xref: /freebsd/sys/cddl/contrib/opensolaris/uts/common/dtrace/dtrace.c (revision de7b456e596ff18032d2cbfdf244c66f36770da4)
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  * Copyright (c) 2012 by Delphix. All rights reserved
27  * Use is subject to license terms.
28  */
29 
30 #pragma ident	"%Z%%M%	%I%	%E% SMI"
31 
32 /*
33  * DTrace - Dynamic Tracing for Solaris
34  *
35  * This is the implementation of the Solaris Dynamic Tracing framework
36  * (DTrace).  The user-visible interface to DTrace is described at length in
37  * the "Solaris Dynamic Tracing Guide".  The interfaces between the libdtrace
38  * library, the in-kernel DTrace framework, and the DTrace providers are
39  * described in the block comments in the <sys/dtrace.h> header file.  The
40  * internal architecture of DTrace is described in the block comments in the
41  * <sys/dtrace_impl.h> header file.  The comments contained within the DTrace
42  * implementation very much assume mastery of all of these sources; if one has
43  * an unanswered question about the implementation, one should consult them
44  * first.
45  *
46  * The functions here are ordered roughly as follows:
47  *
48  *   - Probe context functions
49  *   - Probe hashing functions
50  *   - Non-probe context utility functions
51  *   - Matching functions
52  *   - Provider-to-Framework API functions
53  *   - Probe management functions
54  *   - DIF object functions
55  *   - Format functions
56  *   - Predicate functions
57  *   - ECB functions
58  *   - Buffer functions
59  *   - Enabling functions
60  *   - DOF functions
61  *   - Anonymous enabling functions
62  *   - Consumer state functions
63  *   - Helper functions
64  *   - Hook functions
65  *   - Driver cookbook functions
66  *
67  * Each group of functions begins with a block comment labelled the "DTrace
68  * [Group] Functions", allowing one to find each block by searching forward
69  * on capital-f functions.
70  */
71 #include <sys/errno.h>
72 #if !defined(sun)
73 #include <sys/time.h>
74 #endif
75 #include <sys/stat.h>
76 #include <sys/modctl.h>
77 #include <sys/conf.h>
78 #include <sys/systm.h>
79 #if defined(sun)
80 #include <sys/ddi.h>
81 #include <sys/sunddi.h>
82 #endif
83 #include <sys/cpuvar.h>
84 #include <sys/kmem.h>
85 #if defined(sun)
86 #include <sys/strsubr.h>
87 #endif
88 #include <sys/sysmacros.h>
89 #include <sys/dtrace_impl.h>
90 #include <sys/atomic.h>
91 #include <sys/cmn_err.h>
92 #if defined(sun)
93 #include <sys/mutex_impl.h>
94 #include <sys/rwlock_impl.h>
95 #endif
96 #include <sys/ctf_api.h>
97 #if defined(sun)
98 #include <sys/panic.h>
99 #include <sys/priv_impl.h>
100 #endif
101 #include <sys/policy.h>
102 #if defined(sun)
103 #include <sys/cred_impl.h>
104 #include <sys/procfs_isa.h>
105 #endif
106 #include <sys/taskq.h>
107 #if defined(sun)
108 #include <sys/mkdev.h>
109 #include <sys/kdi.h>
110 #endif
111 #include <sys/zone.h>
112 #include <sys/socket.h>
113 #include <netinet/in.h>
114 
115 /* FreeBSD includes: */
116 #if !defined(sun)
117 #include <sys/callout.h>
118 #include <sys/ctype.h>
119 #include <sys/eventhandler.h>
120 #include <sys/limits.h>
121 #include <sys/kdb.h>
122 #include <sys/kernel.h>
123 #include <sys/malloc.h>
124 #include <sys/sysctl.h>
125 #include <sys/lock.h>
126 #include <sys/mutex.h>
127 #include <sys/rwlock.h>
128 #include <sys/sx.h>
129 #include <sys/dtrace_bsd.h>
130 #include <netinet/in.h>
131 #include "dtrace_cddl.h"
132 #include "dtrace_debug.c"
133 #endif
134 
135 /*
136  * DTrace Tunable Variables
137  *
138  * The following variables may be tuned by adding a line to /etc/system that
139  * includes both the name of the DTrace module ("dtrace") and the name of the
140  * variable.  For example:
141  *
142  *   set dtrace:dtrace_destructive_disallow = 1
143  *
144  * In general, the only variables that one should be tuning this way are those
145  * that affect system-wide DTrace behavior, and for which the default behavior
146  * is undesirable.  Most of these variables are tunable on a per-consumer
147  * basis using DTrace options, and need not be tuned on a system-wide basis.
