof (uint64_t) - diff;
9176 
9177 			aggbase = offs - sizeof (dtrace_aggid_t);
9178 			ASSERT(!(aggbase & (sizeof (uint64_t) - 1)));
9179 		}
9180 
9181 		/*LINTED*/
9182 		if (rec->dtrd_size != 0 && (diff = (offs & (align - 1)))) {
9183 			/*
9184 			 * The current offset is not properly aligned; align it.
9185 			 */
9186 			offs += align - diff;
9187 		}
9188 
9189 		rec->dtrd_offset = offs;
9190 
9191 		if (offs + rec->dtrd_size > ecb->dte_needed) {
9192 			ecb->dte_needed = offs + rec->dtrd_size;
9193 
9194 			if (ecb->dte_needed > state->dts_needed)
9195 				state->dts_needed = ecb->dte_needed;
9196 		}
9197 
9198 		if (DTRACEACT_ISAGG(act->dta_kind)) {
9199 			dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
9200 			dtrace_action_t *first = agg->dtag_first, *prev;
9201 
9202 			ASSERT(rec->dtrd_size != 0 && first != NULL);
9203 			ASSERT(wastuple);
9204 			ASSERT(aggbase != UINT32_MAX);
9205 
9206 			agg->dtag_base = aggbase;
9207 
9208 			while ((prev = first->dta_prev) != NULL &&
9209 			    DTRACEACT_ISAGG(prev->dta_kind)) {
9210 				agg = (dtrace_aggregation_t *)prev;
9211 				first = agg->dtag_first;
9212 			}
9213 
9214 			if (prev != NULL) {
9215 				offs = prev->dta_rec.dtrd_offset +
9216 				    prev->dta_rec.dtrd_size;
9217 			} else {
9218 				offs = sizeof (dtrace_epid_t);
9219 			}
9220 			wastuple = 0;
9221 		} else {
9222 			if (!act->dta_intuple)
9223 				ecb->dte_size = offs + rec->dtrd_size;
9224 
9225 			offs += rec->dtrd_size;
9226 		}
9227 
9228 		wastuple = act->dta_intuple;
9229 	}
9230 
9231 	if ((act = ecb->dte_action) != NULL &&
9232 	    !(act->dta_kind == DTRACEACT_SPECULATE && act->dta_next == NULL) &&
9233 	    ecb->dte_size == sizeof (dtrace_epid_t)) {
9234 		/*
9235 		 * If the size is still sizeof (dtrace_epid_t), then all
9236 		 * actions store no data; set the size to 0.
9237 		 */
9238 		ecb->dte_alignment = maxalign;
9239 		ecb->dte_size = 0;
9240 
9241 		/*
9242 		 * If the needed space is still sizeof (dtrace_epid_t), then
9243 		 * all actions need no additional space; set the needed
9244 		 * size to 0.
9245 		 */
9246 		if (ecb->dte_needed == sizeof (dtrace_epid_t))
9247 			ecb->dte_needed = 0;
9248 
9249 		return;
9250 	}
9251 
9252 	/*
9253 	 * Set our alignment, and make sure that the dte_size and dte_needed
9254 	 * are aligned to the size of an EPID.
9255 	 */
9256 	ecb->dte_alignment = maxalign;
9257 	ecb->dte_size = (ecb->dte_size + (sizeof (dtrace_epid_t) - 1)) &
9258 	    ~(sizeof (dtrace_epid_t) - 1);
9259 	ecb->dte_needed = (ecb->dte_needed + (sizeof (dtrace_epid_t) - 1)) &
9260 	    ~(sizeof (dtrace_epid_t) - 1);
9261 	ASSERT(ecb->dte_size <= ecb->dte_needed);
9262 }
9263 
9264 static dtrace_action_t *
9265 dtrace_ecb_aggregation_create(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
9266 {
9267 	dtrace_aggregation_t *agg;
9268 	size_t size = sizeof (uint64_t);
9269 	int ntuple = desc->dtad_ntuple;
9270 	dtrace_action_t *act;
9271 	dtrace_recdesc_t *frec;
9272 	dtrace_aggid_t aggid;
9273 	dtrace_state_t *state = ecb->dte_state;
9274 
9275 	agg = kmem_zalloc(sizeof (dtrace_aggregation_t), KM_SLEEP);
9276 	agg->dtag_ecb = ecb;
9277 
9278 	ASSERT(DTRACEACT_ISAGG(desc->dtad_kind));
9279 
9280 	switch (desc->dtad_kind) {
9281 	case DTRACEAGG_MIN:
9282 		agg->dtag_initial = INT64_MAX;
9283 		agg->dtag_aggregate = dtrace_aggregate_min;
9284 		break;
9285 
9286 	case DTRACEAGG_MAX:
9287 		agg->dtag_initial = INT64_MIN;
9288 		agg->dtag_aggregate = dtrace_aggregate_max;
9289 		break;
9290 
9291 	case DTRACEAGG_COUNT:
9292 		agg->dtag_aggregate = dtrace_aggregate_count;
9293 		break;
9294 
9295 	case DTRACEAGG_QUANTIZE:
9296 		agg->dtag_aggregate = dtrace_aggregate_quantize;
9297 		size = (((sizeof (uint64_t) * NBBY) - 1) * 2 + 1) *
9298 		    sizeof (uint64_t);
9299 		break;
9300 
9301 	case DTRACEAGG_LQUANTIZE: {
9302 		uint16_t step = DTRACE_LQUANTIZE_STEP(desc->dtad_arg);
9303 		uint16_t levels = DTRACE_LQUANTIZE_LEVELS(desc->dtad_arg);
9304 
9305 		agg->dtag_initial = desc->dtad_arg;
9306 		agg->dtag_aggregate = dtrace_aggregate_lquantize;
9307 
9308 		if (step == 0 || levels == 0)
9309 			goto err;
9310 
9311 		size = levels * sizeof (uint64_t) + 3 * sizeof (uint64_t);
9312 		break;
9313 	}
9314 
9315 	case DTRACEAGG_AVG:
9316 		agg->dtag_aggregate = dtrace_aggregate_avg;
9317 		size = sizeof (uint64_t) * 2;
9318 		break;
9319 
9320 	case DTRACEAGG_STDDEV:
9321 		agg->dtag_aggregate = dtrace_aggregate_stddev;
9322 		size = sizeof (uint64_t) * 4;
9323 		break;
9324 
9325 	case DTRACEAGG_SUM:
9326 		agg->dtag_aggregate = dtrace_aggregate_sum;
9327 		break;
9328 
9329 	default:
9330 		goto err;
9331 	}
9332 
9333 	agg->dtag_action.dta_rec.dtrd_size = size;
9334 
9335 	if (ntuple == 0)
9336 		goto err;
9337 
9338 	/*
9339 	 * We must make sure that we have enough actions for the n-tuple.
9340 	 */
9341 	for (act = ecb->dte_action_last; act != NULL; act = act->dta_prev) {
9342 		if (DTRACEACT_ISAGG(act->dta_kind))
9343 			break;
9344 
9345 		if (--ntuple == 0) {
9346 			/*
9347 			 * This is the action with which our n-tuple begins.
9348 			 */
9349 			agg->dtag_first = act;
9350 			goto success;
9351 		}
9352 	}
9353 
9354 	/*
9355 	 * This n-tuple is short by ntuple elements.  Return failure.
9356 	 */
9357 	ASSERT(ntuple != 0);
9358 err:
9359 	kmem_free(agg, sizeof (dtrace_aggregation_t));
9360 	return (NULL);
9361 
9362 success:
9363 	/*
9364 	 * If the last action in the tuple has a size of zero, it's actually
9365 	 * an expression argument for the aggregating action.
9366 	 */
9367 	ASSERT(ecb->dte_action_last != NULL);
9368 	act = ecb->dte_action_last;
9369 
9370 	if (act->dta_kind == DTRACEACT_DIFEXPR) {
9371 		ASSERT(act->dta_difo != NULL);
9372 
9373 		if (act->dta_difo->dtdo_rtype.dtdt_size == 0)
9374 			agg->dtag_hasarg = 1;
9375 	}
9376 
9377 	/*
9378 	 * We need to allocate an id for this aggregation.
9379 	 */
9380 	aggid = (dtrace_aggid_t)(uintptr_t)vmem_alloc(state->dts_aggid_arena, 1,
9381 	    VM_BESTFIT | VM_SLEEP);
9382 
9383 	if (aggid - 1 >= state->dts_naggregations) {
9384 		dtrace_aggregation_t **oaggs = state->dts_aggregations;
9385 		dtrace_aggregation_t **aggs;
9386 		int naggs = state->dts_naggregations << 1;
9387 		int onaggs = state->dts_naggregations;
9388 
9389 		ASSERT(aggid == state->dts_naggregations + 1);
9390 
9391 		if (naggs == 0) {
9392 			ASSERT(oaggs == NULL);
9393 			naggs = 1;
9394 		}
9395 
9396 		aggs = kmem_zalloc(naggs * sizeof (*aggs), KM_SLEEP);
9397 
9398 		if (oaggs != NULL) {
9399 			bcopy(oaggs, aggs, onaggs * sizeof (*aggs));
9400 			kmem_free(oaggs, onaggs * sizeof (*aggs));
9401 		}
9402 
9403 		state->dts_aggregations = aggs;
9404 		state->dts_naggregations = naggs;
9405 	}
9406 
9407 	ASSERT(state->dts_aggregations[aggid - 1] == NULL);
9408 	state->dts_aggregations[(agg->dtag_id = aggid) - 1] = agg;
9409 
9410 	frec = &agg->dtag_first->dta_rec;
9411 	if (frec->dtrd_alignment < sizeof (dtrace_aggid_t))
9412 		frec->dtrd_alignment = sizeof (dtrace_aggid_t);
9413 
9414 	for (act = agg->dtag_first; act != NULL; act = act->dta_next) {
9415 		ASSERT(!act->dta_intuple);
9416 		act->dta_intuple = 1;
9417 	}
9418 
9419 	return (&agg->dtag_action);
9420 }
9421 
9422 static void
9423 dtrace_ecb_aggregation_destroy(dtrace_ecb_t *ecb, dtrace_action_t *act)
9424 {
9425 	dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
9426 	dtrace_state_t *state = ecb->dte_state;
9427 	dtrace_aggid_t aggid = agg->dtag_id;
9428 
9429 	ASSERT(DTRACEACT_ISAGG(act->dta_kind));
9430 	vmem_free(state->dts_aggid_arena, (void *)(uintptr_t)aggid, 1);
9431 
9432 	ASSERT(state->dts_aggregations[aggid - 1] == agg);
9433 	state->dts_aggregations[aggid - 1] = NULL;
9434 
9435 	kmem_free(agg, sizeof (dtrace_aggregation_t));
9436 }
9437 
9438 static int
9439 dtrace_ecb_action_add(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
9440 {
9441 	dtrace_action_t *action, *last;
9442 	dtrace_difo_t *dp = desc->dtad_difo;
9443 	uint32_t size = 0, align = sizeof (uint8_t), mask;
9444 	uint16_t format = 0;
9445 	dtrace_recdesc_t *rec;
9446 	dtrace_state_t *state = ecb->dte_state;
9447 	dtrace_optval_t *opt = state->dts_options, nframes, strsize;
9448 	uint64_t arg = desc->dtad_arg;
9449 
9450 	ASSERT(MUTEX_HELD(&dtrace_lock));
9451 	ASSERT(ecb->dte_action == NULL || ecb->dte_action->dta_refcnt == 1);
9452 
9453 	if (DTRACEACT_ISAGG(desc->dtad_kind)) {
9454 		/*
9455 		 * If this is an aggregating action, there must be neither
9456 		 * a speculate nor a commit on the action chain.
9457 		 */
9458 		dtrace_action_t *act;
9459 
9460 		for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
9461 			if (act->dta_kind == DTRACEACT_COMMIT)
9462 				return (EINVAL);
9463 
9464 			if (act->dta_kind == DTRACEACT_SPECULATE)
9465 				return (EINVAL);
9466 		}
9467 
9468 		action = dtrace_ecb_aggregation_create(ecb, desc);
9469 
9470 		if (action == NULL)
9471 			return (EINVAL);
9472 	} else {
9473 		if (DTRACEACT_ISDESTRUCTIVE(desc->dtad_kind) ||
9474 		    (desc->dtad_kind == DTRACEACT_DIFEXPR &&
9475 		    dp != NULL && dp->dtdo_destructive)) {
9476 			state->dts_destructive = 1;
9477 		}
9478 
9479 		switch (desc->dtad_kind) {
9480 		case DTRACEACT_PRINTF:
9481 		case DTRACEACT_PRINTA:
9482 		case DTRACEACT_SYSTEM:
9483 		case DTRACEACT_FREOPEN:
9484 			/*
9485 			 * We know that our arg is a string -- turn it into a
9486 			 * format.
9487 			 */
9488 			if (arg == NULL) {
9489 				ASSERT(desc->dtad_kind == DTRACEACT_PRINTA);
9490 				format = 0;
9491 			} else {
9492 				ASSERT(arg != NULL);
9493 				ASSERT(arg > KERNELBASE);
9494 				format = dtrace_format_add(state,
9495 				    (char *)(uintptr_t)arg);
9496 			}
9497 
9498 			/*FALLTHROUGH*/
9499 		case DTRACEACT_LIBACT:
9500 		case DTRACEACT_DIFEXPR:
9501 			if (dp == NULL)
9502 				return (EINVAL);
9503 
9504 			if ((size = dp->dtdo_rtype.dtdt_size) != 0)
9505 				break;
9506 
9507 			if (dp->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING) {
9508 				if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
9509 					return (EINVAL);
9510 
9511 				size = opt[DTRACEOPT_STRSIZE];
9512 			}
9513 
9514 			break;
9515 
9516 		case DTRACEACT_STACK:
9517 			if ((nframes = arg) == 0) {
9518 				nframes = opt[DTRACEOPT_STACKFRAMES];
9519 				ASSERT(nframes > 0);
9520 				arg = nframes;
9521 			}
9522 
9523 			size = nframes * sizeof (pc_t);
9524 			break;
9525 
9526 		case DTRACEACT_JSTACK:
9527 			if ((strsize = DTRACE_USTACK_STRSIZE(arg)) == 0)
9528 				strsize = opt[DTRACEOPT_JSTACKSTRSIZE];
9529 
9530 			if ((nframes = DTRACE_USTACK_NFRAMES(arg)) == 0)
9531 				nframes = opt[DTRACEOPT_JSTACKFRAMES];
9532 
9533 			arg = DTRACE_USTACK_ARG(nframes, strsize);
9534 
9535 			/*FALLTHROUGH*/
9536 		case DTRACEACT_USTACK:
9537 			if (desc->dtad_kind != DTRACEACT_JSTACK &&
9538 			    (nframes = DTRACE_USTACK_NFRAMES(arg)) == 0) {
9539 				strsize = DTRACE_USTACK_STRSIZE(arg);
9540 				nframes = opt[DTRACEOPT_USTACKFRAMES];
9541 				ASSERT(nframes > 0);
9542 				arg = DTRACE_USTACK_ARG(nframes, strsize);
9543 			}
9544 
9545 			/*
9546 			 * Save a slot for the pid.
9547 			 */
9548 			size = (nframes + 1) * sizeof (uint64_t);
9549 			size += DTRACE_USTACK_STRSIZE(arg);
9550 			size = P2ROUNDUP(size, (uint32_t)(sizeof (uintptr_t)));
9551 
9552 			break;
9553 
9554 		case DTRACEACT_SYM:
9555 		case DTRACEACT_MOD:
9556 			if (dp == NULL || ((size = dp->dtdo_rtype.dtdt_size) !=
9557 			    sizeof (uint64_t)) ||
9558 			    (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
9559 				return (EINVAL);
9560 			break;
9561 
9562 		case DTRACEACT_USYM:
9563 		case DTRACEACT_UMOD:
9564 		case DTRACEACT_UADDR:
9565 			if (dp == NULL ||
9566 			    (dp->dtdo_rtype.dtdt_size != sizeof (uint64_t)) ||
9567 			    (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
9568 				return (EINVAL);
9569 
9570 			/*
9571 			 * We have a slot for the pid, plus a slot for the
9572 			 * argument.  To keep things simple (aligned with
9573 			 * bitness-neutral sizing), we store each as a 64-bit
9574 			 * quantity.
9575 			 */
9576 			size = 2 * sizeof (uint64_t);
9577 			break;
9578 
9579 		case DTRACEACT_STOP:
9580 		case DTRACEACT_BREAKPOINT:
9581 		case DTRACEACT_PANIC:
9582 			break;
9583 
9584 		case DTRACEACT_CHILL:
9585 		case DTRACEACT_DISCARD:
9586 		case DTRACEACT_RAISE:
9587 			if (dp == NULL)
9588 				return (EINVAL);
9589 			break;
9590 
9591 		case DTRACEACT_EXIT:
9592 			if (dp == NULL ||
9593 			    (size = dp->dtdo_rtype.dtdt_size) != sizeof (int) ||
9594 			    (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
9595 				return (EINVAL);
9596 			break;
9597 
9598 		case DTRACEACT_SPECULATE:
9599 			if (ecb->dte_size > sizeof (dtrace_epid_t))
9600 				return (EINVAL);
9601 
9602 			if (dp == NULL)
9603 				return (EINVAL);
9604 
9605 			state->dts_speculates = 1;
9606 			break;
9607 
9608 		case DTRACEACT_COMMIT: {
9609 			dtrace_action_t *act = ecb->dte_action;
9610 
9611 			for (; act != NULL; act = act->dta_next) {
9612 				if (act->dta_kind == DTRACEACT_COMMIT)
9613 					return (EINVAL);
9614 			}
9615 
9616 			if (dp == NULL)
9617 				return (EINVAL);
9618 			break;
9619 		}
9620 
9621 		default:
9622 			return (EINVAL);
9623 		}
9624 
9625 		if (size != 0 || desc->dtad_kind == DTRACEACT_SPECULATE) {
9626 			/*
9627 			 * If this is a data-storing action or a speculate,
9628 			 * we must be sure that there isn't a commit on the
9629 			 * action chain.
9630 			 */
9631 			dtrace_action_t *act = ecb->dte_action;
9632 
9633 			for (; act != NULL; act = act->dta_next) {
9634 				if (act->dta_kind == DTRACEACT_COMMIT)
9635 					return (EINVAL);
9636 			}
9637 		}
9638 
9639 		action = kmem_zalloc(sizeof (dtrace_action_t), KM_SLEEP);
9640 		action->dta_rec.dtrd_size = size;
9641 	}
9642 
9643 	action->dta_refcnt = 1;
9644 	rec = &action->dta_rec;
9645 	size = rec->dtrd_size;
9646 
9647 	for (mask = sizeof (uint64_t) - 1; size != 0 && mask > 0; mask >>= 1) {
9648 		if (!(size & mask)) {
9649 			align = mask + 1;
9650 			break;
9651 		}
9652 	}
9653 
9654 	action->dta_kind = desc->dtad_kind;
9655 
9656 	if ((action->dta_difo = dp) != NULL)
9657 		dtrace_difo_hold(dp);
9658 
9659 	rec->dtrd_action = action->dta_kind;
9660 	rec->dtrd_arg = arg;
9661 	rec->dtrd_uarg = desc->dtad_uarg;
9662 	rec->dtrd_alignment = (uint16_t)align;
9663 	rec->dtrd_format = format;
9664 
9665 	if ((last = ecb->dte_action_last) != NULL) {
9666 		ASSERT(ecb->dte_action != NULL);
9667 		action->dta_prev = last;
9668 		last->dta_next = action;
9669 	} else {
9670 		ASSERT(ecb->dte_action == NULL);
9671 		ecb->dte_action = action;
9672 	}
9673 
9674 	ecb->dte_action_last = action;
9675 
9676 	return (0);
9677 }
9678 
9679 static void
9680 dtrace_ecb_action_remove(dtrace_ecb_t *ecb)
9681 {
9682 	dtrace_action_t *act = ecb->dte_action, *next;
9683 	dtrace_vstate_t *vstate = &ecb->dte_state->dts_vstate;
9684 	dtrace_difo_t *dp;
9685 	uint16_t format;
9686 
9687 	if (act != NULL && act->dta_refcnt > 1) {
9688 		ASSERT(act->dta_next == NULL || act->dta_next->dta_refcnt == 1);
9689 		act->dta_refcnt--;
9690 	} else {
9691 		for (; act != NULL; act = next) {
9692 			next = act->dta_next;
9693 			ASSERT(next != NULL || act == ecb->dte_action_last);
9694 			ASSERT(act->dta_refcnt == 1);
9695 
9696 			if ((format = act->dta_rec.dtrd_format) != 0)
9697 				dtrace_format_remove(ecb->dte_state, format);
9698 
9699 			if ((dp = act->dta_difo) != NULL)
9700 				dtrace_difo_release(dp, vstate);
9701 
9702 			if (DTRACEACT_ISAGG(act->dta_kind)) {
9703 				dtrace_ecb_aggregation_destroy(ecb, act);
9704 			} else {
9705 				kmem_free(act, sizeof (dtrace_action_t));
9706 			}
9707 		}
9708 	}
9709 
9710 	ecb->dte_action = NULL;
9711 	ecb->dte_action_last = NULL;
9712 	ecb->dte_size = sizeof (dtrace_epid_t);
9713 }
9714 
9715 static void
9716 dtrace_ecb_disable(dtrace_ecb_t *ecb)
9717 {
9718 	/*
9719 	 * We disable the ECB by removing it from its probe.
9720 	 */
9721 	dtrace_ecb_t *pecb, *prev = NULL;
9722 	dtrace_probe_t *probe = ecb->dte_probe;
9723 
9724 	ASSERT(MUTEX_HELD(&dtrace_lock));
9725 
9726 	if (probe == NULL) {
9727 		/*
9728 		 * This is the NULL probe; there is nothing to disable.
9729 		 */
9730 		return;
9731 	}
9732 
9733 	for (pecb = probe->dtpr_ecb; pecb != NULL; pecb = pecb->dte_next) {
9734 		if (pecb == ecb)
9735 			break;
9736 		prev = pecb;
9737 	}
9738 
9739 	ASSERT(pecb != NULL);
9740 
9741 	if (prev == NULL) {
9742 		probe->dtpr_ecb = ecb->dte_next;
9743 	} else {
9744 		prev->dte_next = ecb->dte_next;
9745 	}
9746 
9747 	if (ecb == probe->dtpr_ecb_last) {
9748 		ASSERT(ecb->dte_next == NULL);
9749 		probe->dtpr_ecb_last = prev;
9750 	}
9751 
9752 	/*
9753 	 * The ECB has been disconnected from the probe; now sync to assure
9754 	 * that all CPUs have seen the change before returning.
9755 	 */
9756 	dtrace_sync();
9757 
9758 	if (probe->dtpr_ecb == NULL) {
9759 		/*
9760 		 * That was the last ECB on the probe; clear the predicate
9761 		 * cache ID for the probe, disable it and sync one more time
9762 		 * to assure that we'll never hit it again.
9763 		 */
9764 		dtrace_provider_t *prov = probe->dtpr_provider;
9765 
9766 		ASSERT(ecb->dte_next == NULL);
9767 		ASSERT(probe->dtpr_ecb_last == NULL);
9768 		probe->dtpr_predcache = DTRACE_CACHEIDNONE;
9769 		prov->dtpv_pops.dtps_disable(prov->dtpv_arg,
9770 		    probe->dtpr_id, probe->dtpr_arg);
9771 		dtrace_sync();
9772 	} else {
9773 		/*
9774 		 * There is at least one ECB remaining on the probe.  If there
9775 		 * is _exactly_ one, set the probe's predicate cache ID to be
9776 		 * the predicate cache ID of the remaining ECB.
9777 		 */
9778 		ASSERT(probe->dtpr_ecb_last != NULL);
9779 		ASSERT(probe->dtpr_predcache == DTRACE_CACHEIDNONE);
9780 
9781 		if (probe->dtpr_ecb == probe->dtpr_ecb_last) {
9782 			dtrace_predicate_t *p = probe->dtpr_ecb->dte_predicate;
9783 
9784 			ASSERT(probe->dtpr_ecb->dte_next == NULL);
9785 
9786 			if (p != NULL)
9787 				probe->dtpr_predcache = p->dtp_cacheid;
9788 		}
9789 
9790 		ecb->dte_next = NULL;
9791 	}
9792 }
9793 
9794 static void
9795 dtrace_ecb_destroy(dtrace_ecb_t *ecb)
9796 {
9797 	dtrace_state_t *state = ecb->dte_state;
9798 	dtrace_vstate_t *vstate = &state->dts_vstate;
9799 	dtrace_predicate_t *pred;
9800 	dtrace_epid_t epid = ecb->dte_epid;
9801 
9802 	ASSERT(MUTEX_HELD(&dtrace_lock));
9803 	ASSERT(ecb->dte_next == NULL);
9804 	ASSERT(ecb->dte_probe == NULL || ecb->dte_probe->dtpr_ecb != ecb);
9805 
9806 	if ((pred = ecb->dte_predicate) != NULL)
9807 		dtrace_predicate_release(pred, vstate);
9808 
9809 	dtrace_ecb_action_remove(ecb);
9810 
9811 	ASSERT(state->dts_ecbs[epid - 1] == ecb);
9812 	state->dts_ecbs[epid - 1] = NULL;
9813 
9814 	kmem_free(ecb, sizeof (dtrace_ecb_t));
9815 }
9816 
9817 static dtrace_ecb_t *
9818 dtrace_ecb_create(dtrace_state_t *state, dtrace_probe_t *probe,
9819     dtrace_enabling_t *enab)
9820 {
9821 	dtrace_ecb_t *ecb;
9822 	dtrace_predicate_t *pred;
9823 	dtrace_actdesc_t *act;
9824 	dtrace_provider_t *prov;
9825 	dtrace_ecbdesc_t *desc = enab->dten_current;
9826 
9827 	ASSERT(MUTEX_HELD(&dtrace_lock));
9828 	ASSERT(state != NULL);
9829 
9830 	ecb = dtrace_ecb_add(state, probe);
9831 	ecb->dte_uarg = desc->dted_uarg;
9832 
9833 	if ((pred = desc->dted_pred.dtpdd_predicate) != NULL) {
9834 		dtrace_predicate_hold(pred);
9835 		ecb->dte_predicate = pred;
9836 	}
9837 
9838 	if (probe != NULL) {
9839 		/*
9840 		 * If the provider shows more leg than the consumer is old
9841 		 * enough to see, we need to enable the appropriate implicit
9842 		 * predicate bits to prevent the ecb from activating at
9843 		 * revealing times.
9844 		 *
9845 		 * Providers specifying DTRACE_PRIV_USER at register time
9846 		 * are stating that they need the /proc-style privilege
9847 		 * model to be enforced, and this is what DTRACE_COND_OWNER
9848 		 * and DTRACE_COND_ZONEOWNER will then do at probe time.
9849 		 */
9850 		prov = probe->dtpr_provider;
9851 		if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLPROC) &&
9852 		    (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
9853 			ecb->dte_cond |= DTRACE_COND_OWNER;
9854 
9855 		if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLZONE) &&
9856 		    (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
9857 			ecb->dte_cond |= DTRACE_COND_ZONEOWNER;
9858 
9859 		/*
9860 		 * If the provider shows us kernel innards and the user
9861 		 * is lacking sufficient privilege, enable the
9862 		 * DTRACE_COND_USERMODE implicit predicate.
9863 		 */
9864 		if (!(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) &&
9865 		    (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_KERNEL))
9866 			ecb->dte_cond |= DTRACE_COND_USERMODE;
9867 	}
9868 
9869 	if (dtrace_ecb_create_cache != NULL) {
9870 		/*
9871 		 * If we have a cached ecb, we'll use its action list instead
9872 		 * of creating our own (saving both time and space).
9873 		 */
9874 		dtrace_ecb_t *cached = dtrace_ecb_create_cache;
9875 		dtrace_action_t *act = cached->dte_action;
9876 
9877 		if (act != NULL) {
9878 			ASSERT(act->dta_refcnt > 0);
9879 			act->dta_refcnt++;
9880 			ecb->dte_action = act;
9881 			ecb->dte_action_last = cached->dte_action_last;
9882 			ecb->dte_needed = cached->dte_needed;
9883 			ecb->dte_size = cached->dte_size;
9884 			ecb->dte_alignment = cached->dte_alignment;
9885 		}
9886 
9887 		return (ecb);
9888 	}
9889 
9890 	for (act = desc->dted_action; act != NULL; act = act->dtad_next) {
9891 		if ((enab->dten_error = dtrace_ecb_action_add(ecb, act)) != 0) {
9892 			dtrace_ecb_destroy(ecb);
9893 			return (NULL);
9894 		}
9895 	}
9896 
9897 	dtrace_ecb_resize(ecb);
9898 
9899 	return (dtrace_ecb_create_cache = ecb);
9900 }
9901 
9902 static int
9903 dtrace_ecb_create_enable(dtrace_probe_t *probe, void *arg)
9904 {
9905 	dtrace_ecb_t *ecb;
9906 	dtrace_enabling_t *enab = arg;
9907 	dtrace_state_t *state = enab->dten_vstate->dtvs_state;
9908 
9909 	ASSERT(state != NULL);
9910 
9911 	if (probe != NULL && probe->dtpr_gen < enab->dten_probegen) {
9912 		/*
9913 		 * This probe was created in a generation for which this
9914 		 * enabling has previously created ECBs; we don't want to
9915 		 * enable it again, so just kick out.
9916 		 */
9917 		return (DTRACE_MATCH_NEXT);
9918 	}
9919 
9920 	if ((ecb = dtrace_ecb_create(state, probe, enab)) == NULL)
9921 		return (DTRACE_MATCH_DONE);
9922 
9923 	dtrace_ecb_enable(ecb);
9924 	return (DTRACE_MATCH_NEXT);
9925 }
9926 
9927 static dtrace_ecb_t *
9928 dtrace_epid2ecb(dtrace_state_t *state, dtrace_epid_t id)
9929 {
9930 	dtrace_ecb_t *ecb;
9931 
9932 	ASSERT(MUTEX_HELD(&dtrace_lock));
9933 
9934 	if (id == 0 || id > state->dts_necbs)
9935 		return (NULL);
9936 
9937 	ASSERT(state->dts_necbs > 0 && state->dts_ecbs != NULL);
9938 	ASSERT((ecb = state->dts_ecbs[id - 1]) == NULL || ecb->dte_epid == id);
9939 
9940 	return (state->dts_ecbs[id - 1]);
9941 }
9942 
9943 static dtrace_aggregation_t *
9944 dtrace_aggid2agg(dtrace_state_t *state, dtrace_aggid_t id)
9945 {
9946 	dtrace_aggregation_t *agg;
9947 
9948 	ASSERT(MUTEX_HELD(&dtrace_lock));
9949 
9950 	if (id == 0 || id > state->dts_naggregations)
9951 		return (NULL);
9952 
9953 	ASSERT(state->dts_naggregations > 0 && state->dts_aggregations != NULL);
9954 	ASSERT((agg = state->dts_aggregations[id - 1]) == NULL ||
9955 	    agg->dtag_id == id);
9956 
9957 	return (state->dts_aggregations[id - 1]);
9958 }
9959 
9960 /*
9961  * DTrace Buffer Functions
9962  *
9963  * The following functions manipulate DTrace buffers.  Most of these functions
9964  * are called in the context of establishing or processing consumer state;
9965  * exceptions are explicitly noted.
