xref: /freebsd/sys/cddl/contrib/opensolaris/uts/common/dtrace/dtrace.c (revision 2e654ff9dfeab072161ca370e2b3f621ed67e393)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  *
21  * $FreeBSD$
22  */
23 
24 /*
25  * Copyright 2008 Sun Microsystems, Inc.  All rights reserved.
26  * Use is subject to license terms.
27  */
28 
29 #pragma ident	"%Z%%M%	%I%	%E% SMI"
30 
31 /*
32  * DTrace - Dynamic Tracing for Solaris
33  *
34  * This is the implementation of the Solaris Dynamic Tracing framework
35  * (DTrace).  The user-visible interface to DTrace is described at length in
36  * the "Solaris Dynamic Tracing Guide".  The interfaces between the libdtrace
37  * library, the in-kernel DTrace framework, and the DTrace providers are
38  * described in the block comments in the <sys/dtrace.h> header file.  The
39  * internal architecture of DTrace is described in the block comments in the
40  * <sys/dtrace_impl.h> header file.  The comments contained within the DTrace
41  * implementation very much assume mastery of all of these sources; if one has
42  * an unanswered question about the implementation, one should consult them
43  * first.
44  *
45  * The functions here are ordered roughly as follows:
46  *
47  *   - Probe context functions
48  *   - Probe hashing functions
49  *   - Non-probe context utility functions
50  *   - Matching functions
51  *   - Provider-to-Framework API functions
52  *   - Probe management functions
53  *   - DIF object functions
54  *   - Format functions
55  *   - Predicate functions
56  *   - ECB functions
57  *   - Buffer functions
58  *   - Enabling functions
59  *   - DOF functions
60  *   - Anonymous enabling functions
61  *   - Consumer state functions
62  *   - Helper functions
63  *   - Hook functions
64  *   - Driver cookbook functions
65  *
66  * Each group of functions begins with a block comment labelled the "DTrace
67  * [Group] Functions", allowing one to find each block by searching forward
68  * on capital-f functions.
69  */
70 #include <sys/errno.h>
71 #if !defined(sun)
72 #include <sys/time.h>
73 #endif
74 #include <sys/stat.h>
75 #include <sys/modctl.h>
76 #include <sys/conf.h>
77 #include <sys/systm.h>
78 #if defined(sun)
79 #include <sys/ddi.h>
80 #include <sys/sunddi.h>
81 #endif
82 #include <sys/cpuvar.h>
83 #include <sys/kmem.h>
84 #if defined(sun)
85 #include <sys/strsubr.h>
86 #endif
87 #include <sys/sysmacros.h>
88 #include <sys/dtrace_impl.h>
89 #include <sys/atomic.h>
90 #include <sys/cmn_err.h>
91 #if defined(sun)
92 #include <sys/mutex_impl.h>
93 #include <sys/rwlock_impl.h>
94 #endif
95 #include <sys/ctf_api.h>
96 #if defined(sun)
97 #include <sys/panic.h>
98 #include <sys/priv_impl.h>
99 #endif
100 #include <sys/policy.h>
101 #if defined(sun)
102 #include <sys/cred_impl.h>
103 #include <sys/procfs_isa.h>
104 #endif
105 #include <sys/taskq.h>
106 #if defined(sun)
107 #include <sys/mkdev.h>
108 #include <sys/kdi.h>
109 #endif
110 #include <sys/zone.h>
111 #include <sys/socket.h>
112 #include <netinet/in.h>
113 
114 /* FreeBSD includes: */
115 #if !defined(sun)
116 #include <sys/callout.h>
117 #include <sys/ctype.h>
118 #include <sys/limits.h>
119 #include <sys/kdb.h>
120 #include <sys/kernel.h>
121 #include <sys/malloc.h>
122 #include <sys/sysctl.h>
123 #include <sys/lock.h>
124 #include <sys/mutex.h>
125 #include <sys/rwlock.h>
126 #include <sys/sx.h>
127 #include <sys/dtrace_bsd.h>
128 #include <netinet/in.h>
129 #include "dtrace_cddl.h"
130 #include "dtrace_debug.c"
131 #endif
132 
133 /*
134  * DTrace Tunable Variables
135  *
136  * The following variables may be tuned by adding a line to /etc/system that
137  * includes both the name of the DTrace module ("dtrace") and the name of the
138  * variable.  For example:
139  *
140  *   set dtrace:dtrace_destructive_disallow = 1
141  *
142  * In general, the only variables that one should be tuning this way are those
143  * that affect system-wide DTrace behavior, and for which the default behavior
144  * is undesirable.  Most of these variables are tunable on a per-consumer
145  * basis using DTrace options, and need not be tuned on a system-wide basis.
146  * When tuning these variables, avoid pathological values; while some attempt
147  * is made to verify the integrity of these variables, they are not considered
148  * part of the supported interface to DTrace, and they are therefore not
149  * checked comprehensively.  Further, these variables should not be tuned
150  * dynamically via "mdb -kw" or other means; they should only be tuned via
151  * /etc/system.
152  */
153 int		dtrace_destructive_disallow = 0;
154 dtrace_optval_t	dtrace_nonroot_maxsize = (16 * 1024 * 1024);
155 size_t		dtrace_difo_maxsize = (256 * 1024);
156 dtrace_optval_t	dtrace_dof_maxsize = (256 * 1024);
157 size_t		dtrace_global_maxsize = (16 * 1024);
158 size_t		dtrace_actions_max = (16 * 1024);
159 size_t		dtrace_retain_max = 1024;
160 dtrace_optval_t	dtrace_helper_actions_max = 128;
161 dtrace_optval_t	dtrace_helper_providers_max = 32;
162 dtrace_optval_t	dtrace_dstate_defsize = (1 * 1024 * 1024);
163 size_t		dtrace_strsize_default = 256;
164 dtrace_optval_t	dtrace_cleanrate_default = 9900990;		/* 101 hz */
165 dtrace_optval_t	dtrace_cleanrate_min = 200000;			/* 5000 hz */
166 dtrace_optval_t	dtrace_cleanrate_max = (uint64_t)60 * NANOSEC;	/* 1/minute */
167 dtrace_optval_t	dtrace_aggrate_default = NANOSEC;		/* 1 hz */
168 dtrace_optval_t	dtrace_statusrate_default = NANOSEC;		/* 1 hz */
169 dtrace_optval_t dtrace_statusrate_max = (hrtime_t)10 * NANOSEC;	 /* 6/minute */
170 dtrace_optval_t	dtrace_switchrate_default = NANOSEC;		/* 1 hz */
171 dtrace_optval_t	dtrace_nspec_default = 1;
172 dtrace_optval_t	dtrace_specsize_default = 32 * 1024;
173 dtrace_optval_t dtrace_stackframes_default = 20;
174 dtrace_optval_t dtrace_ustackframes_default = 20;
175 dtrace_optval_t dtrace_jstackframes_default = 50;
176 dtrace_optval_t dtrace_jstackstrsize_default = 512;
177 int		dtrace_msgdsize_max = 128;
178 hrtime_t	dtrace_chill_max = 500 * (NANOSEC / MILLISEC);	/* 500 ms */
179 hrtime_t	dtrace_chill_interval = NANOSEC;		/* 1000 ms */
180 int		dtrace_devdepth_max = 32;
181 int		dtrace_err_verbose;
182 hrtime_t	dtrace_deadman_interval = NANOSEC;
183 hrtime_t	dtrace_deadman_timeout = (hrtime_t)10 * NANOSEC;
184 hrtime_t	dtrace_deadman_user = (hrtime_t)30 * NANOSEC;
185 hrtime_t	dtrace_unregister_defunct_reap = (hrtime_t)60 * NANOSEC;
186 
187 /*
188  * DTrace External Variables
189  *
190  * As dtrace(7D) is a kernel module, any DTrace variables are obviously
191  * available to DTrace consumers via the backtick (`) syntax.  One of these,
192  * dtrace_zero, is made deliberately so:  it is provided as a source of
193  * well-known, zero-filled memory.  While this variable is not documented,
194  * it is used by some translators as an implementation detail.
195  */
196 const char	dtrace_zero[256] = { 0 };	/* zero-filled memory */
197 
198 /*
199  * DTrace Internal Variables
200  */
201 #if defined(sun)
202 static dev_info_t	*dtrace_devi;		/* device info */
203 #endif
204 #if defined(sun)
205 static vmem_t		*dtrace_arena;		/* probe ID arena */
206 static vmem_t		*dtrace_minor;		/* minor number arena */
207 #else
208 static taskq_t		*dtrace_taskq;		/* task queue */
209 static struct unrhdr	*dtrace_arena;		/* Probe ID number.     */
210 #endif
211 static dtrace_probe_t	**dtrace_probes;	/* array of all probes */
212 static int		dtrace_nprobes;		/* number of probes */
213 static dtrace_provider_t *dtrace_provider;	/* provider list */
214 static dtrace_meta_t	*dtrace_meta_pid;	/* user-land meta provider */
215 static int		dtrace_opens;		/* number of opens */
216 static int		dtrace_helpers;		/* number of helpers */
217 #if defined(sun)
218 static void		*dtrace_softstate;	/* softstate pointer */
219 #endif
220 static dtrace_hash_t	*dtrace_bymod;		/* probes hashed by module */
221 static dtrace_hash_t	*dtrace_byfunc;		/* probes hashed by function */
222 static dtrace_hash_t	*dtrace_byname;		/* probes hashed by name */
223 static dtrace_toxrange_t *dtrace_toxrange;	/* toxic range array */
224 static int		dtrace_toxranges;	/* number of toxic ranges */
225 static int		dtrace_toxranges_max;	/* size of toxic range array */
226 static dtrace_anon_t	dtrace_anon;		/* anonymous enabling */
227 static kmem_cache_t	*dtrace_state_cache;	/* cache for dynamic state */
228 static uint64_t		dtrace_vtime_references; /* number of vtimestamp refs */
229 static kthread_t	*dtrace_panicked;	/* panicking thread */
230 static dtrace_ecb_t	*dtrace_ecb_create_cache; /* cached created ECB */
231 static dtrace_genid_t	dtrace_probegen;	/* current probe generation */
232 static dtrace_helpers_t *dtrace_deferred_pid;	/* deferred helper list */
233 static dtrace_enabling_t *dtrace_retained;	/* list of retained enablings */
234 static dtrace_dynvar_t	dtrace_dynhash_sink;	/* end of dynamic hash chains */
235 #if !defined(sun)
236 static struct mtx	dtrace_unr_mtx;
237 MTX_SYSINIT(dtrace_unr_mtx, &dtrace_unr_mtx, "Unique resource identifier", MTX_DEF);
238 int		dtrace_in_probe;	/* non-zero if executing a probe */
239 #if defined(__i386__) || defined(__amd64__) || defined(__mips__) || defined(__powerpc__)
240 uintptr_t	dtrace_in_probe_addr;	/* Address of invop when already in probe */
241 #endif
242 #endif
243 
244 /*
245  * DTrace Locking
246  * DTrace is protected by three (relatively coarse-grained) locks:
247  *
248  * (1) dtrace_lock is required to manipulate essentially any DTrace state,
249  *     including enabling state, probes, ECBs, consumer state, helper state,
250  *     etc.  Importantly, dtrace_lock is _not_ required when in probe context;
251  *     probe context is lock-free -- synchronization is handled via the
252  *     dtrace_sync() cross call mechanism.
253  *
254  * (2) dtrace_provider_lock is required when manipulating provider state, or
255  *     when provider state must be held constant.
256  *
257  * (3) dtrace_meta_lock is required when manipulating meta provider state, or
258  *     when meta provider state must be held constant.
259  *
260  * The lock ordering between these three locks is dtrace_meta_lock before
261  * dtrace_provider_lock before dtrace_lock.  (In particular, there are
262  * several places where dtrace_provider_lock is held by the framework as it
263  * calls into the providers -- which then call back into the framework,
264  * grabbing dtrace_lock.)
265  *
266  * There are two other locks in the mix:  mod_lock and cpu_lock.  With respect
267  * to dtrace_provider_lock and dtrace_lock, cpu_lock continues its historical
268  * role as a coarse-grained lock; it is acquired before both of these locks.
269  * With respect to dtrace_meta_lock, its behavior is stranger:  cpu_lock must
270  * be acquired _between_ dtrace_meta_lock and any other DTrace locks.
271  * mod_lock is similar with respect to dtrace_provider_lock in that it must be
272  * acquired _between_ dtrace_provider_lock and dtrace_lock.
273  */
274 static kmutex_t		dtrace_lock;		/* probe state lock */
275 static kmutex_t		dtrace_provider_lock;	/* provider state lock */
276 static kmutex_t		dtrace_meta_lock;	/* meta-provider state lock */
277 
278 #if !defined(sun)
279 /* XXX FreeBSD hacks. */
280 static kmutex_t		mod_lock;
281 
282 #define cr_suid		cr_svuid
283 #define cr_sgid		cr_svgid
284 #define	ipaddr_t	in_addr_t
285 #define mod_modname	pathname
286 #define vuprintf	vprintf
287 #define ttoproc(_a)	((_a)->td_proc)
288 #define crgetzoneid(_a)	0
289 #define	NCPU		MAXCPU
290 #define SNOCD		0
291 #define CPU_ON_INTR(_a)	0
292 
293 #define PRIV_EFFECTIVE		(1 << 0)
294 #define PRIV_DTRACE_KERNEL	(1 << 1)
295 #define PRIV_DTRACE_PROC	(1 << 2)
296 #define PRIV_DTRACE_USER	(1 << 3)
297 #define PRIV_PROC_OWNER		(1 << 4)
298 #define PRIV_PROC_ZONE		(1 << 5)
299 #define PRIV_ALL		~0
300 
301 SYSCTL_NODE(_debug, OID_AUTO, dtrace, CTLFLAG_RD, 0, "DTrace Information");
302 #endif
303 
304 #if defined(sun)
305 #define curcpu	CPU->cpu_id
306 #endif
307 
308 
309 /*
310  * DTrace Provider Variables
311  *
312  * These are the variables relating to DTrace as a provider (that is, the
313  * provider of the BEGIN, END, and ERROR probes).
314  */
315 static dtrace_pattr_t	dtrace_provider_attr = {
316 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
317 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
318 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
319 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
320 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
321 };
322 
323 static void
324 dtrace_nullop(void)
325 {}
326 
327 static dtrace_pops_t	dtrace_provider_ops = {
328 	(void (*)(void *, dtrace_probedesc_t *))dtrace_nullop,
329 	(void (*)(void *, modctl_t *))dtrace_nullop,
330 	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
331 	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
332 	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
333 	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
334 	NULL,
335 	NULL,
336 	NULL,
337 	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop
338 };
339 
340 static dtrace_id_t	dtrace_probeid_begin;	/* special BEGIN probe */
341 static dtrace_id_t	dtrace_probeid_end;	/* special END probe */
342 dtrace_id_t		dtrace_probeid_error;	/* special ERROR probe */
343 
344 /*
345  * DTrace Helper Tracing Variables
346  */
347 uint32_t dtrace_helptrace_next = 0;
348 uint32_t dtrace_helptrace_nlocals;
349 char	*dtrace_helptrace_buffer;
350 int	dtrace_helptrace_bufsize = 512 * 1024;
351 
352 #ifdef DEBUG
353 int	dtrace_helptrace_enabled = 1;
354 #else
355 int	dtrace_helptrace_enabled = 0;
356 #endif
357 
358 /*
359  * DTrace Error Hashing
360  *
361  * On DEBUG kernels, DTrace will track the errors that has seen in a hash
362  * table.  This is very useful for checking coverage of tests that are
363  * expected to induce DIF or DOF processing errors, and may be useful for
364  * debugging problems in the DIF code generator or in DOF generation .  The
365  * error hash may be examined with the ::dtrace_errhash MDB dcmd.
366  */
367 #ifdef DEBUG
368 static dtrace_errhash_t	dtrace_errhash[DTRACE_ERRHASHSZ];
369 static const char *dtrace_errlast;
370 static kthread_t *dtrace_errthread;
371 static kmutex_t dtrace_errlock;
372 #endif
373 
374 /*
375  * DTrace Macros and Constants
376  *
377  * These are various macros that are useful in various spots in the
378  * implementation, along with a few random constants that have no meaning
379  * outside of the implementation.  There is no real structure to this cpp
380  * mishmash -- but is there ever?
381  */
382 #define	DTRACE_HASHSTR(hash, probe)	\
383 	dtrace_hash_str(*((char **)((uintptr_t)(probe) + (hash)->dth_stroffs)))
384 
385 #define	DTRACE_HASHNEXT(hash, probe)	\
386 	(dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_nextoffs)
387 
388 #define	DTRACE_HASHPREV(hash, probe)	\
389 	(dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_prevoffs)
390 
391 #define	DTRACE_HASHEQ(hash, lhs, rhs)	\
392 	(strcmp(*((char **)((uintptr_t)(lhs) + (hash)->dth_stroffs)), \
393 	    *((char **)((uintptr_t)(rhs) + (hash)->dth_stroffs))) == 0)
394 
395 #define	DTRACE_AGGHASHSIZE_SLEW		17
396 
397 #define	DTRACE_V4MAPPED_OFFSET		(sizeof (uint32_t) * 3)
398 
399 /*
400  * The key for a thread-local variable consists of the lower 61 bits of the
401  * t_did, plus the 3 bits of the highest active interrupt above LOCK_LEVEL.
402  * We add DIF_VARIABLE_MAX to t_did to assure that the thread key is never
403  * equal to a variable identifier.  This is necessary (but not sufficient) to
404  * assure that global associative arrays never collide with thread-local
405  * variables.  To guarantee that they cannot collide, we must also define the
406  * order for keying dynamic variables.  That order is:
407  *
408  *   [ key0 ] ... [ keyn ] [ variable-key ] [ tls-key ]
409  *
410  * Because the variable-key and the tls-key are in orthogonal spaces, there is
411  * no way for a global variable key signature to match a thread-local key
412  * signature.
413  */
414 #if defined(sun)
415 #define	DTRACE_TLS_THRKEY(where) { \
416 	uint_t intr = 0; \
417 	uint_t actv = CPU->cpu_intr_actv >> (LOCK_LEVEL + 1); \
418 	for (; actv; actv >>= 1) \
419 		intr++; \
420 	ASSERT(intr < (1 << 3)); \
421 	(where) = ((curthread->t_did + DIF_VARIABLE_MAX) & \
422 	    (((uint64_t)1 << 61) - 1)) | ((uint64_t)intr << 61); \
423 }
424 #else
425 #define	DTRACE_TLS_THRKEY(where) { \
426 	solaris_cpu_t *_c = &solaris_cpu[curcpu]; \
427 	uint_t intr = 0; \
428 	uint_t actv = _c->cpu_intr_actv; \
429 	for (; actv; actv >>= 1) \
430 		intr++; \
431 	ASSERT(intr < (1 << 3)); \
432 	(where) = ((curthread->td_tid + DIF_VARIABLE_MAX) & \
433 	    (((uint64_t)1 << 61) - 1)) | ((uint64_t)intr << 61); \
434 }
435 #endif
436 
437 #define	DT_BSWAP_8(x)	((x) & 0xff)
438 #define	DT_BSWAP_16(x)	((DT_BSWAP_8(x) << 8) | DT_BSWAP_8((x) >> 8))
439 #define	DT_BSWAP_32(x)	((DT_BSWAP_16(x) << 16) | DT_BSWAP_16((x) >> 16))
440 #define	DT_BSWAP_64(x)	((DT_BSWAP_32(x) << 32) | DT_BSWAP_32((x) >> 32))
441 
442 #define	DT_MASK_LO 0x00000000FFFFFFFFULL
443 
444 #define	DTRACE_STORE(type, tomax, offset, what) \
445 	*((type *)((uintptr_t)(tomax) + (uintptr_t)offset)) = (type)(what);
446 
447 #ifndef __x86
448 #define	DTRACE_ALIGNCHECK(addr, size, flags)				\
449 	if (addr & (size - 1)) {					\
450 		*flags |= CPU_DTRACE_BADALIGN;				\
451 		cpu_core[curcpu].cpuc_dtrace_illval = addr;	\
452 		return (0);						\
453 	}
454 #else
455 #define	DTRACE_ALIGNCHECK(addr, size, flags)
456 #endif
457 
458 /*
459  * Test whether a range of memory starting at testaddr of size testsz falls
460  * within the range of memory described by addr, sz.  We take care to avoid
461  * problems with overflow and underflow of the unsigned quantities, and
462  * disallow all negative sizes.  Ranges of size 0 are allowed.
463  */
464 #define	DTRACE_INRANGE(testaddr, testsz, baseaddr, basesz) \
465 	((testaddr) - (baseaddr) < (basesz) && \
466 	(testaddr) + (testsz) - (baseaddr) <= (basesz) && \
467 	(testaddr) + (testsz) >= (testaddr))
468 
469 /*
470  * Test whether alloc_sz bytes will fit in the scratch region.  We isolate
471  * alloc_sz on the righthand side of the comparison in order to avoid overflow
472  * or underflow in the comparison with it.  This is simpler than the INRANGE
473  * check above, because we know that the dtms_scratch_ptr is valid in the
474  * range.  Allocations of size zero are allowed.
475  */
476 #define	DTRACE_INSCRATCH(mstate, alloc_sz) \
477 	((mstate)->dtms_scratch_base + (mstate)->dtms_scratch_size - \
478 	(mstate)->dtms_scratch_ptr >= (alloc_sz))
479 
480 #define	DTRACE_LOADFUNC(bits)						\
481 /*CSTYLED*/								\
482 uint##bits##_t								\
483 dtrace_load##bits(uintptr_t addr)					\
484 {									\
485 	size_t size = bits / NBBY;					\
486 	/*CSTYLED*/							\
487 	uint##bits##_t rval;						\
488 	int i;								\
489 	volatile uint16_t *flags = (volatile uint16_t *)		\
490 	    &cpu_core[curcpu].cpuc_dtrace_flags;			\
491 									\
492 	DTRACE_ALIGNCHECK(addr, size, flags);				\
493 									\
494 	for (i = 0; i < dtrace_toxranges; i++) {			\
495 		if (addr >= dtrace_toxrange[i].dtt_limit)		\
496 			continue;					\
497 									\
498 		if (addr + size <= dtrace_toxrange[i].dtt_base)		\
499 			continue;					\
500 									\
501 		/*							\
502 		 * This address falls within a toxic region; return 0.	\
503 		 */							\
504 		*flags |= CPU_DTRACE_BADADDR;				\
505 		cpu_core[curcpu].cpuc_dtrace_illval = addr;		\
506 		return (0);						\
507 	}								\
508 									\
509 	*flags |= CPU_DTRACE_NOFAULT;					\
510 	/*CSTYLED*/							\
511 	rval = *((volatile uint##bits##_t *)addr);			\
512 	*flags &= ~CPU_DTRACE_NOFAULT;					\
513 									\
514 	return (!(*flags & CPU_DTRACE_FAULT) ? rval : 0);		\
515 }
516 
517 #ifdef _LP64
518 #define	dtrace_loadptr	dtrace_load64
519 #else
520 #define	dtrace_loadptr	dtrace_load32
521 #endif
522 
523 #define	DTRACE_DYNHASH_FREE	0
524 #define	DTRACE_DYNHASH_SINK	1
525 #define	DTRACE_DYNHASH_VALID	2
526 
527 #define	DTRACE_MATCH_NEXT	0
528 #define	DTRACE_MATCH_DONE	1
529 #define	DTRACE_ANCHORED(probe)	((probe)->dtpr_func[0] != '\0')
530 #define	DTRACE_STATE_ALIGN	64
531 
532 #define	DTRACE_FLAGS2FLT(flags)						\
533 	(((flags) & CPU_DTRACE_BADADDR) ? DTRACEFLT_BADADDR :		\
534 	((flags) & CPU_DTRACE_ILLOP) ? DTRACEFLT_ILLOP :		\
535 	((flags) & CPU_DTRACE_DIVZERO) ? DTRACEFLT_DIVZERO :		\
536 	((flags) & CPU_DTRACE_KPRIV) ? DTRACEFLT_KPRIV :		\
537 	((flags) & CPU_DTRACE_UPRIV) ? DTRACEFLT_UPRIV :		\
538 	((flags) & CPU_DTRACE_TUPOFLOW) ?  DTRACEFLT_TUPOFLOW :		\
539 	((flags) & CPU_DTRACE_BADALIGN) ?  DTRACEFLT_BADALIGN :		\
540 	((flags) & CPU_DTRACE_NOSCRATCH) ?  DTRACEFLT_NOSCRATCH :	\
541 	((flags) & CPU_DTRACE_BADSTACK) ?  DTRACEFLT_BADSTACK :		\
542 	DTRACEFLT_UNKNOWN)
543 
544 #define	DTRACEACT_ISSTRING(act)						\
545 	((act)->dta_kind == DTRACEACT_DIFEXPR &&			\
546 	(act)->dta_difo->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING)
547 
548 /* Function prototype definitions: */
549 static size_t dtrace_strlen(const char *, size_t);
550 static dtrace_probe_t *dtrace_probe_lookup_id(dtrace_id_t id);
551 static void dtrace_enabling_provide(dtrace_provider_t *);
552 static int dtrace_enabling_match(dtrace_enabling_t *, int *);
553 static void dtrace_enabling_matchall(void);
554 static void dtrace_enabling_reap(void);
555 static dtrace_state_t *dtrace_anon_grab(void);
556 static uint64_t dtrace_helper(int, dtrace_mstate_t *,
557     dtrace_state_t *, uint64_t, uint64_t);
558 static dtrace_helpers_t *dtrace_helpers_create(proc_t *);
559 static void dtrace_buffer_drop(dtrace_buffer_t *);
560 static int dtrace_buffer_consumed(dtrace_buffer_t *, hrtime_t when);
561 static intptr_t dtrace_buffer_reserve(dtrace_buffer_t *, size_t, size_t,
562     dtrace_state_t *, dtrace_mstate_t *);
563 static int dtrace_state_option(dtrace_state_t *, dtrace_optid_t,
564     dtrace_optval_t);
565 static int dtrace_ecb_create_enable(dtrace_probe_t *, void *);
566 static void dtrace_helper_provider_destroy(dtrace_helper_provider_t *);
567 uint16_t dtrace_load16(uintptr_t);
568 uint32_t dtrace_load32(uintptr_t);
569 uint64_t dtrace_load64(uintptr_t);
570 uint8_t dtrace_load8(uintptr_t);
571 void dtrace_dynvar_clean(dtrace_dstate_t *);
572 dtrace_dynvar_t *dtrace_dynvar(dtrace_dstate_t *, uint_t, dtrace_key_t *,
573     size_t, dtrace_dynvar_op_t, dtrace_mstate_t *, dtrace_vstate_t *);
574 uintptr_t dtrace_dif_varstr(uintptr_t, dtrace_state_t *, dtrace_mstate_t *);
575 
576 /*
577  * DTrace Probe Context Functions
578  *
579  * These functions are called from probe context.  Because probe context is
580  * any context in which C may be called, arbitrarily locks may be held,
581  * interrupts may be disabled, we may be in arbitrary dispatched state, etc.
582  * As a result, functions called from probe context may only call other DTrace
583  * support functions -- they may not interact at all with the system at large.
584  * (Note that the ASSERT macro is made probe-context safe by redefining it in
585  * terms of dtrace_assfail(), a probe-context safe function.) If arbitrary
586  * loads are to be performed from probe context, they _must_ be in terms of
587  * the safe dtrace_load*() variants.
588  *
589  * Some functions in this block are not actually called from probe context;
590  * for these functions, there will be a comment above the function reading
591  * "Note:  not called from probe context."
592  */
593 void
594 dtrace_panic(const char *format, ...)
595 {
596 	va_list alist;
597 
598 	va_start(alist, format);
599 	dtrace_vpanic(format, alist);
600 	va_end(alist);
601 }
602 
603 int
604 dtrace_assfail(const char *a, const char *f, int l)
605 {
606 	dtrace_panic("assertion failed: %s, file: %s, line: %d", a, f, l);
607 
608 	/*
609 	 * We just need something here that even the most clever compiler
610 	 * cannot optimize away.
611 	 */
612 	return (a[(uintptr_t)f]);
613 }
614 
615 /*
616  * Atomically increment a specified error counter from probe context.
617  */
618 static void
619 dtrace_error(uint32_t *counter)
620 {
621 	/*
622 	 * Most counters stored to in probe context are per-CPU counters.
623 	 * However, there are some error conditions that are sufficiently
624 	 * arcane that they don't merit per-CPU storage.  If these counters
625 	 * are incremented concurrently on different CPUs, scalability will be
626 	 * adversely affected -- but we don't expect them to be white-hot in a
627 	 * correctly constructed enabling...
628 	 */
629 	uint32_t oval, nval;
630 
631 	do {
632 		oval = *counter;
633 
634 		if ((nval = oval + 1) == 0) {
635 			/*
636 			 * If the counter would wrap, set it to 1 -- assuring
637 			 * that the counter is never zero when we have seen
638 			 * errors.  (The counter must be 32-bits because we
639 			 * aren't guaranteed a 64-bit compare&swap operation.)
640 			 * To save this code both the infamy of being fingered
641 			 * by a priggish news story and the indignity of being
642 			 * the target of a neo-puritan witch trial, we're
643 			 * carefully avoiding any colorful description of the
644 			 * likelihood of this condition -- but suffice it to
645 			 * say that it is only slightly more likely than the
646 			 * overflow of predicate cache IDs, as discussed in
647 			 * dtrace_predicate_create().
648 			 */
649 			nval = 1;
650 		}
651 	} while (dtrace_cas32(counter, oval, nval) != oval);
652 }
653 
654 /*
655  * Use the DTRACE_LOADFUNC macro to define functions for each of loading a
656  * uint8_t, a uint16_t, a uint32_t and a uint64_t.
657  */
658 DTRACE_LOADFUNC(8)
659 DTRACE_LOADFUNC(16)
660 DTRACE_LOADFUNC(32)
661 DTRACE_LOADFUNC(64)
662 
663 static int
664 dtrace_inscratch(uintptr_t dest, size_t size, dtrace_mstate_t *mstate)
665 {
666 	if (dest < mstate->dtms_scratch_base)
667 		return (0);
668 
669 	if (dest + size < dest)
670 		return (0);
671 
672 	if (dest + size > mstate->dtms_scratch_ptr)
673 		return (0);
674 
675 	return (1);
676 }
677 
678 static int
679 dtrace_canstore_statvar(uint64_t addr, size_t sz,
680     dtrace_statvar_t **svars, int nsvars)
681 {
682 	int i;
683 
684 	for (i = 0; i < nsvars; i++) {
685 		dtrace_statvar_t *svar = svars[i];
686 
687 		if (svar == NULL || svar->dtsv_size == 0)
688 			continue;
689 
690 		if (DTRACE_INRANGE(addr, sz, svar->dtsv_data, svar->dtsv_size))
691 			return (1);
692 	}
693 
694 	return (0);
695 }
696 
697 /*
698  * Check to see if the address is within a memory region to which a store may
699  * be issued.  This includes the DTrace scratch areas, and any DTrace variable
700  * region.  The caller of dtrace_canstore() is responsible for performing any
701  * alignment checks that are needed before stores are actually executed.
702  */
703 static int
704 dtrace_canstore(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
705     dtrace_vstate_t *vstate)
706 {
707 	/*
708 	 * First, check to see if the address is in scratch space...
709 	 */
710 	if (DTRACE_INRANGE(addr, sz, mstate->dtms_scratch_base,
711 	    mstate->dtms_scratch_size))
712 		return (1);
713 
714 	/*
715 	 * Now check to see if it's a dynamic variable.  This check will pick
716 	 * up both thread-local variables and any global dynamically-allocated
717 	 * variables.
718 	 */
719 	if (DTRACE_INRANGE(addr, sz, (uintptr_t)vstate->dtvs_dynvars.dtds_base,
720 	    vstate->dtvs_dynvars.dtds_size)) {
721 		dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
722 		uintptr_t base = (uintptr_t)dstate->dtds_base +
723 		    (dstate->dtds_hashsize * sizeof (dtrace_dynhash_t));
724 		uintptr_t chunkoffs;
725 
726 		/*
727 		 * Before we assume that we can store here, we need to make
728 		 * sure that it isn't in our metadata -- storing to our
729 		 * dynamic variable metadata would corrupt our state.  For
730 		 * the range to not include any dynamic variable metadata,
731 		 * it must:
732 		 *
733 		 *	(1) Start above the hash table that is at the base of
734 		 *	the dynamic variable space
735 		 *
736 		 *	(2) Have a starting chunk offset that is beyond the
737 		 *	dtrace_dynvar_t that is at the base of every chunk
738 		 *
739 		 *	(3) Not span a chunk boundary
740 		 *
741 		 */
742 		if (addr < base)
743 			return (0);
744 
745 		chunkoffs = (addr - base) % dstate->dtds_chunksize;
746 
747 		if (chunkoffs < sizeof (dtrace_dynvar_t))
748 			return (0);
749 
750 		if (chunkoffs + sz > dstate->dtds_chunksize)
751 			return (0);
752 
753 		return (1);
754 	}
755 
756 	/*
757 	 * Finally, check the static local and global variables.  These checks
758 	 * take the longest, so we perform them last.
759 	 */
760 	if (dtrace_canstore_statvar(addr, sz,
761 	    vstate->dtvs_locals, vstate->dtvs_nlocals))
762 		return (1);
763 
764 	if (dtrace_canstore_statvar(addr, sz,
765 	    vstate->dtvs_globals, vstate->dtvs_nglobals))
766 		return (1);
767 
768 	return (0);
769 }
770 
771 
772 /*
773  * Convenience routine to check to see if the address is within a memory
774  * region in which a load may be issued given the user's privilege level;
775  * if not, it sets the appropriate error flags and loads 'addr' into the
776  * illegal value slot.
777  *
778  * DTrace subroutines (DIF_SUBR_*) should use this helper to implement
779  * appropriate memory access protection.
780  */
781 static int
782 dtrace_canload(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
783     dtrace_vstate_t *vstate)
784 {
785 	volatile uintptr_t *illval = &cpu_core[curcpu].cpuc_dtrace_illval;
786 
787 	/*
788 	 * If we hold the privilege to read from kernel memory, then
789 	 * everything is readable.
790 	 */
791 	if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
792 		return (1);
793 
794 	/*
795 	 * You can obviously read that which you can store.
796 	 */
797 	if (dtrace_canstore(addr, sz, mstate, vstate))
798 		return (1);
799 
800 	/*
801 	 * We're allowed to read from our own string table.
802 	 */
803 	if (DTRACE_INRANGE(addr, sz, (uintptr_t)mstate->dtms_difo->dtdo_strtab,
804 	    mstate->dtms_difo->dtdo_strlen))
805 		return (1);
806 
807 	DTRACE_CPUFLAG_SET(CPU_DTRACE_KPRIV);
808 	*illval = addr;
809 	return (0);
810 }
811 
812 /*
813  * Convenience routine to check to see if a given string is within a memory
814  * region in which a load may be issued given the user's privilege level;
815  * this exists so that we don't need to issue unnecessary dtrace_strlen()
816  * calls in the event that the user has all privileges.
817  */
818 static int
819 dtrace_strcanload(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
820     dtrace_vstate_t *vstate)
821 {
822 	size_t strsz;
823 
824 	/*
825 	 * If we hold the privilege to read from kernel memory, then
826 	 * everything is readable.
827 	 */
828 	if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
829 		return (1);
830 
831 	strsz = 1 + dtrace_strlen((char *)(uintptr_t)addr, sz);
832 	if (dtrace_canload(addr, strsz, mstate, vstate))
833 		return (1);
834 
835 	return (0);
836 }
837 
838 /*
839  * Convenience routine to check to see if a given variable is within a memory
840  * region in which a load may be issued given the user's privilege level.
841  */
842 static int
843 dtrace_vcanload(void *src, dtrace_diftype_t *type, dtrace_mstate_t *mstate,
844     dtrace_vstate_t *vstate)
845 {
846 	size_t sz;
847 	ASSERT(type->dtdt_flags & DIF_TF_BYREF);
848 
849 	/*
850 	 * If we hold the privilege to read from kernel memory, then
851 	 * everything is readable.
852 	 */
853 	if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
854 		return (1);
855 
856 	if (type->dtdt_kind == DIF_TYPE_STRING)
857 		sz = dtrace_strlen(src,
858 		    vstate->dtvs_state->dts_options[DTRACEOPT_STRSIZE]) + 1;
859 	else
860 		sz = type->dtdt_size;
861 
862 	return (dtrace_canload((uintptr_t)src, sz, mstate, vstate));
863 }
864 
865 /*
866  * Compare two strings using safe loads.
867  */
868 static int
869 dtrace_strncmp(char *s1, char *s2, size_t limit)
870 {
871 	uint8_t c1, c2;
872 	volatile uint16_t *flags;
873 
874 	if (s1 == s2 || limit == 0)
875 		return (0);
876 
877 	flags = (volatile uint16_t *)&cpu_core[curcpu].cpuc_dtrace_flags;
878 
879 	do {
880 		if (s1 == NULL) {
881 			c1 = '\0';
882 		} else {
883 			c1 = dtrace_load8((uintptr_t)s1++);
884 		}
885 
886 		if (s2 == NULL) {
887 			c2 = '\0';
888 		} else {
889 			c2 = dtrace_load8((uintptr_t)s2++);
890 		}
891 
892 		if (c1 != c2)
893 			return (c1 - c2);
894 	} while (--limit && c1 != '\0' && !(*flags & CPU_DTRACE_FAULT));
895 
896 	return (0);
897 }
898 
899 /*
900  * Compute strlen(s) for a string using safe memory accesses.  The additional
901  * len parameter is used to specify a maximum length to ensure completion.
902  */
903 static size_t
904 dtrace_strlen(const char *s, size_t lim)
905 {
906 	uint_t len;
907 
908 	for (len = 0; len != lim; len++) {
909 		if (dtrace_load8((uintptr_t)s++) == '\0')
910 			break;
911 	}
912 
913 	return (len);
914 }
915 
916 /*
917  * Check if an address falls within a toxic region.
918  */
919 static int
920 dtrace_istoxic(uintptr_t kaddr, size_t size)
921 {
922 	uintptr_t taddr, tsize;
923 	int i;
924 
925 	for (i = 0; i < dtrace_toxranges; i++) {
926 		taddr = dtrace_toxrange[i].dtt_base;
927 		tsize = dtrace_toxrange[i].dtt_limit - taddr;
928 
929 		if (kaddr - taddr < tsize) {
930 			DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
931 			cpu_core[curcpu].cpuc_dtrace_illval = kaddr;
932 			return (1);
933 		}
934 
935 		if (taddr - kaddr < size) {
936 			DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
937 			cpu_core[curcpu].cpuc_dtrace_illval = taddr;
938 			return (1);
939 		}
940 	}
941 
942 	return (0);
943 }
944 
945 /*
946  * Copy src to dst using safe memory accesses.  The src is assumed to be unsafe
947  * memory specified by the DIF program.  The dst is assumed to be safe memory
948  * that we can store to directly because it is managed by DTrace.  As with
949  * standard bcopy, overlapping copies are handled properly.
950  */
951 static void
952 dtrace_bcopy(const void *src, void *dst, size_t len)
953 {
954 	if (len != 0) {
955 		uint8_t *s1 = dst;
956 		const uint8_t *s2 = src;
957 
958 		if (s1 <= s2) {
959 			do {
960 				*s1++ = dtrace_load8((uintptr_t)s2++);
961 			} while (--len != 0);
962 		} else {
963 			s2 += len;
964 			s1 += len;
965 
966 			do {
967 				*--s1 = dtrace_load8((uintptr_t)--s2);
968 			} while (--len != 0);
969 		}
970 	}
971 }
972 
973 /*
974  * Copy src to dst using safe memory accesses, up to either the specified
975  * length, or the point that a nul byte is encountered.  The src is assumed to
976  * be unsafe memory specified by the DIF program.  The dst is assumed to be
977  * safe memory that we can store to directly because it is managed by DTrace.
978  * Unlike dtrace_bcopy(), overlapping regions are not handled.
979  */
980 static void
981 dtrace_strcpy(const void *src, void *dst, size_t len)
982 {
983 	if (len != 0) {
984 		uint8_t *s1 = dst, c;
985 		const uint8_t *s2 = src;
986 
987 		do {
988 			*s1++ = c = dtrace_load8((uintptr_t)s2++);
989 		} while (--len != 0 && c != '\0');
990 	}
991 }
992 
993 /*
994  * Copy src to dst, deriving the size and type from the specified (BYREF)
995  * variable type.  The src is assumed to be unsafe memory specified by the DIF
996  * program.  The dst is assumed to be DTrace variable memory that is of the
997  * specified type; we assume that we can store to directly.
998  */
999 static void
1000 dtrace_vcopy(void *src, void *dst, dtrace_diftype_t *type)
1001 {
1002 	ASSERT(type->dtdt_flags & DIF_TF_BYREF);
1003 
1004 	if (type->dtdt_kind == DIF_TYPE_STRING) {
1005 		dtrace_strcpy(src, dst, type->dtdt_size);
1006 	} else {
1007 		dtrace_bcopy(src, dst, type->dtdt_size);
1008 	}
1009 }
1010 
1011 /*
1012  * Compare s1 to s2 using safe memory accesses.  The s1 data is assumed to be
1013  * unsafe memory specified by the DIF program.  The s2 data is assumed to be
1014  * safe memory that we can access directly because it is managed by DTrace.
1015  */
1016 static int
1017 dtrace_bcmp(const void *s1, const void *s2, size_t len)
1018 {
1019 	volatile uint16_t *flags;
1020 
1021 	flags = (volatile uint16_t *)&cpu_core[curcpu].cpuc_dtrace_flags;
1022 
1023 	if (s1 == s2)
1024 		return (0);
1025 
1026 	if (s1 == NULL || s2 == NULL)
1027 		return (1);
1028 
1029 	if (s1 != s2 && len != 0) {
1030 		const uint8_t *ps1 = s1;
1031 		const uint8_t *ps2 = s2;
1032 
1033 		do {
1034 			if (dtrace_load8((uintptr_t)ps1++) != *ps2++)
1035 				return (1);
1036 		} while (--len != 0 && !(*flags & CPU_DTRACE_FAULT));
1037 	}
1038 	return (0);
1039 }
1040 
1041 /*
1042  * Zero the specified region using a simple byte-by-byte loop.  Note that this
1043  * is for safe DTrace-managed memory only.
1044  */
1045 static void
1046 dtrace_bzero(void *dst, size_t len)
1047 {
1048 	uchar_t *cp;
1049 
1050 	for (cp = dst; len != 0; len--)
1051 		*cp++ = 0;
1052 }
1053 
1054 static void
1055 dtrace_add_128(uint64_t *addend1, uint64_t *addend2, uint64_t *sum)
1056 {
1057 	uint64_t result[2];
1058 
1059 	result[0] = addend1[0] + addend2[0];
1060 	result[1] = addend1[1] + addend2[1] +
1061 	    (result[0] < addend1[0] || result[0] < addend2[0] ? 1 : 0);
1062 
1063 	sum[0] = result[0];
1064 	sum[1] = result[1];
1065 }
1066 
1067 /*
1068  * Shift the 128-bit value in a by b. If b is positive, shift left.
1069  * If b is negative, shift right.
1070  */
1071 static void
1072 dtrace_shift_128(uint64_t *a, int b)
1073 {
1074 	uint64_t mask;
1075 
1076 	if (b == 0)
1077 		return;
1078 
1079 	if (b < 0) {
1080 		b = -b;
1081 		if (b >= 64) {
1082 			a[0] = a[1] >> (b - 64);
1083 			a[1] = 0;
1084 		} else {
1085 			a[0] >>= b;
1086 			mask = 1LL << (64 - b);
1087 			mask -= 1;
1088 			a[0] |= ((a[1] & mask) << (64 - b));
1089 			a[1] >>= b;
1090 		}
1091 	} else {
1092 		if (b >= 64) {
1093 			a[1] = a[0] << (b - 64);
1094 			a[0] = 0;
1095 		} else {
1096 			a[1] <<= b;
1097 			mask = a[0] >> (64 - b);
1098 			a[1] |= mask;
1099 			a[0] <<= b;
1100 		}
1101 	}
1102 }
1103 
1104 /*
1105  * The basic idea is to break the 2 64-bit values into 4 32-bit values,
1106  * use native multiplication on those, and then re-combine into the
1107  * resulting 128-bit value.
1108  *
1109  * (hi1 << 32 + lo1) * (hi2 << 32 + lo2) =
1110  *     hi1 * hi2 << 64 +
1111  *     hi1 * lo2 << 32 +
1112  *     hi2 * lo1 << 32 +
1113  *     lo1 * lo2
1114  */
1115 static void
1116 dtrace_multiply_128(uint64_t factor1, uint64_t factor2, uint64_t *product)
1117 {
1118 	uint64_t hi1, hi2, lo1, lo2;
1119 	uint64_t tmp[2];
1120 
1121 	hi1 = factor1 >> 32;
1122 	hi2 = factor2 >> 32;
1123 
1124 	lo1 = factor1 & DT_MASK_LO;
1125 	lo2 = factor2 & DT_MASK_LO;
1126 
1127 	product[0] = lo1 * lo2;
1128 	product[1] = hi1 * hi2;
1129 
1130 	tmp[0] = hi1 * lo2;
1131 	tmp[1] = 0;
1132 	dtrace_shift_128(tmp, 32);
1133 	dtrace_add_128(product, tmp, product);
1134 
1135 	tmp[0] = hi2 * lo1;
1136 	tmp[1] = 0;
1137 	dtrace_shift_128(tmp, 32);
1138 	dtrace_add_128(product, tmp, product);
1139 }
1140 
1141 /*
1142  * This privilege check should be used by actions and subroutines to
1143  * verify that the user credentials of the process that enabled the
1144  * invoking ECB match the target credentials
1145  */
1146 static int
1147 dtrace_priv_proc_common_user(dtrace_state_t *state)
1148 {
1149 	cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1150 
1151 	/*
1152 	 * We should always have a non-NULL state cred here, since if cred
1153 	 * is null (anonymous tracing), we fast-path bypass this routine.
1154 	 */
1155 	ASSERT(s_cr != NULL);
1156 
1157 	if ((cr = CRED()) != NULL &&
1158 	    s_cr->cr_uid == cr->cr_uid &&
1159 	    s_cr->cr_uid == cr->cr_ruid &&
1160 	    s_cr->cr_uid == cr->cr_suid &&
1161 	    s_cr->cr_gid == cr->cr_gid &&
1162 	    s_cr->cr_gid == cr->cr_rgid &&
1163 	    s_cr->cr_gid == cr->cr_sgid)
1164 		return (1);
1165 
1166 	return (0);
1167 }
1168 
1169 /*
1170  * This privilege check should be used by actions and subroutines to
1171  * verify that the zone of the process that enabled the invoking ECB
1172  * matches the target credentials
1173  */
1174 static int
1175 dtrace_priv_proc_common_zone(dtrace_state_t *state)
1176 {
1177 #if defined(sun)
1178 	cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1179 
1180 	/*
1181 	 * We should always have a non-NULL state cred here, since if cred
1182 	 * is null (anonymous tracing), we fast-path bypass this routine.
1183 	 */
1184 	ASSERT(s_cr != NULL);
1185 
1186 	if ((cr = CRED()) != NULL &&
1187 	    s_cr->cr_zone == cr->cr_zone)
1188 		return (1);
1189 
1190 	return (0);
1191 #else
1192 	return (1);
1193 #endif
1194 }
1195 
1196 /*
1197  * This privilege check should be used by actions and subroutines to
1198  * verify that the process has not setuid or changed credentials.
1199  */
1200 static int
1201 dtrace_priv_proc_common_nocd(void)
1202 {
1203 	proc_t *proc;
1204 
1205 	if ((proc = ttoproc(curthread)) != NULL &&
1206 	    !(proc->p_flag & SNOCD))
1207 		return (1);
1208 
1209 	return (0);
1210 }
1211 
1212 static int
1213 dtrace_priv_proc_destructive(dtrace_state_t *state)
1214 {
1215 	int action = state->dts_cred.dcr_action;
1216 
1217 	if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE) == 0) &&
1218 	    dtrace_priv_proc_common_zone(state) == 0)
1219 		goto bad;
1220 
1221 	if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER) == 0) &&
1222 	    dtrace_priv_proc_common_user(state) == 0)
1223 		goto bad;
1224 
1225 	if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG) == 0) &&
1226 	    dtrace_priv_proc_common_nocd() == 0)
1227 		goto bad;
1228 
1229 	return (1);
1230 
1231 bad:
1232 	cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1233 
1234 	return (0);
1235 }
1236 
1237 static int
1238 dtrace_priv_proc_control(dtrace_state_t *state)
1239 {
1240 	if (state->dts_cred.dcr_action & DTRACE_CRA_PROC_CONTROL)
1241 		return (1);
1242 
1243 	if (dtrace_priv_proc_common_zone(state) &&
1244 	    dtrace_priv_proc_common_user(state) &&
1245 	    dtrace_priv_proc_common_nocd())
1246 		return (1);
1247 
1248 	cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1249 
1250 	return (0);
1251 }
1252 
1253 static int
1254 dtrace_priv_proc(dtrace_state_t *state)
1255 {
1256 	if (state->dts_cred.dcr_action & DTRACE_CRA_PROC)
1257 		return (1);
1258 
1259 	cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1260 
1261 	return (0);
1262 }
1263 
1264 static int
1265 dtrace_priv_kernel(dtrace_state_t *state)
1266 {
1267 	if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL)
1268 		return (1);
1269 
1270 	cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1271 
1272 	return (0);
1273 }
1274 
1275 static int
1276 dtrace_priv_kernel_destructive(dtrace_state_t *state)
1277 {
1278 	if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL_DESTRUCTIVE)
1279 		return (1);
1280 
1281 	cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1282 
1283 	return (0);
1284 }
1285 
1286 /*
1287  * Note:  not called from probe context.  This function is called
1288  * asynchronously (and at a regular interval) from outside of probe context to
1289  * clean the dirty dynamic variable lists on all CPUs.  Dynamic variable
1290  * cleaning is explained in detail in <sys/dtrace_impl.h>.
1291  */
1292 void
1293 dtrace_dynvar_clean(dtrace_dstate_t *dstate)
1294 {
1295 	dtrace_dynvar_t *dirty;
1296 	dtrace_dstate_percpu_t *dcpu;
1297 	int i, work = 0;
1298 
1299 	for (i = 0; i < NCPU; i++) {
1300 		dcpu = &dstate->dtds_percpu[i];
1301 
1302 		ASSERT(dcpu->dtdsc_rinsing == NULL);
1303 
1304 		/*
1305 		 * If the dirty list is NULL, there is no dirty work to do.
1306 		 */
1307 		if (dcpu->dtdsc_dirty == NULL)
1308 			continue;
1309 
1310 		/*
1311 		 * If the clean list is non-NULL, then we're not going to do
1312 		 * any work for this CPU -- it means that there has not been
1313 		 * a dtrace_dynvar() allocation on this CPU (or from this CPU)
1314 		 * since the last time we cleaned house.
1315 		 */
1316 		if (dcpu->dtdsc_clean != NULL)
1317 			continue;
1318 
1319 		work = 1;
1320 
1321 		/*
1322 		 * Atomically move the dirty list aside.
1323 		 */
1324 		do {
1325 			dirty = dcpu->dtdsc_dirty;
1326 
1327 			/*
1328 			 * Before we zap the dirty list, set the rinsing list.
1329 			 * (This allows for a potential assertion in
1330 			 * dtrace_dynvar():  if a free dynamic variable appears
1331 			 * on a hash chain, either the dirty list or the
1332 			 * rinsing list for some CPU must be non-NULL.)
1333 			 */
1334 			dcpu->dtdsc_rinsing = dirty;
1335 			dtrace_membar_producer();
1336 		} while (dtrace_casptr(&dcpu->dtdsc_dirty,
1337 		    dirty, NULL) != dirty);
1338 	}
1339 
1340 	if (!work) {
1341 		/*
1342 		 * We have no work to do; we can simply return.
1343 		 */
1344 		return;
1345 	}
1346 
1347 	dtrace_sync();
1348 
1349 	for (i = 0; i < NCPU; i++) {
1350 		dcpu = &dstate->dtds_percpu[i];
1351 
1352 		if (dcpu->dtdsc_rinsing == NULL)
1353 			continue;
1354 
1355 		/*
1356 		 * We are now guaranteed that no hash chain contains a pointer
1357 		 * into this dirty list; we can make it clean.
1358 		 */
1359 		ASSERT(dcpu->dtdsc_clean == NULL);
1360 		dcpu->dtdsc_clean = dcpu->dtdsc_rinsing;
1361 		dcpu->dtdsc_rinsing = NULL;
1362 	}
1363 
1364 	/*
1365 	 * Before we actually set the state to be DTRACE_DSTATE_CLEAN, make
1366 	 * sure that all CPUs have seen all of the dtdsc_clean pointers.
1367 	 * This prevents a race whereby a CPU incorrectly decides that
1368 	 * the state should be something other than DTRACE_DSTATE_CLEAN
1369 	 * after dtrace_dynvar_clean() has completed.
1370 	 */
1371 	dtrace_sync();
1372 
1373 	dstate->dtds_state = DTRACE_DSTATE_CLEAN;
1374 }
1375 
1376 /*
1377  * Depending on the value of the op parameter, this function looks-up,
1378  * allocates or deallocates an arbitrarily-keyed dynamic variable.  If an
1379  * allocation is requested, this function will return a pointer to a
1380  * dtrace_dynvar_t corresponding to the allocated variable -- or NULL if no
1381  * variable can be allocated.  If NULL is returned, the appropriate counter
1382  * will be incremented.
1383  */
1384 dtrace_dynvar_t *
1385 dtrace_dynvar(dtrace_dstate_t *dstate, uint_t nkeys,
1386     dtrace_key_t *key, size_t dsize, dtrace_dynvar_op_t op,
1387     dtrace_mstate_t *mstate, dtrace_vstate_t *vstate)
1388 {
1389 	uint64_t hashval = DTRACE_DYNHASH_VALID;
1390 	dtrace_dynhash_t *hash = dstate->dtds_hash;
1391 	dtrace_dynvar_t *free, *new_free, *next, *dvar, *start, *prev = NULL;
1392 	processorid_t me = curcpu, cpu = me;
1393 	dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[me];
1394 	size_t bucket, ksize;
1395 	size_t chunksize = dstate->dtds_chunksize;
1396 	uintptr_t kdata, lock, nstate;
1397 	uint_t i;
1398 
1399 	ASSERT(nkeys != 0);
1400 
1401 	/*
1402 	 * Hash the key.  As with aggregations, we use Jenkins' "One-at-a-time"
1403 	 * algorithm.  For the by-value portions, we perform the algorithm in
1404 	 * 16-bit chunks (as opposed to 8-bit chunks).  This speeds things up a
1405 	 * bit, and seems to have only a minute effect on distribution.  For
1406 	 * the by-reference data, we perform "One-at-a-time" iterating (safely)
1407 	 * over each referenced byte.  It's painful to do this, but it's much
1408 	 * better than pathological hash distribution.  The efficacy of the
1409 	 * hashing algorithm (and a comparison with other algorithms) may be
1410 	 * found by running the ::dtrace_dynstat MDB dcmd.
1411 	 */
1412 	for (i = 0; i < nkeys; i++) {
1413 		if (key[i].dttk_size == 0) {
1414 			uint64_t val = key[i].dttk_value;
1415 
1416 			hashval += (val >> 48) & 0xffff;
1417 			hashval += (hashval << 10);
1418 			hashval ^= (hashval >> 6);
1419 
1420 			hashval += (val >> 32) & 0xffff;
1421 			hashval += (hashval << 10);
1422 			hashval ^= (hashval >> 6);
1423 
1424 			hashval += (val >> 16) & 0xffff;
1425 			hashval += (hashval << 10);
1426 			hashval ^= (hashval >> 6);
1427 
1428 			hashval += val & 0xffff;
1429 			hashval += (hashval << 10);
1430 			hashval ^= (hashval >> 6);
1431 		} else {
1432 			/*
1433 			 * This is incredibly painful, but it beats the hell
1434 			 * out of the alternative.
1435 			 */
1436 			uint64_t j, size = key[i].dttk_size;
1437 			uintptr_t base = (uintptr_t)key[i].dttk_value;
1438 
1439 			if (!dtrace_canload(base, size, mstate, vstate))
1440 				break;
1441 
1442 			for (j = 0; j < size; j++) {
1443 				hashval += dtrace_load8(base + j);
1444 				hashval += (hashval << 10);
1445 				hashval ^= (hashval >> 6);
1446 			}
1447 		}
1448 	}
1449 
1450 	if (DTRACE_CPUFLAG_ISSET(CPU_DTRACE_FAULT))
1451 		return (NULL);
1452 
1453 	hashval += (hashval << 3);
1454 	hashval ^= (hashval >> 11);
1455 	hashval += (hashval << 15);
1456 
1457 	/*
1458 	 * There is a remote chance (ideally, 1 in 2^31) that our hashval
1459 	 * comes out to be one of our two sentinel hash values.  If this
1460 	 * actually happens, we set the hashval to be a value known to be a
1461 	 * non-sentinel value.
1462 	 */
1463 	if (hashval == DTRACE_DYNHASH_FREE || hashval == DTRACE_DYNHASH_SINK)
1464 		hashval = DTRACE_DYNHASH_VALID;
1465 
1466 	/*
1467 	 * Yes, it's painful to do a divide here.  If the cycle count becomes
1468 	 * important here, tricks can be pulled to reduce it.  (However, it's
1469 	 * critical that hash collisions be kept to an absolute minimum;
1470 	 * they're much more painful than a divide.)  It's better to have a
1471 	 * solution that generates few collisions and still keeps things
1472 	 * relatively simple.
1473 	 */
1474 	bucket = hashval % dstate->dtds_hashsize;
1475 
1476 	if (op == DTRACE_DYNVAR_DEALLOC) {
1477 		volatile uintptr_t *lockp = &hash[bucket].dtdh_lock;
1478 
1479 		for (;;) {
1480 			while ((lock = *lockp) & 1)
1481 				continue;
1482 
1483 			if (dtrace_casptr((volatile void *)lockp,
1484 			    (volatile void *)lock, (volatile void *)(lock + 1)) == (void *)lock)
1485 				break;
1486 		}
1487 
1488 		dtrace_membar_producer();
1489 	}
1490 
1491 top:
1492 	prev = NULL;
1493 	lock = hash[bucket].dtdh_lock;
1494 
1495 	dtrace_membar_consumer();
1496 
1497 	start = hash[bucket].dtdh_chain;
1498 	ASSERT(start != NULL && (start->dtdv_hashval == DTRACE_DYNHASH_SINK ||
1499 	    start->dtdv_hashval != DTRACE_DYNHASH_FREE ||
1500 	    op != DTRACE_DYNVAR_DEALLOC));
1501 
1502 	for (dvar = start; dvar != NULL; dvar = dvar->dtdv_next) {
1503 		dtrace_tuple_t *dtuple = &dvar->dtdv_tuple;
1504 		dtrace_key_t *dkey = &dtuple->dtt_key[0];
1505 
1506 		if (dvar->dtdv_hashval != hashval) {
1507 			if (dvar->dtdv_hashval == DTRACE_DYNHASH_SINK) {
1508 				/*
1509 				 * We've reached the sink, and therefore the
1510 				 * end of the hash chain; we can kick out of
1511 				 * the loop knowing that we have seen a valid
1512 				 * snapshot of state.
1513 				 */
1514 				ASSERT(dvar->dtdv_next == NULL);
1515 				ASSERT(dvar == &dtrace_dynhash_sink);
1516 				break;
1517 			}
1518 
1519 			if (dvar->dtdv_hashval == DTRACE_DYNHASH_FREE) {
1520 				/*
1521 				 * We've gone off the rails:  somewhere along
1522 				 * the line, one of the members of this hash
1523 				 * chain was deleted.  Note that we could also
1524 				 * detect this by simply letting this loop run
1525 				 * to completion, as we would eventually hit
1526 				 * the end of the dirty list.  However, we
1527 				 * want to avoid running the length of the
1528 				 * dirty list unnecessarily (it might be quite
1529 				 * long), so we catch this as early as
1530 				 * possible by detecting the hash marker.  In
1531 				 * this case, we simply set dvar to NULL and
1532 				 * break; the conditional after the loop will
1533 				 * send us back to top.
1534 				 */
1535 				dvar = NULL;
1536 				break;
1537 			}
1538 
1539 			goto next;
1540 		}
1541 
1542 		if (dtuple->dtt_nkeys != nkeys)
1543 			goto next;
1544 
1545 		for (i = 0; i < nkeys; i++, dkey++) {
1546 			if (dkey->dttk_size != key[i].dttk_size)
1547 				goto next; /* size or type mismatch */
1548 
1549 			if (dkey->dttk_size != 0) {
1550 				if (dtrace_bcmp(
1551 				    (void *)(uintptr_t)key[i].dttk_value,
1552 				    (void *)(uintptr_t)dkey->dttk_value,
1553 				    dkey->dttk_size))
1554 					goto next;
1555 			} else {
1556 				if (dkey->dttk_value != key[i].dttk_value)
1557 					goto next;
1558 			}
1559 		}
1560 
1561 		if (op != DTRACE_DYNVAR_DEALLOC)
1562 			return (dvar);
1563 
1564 		ASSERT(dvar->dtdv_next == NULL ||
1565 		    dvar->dtdv_next->dtdv_hashval != DTRACE_DYNHASH_FREE);
1566 
1567 		if (prev != NULL) {
1568 			ASSERT(hash[bucket].dtdh_chain != dvar);
1569 			ASSERT(start != dvar);
1570 			ASSERT(prev->dtdv_next == dvar);
1571 			prev->dtdv_next = dvar->dtdv_next;
1572 		} else {
1573 			if (dtrace_casptr(&hash[bucket].dtdh_chain,
1574 			    start, dvar->dtdv_next) != start) {
1575 				/*
1576 				 * We have failed to atomically swing the
1577 				 * hash table head pointer, presumably because
1578 				 * of a conflicting allocation on another CPU.
1579 				 * We need to reread the hash chain and try
1580 				 * again.
1581 				 */
1582 				goto top;
1583 			}
1584 		}
1585 
1586 		dtrace_membar_producer();
1587 
1588 		/*
1589 		 * Now set the hash value to indicate that it's free.
1590 		 */
1591 		ASSERT(hash[bucket].dtdh_chain != dvar);
1592 		dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
1593 
1594 		dtrace_membar_producer();
1595 
1596 		/*
1597 		 * Set the next pointer to point at the dirty list, and
1598 		 * atomically swing the dirty pointer to the newly freed dvar.
1599 		 */
1600 		do {
1601 			next = dcpu->dtdsc_dirty;
1602 			dvar->dtdv_next = next;
1603 		} while (dtrace_casptr(&dcpu->dtdsc_dirty, next, dvar) != next);
1604 
1605 		/*
1606 		 * Finally, unlock this hash bucket.
1607 		 */
1608 		ASSERT(hash[bucket].dtdh_lock == lock);
1609 		ASSERT(lock & 1);
1610 		hash[bucket].dtdh_lock++;
1611 
1612 		return (NULL);
1613 next:
1614 		prev = dvar;
1615 		continue;
1616 	}
1617 
1618 	if (dvar == NULL) {
1619 		/*
1620 		 * If dvar is NULL, it is because we went off the rails:
1621 		 * one of the elements that we traversed in the hash chain
1622 		 * was deleted while we were traversing it.  In this case,
1623 		 * we assert that we aren't doing a dealloc (deallocs lock
1624 		 * the hash bucket to prevent themselves from racing with
1625 		 * one another), and retry the hash chain traversal.
1626 		 */
1627 		ASSERT(op != DTRACE_DYNVAR_DEALLOC);
1628 		goto top;
1629 	}
1630 
1631 	if (op != DTRACE_DYNVAR_ALLOC) {
1632 		/*
1633 		 * If we are not to allocate a new variable, we want to
1634 		 * return NULL now.  Before we return, check that the value
1635 		 * of the lock word hasn't changed.  If it has, we may have
1636 		 * seen an inconsistent snapshot.
1637 		 */
1638 		if (op == DTRACE_DYNVAR_NOALLOC) {
1639 			if (hash[bucket].dtdh_lock != lock)
1640 				goto top;
1641 		} else {
1642 			ASSERT(op == DTRACE_DYNVAR_DEALLOC);
1643 			ASSERT(hash[bucket].dtdh_lock == lock);
1644 			ASSERT(lock & 1);
1645 			hash[bucket].dtdh_lock++;
1646 		}
1647 
1648 		return (NULL);
1649 	}
1650 
1651 	/*
1652 	 * We need to allocate a new dynamic variable.  The size we need is the
1653 	 * size of dtrace_dynvar plus the size of nkeys dtrace_key_t's plus the
1654 	 * size of any auxiliary key data (rounded up to 8-byte alignment) plus
1655 	 * the size of any referred-to data (dsize).  We then round the final
1656 	 * size up to the chunksize for allocation.
1657 	 */
1658 	for (ksize = 0, i = 0; i < nkeys; i++)
1659 		ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
1660 
1661 	/*
1662 	 * This should be pretty much impossible, but could happen if, say,
1663 	 * strange DIF specified the tuple.  Ideally, this should be an
1664 	 * assertion and not an error condition -- but that requires that the
1665 	 * chunksize calculation in dtrace_difo_chunksize() be absolutely
1666 	 * bullet-proof.  (That is, it must not be able to be fooled by
1667 	 * malicious DIF.)  Given the lack of backwards branches in DIF,
1668 	 * solving this would presumably not amount to solving the Halting
1669 	 * Problem -- but it still seems awfully hard.
1670 	 */
1671 	if (sizeof (dtrace_dynvar_t) + sizeof (dtrace_key_t) * (nkeys - 1) +
1672 	    ksize + dsize > chunksize) {
1673 		dcpu->dtdsc_drops++;
1674 		return (NULL);
1675 	}
1676 
1677 	nstate = DTRACE_DSTATE_EMPTY;
1678 
1679 	do {
1680 retry:
1681 		free = dcpu->dtdsc_free;
1682 
1683 		if (free == NULL) {
1684 			dtrace_dynvar_t *clean = dcpu->dtdsc_clean;
1685 			void *rval;
1686 
1687 			if (clean == NULL) {
1688 				/*
1689 				 * We're out of dynamic variable space on
1690 				 * this CPU.  Unless we have tried all CPUs,
1691 				 * we'll try to allocate from a different
1692 				 * CPU.
1693 				 */
1694 				switch (dstate->dtds_state) {
1695 				case DTRACE_DSTATE_CLEAN: {
1696 					void *sp = &dstate->dtds_state;
1697 
1698 					if (++cpu >= NCPU)
1699 						cpu = 0;
1700 
1701 					if (dcpu->dtdsc_dirty != NULL &&
1702 					    nstate == DTRACE_DSTATE_EMPTY)
1703 						nstate = DTRACE_DSTATE_DIRTY;
1704 
1705 					if (dcpu->dtdsc_rinsing != NULL)
1706 						nstate = DTRACE_DSTATE_RINSING;
1707 
1708 					dcpu = &dstate->dtds_percpu[cpu];
1709 
1710 					if (cpu != me)
1711 						goto retry;
1712 
1713 					(void) dtrace_cas32(sp,
1714 					    DTRACE_DSTATE_CLEAN, nstate);
1715 
1716 					/*
1717 					 * To increment the correct bean
1718 					 * counter, take another lap.
1719 					 */
1720 					goto retry;
1721 				}
1722 
1723 				case DTRACE_DSTATE_DIRTY:
1724 					dcpu->dtdsc_dirty_drops++;
1725 					break;
1726 
1727 				case DTRACE_DSTATE_RINSING:
1728 					dcpu->dtdsc_rinsing_drops++;
1729 					break;
1730 
1731 				case DTRACE_DSTATE_EMPTY:
1732 					dcpu->dtdsc_drops++;
1733 					break;
1734 				}
1735 
1736 				DTRACE_CPUFLAG_SET(CPU_DTRACE_DROP);
1737 				return (NULL);
1738 			}
1739 
1740 			/*
1741 			 * The clean list appears to be non-empty.  We want to
1742 			 * move the clean list to the free list; we start by
1743 			 * moving the clean pointer aside.
1744 			 */
1745 			if (dtrace_casptr(&dcpu->dtdsc_clean,
1746 			    clean, NULL) != clean) {
1747 				/*
1748 				 * We are in one of two situations:
1749 				 *
1750 				 *  (a)	The clean list was switched to the
1751 				 *	free list by another CPU.
1752 				 *
1753 				 *  (b)	The clean list was added to by the
1754 				 *	cleansing cyclic.
1755 				 *
1756 				 * In either of these situations, we can
1757 				 * just reattempt the free list allocation.
1758 				 */
1759 				goto retry;
1760 			}
1761 
1762 			ASSERT(clean->dtdv_hashval == DTRACE_DYNHASH_FREE);
1763 
1764 			/*
1765 			 * Now we'll move the clean list to the free list.
1766 			 * It's impossible for this to fail:  the only way
1767 			 * the free list can be updated is through this
1768 			 * code path, and only one CPU can own the clean list.
1769 			 * Thus, it would only be possible for this to fail if
1770 			 * this code were racing with dtrace_dynvar_clean().
1771 			 * (That is, if dtrace_dynvar_clean() updated the clean
1772 			 * list, and we ended up racing to update the free
1773 			 * list.)  This race is prevented by the dtrace_sync()
1774 			 * in dtrace_dynvar_clean() -- which flushes the
1775 			 * owners of the clean lists out before resetting
1776 			 * the clean lists.
1777 			 */
1778 			rval = dtrace_casptr(&dcpu->dtdsc_free, NULL, clean);
1779 			ASSERT(rval == NULL);
1780 			goto retry;
1781 		}
1782 
1783 		dvar = free;
1784 		new_free = dvar->dtdv_next;
1785 	} while (dtrace_casptr(&dcpu->dtdsc_free, free, new_free) != free);
1786 
1787 	/*
1788 	 * We have now allocated a new chunk.  We copy the tuple keys into the
1789 	 * tuple array and copy any referenced key data into the data space
1790 	 * following the tuple array.  As we do this, we relocate dttk_value
1791 	 * in the final tuple to point to the key data address in the chunk.
1792 	 */
1793 	kdata = (uintptr_t)&dvar->dtdv_tuple.dtt_key[nkeys];
1794 	dvar->dtdv_data = (void *)(kdata + ksize);
1795 	dvar->dtdv_tuple.dtt_nkeys = nkeys;
1796 
1797 	for (i = 0; i < nkeys; i++) {
1798 		dtrace_key_t *dkey = &dvar->dtdv_tuple.dtt_key[i];
1799 		size_t kesize = key[i].dttk_size;
1800 
1801 		if (kesize != 0) {
1802 			dtrace_bcopy(
1803 			    (const void *)(uintptr_t)key[i].dttk_value,
1804 			    (void *)kdata, kesize);
1805 			dkey->dttk_value = kdata;
1806 			kdata += P2ROUNDUP(kesize, sizeof (uint64_t));
1807 		} else {
1808 			dkey->dttk_value = key[i].dttk_value;
1809 		}
1810 
1811 		dkey->dttk_size = kesize;
1812 	}
1813 
1814 	ASSERT(dvar->dtdv_hashval == DTRACE_DYNHASH_FREE);
1815 	dvar->dtdv_hashval = hashval;
1816 	dvar->dtdv_next = start;
1817 
1818 	if (dtrace_casptr(&hash[bucket].dtdh_chain, start, dvar) == start)
1819 		return (dvar);
1820 
1821 	/*
1822 	 * The cas has failed.  Either another CPU is adding an element to
1823 	 * this hash chain, or another CPU is deleting an element from this
1824 	 * hash chain.  The simplest way to deal with both of these cases
1825 	 * (though not necessarily the most efficient) is to free our
1826 	 * allocated block and tail-call ourselves.  Note that the free is
1827 	 * to the dirty list and _not_ to the free list.  This is to prevent
1828 	 * races with allocators, above.
1829 	 */
1830 	dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
1831 
1832 	dtrace_membar_producer();
1833 
1834 	do {
1835 		free = dcpu->dtdsc_dirty;
1836 		dvar->dtdv_next = free;
1837 	} while (dtrace_casptr(&dcpu->dtdsc_dirty, free, dvar) != free);
1838 
1839 	return (dtrace_dynvar(dstate, nkeys, key, dsize, op, mstate, vstate));
1840 }
1841 
1842 /*ARGSUSED*/
1843 static void
1844 dtrace_aggregate_min(uint64_t *oval, uint64_t nval, uint64_t arg)
1845 {
1846 	if ((int64_t)nval < (int64_t)*oval)
1847 		*oval = nval;
1848 }
1849 
1850 /*ARGSUSED*/
1851 static void
1852 dtrace_aggregate_max(uint64_t *oval, uint64_t nval, uint64_t arg)
1853 {
1854 	if ((int64_t)nval > (int64_t)*oval)
1855 		*oval = nval;
1856 }
1857 
1858 static void
1859 dtrace_aggregate_quantize(uint64_t *quanta, uint64_t nval, uint64_t incr)
1860 {
1861 	int i, zero = DTRACE_QUANTIZE_ZEROBUCKET;
1862 	int64_t val = (int64_t)nval;
1863 
1864 	if (val < 0) {
1865 		for (i = 0; i < zero; i++) {
1866 			if (val <= DTRACE_QUANTIZE_BUCKETVAL(i)) {
1867 				quanta[i] += incr;
1868 				return;
1869 			}
1870 		}
1871 	} else {
1872 		for (i = zero + 1; i < DTRACE_QUANTIZE_NBUCKETS; i++) {
1873 			if (val < DTRACE_QUANTIZE_BUCKETVAL(i)) {
1874 				quanta[i - 1] += incr;
1875 				return;
1876 			}
1877 		}
1878 
1879 		quanta[DTRACE_QUANTIZE_NBUCKETS - 1] += incr;
1880 		return;
1881 	}
1882 
1883 	ASSERT(0);
1884 }
1885 
1886 static void
1887 dtrace_aggregate_lquantize(uint64_t *lquanta, uint64_t nval, uint64_t incr)
1888 {
1889 	uint64_t arg = *lquanta++;
1890 	int32_t base = DTRACE_LQUANTIZE_BASE(arg);
1891 	uint16_t step = DTRACE_LQUANTIZE_STEP(arg);
1892 	uint16_t levels = DTRACE_LQUANTIZE_LEVELS(arg);
1893 	int32_t val = (int32_t)nval, level;
1894 
1895 	ASSERT(step != 0);
1896 	ASSERT(levels != 0);
1897 
1898 	if (val < base) {
1899 		/*
1900 		 * This is an underflow.
1901 		 */
1902 		lquanta[0] += incr;
1903 		return;
1904 	}
1905 
1906 	level = (val - base) / step;
1907 
1908 	if (level < levels) {
1909 		lquanta[level + 1] += incr;
1910 		return;
1911 	}
1912 
1913 	/*
1914 	 * This is an overflow.
1915 	 */
1916 	lquanta[levels + 1] += incr;
1917 }
1918 
1919 static int
1920 dtrace_aggregate_llquantize_bucket(uint16_t factor, uint16_t low,
1921     uint16_t high, uint16_t nsteps, int64_t value)
1922 {
1923 	int64_t this = 1, last, next;
1924 	int base = 1, order;
1925 
1926 	ASSERT(factor <= nsteps);
1927 	ASSERT(nsteps % factor == 0);
1928 
1929 	for (order = 0; order < low; order++)
1930 		this *= factor;
1931 
1932 	/*
1933 	 * If our value is less than our factor taken to the power of the
1934 	 * low order of magnitude, it goes into the zeroth bucket.
1935 	 */
1936 	if (value < (last = this))
1937 		return (0);
1938 
1939 	for (this *= factor; order <= high; order++) {
1940 		int nbuckets = this > nsteps ? nsteps : this;
1941 
1942 		if ((next = this * factor) < this) {
1943 			/*
1944 			 * We should not generally get log/linear quantizations
1945 			 * with a high magnitude that allows 64-bits to
1946 			 * overflow, but we nonetheless protect against this
1947 			 * by explicitly checking for overflow, and clamping
1948 			 * our value accordingly.
1949 			 */
1950 			value = this - 1;
1951 		}
1952 
1953 		if (value < this) {
1954 			/*
1955 			 * If our value lies within this order of magnitude,
1956 			 * determine its position by taking the offset within
1957 			 * the order of magnitude, dividing by the bucket
1958 			 * width, and adding to our (accumulated) base.
1959 			 */
1960 			return (base + (value - last) / (this / nbuckets));
1961 		}
1962 
1963 		base += nbuckets - (nbuckets / factor);
1964 		last = this;
1965 		this = next;
1966 	}
1967 
1968 	/*
1969 	 * Our value is greater than or equal to our factor taken to the
1970 	 * power of one plus the high magnitude -- return the top bucket.
1971 	 */
1972 	return (base);
1973 }
1974 
1975 static void
1976 dtrace_aggregate_llquantize(uint64_t *llquanta, uint64_t nval, uint64_t incr)
1977 {
1978 	uint64_t arg = *llquanta++;
1979 	uint16_t factor = DTRACE_LLQUANTIZE_FACTOR(arg);
1980 	uint16_t low = DTRACE_LLQUANTIZE_LOW(arg);
1981 	uint16_t high = DTRACE_LLQUANTIZE_HIGH(arg);
1982 	uint16_t nsteps = DTRACE_LLQUANTIZE_NSTEP(arg);
1983 
1984 	llquanta[dtrace_aggregate_llquantize_bucket(factor,
1985 	    low, high, nsteps, nval)] += incr;
1986 }
1987 
1988 /*ARGSUSED*/
1989 static void
1990 dtrace_aggregate_avg(uint64_t *data, uint64_t nval, uint64_t arg)
1991 {
1992 	data[0]++;
1993 	data[1] += nval;
1994 }
1995 
1996 /*ARGSUSED*/
1997 static void
1998 dtrace_aggregate_stddev(uint64_t *data, uint64_t nval, uint64_t arg)
1999 {
2000 	int64_t snval = (int64_t)nval;
2001 	uint64_t tmp[2];
2002 
2003 	data[0]++;
2004 	data[1] += nval;
2005 
2006 	/*
2007 	 * What we want to say here is:
2008 	 *
2009 	 * data[2] += nval * nval;
2010 	 *
2011 	 * But given that nval is 64-bit, we could easily overflow, so
2012 	 * we do this as 128-bit arithmetic.
2013 	 */
2014 	if (snval < 0)
2015 		snval = -snval;
2016 
2017 	dtrace_multiply_128((uint64_t)snval, (uint64_t)snval, tmp);
2018 	dtrace_add_128(data + 2, tmp, data + 2);
2019 }
2020 
2021 /*ARGSUSED*/
2022 static void
2023 dtrace_aggregate_count(uint64_t *oval, uint64_t nval, uint64_t arg)
2024 {
2025 	*oval = *oval + 1;
2026 }
2027 
2028 /*ARGSUSED*/
2029 static void
2030 dtrace_aggregate_sum(uint64_t *oval, uint64_t nval, uint64_t arg)
2031 {
2032 	*oval += nval;
2033 }
2034 
2035 /*
2036  * Aggregate given the tuple in the principal data buffer, and the aggregating
2037  * action denoted by the specified dtrace_aggregation_t.  The aggregation
2038  * buffer is specified as the buf parameter.  This routine does not return
2039  * failure; if there is no space in the aggregation buffer, the data will be
2040  * dropped, and a corresponding counter incremented.