148  * When tuning these variables, avoid pathological values; while some attempt
149  * is made to verify the integrity of these variables, they are not considered
150  * part of the supported interface to DTrace, and they are therefore not
151  * checked comprehensively.  Further, these variables should not be tuned
152  * dynamically via "mdb -kw" or other means; they should only be tuned via
153  * /etc/system.
154  */
155 int		dtrace_destructive_disallow = 0;
156 dtrace_optval_t	dtrace_nonroot_maxsize = (16 * 1024 * 1024);
157 size_t		dtrace_difo_maxsize = (256 * 1024);
158 dtrace_optval_t	dtrace_dof_maxsize = (256 * 1024);
159 size_t		dtrace_global_maxsize = (16 * 1024);
160 size_t		dtrace_actions_max = (16 * 1024);
161 size_t		dtrace_retain_max = 1024;
162 dtrace_optval_t	dtrace_helper_actions_max = 128;
163 dtrace_optval_t	dtrace_helper_providers_max = 32;
164 dtrace_optval_t	dtrace_dstate_defsize = (1 * 1024 * 1024);
165 size_t		dtrace_strsize_default = 256;
166 dtrace_optval_t	dtrace_cleanrate_default = 9900990;		/* 101 hz */
167 dtrace_optval_t	dtrace_cleanrate_min = 200000;			/* 5000 hz */
168 dtrace_optval_t	dtrace_cleanrate_max = (uint64_t)60 * NANOSEC;	/* 1/minute */
169 dtrace_optval_t	dtrace_aggrate_default = NANOSEC;		/* 1 hz */
170 dtrace_optval_t	dtrace_statusrate_default = NANOSEC;		/* 1 hz */
171 dtrace_optval_t dtrace_statusrate_max = (hrtime_t)10 * NANOSEC;	 /* 6/minute */
172 dtrace_optval_t	dtrace_switchrate_default = NANOSEC;		/* 1 hz */
173 dtrace_optval_t	dtrace_nspec_default = 1;
174 dtrace_optval_t	dtrace_specsize_default = 32 * 1024;
175 dtrace_optval_t dtrace_stackframes_default = 20;
176 dtrace_optval_t dtrace_ustackframes_default = 20;
177 dtrace_optval_t dtrace_jstackframes_default = 50;
178 dtrace_optval_t dtrace_jstackstrsize_default = 512;
179 int		dtrace_msgdsize_max = 128;
180 hrtime_t	dtrace_chill_max = 500 * (NANOSEC / MILLISEC);	/* 500 ms */
181 hrtime_t	dtrace_chill_interval = NANOSEC;		/* 1000 ms */
182 int		dtrace_devdepth_max = 32;
183 int		dtrace_err_verbose;
184 hrtime_t	dtrace_deadman_interval = NANOSEC;
185 hrtime_t	dtrace_deadman_timeout = (hrtime_t)10 * NANOSEC;
186 hrtime_t	dtrace_deadman_user = (hrtime_t)30 * NANOSEC;
187 hrtime_t	dtrace_unregister_defunct_reap = (hrtime_t)60 * NANOSEC;
188 #if !defined(sun)
189 int		dtrace_memstr_max = 4096;
190 #endif
191 
192 /*
193  * DTrace External Variables
194  *
195  * As dtrace(7D) is a kernel module, any DTrace variables are obviously
196  * available to DTrace consumers via the backtick (`) syntax.  One of these,
197  * dtrace_zero, is made deliberately so:  it is provided as a source of
198  * well-known, zero-filled memory.  While this variable is not documented,
199  * it is used by some translators as an implementation detail.