9966  */
9967 
9968 /*
9969  * Note:  called from cross call context.  This function switches the two
9970  * buffers on a given CPU.  The atomicity of this operation is assured by
9971  * disabling interrupts while the actual switch takes place; the disabling of
9972  * interrupts serializes the execution with any execution of dtrace_probe() on
9973  * the same CPU.
9974  */
9975 static void
9976 dtrace_buffer_switch(dtrace_buffer_t *buf)
9977 {
9978 	caddr_t tomax = buf->dtb_tomax;
9979 	caddr_t xamot = buf->dtb_xamot;
9980 	dtrace_icookie_t cookie;
9981 
9982 	ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
9983 	ASSERT(!(buf->dtb_flags & DTRACEBUF_RING));
9984 
9985 	cookie = dtrace_interrupt_disable();
9986 	buf->dtb_tomax = xamot;
9987 	buf->dtb_xamot = tomax;
9988 	buf->dtb_xamot_drops = buf->dtb_drops;
9989 	buf->dtb_xamot_offset = buf->dtb_offset;
9990 	buf->dtb_xamot_errors = buf->dtb_errors;
9991 	buf->dtb_xamot_flags = buf->dtb_flags;
9992 	buf->dtb_offset = 0;
9993 	buf->dtb_drops = 0;
9994 	buf->dtb_errors = 0;
9995 	buf->dtb_flags &= ~(DTRACEBUF_ERROR | DTRACEBUF_DROPPED);
9996 	dtrace_interrupt_enable(cookie);
9997 }
9998 
9999 /*
10000  * Note:  called from cross call context.  This function activates a buffer
10001  * on a CPU.  As with dtrace_buffer_switch(), the atomicity of the operation
10002  * is guaranteed by the disabling of interrupts.
10003  */
10004 static void
10005 dtrace_buffer_activate(dtrace_state_t *state)
10006 {
10007 	dtrace_buffer_t *buf;
10008 	dtrace_icookie_t cookie = dtrace_interrupt_disable();
10009 
10010 	buf = &state->dts_buffer[CPU->cpu_id];
10011 
10012 	if (buf->dtb_tomax != NULL) {
10013 		/*
10014 		 * We might like to assert that the buffer is marked inactive,
10015 		 * but this isn't necessarily true:  the buffer for the CPU
10016 		 * that processes the BEGIN probe has its buffer activated
10017 		 * manually.  In this case, we take the (harmless) action
10018 		 * re-clearing the bit INACTIVE bit.
10019 		 */
10020 		buf->dtb_flags &= ~DTRACEBUF_INACTIVE;
10021 	}
10022 
10023 	dtrace_interrupt_enable(cookie);
10024 }
10025 
10026 static int
10027 dtrace_buffer_alloc(dtrace_buffer_t *bufs, size_t size, int flags,
10028     processorid_t cpu)
10029 {
10030 	cpu_t *cp;
10031 	dtrace_buffer_t *buf;
10032 
10033 	ASSERT(MUTEX_HELD(&cpu_lock));
10034 	ASSERT(MUTEX_HELD(&dtrace_lock));
10035 
10036 	if (size > dtrace_nonroot_maxsize &&
10037 	    !PRIV_POLICY_CHOICE(CRED(), PRIV_ALL, B_FALSE))
10038 		return (EFBIG);
10039 
10040 	cp = cpu_list;
10041 
10042 	do {
10043 		if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
10044 			continue;
10045 
10046 		buf = &bufs[cp->cpu_id];
10047 
10048 		/*
10049 		 * If there is already a buffer allocated for this CPU, it
10050 		 * is only possible that this is a DR event.  In this case,
10051 		 * the buffer size must match our specified size.
10052 		 */
10053 		if (buf->dtb_tomax != NULL) {
10054 			ASSERT(buf->dtb_size == size);
10055 			continue;
10056 		}
10057 
10058 		ASSERT(buf->dtb_xamot == NULL);
10059 
10060 		if ((buf->dtb_tomax = kmem_zalloc(size, KM_NOSLEEP)) == NULL)
10061 			goto err;
10062 
10063 		buf->dtb_size = size;
10064 		buf->dtb_flags = flags;
10065 		buf->dtb_offset = 0;
10066 		buf->dtb_drops = 0;
10067 
10068 		if (flags & DTRACEBUF_NOSWITCH)
10069 			continue;
10070 
10071 		if ((buf->dtb_xamot = kmem_zalloc(size, KM_NOSLEEP)) == NULL)
10072 			goto err;
10073 	} while ((cp = cp->cpu_next) != cpu_list);
10074 
10075 	return (0);
10076 
10077 err:
10078 	cp = cpu_list;
10079 
10080 	do {
10081 		if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
10082 			continue;
10083 
10084 		buf = &bufs[cp->cpu_id];
10085 
10086 		if (buf->dtb_xamot != NULL) {
10087 			ASSERT(buf->dtb_tomax != NULL);
10088 			ASSERT(buf->dtb_size == size);
10089 			kmem_free(buf->dtb_xamot, size);
10090 		}
10091 
10092 		if (buf->dtb_tomax != NULL) {
10093 			ASSERT(buf->dtb_size == size);
10094 			kmem_free(buf->dtb_tomax, size);
10095 		}
10096 
10097 		buf->dtb_tomax = NULL;
10098 		buf->dtb_xamot = NULL;
10099 		buf->dtb_size = 0;
10100 	} while ((cp = cp->cpu_next) != cpu_list);
10101 
10102 	return (ENOMEM);
10103 }
10104 
10105 /*
10106  * Note:  called from probe context.  This function just increments the drop
10107  * count on a buffer.  It has been made a function to allow for the
10108  * possibility of understanding the source of mysterious drop counts.  (A
10109  * problem for which one may be particularly disappointed that DTrace cannot
10110  * be used to understand DTrace.)
10111  */
10112 static void
10113 dtrace_buffer_drop(dtrace_buffer_t *buf)
10114 {
10115 	buf->dtb_drops++;
10116 }
10117 
10118 /*
10119  * Note:  called from probe context.  This function is called to reserve space
10120  * in a buffer.  If mstate is non-NULL, sets the scratch base and size in the
10121  * mstate.  Returns the new offset in the buffer, or a negative value if an
10122  * error has occurred.
10123  */
10124 static intptr_t
10125 dtrace_buffer_reserve(dtrace_buffer_t *buf, size_t needed, size_t align,
10126     dtrace_state_t *state, dtrace_mstate_t *mstate)
10127 {
10128 	intptr_t offs = buf->dtb_offset, soffs;
10129 	intptr_t woffs;
10130 	caddr_t tomax;
10131 	size_t total;
10132 
10133 	if (buf->dtb_flags & DTRACEBUF_INACTIVE)
10134 		return (-1);
10135 
10136 	if ((tomax = buf->dtb_tomax) == NULL) {
10137 		dtrace_buffer_drop(buf);
10138 		return (-1);
10139 	}
10140 
10141 	if (!(buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL))) {
10142 		while (offs & (align - 1)) {
10143 			/*
10144 			 * Assert that our alignment is off by a number which
10145 			 * is itself sizeof (uint32_t) aligned.
10146 			 */
10147 			ASSERT(!((align - (offs & (align - 1))) &
10148 			    (sizeof (uint32_t) - 1)));
10149 			DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
10150 			offs += sizeof (uint32_t);
10151 		}
10152 
10153 		if ((soffs = offs + needed) > buf->dtb_size) {
10154 			dtrace_buffer_drop(buf);
10155 			return (-1);
10156 		}
10157 
10158 		if (mstate == NULL)
10159 			return (offs);
10160 
10161 		mstate->dtms_scratch_base = (uintptr_t)tomax + soffs;
10162 		mstate->dtms_scratch_size = buf->dtb_size - soffs;
10163 		mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
10164 
10165 		return (offs);
10166 	}
10167 
10168 	if (buf->dtb_flags & DTRACEBUF_FILL) {
10169 		if (state->dts_activity != DTRACE_ACTIVITY_COOLDOWN &&
10170 		    (buf->dtb_flags & DTRACEBUF_FULL))
10171 			return (-1);
10172 		goto out;
10173 	}
10174 
10175 	total = needed + (offs & (align - 1));
10176 
10177 	/*
10178 	 * For a ring buffer, life is quite a bit more complicated.  Before
10179 	 * we can store any padding, we need to adjust our wrapping offset.
10180 	 * (If we've never before wrapped or we're not about to, no adjustment
10181 	 * is required.)
10182 	 */
10183 	if ((buf->dtb_flags & DTRACEBUF_WRAPPED) ||
10184 	    offs + total > buf->dtb_size) {
10185 		woffs = buf->dtb_xamot_offset;
10186 
10187 		if (offs + total > buf->dtb_size) {
10188 			/*
10189 			 * We can't fit in the end of the buffer.  First, a
10190 			 * sanity check that we can fit in the buffer at all.
10191 			 */
10192 			if (total > buf->dtb_size) {
10193 				dtrace_buffer_drop(buf);
10194 				return (-1);
10195 			}
10196 
10197 			/*
10198 			 * We're going to be storing at the top of the buffer,
10199 			 * so now we need to deal with the wrapped offset.  We
10200 			 * only reset our wrapped offset to 0 if it is
10201 			 * currently greater than the current offset.  If it
10202 			 * is less than the current offset, it is because a
10203 			 * previous allocation induced a wrap -- but the
10204 			 * allocation didn't subsequently take the space due
10205 			 * to an error or false predicate evaluation.  In this
10206 			 * case, we'll just leave the wrapped offset alone: if
10207 			 * the wrapped offset hasn't been advanced far enough
10208 			 * for this allocation, it will be adjusted in the
10209 			 * lower loop.
10210 			 */
10211 			if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
10212 				if (woffs >= offs)
10213 					woffs = 0;
10214 			} else {
10215 				woffs = 0;
10216 			}
10217 
10218 			/*
10219 			 * Now we know that we're going to be storing to the
10220 			 * top of the buffer and that there is room for us
10221 			 * there.  We need to clear the buffer from the current
10222 			 * offset to the end (there may be old gunk there).
10223 			 */
10224 			while (offs < buf->dtb_size)
10225 				tomax[offs++] = 0;
10226 
10227 			/*
10228 			 * We need to set our offset to zero.  And because we
10229 			 * are wrapping, we need to set the bit indicating as
10230 			 * much.  We can also adjust our needed space back
10231 			 * down to the space required by the ECB -- we know
10232 			 * that the top of the buffer is aligned.
10233 			 */
10234 			offs = 0;
10235 			total = needed;
10236 			buf->dtb_flags |= DTRACEBUF_WRAPPED;
10237 		} else {
10238 			/*
10239 			 * There is room for us in the buffer, so we simply
10240 			 * need to check the wrapped offset.
10241 			 */
10242 			if (woffs < offs) {
10243 				/*
10244 				 * The wrapped offset is less than the offset.
10245 				 * This can happen if we allocated buffer space
10246 				 * that induced a wrap, but then we didn't
10247 				 * subsequently take the space due to an error
10248 				 * or false predicate evaluation.  This is
10249 				 * okay; we know that _this_ allocation isn't
10250 				 * going to induce a wrap.  We still can't
10251 				 * reset the wrapped offset to be zero,
10252 				 * however: the space may have been trashed in
10253 				 * the previous failed probe attempt.  But at
10254 				 * least the wrapped offset doesn't need to
10255 				 * be adjusted at all...
10256 				 */
10257 				goto out;
10258 			}
10259 		}
10260 
10261 		while (offs + total > woffs) {
10262 			dtrace_epid_t epid = *(uint32_t *)(tomax + woffs);
10263 			size_t size;
10264 
10265 			if (epid == DTRACE_EPIDNONE) {
10266 				size = sizeof (uint32_t);
10267 			} else {
10268 				ASSERT(epid <= state->dts_necbs);
10269 				ASSERT(state->dts_ecbs[epid - 1] != NULL);
10270 
10271 				size = state->dts_ecbs[epid - 1]->dte_size;
10272 			}
10273 
10274 			ASSERT(woffs + size <= buf->dtb_size);
10275 			ASSERT(size != 0);
10276 
10277 			if (woffs + size == buf->dtb_size) {
10278 				/*
10279 				 * We've reached the end of the buffer; we want
10280 				 * to set the wrapped offset to 0 and break
10281 				 * out.  However, if the offs is 0, then we're
10282 				 * in a strange edge-condition:  the amount of
10283 				 * space that we want to reserve plus the size
10284 				 * of the record that we're overwriting is
10285 				 * greater than the size of the buffer.  This
10286 				 * is problematic because if we reserve the
10287 				 * space but subsequently don't consume it (due
10288 				 * to a failed predicate or error) the wrapped
10289 				 * offset will be 0 -- yet the EPID at offset 0
10290 				 * will not be committed.  This situation is
10291 				 * relatively easy to deal with:  if we're in
10292 				 * this case, the buffer is indistinguishable
10293 				 * from one that hasn't wrapped; we need only
10294 				 * finish the job by clearing the wrapped bit,
10295 				 * explicitly setting the offset to be 0, and
10296 				 * zero'ing out the old data in the buffer.
10297 				 */
10298 				if (offs == 0) {
10299 					buf->dtb_flags &= ~DTRACEBUF_WRAPPED;
10300 					buf->dtb_offset = 0;
10301 					woffs = total;
10302 
10303 					while (woffs < buf->dtb_size)
10304 						tomax[woffs++] = 0;
10305 				}
10306 
10307 				woffs = 0;
10308 				break;
10309 			}
10310 
10311 			woffs += size;
10312 		}
10313 
10314 		/*
10315 		 * We have a wrapped offset.  It may be that the wrapped offset
10316 		 * has become zero -- that's okay.
10317 		 */
10318 		buf->dtb_xamot_offset = woffs;
10319 	}
10320 
10321 out:
10322 	/*
10323 	 * Now we can plow the buffer with any necessary padding.
10324 	 */
10325 	while (offs & (align - 1)) {
10326 		/*
10327 		 * Assert that our alignment is off by a number which
10328 		 * is itself sizeof (uint32_t) aligned.
10329 		 */
10330 		ASSERT(!((align - (offs & (align - 1))) &
10331 		    (sizeof (uint32_t) - 1)));
10332 		DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
10333 		offs += sizeof (uint32_t);
10334 	}
10335 
10336 	if (buf->dtb_flags & DTRACEBUF_FILL) {
10337 		if (offs + needed > buf->dtb_size - state->dts_reserve) {
10338 			buf->dtb_flags |= DTRACEBUF_FULL;
10339 			return (-1);
10340 		}
10341 	}
10342 
10343 	if (mstate == NULL)
10344 		return (offs);
10345 
10346 	/*
10347 	 * For ring buffers and fill buffers, the scratch space is always
10348 	 * the inactive buffer.
10349 	 */
10350 	mstate->dtms_scratch_base = (uintptr_t)buf->dtb_xamot;
10351 	mstate->dtms_scratch_size = buf->dtb_size;
10352 	mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
10353 
10354 	return (offs);
10355 }
10356 
10357 static void
10358 dtrace_buffer_polish(dtrace_buffer_t *buf)
10359 {
10360 	ASSERT(buf->dtb_flags & DTRACEBUF_RING);
10361 	ASSERT(MUTEX_HELD(&dtrace_lock));
10362 
10363 	if (!(buf->dtb_flags & DTRACEBUF_WRAPPED))
10364 		return;
10365 
10366 	/*
10367 	 * We need to polish the ring buffer.  There are three cases:
10368 	 *
10369 	 * - The first (and presumably most common) is that there is no gap
10370 	 *   between the buffer offset and the wrapped offset.  In this case,
10371 	 *   there is nothing in the buffer that isn't valid data; we can
10372 	 *   mark the buffer as polished and return.
10373 	 *
10374 	 * - The second (less common than the first but still more common
10375 	 *   than the third) is that there is a gap between the buffer offset
10376 	 *   and the wrapped offset, and the wrapped offset is larger than the
10377 	 *   buffer offset.  This can happen because of an alignment issue, or
10378 	 *   can happen because of a call to dtrace_buffer_reserve() that
10379 	 *   didn't subsequently consume the buffer space.  In this case,
10380 	 *   we need to zero the data from the buffer offset to the wrapped
10381 	 *   offset.
10382 	 *
10383 	 * - The third (and least common) is that there is a gap between the
10384 	 *   buffer offset and the wrapped offset, but the wrapped offset is
10385 	 *   _less_ than the buffer offset.  This can only happen because a
10386 	 *   call to dtrace_buffer_reserve() induced a wrap, but the space
10387 	 *   was not subsequently consumed.  In this case, we need to zero the
10388 	 *   space from the offset to the end of the buffer _and_ from the
10389 	 *   top of the buffer to the wrapped offset.
10390 	 */
10391 	if (buf->dtb_offset < buf->dtb_xamot_offset) {
10392 		bzero(buf->dtb_tomax + buf->dtb_offset,
10393 		    buf->dtb_xamot_offset - buf->dtb_offset);
10394 	}
10395 
10396 	if (buf->dtb_offset > buf->dtb_xamot_offset) {
10397 		bzero(buf->dtb_tomax + buf->dtb_offset,
10398 		    buf->dtb_size - buf->dtb_offset);
10399 		bzero(buf->dtb_tomax, buf->dtb_xamot_offset);
10400 	}
10401 }
10402 
10403 static void
10404 dtrace_buffer_free(dtrace_buffer_t *bufs)
10405 {
10406 	int i;
10407 
10408 	for (i = 0; i < NCPU; i++) {
10409 		dtrace_buffer_t *buf = &bufs[i];
10410 
10411 		if (buf->dtb_tomax == NULL) {
10412 			ASSERT(buf->dtb_xamot == NULL);
10413 			ASSERT(buf->dtb_size == 0);
10414 			continue;
10415 		}
10416 
10417 		if (buf->dtb_xamot != NULL) {
10418 			ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
10419 			kmem_free(buf->dtb_xamot, buf->dtb_size);
10420 		}
10421 
10422 		kmem_free(buf->dtb_tomax, buf->dtb_size);
10423 		buf->dtb_size = 0;
10424 		buf->dtb_tomax = NULL;
10425 		buf->dtb_xamot = NULL;
10426 	}
10427 }
10428 
10429 /*
10430  * DTrace Enabling Functions
10431  */
10432 static dtrace_enabling_t *
10433 dtrace_enabling_create(dtrace_vstate_t *vstate)
10434 {
10435 	dtrace_enabling_t *enab;
10436 
10437 	enab = kmem_zalloc(sizeof (dtrace_enabling_t), KM_SLEEP);
10438 	enab->dten_vstate = vstate;
10439 
10440 	return (enab);
10441 }
10442 
10443 static void
10444 dtrace_enabling_add(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb)
10445 {
10446 	dtrace_ecbdesc_t **ndesc;
10447 	size_t osize, nsize;
10448 
10449 	/*
10450 	 * We can't add to enablings after we've enabled them, or after we've
10451 	 * retained them.
10452 	 */
10453 	ASSERT(enab->dten_probegen == 0);
10454 	ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
10455 
10456 	if (enab->dten_ndesc < enab->dten_maxdesc) {
10457 		enab->dten_desc[enab->dten_ndesc++] = ecb;
10458 		return;
10459 	}
10460 
10461 	osize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
10462 
10463 	if (enab->dten_maxdesc == 0) {
10464 		enab->dten_maxdesc = 1;
10465 	} else {
10466 		enab->dten_maxdesc <<= 1;
10467 	}
10468 
10469 	ASSERT(enab->dten_ndesc < enab->dten_maxdesc);
10470 
10471 	nsize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
10472 	ndesc = kmem_zalloc(nsize, KM_SLEEP);
10473 	bcopy(enab->dten_desc, ndesc, osize);
10474 	kmem_free(enab->dten_desc, osize);
10475 
10476 	enab->dten_desc = ndesc;
10477 	enab->dten_desc[enab->dten_ndesc++] = ecb;
10478 }
10479 
10480 static void
10481 dtrace_enabling_addlike(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb,
10482     dtrace_probedesc_t *pd)
10483 {
10484 	dtrace_ecbdesc_t *new;
10485 	dtrace_predicate_t *pred;
10486 	dtrace_actdesc_t *act;
10487 
10488 	/*
10489 	 * We're going to create a new ECB description that matches the
10490 	 * specified ECB in every way, but has the specified probe description.
10491 	 */
10492 	new = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
10493 
10494 	if ((pred = ecb->dted_pred.dtpdd_predicate) != NULL)
10495 		dtrace_predicate_hold(pred);
10496 
10497 	for (act = ecb->dted_action; act != NULL; act = act->dtad_next)
10498 		dtrace_actdesc_hold(act);
10499 
10500 	new->dted_action = ecb->dted_action;
10501 	new->dted_pred = ecb->dted_pred;
10502 	new->dted_probe = *pd;
10503 	new->dted_uarg = ecb->dted_uarg;
10504 
10505 	dtrace_enabling_add(enab, new);
10506 }
10507 
10508 static void
10509 dtrace_enabling_dump(dtrace_enabling_t *enab)
10510 {
10511 	int i;
10512 
10513 	for (i = 0; i < enab->dten_ndesc; i++) {
10514 		dtrace_probedesc_t *desc = &enab->dten_desc[i]->dted_probe;
10515 
10516 		cmn_err(CE_NOTE, "enabling probe %d (%s:%s:%s:%s)", i,
10517 		    desc->dtpd_provider, desc->dtpd_mod,
10518 		    desc->dtpd_func, desc->dtpd_name);
10519 	}
10520 }
10521 
10522 static void
10523 dtrace_enabling_destroy(dtrace_enabling_t *enab)
10524 {
10525 	int i;
10526 	dtrace_ecbdesc_t *ep;
10527 	dtrace_vstate_t *vstate = enab->dten_vstate;
10528 
10529 	ASSERT(MUTEX_HELD(&dtrace_lock));
10530 
10531 	for (i = 0; i < enab->dten_ndesc; i++) {
10532 		dtrace_actdesc_t *act, *next;
10533 		dtrace_predicate_t *pred;
10534 
10535 		ep = enab->dten_desc[i];
10536 
10537 		if ((pred = ep->dted_pred.dtpdd_predicate) != NULL)
10538 			dtrace_predicate_release(pred, vstate);
10539 
10540 		for (act = ep->dted_action; act != NULL; act = next) {
10541 			next = act->dtad_next;
10542 			dtrace_actdesc_release(act, vstate);
10543 		}
10544 
10545 		kmem_free(ep, sizeof (dtrace_ecbdesc_t));
10546 	}
10547 
10548 	kmem_free(enab->dten_desc,
10549 	    enab->dten_maxdesc * sizeof (dtrace_enabling_t *));
10550 
10551 	/*
10552 	 * If this was a retained enabling, decrement the dts_nretained count
10553 	 * and take it off of the dtrace_retained list.
10554 	 */
10555 	if (enab->dten_prev != NULL || enab->dten_next != NULL ||
10556 	    dtrace_retained == enab) {
10557 		ASSERT(enab->dten_vstate->dtvs_state != NULL);
10558 		ASSERT(enab->dten_vstate->dtvs_state->dts_nretained > 0);
10559 		enab->dten_vstate->dtvs_state->dts_nretained--;
10560 	}
10561 
10562 	if (enab->dten_prev == NULL) {
10563 		if (dtrace_retained == enab) {
10564 			dtrace_retained = enab->dten_next;
10565 
10566 			if (dtrace_retained != NULL)
10567 				dtrace_retained->dten_prev = NULL;
10568 		}
10569 	} else {
10570 		ASSERT(enab != dtrace_retained);
10571 		ASSERT(dtrace_retained != NULL);
10572 		enab->dten_prev->dten_next = enab->dten_next;
10573 	}
10574 
10575 	if (enab->dten_next != NULL) {
10576 		ASSERT(dtrace_retained != NULL);
10577 		enab->dten_next->dten_prev = enab->dten_prev;
10578 	}
10579 
10580 	kmem_free(enab, sizeof (dtrace_enabling_t));
10581 }
10582 
10583 static int
10584 dtrace_enabling_retain(dtrace_enabling_t *enab)
10585 {
10586 	dtrace_state_t *state;
10587 
10588 	ASSERT(MUTEX_HELD(&dtrace_lock));
10589 	ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
10590 	ASSERT(enab->dten_vstate != NULL);
10591 
10592 	state = enab->dten_vstate->dtvs_state;
10593 	ASSERT(state != NULL);
10594 
10595 	/*
10596 	 * We only allow each state to retain dtrace_retain_max enablings.
10597 	 */
10598 	if (state->dts_nretained >= dtrace_retain_max)
10599 		return (ENOSPC);
10600 
10601 	state->dts_nretained++;
10602 
10603 	if (dtrace_retained == NULL) {
10604 		dtrace_retained = enab;
10605 		return (0);
10606 	}
10607 
10608 	enab->dten_next = dtrace_retained;
10609 	dtrace_retained->dten_prev = enab;
10610 	dtrace_retained = enab;
10611 
10612 	return (0);
10613 }
10614 
10615 static int
10616 dtrace_enabling_replicate(dtrace_state_t *state, dtrace_probedesc_t *match,
10617     dtrace_probedesc_t *create)
10618 {
10619 	dtrace_enabling_t *new, *enab;
10620 	int found = 0, err = ENOENT;
10621 
10622 	ASSERT(MUTEX_HELD(&dtrace_lock));
10623 	ASSERT(strlen(match->dtpd_provider) < DTRACE_PROVNAMELEN);
10624 	ASSERT(strlen(match->dtpd_mod) < DTRACE_MODNAMELEN);
10625 	ASSERT(strlen(match->dtpd_func) < DTRACE_FUNCNAMELEN);
10626 	ASSERT(strlen(match->dtpd_name) < DTRACE_NAMELEN);
10627 
10628 	new = dtrace_enabling_create(&state->dts_vstate);
10629 
10630 	/*
10631 	 * Iterate over all retained enablings, looking for enablings that
10632 	 * match the specified state.
10633 	 */
10634 	for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
10635 		int i;
10636 
10637 		/*
10638 		 * dtvs_state can only be NULL for helper enablings -- and
10639 		 * helper enablings can't be retained.
10640 		 */
10641 		ASSERT(enab->dten_vstate->dtvs_state != NULL);
10642 
10643 		if (enab->dten_vstate->dtvs_state != state)
10644 			continue;
10645 
10646 		/*
10647 		 * Now iterate over each probe description; we're looking for
10648 		 * an exact match to the specified probe description.
10649 		 */
10650 		for (i = 0; i < enab->dten_ndesc; i++) {
10651 			dtrace_ecbdesc_t *ep = enab->dten_desc[i];
10652 			dtrace_probedesc_t *pd = &ep->dted_probe;
10653 
10654 			if (strcmp(pd->dtpd_provider, match->dtpd_provider))
10655 				continue;
10656 
10657 			if (strcmp(pd->dtpd_mod, match->dtpd_mod))
10658 				continue;
10659 
10660 			if (strcmp(pd->dtpd_func, match->dtpd_func))
10661 				continue;
10662 
10663 			if (strcmp(pd->dtpd_name, match->dtpd_name))
10664 				continue;
10665 
10666 			/*
10667 			 * We have a winning probe!  Add it to our growing
10668 			 * enabling.
10669 			 */
10670 			found = 1;
10671 			dtrace_enabling_addlike(new, ep, create);
10672 		}
10673 	}
10674 
10675 	if (!found || (err = dtrace_enabling_retain(new)) != 0) {
10676 		dtrace_enabling_destroy(new);
10677 		return (err);
10678 	}
10679 
10680 	return (0);
10681 }
10682 
10683 static void
10684 dtrace_enabling_retract(dtrace_state_t *state)
10685 {
10686 	dtrace_enabling_t *enab, *next;
10687 
10688 	ASSERT(MUTEX_HELD(&dtrace_lock));
10689 
10690 	/*
10691 	 * Iterate over all retained enablings, destroy the enablings retained
10692 	 * for the specified state.
10693 	 */
10694 	for (enab = dtrace_retained; enab != NULL; enab = next) {
10695 		next = enab->dten_next;
10696 
10697 		/*
10698 		 * dtvs_state can only be NULL for helper enablings -- and
10699 		 * helper enablings can't be retained.
10700 		 */
10701 		ASSERT(enab->dten_vstate->dtvs_state != NULL);
10702 
10703 		if (enab->dten_vstate->dtvs_state == state) {
10704 			ASSERT(state->dts_nretained > 0);
10705 			dtrace_enabling_destroy(enab);
10706 		}
10707 	}
10708 
10709 	ASSERT(state->dts_nretained == 0);
10710 }
10711 
10712 static int
10713 dtrace_enabling_match(dtrace_enabling_t *enab, int *nmatched)
10714 {
10715 	int i = 0;
10716 	int matched = 0;
10717 
10718 	ASSERT(MUTEX_HELD(&cpu_lock));
10719 	ASSERT(MUTEX_HELD(&dtrace_lock));
10720 
10721 	for (i = 0; i < enab->dten_ndesc; i++) {
10722 		dtrace_ecbdesc_t *ep = enab->dten_desc[i];
10723 
10724 		enab->dten_current = ep;
10725 		enab->dten_error = 0;
10726 
10727 		matched += dtrace_probe_enable(&ep->dted_probe, enab);
10728 
10729 		if (enab->dten_error != 0) {
10730 			/*
10731 			 * If we get an error half-way through enabling the
10732 			 * probes, we kick out -- perhaps with some number of
10733 			 * them enabled.  Leaving enabled probes enabled may
10734 			 * be slightly confusing for user-level, but we expect
10735 			 * that no one will attempt to actually drive on in
10736 			 * the face of such errors.  If this is an anonymous
10737 			 * enabling (indicated with a NULL nmatched pointer),
10738 			 * we cmn_err() a message.  We aren't expecting to
10739 			 * get such an error -- such as it can exist at all,
10740 			 * it would be a result of corrupted DOF in the driver
10741 			 * properties.
10742 			 */
10743 			if (nmatched == NULL) {
10744 				cmn_err(CE_WARN, "dtrace_enabling_match() "
10745 				    "error on %p: %d", (void *)ep,
10746 				    enab->dten_error);
10747 			}
10748 
10749 			return (enab->dten_error);
10750 		}
10751 	}
10752 
10753 	enab->dten_probegen = dtrace_probegen;
10754 	if (nmatched != NULL)
10755 		*nmatched = matched;
10756 
10757 	return (0);
10758 }
10759 
10760 static void
10761 dtrace_enabling_matchall(void)
10762 {
10763 	dtrace_enabling_t *enab;
10764 
10765 	mutex_enter(&cpu_lock);
10766 	mutex_enter(&dtrace_lock);
10767 
10768 	/*
10769 	 * Iterate over all retained enablings to see if any probes match
10770 	 * against them.  We only perform this operation on enablings for which
10771 	 * we have sufficient permissions by virtue of being in the global zone
10772 	 * or in the same zone as the DTrace client.  Because we can be called
10773 	 * after dtrace_detach() has been called, we cannot assert that there
10774 	 * are retained enablings.  We can safely load from dtrace_retained,
10775 	 * however:  the taskq_destroy() at the end of dtrace_detach() will
10776 	 * block pending our completion.