2041  */
2042 static void
2043 dtrace_aggregate(dtrace_aggregation_t *agg, dtrace_buffer_t *dbuf,
2044     intptr_t offset, dtrace_buffer_t *buf, uint64_t expr, uint64_t arg)
2045 {
2046 	dtrace_recdesc_t *rec = &agg->dtag_action.dta_rec;
2047 	uint32_t i, ndx, size, fsize;
2048 	uint32_t align = sizeof (uint64_t) - 1;
2049 	dtrace_aggbuffer_t *agb;
2050 	dtrace_aggkey_t *key;
2051 	uint32_t hashval = 0, limit, isstr;
2052 	caddr_t tomax, data, kdata;
2053 	dtrace_actkind_t action;
2054 	dtrace_action_t *act;
2055 	uintptr_t offs;
2056 
2057 	if (buf == NULL)
2058 		return;
2059 
2060 	if (!agg->dtag_hasarg) {
2061 		/*
2062 		 * Currently, only quantize() and lquantize() take additional
2063 		 * arguments, and they have the same semantics:  an increment
2064 		 * value that defaults to 1 when not present.  If additional
2065 		 * aggregating actions take arguments, the setting of the
2066 		 * default argument value will presumably have to become more
2067 		 * sophisticated...
2068 		 */
2069 		arg = 1;
2070 	}
2071 
2072 	action = agg->dtag_action.dta_kind - DTRACEACT_AGGREGATION;
2073 	size = rec->dtrd_offset - agg->dtag_base;
2074 	fsize = size + rec->dtrd_size;
2075 
2076 	ASSERT(dbuf->dtb_tomax != NULL);
2077 	data = dbuf->dtb_tomax + offset + agg->dtag_base;
2078 
2079 	if ((tomax = buf->dtb_tomax) == NULL) {
2080 		dtrace_buffer_drop(buf);
2081 		return;
2082 	}
2083 
2084 	/*
2085 	 * The metastructure is always at the bottom of the buffer.
2086 	 */
2087 	agb = (dtrace_aggbuffer_t *)(tomax + buf->dtb_size -
2088 	    sizeof (dtrace_aggbuffer_t));
2089 
2090 	if (buf->dtb_offset == 0) {
2091 		/*
2092 		 * We just kludge up approximately 1/8th of the size to be
2093 		 * buckets.  If this guess ends up being routinely
2094 		 * off-the-mark, we may need to dynamically readjust this
2095 		 * based on past performance.
2096 		 */
2097 		uintptr_t hashsize = (buf->dtb_size >> 3) / sizeof (uintptr_t);
2098 
2099 		if ((uintptr_t)agb - hashsize * sizeof (dtrace_aggkey_t *) <
2100 		    (uintptr_t)tomax || hashsize == 0) {
2101 			/*
2102 			 * We've been given a ludicrously small buffer;
2103 			 * increment our drop count and leave.
2104 			 */
2105 			dtrace_buffer_drop(buf);
2106 			return;
2107 		}
2108 
2109 		/*
2110 		 * And now, a pathetic attempt to try to get a an odd (or
2111 		 * perchance, a prime) hash size for better hash distribution.
2112 		 */
2113 		if (hashsize > (DTRACE_AGGHASHSIZE_SLEW << 3))
2114 			hashsize -= DTRACE_AGGHASHSIZE_SLEW;
2115 
2116 		agb->dtagb_hashsize = hashsize;
2117 		agb->dtagb_hash = (dtrace_aggkey_t **)((uintptr_t)agb -
2118 		    agb->dtagb_hashsize * sizeof (dtrace_aggkey_t *));
2119 		agb->dtagb_free = (uintptr_t)agb->dtagb_hash;
2120 
2121 		for (i = 0; i < agb->dtagb_hashsize; i++)
2122 			agb->dtagb_hash[i] = NULL;
2123 	}
2124 
2125 	ASSERT(agg->dtag_first != NULL);
2126 	ASSERT(agg->dtag_first->dta_intuple);
2127 
2128 	/*
2129 	 * Calculate the hash value based on the key.  Note that we _don't_
2130 	 * include the aggid in the hashing (but we will store it as part of
2131 	 * the key).  The hashing algorithm is Bob Jenkins' "One-at-a-time"
2132 	 * algorithm: a simple, quick algorithm that has no known funnels, and
2133 	 * gets good distribution in practice.  The efficacy of the hashing
2134 	 * algorithm (and a comparison with other algorithms) may be found by
2135 	 * running the ::dtrace_aggstat MDB dcmd.
2136 	 */
2137 	for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
2138 		i = act->dta_rec.dtrd_offset - agg->dtag_base;
2139 		limit = i + act->dta_rec.dtrd_size;
2140 		ASSERT(limit <= size);
2141 		isstr = DTRACEACT_ISSTRING(act);
2142 
2143 		for (; i < limit; i++) {
2144 			hashval += data[i];
2145 			hashval += (hashval << 10);
2146 			hashval ^= (hashval >> 6);
2147 
2148 			if (isstr && data[i] == '\0')
2149 				break;
2150 		}
2151 	}
2152 
2153 	hashval += (hashval << 3);
2154 	hashval ^= (hashval >> 11);
2155 	hashval += (hashval << 15);
2156 
2157 	/*
2158 	 * Yes, the divide here is expensive -- but it's generally the least
2159 	 * of the performance issues given the amount of data that we iterate
2160 	 * over to compute hash values, compare data, etc.
2161 	 */
2162 	ndx = hashval % agb->dtagb_hashsize;
2163 
2164 	for (key = agb->dtagb_hash[ndx]; key != NULL; key = key->dtak_next) {
2165 		ASSERT((caddr_t)key >= tomax);
2166 		ASSERT((caddr_t)key < tomax + buf->dtb_size);
2167 
2168 		if (hashval != key->dtak_hashval || key->dtak_size != size)
2169 			continue;
2170 
2171 		kdata = key->dtak_data;
2172 		ASSERT(kdata >= tomax && kdata < tomax + buf->dtb_size);
2173 
2174 		for (act = agg->dtag_first; act->dta_intuple;
2175 		    act = act->dta_next) {
2176 			i = act->dta_rec.dtrd_offset - agg->dtag_base;
2177 			limit = i + act->dta_rec.dtrd_size;
2178 			ASSERT(limit <= size);
2179 			isstr = DTRACEACT_ISSTRING(act);
2180 
2181 			for (; i < limit; i++) {
2182 				if (kdata[i] != data[i])
2183 					goto next;
2184 
2185 				if (isstr && data[i] == '\0')
2186 					break;
2187 			}
2188 		}
2189 
2190 		if (action != key->dtak_action) {
2191 			/*
2192 			 * We are aggregating on the same value in the same
2193 			 * aggregation with two different aggregating actions.
2194 			 * (This should have been picked up in the compiler,
2195 			 * so we may be dealing with errant or devious DIF.)
2196 			 * This is an error condition; we indicate as much,
2197 			 * and return.
2198 			 */
2199 			DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
2200 			return;
2201 		}
2202 
2203 		/*
2204 		 * This is a hit:  we need to apply the aggregator to
2205 		 * the value at this key.
2206 		 */
2207 		agg->dtag_aggregate((uint64_t *)(kdata + size), expr, arg);
2208 		return;
2209 next:
2210 		continue;
2211 	}
2212 
2213 	/*
2214 	 * We didn't find it.  We need to allocate some zero-filled space,
2215 	 * link it into the hash table appropriately, and apply the aggregator
2216 	 * to the (zero-filled) value.
2217 	 */
2218 	offs = buf->dtb_offset;
2219 	while (offs & (align - 1))
2220 		offs += sizeof (uint32_t);
2221 
2222 	/*
2223 	 * If we don't have enough room to both allocate a new key _and_
2224 	 * its associated data, increment the drop count and return.
2225 	 */
2226 	if ((uintptr_t)tomax + offs + fsize >
2227 	    agb->dtagb_free - sizeof (dtrace_aggkey_t)) {
2228 		dtrace_buffer_drop(buf);
2229 		return;
2230 	}
2231 
2232 	/*CONSTCOND*/
2233 	ASSERT(!(sizeof (dtrace_aggkey_t) & (sizeof (uintptr_t) - 1)));
2234 	key = (dtrace_aggkey_t *)(agb->dtagb_free - sizeof (dtrace_aggkey_t));
2235 	agb->dtagb_free -= sizeof (dtrace_aggkey_t);
2236 
2237 	key->dtak_data = kdata = tomax + offs;
2238 	buf->dtb_offset = offs + fsize;
2239 
2240 	/*
2241 	 * Now copy the data across.
2242 	 */
2243 	*((dtrace_aggid_t *)kdata) = agg->dtag_id;
2244 
2245 	for (i = sizeof (dtrace_aggid_t); i < size; i++)
2246 		kdata[i] = data[i];
2247 
2248 	/*
2249 	 * Because strings are not zeroed out by default, we need to iterate
2250 	 * looking for actions that store strings, and we need to explicitly
2251 	 * pad these strings out with zeroes.
2252 	 */
2253 	for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
2254 		int nul;
2255 
2256 		if (!DTRACEACT_ISSTRING(act))
2257 			continue;
2258 
2259 		i = act->dta_rec.dtrd_offset - agg->dtag_base;
2260 		limit = i + act->dta_rec.dtrd_size;
2261 		ASSERT(limit <= size);
2262 
2263 		for (nul = 0; i < limit; i++) {
2264 			if (nul) {
2265 				kdata[i] = '\0';
2266 				continue;
2267 			}
2268 
2269 			if (data[i] != '\0')
2270 				continue;
2271 
2272 			nul = 1;
2273 		}
2274 	}
2275 
2276 	for (i = size; i < fsize; i++)
2277 		kdata[i] = 0;
2278 
2279 	key->dtak_hashval = hashval;
2280 	key->dtak_size = size;
2281 	key->dtak_action = action;
2282 	key->dtak_next = agb->dtagb_hash[ndx];
2283 	agb->dtagb_hash[ndx] = key;
2284 
2285 	/*
2286 	 * Finally, apply the aggregator.
2287 	 */
2288 	*((uint64_t *)(key->dtak_data + size)) = agg->dtag_initial;
2289 	agg->dtag_aggregate((uint64_t *)(key->dtak_data + size), expr, arg);
2290 }
2291 
2292 /*
2293  * Given consumer state, this routine finds a speculation in the INACTIVE
2294  * state and transitions it into the ACTIVE state.  If there is no speculation
2295  * in the INACTIVE state, 0 is returned.  In this case, no error counter is
2296  * incremented -- it is up to the caller to take appropriate action.
2297  */
2298 static int
2299 dtrace_speculation(dtrace_state_t *state)
2300 {
2301 	int i = 0;
2302 	dtrace_speculation_state_t current;
2303 	uint32_t *stat = &state->dts_speculations_unavail, count;
2304 
2305 	while (i < state->dts_nspeculations) {
2306 		dtrace_speculation_t *spec = &state->dts_speculations[i];
2307 
2308 		current = spec->dtsp_state;
2309 
2310 		if (current != DTRACESPEC_INACTIVE) {
2311 			if (current == DTRACESPEC_COMMITTINGMANY ||
2312 			    current == DTRACESPEC_COMMITTING ||
2313 			    current == DTRACESPEC_DISCARDING)
2314 				stat = &state->dts_speculations_busy;
2315 			i++;
2316 			continue;
2317 		}
2318 
2319 		if (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2320 		    current, DTRACESPEC_ACTIVE) == current)
2321 			return (i + 1);
2322 	}
2323 
2324 	/*
2325 	 * We couldn't find a speculation.  If we found as much as a single
2326 	 * busy speculation buffer, we'll attribute this failure as "busy"
2327 	 * instead of "unavail".
2328 	 */
2329 	do {
2330 		count = *stat;
2331 	} while (dtrace_cas32(stat, count, count + 1) != count);
2332 
2333 	return (0);
2334 }
2335 
2336 /*
2337  * This routine commits an active speculation.  If the specified speculation
2338  * is not in a valid state to perform a commit(), this routine will silently do
2339  * nothing.  The state of the specified speculation is transitioned according
2340  * to the state transition diagram outlined in <sys/dtrace_impl.h>
2341  */
2342 static void
2343 dtrace_speculation_commit(dtrace_state_t *state, processorid_t cpu,
2344     dtrace_specid_t which)
2345 {
2346 	dtrace_speculation_t *spec;
2347 	dtrace_buffer_t *src, *dest;
2348 	uintptr_t daddr, saddr, dlimit;
2349 	dtrace_speculation_state_t current, new = 0;
2350 	intptr_t offs;
2351 
2352 	if (which == 0)
2353 		return;
2354 
2355 	if (which > state->dts_nspeculations) {
2356 		cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2357 		return;
2358 	}
2359 
2360 	spec = &state->dts_speculations[which - 1];
2361 	src = &spec->dtsp_buffer[cpu];
2362 	dest = &state->dts_buffer[cpu];
2363 
2364 	do {
2365 		current = spec->dtsp_state;
2366 
2367 		if (current == DTRACESPEC_COMMITTINGMANY)
2368 			break;
2369 
2370 		switch (current) {
2371 		case DTRACESPEC_INACTIVE:
2372 		case DTRACESPEC_DISCARDING:
2373 			return;
2374 
2375 		case DTRACESPEC_COMMITTING:
2376 			/*
2377 			 * This is only possible if we are (a) commit()'ing
2378 			 * without having done a prior speculate() on this CPU
2379 			 * and (b) racing with another commit() on a different
2380 			 * CPU.  There's nothing to do -- we just assert that
2381 			 * our offset is 0.
2382 			 */
2383 			ASSERT(src->dtb_offset == 0);
2384 			return;
2385 
2386 		case DTRACESPEC_ACTIVE:
2387 			new = DTRACESPEC_COMMITTING;
2388 			break;
2389 
2390 		case DTRACESPEC_ACTIVEONE:
2391 			/*
2392 			 * This speculation is active on one CPU.  If our
2393 			 * buffer offset is non-zero, we know that the one CPU
2394 			 * must be us.  Otherwise, we are committing on a
2395 			 * different CPU from the speculate(), and we must
2396 			 * rely on being asynchronously cleaned.
2397 			 */
2398 			if (src->dtb_offset != 0) {
2399 				new = DTRACESPEC_COMMITTING;
2400 				break;
2401 			}
2402 			/*FALLTHROUGH*/
2403 
2404 		case DTRACESPEC_ACTIVEMANY:
2405 			new = DTRACESPEC_COMMITTINGMANY;
2406 			break;
2407 
2408 		default:
2409 			ASSERT(0);
2410 		}
2411 	} while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2412 	    current, new) != current);
2413 
2414 	/*
2415 	 * We have set the state to indicate that we are committing this
2416 	 * speculation.  Now reserve the necessary space in the destination
2417 	 * buffer.
2418 	 */
2419 	if ((offs = dtrace_buffer_reserve(dest, src->dtb_offset,
2420 	    sizeof (uint64_t), state, NULL)) < 0) {
2421 		dtrace_buffer_drop(dest);
2422 		goto out;
2423 	}
2424 
2425 	/*
2426 	 * We have the space; copy the buffer across.  (Note that this is a
2427 	 * highly subobtimal bcopy(); in the unlikely event that this becomes
2428 	 * a serious performance issue, a high-performance DTrace-specific
2429 	 * bcopy() should obviously be invented.)
2430 	 */
2431 	daddr = (uintptr_t)dest->dtb_tomax + offs;
2432 	dlimit = daddr + src->dtb_offset;
2433 	saddr = (uintptr_t)src->dtb_tomax;
2434 
2435 	/*
2436 	 * First, the aligned portion.
2437 	 */
2438 	while (dlimit - daddr >= sizeof (uint64_t)) {
2439 		*((uint64_t *)daddr) = *((uint64_t *)saddr);
2440 
2441 		daddr += sizeof (uint64_t);
2442 		saddr += sizeof (uint64_t);
2443 	}
2444 
2445 	/*
2446 	 * Now any left-over bit...
2447 	 */
2448 	while (dlimit - daddr)
2449 		*((uint8_t *)daddr++) = *((uint8_t *)saddr++);
2450 
2451 	/*
2452 	 * Finally, commit the reserved space in the destination buffer.
2453 	 */
2454 	dest->dtb_offset = offs + src->dtb_offset;
2455 
2456 out:
2457 	/*
2458 	 * If we're lucky enough to be the only active CPU on this speculation
2459 	 * buffer, we can just set the state back to DTRACESPEC_INACTIVE.
2460 	 */
2461 	if (current == DTRACESPEC_ACTIVE ||
2462 	    (current == DTRACESPEC_ACTIVEONE && new == DTRACESPEC_COMMITTING)) {
2463 		uint32_t rval = dtrace_cas32((uint32_t *)&spec->dtsp_state,
2464 		    DTRACESPEC_COMMITTING, DTRACESPEC_INACTIVE);
2465 
2466 		ASSERT(rval == DTRACESPEC_COMMITTING);
2467 	}
2468 
2469 	src->dtb_offset = 0;
2470 	src->dtb_xamot_drops += src->dtb_drops;
2471 	src->dtb_drops = 0;
2472 }
2473 
2474 /*
2475  * This routine discards an active speculation.  If the specified speculation
2476  * is not in a valid state to perform a discard(), this routine will silently
2477  * do nothing.  The state of the specified speculation is transitioned
2478  * according to the state transition diagram outlined in <sys/dtrace_impl.h>
2479  */
2480 static void
2481 dtrace_speculation_discard(dtrace_state_t *state, processorid_t cpu,
2482     dtrace_specid_t which)
2483 {
2484 	dtrace_speculation_t *spec;
2485 	dtrace_speculation_state_t current, new = 0;
2486 	dtrace_buffer_t *buf;
2487 
2488 	if (which == 0)
2489 		return;
2490 
2491 	if (which > state->dts_nspeculations) {
2492 		cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2493 		return;
2494 	}
2495 
2496 	spec = &state->dts_speculations[which - 1];
2497 	buf = &spec->dtsp_buffer[cpu];
2498 
2499 	do {
2500 		current = spec->dtsp_state;
2501 
2502 		switch (current) {
2503 		case DTRACESPEC_INACTIVE:
2504 		case DTRACESPEC_COMMITTINGMANY:
2505 		case DTRACESPEC_COMMITTING:
2506 		case DTRACESPEC_DISCARDING:
2507 			return;
2508 
2509 		case DTRACESPEC_ACTIVE:
2510 		case DTRACESPEC_ACTIVEMANY:
2511 			new = DTRACESPEC_DISCARDING;
2512 			break;
2513 
2514 		case DTRACESPEC_ACTIVEONE:
2515 			if (buf->dtb_offset != 0) {
2516 				new = DTRACESPEC_INACTIVE;
2517 			} else {
2518 				new = DTRACESPEC_DISCARDING;
2519 			}
2520 			break;
2521 
2522 		default:
2523 			ASSERT(0);
2524 		}
2525 	} while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2526 	    current, new) != current);
2527 
2528 	buf->dtb_offset = 0;
2529 	buf->dtb_drops = 0;
2530 }
2531 
2532 /*
2533  * Note:  not called from probe context.  This function is called
2534  * asynchronously from cross call context to clean any speculations that are
2535  * in the COMMITTINGMANY or DISCARDING states.  These speculations may not be
2536  * transitioned back to the INACTIVE state until all CPUs have cleaned the
2537  * speculation.
2538  */
2539 static void
2540 dtrace_speculation_clean_here(dtrace_state_t *state)
2541 {
2542 	dtrace_icookie_t cookie;
2543 	processorid_t cpu = curcpu;
2544 	dtrace_buffer_t *dest = &state->dts_buffer[cpu];
2545 	dtrace_specid_t i;
2546 
2547 	cookie = dtrace_interrupt_disable();
2548 
2549 	if (dest->dtb_tomax == NULL) {
2550 		dtrace_interrupt_enable(cookie);
2551 		return;
2552 	}
2553 
2554 	for (i = 0; i < state->dts_nspeculations; i++) {
2555 		dtrace_speculation_t *spec = &state->dts_speculations[i];
2556 		dtrace_buffer_t *src = &spec->dtsp_buffer[cpu];
2557 
2558 		if (src->dtb_tomax == NULL)
2559 			continue;
2560 
2561 		if (spec->dtsp_state == DTRACESPEC_DISCARDING) {
2562 			src->dtb_offset = 0;
2563 			continue;
2564 		}
2565 
2566 		if (spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
2567 			continue;
2568 
2569 		if (src->dtb_offset == 0)
2570 			continue;
2571 
2572 		dtrace_speculation_commit(state, cpu, i + 1);
2573 	}
2574 
2575 	dtrace_interrupt_enable(cookie);
2576 }
2577 
2578 /*
2579  * Note:  not called from probe context.  This function is called
2580  * asynchronously (and at a regular interval) to clean any speculations that
2581  * are in the COMMITTINGMANY or DISCARDING states.  If it discovers that there
2582  * is work to be done, it cross calls all CPUs to perform that work;
2583  * COMMITMANY and DISCARDING speculations may not be transitioned back to the
2584  * INACTIVE state until they have been cleaned by all CPUs.
2585  */
2586 static void
2587 dtrace_speculation_clean(dtrace_state_t *state)
2588 {
2589 	int work = 0, rv;
2590 	dtrace_specid_t i;
2591 
2592 	for (i = 0; i < state->dts_nspeculations; i++) {
2593 		dtrace_speculation_t *spec = &state->dts_speculations[i];
2594 
2595 		ASSERT(!spec->dtsp_cleaning);
2596 
2597 		if (spec->dtsp_state != DTRACESPEC_DISCARDING &&
2598 		    spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
2599 			continue;
2600 
2601 		work++;
2602 		spec->dtsp_cleaning = 1;
2603 	}
2604 
2605 	if (!work)
2606 		return;
2607 
2608 	dtrace_xcall(DTRACE_CPUALL,
2609 	    (dtrace_xcall_t)dtrace_speculation_clean_here, state);
2610 
2611 	/*
2612 	 * We now know that all CPUs have committed or discarded their
2613 	 * speculation buffers, as appropriate.  We can now set the state
2614 	 * to inactive.
2615 	 */
2616 	for (i = 0; i < state->dts_nspeculations; i++) {
2617 		dtrace_speculation_t *spec = &state->dts_speculations[i];
2618 		dtrace_speculation_state_t current, new;
2619 
2620 		if (!spec->dtsp_cleaning)
2621 			continue;
2622 
2623 		current = spec->dtsp_state;
2624 		ASSERT(current == DTRACESPEC_DISCARDING ||
2625 		    current == DTRACESPEC_COMMITTINGMANY);
2626 
2627 		new = DTRACESPEC_INACTIVE;
2628 
2629 		rv = dtrace_cas32((uint32_t *)&spec->dtsp_state, current, new);
2630 		ASSERT(rv == current);
2631 		spec->dtsp_cleaning = 0;
2632 	}
2633 }
2634 
2635 /*
2636  * Called as part of a speculate() to get the speculative buffer associated
2637  * with a given speculation.  Returns NULL if the specified speculation is not
2638  * in an ACTIVE state.  If the speculation is in the ACTIVEONE state -- and
2639  * the active CPU is not the specified CPU -- the speculation will be
2640  * atomically transitioned into the ACTIVEMANY state.
2641  */
2642 static dtrace_buffer_t *
2643 dtrace_speculation_buffer(dtrace_state_t *state, processorid_t cpuid,
2644     dtrace_specid_t which)
2645 {
2646 	dtrace_speculation_t *spec;
2647 	dtrace_speculation_state_t current, new = 0;
2648 	dtrace_buffer_t *buf;
2649 
2650 	if (which == 0)
2651 		return (NULL);
2652 
2653 	if (which > state->dts_nspeculations) {
2654 		cpu_core[cpuid].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2655 		return (NULL);
2656 	}
2657 
2658 	spec = &state->dts_speculations[which - 1];
2659 	buf = &spec->dtsp_buffer[cpuid];
2660 
2661 	do {
2662 		current = spec->dtsp_state;
2663 
2664 		switch (current) {
2665 		case DTRACESPEC_INACTIVE:
2666 		case DTRACESPEC_COMMITTINGMANY:
2667 		case DTRACESPEC_DISCARDING:
2668 			return (NULL);
2669 
2670 		case DTRACESPEC_COMMITTING:
2671 			ASSERT(buf->dtb_offset == 0);
2672 			return (NULL);
2673 
2674 		case DTRACESPEC_ACTIVEONE:
2675 			/*
2676 			 * This speculation is currently active on one CPU.
2677 			 * Check the offset in the buffer; if it's non-zero,
2678 			 * that CPU must be us (and we leave the state alone).
2679 			 * If it's zero, assume that we're starting on a new
2680 			 * CPU -- and change the state to indicate that the
2681 			 * speculation is active on more than one CPU.
2682 			 */
2683 			if (buf->dtb_offset != 0)
2684 				return (buf);
2685 
2686 			new = DTRACESPEC_ACTIVEMANY;
2687 			break;
2688 
2689 		case DTRACESPEC_ACTIVEMANY:
2690 			return (buf);
2691 
2692 		case DTRACESPEC_ACTIVE:
2693 			new = DTRACESPEC_ACTIVEONE;
2694 			break;
2695 
2696 		default:
2697 			ASSERT(0);
2698 		}
2699 	} while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2700 	    current, new) != current);
2701 
2702 	ASSERT(new == DTRACESPEC_ACTIVEONE || new == DTRACESPEC_ACTIVEMANY);
2703 	return (buf);
2704 }
2705 
2706 /*
2707  * Return a string.  In the event that the user lacks the privilege to access
2708  * arbitrary kernel memory, we copy the string out to scratch memory so that we
2709  * don't fail access checking.
2710  *
2711  * dtrace_dif_variable() uses this routine as a helper for various
2712  * builtin values such as 'execname' and 'probefunc.'
2713  */
2714 uintptr_t
2715 dtrace_dif_varstr(uintptr_t addr, dtrace_state_t *state,
2716     dtrace_mstate_t *mstate)
2717 {
2718 	uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
2719 	uintptr_t ret;
2720 	size_t strsz;
2721 
2722 	/*
2723 	 * The easy case: this probe is allowed to read all of memory, so
2724 	 * we can just return this as a vanilla pointer.
2725 	 */
2726 	if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
2727 		return (addr);
2728 
2729 	/*
2730 	 * This is the tougher case: we copy the string in question from
2731 	 * kernel memory into scratch memory and return it that way: this
2732 	 * ensures that we won't trip up when access checking tests the
2733 	 * BYREF return value.
2734 	 */
2735 	strsz = dtrace_strlen((char *)addr, size) + 1;
2736 
2737 	if (mstate->dtms_scratch_ptr + strsz >
2738 	    mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
2739 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
2740 		return (0);
2741 	}
2742 
2743 	dtrace_strcpy((const void *)addr, (void *)mstate->dtms_scratch_ptr,
2744 	    strsz);
2745 	ret = mstate->dtms_scratch_ptr;
2746 	mstate->dtms_scratch_ptr += strsz;
2747 	return (ret);
2748 }
2749 
2750 /*
2751  * Return a string from a memoy address which is known to have one or
2752  * more concatenated, individually zero terminated, sub-strings.
2753  * In the event that the user lacks the privilege to access
2754  * arbitrary kernel memory, we copy the string out to scratch memory so that we
2755  * don't fail access checking.
2756  *
2757  * dtrace_dif_variable() uses this routine as a helper for various
2758  * builtin values such as 'execargs'.
2759  */
2760 static uintptr_t
2761 dtrace_dif_varstrz(uintptr_t addr, size_t strsz, dtrace_state_t *state,
2762     dtrace_mstate_t *mstate)
2763 {
2764 	char *p;
2765 	size_t i;
2766 	uintptr_t ret;
2767 
2768 	if (mstate->dtms_scratch_ptr + strsz >
2769 	    mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
2770 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
2771 		return (0);
2772 	}
2773 
2774 	dtrace_bcopy((const void *)addr, (void *)mstate->dtms_scratch_ptr,
2775 	    strsz);
2776 
2777 	/* Replace sub-string termination characters with a space. */
2778 	for (p = (char *) mstate->dtms_scratch_ptr, i = 0; i < strsz - 1;
2779 	    p++, i++)
2780 		if (*p == '\0')
2781 			*p = ' ';
2782 
2783 	ret = mstate->dtms_scratch_ptr;
2784 	mstate->dtms_scratch_ptr += strsz;
2785 	return (ret);
2786 }
2787 
2788 /*
2789  * This function implements the DIF emulator's variable lookups.  The emulator
2790  * passes a reserved variable identifier and optional built-in array index.
2791  */
2792 static uint64_t
2793 dtrace_dif_variable(dtrace_mstate_t *mstate, dtrace_state_t *state, uint64_t v,
2794     uint64_t ndx)
2795 {
2796 	/*
2797 	 * If we're accessing one of the uncached arguments, we'll turn this
2798 	 * into a reference in the args array.
2799 	 */
2800 	if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9) {
2801 		ndx = v - DIF_VAR_ARG0;
2802 		v = DIF_VAR_ARGS;
2803 	}
2804 
2805 	switch (v) {
2806 	case DIF_VAR_ARGS:
2807 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_ARGS);
2808 		if (ndx >= sizeof (mstate->dtms_arg) /
2809 		    sizeof (mstate->dtms_arg[0])) {
2810 			int aframes = mstate->dtms_probe->dtpr_aframes + 2;
2811 			dtrace_provider_t *pv;
2812 			uint64_t val;
2813 
2814 			pv = mstate->dtms_probe->dtpr_provider;
2815 			if (pv->dtpv_pops.dtps_getargval != NULL)
2816 				val = pv->dtpv_pops.dtps_getargval(pv->dtpv_arg,
2817 				    mstate->dtms_probe->dtpr_id,
2818 				    mstate->dtms_probe->dtpr_arg, ndx, aframes);
2819 			else
2820 				val = dtrace_getarg(ndx, aframes);
2821 
2822 			/*
2823 			 * This is regrettably required to keep the compiler
2824 			 * from tail-optimizing the call to dtrace_getarg().
2825 			 * The condition always evaluates to true, but the
2826 			 * compiler has no way of figuring that out a priori.
2827 			 * (None of this would be necessary if the compiler
2828 			 * could be relied upon to _always_ tail-optimize
2829 			 * the call to dtrace_getarg() -- but it can't.)
2830 			 */
2831 			if (mstate->dtms_probe != NULL)
2832 				return (val);
2833 
2834 			ASSERT(0);
2835 		}
2836 
2837 		return (mstate->dtms_arg[ndx]);
2838 
2839 #if defined(sun)
2840 	case DIF_VAR_UREGS: {
2841 		klwp_t *lwp;
2842 
2843 		if (!dtrace_priv_proc(state))
2844 			return (0);
2845 
2846 		if ((lwp = curthread->t_lwp) == NULL) {
2847 			DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
2848 			cpu_core[curcpu].cpuc_dtrace_illval = NULL;
2849 			return (0);
2850 		}
2851 
2852 		return (dtrace_getreg(lwp->lwp_regs, ndx));
2853 		return (0);
2854 	}
2855 #else
2856 	case DIF_VAR_UREGS: {
2857 		struct trapframe *tframe;
2858 
2859 		if (!dtrace_priv_proc(state))
2860 			return (0);
2861 
2862 		if ((tframe = curthread->td_frame) == NULL) {
2863 			DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
2864 			cpu_core[curcpu].cpuc_dtrace_illval = 0;
2865 			return (0);
2866 		}
2867 
2868 		return (dtrace_getreg(tframe, ndx));
2869 	}
2870 #endif
2871 
2872 	case DIF_VAR_CURTHREAD:
2873 		if (!dtrace_priv_kernel(state))
2874 			return (0);
2875 		return ((uint64_t)(uintptr_t)curthread);
2876 
2877 	case DIF_VAR_TIMESTAMP:
2878 		if (!(mstate->dtms_present & DTRACE_MSTATE_TIMESTAMP)) {
2879 			mstate->dtms_timestamp = dtrace_gethrtime();
2880 			mstate->dtms_present |= DTRACE_MSTATE_TIMESTAMP;
2881 		}
2882 		return (mstate->dtms_timestamp);
2883 
2884 	case DIF_VAR_VTIMESTAMP:
2885 		ASSERT(dtrace_vtime_references != 0);
2886 		return (curthread->t_dtrace_vtime);
2887 
2888 	case DIF_VAR_WALLTIMESTAMP:
2889 		if (!(mstate->dtms_present & DTRACE_MSTATE_WALLTIMESTAMP)) {
2890 			mstate->dtms_walltimestamp = dtrace_gethrestime();
2891 			mstate->dtms_present |= DTRACE_MSTATE_WALLTIMESTAMP;
2892 		}
2893 		return (mstate->dtms_walltimestamp);
2894 
2895 #if defined(sun)
2896 	case DIF_VAR_IPL:
2897 		if (!dtrace_priv_kernel(state))
2898 			return (0);
2899 		if (!(mstate->dtms_present & DTRACE_MSTATE_IPL)) {
2900 			mstate->dtms_ipl = dtrace_getipl();
2901 			mstate->dtms_present |= DTRACE_MSTATE_IPL;
2902 		}
2903 		return (mstate->dtms_ipl);
2904 #endif
2905 
2906 	case DIF_VAR_EPID:
2907 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_EPID);
2908 		return (mstate->dtms_epid);
2909 
2910 	case DIF_VAR_ID:
2911 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
2912 		return (mstate->dtms_probe->dtpr_id);
2913 
2914 	case DIF_VAR_STACKDEPTH:
2915 		if (!dtrace_priv_kernel(state))
2916 			return (0);
2917 		if (!(mstate->dtms_present & DTRACE_MSTATE_STACKDEPTH)) {
2918 			int aframes = mstate->dtms_probe->dtpr_aframes + 2;
2919 
2920 			mstate->dtms_stackdepth = dtrace_getstackdepth(aframes);
2921 			mstate->dtms_present |= DTRACE_MSTATE_STACKDEPTH;
2922 		}
2923 		return (mstate->dtms_stackdepth);
2924 
2925 	case DIF_VAR_USTACKDEPTH:
2926 		if (!dtrace_priv_proc(state))
2927 			return (0);
2928 		if (!(mstate->dtms_present & DTRACE_MSTATE_USTACKDEPTH)) {
2929 			/*
2930 			 * See comment in DIF_VAR_PID.
2931 			 */
2932 			if (DTRACE_ANCHORED(mstate->dtms_probe) &&
2933 			    CPU_ON_INTR(CPU)) {
2934 				mstate->dtms_ustackdepth = 0;
2935 			} else {
2936 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
2937 				mstate->dtms_ustackdepth =
2938 				    dtrace_getustackdepth();
2939 				DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
2940 			}
2941 			mstate->dtms_present |= DTRACE_MSTATE_USTACKDEPTH;
2942 		}
2943 		return (mstate->dtms_ustackdepth);
2944 
2945 	case DIF_VAR_CALLER:
2946 		if (!dtrace_priv_kernel(state))
2947 			return (0);
2948 		if (!(mstate->dtms_present & DTRACE_MSTATE_CALLER)) {
2949 			int aframes = mstate->dtms_probe->dtpr_aframes + 2;
2950 
2951 			if (!DTRACE_ANCHORED(mstate->dtms_probe)) {
2952 				/*
2953 				 * If this is an unanchored probe, we are
2954 				 * required to go through the slow path:
2955 				 * dtrace_caller() only guarantees correct
2956 				 * results for anchored probes.
2957 				 */
2958 				pc_t caller[2] = {0, 0};
2959 
2960 				dtrace_getpcstack(caller, 2, aframes,
2961 				    (uint32_t *)(uintptr_t)mstate->dtms_arg[0]);
2962 				mstate->dtms_caller = caller[1];
2963 			} else if ((mstate->dtms_caller =
2964 			    dtrace_caller(aframes)) == -1) {
2965 				/*
2966 				 * We have failed to do this the quick way;
2967 				 * we must resort to the slower approach of
2968 				 * calling dtrace_getpcstack().
2969 				 */
2970 				pc_t caller = 0;
2971 
2972 				dtrace_getpcstack(&caller, 1, aframes, NULL);
2973 				mstate->dtms_caller = caller;
2974 			}
2975 
2976 			mstate->dtms_present |= DTRACE_MSTATE_CALLER;
2977 		}
2978 		return (mstate->dtms_caller);
2979 
2980 	case DIF_VAR_UCALLER:
2981 		if (!dtrace_priv_proc(state))
2982 			return (0);
2983 
2984 		if (!(mstate->dtms_present & DTRACE_MSTATE_UCALLER)) {
2985 			uint64_t ustack[3];
2986 
2987 			/*
2988 			 * dtrace_getupcstack() fills in the first uint64_t
2989 			 * with the current PID.  The second uint64_t will
2990 			 * be the program counter at user-level.  The third
2991 			 * uint64_t will contain the caller, which is what
2992 			 * we're after.
2993 			 */
2994 			ustack[2] = 0;
2995 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
2996 			dtrace_getupcstack(ustack, 3);
2997 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
2998 			mstate->dtms_ucaller = ustack[2];
2999 			mstate->dtms_present |= DTRACE_MSTATE_UCALLER;
3000 		}
3001 
3002 		return (mstate->dtms_ucaller);
3003 
3004 	case DIF_VAR_PROBEPROV:
3005 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3006 		return (dtrace_dif_varstr(
3007 		    (uintptr_t)mstate->dtms_probe->dtpr_provider->dtpv_name,
3008 		    state, mstate));
3009 
3010 	case DIF_VAR_PROBEMOD:
3011 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3012 		return (dtrace_dif_varstr(
3013 		    (uintptr_t)mstate->dtms_probe->dtpr_mod,
3014 		    state, mstate));
3015 
3016 	case DIF_VAR_PROBEFUNC:
3017 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3018 		return (dtrace_dif_varstr(
3019 		    (uintptr_t)mstate->dtms_probe->dtpr_func,
3020 		    state, mstate));
3021 
3022 	case DIF_VAR_PROBENAME:
3023 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3024 		return (dtrace_dif_varstr(
3025 		    (uintptr_t)mstate->dtms_probe->dtpr_name,
3026 		    state, mstate));
3027 
3028 	case DIF_VAR_PID:
3029 		if (!dtrace_priv_proc(state))
3030 			return (0);
3031 
3032 #if defined(sun)
3033 		/*
3034 		 * Note that we are assuming that an unanchored probe is
3035 		 * always due to a high-level interrupt.  (And we're assuming
3036 		 * that there is only a single high level interrupt.)
3037 		 */
3038 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3039 			return (pid0.pid_id);
3040 
3041 		/*
3042 		 * It is always safe to dereference one's own t_procp pointer:
3043 		 * it always points to a valid, allocated proc structure.
3044 		 * Further, it is always safe to dereference the p_pidp member
3045 		 * of one's own proc structure.  (These are truisms becuase
3046 		 * threads and processes don't clean up their own state --
3047 		 * they leave that task to whomever reaps them.)
3048 		 */
3049 		return ((uint64_t)curthread->t_procp->p_pidp->pid_id);
3050 #else
3051 		return ((uint64_t)curproc->p_pid);
3052 #endif
3053 
3054 	case DIF_VAR_PPID:
3055 		if (!dtrace_priv_proc(state))
3056 			return (0);
3057 
3058 #if defined(sun)
3059 		/*
3060 		 * See comment in DIF_VAR_PID.
3061 		 */
3062 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3063 			return (pid0.pid_id);
3064 
3065 		/*
3066 		 * It is always safe to dereference one's own t_procp pointer:
3067 		 * it always points to a valid, allocated proc structure.
3068 		 * (This is true because threads don't clean up their own
3069 		 * state -- they leave that task to whomever reaps them.)
3070 		 */
3071 		return ((uint64_t)curthread->t_procp->p_ppid);
3072 #else
3073 		return ((uint64_t)curproc->p_pptr->p_pid);
3074 #endif
3075 
3076 	case DIF_VAR_TID:
3077 #if defined(sun)
3078 		/*
3079 		 * See comment in DIF_VAR_PID.
3080 		 */
3081 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3082 			return (0);
3083 #endif
3084 
3085 		return ((uint64_t)curthread->t_tid);
3086 
3087 	case DIF_VAR_EXECARGS: {
3088 		struct pargs *p_args = curthread->td_proc->p_args;
3089 
3090 		if (p_args == NULL)
3091 			return(0);
3092 
3093 		return (dtrace_dif_varstrz(
3094 		    (uintptr_t) p_args->ar_args, p_args->ar_length, state, mstate));
3095 	}
3096 
3097 	case DIF_VAR_EXECNAME:
3098 #if defined(sun)
3099 		if (!dtrace_priv_proc(state))
3100 			return (0);
3101 
3102 		/*
3103 		 * See comment in DIF_VAR_PID.
3104 		 */
3105 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3106 			return ((uint64_t)(uintptr_t)p0.p_user.u_comm);
3107 
3108 		/*
3109 		 * It is always safe to dereference one's own t_procp pointer:
3110 		 * it always points to a valid, allocated proc structure.
3111 		 * (This is true because threads don't clean up their own
3112 		 * state -- they leave that task to whomever reaps them.)
3113 		 */
3114 		return (dtrace_dif_varstr(
3115 		    (uintptr_t)curthread->t_procp->p_user.u_comm,
3116 		    state, mstate));
3117 #else
3118 		return (dtrace_dif_varstr(
3119 		    (uintptr_t) curthread->td_proc->p_comm, state, mstate));
3120 #endif
3121 
3122 	case DIF_VAR_ZONENAME:
3123 #if defined(sun)
3124 		if (!dtrace_priv_proc(state))
3125 			return (0);
3126 
3127 		/*
3128 		 * See comment in DIF_VAR_PID.
3129 		 */
3130 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3131 			return ((uint64_t)(uintptr_t)p0.p_zone->zone_name);
3132 
3133 		/*
3134 		 * It is always safe to dereference one's own t_procp pointer:
3135 		 * it always points to a valid, allocated proc structure.
3136 		 * (This is true because threads don't clean up their own
3137 		 * state -- they leave that task to whomever reaps them.)
3138 		 */
3139 		return (dtrace_dif_varstr(
3140 		    (uintptr_t)curthread->t_procp->p_zone->zone_name,
3141 		    state, mstate));
3142 #else
3143 		return (0);
3144 #endif
3145 
3146 	case DIF_VAR_UID:
3147 		if (!dtrace_priv_proc(state))
3148 			return (0);
3149 
3150 #if defined(sun)
3151 		/*
3152 		 * See comment in DIF_VAR_PID.
3153 		 */
3154 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3155 			return ((uint64_t)p0.p_cred->cr_uid);
3156 #endif
3157 
3158 		/*
3159 		 * It is always safe to dereference one's own t_procp pointer:
3160 		 * it always points to a valid, allocated proc structure.
3161 		 * (This is true because threads don't clean up their own
3162 		 * state -- they leave that task to whomever reaps them.)
3163 		 *
3164 		 * Additionally, it is safe to dereference one's own process
3165 		 * credential, since this is never NULL after process birth.
3166 		 */
3167 		return ((uint64_t)curthread->t_procp->p_cred->cr_uid);
3168 
3169 	case DIF_VAR_GID:
3170 		if (!dtrace_priv_proc(state))
3171 			return (0);
3172 
3173 #if defined(sun)
3174 		/*
3175 		 * See comment in DIF_VAR_PID.
3176 		 */
3177 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3178 			return ((uint64_t)p0.p_cred->cr_gid);
3179 #endif
3180 
3181 		/*
3182 		 * It is always safe to dereference one's own t_procp pointer:
3183 		 * it always points to a valid, allocated proc structure.
3184 		 * (This is true because threads don't clean up their own
3185 		 * state -- they leave that task to whomever reaps them.)
3186 		 *
3187 		 * Additionally, it is safe to dereference one's own process
3188 		 * credential, since this is never NULL after process birth.
3189 		 */
3190 		return ((uint64_t)curthread->t_procp->p_cred->cr_gid);
3191 
3192 	case DIF_VAR_ERRNO: {
3193 #if defined(sun)
3194 		klwp_t *lwp;
3195 		if (!dtrace_priv_proc(state))
3196 			return (0);
3197 
3198 		/*
3199 		 * See comment in DIF_VAR_PID.
3200 		 */
3201 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3202 			return (0);
3203 
3204 		/*
3205 		 * It is always safe to dereference one's own t_lwp pointer in
3206 		 * the event that this pointer is non-NULL.  (This is true
3207 		 * because threads and lwps don't clean up their own state --
3208 		 * they leave that task to whomever reaps them.)
3209 		 */
3210 		if ((lwp = curthread->t_lwp) == NULL)
3211 			return (0);
3212 
3213 		return ((uint64_t)lwp->lwp_errno);
3214 #else
3215 		return (curthread->td_errno);
3216 #endif
3217 	}
3218 #if !defined(sun)
3219 	case DIF_VAR_CPU: {
3220 		return curcpu;
3221 	}
3222 #endif
3223 	default:
3224 		DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
3225 		return (0);
3226 	}
3227 }
3228 
3229 /*
3230  * Emulate the execution of DTrace ID subroutines invoked by the call opcode.
3231  * Notice that we don't bother validating the proper number of arguments or
3232  * their types in the tuple stack.  This isn't needed because all argument
3233  * interpretation is safe because of our load safety -- the worst that can
3234  * happen is that a bogus program can obtain bogus results.
3235  */
3236 static void
3237 dtrace_dif_subr(uint_t subr, uint_t rd, uint64_t *regs,
3238     dtrace_key_t *tupregs, int nargs,
3239     dtrace_mstate_t *mstate, dtrace_state_t *state)
3240 {
3241 	volatile uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
3242 	volatile uintptr_t *illval = &cpu_core[curcpu].cpuc_dtrace_illval;
3243 	dtrace_vstate_t *vstate = &state->dts_vstate;
3244 
3245 #if defined(sun)
3246 	union {
3247 		mutex_impl_t mi;
3248 		uint64_t mx;
3249 	} m;
3250 
3251 	union {
3252 		krwlock_t ri;
3253 		uintptr_t rw;
3254 	} r;
3255 #else
3256 	struct thread *lowner;
3257 	union {
3258 		struct lock_object *li;
3259 		uintptr_t lx;
3260 	} l;
3261 #endif
3262 
3263 	switch (subr) {
3264 	case DIF_SUBR_RAND:
3265 		regs[rd] = (dtrace_gethrtime() * 2416 + 374441) % 1771875;
3266 		break;
3267 
3268 #if defined(sun)
3269 	case DIF_SUBR_MUTEX_OWNED:
3270 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3271 		    mstate, vstate)) {
3272 			regs[rd] = 0;
3273 			break;
3274 		}
3275 
3276 		m.mx = dtrace_load64(tupregs[0].dttk_value);
3277 		if (MUTEX_TYPE_ADAPTIVE(&m.mi))
3278 			regs[rd] = MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER;
3279 		else
3280 			regs[rd] = LOCK_HELD(&m.mi.m_spin.m_spinlock);
3281 		break;
3282 
3283 	case DIF_SUBR_MUTEX_OWNER:
3284 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3285 		    mstate, vstate)) {
3286 			regs[rd] = 0;
3287 			break;
3288 		}
3289 
3290 		m.mx = dtrace_load64(tupregs[0].dttk_value);
3291 		if (MUTEX_TYPE_ADAPTIVE(&m.mi) &&
3292 		    MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER)
3293 			regs[rd] = (uintptr_t)MUTEX_OWNER(&m.mi);
3294 		else
3295 			regs[rd] = 0;
3296 		break;
3297 
3298 	case DIF_SUBR_MUTEX_TYPE_ADAPTIVE:
3299 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3300 		    mstate, vstate)) {
3301 			regs[rd] = 0;
3302 			break;
3303 		}
3304 
3305 		m.mx = dtrace_load64(tupregs[0].dttk_value);
3306 		regs[rd] = MUTEX_TYPE_ADAPTIVE(&m.mi);
3307 		break;
3308 
3309 	case DIF_SUBR_MUTEX_TYPE_SPIN:
3310 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3311 		    mstate, vstate)) {
3312 			regs[rd] = 0;
3313 			break;
3314 		}
3315 
3316 		m.mx = dtrace_load64(tupregs[0].dttk_value);
3317 		regs[rd] = MUTEX_TYPE_SPIN(&m.mi);
3318 		break;
3319 
3320 	case DIF_SUBR_RW_READ_HELD: {
3321 		uintptr_t tmp;
3322 
3323 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
3324 		    mstate, vstate)) {
3325 			regs[rd] = 0;
3326 			break;
3327 		}
3328 
3329 		r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3330 		regs[rd] = _RW_READ_HELD(&r.ri, tmp);
3331 		break;
3332 	}
3333 
3334 	case DIF_SUBR_RW_WRITE_HELD:
3335 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
3336 		    mstate, vstate)) {
3337 			regs[rd] = 0;
3338 			break;
3339 		}
3340 
3341 		r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3342 		regs[rd] = _RW_WRITE_HELD(&r.ri);
3343 		break;
3344 
3345 	case DIF_SUBR_RW_ISWRITER:
3346 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
3347 		    mstate, vstate)) {
3348 			regs[rd] = 0;
3349 			break;
3350 		}
3351 
3352 		r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3353 		regs[rd] = _RW_ISWRITER(&r.ri);
3354 		break;
3355 
3356 #else
3357 	case DIF_SUBR_MUTEX_OWNED:
3358 		if (!dtrace_canload(tupregs[0].dttk_value,
3359 			sizeof (struct lock_object), mstate, vstate)) {
3360 			regs[rd] = 0;
3361 			break;
3362 		}
3363 		l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3364 		regs[rd] = LOCK_CLASS(l.li)->lc_owner(l.li, &lowner);
3365 		break;
3366 
3367 	case DIF_SUBR_MUTEX_OWNER:
3368 		if (!dtrace_canload(tupregs[0].dttk_value,
3369 			sizeof (struct lock_object), mstate, vstate)) {
3370 			regs[rd] = 0;
3371 			break;
3372 		}
3373 		l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3374 		LOCK_CLASS(l.li)->lc_owner(l.li, &lowner);
3375 		regs[rd] = (uintptr_t)lowner;
3376 		break;
3377 
3378 	case DIF_SUBR_MUTEX_TYPE_ADAPTIVE:
3379 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (struct mtx),
3380 		    mstate, vstate)) {
3381 			regs[rd] = 0;
3382 			break;
3383 		}
3384 		l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3385 		/* XXX - should be only LC_SLEEPABLE? */
3386 		regs[rd] = (LOCK_CLASS(l.li)->lc_flags &
3387 		    (LC_SLEEPLOCK | LC_SLEEPABLE)) != 0;
3388 		break;
3389 
3390 	case DIF_SUBR_MUTEX_TYPE_SPIN:
3391 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (struct mtx),
3392 		    mstate, vstate)) {
3393 			regs[rd] = 0;
3394 			break;
3395 		}
3396 		l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3397 		regs[rd] = (LOCK_CLASS(l.li)->lc_flags & LC_SPINLOCK) != 0;
3398 		break;
3399 
3400 	case DIF_SUBR_RW_READ_HELD:
3401 	case DIF_SUBR_SX_SHARED_HELD:
3402 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
3403 		    mstate, vstate)) {
3404 			regs[rd] = 0;
3405 			break;
3406 		}
3407 		l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3408 		regs[rd] = LOCK_CLASS(l.li)->lc_owner(l.li, &lowner) &&
3409 		    lowner == NULL;
3410 		break;
3411 
3412 	case DIF_SUBR_RW_WRITE_HELD:
3413 	case DIF_SUBR_SX_EXCLUSIVE_HELD:
3414 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
3415 		    mstate, vstate)) {
3416 			regs[rd] = 0;
3417 			break;
3418 		}
3419 		l.lx = dtrace_loadptr(tupregs[0].dttk_value);
3420 		LOCK_CLASS(l.li)->lc_owner(l.li, &lowner);
3421 		regs[rd] = (lowner == curthread);
3422 		break;
3423 
3424 	case DIF_SUBR_RW_ISWRITER:
3425 	case DIF_SUBR_SX_ISEXCLUSIVE:
3426 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
3427 		    mstate, vstate)) {
3428 			regs[rd] = 0;
3429 			break;
3430 		}
3431 		l.lx = dtrace_loadptr(tupregs[0].dttk_value);
3432 		regs[rd] = LOCK_CLASS(l.li)->lc_owner(l.li, &lowner) &&
3433 		    lowner != NULL;
3434 		break;
3435 #endif /* ! defined(sun) */
3436 
3437 	case DIF_SUBR_BCOPY: {
3438 		/*
3439 		 * We need to be sure that the destination is in the scratch
3440 		 * region -- no other region is allowed.
3441 		 */
3442 		uintptr_t src = tupregs[0].dttk_value;
3443 		uintptr_t dest = tupregs[1].dttk_value;
3444 		size_t size = tupregs[2].dttk_value;
3445 
3446 		if (!dtrace_inscratch(dest, size, mstate)) {
3447 			*flags |= CPU_DTRACE_BADADDR;
3448 			*illval = regs[rd];
3449 			break;
3450 		}
3451 
3452 		if (!dtrace_canload(src, size, mstate, vstate)) {
3453 			regs[rd] = 0;
3454 			break;
3455 		}
3456 
3457 		dtrace_bcopy((void *)src, (void *)dest, size);
3458 		break;
3459 	}
3460 
3461 	case DIF_SUBR_ALLOCA:
3462 	case DIF_SUBR_COPYIN: {
3463 		uintptr_t dest = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
3464 		uint64_t size =
3465 		    tupregs[subr == DIF_SUBR_ALLOCA ? 0 : 1].dttk_value;
3466 		size_t scratch_size = (dest - mstate->dtms_scratch_ptr) + size;
3467 
3468 		/*
3469 		 * This action doesn't require any credential checks since
3470 		 * probes will not activate in user contexts to which the
3471 		 * enabling user does not have permissions.
3472 		 */
3473 
3474 		/*
3475 		 * Rounding up the user allocation size could have overflowed
3476 		 * a large, bogus allocation (like -1ULL) to 0.
3477 		 */
3478 		if (scratch_size < size ||
3479 		    !DTRACE_INSCRATCH(mstate, scratch_size)) {
3480 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3481 			regs[rd] = 0;
3482 			break;
3483 		}
3484 
3485 		if (subr == DIF_SUBR_COPYIN) {
3486 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3487 			dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
3488 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3489 		}
3490 
3491 		mstate->dtms_scratch_ptr += scratch_size;
3492 		regs[rd] = dest;
3493 		break;
3494 	}
3495 
3496 	case DIF_SUBR_COPYINTO: {
3497 		uint64_t size = tupregs[1].dttk_value;
3498 		uintptr_t dest = tupregs[2].dttk_value;
3499 
3500 		/*
3501 		 * This action doesn't require any credential checks since
3502 		 * probes will not activate in user contexts to which the
3503 		 * enabling user does not have permissions.
3504 		 */
3505 		if (!dtrace_inscratch(dest, size, mstate)) {
3506 			*flags |= CPU_DTRACE_BADADDR;
3507 			*illval = regs[rd];
3508 			break;
3509 		}
3510 
3511 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3512 		dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
3513 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3514 		break;
3515 	}
3516 
3517 	case DIF_SUBR_COPYINSTR: {
3518 		uintptr_t dest = mstate->dtms_scratch_ptr;
3519 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3520 
3521 		if (nargs > 1 && tupregs[1].dttk_value < size)
3522 			size = tupregs[1].dttk_value + 1;
3523 
3524 		/*
3525 		 * This action doesn't require any credential checks since
3526 		 * probes will not activate in user contexts to which the
3527 		 * enabling user does not have permissions.
3528 		 */
3529 		if (!DTRACE_INSCRATCH(mstate, size)) {
3530 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3531 			regs[rd] = 0;
3532 			break;
3533 		}
3534 
3535 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3536 		dtrace_copyinstr(tupregs[0].dttk_value, dest, size, flags);
3537 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3538 
3539 		((char *)dest)[size - 1] = '\0';
3540 		mstate->dtms_scratch_ptr += size;
3541 		regs[rd] = dest;
3542 		break;
3543 	}
3544 
3545 #if defined(sun)
3546 	case DIF_SUBR_MSGSIZE:
3547 	case DIF_SUBR_MSGDSIZE: {
3548 		uintptr_t baddr = tupregs[0].dttk_value, daddr;
3549 		uintptr_t wptr, rptr;
3550 		size_t count = 0;
3551 		int cont = 0;
3552 
3553 		while (baddr != 0 && !(*flags & CPU_DTRACE_FAULT)) {
3554 
3555 			if (!dtrace_canload(baddr, sizeof (mblk_t), mstate,
3556 			    vstate)) {
3557 				regs[rd] = 0;
3558 				break;
3559 			}
3560 
3561 			wptr = dtrace_loadptr(baddr +
3562 			    offsetof(mblk_t, b_wptr));
3563 
3564 			rptr = dtrace_loadptr(baddr +
3565 			    offsetof(mblk_t, b_rptr));
3566 
3567 			if (wptr < rptr) {
3568 				*flags |= CPU_DTRACE_BADADDR;
3569 				*illval = tupregs[0].dttk_value;
3570 				break;
3571 			}
3572 
3573 			daddr = dtrace_loadptr(baddr +
3574 			    offsetof(mblk_t, b_datap));
3575 
3576 			baddr = dtrace_loadptr(baddr +
3577 			    offsetof(mblk_t, b_cont));
3578 
3579 			/*
3580 			 * We want to prevent against denial-of-service here,
3581 			 * so we're only going to search the list for
3582 			 * dtrace_msgdsize_max mblks.
3583 			 */
3584 			if (cont++ > dtrace_msgdsize_max) {
3585 				*flags |= CPU_DTRACE_ILLOP;
3586 				break;
3587 			}
3588 
3589 			if (subr == DIF_SUBR_MSGDSIZE) {
3590 				if (dtrace_load8(daddr +
3591 				    offsetof(dblk_t, db_type)) != M_DATA)
3592 					continue;
3593 			}
3594 
3595 			count += wptr - rptr;
3596 		}
3597 
3598 		if (!(*flags & CPU_DTRACE_FAULT))
3599 			regs[rd] = count;
3600 
3601 		break;
3602 	}
3603 #endif
3604 
3605 	case DIF_SUBR_PROGENYOF: {
3606 		pid_t pid = tupregs[0].dttk_value;
3607 		proc_t *p;
3608 		int rval = 0;
3609 
3610 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3611 
3612 		for (p = curthread->t_procp; p != NULL; p = p->p_parent) {
3613 #if defined(sun)
3614 			if (p->p_pidp->pid_id == pid) {
3615 #else
3616 			if (p->p_pid == pid) {
3617 #endif
3618 				rval = 1;
3619 				break;
3620 			}
3621 		}
3622 
3623 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3624 
3625 		regs[rd] = rval;
3626 		break;
3627 	}
3628 
3629 	case DIF_SUBR_SPECULATION:
3630 		regs[rd] = dtrace_speculation(state);
3631 		break;
3632 
3633 	case DIF_SUBR_COPYOUT: {
3634 		uintptr_t kaddr = tupregs[0].dttk_value;
3635 		uintptr_t uaddr = tupregs[1].dttk_value;
3636 		uint64_t size = tupregs[2].dttk_value;
3637 
3638 		if (!dtrace_destructive_disallow &&
3639 		    dtrace_priv_proc_control(state) &&
3640 		    !dtrace_istoxic(kaddr, size)) {
3641 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3642 			dtrace_copyout(kaddr, uaddr, size, flags);
3643 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3644 		}
3645 		break;
3646 	}
3647 
3648 	case DIF_SUBR_COPYOUTSTR: {
3649 		uintptr_t kaddr = tupregs[0].dttk_value;
3650 		uintptr_t uaddr = tupregs[1].dttk_value;
3651 		uint64_t size = tupregs[2].dttk_value;
3652 
3653 		if (!dtrace_destructive_disallow &&
3654 		    dtrace_priv_proc_control(state) &&
3655 		    !dtrace_istoxic(kaddr, size)) {
3656 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3657 			dtrace_copyoutstr(kaddr, uaddr, size, flags);
3658 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3659 		}
3660 		break;
3661 	}
3662 
3663 	case DIF_SUBR_STRLEN: {
3664 		size_t sz;
3665 		uintptr_t addr = (uintptr_t)tupregs[0].dttk_value;
3666 		sz = dtrace_strlen((char *)addr,
3667 		    state->dts_options[DTRACEOPT_STRSIZE]);
3668 
3669 		if (!dtrace_canload(addr, sz + 1, mstate, vstate)) {
3670 			regs[rd] = 0;
3671 			break;
3672 		}
3673 
3674 		regs[rd] = sz;
3675 
3676 		break;
3677 	}
3678 
3679 	case DIF_SUBR_STRCHR:
3680 	case DIF_SUBR_STRRCHR: {
3681 		/*
3682 		 * We're going to iterate over the string looking for the
3683 		 * specified character.  We will iterate until we have reached
3684 		 * the string length or we have found the character.  If this
3685 		 * is DIF_SUBR_STRRCHR, we will look for the last occurrence
3686 		 * of the specified character instead of the first.
3687 		 */
3688 		uintptr_t saddr = tupregs[0].dttk_value;
3689 		uintptr_t addr = tupregs[0].dttk_value;
3690 		uintptr_t limit = addr + state->dts_options[DTRACEOPT_STRSIZE];
3691 		char c, target = (char)tupregs[1].dttk_value;
3692 
3693 		for (regs[rd] = 0; addr < limit; addr++) {
3694 			if ((c = dtrace_load8(addr)) == target) {
3695 				regs[rd] = addr;
3696 
3697 				if (subr == DIF_SUBR_STRCHR)
3698 					break;
3699 			}
3700 
3701 			if (c == '\0')
3702 				break;
3703 		}
3704 
3705 		if (!dtrace_canload(saddr, addr - saddr, mstate, vstate)) {
3706 			regs[rd] = 0;
3707 			break;
3708 		}
3709 
3710 		break;
3711 	}
3712 
3713 	case DIF_SUBR_STRSTR:
3714 	case DIF_SUBR_INDEX:
3715 	case DIF_SUBR_RINDEX: {
3716 		/*
3717 		 * We're going to iterate over the string looking for the
3718 		 * specified string.  We will iterate until we have reached
3719 		 * the string length or we have found the string.  (Yes, this
3720 		 * is done in the most naive way possible -- but considering
3721 		 * that the string we're searching for is likely to be
3722 		 * relatively short, the complexity of Rabin-Karp or similar
3723 		 * hardly seems merited.)
3724 		 */
3725 		char *addr = (char *)(uintptr_t)tupregs[0].dttk_value;
3726 		char *substr = (char *)(uintptr_t)tupregs[1].dttk_value;
3727 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3728 		size_t len = dtrace_strlen(addr, size);
3729 		size_t sublen = dtrace_strlen(substr, size);
3730 		char *limit = addr + len, *orig = addr;
3731 		int notfound = subr == DIF_SUBR_STRSTR ? 0 : -1;
3732 		int inc = 1;
3733 
3734 		regs[rd] = notfound;
3735 
3736 		if (!dtrace_canload((uintptr_t)addr, len + 1, mstate, vstate)) {
3737 			regs[rd] = 0;
3738 			break;
3739 		}
3740 
3741 		if (!dtrace_canload((uintptr_t)substr, sublen + 1, mstate,
3742 		    vstate)) {
3743 			regs[rd] = 0;
3744 			break;
3745 		}
3746 
3747 		/*
3748 		 * strstr() and index()/rindex() have similar semantics if
3749 		 * both strings are the empty string: strstr() returns a
3750 		 * pointer to the (empty) string, and index() and rindex()
3751 		 * both return index 0 (regardless of any position argument).
3752 		 */
3753 		if (sublen == 0 && len == 0) {
3754 			if (subr == DIF_SUBR_STRSTR)
3755 				regs[rd] = (uintptr_t)addr;
3756 			else
3757 				regs[rd] = 0;
3758 			break;
3759 		}
3760 
3761 		if (subr != DIF_SUBR_STRSTR) {
3762 			if (subr == DIF_SUBR_RINDEX) {
3763 				limit = orig - 1;
3764 				addr += len;
3765 				inc = -1;
3766 			}
3767 
3768 			/*
3769 			 * Both index() and rindex() take an optional position
3770 			 * argument that denotes the starting position.
3771 			 */
3772 			if (nargs == 3) {
3773 				int64_t pos = (int64_t)tupregs[2].dttk_value;
3774 
3775 				/*
3776 				 * If the position argument to index() is
3777 				 * negative, Perl implicitly clamps it at
3778 				 * zero.  This semantic is a little surprising
3779 				 * given the special meaning of negative
3780 				 * positions to similar Perl functions like
3781 				 * substr(), but it appears to reflect a
3782 				 * notion that index() can start from a
3783 				 * negative index and increment its way up to
3784 				 * the string.  Given this notion, Perl's
3785 				 * rindex() is at least self-consistent in
3786 				 * that it implicitly clamps positions greater
3787 				 * than the string length to be the string
3788 				 * length.  Where Perl completely loses
3789 				 * coherence, however, is when the specified
3790 				 * substring is the empty string ("").  In
3791 				 * this case, even if the position is
3792 				 * negative, rindex() returns 0 -- and even if
3793 				 * the position is greater than the length,
3794 				 * index() returns the string length.  These
3795 				 * semantics violate the notion that index()
3796 				 * should never return a value less than the
3797 				 * specified position and that rindex() should
3798 				 * never return a value greater than the
3799 				 * specified position.  (One assumes that
3800 				 * these semantics are artifacts of Perl's
3801 				 * implementation and not the results of
3802 				 * deliberate design -- it beggars belief that
3803 				 * even Larry Wall could desire such oddness.)
3804 				 * While in the abstract one would wish for
3805 				 * consistent position semantics across
3806 				 * substr(), index() and rindex() -- or at the
3807 				 * very least self-consistent position
3808 				 * semantics for index() and rindex() -- we
3809 				 * instead opt to keep with the extant Perl
3810 				 * semantics, in all their broken glory.  (Do
3811 				 * we have more desire to maintain Perl's
3812 				 * semantics than Perl does?  Probably.)
3813 				 */
3814 				if (subr == DIF_SUBR_RINDEX) {
3815 					if (pos < 0) {
3816 						if (sublen == 0)
3817 							regs[rd] = 0;
3818 						break;
3819 					}
3820 
3821 					if (pos > len)
3822 						pos = len;
3823 				} else {
3824 					if (pos < 0)
3825 						pos = 0;
3826 
3827 					if (pos >= len) {
3828 						if (sublen == 0)
3829 							regs[rd] = len;
3830 						break;
3831 					}
3832 				}
3833 
3834 				addr = orig + pos;
3835 			}
3836 		}
3837 
3838 		for (regs[rd] = notfound; addr != limit; addr += inc) {
3839 			if (dtrace_strncmp(addr, substr, sublen) == 0) {
3840 				if (subr != DIF_SUBR_STRSTR) {
3841 					/*
3842 					 * As D index() and rindex() are
3843 					 * modeled on Perl (and not on awk),
3844 					 * we return a zero-based (and not a
3845 					 * one-based) index.  (For you Perl
3846 					 * weenies: no, we're not going to add
3847 					 * $[ -- and shouldn't you be at a con
3848 					 * or something?)
3849 					 */
3850 					regs[rd] = (uintptr_t)(addr - orig);
3851 					break;
3852 				}
3853 
3854 				ASSERT(subr == DIF_SUBR_STRSTR);
3855 				regs[rd] = (uintptr_t)addr;
3856 				break;
3857 			}
3858 		}
3859 
3860 		break;
3861 	}
3862 
3863 	case DIF_SUBR_STRTOK: {
3864 		uintptr_t addr = tupregs[0].dttk_value;
3865 		uintptr_t tokaddr = tupregs[1].dttk_value;
3866 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3867 		uintptr_t limit, toklimit = tokaddr + size;
3868 		uint8_t c = 0, tokmap[32];	 /* 256 / 8 */
3869 		char *dest = (char *)mstate->dtms_scratch_ptr;
3870 		int i;
3871 
3872 		/*
3873 		 * Check both the token buffer and (later) the input buffer,
3874 		 * since both could be non-scratch addresses.
3875 		 */
3876 		if (!dtrace_strcanload(tokaddr, size, mstate, vstate)) {
3877 			regs[rd] = 0;
3878 			break;
3879 		}
3880 
3881 		if (!DTRACE_INSCRATCH(mstate, size)) {
3882 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3883 			regs[rd] = 0;
3884 			break;
3885 		}
3886 
3887 		if (addr == 0) {
3888 			/*
3889 			 * If the address specified is NULL, we use our saved
3890 			 * strtok pointer from the mstate.  Note that this
3891 			 * means that the saved strtok pointer is _only_
3892 			 * valid within multiple enablings of the same probe --
3893 			 * it behaves like an implicit clause-local variable.
3894 			 */
3895 			addr = mstate->dtms_strtok;
3896 		} else {
3897 			/*
3898 			 * If the user-specified address is non-NULL we must
3899 			 * access check it.  This is the only time we have
3900 			 * a chance to do so, since this address may reside
3901 			 * in the string table of this clause-- future calls
3902 			 * (when we fetch addr from mstate->dtms_strtok)
3903 			 * would fail this access check.
3904 			 */
3905 			if (!dtrace_strcanload(addr, size, mstate, vstate)) {
3906 				regs[rd] = 0;
3907 				break;
3908 			}
3909 		}
3910 
3911 		/*
3912 		 * First, zero the token map, and then process the token
3913 		 * string -- setting a bit in the map for every character
3914 		 * found in the token string.
3915 		 */
3916 		for (i = 0; i < sizeof (tokmap); i++)
3917 			tokmap[i] = 0;
3918 
3919 		for (; tokaddr < toklimit; tokaddr++) {
3920 			if ((c = dtrace_load8(tokaddr)) == '\0')
3921 				break;
3922 
3923 			ASSERT((c >> 3) < sizeof (tokmap));
3924 			tokmap[c >> 3] |= (1 << (c & 0x7));
3925 		}
3926 
3927 		for (limit = addr + size; addr < limit; addr++) {
3928 			/*
3929 			 * We're looking for a character that is _not_ contained
3930 			 * in the token string.
3931 			 */
3932 			if ((c = dtrace_load8(addr)) == '\0')
3933 				break;
3934 
3935 			if (!(tokmap[c >> 3] & (1 << (c & 0x7))))
3936 				break;
3937 		}
3938 
3939 		if (c == '\0') {
3940 			/*
3941 			 * We reached the end of the string without finding
3942 			 * any character that was not in the token string.
3943 			 * We return NULL in this case, and we set the saved
3944 			 * address to NULL as well.
3945 			 */
3946 			regs[rd] = 0;
3947 			mstate->dtms_strtok = 0;
3948 			break;
3949 		}
3950 
3951 		/*
3952 		 * From here on, we're copying into the destination string.
3953 		 */
3954 		for (i = 0; addr < limit && i < size - 1; addr++) {
3955 			if ((c = dtrace_load8(addr)) == '\0')
3956 				break;
3957 
3958 			if (tokmap[c >> 3] & (1 << (c & 0x7)))
3959 				break;
3960 
3961 			ASSERT(i < size);
3962 			dest[i++] = c;
3963 		}
3964 
3965 		ASSERT(i < size);
3966 		dest[i] = '\0';
3967 		regs[rd] = (uintptr_t)dest;
3968 		mstate->dtms_scratch_ptr += size;
3969 		mstate->dtms_strtok = addr;
3970 		break;
3971 	}
3972 
3973 	case DIF_SUBR_SUBSTR: {
3974 		uintptr_t s = tupregs[0].dttk_value;
3975 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3976 		char *d = (char *)mstate->dtms_scratch_ptr;
3977 		int64_t index = (int64_t)tupregs[1].dttk_value;
3978 		int64_t remaining = (int64_t)tupregs[2].dttk_value;
3979 		size_t len = dtrace_strlen((char *)s, size);
3980 		int64_t i = 0;
3981 
3982 		if (!dtrace_canload(s, len + 1, mstate, vstate)) {
3983 			regs[rd] = 0;
3984 			break;
3985 		}
3986 
3987 		if (!DTRACE_INSCRATCH(mstate, size)) {
3988 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3989 			regs[rd] = 0;
3990 			break;
3991 		}
3992 
3993 		if (nargs <= 2)
3994 			remaining = (int64_t)size;
3995 
3996 		if (index < 0) {
3997 			index += len;
3998 
3999 			if (index < 0 && index + remaining > 0) {
4000 				remaining += index;
4001 				index = 0;
4002 			}
4003 		}
4004 
4005 		if (index >= len || index < 0) {
4006 			remaining = 0;
4007 		} else if (remaining < 0) {
4008 			remaining += len - index;
4009 		} else if (index + remaining > size) {
4010 			remaining = size - index;
4011 		}
4012 
4013 		for (i = 0; i < remaining; i++) {
4014 			if ((d[i] = dtrace_load8(s + index + i)) == '\0')
4015 				break;
4016 		}
4017 
4018 		d[i] = '\0';
4019 
4020 		mstate->dtms_scratch_ptr += size;
4021 		regs[rd] = (uintptr_t)d;
4022 		break;
4023 	}
4024 
4025 	case DIF_SUBR_TOUPPER:
4026 	case DIF_SUBR_TOLOWER: {
4027 		uintptr_t s = tupregs[0].dttk_value;
4028 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4029 		char *dest = (char *)mstate->dtms_scratch_ptr, c;
4030 		size_t len = dtrace_strlen((char *)s, size);
4031 		char lower, upper, convert;
4032 		int64_t i;
4033 
4034 		if (subr == DIF_SUBR_TOUPPER) {
4035 			lower = 'a';
4036 			upper = 'z';
4037 			convert = 'A';
4038 		} else {
4039 			lower = 'A';
4040 			upper = 'Z';
4041 			convert = 'a';
4042 		}
4043 
4044 		if (!dtrace_canload(s, len + 1, mstate, vstate)) {
4045 			regs[rd] = 0;
4046 			break;
4047 		}
4048 
4049 		if (!DTRACE_INSCRATCH(mstate, size)) {
4050 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4051 			regs[rd] = 0;
4052 			break;
4053 		}
4054 
4055 		for (i = 0; i < size - 1; i++) {
4056 			if ((c = dtrace_load8(s + i)) == '\0')
4057 				break;
4058 
4059 			if (c >= lower && c <= upper)
4060 				c = convert + (c - lower);
4061 
4062 			dest[i] = c;
4063 		}
4064 
4065 		ASSERT(i < size);
4066 		dest[i] = '\0';
4067 		regs[rd] = (uintptr_t)dest;
4068 		mstate->dtms_scratch_ptr += size;
4069 		break;
4070 	}
4071 
4072 #if defined(sun)
4073 	case DIF_SUBR_GETMAJOR:
4074 #ifdef _LP64
4075 		regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR64) & MAXMAJ64;
4076 #else
4077 		regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR) & MAXMAJ;
4078 #endif
4079 		break;
4080 
4081 	case DIF_SUBR_GETMINOR:
4082 #ifdef _LP64
4083 		regs[rd] = tupregs[0].dttk_value & MAXMIN64;
4084 #else
4085 		regs[rd] = tupregs[0].dttk_value & MAXMIN;
4086 #endif
4087 		break;
4088 
4089 	case DIF_SUBR_DDI_PATHNAME: {
4090 		/*
4091 		 * This one is a galactic mess.  We are going to roughly
4092 		 * emulate ddi_pathname(), but it's made more complicated
4093 		 * by the fact that we (a) want to include the minor name and
4094 		 * (b) must proceed iteratively instead of recursively.
4095 		 */
4096 		uintptr_t dest = mstate->dtms_scratch_ptr;
4097 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4098 		char *start = (char *)dest, *end = start + size - 1;
4099 		uintptr_t daddr = tupregs[0].dttk_value;
4100 		int64_t minor = (int64_t)tupregs[1].dttk_value;
4101 		char *s;
4102 		int i, len, depth = 0;
4103 
4104 		/*
4105 		 * Due to all the pointer jumping we do and context we must
4106 		 * rely upon, we just mandate that the user must have kernel
4107 		 * read privileges to use this routine.
4108 		 */
4109 		if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) == 0) {
4110 			*flags |= CPU_DTRACE_KPRIV;
4111 			*illval = daddr;
4112 			regs[rd] = 0;
4113 		}
4114 
4115 		if (!DTRACE_INSCRATCH(mstate, size)) {
4116 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4117 			regs[rd] = 0;
4118 			break;
4119 		}
4120 
4121 		*end = '\0';
4122 
4123 		/*
4124 		 * We want to have a name for the minor.  In order to do this,
4125 		 * we need to walk the minor list from the devinfo.  We want
4126 		 * to be sure that we don't infinitely walk a circular list,
4127 		 * so we check for circularity by sending a scout pointer
4128 		 * ahead two elements for every element that we iterate over;
4129 		 * if the list is circular, these will ultimately point to the
4130 		 * same element.  You may recognize this little trick as the
4131 		 * answer to a stupid interview question -- one that always
4132 		 * seems to be asked by those who had to have it laboriously
4133 		 * explained to them, and who can't even concisely describe
4134 		 * the conditions under which one would be forced to resort to
4135 		 * this technique.  Needless to say, those conditions are
4136 		 * found here -- and probably only here.  Is this the only use
4137 		 * of this infamous trick in shipping, production code?  If it
4138 		 * isn't, it probably should be...
4139 		 */
4140 		if (minor != -1) {
4141 			uintptr_t maddr = dtrace_loadptr(daddr +
4142 			    offsetof(struct dev_info, devi_minor));
4143 
4144 			uintptr_t next = offsetof(struct ddi_minor_data, next);
4145 			uintptr_t name = offsetof(struct ddi_minor_data,
4146 			    d_minor) + offsetof(struct ddi_minor, name);
4147 			uintptr_t dev = offsetof(struct ddi_minor_data,
4148 			    d_minor) + offsetof(struct ddi_minor, dev);
4149 			uintptr_t scout;
4150 
4151 			if (maddr != NULL)
4152 				scout = dtrace_loadptr(maddr + next);
4153 
4154 			while (maddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
4155 				uint64_t m;
4156 #ifdef _LP64
4157 				m = dtrace_load64(maddr + dev) & MAXMIN64;
4158 #else
4159 				m = dtrace_load32(maddr + dev) & MAXMIN;
4160 #endif
4161 				if (m != minor) {
4162 					maddr = dtrace_loadptr(maddr + next);
4163 
4164 					if (scout == NULL)
4165 						continue;
4166 
4167 					scout = dtrace_loadptr(scout + next);
4168 
4169 					if (scout == NULL)
4170 						continue;
4171 
4172 					scout = dtrace_loadptr(scout + next);
4173 
4174 					if (scout == NULL)
4175 						continue;
4176 
4177 					if (scout == maddr) {
4178 						*flags |= CPU_DTRACE_ILLOP;
4179 						break;
4180 					}
4181 
4182 					continue;
4183 				}
4184 
4185 				/*
4186 				 * We have the minor data.  Now we need to
4187 				 * copy the minor's name into the end of the
4188 				 * pathname.
4189 				 */
4190 				s = (char *)dtrace_loadptr(maddr + name);
4191 				len = dtrace_strlen(s, size);
4192 
4193 				if (*flags & CPU_DTRACE_FAULT)
4194 					break;
4195 
4196 				if (len != 0) {
4197 					if ((end -= (len + 1)) < start)
4198 						break;
4199 
4200 					*end = ':';
4201 				}
4202 
4203 				for (i = 1; i <= len; i++)
4204 					end[i] = dtrace_load8((uintptr_t)s++);
4205 				break;
4206 			}
4207 		}
4208 
4209 		while (daddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
4210 			ddi_node_state_t devi_state;
4211 
4212 			devi_state = dtrace_load32(daddr +
4213 			    offsetof(struct dev_info, devi_node_state));
4214 
4215 			if (*flags & CPU_DTRACE_FAULT)
4216 				break;
4217 
4218 			if (devi_state >= DS_INITIALIZED) {
4219 				s = (char *)dtrace_loadptr(daddr +
4220 				    offsetof(struct dev_info, devi_addr));
4221 				len = dtrace_strlen(s, size);
4222 
4223 				if (*flags & CPU_DTRACE_FAULT)
4224 					break;
4225 
4226 				if (len != 0) {
4227 					if ((end -= (len + 1)) < start)
4228 						break;
4229 
4230 					*end = '@';
4231 				}
4232 
4233 				for (i = 1; i <= len; i++)
4234 					end[i] = dtrace_load8((uintptr_t)s++);
4235 			}
4236 
4237 			/*
4238 			 * Now for the node name...
4239 			 */
4240 			s = (char *)dtrace_loadptr(daddr +
4241 			    offsetof(struct dev_info, devi_node_name));
4242 
4243 			daddr = dtrace_loadptr(daddr +
4244 			    offsetof(struct dev_info, devi_parent));
4245 
4246 			/*
4247 			 * If our parent is NULL (that is, if we're the root
4248 			 * node), we're going to use the special path
4249 			 * "devices".