200  */
201 const char	dtrace_zero[256] = { 0 };	/* zero-filled memory */
202 
203 /*
204  * DTrace Internal Variables
205  */
206 #if defined(sun)
207 static dev_info_t	*dtrace_devi;		/* device info */
208 #endif
209 #if defined(sun)
210 static vmem_t		*dtrace_arena;		/* probe ID arena */
211 static vmem_t		*dtrace_minor;		/* minor number arena */
212 #else
213 static taskq_t		*dtrace_taskq;		/* task queue */
214 static struct unrhdr	*dtrace_arena;		/* Probe ID number.     */
215 #endif
216 static dtrace_probe_t	**dtrace_probes;	/* array of all probes */
217 static int		dtrace_nprobes;		/* number of probes */
218 static dtrace_provider_t *dtrace_provider;	/* provider list */
219 static dtrace_meta_t	*dtrace_meta_pid;	/* user-land meta provider */
220 static int		dtrace_opens;		/* number of opens */
221 static int		dtrace_helpers;		/* number of helpers */
222 #if defined(sun)
223 static void		*dtrace_softstate;	/* softstate pointer */
224 #endif
225 static dtrace_hash_t	*dtrace_bymod;		/* probes hashed by module */
226 static dtrace_hash_t	*dtrace_byfunc;		/* probes hashed by function */
227 static dtrace_hash_t	*dtrace_byname;		/* probes hashed by name */
228 static dtrace_toxrange_t *dtrace_toxrange;	/* toxic range array */
229 static int		dtrace_toxranges;	/* number of toxic ranges */
230 static int		dtrace_toxranges_max;	/* size of toxic range array */
231 static dtrace_anon_t	dtrace_anon;		/* anonymous enabling */
232 static kmem_cache_t	*dtrace_state_cache;	/* cache for dynamic state */
233 static uint64_t		dtrace_vtime_references; /* number of vtimestamp refs */
234 static kthread_t	*dtrace_panicked;	/* panicking thread */
235 static dtrace_ecb_t	*dtrace_ecb_create_cache; /* cached created ECB */
236 static dtrace_genid_t	dtrace_probegen;	/* current probe generation */
237 static dtrace_helpers_t *dtrace_deferred_pid;	/* deferred helper list */
238 static dtrace_enabling_t *dtrace_retained;	/* list of retained enablings */
239 static dtrace_dynvar_t	dtrace_dynhash_sink;	/* end of dynamic hash chains */
240 #if !defined(sun)
241 static struct mtx	dtrace_unr_mtx;
242 MTX_SYSINIT(dtrace_unr_mtx, &dtrace_unr_mtx, "Unique resource identifier", MTX_DEF);
243 int		dtrace_in_probe;	/* non-zero if executing a probe */
244 #if defined(__i386__) || defined(__amd64__) || defined(__mips__) || defined(__powerpc__)
245 uintptr_t	dtrace_in_probe_addr;	/* Address of invop when already in probe */
246 #endif
247 static eventhandler_tag	dtrace_kld_load_tag;
248 static eventhandler_tag	dtrace_kld_unload_try_tag;
249 #endif
250 
251 /*
252  * DTrace Locking
253  * DTrace is protected by three (relatively coarse-grained) locks:
254  *
255  * (1) dtrace_lock is required to manipulate essentially any DTrace state,
256  *     including enabling state, probes, ECBs, consumer state, helper state,
257  *     etc.  Importantly, dtrace_lock is _not_ required when in probe context;
258  *     probe context is lock-free -- synchronization is handled via the
259  *     dtrace_sync() cross call mechanism.
260  *
261  * (2) dtrace_provider_lock is required when manipulating provider state, or
262  *     when provider state must be held constant.
263  *
264  * (3) dtrace_meta_lock is required when manipulating meta provider state, or
265  *     when meta provider state must be held constant.
266  *
267  * The lock ordering between these three locks is dtrace_meta_lock before
268  * dtrace_provider_lock before dtrace_lock.  (In particular, there are
269  * several places where dtrace_provider_lock is held by the framework as it
270  * calls into the providers -- which then call back into the framework,
271  * grabbing dtrace_lock.)
272  *
273  * There are two other locks in the mix:  mod_lock and cpu_lock.  With respect
274  * to dtrace_provider_lock and dtrace_lock, cpu_lock continues its historical
275  * role as a coarse-grained lock; it is acquired before both of these locks.
276  * With respect to dtrace_meta_lock, its behavior is stranger:  cpu_lock must
277  * be acquired _between_ dtrace_meta_lock and any other DTrace locks.