10777 	 */
10778 	for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
10779 		cred_t *cr = enab->dten_vstate->dtvs_state->dts_cred.dcr_cred;
10780 
10781 		if (INGLOBALZONE(curproc) ||
10782 		    cr != NULL && getzoneid() == crgetzoneid(cr))
10783 			(void) dtrace_enabling_match(enab, NULL);
10784 	}
10785 
10786 	mutex_exit(&dtrace_lock);
10787 	mutex_exit(&cpu_lock);
10788 }
10789 
10790 /*
10791  * If an enabling is to be enabled without having matched probes (that is, if
10792  * dtrace_state_go() is to be called on the underlying dtrace_state_t), the
10793  * enabling must be _primed_ by creating an ECB for every ECB description.
10794  * This must be done to assure that we know the number of speculations, the
10795  * number of aggregations, the minimum buffer size needed, etc. before we
10796  * transition out of DTRACE_ACTIVITY_INACTIVE.  To do this without actually
10797  * enabling any probes, we create ECBs for every ECB decription, but with a
10798  * NULL probe -- which is exactly what this function does.
10799  */
10800 static void
10801 dtrace_enabling_prime(dtrace_state_t *state)
10802 {
10803 	dtrace_enabling_t *enab;
10804 	int i;
10805 
10806 	for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
10807 		ASSERT(enab->dten_vstate->dtvs_state != NULL);
10808 
10809 		if (enab->dten_vstate->dtvs_state != state)
10810 			continue;
10811 
10812 		/*
10813 		 * We don't want to prime an enabling more than once, lest
10814 		 * we allow a malicious user to induce resource exhaustion.
10815 		 * (The ECBs that result from priming an enabling aren't
10816 		 * leaked -- but they also aren't deallocated until the
10817 		 * consumer state is destroyed.)
10818 		 */
10819 		if (enab->dten_primed)
10820 			continue;
10821 
10822 		for (i = 0; i < enab->dten_ndesc; i++) {
10823 			enab->dten_current = enab->dten_desc[i];
10824 			(void) dtrace_probe_enable(NULL, enab);
10825 		}
10826 
10827 		enab->dten_primed = 1;
10828 	}
10829 }
10830 
10831 /*
10832  * Called to indicate that probes should be provided due to retained
10833  * enablings.  This is implemented in terms of dtrace_probe_provide(), but it
10834  * must take an initial lap through the enabling calling the dtps_provide()
10835  * entry point explicitly to allow for autocreated probes.
10836  */
10837 static void
10838 dtrace_enabling_provide(dtrace_provider_t *prv)
10839 {
10840 	int i, all = 0;
10841 	dtrace_probedesc_t desc;
10842 
10843 	ASSERT(MUTEX_HELD(&dtrace_lock));
10844 	ASSERT(MUTEX_HELD(&dtrace_provider_lock));
10845 
10846 	if (prv == NULL) {
10847 		all = 1;
10848 		prv = dtrace_provider;
10849 	}
10850 
10851 	do {
10852 		dtrace_enabling_t *enab = dtrace_retained;
10853 		void *parg = prv->dtpv_arg;
10854 
10855 		for (; enab != NULL; enab = enab->dten_next) {
10856 			for (i = 0; i < enab->dten_ndesc; i++) {
10857 				desc = enab->dten_desc[i]->dted_probe;
10858 				mutex_exit(&dtrace_lock);
10859 				prv->dtpv_pops.dtps_provide(parg, &desc);
10860 				mutex_enter(&dtrace_lock);
10861 			}
10862 		}
10863 	} while (all && (prv = prv->dtpv_next) != NULL);
10864 
10865 	mutex_exit(&dtrace_lock);
10866 	dtrace_probe_provide(NULL, all ? NULL : prv);
10867 	mutex_enter(&dtrace_lock);
10868 }
10869 
10870 /*
10871  * DTrace DOF Functions
10872  */
10873 /*ARGSUSED*/
10874 static void
10875 dtrace_dof_error(dof_hdr_t *dof, const char *str)
10876 {
10877 	if (dtrace_err_verbose)
10878 		cmn_err(CE_WARN, "failed to process DOF: %s", str);
10879 
10880 #ifdef DTRACE_ERRDEBUG
10881 	dtrace_errdebug(str);
10882 #endif
10883 }
10884 
10885 /*
10886  * Create DOF out of a currently enabled state.  Right now, we only create
10887  * DOF containing the run-time options -- but this could be expanded to create
10888  * complete DOF representing the enabled state.
10889  */
10890 static dof_hdr_t *
10891 dtrace_dof_create(dtrace_state_t *state)
10892 {
10893 	dof_hdr_t *dof;
10894 	dof_sec_t *sec;
10895 	dof_optdesc_t *opt;
10896 	int i, len = sizeof (dof_hdr_t) +
10897 	    roundup(sizeof (dof_sec_t), sizeof (uint64_t)) +
10898 	    sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
10899 
10900 	ASSERT(MUTEX_HELD(&dtrace_lock));
10901 
10902 	dof = kmem_zalloc(len, KM_SLEEP);
10903 	dof->dofh_ident[DOF_ID_MAG0] = DOF_MAG_MAG0;
10904 	dof->dofh_ident[DOF_ID_MAG1] = DOF_MAG_MAG1;
10905 	dof->dofh_ident[DOF_ID_MAG2] = DOF_MAG_MAG2;
10906 	dof->dofh_ident[DOF_ID_MAG3] = DOF_MAG_MAG3;
10907 
10908 	dof->dofh_ident[DOF_ID_MODEL] = DOF_MODEL_NATIVE;
10909 	dof->dofh_ident[DOF_ID_ENCODING] = DOF_ENCODE_NATIVE;
10910 	dof->dofh_ident[DOF_ID_VERSION] = DOF_VERSION;
10911 	dof->dofh_ident[DOF_ID_DIFVERS] = DIF_VERSION;
10912 	dof->dofh_ident[DOF_ID_DIFIREG] = DIF_DIR_NREGS;
10913 	dof->dofh_ident[DOF_ID_DIFTREG] = DIF_DTR_NREGS;
10914 
10915 	dof->dofh_flags = 0;
10916 	dof->dofh_hdrsize = sizeof (dof_hdr_t);
10917 	dof->dofh_secsize = sizeof (dof_sec_t);
10918 	dof->dofh_secnum = 1;	/* only DOF_SECT_OPTDESC */
10919 	dof->dofh_secoff = sizeof (dof_hdr_t);
10920 	dof->dofh_loadsz = len;
10921 	dof->dofh_filesz = len;
10922 	dof->dofh_pad = 0;
10923 
10924 	/*
10925 	 * Fill in the option section header...
10926 	 */
10927 	sec = (dof_sec_t *)((uintptr_t)dof + sizeof (dof_hdr_t));
10928 	sec->dofs_type = DOF_SECT_OPTDESC;
10929 	sec->dofs_align = sizeof (uint64_t);
10930 	sec->dofs_flags = DOF_SECF_LOAD;
10931 	sec->dofs_entsize = sizeof (dof_optdesc_t);
10932 
10933 	opt = (dof_optdesc_t *)((uintptr_t)sec +
10934 	    roundup(sizeof (dof_sec_t), sizeof (uint64_t)));
10935 
10936 	sec->dofs_offset = (uintptr_t)opt - (uintptr_t)dof;
10937 	sec->dofs_size = sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
10938 
10939 	for (i = 0; i < DTRACEOPT_MAX; i++) {
10940 		opt[i].dofo_option = i;
10941 		opt[i].dofo_strtab = DOF_SECIDX_NONE;
10942 		opt[i].dofo_value = state->dts_options[i];
10943 	}
10944 
10945 	return (dof);
10946 }
10947 
10948 static dof_hdr_t *
10949 dtrace_dof_copyin(uintptr_t uarg, int *errp)
10950 {
10951 	dof_hdr_t hdr, *dof;
10952 
10953 	ASSERT(!MUTEX_HELD(&dtrace_lock));
10954 
10955 	/*
10956 	 * First, we're going to copyin() the sizeof (dof_hdr_t).
10957 	 */
10958 	if (copyin((void *)uarg, &hdr, sizeof (hdr)) != 0) {
10959 		dtrace_dof_error(NULL, "failed to copyin DOF header");
10960 		*errp = EFAULT;
10961 		return (NULL);
10962 	}
10963 
10964 	/*
10965 	 * Now we'll allocate the entire DOF and copy it in -- provided
10966 	 * that the length isn't outrageous.
10967 	 */
10968 	if (hdr.dofh_loadsz >= dtrace_dof_maxsize) {
10969 		dtrace_dof_error(&hdr, "load size exceeds maximum");
10970 		*errp = E2BIG;
10971 		return (NULL);
10972 	}
10973 
10974 	if (hdr.dofh_loadsz < sizeof (hdr)) {
10975 		dtrace_dof_error(&hdr, "invalid load size");
10976 		*errp = EINVAL;
10977 		return (NULL);
10978 	}
10979 
10980 	dof = kmem_alloc(hdr.dofh_loadsz, KM_SLEEP);
10981 
10982 	if (copyin((void *)uarg, dof, hdr.dofh_loadsz) != 0) {
10983 		kmem_free(dof, hdr.dofh_loadsz);
10984 		*errp = EFAULT;
10985 		return (NULL);
10986 	}
10987 
10988 	return (dof);
10989 }
10990 
10991 static dof_hdr_t *
10992 dtrace_dof_property(const char *name)
10993 {
10994 	uchar_t *buf;
10995 	uint64_t loadsz;
10996 	unsigned int len, i;
10997 	dof_hdr_t *dof;
10998 
10999 	/*
11000 	 * Unfortunately, array of values in .conf files are always (and
11001 	 * only) interpreted to be integer arrays.  We must read our DOF
11002 	 * as an integer array, and then squeeze it into a byte array.
11003 	 */
11004 	if (ddi_prop_lookup_int_array(DDI_DEV_T_ANY, dtrace_devi, 0,
11005 	    (char *)name, (int **)&buf, &len) != DDI_PROP_SUCCESS)
11006 		return (NULL);
11007 
11008 	for (i = 0; i < len; i++)
11009 		buf[i] = (uchar_t)(((int *)buf)[i]);
11010 
11011 	if (len < sizeof (dof_hdr_t)) {
11012 		ddi_prop_free(buf);
11013 		dtrace_dof_error(NULL, "truncated header");
11014 		return (NULL);
11015 	}
11016 
11017 	if (len < (loadsz = ((dof_hdr_t *)buf)->dofh_loadsz)) {
11018 		ddi_prop_free(buf);
11019 		dtrace_dof_error(NULL, "truncated DOF");
11020 		return (NULL);
11021 	}
11022 
11023 	if (loadsz >= dtrace_dof_maxsize) {
11024 		ddi_prop_free(buf);
11025 		dtrace_dof_error(NULL, "oversized DOF");
11026 		return (NULL);
11027 	}
11028 
11029 	dof = kmem_alloc(loadsz, KM_SLEEP);
11030 	bcopy(buf, dof, loadsz);
11031 	ddi_prop_free(buf);
11032 
11033 	return (dof);
11034 }
11035 
11036 static void
11037 dtrace_dof_destroy(dof_hdr_t *dof)
11038 {
11039 	kmem_free(dof, dof->dofh_loadsz);
11040 }
11041 
11042 /*
11043  * Return the dof_sec_t pointer corresponding to a given section index.  If the
11044  * index is not valid, dtrace_dof_error() is called and NULL is returned.  If
11045  * a type other than DOF_SECT_NONE is specified, the header is checked against
11046  * this type and NULL is returned if the types do not match.
11047  */
11048 static dof_sec_t *
11049 dtrace_dof_sect(dof_hdr_t *dof, uint32_t type, dof_secidx_t i)
11050 {
11051 	dof_sec_t *sec = (dof_sec_t *)(uintptr_t)
11052 	    ((uintptr_t)dof + dof->dofh_secoff + i * dof->dofh_secsize);
11053 
11054 	if (i >= dof->dofh_secnum) {
11055 		dtrace_dof_error(dof, "referenced section index is invalid");
11056 		return (NULL);
11057 	}
11058 
11059 	if (!(sec->dofs_flags & DOF_SECF_LOAD)) {
11060 		dtrace_dof_error(dof, "referenced section is not loadable");
11061 		return (NULL);
11062 	}
11063 
11064 	if (type != DOF_SECT_NONE && type != sec->dofs_type) {
11065 		dtrace_dof_error(dof, "referenced section is the wrong type");
11066 		return (NULL);
11067 	}
11068 
11069 	return (sec);
11070 }
11071 
11072 static dtrace_probedesc_t *
11073 dtrace_dof_probedesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_probedesc_t *desc)
11074 {
11075 	dof_probedesc_t *probe;
11076 	dof_sec_t *strtab;
11077 	uintptr_t daddr = (uintptr_t)dof;
11078 	uintptr_t str;
11079 	size_t size;
11080 
11081 	if (sec->dofs_type != DOF_SECT_PROBEDESC) {
11082 		dtrace_dof_error(dof, "invalid probe section");
11083 		return (NULL);
11084 	}
11085 
11086 	if (sec->dofs_align != sizeof (dof_secidx_t)) {
11087 		dtrace_dof_error(dof, "bad alignment in probe description");
11088 		return (NULL);
11089 	}
11090 
11091 	if (sec->dofs_offset + sizeof (dof_probedesc_t) > dof->dofh_loadsz) {
11092 		dtrace_dof_error(dof, "truncated probe description");
11093 		return (NULL);
11094 	}
11095 
11096 	probe = (dof_probedesc_t *)(uintptr_t)(daddr + sec->dofs_offset);
11097 	strtab = dtrace_dof_sect(dof, DOF_SECT_STRTAB, probe->dofp_strtab);
11098 
11099 	if (strtab == NULL)
11100 		return (NULL);
11101 
11102 	str = daddr + strtab->dofs_offset;
11103 	size = strtab->dofs_size;
11104 
11105 	if (probe->dofp_provider >= strtab->dofs_size) {
11106 		dtrace_dof_error(dof, "corrupt probe provider");
11107 		return (NULL);
11108 	}
11109 
11110 	(void) strncpy(desc->dtpd_provider,
11111 	    (char *)(str + probe->dofp_provider),
11112 	    MIN(DTRACE_PROVNAMELEN - 1, size - probe->dofp_provider));
11113 
11114 	if (probe->dofp_mod >= strtab->dofs_size) {
11115 		dtrace_dof_error(dof, "corrupt probe module");
11116 		return (NULL);
11117 	}
11118 
11119 	(void) strncpy(desc->dtpd_mod, (char *)(str + probe->dofp_mod),
11120 	    MIN(DTRACE_MODNAMELEN - 1, size - probe->dofp_mod));
11121 
11122 	if (probe->dofp_func >= strtab->dofs_size) {
11123 		dtrace_dof_error(dof, "corrupt probe function");
11124 		return (NULL);
11125 	}
11126 
11127 	(void) strncpy(desc->dtpd_func, (char *)(str + probe->dofp_func),
11128 	    MIN(DTRACE_FUNCNAMELEN - 1, size - probe->dofp_func));
11129 
11130 	if (probe->dofp_name >= strtab->dofs_size) {
11131 		dtrace_dof_error(dof, "corrupt probe name");
11132 		return (NULL);
11133 	}
11134 
11135 	(void) strncpy(desc->dtpd_name, (char *)(str + probe->dofp_name),
11136 	    MIN(DTRACE_NAMELEN - 1, size - probe->dofp_name));
11137 
11138 	return (desc);
11139 }
11140 
11141 static dtrace_difo_t *
11142 dtrace_dof_difo(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
11143     cred_t *cr)
11144 {
11145 	dtrace_difo_t *dp;
11146 	size_t ttl = 0;
11147 	dof_difohdr_t *dofd;
11148 	uintptr_t daddr = (uintptr_t)dof;
11149 	size_t max = dtrace_difo_maxsize;
11150 	int i, l, n;
11151 
11152 	static const struct {
11153 		int section;
11154 		int bufoffs;
11155 		int lenoffs;
11156 		int entsize;
11157 		int align;
11158 		const char *msg;
11159 	} difo[] = {
11160 		{ DOF_SECT_DIF, offsetof(dtrace_difo_t, dtdo_buf),
11161 		offsetof(dtrace_difo_t, dtdo_len), sizeof (dif_instr_t),
11162 		sizeof (dif_instr_t), "multiple DIF sections" },
11163 
11164 		{ DOF_SECT_INTTAB, offsetof(dtrace_difo_t, dtdo_inttab),
11165 		offsetof(dtrace_difo_t, dtdo_intlen), sizeof (uint64_t),
11166 		sizeof (uint64_t), "multiple integer tables" },
11167 
11168 		{ DOF_SECT_STRTAB, offsetof(dtrace_difo_t, dtdo_strtab),
11169 		offsetof(dtrace_difo_t, dtdo_strlen), 0,
11170 		sizeof (char), "multiple string tables" },
11171 
11172 		{ DOF_SECT_VARTAB, offsetof(dtrace_difo_t, dtdo_vartab),
11173 		offsetof(dtrace_difo_t, dtdo_varlen), sizeof (dtrace_difv_t),
11174 		sizeof (uint_t), "multiple variable tables" },
11175 
11176 		{ DOF_SECT_NONE, 0, 0, 0, NULL }
11177 	};
11178 
11179 	if (sec->dofs_type != DOF_SECT_DIFOHDR) {
11180 		dtrace_dof_error(dof, "invalid DIFO header section");
11181 		return (NULL);
11182 	}
11183 
11184 	if (sec->dofs_align != sizeof (dof_secidx_t)) {
11185 		dtrace_dof_error(dof, "bad alignment in DIFO header");
11186 		return (NULL);
11187 	}
11188 
11189 	if (sec->dofs_size < sizeof (dof_difohdr_t) ||
11190 	    sec->dofs_size % sizeof (dof_secidx_t)) {
11191 		dtrace_dof_error(dof, "bad size in DIFO header");
11192 		return (NULL);
11193 	}
11194 
11195 	dofd = (dof_difohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
11196 	n = (sec->dofs_size - sizeof (*dofd)) / sizeof (dof_secidx_t) + 1;
11197 
11198 	dp = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
11199 	dp->dtdo_rtype = dofd->dofd_rtype;
11200 
11201 	for (l = 0; l < n; l++) {
11202 		dof_sec_t *subsec;
11203 		void **bufp;
11204 		uint32_t *lenp;
11205 
11206 		if ((subsec = dtrace_dof_sect(dof, DOF_SECT_NONE,
11207 		    dofd->dofd_links[l])) == NULL)
11208 			goto err; /* invalid section link */
11209 
11210 		if (ttl + subsec->dofs_size > max) {
11211 			dtrace_dof_error(dof, "exceeds maximum size");
11212 			goto err;
11213 		}
11214 
11215 		ttl += subsec->dofs_size;
11216 
11217 		for (i = 0; difo[i].section != DOF_SECT_NONE; i++) {
11218 			if (subsec->dofs_type != difo[i].section)
11219 				continue;
11220 
11221 			if (!(subsec->dofs_flags & DOF_SECF_LOAD)) {
11222 				dtrace_dof_error(dof, "section not loaded");
11223 				goto err;
11224 			}
11225 
11226 			if (subsec->dofs_align != difo[i].align) {
11227 				dtrace_dof_error(dof, "bad alignment");
11228 				goto err;
11229 			}
11230 
11231 			bufp = (void **)((uintptr_t)dp + difo[i].bufoffs);
11232 			lenp = (uint32_t *)((uintptr_t)dp + difo[i].lenoffs);
11233 
11234 			if (*bufp != NULL) {
11235 				dtrace_dof_error(dof, difo[i].msg);
11236 				goto err;
11237 			}
11238 
11239 			if (difo[i].entsize != subsec->dofs_entsize) {
11240 				dtrace_dof_error(dof, "entry size mismatch");
11241 				goto err;
11242 			}
11243 
11244 			if (subsec->dofs_entsize != 0 &&
11245 			    (subsec->dofs_size % subsec->dofs_entsize) != 0) {
11246 				dtrace_dof_error(dof, "corrupt entry size");
11247 				goto err;
11248 			}
11249 
11250 			*lenp = subsec->dofs_size;
11251 			*bufp = kmem_alloc(subsec->dofs_size, KM_SLEEP);
11252 			bcopy((char *)(uintptr_t)(daddr + subsec->dofs_offset),
11253 			    *bufp, subsec->dofs_size);
11254 
11255 			if (subsec->dofs_entsize != 0)
11256 				*lenp /= subsec->dofs_entsize;
11257 
11258 			break;
11259 		}
11260 
11261 		/*
11262 		 * If we encounter a loadable DIFO sub-section that is not
11263 		 * known to us, assume this is a broken program and fail.
11264 		 */
11265 		if (difo[i].section == DOF_SECT_NONE &&
11266 		    (subsec->dofs_flags & DOF_SECF_LOAD)) {
11267 			dtrace_dof_error(dof, "unrecognized DIFO subsection");
11268 			goto err;
11269 		}
11270 	}
11271 
11272 	if (dp->dtdo_buf == NULL) {
11273 		/*
11274 		 * We can't have a DIF object without DIF text.
11275 		 */
11276 		dtrace_dof_error(dof, "missing DIF text");
11277 		goto err;
11278 	}
11279 
11280 	/*
11281 	 * Before we validate the DIF object, run through the variable table
11282 	 * looking for the strings -- if any of their size are under, we'll set
11283 	 * their size to be the system-wide default string size.  Note that
11284 	 * this should _not_ happen if the "strsize" option has been set --
11285 	 * in this case, the compiler should have set the size to reflect the
11286 	 * setting of the option.
11287 	 */
11288 	for (i = 0; i < dp->dtdo_varlen; i++) {
11289 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
11290 		dtrace_diftype_t *t = &v->dtdv_type;
11291 
11292 		if (v->dtdv_id < DIF_VAR_OTHER_UBASE)
11293 			continue;
11294 
11295 		if (t->dtdt_kind == DIF_TYPE_STRING && t->dtdt_size == 0)
11296 			t->dtdt_size = dtrace_strsize_default;
11297 	}
11298 
11299 	if (dtrace_difo_validate(dp, vstate, DIF_DIR_NREGS, cr) != 0)
11300 		goto err;
11301 
11302 	dtrace_difo_init(dp, vstate);
11303 	return (dp);
11304 
11305 err:
11306 	kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
11307 	kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
11308 	kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
11309 	kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
11310 
11311 	kmem_free(dp, sizeof (dtrace_difo_t));
11312 	return (NULL);
11313 }
11314 
11315 static dtrace_predicate_t *
11316 dtrace_dof_predicate(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
11317     cred_t *cr)
11318 {
11319 	dtrace_difo_t *dp;
11320 
11321 	if ((dp = dtrace_dof_difo(dof, sec, vstate, cr)) == NULL)
11322 		return (NULL);
11323 
11324 	return (dtrace_predicate_create(dp));
11325 }
11326 
11327 static dtrace_actdesc_t *
11328 dtrace_dof_actdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
11329     cred_t *cr)
11330 {
11331 	dtrace_actdesc_t *act, *first = NULL, *last = NULL, *next;
11332 	dof_actdesc_t *desc;
11333 	dof_sec_t *difosec;
11334 	size_t offs;
11335 	uintptr_t daddr = (uintptr_t)dof;
11336 	uint64_t arg;
11337 	dtrace_actkind_t kind;
11338 
11339 	if (sec->dofs_type != DOF_SECT_ACTDESC) {
11340 		dtrace_dof_error(dof, "invalid action section");
11341 		return (NULL);
11342 	}
11343 
11344 	if (sec->dofs_offset + sizeof (dof_actdesc_t) > dof->dofh_loadsz) {
11345 		dtrace_dof_error(dof, "truncated action description");
11346 		return (NULL);
11347 	}
11348 
11349 	if (sec->dofs_align != sizeof (uint64_t)) {
11350 		dtrace_dof_error(dof, "bad alignment in action description");
11351 		return (NULL);
11352 	}
11353 
11354 	if (sec->dofs_size < sec->dofs_entsize) {
11355 		dtrace_dof_error(dof, "section entry size exceeds total size");
11356 		return (NULL);
11357 	}
11358 
11359 	if (sec->dofs_entsize != sizeof (dof_actdesc_t)) {
11360 		dtrace_dof_error(dof, "bad entry size in action description");
11361 		return (NULL);
11362 	}
11363 
11364 	if (sec->dofs_size / sec->dofs_entsize > dtrace_actions_max) {
11365 		dtrace_dof_error(dof, "actions exceed dtrace_actions_max");
11366 		return (NULL);
11367 	}
11368 
11369 	for (offs = 0; offs < sec->dofs_size; offs += sec->dofs_entsize) {
11370 		desc = (dof_actdesc_t *)(daddr +
11371 		    (uintptr_t)sec->dofs_offset + offs);
11372 		kind = (dtrace_actkind_t)desc->dofa_kind;
11373 
11374 		if (DTRACEACT_ISPRINTFLIKE(kind) &&
11375 		    (kind != DTRACEACT_PRINTA ||
11376 		    desc->dofa_strtab != DOF_SECIDX_NONE)) {
11377 			dof_sec_t *strtab;
11378 			char *str, *fmt;
11379 			uint64_t i;
11380 
11381 			/*
11382 			 * printf()-like actions must have a format string.
11383 			 */
11384 			if ((strtab = dtrace_dof_sect(dof,
11385 			    DOF_SECT_STRTAB, desc->dofa_strtab)) == NULL)
11386 				goto err;
11387 
11388 			str = (char *)((uintptr_t)dof +
11389 			    (uintptr_t)strtab->dofs_offset);
11390 
11391 			for (i = desc->dofa_arg; i < strtab->dofs_size; i++) {
11392 				if (str[i] == '\0')
11393 					break;
11394 			}
11395 
11396 			if (i >= strtab->dofs_size) {
11397 				dtrace_dof_error(dof, "bogus format string");
11398 				goto err;
11399 			}
11400 
11401 			if (i == desc->dofa_arg) {
11402 				dtrace_dof_error(dof, "empty format string");
11403 				goto err;
11404 			}
11405 
11406 			i -= desc->dofa_arg;
11407 			fmt = kmem_alloc(i + 1, KM_SLEEP);
11408 			bcopy(&str[desc->dofa_arg], fmt, i + 1);
11409 			arg = (uint64_t)(uintptr_t)fmt;
11410 		} else {
11411 			if (kind == DTRACEACT_PRINTA) {
11412 				ASSERT(desc->dofa_strtab == DOF_SECIDX_NONE);
11413 				arg = 0;
11414 			} else {
11415 				arg = desc->dofa_arg;
11416 			}
11417 		}
11418 
11419 		act = dtrace_actdesc_create(kind, desc->dofa_ntuple,
11420 		    desc->dofa_uarg, arg);
11421 
11422 		if (last != NULL) {
11423 			last->dtad_next = act;
11424 		} else {
11425 			first = act;
11426 		}
11427 
11428 		last = act;
11429 
11430 		if (desc->dofa_difo == DOF_SECIDX_NONE)
11431 			continue;
11432 
11433 		if ((difosec = dtrace_dof_sect(dof,
11434 		    DOF_SECT_DIFOHDR, desc->dofa_difo)) == NULL)
11435 			goto err;
11436 
11437 		act->dtad_difo = dtrace_dof_difo(dof, difosec, vstate, cr);
11438 
11439 		if (act->dtad_difo == NULL)
11440 			goto err;
11441 	}
11442 
11443 	ASSERT(first != NULL);
11444 	return (first);
11445 
11446 err:
11447 	for (act = first; act != NULL; act = next) {
11448 		next = act->dtad_next;
11449 		dtrace_actdesc_release(act, vstate);
11450 	}
11451 
11452 	return (NULL);
11453 }
11454 
11455 static dtrace_ecbdesc_t *
11456 dtrace_dof_ecbdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
11457     cred_t *cr)
11458 {
11459 	dtrace_ecbdesc_t *ep;
11460 	dof_ecbdesc_t *ecb;
11461 	dtrace_probedesc_t *desc;
11462 	dtrace_predicate_t *pred = NULL;
11463 
11464 	if (sec->dofs_size < sizeof (dof_ecbdesc_t)) {
11465 		dtrace_dof_error(dof, "truncated ECB description");
11466 		return (NULL);
11467 	}
11468 
11469 	if (sec->dofs_align != sizeof (uint64_t)) {
11470 		dtrace_dof_error(dof, "bad alignment in ECB description");
11471 		return (NULL);
11472 	}
11473 
11474 	ecb = (dof_ecbdesc_t *)((uintptr_t)dof + (uintptr_t)sec->dofs_offset);
11475 	sec = dtrace_dof_sect(dof, DOF_SECT_PROBEDESC, ecb->dofe_probes);
11476 
11477 	if (sec == NULL)
11478 		return (NULL);
11479 
11480 	ep = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
11481 	ep->dted_uarg = ecb->dofe_uarg;
11482 	desc = &ep->dted_probe;
11483 
11484 	if (dtrace_dof_probedesc(dof, sec, desc) == NULL)
11485 		goto err;
11486 
11487 	if (ecb->dofe_pred != DOF_SECIDX_NONE) {
11488 		if ((sec = dtrace_dof_sect(dof,
11489 		    DOF_SECT_DIFOHDR, ecb->dofe_pred)) == NULL)
11490 			goto err;
11491 
11492 		if ((pred = dtrace_dof_predicate(dof, sec, vstate, cr)) == NULL)
11493 			goto err;
11494 
11495 		ep->dted_pred.dtpdd_predicate = pred;
11496 	}
11497 
11498 	if (ecb->dofe_actions != DOF_SECIDX_NONE) {
11499 		if ((sec = dtrace_dof_sect(dof,
11500 		    DOF_SECT_ACTDESC, ecb->dofe_actions)) == NULL)
11501 			goto err;
11502 
11503 		ep->dted_action = dtrace_dof_actdesc(dof, sec, vstate, cr);
11504 
11505 		if (ep->dted_action == NULL)
11506 			goto err;
11507 	}
11508 
11509 	return (ep);
11510 
11511 err:
11512 	if (pred != NULL)
11513 		dtrace_predicate_release(pred, vstate);
11514 	kmem_free(ep, sizeof (dtrace_ecbdesc_t));
11515 	return (NULL);
11516 }
11517 
11518 /*
11519  * Apply the relocations from the specified 'sec' (a DOF_SECT_URELHDR) to the
11520  * specified DOF.  At present, this amounts to simply adding 'ubase' to the
11521  * site of any user SETX relocations to account for load object base address.