4250 			 */
4251 			if (daddr == 0)
4252 				s = "devices";
4253 
4254 			len = dtrace_strlen(s, size);
4255 			if (*flags & CPU_DTRACE_FAULT)
4256 				break;
4257 
4258 			if ((end -= (len + 1)) < start)
4259 				break;
4260 
4261 			for (i = 1; i <= len; i++)
4262 				end[i] = dtrace_load8((uintptr_t)s++);
4263 			*end = '/';
4264 
4265 			if (depth++ > dtrace_devdepth_max) {
4266 				*flags |= CPU_DTRACE_ILLOP;
4267 				break;
4268 			}
4269 		}
4270 
4271 		if (end < start)
4272 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4273 
4274 		if (daddr == 0) {
4275 			regs[rd] = (uintptr_t)end;
4276 			mstate->dtms_scratch_ptr += size;
4277 		}
4278 
4279 		break;
4280 	}
4281 #endif
4282 
4283 	case DIF_SUBR_STRJOIN: {
4284 		char *d = (char *)mstate->dtms_scratch_ptr;
4285 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4286 		uintptr_t s1 = tupregs[0].dttk_value;
4287 		uintptr_t s2 = tupregs[1].dttk_value;
4288 		int i = 0;
4289 
4290 		if (!dtrace_strcanload(s1, size, mstate, vstate) ||
4291 		    !dtrace_strcanload(s2, size, mstate, vstate)) {
4292 			regs[rd] = 0;
4293 			break;
4294 		}
4295 
4296 		if (!DTRACE_INSCRATCH(mstate, size)) {
4297 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4298 			regs[rd] = 0;
4299 			break;
4300 		}
4301 
4302 		for (;;) {
4303 			if (i >= size) {
4304 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4305 				regs[rd] = 0;
4306 				break;
4307 			}
4308 
4309 			if ((d[i++] = dtrace_load8(s1++)) == '\0') {
4310 				i--;
4311 				break;
4312 			}
4313 		}
4314 
4315 		for (;;) {
4316 			if (i >= size) {
4317 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4318 				regs[rd] = 0;
4319 				break;
4320 			}
4321 
4322 			if ((d[i++] = dtrace_load8(s2++)) == '\0')
4323 				break;
4324 		}
4325 
4326 		if (i < size) {
4327 			mstate->dtms_scratch_ptr += i;
4328 			regs[rd] = (uintptr_t)d;
4329 		}
4330 
4331 		break;
4332 	}
4333 
4334 	case DIF_SUBR_LLTOSTR: {
4335 		int64_t i = (int64_t)tupregs[0].dttk_value;
4336 		uint64_t val, digit;
4337 		uint64_t size = 65;	/* enough room for 2^64 in binary */
4338 		char *end = (char *)mstate->dtms_scratch_ptr + size - 1;
4339 		int base = 10;
4340 
4341 		if (nargs > 1) {
4342 			if ((base = tupregs[1].dttk_value) <= 1 ||
4343 			    base > ('z' - 'a' + 1) + ('9' - '0' + 1)) {
4344 				*flags |= CPU_DTRACE_ILLOP;
4345 				break;
4346 			}
4347 		}
4348 
4349 		val = (base == 10 && i < 0) ? i * -1 : i;
4350 
4351 		if (!DTRACE_INSCRATCH(mstate, size)) {
4352 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4353 			regs[rd] = 0;
4354 			break;
4355 		}
4356 
4357 		for (*end-- = '\0'; val; val /= base) {
4358 			if ((digit = val % base) <= '9' - '0') {
4359 				*end-- = '0' + digit;
4360 			} else {
4361 				*end-- = 'a' + (digit - ('9' - '0') - 1);
4362 			}
4363 		}
4364 
4365 		if (i == 0 && base == 16)
4366 			*end-- = '0';
4367 
4368 		if (base == 16)
4369 			*end-- = 'x';
4370 
4371 		if (i == 0 || base == 8 || base == 16)
4372 			*end-- = '0';
4373 
4374 		if (i < 0 && base == 10)
4375 			*end-- = '-';
4376 
4377 		regs[rd] = (uintptr_t)end + 1;
4378 		mstate->dtms_scratch_ptr += size;
4379 		break;
4380 	}
4381 
4382 	case DIF_SUBR_HTONS:
4383 	case DIF_SUBR_NTOHS:
4384 #if BYTE_ORDER == BIG_ENDIAN
4385 		regs[rd] = (uint16_t)tupregs[0].dttk_value;
4386 #else
4387 		regs[rd] = DT_BSWAP_16((uint16_t)tupregs[0].dttk_value);
4388 #endif
4389 		break;
4390 
4391 
4392 	case DIF_SUBR_HTONL:
4393 	case DIF_SUBR_NTOHL:
4394 #if BYTE_ORDER == BIG_ENDIAN
4395 		regs[rd] = (uint32_t)tupregs[0].dttk_value;
4396 #else
4397 		regs[rd] = DT_BSWAP_32((uint32_t)tupregs[0].dttk_value);
4398 #endif
4399 		break;
4400 
4401 
4402 	case DIF_SUBR_HTONLL:
4403 	case DIF_SUBR_NTOHLL:
4404 #if BYTE_ORDER == BIG_ENDIAN
4405 		regs[rd] = (uint64_t)tupregs[0].dttk_value;
4406 #else
4407 		regs[rd] = DT_BSWAP_64((uint64_t)tupregs[0].dttk_value);
4408 #endif
4409 		break;
4410 
4411 
4412 	case DIF_SUBR_DIRNAME:
4413 	case DIF_SUBR_BASENAME: {
4414 		char *dest = (char *)mstate->dtms_scratch_ptr;
4415 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4416 		uintptr_t src = tupregs[0].dttk_value;
4417 		int i, j, len = dtrace_strlen((char *)src, size);
4418 		int lastbase = -1, firstbase = -1, lastdir = -1;
4419 		int start, end;
4420 
4421 		if (!dtrace_canload(src, len + 1, mstate, vstate)) {
4422 			regs[rd] = 0;
4423 			break;
4424 		}
4425 
4426 		if (!DTRACE_INSCRATCH(mstate, size)) {
4427 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4428 			regs[rd] = 0;
4429 			break;
4430 		}
4431 
4432 		/*
4433 		 * The basename and dirname for a zero-length string is
4434 		 * defined to be "."
4435 		 */
4436 		if (len == 0) {
4437 			len = 1;
4438 			src = (uintptr_t)".";
4439 		}
4440 
4441 		/*
4442 		 * Start from the back of the string, moving back toward the
4443 		 * front until we see a character that isn't a slash.  That
4444 		 * character is the last character in the basename.
4445 		 */
4446 		for (i = len - 1; i >= 0; i--) {
4447 			if (dtrace_load8(src + i) != '/')
4448 				break;
4449 		}
4450 
4451 		if (i >= 0)
4452 			lastbase = i;
4453 
4454 		/*
4455 		 * Starting from the last character in the basename, move
4456 		 * towards the front until we find a slash.  The character
4457 		 * that we processed immediately before that is the first
4458 		 * character in the basename.
4459 		 */
4460 		for (; i >= 0; i--) {
4461 			if (dtrace_load8(src + i) == '/')
4462 				break;
4463 		}
4464 
4465 		if (i >= 0)
4466 			firstbase = i + 1;
4467 
4468 		/*
4469 		 * Now keep going until we find a non-slash character.  That
4470 		 * character is the last character in the dirname.
4471 		 */
4472 		for (; i >= 0; i--) {
4473 			if (dtrace_load8(src + i) != '/')
4474 				break;
4475 		}
4476 
4477 		if (i >= 0)
4478 			lastdir = i;
4479 
4480 		ASSERT(!(lastbase == -1 && firstbase != -1));
4481 		ASSERT(!(firstbase == -1 && lastdir != -1));
4482 
4483 		if (lastbase == -1) {
4484 			/*
4485 			 * We didn't find a non-slash character.  We know that
4486 			 * the length is non-zero, so the whole string must be
4487 			 * slashes.  In either the dirname or the basename
4488 			 * case, we return '/'.
4489 			 */
4490 			ASSERT(firstbase == -1);
4491 			firstbase = lastbase = lastdir = 0;
4492 		}
4493 
4494 		if (firstbase == -1) {
4495 			/*
4496 			 * The entire string consists only of a basename
4497 			 * component.  If we're looking for dirname, we need
4498 			 * to change our string to be just "."; if we're
4499 			 * looking for a basename, we'll just set the first
4500 			 * character of the basename to be 0.
4501 			 */
4502 			if (subr == DIF_SUBR_DIRNAME) {
4503 				ASSERT(lastdir == -1);
4504 				src = (uintptr_t)".";
4505 				lastdir = 0;
4506 			} else {
4507 				firstbase = 0;
4508 			}
4509 		}
4510 
4511 		if (subr == DIF_SUBR_DIRNAME) {
4512 			if (lastdir == -1) {
4513 				/*
4514 				 * We know that we have a slash in the name --
4515 				 * or lastdir would be set to 0, above.  And
4516 				 * because lastdir is -1, we know that this
4517 				 * slash must be the first character.  (That
4518 				 * is, the full string must be of the form
4519 				 * "/basename".)  In this case, the last
4520 				 * character of the directory name is 0.
4521 				 */
4522 				lastdir = 0;
4523 			}
4524 
4525 			start = 0;
4526 			end = lastdir;
4527 		} else {
4528 			ASSERT(subr == DIF_SUBR_BASENAME);
4529 			ASSERT(firstbase != -1 && lastbase != -1);
4530 			start = firstbase;
4531 			end = lastbase;
4532 		}
4533 
4534 		for (i = start, j = 0; i <= end && j < size - 1; i++, j++)
4535 			dest[j] = dtrace_load8(src + i);
4536 
4537 		dest[j] = '\0';
4538 		regs[rd] = (uintptr_t)dest;
4539 		mstate->dtms_scratch_ptr += size;
4540 		break;
4541 	}
4542 
4543 	case DIF_SUBR_CLEANPATH: {
4544 		char *dest = (char *)mstate->dtms_scratch_ptr, c;
4545 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4546 		uintptr_t src = tupregs[0].dttk_value;
4547 		int i = 0, j = 0;
4548 
4549 		if (!dtrace_strcanload(src, size, mstate, vstate)) {
4550 			regs[rd] = 0;
4551 			break;
4552 		}
4553 
4554 		if (!DTRACE_INSCRATCH(mstate, size)) {
4555 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4556 			regs[rd] = 0;
4557 			break;
4558 		}
4559 
4560 		/*
4561 		 * Move forward, loading each character.
4562 		 */
4563 		do {
4564 			c = dtrace_load8(src + i++);
4565 next:
4566 			if (j + 5 >= size)	/* 5 = strlen("/..c\0") */
4567 				break;
4568 
4569 			if (c != '/') {
4570 				dest[j++] = c;
4571 				continue;
4572 			}
4573 
4574 			c = dtrace_load8(src + i++);
4575 
4576 			if (c == '/') {
4577 				/*
4578 				 * We have two slashes -- we can just advance
4579 				 * to the next character.
4580 				 */
4581 				goto next;
4582 			}
4583 
4584 			if (c != '.') {
4585 				/*
4586 				 * This is not "." and it's not ".." -- we can
4587 				 * just store the "/" and this character and
4588 				 * drive on.
4589 				 */
4590 				dest[j++] = '/';
4591 				dest[j++] = c;
4592 				continue;
4593 			}
4594 
4595 			c = dtrace_load8(src + i++);
4596 
4597 			if (c == '/') {
4598 				/*
4599 				 * This is a "/./" component.  We're not going
4600 				 * to store anything in the destination buffer;
4601 				 * we're just going to go to the next component.
4602 				 */
4603 				goto next;
4604 			}
4605 
4606 			if (c != '.') {
4607 				/*
4608 				 * This is not ".." -- we can just store the
4609 				 * "/." and this character and continue
4610 				 * processing.
4611 				 */
4612 				dest[j++] = '/';
4613 				dest[j++] = '.';
4614 				dest[j++] = c;
4615 				continue;
4616 			}
4617 
4618 			c = dtrace_load8(src + i++);
4619 
4620 			if (c != '/' && c != '\0') {
4621 				/*
4622 				 * This is not ".." -- it's "..[mumble]".
4623 				 * We'll store the "/.." and this character
4624 				 * and continue processing.
4625 				 */
4626 				dest[j++] = '/';
4627 				dest[j++] = '.';
4628 				dest[j++] = '.';
4629 				dest[j++] = c;
4630 				continue;
4631 			}
4632 
4633 			/*
4634 			 * This is "/../" or "/..\0".  We need to back up
4635 			 * our destination pointer until we find a "/".
4636 			 */
4637 			i--;
4638 			while (j != 0 && dest[--j] != '/')
4639 				continue;
4640 
4641 			if (c == '\0')
4642 				dest[++j] = '/';
4643 		} while (c != '\0');
4644 
4645 		dest[j] = '\0';
4646 		regs[rd] = (uintptr_t)dest;
4647 		mstate->dtms_scratch_ptr += size;
4648 		break;
4649 	}
4650 
4651 	case DIF_SUBR_INET_NTOA:
4652 	case DIF_SUBR_INET_NTOA6:
4653 	case DIF_SUBR_INET_NTOP: {
4654 		size_t size;
4655 		int af, argi, i;
4656 		char *base, *end;
4657 
4658 		if (subr == DIF_SUBR_INET_NTOP) {
4659 			af = (int)tupregs[0].dttk_value;
4660 			argi = 1;
4661 		} else {
4662 			af = subr == DIF_SUBR_INET_NTOA ? AF_INET: AF_INET6;
4663 			argi = 0;
4664 		}
4665 
4666 		if (af == AF_INET) {
4667 			ipaddr_t ip4;
4668 			uint8_t *ptr8, val;
4669 
4670 			/*
4671 			 * Safely load the IPv4 address.
4672 			 */
4673 			ip4 = dtrace_load32(tupregs[argi].dttk_value);
4674 
4675 			/*
4676 			 * Check an IPv4 string will fit in scratch.
4677 			 */
4678 			size = INET_ADDRSTRLEN;
4679 			if (!DTRACE_INSCRATCH(mstate, size)) {
4680 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4681 				regs[rd] = 0;
4682 				break;
4683 			}
4684 			base = (char *)mstate->dtms_scratch_ptr;
4685 			end = (char *)mstate->dtms_scratch_ptr + size - 1;
4686 
4687 			/*
4688 			 * Stringify as a dotted decimal quad.
4689 			 */
4690 			*end-- = '\0';
4691 			ptr8 = (uint8_t *)&ip4;
4692 			for (i = 3; i >= 0; i--) {
4693 				val = ptr8[i];
4694 
4695 				if (val == 0) {
4696 					*end-- = '0';
4697 				} else {
4698 					for (; val; val /= 10) {
4699 						*end-- = '0' + (val % 10);
4700 					}
4701 				}
4702 
4703 				if (i > 0)
4704 					*end-- = '.';
4705 			}
4706 			ASSERT(end + 1 >= base);
4707 
4708 		} else if (af == AF_INET6) {
4709 			struct in6_addr ip6;
4710 			int firstzero, tryzero, numzero, v6end;
4711 			uint16_t val;
4712 			const char digits[] = "0123456789abcdef";
4713 
4714 			/*
4715 			 * Stringify using RFC 1884 convention 2 - 16 bit
4716 			 * hexadecimal values with a zero-run compression.
4717 			 * Lower case hexadecimal digits are used.
4718 			 * 	eg, fe80::214:4fff:fe0b:76c8.
4719 			 * The IPv4 embedded form is returned for inet_ntop,
4720 			 * just the IPv4 string is returned for inet_ntoa6.
4721 			 */
4722 
4723 			/*
4724 			 * Safely load the IPv6 address.
4725 			 */
4726 			dtrace_bcopy(
4727 			    (void *)(uintptr_t)tupregs[argi].dttk_value,
4728 			    (void *)(uintptr_t)&ip6, sizeof (struct in6_addr));
4729 
4730 			/*
4731 			 * Check an IPv6 string will fit in scratch.
4732 			 */
4733 			size = INET6_ADDRSTRLEN;
4734 			if (!DTRACE_INSCRATCH(mstate, size)) {
4735 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4736 				regs[rd] = 0;
4737 				break;
4738 			}
4739 			base = (char *)mstate->dtms_scratch_ptr;
4740 			end = (char *)mstate->dtms_scratch_ptr + size - 1;
4741 			*end-- = '\0';
4742 
4743 			/*
4744 			 * Find the longest run of 16 bit zero values
4745 			 * for the single allowed zero compression - "::".
4746 			 */
4747 			firstzero = -1;
4748 			tryzero = -1;
4749 			numzero = 1;
4750 			for (i = 0; i < sizeof (struct in6_addr); i++) {
4751 #if defined(sun)
4752 				if (ip6._S6_un._S6_u8[i] == 0 &&
4753 #else
4754 				if (ip6.__u6_addr.__u6_addr8[i] == 0 &&
4755 #endif
4756 				    tryzero == -1 && i % 2 == 0) {
4757 					tryzero = i;
4758 					continue;
4759 				}
4760 
4761 				if (tryzero != -1 &&
4762 #if defined(sun)
4763 				    (ip6._S6_un._S6_u8[i] != 0 ||
4764 #else
4765 				    (ip6.__u6_addr.__u6_addr8[i] != 0 ||
4766 #endif
4767 				    i == sizeof (struct in6_addr) - 1)) {
4768 
4769 					if (i - tryzero <= numzero) {
4770 						tryzero = -1;
4771 						continue;
4772 					}
4773 
4774 					firstzero = tryzero;
4775 					numzero = i - i % 2 - tryzero;
4776 					tryzero = -1;
4777 
4778 #if defined(sun)
4779 					if (ip6._S6_un._S6_u8[i] == 0 &&
4780 #else
4781 					if (ip6.__u6_addr.__u6_addr8[i] == 0 &&
4782 #endif
4783 					    i == sizeof (struct in6_addr) - 1)
4784 						numzero += 2;
4785 				}
4786 			}
4787 			ASSERT(firstzero + numzero <= sizeof (struct in6_addr));
4788 
4789 			/*
4790 			 * Check for an IPv4 embedded address.
4791 			 */
4792 			v6end = sizeof (struct in6_addr) - 2;
4793 			if (IN6_IS_ADDR_V4MAPPED(&ip6) ||
4794 			    IN6_IS_ADDR_V4COMPAT(&ip6)) {
4795 				for (i = sizeof (struct in6_addr) - 1;
4796 				    i >= DTRACE_V4MAPPED_OFFSET; i--) {
4797 					ASSERT(end >= base);
4798 
4799 #if defined(sun)
4800 					val = ip6._S6_un._S6_u8[i];
4801 #else
4802 					val = ip6.__u6_addr.__u6_addr8[i];
4803 #endif
4804 
4805 					if (val == 0) {
4806 						*end-- = '0';
4807 					} else {
4808 						for (; val; val /= 10) {
4809 							*end-- = '0' + val % 10;
4810 						}
4811 					}
4812 
4813 					if (i > DTRACE_V4MAPPED_OFFSET)
4814 						*end-- = '.';
4815 				}
4816 
4817 				if (subr == DIF_SUBR_INET_NTOA6)
4818 					goto inetout;
4819 
4820 				/*
4821 				 * Set v6end to skip the IPv4 address that
4822 				 * we have already stringified.
4823 				 */
4824 				v6end = 10;
4825 			}
4826 
4827 			/*
4828 			 * Build the IPv6 string by working through the
4829 			 * address in reverse.
4830 			 */
4831 			for (i = v6end; i >= 0; i -= 2) {
4832 				ASSERT(end >= base);
4833 
4834 				if (i == firstzero + numzero - 2) {
4835 					*end-- = ':';
4836 					*end-- = ':';
4837 					i -= numzero - 2;
4838 					continue;
4839 				}
4840 
4841 				if (i < 14 && i != firstzero - 2)
4842 					*end-- = ':';
4843 
4844 #if defined(sun)
4845 				val = (ip6._S6_un._S6_u8[i] << 8) +
4846 				    ip6._S6_un._S6_u8[i + 1];
4847 #else
4848 				val = (ip6.__u6_addr.__u6_addr8[i] << 8) +
4849 				    ip6.__u6_addr.__u6_addr8[i + 1];
4850 #endif
4851 
4852 				if (val == 0) {
4853 					*end-- = '0';
4854 				} else {
4855 					for (; val; val /= 16) {
4856 						*end-- = digits[val % 16];
4857 					}
4858 				}
4859 			}
4860 			ASSERT(end + 1 >= base);
4861 
4862 		} else {
4863 			/*
4864 			 * The user didn't use AH_INET or AH_INET6.
4865 			 */
4866 			DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
4867 			regs[rd] = 0;
4868 			break;
4869 		}
4870 
4871 inetout:	regs[rd] = (uintptr_t)end + 1;
4872 		mstate->dtms_scratch_ptr += size;
4873 		break;
4874 	}
4875 
4876 	case DIF_SUBR_MEMREF: {
4877 		uintptr_t size = 2 * sizeof(uintptr_t);
4878 		uintptr_t *memref = (uintptr_t *) P2ROUNDUP(mstate->dtms_scratch_ptr, sizeof(uintptr_t));
4879 		size_t scratch_size = ((uintptr_t) memref - mstate->dtms_scratch_ptr) + size;
4880 
4881 		/* address and length */
4882 		memref[0] = tupregs[0].dttk_value;
4883 		memref[1] = tupregs[1].dttk_value;
4884 
4885 		regs[rd] = (uintptr_t) memref;
4886 		mstate->dtms_scratch_ptr += scratch_size;
4887 		break;
4888 	}
4889 
4890 	case DIF_SUBR_TYPEREF: {
4891 		uintptr_t size = 4 * sizeof(uintptr_t);
4892 		uintptr_t *typeref = (uintptr_t *) P2ROUNDUP(mstate->dtms_scratch_ptr, sizeof(uintptr_t));
4893 		size_t scratch_size = ((uintptr_t) typeref - mstate->dtms_scratch_ptr) + size;
4894 
4895 		/* address, num_elements, type_str, type_len */
4896 		typeref[0] = tupregs[0].dttk_value;
4897 		typeref[1] = tupregs[1].dttk_value;
4898 		typeref[2] = tupregs[2].dttk_value;
4899 		typeref[3] = tupregs[3].dttk_value;
4900 
4901 		regs[rd] = (uintptr_t) typeref;
4902 		mstate->dtms_scratch_ptr += scratch_size;
4903 		break;
4904 	}
4905 	}
4906 }
4907 
4908 /*
4909  * Emulate the execution of DTrace IR instructions specified by the given
4910  * DIF object.  This function is deliberately void of assertions as all of
4911  * the necessary checks are handled by a call to dtrace_difo_validate().
4912  */
4913 static uint64_t
4914 dtrace_dif_emulate(dtrace_difo_t *difo, dtrace_mstate_t *mstate,
4915     dtrace_vstate_t *vstate, dtrace_state_t *state)
4916 {
4917 	const dif_instr_t *text = difo->dtdo_buf;
4918 	const uint_t textlen = difo->dtdo_len;
4919 	const char *strtab = difo->dtdo_strtab;
4920 	const uint64_t *inttab = difo->dtdo_inttab;
4921 
4922 	uint64_t rval = 0;
4923 	dtrace_statvar_t *svar;
4924 	dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
4925 	dtrace_difv_t *v;
4926 	volatile uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
4927 	volatile uintptr_t *illval = &cpu_core[curcpu].cpuc_dtrace_illval;
4928 
4929 	dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
4930 	uint64_t regs[DIF_DIR_NREGS];
4931 	uint64_t *tmp;
4932 
4933 	uint8_t cc_n = 0, cc_z = 0, cc_v = 0, cc_c = 0;
4934 	int64_t cc_r;
4935 	uint_t pc = 0, id, opc = 0;
4936 	uint8_t ttop = 0;
4937 	dif_instr_t instr;
4938 	uint_t r1, r2, rd;
4939 
4940 	/*
4941 	 * We stash the current DIF object into the machine state: we need it
4942 	 * for subsequent access checking.
4943 	 */
4944 	mstate->dtms_difo = difo;
4945 
4946 	regs[DIF_REG_R0] = 0; 		/* %r0 is fixed at zero */
4947 
4948 	while (pc < textlen && !(*flags & CPU_DTRACE_FAULT)) {
4949 		opc = pc;
4950 
4951 		instr = text[pc++];
4952 		r1 = DIF_INSTR_R1(instr);
4953 		r2 = DIF_INSTR_R2(instr);
4954 		rd = DIF_INSTR_RD(instr);
4955 
4956 		switch (DIF_INSTR_OP(instr)) {
4957 		case DIF_OP_OR:
4958 			regs[rd] = regs[r1] | regs[r2];
4959 			break;
4960 		case DIF_OP_XOR:
4961 			regs[rd] = regs[r1] ^ regs[r2];
4962 			break;
4963 		case DIF_OP_AND:
4964 			regs[rd] = regs[r1] & regs[r2];
4965 			break;
4966 		case DIF_OP_SLL:
4967 			regs[rd] = regs[r1] << regs[r2];
4968 			break;
4969 		case DIF_OP_SRL:
4970 			regs[rd] = regs[r1] >> regs[r2];
4971 			break;
4972 		case DIF_OP_SUB:
4973 			regs[rd] = regs[r1] - regs[r2];
4974 			break;
4975 		case DIF_OP_ADD:
4976 			regs[rd] = regs[r1] + regs[r2];
4977 			break;
4978 		case DIF_OP_MUL:
4979 			regs[rd] = regs[r1] * regs[r2];
4980 			break;
4981 		case DIF_OP_SDIV:
4982 			if (regs[r2] == 0) {
4983 				regs[rd] = 0;
4984 				*flags |= CPU_DTRACE_DIVZERO;
4985 			} else {
4986 				regs[rd] = (int64_t)regs[r1] /
4987 				    (int64_t)regs[r2];
4988 			}
4989 			break;
4990 
4991 		case DIF_OP_UDIV:
4992 			if (regs[r2] == 0) {
4993 				regs[rd] = 0;
4994 				*flags |= CPU_DTRACE_DIVZERO;
4995 			} else {
4996 				regs[rd] = regs[r1] / regs[r2];
4997 			}
4998 			break;
4999 
5000 		case DIF_OP_SREM:
5001 			if (regs[r2] == 0) {
5002 				regs[rd] = 0;
5003 				*flags |= CPU_DTRACE_DIVZERO;
5004 			} else {
5005 				regs[rd] = (int64_t)regs[r1] %
5006 				    (int64_t)regs[r2];
5007 			}
5008 			break;
5009 
5010 		case DIF_OP_UREM:
5011 			if (regs[r2] == 0) {
5012 				regs[rd] = 0;
5013 				*flags |= CPU_DTRACE_DIVZERO;
5014 			} else {
5015 				regs[rd] = regs[r1] % regs[r2];
5016 			}
5017 			break;
5018 
5019 		case DIF_OP_NOT:
5020 			regs[rd] = ~regs[r1];
5021 			break;
5022 		case DIF_OP_MOV:
5023 			regs[rd] = regs[r1];
5024 			break;
5025 		case DIF_OP_CMP:
5026 			cc_r = regs[r1] - regs[r2];
5027 			cc_n = cc_r < 0;
5028 			cc_z = cc_r == 0;
5029 			cc_v = 0;
5030 			cc_c = regs[r1] < regs[r2];
5031 			break;
5032 		case DIF_OP_TST:
5033 			cc_n = cc_v = cc_c = 0;
5034 			cc_z = regs[r1] == 0;
5035 			break;
5036 		case DIF_OP_BA:
5037 			pc = DIF_INSTR_LABEL(instr);
5038 			break;
5039 		case DIF_OP_BE:
5040 			if (cc_z)
5041 				pc = DIF_INSTR_LABEL(instr);
5042 			break;
5043 		case DIF_OP_BNE:
5044 			if (cc_z == 0)
5045 				pc = DIF_INSTR_LABEL(instr);
5046 			break;
5047 		case DIF_OP_BG:
5048 			if ((cc_z | (cc_n ^ cc_v)) == 0)
5049 				pc = DIF_INSTR_LABEL(instr);
5050 			break;
5051 		case DIF_OP_BGU:
5052 			if ((cc_c | cc_z) == 0)
5053 				pc = DIF_INSTR_LABEL(instr);
5054 			break;
5055 		case DIF_OP_BGE:
5056 			if ((cc_n ^ cc_v) == 0)
5057 				pc = DIF_INSTR_LABEL(instr);
5058 			break;
5059 		case DIF_OP_BGEU:
5060 			if (cc_c == 0)
5061 				pc = DIF_INSTR_LABEL(instr);
5062 			break;
5063 		case DIF_OP_BL:
5064 			if (cc_n ^ cc_v)
5065 				pc = DIF_INSTR_LABEL(instr);
5066 			break;
5067 		case DIF_OP_BLU:
5068 			if (cc_c)
5069 				pc = DIF_INSTR_LABEL(instr);
5070 			break;
5071 		case DIF_OP_BLE:
5072 			if (cc_z | (cc_n ^ cc_v))
5073 				pc = DIF_INSTR_LABEL(instr);
5074 			break;
5075 		case DIF_OP_BLEU:
5076 			if (cc_c | cc_z)
5077 				pc = DIF_INSTR_LABEL(instr);
5078 			break;
5079 		case DIF_OP_RLDSB:
5080 			if (!dtrace_canstore(regs[r1], 1, mstate, vstate)) {
5081 				*flags |= CPU_DTRACE_KPRIV;
5082 				*illval = regs[r1];
5083 				break;
5084 			}
5085 			/*FALLTHROUGH*/
5086 		case DIF_OP_LDSB:
5087 			regs[rd] = (int8_t)dtrace_load8(regs[r1]);
5088 			break;
5089 		case DIF_OP_RLDSH:
5090 			if (!dtrace_canstore(regs[r1], 2, mstate, vstate)) {
5091 				*flags |= CPU_DTRACE_KPRIV;
5092 				*illval = regs[r1];
5093 				break;
5094 			}
5095 			/*FALLTHROUGH*/
5096 		case DIF_OP_LDSH:
5097 			regs[rd] = (int16_t)dtrace_load16(regs[r1]);
5098 			break;
5099 		case DIF_OP_RLDSW:
5100 			if (!dtrace_canstore(regs[r1], 4, mstate, vstate)) {
5101 				*flags |= CPU_DTRACE_KPRIV;
5102 				*illval = regs[r1];
5103 				break;
5104 			}
5105 			/*FALLTHROUGH*/
5106 		case DIF_OP_LDSW:
5107 			regs[rd] = (int32_t)dtrace_load32(regs[r1]);
5108 			break;
5109 		case DIF_OP_RLDUB:
5110 			if (!dtrace_canstore(regs[r1], 1, mstate, vstate)) {
5111 				*flags |= CPU_DTRACE_KPRIV;
5112 				*illval = regs[r1];
5113 				break;
5114 			}
5115 			/*FALLTHROUGH*/
5116 		case DIF_OP_LDUB:
5117 			regs[rd] = dtrace_load8(regs[r1]);
5118 			break;
5119 		case DIF_OP_RLDUH:
5120 			if (!dtrace_canstore(regs[r1], 2, mstate, vstate)) {
5121 				*flags |= CPU_DTRACE_KPRIV;
5122 				*illval = regs[r1];
5123 				break;
5124 			}
5125 			/*FALLTHROUGH*/
5126 		case DIF_OP_LDUH:
5127 			regs[rd] = dtrace_load16(regs[r1]);
5128 			break;
5129 		case DIF_OP_RLDUW:
5130 			if (!dtrace_canstore(regs[r1], 4, mstate, vstate)) {
5131 				*flags |= CPU_DTRACE_KPRIV;
5132 				*illval = regs[r1];
5133 				break;
5134 			}
5135 			/*FALLTHROUGH*/
5136 		case DIF_OP_LDUW:
5137 			regs[rd] = dtrace_load32(regs[r1]);
5138 			break;
5139 		case DIF_OP_RLDX:
5140 			if (!dtrace_canstore(regs[r1], 8, mstate, vstate)) {
5141 				*flags |= CPU_DTRACE_KPRIV;
5142 				*illval = regs[r1];
5143 				break;
5144 			}
5145 			/*FALLTHROUGH*/
5146 		case DIF_OP_LDX:
5147 			regs[rd] = dtrace_load64(regs[r1]);
5148 			break;
5149 		case DIF_OP_ULDSB:
5150 			regs[rd] = (int8_t)
5151 			    dtrace_fuword8((void *)(uintptr_t)regs[r1]);
5152 			break;
5153 		case DIF_OP_ULDSH:
5154 			regs[rd] = (int16_t)
5155 			    dtrace_fuword16((void *)(uintptr_t)regs[r1]);
5156 			break;
5157 		case DIF_OP_ULDSW:
5158 			regs[rd] = (int32_t)
5159 			    dtrace_fuword32((void *)(uintptr_t)regs[r1]);
5160 			break;
5161 		case DIF_OP_ULDUB:
5162 			regs[rd] =
5163 			    dtrace_fuword8((void *)(uintptr_t)regs[r1]);
5164 			break;
5165 		case DIF_OP_ULDUH:
5166 			regs[rd] =
5167 			    dtrace_fuword16((void *)(uintptr_t)regs[r1]);
5168 			break;
5169 		case DIF_OP_ULDUW:
5170 			regs[rd] =
5171 			    dtrace_fuword32((void *)(uintptr_t)regs[r1]);
5172 			break;
5173 		case DIF_OP_ULDX:
5174 			regs[rd] =
5175 			    dtrace_fuword64((void *)(uintptr_t)regs[r1]);
5176 			break;
5177 		case DIF_OP_RET:
5178 			rval = regs[rd];
5179 			pc = textlen;
5180 			break;
5181 		case DIF_OP_NOP:
5182 			break;
5183 		case DIF_OP_SETX:
5184 			regs[rd] = inttab[DIF_INSTR_INTEGER(instr)];
5185 			break;
5186 		case DIF_OP_SETS:
5187 			regs[rd] = (uint64_t)(uintptr_t)
5188 			    (strtab + DIF_INSTR_STRING(instr));
5189 			break;
5190 		case DIF_OP_SCMP: {
5191 			size_t sz = state->dts_options[DTRACEOPT_STRSIZE];
5192 			uintptr_t s1 = regs[r1];
5193 			uintptr_t s2 = regs[r2];
5194 
5195 			if (s1 != 0 &&
5196 			    !dtrace_strcanload(s1, sz, mstate, vstate))
5197 				break;
5198 			if (s2 != 0 &&
5199 			    !dtrace_strcanload(s2, sz, mstate, vstate))
5200 				break;
5201 
5202 			cc_r = dtrace_strncmp((char *)s1, (char *)s2, sz);
5203 
5204 			cc_n = cc_r < 0;
5205 			cc_z = cc_r == 0;
5206 			cc_v = cc_c = 0;
5207 			break;
5208 		}
5209 		case DIF_OP_LDGA:
5210 			regs[rd] = dtrace_dif_variable(mstate, state,
5211 			    r1, regs[r2]);
5212 			break;
5213 		case DIF_OP_LDGS:
5214 			id = DIF_INSTR_VAR(instr);
5215 
5216 			if (id >= DIF_VAR_OTHER_UBASE) {
5217 				uintptr_t a;
5218 
5219 				id -= DIF_VAR_OTHER_UBASE;
5220 				svar = vstate->dtvs_globals[id];
5221 				ASSERT(svar != NULL);
5222 				v = &svar->dtsv_var;
5223 
5224 				if (!(v->dtdv_type.dtdt_flags & DIF_TF_BYREF)) {
5225 					regs[rd] = svar->dtsv_data;
5226 					break;
5227 				}
5228 
5229 				a = (uintptr_t)svar->dtsv_data;
5230 
5231 				if (*(uint8_t *)a == UINT8_MAX) {
5232 					/*
5233 					 * If the 0th byte is set to UINT8_MAX
5234 					 * then this is to be treated as a
5235 					 * reference to a NULL variable.
5236 					 */
5237 					regs[rd] = 0;
5238 				} else {
5239 					regs[rd] = a + sizeof (uint64_t);
5240 				}
5241 
5242 				break;
5243 			}
5244 
5245 			regs[rd] = dtrace_dif_variable(mstate, state, id, 0);
5246 			break;
5247 
5248 		case DIF_OP_STGS:
5249 			id = DIF_INSTR_VAR(instr);
5250 
5251 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
5252 			id -= DIF_VAR_OTHER_UBASE;
5253 
5254 			svar = vstate->dtvs_globals[id];
5255 			ASSERT(svar != NULL);
5256 			v = &svar->dtsv_var;
5257 
5258 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5259 				uintptr_t a = (uintptr_t)svar->dtsv_data;
5260 
5261 				ASSERT(a != 0);
5262 				ASSERT(svar->dtsv_size != 0);
5263 
5264 				if (regs[rd] == 0) {
5265 					*(uint8_t *)a = UINT8_MAX;
5266 					break;
5267 				} else {
5268 					*(uint8_t *)a = 0;
5269 					a += sizeof (uint64_t);
5270 				}
5271 				if (!dtrace_vcanload(
5272 				    (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5273 				    mstate, vstate))
5274 					break;
5275 
5276 				dtrace_vcopy((void *)(uintptr_t)regs[rd],
5277 				    (void *)a, &v->dtdv_type);
5278 				break;
5279 			}
5280 
5281 			svar->dtsv_data = regs[rd];
5282 			break;
5283 
5284 		case DIF_OP_LDTA:
5285 			/*
5286 			 * There are no DTrace built-in thread-local arrays at
5287 			 * present.  This opcode is saved for future work.
5288 			 */
5289 			*flags |= CPU_DTRACE_ILLOP;
5290 			regs[rd] = 0;
5291 			break;
5292 
5293 		case DIF_OP_LDLS:
5294 			id = DIF_INSTR_VAR(instr);
5295 
5296 			if (id < DIF_VAR_OTHER_UBASE) {
5297 				/*
5298 				 * For now, this has no meaning.
5299 				 */
5300 				regs[rd] = 0;
5301 				break;
5302 			}
5303 
5304 			id -= DIF_VAR_OTHER_UBASE;
5305 
5306 			ASSERT(id < vstate->dtvs_nlocals);
5307 			ASSERT(vstate->dtvs_locals != NULL);
5308 
5309 			svar = vstate->dtvs_locals[id];
5310 			ASSERT(svar != NULL);
5311 			v = &svar->dtsv_var;
5312 
5313 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5314 				uintptr_t a = (uintptr_t)svar->dtsv_data;
5315 				size_t sz = v->dtdv_type.dtdt_size;
5316 
5317 				sz += sizeof (uint64_t);
5318 				ASSERT(svar->dtsv_size == NCPU * sz);
5319 				a += curcpu * sz;
5320 
5321 				if (*(uint8_t *)a == UINT8_MAX) {
5322 					/*
5323 					 * If the 0th byte is set to UINT8_MAX
5324 					 * then this is to be treated as a
5325 					 * reference to a NULL variable.
5326 					 */
5327 					regs[rd] = 0;
5328 				} else {
5329 					regs[rd] = a + sizeof (uint64_t);
5330 				}
5331 
5332 				break;
5333 			}
5334 
5335 			ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
5336 			tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
5337 			regs[rd] = tmp[curcpu];
5338 			break;
5339 
5340 		case DIF_OP_STLS:
5341 			id = DIF_INSTR_VAR(instr);
5342 
5343 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
5344 			id -= DIF_VAR_OTHER_UBASE;
5345 			ASSERT(id < vstate->dtvs_nlocals);
5346 
5347 			ASSERT(vstate->dtvs_locals != NULL);
5348 			svar = vstate->dtvs_locals[id];
5349 			ASSERT(svar != NULL);
5350 			v = &svar->dtsv_var;
5351 
5352 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5353 				uintptr_t a = (uintptr_t)svar->dtsv_data;
5354 				size_t sz = v->dtdv_type.dtdt_size;
5355 
5356 				sz += sizeof (uint64_t);
5357 				ASSERT(svar->dtsv_size == NCPU * sz);
5358 				a += curcpu * sz;
5359 
5360 				if (regs[rd] == 0) {
5361 					*(uint8_t *)a = UINT8_MAX;
5362 					break;
5363 				} else {
5364 					*(uint8_t *)a = 0;
5365 					a += sizeof (uint64_t);
5366 				}
5367 
5368 				if (!dtrace_vcanload(
5369 				    (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5370 				    mstate, vstate))
5371 					break;
5372 
5373 				dtrace_vcopy((void *)(uintptr_t)regs[rd],
5374 				    (void *)a, &v->dtdv_type);
5375 				break;
5376 			}
5377 
5378 			ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
5379 			tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
5380 			tmp[curcpu] = regs[rd];
5381 			break;
5382 
5383 		case DIF_OP_LDTS: {
5384 			dtrace_dynvar_t *dvar;
5385 			dtrace_key_t *key;
5386 
5387 			id = DIF_INSTR_VAR(instr);
5388 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
5389 			id -= DIF_VAR_OTHER_UBASE;
5390 			v = &vstate->dtvs_tlocals[id];
5391 
5392 			key = &tupregs[DIF_DTR_NREGS];
5393 			key[0].dttk_value = (uint64_t)id;
5394 			key[0].dttk_size = 0;
5395 			DTRACE_TLS_THRKEY(key[1].dttk_value);
5396 			key[1].dttk_size = 0;
5397 
5398 			dvar = dtrace_dynvar(dstate, 2, key,
5399 			    sizeof (uint64_t), DTRACE_DYNVAR_NOALLOC,
5400 			    mstate, vstate);
5401 
5402 			if (dvar == NULL) {
5403 				regs[rd] = 0;
5404 				break;
5405 			}
5406 
5407 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5408 				regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
5409 			} else {
5410 				regs[rd] = *((uint64_t *)dvar->dtdv_data);
5411 			}
5412 
5413 			break;
5414 		}
5415 
5416 		case DIF_OP_STTS: {
5417 			dtrace_dynvar_t *dvar;
5418 			dtrace_key_t *key;
5419 
5420 			id = DIF_INSTR_VAR(instr);
5421 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
5422 			id -= DIF_VAR_OTHER_UBASE;
5423 
5424 			key = &tupregs[DIF_DTR_NREGS];
5425 			key[0].dttk_value = (uint64_t)id;
5426 			key[0].dttk_size = 0;
5427 			DTRACE_TLS_THRKEY(key[1].dttk_value);
5428 			key[1].dttk_size = 0;
5429 			v = &vstate->dtvs_tlocals[id];
5430 
5431 			dvar = dtrace_dynvar(dstate, 2, key,
5432 			    v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
5433 			    v->dtdv_type.dtdt_size : sizeof (uint64_t),
5434 			    regs[rd] ? DTRACE_DYNVAR_ALLOC :
5435 			    DTRACE_DYNVAR_DEALLOC, mstate, vstate);
5436 
5437 			/*
5438 			 * Given that we're storing to thread-local data,
5439 			 * we need to flush our predicate cache.
5440 			 */
5441 			curthread->t_predcache = 0;
5442 
5443 			if (dvar == NULL)
5444 				break;
5445 
5446 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5447 				if (!dtrace_vcanload(
5448 				    (void *)(uintptr_t)regs[rd],
5449 				    &v->dtdv_type, mstate, vstate))
5450 					break;
5451 
5452 				dtrace_vcopy((void *)(uintptr_t)regs[rd],
5453 				    dvar->dtdv_data, &v->dtdv_type);
5454 			} else {
5455 				*((uint64_t *)dvar->dtdv_data) = regs[rd];
5456 			}
5457 
5458 			break;
5459 		}
5460 
5461 		case DIF_OP_SRA:
5462 			regs[rd] = (int64_t)regs[r1] >> regs[r2];
5463 			break;
5464 
5465 		case DIF_OP_CALL:
5466 			dtrace_dif_subr(DIF_INSTR_SUBR(instr), rd,
5467 			    regs, tupregs, ttop, mstate, state);
5468 			break;
5469 
5470 		case DIF_OP_PUSHTR:
5471 			if (ttop == DIF_DTR_NREGS) {
5472 				*flags |= CPU_DTRACE_TUPOFLOW;
5473 				break;
5474 			}
5475 
5476 			if (r1 == DIF_TYPE_STRING) {
5477 				/*
5478 				 * If this is a string type and the size is 0,
5479 				 * we'll use the system-wide default string
5480 				 * size.  Note that we are _not_ looking at
5481 				 * the value of the DTRACEOPT_STRSIZE option;
5482 				 * had this been set, we would expect to have
5483 				 * a non-zero size value in the "pushtr".
5484 				 */
5485 				tupregs[ttop].dttk_size =
5486 				    dtrace_strlen((char *)(uintptr_t)regs[rd],
5487 				    regs[r2] ? regs[r2] :
5488 				    dtrace_strsize_default) + 1;
5489 			} else {
5490 				tupregs[ttop].dttk_size = regs[r2];
5491 			}
5492 
5493 			tupregs[ttop++].dttk_value = regs[rd];
5494 			break;
5495 
5496 		case DIF_OP_PUSHTV:
5497 			if (ttop == DIF_DTR_NREGS) {
5498 				*flags |= CPU_DTRACE_TUPOFLOW;
5499 				break;
5500 			}
5501 
5502 			tupregs[ttop].dttk_value = regs[rd];
5503 			tupregs[ttop++].dttk_size = 0;
5504 			break;
5505 
5506 		case DIF_OP_POPTS:
5507 			if (ttop != 0)
5508 				ttop--;
5509 			break;
5510 
5511 		case DIF_OP_FLUSHTS:
5512 			ttop = 0;
5513 			break;
5514 
5515 		case DIF_OP_LDGAA:
5516 		case DIF_OP_LDTAA: {
5517 			dtrace_dynvar_t *dvar;
5518 			dtrace_key_t *key = tupregs;
5519 			uint_t nkeys = ttop;
5520 
5521 			id = DIF_INSTR_VAR(instr);
5522 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
5523 			id -= DIF_VAR_OTHER_UBASE;
5524 
5525 			key[nkeys].dttk_value = (uint64_t)id;
5526 			key[nkeys++].dttk_size = 0;
5527 
5528 			if (DIF_INSTR_OP(instr) == DIF_OP_LDTAA) {
5529 				DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
5530 				key[nkeys++].dttk_size = 0;
5531 				v = &vstate->dtvs_tlocals[id];
5532 			} else {
5533 				v = &vstate->dtvs_globals[id]->dtsv_var;
5534 			}
5535 
5536 			dvar = dtrace_dynvar(dstate, nkeys, key,
5537 			    v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
5538 			    v->dtdv_type.dtdt_size : sizeof (uint64_t),
5539 			    DTRACE_DYNVAR_NOALLOC, mstate, vstate);
5540 
5541 			if (dvar == NULL) {
5542 				regs[rd] = 0;
5543 				break;
5544 			}
5545 
5546 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5547 				regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
5548 			} else {
5549 				regs[rd] = *((uint64_t *)dvar->dtdv_data);
5550 			}
5551 
5552 			break;
5553 		}
5554 
5555 		case DIF_OP_STGAA:
5556 		case DIF_OP_STTAA: {
5557 			dtrace_dynvar_t *dvar;
5558 			dtrace_key_t *key = tupregs;
5559 			uint_t nkeys = ttop;
5560 
5561 			id = DIF_INSTR_VAR(instr);
5562 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
5563 			id -= DIF_VAR_OTHER_UBASE;
5564 
5565 			key[nkeys].dttk_value = (uint64_t)id;
5566 			key[nkeys++].dttk_size = 0;
5567 
5568 			if (DIF_INSTR_OP(instr) == DIF_OP_STTAA) {
5569 				DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
5570 				key[nkeys++].dttk_size = 0;
5571 				v = &vstate->dtvs_tlocals[id];
5572 			} else {
5573 				v = &vstate->dtvs_globals[id]->dtsv_var;
5574 			}
5575 
5576 			dvar = dtrace_dynvar(dstate, nkeys, key,
5577 			    v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
5578 			    v->dtdv_type.dtdt_size : sizeof (uint64_t),
5579 			    regs[rd] ? DTRACE_DYNVAR_ALLOC :
5580 			    DTRACE_DYNVAR_DEALLOC, mstate, vstate);
5581 
5582 			if (dvar == NULL)
5583 				break;
5584 
5585 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5586 				if (!dtrace_vcanload(
5587 				    (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5588 				    mstate, vstate))
5589 					break;
5590 
5591 				dtrace_vcopy((void *)(uintptr_t)regs[rd],
5592 				    dvar->dtdv_data, &v->dtdv_type);
5593 			} else {
5594 				*((uint64_t *)dvar->dtdv_data) = regs[rd];
5595 			}
5596 
5597 			break;
5598 		}
5599 
5600 		case DIF_OP_ALLOCS: {
5601 			uintptr_t ptr = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
5602 			size_t size = ptr - mstate->dtms_scratch_ptr + regs[r1];
5603 
5604 			/*
5605 			 * Rounding up the user allocation size could have
5606 			 * overflowed large, bogus allocations (like -1ULL) to
5607 			 * 0.
5608 			 */
5609 			if (size < regs[r1] ||
5610 			    !DTRACE_INSCRATCH(mstate, size)) {
5611 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5612 				regs[rd] = 0;
5613 				break;
5614 			}
5615 
5616 			dtrace_bzero((void *) mstate->dtms_scratch_ptr, size);
5617 			mstate->dtms_scratch_ptr += size;
5618 			regs[rd] = ptr;
5619 			break;
5620 		}
5621 
5622 		case DIF_OP_COPYS:
5623 			if (!dtrace_canstore(regs[rd], regs[r2],
5624 			    mstate, vstate)) {
5625 				*flags |= CPU_DTRACE_BADADDR;
5626 				*illval = regs[rd];
5627 				break;
5628 			}
5629 
5630 			if (!dtrace_canload(regs[r1], regs[r2], mstate, vstate))
5631 				break;
5632 
5633 			dtrace_bcopy((void *)(uintptr_t)regs[r1],
5634 			    (void *)(uintptr_t)regs[rd], (size_t)regs[r2]);
5635 			break;
5636 
5637 		case DIF_OP_STB:
5638 			if (!dtrace_canstore(regs[rd], 1, mstate, vstate)) {
5639 				*flags |= CPU_DTRACE_BADADDR;
5640 				*illval = regs[rd];
5641 				break;
5642 			}
5643 			*((uint8_t *)(uintptr_t)regs[rd]) = (uint8_t)regs[r1];
5644 			break;
5645 
5646 		case DIF_OP_STH:
5647 			if (!dtrace_canstore(regs[rd], 2, mstate, vstate)) {
5648 				*flags |= CPU_DTRACE_BADADDR;
5649 				*illval = regs[rd];
5650 				break;
5651 			}
5652 			if (regs[rd] & 1) {
5653 				*flags |= CPU_DTRACE_BADALIGN;
5654 				*illval = regs[rd];
5655 				break;
5656 			}
5657 			*((uint16_t *)(uintptr_t)regs[rd]) = (uint16_t)regs[r1];
5658 			break;
5659 
5660 		case DIF_OP_STW:
5661 			if (!dtrace_canstore(regs[rd], 4, mstate, vstate)) {
5662 				*flags |= CPU_DTRACE_BADADDR;
5663 				*illval = regs[rd];
5664 				break;
5665 			}
5666 			if (regs[rd] & 3) {
5667 				*flags |= CPU_DTRACE_BADALIGN;
5668 				*illval = regs[rd];
5669 				break;
5670 			}
5671 			*((uint32_t *)(uintptr_t)regs[rd]) = (uint32_t)regs[r1];
5672 			break;
5673 
5674 		case DIF_OP_STX:
5675 			if (!dtrace_canstore(regs[rd], 8, mstate, vstate)) {
5676 				*flags |= CPU_DTRACE_BADADDR;
5677 				*illval = regs[rd];
5678 				break;
5679 			}
5680 			if (regs[rd] & 7) {
5681 				*flags |= CPU_DTRACE_BADALIGN;
5682 				*illval = regs[rd];
5683 				break;
5684 			}
5685 			*((uint64_t *)(uintptr_t)regs[rd]) = regs[r1];
5686 			break;
5687 		}
5688 	}
5689 
5690 	if (!(*flags & CPU_DTRACE_FAULT))
5691 		return (rval);
5692 
5693 	mstate->dtms_fltoffs = opc * sizeof (dif_instr_t);
5694 	mstate->dtms_present |= DTRACE_MSTATE_FLTOFFS;
5695 
5696 	return (0);
5697 }
5698 
5699 static void
5700 dtrace_action_breakpoint(dtrace_ecb_t *ecb)
5701 {
5702 	dtrace_probe_t *probe = ecb->dte_probe;
5703 	dtrace_provider_t *prov = probe->dtpr_provider;
5704 	char c[DTRACE_FULLNAMELEN + 80], *str;
5705 	char *msg = "dtrace: breakpoint action at probe ";
5706 	char *ecbmsg = " (ecb ";
5707 	uintptr_t mask = (0xf << (sizeof (uintptr_t) * NBBY / 4));
5708 	uintptr_t val = (uintptr_t)ecb;
5709 	int shift = (sizeof (uintptr_t) * NBBY) - 4, i = 0;
5710 
5711 	if (dtrace_destructive_disallow)
5712 		return;
5713 
5714 	/*
5715 	 * It's impossible to be taking action on the NULL probe.
5716 	 */
5717 	ASSERT(probe != NULL);
5718 
5719 	/*
5720 	 * This is a poor man's (destitute man's?) sprintf():  we want to
5721 	 * print the provider name, module name, function name and name of
5722 	 * the probe, along with the hex address of the ECB with the breakpoint
5723 	 * action -- all of which we must place in the character buffer by
5724 	 * hand.
5725 	 */
5726 	while (*msg != '\0')
5727 		c[i++] = *msg++;
5728 
5729 	for (str = prov->dtpv_name; *str != '\0'; str++)
5730 		c[i++] = *str;
5731 	c[i++] = ':';
5732 
5733 	for (str = probe->dtpr_mod; *str != '\0'; str++)
5734 		c[i++] = *str;
5735 	c[i++] = ':';
5736 
5737 	for (str = probe->dtpr_func; *str != '\0'; str++)
5738 		c[i++] = *str;
5739 	c[i++] = ':';
5740 
5741 	for (str = probe->dtpr_name; *str != '\0'; str++)
5742 		c[i++] = *str;
5743 
5744 	while (*ecbmsg != '\0')
5745 		c[i++] = *ecbmsg++;
5746 
5747 	while (shift >= 0) {
5748 		mask = (uintptr_t)0xf << shift;
5749 
5750 		if (val >= ((uintptr_t)1 << shift))
5751 			c[i++] = "0123456789abcdef"[(val & mask) >> shift];
5752 		shift -= 4;
5753 	}
5754 
5755 	c[i++] = ')';
5756 	c[i] = '\0';
5757 
5758 #if defined(sun)
5759 	debug_enter(c);
5760 #else
5761 	kdb_enter(KDB_WHY_DTRACE, "breakpoint action");
5762 #endif
5763 }
5764 
5765 static void
5766 dtrace_action_panic(dtrace_ecb_t *ecb)
5767 {
5768 	dtrace_probe_t *probe = ecb->dte_probe;
5769 
5770 	/*
5771 	 * It's impossible to be taking action on the NULL probe.
5772 	 */
5773 	ASSERT(probe != NULL);
5774 
5775 	if (dtrace_destructive_disallow)
5776 		return;
5777 
5778 	if (dtrace_panicked != NULL)
5779 		return;
5780 
5781 	if (dtrace_casptr(&dtrace_panicked, NULL, curthread) != NULL)
5782 		return;
5783 
5784 	/*
5785 	 * We won the right to panic.  (We want to be sure that only one
5786 	 * thread calls panic() from dtrace_probe(), and that panic() is
5787 	 * called exactly once.)
5788 	 */
5789 	dtrace_panic("dtrace: panic action at probe %s:%s:%s:%s (ecb %p)",
5790 	    probe->dtpr_provider->dtpv_name, probe->dtpr_mod,
5791 	    probe->dtpr_func, probe->dtpr_name, (void *)ecb);
5792 }
5793 
5794 static void
5795 dtrace_action_raise(uint64_t sig)
5796 {
5797 	if (dtrace_destructive_disallow)
5798 		return;
5799 
5800 	if (sig >= NSIG) {
5801 		DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
5802 		return;
5803 	}
5804 
5805 #if defined(sun)
5806 	/*
5807 	 * raise() has a queue depth of 1 -- we ignore all subsequent
5808 	 * invocations of the raise() action.
5809 	 */
5810 	if (curthread->t_dtrace_sig == 0)
5811 		curthread->t_dtrace_sig = (uint8_t)sig;
5812 
5813 	curthread->t_sig_check = 1;
5814 	aston(curthread);
5815 #else
5816 	struct proc *p = curproc;
5817 	PROC_LOCK(p);
5818 	kern_psignal(p, sig);
5819 	PROC_UNLOCK(p);
5820 #endif
5821 }
5822 
5823 static void
5824 dtrace_action_stop(void)
5825 {
5826 	if (dtrace_destructive_disallow)
5827 		return;
5828 
5829 #if defined(sun)
5830 	if (!curthread->t_dtrace_stop) {
5831 		curthread->t_dtrace_stop = 1;
5832 		curthread->t_sig_check = 1;
5833 		aston(curthread);
5834 	}
5835 #else
5836 	struct proc *p = curproc;
5837 	PROC_LOCK(p);
5838 	kern_psignal(p, SIGSTOP);
5839 	PROC_UNLOCK(p);
5840 #endif
5841 }
5842 
5843 static void
5844 dtrace_action_chill(dtrace_mstate_t *mstate, hrtime_t val)
5845 {
5846 	hrtime_t now;
5847 	volatile uint16_t *flags;
5848 #if defined(sun)
5849 	cpu_t *cpu = CPU;
5850 #else
5851 	cpu_t *cpu = &solaris_cpu[curcpu];
5852 #endif
5853 
5854 	if (dtrace_destructive_disallow)
5855 		return;
5856 
5857 	flags = (volatile uint16_t *)&cpu_core[cpu->cpu_id].cpuc_dtrace_flags;
5858 
5859 	now = dtrace_gethrtime();
5860 
5861 	if (now - cpu->cpu_dtrace_chillmark > dtrace_chill_interval) {
5862 		/*
5863 		 * We need to advance the mark to the current time.
5864 		 */
5865 		cpu->cpu_dtrace_chillmark = now;
5866 		cpu->cpu_dtrace_chilled = 0;
5867 	}
5868 
5869 	/*
5870 	 * Now check to see if the requested chill time would take us over
5871 	 * the maximum amount of time allowed in the chill interval.  (Or
5872 	 * worse, if the calculation itself induces overflow.)
5873 	 */
5874 	if (cpu->cpu_dtrace_chilled + val > dtrace_chill_max ||
5875 	    cpu->cpu_dtrace_chilled + val < cpu->cpu_dtrace_chilled) {
5876 		*flags |= CPU_DTRACE_ILLOP;
5877 		return;
5878 	}
5879 
5880 	while (dtrace_gethrtime() - now < val)
5881 		continue;
5882 
5883 	/*
5884 	 * Normally, we assure that the value of the variable "timestamp" does
5885 	 * not change within an ECB.  The presence of chill() represents an
5886 	 * exception to this rule, however.
5887 	 */
5888 	mstate->dtms_present &= ~DTRACE_MSTATE_TIMESTAMP;
5889 	cpu->cpu_dtrace_chilled += val;
5890 }
5891 
5892 static void
5893 dtrace_action_ustack(dtrace_mstate_t *mstate, dtrace_state_t *state,
5894     uint64_t *buf, uint64_t arg)
5895 {
5896 	int nframes = DTRACE_USTACK_NFRAMES(arg);
5897 	int strsize = DTRACE_USTACK_STRSIZE(arg);
5898 	uint64_t *pcs = &buf[1], *fps;
5899 	char *str = (char *)&pcs[nframes];
5900 	int size, offs = 0, i, j;
5901 	uintptr_t old = mstate->dtms_scratch_ptr, saved;
5902 	uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
5903 	char *sym;
5904 
5905 	/*
5906 	 * Should be taking a faster path if string space has not been
5907 	 * allocated.
5908 	 */
5909 	ASSERT(strsize != 0);
5910 
5911 	/*
5912 	 * We will first allocate some temporary space for the frame pointers.
5913 	 */
5914 	fps = (uint64_t *)P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
5915 	size = (uintptr_t)fps - mstate->dtms_scratch_ptr +
5916 	    (nframes * sizeof (uint64_t));
5917 
5918 	if (!DTRACE_INSCRATCH(mstate, size)) {
5919 		/*
5920 		 * Not enough room for our frame pointers -- need to indicate
5921 		 * that we ran out of scratch space.
5922 		 */
5923 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5924 		return;
5925 	}
5926 
5927 	mstate->dtms_scratch_ptr += size;
5928 	saved = mstate->dtms_scratch_ptr;
5929 
5930 	/*
5931 	 * Now get a stack with both program counters and frame pointers.
5932 	 */
5933 	DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5934 	dtrace_getufpstack(buf, fps, nframes + 1);
5935 	DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5936 
5937 	/*
5938 	 * If that faulted, we're cooked.
5939 	 */
5940 	if (*flags & CPU_DTRACE_FAULT)
5941 		goto out;
5942 
5943 	/*
5944 	 * Now we want to walk up the stack, calling the USTACK helper.  For
5945 	 * each iteration, we restore the scratch pointer.
5946 	 */
5947 	for (i = 0; i < nframes; i++) {
5948 		mstate->dtms_scratch_ptr = saved;
5949 
5950 		if (offs >= strsize)
5951 			break;
5952 
5953 		sym = (char *)(uintptr_t)dtrace_helper(
5954 		    DTRACE_HELPER_ACTION_USTACK,
5955 		    mstate, state, pcs[i], fps[i]);
5956 
5957 		/*
5958 		 * If we faulted while running the helper, we're going to
5959 		 * clear the fault and null out the corresponding string.
5960 		 */
5961 		if (*flags & CPU_DTRACE_FAULT) {
5962 			*flags &= ~CPU_DTRACE_FAULT;
5963 			str[offs++] = '\0';
5964 			continue;
5965 		}
5966 
5967 		if (sym == NULL) {
5968 			str[offs++] = '\0';
5969 			continue;
5970 		}
5971 
5972 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5973 
5974 		/*
5975 		 * Now copy in the string that the helper returned to us.
5976 		 */
5977 		for (j = 0; offs + j < strsize; j++) {
5978 			if ((str[offs + j] = sym[j]) == '\0')
5979 				break;
5980 		}
5981 
5982 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5983 
5984 		offs += j + 1;
5985 	}
5986 
5987 	if (offs >= strsize) {
5988 		/*
5989 		 * If we didn't have room for all of the strings, we don't
5990 		 * abort processing -- this needn't be a fatal error -- but we
5991 		 * still want to increment a counter (dts_stkstroverflows) to
5992 		 * allow this condition to be warned about.  (If this is from
5993 		 * a jstack() action, it is easily tuned via jstackstrsize.)
5994 		 */
5995 		dtrace_error(&state->dts_stkstroverflows);
5996 	}
5997 
5998 	while (offs < strsize)
5999 		str[offs++] = '\0';
6000 
6001 out:
6002 	mstate->dtms_scratch_ptr = old;
6003 }
6004 
6005 /*
6006  * If you're looking for the epicenter of DTrace, you just found it.  This
6007  * is the function called by the provider to fire a probe -- from which all
6008  * subsequent probe-context DTrace activity emanates.
6009  */
6010 void
6011 dtrace_probe(dtrace_id_t id, uintptr_t arg0, uintptr_t arg1,
6012     uintptr_t arg2, uintptr_t arg3, uintptr_t arg4)
6013 {
6014 	processorid_t cpuid;
6015 	dtrace_icookie_t cookie;
6016 	dtrace_probe_t *probe;
6017 	dtrace_mstate_t mstate;
6018 	dtrace_ecb_t *ecb;
6019 	dtrace_action_t *act;
6020 	intptr_t offs;
6021 	size_t size;
6022 	int vtime, onintr;
6023 	volatile uint16_t *flags;
6024 	hrtime_t now;
6025 
6026 	if (panicstr != NULL)
6027 		return;
6028 
6029 #if defined(sun)
6030 	/*
6031 	 * Kick out immediately if this CPU is still being born (in which case
6032 	 * curthread will be set to -1) or the current thread can't allow
6033 	 * probes in its current context.
6034 	 */
6035 	if (((uintptr_t)curthread & 1) || (curthread->t_flag & T_DONTDTRACE))
6036 		return;
6037 #endif
6038 
6039 	cookie = dtrace_interrupt_disable();
6040 	probe = dtrace_probes[id - 1];
6041 	cpuid = curcpu;
6042 	onintr = CPU_ON_INTR(CPU);
6043 
6044 	if (!onintr && probe->dtpr_predcache != DTRACE_CACHEIDNONE &&
6045 	    probe->dtpr_predcache == curthread->t_predcache) {
6046 		/*
6047 		 * We have hit in the predicate cache; we know that
6048 		 * this predicate would evaluate to be false.
6049 		 */
6050 		dtrace_interrupt_enable(cookie);
6051 		return;
6052 	}
6053 
6054 #if defined(sun)
6055 	if (panic_quiesce) {
6056 #else
6057 	if (panicstr != NULL) {
6058 #endif
6059 		/*
6060 		 * We don't trace anything if we're panicking.
6061 		 */
6062 		dtrace_interrupt_enable(cookie);
6063 		return;
6064 	}
6065 
6066 	now = dtrace_gethrtime();
6067 	vtime = dtrace_vtime_references != 0;
6068 
6069 	if (vtime && curthread->t_dtrace_start)
6070 		curthread->t_dtrace_vtime += now - curthread->t_dtrace_start;
6071 
6072 	mstate.dtms_difo = NULL;
6073 	mstate.dtms_probe = probe;
6074 	mstate.dtms_strtok = 0;
6075 	mstate.dtms_arg[0] = arg0;
6076 	mstate.dtms_arg[1] = arg1;
6077 	mstate.dtms_arg[2] = arg2;
6078 	mstate.dtms_arg[3] = arg3;
6079 	mstate.dtms_arg[4] = arg4;
6080 
6081 	flags = (volatile uint16_t *)&cpu_core[cpuid].cpuc_dtrace_flags;
6082 
6083 	for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
6084 		dtrace_predicate_t *pred = ecb->dte_predicate;
6085 		dtrace_state_t *state = ecb->dte_state;
6086 		dtrace_buffer_t *buf = &state->dts_buffer[cpuid];
6087 		dtrace_buffer_t *aggbuf = &state->dts_aggbuffer[cpuid];
6088 		dtrace_vstate_t *vstate = &state->dts_vstate;
6089 		dtrace_provider_t *prov = probe->dtpr_provider;
6090 		uint64_t tracememsize = 0;
6091 		int committed = 0;
6092 		caddr_t tomax;
6093 
6094 		/*
6095 		 * A little subtlety with the following (seemingly innocuous)
6096 		 * declaration of the automatic 'val':  by looking at the
6097 		 * code, you might think that it could be declared in the
6098 		 * action processing loop, below.  (That is, it's only used in
6099 		 * the action processing loop.)  However, it must be declared
6100 		 * out of that scope because in the case of DIF expression
6101 		 * arguments to aggregating actions, one iteration of the
6102 		 * action loop will use the last iteration's value.
6103 		 */
6104 		uint64_t val = 0;
6105 
6106 		mstate.dtms_present = DTRACE_MSTATE_ARGS | DTRACE_MSTATE_PROBE;
6107 		*flags &= ~CPU_DTRACE_ERROR;
6108 
6109 		if (prov == dtrace_provider) {
6110 			/*
6111 			 * If dtrace itself is the provider of this probe,
6112 			 * we're only going to continue processing the ECB if
6113 			 * arg0 (the dtrace_state_t) is equal to the ECB's
6114 			 * creating state.  (This prevents disjoint consumers
6115 			 * from seeing one another's metaprobes.)
6116 			 */
6117 			if (arg0 != (uint64_t)(uintptr_t)state)
6118 				continue;
6119 		}
6120 
6121 		if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE) {
6122 			/*
6123 			 * We're not currently active.  If our provider isn't
6124 			 * the dtrace pseudo provider, we're not interested.
6125 			 */
6126 			if (prov != dtrace_provider)
6127 				continue;
6128 
6129 			/*
6130 			 * Now we must further check if we are in the BEGIN
6131 			 * probe.  If we are, we will only continue processing
6132 			 * if we're still in WARMUP -- if one BEGIN enabling
6133 			 * has invoked the exit() action, we don't want to
6134 			 * evaluate subsequent BEGIN enablings.
6135 			 */
6136 			if (probe->dtpr_id == dtrace_probeid_begin &&
6137 			    state->dts_activity != DTRACE_ACTIVITY_WARMUP) {
6138 				ASSERT(state->dts_activity ==
6139 				    DTRACE_ACTIVITY_DRAINING);
6140 				continue;
6141 			}
6142 		}
6143 
6144 		if (ecb->dte_cond) {
6145 			/*
6146 			 * If the dte_cond bits indicate that this
6147 			 * consumer is only allowed to see user-mode firings
6148 			 * of this probe, call the provider's dtps_usermode()
6149 			 * entry point to check that the probe was fired
6150 			 * while in a user context. Skip this ECB if that's
6151 			 * not the case.
6152 			 */
6153 			if ((ecb->dte_cond & DTRACE_COND_USERMODE) &&
6154 			    prov->dtpv_pops.dtps_usermode(prov->dtpv_arg,
6155 			    probe->dtpr_id, probe->dtpr_arg) == 0)
6156 				continue;
6157 
6158 #if defined(sun)
6159 			/*
6160 			 * This is more subtle than it looks. We have to be
6161 			 * absolutely certain that CRED() isn't going to
6162 			 * change out from under us so it's only legit to
6163 			 * examine that structure if we're in constrained
6164 			 * situations. Currently, the only times we'll this
6165 			 * check is if a non-super-user has enabled the
6166 			 * profile or syscall providers -- providers that
6167 			 * allow visibility of all processes. For the
6168 			 * profile case, the check above will ensure that
6169 			 * we're examining a user context.
6170 			 */
6171 			if (ecb->dte_cond & DTRACE_COND_OWNER) {
6172 				cred_t *cr;
6173 				cred_t *s_cr =
6174 				    ecb->dte_state->dts_cred.dcr_cred;
6175 				proc_t *proc;
6176 
6177 				ASSERT(s_cr != NULL);
6178 
6179 				if ((cr = CRED()) == NULL ||
6180 				    s_cr->cr_uid != cr->cr_uid ||
6181 				    s_cr->cr_uid != cr->cr_ruid ||
6182 				    s_cr->cr_uid != cr->cr_suid ||
6183 				    s_cr->cr_gid != cr->cr_gid ||
6184 				    s_cr->cr_gid != cr->cr_rgid ||
6185 				    s_cr->cr_gid != cr->cr_sgid ||
6186 				    (proc = ttoproc(curthread)) == NULL ||
6187 				    (proc->p_flag & SNOCD))
6188 					continue;
6189 			}
6190 
6191 			if (ecb->dte_cond & DTRACE_COND_ZONEOWNER) {
6192 				cred_t *cr;
6193 				cred_t *s_cr =
6194 				    ecb->dte_state->dts_cred.dcr_cred;
6195 
6196 				ASSERT(s_cr != NULL);
6197 
6198 				if ((cr = CRED()) == NULL ||
6199 				    s_cr->cr_zone->zone_id !=
6200 				    cr->cr_zone->zone_id)
6201 					continue;
6202 			}
6203 #endif
6204 		}
6205 
6206 		if (now - state->dts_alive > dtrace_deadman_timeout) {
6207 			/*
6208 			 * We seem to be dead.  Unless we (a) have kernel
6209 			 * destructive permissions (b) have expicitly enabled
6210 			 * destructive actions and (c) destructive actions have
6211 			 * not been disabled, we're going to transition into
6212 			 * the KILLED state, from which no further processing
6213 			 * on this state will be performed.
6214 			 */
6215 			if (!dtrace_priv_kernel_destructive(state) ||
6216 			    !state->dts_cred.dcr_destructive ||
6217 			    dtrace_destructive_disallow) {
6218 				void *activity = &state->dts_activity;
6219 				dtrace_activity_t current;
6220 
6221 				do {
6222 					current = state->dts_activity;
6223 				} while (dtrace_cas32(activity, current,
6224 				    DTRACE_ACTIVITY_KILLED) != current);
6225 
6226 				continue;
6227 			}
6228 		}
6229 
6230 		if ((offs = dtrace_buffer_reserve(buf, ecb->dte_needed,
6231 		    ecb->dte_alignment, state, &mstate)) < 0)
6232 			continue;
6233 
6234 		tomax = buf->dtb_tomax;
6235 		ASSERT(tomax != NULL);
6236 
6237 		if (ecb->dte_size != 0)
6238 			DTRACE_STORE(uint32_t, tomax, offs, ecb->dte_epid);
6239 
6240 		mstate.dtms_epid = ecb->dte_epid;
6241 		mstate.dtms_present |= DTRACE_MSTATE_EPID;
6242 
6243 		if (state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)
6244 			mstate.dtms_access = DTRACE_ACCESS_KERNEL;
6245 		else
6246 			mstate.dtms_access = 0;
6247 
6248 		if (pred != NULL) {
6249 			dtrace_difo_t *dp = pred->dtp_difo;
6250 			int rval;
6251 
6252 			rval = dtrace_dif_emulate(dp, &mstate, vstate, state);
6253 
6254 			if (!(*flags & CPU_DTRACE_ERROR) && !rval) {
6255 				dtrace_cacheid_t cid = probe->dtpr_predcache;
6256 
6257 				if (cid != DTRACE_CACHEIDNONE && !onintr) {
6258 					/*
6259 					 * Update the predicate cache...
6260 					 */
6261 					ASSERT(cid == pred->dtp_cacheid);
6262 					curthread->t_predcache = cid;
6263 				}
6264 
6265 				continue;
6266 			}
6267 		}
6268 
6269 		for (act = ecb->dte_action; !(*flags & CPU_DTRACE_ERROR) &&
6270 		    act != NULL; act = act->dta_next) {
6271 			size_t valoffs;
6272 			dtrace_difo_t *dp;
6273 			dtrace_recdesc_t *rec = &act->dta_rec;
6274 
6275 			size = rec->dtrd_size;
6276 			valoffs = offs + rec->dtrd_offset;
6277 
6278 			if (DTRACEACT_ISAGG(act->dta_kind)) {
6279 				uint64_t v = 0xbad;
6280 				dtrace_aggregation_t *agg;
6281 
6282 				agg = (dtrace_aggregation_t *)act;
6283 
6284 				if ((dp = act->dta_difo) != NULL)
6285 					v = dtrace_dif_emulate(dp,
6286 					    &mstate, vstate, state);
6287 
6288 				if (*flags & CPU_DTRACE_ERROR)
6289 					continue;
6290 
6291 				/*
6292 				 * Note that we always pass the expression
6293 				 * value from the previous iteration of the
6294 				 * action loop.  This value will only be used
6295 				 * if there is an expression argument to the
6296 				 * aggregating action, denoted by the
6297 				 * dtag_hasarg field.
6298 				 */
6299 				dtrace_aggregate(agg, buf,
6300 				    offs, aggbuf, v, val);
6301 				continue;
6302 			}
6303 
6304 			switch (act->dta_kind) {
6305 			case DTRACEACT_STOP:
6306 				if (dtrace_priv_proc_destructive(state))
6307 					dtrace_action_stop();
6308 				continue;
6309 
6310 			case DTRACEACT_BREAKPOINT:
6311 				if (dtrace_priv_kernel_destructive(state))
6312 					dtrace_action_breakpoint(ecb);
6313 				continue;
6314 
6315 			case DTRACEACT_PANIC:
6316 				if (dtrace_priv_kernel_destructive(state))
6317 					dtrace_action_panic(ecb);
6318 				continue;
6319 
6320 			case DTRACEACT_STACK:
6321 				if (!dtrace_priv_kernel(state))
6322 					continue;
6323 
6324 				dtrace_getpcstack((pc_t *)(tomax + valoffs),
6325 				    size / sizeof (pc_t), probe->dtpr_aframes,
6326 				    DTRACE_ANCHORED(probe) ? NULL :
6327 				    (uint32_t *)arg0);
6328 				continue;
6329 
6330 			case DTRACEACT_JSTACK:
6331 			case DTRACEACT_USTACK:
6332 				if (!dtrace_priv_proc(state))
6333 					continue;
6334 
6335 				/*
6336 				 * See comment in DIF_VAR_PID.
6337 				 */
6338 				if (DTRACE_ANCHORED(mstate.dtms_probe) &&
6339 				    CPU_ON_INTR(CPU)) {
6340 					int depth = DTRACE_USTACK_NFRAMES(
6341 					    rec->dtrd_arg) + 1;
6342 
6343 					dtrace_bzero((void *)(tomax + valoffs),
6344 					    DTRACE_USTACK_STRSIZE(rec->dtrd_arg)
6345 					    + depth * sizeof (uint64_t));
6346 
6347 					continue;
6348 				}
6349 
6350 				if (DTRACE_USTACK_STRSIZE(rec->dtrd_arg) != 0 &&
6351 				    curproc->p_dtrace_helpers != NULL) {
6352 					/*
6353 					 * This is the slow path -- we have
6354 					 * allocated string space, and we're
6355 					 * getting the stack of a process that
6356 					 * has helpers.  Call into a separate
6357 					 * routine to perform this processing.
6358 					 */
6359 					dtrace_action_ustack(&mstate, state,
6360 					    (uint64_t *)(tomax + valoffs),
6361 					    rec->dtrd_arg);
6362 					continue;
6363 				}
6364 
6365 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6366 				dtrace_getupcstack((uint64_t *)
6367 				    (tomax + valoffs),
6368 				    DTRACE_USTACK_NFRAMES(rec->dtrd_arg) + 1);
6369 				DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6370 				continue;
6371 
6372 			default:
6373 				break;
6374 			}
6375 
6376 			dp = act->dta_difo;
6377 			ASSERT(dp != NULL);
6378 
6379 			val = dtrace_dif_emulate(dp, &mstate, vstate, state);
6380 
6381 			if (*flags & CPU_DTRACE_ERROR)
6382 				continue;
6383 
6384 			switch (act->dta_kind) {
6385 			case DTRACEACT_SPECULATE:
6386 				ASSERT(buf == &state->dts_buffer[cpuid]);
6387 				buf = dtrace_speculation_buffer(state,
6388 				    cpuid, val);
6389 
6390 				if (buf == NULL) {
6391 					*flags |= CPU_DTRACE_DROP;
6392 					continue;
6393 				}
6394 
6395 				offs = dtrace_buffer_reserve(buf,
6396 				    ecb->dte_needed, ecb->dte_alignment,
6397 				    state, NULL);
6398 
6399 				if (offs < 0) {
6400 					*flags |= CPU_DTRACE_DROP;
6401 					continue;
6402 				}
6403 
6404 				tomax = buf->dtb_tomax;
6405 				ASSERT(tomax != NULL);
6406 
6407 				if (ecb->dte_size != 0)
6408 					DTRACE_STORE(uint32_t, tomax, offs,
6409 					    ecb->dte_epid);
6410 				continue;
6411 
6412 			case DTRACEACT_PRINTM: {
6413 				/* The DIF returns a 'memref'. */
6414 				uintptr_t *memref = (uintptr_t *)(uintptr_t) val;
6415 
6416 				/* Get the size from the memref. */
6417 				size = memref[1];
6418 
6419 				/*
6420 				 * Check if the size exceeds the allocated
6421 				 * buffer size.
6422 				 */
6423 				if (size + sizeof(uintptr_t) > dp->dtdo_rtype.dtdt_size) {
6424 					/* Flag a drop! */
6425 					*flags |= CPU_DTRACE_DROP;
6426 					continue;
6427 				}
6428 
6429 				/* Store the size in the buffer first. */
6430 				DTRACE_STORE(uintptr_t, tomax,
6431 				    valoffs, size);
6432 
6433 				/*
6434 				 * Offset the buffer address to the start
6435 				 * of the data.
6436 				 */
6437 				valoffs += sizeof(uintptr_t);
6438 
6439 				/*
6440 				 * Reset to the memory address rather than
6441 				 * the memref array, then let the BYREF
6442 				 * code below do the work to store the
6443 				 * memory data in the buffer.
6444 				 */
6445 				val = memref[0];
6446 				break;
6447 			}
6448 
6449 			case DTRACEACT_PRINTT: {
6450 				/* The DIF returns a 'typeref'. */
6451 				uintptr_t *typeref = (uintptr_t *)(uintptr_t) val;
6452 				char c = '\0' + 1;
6453 				size_t s;
6454 
6455 				/*
6456 				 * Get the type string length and round it
6457 				 * up so that the data that follows is
6458 				 * aligned for easy access.
6459 				 */
6460 				size_t typs = strlen((char *) typeref[2]) + 1;
6461 				typs = roundup(typs,  sizeof(uintptr_t));
6462 
6463 				/*
6464 				 *Get the size from the typeref using the
6465 				 * number of elements and the type size.
6466 				 */
6467 				size = typeref[1] * typeref[3];
6468 
6469 				/*
6470 				 * Check if the size exceeds the allocated
6471 				 * buffer size.
6472 				 */
6473 				if (size + typs + 2 * sizeof(uintptr_t) > dp->dtdo_rtype.dtdt_size) {
6474 					/* Flag a drop! */
6475 					*flags |= CPU_DTRACE_DROP;
6476 
6477 				}
6478 
6479 				/* Store the size in the buffer first. */
6480 				DTRACE_STORE(uintptr_t, tomax,
6481 				    valoffs, size);
6482 				valoffs += sizeof(uintptr_t);
6483 
6484 				/* Store the type size in the buffer. */
6485 				DTRACE_STORE(uintptr_t, tomax,
6486 				    valoffs, typeref[3]);
6487 				valoffs += sizeof(uintptr_t);
6488 
6489 				val = typeref[2];
6490 
6491 				for (s = 0; s < typs; s++) {
6492 					if (c != '\0')
6493 						c = dtrace_load8(val++);
6494 
6495 					DTRACE_STORE(uint8_t, tomax,
6496 					    valoffs++, c);
6497 				}
6498 
6499 				/*
6500 				 * Reset to the memory address rather than
6501 				 * the typeref array, then let the BYREF
6502 				 * code below do the work to store the
6503 				 * memory data in the buffer.
6504 				 */
6505 				val = typeref[0];
6506 				break;
6507 			}
6508 
6509 			case DTRACEACT_CHILL:
6510 				if (dtrace_priv_kernel_destructive(state))
6511 					dtrace_action_chill(&mstate, val);
6512 				continue;
6513 
6514 			case DTRACEACT_RAISE:
6515 				if (dtrace_priv_proc_destructive(state))
6516 					dtrace_action_raise(val);
6517 				continue;
6518 
6519 			case DTRACEACT_COMMIT:
6520 				ASSERT(!committed);
6521 
6522 				/*
6523 				 * We need to commit our buffer state.
6524 				 */
6525 				if (ecb->dte_size)
6526 					buf->dtb_offset = offs + ecb->dte_size;
6527 				buf = &state->dts_buffer[cpuid];
6528 				dtrace_speculation_commit(state, cpuid, val);
6529 				committed = 1;
6530 				continue;
6531 
6532 			case DTRACEACT_DISCARD:
6533 				dtrace_speculation_discard(state, cpuid, val);
6534 				continue;
6535 
6536 			case DTRACEACT_DIFEXPR:
6537 			case DTRACEACT_LIBACT:
6538 			case DTRACEACT_PRINTF:
6539 			case DTRACEACT_PRINTA:
6540 			case DTRACEACT_SYSTEM:
6541 			case DTRACEACT_FREOPEN:
6542 			case DTRACEACT_TRACEMEM:
6543 				break;
6544 
6545 			case DTRACEACT_TRACEMEM_DYNSIZE:
6546 				tracememsize = val;
6547 				break;
6548 
6549 			case DTRACEACT_SYM:
6550 			case DTRACEACT_MOD:
6551 				if (!dtrace_priv_kernel(state))
6552 					continue;
6553 				break;
6554 
6555 			case DTRACEACT_USYM:
6556 			case DTRACEACT_UMOD:
6557 			case DTRACEACT_UADDR: {
6558 #if defined(sun)
6559 				struct pid *pid = curthread->t_procp->p_pidp;
6560 #endif
6561 
6562 				if (!dtrace_priv_proc(state))
6563 					continue;
6564 
6565 				DTRACE_STORE(uint64_t, tomax,
6566 #if defined(sun)
6567 				    valoffs, (uint64_t)pid->pid_id);
6568 #else
6569 				    valoffs, (uint64_t) curproc->p_pid);
6570 #endif
6571 				DTRACE_STORE(uint64_t, tomax,
6572 				    valoffs + sizeof (uint64_t), val);
6573 
6574 				continue;
6575 			}
6576 
6577 			case DTRACEACT_EXIT: {
6578 				/*
6579 				 * For the exit action, we are going to attempt
6580 				 * to atomically set our activity to be
6581 				 * draining.  If this fails (either because
6582 				 * another CPU has beat us to the exit action,
6583 				 * or because our current activity is something
6584 				 * other than ACTIVE or WARMUP), we will
6585 				 * continue.  This assures that the exit action
6586 				 * can be successfully recorded at most once
6587 				 * when we're in the ACTIVE state.  If we're
6588 				 * encountering the exit() action while in
6589 				 * COOLDOWN, however, we want to honor the new
6590 				 * status code.  (We know that we're the only
6591 				 * thread in COOLDOWN, so there is no race.)
6592 				 */
6593 				void *activity = &state->dts_activity;
6594 				dtrace_activity_t current = state->dts_activity;
6595 
6596 				if (current == DTRACE_ACTIVITY_COOLDOWN)
6597 					break;
6598 
6599 				if (current != DTRACE_ACTIVITY_WARMUP)
6600 					current = DTRACE_ACTIVITY_ACTIVE;
6601 
6602 				if (dtrace_cas32(activity, current,
6603 				    DTRACE_ACTIVITY_DRAINING) != current) {
6604 					*flags |= CPU_DTRACE_DROP;
6605 					continue;
6606 				}
6607 
6608 				break;
6609 			}
6610 
6611 			default:
6612 				ASSERT(0);
6613 			}
6614 
6615 			if (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF) {
6616 				uintptr_t end = valoffs + size;
6617 
6618 				if (tracememsize != 0 &&
6619 				    valoffs + tracememsize < end) {
6620 					end = valoffs + tracememsize;
6621 					tracememsize = 0;
6622 				}
6623 
6624 				if (!dtrace_vcanload((void *)(uintptr_t)val,
6625 				    &dp->dtdo_rtype, &mstate, vstate))
6626 					continue;
6627 
6628 				/*
6629 				 * If this is a string, we're going to only
6630 				 * load until we find the zero byte -- after
6631 				 * which we'll store zero bytes.
6632 				 */
6633 				if (dp->dtdo_rtype.dtdt_kind ==
6634 				    DIF_TYPE_STRING) {
6635 					char c = '\0' + 1;
6636 					int intuple = act->dta_intuple;
6637 					size_t s;
6638 
6639 					for (s = 0; s < size; s++) {
6640 						if (c != '\0')
6641 							c = dtrace_load8(val++);
6642 
6643 						DTRACE_STORE(uint8_t, tomax,
6644 						    valoffs++, c);
6645 
6646 						if (c == '\0' && intuple)
6647 							break;
6648 					}
6649 
6650 					continue;
6651 				}
6652 
6653 				while (valoffs < end) {
6654 					DTRACE_STORE(uint8_t, tomax, valoffs++,
6655 					    dtrace_load8(val++));
6656 				}
6657 
6658 				continue;
6659 			}
6660 
6661 			switch (size) {
6662 			case 0:
6663 				break;
6664 
6665 			case sizeof (uint8_t):
6666 				DTRACE_STORE(uint8_t, tomax, valoffs, val);
6667 				break;
6668 			case sizeof (uint16_t):
6669 				DTRACE_STORE(uint16_t, tomax, valoffs, val);
6670 				break;
6671 			case sizeof (uint32_t):
6672 				DTRACE_STORE(uint32_t, tomax, valoffs, val);
6673 				break;
6674 			case sizeof (uint64_t):
6675 				DTRACE_STORE(uint64_t, tomax, valoffs, val);
6676 				break;
6677 			default:
6678 				/*
6679 				 * Any other size should have been returned by
6680 				 * reference, not by value.
6681 				 */
6682 				ASSERT(0);
6683 				break;
6684 			}
6685 		}
6686 
6687 		if (*flags & CPU_DTRACE_DROP)
6688 			continue;
6689 
6690 		if (*flags & CPU_DTRACE_FAULT) {
6691 			int ndx;
6692 			dtrace_action_t *err;
6693 
6694 			buf->dtb_errors++;
6695 
6696 			if (probe->dtpr_id == dtrace_probeid_error) {
6697 				/*
6698 				 * There's nothing we can do -- we had an
6699 				 * error on the error probe.  We bump an
6700 				 * error counter to at least indicate that
6701 				 * this condition happened.
6702 				 */
6703 				dtrace_error(&state->dts_dblerrors);
6704 				continue;
6705 			}
6706 
6707 			if (vtime) {
6708 				/*
6709 				 * Before recursing on dtrace_probe(), we
6710 				 * need to explicitly clear out our start
6711 				 * time to prevent it from being accumulated
6712 				 * into t_dtrace_vtime.
6713 				 */
6714 				curthread->t_dtrace_start = 0;
6715 			}
6716 
6717 			/*
6718 			 * Iterate over the actions to figure out which action
6719 			 * we were processing when we experienced the error.
6720 			 * Note that act points _past_ the faulting action; if
6721 			 * act is ecb->dte_action, the fault was in the
6722 			 * predicate, if it's ecb->dte_action->dta_next it's
6723 			 * in action #1, and so on.
6724 			 */
6725 			for (err = ecb->dte_action, ndx = 0;
6726 			    err != act; err = err->dta_next, ndx++)
6727 				continue;
6728 
6729 			dtrace_probe_error(state, ecb->dte_epid, ndx,
6730 			    (mstate.dtms_present & DTRACE_MSTATE_FLTOFFS) ?
6731 			    mstate.dtms_fltoffs : -1, DTRACE_FLAGS2FLT(*flags),
6732 			    cpu_core[cpuid].cpuc_dtrace_illval);
6733 
6734 			continue;
6735 		}
6736 
6737 		if (!committed)
6738 			buf->dtb_offset = offs + ecb->dte_size;
6739 	}
6740 
6741 	if (vtime)
6742 		curthread->t_dtrace_start = dtrace_gethrtime();
6743 
6744 	dtrace_interrupt_enable(cookie);
6745 }
6746 
6747 /*
6748  * DTrace Probe Hashing Functions
6749  *
6750  * The functions in this section (and indeed, the functions in remaining
6751  * sections) are not _called_ from probe context.  (Any exceptions to this are
6752  * marked with a "Note:".)  Rather, they are called from elsewhere in the
6753  * DTrace framework to look-up probes in, add probes to and remove probes from
6754  * the DTrace probe hashes.  (Each probe is hashed by each element of the
6755  * probe tuple -- allowing for fast lookups, regardless of what was
6756  * specified.)
6757  */
6758 static uint_t
6759 dtrace_hash_str(const char *p)
6760 {
6761 	unsigned int g;
6762 	uint_t hval = 0;
6763 
6764 	while (*p) {
6765 		hval = (hval << 4) + *p++;
6766 		if ((g = (hval & 0xf0000000)) != 0)
6767 			hval ^= g >> 24;
6768 		hval &= ~g;
6769 	}
6770 	return (hval);
6771 }
6772 
6773 static dtrace_hash_t *
6774 dtrace_hash_create(uintptr_t stroffs, uintptr_t nextoffs, uintptr_t prevoffs)
6775 {
6776 	dtrace_hash_t *hash = kmem_zalloc(sizeof (dtrace_hash_t), KM_SLEEP);
6777 
6778 	hash->dth_stroffs = stroffs;
6779 	hash->dth_nextoffs = nextoffs;
6780 	hash->dth_prevoffs = prevoffs;
6781 
6782 	hash->dth_size = 1;
6783 	hash->dth_mask = hash->dth_size - 1;
6784 
6785 	hash->dth_tab = kmem_zalloc(hash->dth_size *
6786 	    sizeof (dtrace_hashbucket_t *), KM_SLEEP);
6787 
6788 	return (hash);
6789 }
6790 
6791 static void
6792 dtrace_hash_destroy(dtrace_hash_t *hash)
6793 {
6794 #ifdef DEBUG
6795 	int i;
6796 
6797 	for (i = 0; i < hash->dth_size; i++)
6798 		ASSERT(hash->dth_tab[i] == NULL);
6799 #endif
6800 
6801 	kmem_free(hash->dth_tab,
6802 	    hash->dth_size * sizeof (dtrace_hashbucket_t *));
6803 	kmem_free(hash, sizeof (dtrace_hash_t));
6804 }
6805 
6806 static void
6807 dtrace_hash_resize(dtrace_hash_t *hash)
6808 {
6809 	int size = hash->dth_size, i, ndx;
6810 	int new_size = hash->dth_size << 1;
6811 	int new_mask = new_size - 1;
6812 	dtrace_hashbucket_t **new_tab, *bucket, *next;
6813 
6814 	ASSERT((new_size & new_mask) == 0);
6815 
6816 	new_tab = kmem_zalloc(new_size * sizeof (void *), KM_SLEEP);
6817 
6818 	for (i = 0; i < size; i++) {
6819 		for (bucket = hash->dth_tab[i]; bucket != NULL; bucket = next) {
6820 			dtrace_probe_t *probe = bucket->dthb_chain;
6821 
6822 			ASSERT(probe != NULL);
6823 			ndx = DTRACE_HASHSTR(hash, probe) & new_mask;
6824 
6825 			next = bucket->dthb_next;
6826 			bucket->dthb_next = new_tab[ndx];
6827 			new_tab[ndx] = bucket;
6828 		}
6829 	}
6830 
6831 	kmem_free(hash->dth_tab, hash->dth_size * sizeof (void *));
6832 	hash->dth_tab = new_tab;
6833 	hash->dth_size = new_size;
6834 	hash->dth_mask = new_mask;
6835 }
6836 
6837 static void
6838 dtrace_hash_add(dtrace_hash_t *hash, dtrace_probe_t *new)
6839 {
6840 	int hashval = DTRACE_HASHSTR(hash, new);
6841 	int ndx = hashval & hash->dth_mask;
6842 	dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6843 	dtrace_probe_t **nextp, **prevp;
6844 
6845 	for (; bucket != NULL; bucket = bucket->dthb_next) {
6846 		if (DTRACE_HASHEQ(hash, bucket->dthb_chain, new))
6847 			goto add;
6848 	}
6849 
6850 	if ((hash->dth_nbuckets >> 1) > hash->dth_size) {
6851 		dtrace_hash_resize(hash);
6852 		dtrace_hash_add(hash, new);
6853 		return;
6854 	}
6855 
6856 	bucket = kmem_zalloc(sizeof (dtrace_hashbucket_t), KM_SLEEP);
6857 	bucket->dthb_next = hash->dth_tab[ndx];
6858 	hash->dth_tab[ndx] = bucket;
6859 	hash->dth_nbuckets++;
6860 
6861 add:
6862 	nextp = DTRACE_HASHNEXT(hash, new);
6863 	ASSERT(*nextp == NULL && *(DTRACE_HASHPREV(hash, new)) == NULL);
6864 	*nextp = bucket->dthb_chain;
6865 
6866 	if (bucket->dthb_chain != NULL) {
6867 		prevp = DTRACE_HASHPREV(hash, bucket->dthb_chain);
6868 		ASSERT(*prevp == NULL);
6869 		*prevp = new;
6870 	}
6871 
6872 	bucket->dthb_chain = new;
6873 	bucket->dthb_len++;
6874 }
6875 
6876 static dtrace_probe_t *
6877 dtrace_hash_lookup(dtrace_hash_t *hash, dtrace_probe_t *template)
6878 {
6879 	int hashval = DTRACE_HASHSTR(hash, template);
6880 	int ndx = hashval & hash->dth_mask;
6881 	dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6882 
6883 	for (; bucket != NULL; bucket = bucket->dthb_next) {
6884 		if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
6885 			return (bucket->dthb_chain);
6886 	}
6887 
6888 	return (NULL);
6889 }
6890 
6891 static int
6892 dtrace_hash_collisions(dtrace_hash_t *hash, dtrace_probe_t *template)
6893 {
6894 	int hashval = DTRACE_HASHSTR(hash, template);
6895 	int ndx = hashval & hash->dth_mask;
6896 	dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6897 
6898 	for (; bucket != NULL; bucket = bucket->dthb_next) {
6899 		if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
6900 			return (bucket->dthb_len);
6901 	}
6902 
6903 	return (0);
6904 }
6905 
6906 static void
6907 dtrace_hash_remove(dtrace_hash_t *hash, dtrace_probe_t *probe)
6908 {
6909 	int ndx = DTRACE_HASHSTR(hash, probe) & hash->dth_mask;
6910 	dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6911 
6912 	dtrace_probe_t **prevp = DTRACE_HASHPREV(hash, probe);
6913 	dtrace_probe_t **nextp = DTRACE_HASHNEXT(hash, probe);
6914 
6915 	/*
6916 	 * Find the bucket that we're removing this probe from.
6917 	 */
6918 	for (; bucket != NULL; bucket = bucket->dthb_next) {
6919 		if (DTRACE_HASHEQ(hash, bucket->dthb_chain, probe))
6920 			break;
6921 	}
6922 
6923 	ASSERT(bucket != NULL);
6924 
6925 	if (*prevp == NULL) {
6926 		if (*nextp == NULL) {
6927 			/*
6928 			 * The removed probe was the only probe on this
6929 			 * bucket; we need to remove the bucket.
6930 			 */
6931 			dtrace_hashbucket_t *b = hash->dth_tab[ndx];
6932 
6933 			ASSERT(bucket->dthb_chain == probe);
6934 			ASSERT(b != NULL);
6935 
6936 			if (b == bucket) {
6937 				hash->dth_tab[ndx] = bucket->dthb_next;
6938 			} else {
6939 				while (b->dthb_next != bucket)
6940 					b = b->dthb_next;
6941 				b->dthb_next = bucket->dthb_next;
6942 			}
6943 
6944 			ASSERT(hash->dth_nbuckets > 0);
6945 			hash->dth_nbuckets--;
6946 			kmem_free(bucket, sizeof (dtrace_hashbucket_t));
6947 			return;
6948 		}
6949 
6950 		bucket->dthb_chain = *nextp;
6951 	} else {
6952 		*(DTRACE_HASHNEXT(hash, *prevp)) = *nextp;
6953 	}
6954 
6955 	if (*nextp != NULL)
6956 		*(DTRACE_HASHPREV(hash, *nextp)) = *prevp;
6957 }
6958 
6959 /*
6960  * DTrace Utility Functions
6961  *
6962  * These are random utility functions that are _not_ called from probe context.
6963  */
6964 static int
6965 dtrace_badattr(const dtrace_attribute_t *a)
6966 {
6967 	return (a->dtat_name > DTRACE_STABILITY_MAX ||
6968 	    a->dtat_data > DTRACE_STABILITY_MAX ||
6969 	    a->dtat_class > DTRACE_CLASS_MAX);
6970 }
6971 
6972 /*
6973  * Return a duplicate copy of a string.  If the specified string is NULL,
6974  * this function returns a zero-length string.
6975  */
6976 static char *
6977 dtrace_strdup(const char *str)
6978 {
6979 	char *new = kmem_zalloc((str != NULL ? strlen(str) : 0) + 1, KM_SLEEP);
6980 
6981 	if (str != NULL)
6982 		(void) strcpy(new, str);
6983 
6984 	return (new);
6985 }
6986 
6987 #define	DTRACE_ISALPHA(c)	\
6988 	(((c) >= 'a' && (c) <= 'z') || ((c) >= 'A' && (c) <= 'Z'))
6989 
6990 static int
6991 dtrace_badname(const char *s)
6992 {
6993 	char c;
6994 
6995 	if (s == NULL || (c = *s++) == '\0')
6996 		return (0);
6997 
6998 	if (!DTRACE_ISALPHA(c) && c != '-' && c != '_' && c != '.')
6999 		return (1);
7000 
7001 	while ((c = *s++) != '\0') {
7002 		if (!DTRACE_ISALPHA(c) && (c < '0' || c > '9') &&
7003 		    c != '-' && c != '_' && c != '.' && c != '`')
7004 			return (1);
7005 	}
7006 
7007 	return (0);
7008 }
7009 
7010 static void
7011 dtrace_cred2priv(cred_t *cr, uint32_t *privp, uid_t *uidp, zoneid_t *zoneidp)
7012 {
7013 	uint32_t priv;
7014 
7015 #if defined(sun)
7016 	if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
7017 		/*
7018 		 * For DTRACE_PRIV_ALL, the uid and zoneid don't matter.
7019 		 */
7020 		priv = DTRACE_PRIV_ALL;
7021 	} else {
7022 		*uidp = crgetuid(cr);
7023 		*zoneidp = crgetzoneid(cr);
7024 
7025 		priv = 0;
7026 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE))
7027 			priv |= DTRACE_PRIV_KERNEL | DTRACE_PRIV_USER;
7028 		else if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE))
7029 			priv |= DTRACE_PRIV_USER;
7030 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE))
7031 			priv |= DTRACE_PRIV_PROC;
7032 		if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
7033 			priv |= DTRACE_PRIV_OWNER;
7034 		if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
7035 			priv |= DTRACE_PRIV_ZONEOWNER;
7036 	}
7037 #else
7038 	priv = DTRACE_PRIV_ALL;
7039 #endif
7040 
7041 	*privp = priv;
7042 }
7043 
7044 #ifdef DTRACE_ERRDEBUG
7045 static void
7046 dtrace_errdebug(const char *str)
7047 {
7048 	int hval = dtrace_hash_str(str) % DTRACE_ERRHASHSZ;
7049 	int occupied = 0;
7050 
7051 	mutex_enter(&dtrace_errlock);
7052 	dtrace_errlast = str;
7053 	dtrace_errthread = curthread;
7054 
7055 	while (occupied++ < DTRACE_ERRHASHSZ) {
7056 		if (dtrace_errhash[hval].dter_msg == str) {
7057 			dtrace_errhash[hval].dter_count++;
7058 			goto out;
7059 		}
7060 
7061 		if (dtrace_errhash[hval].dter_msg != NULL) {
7062 			hval = (hval + 1) % DTRACE_ERRHASHSZ;
7063 			continue;
7064 		}
7065 
7066 		dtrace_errhash[hval].dter_msg = str;
7067 		dtrace_errhash[hval].dter_count = 1;
7068 		goto out;
7069 	}
7070 
7071 	panic("dtrace: undersized error hash");
7072 out:
7073 	mutex_exit(&dtrace_errlock);
7074 }
7075 #endif
7076 
7077 /*
7078  * DTrace Matching Functions
7079  *
7080  * These functions are used to match groups of probes, given some elements of
7081  * a probe tuple, or some globbed expressions for elements of a probe tuple.
7082  */
7083 static int
7084 dtrace_match_priv(const dtrace_probe_t *prp, uint32_t priv, uid_t uid,
7085     zoneid_t zoneid)
7086 {
7087 	if (priv != DTRACE_PRIV_ALL) {
7088 		uint32_t ppriv = prp->dtpr_provider->dtpv_priv.dtpp_flags;
7089 		uint32_t match = priv & ppriv;
7090 
7091 		/*
7092 		 * No PRIV_DTRACE_* privileges...
7093 		 */
7094 		if ((priv & (DTRACE_PRIV_PROC | DTRACE_PRIV_USER |
7095 		    DTRACE_PRIV_KERNEL)) == 0)
7096 			return (0);
7097 
7098 		/*
7099 		 * No matching bits, but there were bits to match...
7100 		 */
7101 		if (match == 0 && ppriv != 0)
7102 			return (0);
7103 
7104 		/*
7105 		 * Need to have permissions to the process, but don't...
7106 		 */
7107 		if (((ppriv & ~match) & DTRACE_PRIV_OWNER) != 0 &&
7108 		    uid != prp->dtpr_provider->dtpv_priv.dtpp_uid) {
7109 			return (0);
7110 		}
7111 
7112 		/*
7113 		 * Need to be in the same zone unless we possess the
7114 		 * privilege to examine all zones.
7115 		 */
7116 		if (((ppriv & ~match) & DTRACE_PRIV_ZONEOWNER) != 0 &&
7117 		    zoneid != prp->dtpr_provider->dtpv_priv.dtpp_zoneid) {
7118 			return (0);
7119 		}
7120 	}
7121 
7122 	return (1);
7123 }
7124 
7125 /*
7126  * dtrace_match_probe compares a dtrace_probe_t to a pre-compiled key, which
7127  * consists of input pattern strings and an ops-vector to evaluate them.
7128  * This function returns >0 for match, 0 for no match, and <0 for error.
7129  */
7130 static int
7131 dtrace_match_probe(const dtrace_probe_t *prp, const dtrace_probekey_t *pkp,
7132     uint32_t priv, uid_t uid, zoneid_t zoneid)
7133 {
7134 	dtrace_provider_t *pvp = prp->dtpr_provider;
7135 	int rv;
7136 
7137 	if (pvp->dtpv_defunct)
7138 		return (0);
7139 
7140 	if ((rv = pkp->dtpk_pmatch(pvp->dtpv_name, pkp->dtpk_prov, 0)) <= 0)
7141 		return (rv);
7142 
7143 	if ((rv = pkp->dtpk_mmatch(prp->dtpr_mod, pkp->dtpk_mod, 0)) <= 0)
7144 		return (rv);
7145 
7146 	if ((rv = pkp->dtpk_fmatch(prp->dtpr_func, pkp->dtpk_func, 0)) <= 0)
7147 		return (rv);
7148 
7149 	if ((rv = pkp->dtpk_nmatch(prp->dtpr_name, pkp->dtpk_name, 0)) <= 0)
7150 		return (rv);
7151 
7152 	if (dtrace_match_priv(prp, priv, uid, zoneid) == 0)
7153 		return (0);
7154 
7155 	return (rv);
7156 }
7157 
7158 /*
7159  * dtrace_match_glob() is a safe kernel implementation of the gmatch(3GEN)
7160  * interface for matching a glob pattern 'p' to an input string 's'.  Unlike
7161  * libc's version, the kernel version only applies to 8-bit ASCII strings.
7162  * In addition, all of the recursion cases except for '*' matching have been
7163  * unwound.  For '*', we still implement recursive evaluation, but a depth
7164  * counter is maintained and matching is aborted if we recurse too deep.
7165  * The function returns 0 if no match, >0 if match, and <0 if recursion error.
7166  */
7167 static int
7168 dtrace_match_glob(const char *s, const char *p, int depth)
7169 {
7170 	const char *olds;
7171 	char s1, c;
7172 	int gs;
7173 
7174 	if (depth > DTRACE_PROBEKEY_MAXDEPTH)
7175 		return (-1);
7176 
7177 	if (s == NULL)
7178 		s = ""; /* treat NULL as empty string */
7179 
7180 top:
7181 	olds = s;
7182 	s1 = *s++;
7183 
7184 	if (p == NULL)
7185 		return (0);
7186 
7187 	if ((c = *p++) == '\0')
7188 		return (s1 == '\0');
7189 
7190 	switch (c) {
7191 	case '[': {
7192 		int ok = 0, notflag = 0;
7193 		char lc = '\0';
7194 
7195 		if (s1 == '\0')
7196 			return (0);
7197 
7198 		if (*p == '!') {
7199 			notflag = 1;
7200 			p++;
7201 		}
7202 
7203 		if ((c = *p++) == '\0')
7204 			return (0);
7205 
7206 		do {
7207 			if (c == '-' && lc != '\0' && *p != ']') {
7208 				if ((c = *p++) == '\0')
7209 					return (0);
7210 				if (c == '\\' && (c = *p++) == '\0')
7211 					return (0);
7212 
7213 				if (notflag) {
7214 					if (s1 < lc || s1 > c)
7215 						ok++;
7216 					else
7217 						return (0);
7218 				} else if (lc <= s1 && s1 <= c)
7219 					ok++;
7220 
7221 			} else if (c == '\\' && (c = *p++) == '\0')
7222 				return (0);
7223 
7224 			lc = c; /* save left-hand 'c' for next iteration */
7225 
7226 			if (notflag) {
7227 				if (s1 != c)
7228 					ok++;
7229 				else
7230 					return (0);
7231 			} else if (s1 == c)
7232 				ok++;
7233 
7234 			if ((c = *p++) == '\0')
7235 				return (0);
7236 
7237 		} while (c != ']');
7238 
7239 		if (ok)
7240 			goto top;
7241 
7242 		return (0);
7243 	}
7244 
7245 	case '\\':
7246 		if ((c = *p++) == '\0')
7247 			return (0);
7248 		/*FALLTHRU*/
7249 
7250 	default:
7251 		if (c != s1)
7252 			return (0);
7253 		/*FALLTHRU*/
7254 
7255 	case '?':
7256 		if (s1 != '\0')
7257 			goto top;
7258 		return (0);
7259 
7260 	case '*':
7261 		while (*p == '*')
7262 			p++; /* consecutive *'s are identical to a single one */
7263 
7264 		if (*p == '\0')
7265 			return (1);
7266 
7267 		for (s = olds; *s != '\0'; s++) {
7268 			if ((gs = dtrace_match_glob(s, p, depth + 1)) != 0)
7269 				return (gs);
7270 		}
7271 
7272 		return (0);
7273 	}
7274 }
7275 
7276 /*ARGSUSED*/
7277 static int
7278 dtrace_match_string(const char *s, const char *p, int depth)
7279 {
7280 	return (s != NULL && strcmp(s, p) == 0);
7281 }
7282 
7283 /*ARGSUSED*/
7284 static int
7285 dtrace_match_nul(const char *s, const char *p, int depth)
7286 {
7287 	return (1); /* always match the empty pattern */
7288 }
7289 
7290 /*ARGSUSED*/
7291 static int
7292 dtrace_match_nonzero(const char *s, const char *p, int depth)
7293 {
7294 	return (s != NULL && s[0] != '\0');
7295 }
7296 
7297 static int
7298 dtrace_match(const dtrace_probekey_t *pkp, uint32_t priv, uid_t uid,
7299     zoneid_t zoneid, int (*matched)(dtrace_probe_t *, void *), void *arg)
7300 {
7301 	dtrace_probe_t template, *probe;
7302 	dtrace_hash_t *hash = NULL;
7303 	int len, best = INT_MAX, nmatched = 0;
7304 	dtrace_id_t i;
7305 
7306 	ASSERT(MUTEX_HELD(&dtrace_lock));
7307 
7308 	/*
7309 	 * If the probe ID is specified in the key, just lookup by ID and
7310 	 * invoke the match callback once if a matching probe is found.
7311 	 */
7312 	if (pkp->dtpk_id != DTRACE_IDNONE) {
7313 		if ((probe = dtrace_probe_lookup_id(pkp->dtpk_id)) != NULL &&
7314 		    dtrace_match_probe(probe, pkp, priv, uid, zoneid) > 0) {
7315 			(void) (*matched)(probe, arg);
7316 			nmatched++;
7317 		}
7318 		return (nmatched);
7319 	}
7320 
7321 	template.dtpr_mod = (char *)pkp->dtpk_mod;
7322 	template.dtpr_func = (char *)pkp->dtpk_func;
7323 	template.dtpr_name = (char *)pkp->dtpk_name;
7324 
7325 	/*
7326 	 * We want to find the most distinct of the module name, function
7327 	 * name, and name.  So for each one that is not a glob pattern or
7328 	 * empty string, we perform a lookup in the corresponding hash and
7329 	 * use the hash table with the fewest collisions to do our search.
7330 	 */
7331 	if (pkp->dtpk_mmatch == &dtrace_match_string &&
7332 	    (len = dtrace_hash_collisions(dtrace_bymod, &template)) < best) {
7333 		best = len;
7334 		hash = dtrace_bymod;
7335 	}
7336 
7337 	if (pkp->dtpk_fmatch == &dtrace_match_string &&
7338 	    (len = dtrace_hash_collisions(dtrace_byfunc, &template)) < best) {
7339 		best = len;
7340 		hash = dtrace_byfunc;
7341 	}
7342 
7343 	if (pkp->dtpk_nmatch == &dtrace_match_string &&
7344 	    (len = dtrace_hash_collisions(dtrace_byname, &template)) < best) {
7345 		best = len;
7346 		hash = dtrace_byname;
7347 	}
7348 
7349 	/*
7350 	 * If we did not select a hash table, iterate over every probe and
7351 	 * invoke our callback for each one that matches our input probe key.
7352 	 */
7353 	if (hash == NULL) {
7354 		for (i = 0; i < dtrace_nprobes; i++) {
7355 			if ((probe = dtrace_probes[i]) == NULL ||
7356 			    dtrace_match_probe(probe, pkp, priv, uid,
7357 			    zoneid) <= 0)
7358 				continue;
7359 
7360 			nmatched++;
7361 
7362 			if ((*matched)(probe, arg) != DTRACE_MATCH_NEXT)
7363 				break;
7364 		}
7365 
7366 		return (nmatched);
7367 	}
7368 
7369 	/*
7370 	 * If we selected a hash table, iterate over each probe of the same key
7371 	 * name and invoke the callback for every probe that matches the other
7372 	 * attributes of our input probe key.
7373 	 */
7374 	for (probe = dtrace_hash_lookup(hash, &template); probe != NULL;
7375 	    probe = *(DTRACE_HASHNEXT(hash, probe))) {
7376 
7377 		if (dtrace_match_probe(probe, pkp, priv, uid, zoneid) <= 0)
7378 			continue;
7379 
7380 		nmatched++;
7381 
7382 		if ((*matched)(probe, arg) != DTRACE_MATCH_NEXT)
7383 			break;
7384 	}
7385 
7386 	return (nmatched);
7387 }
7388 
7389 /*
7390  * Return the function pointer dtrace_probecmp() should use to compare the
7391  * specified pattern with a string.  For NULL or empty patterns, we select
7392  * dtrace_match_nul().  For glob pattern strings, we use dtrace_match_glob().
7393  * For non-empty non-glob strings, we use dtrace_match_string().
7394  */
7395 static dtrace_probekey_f *
7396 dtrace_probekey_func(const char *p)
7397 {
7398 	char c;
7399 
7400 	if (p == NULL || *p == '\0')
7401 		return (&dtrace_match_nul);
7402 
7403 	while ((c = *p++) != '\0') {
7404 		if (c == '[' || c == '?' || c == '*' || c == '\\')
7405 			return (&dtrace_match_glob);
7406 	}
7407 
7408 	return (&dtrace_match_string);
7409 }
7410 
7411 /*
7412  * Build a probe comparison key for use with dtrace_match_probe() from the
7413  * given probe description.  By convention, a null key only matches anchored
7414  * probes: if each field is the empty string, reset dtpk_fmatch to
7415  * dtrace_match_nonzero().
7416  */
7417 static void
7418 dtrace_probekey(dtrace_probedesc_t *pdp, dtrace_probekey_t *pkp)
7419 {
7420 	pkp->dtpk_prov = pdp->dtpd_provider;
7421 	pkp->dtpk_pmatch = dtrace_probekey_func(pdp->dtpd_provider);
7422 
7423 	pkp->dtpk_mod = pdp->dtpd_mod;
7424 	pkp->dtpk_mmatch = dtrace_probekey_func(pdp->dtpd_mod);
7425 
7426 	pkp->dtpk_func = pdp->dtpd_func;
7427 	pkp->dtpk_fmatch = dtrace_probekey_func(pdp->dtpd_func);
7428 
7429 	pkp->dtpk_name = pdp->dtpd_name;
7430 	pkp->dtpk_nmatch = dtrace_probekey_func(pdp->dtpd_name);
7431 
7432 	pkp->dtpk_id = pdp->dtpd_id;
7433 
7434 	if (pkp->dtpk_id == DTRACE_IDNONE &&
7435 	    pkp->dtpk_pmatch == &dtrace_match_nul &&
7436 	    pkp->dtpk_mmatch == &dtrace_match_nul &&
7437 	    pkp->dtpk_fmatch == &dtrace_match_nul &&
7438 	    pkp->dtpk_nmatch == &dtrace_match_nul)
7439 		pkp->dtpk_fmatch = &dtrace_match_nonzero;
7440 }
7441 
7442 /*
7443  * DTrace Provider-to-Framework API Functions
7444  *
7445  * These functions implement much of the Provider-to-Framework API, as
7446  * described in <sys/dtrace.h>.  The parts of the API not in this section are
7447  * the functions in the API for probe management (found below), and
7448  * dtrace_probe() itself (found above).
7449  */
7450 
7451 /*
7452  * Register the calling provider with the DTrace framework.  This should
7453  * generally be called by DTrace providers in their attach(9E) entry point.
7454  */
7455 int
7456 dtrace_register(const char *name, const dtrace_pattr_t *pap, uint32_t priv,
7457     cred_t *cr, const dtrace_pops_t *pops, void *arg, dtrace_provider_id_t *idp)
7458 {
7459 	dtrace_provider_t *provider;
7460 
7461 	if (name == NULL || pap == NULL || pops == NULL || idp == NULL) {
7462 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7463 		    "arguments", name ? name : "<NULL>");
7464 		return (EINVAL);
7465 	}
7466 
7467 	if (name[0] == '\0' || dtrace_badname(name)) {
7468 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7469 		    "provider name", name);
7470 		return (EINVAL);
7471 	}
7472 
7473 	if ((pops->dtps_provide == NULL && pops->dtps_provide_module == NULL) ||
7474 	    pops->dtps_enable == NULL || pops->dtps_disable == NULL ||
7475 	    pops->dtps_destroy == NULL ||
7476 	    ((pops->dtps_resume == NULL) != (pops->dtps_suspend == NULL))) {
7477 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7478 		    "provider ops", name);
7479 		return (EINVAL);
7480 	}
7481 
7482 	if (dtrace_badattr(&pap->dtpa_provider) ||
7483 	    dtrace_badattr(&pap->dtpa_mod) ||
7484 	    dtrace_badattr(&pap->dtpa_func) ||
7485 	    dtrace_badattr(&pap->dtpa_name) ||
7486 	    dtrace_badattr(&pap->dtpa_args)) {
7487 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7488 		    "provider attributes", name);
7489 		return (EINVAL);
7490 	}
7491 
7492 	if (priv & ~DTRACE_PRIV_ALL) {
7493 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7494 		    "privilege attributes", name);
7495 		return (EINVAL);
7496 	}
7497 
7498 	if ((priv & DTRACE_PRIV_KERNEL) &&
7499 	    (priv & (DTRACE_PRIV_USER | DTRACE_PRIV_OWNER)) &&
7500 	    pops->dtps_usermode == NULL) {
7501 		cmn_err(CE_WARN, "failed to register provider '%s': need "
7502 		    "dtps_usermode() op for given privilege attributes", name);
7503 		return (EINVAL);
7504 	}
7505 
7506 	provider = kmem_zalloc(sizeof (dtrace_provider_t), KM_SLEEP);
7507 	provider->dtpv_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
7508 	(void) strcpy(provider->dtpv_name, name);
7509 
7510 	provider->dtpv_attr = *pap;
7511 	provider->dtpv_priv.dtpp_flags = priv;
7512 	if (cr != NULL) {
7513 		provider->dtpv_priv.dtpp_uid = crgetuid(cr);
7514 		provider->dtpv_priv.dtpp_zoneid = crgetzoneid(cr);
7515 	}
7516 	provider->dtpv_pops = *pops;
7517 
7518 	if (pops->dtps_provide == NULL) {
7519 		ASSERT(pops->dtps_provide_module != NULL);
7520 		provider->dtpv_pops.dtps_provide =
7521 		    (void (*)(void *, dtrace_probedesc_t *))dtrace_nullop;
7522 	}
7523 
7524 	if (pops->dtps_provide_module == NULL) {
7525 		ASSERT(pops->dtps_provide != NULL);
7526 		provider->dtpv_pops.dtps_provide_module =
7527 		    (void (*)(void *, modctl_t *))dtrace_nullop;
7528 	}
7529 
7530 	if (pops->dtps_suspend == NULL) {
7531 		ASSERT(pops->dtps_resume == NULL);
7532 		provider->dtpv_pops.dtps_suspend =
7533 		    (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
7534 		provider->dtpv_pops.dtps_resume =
7535 		    (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
7536 	}
7537 
7538 	provider->dtpv_arg = arg;
7539 	*idp = (dtrace_provider_id_t)provider;
7540 
7541 	if (pops == &dtrace_provider_ops) {
7542 		ASSERT(MUTEX_HELD(&dtrace_provider_lock));
7543 		ASSERT(MUTEX_HELD(&dtrace_lock));
7544 		ASSERT(dtrace_anon.dta_enabling == NULL);
7545 
7546 		/*
7547 		 * We make sure that the DTrace provider is at the head of
7548 		 * the provider chain.
7549 		 */
7550 		provider->dtpv_next = dtrace_provider;
7551 		dtrace_provider = provider;
7552 		return (0);
7553 	}
7554 
7555 	mutex_enter(&dtrace_provider_lock);
7556 	mutex_enter(&dtrace_lock);
7557 
7558 	/*
7559 	 * If there is at least one provider registered, we'll add this
7560 	 * provider after the first provider.
7561 	 */
7562 	if (dtrace_provider != NULL) {
7563 		provider->dtpv_next = dtrace_provider->dtpv_next;
7564 		dtrace_provider->dtpv_next = provider;
7565 	} else {
7566 		dtrace_provider = provider;
7567 	}
7568 
7569 	if (dtrace_retained != NULL) {
7570 		dtrace_enabling_provide(provider);
7571 
7572 		/*
7573 		 * Now we need to call dtrace_enabling_matchall() -- which
7574 		 * will acquire cpu_lock and dtrace_lock.  We therefore need
7575 		 * to drop all of our locks before calling into it...
7576 		 */
7577 		mutex_exit(&dtrace_lock);
7578 		mutex_exit(&dtrace_provider_lock);
7579 		dtrace_enabling_matchall();
7580 
7581 		return (0);
7582 	}
7583 
7584 	mutex_exit(&dtrace_lock);
7585 	mutex_exit(&dtrace_provider_lock);
7586 
7587 	return (0);
7588 }
7589 
7590 /*
7591  * Unregister the specified provider from the DTrace framework.  This should
7592  * generally be called by DTrace providers in their detach(9E) entry point.
7593  */
7594 int
7595 dtrace_unregister(dtrace_provider_id_t id)
7596 {
7597 	dtrace_provider_t *old = (dtrace_provider_t *)id;
7598 	dtrace_provider_t *prev = NULL;
7599 	int i, self = 0, noreap = 0;
7600 	dtrace_probe_t *probe, *first = NULL;
7601 
7602 	if (old->dtpv_pops.dtps_enable ==
7603 	    (void (*)(void *, dtrace_id_t, void *))dtrace_nullop) {
7604 		/*
7605 		 * If DTrace itself is the provider, we're called with locks
7606 		 * already held.
7607 		 */
7608 		ASSERT(old == dtrace_provider);
7609 #if defined(sun)
7610 		ASSERT(dtrace_devi != NULL);
7611 #endif
7612 		ASSERT(MUTEX_HELD(&dtrace_provider_lock));
7613 		ASSERT(MUTEX_HELD(&dtrace_lock));
7614 		self = 1;
7615 
7616 		if (dtrace_provider->dtpv_next != NULL) {
7617 			/*
7618 			 * There's another provider here; return failure.
7619 			 */
7620 			return (EBUSY);
7621 		}
7622 	} else {
7623 		mutex_enter(&dtrace_provider_lock);
7624 		mutex_enter(&mod_lock);
7625 		mutex_enter(&dtrace_lock);
7626 	}
7627 
7628 	/*
7629 	 * If anyone has /dev/dtrace open, or if there are anonymous enabled
7630 	 * probes, we refuse to let providers slither away, unless this
7631 	 * provider has already been explicitly invalidated.
7632 	 */
7633 	if (!old->dtpv_defunct &&
7634 	    (dtrace_opens || (dtrace_anon.dta_state != NULL &&
7635 	    dtrace_anon.dta_state->dts_necbs > 0))) {
7636 		if (!self) {
7637 			mutex_exit(&dtrace_lock);
7638 			mutex_exit(&mod_lock);
7639 			mutex_exit(&dtrace_provider_lock);
7640 		}
7641 		return (EBUSY);
7642 	}
7643 
7644 	/*
7645 	 * Attempt to destroy the probes associated with this provider.
7646 	 */
7647 	for (i = 0; i < dtrace_nprobes; i++) {
7648 		if ((probe = dtrace_probes[i]) == NULL)
7649 			continue;
7650 
7651 		if (probe->dtpr_provider != old)
7652 			continue;
7653 
7654 		if (probe->dtpr_ecb == NULL)
7655 			continue;
7656 
7657 		/*
7658 		 * If we are trying to unregister a defunct provider, and the
7659 		 * provider was made defunct within the interval dictated by
7660 		 * dtrace_unregister_defunct_reap, we'll (asynchronously)
7661 		 * attempt to reap our enablings.  To denote that the provider
7662 		 * should reattempt to unregister itself at some point in the
7663 		 * future, we will return a differentiable error code (EAGAIN
7664 		 * instead of EBUSY) in this case.
7665 		 */
7666 		if (dtrace_gethrtime() - old->dtpv_defunct >
7667 		    dtrace_unregister_defunct_reap)
7668 			noreap = 1;
7669 
7670 		if (!self) {
7671 			mutex_exit(&dtrace_lock);
7672 			mutex_exit(&mod_lock);
7673 			mutex_exit(&dtrace_provider_lock);
7674 		}
7675 
7676 		if (noreap)
7677 			return (EBUSY);
7678 
7679 		(void) taskq_dispatch(dtrace_taskq,
7680 		    (task_func_t *)dtrace_enabling_reap, NULL, TQ_SLEEP);
7681 
7682 		return (EAGAIN);
7683 	}
7684 
7685 	/*
7686 	 * All of the probes for this provider are disabled; we can safely
7687 	 * remove all of them from their hash chains and from the probe array.
7688 	 */
7689 	for (i = 0; i < dtrace_nprobes; i++) {
7690 		if ((probe = dtrace_probes[i]) == NULL)
7691 			continue;
7692 
7693 		if (probe->dtpr_provider != old)
7694 			continue;
7695 
7696 		dtrace_probes[i] = NULL;
7697 
7698 		dtrace_hash_remove(dtrace_bymod, probe);
7699 		dtrace_hash_remove(dtrace_byfunc, probe);
7700 		dtrace_hash_remove(dtrace_byname, probe);
7701 
7702 		if (first == NULL) {
7703 			first = probe;
7704 			probe->dtpr_nextmod = NULL;
7705 		} else {
7706 			probe->dtpr_nextmod = first;
7707 			first = probe;
7708 		}
7709 	}
7710 
7711 	/*
7712 	 * The provider's probes have been removed from the hash chains and
7713 	 * from the probe array.  Now issue a dtrace_sync() to be sure that
7714 	 * everyone has cleared out from any probe array processing.
7715 	 */
7716 	dtrace_sync();
7717 
7718 	for (probe = first; probe != NULL; probe = first) {
7719 		first = probe->dtpr_nextmod;
7720 
7721 		old->dtpv_pops.dtps_destroy(old->dtpv_arg, probe->dtpr_id,
7722 		    probe->dtpr_arg);
7723 		kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
7724 		kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
7725 		kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
7726 #if defined(sun)
7727 		vmem_free(dtrace_arena, (void *)(uintptr_t)(probe->dtpr_id), 1);
7728 #else
7729 		free_unr(dtrace_arena, probe->dtpr_id);
7730 #endif
7731 		kmem_free(probe, sizeof (dtrace_probe_t));
7732 	}
7733 
7734 	if ((prev = dtrace_provider) == old) {
7735 #if defined(sun)
7736 		ASSERT(self || dtrace_devi == NULL);
7737 		ASSERT(old->dtpv_next == NULL || dtrace_devi == NULL);
7738 #endif
7739 		dtrace_provider = old->dtpv_next;
7740 	} else {
7741 		while (prev != NULL && prev->dtpv_next != old)
7742 			prev = prev->dtpv_next;
7743 
7744 		if (prev == NULL) {
7745 			panic("attempt to unregister non-existent "
7746 			    "dtrace provider %p\n", (void *)id);
7747 		}
7748 
7749 		prev->dtpv_next = old->dtpv_next;
7750 	}
7751 
7752 	if (!self) {
7753 		mutex_exit(&dtrace_lock);
7754 		mutex_exit(&mod_lock);
7755 		mutex_exit(&dtrace_provider_lock);
7756 	}
7757 
7758 	kmem_free(old->dtpv_name, strlen(old->dtpv_name) + 1);
7759 	kmem_free(old, sizeof (dtrace_provider_t));
7760 
7761 	return (0);
7762 }
7763 
7764 /*
7765  * Invalidate the specified provider.  All subsequent probe lookups for the
7766  * specified provider will fail, but its probes will not be removed.
7767  */
7768 void
7769 dtrace_invalidate(dtrace_provider_id_t id)
7770 {
7771 	dtrace_provider_t *pvp = (dtrace_provider_t *)id;
7772 
7773 	ASSERT(pvp->dtpv_pops.dtps_enable !=
7774 	    (void (*)(void *, dtrace_id_t, void *))dtrace_nullop);
7775 
7776 	mutex_enter(&dtrace_provider_lock);
7777 	mutex_enter(&dtrace_lock);
7778 
7779 	pvp->dtpv_defunct = dtrace_gethrtime();
7780 
7781 	mutex_exit(&dtrace_lock);
7782 	mutex_exit(&dtrace_provider_lock);
7783 }
7784 
7785 /*
7786  * Indicate whether or not DTrace has attached.
7787  */
7788 int
7789 dtrace_attached(void)
7790 {
7791 	/*
7792 	 * dtrace_provider will be non-NULL iff the DTrace driver has
7793 	 * attached.  (It's non-NULL because DTrace is always itself a
7794 	 * provider.)
7795 	 */
7796 	return (dtrace_provider != NULL);
7797 }
7798 
7799 /*
7800  * Remove all the unenabled probes for the given provider.  This function is
7801  * not unlike dtrace_unregister(), except that it doesn't remove the provider
7802  * -- just as many of its associated probes as it can.
7803  */
7804 int
7805 dtrace_condense(dtrace_provider_id_t id)
7806 {
7807 	dtrace_provider_t *prov = (dtrace_provider_t *)id;
7808 	int i;
7809 	dtrace_probe_t *probe;
7810 
7811 	/*
7812 	 * Make sure this isn't the dtrace provider itself.
7813 	 */
7814 	ASSERT(prov->dtpv_pops.dtps_enable !=
7815 	    (void (*)(void *, dtrace_id_t, void *))dtrace_nullop);
7816 
7817 	mutex_enter(&dtrace_provider_lock);
7818 	mutex_enter(&dtrace_lock);
7819 
7820 	/*
7821 	 * Attempt to destroy the probes associated with this provider.
7822 	 */
7823 	for (i = 0; i < dtrace_nprobes; i++) {
7824 		if ((probe = dtrace_probes[i]) == NULL)
7825 			continue;
7826 
7827 		if (probe->dtpr_provider != prov)
7828 			continue;
7829 
7830 		if (probe->dtpr_ecb != NULL)
7831 			continue;
7832 
7833 		dtrace_probes[i] = NULL;
7834 
7835 		dtrace_hash_remove(dtrace_bymod, probe);
7836 		dtrace_hash_remove(dtrace_byfunc, probe);
7837 		dtrace_hash_remove(dtrace_byname, probe);
7838 
7839 		prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, i + 1,
7840 		    probe->dtpr_arg);
7841 		kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
7842 		kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
7843 		kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
7844 		kmem_free(probe, sizeof (dtrace_probe_t));
7845 #if defined(sun)
7846 		vmem_free(dtrace_arena, (void *)((uintptr_t)i + 1), 1);
7847 #else
7848 		free_unr(dtrace_arena, i + 1);
7849 #endif
7850 	}
7851 
7852 	mutex_exit(&dtrace_lock);
7853 	mutex_exit(&dtrace_provider_lock);
7854 
7855 	return (0);
7856 }
7857 
7858 /*
7859  * DTrace Probe Management Functions
7860  *
7861  * The functions in this section perform the DTrace probe management,
7862  * including functions to create probes, look-up probes, and call into the
7863  * providers to request that probes be provided.  Some of these functions are
7864  * in the Provider-to-Framework API; these functions can be identified by the
7865  * fact that they are not declared "static".
7866  */
7867 
7868 /*
7869  * Create a probe with the specified module name, function name, and name.
7870  */
7871 dtrace_id_t
7872 dtrace_probe_create(dtrace_provider_id_t prov, const char *mod,
7873     const char *func, const char *name, int aframes, void *arg)
7874 {
7875 	dtrace_probe_t *probe, **probes;
7876 	dtrace_provider_t *provider = (dtrace_provider_t *)prov;
7877 	dtrace_id_t id;
7878 
7879 	if (provider == dtrace_provider) {
7880 		ASSERT(MUTEX_HELD(&dtrace_lock));
7881 	} else {
7882 		mutex_enter(&dtrace_lock);
7883 	}
7884 
7885 #if defined(sun)
7886 	id = (dtrace_id_t)(uintptr_t)vmem_alloc(dtrace_arena, 1,
7887 	    VM_BESTFIT | VM_SLEEP);
7888 #else
7889 	id = alloc_unr(dtrace_arena);
7890 #endif
7891 	probe = kmem_zalloc(sizeof (dtrace_probe_t), KM_SLEEP);
7892 
7893 	probe->dtpr_id = id;
7894 	probe->dtpr_gen = dtrace_probegen++;
7895 	probe->dtpr_mod = dtrace_strdup(mod);
7896 	probe->dtpr_func = dtrace_strdup(func);
7897 	probe->dtpr_name = dtrace_strdup(name);
7898 	probe->dtpr_arg = arg;
7899 	probe->dtpr_aframes = aframes;
7900 	probe->dtpr_provider = provider;
7901 
7902 	dtrace_hash_add(dtrace_bymod, probe);
7903 	dtrace_hash_add(dtrace_byfunc, probe);
7904 	dtrace_hash_add(dtrace_byname, probe);
7905 
7906 	if (id - 1 >= dtrace_nprobes) {
7907 		size_t osize = dtrace_nprobes * sizeof (dtrace_probe_t *);
7908 		size_t nsize = osize << 1;
7909 
7910 		if (nsize == 0) {
7911 			ASSERT(osize == 0);
7912 			ASSERT(dtrace_probes == NULL);
7913 			nsize = sizeof (dtrace_probe_t *);
7914 		}
7915 
7916 		probes = kmem_zalloc(nsize, KM_SLEEP);
7917 
7918 		if (dtrace_probes == NULL) {
7919 			ASSERT(osize == 0);
7920 			dtrace_probes = probes;
7921 			dtrace_nprobes = 1;
7922 		} else {
7923 			dtrace_probe_t **oprobes = dtrace_probes;
7924 
7925 			bcopy(oprobes, probes, osize);
7926 			dtrace_membar_producer();
7927 			dtrace_probes = probes;
7928 
7929 			dtrace_sync();
7930 
7931 			/*
7932 			 * All CPUs are now seeing the new probes array; we can
7933 			 * safely free the old array.
7934 			 */
7935 			kmem_free(oprobes, osize);
7936 			dtrace_nprobes <<= 1;
7937 		}
7938 
7939 		ASSERT(id - 1 < dtrace_nprobes);
7940 	}
7941 
7942 	ASSERT(dtrace_probes[id - 1] == NULL);
7943 	dtrace_probes[id - 1] = probe;
7944 
7945 	if (provider != dtrace_provider)
7946 		mutex_exit(&dtrace_lock);
7947 
7948 	return (id);
7949 }
7950 
7951 static dtrace_probe_t *
7952 dtrace_probe_lookup_id(dtrace_id_t id)
7953 {
7954 	ASSERT(MUTEX_HELD(&dtrace_lock));
7955 
7956 	if (id == 0 || id > dtrace_nprobes)
7957 		return (NULL);
7958 
7959 	return (dtrace_probes[id - 1]);
7960 }
7961 
7962 static int
7963 dtrace_probe_lookup_match(dtrace_probe_t *probe, void *arg)
7964 {
7965 	*((dtrace_id_t *)arg) = probe->dtpr_id;
7966 
7967 	return (DTRACE_MATCH_DONE);
7968 }
7969 
7970 /*
7971  * Look up a probe based on provider and one or more of module name, function
7972  * name and probe name.
7973  */
7974 dtrace_id_t
7975 dtrace_probe_lookup(dtrace_provider_id_t prid, char *mod,
7976     char *func, char *name)
7977 {
7978 	dtrace_probekey_t pkey;
7979 	dtrace_id_t id;
7980 	int match;
7981 
7982 	pkey.dtpk_prov = ((dtrace_provider_t *)prid)->dtpv_name;
7983 	pkey.dtpk_pmatch = &dtrace_match_string;
7984 	pkey.dtpk_mod = mod;
7985 	pkey.dtpk_mmatch = mod ? &dtrace_match_string : &dtrace_match_nul;
7986 	pkey.dtpk_func = func;
7987 	pkey.dtpk_fmatch = func ? &dtrace_match_string : &dtrace_match_nul;
7988 	pkey.dtpk_name = name;
7989 	pkey.dtpk_nmatch = name ? &dtrace_match_string : &dtrace_match_nul;
7990 	pkey.dtpk_id = DTRACE_IDNONE;
7991 
7992 	mutex_enter(&dtrace_lock);
7993 	match = dtrace_match(&pkey, DTRACE_PRIV_ALL, 0, 0,
7994 	    dtrace_probe_lookup_match, &id);
7995 	mutex_exit(&dtrace_lock);
7996 
7997 	ASSERT(match == 1 || match == 0);
7998 	return (match ? id : 0);
7999 }
8000 
8001 /*
8002  * Returns the probe argument associated with the specified probe.
8003  */
8004 void *
8005 dtrace_probe_arg(dtrace_provider_id_t id, dtrace_id_t pid)
8006 {
8007 	dtrace_probe_t *probe;
8008 	void *rval = NULL;
8009 
8010 	mutex_enter(&dtrace_lock);
8011 
8012 	if ((probe = dtrace_probe_lookup_id(pid)) != NULL &&
8013 	    probe->dtpr_provider == (dtrace_provider_t *)id)
8014 		rval = probe->dtpr_arg;
8015 
8016 	mutex_exit(&dtrace_lock);
8017 
8018 	return (rval);
8019 }
8020 
8021 /*
8022  * Copy a probe into a probe description.
8023  */
8024 static void
8025 dtrace_probe_description(const dtrace_probe_t *prp, dtrace_probedesc_t *pdp)
8026 {
8027 	bzero(pdp, sizeof (dtrace_probedesc_t));
8028 	pdp->dtpd_id = prp->dtpr_id;
8029 
8030 	(void) strncpy(pdp->dtpd_provider,
8031 	    prp->dtpr_provider->dtpv_name, DTRACE_PROVNAMELEN - 1);
8032 
8033 	(void) strncpy(pdp->dtpd_mod, prp->dtpr_mod, DTRACE_MODNAMELEN - 1);
8034 	(void) strncpy(pdp->dtpd_func, prp->dtpr_func, DTRACE_FUNCNAMELEN - 1);
8035 	(void) strncpy(pdp->dtpd_name, prp->dtpr_name, DTRACE_NAMELEN - 1);
8036 }
8037 
8038 #if !defined(sun)
8039 static int
8040 dtrace_probe_provide_cb(linker_file_t lf, void *arg)
8041 {
8042 	dtrace_provider_t *prv = (dtrace_provider_t *) arg;
8043 
8044 	prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, lf);
8045 
8046 	return(0);
8047 }
8048 #endif
8049 
8050 
8051 /*
8052  * Called to indicate that a probe -- or probes -- should be provided by a
8053  * specfied provider.  If the specified description is NULL, the provider will
8054  * be told to provide all of its probes.  (This is done whenever a new
8055  * consumer comes along, or whenever a retained enabling is to be matched.) If
8056  * the specified description is non-NULL, the provider is given the
8057  * opportunity to dynamically provide the specified probe, allowing providers
8058  * to support the creation of probes on-the-fly.  (So-called _autocreated_
8059  * probes.)  If the provider is NULL, the operations will be applied to all
8060  * providers; if the provider is non-NULL the operations will only be applied
8061  * to the specified provider.  The dtrace_provider_lock must be held, and the
8062  * dtrace_lock must _not_ be held -- the provider's dtps_provide() operation
8063  * will need to grab the dtrace_lock when it reenters the framework through
8064  * dtrace_probe_lookup(), dtrace_probe_create(), etc.
8065  */
8066 static void
8067 dtrace_probe_provide(dtrace_probedesc_t *desc, dtrace_provider_t *prv)
8068 {
8069 #if defined(sun)
8070 	modctl_t *ctl;
8071 #endif
8072 	int all = 0;
8073 
8074 	ASSERT(MUTEX_HELD(&dtrace_provider_lock));
8075 
8076 	if (prv == NULL) {
8077 		all = 1;
8078 		prv = dtrace_provider;
8079 	}
8080 
8081 	do {
8082 		/*
8083 		 * First, call the blanket provide operation.
8084 		 */
8085 		prv->dtpv_pops.dtps_provide(prv->dtpv_arg, desc);
8086 
8087 		/*
8088 		 * Now call the per-module provide operation.  We will grab
8089 		 * mod_lock to prevent the list from being modified.  Note
8090 		 * that this also prevents the mod_busy bits from changing.
8091 		 * (mod_busy can only be changed with mod_lock held.)
8092 		 */
8093 		mutex_enter(&mod_lock);
8094 
8095 #if defined(sun)
8096 		ctl = &modules;
8097 		do {
8098 			if (ctl->mod_busy || ctl->mod_mp == NULL)
8099 				continue;
8100 
8101 			prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
8102 
8103 		} while ((ctl = ctl->mod_next) != &modules);
8104 #else
8105 		(void) linker_file_foreach(dtrace_probe_provide_cb, prv);
8106 #endif
8107 
8108 		mutex_exit(&mod_lock);
8109 	} while (all && (prv = prv->dtpv_next) != NULL);
8110 }
8111 
8112 #if defined(sun)
8113 /*
8114  * Iterate over each probe, and call the Framework-to-Provider API function
8115  * denoted by offs.
8116  */
8117 static void
8118 dtrace_probe_foreach(uintptr_t offs)
8119 {
8120 	dtrace_provider_t *prov;
8121 	void (*func)(void *, dtrace_id_t, void *);
8122 	dtrace_probe_t *probe;
8123 	dtrace_icookie_t cookie;
8124 	int i;
8125 
8126 	/*
8127 	 * We disable interrupts to walk through the probe array.  This is
8128 	 * safe -- the dtrace_sync() in dtrace_unregister() assures that we
8129 	 * won't see stale data.
8130 	 */
8131 	cookie = dtrace_interrupt_disable();
8132 
8133 	for (i = 0; i < dtrace_nprobes; i++) {
8134 		if ((probe = dtrace_probes[i]) == NULL)
8135 			continue;
8136 
8137 		if (probe->dtpr_ecb == NULL) {
8138 			/*
8139 			 * This probe isn't enabled -- don't call the function.
8140 			 */
8141 			continue;
8142 		}
8143 
8144 		prov = probe->dtpr_provider;
8145 		func = *((void(**)(void *, dtrace_id_t, void *))
8146 		    ((uintptr_t)&prov->dtpv_pops + offs));
8147 
8148 		func(prov->dtpv_arg, i + 1, probe->dtpr_arg);
8149 	}
8150 
8151 	dtrace_interrupt_enable(cookie);
8152 }
8153 #endif
8154 
8155 static int
8156 dtrace_probe_enable(dtrace_probedesc_t *desc, dtrace_enabling_t *enab)
8157 {
8158 	dtrace_probekey_t pkey;
8159 	uint32_t priv;
8160 	uid_t uid;
8161 	zoneid_t zoneid;
8162 
8163 	ASSERT(MUTEX_HELD(&dtrace_lock));
8164 	dtrace_ecb_create_cache = NULL;
8165 
8166 	if (desc == NULL) {
8167 		/*
8168 		 * If we're passed a NULL description, we're being asked to
8169 		 * create an ECB with a NULL probe.
8170 		 */
8171 		(void) dtrace_ecb_create_enable(NULL, enab);
8172 		return (0);
8173 	}
8174 
8175 	dtrace_probekey(desc, &pkey);
8176 	dtrace_cred2priv(enab->dten_vstate->dtvs_state->dts_cred.dcr_cred,
8177 	    &priv, &uid, &zoneid);
8178 
8179 	return (dtrace_match(&pkey, priv, uid, zoneid, dtrace_ecb_create_enable,
8180 	    enab));
8181 }
8182 
8183 /*
8184  * DTrace Helper Provider Functions
8185  */
8186 static void
8187 dtrace_dofattr2attr(dtrace_attribute_t *attr, const dof_attr_t dofattr)
8188 {
8189 	attr->dtat_name = DOF_ATTR_NAME(dofattr);
8190 	attr->dtat_data = DOF_ATTR_DATA(dofattr);
8191 	attr->dtat_class = DOF_ATTR_CLASS(dofattr);
8192 }
8193 
8194 static void
8195 dtrace_dofprov2hprov(dtrace_helper_provdesc_t *hprov,
8196     const dof_provider_t *dofprov, char *strtab)
8197 {
8198 	hprov->dthpv_provname = strtab + dofprov->dofpv_name;
8199 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_provider,
8200 	    dofprov->dofpv_provattr);
8201 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_mod,
8202 	    dofprov->dofpv_modattr);
8203 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_func,
8204 	    dofprov->dofpv_funcattr);
8205 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_name,
8206 	    dofprov->dofpv_nameattr);
8207 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_args,
8208 	    dofprov->dofpv_argsattr);
8209 }
8210 
8211 static void
8212 dtrace_helper_provide_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
8213 {
8214 	uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8215 	dof_hdr_t *dof = (dof_hdr_t *)daddr;
8216 	dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
8217 	dof_provider_t *provider;
8218 	dof_probe_t *probe;
8219 	uint32_t *off, *enoff;
8220 	uint8_t *arg;
8221 	char *strtab;
8222 	uint_t i, nprobes;
8223 	dtrace_helper_provdesc_t dhpv;
8224 	dtrace_helper_probedesc_t dhpb;
8225 	dtrace_meta_t *meta = dtrace_meta_pid;
8226 	dtrace_mops_t *mops = &meta->dtm_mops;
8227 	void *parg;
8228 
8229 	provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
8230 	str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8231 	    provider->dofpv_strtab * dof->dofh_secsize);
8232 	prb_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8233 	    provider->dofpv_probes * dof->dofh_secsize);
8234 	arg_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8235 	    provider->dofpv_prargs * dof->dofh_secsize);
8236 	off_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8237 	    provider->dofpv_proffs * dof->dofh_secsize);
8238 
8239 	strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
8240 	off = (uint32_t *)(uintptr_t)(daddr + off_sec->dofs_offset);
8241 	arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
8242 	enoff = NULL;
8243 
8244 	/*
8245 	 * See dtrace_helper_provider_validate().
8246 	 */
8247 	if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
8248 	    provider->dofpv_prenoffs != DOF_SECT_NONE) {
8249 		enoff_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8250 		    provider->dofpv_prenoffs * dof->dofh_secsize);
8251 		enoff = (uint32_t *)(uintptr_t)(daddr + enoff_sec->dofs_offset);
8252 	}
8253 
8254 	nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
8255 
8256 	/*
8257 	 * Create the provider.
8258 	 */
8259 	dtrace_dofprov2hprov(&dhpv, provider, strtab);
8260 
8261 	if ((parg = mops->dtms_provide_pid(meta->dtm_arg, &dhpv, pid)) == NULL)
8262 		return;
8263 
8264 	meta->dtm_count++;
8265 
8266 	/*
8267 	 * Create the probes.
8268 	 */
8269 	for (i = 0; i < nprobes; i++) {
8270 		probe = (dof_probe_t *)(uintptr_t)(daddr +
8271 		    prb_sec->dofs_offset + i * prb_sec->dofs_entsize);
8272 
8273 		dhpb.dthpb_mod = dhp->dofhp_mod;
8274 		dhpb.dthpb_func = strtab + probe->dofpr_func;
8275 		dhpb.dthpb_name = strtab + probe->dofpr_name;
8276 		dhpb.dthpb_base = probe->dofpr_addr;
8277 		dhpb.dthpb_offs = off + probe->dofpr_offidx;
8278 		dhpb.dthpb_noffs = probe->dofpr_noffs;
8279 		if (enoff != NULL) {
8280 			dhpb.dthpb_enoffs = enoff + probe->dofpr_enoffidx;
8281 			dhpb.dthpb_nenoffs = probe->dofpr_nenoffs;
8282 		} else {
8283 			dhpb.dthpb_enoffs = NULL;
8284 			dhpb.dthpb_nenoffs = 0;
8285 		}
8286 		dhpb.dthpb_args = arg + probe->dofpr_argidx;
8287 		dhpb.dthpb_nargc = probe->dofpr_nargc;
8288 		dhpb.dthpb_xargc = probe->dofpr_xargc;
8289 		dhpb.dthpb_ntypes = strtab + probe->dofpr_nargv;
8290 		dhpb.dthpb_xtypes = strtab + probe->dofpr_xargv;
8291 
8292 		mops->dtms_create_probe(meta->dtm_arg, parg, &dhpb);
8293 	}
8294 }
8295 
8296 static void
8297 dtrace_helper_provide(dof_helper_t *dhp, pid_t pid)
8298 {
8299 	uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8300 	dof_hdr_t *dof = (dof_hdr_t *)daddr;
8301 	int i;
8302 
8303 	ASSERT(MUTEX_HELD(&dtrace_meta_lock));
8304 
8305 	for (i = 0; i < dof->dofh_secnum; i++) {
8306 		dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
8307 		    dof->dofh_secoff + i * dof->dofh_secsize);
8308 
8309 		if (sec->dofs_type != DOF_SECT_PROVIDER)
8310 			continue;
8311 
8312 		dtrace_helper_provide_one(dhp, sec, pid);
8313 	}
8314 
8315 	/*
8316 	 * We may have just created probes, so we must now rematch against
8317 	 * any retained enablings.  Note that this call will acquire both
8318 	 * cpu_lock and dtrace_lock; the fact that we are holding
8319 	 * dtrace_meta_lock now is what defines the ordering with respect to
8320 	 * these three locks.
8321 	 */
8322 	dtrace_enabling_matchall();
8323 }
8324 
8325 static void
8326 dtrace_helper_provider_remove_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
8327 {
8328 	uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8329 	dof_hdr_t *dof = (dof_hdr_t *)daddr;
8330 	dof_sec_t *str_sec;
8331 	dof_provider_t *provider;
8332 	char *strtab;
8333 	dtrace_helper_provdesc_t dhpv;
8334 	dtrace_meta_t *meta = dtrace_meta_pid;
8335 	dtrace_mops_t *mops = &meta->dtm_mops;
8336 
8337 	provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
8338 	str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8339 	    provider->dofpv_strtab * dof->dofh_secsize);
8340 
8341 	strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
8342 
8343 	/*
8344 	 * Create the provider.
8345 	 */
8346 	dtrace_dofprov2hprov(&dhpv, provider, strtab);
8347 
8348 	mops->dtms_remove_pid(meta->dtm_arg, &dhpv, pid);
8349 
8350 	meta->dtm_count--;
8351 }
8352 
8353 static void
8354 dtrace_helper_provider_remove(dof_helper_t *dhp, pid_t pid)
8355 {
8356 	uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8357 	dof_hdr_t *dof = (dof_hdr_t *)daddr;
8358 	int i;
8359 
8360 	ASSERT(MUTEX_HELD(&dtrace_meta_lock));
8361 
8362 	for (i = 0; i < dof->dofh_secnum; i++) {
8363 		dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
8364 		    dof->dofh_secoff + i * dof->dofh_secsize);
8365 
8366 		if (sec->dofs_type != DOF_SECT_PROVIDER)
8367 			continue;
8368 
8369 		dtrace_helper_provider_remove_one(dhp, sec, pid);
8370 	}
8371 }
8372 
8373 /*
8374  * DTrace Meta Provider-to-Framework API Functions
8375  *
8376  * These functions implement the Meta Provider-to-Framework API, as described
8377  * in <sys/dtrace.h>.
8378  */
8379 int
8380 dtrace_meta_register(const char *name, const dtrace_mops_t *mops, void *arg,
8381     dtrace_meta_provider_id_t *idp)
8382 {
8383 	dtrace_meta_t *meta;
8384 	dtrace_helpers_t *help, *next;
8385 	int i;
8386 
8387 	*idp = DTRACE_METAPROVNONE;
8388 
8389 	/*
8390 	 * We strictly don't need the name, but we hold onto it for
8391 	 * debuggability. All hail error queues!
8392 	 */
8393 	if (name == NULL) {
8394 		cmn_err(CE_WARN, "failed to register meta-provider: "
8395 		    "invalid name");
8396 		return (EINVAL);
8397 	}
8398 
8399 	if (mops == NULL ||
8400 	    mops->dtms_create_probe == NULL ||
8401 	    mops->dtms_provide_pid == NULL ||
8402 	    mops->dtms_remove_pid == NULL) {
8403 		cmn_err(CE_WARN, "failed to register meta-register %s: "
8404 		    "invalid ops", name);
8405 		return (EINVAL);
8406 	}
8407 
8408 	meta = kmem_zalloc(sizeof (dtrace_meta_t), KM_SLEEP);
8409 	meta->dtm_mops = *mops;
8410 	meta->dtm_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
8411 	(void) strcpy(meta->dtm_name, name);
8412 	meta->dtm_arg = arg;
8413 
8414 	mutex_enter(&dtrace_meta_lock);
8415 	mutex_enter(&dtrace_lock);
8416 
8417 	if (dtrace_meta_pid != NULL) {
8418 		mutex_exit(&dtrace_lock);
8419 		mutex_exit(&dtrace_meta_lock);
8420 		cmn_err(CE_WARN, "failed to register meta-register %s: "
8421 		    "user-land meta-provider exists", name);
8422 		kmem_free(meta->dtm_name, strlen(meta->dtm_name) + 1);
8423 		kmem_free(meta, sizeof (dtrace_meta_t));
8424 		return (EINVAL);
8425 	}
8426 
8427 	dtrace_meta_pid = meta;
8428 	*idp = (dtrace_meta_provider_id_t)meta;
8429 
8430 	/*
8431 	 * If there are providers and probes ready to go, pass them
8432 	 * off to the new meta provider now.
8433 	 */
8434 
8435 	help = dtrace_deferred_pid;
8436 	dtrace_deferred_pid = NULL;
8437 
8438 	mutex_exit(&dtrace_lock);
8439 
8440 	while (help != NULL) {
8441 		for (i = 0; i < help->dthps_nprovs; i++) {
8442 			dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
8443 			    help->dthps_pid);
8444 		}
8445 
8446 		next = help->dthps_next;
8447 		help->dthps_next = NULL;
8448 		help->dthps_prev = NULL;
8449 		help->dthps_deferred = 0;
8450 		help = next;
8451 	}
8452 
8453 	mutex_exit(&dtrace_meta_lock);
8454 
8455 	return (0);
8456 }
8457 
8458 int
8459 dtrace_meta_unregister(dtrace_meta_provider_id_t id)
8460 {
8461 	dtrace_meta_t **pp, *old = (dtrace_meta_t *)id;
8462 
8463 	mutex_enter(&dtrace_meta_lock);
8464 	mutex_enter(&dtrace_lock);
8465 
8466 	if (old == dtrace_meta_pid) {
8467 		pp = &dtrace_meta_pid;
8468 	} else {
8469 		panic("attempt to unregister non-existent "
8470 		    "dtrace meta-provider %p\n", (void *)old);
8471 	}
8472 
8473 	if (old->dtm_count != 0) {
8474 		mutex_exit(&dtrace_lock);
8475 		mutex_exit(&dtrace_meta_lock);
8476 		return (EBUSY);
8477 	}
8478 
8479 	*pp = NULL;
8480 
8481 	mutex_exit(&dtrace_lock);
8482 	mutex_exit(&dtrace_meta_lock);
8483 
8484 	kmem_free(old->dtm_name, strlen(old->dtm_name) + 1);
8485 	kmem_free(old, sizeof (dtrace_meta_t));
8486 
8487 	return (0);
8488 }
8489 
8490 
8491 /*
8492  * DTrace DIF Object Functions
8493  */
8494 static int
8495 dtrace_difo_err(uint_t pc, const char *format, ...)