11522  * In the future, if we need other relocations, this function can be extended.
11523  */
11524 static int
11525 dtrace_dof_relocate(dof_hdr_t *dof, dof_sec_t *sec, uint64_t ubase)
11526 {
11527 	uintptr_t daddr = (uintptr_t)dof;
11528 	dof_relohdr_t *dofr =
11529 	    (dof_relohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
11530 	dof_sec_t *ss, *rs, *ts;
11531 	dof_relodesc_t *r;
11532 	uint_t i, n;
11533 
11534 	if (sec->dofs_size < sizeof (dof_relohdr_t) ||
11535 	    sec->dofs_align != sizeof (dof_secidx_t)) {
11536 		dtrace_dof_error(dof, "invalid relocation header");
11537 		return (-1);
11538 	}
11539 
11540 	ss = dtrace_dof_sect(dof, DOF_SECT_STRTAB, dofr->dofr_strtab);
11541 	rs = dtrace_dof_sect(dof, DOF_SECT_RELTAB, dofr->dofr_relsec);
11542 	ts = dtrace_dof_sect(dof, DOF_SECT_NONE, dofr->dofr_tgtsec);
11543 
11544 	if (ss == NULL || rs == NULL || ts == NULL)
11545 		return (-1); /* dtrace_dof_error() has been called already */
11546 
11547 	if (rs->dofs_entsize < sizeof (dof_relodesc_t) ||
11548 	    rs->dofs_align != sizeof (uint64_t)) {
11549 		dtrace_dof_error(dof, "invalid relocation section");
11550 		return (-1);
11551 	}
11552 
11553 	r = (dof_relodesc_t *)(uintptr_t)(daddr + rs->dofs_offset);
11554 	n = rs->dofs_size / rs->dofs_entsize;
11555 
11556 	for (i = 0; i < n; i++) {
11557 		uintptr_t taddr = daddr + ts->dofs_offset + r->dofr_offset;
11558 
11559 		switch (r->dofr_type) {
11560 		case DOF_RELO_NONE:
11561 			break;
11562 		case DOF_RELO_SETX:
11563 			if (r->dofr_offset >= ts->dofs_size || r->dofr_offset +
11564 			    sizeof (uint64_t) > ts->dofs_size) {
11565 				dtrace_dof_error(dof, "bad relocation offset");
11566 				return (-1);
11567 			}
11568 
11569 			if (!IS_P2ALIGNED(taddr, sizeof (uint64_t))) {
11570 				dtrace_dof_error(dof, "misaligned setx relo");
11571 				return (-1);
11572 			}
11573 
11574 			*(uint64_t *)taddr += ubase;
11575 			break;
11576 		default:
11577 			dtrace_dof_error(dof, "invalid relocation type");
11578 			return (-1);
11579 		}
11580 
11581 		r = (dof_relodesc_t *)((uintptr_t)r + rs->dofs_entsize);
11582 	}
11583 
11584 	return (0);
11585 }
11586 
11587 /*
11588  * The dof_hdr_t passed to dtrace_dof_slurp() should be a partially validated
11589  * header:  it should be at the front of a memory region that is at least
11590  * sizeof (dof_hdr_t) in size -- and then at least dof_hdr.dofh_loadsz in
11591  * size.  It need not be validated in any other way.
11592  */
11593 static int
11594 dtrace_dof_slurp(dof_hdr_t *dof, dtrace_vstate_t *vstate, cred_t *cr,
11595     dtrace_enabling_t **enabp, uint64_t ubase, int noprobes)
11596 {
11597 	uint64_t len = dof->dofh_loadsz, seclen;
11598 	uintptr_t daddr = (uintptr_t)dof;
11599 	dtrace_ecbdesc_t *ep;
11600 	dtrace_enabling_t *enab;
11601 	uint_t i;
11602 
11603 	ASSERT(MUTEX_HELD(&dtrace_lock));
11604 	ASSERT(dof->dofh_loadsz >= sizeof (dof_hdr_t));
11605 
11606 	/*
11607 	 * Check the DOF header identification bytes.  In addition to checking
11608 	 * valid settings, we also verify that unused bits/bytes are zeroed so
11609 	 * we can use them later without fear of regressing existing binaries.
11610 	 */
11611 	if (bcmp(&dof->dofh_ident[DOF_ID_MAG0],
11612 	    DOF_MAG_STRING, DOF_MAG_STRLEN) != 0) {
11613 		dtrace_dof_error(dof, "DOF magic string mismatch");
11614 		return (-1);
11615 	}
11616 
11617 	if (dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_ILP32 &&
11618 	    dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_LP64) {
11619 		dtrace_dof_error(dof, "DOF has invalid data model");
11620 		return (-1);
11621 	}
11622 
11623 	if (dof->dofh_ident[DOF_ID_ENCODING] != DOF_ENCODE_NATIVE) {
11624 		dtrace_dof_error(dof, "DOF encoding mismatch");
11625 		return (-1);
11626 	}
11627 
11628 	if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
11629 	    dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_2) {
11630 		dtrace_dof_error(dof, "DOF version mismatch");
11631 		return (-1);
11632 	}
11633 
11634 	if (dof->dofh_ident[DOF_ID_DIFVERS] != DIF_VERSION_2) {
11635 		dtrace_dof_error(dof, "DOF uses unsupported instruction set");
11636 		return (-1);
11637 	}
11638 
11639 	if (dof->dofh_ident[DOF_ID_DIFIREG] > DIF_DIR_NREGS) {
11640 		dtrace_dof_error(dof, "DOF uses too many integer registers");
11641 		return (-1);
11642 	}
11643 
11644 	if (dof->dofh_ident[DOF_ID_DIFTREG] > DIF_DTR_NREGS) {
11645 		dtrace_dof_error(dof, "DOF uses too many tuple registers");
11646 		return (-1);
11647 	}
11648 
11649 	for (i = DOF_ID_PAD; i < DOF_ID_SIZE; i++) {
11650 		if (dof->dofh_ident[i] != 0) {
11651 			dtrace_dof_error(dof, "DOF has invalid ident byte set");
11652 			return (-1);
11653 		}
11654 	}
11655 
11656 	if (dof->dofh_flags & ~DOF_FL_VALID) {
11657 		dtrace_dof_error(dof, "DOF has invalid flag bits set");
11658 		return (-1);
11659 	}
11660 
11661 	if (dof->dofh_secsize == 0) {
11662 		dtrace_dof_error(dof, "zero section header size");
11663 		return (-1);
11664 	}
11665 
11666 	/*
11667 	 * Check that the section headers don't exceed the amount of DOF
11668 	 * data.  Note that we cast the section size and number of sections
11669 	 * to uint64_t's to prevent possible overflow in the multiplication.
11670 	 */
11671 	seclen = (uint64_t)dof->dofh_secnum * (uint64_t)dof->dofh_secsize;
11672 
11673 	if (dof->dofh_secoff > len || seclen > len ||
11674 	    dof->dofh_secoff + seclen > len) {
11675 		dtrace_dof_error(dof, "truncated section headers");
11676 		return (-1);
11677 	}
11678 
11679 	if (!IS_P2ALIGNED(dof->dofh_secoff, sizeof (uint64_t))) {
11680 		dtrace_dof_error(dof, "misaligned section headers");
11681 		return (-1);
11682 	}
11683 
11684 	if (!IS_P2ALIGNED(dof->dofh_secsize, sizeof (uint64_t))) {
11685 		dtrace_dof_error(dof, "misaligned section size");
11686 		return (-1);
11687 	}
11688 
11689 	/*
11690 	 * Take an initial pass through the section headers to be sure that
11691 	 * the headers don't have stray offsets.  If the 'noprobes' flag is
11692 	 * set, do not permit sections relating to providers, probes, or args.
11693 	 */
11694 	for (i = 0; i < dof->dofh_secnum; i++) {
11695 		dof_sec_t *sec = (dof_sec_t *)(daddr +
11696 		    (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
11697 
11698 		if (noprobes) {
11699 			switch (sec->dofs_type) {
11700 			case DOF_SECT_PROVIDER:
11701 			case DOF_SECT_PROBES:
11702 			case DOF_SECT_PRARGS:
11703 			case DOF_SECT_PROFFS:
11704 				dtrace_dof_error(dof, "illegal sections "
11705 				    "for enabling");
11706 				return (-1);
11707 			}
11708 		}
11709 
11710 		if (!(sec->dofs_flags & DOF_SECF_LOAD))
11711 			continue; /* just ignore non-loadable sections */
11712 
11713 		if (sec->dofs_align & (sec->dofs_align - 1)) {
11714 			dtrace_dof_error(dof, "bad section alignment");
11715 			return (-1);
11716 		}
11717 
11718 		if (sec->dofs_offset & (sec->dofs_align - 1)) {
11719 			dtrace_dof_error(dof, "misaligned section");
11720 			return (-1);
11721 		}
11722 
11723 		if (sec->dofs_offset > len || sec->dofs_size > len ||
11724 		    sec->dofs_offset + sec->dofs_size > len) {
11725 			dtrace_dof_error(dof, "corrupt section header");
11726 			return (-1);
11727 		}
11728 
11729 		if (sec->dofs_type == DOF_SECT_STRTAB && *((char *)daddr +
11730 		    sec->dofs_offset + sec->dofs_size - 1) != '\0') {
11731 			dtrace_dof_error(dof, "non-terminating string table");
11732 			return (-1);
11733 		}
11734 	}
11735 
11736 	/*
11737 	 * Take a second pass through the sections and locate and perform any
11738 	 * relocations that are present.  We do this after the first pass to
11739 	 * be sure that all sections have had their headers validated.
11740 	 */
11741 	for (i = 0; i < dof->dofh_secnum; i++) {
11742 		dof_sec_t *sec = (dof_sec_t *)(daddr +
11743 		    (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
11744 
11745 		if (!(sec->dofs_flags & DOF_SECF_LOAD))
11746 			continue; /* skip sections that are not loadable */
11747 
11748 		switch (sec->dofs_type) {
11749 		case DOF_SECT_URELHDR:
11750 			if (dtrace_dof_relocate(dof, sec, ubase) != 0)
11751 				return (-1);
11752 			break;
11753 		}
11754 	}
11755 
11756 	if ((enab = *enabp) == NULL)
11757 		enab = *enabp = dtrace_enabling_create(vstate);
11758 
11759 	for (i = 0; i < dof->dofh_secnum; i++) {
11760 		dof_sec_t *sec = (dof_sec_t *)(daddr +
11761 		    (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
11762 
11763 		if (sec->dofs_type != DOF_SECT_ECBDESC)
11764 			continue;
11765 
11766 		if ((ep = dtrace_dof_ecbdesc(dof, sec, vstate, cr)) == NULL) {
11767 			dtrace_enabling_destroy(enab);
11768 			*enabp = NULL;
11769 			return (-1);
11770 		}
11771 
11772 		dtrace_enabling_add(enab, ep);
11773 	}
11774 
11775 	return (0);
11776 }
11777 
11778 /*
11779  * Process DOF for any options.  This routine assumes that the DOF has been
11780  * at least processed by dtrace_dof_slurp().
11781  */
11782 static int
11783 dtrace_dof_options(dof_hdr_t *dof, dtrace_state_t *state)
11784 {
11785 	int i, rval;
11786 	uint32_t entsize;
11787 	size_t offs;
11788 	dof_optdesc_t *desc;
11789 
11790 	for (i = 0; i < dof->dofh_secnum; i++) {
11791 		dof_sec_t *sec = (dof_sec_t *)((uintptr_t)dof +
11792 		    (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
11793 
11794 		if (sec->dofs_type != DOF_SECT_OPTDESC)
11795 			continue;
11796 
11797 		if (sec->dofs_align != sizeof (uint64_t)) {
11798 			dtrace_dof_error(dof, "bad alignment in "
11799 			    "option description");
11800 			return (EINVAL);
11801 		}
11802 
11803 		if ((entsize = sec->dofs_entsize) == 0) {
11804 			dtrace_dof_error(dof, "zeroed option entry size");
11805 			return (EINVAL);
11806 		}
11807 
11808 		if (entsize < sizeof (dof_optdesc_t)) {
11809 			dtrace_dof_error(dof, "bad option entry size");
11810 			return (EINVAL);
11811 		}
11812 
11813 		for (offs = 0; offs < sec->dofs_size; offs += entsize) {
11814 			desc = (dof_optdesc_t *)((uintptr_t)dof +
11815 			    (uintptr_t)sec->dofs_offset + offs);
11816 
11817 			if (desc->dofo_strtab != DOF_SECIDX_NONE) {
11818 				dtrace_dof_error(dof, "non-zero option string");
11819 				return (EINVAL);
11820 			}
11821 
11822 			if (desc->dofo_value == DTRACEOPT_UNSET) {
11823 				dtrace_dof_error(dof, "unset option");
11824 				return (EINVAL);
11825 			}
11826 
11827 			if ((rval = dtrace_state_option(state,
11828 			    desc->dofo_option, desc->dofo_value)) != 0) {
11829 				dtrace_dof_error(dof, "rejected option");
11830 				return (rval);
11831 			}
11832 		}
11833 	}
11834 
11835 	return (0);
11836 }
11837 
11838 /*
11839  * DTrace Consumer State Functions
11840  */
11841 int
11842 dtrace_dstate_init(dtrace_dstate_t *dstate, size_t size)
11843 {
11844 	size_t hashsize, maxper, min, chunksize = dstate->dtds_chunksize;
11845 	void *base;
11846 	uintptr_t limit;
11847 	dtrace_dynvar_t *dvar, *next, *start;
11848 	int i;
11849 
11850 	ASSERT(MUTEX_HELD(&dtrace_lock));
11851 	ASSERT(dstate->dtds_base == NULL && dstate->dtds_percpu == NULL);
11852 
11853 	bzero(dstate, sizeof (dtrace_dstate_t));
11854 
11855 	if ((dstate->dtds_chunksize = chunksize) == 0)
11856 		dstate->dtds_chunksize = DTRACE_DYNVAR_CHUNKSIZE;
11857 
11858 	if (size < (min = dstate->dtds_chunksize + sizeof (dtrace_dynhash_t)))
11859 		size = min;
11860 
11861 	if ((base = kmem_zalloc(size, KM_NOSLEEP)) == NULL)
11862 		return (ENOMEM);
11863 
11864 	dstate->dtds_size = size;
11865 	dstate->dtds_base = base;
11866 	dstate->dtds_percpu = kmem_cache_alloc(dtrace_state_cache, KM_SLEEP);
11867 	bzero(dstate->dtds_percpu, NCPU * sizeof (dtrace_dstate_percpu_t));
11868 
11869 	hashsize = size / (dstate->dtds_chunksize + sizeof (dtrace_dynhash_t));
11870 
11871 	if (hashsize != 1 && (hashsize & 1))
11872 		hashsize--;
11873 
11874 	dstate->dtds_hashsize = hashsize;
11875 	dstate->dtds_hash = dstate->dtds_base;
11876 
11877 	/*
11878 	 * Set all of our hash buckets to point to the single sink, and (if
11879 	 * it hasn't already been set), set the sink's hash value to be the
11880 	 * sink sentinel value.  The sink is needed for dynamic variable
11881 	 * lookups to know that they have iterated over an entire, valid hash
11882 	 * chain.
11883 	 */
11884 	for (i = 0; i < hashsize; i++)
11885 		dstate->dtds_hash[i].dtdh_chain = &dtrace_dynhash_sink;
11886 
11887 	if (dtrace_dynhash_sink.dtdv_hashval != DTRACE_DYNHASH_SINK)
11888 		dtrace_dynhash_sink.dtdv_hashval = DTRACE_DYNHASH_SINK;
11889 
11890 	/*
11891 	 * Determine number of active CPUs.  Divide free list evenly among
11892 	 * active CPUs.
11893 	 */
11894 	start = (dtrace_dynvar_t *)
11895 	    ((uintptr_t)base + hashsize * sizeof (dtrace_dynhash_t));
11896 	limit = (uintptr_t)base + size;
11897 
11898 	maxper = (limit - (uintptr_t)start) / NCPU;
11899 	maxper = (maxper / dstate->dtds_chunksize) * dstate->dtds_chunksize;
11900 
11901 	for (i = 0; i < NCPU; i++) {
11902 		dstate->dtds_percpu[i].dtdsc_free = dvar = start;
11903 
11904 		/*
11905 		 * If we don't even have enough chunks to make it once through
11906 		 * NCPUs, we're just going to allocate everything to the first
11907 		 * CPU.  And if we're on the last CPU, we're going to allocate
11908 		 * whatever is left over.  In either case, we set the limit to
11909 		 * be the limit of the dynamic variable space.
11910 		 */
11911 		if (maxper == 0 || i == NCPU - 1) {
11912 			limit = (uintptr_t)base + size;
11913 			start = NULL;
11914 		} else {
11915 			limit = (uintptr_t)start + maxper;
11916 			start = (dtrace_dynvar_t *)limit;
11917 		}
11918 
11919 		ASSERT(limit <= (uintptr_t)base + size);
11920 
11921 		for (;;) {
11922 			next = (dtrace_dynvar_t *)((uintptr_t)dvar +
11923 			    dstate->dtds_chunksize);
11924 
11925 			if ((uintptr_t)next + dstate->dtds_chunksize >= limit)
11926 				break;
11927 
11928 			dvar->dtdv_next = next;
11929 			dvar = next;
11930 		}
11931 
11932 		if (maxper == 0)
11933 			break;
11934 	}
11935 
11936 	return (0);
11937 }
11938 
11939 void
11940 dtrace_dstate_fini(dtrace_dstate_t *dstate)
11941 {
11942 	ASSERT(MUTEX_HELD(&cpu_lock));
11943 
11944 	if (dstate->dtds_base == NULL)
11945 		return;
11946 
11947 	kmem_free(dstate->dtds_base, dstate->dtds_size);
11948 	kmem_cache_free(dtrace_state_cache, dstate->dtds_percpu);
11949 }
11950 
11951 static void
11952 dtrace_vstate_fini(dtrace_vstate_t *vstate)
11953 {
11954 	/*
11955 	 * Logical XOR, where are you?
11956 	 */
11957 	ASSERT((vstate->dtvs_nglobals == 0) ^ (vstate->dtvs_globals != NULL));
11958 
11959 	if (vstate->dtvs_nglobals > 0) {
11960 		kmem_free(vstate->dtvs_globals, vstate->dtvs_nglobals *
11961 		    sizeof (dtrace_statvar_t *));
11962 	}
11963 
11964 	if (vstate->dtvs_ntlocals > 0) {
11965 		kmem_free(vstate->dtvs_tlocals, vstate->dtvs_ntlocals *
11966 		    sizeof (dtrace_difv_t));
11967 	}
11968 
11969 	ASSERT((vstate->dtvs_nlocals == 0) ^ (vstate->dtvs_locals != NULL));
11970 
11971 	if (vstate->dtvs_nlocals > 0) {
11972 		kmem_free(vstate->dtvs_locals, vstate->dtvs_nlocals *
11973 		    sizeof (dtrace_statvar_t *));
11974 	}
11975 }
11976 
11977 static void
11978 dtrace_state_clean(dtrace_state_t *state)
11979 {
11980 	if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE)
11981 		return;
11982 
11983 	dtrace_dynvar_clean(&state->dts_vstate.dtvs_dynvars);
11984 	dtrace_speculation_clean(state);
11985 }
11986 
11987 static void
11988 dtrace_state_deadman(dtrace_state_t *state)
11989 {
11990 	hrtime_t now;
11991 
11992 	dtrace_sync();
11993 
11994 	now = dtrace_gethrtime();
11995 
11996 	if (state != dtrace_anon.dta_state &&
11997 	    now - state->dts_laststatus >= dtrace_deadman_user)
11998 		return;
11999 
12000 	/*
12001 	 * We must be sure that dts_alive never appears to be less than the
12002 	 * value upon entry to dtrace_state_deadman(), and because we lack a
12003 	 * dtrace_cas64(), we cannot store to it atomically.  We thus instead
12004 	 * store INT64_MAX to it, followed by a memory barrier, followed by
12005 	 * the new value.  This assures that dts_alive never appears to be
12006 	 * less than its true value, regardless of the order in which the
12007 	 * stores to the underlying storage are issued.
12008 	 */
12009 	state->dts_alive = INT64_MAX;
12010 	dtrace_membar_producer();
12011 	state->dts_alive = now;
12012 }
12013 
12014 dtrace_state_t *
12015 dtrace_state_create(dev_t *devp, cred_t *cr)
12016 {
12017 	minor_t minor;
12018 	major_t major;
12019 	char c[30];
12020 	dtrace_state_t *state;
12021 	dtrace_optval_t *opt;
12022 	int bufsize = NCPU * sizeof (dtrace_buffer_t), i;
12023 
12024 	ASSERT(MUTEX_HELD(&dtrace_lock));
12025 	ASSERT(MUTEX_HELD(&cpu_lock));
12026 
12027 	minor = (minor_t)(uintptr_t)vmem_alloc(dtrace_minor, 1,
12028 	    VM_BESTFIT | VM_SLEEP);
12029 
12030 	if (ddi_soft_state_zalloc(dtrace_softstate, minor) != DDI_SUCCESS) {
12031 		vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
12032 		return (NULL);
12033 	}
12034 
12035 	state = ddi_get_soft_state(dtrace_softstate, minor);
12036 	state->dts_epid = DTRACE_EPIDNONE + 1;
12037 
12038 	(void) snprintf(c, sizeof (c), "dtrace_aggid_%d", minor);
12039 	state->dts_aggid_arena = vmem_create(c, (void *)1, UINT32_MAX, 1,
12040 	    NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
12041 
12042 	if (devp != NULL) {
12043 		major = getemajor(*devp);
12044 	} else {
12045 		major = ddi_driver_major(dtrace_devi);
12046 	}
12047 
12048 	state->dts_dev = makedevice(major, minor);
12049 
12050 	if (devp != NULL)
12051 		*devp = state->dts_dev;
12052 
12053 	/*
12054 	 * We allocate NCPU buffers.  On the one hand, this can be quite
12055 	 * a bit of memory per instance (nearly 36K on a Starcat).  On the
12056 	 * other hand, it saves an additional memory reference in the probe
12057 	 * path.
12058 	 */
12059 	state->dts_buffer = kmem_zalloc(bufsize, KM_SLEEP);
12060 	state->dts_aggbuffer = kmem_zalloc(bufsize, KM_SLEEP);
12061 	state->dts_cleaner = CYCLIC_NONE;
12062 	state->dts_deadman = CYCLIC_NONE;
12063 	state->dts_vstate.dtvs_state = state;
12064 
12065 	for (i = 0; i < DTRACEOPT_MAX; i++)
12066 		state->dts_options[i] = DTRACEOPT_UNSET;
12067 
12068 	/*
12069 	 * Set the default options.
12070 	 */
12071 	opt = state->dts_options;
12072 	opt[DTRACEOPT_BUFPOLICY] = DTRACEOPT_BUFPOLICY_SWITCH;
12073 	opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_AUTO;
12074 	opt[DTRACEOPT_NSPEC] = dtrace_nspec_default;
12075 	opt[DTRACEOPT_SPECSIZE] = dtrace_specsize_default;
12076 	opt[DTRACEOPT_CPU] = (dtrace_optval_t)DTRACE_CPUALL;
12077 	opt[DTRACEOPT_STRSIZE] = dtrace_strsize_default;
12078 	opt[DTRACEOPT_STACKFRAMES] = dtrace_stackframes_default;
12079 	opt[DTRACEOPT_USTACKFRAMES] = dtrace_ustackframes_default;
12080 	opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_default;
12081 	opt[DTRACEOPT_AGGRATE] = dtrace_aggrate_default;
12082 	opt[DTRACEOPT_SWITCHRATE] = dtrace_switchrate_default;
12083 	opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_default;
12084 	opt[DTRACEOPT_JSTACKFRAMES] = dtrace_jstackframes_default;
12085 	opt[DTRACEOPT_JSTACKSTRSIZE] = dtrace_jstackstrsize_default;
12086 
12087 	state->dts_activity = DTRACE_ACTIVITY_INACTIVE;
12088 
12089 	/*
12090 	 * Depending on the user credentials, we set flag bits which alter probe
12091 	 * visibility or the amount of destructiveness allowed.  In the case of
12092 	 * actual anonymous tracing, or the possession of all privileges, all of
12093 	 * the normal checks are bypassed.
12094 	 */
12095 	if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
12096 		state->dts_cred.dcr_visible = DTRACE_CRV_ALL;
12097 		state->dts_cred.dcr_action = DTRACE_CRA_ALL;
12098 	} else {
12099 		/*
12100 		 * Set up the credentials for this instantiation.  We take a
12101 		 * hold on the credential to prevent it from disappearing on
12102 		 * us; this in turn prevents the zone_t referenced by this
12103 		 * credential from disappearing.  This means that we can
12104 		 * examine the credential and the zone from probe context.
12105 		 */
12106 		crhold(cr);
12107 		state->dts_cred.dcr_cred = cr;
12108 
12109 		/*
12110 		 * CRA_PROC means "we have *some* privilege for dtrace" and
12111 		 * unlocks the use of variables like pid, zonename, etc.
12112 		 */
12113 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE) ||
12114 		    PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
12115 			state->dts_cred.dcr_action |= DTRACE_CRA_PROC;
12116 		}
12117 
12118 		/*
12119 		 * dtrace_user allows use of syscall and profile providers.
12120 		 * If the user also has proc_owner and/or proc_zone, we
12121 		 * extend the scope to include additional visibility and
12122 		 * destructive power.
12123 		 */
12124 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE)) {
12125 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE)) {
12126 				state->dts_cred.dcr_visible |=
12127 				    DTRACE_CRV_ALLPROC;
12128 
12129 				state->dts_cred.dcr_action |=
12130 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
12131 			}
12132 
12133 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE)) {
12134 				state->dts_cred.dcr_visible |=
12135 				    DTRACE_CRV_ALLZONE;
12136 
12137 				state->dts_cred.dcr_action |=
12138 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
12139 			}
12140 
12141 			/*
12142 			 * If we have all privs in whatever zone this is,
12143 			 * we can do destructive things to processes which
12144 			 * have altered credentials.
12145 			 */
12146 			if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
12147 			    cr->cr_zone->zone_privset)) {
12148 				state->dts_cred.dcr_action |=
12149 				    DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
12150 			}
12151 		}
12152 
12153 		/*
12154 		 * Holding the dtrace_kernel privilege also implies that
12155 		 * the user has the dtrace_user privilege from a visibility
12156 		 * perspective.  But without further privileges, some
12157 		 * destructive actions are not available.
12158 		 */
12159 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE)) {
12160 			/*
12161 			 * Make all probes in all zones visible.  However,
12162 			 * this doesn't mean that all actions become available
12163 			 * to all zones.
12164 			 */
12165 			state->dts_cred.dcr_visible |= DTRACE_CRV_KERNEL |
12166 			    DTRACE_CRV_ALLPROC | DTRACE_CRV_ALLZONE;
12167 
12168 			state->dts_cred.dcr_action |= DTRACE_CRA_KERNEL |
12169 			    DTRACE_CRA_PROC;
12170 			/*
12171 			 * Holding proc_owner means that destructive actions
12172 			 * for *this* zone are allowed.
12173 			 */
12174 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
12175 				state->dts_cred.dcr_action |=
12176 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
12177 
12178 			/*
12179 			 * Holding proc_zone means that destructive actions
12180 			 * for this user/group ID in all zones is allowed.
12181 			 */
12182 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
12183 				state->dts_cred.dcr_action |=
12184 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
12185 
12186 			/*
12187 			 * If we have all privs in whatever zone this is,
12188 			 * we can do destructive things to processes which
12189 			 * have altered credentials.
12190 			 */
12191 			if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
12192 			    cr->cr_zone->zone_privset)) {
12193 				state->dts_cred.dcr_action |=
12194 				    DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
12195 			}
12196 		}
12197 
12198 		/*
12199 		 * Holding the dtrace_proc privilege gives control over fasttrap
12200 		 * and pid providers.  We need to grant wider destructive
12201 		 * privileges in the event that the user has proc_owner and/or
12202 		 * proc_zone.
12203 		 */
12204 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
12205 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
12206 				state->dts_cred.dcr_action |=
12207 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
12208 
12209 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
12210 				state->dts_cred.dcr_action |=
12211 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
12212 		}
12213 	}
12214 
12215 	return (state);
12216 }
12217 
12218 static int
12219 dtrace_state_buffer(dtrace_state_t *state, dtrace_buffer_t *buf, int which)
12220 {
12221 	dtrace_optval_t *opt = state->dts_options, size;
12222 	processorid_t cpu;
12223 	int flags = 0, rval;
12224 
12225 	ASSERT(MUTEX_HELD(&dtrace_lock));
12226 	ASSERT(MUTEX_HELD(&cpu_lock));
12227 	ASSERT(which < DTRACEOPT_MAX);
12228 	ASSERT(state->dts_activity == DTRACE_ACTIVITY_INACTIVE ||
12229 	    (state == dtrace_anon.dta_state &&
12230 	    state->dts_activity == DTRACE_ACTIVITY_ACTIVE));
12231 
12232 	if (opt[which] == DTRACEOPT_UNSET || opt[which] == 0)
12233 		return (0);
12234 
12235 	if (opt[DTRACEOPT_CPU] != DTRACEOPT_UNSET)
12236 		cpu = opt[DTRACEOPT_CPU];
12237 
12238 	if (which == DTRACEOPT_SPECSIZE)
12239 		flags |= DTRACEBUF_NOSWITCH;
12240 
12241 	if (which == DTRACEOPT_BUFSIZE) {
12242 		if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_RING)
12243 			flags |= DTRACEBUF_RING;
12244 
12245 		if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_FILL)
12246 			flags |= DTRACEBUF_FILL;
12247 
12248 		if (state != dtrace_anon.dta_state ||
12249 		    state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
12250 			flags |= DTRACEBUF_INACTIVE;
12251 	}
12252 
12253 	for (size = opt[which]; size >= sizeof (uint64_t); size >>= 1) {
12254 		/*
12255 		 * The size must be 8-byte aligned.  If the size is not 8-byte
12256 		 * aligned, drop it down by the difference.
12257 		 */
12258 		if (size & (sizeof (uint64_t) - 1))
12259 			size -= size & (sizeof (uint64_t) - 1);
12260 
12261 		if (size < state->dts_reserve) {
12262 			/*
12263 			 * Buffers always must be large enough to accommodate
12264 			 * their prereserved space.  We return E2BIG instead
12265 			 * of ENOMEM in this case to allow for user-level
12266 			 * software to differentiate the cases.