8496 {
8497 	if (dtrace_err_verbose) {
8498 		va_list alist;
8499 
8500 		(void) uprintf("dtrace DIF object error: [%u]: ", pc);
8501 		va_start(alist, format);
8502 		(void) vuprintf(format, alist);
8503 		va_end(alist);
8504 	}
8505 
8506 #ifdef DTRACE_ERRDEBUG
8507 	dtrace_errdebug(format);
8508 #endif
8509 	return (1);
8510 }
8511 
8512 /*
8513  * Validate a DTrace DIF object by checking the IR instructions.  The following
8514  * rules are currently enforced by dtrace_difo_validate():
8515  *
8516  * 1. Each instruction must have a valid opcode
8517  * 2. Each register, string, variable, or subroutine reference must be valid
8518  * 3. No instruction can modify register %r0 (must be zero)
8519  * 4. All instruction reserved bits must be set to zero
8520  * 5. The last instruction must be a "ret" instruction
8521  * 6. All branch targets must reference a valid instruction _after_ the branch
8522  */
8523 static int
8524 dtrace_difo_validate(dtrace_difo_t *dp, dtrace_vstate_t *vstate, uint_t nregs,
8525     cred_t *cr)
8526 {
8527 	int err = 0, i;
8528 	int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
8529 	int kcheckload;
8530 	uint_t pc;
8531 
8532 	kcheckload = cr == NULL ||
8533 	    (vstate->dtvs_state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) == 0;
8534 
8535 	dp->dtdo_destructive = 0;
8536 
8537 	for (pc = 0; pc < dp->dtdo_len && err == 0; pc++) {
8538 		dif_instr_t instr = dp->dtdo_buf[pc];
8539 
8540 		uint_t r1 = DIF_INSTR_R1(instr);
8541 		uint_t r2 = DIF_INSTR_R2(instr);
8542 		uint_t rd = DIF_INSTR_RD(instr);
8543 		uint_t rs = DIF_INSTR_RS(instr);
8544 		uint_t label = DIF_INSTR_LABEL(instr);
8545 		uint_t v = DIF_INSTR_VAR(instr);
8546 		uint_t subr = DIF_INSTR_SUBR(instr);
8547 		uint_t type = DIF_INSTR_TYPE(instr);
8548 		uint_t op = DIF_INSTR_OP(instr);
8549 
8550 		switch (op) {
8551 		case DIF_OP_OR:
8552 		case DIF_OP_XOR:
8553 		case DIF_OP_AND:
8554 		case DIF_OP_SLL:
8555 		case DIF_OP_SRL:
8556 		case DIF_OP_SRA:
8557 		case DIF_OP_SUB:
8558 		case DIF_OP_ADD:
8559 		case DIF_OP_MUL:
8560 		case DIF_OP_SDIV:
8561 		case DIF_OP_UDIV:
8562 		case DIF_OP_SREM:
8563 		case DIF_OP_UREM:
8564 		case DIF_OP_COPYS:
8565 			if (r1 >= nregs)
8566 				err += efunc(pc, "invalid register %u\n", r1);
8567 			if (r2 >= nregs)
8568 				err += efunc(pc, "invalid register %u\n", r2);
8569 			if (rd >= nregs)
8570 				err += efunc(pc, "invalid register %u\n", rd);
8571 			if (rd == 0)
8572 				err += efunc(pc, "cannot write to %r0\n");
8573 			break;
8574 		case DIF_OP_NOT:
8575 		case DIF_OP_MOV:
8576 		case DIF_OP_ALLOCS:
8577 			if (r1 >= nregs)
8578 				err += efunc(pc, "invalid register %u\n", r1);
8579 			if (r2 != 0)
8580 				err += efunc(pc, "non-zero reserved bits\n");
8581 			if (rd >= nregs)
8582 				err += efunc(pc, "invalid register %u\n", rd);
8583 			if (rd == 0)
8584 				err += efunc(pc, "cannot write to %r0\n");
8585 			break;
8586 		case DIF_OP_LDSB:
8587 		case DIF_OP_LDSH:
8588 		case DIF_OP_LDSW:
8589 		case DIF_OP_LDUB:
8590 		case DIF_OP_LDUH:
8591 		case DIF_OP_LDUW:
8592 		case DIF_OP_LDX:
8593 			if (r1 >= nregs)
8594 				err += efunc(pc, "invalid register %u\n", r1);
8595 			if (r2 != 0)
8596 				err += efunc(pc, "non-zero reserved bits\n");
8597 			if (rd >= nregs)
8598 				err += efunc(pc, "invalid register %u\n", rd);
8599 			if (rd == 0)
8600 				err += efunc(pc, "cannot write to %r0\n");
8601 			if (kcheckload)
8602 				dp->dtdo_buf[pc] = DIF_INSTR_LOAD(op +
8603 				    DIF_OP_RLDSB - DIF_OP_LDSB, r1, rd);
8604 			break;
8605 		case DIF_OP_RLDSB:
8606 		case DIF_OP_RLDSH:
8607 		case DIF_OP_RLDSW:
8608 		case DIF_OP_RLDUB:
8609 		case DIF_OP_RLDUH:
8610 		case DIF_OP_RLDUW:
8611 		case DIF_OP_RLDX:
8612 			if (r1 >= nregs)
8613 				err += efunc(pc, "invalid register %u\n", r1);
8614 			if (r2 != 0)
8615 				err += efunc(pc, "non-zero reserved bits\n");
8616 			if (rd >= nregs)
8617 				err += efunc(pc, "invalid register %u\n", rd);
8618 			if (rd == 0)
8619 				err += efunc(pc, "cannot write to %r0\n");
8620 			break;
8621 		case DIF_OP_ULDSB:
8622 		case DIF_OP_ULDSH:
8623 		case DIF_OP_ULDSW:
8624 		case DIF_OP_ULDUB:
8625 		case DIF_OP_ULDUH:
8626 		case DIF_OP_ULDUW:
8627 		case DIF_OP_ULDX:
8628 			if (r1 >= nregs)
8629 				err += efunc(pc, "invalid register %u\n", r1);
8630 			if (r2 != 0)
8631 				err += efunc(pc, "non-zero reserved bits\n");
8632 			if (rd >= nregs)
8633 				err += efunc(pc, "invalid register %u\n", rd);
8634 			if (rd == 0)
8635 				err += efunc(pc, "cannot write to %r0\n");
8636 			break;
8637 		case DIF_OP_STB:
8638 		case DIF_OP_STH:
8639 		case DIF_OP_STW:
8640 		case DIF_OP_STX:
8641 			if (r1 >= nregs)
8642 				err += efunc(pc, "invalid register %u\n", r1);
8643 			if (r2 != 0)
8644 				err += efunc(pc, "non-zero reserved bits\n");
8645 			if (rd >= nregs)
8646 				err += efunc(pc, "invalid register %u\n", rd);
8647 			if (rd == 0)
8648 				err += efunc(pc, "cannot write to 0 address\n");
8649 			break;
8650 		case DIF_OP_CMP:
8651 		case DIF_OP_SCMP:
8652 			if (r1 >= nregs)
8653 				err += efunc(pc, "invalid register %u\n", r1);
8654 			if (r2 >= nregs)
8655 				err += efunc(pc, "invalid register %u\n", r2);
8656 			if (rd != 0)
8657 				err += efunc(pc, "non-zero reserved bits\n");
8658 			break;
8659 		case DIF_OP_TST:
8660 			if (r1 >= nregs)
8661 				err += efunc(pc, "invalid register %u\n", r1);
8662 			if (r2 != 0 || rd != 0)
8663 				err += efunc(pc, "non-zero reserved bits\n");
8664 			break;
8665 		case DIF_OP_BA:
8666 		case DIF_OP_BE:
8667 		case DIF_OP_BNE:
8668 		case DIF_OP_BG:
8669 		case DIF_OP_BGU:
8670 		case DIF_OP_BGE:
8671 		case DIF_OP_BGEU:
8672 		case DIF_OP_BL:
8673 		case DIF_OP_BLU:
8674 		case DIF_OP_BLE:
8675 		case DIF_OP_BLEU:
8676 			if (label >= dp->dtdo_len) {
8677 				err += efunc(pc, "invalid branch target %u\n",
8678 				    label);
8679 			}
8680 			if (label <= pc) {
8681 				err += efunc(pc, "backward branch to %u\n",
8682 				    label);
8683 			}
8684 			break;
8685 		case DIF_OP_RET:
8686 			if (r1 != 0 || r2 != 0)
8687 				err += efunc(pc, "non-zero reserved bits\n");
8688 			if (rd >= nregs)
8689 				err += efunc(pc, "invalid register %u\n", rd);
8690 			break;
8691 		case DIF_OP_NOP:
8692 		case DIF_OP_POPTS:
8693 		case DIF_OP_FLUSHTS:
8694 			if (r1 != 0 || r2 != 0 || rd != 0)
8695 				err += efunc(pc, "non-zero reserved bits\n");
8696 			break;
8697 		case DIF_OP_SETX:
8698 			if (DIF_INSTR_INTEGER(instr) >= dp->dtdo_intlen) {
8699 				err += efunc(pc, "invalid integer ref %u\n",
8700 				    DIF_INSTR_INTEGER(instr));
8701 			}
8702 			if (rd >= nregs)
8703 				err += efunc(pc, "invalid register %u\n", rd);
8704 			if (rd == 0)
8705 				err += efunc(pc, "cannot write to %r0\n");
8706 			break;
8707 		case DIF_OP_SETS:
8708 			if (DIF_INSTR_STRING(instr) >= dp->dtdo_strlen) {
8709 				err += efunc(pc, "invalid string ref %u\n",
8710 				    DIF_INSTR_STRING(instr));
8711 			}
8712 			if (rd >= nregs)
8713 				err += efunc(pc, "invalid register %u\n", rd);
8714 			if (rd == 0)
8715 				err += efunc(pc, "cannot write to %r0\n");
8716 			break;
8717 		case DIF_OP_LDGA:
8718 		case DIF_OP_LDTA:
8719 			if (r1 > DIF_VAR_ARRAY_MAX)
8720 				err += efunc(pc, "invalid array %u\n", r1);
8721 			if (r2 >= nregs)
8722 				err += efunc(pc, "invalid register %u\n", r2);
8723 			if (rd >= nregs)
8724 				err += efunc(pc, "invalid register %u\n", rd);
8725 			if (rd == 0)
8726 				err += efunc(pc, "cannot write to %r0\n");
8727 			break;
8728 		case DIF_OP_LDGS:
8729 		case DIF_OP_LDTS:
8730 		case DIF_OP_LDLS:
8731 		case DIF_OP_LDGAA:
8732 		case DIF_OP_LDTAA:
8733 			if (v < DIF_VAR_OTHER_MIN || v > DIF_VAR_OTHER_MAX)
8734 				err += efunc(pc, "invalid variable %u\n", v);
8735 			if (rd >= nregs)
8736 				err += efunc(pc, "invalid register %u\n", rd);
8737 			if (rd == 0)
8738 				err += efunc(pc, "cannot write to %r0\n");
8739 			break;
8740 		case DIF_OP_STGS:
8741 		case DIF_OP_STTS:
8742 		case DIF_OP_STLS:
8743 		case DIF_OP_STGAA:
8744 		case DIF_OP_STTAA:
8745 			if (v < DIF_VAR_OTHER_UBASE || v > DIF_VAR_OTHER_MAX)
8746 				err += efunc(pc, "invalid variable %u\n", v);
8747 			if (rs >= nregs)
8748 				err += efunc(pc, "invalid register %u\n", rd);
8749 			break;
8750 		case DIF_OP_CALL:
8751 			if (subr > DIF_SUBR_MAX)
8752 				err += efunc(pc, "invalid subr %u\n", subr);
8753 			if (rd >= nregs)
8754 				err += efunc(pc, "invalid register %u\n", rd);
8755 			if (rd == 0)
8756 				err += efunc(pc, "cannot write to %r0\n");
8757 
8758 			if (subr == DIF_SUBR_COPYOUT ||
8759 			    subr == DIF_SUBR_COPYOUTSTR) {
8760 				dp->dtdo_destructive = 1;
8761 			}
8762 			break;
8763 		case DIF_OP_PUSHTR:
8764 			if (type != DIF_TYPE_STRING && type != DIF_TYPE_CTF)
8765 				err += efunc(pc, "invalid ref type %u\n", type);
8766 			if (r2 >= nregs)
8767 				err += efunc(pc, "invalid register %u\n", r2);
8768 			if (rs >= nregs)
8769 				err += efunc(pc, "invalid register %u\n", rs);
8770 			break;
8771 		case DIF_OP_PUSHTV:
8772 			if (type != DIF_TYPE_CTF)
8773 				err += efunc(pc, "invalid val type %u\n", type);
8774 			if (r2 >= nregs)
8775 				err += efunc(pc, "invalid register %u\n", r2);
8776 			if (rs >= nregs)
8777 				err += efunc(pc, "invalid register %u\n", rs);
8778 			break;
8779 		default:
8780 			err += efunc(pc, "invalid opcode %u\n",
8781 			    DIF_INSTR_OP(instr));
8782 		}
8783 	}
8784 
8785 	if (dp->dtdo_len != 0 &&
8786 	    DIF_INSTR_OP(dp->dtdo_buf[dp->dtdo_len - 1]) != DIF_OP_RET) {
8787 		err += efunc(dp->dtdo_len - 1,
8788 		    "expected 'ret' as last DIF instruction\n");
8789 	}
8790 
8791 	if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF)) {
8792 		/*
8793 		 * If we're not returning by reference, the size must be either
8794 		 * 0 or the size of one of the base types.
8795 		 */
8796 		switch (dp->dtdo_rtype.dtdt_size) {
8797 		case 0:
8798 		case sizeof (uint8_t):
8799 		case sizeof (uint16_t):
8800 		case sizeof (uint32_t):
8801 		case sizeof (uint64_t):
8802 			break;
8803 
8804 		default:
8805 			err += efunc(dp->dtdo_len - 1, "bad return size");
8806 		}
8807 	}
8808 
8809 	for (i = 0; i < dp->dtdo_varlen && err == 0; i++) {
8810 		dtrace_difv_t *v = &dp->dtdo_vartab[i], *existing = NULL;
8811 		dtrace_diftype_t *vt, *et;
8812 		uint_t id, ndx;
8813 
8814 		if (v->dtdv_scope != DIFV_SCOPE_GLOBAL &&
8815 		    v->dtdv_scope != DIFV_SCOPE_THREAD &&
8816 		    v->dtdv_scope != DIFV_SCOPE_LOCAL) {
8817 			err += efunc(i, "unrecognized variable scope %d\n",
8818 			    v->dtdv_scope);
8819 			break;
8820 		}
8821 
8822 		if (v->dtdv_kind != DIFV_KIND_ARRAY &&
8823 		    v->dtdv_kind != DIFV_KIND_SCALAR) {
8824 			err += efunc(i, "unrecognized variable type %d\n",
8825 			    v->dtdv_kind);
8826 			break;
8827 		}
8828 
8829 		if ((id = v->dtdv_id) > DIF_VARIABLE_MAX) {
8830 			err += efunc(i, "%d exceeds variable id limit\n", id);
8831 			break;
8832 		}
8833 
8834 		if (id < DIF_VAR_OTHER_UBASE)
8835 			continue;
8836 
8837 		/*
8838 		 * For user-defined variables, we need to check that this
8839 		 * definition is identical to any previous definition that we
8840 		 * encountered.
8841 		 */
8842 		ndx = id - DIF_VAR_OTHER_UBASE;
8843 
8844 		switch (v->dtdv_scope) {
8845 		case DIFV_SCOPE_GLOBAL:
8846 			if (ndx < vstate->dtvs_nglobals) {
8847 				dtrace_statvar_t *svar;
8848 
8849 				if ((svar = vstate->dtvs_globals[ndx]) != NULL)
8850 					existing = &svar->dtsv_var;
8851 			}
8852 
8853 			break;
8854 
8855 		case DIFV_SCOPE_THREAD:
8856 			if (ndx < vstate->dtvs_ntlocals)
8857 				existing = &vstate->dtvs_tlocals[ndx];
8858 			break;
8859 
8860 		case DIFV_SCOPE_LOCAL:
8861 			if (ndx < vstate->dtvs_nlocals) {
8862 				dtrace_statvar_t *svar;
8863 
8864 				if ((svar = vstate->dtvs_locals[ndx]) != NULL)
8865 					existing = &svar->dtsv_var;
8866 			}
8867 
8868 			break;
8869 		}
8870 
8871 		vt = &v->dtdv_type;
8872 
8873 		if (vt->dtdt_flags & DIF_TF_BYREF) {
8874 			if (vt->dtdt_size == 0) {
8875 				err += efunc(i, "zero-sized variable\n");
8876 				break;
8877 			}
8878 
8879 			if (v->dtdv_scope == DIFV_SCOPE_GLOBAL &&
8880 			    vt->dtdt_size > dtrace_global_maxsize) {
8881 				err += efunc(i, "oversized by-ref global\n");
8882 				break;
8883 			}
8884 		}
8885 
8886 		if (existing == NULL || existing->dtdv_id == 0)
8887 			continue;
8888 
8889 		ASSERT(existing->dtdv_id == v->dtdv_id);
8890 		ASSERT(existing->dtdv_scope == v->dtdv_scope);
8891 
8892 		if (existing->dtdv_kind != v->dtdv_kind)
8893 			err += efunc(i, "%d changed variable kind\n", id);
8894 
8895 		et = &existing->dtdv_type;
8896 
8897 		if (vt->dtdt_flags != et->dtdt_flags) {
8898 			err += efunc(i, "%d changed variable type flags\n", id);
8899 			break;
8900 		}
8901 
8902 		if (vt->dtdt_size != 0 && vt->dtdt_size != et->dtdt_size) {
8903 			err += efunc(i, "%d changed variable type size\n", id);
8904 			break;
8905 		}
8906 	}
8907 
8908 	return (err);
8909 }
8910 
8911 /*
8912  * Validate a DTrace DIF object that it is to be used as a helper.  Helpers
8913  * are much more constrained than normal DIFOs.  Specifically, they may
8914  * not:
8915  *
8916  * 1. Make calls to subroutines other than copyin(), copyinstr() or
8917  *    miscellaneous string routines
8918  * 2. Access DTrace variables other than the args[] array, and the
8919  *    curthread, pid, ppid, tid, execname, zonename, uid and gid variables.
8920  * 3. Have thread-local variables.
8921  * 4. Have dynamic variables.
8922  */
8923 static int
8924 dtrace_difo_validate_helper(dtrace_difo_t *dp)
8925 {
8926 	int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
8927 	int err = 0;
8928 	uint_t pc;
8929 
8930 	for (pc = 0; pc < dp->dtdo_len; pc++) {
8931 		dif_instr_t instr = dp->dtdo_buf[pc];
8932 
8933 		uint_t v = DIF_INSTR_VAR(instr);
8934 		uint_t subr = DIF_INSTR_SUBR(instr);
8935 		uint_t op = DIF_INSTR_OP(instr);
8936 
8937 		switch (op) {
8938 		case DIF_OP_OR:
8939 		case DIF_OP_XOR:
8940 		case DIF_OP_AND:
8941 		case DIF_OP_SLL:
8942 		case DIF_OP_SRL:
8943 		case DIF_OP_SRA:
8944 		case DIF_OP_SUB:
8945 		case DIF_OP_ADD:
8946 		case DIF_OP_MUL:
8947 		case DIF_OP_SDIV:
8948 		case DIF_OP_UDIV:
8949 		case DIF_OP_SREM:
8950 		case DIF_OP_UREM:
8951 		case DIF_OP_COPYS:
8952 		case DIF_OP_NOT:
8953 		case DIF_OP_MOV:
8954 		case DIF_OP_RLDSB:
8955 		case DIF_OP_RLDSH:
8956 		case DIF_OP_RLDSW:
8957 		case DIF_OP_RLDUB:
8958 		case DIF_OP_RLDUH:
8959 		case DIF_OP_RLDUW:
8960 		case DIF_OP_RLDX:
8961 		case DIF_OP_ULDSB:
8962 		case DIF_OP_ULDSH:
8963 		case DIF_OP_ULDSW:
8964 		case DIF_OP_ULDUB:
8965 		case DIF_OP_ULDUH:
8966 		case DIF_OP_ULDUW:
8967 		case DIF_OP_ULDX:
8968 		case DIF_OP_STB:
8969 		case DIF_OP_STH:
8970 		case DIF_OP_STW:
8971 		case DIF_OP_STX:
8972 		case DIF_OP_ALLOCS:
8973 		case DIF_OP_CMP:
8974 		case DIF_OP_SCMP:
8975 		case DIF_OP_TST:
8976 		case DIF_OP_BA:
8977 		case DIF_OP_BE:
8978 		case DIF_OP_BNE:
8979 		case DIF_OP_BG:
8980 		case DIF_OP_BGU:
8981 		case DIF_OP_BGE:
8982 		case DIF_OP_BGEU:
8983 		case DIF_OP_BL:
8984 		case DIF_OP_BLU:
8985 		case DIF_OP_BLE:
8986 		case DIF_OP_BLEU:
8987 		case DIF_OP_RET:
8988 		case DIF_OP_NOP:
8989 		case DIF_OP_POPTS:
8990 		case DIF_OP_FLUSHTS:
8991 		case DIF_OP_SETX:
8992 		case DIF_OP_SETS:
8993 		case DIF_OP_LDGA:
8994 		case DIF_OP_LDLS:
8995 		case DIF_OP_STGS:
8996 		case DIF_OP_STLS:
8997 		case DIF_OP_PUSHTR:
8998 		case DIF_OP_PUSHTV:
8999 			break;
9000 
9001 		case DIF_OP_LDGS:
9002 			if (v >= DIF_VAR_OTHER_UBASE)
9003 				break;
9004 
9005 			if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9)
9006 				break;
9007 
9008 			if (v == DIF_VAR_CURTHREAD || v == DIF_VAR_PID ||
9009 			    v == DIF_VAR_PPID || v == DIF_VAR_TID ||
9010 			    v == DIF_VAR_EXECARGS ||
9011 			    v == DIF_VAR_EXECNAME || v == DIF_VAR_ZONENAME ||
9012 			    v == DIF_VAR_UID || v == DIF_VAR_GID)
9013 				break;
9014 
9015 			err += efunc(pc, "illegal variable %u\n", v);
9016 			break;
9017 
9018 		case DIF_OP_LDTA:
9019 		case DIF_OP_LDTS:
9020 		case DIF_OP_LDGAA:
9021 		case DIF_OP_LDTAA:
9022 			err += efunc(pc, "illegal dynamic variable load\n");
9023 			break;
9024 
9025 		case DIF_OP_STTS:
9026 		case DIF_OP_STGAA:
9027 		case DIF_OP_STTAA:
9028 			err += efunc(pc, "illegal dynamic variable store\n");
9029 			break;
9030 
9031 		case DIF_OP_CALL:
9032 			if (subr == DIF_SUBR_ALLOCA ||
9033 			    subr == DIF_SUBR_BCOPY ||
9034 			    subr == DIF_SUBR_COPYIN ||
9035 			    subr == DIF_SUBR_COPYINTO ||
9036 			    subr == DIF_SUBR_COPYINSTR ||
9037 			    subr == DIF_SUBR_INDEX ||
9038 			    subr == DIF_SUBR_INET_NTOA ||
9039 			    subr == DIF_SUBR_INET_NTOA6 ||
9040 			    subr == DIF_SUBR_INET_NTOP ||
9041 			    subr == DIF_SUBR_LLTOSTR ||
9042 			    subr == DIF_SUBR_RINDEX ||
9043 			    subr == DIF_SUBR_STRCHR ||
9044 			    subr == DIF_SUBR_STRJOIN ||
9045 			    subr == DIF_SUBR_STRRCHR ||
9046 			    subr == DIF_SUBR_STRSTR ||
9047 			    subr == DIF_SUBR_HTONS ||
9048 			    subr == DIF_SUBR_HTONL ||
9049 			    subr == DIF_SUBR_HTONLL ||
9050 			    subr == DIF_SUBR_NTOHS ||
9051 			    subr == DIF_SUBR_NTOHL ||
9052 			    subr == DIF_SUBR_NTOHLL ||
9053 			    subr == DIF_SUBR_MEMREF ||
9054 			    subr == DIF_SUBR_TYPEREF)
9055 				break;
9056 
9057 			err += efunc(pc, "invalid subr %u\n", subr);
9058 			break;
9059 
9060 		default:
9061 			err += efunc(pc, "invalid opcode %u\n",
9062 			    DIF_INSTR_OP(instr));
9063 		}
9064 	}
9065 
9066 	return (err);
9067 }
9068 
9069 /*
9070  * Returns 1 if the expression in the DIF object can be cached on a per-thread
9071  * basis; 0 if not.
9072  */
9073 static int
9074 dtrace_difo_cacheable(dtrace_difo_t *dp)
9075 {
9076 	int i;
9077 
9078 	if (dp == NULL)
9079 		return (0);
9080 
9081 	for (i = 0; i < dp->dtdo_varlen; i++) {
9082 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
9083 
9084 		if (v->dtdv_scope != DIFV_SCOPE_GLOBAL)
9085 			continue;
9086 
9087 		switch (v->dtdv_id) {
9088 		case DIF_VAR_CURTHREAD:
9089 		case DIF_VAR_PID:
9090 		case DIF_VAR_TID:
9091 		case DIF_VAR_EXECARGS:
9092 		case DIF_VAR_EXECNAME:
9093 		case DIF_VAR_ZONENAME:
9094 			break;
9095 
9096 		default:
9097 			return (0);
9098 		}
9099 	}
9100 
9101 	/*
9102 	 * This DIF object may be cacheable.  Now we need to look for any
9103 	 * array loading instructions, any memory loading instructions, or
9104 	 * any stores to thread-local variables.
9105 	 */
9106 	for (i = 0; i < dp->dtdo_len; i++) {
9107 		uint_t op = DIF_INSTR_OP(dp->dtdo_buf[i]);
9108 
9109 		if ((op >= DIF_OP_LDSB && op <= DIF_OP_LDX) ||
9110 		    (op >= DIF_OP_ULDSB && op <= DIF_OP_ULDX) ||
9111 		    (op >= DIF_OP_RLDSB && op <= DIF_OP_RLDX) ||
9112 		    op == DIF_OP_LDGA || op == DIF_OP_STTS)
9113 			return (0);
9114 	}
9115 
9116 	return (1);
9117 }
9118 
9119 static void
9120 dtrace_difo_hold(dtrace_difo_t *dp)
9121 {
9122 	int i;
9123 
9124 	ASSERT(MUTEX_HELD(&dtrace_lock));
9125 
9126 	dp->dtdo_refcnt++;
9127 	ASSERT(dp->dtdo_refcnt != 0);
9128 
9129 	/*
9130 	 * We need to check this DIF object for references to the variable
9131 	 * DIF_VAR_VTIMESTAMP.
9132 	 */
9133 	for (i = 0; i < dp->dtdo_varlen; i++) {
9134 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
9135 
9136 		if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
9137 			continue;
9138 
9139 		if (dtrace_vtime_references++ == 0)
9140 			dtrace_vtime_enable();
9141 	}
9142 }
9143 
9144 /*
9145  * This routine calculates the dynamic variable chunksize for a given DIF
9146  * object.  The calculation is not fool-proof, and can probably be tricked by
9147  * malicious DIF -- but it works for all compiler-generated DIF.  Because this
9148  * calculation is likely imperfect, dtrace_dynvar() is able to gracefully fail
9149  * if a dynamic variable size exceeds the chunksize.
9150  */
9151 static void
9152 dtrace_difo_chunksize(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9153 {
9154 	uint64_t sval = 0;
9155 	dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
9156 	const dif_instr_t *text = dp->dtdo_buf;
9157 	uint_t pc, srd = 0;
9158 	uint_t ttop = 0;
9159 	size_t size, ksize;
9160 	uint_t id, i;
9161 
9162 	for (pc = 0; pc < dp->dtdo_len; pc++) {
9163 		dif_instr_t instr = text[pc];
9164 		uint_t op = DIF_INSTR_OP(instr);
9165 		uint_t rd = DIF_INSTR_RD(instr);
9166 		uint_t r1 = DIF_INSTR_R1(instr);
9167 		uint_t nkeys = 0;
9168 		uchar_t scope = 0;
9169 
9170 		dtrace_key_t *key = tupregs;
9171 
9172 		switch (op) {
9173 		case DIF_OP_SETX:
9174 			sval = dp->dtdo_inttab[DIF_INSTR_INTEGER(instr)];
9175 			srd = rd;
9176 			continue;
9177 
9178 		case DIF_OP_STTS:
9179 			key = &tupregs[DIF_DTR_NREGS];
9180 			key[0].dttk_size = 0;
9181 			key[1].dttk_size = 0;
9182 			nkeys = 2;
9183 			scope = DIFV_SCOPE_THREAD;
9184 			break;
9185 
9186 		case DIF_OP_STGAA:
9187 		case DIF_OP_STTAA:
9188 			nkeys = ttop;
9189 
9190 			if (DIF_INSTR_OP(instr) == DIF_OP_STTAA)
9191 				key[nkeys++].dttk_size = 0;
9192 
9193 			key[nkeys++].dttk_size = 0;
9194 
9195 			if (op == DIF_OP_STTAA) {
9196 				scope = DIFV_SCOPE_THREAD;
9197 			} else {
9198 				scope = DIFV_SCOPE_GLOBAL;
9199 			}
9200 
9201 			break;
9202 
9203 		case DIF_OP_PUSHTR:
9204 			if (ttop == DIF_DTR_NREGS)
9205 				return;
9206 
9207 			if ((srd == 0 || sval == 0) && r1 == DIF_TYPE_STRING) {
9208 				/*
9209 				 * If the register for the size of the "pushtr"
9210 				 * is %r0 (or the value is 0) and the type is
9211 				 * a string, we'll use the system-wide default
9212 				 * string size.
9213 				 */
9214 				tupregs[ttop++].dttk_size =
9215 				    dtrace_strsize_default;
9216 			} else {
9217 				if (srd == 0)
9218 					return;
9219 
9220 				tupregs[ttop++].dttk_size = sval;
9221 			}
9222 
9223 			break;
9224 
9225 		case DIF_OP_PUSHTV:
9226 			if (ttop == DIF_DTR_NREGS)
9227 				return;
9228 
9229 			tupregs[ttop++].dttk_size = 0;
9230 			break;
9231 
9232 		case DIF_OP_FLUSHTS:
9233 			ttop = 0;
9234 			break;
9235 
9236 		case DIF_OP_POPTS:
9237 			if (ttop != 0)
9238 				ttop--;
9239 			break;
9240 		}
9241 
9242 		sval = 0;
9243 		srd = 0;
9244 
9245 		if (nkeys == 0)
9246 			continue;
9247 
9248 		/*
9249 		 * We have a dynamic variable allocation; calculate its size.
9250 		 */
9251 		for (ksize = 0, i = 0; i < nkeys; i++)
9252 			ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
9253 
9254 		size = sizeof (dtrace_dynvar_t);
9255 		size += sizeof (dtrace_key_t) * (nkeys - 1);
9256 		size += ksize;
9257 
9258 		/*
9259 		 * Now we need to determine the size of the stored data.
9260 		 */
9261 		id = DIF_INSTR_VAR(instr);
9262 
9263 		for (i = 0; i < dp->dtdo_varlen; i++) {
9264 			dtrace_difv_t *v = &dp->dtdo_vartab[i];
9265 
9266 			if (v->dtdv_id == id && v->dtdv_scope == scope) {
9267 				size += v->dtdv_type.dtdt_size;
9268 				break;
9269 			}
9270 		}
9271 
9272 		if (i == dp->dtdo_varlen)
9273 			return;
9274 
9275 		/*
9276 		 * We have the size.  If this is larger than the chunk size
9277 		 * for our dynamic variable state, reset the chunk size.
9278 		 */
9279 		size = P2ROUNDUP(size, sizeof (uint64_t));
9280 
9281 		if (size > vstate->dtvs_dynvars.dtds_chunksize)
9282 			vstate->dtvs_dynvars.dtds_chunksize = size;
9283 	}
9284 }
9285 
9286 static void
9287 dtrace_difo_init(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9288 {
9289 	int i, oldsvars, osz, nsz, otlocals, ntlocals;
9290 	uint_t id;
9291 
9292 	ASSERT(MUTEX_HELD(&dtrace_lock));
9293 	ASSERT(dp->dtdo_buf != NULL && dp->dtdo_len != 0);
9294 
9295 	for (i = 0; i < dp->dtdo_varlen; i++) {
9296 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
9297 		dtrace_statvar_t *svar, ***svarp = NULL;
9298 		size_t dsize = 0;
9299 		uint8_t scope = v->dtdv_scope;
9300 		int *np = NULL;
9301 
9302 		if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
9303 			continue;
9304 
9305 		id -= DIF_VAR_OTHER_UBASE;
9306 
9307 		switch (scope) {
9308 		case DIFV_SCOPE_THREAD:
9309 			while (id >= (otlocals = vstate->dtvs_ntlocals)) {
9310 				dtrace_difv_t *tlocals;
9311 
9312 				if ((ntlocals = (otlocals << 1)) == 0)
9313 					ntlocals = 1;
9314 
9315 				osz = otlocals * sizeof (dtrace_difv_t);
9316 				nsz = ntlocals * sizeof (dtrace_difv_t);
9317 
9318 				tlocals = kmem_zalloc(nsz, KM_SLEEP);
9319 
9320 				if (osz != 0) {
9321 					bcopy(vstate->dtvs_tlocals,
9322 					    tlocals, osz);
9323 					kmem_free(vstate->dtvs_tlocals, osz);
9324 				}
9325 
9326 				vstate->dtvs_tlocals = tlocals;
9327 				vstate->dtvs_ntlocals = ntlocals;
9328 			}
9329 
9330 			vstate->dtvs_tlocals[id] = *v;
9331 			continue;
9332 
9333 		case DIFV_SCOPE_LOCAL:
9334 			np = &vstate->dtvs_nlocals;
9335 			svarp = &vstate->dtvs_locals;
9336 
9337 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
9338 				dsize = NCPU * (v->dtdv_type.dtdt_size +
9339 				    sizeof (uint64_t));
9340 			else
9341 				dsize = NCPU * sizeof (uint64_t);
9342 
9343 			break;
9344 
9345 		case DIFV_SCOPE_GLOBAL:
9346 			np = &vstate->dtvs_nglobals;
9347 			svarp = &vstate->dtvs_globals;
9348 
9349 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
9350 				dsize = v->dtdv_type.dtdt_size +
9351 				    sizeof (uint64_t);
9352 
9353 			break;
9354 
9355 		default:
9356 			ASSERT(0);
9357 		}
9358 
9359 		while (id >= (oldsvars = *np)) {
9360 			dtrace_statvar_t **statics;
9361 			int newsvars, oldsize, newsize;
9362 
9363 			if ((newsvars = (oldsvars << 1)) == 0)
9364 				newsvars = 1;
9365 
9366 			oldsize = oldsvars * sizeof (dtrace_statvar_t *);
9367 			newsize = newsvars * sizeof (dtrace_statvar_t *);
9368 
9369 			statics = kmem_zalloc(newsize, KM_SLEEP);
9370 
9371 			if (oldsize != 0) {
9372 				bcopy(*svarp, statics, oldsize);
9373 				kmem_free(*svarp, oldsize);
9374 			}
9375 
9376 			*svarp = statics;
9377 			*np = newsvars;
9378 		}
9379 
9380 		if ((svar = (*svarp)[id]) == NULL) {
9381 			svar = kmem_zalloc(sizeof (dtrace_statvar_t), KM_SLEEP);
9382 			svar->dtsv_var = *v;
9383 
9384 			if ((svar->dtsv_size = dsize) != 0) {
9385 				svar->dtsv_data = (uint64_t)(uintptr_t)
9386 				    kmem_zalloc(dsize, KM_SLEEP);
9387 			}
9388 
9389 			(*svarp)[id] = svar;
9390 		}
9391 
9392 		svar->dtsv_refcnt++;
9393 	}
9394 
9395 	dtrace_difo_chunksize(dp, vstate);
9396 	dtrace_difo_hold(dp);
9397 }
9398 
9399 static dtrace_difo_t *
9400 dtrace_difo_duplicate(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9401 {
9402 	dtrace_difo_t *new;
9403 	size_t sz;
9404 
9405 	ASSERT(dp->dtdo_buf != NULL);
9406 	ASSERT(dp->dtdo_refcnt != 0);
9407 
9408 	new = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
9409 
9410 	ASSERT(dp->dtdo_buf != NULL);
9411 	sz = dp->dtdo_len * sizeof (dif_instr_t);
9412 	new->dtdo_buf = kmem_alloc(sz, KM_SLEEP);
9413 	bcopy(dp->dtdo_buf, new->dtdo_buf, sz);
9414 	new->dtdo_len = dp->dtdo_len;
9415 
9416 	if (dp->dtdo_strtab != NULL) {
9417 		ASSERT(dp->dtdo_strlen != 0);
9418 		new->dtdo_strtab = kmem_alloc(dp->dtdo_strlen, KM_SLEEP);
9419 		bcopy(dp->dtdo_strtab, new->dtdo_strtab, dp->dtdo_strlen);
9420 		new->dtdo_strlen = dp->dtdo_strlen;
9421 	}
9422 
9423 	if (dp->dtdo_inttab != NULL) {
9424 		ASSERT(dp->dtdo_intlen != 0);
9425 		sz = dp->dtdo_intlen * sizeof (uint64_t);
9426 		new->dtdo_inttab = kmem_alloc(sz, KM_SLEEP);
9427 		bcopy(dp->dtdo_inttab, new->dtdo_inttab, sz);
9428 		new->dtdo_intlen = dp->dtdo_intlen;
9429 	}
9430 
9431 	if (dp->dtdo_vartab != NULL) {
9432 		ASSERT(dp->dtdo_varlen != 0);
9433 		sz = dp->dtdo_varlen * sizeof (dtrace_difv_t);
9434 		new->dtdo_vartab = kmem_alloc(sz, KM_SLEEP);
9435 		bcopy(dp->dtdo_vartab, new->dtdo_vartab, sz);
9436 		new->dtdo_varlen = dp->dtdo_varlen;
9437 	}
9438 
9439 	dtrace_difo_init(new, vstate);
9440 	return (new);
9441 }
9442 
9443 static void
9444 dtrace_difo_destroy(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9445 {
9446 	int i;
9447 
9448 	ASSERT(dp->dtdo_refcnt == 0);
9449 
9450 	for (i = 0; i < dp->dtdo_varlen; i++) {
9451 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
9452 		dtrace_statvar_t *svar, **svarp = NULL;
9453 		uint_t id;
9454 		uint8_t scope = v->dtdv_scope;
9455 		int *np = NULL;
9456 
9457 		switch (scope) {
9458 		case DIFV_SCOPE_THREAD:
9459 			continue;
9460 
9461 		case DIFV_SCOPE_LOCAL:
9462 			np = &vstate->dtvs_nlocals;
9463 			svarp = vstate->dtvs_locals;
9464 			break;
9465 
9466 		case DIFV_SCOPE_GLOBAL:
9467 			np = &vstate->dtvs_nglobals;
9468 			svarp = vstate->dtvs_globals;
9469 			break;
9470 
9471 		default:
9472 			ASSERT(0);
9473 		}
9474 
9475 		if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
9476 			continue;
9477 
9478 		id -= DIF_VAR_OTHER_UBASE;
9479 		ASSERT(id < *np);
9480 
9481 		svar = svarp[id];
9482 		ASSERT(svar != NULL);
9483 		ASSERT(svar->dtsv_refcnt > 0);
9484 
9485 		if (--svar->dtsv_refcnt > 0)
9486 			continue;
9487 
9488 		if (svar->dtsv_size != 0) {
9489 			ASSERT(svar->dtsv_data != 0);
9490 			kmem_free((void *)(uintptr_t)svar->dtsv_data,
9491 			    svar->dtsv_size);
9492 		}
9493 
9494 		kmem_free(svar, sizeof (dtrace_statvar_t));
9495 		svarp[id] = NULL;
9496 	}
9497 
9498 	if (dp->dtdo_buf != NULL)
9499 		kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
9500 	if (dp->dtdo_inttab != NULL)
9501 		kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
9502 	if (dp->dtdo_strtab != NULL)
9503 		kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
9504 	if (dp->dtdo_vartab != NULL)
9505 		kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
9506 
9507 	kmem_free(dp, sizeof (dtrace_difo_t));
9508 }
9509 
9510 static void
9511 dtrace_difo_release(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9512 {
9513 	int i;
9514 
9515 	ASSERT(MUTEX_HELD(&dtrace_lock));
9516 	ASSERT(dp->dtdo_refcnt != 0);
9517 
9518 	for (i = 0; i < dp->dtdo_varlen; i++) {
9519 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
9520 
9521 		if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
9522 			continue;
9523 
9524 		ASSERT(dtrace_vtime_references > 0);
9525 		if (--dtrace_vtime_references == 0)
9526 			dtrace_vtime_disable();
9527 	}
9528 
9529 	if (--dp->dtdo_refcnt == 0)
9530 		dtrace_difo_destroy(dp, vstate);
9531 }
9532 
9533 /*
9534  * DTrace Format Functions
9535  */
9536 static uint16_t
9537 dtrace_format_add(dtrace_state_t *state, char *str)
9538 {
9539 	char *fmt, **new;
9540 	uint16_t ndx, len = strlen(str) + 1;
9541 
9542 	fmt = kmem_zalloc(len, KM_SLEEP);
9543 	bcopy(str, fmt, len);
9544 
9545 	for (ndx = 0; ndx < state->dts_nformats; ndx++) {
9546 		if (state->dts_formats[ndx] == NULL) {
9547 			state->dts_formats[ndx] = fmt;
9548 			return (ndx + 1);
9549 		}
9550 	}
9551 
9552 	if (state->dts_nformats == USHRT_MAX) {
9553 		/*
9554 		 * This is only likely if a denial-of-service attack is being
9555 		 * attempted.  As such, it's okay to fail silently here.
9556 		 */
9557 		kmem_free(fmt, len);
9558 		return (0);
9559 	}
9560 
9561 	/*
9562 	 * For simplicity, we always resize the formats array to be exactly the
9563 	 * number of formats.
9564 	 */
9565 	ndx = state->dts_nformats++;
9566 	new = kmem_alloc((ndx + 1) * sizeof (char *), KM_SLEEP);
9567 
9568 	if (state->dts_formats != NULL) {
9569 		ASSERT(ndx != 0);
9570 		bcopy(state->dts_formats, new, ndx * sizeof (char *));
9571 		kmem_free(state->dts_formats, ndx * sizeof (char *));
9572 	}
9573 
9574 	state->dts_formats = new;
9575 	state->dts_formats[ndx] = fmt;
9576 
9577 	return (ndx + 1);
9578 }
9579 
9580 static void
9581 dtrace_format_remove(dtrace_state_t *state, uint16_t format)
9582 {
9583 	char *fmt;
9584 
9585 	ASSERT(state->dts_formats != NULL);
9586 	ASSERT(format <= state->dts_nformats);
9587 	ASSERT(state->dts_formats[format - 1] != NULL);
9588 
9589 	fmt = state->dts_formats[format - 1];
9590 	kmem_free(fmt, strlen(fmt) + 1);
9591 	state->dts_formats[format - 1] = NULL;
9592 }
9593 
9594 static void
9595 dtrace_format_destroy(dtrace_state_t *state)
9596 {
9597 	int i;
9598 
9599 	if (state->dts_nformats == 0) {
9600 		ASSERT(state->dts_formats == NULL);
9601 		return;
9602 	}
9603 
9604 	ASSERT(state->dts_formats != NULL);
9605 
9606 	for (i = 0; i < state->dts_nformats; i++) {
9607 		char *fmt = state->dts_formats[i];
9608 
9609 		if (fmt == NULL)
9610 			continue;
9611 
9612 		kmem_free(fmt, strlen(fmt) + 1);
9613 	}
9614 
9615 	kmem_free(state->dts_formats, state->dts_nformats * sizeof (char *));
9616 	state->dts_nformats = 0;
9617 	state->dts_formats = NULL;
9618 }
9619 
9620 /*
9621  * DTrace Predicate Functions
9622  */
9623 static dtrace_predicate_t *
9624 dtrace_predicate_create(dtrace_difo_t *dp)
9625 {
9626 	dtrace_predicate_t *pred;
9627 
9628 	ASSERT(MUTEX_HELD(&dtrace_lock));
9629 	ASSERT(dp->dtdo_refcnt != 0);
9630 
9631 	pred = kmem_zalloc(sizeof (dtrace_predicate_t), KM_SLEEP);
9632 	pred->dtp_difo = dp;
9633 	pred->dtp_refcnt = 1;
9634 
9635 	if (!dtrace_difo_cacheable(dp))
9636 		return (pred);
9637 
9638 	if (dtrace_predcache_id == DTRACE_CACHEIDNONE) {
9639 		/*
9640 		 * This is only theoretically possible -- we have had 2^32
9641 		 * cacheable predicates on this machine.  We cannot allow any
9642 		 * more predicates to become cacheable:  as unlikely as it is,
9643 		 * there may be a thread caching a (now stale) predicate cache
9644 		 * ID. (N.B.: the temptation is being successfully resisted to
9645 		 * have this cmn_err() "Holy shit -- we executed this code!")
9646 		 */
9647 		return (pred);
9648 	}
9649 
9650 	pred->dtp_cacheid = dtrace_predcache_id++;
9651 
9652 	return (pred);
9653 }
9654 
9655 static void
9656 dtrace_predicate_hold(dtrace_predicate_t *pred)
9657 {
9658 	ASSERT(MUTEX_HELD(&dtrace_lock));
9659 	ASSERT(pred->dtp_difo != NULL && pred->dtp_difo->dtdo_refcnt != 0);
9660 	ASSERT(pred->dtp_refcnt > 0);
9661 
9662 	pred->dtp_refcnt++;
9663 }
9664 
9665 static void
9666 dtrace_predicate_release(dtrace_predicate_t *pred, dtrace_vstate_t *vstate)
9667 {
9668 	dtrace_difo_t *dp = pred->dtp_difo;
9669 
9670 	ASSERT(MUTEX_HELD(&dtrace_lock));
9671 	ASSERT(dp != NULL && dp->dtdo_refcnt != 0);
9672 	ASSERT(pred->dtp_refcnt > 0);
9673 
9674 	if (--pred->dtp_refcnt == 0) {
9675 		dtrace_difo_release(pred->dtp_difo, vstate);
9676 		kmem_free(pred, sizeof (dtrace_predicate_t));
9677 	}
9678 }
9679 
9680 /*
9681  * DTrace Action Description Functions
9682  */
9683 static dtrace_actdesc_t *
9684 dtrace_actdesc_create(dtrace_actkind_t kind, uint32_t ntuple,
9685     uint64_t uarg, uint64_t arg)
9686 {
9687 	dtrace_actdesc_t *act;
9688 
9689 #if defined(sun)
9690 	ASSERT(!DTRACEACT_ISPRINTFLIKE(kind) || (arg != NULL &&
9691 	    arg >= KERNELBASE) || (arg == NULL && kind == DTRACEACT_PRINTA));
9692 #endif
9693 
9694 	act = kmem_zalloc(sizeof (dtrace_actdesc_t), KM_SLEEP);
9695 	act->dtad_kind = kind;
9696 	act->dtad_ntuple = ntuple;
9697 	act->dtad_uarg = uarg;
9698 	act->dtad_arg = arg;
9699 	act->dtad_refcnt = 1;
9700 
9701 	return (act);
9702 }
9703 
9704 static void
9705 dtrace_actdesc_hold(dtrace_actdesc_t *act)
9706 {
9707 	ASSERT(act->dtad_refcnt >= 1);
9708 	act->dtad_refcnt++;
9709 }
9710 
9711 static void
9712 dtrace_actdesc_release(dtrace_actdesc_t *act, dtrace_vstate_t *vstate)
9713 {
9714 	dtrace_actkind_t kind = act->dtad_kind;
9715 	dtrace_difo_t *dp;
9716 
9717 	ASSERT(act->dtad_refcnt >= 1);
9718 
9719 	if (--act->dtad_refcnt != 0)
9720 		return;
9721 
9722 	if ((dp = act->dtad_difo) != NULL)
9723 		dtrace_difo_release(dp, vstate);
9724 
9725 	if (DTRACEACT_ISPRINTFLIKE(kind)) {
9726 		char *str = (char *)(uintptr_t)act->dtad_arg;
9727 
9728 #if defined(sun)
9729 		ASSERT((str != NULL && (uintptr_t)str >= KERNELBASE) ||
9730 		    (str == NULL && act->dtad_kind == DTRACEACT_PRINTA));
9731 #endif
9732 
9733 		if (str != NULL)
9734 			kmem_free(str, strlen(str) + 1);
9735 	}
9736 
9737 	kmem_free(act, sizeof (dtrace_actdesc_t));
9738 }
9739 
9740 /*
9741  * DTrace ECB Functions
9742  */
9743 static dtrace_ecb_t *
9744 dtrace_ecb_add(dtrace_state_t *state, dtrace_probe_t *probe)
9745 {
9746 	dtrace_ecb_t *ecb;
9747 	dtrace_epid_t epid;
9748 
9749 	ASSERT(MUTEX_HELD(&dtrace_lock));
9750 
9751 	ecb = kmem_zalloc(sizeof (dtrace_ecb_t), KM_SLEEP);
9752 	ecb->dte_predicate = NULL;
9753 	ecb->dte_probe = probe;
9754 
9755 	/*
9756 	 * The default size is the size of the default action: recording
9757 	 * the epid.
9758 	 */
9759 	ecb->dte_size = ecb->dte_needed = sizeof (dtrace_epid_t);
9760 	ecb->dte_alignment = sizeof (dtrace_epid_t);
9761 
9762 	epid = state->dts_epid++;
9763 
9764 	if (epid - 1 >= state->dts_necbs) {
9765 		dtrace_ecb_t **oecbs = state->dts_ecbs, **ecbs;
9766 		int necbs = state->dts_necbs << 1;
9767 
9768 		ASSERT(epid == state->dts_necbs + 1);
9769 
9770 		if (necbs == 0) {
9771 			ASSERT(oecbs == NULL);
9772 			necbs = 1;
9773 		}
9774 
9775 		ecbs = kmem_zalloc(necbs * sizeof (*ecbs), KM_SLEEP);
9776 
9777 		if (oecbs != NULL)
9778 			bcopy(oecbs, ecbs, state->dts_necbs * sizeof (*ecbs));
9779 
9780 		dtrace_membar_producer();
9781 		state->dts_ecbs = ecbs;
9782 
9783 		if (oecbs != NULL) {
9784 			/*
9785 			 * If this state is active, we must dtrace_sync()
9786 			 * before we can free the old dts_ecbs array:  we're
9787 			 * coming in hot, and there may be active ring
9788 			 * buffer processing (which indexes into the dts_ecbs
9789 			 * array) on another CPU.
9790 			 */
9791 			if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
9792 				dtrace_sync();
9793 
9794 			kmem_free(oecbs, state->dts_necbs * sizeof (*ecbs));
9795 		}
9796 
9797 		dtrace_membar_producer();
9798 		state->dts_necbs = necbs;
9799 	}
9800 
9801 	ecb->dte_state = state;
9802 
9803 	ASSERT(state->dts_ecbs[epid - 1] == NULL);
9804 	dtrace_membar_producer();
9805 	state->dts_ecbs[(ecb->dte_epid = epid) - 1] = ecb;
9806 
9807 	return (ecb);
9808 }
9809 
9810 static void
9811 dtrace_ecb_enable(dtrace_ecb_t *ecb)
9812 {
9813 	dtrace_probe_t *probe = ecb->dte_probe;
9814 
9815 	ASSERT(MUTEX_HELD(&cpu_lock));
9816 	ASSERT(MUTEX_HELD(&dtrace_lock));
9817 	ASSERT(ecb->dte_next == NULL);
9818 
9819 	if (probe == NULL) {
9820 		/*
9821 		 * This is the NULL probe -- there's nothing to do.
9822 		 */
9823 		return;
9824 	}
9825 
9826 	if (probe->dtpr_ecb == NULL) {
9827 		dtrace_provider_t *prov = probe->dtpr_provider;
9828 
9829 		/*
9830 		 * We're the first ECB on this probe.
9831 		 */
9832 		probe->dtpr_ecb = probe->dtpr_ecb_last = ecb;
9833 
9834 		if (ecb->dte_predicate != NULL)
9835 			probe->dtpr_predcache = ecb->dte_predicate->dtp_cacheid;
9836 
9837 		prov->dtpv_pops.dtps_enable(prov->dtpv_arg,
9838 		    probe->dtpr_id, probe->dtpr_arg);
9839 	} else {
9840 		/*
9841 		 * This probe is already active.  Swing the last pointer to
9842 		 * point to the new ECB, and issue a dtrace_sync() to assure
9843 		 * that all CPUs have seen the change.
9844 		 */
9845 		ASSERT(probe->dtpr_ecb_last != NULL);
9846 		probe->dtpr_ecb_last->dte_next = ecb;
9847 		probe->dtpr_ecb_last = ecb;
9848 		probe->dtpr_predcache = 0;
9849 
9850 		dtrace_sync();
9851 	}
9852 }
9853 
9854 static void
9855 dtrace_ecb_resize(dtrace_ecb_t *ecb)
9856 {
9857 	uint32_t maxalign = sizeof (dtrace_epid_t);
9858 	uint32_t align = sizeof (uint8_t), offs, diff;
9859 	dtrace_action_t *act;
9860 	int wastuple = 0;
9861 	uint32_t aggbase = UINT32_MAX;
9862 	dtrace_state_t *state = ecb->dte_state;
9863 
9864 	/*
9865 	 * If we record anything, we always record the epid.  (And we always
9866 	 * record it first.)
9867 	 */
9868 	offs = sizeof (dtrace_epid_t);
9869 	ecb->dte_size = ecb->dte_needed = sizeof (dtrace_epid_t);
9870 
9871 	for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
9872 		dtrace_recdesc_t *rec = &act->dta_rec;
9873 
9874 		if ((align = rec->dtrd_alignment) > maxalign)
9875 			maxalign = align;
9876 
9877 		if (!wastuple && act->dta_intuple) {
9878 			/*
9879 			 * This is the first record in a tuple.  Align the
9880 			 * offset to be at offset 4 in an 8-byte aligned
9881 			 * block.
9882 			 */
9883 			diff = offs + sizeof (dtrace_aggid_t);
9884 
9885 			if ((diff = (diff & (sizeof (uint64_t) - 1))))
9886 				offs += sizeof (uint64_t) - diff;
9887 
9888 			aggbase = offs - sizeof (dtrace_aggid_t);
9889 			ASSERT(!(aggbase & (sizeof (uint64_t) - 1)));
9890 		}
9891 
9892 		/*LINTED*/
9893 		if (rec->dtrd_size != 0 && (diff = (offs & (align - 1)))) {
9894 			/*
9895 			 * The current offset is not properly aligned; align it.
9896 			 */
9897 			offs += align - diff;
9898 		}
9899 
9900 		rec->dtrd_offset = offs;
9901 
9902 		if (offs + rec->dtrd_size > ecb->dte_needed) {
9903 			ecb->dte_needed = offs + rec->dtrd_size;
9904 
9905 			if (ecb->dte_needed > state->dts_needed)
9906 				state->dts_needed = ecb->dte_needed;
9907 		}
9908 
9909 		if (DTRACEACT_ISAGG(act->dta_kind)) {
9910 			dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
9911 			dtrace_action_t *first = agg->dtag_first, *prev;
9912 
9913 			ASSERT(rec->dtrd_size != 0 && first != NULL);
9914 			ASSERT(wastuple);
9915 			ASSERT(aggbase != UINT32_MAX);
9916 
9917 			agg->dtag_base = aggbase;
9918 
9919 			while ((prev = first->dta_prev) != NULL &&
9920 			    DTRACEACT_ISAGG(prev->dta_kind)) {
9921 				agg = (dtrace_aggregation_t *)prev;
9922 				first = agg->dtag_first;
9923 			}
9924 
9925 			if (prev != NULL) {
9926 				offs = prev->dta_rec.dtrd_offset +
9927 				    prev->dta_rec.dtrd_size;
9928 			} else {
9929 				offs = sizeof (dtrace_epid_t);
9930 			}
9931 			wastuple = 0;
9932 		} else {
9933 			if (!act->dta_intuple)
9934 				ecb->dte_size = offs + rec->dtrd_size;
9935 
9936 			offs += rec->dtrd_size;
9937 		}
9938 
9939 		wastuple = act->dta_intuple;
9940 	}
9941 
9942 	if ((act = ecb->dte_action) != NULL &&
9943 	    !(act->dta_kind == DTRACEACT_SPECULATE && act->dta_next == NULL) &&
9944 	    ecb->dte_size == sizeof (dtrace_epid_t)) {
9945 		/*
9946 		 * If the size is still sizeof (dtrace_epid_t), then all
9947 		 * actions store no data; set the size to 0.
9948 		 */
9949 		ecb->dte_alignment = maxalign;
9950 		ecb->dte_size = 0;
9951 
9952 		/*
9953 		 * If the needed space is still sizeof (dtrace_epid_t), then
9954 		 * all actions need no additional space; set the needed
9955 		 * size to 0.
9956 		 */
9957 		if (ecb->dte_needed == sizeof (dtrace_epid_t))
9958 			ecb->dte_needed = 0;
9959 
9960 		return;
9961 	}
9962 
9963 	/*
9964 	 * Set our alignment, and make sure that the dte_size and dte_needed
9965 	 * are aligned to the size of an EPID.
9966 	 */
9967 	ecb->dte_alignment = maxalign;
9968 	ecb->dte_size = (ecb->dte_size + (sizeof (dtrace_epid_t) - 1)) &
9969 	    ~(sizeof (dtrace_epid_t) - 1);
9970 	ecb->dte_needed = (ecb->dte_needed + (sizeof (dtrace_epid_t) - 1)) &
9971 	    ~(sizeof (dtrace_epid_t) - 1);
9972 	ASSERT(ecb->dte_size <= ecb->dte_needed);
9973 }
9974 
9975 static dtrace_action_t *
9976 dtrace_ecb_aggregation_create(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
9977 {
9978 	dtrace_aggregation_t *agg;
9979 	size_t size = sizeof (uint64_t);
9980 	int ntuple = desc->dtad_ntuple;
9981 	dtrace_action_t *act;
9982 	dtrace_recdesc_t *frec;
9983 	dtrace_aggid_t aggid;
9984 	dtrace_state_t *state = ecb->dte_state;
9985 
9986 	agg = kmem_zalloc(sizeof (dtrace_aggregation_t), KM_SLEEP);
9987 	agg->dtag_ecb = ecb;
9988 
9989 	ASSERT(DTRACEACT_ISAGG(desc->dtad_kind));
9990 
9991 	switch (desc->dtad_kind) {
9992 	case DTRACEAGG_MIN:
9993 		agg->dtag_initial = INT64_MAX;
9994 		agg->dtag_aggregate = dtrace_aggregate_min;
9995 		break;
9996 
9997 	case DTRACEAGG_MAX:
9998 		agg->dtag_initial = INT64_MIN;
9999 		agg->dtag_aggregate = dtrace_aggregate_max;
10000 		break;
10001 
10002 	case DTRACEAGG_COUNT:
10003 		agg->dtag_aggregate = dtrace_aggregate_count;
10004 		break;
10005 
10006 	case DTRACEAGG_QUANTIZE:
10007 		agg->dtag_aggregate = dtrace_aggregate_quantize;
10008 		size = (((sizeof (uint64_t) * NBBY) - 1) * 2 + 1) *
10009 		    sizeof (uint64_t);
10010 		break;
10011 
10012 	case DTRACEAGG_LQUANTIZE: {
10013 		uint16_t step = DTRACE_LQUANTIZE_STEP(desc->dtad_arg);
10014 		uint16_t levels = DTRACE_LQUANTIZE_LEVELS(desc->dtad_arg);
10015 
10016 		agg->dtag_initial = desc->dtad_arg;
10017 		agg->dtag_aggregate = dtrace_aggregate_lquantize;
10018 
10019 		if (step == 0 || levels == 0)
10020 			goto err;
10021 
10022 		size = levels * sizeof (uint64_t) + 3 * sizeof (uint64_t);
10023 		break;
10024 	}
10025 
10026 	case DTRACEAGG_LLQUANTIZE: {
10027 		uint16_t factor = DTRACE_LLQUANTIZE_FACTOR(desc->dtad_arg);
10028 		uint16_t low = DTRACE_LLQUANTIZE_LOW(desc->dtad_arg);
10029 		uint16_t high = DTRACE_LLQUANTIZE_HIGH(desc->dtad_arg);
10030 		uint16_t nsteps = DTRACE_LLQUANTIZE_NSTEP(desc->dtad_arg);
10031 		int64_t v;
10032 
10033 		agg->dtag_initial = desc->dtad_arg;
10034 		agg->dtag_aggregate = dtrace_aggregate_llquantize;
10035 
10036 		if (factor < 2 || low >= high || nsteps < factor)
10037 			goto err;
10038 
10039 		/*
10040 		 * Now check that the number of steps evenly divides a power
10041 		 * of the factor.  (This assures both integer bucket size and
10042 		 * linearity within each magnitude.)
10043 		 */
10044 		for (v = factor; v < nsteps; v *= factor)
10045 			continue;
10046 
10047 		if ((v % nsteps) || (nsteps % factor))
10048 			goto err;
10049 
10050 		size = (dtrace_aggregate_llquantize_bucket(factor,
10051 		    low, high, nsteps, INT64_MAX) + 2) * sizeof (uint64_t);
10052 		break;
10053 	}
10054 
10055 	case DTRACEAGG_AVG:
10056 		agg->dtag_aggregate = dtrace_aggregate_avg;
10057 		size = sizeof (uint64_t) * 2;
10058 		break;
10059 
10060 	case DTRACEAGG_STDDEV:
10061 		agg->dtag_aggregate = dtrace_aggregate_stddev;
10062 		size = sizeof (uint64_t) * 4;
10063 		break;
10064 
10065 	case DTRACEAGG_SUM:
10066 		agg->dtag_aggregate = dtrace_aggregate_sum;
10067 		break;
10068 
10069 	default:
10070 		goto err;
10071 	}
10072 
10073 	agg->dtag_action.dta_rec.dtrd_size = size;
10074 
10075 	if (ntuple == 0)
10076 		goto err;
10077 
10078 	/*
10079 	 * We must make sure that we have enough actions for the n-tuple.
10080 	 */
10081 	for (act = ecb->dte_action_last; act != NULL; act = act->dta_prev) {
10082 		if (DTRACEACT_ISAGG(act->dta_kind))
10083 			break;
10084 
10085 		if (--ntuple == 0) {
10086 			/*
10087 			 * This is the action with which our n-tuple begins.
10088 			 */
10089 			agg->dtag_first = act;
10090 			goto success;
10091 		}
10092 	}
10093 
10094 	/*
10095 	 * This n-tuple is short by ntuple elements.  Return failure.
10096 	 */
10097 	ASSERT(ntuple != 0);
10098 err:
10099 	kmem_free(agg, sizeof (dtrace_aggregation_t));
10100 	return (NULL);
10101 
10102 success:
10103 	/*
10104 	 * If the last action in the tuple has a size of zero, it's actually
10105 	 * an expression argument for the aggregating action.
10106 	 */
10107 	ASSERT(ecb->dte_action_last != NULL);
10108 	act = ecb->dte_action_last;
10109 
10110 	if (act->dta_kind == DTRACEACT_DIFEXPR) {
10111 		ASSERT(act->dta_difo != NULL);
10112 
10113 		if (act->dta_difo->dtdo_rtype.dtdt_size == 0)
10114 			agg->dtag_hasarg = 1;
10115 	}
10116 
10117 	/*
10118 	 * We need to allocate an id for this aggregation.
10119 	 */
10120 #if defined(sun)
10121 	aggid = (dtrace_aggid_t)(uintptr_t)vmem_alloc(state->dts_aggid_arena, 1,
10122 	    VM_BESTFIT | VM_SLEEP);
10123 #else
10124 	aggid = alloc_unr(state->dts_aggid_arena);
10125 #endif
10126 
10127 	if (aggid - 1 >= state->dts_naggregations) {
10128 		dtrace_aggregation_t **oaggs = state->dts_aggregations;
10129 		dtrace_aggregation_t **aggs;
10130 		int naggs = state->dts_naggregations << 1;
10131 		int onaggs = state->dts_naggregations;
10132 
10133 		ASSERT(aggid == state->dts_naggregations + 1);
10134 
10135 		if (naggs == 0) {
10136 			ASSERT(oaggs == NULL);
10137 			naggs = 1;
10138 		}
10139 
10140 		aggs = kmem_zalloc(naggs * sizeof (*aggs), KM_SLEEP);
10141 
10142 		if (oaggs != NULL) {
10143 			bcopy(oaggs, aggs, onaggs * sizeof (*aggs));
10144 			kmem_free(oaggs, onaggs * sizeof (*aggs));
10145 		}
10146 
10147 		state->dts_aggregations = aggs;
10148 		state->dts_naggregations = naggs;
10149 	}
10150 
10151 	ASSERT(state->dts_aggregations[aggid - 1] == NULL);
10152 	state->dts_aggregations[(agg->dtag_id = aggid) - 1] = agg;
10153 
10154 	frec = &agg->dtag_first->dta_rec;
10155 	if (frec->dtrd_alignment < sizeof (dtrace_aggid_t))
10156 		frec->dtrd_alignment = sizeof (dtrace_aggid_t);
10157 
10158 	for (act = agg->dtag_first; act != NULL; act = act->dta_next) {
10159 		ASSERT(!act->dta_intuple);
10160 		act->dta_intuple = 1;
10161 	}
10162 
10163 	return (&agg->dtag_action);
10164 }
10165 
10166 static void
10167 dtrace_ecb_aggregation_destroy(dtrace_ecb_t *ecb, dtrace_action_t *act)
10168 {
10169 	dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
10170 	dtrace_state_t *state = ecb->dte_state;
10171 	dtrace_aggid_t aggid = agg->dtag_id;
10172 
10173 	ASSERT(DTRACEACT_ISAGG(act->dta_kind));
10174 #if defined(sun)
10175 	vmem_free(state->dts_aggid_arena, (void *)(uintptr_t)aggid, 1);
10176 #else
10177 	free_unr(state->dts_aggid_arena, aggid);
10178 #endif
10179 
10180 	ASSERT(state->dts_aggregations[aggid - 1] == agg);
10181 	state->dts_aggregations[aggid - 1] = NULL;
10182 
10183 	kmem_free(agg, sizeof (dtrace_aggregation_t));
10184 }
10185 
10186 static int
10187 dtrace_ecb_action_add(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
10188 {
10189 	dtrace_action_t *action, *last;
10190 	dtrace_difo_t *dp = desc->dtad_difo;
10191 	uint32_t size = 0, align = sizeof (uint8_t), mask;
10192 	uint16_t format = 0;
10193 	dtrace_recdesc_t *rec;
10194 	dtrace_state_t *state = ecb->dte_state;
10195 	dtrace_optval_t *opt = state->dts_options, nframes = 0, strsize;
10196 	uint64_t arg = desc->dtad_arg;
10197 
10198 	ASSERT(MUTEX_HELD(&dtrace_lock));
10199 	ASSERT(ecb->dte_action == NULL || ecb->dte_action->dta_refcnt == 1);
10200 
10201 	if (DTRACEACT_ISAGG(desc->dtad_kind)) {
10202 		/*
10203 		 * If this is an aggregating action, there must be neither
10204 		 * a speculate nor a commit on the action chain.
10205 		 */
10206 		dtrace_action_t *act;
10207 
10208 		for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
10209 			if (act->dta_kind == DTRACEACT_COMMIT)
10210 				return (EINVAL);
10211 
10212 			if (act->dta_kind == DTRACEACT_SPECULATE)
10213 				return (EINVAL);
10214 		}
10215 
10216 		action = dtrace_ecb_aggregation_create(ecb, desc);
10217 
10218 		if (action == NULL)
10219 			return (EINVAL);
10220 	} else {
10221 		if (DTRACEACT_ISDESTRUCTIVE(desc->dtad_kind) ||
10222 		    (desc->dtad_kind == DTRACEACT_DIFEXPR &&
10223 		    dp != NULL && dp->dtdo_destructive)) {
10224 			state->dts_destructive = 1;
10225 		}
10226 
10227 		switch (desc->dtad_kind) {
10228 		case DTRACEACT_PRINTF:
10229 		case DTRACEACT_PRINTA:
10230 		case DTRACEACT_SYSTEM:
10231 		case DTRACEACT_FREOPEN:
10232 		case DTRACEACT_DIFEXPR:
10233 			/*
10234 			 * We know that our arg is a string -- turn it into a
10235 			 * format.
10236 			 */
10237 			if (arg == 0) {
10238 				ASSERT(desc->dtad_kind == DTRACEACT_PRINTA ||
10239 				    desc->dtad_kind == DTRACEACT_DIFEXPR);
10240 				format = 0;
10241 			} else {
10242 				ASSERT(arg != 0);
10243 #if defined(sun)
10244 				ASSERT(arg > KERNELBASE);
10245 #endif
10246 				format = dtrace_format_add(state,
10247 				    (char *)(uintptr_t)arg);
10248 			}
10249 
10250 			/*FALLTHROUGH*/
10251 		case DTRACEACT_LIBACT:
10252 		case DTRACEACT_TRACEMEM:
10253 		case DTRACEACT_TRACEMEM_DYNSIZE:
10254 			if (dp == NULL)
10255 				return (EINVAL);
10256 
10257 			if ((size = dp->dtdo_rtype.dtdt_size) != 0)
10258 				break;
10259 
10260 			if (dp->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING) {
10261 				if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10262 					return (EINVAL);
10263 
10264 				size = opt[DTRACEOPT_STRSIZE];
10265 			}
10266 
10267 			break;
10268 
10269 		case DTRACEACT_STACK:
10270 			if ((nframes = arg) == 0) {
10271 				nframes = opt[DTRACEOPT_STACKFRAMES];
10272 				ASSERT(nframes > 0);
10273 				arg = nframes;
10274 			}
10275 
10276 			size = nframes * sizeof (pc_t);
10277 			break;
10278 
10279 		case DTRACEACT_JSTACK:
10280 			if ((strsize = DTRACE_USTACK_STRSIZE(arg)) == 0)
10281 				strsize = opt[DTRACEOPT_JSTACKSTRSIZE];
10282 
10283 			if ((nframes = DTRACE_USTACK_NFRAMES(arg)) == 0)
10284 				nframes = opt[DTRACEOPT_JSTACKFRAMES];
10285 
10286 			arg = DTRACE_USTACK_ARG(nframes, strsize);
10287 
10288 			/*FALLTHROUGH*/
10289 		case DTRACEACT_USTACK:
10290 			if (desc->dtad_kind != DTRACEACT_JSTACK &&
10291 			    (nframes = DTRACE_USTACK_NFRAMES(arg)) == 0) {
10292 				strsize = DTRACE_USTACK_STRSIZE(arg);
10293 				nframes = opt[DTRACEOPT_USTACKFRAMES];
10294 				ASSERT(nframes > 0);
10295 				arg = DTRACE_USTACK_ARG(nframes, strsize);
10296 			}
10297 
10298 			/*
10299 			 * Save a slot for the pid.
10300 			 */
10301 			size = (nframes + 1) * sizeof (uint64_t);
10302 			size += DTRACE_USTACK_STRSIZE(arg);
10303 			size = P2ROUNDUP(size, (uint32_t)(sizeof (uintptr_t)));
10304 
10305 			break;
10306 
10307 		case DTRACEACT_SYM:
10308 		case DTRACEACT_MOD:
10309 			if (dp == NULL || ((size = dp->dtdo_rtype.dtdt_size) !=
10310 			    sizeof (uint64_t)) ||
10311 			    (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10312 				return (EINVAL);
10313 			break;
10314 
10315 		case DTRACEACT_USYM:
10316 		case DTRACEACT_UMOD:
10317 		case DTRACEACT_UADDR:
10318 			if (dp == NULL ||
10319 			    (dp->dtdo_rtype.dtdt_size != sizeof (uint64_t)) ||
10320 			    (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10321 				return (EINVAL);
10322 
10323 			/*
10324 			 * We have a slot for the pid, plus a slot for the
10325 			 * argument.  To keep things simple (aligned with
10326 			 * bitness-neutral sizing), we store each as a 64-bit
10327 			 * quantity.
10328 			 */
10329 			size = 2 * sizeof (uint64_t);
10330 			break;
10331 
10332 		case DTRACEACT_STOP:
10333 		case DTRACEACT_BREAKPOINT:
10334 		case DTRACEACT_PANIC:
10335 			break;
10336 
10337 		case DTRACEACT_CHILL:
10338 		case DTRACEACT_DISCARD:
10339 		case DTRACEACT_RAISE:
10340 			if (dp == NULL)
10341 				return (EINVAL);
10342 			break;
10343 
10344 		case DTRACEACT_EXIT:
10345 			if (dp == NULL ||
10346 			    (size = dp->dtdo_rtype.dtdt_size) != sizeof (int) ||
10347 			    (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10348 				return (EINVAL);
10349 			break;
10350 
10351 		case DTRACEACT_SPECULATE:
10352 			if (ecb->dte_size > sizeof (dtrace_epid_t))
10353 				return (EINVAL);
10354 
10355 			if (dp == NULL)
10356 				return (EINVAL);
10357 
10358 			state->dts_speculates = 1;
10359 			break;
10360 
10361 		case DTRACEACT_PRINTM:
10362 		    	size = dp->dtdo_rtype.dtdt_size;
10363 			break;
10364 
10365 		case DTRACEACT_PRINTT:
10366 		    	size = dp->dtdo_rtype.dtdt_size;
10367 			break;
10368 
10369 		case DTRACEACT_COMMIT: {
10370 			dtrace_action_t *act = ecb->dte_action;
10371 
10372 			for (; act != NULL; act = act->dta_next) {
10373 				if (act->dta_kind == DTRACEACT_COMMIT)
10374 					return (EINVAL);
10375 			}
10376 
10377 			if (dp == NULL)
10378 				return (EINVAL);
10379 			break;
10380 		}
10381 
10382 		default:
10383 			return (EINVAL);
10384 		}
10385 
10386 		if (size != 0 || desc->dtad_kind == DTRACEACT_SPECULATE) {
10387 			/*
10388 			 * If this is a data-storing action or a speculate,
10389 			 * we must be sure that there isn't a commit on the
10390 			 * action chain.
10391 			 */
10392 			dtrace_action_t *act = ecb->dte_action;
10393 
10394 			for (; act != NULL; act = act->dta_next) {
10395 				if (act->dta_kind == DTRACEACT_COMMIT)
10396 					return (EINVAL);
10397 			}
10398 		}
10399 
10400 		action = kmem_zalloc(sizeof (dtrace_action_t), KM_SLEEP);
10401 		action->dta_rec.dtrd_size = size;
10402 	}
10403 
10404 	action->dta_refcnt = 1;
10405 	rec = &action->dta_rec;
10406 	size = rec->dtrd_size;
10407 
10408 	for (mask = sizeof (uint64_t) - 1; size != 0 && mask > 0; mask >>= 1) {
10409 		if (!(size & mask)) {
10410 			align = mask + 1;
10411 			break;
10412 		}
10413 	}
10414 
10415 	action->dta_kind = desc->dtad_kind;
10416 
10417 	if ((action->dta_difo = dp) != NULL)
10418 		dtrace_difo_hold(dp);
10419 
10420 	rec->dtrd_action = action->dta_kind;
10421 	rec->dtrd_arg = arg;
10422 	rec->dtrd_uarg = desc->dtad_uarg;
10423 	rec->dtrd_alignment = (uint16_t)align;
10424 	rec->dtrd_format = format;
10425 
10426 	if ((last = ecb->dte_action_last) != NULL) {
10427 		ASSERT(ecb->dte_action != NULL);
10428 		action->dta_prev = last;
10429 		last->dta_next = action;
10430 	} else {
10431 		ASSERT(ecb->dte_action == NULL);
10432 		ecb->dte_action = action;
10433 	}
10434 
10435 	ecb->dte_action_last = action;
10436 
10437 	return (0);
10438 }
10439 
10440 static void
10441 dtrace_ecb_action_remove(dtrace_ecb_t *ecb)
10442 {
10443 	dtrace_action_t *act = ecb->dte_action, *next;
10444 	dtrace_vstate_t *vstate = &ecb->dte_state->dts_vstate;
10445 	dtrace_difo_t *dp;
10446 	uint16_t format;
10447 
10448 	if (act != NULL && act->dta_refcnt > 1) {
10449 		ASSERT(act->dta_next == NULL || act->dta_next->dta_refcnt == 1);
10450 		act->dta_refcnt--;
10451 	} else {
10452 		for (; act != NULL; act = next) {
10453 			next = act->dta_next;
10454 			ASSERT(next != NULL || act == ecb->dte_action_last);
10455 			ASSERT(act->dta_refcnt == 1);
10456 
10457 			if ((format = act->dta_rec.dtrd_format) != 0)
10458 				dtrace_format_remove(ecb->dte_state, format);
10459 
10460 			if ((dp = act->dta_difo) != NULL)
10461 				dtrace_difo_release(dp, vstate);
10462 
10463 			if (DTRACEACT_ISAGG(act->dta_kind)) {
10464 				dtrace_ecb_aggregation_destroy(ecb, act);
10465 			} else {
10466 				kmem_free(act, sizeof (dtrace_action_t));
10467 			}
10468 		}
10469 	}
10470 
10471 	ecb->dte_action = NULL;
10472 	ecb->dte_action_last = NULL;
10473 	ecb->dte_size = sizeof (dtrace_epid_t);
10474 }
10475 
10476 static void
10477 dtrace_ecb_disable(dtrace_ecb_t *ecb)
10478 {
10479 	/*
10480 	 * We disable the ECB by removing it from its probe.
10481 	 */
10482 	dtrace_ecb_t *pecb, *prev = NULL;
10483 	dtrace_probe_t *probe = ecb->dte_probe;
10484 
10485 	ASSERT(MUTEX_HELD(&dtrace_lock));
10486 
10487 	if (probe == NULL) {
10488 		/*
10489 		 * This is the NULL probe; there is nothing to disable.
10490 		 */
10491 		return;
10492 	}
10493 
10494 	for (pecb = probe->dtpr_ecb; pecb != NULL; pecb = pecb->dte_next) {
10495 		if (pecb == ecb)
10496 			break;
10497 		prev = pecb;
10498 	}
10499 
10500 	ASSERT(pecb != NULL);
10501 
10502 	if (prev == NULL) {
10503 		probe->dtpr_ecb = ecb->dte_next;
10504 	} else {
10505 		prev->dte_next = ecb->dte_next;
10506 	}
10507 
10508 	if (ecb == probe->dtpr_ecb_last) {
10509 		ASSERT(ecb->dte_next == NULL);
10510 		probe->dtpr_ecb_last = prev;
10511 	}
10512 
10513 	/*
10514 	 * The ECB has been disconnected from the probe; now sync to assure
10515 	 * that all CPUs have seen the change before returning.
10516 	 */
10517 	dtrace_sync();
10518 
10519 	if (probe->dtpr_ecb == NULL) {
10520 		/*
10521 		 * That was the last ECB on the probe; clear the predicate
10522 		 * cache ID for the probe, disable it and sync one more time
10523 		 * to assure that we'll never hit it again.
10524 		 */
10525 		dtrace_provider_t *prov = probe->dtpr_provider;
10526 
10527 		ASSERT(ecb->dte_next == NULL);
10528 		ASSERT(probe->dtpr_ecb_last == NULL);
10529 		probe->dtpr_predcache = DTRACE_CACHEIDNONE;
10530 		prov->dtpv_pops.dtps_disable(prov->dtpv_arg,
10531 		    probe->dtpr_id, probe->dtpr_arg);
10532 		dtrace_sync();
10533 	} else {
10534 		/*
10535 		 * There is at least one ECB remaining on the probe.  If there
10536 		 * is _exactly_ one, set the probe's predicate cache ID to be
10537 		 * the predicate cache ID of the remaining ECB.
10538 		 */
10539 		ASSERT(probe->dtpr_ecb_last != NULL);
10540 		ASSERT(probe->dtpr_predcache == DTRACE_CACHEIDNONE);
10541 
10542 		if (probe->dtpr_ecb == probe->dtpr_ecb_last) {
10543 			dtrace_predicate_t *p = probe->dtpr_ecb->dte_predicate;
10544 
10545 			ASSERT(probe->dtpr_ecb->dte_next == NULL);
10546 
10547 			if (p != NULL)
10548 				probe->dtpr_predcache = p->dtp_cacheid;
10549 		}
10550 
10551 		ecb->dte_next = NULL;
10552 	}
10553 }
10554 
10555 static void
10556 dtrace_ecb_destroy(dtrace_ecb_t *ecb)
10557 {
10558 	dtrace_state_t *state = ecb->dte_state;
10559 	dtrace_vstate_t *vstate = &state->dts_vstate;
10560 	dtrace_predicate_t *pred;
10561 	dtrace_epid_t epid = ecb->dte_epid;
10562 
10563 	ASSERT(MUTEX_HELD(&dtrace_lock));
10564 	ASSERT(ecb->dte_next == NULL);
10565 	ASSERT(ecb->dte_probe == NULL || ecb->dte_probe->dtpr_ecb != ecb);
10566 
10567 	if ((pred = ecb->dte_predicate) != NULL)
10568 		dtrace_predicate_release(pred, vstate);
10569 
10570 	dtrace_ecb_action_remove(ecb);
10571 
10572 	ASSERT(state->dts_ecbs[epid - 1] == ecb);
10573 	state->dts_ecbs[epid - 1] = NULL;
10574 
10575 	kmem_free(ecb, sizeof (dtrace_ecb_t));
10576 }
10577 
10578 static dtrace_ecb_t *
10579 dtrace_ecb_create(dtrace_state_t *state, dtrace_probe_t *probe,
10580     dtrace_enabling_t *enab)
10581 {
10582 	dtrace_ecb_t *ecb;
10583 	dtrace_predicate_t *pred;
10584 	dtrace_actdesc_t *act;
10585 	dtrace_provider_t *prov;
10586 	dtrace_ecbdesc_t *desc = enab->dten_current;
10587 
10588 	ASSERT(MUTEX_HELD(&dtrace_lock));
10589 	ASSERT(state != NULL);
10590 
10591 	ecb = dtrace_ecb_add(state, probe);
10592 	ecb->dte_uarg = desc->dted_uarg;
10593 
10594 	if ((pred = desc->dted_pred.dtpdd_predicate) != NULL) {
10595 		dtrace_predicate_hold(pred);
10596 		ecb->dte_predicate = pred;
10597 	}
10598 
10599 	if (probe != NULL) {
10600 		/*
10601 		 * If the provider shows more leg than the consumer is old
10602 		 * enough to see, we need to enable the appropriate implicit
10603 		 * predicate bits to prevent the ecb from activating at
10604 		 * revealing times.
10605 		 *
10606 		 * Providers specifying DTRACE_PRIV_USER at register time
10607 		 * are stating that they need the /proc-style privilege
10608 		 * model to be enforced, and this is what DTRACE_COND_OWNER
10609 		 * and DTRACE_COND_ZONEOWNER will then do at probe time.