12267 			 */
12268 			return (E2BIG);
12269 		}
12270 
12271 		rval = dtrace_buffer_alloc(buf, size, flags, cpu);
12272 
12273 		if (rval != ENOMEM) {
12274 			opt[which] = size;
12275 			return (rval);
12276 		}
12277 
12278 		if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
12279 			return (rval);
12280 	}
12281 
12282 	return (ENOMEM);
12283 }
12284 
12285 static int
12286 dtrace_state_buffers(dtrace_state_t *state)
12287 {
12288 	dtrace_speculation_t *spec = state->dts_speculations;
12289 	int rval, i;
12290 
12291 	if ((rval = dtrace_state_buffer(state, state->dts_buffer,
12292 	    DTRACEOPT_BUFSIZE)) != 0)
12293 		return (rval);
12294 
12295 	if ((rval = dtrace_state_buffer(state, state->dts_aggbuffer,
12296 	    DTRACEOPT_AGGSIZE)) != 0)
12297 		return (rval);
12298 
12299 	for (i = 0; i < state->dts_nspeculations; i++) {
12300 		if ((rval = dtrace_state_buffer(state,
12301 		    spec[i].dtsp_buffer, DTRACEOPT_SPECSIZE)) != 0)
12302 			return (rval);
12303 	}
12304 
12305 	return (0);
12306 }
12307 
12308 static void
12309 dtrace_state_prereserve(dtrace_state_t *state)
12310 {
12311 	dtrace_ecb_t *ecb;
12312 	dtrace_probe_t *probe;
12313 
12314 	state->dts_reserve = 0;
12315 
12316 	if (state->dts_options[DTRACEOPT_BUFPOLICY] != DTRACEOPT_BUFPOLICY_FILL)
12317 		return;
12318 
12319 	/*
12320 	 * If our buffer policy is a "fill" buffer policy, we need to set the
12321 	 * prereserved space to be the space required by the END probes.
12322 	 */
12323 	probe = dtrace_probes[dtrace_probeid_end - 1];
12324 	ASSERT(probe != NULL);
12325 
12326 	for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
12327 		if (ecb->dte_state != state)
12328 			continue;
12329 
12330 		state->dts_reserve += ecb->dte_needed + ecb->dte_alignment;
12331 	}
12332 }
12333 
12334 static int
12335 dtrace_state_go(dtrace_state_t *state, processorid_t *cpu)
12336 {
12337 	dtrace_optval_t *opt = state->dts_options, sz, nspec;
12338 	dtrace_speculation_t *spec;
12339 	dtrace_buffer_t *buf;
12340 	cyc_handler_t hdlr;
12341 	cyc_time_t when;
12342 	int rval = 0, i, bufsize = NCPU * sizeof (dtrace_buffer_t);
12343 	dtrace_icookie_t cookie;
12344 
12345 	mutex_enter(&cpu_lock);
12346 	mutex_enter(&dtrace_lock);
12347 
12348 	if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
12349 		rval = EBUSY;
12350 		goto out;
12351 	}
12352 
12353 	/*
12354 	 * Before we can perform any checks, we must prime all of the
12355 	 * retained enablings that correspond to this state.
12356 	 */
12357 	dtrace_enabling_prime(state);
12358 
12359 	if (state->dts_destructive && !state->dts_cred.dcr_destructive) {
12360 		rval = EACCES;
12361 		goto out;
12362 	}
12363 
12364 	dtrace_state_prereserve(state);
12365 
12366 	/*
12367 	 * Now we want to do is try to allocate our speculations.
12368 	 * We do not automatically resize the number of speculations; if
12369 	 * this fails, we will fail the operation.
12370 	 */
12371 	nspec = opt[DTRACEOPT_NSPEC];
12372 	ASSERT(nspec != DTRACEOPT_UNSET);
12373 
12374 	if (nspec > INT_MAX) {
12375 		rval = ENOMEM;
12376 		goto out;
12377 	}
12378 
12379 	spec = kmem_zalloc(nspec * sizeof (dtrace_speculation_t), KM_NOSLEEP);
12380 
12381 	if (spec == NULL) {
12382 		rval = ENOMEM;
12383 		goto out;
12384 	}
12385 
12386 	state->dts_speculations = spec;
12387 	state->dts_nspeculations = (int)nspec;
12388 
12389 	for (i = 0; i < nspec; i++) {
12390 		if ((buf = kmem_zalloc(bufsize, KM_NOSLEEP)) == NULL) {
12391 			rval = ENOMEM;
12392 			goto err;
12393 		}
12394 
12395 		spec[i].dtsp_buffer = buf;
12396 	}
12397 
12398 	if (opt[DTRACEOPT_GRABANON] != DTRACEOPT_UNSET) {
12399 		if (dtrace_anon.dta_state == NULL) {
12400 			rval = ENOENT;
12401 			goto out;
12402 		}
12403 
12404 		if (state->dts_necbs != 0) {
12405 			rval = EALREADY;
12406 			goto out;
12407 		}
12408 
12409 		state->dts_anon = dtrace_anon_grab();
12410 		ASSERT(state->dts_anon != NULL);
12411 		state = state->dts_anon;
12412 
12413 		/*
12414 		 * We want "grabanon" to be set in the grabbed state, so we'll
12415 		 * copy that option value from the grabbing state into the
12416 		 * grabbed state.
12417 		 */
12418 		state->dts_options[DTRACEOPT_GRABANON] =
12419 		    opt[DTRACEOPT_GRABANON];
12420 
12421 		*cpu = dtrace_anon.dta_beganon;
12422 
12423 		/*
12424 		 * If the anonymous state is active (as it almost certainly
12425 		 * is if the anonymous enabling ultimately matched anything),
12426 		 * we don't allow any further option processing -- but we
12427 		 * don't return failure.
12428 		 */
12429 		if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
12430 			goto out;
12431 	}
12432 
12433 	if (opt[DTRACEOPT_AGGSIZE] != DTRACEOPT_UNSET &&
12434 	    opt[DTRACEOPT_AGGSIZE] != 0) {
12435 		if (state->dts_aggregations == NULL) {
12436 			/*
12437 			 * We're not going to create an aggregation buffer
12438 			 * because we don't have any ECBs that contain
12439 			 * aggregations -- set this option to 0.
12440 			 */
12441 			opt[DTRACEOPT_AGGSIZE] = 0;
12442 		} else {
12443 			/*
12444 			 * If we have an aggregation buffer, we must also have
12445 			 * a buffer to use as scratch.
12446 			 */
12447 			if (opt[DTRACEOPT_BUFSIZE] == DTRACEOPT_UNSET ||
12448 			    opt[DTRACEOPT_BUFSIZE] < state->dts_needed) {
12449 				opt[DTRACEOPT_BUFSIZE] = state->dts_needed;
12450 			}
12451 		}
12452 	}
12453 
12454 	if (opt[DTRACEOPT_SPECSIZE] != DTRACEOPT_UNSET &&
12455 	    opt[DTRACEOPT_SPECSIZE] != 0) {
12456 		if (!state->dts_speculates) {
12457 			/*
12458 			 * We're not going to create speculation buffers
12459 			 * because we don't have any ECBs that actually
12460 			 * speculate -- set the speculation size to 0.
12461 			 */
12462 			opt[DTRACEOPT_SPECSIZE] = 0;
12463 		}
12464 	}
12465 
12466 	/*
12467 	 * The bare minimum size for any buffer that we're actually going to
12468 	 * do anything to is sizeof (uint64_t).
12469 	 */
12470 	sz = sizeof (uint64_t);
12471 
12472 	if ((state->dts_needed != 0 && opt[DTRACEOPT_BUFSIZE] < sz) ||
12473 	    (state->dts_speculates && opt[DTRACEOPT_SPECSIZE] < sz) ||
12474 	    (state->dts_aggregations != NULL && opt[DTRACEOPT_AGGSIZE] < sz)) {
12475 		/*
12476 		 * A buffer size has been explicitly set to 0 (or to a size
12477 		 * that will be adjusted to 0) and we need the space -- we
12478 		 * need to return failure.  We return ENOSPC to differentiate
12479 		 * it from failing to allocate a buffer due to failure to meet
12480 		 * the reserve (for which we return E2BIG).
12481 		 */
12482 		rval = ENOSPC;
12483 		goto out;
12484 	}
12485 
12486 	if ((rval = dtrace_state_buffers(state)) != 0)
12487 		goto err;
12488 
12489 	if ((sz = opt[DTRACEOPT_DYNVARSIZE]) == DTRACEOPT_UNSET)
12490 		sz = dtrace_dstate_defsize;
12491 
12492 	do {
12493 		rval = dtrace_dstate_init(&state->dts_vstate.dtvs_dynvars, sz);
12494 
12495 		if (rval == 0)
12496 			break;
12497 
12498 		if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
12499 			goto err;
12500 	} while (sz >>= 1);
12501 
12502 	opt[DTRACEOPT_DYNVARSIZE] = sz;
12503 
12504 	if (rval != 0)
12505 		goto err;
12506 
12507 	if (opt[DTRACEOPT_STATUSRATE] > dtrace_statusrate_max)
12508 		opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_max;
12509 
12510 	if (opt[DTRACEOPT_CLEANRATE] == 0)
12511 		opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
12512 
12513 	if (opt[DTRACEOPT_CLEANRATE] < dtrace_cleanrate_min)
12514 		opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_min;
12515 
12516 	if (opt[DTRACEOPT_CLEANRATE] > dtrace_cleanrate_max)
12517 		opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
12518 
12519 	hdlr.cyh_func = (cyc_func_t)dtrace_state_clean;
12520 	hdlr.cyh_arg = state;
12521 	hdlr.cyh_level = CY_LOW_LEVEL;
12522 
12523 	when.cyt_when = 0;
12524 	when.cyt_interval = opt[DTRACEOPT_CLEANRATE];
12525 
12526 	state->dts_cleaner = cyclic_add(&hdlr, &when);
12527 
12528 	hdlr.cyh_func = (cyc_func_t)dtrace_state_deadman;
12529 	hdlr.cyh_arg = state;
12530 	hdlr.cyh_level = CY_LOW_LEVEL;
12531 
12532 	when.cyt_when = 0;
12533 	when.cyt_interval = dtrace_deadman_interval;
12534 
12535 	state->dts_alive = state->dts_laststatus = dtrace_gethrtime();
12536 	state->dts_deadman = cyclic_add(&hdlr, &when);
12537 
12538 	state->dts_activity = DTRACE_ACTIVITY_WARMUP;
12539 
12540 	/*
12541 	 * Now it's time to actually fire the BEGIN probe.  We need to disable
12542 	 * interrupts here both to record the CPU on which we fired the BEGIN
12543 	 * probe (the data from this CPU will be processed first at user
12544 	 * level) and to manually activate the buffer for this CPU.
12545 	 */
12546 	cookie = dtrace_interrupt_disable();
12547 	*cpu = CPU->cpu_id;
12548 	ASSERT(state->dts_buffer[*cpu].dtb_flags & DTRACEBUF_INACTIVE);
12549 	state->dts_buffer[*cpu].dtb_flags &= ~DTRACEBUF_INACTIVE;
12550 
12551 	dtrace_probe(dtrace_probeid_begin,
12552 	    (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
12553 	dtrace_interrupt_enable(cookie);
12554 	/*
12555 	 * We may have had an exit action from a BEGIN probe; only change our
12556 	 * state to ACTIVE if we're still in WARMUP.
12557 	 */
12558 	ASSERT(state->dts_activity == DTRACE_ACTIVITY_WARMUP ||
12559 	    state->dts_activity == DTRACE_ACTIVITY_DRAINING);
12560 
12561 	if (state->dts_activity == DTRACE_ACTIVITY_WARMUP)
12562 		state->dts_activity = DTRACE_ACTIVITY_ACTIVE;
12563 
12564 	/*
12565 	 * Regardless of whether or not now we're in ACTIVE or DRAINING, we
12566 	 * want each CPU to transition its principal buffer out of the
12567 	 * INACTIVE state.  Doing this assures that no CPU will suddenly begin
12568 	 * processing an ECB halfway down a probe's ECB chain; all CPUs will
12569 	 * atomically transition from processing none of a state's ECBs to
12570 	 * processing all of them.
12571 	 */
12572 	dtrace_xcall(DTRACE_CPUALL,
12573 	    (dtrace_xcall_t)dtrace_buffer_activate, state);
12574 	goto out;
12575 
12576 err:
12577 	dtrace_buffer_free(state->dts_buffer);
12578 	dtrace_buffer_free(state->dts_aggbuffer);
12579 
12580 	if ((nspec = state->dts_nspeculations) == 0) {
12581 		ASSERT(state->dts_speculations == NULL);
12582 		goto out;
12583 	}
12584 
12585 	spec = state->dts_speculations;
12586 	ASSERT(spec != NULL);
12587 
12588 	for (i = 0; i < state->dts_nspeculations; i++) {
12589 		if ((buf = spec[i].dtsp_buffer) == NULL)
12590 			break;
12591 
12592 		dtrace_buffer_free(buf);
12593 		kmem_free(buf, bufsize);
12594 	}
12595 
12596 	kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
12597 	state->dts_nspeculations = 0;
12598 	state->dts_speculations = NULL;
12599 
12600 out:
12601 	mutex_exit(&dtrace_lock);
12602 	mutex_exit(&cpu_lock);
12603 
12604 	return (rval);
12605 }
12606 
12607 static int
12608 dtrace_state_stop(dtrace_state_t *state, processorid_t *cpu)
12609 {
12610 	dtrace_icookie_t cookie;
12611 
12612 	ASSERT(MUTEX_HELD(&dtrace_lock));
12613 
12614 	if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE &&
12615 	    state->dts_activity != DTRACE_ACTIVITY_DRAINING)
12616 		return (EINVAL);
12617 
12618 	/*
12619 	 * We'll set the activity to DTRACE_ACTIVITY_DRAINING, and issue a sync
12620 	 * to be sure that every CPU has seen it.  See below for the details
12621 	 * on why this is done.
12622 	 */
12623 	state->dts_activity = DTRACE_ACTIVITY_DRAINING;
12624 	dtrace_sync();
12625 
12626 	/*
12627 	 * By this point, it is impossible for any CPU to be still processing
12628 	 * with DTRACE_ACTIVITY_ACTIVE.  We can thus set our activity to
12629 	 * DTRACE_ACTIVITY_COOLDOWN and know that we're not racing with any
12630 	 * other CPU in dtrace_buffer_reserve().  This allows dtrace_probe()
12631 	 * and callees to know that the activity is DTRACE_ACTIVITY_COOLDOWN
12632 	 * iff we're in the END probe.
12633 	 */
12634 	state->dts_activity = DTRACE_ACTIVITY_COOLDOWN;
12635 	dtrace_sync();
12636 	ASSERT(state->dts_activity == DTRACE_ACTIVITY_COOLDOWN);
12637 
12638 	/*
12639 	 * Finally, we can release the reserve and call the END probe.  We
12640 	 * disable interrupts across calling the END probe to allow us to
12641 	 * return the CPU on which we actually called the END probe.  This
12642 	 * allows user-land to be sure that this CPU's principal buffer is
12643 	 * processed last.
12644 	 */
12645 	state->dts_reserve = 0;
12646 
12647 	cookie = dtrace_interrupt_disable();
12648 	*cpu = CPU->cpu_id;
12649 	dtrace_probe(dtrace_probeid_end,
12650 	    (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
12651 	dtrace_interrupt_enable(cookie);
12652 
12653 	state->dts_activity = DTRACE_ACTIVITY_STOPPED;
12654 	dtrace_sync();
12655 
12656 	return (0);
12657 }
12658 
12659 static int
12660 dtrace_state_option(dtrace_state_t *state, dtrace_optid_t option,
12661     dtrace_optval_t val)
12662 {
12663 	ASSERT(MUTEX_HELD(&dtrace_lock));
12664 
12665 	if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
12666 		return (EBUSY);
12667 
12668 	if (option >= DTRACEOPT_MAX)
12669 		return (EINVAL);
12670 
12671 	if (option != DTRACEOPT_CPU && val < 0)
12672 		return (EINVAL);
12673 
12674 	switch (option) {
12675 	case DTRACEOPT_DESTRUCTIVE:
12676 		if (dtrace_destructive_disallow)
12677 			return (EACCES);
12678 
12679 		state->dts_cred.dcr_destructive = 1;
12680 		break;
12681 
12682 	case DTRACEOPT_BUFSIZE:
12683 	case DTRACEOPT_DYNVARSIZE:
12684 	case DTRACEOPT_AGGSIZE:
12685 	case DTRACEOPT_SPECSIZE:
12686 	case DTRACEOPT_STRSIZE:
12687 		if (val < 0)
12688 			return (EINVAL);
12689 
12690 		if (val >= LONG_MAX) {
12691 			/*
12692 			 * If this is an otherwise negative value, set it to
12693 			 * the highest multiple of 128m less than LONG_MAX.
12694 			 * Technically, we're adjusting the size without
12695 			 * regard to the buffer resizing policy, but in fact,
12696 			 * this has no effect -- if we set the buffer size to
12697 			 * ~LONG_MAX and the buffer policy is ultimately set to
12698 			 * be "manual", the buffer allocation is guaranteed to
12699 			 * fail, if only because the allocation requires two
12700 			 * buffers.  (We set the the size to the highest
12701 			 * multiple of 128m because it ensures that the size
12702 			 * will remain a multiple of a megabyte when
12703 			 * repeatedly halved -- all the way down to 15m.)
12704 			 */
12705 			val = LONG_MAX - (1 << 27) + 1;
12706 		}
12707 	}
12708 
12709 	state->dts_options[option] = val;
12710 
12711 	return (0);
12712 }
12713 
12714 static void
12715 dtrace_state_destroy(dtrace_state_t *state)
12716 {
12717 	dtrace_ecb_t *ecb;
12718 	dtrace_vstate_t *vstate = &state->dts_vstate;
12719 	minor_t minor = getminor(state->dts_dev);
12720 	int i, bufsize = NCPU * sizeof (dtrace_buffer_t);
12721 	dtrace_speculation_t *spec = state->dts_speculations;
12722 	int nspec = state->dts_nspeculations;
12723 	uint32_t match;
12724 
12725 	ASSERT(MUTEX_HELD(&dtrace_lock));
12726 	ASSERT(MUTEX_HELD(&cpu_lock));
12727 
12728 	/*
12729 	 * First, retract any retained enablings for this state.
12730 	 */
12731 	dtrace_enabling_retract(state);
12732 	ASSERT(state->dts_nretained == 0);
12733 
12734 	if (state->dts_activity == DTRACE_ACTIVITY_ACTIVE ||
12735 	    state->dts_activity == DTRACE_ACTIVITY_DRAINING) {
12736 		/*
12737 		 * We have managed to come into dtrace_state_destroy() on a
12738 		 * hot enabling -- almost certainly because of a disorderly
12739 		 * shutdown of a consumer.  (That is, a consumer that is
12740 		 * exiting without having called dtrace_stop().) In this case,
12741 		 * we're going to set our activity to be KILLED, and then
12742 		 * issue a sync to be sure that everyone is out of probe
12743 		 * context before we start blowing away ECBs.
12744 		 */
12745 		state->dts_activity = DTRACE_ACTIVITY_KILLED;
12746 		dtrace_sync();
12747 	}
12748 
12749 	/*
12750 	 * Release the credential hold we took in dtrace_state_create().
12751 	 */
12752 	if (state->dts_cred.dcr_cred != NULL)
12753 		crfree(state->dts_cred.dcr_cred);
12754 
12755 	/*
12756 	 * Now we can safely disable and destroy any enabled probes.  Because
12757 	 * any DTRACE_PRIV_KERNEL probes may actually be slowing our progress
12758 	 * (especially if they're all enabled), we take two passes through the
12759 	 * ECBs:  in the first, we disable just DTRACE_PRIV_KERNEL probes, and
12760 	 * in the second we disable whatever is left over.
12761 	 */
12762 	for (match = DTRACE_PRIV_KERNEL; ; match = 0) {
12763 		for (i = 0; i < state->dts_necbs; i++) {
12764 			if ((ecb = state->dts_ecbs[i]) == NULL)
12765 				continue;
12766 
12767 			if (match && ecb->dte_probe != NULL) {
12768 				dtrace_probe_t *probe = ecb->dte_probe;
12769 				dtrace_provider_t *prov = probe->dtpr_provider;
12770 
12771 				if (!(prov->dtpv_priv.dtpp_flags & match))
12772 					continue;
12773 			}
12774 
12775 			dtrace_ecb_disable(ecb);
12776 			dtrace_ecb_destroy(ecb);
12777 		}
12778 
12779 		if (!match)
12780 			break;
12781 	}
12782 
12783 	/*
12784 	 * Before we free the buffers, perform one more sync to assure that
12785 	 * every CPU is out of probe context.
12786 	 */
12787 	dtrace_sync();
12788 
12789 	dtrace_buffer_free(state->dts_buffer);
12790 	dtrace_buffer_free(state->dts_aggbuffer);
12791 
12792 	for (i = 0; i < nspec; i++)
12793 		dtrace_buffer_free(spec[i].dtsp_buffer);
12794 
12795 	if (state->dts_cleaner != CYCLIC_NONE)
12796 		cyclic_remove(state->dts_cleaner);
12797 
12798 	if (state->dts_deadman != CYCLIC_NONE)
12799 		cyclic_remove(state->dts_deadman);
12800 
12801 	dtrace_dstate_fini(&vstate->dtvs_dynvars);
12802 	dtrace_vstate_fini(vstate);
12803 	kmem_free(state->dts_ecbs, state->dts_necbs * sizeof (dtrace_ecb_t *));
12804 
12805 	if (state->dts_aggregations != NULL) {
12806 #ifdef DEBUG
12807 		for (i = 0; i < state->dts_naggregations; i++)
12808 			ASSERT(state->dts_aggregations[i] == NULL);
12809 #endif
12810 		ASSERT(state->dts_naggregations > 0);
12811 		kmem_free(state->dts_aggregations,
12812 		    state->dts_naggregations * sizeof (dtrace_aggregation_t *));
12813 	}
12814 
12815 	kmem_free(state->dts_buffer, bufsize);
12816 	kmem_free(state->dts_aggbuffer, bufsize);
12817 
12818 	for (i = 0; i < nspec; i++)
12819 		kmem_free(spec[i].dtsp_buffer, bufsize);
12820 
12821 	kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
12822 
12823 	dtrace_format_destroy(state);
12824 
12825 	vmem_destroy(state->dts_aggid_arena);
12826 	ddi_soft_state_free(dtrace_softstate, minor);
12827 	vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
12828 }
12829 
12830 /*
12831  * DTrace Anonymous Enabling Functions
12832  */
12833 static dtrace_state_t *
12834 dtrace_anon_grab(void)
12835 {
12836 	dtrace_state_t *state;
12837 
12838 	ASSERT(MUTEX_HELD(&dtrace_lock));
12839 
12840 	if ((state = dtrace_anon.dta_state) == NULL) {
12841 		ASSERT(dtrace_anon.dta_enabling == NULL);
12842 		return (NULL);
12843 	}
12844 
12845 	ASSERT(dtrace_anon.dta_enabling != NULL);
12846 	ASSERT(dtrace_retained != NULL);
12847 
12848 	dtrace_enabling_destroy(dtrace_anon.dta_enabling);
12849 	dtrace_anon.dta_enabling = NULL;
12850 	dtrace_anon.dta_state = NULL;
12851 
12852 	return (state);
12853 }
12854 
12855 static void
12856 dtrace_anon_property(void)
12857 {
12858 	int i, rv;
12859 	dtrace_state_t *state;
12860 	dof_hdr_t *dof;
12861 	char c[32];		/* enough for "dof-data-" + digits */
12862 
12863 	ASSERT(MUTEX_HELD(&dtrace_lock));
12864 	ASSERT(MUTEX_HELD(&cpu_lock));
12865 
12866 	for (i = 0; ; i++) {
12867 		(void) snprintf(c, sizeof (c), "dof-data-%d", i);
12868 
12869 		dtrace_err_verbose = 1;
12870 
12871 		if ((dof = dtrace_dof_property(c)) == NULL) {
12872 			dtrace_err_verbose = 0;
12873 			break;
12874 		}
12875 
12876 		/*
12877 		 * We want to create anonymous state, so we need to transition
12878 		 * the kernel debugger to indicate that DTrace is active.  If
12879 		 * this fails (e.g. because the debugger has modified text in
12880 		 * some way), we won't continue with the processing.
12881 		 */
12882 		if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
12883 			cmn_err(CE_NOTE, "kernel debugger active; anonymous "
12884 			    "enabling ignored.");
12885 			dtrace_dof_destroy(dof);
12886 			break;
12887 		}
12888 
12889 		/*
12890 		 * If we haven't allocated an anonymous state, we'll do so now.
12891 		 */
12892 		if ((state = dtrace_anon.dta_state) == NULL) {
12893 			state = dtrace_state_create(NULL, NULL);
12894 			dtrace_anon.dta_state = state;
12895 
12896 			if (state == NULL) {
12897 				/*
12898 				 * This basically shouldn't happen:  the only
12899 				 * failure mode from dtrace_state_create() is a
12900 				 * failure of ddi_soft_state_zalloc() that
12901 				 * itself should never happen.  Still, the
12902 				 * interface allows for a failure mode, and
12903 				 * we want to fail as gracefully as possible:
12904 				 * we'll emit an error message and cease
12905 				 * processing anonymous state in this case.
12906 				 */
12907 				cmn_err(CE_WARN, "failed to create "
12908 				    "anonymous state");
12909 				dtrace_dof_destroy(dof);
12910 				break;
12911 			}
12912 		}
12913 
12914 		rv = dtrace_dof_slurp(dof, &state->dts_vstate, CRED(),
12915 		    &dtrace_anon.dta_enabling, 0, B_TRUE);
12916 
12917 		if (rv == 0)
12918 			rv = dtrace_dof_options(dof, state);
12919 
12920 		dtrace_err_verbose = 0;
12921 		dtrace_dof_destroy(dof);
12922 
12923 		if (rv != 0) {
12924 			/*
12925 			 * This is malformed DOF; chuck any anonymous state
12926 			 * that we created.
12927 			 */
12928 			ASSERT(dtrace_anon.dta_enabling == NULL);
12929 			dtrace_state_destroy(state);
12930 			dtrace_anon.dta_state = NULL;
12931 			break;
12932 		}
12933 
12934 		ASSERT(dtrace_anon.dta_enabling != NULL);
12935 	}
12936 
12937 	if (dtrace_anon.dta_enabling != NULL) {
12938 		int rval;
12939 
12940 		/*
12941 		 * dtrace_enabling_retain() can only fail because we are
12942 		 * trying to retain more enablings than are allowed -- but
12943 		 * we only have one anonymous enabling, and we are guaranteed
12944 		 * to be allowed at least one retained enabling; we assert
12945 		 * that dtrace_enabling_retain() returns success.
12946 		 */
12947 		rval = dtrace_enabling_retain(dtrace_anon.dta_enabling);
12948 		ASSERT(rval == 0);
12949 
12950 		dtrace_enabling_dump(dtrace_anon.dta_enabling);
12951 	}
12952 }
12953 
12954 /*
12955  * DTrace Helper Functions
12956  */
12957 static void
12958 dtrace_helper_trace(dtrace_helper_action_t *helper,
12959     dtrace_mstate_t *mstate, dtrace_vstate_t *vstate, int where)
12960 {
12961 	uint32_t size, next, nnext, i;
12962 	dtrace_helptrace_t *ent;
12963 	uint16_t flags = cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
12964 
12965 	if (!dtrace_helptrace_enabled)
12966 		return;
12967 
12968 	ASSERT(vstate->dtvs_nlocals <= dtrace_helptrace_nlocals);
12969 
12970 	/*
12971 	 * What would a tracing framework be without its own tracing
12972 	 * framework?  (Well, a hell of a lot simpler, for starters...)
12973 	 */
12974 	size = sizeof (dtrace_helptrace_t) + dtrace_helptrace_nlocals *
12975 	    sizeof (uint64_t) - sizeof (uint64_t);
12976 
12977 	/*
12978 	 * Iterate until we can allocate a slot in the trace buffer.
12979 	 */
12980 	do {
12981 		next = dtrace_helptrace_next;
12982 
12983 		if (next + size < dtrace_helptrace_bufsize) {
12984 			nnext = next + size;
12985 		} else {
12986 			nnext = size;
12987 		}
12988 	} while (dtrace_cas32(&dtrace_helptrace_next, next, nnext) != next);
12989 
12990 	/*
12991 	 * We have our slot; fill it in.
12992 	 */
12993 	if (nnext == size)
12994 		next = 0;
12995 
12996 	ent = (dtrace_helptrace_t *)&dtrace_helptrace_buffer[next];
12997 	ent->dtht_helper = helper;
12998 	ent->dtht_where = where;
12999 	ent->dtht_nlocals = vstate->dtvs_nlocals;
13000 
13001 	ent->dtht_fltoffs = (mstate->dtms_present & DTRACE_MSTATE_FLTOFFS) ?
13002 	    mstate->dtms_fltoffs : -1;
13003 	ent->dtht_fault = DTRACE_FLAGS2FLT(flags);
13004 	ent->dtht_illval = cpu_core[CPU->cpu_id].cpuc_dtrace_illval;
13005 
13006 	for (i = 0; i < vstate->dtvs_nlocals; i++) {
13007 		dtrace_statvar_t *svar;
13008 
13009 		if ((svar = vstate->dtvs_locals[i]) == NULL)
13010 			continue;
13011 
13012 		ASSERT(svar->dtsv_size >= NCPU * sizeof (uint64_t));
13013 		ent->dtht_locals[i] =
13014 		    ((uint64_t *)(uintptr_t)svar->dtsv_data)[CPU->cpu_id];
13015 	}
13016 }
13017 
13018 static uint64_t
13019 dtrace_helper(int which, dtrace_mstate_t *mstate,
13020     dtrace_state_t *state, uint64_t arg0, uint64_t arg1)
13021 {
13022 	uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
13023 	uint64_t sarg0 = mstate->dtms_arg[0];
13024 	uint64_t sarg1 = mstate->dtms_arg[1];
13025 	uint64_t rval;
13026 	dtrace_helpers_t *helpers = curproc->p_dtrace_helpers;
13027 	dtrace_helper_action_t *helper;
13028 	dtrace_vstate_t *vstate;
13029 	dtrace_difo_t *pred;
13030 	int i, trace = dtrace_helptrace_enabled;
13031 
13032 	ASSERT(which >= 0 && which < DTRACE_NHELPER_ACTIONS);
13033 
13034 	if (helpers == NULL)
13035 		return (0);
13036 
13037 	if ((helper = helpers->dthps_actions[which]) == NULL)
13038 		return (0);
13039 
13040 	vstate = &helpers->dthps_vstate;
13041 	mstate->dtms_arg[0] = arg0;
13042 	mstate->dtms_arg[1] = arg1;
13043 
13044 	/*
13045 	 * Now iterate over each helper.  If its predicate evaluates to 'true',
13046 	 * we'll call the corresponding actions.  Note that the below calls
13047 	 * to dtrace_dif_emulate() may set faults in machine state.  This is
13048 	 * okay:  our caller (the outer dtrace_dif_emulate()) will simply plow
13049 	 * the stored DIF offset with its own (which is the desired behavior).
13050 	 * Also, note the calls to dtrace_dif_emulate() may allocate scratch
13051 	 * from machine state; this is okay, too.
13052 	 */
13053 	for (; helper != NULL; helper = helper->dtha_next) {
13054 		if ((pred = helper->dtha_predicate) != NULL) {
13055 			if (trace)
13056 				dtrace_helper_trace(helper, mstate, vstate, 0);
13057 
13058 			if (!dtrace_dif_emulate(pred, mstate, vstate, state))
13059 				goto next;
13060 
13061 			if (*flags & CPU_DTRACE_FAULT)
13062 				goto err;
13063 		}
13064 
13065 		for (i = 0; i < helper->dtha_nactions; i++) {
13066 			if (trace)
13067 				dtrace_helper_trace(helper,
13068 				    mstate, vstate, i + 1);
13069 
13070 			rval = dtrace_dif_emulate(helper->dtha_actions[i],
13071 			    mstate, vstate, state);
13072 
13073 			if (*flags & CPU_DTRACE_FAULT)
13074 				goto err;
13075 		}
13076 
13077 next:
13078 		if (trace)
13079 			dtrace_helper_trace(helper, mstate, vstate,
13080 			    DTRACE_HELPTRACE_NEXT);
13081 	}
13082 
13083 	if (trace)
13084 		dtrace_helper_trace(helper, mstate, vstate,
13085 		    DTRACE_HELPTRACE_DONE);
13086 
13087 	/*
13088 	 * Restore the arg0 that we saved upon entry.
13089 	 */
13090 	mstate->dtms_arg[0] = sarg0;
13091 	mstate->dtms_arg[1] = sarg1;
13092 
13093 	return (rval);
13094 
13095 err:
13096 	if (trace)
13097 		dtrace_helper_trace(helper, mstate, vstate,
13098 		    DTRACE_HELPTRACE_ERR);
13099 
13100 	/*
13101 	 * Restore the arg0 that we saved upon entry.
13102 	 */
13103 	mstate->dtms_arg[0] = sarg0;
13104 	mstate->dtms_arg[1] = sarg1;
13105 
13106 	return (NULL);
13107 }
13108 
13109 static void
13110 dtrace_helper_action_destroy(dtrace_helper_action_t *helper,
13111     dtrace_vstate_t *vstate)
13112 {
13113 	int i;
13114 
13115 	if (helper->dtha_predicate != NULL)
13116 		dtrace_difo_release(helper->dtha_predicate, vstate);
13117 
13118 	for (i = 0; i < helper->dtha_nactions; i++) {
13119 		ASSERT(helper->dtha_actions[i] != NULL);
13120 		dtrace_difo_release(helper->dtha_actions[i], vstate);
13121 	}
13122 
13123 	kmem_free(helper->dtha_actions,
13124 	    helper->dtha_nactions * sizeof (dtrace_difo_t *));
13125 	kmem_free(helper, sizeof (dtrace_helper_action_t));
13126 }
13127 
13128 static int
13129 dtrace_helper_destroygen(int gen)
13130 {
13131 	proc_t *p = curproc;
13132 	dtrace_helpers_t *help = p->p_dtrace_helpers;
13133 	dtrace_vstate_t *vstate;
13134 	int i;
13135 
13136 	ASSERT(MUTEX_HELD(&dtrace_lock));
13137 
13138 	if (help == NULL || gen > help->dthps_generation)
13139 		return (EINVAL);
13140 
13141 	vstate = &help->dthps_vstate;
13142 
13143 	for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
13144 		dtrace_helper_action_t *last = NULL, *h, *next;
13145 
13146 		for (h = help->dthps_actions[i]; h != NULL; h = next) {
13147 			next = h->dtha_next;
13148 
13149 			if (h->dtha_generation == gen) {
13150 				if (last != NULL) {
13151 					last->dtha_next = next;
13152 				} else {
13153 					help->dthps_actions[i] = next;
13154 				}
13155 
13156 				dtrace_helper_action_destroy(h, vstate);
13157 			} else {
13158 				last = h;
13159 			}
13160 		}
13161 	}
13162 
13163 	/*
13164 	 * Interate until we've cleared out all helper providers with the
13165 	 * given generation number.
13166 	 */
13167 	for (;;) {
13168 		dtrace_helper_provider_t *prov;
13169 
13170 		/*
13171 		 * Look for a helper provider with the right generation. We
13172 		 * have to start back at the beginning of the list each time
13173 		 * because we drop dtrace_lock. It's unlikely that we'll make
13174 		 * more than two passes.
13175 		 */
13176 		for (i = 0; i < help->dthps_nprovs; i++) {
13177 			prov = help->dthps_provs[i];
13178 
13179 			if (prov->dthp_generation == gen)
13180 				break;
13181 		}
13182 
13183 		/*
13184 		 * If there were no matches, we're done.
13185 		 */
13186 		if (i == help->dthps_nprovs)
13187 			break;
13188 
13189 		/*
13190 		 * Move the last helper provider into this slot.
13191 		 */
13192 		help->dthps_nprovs--;
13193 		help->dthps_provs[i] = help->dthps_provs[help->dthps_nprovs];
13194 		help->dthps_provs[help->dthps_nprovs] = NULL;
13195 
13196 		mutex_exit(&dtrace_lock);
13197 
13198 		/*
13199 		 * If we have a meta provider, remove this helper provider.
13200 		 */
13201 		mutex_enter(&dtrace_meta_lock);
13202 		if (dtrace_meta_pid != NULL) {
13203 			ASSERT(dtrace_deferred_pid == NULL);
13204 			dtrace_helper_provider_remove(&prov->dthp_prov,
13205 			    p->p_pid);
13206 		}
13207 		mutex_exit(&dtrace_meta_lock);
13208 
13209 		dtrace_helper_provider_destroy(prov);
13210 
13211 		mutex_enter(&dtrace_lock);
13212 	}
13213 
13214 	return (0);
13215 }
13216 
13217 static int
13218 dtrace_helper_validate(dtrace_helper_action_t *helper)
13219 {
13220 	int err = 0, i;
13221 	dtrace_difo_t *dp;
13222 
13223 	if ((dp = helper->dtha_predicate) != NULL)
13224 		err += dtrace_difo_validate_helper(dp);
13225 
13226 	for (i = 0; i < helper->dtha_nactions; i++)
13227 		err += dtrace_difo_validate_helper(helper->dtha_actions[i]);
13228 
13229 	return (err == 0);
13230 }
13231 
13232 static int
13233 dtrace_helper_action_add(int which, dtrace_ecbdesc_t *ep)
13234 {
13235 	dtrace_helpers_t *help;
13236 	dtrace_helper_action_t *helper, *last;
13237 	dtrace_actdesc_t *act;
13238 	dtrace_vstate_t *vstate;
13239 	dtrace_predicate_t *pred;
13240 	int count = 0, nactions = 0, i;
13241 
13242 	if (which < 0 || which >= DTRACE_NHELPER_ACTIONS)
13243 		return (EINVAL);
13244 
13245 	help = curproc->p_dtrace_helpers;
13246 	last = help->dthps_actions[which];
13247 	vstate = &help->dthps_vstate;
13248 
13249 	for (count = 0; last != NULL; last = last->dtha_next) {
13250 		count++;
13251 		if (last->dtha_next == NULL)
13252 			break;
13253 	}
13254 
13255 	/*
13256 	 * If we already have dtrace_helper_actions_max helper actions for this
13257 	 * helper action type, we'll refuse to add a new one.
13258 	 */
13259 	if (count >= dtrace_helper_actions_max)
13260 		return (ENOSPC);
13261 
13262 	helper = kmem_zalloc(sizeof (dtrace_helper_action_t), KM_SLEEP);
13263 	helper->dtha_generation = help->dthps_generation;
13264 
13265 	if ((pred = ep->dted_pred.dtpdd_predicate) != NULL) {
13266 		ASSERT(pred->dtp_difo != NULL);
13267 		dtrace_difo_hold(pred->dtp_difo);
13268 		helper->dtha_predicate = pred->dtp_difo;
13269 	}
13270 
13271 	for (act = ep->dted_action; act != NULL; act = act->dtad_next) {
13272 		if (act->dtad_kind != DTRACEACT_DIFEXPR)
13273 			goto err;
13274 
13275 		if (act->dtad_difo == NULL)
13276 			goto err;
13277 
13278 		nactions++;
13279 	}
13280 
13281 	helper->dtha_actions = kmem_zalloc(sizeof (dtrace_difo_t *) *
13282 	    (helper->dtha_nactions = nactions), KM_SLEEP);
13283 
13284 	for (act = ep->dted_action, i = 0; act != NULL; act = act->dtad_next) {
13285 		dtrace_difo_hold(act->dtad_difo);
13286 		helper->dtha_actions[i++] = act->dtad_difo;
13287 	}
13288 
13289 	if (!dtrace_helper_validate(helper))
13290 		goto err;
13291 
13292 	if (last == NULL) {
13293 		help->dthps_actions[which] = helper;
13294 	} else {
13295 		last->dtha_next = helper;
13296 	}
13297 
13298 	if (vstate->dtvs_nlocals > dtrace_helptrace_nlocals) {
13299 		dtrace_helptrace_nlocals = vstate->dtvs_nlocals;
13300 		dtrace_helptrace_next = 0;
13301 	}
13302 
13303 	return (0);
13304 err:
13305 	dtrace_helper_action_destroy(helper, vstate);
13306 	return (EINVAL);
13307 }
13308 
13309 static void
13310 dtrace_helper_provider_register(proc_t *p, dtrace_helpers_t *help,
13311     dof_helper_t *dofhp)
13312 {
13313 	ASSERT(MUTEX_NOT_HELD(&dtrace_lock));
13314 
13315 	mutex_enter(&dtrace_meta_lock);
13316 	mutex_enter(&dtrace_lock);
13317 
13318 	if (!dtrace_attached() || dtrace_meta_pid == NULL) {
13319 		/*
13320 		 * If the dtrace module is loaded but not attached, or if
13321 		 * there aren't isn't a meta provider registered to deal with
13322 		 * these provider descriptions, we need to postpone creating
13323 		 * the actual providers until later.
13324 		 */
13325 
13326 		if (help->dthps_next == NULL && help->dthps_prev == NULL &&
13327 		    dtrace_deferred_pid != help) {
13328 			help->dthps_deferred = 1;
13329 			help->dthps_pid = p->p_pid;
13330 			help->dthps_next = dtrace_deferred_pid;
13331 			help->dthps_prev = NULL;
13332 			if (dtrace_deferred_pid != NULL)
13333 				dtrace_deferred_pid->dthps_prev = help;
13334 			dtrace_deferred_pid = help;
13335 		}
13336 
13337 		mutex_exit(&dtrace_lock);
13338 
13339 	} else if (dofhp != NULL) {
13340 		/*
13341 		 * If the dtrace module is loaded and we have a particular
13342 		 * helper provider description, pass that off to the
13343 		 * meta provider.
13344 		 */
13345 
13346 		mutex_exit(&dtrace_lock);
13347 
13348 		dtrace_helper_provide(dofhp, p->p_pid);
13349 
13350 	} else {
13351 		/*
13352 		 * Otherwise, just pass all the helper provider descriptions
13353 		 * off to the meta provider.
13354 		 */
13355 
13356 		int i;
13357 		mutex_exit(&dtrace_lock);
13358 
13359 		for (i = 0; i < help->dthps_nprovs; i++) {
13360 			dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
13361 			    p->p_pid);
13362 		}
13363 	}
13364 
13365 	mutex_exit(&dtrace_meta_lock);
13366 }
13367 
13368 static int
13369 dtrace_helper_provider_add(dof_helper_t *dofhp, int gen)
13370 {
13371 	dtrace_helpers_t *help;
13372 	dtrace_helper_provider_t *hprov, **tmp_provs;
13373 	uint_t tmp_maxprovs, i;
13374 
13375 	ASSERT(MUTEX_HELD(&dtrace_lock));
13376 
13377 	help = curproc->p_dtrace_helpers;
13378 	ASSERT(help != NULL);
13379 
13380 	/*
13381 	 * If we already have dtrace_helper_providers_max helper providers,
13382 	 * we're refuse to add a new one.
13383 	 */
13384 	if (help->dthps_nprovs >= dtrace_helper_providers_max)
13385 		return (ENOSPC);
13386 
13387 	/*
13388 	 * Check to make sure this isn't a duplicate.
13389 	 */
13390 	for (i = 0; i < help->dthps_nprovs; i++) {
13391 		if (dofhp->dofhp_addr ==
13392 		    help->dthps_provs[i]->dthp_prov.dofhp_addr)
13393 			return (EALREADY);
13394 	}
13395 
13396 	hprov = kmem_zalloc(sizeof (dtrace_helper_provider_t), KM_SLEEP);
13397 	hprov->dthp_prov = *dofhp;
13398 	hprov->dthp_ref = 1;
13399 	hprov->dthp_generation = gen;
13400 
13401 	/*
13402 	 * Allocate a bigger table for helper providers if it's already full.
13403 	 */
13404 	if (help->dthps_maxprovs == help->dthps_nprovs) {
13405 		tmp_maxprovs = help->dthps_maxprovs;
13406 		tmp_provs = help->dthps_provs;
13407 
13408 		if (help->dthps_maxprovs == 0)
13409 			help->dthps_maxprovs = 2;
13410 		else
13411 			help->dthps_maxprovs *= 2;
13412 		if (help->dthps_maxprovs > dtrace_helper_providers_max)
13413 			help->dthps_maxprovs = dtrace_helper_providers_max;
13414 
13415 		ASSERT(tmp_maxprovs < help->dthps_maxprovs);
13416 
13417 		help->dthps_provs = kmem_zalloc(help->dthps_maxprovs *
13418 		    sizeof (dtrace_helper_provider_t *), KM_SLEEP);
13419 
13420 		if (tmp_provs != NULL) {
13421 			bcopy(tmp_provs, help->dthps_provs, tmp_maxprovs *
13422 			    sizeof (dtrace_helper_provider_t *));
13423 			kmem_free(tmp_provs, tmp_maxprovs *
13424 			    sizeof (dtrace_helper_provider_t *));
13425 		}
13426 	}
13427 
13428 	help->dthps_provs[help->dthps_nprovs] = hprov;
13429 	help->dthps_nprovs++;
13430 
13431 	return (0);
13432 }
13433 
13434 static void
13435 dtrace_helper_provider_destroy(dtrace_helper_provider_t *hprov)
13436 {
13437 	mutex_enter(&dtrace_lock);
13438 
13439 	if (--hprov->dthp_ref == 0) {
13440 		dof_hdr_t *dof;
13441 		mutex_exit(&dtrace_lock);
13442 		dof = (dof_hdr_t *)(uintptr_t)hprov->dthp_prov.dofhp_dof;
13443 		dtrace_dof_destroy(dof);
13444 		kmem_free(hprov, sizeof (dtrace_helper_provider_t));
13445 	} else {
13446 		mutex_exit(&dtrace_lock);
13447 	}
13448 }
13449 
13450 static int
13451 dtrace_helper_provider_validate(dof_hdr_t *dof, dof_sec_t *sec)
13452 {
13453 	uintptr_t daddr = (uintptr_t)dof;
13454 	dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
13455 	dof_provider_t *provider;
13456 	dof_probe_t *probe;
13457 	uint8_t *arg;
13458 	char *strtab, *typestr;
13459 	dof_stridx_t typeidx;
13460 	size_t typesz;
13461 	uint_t nprobes, j, k;
13462 
13463 	ASSERT(sec->dofs_type == DOF_SECT_PROVIDER);
13464 
13465 	if (sec->dofs_offset & (sizeof (uint_t) - 1)) {
13466 		dtrace_dof_error(dof, "misaligned section offset");
13467 		return (-1);
13468 	}
13469 
13470 	/*
13471 	 * The section needs to be large enough to contain the DOF provider
13472 	 * structure appropriate for the given version.
13473 	 */
13474 	if (sec->dofs_size <
13475 	    ((dof->dofh_ident[DOF_ID_VERSION] == DOF_VERSION_1) ?
13476 	    offsetof(dof_provider_t, dofpv_prenoffs) :
13477 	    sizeof (dof_provider_t))) {
13478 		dtrace_dof_error(dof, "provider section too small");
13479 		return (-1);
13480 	}
13481 
13482 	provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
13483 	str_sec = dtrace_dof_sect(dof, DOF_SECT_STRTAB, provider->dofpv_strtab);
13484 	prb_sec = dtrace_dof_sect(dof, DOF_SECT_PROBES, provider->dofpv_probes);
13485 	arg_sec = dtrace_dof_sect(dof, DOF_SECT_PRARGS, provider->dofpv_prargs);
13486 	off_sec = dtrace_dof_sect(dof, DOF_SECT_PROFFS, provider->dofpv_proffs);
13487 
13488 	if (str_sec == NULL || prb_sec == NULL ||
13489 	    arg_sec == NULL || off_sec == NULL)
13490 		return (-1);
13491 
13492 	enoff_sec = NULL;
13493 
13494 	if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
13495 	    provider->dofpv_prenoffs != DOF_SECT_NONE &&
13496 	    (enoff_sec = dtrace_dof_sect(dof, DOF_SECT_PRENOFFS,
13497 	    provider->dofpv_prenoffs)) == NULL)
13498 		return (-1);
13499 
13500 	strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
13501 
13502 	if (provider->dofpv_name >= str_sec->dofs_size ||
13503 	    strlen(strtab + provider->dofpv_name) >= DTRACE_PROVNAMELEN) {
13504 		dtrace_dof_error(dof, "invalid provider name");
13505 		return (-1);
13506 	}
13507 
13508 	if (prb_sec->dofs_entsize == 0 ||
13509 	    prb_sec->dofs_entsize > prb_sec->dofs_size) {
13510 		dtrace_dof_error(dof, "invalid entry size");
13511 		return (-1);
13512 	}
13513 
13514 	if (prb_sec->dofs_entsize & (sizeof (uintptr_t) - 1)) {
13515 		dtrace_dof_error(dof, "misaligned entry size");
13516 		return (-1);
13517 	}
13518 
13519 	if (off_sec->dofs_entsize != sizeof (uint32_t)) {
13520 		dtrace_dof_error(dof, "invalid entry size");
13521 		return (-1);
13522 	}
13523 
13524 	if (off_sec->dofs_offset & (sizeof (uint32_t) - 1)) {
13525 		dtrace_dof_error(dof, "misaligned section offset");
13526 		return (-1);
13527 	}
13528 
13529 	if (arg_sec->dofs_entsize != sizeof (uint8_t)) {
13530 		dtrace_dof_error(dof, "invalid entry size");
13531 		return (-1);
13532 	}
13533 
13534 	arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
13535 
13536 	nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
13537 
13538 	/*
13539 	 * Take a pass through the probes to check for errors.
13540 	 */
13541 	for (j = 0; j < nprobes; j++) {
13542 		probe = (dof_probe_t *)(uintptr_t)(daddr +
13543 		    prb_sec->dofs_offset + j * prb_sec->dofs_entsize);
13544 
13545 		if (probe->dofpr_func >= str_sec->dofs_size) {
13546 			dtrace_dof_error(dof, "invalid function name");
13547 			return (-1);
13548 		}
13549 
13550 		if (strlen(strtab + probe->dofpr_func) >= DTRACE_FUNCNAMELEN) {
13551 			dtrace_dof_error(dof, "function name too long");
13552 			return (-1);
13553 		}
13554 
13555 		if (probe->dofpr_name >= str_sec->dofs_size ||
13556 		    strlen(strtab + probe->dofpr_name) >= DTRACE_NAMELEN) {
13557 			dtrace_dof_error(dof, "invalid probe name");
13558 			return (-1);
13559 		}
13560 
13561 		/*
13562 		 * The offset count must not wrap the index, and the offsets
13563 		 * must also not overflow the section's data.
13564 		 */
13565 		if (probe->dofpr_offidx + probe->dofpr_noffs <
13566 		    probe->dofpr_offidx ||
13567 		    (probe->dofpr_offidx + probe->dofpr_noffs) *
13568 		    off_sec->dofs_entsize > off_sec->dofs_size) {
13569 			dtrace_dof_error(dof, "invalid probe offset");
13570 			return (-1);
13571 		}
13572 
13573 		if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1) {
13574 			/*
13575 			 * If there's no is-enabled offset section, make sure
13576 			 * there aren't any is-enabled offsets. Otherwise
13577 			 * perform the same checks as for probe offsets
13578 			 * (immediately above).
13579 			 */
13580 			if (enoff_sec == NULL) {
13581 				if (probe->dofpr_enoffidx != 0 ||
13582 				    probe->dofpr_nenoffs != 0) {
13583 					dtrace_dof_error(dof, "is-enabled "
13584 					    "offsets with null section");
13585 					return (-1);
13586 				}
13587 			} else if (probe->dofpr_enoffidx +
13588 			    probe->dofpr_nenoffs < probe->dofpr_enoffidx ||
13589 			    (probe->dofpr_enoffidx + probe->dofpr_nenoffs) *
13590 			    enoff_sec->dofs_entsize > enoff_sec->dofs_size) {
13591 				dtrace_dof_error(dof, "invalid is-enabled "
13592 				    "offset");
13593 				return (-1);
13594 			}
13595 
13596 			if (probe->dofpr_noffs + probe->dofpr_nenoffs == 0) {
13597 				dtrace_dof_error(dof, "zero probe and "
13598 				    "is-enabled offsets");
13599 				return (-1);
13600 			}
13601 		} else if (probe->dofpr_noffs == 0) {
13602 			dtrace_dof_error(dof, "zero probe offsets");
13603 			return (-1);
13604 		}
13605 
13606 		if (probe->dofpr_argidx + probe->dofpr_xargc <
13607 		    probe->dofpr_argidx ||
13608 		    (probe->dofpr_argidx + probe->dofpr_xargc) *
13609 		    arg_sec->dofs_entsize > arg_sec->dofs_size) {
13610 			dtrace_dof_error(dof, "invalid args");
13611 			return (-1);
13612 		}
13613 
13614 		typeidx = probe->dofpr_nargv;
13615 		typestr = strtab + probe->dofpr_nargv;
13616 		for (k = 0; k < probe->dofpr_nargc; k++) {
13617 			if (typeidx >= str_sec->dofs_size) {
13618 				dtrace_dof_error(dof, "bad "
13619 				    "native argument type");
13620 				return (-1);
13621 			}
13622 
13623 			typesz = strlen(typestr) + 1;
13624 			if (typesz > DTRACE_ARGTYPELEN) {
13625 				dtrace_dof_error(dof, "native "
13626 				    "argument type too long");
13627 				return (-1);
13628 			}
13629 			typeidx += typesz;
13630 			typestr += typesz;
13631 		}
13632 
13633 		typeidx = probe->dofpr_xargv;
13634 		typestr = strtab + probe->dofpr_xargv;
13635 		for (k = 0; k < probe->dofpr_xargc; k++) {
13636 			if (arg[probe->dofpr_argidx + k] > probe->dofpr_nargc) {
13637 				dtrace_dof_error(dof, "bad "
13638 				    "native argument index");
13639 				return (-1);
13640 			}
13641 
13642 			if (typeidx >= str_sec->dofs_size) {
13643 				dtrace_dof_error(dof, "bad "
13644 				    "translated argument type");
13645 				return (-1);
13646 			}
13647 
13648 			typesz = strlen(typestr) + 1;
13649 			if (typesz > DTRACE_ARGTYPELEN) {
13650 				dtrace_dof_error(dof, "translated argument "
13651 				    "type too long");
13652 				return (-1);
13653 			}
13654 
13655 			typeidx += typesz;
13656 			typestr += typesz;
13657 		}
13658 	}
13659 
13660 	return (0);
13661 }
13662 
13663 static int
13664 dtrace_helper_slurp(dof_hdr_t *dof, dof_helper_t *dhp)
13665 {
13666 	dtrace_helpers_t *help;
13667 	dtrace_vstate_t *vstate;
13668 	dtrace_enabling_t *enab = NULL;
13669 	int i, gen, rv, nhelpers = 0, nprovs = 0, destroy = 1;
13670 	uintptr_t daddr = (uintptr_t)dof;
13671 
13672 	ASSERT(MUTEX_HELD(&dtrace_lock));
13673 
13674 	if ((help = curproc->p_dtrace_helpers) == NULL)
13675 		help = dtrace_helpers_create(curproc);
13676 
13677 	vstate = &help->dthps_vstate;
13678 
13679 	if ((rv = dtrace_dof_slurp(dof, vstate, NULL, &enab,
13680 	    dhp != NULL ? dhp->dofhp_addr : 0, B_FALSE)) != 0) {
13681 		dtrace_dof_destroy(dof);
13682 		return (rv);
13683 	}
13684 
13685 	/*
13686 	 * Look for helper providers and validate their descriptions.
13687 	 */
13688 	if (dhp != NULL) {
13689 		for (i = 0; i < dof->dofh_secnum; i++) {
13690 			dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
13691 			    dof->dofh_secoff + i * dof->dofh_secsize);
13692 
13693 			if (sec->dofs_type != DOF_SECT_PROVIDER)
13694 				continue;
13695 
13696 			if (dtrace_helper_provider_validate(dof, sec) != 0) {
13697 				dtrace_enabling_destroy(enab);
13698 				dtrace_dof_destroy(dof);
13699 				return (-1);
13700 			}
13701 
13702 			nprovs++;
13703 		}
13704 	}
13705 
13706 	/*
13707 	 * Now we need to walk through the ECB descriptions in the enabling.
13708 	 */
13709 	for (i = 0; i < enab->dten_ndesc; i++) {
13710 		dtrace_ecbdesc_t *ep = enab->dten_desc[i];
13711 		dtrace_probedesc_t *desc = &ep->dted_probe;
13712 
13713 		if (strcmp(desc->dtpd_provider, "dtrace") != 0)
13714 			continue;
13715 
13716 		if (strcmp(desc->dtpd_mod, "helper") != 0)
13717 			continue;
13718 
13719 		if (strcmp(desc->dtpd_func, "ustack") != 0)
13720 			continue;
13721 
13722 		if ((rv = dtrace_helper_action_add(DTRACE_HELPER_ACTION_USTACK,
13723 		    ep)) != 0) {
13724 			/*
13725 			 * Adding this helper action failed -- we are now going
13726 			 * to rip out the entire generation and return failure.
13727 			 */
13728 			(void) dtrace_helper_destroygen(help->dthps_generation);
13729 			dtrace_enabling_destroy(enab);
13730 			dtrace_dof_destroy(dof);
13731 			return (-1);
13732 		}
13733 
13734 		nhelpers++;
13735 	}
13736 
13737 	if (nhelpers < enab->dten_ndesc)
13738 		dtrace_dof_error(dof, "unmatched helpers");
13739 
13740 	gen = help->dthps_generation++;
13741 	dtrace_enabling_destroy(enab);
13742 
13743 	if (dhp != NULL && nprovs > 0) {
13744 		dhp->dofhp_dof = (uint64_t)(uintptr_t)dof;
13745 		if (dtrace_helper_provider_add(dhp, gen) == 0) {
13746 			mutex_exit(&dtrace_lock);
13747 			dtrace_helper_provider_register(curproc, help, dhp);
13748 			mutex_enter(&dtrace_lock);
13749 
13750 			destroy = 0;
13751 		}
13752 	}
13753 
13754 	if (destroy)
13755 		dtrace_dof_destroy(dof);
13756 
13757 	return (gen);
13758 }
13759 
13760 static dtrace_helpers_t *
13761 dtrace_helpers_create(proc_t *p)
13762 {
13763 	dtrace_helpers_t *help;
13764 
13765 	ASSERT(MUTEX_HELD(&dtrace_lock));
13766 	ASSERT(p->p_dtrace_helpers == NULL);
13767 
13768 	help = kmem_zalloc(sizeof (dtrace_helpers_t), KM_SLEEP);
13769 	help->dthps_actions = kmem_zalloc(sizeof (dtrace_helper_action_t *) *
13770 	    DTRACE_NHELPER_ACTIONS, KM_SLEEP);
13771 
13772 	p->p_dtrace_helpers = help;
13773 	dtrace_helpers++;
13774 
13775 	return (help);
13776 }
13777 
13778 static void
13779 dtrace_helpers_destroy(void)
13780 {
13781 	dtrace_helpers_t *help;
13782 	dtrace_vstate_t *vstate;
13783 	proc_t *p = curproc;
13784 	int i;
13785 
13786 	mutex_enter(&dtrace_lock);
13787 
13788 	ASSERT(p->p_dtrace_helpers != NULL);
13789 	ASSERT(dtrace_helpers > 0);
13790 
13791 	help = p->p_dtrace_helpers;
13792 	vstate = &help->dthps_vstate;
13793 
13794 	/*
13795 	 * We're now going to lose the help from this process.
13796 	 */
13797 	p->p_dtrace_helpers = NULL;
13798 	dtrace_sync();
13799 
13800 	/*
13801 	 * Destory the helper actions.
13802 	 */
13803 	for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
13804 		dtrace_helper_action_t *h, *next;
13805 
13806 		for (h = help->dthps_actions[i]; h != NULL; h = next) {
13807 			next = h->dtha_next;
13808 			dtrace_helper_action_destroy(h, vstate);
13809 			h = next;
13810 		}
13811 	}
13812 
13813 	mutex_exit(&dtrace_lock);
13814 
13815 	/*
13816 	 * Destroy the helper providers.
13817 	 */
13818 	if (help->dthps_maxprovs > 0) {
13819 		mutex_enter(&dtrace_meta_lock);
13820 		if (dtrace_meta_pid != NULL) {
13821 			ASSERT(dtrace_deferred_pid == NULL);
13822 
13823 			for (i = 0; i < help->dthps_nprovs; i++) {
13824 				dtrace_helper_provider_remove(
13825 				    &help->dthps_provs[i]->dthp_prov, p->p_pid);
13826 			}
13827 		} else {
13828 			mutex_enter(&dtrace_lock);
13829 			ASSERT(help->dthps_deferred == 0 ||
13830 			    help->dthps_next != NULL ||
13831 			    help->dthps_prev != NULL ||
13832 			    help == dtrace_deferred_pid);
13833 
13834 			/*
13835 			 * Remove the helper from the deferred list.