10610 		 */
10611 		prov = probe->dtpr_provider;
10612 		if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLPROC) &&
10613 		    (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
10614 			ecb->dte_cond |= DTRACE_COND_OWNER;
10615 
10616 		if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLZONE) &&
10617 		    (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
10618 			ecb->dte_cond |= DTRACE_COND_ZONEOWNER;
10619 
10620 		/*
10621 		 * If the provider shows us kernel innards and the user
10622 		 * is lacking sufficient privilege, enable the
10623 		 * DTRACE_COND_USERMODE implicit predicate.
10624 		 */
10625 		if (!(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) &&
10626 		    (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_KERNEL))
10627 			ecb->dte_cond |= DTRACE_COND_USERMODE;
10628 	}
10629 
10630 	if (dtrace_ecb_create_cache != NULL) {
10631 		/*
10632 		 * If we have a cached ecb, we'll use its action list instead
10633 		 * of creating our own (saving both time and space).
10634 		 */
10635 		dtrace_ecb_t *cached = dtrace_ecb_create_cache;
10636 		dtrace_action_t *act = cached->dte_action;
10637 
10638 		if (act != NULL) {
10639 			ASSERT(act->dta_refcnt > 0);
10640 			act->dta_refcnt++;
10641 			ecb->dte_action = act;
10642 			ecb->dte_action_last = cached->dte_action_last;
10643 			ecb->dte_needed = cached->dte_needed;
10644 			ecb->dte_size = cached->dte_size;
10645 			ecb->dte_alignment = cached->dte_alignment;
10646 		}
10647 
10648 		return (ecb);
10649 	}
10650 
10651 	for (act = desc->dted_action; act != NULL; act = act->dtad_next) {
10652 		if ((enab->dten_error = dtrace_ecb_action_add(ecb, act)) != 0) {
10653 			dtrace_ecb_destroy(ecb);
10654 			return (NULL);
10655 		}
10656 	}
10657 
10658 	dtrace_ecb_resize(ecb);
10659 
10660 	return (dtrace_ecb_create_cache = ecb);
10661 }
10662 
10663 static int
10664 dtrace_ecb_create_enable(dtrace_probe_t *probe, void *arg)
10665 {
10666 	dtrace_ecb_t *ecb;
10667 	dtrace_enabling_t *enab = arg;
10668 	dtrace_state_t *state = enab->dten_vstate->dtvs_state;
10669 
10670 	ASSERT(state != NULL);
10671 
10672 	if (probe != NULL && probe->dtpr_gen < enab->dten_probegen) {
10673 		/*
10674 		 * This probe was created in a generation for which this
10675 		 * enabling has previously created ECBs; we don't want to
10676 		 * enable it again, so just kick out.
10677 		 */
10678 		return (DTRACE_MATCH_NEXT);
10679 	}
10680 
10681 	if ((ecb = dtrace_ecb_create(state, probe, enab)) == NULL)
10682 		return (DTRACE_MATCH_DONE);
10683 
10684 	dtrace_ecb_enable(ecb);
10685 	return (DTRACE_MATCH_NEXT);
10686 }
10687 
10688 static dtrace_ecb_t *
10689 dtrace_epid2ecb(dtrace_state_t *state, dtrace_epid_t id)
10690 {
10691 	dtrace_ecb_t *ecb;
10692 
10693 	ASSERT(MUTEX_HELD(&dtrace_lock));
10694 
10695 	if (id == 0 || id > state->dts_necbs)
10696 		return (NULL);
10697 
10698 	ASSERT(state->dts_necbs > 0 && state->dts_ecbs != NULL);
10699 	ASSERT((ecb = state->dts_ecbs[id - 1]) == NULL || ecb->dte_epid == id);
10700 
10701 	return (state->dts_ecbs[id - 1]);
10702 }
10703 
10704 static dtrace_aggregation_t *
10705 dtrace_aggid2agg(dtrace_state_t *state, dtrace_aggid_t id)
10706 {
10707 	dtrace_aggregation_t *agg;
10708 
10709 	ASSERT(MUTEX_HELD(&dtrace_lock));
10710 
10711 	if (id == 0 || id > state->dts_naggregations)
10712 		return (NULL);
10713 
10714 	ASSERT(state->dts_naggregations > 0 && state->dts_aggregations != NULL);
10715 	ASSERT((agg = state->dts_aggregations[id - 1]) == NULL ||
10716 	    agg->dtag_id == id);
10717 
10718 	return (state->dts_aggregations[id - 1]);
10719 }
10720 
10721 /*
10722  * DTrace Buffer Functions
10723  *
10724  * The following functions manipulate DTrace buffers.  Most of these functions
10725  * are called in the context of establishing or processing consumer state;
10726  * exceptions are explicitly noted.
10727  */
10728 
10729 /*
10730  * Note:  called from cross call context.  This function switches the two
10731  * buffers on a given CPU.  The atomicity of this operation is assured by
10732  * disabling interrupts while the actual switch takes place; the disabling of
10733  * interrupts serializes the execution with any execution of dtrace_probe() on
10734  * the same CPU.
10735  */
10736 static void
10737 dtrace_buffer_switch(dtrace_buffer_t *buf)
10738 {
10739 	caddr_t tomax = buf->dtb_tomax;
10740 	caddr_t xamot = buf->dtb_xamot;
10741 	dtrace_icookie_t cookie;
10742 	hrtime_t now = dtrace_gethrtime();
10743 
10744 	ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
10745 	ASSERT(!(buf->dtb_flags & DTRACEBUF_RING));
10746 
10747 	cookie = dtrace_interrupt_disable();
10748 	buf->dtb_tomax = xamot;
10749 	buf->dtb_xamot = tomax;
10750 	buf->dtb_xamot_drops = buf->dtb_drops;
10751 	buf->dtb_xamot_offset = buf->dtb_offset;
10752 	buf->dtb_xamot_errors = buf->dtb_errors;
10753 	buf->dtb_xamot_flags = buf->dtb_flags;
10754 	buf->dtb_offset = 0;
10755 	buf->dtb_drops = 0;
10756 	buf->dtb_errors = 0;
10757 	buf->dtb_flags &= ~(DTRACEBUF_ERROR | DTRACEBUF_DROPPED);
10758 	buf->dtb_interval = now - buf->dtb_switched;
10759 	buf->dtb_switched = now;
10760 	dtrace_interrupt_enable(cookie);
10761 }
10762 
10763 /*
10764  * Note:  called from cross call context.  This function activates a buffer
10765  * on a CPU.  As with dtrace_buffer_switch(), the atomicity of the operation
10766  * is guaranteed by the disabling of interrupts.
10767  */
10768 static void
10769 dtrace_buffer_activate(dtrace_state_t *state)
10770 {
10771 	dtrace_buffer_t *buf;
10772 	dtrace_icookie_t cookie = dtrace_interrupt_disable();
10773 
10774 	buf = &state->dts_buffer[curcpu];
10775 
10776 	if (buf->dtb_tomax != NULL) {
10777 		/*
10778 		 * We might like to assert that the buffer is marked inactive,
10779 		 * but this isn't necessarily true:  the buffer for the CPU
10780 		 * that processes the BEGIN probe has its buffer activated
10781 		 * manually.  In this case, we take the (harmless) action
10782 		 * re-clearing the bit INACTIVE bit.
10783 		 */
10784 		buf->dtb_flags &= ~DTRACEBUF_INACTIVE;
10785 	}
10786 
10787 	dtrace_interrupt_enable(cookie);
10788 }
10789 
10790 static int
10791 dtrace_buffer_alloc(dtrace_buffer_t *bufs, size_t size, int flags,
10792     processorid_t cpu)
10793 {
10794 #if defined(sun)
10795 	cpu_t *cp;
10796 #endif
10797 	dtrace_buffer_t *buf;
10798 
10799 #if defined(sun)
10800 	ASSERT(MUTEX_HELD(&cpu_lock));
10801 	ASSERT(MUTEX_HELD(&dtrace_lock));
10802 
10803 	if (size > dtrace_nonroot_maxsize &&
10804 	    !PRIV_POLICY_CHOICE(CRED(), PRIV_ALL, B_FALSE))
10805 		return (EFBIG);
10806 
10807 	cp = cpu_list;
10808 
10809 	do {
10810 		if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
10811 			continue;
10812 
10813 		buf = &bufs[cp->cpu_id];
10814 
10815 		/*
10816 		 * If there is already a buffer allocated for this CPU, it
10817 		 * is only possible that this is a DR event.  In this case,
10818 		 */
10819 		if (buf->dtb_tomax != NULL) {
10820 			ASSERT(buf->dtb_size == size);
10821 			continue;
10822 		}
10823 
10824 		ASSERT(buf->dtb_xamot == NULL);
10825 
10826 		if ((buf->dtb_tomax = kmem_zalloc(size, KM_NOSLEEP)) == NULL)
10827 			goto err;
10828 
10829 		buf->dtb_size = size;
10830 		buf->dtb_flags = flags;
10831 		buf->dtb_offset = 0;
10832 		buf->dtb_drops = 0;
10833 
10834 		if (flags & DTRACEBUF_NOSWITCH)
10835 			continue;
10836 
10837 		if ((buf->dtb_xamot = kmem_zalloc(size, KM_NOSLEEP)) == NULL)
10838 			goto err;
10839 	} while ((cp = cp->cpu_next) != cpu_list);
10840 
10841 	return (0);
10842 
10843 err:
10844 	cp = cpu_list;
10845 
10846 	do {
10847 		if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
10848 			continue;
10849 
10850 		buf = &bufs[cp->cpu_id];
10851 
10852 		if (buf->dtb_xamot != NULL) {
10853 			ASSERT(buf->dtb_tomax != NULL);
10854 			ASSERT(buf->dtb_size == size);
10855 			kmem_free(buf->dtb_xamot, size);
10856 		}
10857 
10858 		if (buf->dtb_tomax != NULL) {
10859 			ASSERT(buf->dtb_size == size);
10860 			kmem_free(buf->dtb_tomax, size);
10861 		}
10862 
10863 		buf->dtb_tomax = NULL;
10864 		buf->dtb_xamot = NULL;
10865 		buf->dtb_size = 0;
10866 	} while ((cp = cp->cpu_next) != cpu_list);
10867 
10868 	return (ENOMEM);
10869 #else
10870 	int i;
10871 
10872 #if defined(__amd64__) || defined(__mips__) || defined(__powerpc__)
10873 	/*
10874 	 * FreeBSD isn't good at limiting the amount of memory we
10875 	 * ask to malloc, so let's place a limit here before trying
10876 	 * to do something that might well end in tears at bedtime.
10877 	 */
10878 	if (size > physmem * PAGE_SIZE / (128 * (mp_maxid + 1)))
10879 		return(ENOMEM);
10880 #endif
10881 
10882 	ASSERT(MUTEX_HELD(&dtrace_lock));
10883 	CPU_FOREACH(i) {
10884 		if (cpu != DTRACE_CPUALL && cpu != i)
10885 			continue;
10886 
10887 		buf = &bufs[i];
10888 
10889 		/*
10890 		 * If there is already a buffer allocated for this CPU, it
10891 		 * is only possible that this is a DR event.  In this case,
10892 		 * the buffer size must match our specified size.
10893 		 */
10894 		if (buf->dtb_tomax != NULL) {
10895 			ASSERT(buf->dtb_size == size);
10896 			continue;
10897 		}
10898 
10899 		ASSERT(buf->dtb_xamot == NULL);
10900 
10901 		if ((buf->dtb_tomax = kmem_zalloc(size, KM_NOSLEEP)) == NULL)
10902 			goto err;
10903 
10904 		buf->dtb_size = size;
10905 		buf->dtb_flags = flags;
10906 		buf->dtb_offset = 0;
10907 		buf->dtb_drops = 0;
10908 
10909 		if (flags & DTRACEBUF_NOSWITCH)
10910 			continue;
10911 
10912 		if ((buf->dtb_xamot = kmem_zalloc(size, KM_NOSLEEP)) == NULL)
10913 			goto err;
10914 	}
10915 
10916 	return (0);
10917 
10918 err:
10919 	/*
10920 	 * Error allocating memory, so free the buffers that were
10921 	 * allocated before the failed allocation.
10922 	 */
10923 	CPU_FOREACH(i) {
10924 		if (cpu != DTRACE_CPUALL && cpu != i)
10925 			continue;
10926 
10927 		buf = &bufs[i];
10928 
10929 		if (buf->dtb_xamot != NULL) {
10930 			ASSERT(buf->dtb_tomax != NULL);
10931 			ASSERT(buf->dtb_size == size);
10932 			kmem_free(buf->dtb_xamot, size);
10933 		}
10934 
10935 		if (buf->dtb_tomax != NULL) {
10936 			ASSERT(buf->dtb_size == size);
10937 			kmem_free(buf->dtb_tomax, size);
10938 		}
10939 
10940 		buf->dtb_tomax = NULL;
10941 		buf->dtb_xamot = NULL;
10942 		buf->dtb_size = 0;
10943 
10944 	}
10945 
10946 	return (ENOMEM);
10947 #endif
10948 }
10949 
10950 /*
10951  * Note:  called from probe context.  This function just increments the drop
10952  * count on a buffer.  It has been made a function to allow for the
10953  * possibility of understanding the source of mysterious drop counts.  (A
10954  * problem for which one may be particularly disappointed that DTrace cannot
10955  * be used to understand DTrace.)
10956  */
10957 static void
10958 dtrace_buffer_drop(dtrace_buffer_t *buf)
10959 {
10960 	buf->dtb_drops++;
10961 }
10962 
10963 /*
10964  * Note:  called from probe context.  This function is called to reserve space
10965  * in a buffer.  If mstate is non-NULL, sets the scratch base and size in the
10966  * mstate.  Returns the new offset in the buffer, or a negative value if an
10967  * error has occurred.
10968  */
10969 static intptr_t
10970 dtrace_buffer_reserve(dtrace_buffer_t *buf, size_t needed, size_t align,
10971     dtrace_state_t *state, dtrace_mstate_t *mstate)
10972 {
10973 	intptr_t offs = buf->dtb_offset, soffs;
10974 	intptr_t woffs;
10975 	caddr_t tomax;
10976 	size_t total;
10977 
10978 	if (buf->dtb_flags & DTRACEBUF_INACTIVE)
10979 		return (-1);
10980 
10981 	if ((tomax = buf->dtb_tomax) == NULL) {
10982 		dtrace_buffer_drop(buf);
10983 		return (-1);
10984 	}
10985 
10986 	if (!(buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL))) {
10987 		while (offs & (align - 1)) {
10988 			/*
10989 			 * Assert that our alignment is off by a number which
10990 			 * is itself sizeof (uint32_t) aligned.
10991 			 */
10992 			ASSERT(!((align - (offs & (align - 1))) &
10993 			    (sizeof (uint32_t) - 1)));
10994 			DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
10995 			offs += sizeof (uint32_t);
10996 		}
10997 
10998 		if ((soffs = offs + needed) > buf->dtb_size) {
10999 			dtrace_buffer_drop(buf);
11000 			return (-1);
11001 		}
11002 
11003 		if (mstate == NULL)
11004 			return (offs);
11005 
11006 		mstate->dtms_scratch_base = (uintptr_t)tomax + soffs;
11007 		mstate->dtms_scratch_size = buf->dtb_size - soffs;
11008 		mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
11009 
11010 		return (offs);
11011 	}
11012 
11013 	if (buf->dtb_flags & DTRACEBUF_FILL) {
11014 		if (state->dts_activity != DTRACE_ACTIVITY_COOLDOWN &&
11015 		    (buf->dtb_flags & DTRACEBUF_FULL))
11016 			return (-1);
11017 		goto out;
11018 	}
11019 
11020 	total = needed + (offs & (align - 1));
11021 
11022 	/*
11023 	 * For a ring buffer, life is quite a bit more complicated.  Before
11024 	 * we can store any padding, we need to adjust our wrapping offset.
11025 	 * (If we've never before wrapped or we're not about to, no adjustment
11026 	 * is required.)
11027 	 */
11028 	if ((buf->dtb_flags & DTRACEBUF_WRAPPED) ||
11029 	    offs + total > buf->dtb_size) {
11030 		woffs = buf->dtb_xamot_offset;
11031 
11032 		if (offs + total > buf->dtb_size) {
11033 			/*
11034 			 * We can't fit in the end of the buffer.  First, a
11035 			 * sanity check that we can fit in the buffer at all.
11036 			 */
11037 			if (total > buf->dtb_size) {
11038 				dtrace_buffer_drop(buf);
11039 				return (-1);
11040 			}
11041 
11042 			/*
11043 			 * We're going to be storing at the top of the buffer,
11044 			 * so now we need to deal with the wrapped offset.  We
11045 			 * only reset our wrapped offset to 0 if it is
11046 			 * currently greater than the current offset.  If it
11047 			 * is less than the current offset, it is because a
11048 			 * previous allocation induced a wrap -- but the
11049 			 * allocation didn't subsequently take the space due
11050 			 * to an error or false predicate evaluation.  In this
11051 			 * case, we'll just leave the wrapped offset alone: if
11052 			 * the wrapped offset hasn't been advanced far enough
11053 			 * for this allocation, it will be adjusted in the
11054 			 * lower loop.
11055 			 */
11056 			if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
11057 				if (woffs >= offs)
11058 					woffs = 0;
11059 			} else {
11060 				woffs = 0;
11061 			}
11062 
11063 			/*
11064 			 * Now we know that we're going to be storing to the
11065 			 * top of the buffer and that there is room for us
11066 			 * there.  We need to clear the buffer from the current
11067 			 * offset to the end (there may be old gunk there).
11068 			 */
11069 			while (offs < buf->dtb_size)
11070 				tomax[offs++] = 0;
11071 
11072 			/*
11073 			 * We need to set our offset to zero.  And because we
11074 			 * are wrapping, we need to set the bit indicating as
11075 			 * much.  We can also adjust our needed space back
11076 			 * down to the space required by the ECB -- we know
11077 			 * that the top of the buffer is aligned.
11078 			 */
11079 			offs = 0;
11080 			total = needed;
11081 			buf->dtb_flags |= DTRACEBUF_WRAPPED;
11082 		} else {
11083 			/*
11084 			 * There is room for us in the buffer, so we simply
11085 			 * need to check the wrapped offset.
11086 			 */
11087 			if (woffs < offs) {
11088 				/*
11089 				 * The wrapped offset is less than the offset.
11090 				 * This can happen if we allocated buffer space
11091 				 * that induced a wrap, but then we didn't
11092 				 * subsequently take the space due to an error
11093 				 * or false predicate evaluation.  This is
11094 				 * okay; we know that _this_ allocation isn't
11095 				 * going to induce a wrap.  We still can't
11096 				 * reset the wrapped offset to be zero,
11097 				 * however: the space may have been trashed in
11098 				 * the previous failed probe attempt.  But at
11099 				 * least the wrapped offset doesn't need to
11100 				 * be adjusted at all...
11101 				 */
11102 				goto out;
11103 			}
11104 		}
11105 
11106 		while (offs + total > woffs) {
11107 			dtrace_epid_t epid = *(uint32_t *)(tomax + woffs);
11108 			size_t size;
11109 
11110 			if (epid == DTRACE_EPIDNONE) {
11111 				size = sizeof (uint32_t);
11112 			} else {
11113 				ASSERT(epid <= state->dts_necbs);
11114 				ASSERT(state->dts_ecbs[epid - 1] != NULL);
11115 
11116 				size = state->dts_ecbs[epid - 1]->dte_size;
11117 			}
11118 
11119 			ASSERT(woffs + size <= buf->dtb_size);
11120 			ASSERT(size != 0);
11121 
11122 			if (woffs + size == buf->dtb_size) {
11123 				/*
11124 				 * We've reached the end of the buffer; we want
11125 				 * to set the wrapped offset to 0 and break
11126 				 * out.  However, if the offs is 0, then we're
11127 				 * in a strange edge-condition:  the amount of
11128 				 * space that we want to reserve plus the size
11129 				 * of the record that we're overwriting is
11130 				 * greater than the size of the buffer.  This
11131 				 * is problematic because if we reserve the
11132 				 * space but subsequently don't consume it (due
11133 				 * to a failed predicate or error) the wrapped
11134 				 * offset will be 0 -- yet the EPID at offset 0
11135 				 * will not be committed.  This situation is
11136 				 * relatively easy to deal with:  if we're in
11137 				 * this case, the buffer is indistinguishable
11138 				 * from one that hasn't wrapped; we need only
11139 				 * finish the job by clearing the wrapped bit,
11140 				 * explicitly setting the offset to be 0, and
11141 				 * zero'ing out the old data in the buffer.
11142 				 */
11143 				if (offs == 0) {
11144 					buf->dtb_flags &= ~DTRACEBUF_WRAPPED;
11145 					buf->dtb_offset = 0;
11146 					woffs = total;
11147 
11148 					while (woffs < buf->dtb_size)
11149 						tomax[woffs++] = 0;
11150 				}
11151 
11152 				woffs = 0;
11153 				break;
11154 			}
11155 
11156 			woffs += size;
11157 		}
11158 
11159 		/*
11160 		 * We have a wrapped offset.  It may be that the wrapped offset
11161 		 * has become zero -- that's okay.
11162 		 */
11163 		buf->dtb_xamot_offset = woffs;
11164 	}
11165 
11166 out:
11167 	/*
11168 	 * Now we can plow the buffer with any necessary padding.
11169 	 */
11170 	while (offs & (align - 1)) {
11171 		/*
11172 		 * Assert that our alignment is off by a number which
11173 		 * is itself sizeof (uint32_t) aligned.
11174 		 */
11175 		ASSERT(!((align - (offs & (align - 1))) &
11176 		    (sizeof (uint32_t) - 1)));
11177 		DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
11178 		offs += sizeof (uint32_t);
11179 	}
11180 
11181 	if (buf->dtb_flags & DTRACEBUF_FILL) {
11182 		if (offs + needed > buf->dtb_size - state->dts_reserve) {
11183 			buf->dtb_flags |= DTRACEBUF_FULL;
11184 			return (-1);
11185 		}
11186 	}
11187 
11188 	if (mstate == NULL)
11189 		return (offs);
11190 
11191 	/*
11192 	 * For ring buffers and fill buffers, the scratch space is always
11193 	 * the inactive buffer.
11194 	 */
11195 	mstate->dtms_scratch_base = (uintptr_t)buf->dtb_xamot;
11196 	mstate->dtms_scratch_size = buf->dtb_size;
11197 	mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
11198 
11199 	return (offs);
11200 }
11201 
11202 static void
11203 dtrace_buffer_polish(dtrace_buffer_t *buf)
11204 {
11205 	ASSERT(buf->dtb_flags & DTRACEBUF_RING);
11206 	ASSERT(MUTEX_HELD(&dtrace_lock));
11207 
11208 	if (!(buf->dtb_flags & DTRACEBUF_WRAPPED))
11209 		return;
11210 
11211 	/*
11212 	 * We need to polish the ring buffer.  There are three cases:
11213 	 *
11214 	 * - The first (and presumably most common) is that there is no gap
11215 	 *   between the buffer offset and the wrapped offset.  In this case,
11216 	 *   there is nothing in the buffer that isn't valid data; we can
11217 	 *   mark the buffer as polished and return.
11218 	 *
11219 	 * - The second (less common than the first but still more common
11220 	 *   than the third) is that there is a gap between the buffer offset
11221 	 *   and the wrapped offset, and the wrapped offset is larger than the
11222 	 *   buffer offset.  This can happen because of an alignment issue, or
11223 	 *   can happen because of a call to dtrace_buffer_reserve() that
11224 	 *   didn't subsequently consume the buffer space.  In this case,
11225 	 *   we need to zero the data from the buffer offset to the wrapped
11226 	 *   offset.
11227 	 *
11228 	 * - The third (and least common) is that there is a gap between the
11229 	 *   buffer offset and the wrapped offset, but the wrapped offset is
11230 	 *   _less_ than the buffer offset.  This can only happen because a
11231 	 *   call to dtrace_buffer_reserve() induced a wrap, but the space
11232 	 *   was not subsequently consumed.  In this case, we need to zero the
11233 	 *   space from the offset to the end of the buffer _and_ from the
11234 	 *   top of the buffer to the wrapped offset.
11235 	 */
11236 	if (buf->dtb_offset < buf->dtb_xamot_offset) {
11237 		bzero(buf->dtb_tomax + buf->dtb_offset,
11238 		    buf->dtb_xamot_offset - buf->dtb_offset);
11239 	}
11240 
11241 	if (buf->dtb_offset > buf->dtb_xamot_offset) {
11242 		bzero(buf->dtb_tomax + buf->dtb_offset,
11243 		    buf->dtb_size - buf->dtb_offset);
11244 		bzero(buf->dtb_tomax, buf->dtb_xamot_offset);
11245 	}
11246 }
11247 
11248 /*
11249  * This routine determines if data generated at the specified time has likely
11250  * been entirely consumed at user-level.  This routine is called to determine
11251  * if an ECB on a defunct probe (but for an active enabling) can be safely
11252  * disabled and destroyed.
11253  */
11254 static int
11255 dtrace_buffer_consumed(dtrace_buffer_t *bufs, hrtime_t when)
11256 {
11257 	int i;
11258 
11259 	for (i = 0; i < NCPU; i++) {
11260 		dtrace_buffer_t *buf = &bufs[i];
11261 
11262 		if (buf->dtb_size == 0)
11263 			continue;
11264 
11265 		if (buf->dtb_flags & DTRACEBUF_RING)
11266 			return (0);
11267 
11268 		if (!buf->dtb_switched && buf->dtb_offset != 0)
11269 			return (0);
11270 
11271 		if (buf->dtb_switched - buf->dtb_interval < when)
11272 			return (0);
11273 	}
11274 
11275 	return (1);
11276 }
11277 
11278 static void
11279 dtrace_buffer_free(dtrace_buffer_t *bufs)
11280 {
11281 	int i;
11282 
11283 	for (i = 0; i < NCPU; i++) {
11284 		dtrace_buffer_t *buf = &bufs[i];
11285 
11286 		if (buf->dtb_tomax == NULL) {
11287 			ASSERT(buf->dtb_xamot == NULL);
11288 			ASSERT(buf->dtb_size == 0);
11289 			continue;
11290 		}
11291 
11292 		if (buf->dtb_xamot != NULL) {
11293 			ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
11294 			kmem_free(buf->dtb_xamot, buf->dtb_size);
11295 		}
11296 
11297 		kmem_free(buf->dtb_tomax, buf->dtb_size);
11298 		buf->dtb_size = 0;
11299 		buf->dtb_tomax = NULL;
11300 		buf->dtb_xamot = NULL;
11301 	}
11302 }
11303 
11304 /*
11305  * DTrace Enabling Functions
11306  */
11307 static dtrace_enabling_t *
11308 dtrace_enabling_create(dtrace_vstate_t *vstate)
11309 {
11310 	dtrace_enabling_t *enab;
11311 
11312 	enab = kmem_zalloc(sizeof (dtrace_enabling_t), KM_SLEEP);
11313 	enab->dten_vstate = vstate;
11314 
11315 	return (enab);
11316 }
11317 
11318 static void
11319 dtrace_enabling_add(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb)
11320 {
11321 	dtrace_ecbdesc_t **ndesc;
11322 	size_t osize, nsize;
11323 
11324 	/*
11325 	 * We can't add to enablings after we've enabled them, or after we've
11326 	 * retained them.
11327 	 */
11328 	ASSERT(enab->dten_probegen == 0);
11329 	ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
11330 
11331 	if (enab->dten_ndesc < enab->dten_maxdesc) {
11332 		enab->dten_desc[enab->dten_ndesc++] = ecb;
11333 		return;
11334 	}
11335 
11336 	osize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
11337 
11338 	if (enab->dten_maxdesc == 0) {
11339 		enab->dten_maxdesc = 1;
11340 	} else {
11341 		enab->dten_maxdesc <<= 1;
11342 	}
11343 
11344 	ASSERT(enab->dten_ndesc < enab->dten_maxdesc);
11345 
11346 	nsize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
11347 	ndesc = kmem_zalloc(nsize, KM_SLEEP);
11348 	bcopy(enab->dten_desc, ndesc, osize);
11349 	if (enab->dten_desc != NULL)
11350 		kmem_free(enab->dten_desc, osize);
11351 
11352 	enab->dten_desc = ndesc;
11353 	enab->dten_desc[enab->dten_ndesc++] = ecb;
11354 }
11355 
11356 static void
11357 dtrace_enabling_addlike(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb,
11358     dtrace_probedesc_t *pd)
11359 {
11360 	dtrace_ecbdesc_t *new;
11361 	dtrace_predicate_t *pred;
11362 	dtrace_actdesc_t *act;
11363 
11364 	/*
11365 	 * We're going to create a new ECB description that matches the
11366 	 * specified ECB in every way, but has the specified probe description.
11367 	 */
11368 	new = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
11369 
11370 	if ((pred = ecb->dted_pred.dtpdd_predicate) != NULL)
11371 		dtrace_predicate_hold(pred);
11372 
11373 	for (act = ecb->dted_action; act != NULL; act = act->dtad_next)
11374 		dtrace_actdesc_hold(act);
11375 
11376 	new->dted_action = ecb->dted_action;
11377 	new->dted_pred = ecb->dted_pred;
11378 	new->dted_probe = *pd;
11379 	new->dted_uarg = ecb->dted_uarg;
11380 
11381 	dtrace_enabling_add(enab, new);
11382 }
11383 
11384 static void
11385 dtrace_enabling_dump(dtrace_enabling_t *enab)
11386 {
11387 	int i;
11388 
11389 	for (i = 0; i < enab->dten_ndesc; i++) {
11390 		dtrace_probedesc_t *desc = &enab->dten_desc[i]->dted_probe;
11391 
11392 		cmn_err(CE_NOTE, "enabling probe %d (%s:%s:%s:%s)", i,
11393 		    desc->dtpd_provider, desc->dtpd_mod,
11394 		    desc->dtpd_func, desc->dtpd_name);
11395 	}
11396 }
11397 
11398 static void
11399 dtrace_enabling_destroy(dtrace_enabling_t *enab)
11400 {
11401 	int i;
11402 	dtrace_ecbdesc_t *ep;
11403 	dtrace_vstate_t *vstate = enab->dten_vstate;
11404 
11405 	ASSERT(MUTEX_HELD(&dtrace_lock));
11406 
11407 	for (i = 0; i < enab->dten_ndesc; i++) {
11408 		dtrace_actdesc_t *act, *next;
11409 		dtrace_predicate_t *pred;
11410 
11411 		ep = enab->dten_desc[i];
11412 
11413 		if ((pred = ep->dted_pred.dtpdd_predicate) != NULL)
11414 			dtrace_predicate_release(pred, vstate);
11415 
11416 		for (act = ep->dted_action; act != NULL; act = next) {
11417 			next = act->dtad_next;
11418 			dtrace_actdesc_release(act, vstate);
11419 		}
11420 
11421 		kmem_free(ep, sizeof (dtrace_ecbdesc_t));
11422 	}
11423 
11424 	if (enab->dten_desc != NULL)
11425 		kmem_free(enab->dten_desc,
11426 		    enab->dten_maxdesc * sizeof (dtrace_enabling_t *));
11427 
11428 	/*
11429 	 * If this was a retained enabling, decrement the dts_nretained count
11430 	 * and take it off of the dtrace_retained list.
11431 	 */
11432 	if (enab->dten_prev != NULL || enab->dten_next != NULL ||
11433 	    dtrace_retained == enab) {
11434 		ASSERT(enab->dten_vstate->dtvs_state != NULL);
11435 		ASSERT(enab->dten_vstate->dtvs_state->dts_nretained > 0);
11436 		enab->dten_vstate->dtvs_state->dts_nretained--;
11437 	}
11438 
11439 	if (enab->dten_prev == NULL) {
11440 		if (dtrace_retained == enab) {
11441 			dtrace_retained = enab->dten_next;
11442 
11443 			if (dtrace_retained != NULL)
11444 				dtrace_retained->dten_prev = NULL;
11445 		}
11446 	} else {
11447 		ASSERT(enab != dtrace_retained);
11448 		ASSERT(dtrace_retained != NULL);
11449 		enab->dten_prev->dten_next = enab->dten_next;
11450 	}
11451 
11452 	if (enab->dten_next != NULL) {
11453 		ASSERT(dtrace_retained != NULL);
11454 		enab->dten_next->dten_prev = enab->dten_prev;
11455 	}
11456 
11457 	kmem_free(enab, sizeof (dtrace_enabling_t));
11458 }
11459 
11460 static int
11461 dtrace_enabling_retain(dtrace_enabling_t *enab)
11462 {
11463 	dtrace_state_t *state;
11464 
11465 	ASSERT(MUTEX_HELD(&dtrace_lock));
11466 	ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
11467 	ASSERT(enab->dten_vstate != NULL);
11468 
11469 	state = enab->dten_vstate->dtvs_state;
11470 	ASSERT(state != NULL);
11471 
11472 	/*
11473 	 * We only allow each state to retain dtrace_retain_max enablings.
11474 	 */
11475 	if (state->dts_nretained >= dtrace_retain_max)
11476 		return (ENOSPC);
11477 
11478 	state->dts_nretained++;
11479 
11480 	if (dtrace_retained == NULL) {
11481 		dtrace_retained = enab;
11482 		return (0);
11483 	}
11484 
11485 	enab->dten_next = dtrace_retained;
11486 	dtrace_retained->dten_prev = enab;
11487 	dtrace_retained = enab;
11488 
11489 	return (0);
11490 }
11491 
11492 static int
11493 dtrace_enabling_replicate(dtrace_state_t *state, dtrace_probedesc_t *match,
11494     dtrace_probedesc_t *create)
11495 {
11496 	dtrace_enabling_t *new, *enab;
11497 	int found = 0, err = ENOENT;
11498 
11499 	ASSERT(MUTEX_HELD(&dtrace_lock));
11500 	ASSERT(strlen(match->dtpd_provider) < DTRACE_PROVNAMELEN);
11501 	ASSERT(strlen(match->dtpd_mod) < DTRACE_MODNAMELEN);
11502 	ASSERT(strlen(match->dtpd_func) < DTRACE_FUNCNAMELEN);
11503 	ASSERT(strlen(match->dtpd_name) < DTRACE_NAMELEN);
11504 
11505 	new = dtrace_enabling_create(&state->dts_vstate);
11506 
11507 	/*
11508 	 * Iterate over all retained enablings, looking for enablings that
11509 	 * match the specified state.
11510 	 */
11511 	for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11512 		int i;
11513 
11514 		/*
11515 		 * dtvs_state can only be NULL for helper enablings -- and
11516 		 * helper enablings can't be retained.
11517 		 */
11518 		ASSERT(enab->dten_vstate->dtvs_state != NULL);
11519 
11520 		if (enab->dten_vstate->dtvs_state != state)
11521 			continue;
11522 
11523 		/*
11524 		 * Now iterate over each probe description; we're looking for
11525 		 * an exact match to the specified probe description.
11526 		 */
11527 		for (i = 0; i < enab->dten_ndesc; i++) {
11528 			dtrace_ecbdesc_t *ep = enab->dten_desc[i];
11529 			dtrace_probedesc_t *pd = &ep->dted_probe;
11530 
11531 			if (strcmp(pd->dtpd_provider, match->dtpd_provider))
11532 				continue;
11533 
11534 			if (strcmp(pd->dtpd_mod, match->dtpd_mod))
11535 				continue;
11536 
11537 			if (strcmp(pd->dtpd_func, match->dtpd_func))
11538 				continue;
11539 
11540 			if (strcmp(pd->dtpd_name, match->dtpd_name))
11541 				continue;
11542 
11543 			/*
11544 			 * We have a winning probe!  Add it to our growing
11545 			 * enabling.
11546 			 */
11547 			found = 1;
11548 			dtrace_enabling_addlike(new, ep, create);
11549 		}
11550 	}
11551 
11552 	if (!found || (err = dtrace_enabling_retain(new)) != 0) {
11553 		dtrace_enabling_destroy(new);
11554 		return (err);
11555 	}
11556 
11557 	return (0);
11558 }
11559 
11560 static void
11561 dtrace_enabling_retract(dtrace_state_t *state)
11562 {
11563 	dtrace_enabling_t *enab, *next;
11564 
11565 	ASSERT(MUTEX_HELD(&dtrace_lock));
11566 
11567 	/*
11568 	 * Iterate over all retained enablings, destroy the enablings retained
11569 	 * for the specified state.
11570 	 */
11571 	for (enab = dtrace_retained; enab != NULL; enab = next) {
11572 		next = enab->dten_next;
11573 
11574 		/*
11575 		 * dtvs_state can only be NULL for helper enablings -- and
11576 		 * helper enablings can't be retained.
11577 		 */
11578 		ASSERT(enab->dten_vstate->dtvs_state != NULL);
11579 
11580 		if (enab->dten_vstate->dtvs_state == state) {
11581 			ASSERT(state->dts_nretained > 0);
11582 			dtrace_enabling_destroy(enab);
11583 		}
11584 	}
11585 
11586 	ASSERT(state->dts_nretained == 0);
11587 }
11588 
11589 static int
11590 dtrace_enabling_match(dtrace_enabling_t *enab, int *nmatched)
11591 {
11592 	int i = 0;
11593 	int matched = 0;
11594 
11595 	ASSERT(MUTEX_HELD(&cpu_lock));
11596 	ASSERT(MUTEX_HELD(&dtrace_lock));
11597 
11598 	for (i = 0; i < enab->dten_ndesc; i++) {
11599 		dtrace_ecbdesc_t *ep = enab->dten_desc[i];
11600 
11601 		enab->dten_current = ep;
11602 		enab->dten_error = 0;
11603 
11604 		matched += dtrace_probe_enable(&ep->dted_probe, enab);
11605 
11606 		if (enab->dten_error != 0) {
11607 			/*
11608 			 * If we get an error half-way through enabling the
11609 			 * probes, we kick out -- perhaps with some number of
11610 			 * them enabled.  Leaving enabled probes enabled may
11611 			 * be slightly confusing for user-level, but we expect
11612 			 * that no one will attempt to actually drive on in
11613 			 * the face of such errors.  If this is an anonymous
11614 			 * enabling (indicated with a NULL nmatched pointer),
11615 			 * we cmn_err() a message.  We aren't expecting to
11616 			 * get such an error -- such as it can exist at all,
11617 			 * it would be a result of corrupted DOF in the driver
11618 			 * properties.
11619 			 */
11620 			if (nmatched == NULL) {
11621 				cmn_err(CE_WARN, "dtrace_enabling_match() "
11622 				    "error on %p: %d", (void *)ep,
11623 				    enab->dten_error);
11624 			}
11625 
11626 			return (enab->dten_error);
11627 		}
11628 	}
11629 
11630 	enab->dten_probegen = dtrace_probegen;
11631 	if (nmatched != NULL)
11632 		*nmatched = matched;
11633 
11634 	return (0);
11635 }
11636 
11637 static void
11638 dtrace_enabling_matchall(void)
11639 {
11640 	dtrace_enabling_t *enab;
11641 
11642 	mutex_enter(&cpu_lock);
11643 	mutex_enter(&dtrace_lock);
11644 
11645 	/*
11646 	 * Iterate over all retained enablings to see if any probes match
11647 	 * against them.  We only perform this operation on enablings for which
11648 	 * we have sufficient permissions by virtue of being in the global zone
11649 	 * or in the same zone as the DTrace client.  Because we can be called
11650 	 * after dtrace_detach() has been called, we cannot assert that there
11651 	 * are retained enablings.  We can safely load from dtrace_retained,
11652 	 * however:  the taskq_destroy() at the end of dtrace_detach() will
11653 	 * block pending our completion.
11654 	 */
11655 	for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11656 #if defined(sun)
11657 		cred_t *cr = enab->dten_vstate->dtvs_state->dts_cred.dcr_cred;
11658 
11659 		if (INGLOBALZONE(curproc) || getzoneid() == crgetzoneid(cr))
11660 #endif
11661 			(void) dtrace_enabling_match(enab, NULL);
11662 	}
11663 
11664 	mutex_exit(&dtrace_lock);
11665 	mutex_exit(&cpu_lock);
11666 }
11667 
11668 /*
11669  * If an enabling is to be enabled without having matched probes (that is, if
11670  * dtrace_state_go() is to be called on the underlying dtrace_state_t), the
11671  * enabling must be _primed_ by creating an ECB for every ECB description.
11672  * This must be done to assure that we know the number of speculations, the
11673  * number of aggregations, the minimum buffer size needed, etc. before we
11674  * transition out of DTRACE_ACTIVITY_INACTIVE.  To do this without actually
11675  * enabling any probes, we create ECBs for every ECB decription, but with a
11676  * NULL probe -- which is exactly what this function does.
11677  */
11678 static void
11679 dtrace_enabling_prime(dtrace_state_t *state)
11680 {
11681 	dtrace_enabling_t *enab;
11682 	int i;
11683 
11684 	for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11685 		ASSERT(enab->dten_vstate->dtvs_state != NULL);
11686 
11687 		if (enab->dten_vstate->dtvs_state != state)
11688 			continue;
11689 
11690 		/*
11691 		 * We don't want to prime an enabling more than once, lest
11692 		 * we allow a malicious user to induce resource exhaustion.
11693 		 * (The ECBs that result from priming an enabling aren't
11694 		 * leaked -- but they also aren't deallocated until the
11695 		 * consumer state is destroyed.)
11696 		 */
11697 		if (enab->dten_primed)
11698 			continue;
11699 
11700 		for (i = 0; i < enab->dten_ndesc; i++) {
11701 			enab->dten_current = enab->dten_desc[i];
11702 			(void) dtrace_probe_enable(NULL, enab);
11703 		}
11704 
11705 		enab->dten_primed = 1;
11706 	}
11707 }
11708 
11709 /*
11710  * Called to indicate that probes should be provided due to retained
11711  * enablings.  This is implemented in terms of dtrace_probe_provide(), but it
11712  * must take an initial lap through the enabling calling the dtps_provide()
11713  * entry point explicitly to allow for autocreated probes.
11714  */
11715 static void
11716 dtrace_enabling_provide(dtrace_provider_t *prv)
11717 {
11718 	int i, all = 0;
11719 	dtrace_probedesc_t desc;
11720 
11721 	ASSERT(MUTEX_HELD(&dtrace_lock));
11722 	ASSERT(MUTEX_HELD(&dtrace_provider_lock));
11723 
11724 	if (prv == NULL) {
11725 		all = 1;
11726 		prv = dtrace_provider;
11727 	}
11728 
11729 	do {
11730 		dtrace_enabling_t *enab = dtrace_retained;
11731 		void *parg = prv->dtpv_arg;
11732 
11733 		for (; enab != NULL; enab = enab->dten_next) {
11734 			for (i = 0; i < enab->dten_ndesc; i++) {
11735 				desc = enab->dten_desc[i]->dted_probe;
11736 				mutex_exit(&dtrace_lock);
11737 				prv->dtpv_pops.dtps_provide(parg, &desc);
11738 				mutex_enter(&dtrace_lock);
11739 			}
11740 		}
11741 	} while (all && (prv = prv->dtpv_next) != NULL);
11742 
11743 	mutex_exit(&dtrace_lock);
11744 	dtrace_probe_provide(NULL, all ? NULL : prv);
11745 	mutex_enter(&dtrace_lock);
11746 }
11747 
11748 /*
11749  * Called to reap ECBs that are attached to probes from defunct providers.
11750  */
11751 static void
11752 dtrace_enabling_reap(void)
11753 {
11754 	dtrace_provider_t *prov;
11755 	dtrace_probe_t *probe;
11756 	dtrace_ecb_t *ecb;
11757 	hrtime_t when;
11758 	int i;
11759 
11760 	mutex_enter(&cpu_lock);
11761 	mutex_enter(&dtrace_lock);
11762 
11763 	for (i = 0; i < dtrace_nprobes; i++) {
11764 		if ((probe = dtrace_probes[i]) == NULL)
11765 			continue;
11766 
11767 		if (probe->dtpr_ecb == NULL)
11768 			continue;
11769 
11770 		prov = probe->dtpr_provider;
11771 
11772 		if ((when = prov->dtpv_defunct) == 0)
11773 			continue;
11774 
11775 		/*
11776 		 * We have ECBs on a defunct provider:  we want to reap these
11777 		 * ECBs to allow the provider to unregister.  The destruction
11778 		 * of these ECBs must be done carefully:  if we destroy the ECB
11779 		 * and the consumer later wishes to consume an EPID that
11780 		 * corresponds to the destroyed ECB (and if the EPID metadata
11781 		 * has not been previously consumed), the consumer will abort
11782 		 * processing on the unknown EPID.  To reduce (but not, sadly,
11783 		 * eliminate) the possibility of this, we will only destroy an
11784 		 * ECB for a defunct provider if, for the state that
11785 		 * corresponds to the ECB:
11786 		 *
11787 		 *  (a)	There is no speculative tracing (which can effectively
11788 		 *	cache an EPID for an arbitrary amount of time).
11789 		 *
11790 		 *  (b)	The principal buffers have been switched twice since the
11791 		 *	provider became defunct.
11792 		 *
11793 		 *  (c)	The aggregation buffers are of zero size or have been
11794 		 *	switched twice since the provider became defunct.
11795 		 *
11796 		 * We use dts_speculates to determine (a) and call a function
11797 		 * (dtrace_buffer_consumed()) to determine (b) and (c).  Note
11798 		 * that as soon as we've been unable to destroy one of the ECBs
11799 		 * associated with the probe, we quit trying -- reaping is only
11800 		 * fruitful in as much as we can destroy all ECBs associated
11801 		 * with the defunct provider's probes.
11802 		 */
11803 		while ((ecb = probe->dtpr_ecb) != NULL) {
11804 			dtrace_state_t *state = ecb->dte_state;
11805 			dtrace_buffer_t *buf = state->dts_buffer;
11806 			dtrace_buffer_t *aggbuf = state->dts_aggbuffer;
11807 
11808 			if (state->dts_speculates)
11809 				break;
11810 
11811 			if (!dtrace_buffer_consumed(buf, when))
11812 				break;
11813 
11814 			if (!dtrace_buffer_consumed(aggbuf, when))
11815 				break;
11816 
11817 			dtrace_ecb_disable(ecb);
11818 			ASSERT(probe->dtpr_ecb != ecb);
11819 			dtrace_ecb_destroy(ecb);
11820 		}
11821 	}
11822 
11823 	mutex_exit(&dtrace_lock);
11824 	mutex_exit(&cpu_lock);
11825 }
11826 
11827 /*
11828  * DTrace DOF Functions
11829  */
11830 /*ARGSUSED*/
11831 static void
11832 dtrace_dof_error(dof_hdr_t *dof, const char *str)
11833 {
11834 	if (dtrace_err_verbose)
11835 		cmn_err(CE_WARN, "failed to process DOF: %s", str);
11836 
11837 #ifdef DTRACE_ERRDEBUG
11838 	dtrace_errdebug(str);
11839 #endif
11840 }
11841 
11842 /*
11843  * Create DOF out of a currently enabled state.  Right now, we only create
11844  * DOF containing the run-time options -- but this could be expanded to create
11845  * complete DOF representing the enabled state.
11846  */
11847 static dof_hdr_t *
11848 dtrace_dof_create(dtrace_state_t *state)
11849 {
11850 	dof_hdr_t *dof;
11851 	dof_sec_t *sec;
11852 	dof_optdesc_t *opt;
11853 	int i, len = sizeof (dof_hdr_t) +
11854 	    roundup(sizeof (dof_sec_t), sizeof (uint64_t)) +
11855 	    sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
11856 
11857 	ASSERT(MUTEX_HELD(&dtrace_lock));
11858 
11859 	dof = kmem_zalloc(len, KM_SLEEP);
11860 	dof->dofh_ident[DOF_ID_MAG0] = DOF_MAG_MAG0;
11861 	dof->dofh_ident[DOF_ID_MAG1] = DOF_MAG_MAG1;
11862 	dof->dofh_ident[DOF_ID_MAG2] = DOF_MAG_MAG2;
11863 	dof->dofh_ident[DOF_ID_MAG3] = DOF_MAG_MAG3;
11864 
11865 	dof->dofh_ident[DOF_ID_MODEL] = DOF_MODEL_NATIVE;
11866 	dof->dofh_ident[DOF_ID_ENCODING] = DOF_ENCODE_NATIVE;
11867 	dof->dofh_ident[DOF_ID_VERSION] = DOF_VERSION;
11868 	dof->dofh_ident[DOF_ID_DIFVERS] = DIF_VERSION;
11869 	dof->dofh_ident[DOF_ID_DIFIREG] = DIF_DIR_NREGS;
11870 	dof->dofh_ident[DOF_ID_DIFTREG] = DIF_DTR_NREGS;
11871 
11872 	dof->dofh_flags = 0;
11873 	dof->dofh_hdrsize = sizeof (dof_hdr_t);
11874 	dof->dofh_secsize = sizeof (dof_sec_t);
11875 	dof->dofh_secnum = 1;	/* only DOF_SECT_OPTDESC */
11876 	dof->dofh_secoff = sizeof (dof_hdr_t);
11877 	dof->dofh_loadsz = len;
11878 	dof->dofh_filesz = len;
11879 	dof->dofh_pad = 0;
11880 
11881 	/*
11882 	 * Fill in the option section header...
11883 	 */
11884 	sec = (dof_sec_t *)((uintptr_t)dof + sizeof (dof_hdr_t));
11885 	sec->dofs_type = DOF_SECT_OPTDESC;
11886 	sec->dofs_align = sizeof (uint64_t);
11887 	sec->dofs_flags = DOF_SECF_LOAD;
11888 	sec->dofs_entsize = sizeof (dof_optdesc_t);
11889 
11890 	opt = (dof_optdesc_t *)((uintptr_t)sec +
11891 	    roundup(sizeof (dof_sec_t), sizeof (uint64_t)));
11892 
11893 	sec->dofs_offset = (uintptr_t)opt - (uintptr_t)dof;
11894 	sec->dofs_size = sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
11895 
11896 	for (i = 0; i < DTRACEOPT_MAX; i++) {
11897 		opt[i].dofo_option = i;
11898 		opt[i].dofo_strtab = DOF_SECIDX_NONE;
11899 		opt[i].dofo_value = state->dts_options[i];
11900 	}
11901 
11902 	return (dof);
11903 }
11904 
11905 static dof_hdr_t *
11906 dtrace_dof_copyin(uintptr_t uarg, int *errp)
11907 {
11908 	dof_hdr_t hdr, *dof;
11909 
11910 	ASSERT(!MUTEX_HELD(&dtrace_lock));
11911 
11912 	/*
11913 	 * First, we're going to copyin() the sizeof (dof_hdr_t).
11914 	 */
11915 	if (copyin((void *)uarg, &hdr, sizeof (hdr)) != 0) {
11916 		dtrace_dof_error(NULL, "failed to copyin DOF header");
11917 		*errp = EFAULT;
11918 		return (NULL);
11919 	}
11920 
11921 	/*
11922 	 * Now we'll allocate the entire DOF and copy it in -- provided
11923 	 * that the length isn't outrageous.
11924 	 */
11925 	if (hdr.dofh_loadsz >= dtrace_dof_maxsize) {
11926 		dtrace_dof_error(&hdr, "load size exceeds maximum");
11927 		*errp = E2BIG;
11928 		return (NULL);
11929 	}
11930 
11931 	if (hdr.dofh_loadsz < sizeof (hdr)) {
11932 		dtrace_dof_error(&hdr, "invalid load size");
11933 		*errp = EINVAL;
11934 		return (NULL);
11935 	}
11936 
11937 	dof = kmem_alloc(hdr.dofh_loadsz, KM_SLEEP);
11938 
11939 	if (copyin((void *)uarg, dof, hdr.dofh_loadsz) != 0) {
11940 		kmem_free(dof, hdr.dofh_loadsz);
11941 		*errp = EFAULT;
11942 		return (NULL);
11943 	}
11944 
11945 	return (dof);
11946 }
11947 
11948 #if !defined(sun)
11949 static __inline uchar_t
11950 dtrace_dof_char(char c) {
11951 	switch (c) {
11952 	case '0':
11953 	case '1':
11954 	case '2':
11955 	case '3':
11956 	case '4':
11957 	case '5':
11958 	case '6':
11959 	case '7':
11960 	case '8':
11961 	case '9':
11962 		return (c - '0');
11963 	case 'A':
11964 	case 'B':
11965 	case 'C':
11966 	case 'D':
11967 	case 'E':
11968 	case 'F':
11969 		return (c - 'A' + 10);
11970 	case 'a':
11971 	case 'b':
11972 	case 'c':
11973 	case 'd':
11974 	case 'e':
11975 	case 'f':
11976 		return (c - 'a' + 10);
11977 	}
11978 	/* Should not reach here. */
11979 	return (0);
11980 }
11981 #endif
11982 
11983 static dof_hdr_t *
11984 dtrace_dof_property(const char *name)
11985 {
11986 	uchar_t *buf;
11987 	uint64_t loadsz;
11988 	unsigned int len, i;
11989 	dof_hdr_t *dof;
11990 
11991 #if defined(sun)
11992 	/*
11993 	 * Unfortunately, array of values in .conf files are always (and
11994 	 * only) interpreted to be integer arrays.  We must read our DOF
11995 	 * as an integer array, and then squeeze it into a byte array.
11996 	 */
11997 	if (ddi_prop_lookup_int_array(DDI_DEV_T_ANY, dtrace_devi, 0,
11998 	    (char *)name, (int **)&buf, &len) != DDI_PROP_SUCCESS)
11999 		return (NULL);
12000 
12001 	for (i = 0; i < len; i++)
12002 		buf[i] = (uchar_t)(((int *)buf)[i]);
12003 
12004 	if (len < sizeof (dof_hdr_t)) {
12005 		ddi_prop_free(buf);
12006 		dtrace_dof_error(NULL, "truncated header");
12007 		return (NULL);
12008 	}
12009 
12010 	if (len < (loadsz = ((dof_hdr_t *)buf)->dofh_loadsz)) {
12011 		ddi_prop_free(buf);
12012 		dtrace_dof_error(NULL, "truncated DOF");
12013 		return (NULL);
12014 	}
12015 
12016 	if (loadsz >= dtrace_dof_maxsize) {
12017 		ddi_prop_free(buf);
12018 		dtrace_dof_error(NULL, "oversized DOF");
12019 		return (NULL);
12020 	}
12021 
12022 	dof = kmem_alloc(loadsz, KM_SLEEP);
12023 	bcopy(buf, dof, loadsz);
12024 	ddi_prop_free(buf);
12025 #else
12026 	char *p;
12027 	char *p_env;
12028 
12029 	if ((p_env = getenv(name)) == NULL)
12030 		return (NULL);
12031 
12032 	len = strlen(p_env) / 2;
12033 
12034 	buf = kmem_alloc(len, KM_SLEEP);
12035 
12036 	dof = (dof_hdr_t *) buf;
12037 
12038 	p = p_env;
12039 
12040 	for (i = 0; i < len; i++) {
12041 		buf[i] = (dtrace_dof_char(p[0]) << 4) |
12042 		     dtrace_dof_char(p[1]);
12043 		p += 2;
12044 	}
12045 
12046 	freeenv(p_env);
12047 
12048 	if (len < sizeof (dof_hdr_t)) {
12049 		kmem_free(buf, 0);
12050 		dtrace_dof_error(NULL, "truncated header");
12051 		return (NULL);
12052 	}
12053 
12054 	if (len < (loadsz = dof->dofh_loadsz)) {
12055 		kmem_free(buf, 0);
12056 		dtrace_dof_error(NULL, "truncated DOF");
12057 		return (NULL);
12058 	}
12059 
12060 	if (loadsz >= dtrace_dof_maxsize) {
12061 		kmem_free(buf, 0);
12062 		dtrace_dof_error(NULL, "oversized DOF");
12063 		return (NULL);
12064 	}
12065 #endif
12066 
12067 	return (dof);
12068 }
12069 
12070 static void
12071 dtrace_dof_destroy(dof_hdr_t *dof)
12072 {
12073 	kmem_free(dof, dof->dofh_loadsz);
12074 }
12075 
12076 /*
12077  * Return the dof_sec_t pointer corresponding to a given section index.  If the
12078  * index is not valid, dtrace_dof_error() is called and NULL is returned.  If
12079  * a type other than DOF_SECT_NONE is specified, the header is checked against
12080  * this type and NULL is returned if the types do not match.
12081  */
12082 static dof_sec_t *
12083 dtrace_dof_sect(dof_hdr_t *dof, uint32_t type, dof_secidx_t i)
12084 {
12085 	dof_sec_t *sec = (dof_sec_t *)(uintptr_t)
12086 	    ((uintptr_t)dof + dof->dofh_secoff + i * dof->dofh_secsize);
12087 
12088 	if (i >= dof->dofh_secnum) {
12089 		dtrace_dof_error(dof, "referenced section index is invalid");
12090 		return (NULL);
12091 	}
12092 
12093 	if (!(sec->dofs_flags & DOF_SECF_LOAD)) {
12094 		dtrace_dof_error(dof, "referenced section is not loadable");
12095 		return (NULL);
12096 	}
12097 
12098 	if (type != DOF_SECT_NONE && type != sec->dofs_type) {
12099 		dtrace_dof_error(dof, "referenced section is the wrong type");
12100 		return (NULL);
12101 	}
12102 
12103 	return (sec);
12104 }
12105 
12106 static dtrace_probedesc_t *
12107 dtrace_dof_probedesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_probedesc_t *desc)
12108 {
12109 	dof_probedesc_t *probe;
12110 	dof_sec_t *strtab;
12111 	uintptr_t daddr = (uintptr_t)dof;
12112 	uintptr_t str;
12113 	size_t size;
12114 
12115 	if (sec->dofs_type != DOF_SECT_PROBEDESC) {
12116 		dtrace_dof_error(dof, "invalid probe section");
12117 		return (NULL);
12118 	}
12119 
12120 	if (sec->dofs_align != sizeof (dof_secidx_t)) {
12121 		dtrace_dof_error(dof, "bad alignment in probe description");
12122 		return (NULL);
12123 	}
12124 
12125 	if (sec->dofs_offset + sizeof (dof_probedesc_t) > dof->dofh_loadsz) {
12126 		dtrace_dof_error(dof, "truncated probe description");
12127 		return (NULL);
12128 	}
12129 
12130 	probe = (dof_probedesc_t *)(uintptr_t)(daddr + sec->dofs_offset);
12131 	strtab = dtrace_dof_sect(dof, DOF_SECT_STRTAB, probe->dofp_strtab);
12132 
12133 	if (strtab == NULL)
12134 		return (NULL);
12135 
12136 	str = daddr + strtab->dofs_offset;
12137 	size = strtab->dofs_size;
12138 
12139 	if (probe->dofp_provider >= strtab->dofs_size) {
12140 		dtrace_dof_error(dof, "corrupt probe provider");
12141 		return (NULL);
12142 	}
12143 
12144 	(void) strncpy(desc->dtpd_provider,
12145 	    (char *)(str + probe->dofp_provider),
12146 	    MIN(DTRACE_PROVNAMELEN - 1, size - probe->dofp_provider));
12147 
12148 	if (probe->dofp_mod >= strtab->dofs_size) {
12149 		dtrace_dof_error(dof, "corrupt probe module");
12150 		return (NULL);
12151 	}
12152 
12153 	(void) strncpy(desc->dtpd_mod, (char *)(str + probe->dofp_mod),
12154 	    MIN(DTRACE_MODNAMELEN - 1, size - probe->dofp_mod));
12155 
12156 	if (probe->dofp_func >= strtab->dofs_size) {
12157 		dtrace_dof_error(dof, "corrupt probe function");
12158 		return (NULL);
12159 	}
12160 
12161 	(void) strncpy(desc->dtpd_func, (char *)(str + probe->dofp_func),
12162 	    MIN(DTRACE_FUNCNAMELEN - 1, size - probe->dofp_func));
12163 
12164 	if (probe->dofp_name >= strtab->dofs_size) {
12165 		dtrace_dof_error(dof, "corrupt probe name");
12166 		return (NULL);
12167 	}
12168 
12169 	(void) strncpy(desc->dtpd_name, (char *)(str + probe->dofp_name),
12170 	    MIN(DTRACE_NAMELEN - 1, size - probe->dofp_name));
12171 
12172 	return (desc);
12173 }
12174 
12175 static dtrace_difo_t *
12176 dtrace_dof_difo(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12177     cred_t *cr)
12178 {
12179 	dtrace_difo_t *dp;
12180 	size_t ttl = 0;
12181 	dof_difohdr_t *dofd;
12182 	uintptr_t daddr = (uintptr_t)dof;
12183 	size_t max = dtrace_difo_maxsize;
12184 	int i, l, n;
12185 
12186 	static const struct {
12187 		int section;
12188 		int bufoffs;
12189 		int lenoffs;
12190 		int entsize;
12191 		int align;
12192 		const char *msg;
12193 	} difo[] = {
12194 		{ DOF_SECT_DIF, offsetof(dtrace_difo_t, dtdo_buf),
12195 		offsetof(dtrace_difo_t, dtdo_len), sizeof (dif_instr_t),
12196 		sizeof (dif_instr_t), "multiple DIF sections" },
12197 
12198 		{ DOF_SECT_INTTAB, offsetof(dtrace_difo_t, dtdo_inttab),
12199 		offsetof(dtrace_difo_t, dtdo_intlen), sizeof (uint64_t),
12200 		sizeof (uint64_t), "multiple integer tables" },
12201 
12202 		{ DOF_SECT_STRTAB, offsetof(dtrace_difo_t, dtdo_strtab),
12203 		offsetof(dtrace_difo_t, dtdo_strlen), 0,
12204 		sizeof (char), "multiple string tables" },
12205 
12206 		{ DOF_SECT_VARTAB, offsetof(dtrace_difo_t, dtdo_vartab),
12207 		offsetof(dtrace_difo_t, dtdo_varlen), sizeof (dtrace_difv_t),
12208 		sizeof (uint_t), "multiple variable tables" },
12209 
12210 		{ DOF_SECT_NONE, 0, 0, 0, 0, NULL }
12211 	};
12212 
12213 	if (sec->dofs_type != DOF_SECT_DIFOHDR) {
12214 		dtrace_dof_error(dof, "invalid DIFO header section");
12215 		return (NULL);
12216 	}
12217 
12218 	if (sec->dofs_align != sizeof (dof_secidx_t)) {
12219 		dtrace_dof_error(dof, "bad alignment in DIFO header");
12220 		return (NULL);
12221 	}
12222 
12223 	if (sec->dofs_size < sizeof (dof_difohdr_t) ||
12224 	    sec->dofs_size % sizeof (dof_secidx_t)) {
12225 		dtrace_dof_error(dof, "bad size in DIFO header");
12226 		return (NULL);
12227 	}
12228 
12229 	dofd = (dof_difohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
12230 	n = (sec->dofs_size - sizeof (*dofd)) / sizeof (dof_secidx_t) + 1;
12231 
12232 	dp = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
12233 	dp->dtdo_rtype = dofd->dofd_rtype;
12234 
12235 	for (l = 0; l < n; l++) {
12236 		dof_sec_t *subsec;
12237 		void **bufp;
12238 		uint32_t *lenp;
12239 
12240 		if ((subsec = dtrace_dof_sect(dof, DOF_SECT_NONE,
12241 		    dofd->dofd_links[l])) == NULL)
12242 			goto err; /* invalid section link */
12243 
12244 		if (ttl + subsec->dofs_size > max) {
12245 			dtrace_dof_error(dof, "exceeds maximum size");
12246 			goto err;
12247 		}
12248 
12249 		ttl += subsec->dofs_size;
12250 
12251 		for (i = 0; difo[i].section != DOF_SECT_NONE; i++) {
12252 			if (subsec->dofs_type != difo[i].section)
12253 				continue;
12254 
12255 			if (!(subsec->dofs_flags & DOF_SECF_LOAD)) {
12256 				dtrace_dof_error(dof, "section not loaded");
12257 				goto err;
12258 			}
12259 
12260 			if (subsec->dofs_align != difo[i].align) {
12261 				dtrace_dof_error(dof, "bad alignment");
12262 				goto err;
12263 			}
12264 
12265 			bufp = (void **)((uintptr_t)dp + difo[i].bufoffs);
12266 			lenp = (uint32_t *)((uintptr_t)dp + difo[i].lenoffs);
12267 
12268 			if (*bufp != NULL) {
12269 				dtrace_dof_error(dof, difo[i].msg);
12270 				goto err;
12271 			}
12272 
12273 			if (difo[i].entsize != subsec->dofs_entsize) {
12274 				dtrace_dof_error(dof, "entry size mismatch");
12275 				goto err;
12276 			}
12277 
12278 			if (subsec->dofs_entsize != 0 &&
12279 			    (subsec->dofs_size % subsec->dofs_entsize) != 0) {
12280 				dtrace_dof_error(dof, "corrupt entry size");
12281 				goto err;
12282 			}
12283 
12284 			*lenp = subsec->dofs_size;
12285 			*bufp = kmem_alloc(subsec->dofs_size, KM_SLEEP);
12286 			bcopy((char *)(uintptr_t)(daddr + subsec->dofs_offset),
12287 			    *bufp, subsec->dofs_size);
12288 
12289 			if (subsec->dofs_entsize != 0)
12290 				*lenp /= subsec->dofs_entsize;
12291 
12292 			break;
12293 		}
12294 
12295 		/*
12296 		 * If we encounter a loadable DIFO sub-section that is not
12297 		 * known to us, assume this is a broken program and fail.
12298 		 */
12299 		if (difo[i].section == DOF_SECT_NONE &&
12300 		    (subsec->dofs_flags & DOF_SECF_LOAD)) {
12301 			dtrace_dof_error(dof, "unrecognized DIFO subsection");
12302 			goto err;
12303 		}
12304 	}
12305 
12306 	if (dp->dtdo_buf == NULL) {
12307 		/*
12308 		 * We can't have a DIF object without DIF text.
12309 		 */
12310 		dtrace_dof_error(dof, "missing DIF text");
12311 		goto err;
12312 	}
12313 
12314 	/*
12315 	 * Before we validate the DIF object, run through the variable table
12316 	 * looking for the strings -- if any of their size are under, we'll set
12317 	 * their size to be the system-wide default string size.  Note that
12318 	 * this should _not_ happen if the "strsize" option has been set --
12319 	 * in this case, the compiler should have set the size to reflect the
12320 	 * setting of the option.
12321 	 */
12322 	for (i = 0; i < dp->dtdo_varlen; i++) {
12323 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
12324 		dtrace_diftype_t *t = &v->dtdv_type;
12325 
12326 		if (v->dtdv_id < DIF_VAR_OTHER_UBASE)
12327 			continue;
12328 
12329 		if (t->dtdt_kind == DIF_TYPE_STRING && t->dtdt_size == 0)
12330 			t->dtdt_size = dtrace_strsize_default;
12331 	}
12332 
12333 	if (dtrace_difo_validate(dp, vstate, DIF_DIR_NREGS, cr) != 0)
12334 		goto err;
12335 
12336 	dtrace_difo_init(dp, vstate);
12337 	return (dp);
12338 
12339 err:
12340 	kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
12341 	kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
12342 	kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
12343 	kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
12344 
12345 	kmem_free(dp, sizeof (dtrace_difo_t));
12346 	return (NULL);
12347 }
12348 
12349 static dtrace_predicate_t *
12350 dtrace_dof_predicate(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12351     cred_t *cr)
12352 {
12353 	dtrace_difo_t *dp;
12354 
12355 	if ((dp = dtrace_dof_difo(dof, sec, vstate, cr)) == NULL)
12356 		return (NULL);
12357 
12358 	return (dtrace_predicate_create(dp));
12359 }
12360 
12361 static dtrace_actdesc_t *
12362 dtrace_dof_actdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12363     cred_t *cr)
12364 {
12365 	dtrace_actdesc_t *act, *first = NULL, *last = NULL, *next;
12366 	dof_actdesc_t *desc;
12367 	dof_sec_t *difosec;
12368 	size_t offs;
12369 	uintptr_t daddr = (uintptr_t)dof;
12370 	uint64_t arg;
12371 	dtrace_actkind_t kind;
12372 
12373 	if (sec->dofs_type != DOF_SECT_ACTDESC) {
12374 		dtrace_dof_error(dof, "invalid action section");
12375 		return (NULL);
12376 	}
12377 
12378 	if (sec->dofs_offset + sizeof (dof_actdesc_t) > dof->dofh_loadsz) {
12379 		dtrace_dof_error(dof, "truncated action description");
12380 		return (NULL);
12381 	}
12382 
12383 	if (sec->dofs_align != sizeof (uint64_t)) {
12384 		dtrace_dof_error(dof, "bad alignment in action description");
12385 		return (NULL);
12386 	}
12387 
12388 	if (sec->dofs_size < sec->dofs_entsize) {
12389 		dtrace_dof_error(dof, "section entry size exceeds total size");
12390 		return (NULL);
12391 	}
12392 
12393 	if (sec->dofs_entsize != sizeof (dof_actdesc_t)) {
12394 		dtrace_dof_error(dof, "bad entry size in action description");
12395 		return (NULL);
12396 	}
12397 
12398 	if (sec->dofs_size / sec->dofs_entsize > dtrace_actions_max) {
12399 		dtrace_dof_error(dof, "actions exceed dtrace_actions_max");
12400 		return (NULL);
12401 	}
12402 
12403 	for (offs = 0; offs < sec->dofs_size; offs += sec->dofs_entsize) {
12404 		desc = (dof_actdesc_t *)(daddr +
12405 		    (uintptr_t)sec->dofs_offset + offs);
12406 		kind = (dtrace_actkind_t)desc->dofa_kind;
12407 
12408 		if ((DTRACEACT_ISPRINTFLIKE(kind) &&
12409 		    (kind != DTRACEACT_PRINTA ||
12410 		    desc->dofa_strtab != DOF_SECIDX_NONE)) ||
12411 		    (kind == DTRACEACT_DIFEXPR &&
12412 		    desc->dofa_strtab != DOF_SECIDX_NONE)) {
12413 			dof_sec_t *strtab;
12414 			char *str, *fmt;
12415 			uint64_t i;
12416 
12417 			/*
12418 			 * The argument to these actions is an index into the
12419 			 * DOF string table.  For printf()-like actions, this
12420 			 * is the format string.  For print(), this is the
12421 			 * CTF type of the expression result.
12422 			 */
12423 			if ((strtab = dtrace_dof_sect(dof,
12424 			    DOF_SECT_STRTAB, desc->dofa_strtab)) == NULL)
12425 				goto err;
12426 
12427 			str = (char *)((uintptr_t)dof +
12428 			    (uintptr_t)strtab->dofs_offset);
12429 
12430 			for (i = desc->dofa_arg; i < strtab->dofs_size; i++) {
12431 				if (str[i] == '\0')
12432 					break;
12433 			}
12434 
12435 			if (i >= strtab->dofs_size) {
12436 				dtrace_dof_error(dof, "bogus format string");
12437 				goto err;
12438 			}
12439 
12440 			if (i == desc->dofa_arg) {
12441 				dtrace_dof_error(dof, "empty format string");
12442 				goto err;
12443 			}
12444 
12445 			i -= desc->dofa_arg;
12446 			fmt = kmem_alloc(i + 1, KM_SLEEP);
12447 			bcopy(&str[desc->dofa_arg], fmt, i + 1);
12448 			arg = (uint64_t)(uintptr_t)fmt;
12449 		} else {
12450 			if (kind == DTRACEACT_PRINTA) {
12451 				ASSERT(desc->dofa_strtab == DOF_SECIDX_NONE);
12452 				arg = 0;
12453 			} else {
12454 				arg = desc->dofa_arg;
12455 			}
12456 		}
12457 
12458 		act = dtrace_actdesc_create(kind, desc->dofa_ntuple,
12459 		    desc->dofa_uarg, arg);
12460 
12461 		if (last != NULL) {
12462 			last->dtad_next = act;
12463 		} else {
12464 			first = act;
12465 		}
12466 
12467 		last = act;
12468 
12469 		if (desc->dofa_difo == DOF_SECIDX_NONE)
12470 			continue;
12471 
12472 		if ((difosec = dtrace_dof_sect(dof,
12473 		    DOF_SECT_DIFOHDR, desc->dofa_difo)) == NULL)
12474 			goto err;
12475 
12476 		act->dtad_difo = dtrace_dof_difo(dof, difosec, vstate, cr);
12477 
12478 		if (act->dtad_difo == NULL)
12479 			goto err;
12480 	}
12481 
12482 	ASSERT(first != NULL);
12483 	return (first);
12484 
12485 err:
12486 	for (act = first; act != NULL; act = next) {
12487 		next = act->dtad_next;
12488 		dtrace_actdesc_release(act, vstate);
12489 	}
12490 
12491 	return (NULL);
12492 }
12493 
12494 static dtrace_ecbdesc_t *
12495 dtrace_dof_ecbdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12496     cred_t *cr)
12497 {
12498 	dtrace_ecbdesc_t *ep;
12499 	dof_ecbdesc_t *ecb;
12500 	dtrace_probedesc_t *desc;
12501 	dtrace_predicate_t *pred = NULL;
12502 
12503 	if (sec->dofs_size < sizeof (dof_ecbdesc_t)) {
12504 		dtrace_dof_error(dof, "truncated ECB description");
12505 		return (NULL);
12506 	}
12507 
12508 	if (sec->dofs_align != sizeof (uint64_t)) {
12509 		dtrace_dof_error(dof, "bad alignment in ECB description");
12510 		return (NULL);
12511 	}
12512 
12513 	ecb = (dof_ecbdesc_t *)((uintptr_t)dof + (uintptr_t)sec->dofs_offset);
12514 	sec = dtrace_dof_sect(dof, DOF_SECT_PROBEDESC, ecb->dofe_probes);
12515 
12516 	if (sec == NULL)
12517 		return (NULL);
12518 
12519 	ep = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
12520 	ep->dted_uarg = ecb->dofe_uarg;
12521 	desc = &ep->dted_probe;
12522 
12523 	if (dtrace_dof_probedesc(dof, sec, desc) == NULL)
12524 		goto err;
12525 
12526 	if (ecb->dofe_pred != DOF_SECIDX_NONE) {
12527 		if ((sec = dtrace_dof_sect(dof,
12528 		    DOF_SECT_DIFOHDR, ecb->dofe_pred)) == NULL)
12529 			goto err;
12530 
12531 		if ((pred = dtrace_dof_predicate(dof, sec, vstate, cr)) == NULL)
12532 			goto err;
12533 
12534 		ep->dted_pred.dtpdd_predicate = pred;
12535 	}
12536 
12537 	if (ecb->dofe_actions != DOF_SECIDX_NONE) {
12538 		if ((sec = dtrace_dof_sect(dof,
12539 		    DOF_SECT_ACTDESC, ecb->dofe_actions)) == NULL)
12540 			goto err;
12541 
12542 		ep->dted_action = dtrace_dof_actdesc(dof, sec, vstate, cr);
12543 
12544 		if (ep->dted_action == NULL)
12545 			goto err;
12546 	}
12547 
12548 	return (ep);
12549 
12550 err:
12551 	if (pred != NULL)
12552 		dtrace_predicate_release(pred, vstate);
12553 	kmem_free(ep, sizeof (dtrace_ecbdesc_t));
12554 	return (NULL);
12555 }
12556 
12557 /*
12558  * Apply the relocations from the specified 'sec' (a DOF_SECT_URELHDR) to the
12559  * specified DOF.  At present, this amounts to simply adding 'ubase' to the
12560  * site of any user SETX relocations to account for load object base address.
12561  * In the future, if we need other relocations, this function can be extended.
12562  */
12563 static int
12564 dtrace_dof_relocate(dof_hdr_t *dof, dof_sec_t *sec, uint64_t ubase)
12565 {
12566 	uintptr_t daddr = (uintptr_t)dof;
12567 	dof_relohdr_t *dofr =
12568 	    (dof_relohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
12569 	dof_sec_t *ss, *rs, *ts;
12570 	dof_relodesc_t *r;
12571 	uint_t i, n;
12572 
12573 	if (sec->dofs_size < sizeof (dof_relohdr_t) ||
12574 	    sec->dofs_align != sizeof (dof_secidx_t)) {
12575 		dtrace_dof_error(dof, "invalid relocation header");
12576 		return (-1);
12577 	}
12578 
12579 	ss = dtrace_dof_sect(dof, DOF_SECT_STRTAB, dofr->dofr_strtab);
12580 	rs = dtrace_dof_sect(dof, DOF_SECT_RELTAB, dofr->dofr_relsec);
12581 	ts = dtrace_dof_sect(dof, DOF_SECT_NONE, dofr->dofr_tgtsec);
12582 
12583 	if (ss == NULL || rs == NULL || ts == NULL)
12584 		return (-1); /* dtrace_dof_error() has been called already */
12585 
12586 	if (rs->dofs_entsize < sizeof (dof_relodesc_t) ||
12587 	    rs->dofs_align != sizeof (uint64_t)) {
12588 		dtrace_dof_error(dof, "invalid relocation section");
12589 		return (-1);
12590 	}
12591 
12592 	r = (dof_relodesc_t *)(uintptr_t)(daddr + rs->dofs_offset);
12593 	n = rs->dofs_size / rs->dofs_entsize;
12594 
12595 	for (i = 0; i < n; i++) {
12596 		uintptr_t taddr = daddr + ts->dofs_offset + r->dofr_offset;
12597 
12598 		switch (r->dofr_type) {
12599 		case DOF_RELO_NONE:
12600 			break;
12601 		case DOF_RELO_SETX:
12602 			if (r->dofr_offset >= ts->dofs_size || r->dofr_offset +
12603 			    sizeof (uint64_t) > ts->dofs_size) {
12604 				dtrace_dof_error(dof, "bad relocation offset");
12605 				return (-1);
12606 			}
12607 
12608 			if (!IS_P2ALIGNED(taddr, sizeof (uint64_t))) {
12609 				dtrace_dof_error(dof, "misaligned setx relo");
12610 				return (-1);
12611 			}
12612 
12613 			*(uint64_t *)taddr += ubase;
12614 			break;
12615 		default:
12616 			dtrace_dof_error(dof, "invalid relocation type");
12617 			return (-1);
12618 		}
12619 
12620 		r = (dof_relodesc_t *)((uintptr_t)r + rs->dofs_entsize);
12621 	}
12622 
12623 	return (0);
12624 }
12625 
12626 /*
12627  * The dof_hdr_t passed to dtrace_dof_slurp() should be a partially validated
12628  * header:  it should be at the front of a memory region that is at least
12629  * sizeof (dof_hdr_t) in size -- and then at least dof_hdr.dofh_loadsz in
12630  * size.  It need not be validated in any other way.
12631  */
12632 static int
12633 dtrace_dof_slurp(dof_hdr_t *dof, dtrace_vstate_t *vstate, cred_t *cr,
12634     dtrace_enabling_t **enabp, uint64_t ubase, int noprobes)
12635 {
12636 	uint64_t len = dof->dofh_loadsz, seclen;
12637 	uintptr_t daddr = (uintptr_t)dof;
12638 	dtrace_ecbdesc_t *ep;
12639 	dtrace_enabling_t *enab;
12640 	uint_t i;
12641 
12642 	ASSERT(MUTEX_HELD(&dtrace_lock));
12643 	ASSERT(dof->dofh_loadsz >= sizeof (dof_hdr_t));
12644 
12645 	/*
12646 	 * Check the DOF header identification bytes.  In addition to checking
12647 	 * valid settings, we also verify that unused bits/bytes are zeroed so
12648 	 * we can use them later without fear of regressing existing binaries.
12649 	 */
12650 	if (bcmp(&dof->dofh_ident[DOF_ID_MAG0],
12651 	    DOF_MAG_STRING, DOF_MAG_STRLEN) != 0) {
12652 		dtrace_dof_error(dof, "DOF magic string mismatch");
12653 		return (-1);
12654 	}
12655 
12656 	if (dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_ILP32 &&
12657 	    dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_LP64) {
12658 		dtrace_dof_error(dof, "DOF has invalid data model");
12659 		return (-1);
12660 	}
12661 
12662 	if (dof->dofh_ident[DOF_ID_ENCODING] != DOF_ENCODE_NATIVE) {
12663 		dtrace_dof_error(dof, "DOF encoding mismatch");
12664 		return (-1);
12665 	}
12666 
12667 	if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
12668 	    dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_2) {
12669 		dtrace_dof_error(dof, "DOF version mismatch");
12670 		return (-1);
12671 	}
12672 
12673 	if (dof->dofh_ident[DOF_ID_DIFVERS] != DIF_VERSION_2) {
12674 		dtrace_dof_error(dof, "DOF uses unsupported instruction set");
12675 		return (-1);
12676 	}
12677 
12678 	if (dof->dofh_ident[DOF_ID_DIFIREG] > DIF_DIR_NREGS) {
12679 		dtrace_dof_error(dof, "DOF uses too many integer registers");
12680 		return (-1);
12681 	}
12682 
12683 	if (dof->dofh_ident[DOF_ID_DIFTREG] > DIF_DTR_NREGS) {
12684 		dtrace_dof_error(dof, "DOF uses too many tuple registers");
12685 		return (-1);
12686 	}
12687 
12688 	for (i = DOF_ID_PAD; i < DOF_ID_SIZE; i++) {
12689 		if (dof->dofh_ident[i] != 0) {
12690 			dtrace_dof_error(dof, "DOF has invalid ident byte set");
12691 			return (-1);
12692 		}
12693 	}
12694 
12695 	if (dof->dofh_flags & ~DOF_FL_VALID) {
12696 		dtrace_dof_error(dof, "DOF has invalid flag bits set");
12697 		return (-1);
12698 	}
12699 
12700 	if (dof->dofh_secsize == 0) {
12701 		dtrace_dof_error(dof, "zero section header size");
12702 		return (-1);
12703 	}
12704 
12705 	/*
12706 	 * Check that the section headers don't exceed the amount of DOF
12707 	 * data.  Note that we cast the section size and number of sections
12708 	 * to uint64_t's to prevent possible overflow in the multiplication.
12709 	 */
12710 	seclen = (uint64_t)dof->dofh_secnum * (uint64_t)dof->dofh_secsize;
12711 
12712 	if (dof->dofh_secoff > len || seclen > len ||
12713 	    dof->dofh_secoff + seclen > len) {
12714 		dtrace_dof_error(dof, "truncated section headers");
12715 		return (-1);
12716 	}
12717 
12718 	if (!IS_P2ALIGNED(dof->dofh_secoff, sizeof (uint64_t))) {
12719 		dtrace_dof_error(dof, "misaligned section headers");
12720 		return (-1);
12721 	}
12722 
12723 	if (!IS_P2ALIGNED(dof->dofh_secsize, sizeof (uint64_t))) {
12724 		dtrace_dof_error(dof, "misaligned section size");
12725 		return (-1);
12726 	}
12727 
12728 	/*
12729 	 * Take an initial pass through the section headers to be sure that
12730 	 * the headers don't have stray offsets.  If the 'noprobes' flag is
12731 	 * set, do not permit sections relating to providers, probes, or args.
12732 	 */
12733 	for (i = 0; i < dof->dofh_secnum; i++) {
12734 		dof_sec_t *sec = (dof_sec_t *)(daddr +
12735 		    (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12736 
12737 		if (noprobes) {
12738 			switch (sec->dofs_type) {
12739 			case DOF_SECT_PROVIDER:
12740 			case DOF_SECT_PROBES:
12741 			case DOF_SECT_PRARGS:
12742 			case DOF_SECT_PROFFS:
12743 				dtrace_dof_error(dof, "illegal sections "
12744 				    "for enabling");
12745 				return (-1);
12746 			}
12747 		}
12748 
12749 		if (!(sec->dofs_flags & DOF_SECF_LOAD))
12750 			continue; /* just ignore non-loadable sections */
12751 
12752 		if (sec->dofs_align & (sec->dofs_align - 1)) {
12753 			dtrace_dof_error(dof, "bad section alignment");
12754 			return (-1);
12755 		}
12756 
12757 		if (sec->dofs_offset & (sec->dofs_align - 1)) {
12758 			dtrace_dof_error(dof, "misaligned section");
12759 			return (-1);
12760 		}
12761 
12762 		if (sec->dofs_offset > len || sec->dofs_size > len ||
12763 		    sec->dofs_offset + sec->dofs_size > len) {
12764 			dtrace_dof_error(dof, "corrupt section header");
12765 			return (-1);
12766 		}
12767 
12768 		if (sec->dofs_type == DOF_SECT_STRTAB && *((char *)daddr +
12769 		    sec->dofs_offset + sec->dofs_size - 1) != '\0') {
12770 			dtrace_dof_error(dof, "non-terminating string table");
12771 			return (-1);
12772 		}
12773 	}
12774 
12775 	/*
12776 	 * Take a second pass through the sections and locate and perform any
12777 	 * relocations that are present.  We do this after the first pass to
12778 	 * be sure that all sections have had their headers validated.
12779 	 */
12780 	for (i = 0; i < dof->dofh_secnum; i++) {
12781 		dof_sec_t *sec = (dof_sec_t *)(daddr +
12782 		    (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12783 
12784 		if (!(sec->dofs_flags & DOF_SECF_LOAD))
12785 			continue; /* skip sections that are not loadable */
12786 
12787 		switch (sec->dofs_type) {
12788 		case DOF_SECT_URELHDR:
12789 			if (dtrace_dof_relocate(dof, sec, ubase) != 0)
12790 				return (-1);
12791 			break;
12792 		}
12793 	}
12794 
12795 	if ((enab = *enabp) == NULL)
12796 		enab = *enabp = dtrace_enabling_create(vstate);
12797 
12798 	for (i = 0; i < dof->dofh_secnum; i++) {
12799 		dof_sec_t *sec = (dof_sec_t *)(daddr +
12800 		    (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12801 
12802 		if (sec->dofs_type != DOF_SECT_ECBDESC)
12803 			continue;
12804 
12805 		if ((ep = dtrace_dof_ecbdesc(dof, sec, vstate, cr)) == NULL) {
12806 			dtrace_enabling_destroy(enab);
12807 			*enabp = NULL;
12808 			return (-1);
12809 		}
12810 
12811 		dtrace_enabling_add(enab, ep);
12812 	}
12813 
12814 	return (0);
12815 }
12816 
12817 /*
12818  * Process DOF for any options.  This routine assumes that the DOF has been
12819  * at least processed by dtrace_dof_slurp().
12820  */
12821 static int
12822 dtrace_dof_options(dof_hdr_t *dof, dtrace_state_t *state)
12823 {
12824 	int i, rval;
12825 	uint32_t entsize;
12826 	size_t offs;
12827 	dof_optdesc_t *desc;
12828 
12829 	for (i = 0; i < dof->dofh_secnum; i++) {
12830 		dof_sec_t *sec = (dof_sec_t *)((uintptr_t)dof +
12831 		    (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12832 
12833 		if (sec->dofs_type != DOF_SECT_OPTDESC)
12834 			continue;
12835 
12836 		if (sec->dofs_align != sizeof (uint64_t)) {
12837 			dtrace_dof_error(dof, "bad alignment in "
12838 			    "option description");
12839 			return (EINVAL);
12840 		}
12841 
12842 		if ((entsize = sec->dofs_entsize) == 0) {
12843 			dtrace_dof_error(dof, "zeroed option entry size");
12844 			return (EINVAL);
12845 		}
12846 
12847 		if (entsize < sizeof (dof_optdesc_t)) {
12848 			dtrace_dof_error(dof, "bad option entry size");
12849 			return (EINVAL);
12850 		}
12851 
12852 		for (offs = 0; offs < sec->dofs_size; offs += entsize) {
12853 			desc = (dof_optdesc_t *)((uintptr_t)dof +
12854 			    (uintptr_t)sec->dofs_offset + offs);
12855 
12856 			if (desc->dofo_strtab != DOF_SECIDX_NONE) {
12857 				dtrace_dof_error(dof, "non-zero option string");
12858 				return (EINVAL);
12859 			}
12860 
12861 			if (desc->dofo_value == DTRACEOPT_UNSET) {
12862 				dtrace_dof_error(dof, "unset option");
12863 				return (EINVAL);
12864 			}
12865 
12866 			if ((rval = dtrace_state_option(state,
12867 			    desc->dofo_option, desc->dofo_value)) != 0) {
12868 				dtrace_dof_error(dof, "rejected option");
12869 				return (rval);
12870 			}
12871 		}
12872 	}
12873 
12874 	return (0);
12875 }
12876 
12877 /*
12878  * DTrace Consumer State Functions
12879  */
12880 static int
12881 dtrace_dstate_init(dtrace_dstate_t *dstate, size_t size)
12882 {
12883 	size_t hashsize, maxper, min, chunksize = dstate->dtds_chunksize;
12884 	void *base;
12885 	uintptr_t limit;
12886 	dtrace_dynvar_t *dvar, *next, *start;
12887 	int i;
12888 
12889 	ASSERT(MUTEX_HELD(&dtrace_lock));
12890 	ASSERT(dstate->dtds_base == NULL && dstate->dtds_percpu == NULL);
12891 
12892 	bzero(dstate, sizeof (dtrace_dstate_t));
12893 
12894 	if ((dstate->dtds_chunksize = chunksize) == 0)
12895 		dstate->dtds_chunksize = DTRACE_DYNVAR_CHUNKSIZE;
12896 
12897 	if (size < (min = dstate->dtds_chunksize + sizeof (dtrace_dynhash_t)))
12898 		size = min;
12899 
12900 	if ((base = kmem_zalloc(size, KM_NOSLEEP)) == NULL)
12901 		return (ENOMEM);
12902 
12903 	dstate->dtds_size = size;
12904 	dstate->dtds_base = base;
12905 	dstate->dtds_percpu = kmem_cache_alloc(dtrace_state_cache, KM_SLEEP);
12906 	bzero(dstate->dtds_percpu, NCPU * sizeof (dtrace_dstate_percpu_t));
12907 
12908 	hashsize = size / (dstate->dtds_chunksize + sizeof (dtrace_dynhash_t));
12909 
12910 	if (hashsize != 1 && (hashsize & 1))
12911 		hashsize--;
12912 
12913 	dstate->dtds_hashsize = hashsize;
12914 	dstate->dtds_hash = dstate->dtds_base;
12915 
12916 	/*
12917 	 * Set all of our hash buckets to point to the single sink, and (if
12918 	 * it hasn't already been set), set the sink's hash value to be the
12919 	 * sink sentinel value.  The sink is needed for dynamic variable
12920 	 * lookups to know that they have iterated over an entire, valid hash
12921 	 * chain.
12922 	 */
12923 	for (i = 0; i < hashsize; i++)
12924 		dstate->dtds_hash[i].dtdh_chain = &dtrace_dynhash_sink;
12925 
12926 	if (dtrace_dynhash_sink.dtdv_hashval != DTRACE_DYNHASH_SINK)
12927 		dtrace_dynhash_sink.dtdv_hashval = DTRACE_DYNHASH_SINK;
12928 
12929 	/*
12930 	 * Determine number of active CPUs.  Divide free list evenly among
12931 	 * active CPUs.
12932 	 */
12933 	start = (dtrace_dynvar_t *)
12934 	    ((uintptr_t)base + hashsize * sizeof (dtrace_dynhash_t));
12935 	limit = (uintptr_t)base + size;
12936 
12937 	maxper = (limit - (uintptr_t)start) / NCPU;
12938 	maxper = (maxper / dstate->dtds_chunksize) * dstate->dtds_chunksize;
12939 
12940 #if !defined(sun)
12941 	CPU_FOREACH(i) {
12942 #else
12943 	for (i = 0; i < NCPU; i++) {
12944 #endif
12945 		dstate->dtds_percpu[i].dtdsc_free = dvar = start;
12946 
12947 		/*
12948 		 * If we don't even have enough chunks to make it once through
12949 		 * NCPUs, we're just going to allocate everything to the first
12950 		 * CPU.  And if we're on the last CPU, we're going to allocate
12951 		 * whatever is left over.  In either case, we set the limit to
12952 		 * be the limit of the dynamic variable space.
12953 		 */
12954 		if (maxper == 0 || i == NCPU - 1) {
12955 			limit = (uintptr_t)base + size;
12956 			start = NULL;
12957 		} else {
12958 			limit = (uintptr_t)start + maxper;
12959 			start = (dtrace_dynvar_t *)limit;
12960 		}
12961 
12962 		ASSERT(limit <= (uintptr_t)base + size);
12963 
12964 		for (;;) {
12965 			next = (dtrace_dynvar_t *)((uintptr_t)dvar +
12966 			    dstate->dtds_chunksize);
12967 
12968 			if ((uintptr_t)next + dstate->dtds_chunksize >= limit)
12969 				break;
12970 
12971 			dvar->dtdv_next = next;
12972 			dvar = next;
12973 		}
12974 
12975 		if (maxper == 0)
12976 			break;
12977 	}
12978 
12979 	return (0);
12980 }
12981 
12982 static void
12983 dtrace_dstate_fini(dtrace_dstate_t *dstate)
12984 {
12985 	ASSERT(MUTEX_HELD(&cpu_lock));
12986 
12987 	if (dstate->dtds_base == NULL)
12988 		return;
12989 
12990 	kmem_free(dstate->dtds_base, dstate->dtds_size);
12991 	kmem_cache_free(dtrace_state_cache, dstate->dtds_percpu);
12992 }
12993 
12994 static void
12995 dtrace_vstate_fini(dtrace_vstate_t *vstate)
12996 {
12997 	/*
12998 	 * Logical XOR, where are you?
12999 	 */
13000 	ASSERT((vstate->dtvs_nglobals == 0) ^ (vstate->dtvs_globals != NULL));
13001 
13002 	if (vstate->dtvs_nglobals > 0) {
13003 		kmem_free(vstate->dtvs_globals, vstate->dtvs_nglobals *
13004 		    sizeof (dtrace_statvar_t *));
13005 	}
13006 
13007 	if (vstate->dtvs_ntlocals > 0) {
13008 		kmem_free(vstate->dtvs_tlocals, vstate->dtvs_ntlocals *
13009 		    sizeof (dtrace_difv_t));
13010 	}
13011 
13012 	ASSERT((vstate->dtvs_nlocals == 0) ^ (vstate->dtvs_locals != NULL));
13013 
13014 	if (vstate->dtvs_nlocals > 0) {
13015 		kmem_free(vstate->dtvs_locals, vstate->dtvs_nlocals *
13016 		    sizeof (dtrace_statvar_t *));
13017 	}
13018 }
13019 
13020 #if defined(sun)
13021 static void
13022 dtrace_state_clean(dtrace_state_t *state)
13023 {
13024 	if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE)
13025 		return;
13026 
13027 	dtrace_dynvar_clean(&state->dts_vstate.dtvs_dynvars);
13028 	dtrace_speculation_clean(state);
13029 }
13030 
13031 static void
13032 dtrace_state_deadman(dtrace_state_t *state)
13033 {
13034 	hrtime_t now;
13035 
13036 	dtrace_sync();
13037 
13038 	now = dtrace_gethrtime();
13039 
13040 	if (state != dtrace_anon.dta_state &&
13041 	    now - state->dts_laststatus >= dtrace_deadman_user)
13042 		return;
13043 
13044 	/*
13045 	 * We must be sure that dts_alive never appears to be less than the
13046 	 * value upon entry to dtrace_state_deadman(), and because we lack a
13047 	 * dtrace_cas64(), we cannot store to it atomically.  We thus instead
13048 	 * store INT64_MAX to it, followed by a memory barrier, followed by
13049 	 * the new value.  This assures that dts_alive never appears to be
13050 	 * less than its true value, regardless of the order in which the
13051 	 * stores to the underlying storage are issued.
13052 	 */
13053 	state->dts_alive = INT64_MAX;
13054 	dtrace_membar_producer();
13055 	state->dts_alive = now;
13056 }
13057 #else
13058 static void
13059 dtrace_state_clean(void *arg)
13060 {
13061 	dtrace_state_t *state = arg;
13062 	dtrace_optval_t *opt = state->dts_options;
13063 
13064 	if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE)
13065 		return;
13066 
13067 	dtrace_dynvar_clean(&state->dts_vstate.dtvs_dynvars);
13068 	dtrace_speculation_clean(state);
13069 
13070 	callout_reset(&state->dts_cleaner, hz * opt[DTRACEOPT_CLEANRATE] / NANOSEC,
13071 	    dtrace_state_clean, state);
13072 }
13073 
13074 static void
13075 dtrace_state_deadman(void *arg)
13076 {
13077 	dtrace_state_t *state = arg;
13078 	hrtime_t now;
13079 
13080 	dtrace_sync();
13081 
13082 	dtrace_debug_output();
13083 
13084 	now = dtrace_gethrtime();
13085 
13086 	if (state != dtrace_anon.dta_state &&
13087 	    now - state->dts_laststatus >= dtrace_deadman_user)
13088 		return;
13089 
13090 	/*
13091 	 * We must be sure that dts_alive never appears to be less than the
13092 	 * value upon entry to dtrace_state_deadman(), and because we lack a
13093 	 * dtrace_cas64(), we cannot store to it atomically.  We thus instead
13094 	 * store INT64_MAX to it, followed by a memory barrier, followed by
13095 	 * the new value.  This assures that dts_alive never appears to be
13096 	 * less than its true value, regardless of the order in which the
13097 	 * stores to the underlying storage are issued.
13098 	 */
13099 	state->dts_alive = INT64_MAX;
13100 	dtrace_membar_producer();
13101 	state->dts_alive = now;
13102 
13103 	callout_reset(&state->dts_deadman, hz * dtrace_deadman_interval / NANOSEC,
13104 	    dtrace_state_deadman, state);
13105 }
13106 #endif
13107 
13108 static dtrace_state_t *
13109 #if defined(sun)
13110 dtrace_state_create(dev_t *devp, cred_t *cr)
13111 #else
13112 dtrace_state_create(struct cdev *dev)
13113 #endif
13114 {
13115 #if defined(sun)
13116 	minor_t minor;
13117 	major_t major;
13118 #else
13119 	cred_t *cr = NULL;
13120 	int m = 0;
13121 #endif
13122 	char c[30];
13123 	dtrace_state_t *state;
13124 	dtrace_optval_t *opt;
13125 	int bufsize = NCPU * sizeof (dtrace_buffer_t), i;
13126 
13127 	ASSERT(MUTEX_HELD(&dtrace_lock));
13128 	ASSERT(MUTEX_HELD(&cpu_lock));
13129 
13130 #if defined(sun)
13131 	minor = (minor_t)(uintptr_t)vmem_alloc(dtrace_minor, 1,
13132 	    VM_BESTFIT | VM_SLEEP);
13133 
13134 	if (ddi_soft_state_zalloc(dtrace_softstate, minor) != DDI_SUCCESS) {
13135 		vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
13136 		return (NULL);
13137 	}
13138 
13139 	state = ddi_get_soft_state(dtrace_softstate, minor);
13140 #else
13141 	if (dev != NULL) {
13142 		cr = dev->si_cred;
13143 		m = dev2unit(dev);
13144 		}
13145 
13146 	/* Allocate memory for the state. */
13147 	state = kmem_zalloc(sizeof(dtrace_state_t), KM_SLEEP);
13148 #endif
13149 
13150 	state->dts_epid = DTRACE_EPIDNONE + 1;
13151 
13152 	(void) snprintf(c, sizeof (c), "dtrace_aggid_%d", m);
13153 #if defined(sun)
13154 	state->dts_aggid_arena = vmem_create(c, (void *)1, UINT32_MAX, 1,
13155 	    NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
13156 
13157 	if (devp != NULL) {
13158 		major = getemajor(*devp);
13159 	} else {
13160 		major = ddi_driver_major(dtrace_devi);
13161 	}
13162 
13163 	state->dts_dev = makedevice(major, minor);
13164 
13165 	if (devp != NULL)
13166 		*devp = state->dts_dev;
13167 #else
13168 	state->dts_aggid_arena = new_unrhdr(1, INT_MAX, &dtrace_unr_mtx);
13169 	state->dts_dev = dev;
13170 #endif
13171 
13172 	/*
13173 	 * We allocate NCPU buffers.  On the one hand, this can be quite
13174 	 * a bit of memory per instance (nearly 36K on a Starcat).  On the
13175 	 * other hand, it saves an additional memory reference in the probe
13176 	 * path.
13177 	 */
13178 	state->dts_buffer = kmem_zalloc(bufsize, KM_SLEEP);
13179 	state->dts_aggbuffer = kmem_zalloc(bufsize, KM_SLEEP);
13180 
13181 #if defined(sun)
13182 	state->dts_cleaner = CYCLIC_NONE;
13183 	state->dts_deadman = CYCLIC_NONE;
13184 #else
13185 	callout_init(&state->dts_cleaner, CALLOUT_MPSAFE);
13186 	callout_init(&state->dts_deadman, CALLOUT_MPSAFE);
13187 #endif
13188 	state->dts_vstate.dtvs_state = state;
13189 
13190 	for (i = 0; i < DTRACEOPT_MAX; i++)
13191 		state->dts_options[i] = DTRACEOPT_UNSET;
13192 
13193 	/*
13194 	 * Set the default options.
13195 	 */
13196 	opt = state->dts_options;
13197 	opt[DTRACEOPT_BUFPOLICY] = DTRACEOPT_BUFPOLICY_SWITCH;
13198 	opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_AUTO;
13199 	opt[DTRACEOPT_NSPEC] = dtrace_nspec_default;
13200 	opt[DTRACEOPT_SPECSIZE] = dtrace_specsize_default;
13201 	opt[DTRACEOPT_CPU] = (dtrace_optval_t)DTRACE_CPUALL;
13202 	opt[DTRACEOPT_STRSIZE] = dtrace_strsize_default;
13203 	opt[DTRACEOPT_STACKFRAMES] = dtrace_stackframes_default;
13204 	opt[DTRACEOPT_USTACKFRAMES] = dtrace_ustackframes_default;
13205 	opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_default;
13206 	opt[DTRACEOPT_AGGRATE] = dtrace_aggrate_default;
13207 	opt[DTRACEOPT_SWITCHRATE] = dtrace_switchrate_default;
13208 	opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_default;
13209 	opt[DTRACEOPT_JSTACKFRAMES] = dtrace_jstackframes_default;
13210 	opt[DTRACEOPT_JSTACKSTRSIZE] = dtrace_jstackstrsize_default;
13211 
13212 	state->dts_activity = DTRACE_ACTIVITY_INACTIVE;
13213 
13214 	/*
13215 	 * Depending on the user credentials, we set flag bits which alter probe
13216 	 * visibility or the amount of destructiveness allowed.  In the case of
13217 	 * actual anonymous tracing, or the possession of all privileges, all of
13218 	 * the normal checks are bypassed.
13219 	 */
13220 	if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
13221 		state->dts_cred.dcr_visible = DTRACE_CRV_ALL;
13222 		state->dts_cred.dcr_action = DTRACE_CRA_ALL;
13223 	} else {
13224 		/*
13225 		 * Set up the credentials for this instantiation.  We take a
13226 		 * hold on the credential to prevent it from disappearing on
13227 		 * us; this in turn prevents the zone_t referenced by this
13228 		 * credential from disappearing.  This means that we can
13229 		 * examine the credential and the zone from probe context.
13230 		 */
13231 		crhold(cr);
13232 		state->dts_cred.dcr_cred = cr;
13233 
13234 		/*
13235 		 * CRA_PROC means "we have *some* privilege for dtrace" and
13236 		 * unlocks the use of variables like pid, zonename, etc.
13237 		 */
13238 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE) ||
13239 		    PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
13240 			state->dts_cred.dcr_action |= DTRACE_CRA_PROC;
13241 		}
13242 
13243 		/*
13244 		 * dtrace_user allows use of syscall and profile providers.
13245 		 * If the user also has proc_owner and/or proc_zone, we
13246 		 * extend the scope to include additional visibility and
13247 		 * destructive power.
13248 		 */
13249 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE)) {
13250 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE)) {
13251 				state->dts_cred.dcr_visible |=
13252 				    DTRACE_CRV_ALLPROC;
13253 
13254 				state->dts_cred.dcr_action |=
13255 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
13256 			}
13257 
13258 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE)) {
13259 				state->dts_cred.dcr_visible |=
13260 				    DTRACE_CRV_ALLZONE;
13261 
13262 				state->dts_cred.dcr_action |=
13263 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
13264 			}
13265 
13266 			/*
13267 			 * If we have all privs in whatever zone this is,
13268 			 * we can do destructive things to processes which
13269 			 * have altered credentials.
13270 			 */
13271 #if defined(sun)
13272 			if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
13273 			    cr->cr_zone->zone_privset)) {
13274 				state->dts_cred.dcr_action |=
13275 				    DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
13276 			}
13277 #endif
13278 		}
13279 
13280 		/*
13281 		 * Holding the dtrace_kernel privilege also implies that
13282 		 * the user has the dtrace_user privilege from a visibility
13283 		 * perspective.  But without further privileges, some
13284 		 * destructive actions are not available.
13285 		 */
13286 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE)) {
13287 			/*
13288 			 * Make all probes in all zones visible.  However,
13289 			 * this doesn't mean that all actions become available
13290 			 * to all zones.
13291 			 */
13292 			state->dts_cred.dcr_visible |= DTRACE_CRV_KERNEL |
13293 			    DTRACE_CRV_ALLPROC | DTRACE_CRV_ALLZONE;
13294 
13295 			state->dts_cred.dcr_action |= DTRACE_CRA_KERNEL |
13296 			    DTRACE_CRA_PROC;
13297 			/*
13298 			 * Holding proc_owner means that destructive actions
13299 			 * for *this* zone are allowed.
13300 			 */
13301 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
13302 				state->dts_cred.dcr_action |=
13303 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
13304 
13305 			/*
13306 			 * Holding proc_zone means that destructive actions
13307 			 * for this user/group ID in all zones is allowed.
13308 			 */
13309 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
13310 				state->dts_cred.dcr_action |=
13311 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
13312 
13313 #if defined(sun)
13314 			/*
13315 			 * If we have all privs in whatever zone this is,
13316 			 * we can do destructive things to processes which
13317 			 * have altered credentials.
13318 			 */
13319 			if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
13320 			    cr->cr_zone->zone_privset)) {
13321 				state->dts_cred.dcr_action |=
13322 				    DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
13323 			}
13324 #endif
13325 		}
13326 
13327 		/*
13328 		 * Holding the dtrace_proc privilege gives control over fasttrap
13329 		 * and pid providers.  We need to grant wider destructive
13330 		 * privileges in the event that the user has proc_owner and/or
13331 		 * proc_zone.
13332 		 */
13333 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
13334 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
13335 				state->dts_cred.dcr_action |=
13336 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
13337 
13338 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
13339 				state->dts_cred.dcr_action |=
13340 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
13341 		}
13342 	}
13343 
13344 	return (state);
13345 }
13346 
13347 static int
13348 dtrace_state_buffer(dtrace_state_t *state, dtrace_buffer_t *buf, int which)
13349 {
13350 	dtrace_optval_t *opt = state->dts_options, size;
13351 	processorid_t cpu = 0;;
13352 	int flags = 0, rval;
13353 
13354 	ASSERT(MUTEX_HELD(&dtrace_lock));
13355 	ASSERT(MUTEX_HELD(&cpu_lock));
13356 	ASSERT(which < DTRACEOPT_MAX);
13357 	ASSERT(state->dts_activity == DTRACE_ACTIVITY_INACTIVE ||
13358 	    (state == dtrace_anon.dta_state &&
13359 	    state->dts_activity == DTRACE_ACTIVITY_ACTIVE));
13360 
13361 	if (opt[which] == DTRACEOPT_UNSET || opt[which] == 0)
13362 		return (0);
13363 
13364 	if (opt[DTRACEOPT_CPU] != DTRACEOPT_UNSET)
13365 		cpu = opt[DTRACEOPT_CPU];
13366 
13367 	if (which == DTRACEOPT_SPECSIZE)
13368 		flags |= DTRACEBUF_NOSWITCH;
13369 
13370 	if (which == DTRACEOPT_BUFSIZE) {
13371 		if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_RING)
13372 			flags |= DTRACEBUF_RING;
13373 
13374 		if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_FILL)
13375 			flags |= DTRACEBUF_FILL;
13376 
13377 		if (state != dtrace_anon.dta_state ||
13378 		    state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
13379 			flags |= DTRACEBUF_INACTIVE;
13380 	}
13381 
13382 	for (size = opt[which]; size >= sizeof (uint64_t); size >>= 1) {
13383 		/*
13384 		 * The size must be 8-byte aligned.  If the size is not 8-byte
13385 		 * aligned, drop it down by the difference.
13386 		 */
13387 		if (size & (sizeof (uint64_t) - 1))
13388 			size -= size & (sizeof (uint64_t) - 1);
13389 
13390 		if (size < state->dts_reserve) {
13391 			/*
13392 			 * Buffers always must be large enough to accommodate
13393 			 * their prereserved space.  We return E2BIG instead
13394 			 * of ENOMEM in this case to allow for user-level
13395 			 * software to differentiate the cases.
13396 			 */
13397 			return (E2BIG);
13398 		}
13399 
13400 		rval = dtrace_buffer_alloc(buf, size, flags, cpu);
13401 
13402 		if (rval != ENOMEM) {
13403 			opt[which] = size;
13404 			return (rval);
13405 		}
13406 
13407 		if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
13408 			return (rval);
13409 	}
13410 
13411 	return (ENOMEM);
13412 }
13413 
13414 static int
13415 dtrace_state_buffers(dtrace_state_t *state)
13416 {
13417 	dtrace_speculation_t *spec = state->dts_speculations;
13418 	int rval, i;
13419 
13420 	if ((rval = dtrace_state_buffer(state, state->dts_buffer,
13421 	    DTRACEOPT_BUFSIZE)) != 0)
13422 		return (rval);
13423 
13424 	if ((rval = dtrace_state_buffer(state, state->dts_aggbuffer,
13425 	    DTRACEOPT_AGGSIZE)) != 0)
13426 		return (rval);
13427 
13428 	for (i = 0; i < state->dts_nspeculations; i++) {
13429 		if ((rval = dtrace_state_buffer(state,
13430 		    spec[i].dtsp_buffer, DTRACEOPT_SPECSIZE)) != 0)
13431 			return (rval);
13432 	}
13433 
13434 	return (0);
13435 }
13436 
13437 static void
13438 dtrace_state_prereserve(dtrace_state_t *state)
13439 {
13440 	dtrace_ecb_t *ecb;
13441 	dtrace_probe_t *probe;
13442 
13443 	state->dts_reserve = 0;
13444 
13445 	if (state->dts_options[DTRACEOPT_BUFPOLICY] != DTRACEOPT_BUFPOLICY_FILL)
13446 		return;
13447 
13448 	/*
13449 	 * If our buffer policy is a "fill" buffer policy, we need to set the
13450 	 * prereserved space to be the space required by the END probes.
13451 	 */
13452 	probe = dtrace_probes[dtrace_probeid_end - 1];
13453 	ASSERT(probe != NULL);
13454 
13455 	for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
13456 		if (ecb->dte_state != state)
13457 			continue;
13458 
13459 		state->dts_reserve += ecb->dte_needed + ecb->dte_alignment;
13460 	}
13461 }
13462 
13463 static int
13464 dtrace_state_go(dtrace_state_t *state, processorid_t *cpu)
13465 {
13466 	dtrace_optval_t *opt = state->dts_options, sz, nspec;
13467 	dtrace_speculation_t *spec;
13468 	dtrace_buffer_t *buf;
13469 #if defined(sun)
13470 	cyc_handler_t hdlr;
13471 	cyc_time_t when;
13472 #endif
13473 	int rval = 0, i, bufsize = NCPU * sizeof (dtrace_buffer_t);
13474 	dtrace_icookie_t cookie;
13475 
13476 	mutex_enter(&cpu_lock);
13477 	mutex_enter(&dtrace_lock);
13478 
13479 	if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
13480 		rval = EBUSY;
13481 		goto out;
13482 	}
13483 
13484 	/*
13485 	 * Before we can perform any checks, we must prime all of the
13486 	 * retained enablings that correspond to this state.
13487 	 */
13488 	dtrace_enabling_prime(state);
13489 
13490 	if (state->dts_destructive && !state->dts_cred.dcr_destructive) {
13491 		rval = EACCES;
13492 		goto out;
13493 	}
13494 
13495 	dtrace_state_prereserve(state);
13496 
13497 	/*
13498 	 * Now we want to do is try to allocate our speculations.
13499 	 * We do not automatically resize the number of speculations; if
13500 	 * this fails, we will fail the operation.
13501 	 */
13502 	nspec = opt[DTRACEOPT_NSPEC];
13503 	ASSERT(nspec != DTRACEOPT_UNSET);
13504 
13505 	if (nspec > INT_MAX) {
13506 		rval = ENOMEM;
13507 		goto out;
13508 	}
13509 
13510 	spec = kmem_zalloc(nspec * sizeof (dtrace_speculation_t), KM_NOSLEEP);
13511 
13512 	if (spec == NULL) {
13513 		rval = ENOMEM;
13514 		goto out;
13515 	}
13516 
13517 	state->dts_speculations = spec;
13518 	state->dts_nspeculations = (int)nspec;
13519 
13520 	for (i = 0; i < nspec; i++) {
13521 		if ((buf = kmem_zalloc(bufsize, KM_NOSLEEP)) == NULL) {
13522 			rval = ENOMEM;
13523 			goto err;
13524 		}
13525 
13526 		spec[i].dtsp_buffer = buf;
13527 	}
13528 
13529 	if (opt[DTRACEOPT_GRABANON] != DTRACEOPT_UNSET) {
13530 		if (dtrace_anon.dta_state == NULL) {
13531 			rval = ENOENT;
13532 			goto out;
13533 		}
13534 
13535 		if (state->dts_necbs != 0) {
13536 			rval = EALREADY;
13537 			goto out;
13538 		}
13539 
13540 		state->dts_anon = dtrace_anon_grab();
13541 		ASSERT(state->dts_anon != NULL);
13542 		state = state->dts_anon;
13543 
13544 		/*
13545 		 * We want "grabanon" to be set in the grabbed state, so we'll
13546 		 * copy that option value from the grabbing state into the
13547 		 * grabbed state.
13548 		 */
13549 		state->dts_options[DTRACEOPT_GRABANON] =
13550 		    opt[DTRACEOPT_GRABANON];
13551 
13552 		*cpu = dtrace_anon.dta_beganon;
13553 
13554 		/*
13555 		 * If the anonymous state is active (as it almost certainly
13556 		 * is if the anonymous enabling ultimately matched anything),
13557 		 * we don't allow any further option processing -- but we
13558 		 * don't return failure.
13559 		 */
13560 		if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
13561 			goto out;
13562 	}
13563 
13564 	if (opt[DTRACEOPT_AGGSIZE] != DTRACEOPT_UNSET &&
13565 	    opt[DTRACEOPT_AGGSIZE] != 0) {
13566 		if (state->dts_aggregations == NULL) {
13567 			/*
13568 			 * We're not going to create an aggregation buffer
13569 			 * because we don't have any ECBs that contain
13570 			 * aggregations -- set this option to 0.
13571 			 */
13572 			opt[DTRACEOPT_AGGSIZE] = 0;
13573 		} else {
13574 			/*
13575 			 * If we have an aggregation buffer, we must also have
13576 			 * a buffer to use as scratch.
13577 			 */
13578 			if (opt[DTRACEOPT_BUFSIZE] == DTRACEOPT_UNSET ||
13579 			    opt[DTRACEOPT_BUFSIZE] < state->dts_needed) {
13580 				opt[DTRACEOPT_BUFSIZE] = state->dts_needed;
13581 			}
13582 		}
13583 	}
13584 
13585 	if (opt[DTRACEOPT_SPECSIZE] != DTRACEOPT_UNSET &&
13586 	    opt[DTRACEOPT_SPECSIZE] != 0) {
13587 		if (!state->dts_speculates) {
13588 			/*
13589 			 * We're not going to create speculation buffers
13590 			 * because we don't have any ECBs that actually
13591 			 * speculate -- set the speculation size to 0.
13592 			 */
13593 			opt[DTRACEOPT_SPECSIZE] = 0;
13594 		}
13595 	}
13596 
13597 	/*
13598 	 * The bare minimum size for any buffer that we're actually going to
13599 	 * do anything to is sizeof (uint64_t).
13600 	 */
13601 	sz = sizeof (uint64_t);
13602 
13603 	if ((state->dts_needed != 0 && opt[DTRACEOPT_BUFSIZE] < sz) ||
13604 	    (state->dts_speculates && opt[DTRACEOPT_SPECSIZE] < sz) ||
13605 	    (state->dts_aggregations != NULL && opt[DTRACEOPT_AGGSIZE] < sz)) {
13606 		/*
13607 		 * A buffer size has been explicitly set to 0 (or to a size
13608 		 * that will be adjusted to 0) and we need the space -- we
13609 		 * need to return failure.  We return ENOSPC to differentiate
13610 		 * it from failing to allocate a buffer due to failure to meet
13611 		 * the reserve (for which we return E2BIG).
13612 		 */
13613 		rval = ENOSPC;
13614 		goto out;
13615 	}
13616 
13617 	if ((rval = dtrace_state_buffers(state)) != 0)
13618 		goto err;
13619 
13620 	if ((sz = opt[DTRACEOPT_DYNVARSIZE]) == DTRACEOPT_UNSET)
13621 		sz = dtrace_dstate_defsize;
13622 
13623 	do {
13624 		rval = dtrace_dstate_init(&state->dts_vstate.dtvs_dynvars, sz);
13625 
13626 		if (rval == 0)
13627 			break;
13628 
13629 		if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
13630 			goto err;
13631 	} while (sz >>= 1);
13632 
13633 	opt[DTRACEOPT_DYNVARSIZE] = sz;
13634 
13635 	if (rval != 0)
13636 		goto err;
13637 
13638 	if (opt[DTRACEOPT_STATUSRATE] > dtrace_statusrate_max)
13639 		opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_max;
13640 
13641 	if (opt[DTRACEOPT_CLEANRATE] == 0)
13642 		opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
13643 
13644 	if (opt[DTRACEOPT_CLEANRATE] < dtrace_cleanrate_min)
13645 		opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_min;
13646 
13647 	if (opt[DTRACEOPT_CLEANRATE] > dtrace_cleanrate_max)
13648 		opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
13649 
13650 	state->dts_alive = state->dts_laststatus = dtrace_gethrtime();
13651 #if defined(sun)
13652 	hdlr.cyh_func = (cyc_func_t)dtrace_state_clean;
13653 	hdlr.cyh_arg = state;
13654 	hdlr.cyh_level = CY_LOW_LEVEL;
13655 
13656 	when.cyt_when = 0;
13657 	when.cyt_interval = opt[DTRACEOPT_CLEANRATE];
13658 
13659 	state->dts_cleaner = cyclic_add(&hdlr, &when);
13660 
13661 	hdlr.cyh_func = (cyc_func_t)dtrace_state_deadman;
13662 	hdlr.cyh_arg = state;
13663 	hdlr.cyh_level = CY_LOW_LEVEL;
13664 
13665 	when.cyt_when = 0;
13666 	when.cyt_interval = dtrace_deadman_interval;
13667 
13668 	state->dts_deadman = cyclic_add(&hdlr, &when);
13669 #else
13670 	callout_reset(&state->dts_cleaner, hz * opt[DTRACEOPT_CLEANRATE] / NANOSEC,
13671 	    dtrace_state_clean, state);
13672 	callout_reset(&state->dts_deadman, hz * dtrace_deadman_interval / NANOSEC,
13673 	    dtrace_state_deadman, state);
13674 #endif
13675 
13676 	state->dts_activity = DTRACE_ACTIVITY_WARMUP;
13677 
13678 	/*
13679 	 * Now it's time to actually fire the BEGIN probe.  We need to disable
13680 	 * interrupts here both to record the CPU on which we fired the BEGIN
13681 	 * probe (the data from this CPU will be processed first at user
13682 	 * level) and to manually activate the buffer for this CPU.
13683 	 */
13684 	cookie = dtrace_interrupt_disable();
13685 	*cpu = curcpu;
13686 	ASSERT(state->dts_buffer[*cpu].dtb_flags & DTRACEBUF_INACTIVE);
13687 	state->dts_buffer[*cpu].dtb_flags &= ~DTRACEBUF_INACTIVE;
13688 
13689 	dtrace_probe(dtrace_probeid_begin,
13690 	    (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
13691 	dtrace_interrupt_enable(cookie);
13692 	/*
13693 	 * We may have had an exit action from a BEGIN probe; only change our
13694 	 * state to ACTIVE if we're still in WARMUP.
13695 	 */
13696 	ASSERT(state->dts_activity == DTRACE_ACTIVITY_WARMUP ||
13697 	    state->dts_activity == DTRACE_ACTIVITY_DRAINING);
13698 
13699 	if (state->dts_activity == DTRACE_ACTIVITY_WARMUP)
13700 		state->dts_activity = DTRACE_ACTIVITY_ACTIVE;
13701 
13702 	/*
13703 	 * Regardless of whether or not now we're in ACTIVE or DRAINING, we
13704 	 * want each CPU to transition its principal buffer out of the
13705 	 * INACTIVE state.  Doing this assures that no CPU will suddenly begin
13706 	 * processing an ECB halfway down a probe's ECB chain; all CPUs will
13707 	 * atomically transition from processing none of a state's ECBs to
13708 	 * processing all of them.
13709 	 */
13710 	dtrace_xcall(DTRACE_CPUALL,
13711 	    (dtrace_xcall_t)dtrace_buffer_activate, state);
13712 	goto out;
13713 
13714 err:
13715 	dtrace_buffer_free(state->dts_buffer);
13716 	dtrace_buffer_free(state->dts_aggbuffer);
13717 
13718 	if ((nspec = state->dts_nspeculations) == 0) {
13719 		ASSERT(state->dts_speculations == NULL);
13720 		goto out;
13721 	}
13722 
13723 	spec = state->dts_speculations;
13724 	ASSERT(spec != NULL);
13725 
13726 	for (i = 0; i < state->dts_nspeculations; i++) {
13727 		if ((buf = spec[i].dtsp_buffer) == NULL)
13728 			break;
13729 
13730 		dtrace_buffer_free(buf);
13731 		kmem_free(buf, bufsize);
13732 	}
13733 
13734 	kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
13735 	state->dts_nspeculations = 0;
13736 	state->dts_speculations = NULL;
13737 
13738 out:
13739 	mutex_exit(&dtrace_lock);
13740 	mutex_exit(&cpu_lock);
13741 
13742 	return (rval);
13743 }
13744 
13745 static int
13746 dtrace_state_stop(dtrace_state_t *state, processorid_t *cpu)
13747 {
13748 	dtrace_icookie_t cookie;
13749 
13750 	ASSERT(MUTEX_HELD(&dtrace_lock));
13751 
13752 	if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE &&
13753 	    state->dts_activity != DTRACE_ACTIVITY_DRAINING)
13754 		return (EINVAL);
13755 
13756 	/*
13757 	 * We'll set the activity to DTRACE_ACTIVITY_DRAINING, and issue a sync
13758 	 * to be sure that every CPU has seen it.  See below for the details
13759 	 * on why this is done.
13760 	 */
13761 	state->dts_activity = DTRACE_ACTIVITY_DRAINING;
13762 	dtrace_sync();
13763 
13764 	/*
13765 	 * By this point, it is impossible for any CPU to be still processing
13766 	 * with DTRACE_ACTIVITY_ACTIVE.  We can thus set our activity to
13767 	 * DTRACE_ACTIVITY_COOLDOWN and know that we're not racing with any
13768 	 * other CPU in dtrace_buffer_reserve().  This allows dtrace_probe()
13769 	 * and callees to know that the activity is DTRACE_ACTIVITY_COOLDOWN
13770 	 * iff we're in the END probe.
13771 	 */
13772 	state->dts_activity = DTRACE_ACTIVITY_COOLDOWN;
13773 	dtrace_sync();
13774 	ASSERT(state->dts_activity == DTRACE_ACTIVITY_COOLDOWN);
13775 
13776 	/*
13777 	 * Finally, we can release the reserve and call the END probe.  We
13778 	 * disable interrupts across calling the END probe to allow us to
13779 	 * return the CPU on which we actually called the END probe.  This
13780 	 * allows user-land to be sure that this CPU's principal buffer is
13781 	 * processed last.
13782 	 */
13783 	state->dts_reserve = 0;
13784 
13785 	cookie = dtrace_interrupt_disable();
13786 	*cpu = curcpu;
13787 	dtrace_probe(dtrace_probeid_end,
13788 	    (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
13789 	dtrace_interrupt_enable(cookie);
13790 
13791 	state->dts_activity = DTRACE_ACTIVITY_STOPPED;
13792 	dtrace_sync();
13793 
13794 	return (0);
13795 }
13796 
13797 static int
13798 dtrace_state_option(dtrace_state_t *state, dtrace_optid_t option,
13799     dtrace_optval_t val)
13800 {
13801 	ASSERT(MUTEX_HELD(&dtrace_lock));
13802 
13803 	if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
13804 		return (EBUSY);
13805 
13806 	if (option >= DTRACEOPT_MAX)
13807 		return (EINVAL);
13808 
13809 	if (option != DTRACEOPT_CPU && val < 0)
13810 		return (EINVAL);
13811 
13812 	switch (option) {
13813 	case DTRACEOPT_DESTRUCTIVE:
13814 		if (dtrace_destructive_disallow)
13815 			return (EACCES);
13816 
13817 		state->dts_cred.dcr_destructive = 1;
13818 		break;
13819 
13820 	case DTRACEOPT_BUFSIZE:
13821 	case DTRACEOPT_DYNVARSIZE:
13822 	case DTRACEOPT_AGGSIZE:
13823 	case DTRACEOPT_SPECSIZE:
13824 	case DTRACEOPT_STRSIZE:
13825 		if (val < 0)
13826 			return (EINVAL);
13827 
13828 		if (val >= LONG_MAX) {
13829 			/*
13830 			 * If this is an otherwise negative value, set it to
13831 			 * the highest multiple of 128m less than LONG_MAX.
13832 			 * Technically, we're adjusting the size without
13833 			 * regard to the buffer resizing policy, but in fact,
13834 			 * this has no effect -- if we set the buffer size to
13835 			 * ~LONG_MAX and the buffer policy is ultimately set to
13836 			 * be "manual", the buffer allocation is guaranteed to
13837 			 * fail, if only because the allocation requires two
13838 			 * buffers.  (We set the the size to the highest
13839 			 * multiple of 128m because it ensures that the size
13840 			 * will remain a multiple of a megabyte when
13841 			 * repeatedly halved -- all the way down to 15m.)
13842 			 */
13843 			val = LONG_MAX - (1 << 27) + 1;
13844 		}
13845 	}
13846 
13847 	state->dts_options[option] = val;
13848 
13849 	return (0);
13850 }
13851 
13852 static void
13853 dtrace_state_destroy(dtrace_state_t *state)
13854 {
13855 	dtrace_ecb_t *ecb;
13856 	dtrace_vstate_t *vstate = &state->dts_vstate;
13857 #if defined(sun)
13858 	minor_t minor = getminor(state->dts_dev);
13859 #endif
13860 	int i, bufsize = NCPU * sizeof (dtrace_buffer_t);
13861 	dtrace_speculation_t *spec = state->dts_speculations;
13862 	int nspec = state->dts_nspeculations;
13863 	uint32_t match;
13864 
13865 	ASSERT(MUTEX_HELD(&dtrace_lock));
13866 	ASSERT(MUTEX_HELD(&cpu_lock));
13867 
13868 	/*
13869 	 * First, retract any retained enablings for this state.
13870 	 */
13871 	dtrace_enabling_retract(state);
13872 	ASSERT(state->dts_nretained == 0);
13873 
13874 	if (state->dts_activity == DTRACE_ACTIVITY_ACTIVE ||
13875 	    state->dts_activity == DTRACE_ACTIVITY_DRAINING) {
13876 		/*
13877 		 * We have managed to come into dtrace_state_destroy() on a
13878 		 * hot enabling -- almost certainly because of a disorderly
13879 		 * shutdown of a consumer.  (That is, a consumer that is
13880 		 * exiting without having called dtrace_stop().) In this case,
13881 		 * we're going to set our activity to be KILLED, and then
13882 		 * issue a sync to be sure that everyone is out of probe
13883 		 * context before we start blowing away ECBs.
13884 		 */
13885 		state->dts_activity = DTRACE_ACTIVITY_KILLED;
13886 		dtrace_sync();
13887 	}
13888 
13889 	/*
13890 	 * Release the credential hold we took in dtrace_state_create().
13891 	 */
13892 	if (state->dts_cred.dcr_cred != NULL)
13893 		crfree(state->dts_cred.dcr_cred);
13894 
13895 	/*
13896 	 * Now we can safely disable and destroy any enabled probes.  Because
13897 	 * any DTRACE_PRIV_KERNEL probes may actually be slowing our progress
13898 	 * (especially if they're all enabled), we take two passes through the
13899 	 * ECBs:  in the first, we disable just DTRACE_PRIV_KERNEL probes, and
13900 	 * in the second we disable whatever is left over.
13901 	 */
13902 	for (match = DTRACE_PRIV_KERNEL; ; match = 0) {
13903 		for (i = 0; i < state->dts_necbs; i++) {
13904 			if ((ecb = state->dts_ecbs[i]) == NULL)
13905 				continue;
13906 
13907 			if (match && ecb->dte_probe != NULL) {
13908 				dtrace_probe_t *probe = ecb->dte_probe;
13909 				dtrace_provider_t *prov = probe->dtpr_provider;
13910 
13911 				if (!(prov->dtpv_priv.dtpp_flags & match))
13912 					continue;
13913 			}
13914 
13915 			dtrace_ecb_disable(ecb);
13916 			dtrace_ecb_destroy(ecb);
13917 		}
13918 
13919 		if (!match)
13920 			break;
13921 	}
13922 
13923 	/*
13924 	 * Before we free the buffers, perform one more sync to assure that
13925 	 * every CPU is out of probe context.
13926 	 */
13927 	dtrace_sync();
13928 
13929 	dtrace_buffer_free(state->dts_buffer);
13930 	dtrace_buffer_free(state->dts_aggbuffer);
13931 
13932 	for (i = 0; i < nspec; i++)
13933 		dtrace_buffer_free(spec[i].dtsp_buffer);
13934 
13935 #if defined(sun)
13936 	if (state->dts_cleaner != CYCLIC_NONE)
13937 		cyclic_remove(state->dts_cleaner);
13938 
13939 	if (state->dts_deadman != CYCLIC_NONE)
13940 		cyclic_remove(state->dts_deadman);
13941 #else
13942 	callout_stop(&state->dts_cleaner);
13943 	callout_drain(&state->dts_cleaner);
13944 	callout_stop(&state->dts_deadman);
13945 	callout_drain(&state->dts_deadman);
13946 #endif
13947 
13948 	dtrace_dstate_fini(&vstate->dtvs_dynvars);
13949 	dtrace_vstate_fini(vstate);
13950 	if (state->dts_ecbs != NULL)
13951 		kmem_free(state->dts_ecbs, state->dts_necbs * sizeof (dtrace_ecb_t *));
13952 
13953 	if (state->dts_aggregations != NULL) {
13954 #ifdef DEBUG
13955 		for (i = 0; i < state->dts_naggregations; i++)
13956 			ASSERT(state->dts_aggregations[i] == NULL);
13957 #endif
13958 		ASSERT(state->dts_naggregations > 0);
13959 		kmem_free(state->dts_aggregations,
13960 		    state->dts_naggregations * sizeof (dtrace_aggregation_t *));
13961 	}
13962 
13963 	kmem_free(state->dts_buffer, bufsize);
13964 	kmem_free(state->dts_aggbuffer, bufsize);
13965 
13966 	for (i = 0; i < nspec; i++)
13967 		kmem_free(spec[i].dtsp_buffer, bufsize);
13968 
13969 	if (spec != NULL)
13970 		kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
13971 
13972 	dtrace_format_destroy(state);
13973 
13974 	if (state->dts_aggid_arena != NULL) {
13975 #if defined(sun)
13976 		vmem_destroy(state->dts_aggid_arena);
13977 #else
13978 		delete_unrhdr(state->dts_aggid_arena);
13979 #endif
13980 		state->dts_aggid_arena = NULL;
13981 	}
13982 #if defined(sun)
13983 	ddi_soft_state_free(dtrace_softstate, minor);
13984 	vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
13985 #endif
13986 }
13987 
13988 /*
13989  * DTrace Anonymous Enabling Functions
13990  */
13991 static dtrace_state_t *
13992 dtrace_anon_grab(void)
13993 {
13994 	dtrace_state_t *state;
13995 
13996 	ASSERT(MUTEX_HELD(&dtrace_lock));
13997 
13998 	if ((state = dtrace_anon.dta_state) == NULL) {
13999 		ASSERT(dtrace_anon.dta_enabling == NULL);
14000 		return (NULL);
14001 	}
14002 
14003 	ASSERT(dtrace_anon.dta_enabling != NULL);
14004 	ASSERT(dtrace_retained != NULL);
14005 
14006 	dtrace_enabling_destroy(dtrace_anon.dta_enabling);
14007 	dtrace_anon.dta_enabling = NULL;
14008 	dtrace_anon.dta_state = NULL;
14009 
14010 	return (state);
14011 }
14012 
14013 static void
14014 dtrace_anon_property(void)
14015 {
14016 	int i, rv;
14017 	dtrace_state_t *state;
14018 	dof_hdr_t *dof;
14019 	char c[32];		/* enough for "dof-data-" + digits */
14020 
14021 	ASSERT(MUTEX_HELD(&dtrace_lock));
14022 	ASSERT(MUTEX_HELD(&cpu_lock));
14023 
14024 	for (i = 0; ; i++) {
14025 		(void) snprintf(c, sizeof (c), "dof-data-%d", i);
14026 
14027 		dtrace_err_verbose = 1;
14028 
14029 		if ((dof = dtrace_dof_property(c)) == NULL) {
14030 			dtrace_err_verbose = 0;
14031 			break;
14032 		}
14033 
14034 #if defined(sun)
14035 		/*
14036 		 * We want to create anonymous state, so we need to transition
14037 		 * the kernel debugger to indicate that DTrace is active.  If
14038 		 * this fails (e.g. because the debugger has modified text in
14039 		 * some way), we won't continue with the processing.
14040 		 */
14041 		if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
14042 			cmn_err(CE_NOTE, "kernel debugger active; anonymous "
14043 			    "enabling ignored.");
14044 			dtrace_dof_destroy(dof);
14045 			break;
14046 		}
14047 #endif
14048 
14049 		/*
14050 		 * If we haven't allocated an anonymous state, we'll do so now.
14051 		 */
14052 		if ((state = dtrace_anon.dta_state) == NULL) {
14053 #if defined(sun)
14054 			state = dtrace_state_create(NULL, NULL);
14055 #else
14056 			state = dtrace_state_create(NULL);
14057 #endif
14058 			dtrace_anon.dta_state = state;
14059 
14060 			if (state == NULL) {
14061 				/*
14062 				 * This basically shouldn't happen:  the only
14063 				 * failure mode from dtrace_state_create() is a
14064 				 * failure of ddi_soft_state_zalloc() that
14065 				 * itself should never happen.  Still, the
14066 				 * interface allows for a failure mode, and
14067 				 * we want to fail as gracefully as possible:
14068 				 * we'll emit an error message and cease
14069 				 * processing anonymous state in this case.
14070 				 */
14071 				cmn_err(CE_WARN, "failed to create "
14072 				    "anonymous state");
14073 				dtrace_dof_destroy(dof);
14074 				break;
14075 			}
14076 		}
14077 
14078 		rv = dtrace_dof_slurp(dof, &state->dts_vstate, CRED(),
14079 		    &dtrace_anon.dta_enabling, 0, B_TRUE);
14080 
14081 		if (rv == 0)
14082 			rv = dtrace_dof_options(dof, state);
14083 
14084 		dtrace_err_verbose = 0;
14085 		dtrace_dof_destroy(dof);
14086 
14087 		if (rv != 0) {
14088 			/*
14089 			 * This is malformed DOF; chuck any anonymous state
14090 			 * that we created.
14091 			 */
14092 			ASSERT(dtrace_anon.dta_enabling == NULL);
14093 			dtrace_state_destroy(state);
14094 			dtrace_anon.dta_state = NULL;
14095 			break;
14096 		}
14097 
14098 		ASSERT(dtrace_anon.dta_enabling != NULL);
14099 	}
14100 
14101 	if (dtrace_anon.dta_enabling != NULL) {
14102 		int rval;
14103 
14104 		/*
14105 		 * dtrace_enabling_retain() can only fail because we are
14106 		 * trying to retain more enablings than are allowed -- but
14107 		 * we only have one anonymous enabling, and we are guaranteed
14108 		 * to be allowed at least one retained enabling; we assert
14109 		 * that dtrace_enabling_retain() returns success.
14110 		 */
14111 		rval = dtrace_enabling_retain(dtrace_anon.dta_enabling);
14112 		ASSERT(rval == 0);
14113 
14114 		dtrace_enabling_dump(dtrace_anon.dta_enabling);
14115 	}
14116 }
14117 
14118 /*
14119  * DTrace Helper Functions
14120  */
14121 static void
14122 dtrace_helper_trace(dtrace_helper_action_t *helper,
14123     dtrace_mstate_t *mstate, dtrace_vstate_t *vstate, int where)
14124 {
14125 	uint32_t size, next, nnext, i;
14126 	dtrace_helptrace_t *ent;
14127 	uint16_t flags = cpu_core[curcpu].cpuc_dtrace_flags;
14128 
14129 	if (!dtrace_helptrace_enabled)
14130 		return;
14131 
14132 	ASSERT(vstate->dtvs_nlocals <= dtrace_helptrace_nlocals);
14133 
14134 	/*
14135 	 * What would a tracing framework be without its own tracing
14136 	 * framework?  (Well, a hell of a lot simpler, for starters...)
14137 	 */
14138 	size = sizeof (dtrace_helptrace_t) + dtrace_helptrace_nlocals *
14139 	    sizeof (uint64_t) - sizeof (uint64_t);
14140 
14141 	/*
14142 	 * Iterate until we can allocate a slot in the trace buffer.
14143 	 */
14144 	do {
14145 		next = dtrace_helptrace_next;
14146 
14147 		if (next + size < dtrace_helptrace_bufsize) {
14148 			nnext = next + size;
14149 		} else {
14150 			nnext = size;
14151 		}
14152 	} while (dtrace_cas32(&dtrace_helptrace_next, next, nnext) != next);
14153 
14154 	/*
14155 	 * We have our slot; fill it in.
14156 	 */
14157 	if (nnext == size)
14158 		next = 0;
14159 
14160 	ent = (dtrace_helptrace_t *)&dtrace_helptrace_buffer[next];
14161 	ent->dtht_helper = helper;
14162 	ent->dtht_where = where;
14163 	ent->dtht_nlocals = vstate->dtvs_nlocals;
14164 
14165 	ent->dtht_fltoffs = (mstate->dtms_present & DTRACE_MSTATE_FLTOFFS) ?
14166 	    mstate->dtms_fltoffs : -1;
14167 	ent->dtht_fault = DTRACE_FLAGS2FLT(flags);
14168 	ent->dtht_illval = cpu_core[curcpu].cpuc_dtrace_illval;
14169 
14170 	for (i = 0; i < vstate->dtvs_nlocals; i++) {
14171 		dtrace_statvar_t *svar;
14172 
14173 		if ((svar = vstate->dtvs_locals[i]) == NULL)
14174 			continue;
14175 
14176 		ASSERT(svar->dtsv_size >= NCPU * sizeof (uint64_t));
14177 		ent->dtht_locals[i] =
14178 		    ((uint64_t *)(uintptr_t)svar->dtsv_data)[curcpu];
14179 	}
14180 }
14181 
14182 static uint64_t
14183 dtrace_helper(int which, dtrace_mstate_t *mstate,
14184     dtrace_state_t *state, uint64_t arg0, uint64_t arg1)
14185 {
14186 	uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
14187 	uint64_t sarg0 = mstate->dtms_arg[0];
14188 	uint64_t sarg1 = mstate->dtms_arg[1];
14189 	uint64_t rval = 0;
14190 	dtrace_helpers_t *helpers = curproc->p_dtrace_helpers;
14191 	dtrace_helper_action_t *helper;
14192 	dtrace_vstate_t *vstate;
14193 	dtrace_difo_t *pred;
14194 	int i, trace = dtrace_helptrace_enabled;
14195 
14196 	ASSERT(which >= 0 && which < DTRACE_NHELPER_ACTIONS);
14197 
14198 	if (helpers == NULL)
14199 		return (0);
14200 
14201 	if ((helper = helpers->dthps_actions[which]) == NULL)
14202 		return (0);
14203 
14204 	vstate = &helpers->dthps_vstate;
14205 	mstate->dtms_arg[0] = arg0;
14206 	mstate->dtms_arg[1] = arg1;
14207 
14208 	/*
14209 	 * Now iterate over each helper.  If its predicate evaluates to 'true',
14210 	 * we'll call the corresponding actions.  Note that the below calls
14211 	 * to dtrace_dif_emulate() may set faults in machine state.  This is
14212 	 * okay:  our caller (the outer dtrace_dif_emulate()) will simply plow
14213 	 * the stored DIF offset with its own (which is the desired behavior).
14214 	 * Also, note the calls to dtrace_dif_emulate() may allocate scratch
14215 	 * from machine state; this is okay, too.
14216 	 */
14217 	for (; helper != NULL; helper = helper->dtha_next) {
14218 		if ((pred = helper->dtha_predicate) != NULL) {
14219 			if (trace)
14220 				dtrace_helper_trace(helper, mstate, vstate, 0);
14221 
14222 			if (!dtrace_dif_emulate(pred, mstate, vstate, state))
14223 				goto next;
14224 
14225 			if (*flags & CPU_DTRACE_FAULT)
14226 				goto err;
14227 		}
14228 
14229 		for (i = 0; i < helper->dtha_nactions; i++) {
14230 			if (trace)
14231 				dtrace_helper_trace(helper,
14232 				    mstate, vstate, i + 1);
14233 
14234 			rval = dtrace_dif_emulate(helper->dtha_actions[i],
14235 			    mstate, vstate, state);
14236 
14237 			if (*flags & CPU_DTRACE_FAULT)
14238 				goto err;
14239 		}
14240 
14241 next:
14242 		if (trace)
14243 			dtrace_helper_trace(helper, mstate, vstate,
14244 			    DTRACE_HELPTRACE_NEXT);
14245 	}
14246 
14247 	if (trace)
14248 		dtrace_helper_trace(helper, mstate, vstate,
14249 		    DTRACE_HELPTRACE_DONE);
14250 
14251 	/*
14252 	 * Restore the arg0 that we saved upon entry.
14253 	 */
14254 	mstate->dtms_arg[0] = sarg0;
14255 	mstate->dtms_arg[1] = sarg1;
14256 
14257 	return (rval);
14258 
14259 err:
14260 	if (trace)
14261 		dtrace_helper_trace(helper, mstate, vstate,
14262 		    DTRACE_HELPTRACE_ERR);
14263 
14264 	/*
14265 	 * Restore the arg0 that we saved upon entry.
14266 	 */
14267 	mstate->dtms_arg[0] = sarg0;
14268 	mstate->dtms_arg[1] = sarg1;
14269 
14270 	return (0);
14271 }
14272 
14273 static void
14274 dtrace_helper_action_destroy(dtrace_helper_action_t *helper,
14275     dtrace_vstate_t *vstate)
14276 {
14277 	int i;
14278 
14279 	if (helper->dtha_predicate != NULL)
14280 		dtrace_difo_release(helper->dtha_predicate, vstate);
14281 
14282 	for (i = 0; i < helper->dtha_nactions; i++) {
14283 		ASSERT(helper->dtha_actions[i] != NULL);
14284 		dtrace_difo_release(helper->dtha_actions[i], vstate);
14285 	}
14286 
14287 	kmem_free(helper->dtha_actions,
14288 	    helper->dtha_nactions * sizeof (dtrace_difo_t *));
14289 	kmem_free(helper, sizeof (dtrace_helper_action_t));
14290 }
14291 
14292 static int
14293 dtrace_helper_destroygen(int gen)
14294 {
14295 	proc_t *p = curproc;
14296 	dtrace_helpers_t *help = p->p_dtrace_helpers;
14297 	dtrace_vstate_t *vstate;
14298 	int i;
14299 
14300 	ASSERT(MUTEX_HELD(&dtrace_lock));
14301 
14302 	if (help == NULL || gen > help->dthps_generation)
14303 		return (EINVAL);
14304 
14305 	vstate = &help->dthps_vstate;
14306 
14307 	for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
14308 		dtrace_helper_action_t *last = NULL, *h, *next;
14309 
14310 		for (h = help->dthps_actions[i]; h != NULL; h = next) {
14311 			next = h->dtha_next;
14312 
14313 			if (h->dtha_generation == gen) {
14314 				if (last != NULL) {
14315 					last->dtha_next = next;
14316 				} else {
14317 					help->dthps_actions[i] = next;
14318 				}
14319 
14320 				dtrace_helper_action_destroy(h, vstate);
14321 			} else {
14322 				last = h;
14323 			}
14324 		}
14325 	}
14326 
14327 	/*
14328 	 * Interate until we've cleared out all helper providers with the
14329 	 * given generation number.
14330 	 */
14331 	for (;;) {
14332 		dtrace_helper_provider_t *prov;
14333 
14334 		/*
14335 		 * Look for a helper provider with the right generation. We
14336 		 * have to start back at the beginning of the list each time
14337 		 * because we drop dtrace_lock. It's unlikely that we'll make
14338 		 * more than two passes.
14339 		 */
14340 		for (i = 0; i < help->dthps_nprovs; i++) {
14341 			prov = help->dthps_provs[i];
14342 
14343 			if (prov->dthp_generation == gen)
14344 				break;
14345 		}
14346 
14347 		/*
14348 		 * If there were no matches, we're done.
14349 		 */
14350 		if (i == help->dthps_nprovs)
14351 			break;
14352 
14353 		/*
14354 		 * Move the last helper provider into this slot.
14355 		 */
14356 		help->dthps_nprovs--;
14357 		help->dthps_provs[i] = help->dthps_provs[help->dthps_nprovs];
14358 		help->dthps_provs[help->dthps_nprovs] = NULL;
14359 
14360 		mutex_exit(&dtrace_lock);
14361 
14362 		/*
14363 		 * If we have a meta provider, remove this helper provider.
14364 		 */
14365 		mutex_enter(&dtrace_meta_lock);
14366 		if (dtrace_meta_pid != NULL) {
14367 			ASSERT(dtrace_deferred_pid == NULL);
14368 			dtrace_helper_provider_remove(&prov->dthp_prov,
14369 			    p->p_pid);
14370 		}
14371 		mutex_exit(&dtrace_meta_lock);
14372 
14373 		dtrace_helper_provider_destroy(prov);
14374 
14375 		mutex_enter(&dtrace_lock);
14376 	}
14377 
14378 	return (0);
14379 }
14380 
14381 static int
14382 dtrace_helper_validate(dtrace_helper_action_t *helper)
14383 {
14384 	int err = 0, i;
14385 	dtrace_difo_t *dp;
14386 
14387 	if ((dp = helper->dtha_predicate) != NULL)
14388 		err += dtrace_difo_validate_helper(dp);
14389 
14390 	for (i = 0; i < helper->dtha_nactions; i++)
14391 		err += dtrace_difo_validate_helper(helper->dtha_actions[i]);
14392 
14393 	return (err == 0);
14394 }
14395 
14396 static int
14397 dtrace_helper_action_add(int which, dtrace_ecbdesc_t *ep)
14398 {
14399 	dtrace_helpers_t *help;
14400 	dtrace_helper_action_t *helper, *last;
14401 	dtrace_actdesc_t *act;
14402 	dtrace_vstate_t *vstate;
14403 	dtrace_predicate_t *pred;
14404 	int count = 0, nactions = 0, i;
14405 
14406 	if (which < 0 || which >= DTRACE_NHELPER_ACTIONS)
14407 		return (EINVAL);
14408 
14409 	help = curproc->p_dtrace_helpers;
14410 	last = help->dthps_actions[which];
14411 	vstate = &help->dthps_vstate;
14412 
14413 	for (count = 0; last != NULL; last = last->dtha_next) {
14414 		count++;
14415 		if (last->dtha_next == NULL)
14416 			break;
14417 	}
14418 
14419 	/*
14420 	 * If we already have dtrace_helper_actions_max helper actions for this
14421 	 * helper action type, we'll refuse to add a new one.
14422 	 */
14423 	if (count >= dtrace_helper_actions_max)
14424 		return (ENOSPC);
14425 
14426 	helper = kmem_zalloc(sizeof (dtrace_helper_action_t), KM_SLEEP);
14427 	helper->dtha_generation = help->dthps_generation;
14428 
14429 	if ((pred = ep->dted_pred.dtpdd_predicate) != NULL) {
14430 		ASSERT(pred->dtp_difo != NULL);
14431 		dtrace_difo_hold(pred->dtp_difo);
14432 		helper->dtha_predicate = pred->dtp_difo;
14433 	}
14434 
14435 	for (act = ep->dted_action; act != NULL; act = act->dtad_next) {
14436 		if (act->dtad_kind != DTRACEACT_DIFEXPR)
14437 			goto err;
14438 
14439 		if (act->dtad_difo == NULL)
14440 			goto err;
14441 
14442 		nactions++;
14443 	}
14444 
14445 	helper->dtha_actions = kmem_zalloc(sizeof (dtrace_difo_t *) *
14446 	    (helper->dtha_nactions = nactions), KM_SLEEP);
14447 
14448 	for (act = ep->dted_action, i = 0; act != NULL; act = act->dtad_next) {
14449 		dtrace_difo_hold(act->dtad_difo);
14450 		helper->dtha_actions[i++] = act->dtad_difo;
14451 	}
14452 
14453 	if (!dtrace_helper_validate(helper))
14454 		goto err;
14455 
14456 	if (last == NULL) {
14457 		help->dthps_actions[which] = helper;
14458 	} else {
14459 		last->dtha_next = helper;
14460 	}
14461 
14462 	if (vstate->dtvs_nlocals > dtrace_helptrace_nlocals) {
14463 		dtrace_helptrace_nlocals = vstate->dtvs_nlocals;
14464 		dtrace_helptrace_next = 0;
14465 	}
14466 
14467 	return (0);
14468 err:
14469 	dtrace_helper_action_destroy(helper, vstate);
14470 	return (EINVAL);
14471 }
14472 
14473 static void
14474 dtrace_helper_provider_register(proc_t *p, dtrace_helpers_t *help,
14475     dof_helper_t *dofhp)
14476 {
14477 	ASSERT(MUTEX_NOT_HELD(&dtrace_lock));
14478 
14479 	mutex_enter(&dtrace_meta_lock);
14480 	mutex_enter(&dtrace_lock);
14481 
14482 	if (!dtrace_attached() || dtrace_meta_pid == NULL) {
14483 		/*
14484 		 * If the dtrace module is loaded but not attached, or if
14485 		 * there aren't isn't a meta provider registered to deal with
14486 		 * these provider descriptions, we need to postpone creating
14487 		 * the actual providers until later.
14488 		 */
14489 
14490 		if (help->dthps_next == NULL && help->dthps_prev == NULL &&
14491 		    dtrace_deferred_pid != help) {
14492 			help->dthps_deferred = 1;
14493 			help->dthps_pid = p->p_pid;
14494 			help->dthps_next = dtrace_deferred_pid;
14495 			help->dthps_prev = NULL;
14496 			if (dtrace_deferred_pid != NULL)
14497 				dtrace_deferred_pid->dthps_prev = help;
14498 			dtrace_deferred_pid = help;
14499 		}
14500 
14501 		mutex_exit(&dtrace_lock);
14502 
14503 	} else if (dofhp != NULL) {
14504 		/*
14505 		 * If the dtrace module is loaded and we have a particular
14506 		 * helper provider description, pass that off to the
14507 		 * meta provider.
14508 		 */
14509 
14510 		mutex_exit(&dtrace_lock);
14511 
14512 		dtrace_helper_provide(dofhp, p->p_pid);
14513 
14514 	} else {
14515 		/*
14516 		 * Otherwise, just pass all the helper provider descriptions
14517 		 * off to the meta provider.
14518 		 */
14519 
14520 		int i;
14521 		mutex_exit(&dtrace_lock);
14522 
14523 		for (i = 0; i < help->dthps_nprovs; i++) {
14524 			dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
14525 			    p->p_pid);
14526 		}
14527 	}
14528 
14529 	mutex_exit(&dtrace_meta_lock);
14530 }
14531 
14532 static int
14533 dtrace_helper_provider_add(dof_helper_t *dofhp, int gen)
14534 {
14535 	dtrace_helpers_t *help;
14536 	dtrace_helper_provider_t *hprov, **tmp_provs;
14537 	uint_t tmp_maxprovs, i;
14538 
14539 	ASSERT(MUTEX_HELD(&dtrace_lock));
14540 
14541 	help = curproc->p_dtrace_helpers;
14542 	ASSERT(help != NULL);
14543 
14544 	/*
14545 	 * If we already have dtrace_helper_providers_max helper providers,
14546 	 * we're refuse to add a new one.
14547 	 */
14548 	if (help->dthps_nprovs >= dtrace_helper_providers_max)
14549 		return (ENOSPC);
14550 
14551 	/*
14552 	 * Check to make sure this isn't a duplicate.
14553 	 */
14554 	for (i = 0; i < help->dthps_nprovs; i++) {
14555 		if (dofhp->dofhp_addr ==
14556 		    help->dthps_provs[i]->dthp_prov.dofhp_addr)
14557 			return (EALREADY);
14558 	}
14559 
14560 	hprov = kmem_zalloc(sizeof (dtrace_helper_provider_t), KM_SLEEP);
14561 	hprov->dthp_prov = *dofhp;
14562 	hprov->dthp_ref = 1;
14563 	hprov->dthp_generation = gen;
14564 
14565 	/*
14566 	 * Allocate a bigger table for helper providers if it's already full.
14567 	 */
14568 	if (help->dthps_maxprovs == help->dthps_nprovs) {
14569 		tmp_maxprovs = help->dthps_maxprovs;
14570 		tmp_provs = help->dthps_provs;
14571 
14572 		if (help->dthps_maxprovs == 0)
14573 			help->dthps_maxprovs = 2;
14574 		else
14575 			help->dthps_maxprovs *= 2;
14576 		if (help->dthps_maxprovs > dtrace_helper_providers_max)
14577 			help->dthps_maxprovs = dtrace_helper_providers_max;
14578 
14579 		ASSERT(tmp_maxprovs < help->dthps_maxprovs);
14580 
14581 		help->dthps_provs = kmem_zalloc(help->dthps_maxprovs *
14582 		    sizeof (dtrace_helper_provider_t *), KM_SLEEP);
14583 
14584 		if (tmp_provs != NULL) {
14585 			bcopy(tmp_provs, help->dthps_provs, tmp_maxprovs *
14586 			    sizeof (dtrace_helper_provider_t *));
14587 			kmem_free(tmp_provs, tmp_maxprovs *
14588 			    sizeof (dtrace_helper_provider_t *));
14589 		}
14590 	}
14591 
14592 	help->dthps_provs[help->dthps_nprovs] = hprov;
14593 	help->dthps_nprovs++;
14594 
14595 	return (0);
14596 }
14597 
14598 static void
14599 dtrace_helper_provider_destroy(dtrace_helper_provider_t *hprov)
14600 {
14601 	mutex_enter(&dtrace_lock);
14602 
14603 	if (--hprov->dthp_ref == 0) {
14604 		dof_hdr_t *dof;
14605 		mutex_exit(&dtrace_lock);
14606 		dof = (dof_hdr_t *)(uintptr_t)hprov->dthp_prov.dofhp_dof;
14607 		dtrace_dof_destroy(dof);
14608 		kmem_free(hprov, sizeof (dtrace_helper_provider_t));
14609 	} else {
14610 		mutex_exit(&dtrace_lock);
14611 	}
14612 }
14613 
14614 static int
14615 dtrace_helper_provider_validate(dof_hdr_t *dof, dof_sec_t *sec)
14616 {
14617 	uintptr_t daddr = (uintptr_t)dof;
14618 	dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
14619 	dof_provider_t *provider;
14620 	dof_probe_t *probe;
14621 	uint8_t *arg;
14622 	char *strtab, *typestr;
14623 	dof_stridx_t typeidx;
14624 	size_t typesz;
14625 	uint_t nprobes, j, k;
14626 
14627 	ASSERT(sec->dofs_type == DOF_SECT_PROVIDER);
14628 
14629 	if (sec->dofs_offset & (sizeof (uint_t) - 1)) {
14630 		dtrace_dof_error(dof, "misaligned section offset");
14631 		return (-1);
14632 	}
14633 
14634 	/*
14635 	 * The section needs to be large enough to contain the DOF provider
14636 	 * structure appropriate for the given version.
14637 	 */
14638 	if (sec->dofs_size <
14639 	    ((dof->dofh_ident[DOF_ID_VERSION] == DOF_VERSION_1) ?
14640 	    offsetof(dof_provider_t, dofpv_prenoffs) :
14641 	    sizeof (dof_provider_t))) {
14642 		dtrace_dof_error(dof, "provider section too small");
14643 		return (-1);
14644 	}
14645 
14646 	provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
14647 	str_sec = dtrace_dof_sect(dof, DOF_SECT_STRTAB, provider->dofpv_strtab);
14648 	prb_sec = dtrace_dof_sect(dof, DOF_SECT_PROBES, provider->dofpv_probes);
14649 	arg_sec = dtrace_dof_sect(dof, DOF_SECT_PRARGS, provider->dofpv_prargs);
14650 	off_sec = dtrace_dof_sect(dof, DOF_SECT_PROFFS, provider->dofpv_proffs);
14651 
14652 	if (str_sec == NULL || prb_sec == NULL ||
14653 	    arg_sec == NULL || off_sec == NULL)
14654 		return (-1);
14655 
14656 	enoff_sec = NULL;
14657 
14658 	if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
14659 	    provider->dofpv_prenoffs != DOF_SECT_NONE &&
14660 	    (enoff_sec = dtrace_dof_sect(dof, DOF_SECT_PRENOFFS,
14661 	    provider->dofpv_prenoffs)) == NULL)
14662 		return (-1);
14663 
14664 	strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
14665 
14666 	if (provider->dofpv_name >= str_sec->dofs_size ||
14667 	    strlen(strtab + provider->dofpv_name) >= DTRACE_PROVNAMELEN) {
14668 		dtrace_dof_error(dof, "invalid provider name");
14669 		return (-1);
14670 	}
14671 
14672 	if (prb_sec->dofs_entsize == 0 ||
14673 	    prb_sec->dofs_entsize > prb_sec->dofs_size) {
14674 		dtrace_dof_error(dof, "invalid entry size");
14675 		return (-1);
14676 	}
14677 
14678 	if (prb_sec->dofs_entsize & (sizeof (uintptr_t) - 1)) {
14679 		dtrace_dof_error(dof, "misaligned entry size");
14680 		return (-1);
14681 	}
14682 
14683 	if (off_sec->dofs_entsize != sizeof (uint32_t)) {
14684 		dtrace_dof_error(dof, "invalid entry size");
14685 		return (-1);
14686 	}
14687 
14688 	if (off_sec->dofs_offset & (sizeof (uint32_t) - 1)) {
14689 		dtrace_dof_error(dof, "misaligned section offset");
14690 		return (-1);
14691 	}
14692 
14693 	if (arg_sec->dofs_entsize != sizeof (uint8_t)) {
14694 		dtrace_dof_error(dof, "invalid entry size");
14695 		return (-1);
14696 	}
14697 
14698 	arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
14699 
14700 	nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
14701 
14702 	/*
14703 	 * Take a pass through the probes to check for errors.
14704 	 */
14705 	for (j = 0; j < nprobes; j++) {
14706 		probe = (dof_probe_t *)(uintptr_t)(daddr +
14707 		    prb_sec->dofs_offset + j * prb_sec->dofs_entsize);
14708 
14709 		if (probe->dofpr_func >= str_sec->dofs_size) {
14710 			dtrace_dof_error(dof, "invalid function name");
14711 			return (-1);
14712 		}
14713 
14714 		if (strlen(strtab + probe->dofpr_func) >= DTRACE_FUNCNAMELEN) {
14715 			dtrace_dof_error(dof, "function name too long");
14716 			return (-1);
14717 		}
14718 
14719 		if (probe->dofpr_name >= str_sec->dofs_size ||
14720 		    strlen(strtab + probe->dofpr_name) >= DTRACE_NAMELEN) {
14721 			dtrace_dof_error(dof, "invalid probe name");
14722 			return (-1);
14723 		}
14724 
14725 		/*
14726 		 * The offset count must not wrap the index, and the offsets
14727 		 * must also not overflow the section's data.