13836 			 */
13837 			if (help->dthps_next != NULL)
13838 				help->dthps_next->dthps_prev = help->dthps_prev;
13839 			if (help->dthps_prev != NULL)
13840 				help->dthps_prev->dthps_next = help->dthps_next;
13841 			if (dtrace_deferred_pid == help) {
13842 				dtrace_deferred_pid = help->dthps_next;
13843 				ASSERT(help->dthps_prev == NULL);
13844 			}
13845 
13846 			mutex_exit(&dtrace_lock);
13847 		}
13848 
13849 		mutex_exit(&dtrace_meta_lock);
13850 
13851 		for (i = 0; i < help->dthps_nprovs; i++) {
13852 			dtrace_helper_provider_destroy(help->dthps_provs[i]);
13853 		}
13854 
13855 		kmem_free(help->dthps_provs, help->dthps_maxprovs *
13856 		    sizeof (dtrace_helper_provider_t *));
13857 	}
13858 
13859 	mutex_enter(&dtrace_lock);
13860 
13861 	dtrace_vstate_fini(&help->dthps_vstate);
13862 	kmem_free(help->dthps_actions,
13863 	    sizeof (dtrace_helper_action_t *) * DTRACE_NHELPER_ACTIONS);
13864 	kmem_free(help, sizeof (dtrace_helpers_t));
13865 
13866 	--dtrace_helpers;
13867 	mutex_exit(&dtrace_lock);
13868 }
13869 
13870 static void
13871 dtrace_helpers_duplicate(proc_t *from, proc_t *to)
13872 {
13873 	dtrace_helpers_t *help, *newhelp;
13874 	dtrace_helper_action_t *helper, *new, *last;
13875 	dtrace_difo_t *dp;
13876 	dtrace_vstate_t *vstate;
13877 	int i, j, sz, hasprovs = 0;
13878 
13879 	mutex_enter(&dtrace_lock);
13880 	ASSERT(from->p_dtrace_helpers != NULL);
13881 	ASSERT(dtrace_helpers > 0);
13882 
13883 	help = from->p_dtrace_helpers;
13884 	newhelp = dtrace_helpers_create(to);
13885 	ASSERT(to->p_dtrace_helpers != NULL);
13886 
13887 	newhelp->dthps_generation = help->dthps_generation;
13888 	vstate = &newhelp->dthps_vstate;
13889 
13890 	/*
13891 	 * Duplicate the helper actions.
13892 	 */
13893 	for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
13894 		if ((helper = help->dthps_actions[i]) == NULL)
13895 			continue;
13896 
13897 		for (last = NULL; helper != NULL; helper = helper->dtha_next) {
13898 			new = kmem_zalloc(sizeof (dtrace_helper_action_t),
13899 			    KM_SLEEP);
13900 			new->dtha_generation = helper->dtha_generation;
13901 
13902 			if ((dp = helper->dtha_predicate) != NULL) {
13903 				dp = dtrace_difo_duplicate(dp, vstate);
13904 				new->dtha_predicate = dp;
13905 			}
13906 
13907 			new->dtha_nactions = helper->dtha_nactions;
13908 			sz = sizeof (dtrace_difo_t *) * new->dtha_nactions;
13909 			new->dtha_actions = kmem_alloc(sz, KM_SLEEP);
13910 
13911 			for (j = 0; j < new->dtha_nactions; j++) {
13912 				dtrace_difo_t *dp = helper->dtha_actions[j];
13913 
13914 				ASSERT(dp != NULL);
13915 				dp = dtrace_difo_duplicate(dp, vstate);
13916 				new->dtha_actions[j] = dp;
13917 			}
13918 
13919 			if (last != NULL) {
13920 				last->dtha_next = new;
13921 			} else {
13922 				newhelp->dthps_actions[i] = new;
13923 			}
13924 
13925 			last = new;
13926 		}
13927 	}
13928 
13929 	/*
13930 	 * Duplicate the helper providers and register them with the
13931 	 * DTrace framework.
13932 	 */
13933 	if (help->dthps_nprovs > 0) {
13934 		newhelp->dthps_nprovs = help->dthps_nprovs;
13935 		newhelp->dthps_maxprovs = help->dthps_nprovs;
13936 		newhelp->dthps_provs = kmem_alloc(newhelp->dthps_nprovs *
13937 		    sizeof (dtrace_helper_provider_t *), KM_SLEEP);
13938 		for (i = 0; i < newhelp->dthps_nprovs; i++) {
13939 			newhelp->dthps_provs[i] = help->dthps_provs[i];
13940 			newhelp->dthps_provs[i]->dthp_ref++;
13941 		}
13942 
13943 		hasprovs = 1;
13944 	}
13945 
13946 	mutex_exit(&dtrace_lock);
13947 
13948 	if (hasprovs)
13949 		dtrace_helper_provider_register(to, newhelp, NULL);
13950 }
13951 
13952 /*
13953  * DTrace Hook Functions
13954  */
13955 static void
13956 dtrace_module_loaded(struct modctl *ctl)
13957 {
13958 	dtrace_provider_t *prv;
13959 
13960 	mutex_enter(&dtrace_provider_lock);
13961 	mutex_enter(&mod_lock);
13962 
13963 	ASSERT(ctl->mod_busy);
13964 
13965 	/*
13966 	 * We're going to call each providers per-module provide operation
13967 	 * specifying only this module.
13968 	 */
13969 	for (prv = dtrace_provider; prv != NULL; prv = prv->dtpv_next)
13970 		prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
13971 
13972 	mutex_exit(&mod_lock);
13973 	mutex_exit(&dtrace_provider_lock);
13974 
13975 	/*
13976 	 * If we have any retained enablings, we need to match against them.
13977 	 * Enabling probes requires that cpu_lock be held, and we cannot hold
13978 	 * cpu_lock here -- it is legal for cpu_lock to be held when loading a
13979 	 * module.  (In particular, this happens when loading scheduling
13980 	 * classes.)  So if we have any retained enablings, we need to dispatch
13981 	 * our task queue to do the match for us.
13982 	 */
13983 	mutex_enter(&dtrace_lock);
13984 
13985 	if (dtrace_retained == NULL) {
13986 		mutex_exit(&dtrace_lock);
13987 		return;
13988 	}
13989 
13990 	(void) taskq_dispatch(dtrace_taskq,
13991 	    (task_func_t *)dtrace_enabling_matchall, NULL, TQ_SLEEP);
13992 
13993 	mutex_exit(&dtrace_lock);
13994 
13995 	/*
13996 	 * And now, for a little heuristic sleaze:  in general, we want to
13997 	 * match modules as soon as they load.  However, we cannot guarantee
13998 	 * this, because it would lead us to the lock ordering violation
13999 	 * outlined above.  The common case, of course, is that cpu_lock is
14000 	 * _not_ held -- so we delay here for a clock tick, hoping that that's
14001 	 * long enough for the task queue to do its work.  If it's not, it's
14002 	 * not a serious problem -- it just means that the module that we
14003 	 * just loaded may not be immediately instrumentable.
14004 	 */
14005 	delay(1);
14006 }
14007 
14008 static void
14009 dtrace_module_unloaded(struct modctl *ctl)
14010 {
14011 	dtrace_probe_t template, *probe, *first, *next;
14012 	dtrace_provider_t *prov;
14013 
14014 	template.dtpr_mod = ctl->mod_modname;
14015 
14016 	mutex_enter(&dtrace_provider_lock);
14017 	mutex_enter(&mod_lock);
14018 	mutex_enter(&dtrace_lock);
14019 
14020 	if (dtrace_bymod == NULL) {
14021 		/*
14022 		 * The DTrace module is loaded (obviously) but not attached;
14023 		 * we don't have any work to do.
14024 		 */
14025 		mutex_exit(&dtrace_provider_lock);
14026 		mutex_exit(&mod_lock);
14027 		mutex_exit(&dtrace_lock);
14028 		return;
14029 	}
14030 
14031 	for (probe = first = dtrace_hash_lookup(dtrace_bymod, &template);
14032 	    probe != NULL; probe = probe->dtpr_nextmod) {
14033 		if (probe->dtpr_ecb != NULL) {
14034 			mutex_exit(&dtrace_provider_lock);
14035 			mutex_exit(&mod_lock);
14036 			mutex_exit(&dtrace_lock);
14037 
14038 			/*
14039 			 * This shouldn't _actually_ be possible -- we're
14040 			 * unloading a module that has an enabled probe in it.
14041 			 * (It's normally up to the provider to make sure that
14042 			 * this can't happen.)  However, because dtps_enable()
14043 			 * doesn't have a failure mode, there can be an
14044 			 * enable/unload race.  Upshot:  we don't want to
14045 			 * assert, but we're not going to disable the
14046 			 * probe, either.
14047 			 */
14048 			if (dtrace_err_verbose) {
14049 				cmn_err(CE_WARN, "unloaded module '%s' had "
14050 				    "enabled probes", ctl->mod_modname);
14051 			}
14052 
14053 			return;
14054 		}
14055 	}
14056 
14057 	probe = first;
14058 
14059 	for (first = NULL; probe != NULL; probe = next) {
14060 		ASSERT(dtrace_probes[probe->dtpr_id - 1] == probe);
14061 
14062 		dtrace_probes[probe->dtpr_id - 1] = NULL;
14063 
14064 		next = probe->dtpr_nextmod;
14065 		dtrace_hash_remove(dtrace_bymod, probe);
14066 		dtrace_hash_remove(dtrace_byfunc, probe);
14067 		dtrace_hash_remove(dtrace_byname, probe);
14068 
14069 		if (first == NULL) {
14070 			first = probe;
14071 			probe->dtpr_nextmod = NULL;
14072 		} else {
14073 			probe->dtpr_nextmod = first;
14074 			first = probe;
14075 		}
14076 	}
14077 
14078 	/*
14079 	 * We've removed all of the module's probes from the hash chains and
14080 	 * from the probe array.  Now issue a dtrace_sync() to be sure that
14081 	 * everyone has cleared out from any probe array processing.
14082 	 */
14083 	dtrace_sync();
14084 
14085 	for (probe = first; probe != NULL; probe = first) {
14086 		first = probe->dtpr_nextmod;
14087 		prov = probe->dtpr_provider;
14088 		prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, probe->dtpr_id,
14089 		    probe->dtpr_arg);
14090 		kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
14091 		kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
14092 		kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
14093 		vmem_free(dtrace_arena, (void *)(uintptr_t)probe->dtpr_id, 1);
14094 		kmem_free(probe, sizeof (dtrace_probe_t));
14095 	}
14096 
14097 	mutex_exit(&dtrace_lock);
14098 	mutex_exit(&mod_lock);
14099 	mutex_exit(&dtrace_provider_lock);
14100 }
14101 
14102 void
14103 dtrace_suspend(void)
14104 {
14105 	dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_suspend));
14106 }
14107 
14108 void
14109 dtrace_resume(void)
14110 {
14111 	dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_resume));
14112 }
14113 
14114 static int
14115 dtrace_cpu_setup(cpu_setup_t what, processorid_t cpu)
14116 {
14117 	ASSERT(MUTEX_HELD(&cpu_lock));
14118 	mutex_enter(&dtrace_lock);
14119 
14120 	switch (what) {
14121 	case CPU_CONFIG: {
14122 		dtrace_state_t *state;
14123 		dtrace_optval_t *opt, rs, c;
14124 
14125 		/*
14126 		 * For now, we only allocate a new buffer for anonymous state.
14127 		 */
14128 		if ((state = dtrace_anon.dta_state) == NULL)
14129 			break;
14130 
14131 		if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
14132 			break;
14133 
14134 		opt = state->dts_options;
14135 		c = opt[DTRACEOPT_CPU];
14136 
14137 		if (c != DTRACE_CPUALL && c != DTRACEOPT_UNSET && c != cpu)
14138 			break;
14139 
14140 		/*
14141 		 * Regardless of what the actual policy is, we're going to
14142 		 * temporarily set our resize policy to be manual.  We're
14143 		 * also going to temporarily set our CPU option to denote
14144 		 * the newly configured CPU.
14145 		 */
14146 		rs = opt[DTRACEOPT_BUFRESIZE];
14147 		opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_MANUAL;
14148 		opt[DTRACEOPT_CPU] = (dtrace_optval_t)cpu;
14149 
14150 		(void) dtrace_state_buffers(state);
14151 
14152 		opt[DTRACEOPT_BUFRESIZE] = rs;
14153 		opt[DTRACEOPT_CPU] = c;
14154 
14155 		break;
14156 	}
14157 
14158 	case CPU_UNCONFIG:
14159 		/*
14160 		 * We don't free the buffer in the CPU_UNCONFIG case.  (The
14161 		 * buffer will be freed when the consumer exits.)
14162 		 */
14163 		break;
14164 
14165 	default:
14166 		break;
14167 	}
14168 
14169 	mutex_exit(&dtrace_lock);
14170 	return (0);
14171 }
14172 
14173 static void
14174 dtrace_cpu_setup_initial(processorid_t cpu)
14175 {
14176 	(void) dtrace_cpu_setup(CPU_CONFIG, cpu);
14177 }
14178 
14179 static void
14180 dtrace_toxrange_add(uintptr_t base, uintptr_t limit)
14181 {
14182 	if (dtrace_toxranges >= dtrace_toxranges_max) {
14183 		int osize, nsize;
14184 		dtrace_toxrange_t *range;
14185 
14186 		osize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
14187 
14188 		if (osize == 0) {
14189 			ASSERT(dtrace_toxrange == NULL);
14190 			ASSERT(dtrace_toxranges_max == 0);
14191 			dtrace_toxranges_max = 1;
14192 		} else {
14193 			dtrace_toxranges_max <<= 1;
14194 		}
14195 
14196 		nsize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
14197 		range = kmem_zalloc(nsize, KM_SLEEP);
14198 
14199 		if (dtrace_toxrange != NULL) {
14200 			ASSERT(osize != 0);
14201 			bcopy(dtrace_toxrange, range, osize);
14202 			kmem_free(dtrace_toxrange, osize);
14203 		}
14204 
14205 		dtrace_toxrange = range;
14206 	}
14207 
14208 	ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_base == NULL);
14209 	ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_limit == NULL);
14210 
14211 	dtrace_toxrange[dtrace_toxranges].dtt_base = base;
14212 	dtrace_toxrange[dtrace_toxranges].dtt_limit = limit;
14213 	dtrace_toxranges++;
14214 }
14215 
14216 /*
14217  * DTrace Driver Cookbook Functions
14218  */
14219 /*ARGSUSED*/
14220 static int
14221 dtrace_attach(dev_info_t *devi, ddi_attach_cmd_t cmd)
14222 {
14223 	dtrace_provider_id_t id;
14224 	dtrace_state_t *state = NULL;
14225 	dtrace_enabling_t *enab;
14226 
14227 	mutex_enter(&cpu_lock);
14228 	mutex_enter(&dtrace_provider_lock);
14229 	mutex_enter(&dtrace_lock);
14230 
14231 	if (ddi_soft_state_init(&dtrace_softstate,
14232 	    sizeof (dtrace_state_t), 0) != 0) {
14233 		cmn_err(CE_NOTE, "/dev/dtrace failed to initialize soft state");
14234 		mutex_exit(&cpu_lock);
14235 		mutex_exit(&dtrace_provider_lock);
14236 		mutex_exit(&dtrace_lock);
14237 		return (DDI_FAILURE);
14238 	}
14239 
14240 	if (ddi_create_minor_node(devi, DTRACEMNR_DTRACE, S_IFCHR,
14241 	    DTRACEMNRN_DTRACE, DDI_PSEUDO, NULL) == DDI_FAILURE ||
14242 	    ddi_create_minor_node(devi, DTRACEMNR_HELPER, S_IFCHR,
14243 	    DTRACEMNRN_HELPER, DDI_PSEUDO, NULL) == DDI_FAILURE) {
14244 		cmn_err(CE_NOTE, "/dev/dtrace couldn't create minor nodes");
14245 		ddi_remove_minor_node(devi, NULL);
14246 		ddi_soft_state_fini(&dtrace_softstate);
14247 		mutex_exit(&cpu_lock);
14248 		mutex_exit(&dtrace_provider_lock);
14249 		mutex_exit(&dtrace_lock);
14250 		return (DDI_FAILURE);
14251 	}
14252 
14253 	ddi_report_dev(devi);
14254 	dtrace_devi = devi;
14255 
14256 	dtrace_modload = dtrace_module_loaded;
14257 	dtrace_modunload = dtrace_module_unloaded;
14258 	dtrace_cpu_init = dtrace_cpu_setup_initial;
14259 	dtrace_helpers_cleanup = dtrace_helpers_destroy;
14260 	dtrace_helpers_fork = dtrace_helpers_duplicate;
14261 	dtrace_cpustart_init = dtrace_suspend;
14262 	dtrace_cpustart_fini = dtrace_resume;
14263 	dtrace_debugger_init = dtrace_suspend;
14264 	dtrace_debugger_fini = dtrace_resume;
14265 
14266 	register_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
14267 
14268 	ASSERT(MUTEX_HELD(&cpu_lock));
14269 
14270 	dtrace_arena = vmem_create("dtrace", (void *)1, UINT32_MAX, 1,
14271 	    NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
14272 	dtrace_minor = vmem_create("dtrace_minor", (void *)DTRACEMNRN_CLONE,
14273 	    UINT32_MAX - DTRACEMNRN_CLONE, 1, NULL, NULL, NULL, 0,
14274 	    VM_SLEEP | VMC_IDENTIFIER);
14275 	dtrace_taskq = taskq_create("dtrace_taskq", 1, maxclsyspri,
14276 	    1, INT_MAX, 0);
14277 
14278 	dtrace_state_cache = kmem_cache_create("dtrace_state_cache",
14279 	    sizeof (dtrace_dstate_percpu_t) * NCPU, DTRACE_STATE_ALIGN,
14280 	    NULL, NULL, NULL, NULL, NULL, 0);
14281 
14282 	ASSERT(MUTEX_HELD(&cpu_lock));
14283 	dtrace_bymod = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_mod),
14284 	    offsetof(dtrace_probe_t, dtpr_nextmod),
14285 	    offsetof(dtrace_probe_t, dtpr_prevmod));
14286 
14287 	dtrace_byfunc = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_func),
14288 	    offsetof(dtrace_probe_t, dtpr_nextfunc),
14289 	    offsetof(dtrace_probe_t, dtpr_prevfunc));
14290 
14291 	dtrace_byname = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_name),
14292 	    offsetof(dtrace_probe_t, dtpr_nextname),
14293 	    offsetof(dtrace_probe_t, dtpr_prevname));
14294 
14295 	if (dtrace_retain_max < 1) {
14296 		cmn_err(CE_WARN, "illegal value (%lu) for dtrace_retain_max; "
14297 		    "setting to 1", dtrace_retain_max);
14298 		dtrace_retain_max = 1;
14299 	}
14300 
14301 	/*
14302 	 * Now discover our toxic ranges.
14303 	 */
14304 	dtrace_toxic_ranges(dtrace_toxrange_add);
14305 
14306 	/*
14307 	 * Before we register ourselves as a provider to our own framework,
14308 	 * we would like to assert that dtrace_provider is NULL -- but that's
14309 	 * not true if we were loaded as a dependency of a DTrace provider.
14310 	 * Once we've registered, we can assert that dtrace_provider is our
14311 	 * pseudo provider.
14312 	 */
14313 	(void) dtrace_register("dtrace", &dtrace_provider_attr,
14314 	    DTRACE_PRIV_NONE, 0, &dtrace_provider_ops, NULL, &id);
14315 
14316 	ASSERT(dtrace_provider != NULL);
14317 	ASSERT((dtrace_provider_id_t)dtrace_provider == id);
14318 
14319 	dtrace_probeid_begin = dtrace_probe_create((dtrace_provider_id_t)
14320 	    dtrace_provider, NULL, NULL, "BEGIN", 0, NULL);
14321 	dtrace_probeid_end = dtrace_probe_create((dtrace_provider_id_t)
14322 	    dtrace_provider, NULL, NULL, "END", 0, NULL);
14323 	dtrace_probeid_error = dtrace_probe_create((dtrace_provider_id_t)
14324 	    dtrace_provider, NULL, NULL, "ERROR", 1, NULL);
14325 
14326 	dtrace_anon_property();
14327 	mutex_exit(&cpu_lock);
14328 
14329 	/*
14330 	 * If DTrace helper tracing is enabled, we need to allocate the
14331 	 * trace buffer and initialize the values.
14332 	 */
14333 	if (dtrace_helptrace_enabled) {
14334 		ASSERT(dtrace_helptrace_buffer == NULL);
14335 		dtrace_helptrace_buffer =
14336 		    kmem_zalloc(dtrace_helptrace_bufsize, KM_SLEEP);
14337 		dtrace_helptrace_next = 0;
14338 	}
14339 
14340 	/*
14341 	 * If there are already providers, we must ask them to provide their
14342 	 * probes, and then match any anonymous enabling against them.  Note
14343 	 * that there should be no other retained enablings at this time:
14344 	 * the only retained enablings at this time should be the anonymous
14345 	 * enabling.
14346 	 */
14347 	if (dtrace_anon.dta_enabling != NULL) {
14348 		ASSERT(dtrace_retained == dtrace_anon.dta_enabling);
14349 
14350 		dtrace_enabling_provide(NULL);
14351 		state = dtrace_anon.dta_state;
14352 
14353 		/*
14354 		 * We couldn't hold cpu_lock across the above call to
14355 		 * dtrace_enabling_provide(), but we must hold it to actually
14356 		 * enable the probes.  We have to drop all of our locks, pick
14357 		 * up cpu_lock, and regain our locks before matching the
14358 		 * retained anonymous enabling.
14359 		 */
14360 		mutex_exit(&dtrace_lock);
14361 		mutex_exit(&dtrace_provider_lock);
14362 
14363 		mutex_enter(&cpu_lock);
14364 		mutex_enter(&dtrace_provider_lock);
14365 		mutex_enter(&dtrace_lock);
14366 
14367 		if ((enab = dtrace_anon.dta_enabling) != NULL)
14368 			(void) dtrace_enabling_match(enab, NULL);
14369 
14370 		mutex_exit(&cpu_lock);
14371 	}
14372 
14373 	mutex_exit(&dtrace_lock);
14374 	mutex_exit(&dtrace_provider_lock);
14375 
14376 	if (state != NULL) {
14377 		/*
14378 		 * If we created any anonymous state, set it going now.
14379 		 */
14380 		(void) dtrace_state_go(state, &dtrace_anon.dta_beganon);
14381 	}
14382 
14383 	return (DDI_SUCCESS);
14384 }
14385 
14386 /*ARGSUSED*/
14387 static int
14388 dtrace_open(dev_t *devp, int flag, int otyp, cred_t *cred_p)
14389 {
14390 	dtrace_state_t *state;
14391 	uint32_t priv;
14392 	uid_t uid;
14393 	zoneid_t zoneid;
14394 
14395 	if (getminor(*devp) == DTRACEMNRN_HELPER)
14396 		return (0);
14397 
14398 	/*
14399 	 * If this wasn't an open with the "helper" minor, then it must be
14400 	 * the "dtrace" minor.
14401 	 */
14402 	if (getminor(*devp) != DTRACEMNRN_DTRACE)
14403 		return (ENXIO);
14404 
14405 	/*
14406 	 * If no DTRACE_PRIV_* bits are set in the credential, then the
14407 	 * caller lacks sufficient permission to do anything with DTrace.
14408 	 */
14409 	dtrace_cred2priv(cred_p, &priv, &uid, &zoneid);
14410 	if (priv == DTRACE_PRIV_NONE)
14411 		return (EACCES);
14412 
14413 	/*
14414 	 * Ask all providers to provide all their probes.
14415 	 */
14416 	mutex_enter(&dtrace_provider_lock);
14417 	dtrace_probe_provide(NULL, NULL);
14418 	mutex_exit(&dtrace_provider_lock);
14419 
14420 	mutex_enter(&cpu_lock);
14421 	mutex_enter(&dtrace_lock);
14422 	dtrace_opens++;
14423 	dtrace_membar_producer();
14424 
14425 	/*
14426 	 * If the kernel debugger is active (that is, if the kernel debugger
14427 	 * modified text in some way), we won't allow the open.
14428 	 */
14429 	if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
14430 		dtrace_opens--;
14431 		mutex_exit(&cpu_lock);
14432 		mutex_exit(&dtrace_lock);
14433 		return (EBUSY);
14434 	}
14435 
14436 	state = dtrace_state_create(devp, cred_p);
14437 	mutex_exit(&cpu_lock);
14438 
14439 	if (state == NULL) {
14440 		if (--dtrace_opens == 0)
14441 			(void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
14442 		mutex_exit(&dtrace_lock);
14443 		return (EAGAIN);
14444 	}
14445 
14446 	mutex_exit(&dtrace_lock);
14447 
14448 	return (0);
14449 }
14450 
14451 /*ARGSUSED*/
14452 static int
14453 dtrace_close(dev_t dev, int flag, int otyp, cred_t *cred_p)
14454 {
14455 	minor_t minor = getminor(dev);
14456 	dtrace_state_t *state;
14457 
14458 	if (minor == DTRACEMNRN_HELPER)
14459 		return (0);
14460 
14461 	state = ddi_get_soft_state(dtrace_softstate, minor);
14462 
14463 	mutex_enter(&cpu_lock);
14464 	mutex_enter(&dtrace_lock);
14465 
14466 	if (state->dts_anon) {
14467 		/*
14468 		 * There is anonymous state. Destroy that first.
14469 		 */
14470 		ASSERT(dtrace_anon.dta_state == NULL);
14471 		dtrace_state_destroy(state->dts_anon);
14472 	}
14473 
14474 	dtrace_state_destroy(state);
14475 	ASSERT(dtrace_opens > 0);
14476 	if (--dtrace_opens == 0)
14477 		(void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
14478 
14479 	mutex_exit(&dtrace_lock);
14480 	mutex_exit(&cpu_lock);
14481 
14482 	return (0);
14483 }
14484 
14485 /*ARGSUSED*/
14486 static int
14487 dtrace_ioctl_helper(int cmd, intptr_t arg, int *rv)
14488 {
14489 	int rval;
14490 	dof_helper_t help, *dhp = NULL;
14491 
14492 	switch (cmd) {
14493 	case DTRACEHIOC_ADDDOF:
14494 		if (copyin((void *)arg, &help, sizeof (help)) != 0) {
14495 			dtrace_dof_error(NULL, "failed to copyin DOF helper");
14496 			return (EFAULT);
14497 		}
14498 
14499 		dhp = &help;
14500 		arg = (intptr_t)help.dofhp_dof;
14501 		/*FALLTHROUGH*/
14502 
14503 	case DTRACEHIOC_ADD: {
14504 		dof_hdr_t *dof = dtrace_dof_copyin(arg, &rval);
14505 
14506 		if (dof == NULL)
14507 			return (rval);
14508 
14509 		mutex_enter(&dtrace_lock);
14510 
14511 		/*
14512 		 * dtrace_helper_slurp() takes responsibility for the dof --
14513 		 * it may free it now or it may save it and free it later.
14514 		 */
14515 		if ((rval = dtrace_helper_slurp(dof, dhp)) != -1) {
14516 			*rv = rval;
14517 			rval = 0;
14518 		} else {
14519 			rval = EINVAL;
14520 		}
14521 
14522 		mutex_exit(&dtrace_lock);
14523 		return (rval);
14524 	}
14525 
14526 	case DTRACEHIOC_REMOVE: {
14527 		mutex_enter(&dtrace_lock);
14528 		rval = dtrace_helper_destroygen(arg);
14529 		mutex_exit(&dtrace_lock);
14530 
14531 		return (rval);
14532 	}
14533 
14534 	default:
14535 		break;
14536 	}
14537 
14538 	return (ENOTTY);
14539 }
14540 
14541 /*ARGSUSED*/
14542 static int
14543 dtrace_ioctl(dev_t dev, int cmd, intptr_t arg, int md, cred_t *cr, int *rv)
14544 {
14545 	minor_t minor = getminor(dev);
14546 	dtrace_state_t *state;
14547 	int rval;
14548 
14549 	if (minor == DTRACEMNRN_HELPER)
14550 		return (dtrace_ioctl_helper(cmd, arg, rv));
14551 
14552 	state = ddi_get_soft_state(dtrace_softstate, minor);
14553 
14554 	if (state->dts_anon) {
14555 		ASSERT(dtrace_anon.dta_state == NULL);
14556 		state = state->dts_anon;
14557 	}
14558 
14559 	switch (cmd) {
14560 	case DTRACEIOC_PROVIDER: {
14561 		dtrace_providerdesc_t pvd;
14562 		dtrace_provider_t *pvp;
14563 
14564 		if (copyin((void *)arg, &pvd, sizeof (pvd)) != 0)
14565 			return (EFAULT);
14566 
14567 		pvd.dtvd_name[DTRACE_PROVNAMELEN - 1] = '\0';
14568 		mutex_enter(&dtrace_provider_lock);
14569 
14570 		for (pvp = dtrace_provider; pvp != NULL; pvp = pvp->dtpv_next) {
14571 			if (strcmp(pvp->dtpv_name, pvd.dtvd_name) == 0)
14572 				break;
14573 		}
14574 
14575 		mutex_exit(&dtrace_provider_lock);
14576 
14577 		if (pvp == NULL)
14578 			return (ESRCH);
14579 
14580 		bcopy(&pvp->dtpv_priv, &pvd.dtvd_priv, sizeof (dtrace_ppriv_t));
14581 		bcopy(&pvp->dtpv_attr, &pvd.dtvd_attr, sizeof (dtrace_pattr_t));
14582 		if (copyout(&pvd, (void *)arg, sizeof (pvd)) != 0)
14583 			return (EFAULT);
14584 
14585 		return (0);
14586 	}
14587 
14588 	case DTRACEIOC_EPROBE: {
14589 		dtrace_eprobedesc_t epdesc;
14590 		dtrace_ecb_t *ecb;
14591 		dtrace_action_t *act;
14592 		void *buf;
14593 		size_t size;
14594 		uintptr_t dest;
14595 		int nrecs;
14596 
14597 		if (copyin((void *)arg, &epdesc, sizeof (epdesc)) != 0)
14598 			return (EFAULT);
14599 
14600 		mutex_enter(&dtrace_lock);
14601 
14602 		if ((ecb = dtrace_epid2ecb(state, epdesc.dtepd_epid)) == NULL) {
14603 			mutex_exit(&dtrace_lock);
14604 			return (EINVAL);
14605 		}
14606 
14607 		if (ecb->dte_probe == NULL) {
14608 			mutex_exit(&dtrace_lock);
14609 			return (EINVAL);
14610 		}
14611 
14612 		epdesc.dtepd_probeid = ecb->dte_probe->dtpr_id;
14613 		epdesc.dtepd_uarg = ecb->dte_uarg;
14614 		epdesc.dtepd_size = ecb->dte_size;
14615 
14616 		nrecs = epdesc.dtepd_nrecs;
14617 		epdesc.dtepd_nrecs = 0;
14618 		for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
14619 			if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
14620 				continue;
14621 
14622 			epdesc.dtepd_nrecs++;
14623 		}
14624 
14625 		/*
14626 		 * Now that we have the size, we need to allocate a temporary
14627 		 * buffer in which to store the complete description.  We need
14628 		 * the temporary buffer to be able to drop dtrace_lock()
14629 		 * across the copyout(), below.
14630 		 */
14631 		size = sizeof (dtrace_eprobedesc_t) +
14632 		    (epdesc.dtepd_nrecs * sizeof (dtrace_recdesc_t));
14633 
14634 		buf = kmem_alloc(size, KM_SLEEP);
14635 		dest = (uintptr_t)buf;
14636 
14637 		bcopy(&epdesc, (void *)dest, sizeof (epdesc));
14638 		dest += offsetof(dtrace_eprobedesc_t, dtepd_rec[0]);
14639 
14640 		for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
14641 			if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
14642 				continue;
14643 
14644 			if (nrecs-- == 0)
14645 				break;
14646 
14647 			bcopy(&act->dta_rec, (void *)dest,
14648 			    sizeof (dtrace_recdesc_t));
14649 			dest += sizeof (dtrace_recdesc_t);
14650 		}
14651 
14652 		mutex_exit(&dtrace_lock);
14653 
14654 		if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
14655 			kmem_free(buf, size);
14656 			return (EFAULT);
14657 		}
14658 
14659 		kmem_free(buf, size);
14660 		return (0);
14661 	}
14662 
14663 	case DTRACEIOC_AGGDESC: {
14664 		dtrace_aggdesc_t aggdesc;
14665 		dtrace_action_t *act;
14666 		dtrace_aggregation_t *agg;
14667 		int nrecs;
14668 		uint32_t offs;
14669 		dtrace_recdesc_t *lrec;
14670 		void *buf;
14671 		size_t size;
14672 		uintptr_t dest;
14673 
14674 		if (copyin((void *)arg, &aggdesc, sizeof (aggdesc)) != 0)
14675 			return (EFAULT);
14676 
14677 		mutex_enter(&dtrace_lock);
14678 
14679 		if ((agg = dtrace_aggid2agg(state, aggdesc.dtagd_id)) == NULL) {
14680 			mutex_exit(&dtrace_lock);
14681 			return (EINVAL);
14682 		}
14683 
14684 		aggdesc.dtagd_epid = agg->dtag_ecb->dte_epid;
14685 
14686 		nrecs = aggdesc.dtagd_nrecs;
14687 		aggdesc.dtagd_nrecs = 0;
14688 
14689 		offs = agg->dtag_base;
14690 		lrec = &agg->dtag_action.dta_rec;
14691 		aggdesc.dtagd_size = lrec->dtrd_offset + lrec->dtrd_size - offs;
14692 
14693 		for (act = agg->dtag_first; ; act = act->dta_next) {
14694 			ASSERT(act->dta_intuple ||
14695 			    DTRACEACT_ISAGG(act->dta_kind));
14696 
14697 			/*
14698 			 * If this action has a record size of zero, it
14699 			 * denotes an argument to the aggregating action.
14700 			 * Because the presence of this record doesn't (or
14701 			 * shouldn't) affect the way the data is interpreted,
14702 			 * we don't copy it out to save user-level the
14703 			 * confusion of dealing with a zero-length record.
14704 			 */
14705 			if (act->dta_rec.dtrd_size == 0) {
14706 				ASSERT(agg->dtag_hasarg);
14707 				continue;
14708 			}
14709 
14710 			aggdesc.dtagd_nrecs++;
14711 
14712 			if (act == &agg->dtag_action)
14713 				break;
14714 		}
14715 
14716 		/*
14717 		 * Now that we have the size, we need to allocate a temporary
14718 		 * buffer in which to store the complete description.  We need
14719 		 * the temporary buffer to be able to drop dtrace_lock()
14720 		 * across the copyout(), below.
14721 		 */
14722 		size = sizeof (dtrace_aggdesc_t) +
14723 		    (aggdesc.dtagd_nrecs * sizeof (dtrace_recdesc_t));
14724 
14725 		buf = kmem_alloc(size, KM_SLEEP);
14726 		dest = (uintptr_t)buf;
14727 
14728 		bcopy(&aggdesc, (void *)dest, sizeof (aggdesc));
14729 		dest += offsetof(dtrace_aggdesc_t, dtagd_rec[0]);
14730 
14731 		for (act = agg->dtag_first; ; act = act->dta_next) {
14732 			dtrace_recdesc_t rec = act->dta_rec;
14733 
14734 			/*
14735 			 * See the comment in the above loop for why we pass
14736 			 * over zero-length records.
14737 			 */
14738 			if (rec.dtrd_size == 0) {
14739 				ASSERT(agg->dtag_hasarg);
14740 				continue;
14741 			}
14742 
14743 			if (nrecs-- == 0)
14744 				break;
14745 
14746 			rec.dtrd_offset -= offs;
14747 			bcopy(&rec, (void *)dest, sizeof (rec));
14748 			dest += sizeof (dtrace_recdesc_t);
14749 
14750 			if (act == &agg->dtag_action)
14751 				break;
14752 		}
14753 
14754 		mutex_exit(&dtrace_lock);
14755 
14756 		if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
14757 			kmem_free(buf, size);
14758 			return (EFAULT);
14759 		}
14760 
14761 		kmem_free(buf, size);
14762 		return (0);
14763 	}
14764 
14765 	case DTRACEIOC_ENABLE: {
14766 		dof_hdr_t *dof;
14767 		dtrace_enabling_t *enab = NULL;
14768 		dtrace_vstate_t *vstate;
14769 		int err = 0;
14770 
14771 		*rv = 0;
14772 
14773 		/*
14774 		 * If a NULL argument has been passed, we take this as our
14775 		 * cue to reevaluate our enablings.
14776 		 */
14777 		if (arg == NULL) {
14778 			dtrace_enabling_matchall();
14779 
14780 			return (0);
14781 		}
14782 
14783 		if ((dof = dtrace_dof_copyin(arg, &rval)) == NULL)
14784 			return (rval);
14785 
14786 		mutex_enter(&cpu_lock);
14787 		mutex_enter(&dtrace_lock);
14788 		vstate = &state->dts_vstate;
14789 
14790 		if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
14791 			mutex_exit(&dtrace_lock);
14792 			mutex_exit(&cpu_lock);
14793 			dtrace_dof_destroy(dof);
14794 			return (EBUSY);
14795 		}
14796 
14797 		if (dtrace_dof_slurp(dof, vstate, cr, &enab, 0, B_TRUE) != 0) {
14798 			mutex_exit(&dtrace_lock);
14799 			mutex_exit(&cpu_lock);
14800 			dtrace_dof_destroy(dof);
14801 			return (EINVAL);
14802 		}
14803 
14804 		if ((rval = dtrace_dof_options(dof, state)) != 0) {
14805 			dtrace_enabling_destroy(enab);
14806 			mutex_exit(&dtrace_lock);
14807 			mutex_exit(&cpu_lock);
14808 			dtrace_dof_destroy(dof);
14809 			return (rval);
14810 		}
14811 
14812 		if ((err = dtrace_enabling_match(enab, rv)) == 0) {
14813 			err = dtrace_enabling_retain(enab);
14814 		} else {
14815 			dtrace_enabling_destroy(enab);
14816 		}
14817 
14818 		mutex_exit(&cpu_lock);
14819 		mutex_exit(&dtrace_lock);
14820 		dtrace_dof_destroy(dof);
14821 
14822 		return (err);
14823 	}
14824 
14825 	case DTRACEIOC_REPLICATE: {
14826 		dtrace_repldesc_t desc;
14827 		dtrace_probedesc_t *match = &desc.dtrpd_match;
14828 		dtrace_probedesc_t *create = &desc.dtrpd_create;
14829 		int err;
14830 
14831 		if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
14832 			return (EFAULT);
14833 
14834 		match->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
14835 		match->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
14836 		match->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
14837 		match->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
14838 
14839 		create->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
14840 		create->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
14841 		create->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
14842 		create->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
14843 
14844 		mutex_enter(&dtrace_lock);
14845 		err = dtrace_enabling_replicate(state, match, create);
14846 		mutex_exit(&dtrace_lock);
14847 
14848 		return (err);
14849 	}
14850 
14851 	case DTRACEIOC_PROBEMATCH:
14852 	case DTRACEIOC_PROBES: {
14853 		dtrace_probe_t *probe = NULL;
14854 		dtrace_probedesc_t desc;
14855 		dtrace_probekey_t pkey;
14856 		dtrace_id_t i;
14857 		int m = 0;
14858 		uint32_t priv;
14859 		uid_t uid;
14860 		zoneid_t zoneid;
14861 
14862 		if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
14863 			return (EFAULT);
14864 
14865 		desc.dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
14866 		desc.dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
14867 		desc.dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
14868 		desc.dtpd_name[DTRACE_NAMELEN - 1] = '\0';
14869 
14870 		/*
14871 		 * Before we attempt to match this probe, we want to give
14872 		 * all providers the opportunity to provide it.
14873 		 */
14874 		if (desc.dtpd_id == DTRACE_IDNONE) {
14875 			mutex_enter(&dtrace_provider_lock);
14876 			dtrace_probe_provide(&desc, NULL);
14877 			mutex_exit(&dtrace_provider_lock);
14878 			desc.dtpd_id++;
14879 		}
14880 
14881 		if (cmd == DTRACEIOC_PROBEMATCH)  {
14882 			dtrace_probekey(&desc, &pkey);
14883 			pkey.dtpk_id = DTRACE_IDNONE;
14884 		}
14885 
14886 		dtrace_cred2priv(cr, &priv, &uid, &zoneid);
14887 
14888 		mutex_enter(&dtrace_lock);
14889 
14890 		if (cmd == DTRACEIOC_PROBEMATCH) {
14891 			for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
14892 				if ((probe = dtrace_probes[i - 1]) != NULL &&
14893 				    (m = dtrace_match_probe(probe, &pkey,
14894 				    priv, uid, zoneid)) != 0)
14895 					break;
14896 			}
14897 
14898 			if (m < 0) {
14899 				mutex_exit(&dtrace_lock);
14900 				return (EINVAL);
14901 			}
14902 
14903 		} else {
14904 			for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
14905 				if ((probe = dtrace_probes[i - 1]) != NULL &&
14906 				    dtrace_match_priv(probe, priv, uid, zoneid))
14907 					break;
14908 			}
14909 		}
14910 
14911 		if (probe == NULL) {
14912 			mutex_exit(&dtrace_lock);
14913 			return (ESRCH);
14914 		}
14915 
14916 		dtrace_probe_description(probe, &desc);
14917 		mutex_exit(&dtrace_lock);
14918 
14919 		if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
14920 			return (EFAULT);
14921 
14922 		return (0);
14923 	}
14924 
14925 	case DTRACEIOC_PROBEARG: {
14926 		dtrace_argdesc_t desc;
14927 		dtrace_probe_t *probe;
14928 		dtrace_provider_t *prov;
14929 
14930 		if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
14931 			return (EFAULT);
14932 
14933 		if (desc.dtargd_id == DTRACE_IDNONE)
14934 			return (EINVAL);
14935 
14936 		if (desc.dtargd_ndx == DTRACE_ARGNONE)
14937 			return (EINVAL);
14938 
14939 		mutex_enter(&dtrace_provider_lock);
14940 		mutex_enter(&mod_lock);
14941 		mutex_enter(&dtrace_lock);
14942 
14943 		if (desc.dtargd_id > dtrace_nprobes) {
14944 			mutex_exit(&dtrace_lock);
14945 			mutex_exit(&mod_lock);
14946 			mutex_exit(&dtrace_provider_lock);
14947 			return (EINVAL);
14948 		}
14949 
14950 		if ((probe = dtrace_probes[desc.dtargd_id - 1]) == NULL) {
14951 			mutex_exit(&dtrace_lock);
14952 			mutex_exit(&mod_lock);
14953 			mutex_exit(&dtrace_provider_lock);
14954 			return (EINVAL);
14955 		}
14956 
14957 		mutex_exit(&dtrace_lock);
14958 
14959 		prov = probe->dtpr_provider;
14960 
14961 		if (prov->dtpv_pops.dtps_getargdesc == NULL) {
14962 			/*
14963 			 * There isn't any typed information for this probe.
14964 			 * Set the argument number to DTRACE_ARGNONE.
14965 			 */
14966 			desc.dtargd_ndx = DTRACE_ARGNONE;
14967 		} else {
14968 			desc.dtargd_native[0] = '\0';
14969 			desc.dtargd_xlate[0] = '\0';
14970 			desc.dtargd_mapping = desc.dtargd_ndx;
14971 
14972 			prov->dtpv_pops.dtps_getargdesc(prov->dtpv_arg,
14973 			    probe->dtpr_id, probe->dtpr_arg, &desc);
14974 		}
14975 
14976 		mutex_exit(&mod_lock);
14977 		mutex_exit(&dtrace_provider_lock);
14978 
14979 		if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
14980 			return (EFAULT);
14981 
14982 		return (0);
14983 	}
14984 
14985 	case DTRACEIOC_GO: {
14986 		processorid_t cpuid;
14987 		rval = dtrace_state_go(state, &cpuid);
14988 
14989 		if (rval != 0)
14990 			return (rval);
14991 
14992 		if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
14993 			return (EFAULT);
14994 
14995 		return (0);
14996 	}
14997 
14998 	case DTRACEIOC_STOP: {
14999 		processorid_t cpuid;
15000 
15001 		mutex_enter(&dtrace_lock);
15002 		rval = dtrace_state_stop(state, &cpuid);
15003 		mutex_exit(&dtrace_lock);
15004 
15005 		if (rval != 0)
15006 			return (rval);
15007 
15008 		if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
15009 			return (EFAULT);
15010 
15011 		return (0);
15012 	}
15013 
15014 	case DTRACEIOC_DOFGET: {
15015 		dof_hdr_t hdr, *dof;
15016 		uint64_t len;
15017 
15018 		if (copyin((void *)arg, &hdr, sizeof (hdr)) != 0)
15019 			return (EFAULT);
15020 
15021 		mutex_enter(&dtrace_lock);
15022 		dof = dtrace_dof_create(state);
15023 		mutex_exit(&dtrace_lock);
15024 
15025 		len = MIN(hdr.dofh_loadsz, dof->dofh_loadsz);
15026 		rval = copyout(dof, (void *)arg, len);
15027 		dtrace_dof_destroy(dof);
15028 
15029 		return (rval == 0 ? 0 : EFAULT);
15030 	}
15031 
15032 	case DTRACEIOC_AGGSNAP:
15033 	case DTRACEIOC_BUFSNAP: {
15034 		dtrace_bufdesc_t desc;
15035 		caddr_t cached;
15036 		dtrace_buffer_t *buf;
15037 
15038 		if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
15039 			return (EFAULT);
15040 
15041 		if (desc.dtbd_cpu < 0 || desc.dtbd_cpu >= NCPU)
15042 			return (EINVAL);
15043 
15044 		mutex_enter(&dtrace_lock);
15045 
15046 		if (cmd == DTRACEIOC_BUFSNAP) {
15047 			buf = &state->dts_buffer[desc.dtbd_cpu];
15048 		} else {
15049 			buf = &state->dts_aggbuffer[desc.dtbd_cpu];
15050 		}
15051 
15052 		if (buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL)) {
15053 			size_t sz = buf->dtb_offset;
15054 
15055 			if (state->dts_activity != DTRACE_ACTIVITY_STOPPED) {
15056 				mutex_exit(&dtrace_lock);
15057 				return (EBUSY);
15058 			}
15059 
15060 			/*
15061 			 * If this buffer has already been consumed, we're
15062 			 * going to indicate that there's nothing left here
15063 			 * to consume.
15064 			 */
15065 			if (buf->dtb_flags & DTRACEBUF_CONSUMED) {
15066 				mutex_exit(&dtrace_lock);
15067 
15068 				desc.dtbd_size = 0;
15069 				desc.dtbd_drops = 0;
15070 				desc.dtbd_errors = 0;
15071 				desc.dtbd_oldest = 0;
15072 				sz = sizeof (desc);
15073 
15074 				if (copyout(&desc, (void *)arg, sz) != 0)
15075 					return (EFAULT);
15076 
15077 				return (0);
15078 			}
15079 
15080 			/*
15081 			 * If this is a ring buffer that has wrapped, we want
15082 			 * to copy the whole thing out.
15083 			 */
15084 			if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
15085 				dtrace_buffer_polish(buf);
15086 				sz = buf->dtb_size;
15087 			}
15088 
15089 			if (copyout(buf->dtb_tomax, desc.dtbd_data, sz) != 0) {
15090 				mutex_exit(&dtrace_lock);
15091 				return (EFAULT);
15092 			}
15093 
15094 			desc.dtbd_size = sz;
15095 			desc.dtbd_drops = buf->dtb_drops;
15096 			desc.dtbd_errors = buf->dtb_errors;
15097 			desc.dtbd_oldest = buf->dtb_xamot_offset;
15098 
15099 			mutex_exit(&dtrace_lock);
15100 
15101 			if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
15102 				return (EFAULT);
15103 
15104 			buf->dtb_flags |= DTRACEBUF_CONSUMED;
15105 
15106 			return (0);
15107 		}
15108 
15109 		if (buf->dtb_tomax == NULL) {
15110 			ASSERT(buf->dtb_xamot == NULL);
15111 			mutex_exit(&dtrace_lock);
15112 			return (ENOENT);
15113 		}
15114 
15115 		cached = buf->dtb_tomax;
15116 		ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
15117 
15118 		dtrace_xcall(desc.dtbd_cpu,
15119 		    (dtrace_xcall_t)dtrace_buffer_switch, buf);
15120 
15121 		state->dts_errors += buf->dtb_xamot_errors;
15122 
15123 		/*
15124 		 * If the buffers did not actually switch, then the cross call
15125 		 * did not take place -- presumably because the given CPU is
15126 		 * not in the ready set.  If this is the case, we'll return
15127 		 * ENOENT.
15128 		 */
15129 		if (buf->dtb_tomax == cached) {
15130 			ASSERT(buf->dtb_xamot != cached);
15131 			mutex_exit(&dtrace_lock);
15132 			return (ENOENT);
15133 		}
15134 
15135 		ASSERT(cached == buf->dtb_xamot);
15136 
15137 		/*
15138 		 * We have our snapshot; now copy it out.
15139 		 */
15140 		if (copyout(buf->dtb_xamot, desc.dtbd_data,
15141 		    buf->dtb_xamot_offset) != 0) {
15142 			mutex_exit(&dtrace_lock);
15143 			return (EFAULT);
15144 		}
15145 
15146 		desc.dtbd_size = buf->dtb_xamot_offset;
15147 		desc.dtbd_drops = buf->dtb_xamot_drops;
15148 		desc.dtbd_errors = buf->dtb_xamot_errors;
15149 		desc.dtbd_oldest = 0;
15150 
15151 		mutex_exit(&dtrace_lock);
15152 
15153 		/*
15154 		 * Finally, copy out the buffer description.
15155 		 */
15156 		if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
15157 			return (EFAULT);
15158 
15159 		return (0);
15160 	}
15161 
15162 	case DTRACEIOC_CONF: {
15163 		dtrace_conf_t conf;
15164 
15165 		bzero(&conf, sizeof (conf));
15166 		conf.dtc_difversion = DIF_VERSION;
15167 		conf.dtc_difintregs = DIF_DIR_NREGS;
15168 		conf.dtc_diftupregs = DIF_DTR_NREGS;
15169 		conf.dtc_ctfmodel = CTF_MODEL_NATIVE;
15170 
15171 		if (copyout(&conf, (void *)arg, sizeof (conf)) != 0)
15172 			return (EFAULT);
15173 
15174 		return (0);
15175 	}
15176 
15177 	case DTRACEIOC_STATUS: {
15178 		dtrace_status_t stat;
15179 		dtrace_dstate_t *dstate;
15180 		int i, j;
15181 		uint64_t nerrs;
15182 
15183 		/*
15184 		 * See the comment in dtrace_state_deadman() for the reason
15185 		 * for setting dts_laststatus to INT64_MAX before setting
15186 		 * it to the correct value.
15187 		 */
15188 		state->dts_laststatus = INT64_MAX;
15189 		dtrace_membar_producer();
15190 		state->dts_laststatus = dtrace_gethrtime();
15191 
15192 		bzero(&stat, sizeof (stat));
15193 
15194 		mutex_enter(&dtrace_lock);
15195 
15196 		if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE) {
15197 			mutex_exit(&dtrace_lock);
15198 			return (ENOENT);
15199 		}
15200 
15201 		if (state->dts_activity == DTRACE_ACTIVITY_DRAINING)
15202 			stat.dtst_exiting = 1;
15203 
15204 		nerrs = state->dts_errors;
15205 		dstate = &state->dts_vstate.dtvs_dynvars;
15206 
15207 		for (i = 0; i < NCPU; i++) {
15208 			dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[i];
15209 
15210 			stat.dtst_dyndrops += dcpu->dtdsc_drops;
15211 			stat.dtst_dyndrops_dirty += dcpu->dtdsc_dirty_drops;
15212 			stat.dtst_dyndrops_rinsing += dcpu->dtdsc_rinsing_drops;
15213 
15214 			if (state->dts_buffer[i].dtb_flags & DTRACEBUF_FULL)
15215 				stat.dtst_filled++;
15216 
15217 			nerrs += state->dts_buffer[i].dtb_errors;
15218 
15219 			for (j = 0; j < state->dts_nspeculations; j++) {
15220 				dtrace_speculation_t *spec;
15221 				dtrace_buffer_t *buf;
15222 
15223 				spec = &state->dts_speculations[j];
15224 				buf = &spec->dtsp_buffer[i];
15225 				stat.dtst_specdrops += buf->dtb_xamot_drops;
15226 			}
15227 		}
15228 
15229 		stat.dtst_specdrops_busy = state->dts_speculations_busy;
15230 		stat.dtst_specdrops_unavail = state->dts_speculations_unavail;
15231 		stat.dtst_stkstroverflows = state->dts_stkstroverflows;
15232 		stat.dtst_dblerrors = state->dts_dblerrors;
15233 		stat.dtst_killed =
15234 		    (state->dts_activity == DTRACE_ACTIVITY_KILLED);
15235 		stat.dtst_errors = nerrs;
15236 
15237 		mutex_exit(&dtrace_lock);
15238 
15239 		if (copyout(&stat, (void *)arg, sizeof (stat)) != 0)
15240 			return (EFAULT);
15241 
15242 		return (0);
15243 	}
15244 
15245 	case DTRACEIOC_FORMAT: {
15246 		dtrace_fmtdesc_t fmt;
15247 		char *str;
15248 		int len;
15249 
15250 		if (copyin((void *)arg, &fmt, sizeof (fmt)) != 0)
15251 			return (EFAULT);
15252 
15253 		mutex_enter(&dtrace_lock);
15254 
15255 		if (fmt.dtfd_format == 0 ||
15256 		    fmt.dtfd_format > state->dts_nformats) {
15257 			mutex_exit(&dtrace_lock);
15258 			return (EINVAL);
15259 		}
15260 
15261 		/*
15262 		 * Format strings are allocated contiguously and they are
15263 		 * never freed; if a format index is less than the number
15264 		 * of formats, we can assert that the format map is non-NULL
15265 		 * and that the format for the specified index is non-NULL.
15266 		 */
15267 		ASSERT(state->dts_formats != NULL);
15268 		str = state->dts_formats[fmt.dtfd_format - 1];
15269 		ASSERT(str != NULL);
15270 
15271 		len = strlen(str) + 1;
15272 
15273 		if (len > fmt.dtfd_length) {
15274 			fmt.dtfd_length = len;
15275 
15276 			if (copyout(&fmt, (void *)arg, sizeof (fmt)) != 0) {
15277 				mutex_exit(&dtrace_lock);
15278 				return (EINVAL);
15279 			}
15280 		} else {
15281 			if (copyout(str, fmt.dtfd_string, len) != 0) {
15282 				mutex_exit(&dtrace_lock);
15283 				return (EINVAL);
15284 			}
15285 		}
15286 
15287 		mutex_exit(&dtrace_lock);
15288 		return (0);
15289 	}
15290 
15291 	default:
15292 		break;
15293 	}
15294 
15295 	return (ENOTTY);
15296 }
15297 
15298 /*ARGSUSED*/
15299 static int
15300 dtrace_detach(dev_info_t *dip, ddi_detach_cmd_t cmd)
15301 {
15302 	dtrace_state_t *state;
15303 
15304 	switch (cmd) {
15305 	case DDI_DETACH:
15306 		break;
15307 
15308 	case DDI_SUSPEND:
15309 		return (DDI_SUCCESS);
15310 
15311 	default:
15312 		return (DDI_FAILURE);
15313 	}
15314 
15315 	mutex_enter(&cpu_lock);
15316 	mutex_enter(&dtrace_provider_lock);
15317 	mutex_enter(&dtrace_lock);
15318 
15319 	ASSERT(dtrace_opens == 0);
15320 
15321 	if (dtrace_helpers > 0) {
15322 		mutex_exit(&dtrace_provider_lock);
15323 		mutex_exit(&dtrace_lock);
15324 		mutex_exit(&cpu_lock);
15325 		return (DDI_FAILURE);
15326 	}
15327 
15328 	if (dtrace_unregister((dtrace_provider_id_t)dtrace_provider) != 0) {
15329 		mutex_exit(&dtrace_provider_lock);
15330 		mutex_exit(&dtrace_lock);
15331 		mutex_exit(&cpu_lock);
15332 		return (DDI_FAILURE);
15333 	}
15334 
15335 	dtrace_provider = NULL;
15336 
15337 	if ((state = dtrace_anon_grab()) != NULL) {
15338 		/*
15339 		 * If there were ECBs on this state, the provider should
15340 		 * have not been allowed to detach; assert that there is
15341 		 * none.
15342 		 */
15343 		ASSERT(state->dts_necbs == 0);
15344 		dtrace_state_destroy(state);
15345 
15346 		/*
15347 		 * If we're being detached with anonymous state, we need to
15348 		 * indicate to the kernel debugger that DTrace is now inactive.
15349 		 */
15350 		(void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
15351 	}
15352 
15353 	bzero(&dtrace_anon, sizeof (dtrace_anon_t));
15354 	unregister_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
15355 	dtrace_cpu_init = NULL;
15356 	dtrace_helpers_cleanup = NULL;
15357 	dtrace_helpers_fork = NULL;
15358 	dtrace_cpustart_init = NULL;
15359 	dtrace_cpustart_fini = NULL;
15360 	dtrace_debugger_init = NULL;
15361 	dtrace_debugger_fini = NULL;
15362 	dtrace_modload = NULL;
15363 	dtrace_modunload = NULL;
15364 
15365 	mutex_exit(&cpu_lock);
15366 
15367 	if (dtrace_helptrace_enabled) {
15368 		kmem_free(dtrace_helptrace_buffer, dtrace_helptrace_bufsize);
15369 		dtrace_helptrace_buffer = NULL;
15370 	}
15371 
15372 	kmem_free(dtrace_probes, dtrace_nprobes * sizeof (dtrace_probe_t *));
15373 	dtrace_probes = NULL;
15374 	dtrace_nprobes = 0;
15375 
15376 	dtrace_hash_destroy(dtrace_bymod);
15377 	dtrace_hash_destroy(dtrace_byfunc);
15378 	dtrace_hash_destroy(dtrace_byname);
15379 	dtrace_bymod = NULL;
15380 	dtrace_byfunc = NULL;
15381 	dtrace_byname = NULL;
15382 
15383 	kmem_cache_destroy(dtrace_state_cache);
15384 	vmem_destroy(dtrace_minor);
15385 	vmem_destroy(dtrace_arena);
15386 
15387 	if (dtrace_toxrange != NULL) {
15388 		kmem_free(dtrace_toxrange,
15389 		    dtrace_toxranges_max * sizeof (dtrace_toxrange_t));
15390 		dtrace_toxrange = NULL;
15391 		dtrace_toxranges = 0;
15392 		dtrace_toxranges_max = 0;
15393 	}
15394 
15395 	ddi_remove_minor_node(dtrace_devi, NULL);
15396 	dtrace_devi = NULL;
15397 
15398 	ddi_soft_state_fini(&dtrace_softstate);
15399 
15400 	ASSERT(dtrace_vtime_references == 0);
15401 	ASSERT(dtrace_opens == 0);
15402 	ASSERT(dtrace_retained == NULL);
15403 
15404 	mutex_exit(&dtrace_lock);
15405 	mutex_exit(&dtrace_provider_lock);
15406 
15407 	/*
15408 	 * We don't destroy the task queue until after we have dropped our
15409 	 * locks (taskq_destroy() may block on running tasks).  To prevent
15410 	 * attempting to do work after we have effectively detached but before
15411 	 * the task queue has been destroyed, all tasks dispatched via the
15412 	 * task queue must check that DTrace is still attached before
15413 	 * performing any operation.
15414 	 */
15415 	taskq_destroy(dtrace_taskq);
15416 	dtrace_taskq = NULL;
15417 
15418 	return (DDI_SUCCESS);
15419 }
15420 
15421 /*ARGSUSED*/
15422 static int
15423 dtrace_info(dev_info_t *dip, ddi_info_cmd_t infocmd, void *arg, void **result)
15424 {
15425 	int error;
15426 
15427 	switch (infocmd) {
15428 	case DDI_INFO_DEVT2DEVINFO:
15429 		*result = (void *)dtrace_devi;
15430 		error = DDI_SUCCESS;
15431 		break;
15432 	case DDI_INFO_DEVT2INSTANCE:
15433 		*result = (void *)0;
15434 		error = DDI_SUCCESS;
15435 		break;
15436 	default:
15437 		error = DDI_FAILURE;
15438 	}
15439 	return (error);
15440 }
15441 
15442 static struct cb_ops dtrace_cb_ops = {
15443 	dtrace_open,		/* open */
15444 	dtrace_close,		/* close */
15445 	nulldev,		/* strategy */
15446 	nulldev,		/* print */
15447 	nodev,			/* dump */
15448 	nodev,			/* read */
15449 	nodev,			/* write */
15450 	dtrace_ioctl,		/* ioctl */
15451 	nodev,			/* devmap */
15452 	nodev,			/* mmap */
15453 	nodev,			/* segmap */
15454 	nochpoll,		/* poll */
15455 	ddi_prop_op,		/* cb_prop_op */
15456 	0,			/* streamtab  */
15457 	D_NEW | D_MP		/* Driver compatibility flag */
15458 };
15459 
15460 static struct dev_ops dtrace_ops = {
15461 	DEVO_REV,		/* devo_rev */
15462 	0,			/* refcnt */
15463 	dtrace_info,		/* get_dev_info */
15464 	nulldev,		/* identify */
15465 	nulldev,		/* probe */
15466 	dtrace_attach,		/* attach */
15467 	dtrace_detach,		/* detach */
15468 	nodev,			/* reset */
15469 	&dtrace_cb_ops,		/* driver operations */
15470 	NULL,			/* bus operations */
15471 	nodev			/* dev power */
15472 };
15473 
15474 static struct modldrv modldrv = {
15475 	&mod_driverops,		/* module type (this is a pseudo driver) */
15476 	"Dynamic Tracing",	/* name of module */
15477 	&dtrace_ops,		/* driver ops */
15478 };
15479 
15480 static struct modlinkage modlinkage = {
15481 	MODREV_1,
15482 	(void *)&modldrv,
15483 	NULL
15484 };
15485 
15486 int
15487 _init(void)
15488 {
15489 	return (mod_install(&modlinkage));
15490 }
15491 
15492 int
15493 _info(struct modinfo *modinfop)
15494 {
15495 	return (mod_info(&modlinkage, modinfop));
15496 }
15497 
15498 int
15499 _fini(void)
15500 {
15501 	return (mod_remove(&modlinkage));
15502 }
15503