14728 		 */
14729 		if (probe->dofpr_offidx + probe->dofpr_noffs <
14730 		    probe->dofpr_offidx ||
14731 		    (probe->dofpr_offidx + probe->dofpr_noffs) *
14732 		    off_sec->dofs_entsize > off_sec->dofs_size) {
14733 			dtrace_dof_error(dof, "invalid probe offset");
14734 			return (-1);
14735 		}
14736 
14737 		if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1) {
14738 			/*
14739 			 * If there's no is-enabled offset section, make sure
14740 			 * there aren't any is-enabled offsets. Otherwise
14741 			 * perform the same checks as for probe offsets
14742 			 * (immediately above).
14743 			 */
14744 			if (enoff_sec == NULL) {
14745 				if (probe->dofpr_enoffidx != 0 ||
14746 				    probe->dofpr_nenoffs != 0) {
14747 					dtrace_dof_error(dof, "is-enabled "
14748 					    "offsets with null section");
14749 					return (-1);
14750 				}
14751 			} else if (probe->dofpr_enoffidx +
14752 			    probe->dofpr_nenoffs < probe->dofpr_enoffidx ||
14753 			    (probe->dofpr_enoffidx + probe->dofpr_nenoffs) *
14754 			    enoff_sec->dofs_entsize > enoff_sec->dofs_size) {
14755 				dtrace_dof_error(dof, "invalid is-enabled "
14756 				    "offset");
14757 				return (-1);
14758 			}
14759 
14760 			if (probe->dofpr_noffs + probe->dofpr_nenoffs == 0) {
14761 				dtrace_dof_error(dof, "zero probe and "
14762 				    "is-enabled offsets");
14763 				return (-1);
14764 			}
14765 		} else if (probe->dofpr_noffs == 0) {
14766 			dtrace_dof_error(dof, "zero probe offsets");
14767 			return (-1);
14768 		}
14769 
14770 		if (probe->dofpr_argidx + probe->dofpr_xargc <
14771 		    probe->dofpr_argidx ||
14772 		    (probe->dofpr_argidx + probe->dofpr_xargc) *
14773 		    arg_sec->dofs_entsize > arg_sec->dofs_size) {
14774 			dtrace_dof_error(dof, "invalid args");
14775 			return (-1);
14776 		}
14777 
14778 		typeidx = probe->dofpr_nargv;
14779 		typestr = strtab + probe->dofpr_nargv;
14780 		for (k = 0; k < probe->dofpr_nargc; k++) {
14781 			if (typeidx >= str_sec->dofs_size) {
14782 				dtrace_dof_error(dof, "bad "
14783 				    "native argument type");
14784 				return (-1);
14785 			}
14786 
14787 			typesz = strlen(typestr) + 1;
14788 			if (typesz > DTRACE_ARGTYPELEN) {
14789 				dtrace_dof_error(dof, "native "
14790 				    "argument type too long");
14791 				return (-1);
14792 			}
14793 			typeidx += typesz;
14794 			typestr += typesz;
14795 		}
14796 
14797 		typeidx = probe->dofpr_xargv;
14798 		typestr = strtab + probe->dofpr_xargv;
14799 		for (k = 0; k < probe->dofpr_xargc; k++) {
14800 			if (arg[probe->dofpr_argidx + k] > probe->dofpr_nargc) {
14801 				dtrace_dof_error(dof, "bad "
14802 				    "native argument index");
14803 				return (-1);
14804 			}
14805 
14806 			if (typeidx >= str_sec->dofs_size) {
14807 				dtrace_dof_error(dof, "bad "
14808 				    "translated argument type");
14809 				return (-1);
14810 			}
14811 
14812 			typesz = strlen(typestr) + 1;
14813 			if (typesz > DTRACE_ARGTYPELEN) {
14814 				dtrace_dof_error(dof, "translated argument "
14815 				    "type too long");
14816 				return (-1);
14817 			}
14818 
14819 			typeidx += typesz;
14820 			typestr += typesz;
14821 		}
14822 	}
14823 
14824 	return (0);
14825 }
14826 
14827 static int
14828 dtrace_helper_slurp(dof_hdr_t *dof, dof_helper_t *dhp)
14829 {
14830 	dtrace_helpers_t *help;
14831 	dtrace_vstate_t *vstate;
14832 	dtrace_enabling_t *enab = NULL;
14833 	int i, gen, rv, nhelpers = 0, nprovs = 0, destroy = 1;
14834 	uintptr_t daddr = (uintptr_t)dof;
14835 
14836 	ASSERT(MUTEX_HELD(&dtrace_lock));
14837 
14838 	if ((help = curproc->p_dtrace_helpers) == NULL)
14839 		help = dtrace_helpers_create(curproc);
14840 
14841 	vstate = &help->dthps_vstate;
14842 
14843 	if ((rv = dtrace_dof_slurp(dof, vstate, NULL, &enab,
14844 	    dhp != NULL ? dhp->dofhp_addr : 0, B_FALSE)) != 0) {
14845 		dtrace_dof_destroy(dof);
14846 		return (rv);
14847 	}
14848 
14849 	/*
14850 	 * Look for helper providers and validate their descriptions.
14851 	 */
14852 	if (dhp != NULL) {
14853 		for (i = 0; i < dof->dofh_secnum; i++) {
14854 			dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
14855 			    dof->dofh_secoff + i * dof->dofh_secsize);
14856 
14857 			if (sec->dofs_type != DOF_SECT_PROVIDER)
14858 				continue;
14859 
14860 			if (dtrace_helper_provider_validate(dof, sec) != 0) {
14861 				dtrace_enabling_destroy(enab);
14862 				dtrace_dof_destroy(dof);
14863 				return (-1);
14864 			}
14865 
14866 			nprovs++;
14867 		}
14868 	}
14869 
14870 	/*
14871 	 * Now we need to walk through the ECB descriptions in the enabling.
14872 	 */
14873 	for (i = 0; i < enab->dten_ndesc; i++) {
14874 		dtrace_ecbdesc_t *ep = enab->dten_desc[i];
14875 		dtrace_probedesc_t *desc = &ep->dted_probe;
14876 
14877 		if (strcmp(desc->dtpd_provider, "dtrace") != 0)
14878 			continue;
14879 
14880 		if (strcmp(desc->dtpd_mod, "helper") != 0)
14881 			continue;
14882 
14883 		if (strcmp(desc->dtpd_func, "ustack") != 0)
14884 			continue;
14885 
14886 		if ((rv = dtrace_helper_action_add(DTRACE_HELPER_ACTION_USTACK,
14887 		    ep)) != 0) {
14888 			/*
14889 			 * Adding this helper action failed -- we are now going
14890 			 * to rip out the entire generation and return failure.
14891 			 */
14892 			(void) dtrace_helper_destroygen(help->dthps_generation);
14893 			dtrace_enabling_destroy(enab);
14894 			dtrace_dof_destroy(dof);
14895 			return (-1);
14896 		}
14897 
14898 		nhelpers++;
14899 	}
14900 
14901 	if (nhelpers < enab->dten_ndesc)
14902 		dtrace_dof_error(dof, "unmatched helpers");
14903 
14904 	gen = help->dthps_generation++;
14905 	dtrace_enabling_destroy(enab);
14906 
14907 	if (dhp != NULL && nprovs > 0) {
14908 		dhp->dofhp_dof = (uint64_t)(uintptr_t)dof;
14909 		if (dtrace_helper_provider_add(dhp, gen) == 0) {
14910 			mutex_exit(&dtrace_lock);
14911 			dtrace_helper_provider_register(curproc, help, dhp);
14912 			mutex_enter(&dtrace_lock);
14913 
14914 			destroy = 0;
14915 		}
14916 	}
14917 
14918 	if (destroy)
14919 		dtrace_dof_destroy(dof);
14920 
14921 	return (gen);
14922 }
14923 
14924 static dtrace_helpers_t *
14925 dtrace_helpers_create(proc_t *p)
14926 {
14927 	dtrace_helpers_t *help;
14928 
14929 	ASSERT(MUTEX_HELD(&dtrace_lock));
14930 	ASSERT(p->p_dtrace_helpers == NULL);
14931 
14932 	help = kmem_zalloc(sizeof (dtrace_helpers_t), KM_SLEEP);
14933 	help->dthps_actions = kmem_zalloc(sizeof (dtrace_helper_action_t *) *
14934 	    DTRACE_NHELPER_ACTIONS, KM_SLEEP);
14935 
14936 	p->p_dtrace_helpers = help;
14937 	dtrace_helpers++;
14938 
14939 	return (help);
14940 }
14941 
14942 #if defined(sun)
14943 static
14944 #endif
14945 void
14946 dtrace_helpers_destroy(proc_t *p)
14947 {
14948 	dtrace_helpers_t *help;
14949 	dtrace_vstate_t *vstate;
14950 #if defined(sun)
14951 	proc_t *p = curproc;
14952 #endif
14953 	int i;
14954 
14955 	mutex_enter(&dtrace_lock);
14956 
14957 	ASSERT(p->p_dtrace_helpers != NULL);
14958 	ASSERT(dtrace_helpers > 0);
14959 
14960 	help = p->p_dtrace_helpers;
14961 	vstate = &help->dthps_vstate;
14962 
14963 	/*
14964 	 * We're now going to lose the help from this process.
14965 	 */
14966 	p->p_dtrace_helpers = NULL;
14967 	dtrace_sync();
14968 
14969 	/*
14970 	 * Destory the helper actions.
14971 	 */
14972 	for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
14973 		dtrace_helper_action_t *h, *next;
14974 
14975 		for (h = help->dthps_actions[i]; h != NULL; h = next) {
14976 			next = h->dtha_next;
14977 			dtrace_helper_action_destroy(h, vstate);
14978 			h = next;
14979 		}
14980 	}
14981 
14982 	mutex_exit(&dtrace_lock);
14983 
14984 	/*
14985 	 * Destroy the helper providers.
14986 	 */
14987 	if (help->dthps_maxprovs > 0) {
14988 		mutex_enter(&dtrace_meta_lock);
14989 		if (dtrace_meta_pid != NULL) {
14990 			ASSERT(dtrace_deferred_pid == NULL);
14991 
14992 			for (i = 0; i < help->dthps_nprovs; i++) {
14993 				dtrace_helper_provider_remove(
14994 				    &help->dthps_provs[i]->dthp_prov, p->p_pid);
14995 			}
14996 		} else {
14997 			mutex_enter(&dtrace_lock);
14998 			ASSERT(help->dthps_deferred == 0 ||
14999 			    help->dthps_next != NULL ||
15000 			    help->dthps_prev != NULL ||
15001 			    help == dtrace_deferred_pid);
15002 
15003 			/*
15004 			 * Remove the helper from the deferred list.
15005 			 */
15006 			if (help->dthps_next != NULL)
15007 				help->dthps_next->dthps_prev = help->dthps_prev;
15008 			if (help->dthps_prev != NULL)
15009 				help->dthps_prev->dthps_next = help->dthps_next;
15010 			if (dtrace_deferred_pid == help) {
15011 				dtrace_deferred_pid = help->dthps_next;
15012 				ASSERT(help->dthps_prev == NULL);
15013 			}
15014 
15015 			mutex_exit(&dtrace_lock);
15016 		}
15017 
15018 		mutex_exit(&dtrace_meta_lock);
15019 
15020 		for (i = 0; i < help->dthps_nprovs; i++) {
15021 			dtrace_helper_provider_destroy(help->dthps_provs[i]);
15022 		}
15023 
15024 		kmem_free(help->dthps_provs, help->dthps_maxprovs *
15025 		    sizeof (dtrace_helper_provider_t *));
15026 	}
15027 
15028 	mutex_enter(&dtrace_lock);
15029 
15030 	dtrace_vstate_fini(&help->dthps_vstate);
15031 	kmem_free(help->dthps_actions,
15032 	    sizeof (dtrace_helper_action_t *) * DTRACE_NHELPER_ACTIONS);
15033 	kmem_free(help, sizeof (dtrace_helpers_t));
15034 
15035 	--dtrace_helpers;
15036 	mutex_exit(&dtrace_lock);
15037 }
15038 
15039 #if defined(sun)
15040 static
15041 #endif
15042 void
15043 dtrace_helpers_duplicate(proc_t *from, proc_t *to)
15044 {
15045 	dtrace_helpers_t *help, *newhelp;
15046 	dtrace_helper_action_t *helper, *new, *last;
15047 	dtrace_difo_t *dp;
15048 	dtrace_vstate_t *vstate;
15049 	int i, j, sz, hasprovs = 0;
15050 
15051 	mutex_enter(&dtrace_lock);
15052 	ASSERT(from->p_dtrace_helpers != NULL);
15053 	ASSERT(dtrace_helpers > 0);
15054 
15055 	help = from->p_dtrace_helpers;
15056 	newhelp = dtrace_helpers_create(to);
15057 	ASSERT(to->p_dtrace_helpers != NULL);
15058 
15059 	newhelp->dthps_generation = help->dthps_generation;
15060 	vstate = &newhelp->dthps_vstate;
15061 
15062 	/*
15063 	 * Duplicate the helper actions.
15064 	 */
15065 	for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
15066 		if ((helper = help->dthps_actions[i]) == NULL)
15067 			continue;
15068 
15069 		for (last = NULL; helper != NULL; helper = helper->dtha_next) {
15070 			new = kmem_zalloc(sizeof (dtrace_helper_action_t),
15071 			    KM_SLEEP);
15072 			new->dtha_generation = helper->dtha_generation;
15073 
15074 			if ((dp = helper->dtha_predicate) != NULL) {
15075 				dp = dtrace_difo_duplicate(dp, vstate);
15076 				new->dtha_predicate = dp;
15077 			}
15078 
15079 			new->dtha_nactions = helper->dtha_nactions;
15080 			sz = sizeof (dtrace_difo_t *) * new->dtha_nactions;
15081 			new->dtha_actions = kmem_alloc(sz, KM_SLEEP);
15082 
15083 			for (j = 0; j < new->dtha_nactions; j++) {
15084 				dtrace_difo_t *dp = helper->dtha_actions[j];
15085 
15086 				ASSERT(dp != NULL);
15087 				dp = dtrace_difo_duplicate(dp, vstate);
15088 				new->dtha_actions[j] = dp;
15089 			}
15090 
15091 			if (last != NULL) {
15092 				last->dtha_next = new;
15093 			} else {
15094 				newhelp->dthps_actions[i] = new;
15095 			}
15096 
15097 			last = new;
15098 		}
15099 	}
15100 
15101 	/*
15102 	 * Duplicate the helper providers and register them with the
15103 	 * DTrace framework.
15104 	 */
15105 	if (help->dthps_nprovs > 0) {
15106 		newhelp->dthps_nprovs = help->dthps_nprovs;
15107 		newhelp->dthps_maxprovs = help->dthps_nprovs;
15108 		newhelp->dthps_provs = kmem_alloc(newhelp->dthps_nprovs *
15109 		    sizeof (dtrace_helper_provider_t *), KM_SLEEP);
15110 		for (i = 0; i < newhelp->dthps_nprovs; i++) {
15111 			newhelp->dthps_provs[i] = help->dthps_provs[i];
15112 			newhelp->dthps_provs[i]->dthp_ref++;
15113 		}
15114 
15115 		hasprovs = 1;
15116 	}
15117 
15118 	mutex_exit(&dtrace_lock);
15119 
15120 	if (hasprovs)
15121 		dtrace_helper_provider_register(to, newhelp, NULL);
15122 }
15123 
15124 #if defined(sun)
15125 /*
15126  * DTrace Hook Functions
15127  */
15128 static void
15129 dtrace_module_loaded(modctl_t *ctl)
15130 {
15131 	dtrace_provider_t *prv;
15132 
15133 	mutex_enter(&dtrace_provider_lock);
15134 	mutex_enter(&mod_lock);
15135 
15136 	ASSERT(ctl->mod_busy);
15137 
15138 	/*
15139 	 * We're going to call each providers per-module provide operation
15140 	 * specifying only this module.
15141 	 */
15142 	for (prv = dtrace_provider; prv != NULL; prv = prv->dtpv_next)
15143 		prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
15144 
15145 	mutex_exit(&mod_lock);
15146 	mutex_exit(&dtrace_provider_lock);
15147 
15148 	/*
15149 	 * If we have any retained enablings, we need to match against them.
15150 	 * Enabling probes requires that cpu_lock be held, and we cannot hold
15151 	 * cpu_lock here -- it is legal for cpu_lock to be held when loading a
15152 	 * module.  (In particular, this happens when loading scheduling
15153 	 * classes.)  So if we have any retained enablings, we need to dispatch
15154 	 * our task queue to do the match for us.
15155 	 */
15156 	mutex_enter(&dtrace_lock);
15157 
15158 	if (dtrace_retained == NULL) {
15159 		mutex_exit(&dtrace_lock);
15160 		return;
15161 	}
15162 
15163 	(void) taskq_dispatch(dtrace_taskq,
15164 	    (task_func_t *)dtrace_enabling_matchall, NULL, TQ_SLEEP);
15165 
15166 	mutex_exit(&dtrace_lock);
15167 
15168 	/*
15169 	 * And now, for a little heuristic sleaze:  in general, we want to
15170 	 * match modules as soon as they load.  However, we cannot guarantee
15171 	 * this, because it would lead us to the lock ordering violation
15172 	 * outlined above.  The common case, of course, is that cpu_lock is
15173 	 * _not_ held -- so we delay here for a clock tick, hoping that that's
15174 	 * long enough for the task queue to do its work.  If it's not, it's
15175 	 * not a serious problem -- it just means that the module that we
15176 	 * just loaded may not be immediately instrumentable.
15177 	 */
15178 	delay(1);
15179 }
15180 
15181 static void
15182 dtrace_module_unloaded(modctl_t *ctl)
15183 {
15184 	dtrace_probe_t template, *probe, *first, *next;
15185 	dtrace_provider_t *prov;
15186 
15187 	template.dtpr_mod = ctl->mod_modname;
15188 
15189 	mutex_enter(&dtrace_provider_lock);
15190 	mutex_enter(&mod_lock);
15191 	mutex_enter(&dtrace_lock);
15192 
15193 	if (dtrace_bymod == NULL) {
15194 		/*
15195 		 * The DTrace module is loaded (obviously) but not attached;
15196 		 * we don't have any work to do.
15197 		 */
15198 		mutex_exit(&dtrace_provider_lock);
15199 		mutex_exit(&mod_lock);
15200 		mutex_exit(&dtrace_lock);
15201 		return;
15202 	}
15203 
15204 	for (probe = first = dtrace_hash_lookup(dtrace_bymod, &template);
15205 	    probe != NULL; probe = probe->dtpr_nextmod) {
15206 		if (probe->dtpr_ecb != NULL) {
15207 			mutex_exit(&dtrace_provider_lock);
15208 			mutex_exit(&mod_lock);
15209 			mutex_exit(&dtrace_lock);
15210 
15211 			/*
15212 			 * This shouldn't _actually_ be possible -- we're
15213 			 * unloading a module that has an enabled probe in it.
15214 			 * (It's normally up to the provider to make sure that
15215 			 * this can't happen.)  However, because dtps_enable()
15216 			 * doesn't have a failure mode, there can be an
15217 			 * enable/unload race.  Upshot:  we don't want to
15218 			 * assert, but we're not going to disable the
15219 			 * probe, either.
15220 			 */
15221 			if (dtrace_err_verbose) {
15222 				cmn_err(CE_WARN, "unloaded module '%s' had "
15223 				    "enabled probes", ctl->mod_modname);
15224 			}
15225 
15226 			return;
15227 		}
15228 	}
15229 
15230 	probe = first;
15231 
15232 	for (first = NULL; probe != NULL; probe = next) {
15233 		ASSERT(dtrace_probes[probe->dtpr_id - 1] == probe);
15234 
15235 		dtrace_probes[probe->dtpr_id - 1] = NULL;
15236 
15237 		next = probe->dtpr_nextmod;
15238 		dtrace_hash_remove(dtrace_bymod, probe);
15239 		dtrace_hash_remove(dtrace_byfunc, probe);
15240 		dtrace_hash_remove(dtrace_byname, probe);
15241 
15242 		if (first == NULL) {
15243 			first = probe;
15244 			probe->dtpr_nextmod = NULL;
15245 		} else {
15246 			probe->dtpr_nextmod = first;
15247 			first = probe;
15248 		}
15249 	}
15250 
15251 	/*
15252 	 * We've removed all of the module's probes from the hash chains and
15253 	 * from the probe array.  Now issue a dtrace_sync() to be sure that
15254 	 * everyone has cleared out from any probe array processing.
15255 	 */
15256 	dtrace_sync();
15257 
15258 	for (probe = first; probe != NULL; probe = first) {
15259 		first = probe->dtpr_nextmod;
15260 		prov = probe->dtpr_provider;
15261 		prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, probe->dtpr_id,
15262 		    probe->dtpr_arg);
15263 		kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
15264 		kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
15265 		kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
15266 		vmem_free(dtrace_arena, (void *)(uintptr_t)probe->dtpr_id, 1);
15267 		kmem_free(probe, sizeof (dtrace_probe_t));
15268 	}
15269 
15270 	mutex_exit(&dtrace_lock);
15271 	mutex_exit(&mod_lock);
15272 	mutex_exit(&dtrace_provider_lock);
15273 }
15274 
15275 static void
15276 dtrace_suspend(void)
15277 {
15278 	dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_suspend));
15279 }
15280 
15281 static void
15282 dtrace_resume(void)
15283 {
15284 	dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_resume));
15285 }
15286 #endif
15287 
15288 static int
15289 dtrace_cpu_setup(cpu_setup_t what, processorid_t cpu)
15290 {
15291 	ASSERT(MUTEX_HELD(&cpu_lock));
15292 	mutex_enter(&dtrace_lock);
15293 
15294 	switch (what) {
15295 	case CPU_CONFIG: {
15296 		dtrace_state_t *state;
15297 		dtrace_optval_t *opt, rs, c;
15298 
15299 		/*
15300 		 * For now, we only allocate a new buffer for anonymous state.
15301 		 */
15302 		if ((state = dtrace_anon.dta_state) == NULL)
15303 			break;
15304 
15305 		if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
15306 			break;
15307 
15308 		opt = state->dts_options;
15309 		c = opt[DTRACEOPT_CPU];
15310 
15311 		if (c != DTRACE_CPUALL && c != DTRACEOPT_UNSET && c != cpu)
15312 			break;
15313 
15314 		/*
15315 		 * Regardless of what the actual policy is, we're going to
15316 		 * temporarily set our resize policy to be manual.  We're
15317 		 * also going to temporarily set our CPU option to denote
15318 		 * the newly configured CPU.
15319 		 */
15320 		rs = opt[DTRACEOPT_BUFRESIZE];
15321 		opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_MANUAL;
15322 		opt[DTRACEOPT_CPU] = (dtrace_optval_t)cpu;
15323 
15324 		(void) dtrace_state_buffers(state);
15325 
15326 		opt[DTRACEOPT_BUFRESIZE] = rs;
15327 		opt[DTRACEOPT_CPU] = c;
15328 
15329 		break;
15330 	}
15331 
15332 	case CPU_UNCONFIG:
15333 		/*
15334 		 * We don't free the buffer in the CPU_UNCONFIG case.  (The
15335 		 * buffer will be freed when the consumer exits.)
15336 		 */
15337 		break;
15338 
15339 	default:
15340 		break;
15341 	}
15342 
15343 	mutex_exit(&dtrace_lock);
15344 	return (0);
15345 }
15346 
15347 #if defined(sun)
15348 static void
15349 dtrace_cpu_setup_initial(processorid_t cpu)
15350 {
15351 	(void) dtrace_cpu_setup(CPU_CONFIG, cpu);
15352 }
15353 #endif
15354 
15355 static void
15356 dtrace_toxrange_add(uintptr_t base, uintptr_t limit)
15357 {
15358 	if (dtrace_toxranges >= dtrace_toxranges_max) {
15359 		int osize, nsize;
15360 		dtrace_toxrange_t *range;
15361 
15362 		osize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
15363 
15364 		if (osize == 0) {
15365 			ASSERT(dtrace_toxrange == NULL);
15366 			ASSERT(dtrace_toxranges_max == 0);
15367 			dtrace_toxranges_max = 1;
15368 		} else {
15369 			dtrace_toxranges_max <<= 1;
15370 		}
15371 
15372 		nsize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
15373 		range = kmem_zalloc(nsize, KM_SLEEP);
15374 
15375 		if (dtrace_toxrange != NULL) {
15376 			ASSERT(osize != 0);
15377 			bcopy(dtrace_toxrange, range, osize);
15378 			kmem_free(dtrace_toxrange, osize);
15379 		}
15380 
15381 		dtrace_toxrange = range;
15382 	}
15383 
15384 	ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_base == 0);
15385 	ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_limit == 0);
15386 
15387 	dtrace_toxrange[dtrace_toxranges].dtt_base = base;
15388 	dtrace_toxrange[dtrace_toxranges].dtt_limit = limit;
15389 	dtrace_toxranges++;
15390 }
15391 
15392 /*
15393  * DTrace Driver Cookbook Functions
15394  */
15395 #if defined(sun)
15396 /*ARGSUSED*/
15397 static int
15398 dtrace_attach(dev_info_t *devi, ddi_attach_cmd_t cmd)
15399 {
15400 	dtrace_provider_id_t id;
15401 	dtrace_state_t *state = NULL;
15402 	dtrace_enabling_t *enab;
15403 
15404 	mutex_enter(&cpu_lock);
15405 	mutex_enter(&dtrace_provider_lock);
15406 	mutex_enter(&dtrace_lock);
15407 
15408 	if (ddi_soft_state_init(&dtrace_softstate,
15409 	    sizeof (dtrace_state_t), 0) != 0) {
15410 		cmn_err(CE_NOTE, "/dev/dtrace failed to initialize soft state");
15411 		mutex_exit(&cpu_lock);
15412 		mutex_exit(&dtrace_provider_lock);
15413 		mutex_exit(&dtrace_lock);
15414 		return (DDI_FAILURE);
15415 	}
15416 
15417 	if (ddi_create_minor_node(devi, DTRACEMNR_DTRACE, S_IFCHR,
15418 	    DTRACEMNRN_DTRACE, DDI_PSEUDO, NULL) == DDI_FAILURE ||
15419 	    ddi_create_minor_node(devi, DTRACEMNR_HELPER, S_IFCHR,
15420 	    DTRACEMNRN_HELPER, DDI_PSEUDO, NULL) == DDI_FAILURE) {
15421 		cmn_err(CE_NOTE, "/dev/dtrace couldn't create minor nodes");
15422 		ddi_remove_minor_node(devi, NULL);
15423 		ddi_soft_state_fini(&dtrace_softstate);
15424 		mutex_exit(&cpu_lock);
15425 		mutex_exit(&dtrace_provider_lock);
15426 		mutex_exit(&dtrace_lock);
15427 		return (DDI_FAILURE);
15428 	}
15429 
15430 	ddi_report_dev(devi);
15431 	dtrace_devi = devi;
15432 
15433 	dtrace_modload = dtrace_module_loaded;
15434 	dtrace_modunload = dtrace_module_unloaded;
15435 	dtrace_cpu_init = dtrace_cpu_setup_initial;
15436 	dtrace_helpers_cleanup = dtrace_helpers_destroy;
15437 	dtrace_helpers_fork = dtrace_helpers_duplicate;
15438 	dtrace_cpustart_init = dtrace_suspend;
15439 	dtrace_cpustart_fini = dtrace_resume;
15440 	dtrace_debugger_init = dtrace_suspend;
15441 	dtrace_debugger_fini = dtrace_resume;
15442 
15443 	register_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
15444 
15445 	ASSERT(MUTEX_HELD(&cpu_lock));
15446 
15447 	dtrace_arena = vmem_create("dtrace", (void *)1, UINT32_MAX, 1,
15448 	    NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
15449 	dtrace_minor = vmem_create("dtrace_minor", (void *)DTRACEMNRN_CLONE,
15450 	    UINT32_MAX - DTRACEMNRN_CLONE, 1, NULL, NULL, NULL, 0,
15451 	    VM_SLEEP | VMC_IDENTIFIER);
15452 	dtrace_taskq = taskq_create("dtrace_taskq", 1, maxclsyspri,
15453 	    1, INT_MAX, 0);
15454 
15455 	dtrace_state_cache = kmem_cache_create("dtrace_state_cache",
15456 	    sizeof (dtrace_dstate_percpu_t) * NCPU, DTRACE_STATE_ALIGN,
15457 	    NULL, NULL, NULL, NULL, NULL, 0);
15458 
15459 	ASSERT(MUTEX_HELD(&cpu_lock));
15460 	dtrace_bymod = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_mod),
15461 	    offsetof(dtrace_probe_t, dtpr_nextmod),
15462 	    offsetof(dtrace_probe_t, dtpr_prevmod));
15463 
15464 	dtrace_byfunc = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_func),
15465 	    offsetof(dtrace_probe_t, dtpr_nextfunc),
15466 	    offsetof(dtrace_probe_t, dtpr_prevfunc));
15467 
15468 	dtrace_byname = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_name),
15469 	    offsetof(dtrace_probe_t, dtpr_nextname),
15470 	    offsetof(dtrace_probe_t, dtpr_prevname));
15471 
15472 	if (dtrace_retain_max < 1) {
15473 		cmn_err(CE_WARN, "illegal value (%lu) for dtrace_retain_max; "
15474 		    "setting to 1", dtrace_retain_max);
15475 		dtrace_retain_max = 1;
15476 	}
15477 
15478 	/*
15479 	 * Now discover our toxic ranges.
15480 	 */
15481 	dtrace_toxic_ranges(dtrace_toxrange_add);
15482 
15483 	/*
15484 	 * Before we register ourselves as a provider to our own framework,
15485 	 * we would like to assert that dtrace_provider is NULL -- but that's
15486 	 * not true if we were loaded as a dependency of a DTrace provider.
15487 	 * Once we've registered, we can assert that dtrace_provider is our
15488 	 * pseudo provider.
15489 	 */
15490 	(void) dtrace_register("dtrace", &dtrace_provider_attr,
15491 	    DTRACE_PRIV_NONE, 0, &dtrace_provider_ops, NULL, &id);
15492 
15493 	ASSERT(dtrace_provider != NULL);
15494 	ASSERT((dtrace_provider_id_t)dtrace_provider == id);
15495 
15496 	dtrace_probeid_begin = dtrace_probe_create((dtrace_provider_id_t)
15497 	    dtrace_provider, NULL, NULL, "BEGIN", 0, NULL);
15498 	dtrace_probeid_end = dtrace_probe_create((dtrace_provider_id_t)
15499 	    dtrace_provider, NULL, NULL, "END", 0, NULL);
15500 	dtrace_probeid_error = dtrace_probe_create((dtrace_provider_id_t)
15501 	    dtrace_provider, NULL, NULL, "ERROR", 1, NULL);
15502 
15503 	dtrace_anon_property();
15504 	mutex_exit(&cpu_lock);
15505 
15506 	/*
15507 	 * If DTrace helper tracing is enabled, we need to allocate the
15508 	 * trace buffer and initialize the values.
15509 	 */
15510 	if (dtrace_helptrace_enabled) {
15511 		ASSERT(dtrace_helptrace_buffer == NULL);
15512 		dtrace_helptrace_buffer =
15513 		    kmem_zalloc(dtrace_helptrace_bufsize, KM_SLEEP);
15514 		dtrace_helptrace_next = 0;
15515 	}
15516 
15517 	/*
15518 	 * If there are already providers, we must ask them to provide their
15519 	 * probes, and then match any anonymous enabling against them.  Note
15520 	 * that there should be no other retained enablings at this time:
15521 	 * the only retained enablings at this time should be the anonymous
15522 	 * enabling.
15523 	 */
15524 	if (dtrace_anon.dta_enabling != NULL) {
15525 		ASSERT(dtrace_retained == dtrace_anon.dta_enabling);
15526 
15527 		dtrace_enabling_provide(NULL);
15528 		state = dtrace_anon.dta_state;
15529 
15530 		/*
15531 		 * We couldn't hold cpu_lock across the above call to
15532 		 * dtrace_enabling_provide(), but we must hold it to actually
15533 		 * enable the probes.  We have to drop all of our locks, pick
15534 		 * up cpu_lock, and regain our locks before matching the
15535 		 * retained anonymous enabling.
15536 		 */
15537 		mutex_exit(&dtrace_lock);
15538 		mutex_exit(&dtrace_provider_lock);
15539 
15540 		mutex_enter(&cpu_lock);
15541 		mutex_enter(&dtrace_provider_lock);
15542 		mutex_enter(&dtrace_lock);
15543 
15544 		if ((enab = dtrace_anon.dta_enabling) != NULL)
15545 			(void) dtrace_enabling_match(enab, NULL);
15546 
15547 		mutex_exit(&cpu_lock);
15548 	}
15549 
15550 	mutex_exit(&dtrace_lock);
15551 	mutex_exit(&dtrace_provider_lock);
15552 
15553 	if (state != NULL) {
15554 		/*
15555 		 * If we created any anonymous state, set it going now.
15556 		 */
15557 		(void) dtrace_state_go(state, &dtrace_anon.dta_beganon);
15558 	}
15559 
15560 	return (DDI_SUCCESS);
15561 }
15562 #endif
15563 
15564 #if !defined(sun)
15565 #if __FreeBSD_version >= 800039
15566 static void dtrace_dtr(void *);
15567 #endif
15568 #endif
15569 
15570 /*ARGSUSED*/
15571 static int
15572 #if defined(sun)
15573 dtrace_open(dev_t *devp, int flag, int otyp, cred_t *cred_p)
15574 #else
15575 dtrace_open(struct cdev *dev, int oflags, int devtype, struct thread *td)
15576 #endif
15577 {
15578 	dtrace_state_t *state;
15579 	uint32_t priv;
15580 	uid_t uid;
15581 	zoneid_t zoneid;
15582 
15583 #if defined(sun)
15584 	if (getminor(*devp) == DTRACEMNRN_HELPER)
15585 		return (0);
15586 
15587 	/*
15588 	 * If this wasn't an open with the "helper" minor, then it must be
15589 	 * the "dtrace" minor.
15590 	 */
15591 	ASSERT(getminor(*devp) == DTRACEMNRN_DTRACE);
15592 #else
15593 	cred_t *cred_p = NULL;
15594 
15595 #if __FreeBSD_version < 800039
15596 	/*
15597 	 * The first minor device is the one that is cloned so there is
15598 	 * nothing more to do here.
15599 	 */
15600 	if (dev2unit(dev) == 0)
15601 		return 0;
15602 
15603 	/*
15604 	 * Devices are cloned, so if the DTrace state has already
15605 	 * been allocated, that means this device belongs to a
15606 	 * different client. Each client should open '/dev/dtrace'
15607 	 * to get a cloned device.
15608 	 */
15609 	if (dev->si_drv1 != NULL)
15610 		return (EBUSY);
15611 #endif
15612 
15613 	cred_p = dev->si_cred;
15614 #endif
15615 
15616 	/*
15617 	 * If no DTRACE_PRIV_* bits are set in the credential, then the
15618 	 * caller lacks sufficient permission to do anything with DTrace.
15619 	 */
15620 	dtrace_cred2priv(cred_p, &priv, &uid, &zoneid);
15621 	if (priv == DTRACE_PRIV_NONE) {
15622 #if !defined(sun)
15623 #if __FreeBSD_version < 800039
15624 		/* Destroy the cloned device. */
15625                 destroy_dev(dev);
15626 #endif
15627 #endif
15628 
15629 		return (EACCES);
15630 	}
15631 
15632 	/*
15633 	 * Ask all providers to provide all their probes.
15634 	 */
15635 	mutex_enter(&dtrace_provider_lock);
15636 	dtrace_probe_provide(NULL, NULL);
15637 	mutex_exit(&dtrace_provider_lock);
15638 
15639 	mutex_enter(&cpu_lock);
15640 	mutex_enter(&dtrace_lock);
15641 	dtrace_opens++;
15642 	dtrace_membar_producer();
15643 
15644 #if defined(sun)
15645 	/*
15646 	 * If the kernel debugger is active (that is, if the kernel debugger
15647 	 * modified text in some way), we won't allow the open.
15648 	 */
15649 	if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
15650 		dtrace_opens--;
15651 		mutex_exit(&cpu_lock);
15652 		mutex_exit(&dtrace_lock);
15653 		return (EBUSY);
15654 	}
15655 
15656 	state = dtrace_state_create(devp, cred_p);
15657 #else
15658 	state = dtrace_state_create(dev);
15659 #if __FreeBSD_version < 800039
15660 	dev->si_drv1 = state;
15661 #else
15662 	devfs_set_cdevpriv(state, dtrace_dtr);
15663 #endif
15664 	/* This code actually belongs in dtrace_attach() */
15665 	if (dtrace_opens == 1)
15666 		dtrace_taskq = taskq_create("dtrace_taskq", 1, maxclsyspri,
15667 		    1, INT_MAX, 0);
15668 #endif
15669 
15670 	mutex_exit(&cpu_lock);
15671 
15672 	if (state == NULL) {
15673 #if defined(sun)
15674 		if (--dtrace_opens == 0)
15675 			(void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
15676 #else
15677 		--dtrace_opens;
15678 #endif
15679 		mutex_exit(&dtrace_lock);
15680 #if !defined(sun)
15681 #if __FreeBSD_version < 800039
15682 		/* Destroy the cloned device. */
15683                 destroy_dev(dev);
15684 #endif
15685 #endif
15686 		return (EAGAIN);
15687 	}
15688 
15689 	mutex_exit(&dtrace_lock);
15690 
15691 	return (0);
15692 }
15693 
15694 /*ARGSUSED*/
15695 #if defined(sun)
15696 static int
15697 dtrace_close(dev_t dev, int flag, int otyp, cred_t *cred_p)
15698 #elif __FreeBSD_version < 800039
15699 static int
15700 dtrace_close(struct cdev *dev, int flags, int fmt __unused, struct thread *td)
15701 #else
15702 static void
15703 dtrace_dtr(void *data)
15704 #endif
15705 {
15706 #if defined(sun)
15707 	minor_t minor = getminor(dev);
15708 	dtrace_state_t *state;
15709 
15710 	if (minor == DTRACEMNRN_HELPER)
15711 		return (0);
15712 
15713 	state = ddi_get_soft_state(dtrace_softstate, minor);
15714 #else
15715 #if __FreeBSD_version < 800039
15716 	dtrace_state_t *state = dev->si_drv1;
15717 
15718 	/* Check if this is not a cloned device. */
15719 	if (dev2unit(dev) == 0)
15720 		return (0);
15721 #else
15722 	dtrace_state_t *state = data;
15723 #endif
15724 
15725 #endif
15726 
15727 	mutex_enter(&cpu_lock);
15728 	mutex_enter(&dtrace_lock);
15729 
15730 	if (state != NULL) {
15731 		if (state->dts_anon) {
15732 			/*
15733 			 * There is anonymous state. Destroy that first.
15734 			 */
15735 			ASSERT(dtrace_anon.dta_state == NULL);
15736 			dtrace_state_destroy(state->dts_anon);
15737 		}
15738 
15739 		dtrace_state_destroy(state);
15740 
15741 #if !defined(sun)
15742 		kmem_free(state, 0);
15743 #if __FreeBSD_version < 800039
15744 		dev->si_drv1 = NULL;
15745 #endif
15746 #endif
15747 	}
15748 
15749 	ASSERT(dtrace_opens > 0);
15750 #if defined(sun)
15751 	if (--dtrace_opens == 0)
15752 		(void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
15753 #else
15754 	--dtrace_opens;
15755 	/* This code actually belongs in dtrace_detach() */
15756 	if ((dtrace_opens == 0) && (dtrace_taskq != NULL)) {
15757 		taskq_destroy(dtrace_taskq);
15758 		dtrace_taskq = NULL;
15759 	}
15760 #endif
15761 
15762 	mutex_exit(&dtrace_lock);
15763 	mutex_exit(&cpu_lock);
15764 
15765 #if __FreeBSD_version < 800039
15766 	/* Schedule this cloned device to be destroyed. */
15767 	destroy_dev_sched(dev);
15768 #endif
15769 
15770 #if defined(sun) || __FreeBSD_version < 800039
15771 	return (0);
15772 #endif
15773 }
15774 
15775 #if defined(sun)
15776 /*ARGSUSED*/
15777 static int
15778 dtrace_ioctl_helper(int cmd, intptr_t arg, int *rv)
15779 {
15780 	int rval;
15781 	dof_helper_t help, *dhp = NULL;
15782 
15783 	switch (cmd) {
15784 	case DTRACEHIOC_ADDDOF:
15785 		if (copyin((void *)arg, &help, sizeof (help)) != 0) {
15786 			dtrace_dof_error(NULL, "failed to copyin DOF helper");
15787 			return (EFAULT);
15788 		}
15789 
15790 		dhp = &help;
15791 		arg = (intptr_t)help.dofhp_dof;
15792 		/*FALLTHROUGH*/
15793 
15794 	case DTRACEHIOC_ADD: {
15795 		dof_hdr_t *dof = dtrace_dof_copyin(arg, &rval);
15796 
15797 		if (dof == NULL)
15798 			return (rval);
15799 
15800 		mutex_enter(&dtrace_lock);
15801 
15802 		/*
15803 		 * dtrace_helper_slurp() takes responsibility for the dof --
15804 		 * it may free it now or it may save it and free it later.
15805 		 */
15806 		if ((rval = dtrace_helper_slurp(dof, dhp)) != -1) {
15807 			*rv = rval;
15808 			rval = 0;
15809 		} else {
15810 			rval = EINVAL;
15811 		}
15812 
15813 		mutex_exit(&dtrace_lock);
15814 		return (rval);
15815 	}
15816 
15817 	case DTRACEHIOC_REMOVE: {
15818 		mutex_enter(&dtrace_lock);
15819 		rval = dtrace_helper_destroygen(arg);
15820 		mutex_exit(&dtrace_lock);
15821 
15822 		return (rval);
15823 	}
15824 
15825 	default:
15826 		break;
15827 	}
15828 
15829 	return (ENOTTY);
15830 }
15831 
15832 /*ARGSUSED*/
15833 static int
15834 dtrace_ioctl(dev_t dev, int cmd, intptr_t arg, int md, cred_t *cr, int *rv)
15835 {
15836 	minor_t minor = getminor(dev);
15837 	dtrace_state_t *state;
15838 	int rval;
15839 
15840 	if (minor == DTRACEMNRN_HELPER)
15841 		return (dtrace_ioctl_helper(cmd, arg, rv));
15842 
15843 	state = ddi_get_soft_state(dtrace_softstate, minor);
15844 
15845 	if (state->dts_anon) {
15846 		ASSERT(dtrace_anon.dta_state == NULL);
15847 		state = state->dts_anon;
15848 	}
15849 
15850 	switch (cmd) {
15851 	case DTRACEIOC_PROVIDER: {
15852 		dtrace_providerdesc_t pvd;
15853 		dtrace_provider_t *pvp;
15854 
15855 		if (copyin((void *)arg, &pvd, sizeof (pvd)) != 0)
15856 			return (EFAULT);
15857 
15858 		pvd.dtvd_name[DTRACE_PROVNAMELEN - 1] = '\0';
15859 		mutex_enter(&dtrace_provider_lock);
15860 
15861 		for (pvp = dtrace_provider; pvp != NULL; pvp = pvp->dtpv_next) {
15862 			if (strcmp(pvp->dtpv_name, pvd.dtvd_name) == 0)
15863 				break;
15864 		}
15865 
15866 		mutex_exit(&dtrace_provider_lock);
15867 
15868 		if (pvp == NULL)
15869 			return (ESRCH);
15870 
15871 		bcopy(&pvp->dtpv_priv, &pvd.dtvd_priv, sizeof (dtrace_ppriv_t));
15872 		bcopy(&pvp->dtpv_attr, &pvd.dtvd_attr, sizeof (dtrace_pattr_t));
15873 
15874 		if (copyout(&pvd, (void *)arg, sizeof (pvd)) != 0)
15875 			return (EFAULT);
15876 
15877 		return (0);
15878 	}
15879 
15880 	case DTRACEIOC_EPROBE: {
15881 		dtrace_eprobedesc_t epdesc;
15882 		dtrace_ecb_t *ecb;
15883 		dtrace_action_t *act;
15884 		void *buf;
15885 		size_t size;
15886 		uintptr_t dest;
15887 		int nrecs;
15888 
15889 		if (copyin((void *)arg, &epdesc, sizeof (epdesc)) != 0)
15890 			return (EFAULT);
15891 
15892 		mutex_enter(&dtrace_lock);
15893 
15894 		if ((ecb = dtrace_epid2ecb(state, epdesc.dtepd_epid)) == NULL) {
15895 			mutex_exit(&dtrace_lock);
15896 			return (EINVAL);
15897 		}
15898 
15899 		if (ecb->dte_probe == NULL) {
15900 			mutex_exit(&dtrace_lock);
15901 			return (EINVAL);
15902 		}
15903 
15904 		epdesc.dtepd_probeid = ecb->dte_probe->dtpr_id;
15905 		epdesc.dtepd_uarg = ecb->dte_uarg;
15906 		epdesc.dtepd_size = ecb->dte_size;
15907 
15908 		nrecs = epdesc.dtepd_nrecs;
15909 		epdesc.dtepd_nrecs = 0;
15910 		for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
15911 			if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
15912 				continue;
15913 
15914 			epdesc.dtepd_nrecs++;
15915 		}
15916 
15917 		/*
15918 		 * Now that we have the size, we need to allocate a temporary
15919 		 * buffer in which to store the complete description.  We need
15920 		 * the temporary buffer to be able to drop dtrace_lock()
15921 		 * across the copyout(), below.
15922 		 */
15923 		size = sizeof (dtrace_eprobedesc_t) +
15924 		    (epdesc.dtepd_nrecs * sizeof (dtrace_recdesc_t));
15925 
15926 		buf = kmem_alloc(size, KM_SLEEP);
15927 		dest = (uintptr_t)buf;
15928 
15929 		bcopy(&epdesc, (void *)dest, sizeof (epdesc));
15930 		dest += offsetof(dtrace_eprobedesc_t, dtepd_rec[0]);
15931 
15932 		for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
15933 			if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
15934 				continue;
15935 
15936 			if (nrecs-- == 0)
15937 				break;
15938 
15939 			bcopy(&act->dta_rec, (void *)dest,
15940 			    sizeof (dtrace_recdesc_t));
15941 			dest += sizeof (dtrace_recdesc_t);
15942 		}
15943 
15944 		mutex_exit(&dtrace_lock);
15945 
15946 		if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
15947 			kmem_free(buf, size);
15948 			return (EFAULT);
15949 		}
15950 
15951 		kmem_free(buf, size);
15952 		return (0);
15953 	}
15954 
15955 	case DTRACEIOC_AGGDESC: {
15956 		dtrace_aggdesc_t aggdesc;
15957 		dtrace_action_t *act;
15958 		dtrace_aggregation_t *agg;
15959 		int nrecs;
15960 		uint32_t offs;
15961 		dtrace_recdesc_t *lrec;
15962 		void *buf;
15963 		size_t size;
15964 		uintptr_t dest;
15965 
15966 		if (copyin((void *)arg, &aggdesc, sizeof (aggdesc)) != 0)
15967 			return (EFAULT);
15968 
15969 		mutex_enter(&dtrace_lock);
15970 
15971 		if ((agg = dtrace_aggid2agg(state, aggdesc.dtagd_id)) == NULL) {
15972 			mutex_exit(&dtrace_lock);
15973 			return (EINVAL);
15974 		}
15975 
15976 		aggdesc.dtagd_epid = agg->dtag_ecb->dte_epid;
15977 
15978 		nrecs = aggdesc.dtagd_nrecs;
15979 		aggdesc.dtagd_nrecs = 0;
15980 
15981 		offs = agg->dtag_base;
15982 		lrec = &agg->dtag_action.dta_rec;
15983 		aggdesc.dtagd_size = lrec->dtrd_offset + lrec->dtrd_size - offs;
15984 
15985 		for (act = agg->dtag_first; ; act = act->dta_next) {
15986 			ASSERT(act->dta_intuple ||
15987 			    DTRACEACT_ISAGG(act->dta_kind));
15988 
15989 			/*
15990 			 * If this action has a record size of zero, it
15991 			 * denotes an argument to the aggregating action.
15992 			 * Because the presence of this record doesn't (or
15993 			 * shouldn't) affect the way the data is interpreted,
15994 			 * we don't copy it out to save user-level the
15995 			 * confusion of dealing with a zero-length record.
15996 			 */
15997 			if (act->dta_rec.dtrd_size == 0) {
15998 				ASSERT(agg->dtag_hasarg);
15999 				continue;
16000 			}
16001 
16002 			aggdesc.dtagd_nrecs++;
16003 
16004 			if (act == &agg->dtag_action)
16005 				break;
16006 		}
16007 
16008 		/*
16009 		 * Now that we have the size, we need to allocate a temporary
16010 		 * buffer in which to store the complete description.  We need
16011 		 * the temporary buffer to be able to drop dtrace_lock()
16012 		 * across the copyout(), below.
16013 		 */
16014 		size = sizeof (dtrace_aggdesc_t) +
16015 		    (aggdesc.dtagd_nrecs * sizeof (dtrace_recdesc_t));
16016 
16017 		buf = kmem_alloc(size, KM_SLEEP);
16018 		dest = (uintptr_t)buf;
16019 
16020 		bcopy(&aggdesc, (void *)dest, sizeof (aggdesc));
16021 		dest += offsetof(dtrace_aggdesc_t, dtagd_rec[0]);
16022 
16023 		for (act = agg->dtag_first; ; act = act->dta_next) {
16024 			dtrace_recdesc_t rec = act->dta_rec;
16025 
16026 			/*
16027 			 * See the comment in the above loop for why we pass
16028 			 * over zero-length records.
16029 			 */
16030 			if (rec.dtrd_size == 0) {
16031 				ASSERT(agg->dtag_hasarg);
16032 				continue;
16033 			}
16034 
16035 			if (nrecs-- == 0)
16036 				break;
16037 
16038 			rec.dtrd_offset -= offs;
16039 			bcopy(&rec, (void *)dest, sizeof (rec));
16040 			dest += sizeof (dtrace_recdesc_t);
16041 
16042 			if (act == &agg->dtag_action)
16043 				break;
16044 		}
16045 
16046 		mutex_exit(&dtrace_lock);
16047 
16048 		if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
16049 			kmem_free(buf, size);
16050 			return (EFAULT);
16051 		}
16052 
16053 		kmem_free(buf, size);
16054 		return (0);
16055 	}
16056 
16057 	case DTRACEIOC_ENABLE: {
16058 		dof_hdr_t *dof;
16059 		dtrace_enabling_t *enab = NULL;
16060 		dtrace_vstate_t *vstate;
16061 		int err = 0;
16062 
16063 		*rv = 0;
16064 
16065 		/*
16066 		 * If a NULL argument has been passed, we take this as our
16067 		 * cue to reevaluate our enablings.
16068 		 */
16069 		if (arg == NULL) {
16070 			dtrace_enabling_matchall();
16071 
16072 			return (0);
16073 		}
16074 
16075 		if ((dof = dtrace_dof_copyin(arg, &rval)) == NULL)
16076 			return (rval);
16077 
16078 		mutex_enter(&cpu_lock);
16079 		mutex_enter(&dtrace_lock);
16080 		vstate = &state->dts_vstate;
16081 
16082 		if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
16083 			mutex_exit(&dtrace_lock);
16084 			mutex_exit(&cpu_lock);
16085 			dtrace_dof_destroy(dof);
16086 			return (EBUSY);
16087 		}
16088 
16089 		if (dtrace_dof_slurp(dof, vstate, cr, &enab, 0, B_TRUE) != 0) {
16090 			mutex_exit(&dtrace_lock);
16091 			mutex_exit(&cpu_lock);
16092 			dtrace_dof_destroy(dof);
16093 			return (EINVAL);
16094 		}
16095 
16096 		if ((rval = dtrace_dof_options(dof, state)) != 0) {
16097 			dtrace_enabling_destroy(enab);
16098 			mutex_exit(&dtrace_lock);
16099 			mutex_exit(&cpu_lock);
16100 			dtrace_dof_destroy(dof);
16101 			return (rval);
16102 		}
16103 
16104 		if ((err = dtrace_enabling_match(enab, rv)) == 0) {
16105 			err = dtrace_enabling_retain(enab);
16106 		} else {
16107 			dtrace_enabling_destroy(enab);
16108 		}
16109 
16110 		mutex_exit(&cpu_lock);
16111 		mutex_exit(&dtrace_lock);
16112 		dtrace_dof_destroy(dof);
16113 
16114 		return (err);
16115 	}
16116 
16117 	case DTRACEIOC_REPLICATE: {
16118 		dtrace_repldesc_t desc;
16119 		dtrace_probedesc_t *match = &desc.dtrpd_match;
16120 		dtrace_probedesc_t *create = &desc.dtrpd_create;
16121 		int err;
16122 
16123 		if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16124 			return (EFAULT);
16125 
16126 		match->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
16127 		match->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
16128 		match->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
16129 		match->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
16130 
16131 		create->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
16132 		create->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
16133 		create->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
16134 		create->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
16135 
16136 		mutex_enter(&dtrace_lock);
16137 		err = dtrace_enabling_replicate(state, match, create);
16138 		mutex_exit(&dtrace_lock);
16139 
16140 		return (err);
16141 	}
16142 
16143 	case DTRACEIOC_PROBEMATCH:
16144 	case DTRACEIOC_PROBES: {
16145 		dtrace_probe_t *probe = NULL;
16146 		dtrace_probedesc_t desc;
16147 		dtrace_probekey_t pkey;
16148 		dtrace_id_t i;
16149 		int m = 0;
16150 		uint32_t priv;
16151 		uid_t uid;
16152 		zoneid_t zoneid;
16153 
16154 		if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16155 			return (EFAULT);
16156 
16157 		desc.dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
16158 		desc.dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
16159 		desc.dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
16160 		desc.dtpd_name[DTRACE_NAMELEN - 1] = '\0';
16161 
16162 		/*
16163 		 * Before we attempt to match this probe, we want to give
16164 		 * all providers the opportunity to provide it.
16165 		 */
16166 		if (desc.dtpd_id == DTRACE_IDNONE) {
16167 			mutex_enter(&dtrace_provider_lock);
16168 			dtrace_probe_provide(&desc, NULL);
16169 			mutex_exit(&dtrace_provider_lock);
16170 			desc.dtpd_id++;
16171 		}
16172 
16173 		if (cmd == DTRACEIOC_PROBEMATCH)  {
16174 			dtrace_probekey(&desc, &pkey);
16175 			pkey.dtpk_id = DTRACE_IDNONE;
16176 		}
16177 
16178 		dtrace_cred2priv(cr, &priv, &uid, &zoneid);
16179 
16180 		mutex_enter(&dtrace_lock);
16181 
16182 		if (cmd == DTRACEIOC_PROBEMATCH) {
16183 			for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
16184 				if ((probe = dtrace_probes[i - 1]) != NULL &&
16185 				    (m = dtrace_match_probe(probe, &pkey,
16186 				    priv, uid, zoneid)) != 0)
16187 					break;
16188 			}
16189 
16190 			if (m < 0) {
16191 				mutex_exit(&dtrace_lock);
16192 				return (EINVAL);
16193 			}
16194 
16195 		} else {
16196 			for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
16197 				if ((probe = dtrace_probes[i - 1]) != NULL &&
16198 				    dtrace_match_priv(probe, priv, uid, zoneid))
16199 					break;
16200 			}
16201 		}
16202 
16203 		if (probe == NULL) {
16204 			mutex_exit(&dtrace_lock);
16205 			return (ESRCH);
16206 		}
16207 
16208 		dtrace_probe_description(probe, &desc);
16209 		mutex_exit(&dtrace_lock);
16210 
16211 		if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16212 			return (EFAULT);
16213 
16214 		return (0);
16215 	}
16216 
16217 	case DTRACEIOC_PROBEARG: {
16218 		dtrace_argdesc_t desc;
16219 		dtrace_probe_t *probe;
16220 		dtrace_provider_t *prov;
16221 
16222 		if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16223 			return (EFAULT);
16224 
16225 		if (desc.dtargd_id == DTRACE_IDNONE)
16226 			return (EINVAL);
16227 
16228 		if (desc.dtargd_ndx == DTRACE_ARGNONE)
16229 			return (EINVAL);
16230 
16231 		mutex_enter(&dtrace_provider_lock);
16232 		mutex_enter(&mod_lock);
16233 		mutex_enter(&dtrace_lock);
16234 
16235 		if (desc.dtargd_id > dtrace_nprobes) {
16236 			mutex_exit(&dtrace_lock);
16237 			mutex_exit(&mod_lock);
16238 			mutex_exit(&dtrace_provider_lock);
16239 			return (EINVAL);
16240 		}
16241 
16242 		if ((probe = dtrace_probes[desc.dtargd_id - 1]) == NULL) {
16243 			mutex_exit(&dtrace_lock);
16244 			mutex_exit(&mod_lock);
16245 			mutex_exit(&dtrace_provider_lock);
16246 			return (EINVAL);
16247 		}
16248 
16249 		mutex_exit(&dtrace_lock);
16250 
16251 		prov = probe->dtpr_provider;
16252 
16253 		if (prov->dtpv_pops.dtps_getargdesc == NULL) {
16254 			/*
16255 			 * There isn't any typed information for this probe.
16256 			 * Set the argument number to DTRACE_ARGNONE.
16257 			 */
16258 			desc.dtargd_ndx = DTRACE_ARGNONE;
16259 		} else {
16260 			desc.dtargd_native[0] = '\0';
16261 			desc.dtargd_xlate[0] = '\0';
16262 			desc.dtargd_mapping = desc.dtargd_ndx;
16263 
16264 			prov->dtpv_pops.dtps_getargdesc(prov->dtpv_arg,
16265 			    probe->dtpr_id, probe->dtpr_arg, &desc);
16266 		}
16267 
16268 		mutex_exit(&mod_lock);
16269 		mutex_exit(&dtrace_provider_lock);
16270 
16271 		if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16272 			return (EFAULT);
16273 
16274 		return (0);
16275 	}
16276 
16277 	case DTRACEIOC_GO: {
16278 		processorid_t cpuid;
16279 		rval = dtrace_state_go(state, &cpuid);
16280 
16281 		if (rval != 0)
16282 			return (rval);
16283 
16284 		if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
16285 			return (EFAULT);
16286 
16287 		return (0);
16288 	}
16289 
16290 	case DTRACEIOC_STOP: {
16291 		processorid_t cpuid;
16292 
16293 		mutex_enter(&dtrace_lock);
16294 		rval = dtrace_state_stop(state, &cpuid);
16295 		mutex_exit(&dtrace_lock);
16296 
16297 		if (rval != 0)
16298 			return (rval);
16299 
16300 		if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
16301 			return (EFAULT);
16302 
16303 		return (0);
16304 	}
16305 
16306 	case DTRACEIOC_DOFGET: {
16307 		dof_hdr_t hdr, *dof;
16308 		uint64_t len;
16309 
16310 		if (copyin((void *)arg, &hdr, sizeof (hdr)) != 0)
16311 			return (EFAULT);
16312 
16313 		mutex_enter(&dtrace_lock);
16314 		dof = dtrace_dof_create(state);
16315 		mutex_exit(&dtrace_lock);
16316 
16317 		len = MIN(hdr.dofh_loadsz, dof->dofh_loadsz);
16318 		rval = copyout(dof, (void *)arg, len);
16319 		dtrace_dof_destroy(dof);
16320 
16321 		return (rval == 0 ? 0 : EFAULT);
16322 	}
16323 
16324 	case DTRACEIOC_AGGSNAP:
16325 	case DTRACEIOC_BUFSNAP: {
16326 		dtrace_bufdesc_t desc;
16327 		caddr_t cached;
16328 		dtrace_buffer_t *buf;
16329 
16330 		if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16331 			return (EFAULT);
16332 
16333 		if (desc.dtbd_cpu < 0 || desc.dtbd_cpu >= NCPU)
16334 			return (EINVAL);
16335 
16336 		mutex_enter(&dtrace_lock);
16337 
16338 		if (cmd == DTRACEIOC_BUFSNAP) {
16339 			buf = &state->dts_buffer[desc.dtbd_cpu];
16340 		} else {
16341 			buf = &state->dts_aggbuffer[desc.dtbd_cpu];
16342 		}
16343 
16344 		if (buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL)) {
16345 			size_t sz = buf->dtb_offset;
16346 
16347 			if (state->dts_activity != DTRACE_ACTIVITY_STOPPED) {
16348 				mutex_exit(&dtrace_lock);
16349 				return (EBUSY);
16350 			}
16351 
16352 			/*
16353 			 * If this buffer has already been consumed, we're
16354 			 * going to indicate that there's nothing left here
16355 			 * to consume.
16356 			 */
16357 			if (buf->dtb_flags & DTRACEBUF_CONSUMED) {
16358 				mutex_exit(&dtrace_lock);
16359 
16360 				desc.dtbd_size = 0;
16361 				desc.dtbd_drops = 0;
16362 				desc.dtbd_errors = 0;
16363 				desc.dtbd_oldest = 0;
16364 				sz = sizeof (desc);
16365 
16366 				if (copyout(&desc, (void *)arg, sz) != 0)
16367 					return (EFAULT);
16368 
16369 				return (0);
16370 			}
16371 
16372 			/*
16373 			 * If this is a ring buffer that has wrapped, we want
16374 			 * to copy the whole thing out.
16375 			 */
16376 			if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
16377 				dtrace_buffer_polish(buf);
16378 				sz = buf->dtb_size;
16379 			}
16380 
16381 			if (copyout(buf->dtb_tomax, desc.dtbd_data, sz) != 0) {
16382 				mutex_exit(&dtrace_lock);
16383 				return (EFAULT);
16384 			}
16385 
16386 			desc.dtbd_size = sz;
16387 			desc.dtbd_drops = buf->dtb_drops;
16388 			desc.dtbd_errors = buf->dtb_errors;
16389 			desc.dtbd_oldest = buf->dtb_xamot_offset;
16390 
16391 			mutex_exit(&dtrace_lock);
16392 
16393 			if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16394 				return (EFAULT);
16395 
16396 			buf->dtb_flags |= DTRACEBUF_CONSUMED;
16397 
16398 			return (0);
16399 		}
16400 
16401 		if (buf->dtb_tomax == NULL) {
16402 			ASSERT(buf->dtb_xamot == NULL);
16403 			mutex_exit(&dtrace_lock);
16404 			return (ENOENT);
16405 		}
16406 
16407 		cached = buf->dtb_tomax;
16408 		ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
16409 
16410 		dtrace_xcall(desc.dtbd_cpu,
16411 		    (dtrace_xcall_t)dtrace_buffer_switch, buf);
16412 
16413 		state->dts_errors += buf->dtb_xamot_errors;
16414 
16415 		/*
16416 		 * If the buffers did not actually switch, then the cross call
16417 		 * did not take place -- presumably because the given CPU is
16418 		 * not in the ready set.  If this is the case, we'll return
16419 		 * ENOENT.
16420 		 */
16421 		if (buf->dtb_tomax == cached) {
16422 			ASSERT(buf->dtb_xamot != cached);
16423 			mutex_exit(&dtrace_lock);
16424 			return (ENOENT);
16425 		}
16426 
16427 		ASSERT(cached == buf->dtb_xamot);
16428 
16429 		/*
16430 		 * We have our snapshot; now copy it out.
16431 		 */
16432 		if (copyout(buf->dtb_xamot, desc.dtbd_data,
16433 		    buf->dtb_xamot_offset) != 0) {
16434 			mutex_exit(&dtrace_lock);
16435 			return (EFAULT);
16436 		}
16437 
16438 		desc.dtbd_size = buf->dtb_xamot_offset;
16439 		desc.dtbd_drops = buf->dtb_xamot_drops;
16440 		desc.dtbd_errors = buf->dtb_xamot_errors;
16441 		desc.dtbd_oldest = 0;
16442 
16443 		mutex_exit(&dtrace_lock);
16444 
16445 		/*
16446 		 * Finally, copy out the buffer description.
16447 		 */
16448 		if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16449 			return (EFAULT);
16450 
16451 		return (0);
16452 	}
16453 
16454 	case DTRACEIOC_CONF: {
16455 		dtrace_conf_t conf;
16456 
16457 		bzero(&conf, sizeof (conf));
16458 		conf.dtc_difversion = DIF_VERSION;
16459 		conf.dtc_difintregs = DIF_DIR_NREGS;
16460 		conf.dtc_diftupregs = DIF_DTR_NREGS;
16461 		conf.dtc_ctfmodel = CTF_MODEL_NATIVE;
16462 
16463 		if (copyout(&conf, (void *)arg, sizeof (conf)) != 0)
16464 			return (EFAULT);
16465 
16466 		return (0);
16467 	}
16468 
16469 	case DTRACEIOC_STATUS: {
16470 		dtrace_status_t stat;
16471 		dtrace_dstate_t *dstate;
16472 		int i, j;
16473 		uint64_t nerrs;
16474 
16475 		/*
16476 		 * See the comment in dtrace_state_deadman() for the reason
16477 		 * for setting dts_laststatus to INT64_MAX before setting
16478 		 * it to the correct value.
16479 		 */
16480 		state->dts_laststatus = INT64_MAX;
16481 		dtrace_membar_producer();
16482 		state->dts_laststatus = dtrace_gethrtime();
16483 
16484 		bzero(&stat, sizeof (stat));
16485 
16486 		mutex_enter(&dtrace_lock);
16487 
16488 		if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE) {
16489 			mutex_exit(&dtrace_lock);
16490 			return (ENOENT);
16491 		}
16492 
16493 		if (state->dts_activity == DTRACE_ACTIVITY_DRAINING)
16494 			stat.dtst_exiting = 1;
16495 
16496 		nerrs = state->dts_errors;
16497 		dstate = &state->dts_vstate.dtvs_dynvars;
16498 
16499 		for (i = 0; i < NCPU; i++) {
16500 			dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[i];
16501 
16502 			stat.dtst_dyndrops += dcpu->dtdsc_drops;
16503 			stat.dtst_dyndrops_dirty += dcpu->dtdsc_dirty_drops;
16504 			stat.dtst_dyndrops_rinsing += dcpu->dtdsc_rinsing_drops;
16505 
16506 			if (state->dts_buffer[i].dtb_flags & DTRACEBUF_FULL)
16507 				stat.dtst_filled++;
16508 
16509 			nerrs += state->dts_buffer[i].dtb_errors;
16510 
16511 			for (j = 0; j < state->dts_nspeculations; j++) {
16512 				dtrace_speculation_t *spec;
16513 				dtrace_buffer_t *buf;
16514 
16515 				spec = &state->dts_speculations[j];
16516 				buf = &spec->dtsp_buffer[i];
16517 				stat.dtst_specdrops += buf->dtb_xamot_drops;
16518 			}
16519 		}
16520 
16521 		stat.dtst_specdrops_busy = state->dts_speculations_busy;
16522 		stat.dtst_specdrops_unavail = state->dts_speculations_unavail;
16523 		stat.dtst_stkstroverflows = state->dts_stkstroverflows;
16524 		stat.dtst_dblerrors = state->dts_dblerrors;
16525 		stat.dtst_killed =
16526 		    (state->dts_activity == DTRACE_ACTIVITY_KILLED);
16527 		stat.dtst_errors = nerrs;
16528 
16529 		mutex_exit(&dtrace_lock);
16530 
16531 		if (copyout(&stat, (void *)arg, sizeof (stat)) != 0)
16532 			return (EFAULT);
16533 
16534 		return (0);
16535 	}
16536 
16537 	case DTRACEIOC_FORMAT: {
16538 		dtrace_fmtdesc_t fmt;
16539 		char *str;
16540 		int len;
16541 
16542 		if (copyin((void *)arg, &fmt, sizeof (fmt)) != 0)
16543 			return (EFAULT);
16544 
16545 		mutex_enter(&dtrace_lock);
16546 
16547 		if (fmt.dtfd_format == 0 ||
16548 		    fmt.dtfd_format > state->dts_nformats) {
16549 			mutex_exit(&dtrace_lock);
16550 			return (EINVAL);
16551 		}
16552 
16553 		/*
16554 		 * Format strings are allocated contiguously and they are
16555 		 * never freed; if a format index is less than the number
16556 		 * of formats, we can assert that the format map is non-NULL
16557 		 * and that the format for the specified index is non-NULL.
16558 		 */
16559 		ASSERT(state->dts_formats != NULL);
16560 		str = state->dts_formats[fmt.dtfd_format - 1];
16561 		ASSERT(str != NULL);
16562 
16563 		len = strlen(str) + 1;
16564 
16565 		if (len > fmt.dtfd_length) {
16566 			fmt.dtfd_length = len;
16567 
16568 			if (copyout(&fmt, (void *)arg, sizeof (fmt)) != 0) {
16569 				mutex_exit(&dtrace_lock);
16570 				return (EINVAL);
16571 			}
16572 		} else {
16573 			if (copyout(str, fmt.dtfd_string, len) != 0) {
16574 				mutex_exit(&dtrace_lock);
16575 				return (EINVAL);
16576 			}
16577 		}
16578 
16579 		mutex_exit(&dtrace_lock);
16580 		return (0);
16581 	}
16582 
16583 	default:
16584 		break;
16585 	}
16586 
16587 	return (ENOTTY);
16588 }
16589 
16590 /*ARGSUSED*/
16591 static int
16592 dtrace_detach(dev_info_t *dip, ddi_detach_cmd_t cmd)
16593 {
16594 	dtrace_state_t *state;
16595 
16596 	switch (cmd) {
16597 	case DDI_DETACH:
16598 		break;
16599 
16600 	case DDI_SUSPEND:
16601 		return (DDI_SUCCESS);
16602 
16603 	default:
16604 		return (DDI_FAILURE);
16605 	}
16606 
16607 	mutex_enter(&cpu_lock);
16608 	mutex_enter(&dtrace_provider_lock);
16609 	mutex_enter(&dtrace_lock);
16610 
16611 	ASSERT(dtrace_opens == 0);
16612 
16613 	if (dtrace_helpers > 0) {
16614 		mutex_exit(&dtrace_provider_lock);
16615 		mutex_exit(&dtrace_lock);
16616 		mutex_exit(&cpu_lock);
16617 		return (DDI_FAILURE);
16618 	}
16619 
16620 	if (dtrace_unregister((dtrace_provider_id_t)dtrace_provider) != 0) {
16621 		mutex_exit(&dtrace_provider_lock);
16622 		mutex_exit(&dtrace_lock);
16623 		mutex_exit(&cpu_lock);
16624 		return (DDI_FAILURE);
16625 	}
16626 
16627 	dtrace_provider = NULL;
16628 
16629 	if ((state = dtrace_anon_grab()) != NULL) {
16630 		/*
16631 		 * If there were ECBs on this state, the provider should
16632 		 * have not been allowed to detach; assert that there is
16633 		 * none.
16634 		 */
16635 		ASSERT(state->dts_necbs == 0);
16636 		dtrace_state_destroy(state);
16637 
16638 		/*
16639 		 * If we're being detached with anonymous state, we need to
16640 		 * indicate to the kernel debugger that DTrace is now inactive.
16641 		 */
16642 		(void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
16643 	}
16644 
16645 	bzero(&dtrace_anon, sizeof (dtrace_anon_t));
16646 	unregister_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
16647 	dtrace_cpu_init = NULL;
16648 	dtrace_helpers_cleanup = NULL;
16649 	dtrace_helpers_fork = NULL;
16650 	dtrace_cpustart_init = NULL;
16651 	dtrace_cpustart_fini = NULL;
16652 	dtrace_debugger_init = NULL;
16653 	dtrace_debugger_fini = NULL;
16654 	dtrace_modload = NULL;
16655 	dtrace_modunload = NULL;
16656 
16657 	mutex_exit(&cpu_lock);
16658 
16659 	if (dtrace_helptrace_enabled) {
16660 		kmem_free(dtrace_helptrace_buffer, dtrace_helptrace_bufsize);
16661 		dtrace_helptrace_buffer = NULL;
16662 	}
16663 
16664 	kmem_free(dtrace_probes, dtrace_nprobes * sizeof (dtrace_probe_t *));
16665 	dtrace_probes = NULL;
16666 	dtrace_nprobes = 0;
16667 
16668 	dtrace_hash_destroy(dtrace_bymod);
16669 	dtrace_hash_destroy(dtrace_byfunc);
16670 	dtrace_hash_destroy(dtrace_byname);
16671 	dtrace_bymod = NULL;
16672 	dtrace_byfunc = NULL;
16673 	dtrace_byname = NULL;
16674 
16675 	kmem_cache_destroy(dtrace_state_cache);
16676 	vmem_destroy(dtrace_minor);
16677 	vmem_destroy(dtrace_arena);
16678 
16679 	if (dtrace_toxrange != NULL) {
16680 		kmem_free(dtrace_toxrange,
16681 		    dtrace_toxranges_max * sizeof (dtrace_toxrange_t));
16682 		dtrace_toxrange = NULL;
16683 		dtrace_toxranges = 0;
16684 		dtrace_toxranges_max = 0;
16685 	}
16686 
16687 	ddi_remove_minor_node(dtrace_devi, NULL);
16688 	dtrace_devi = NULL;
16689 
16690 	ddi_soft_state_fini(&dtrace_softstate);
16691 
16692 	ASSERT(dtrace_vtime_references == 0);
16693 	ASSERT(dtrace_opens == 0);
16694 	ASSERT(dtrace_retained == NULL);
16695 
16696 	mutex_exit(&dtrace_lock);
16697 	mutex_exit(&dtrace_provider_lock);
16698 
16699 	/*
16700 	 * We don't destroy the task queue until after we have dropped our
16701 	 * locks (taskq_destroy() may block on running tasks).  To prevent
16702 	 * attempting to do work after we have effectively detached but before
16703 	 * the task queue has been destroyed, all tasks dispatched via the
16704 	 * task queue must check that DTrace is still attached before
16705 	 * performing any operation.
16706 	 */
16707 	taskq_destroy(dtrace_taskq);
16708 	dtrace_taskq = NULL;
16709 
16710 	return (DDI_SUCCESS);
16711 }
16712 #endif
16713 
16714 #if defined(sun)
16715 /*ARGSUSED*/
16716 static int
16717 dtrace_info(dev_info_t *dip, ddi_info_cmd_t infocmd, void *arg, void **result)
16718 {
16719 	int error;
16720 
16721 	switch (infocmd) {
16722 	case DDI_INFO_DEVT2DEVINFO:
16723 		*result = (void *)dtrace_devi;
16724 		error = DDI_SUCCESS;
16725 		break;
16726 	case DDI_INFO_DEVT2INSTANCE:
16727 		*result = (void *)0;
16728 		error = DDI_SUCCESS;
16729 		break;
16730 	default:
16731 		error = DDI_FAILURE;
16732 	}
16733 	return (error);
16734 }
16735 #endif
16736 
16737 #if defined(sun)
16738 static struct cb_ops dtrace_cb_ops = {
16739 	dtrace_open,		/* open */
16740 	dtrace_close,		/* close */
16741 	nulldev,		/* strategy */
16742 	nulldev,		/* print */
16743 	nodev,			/* dump */
16744 	nodev,			/* read */
16745 	nodev,			/* write */
16746 	dtrace_ioctl,		/* ioctl */
16747 	nodev,			/* devmap */
16748 	nodev,			/* mmap */
16749 	nodev,			/* segmap */
16750 	nochpoll,		/* poll */
16751 	ddi_prop_op,		/* cb_prop_op */
16752 	0,			/* streamtab  */
16753 	D_NEW | D_MP		/* Driver compatibility flag */
16754 };
16755 
16756 static struct dev_ops dtrace_ops = {
16757 	DEVO_REV,		/* devo_rev */
16758 	0,			/* refcnt */
16759 	dtrace_info,		/* get_dev_info */
16760 	nulldev,		/* identify */
16761 	nulldev,		/* probe */
16762 	dtrace_attach,		/* attach */
16763 	dtrace_detach,		/* detach */
16764 	nodev,			/* reset */
16765 	&dtrace_cb_ops,		/* driver operations */
16766 	NULL,			/* bus operations */
16767 	nodev			/* dev power */
16768 };
16769 
16770 static struct modldrv modldrv = {
16771 	&mod_driverops,		/* module type (this is a pseudo driver) */
16772 	"Dynamic Tracing",	/* name of module */
16773 	&dtrace_ops,		/* driver ops */
16774 };
16775 
16776 static struct modlinkage modlinkage = {
16777 	MODREV_1,
16778 	(void *)&modldrv,
16779 	NULL
16780 };
16781 
16782 int
16783 _init(void)
16784 {
16785 	return (mod_install(&modlinkage));
16786 }
16787 
16788 int
16789 _info(struct modinfo *modinfop)
16790 {
16791 	return (mod_info(&modlinkage, modinfop));
16792 }
16793 
16794 int
16795 _fini(void)
16796 {
16797 	return (mod_remove(&modlinkage));
16798 }
16799 #else
16800 
16801 static d_ioctl_t	dtrace_ioctl;
16802 static d_ioctl_t	dtrace_ioctl_helper;
16803 static void		dtrace_load(void *);
16804 static int		dtrace_unload(void);
16805 #if __FreeBSD_version < 800039
16806 static void		dtrace_clone(void *, struct ucred *, char *, int , struct cdev **);
16807 static struct clonedevs	*dtrace_clones;		/* Ptr to the array of cloned devices. */
16808 static eventhandler_tag	eh_tag;			/* Event handler tag. */
16809 #else
16810 static struct cdev	*dtrace_dev;
16811 static struct cdev	*helper_dev;
16812 #endif
16813 
16814 void dtrace_invop_init(void);
16815 void dtrace_invop_uninit(void);
16816 
16817 static struct cdevsw dtrace_cdevsw = {
16818 	.d_version	= D_VERSION,
16819 #if __FreeBSD_version < 800039
16820 	.d_flags	= D_TRACKCLOSE | D_NEEDMINOR,
16821 	.d_close	= dtrace_close,
16822 #endif
16823 	.d_ioctl	= dtrace_ioctl,
16824 	.d_open		= dtrace_open,
16825 	.d_name		= "dtrace",
16826 };
16827 
16828 static struct cdevsw helper_cdevsw = {
16829 	.d_version	= D_VERSION,
16830 	.d_ioctl	= dtrace_ioctl_helper,
16831 	.d_name		= "helper",
16832 };
16833 
16834 #include <dtrace_anon.c>
16835 #if __FreeBSD_version < 800039
16836 #include <dtrace_clone.c>
16837 #endif
16838 #include <dtrace_ioctl.c>
16839 #include <dtrace_load.c>
16840 #include <dtrace_modevent.c>
16841 #include <dtrace_sysctl.c>
16842 #include <dtrace_unload.c>
16843 #include <dtrace_vtime.c>
16844 #include <dtrace_hacks.c>
16845 #include <dtrace_isa.c>
16846 
16847 SYSINIT(dtrace_load, SI_SUB_DTRACE, SI_ORDER_FIRST, dtrace_load, NULL);
16848 SYSUNINIT(dtrace_unload, SI_SUB_DTRACE, SI_ORDER_FIRST, dtrace_unload, NULL);
16849 SYSINIT(dtrace_anon_init, SI_SUB_DTRACE_ANON, SI_ORDER_FIRST, dtrace_anon_init, NULL);
16850 
16851 DEV_MODULE(dtrace, dtrace_modevent, NULL);
16852 MODULE_VERSION(dtrace, 1);
16853 MODULE_DEPEND(dtrace, cyclic, 1, 1, 1);
16854 MODULE_DEPEND(dtrace, opensolaris, 1, 1, 1);
16855 #endif
16856