278  * mod_lock is similar with respect to dtrace_provider_lock in that it must be
279  * acquired _between_ dtrace_provider_lock and dtrace_lock.
280  */
281 static kmutex_t		dtrace_lock;		/* probe state lock */
282 static kmutex_t		dtrace_provider_lock;	/* provider state lock */
283 static kmutex_t		dtrace_meta_lock;	/* meta-provider state lock */
284 
285 #if !defined(sun)
286 /* XXX FreeBSD hacks. */
287 #define cr_suid		cr_svuid
288 #define cr_sgid		cr_svgid
289 #define	ipaddr_t	in_addr_t
290 #define mod_modname	pathname
291 #define vuprintf	vprintf
292 #define ttoproc(_a)	((_a)->td_proc)
293 #define crgetzoneid(_a)	0
294 #define	NCPU		MAXCPU
295 #define SNOCD		0
296 #define CPU_ON_INTR(_a)	0
297 
298 #define PRIV_EFFECTIVE		(1 << 0)
299 #define PRIV_DTRACE_KERNEL	(1 << 1)
300 #define PRIV_DTRACE_PROC	(1 << 2)
301 #define PRIV_DTRACE_USER	(1 << 3)
302 #define PRIV_PROC_OWNER		(1 << 4)
303 #define PRIV_PROC_ZONE		(1 << 5)
304 #define PRIV_ALL		~0
305 
306 SYSCTL_NODE(_debug, OID_AUTO, dtrace, CTLFLAG_RD, 0, "DTrace Information");
307 #endif
308 
309 #if defined(sun)
310 #define curcpu	CPU->cpu_id
311 #endif
312 
313 
314 /*
315  * DTrace Provider Variables
316  *
317  * These are the variables relating to DTrace as a provider (that is, the
318  * provider of the BEGIN, END, and ERROR probes).
319  */
320 static dtrace_pattr_t	dtrace_provider_attr = {
321 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
322 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
323 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
324 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
325 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
326 };
327 
328 static void
329 dtrace_nullop(void)
330 {}
331 
332 static dtrace_pops_t	dtrace_provider_ops = {
333 	(void (*)(void *, dtrace_probedesc_t *))dtrace_nullop,
334 	(void (*)(void *, modctl_t *))dtrace_nullop,
335 	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
336 	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
337 	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
338 	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
339 	NULL,
340 	NULL,
341 	NULL,
342 	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop
343 };
344 
345 static dtrace_id_t	dtrace_probeid_begin;	/* special BEGIN probe */
346 static dtrace_id_t	dtrace_probeid_end;	/* special END probe */
347 dtrace_id_t		dtrace_probeid_error;	/* special ERROR probe */
348 
349 /*
350  * DTrace Helper Tracing Variables
351  */
352 uint32_t dtrace_helptrace_next = 0;
353 uint32_t dtrace_helptrace_nlocals;
354 char	*dtrace_helptrace_buffer;
355 int	dtrace_helptrace_bufsize = 512 * 1024;
356 
357 #ifdef DEBUG
358 int	dtrace_helptrace_enabled = 1;
359 #else
360 int	dtrace_helptrace_enabled = 0;
361 #endif
362 
363 /*
364  * DTrace Error Hashing
365  *
366  * On DEBUG kernels, DTrace will track the errors that has seen in a hash
367  * table.  This is very useful for checking coverage of tests that are
368  * expected to induce DIF or DOF processing errors, and may be useful for
369  * debugging problems in the DIF code generator or in DOF generation .  The
370  * error hash may be examined with the ::dtrace_errhash MDB dcmd.
371  */
372 #ifdef DEBUG
373 static dtrace_errhash_t	dtrace_errhash[DTRACE_ERRHASHSZ];
374 static const char *dtrace_errlast;
375 static kthread_t *dtrace_errthread;
376 static kmutex_t dtrace_errlock;
377 #endif
378 
379 /*
380  * DTrace Macros and Constants
381  *
382  * These are various macros that are useful in various spots in the
383  * implementation, along with a few random constants that have no meaning
384  * outside of the implementation.  There is no real structure to this cpp
385  * mishmash -- but is there ever?
386  */
387 #define	DTRACE_HASHSTR(hash, probe)	\
388 	dtrace_hash_str(*((char **)((uintptr_t)(probe) + (hash)->dth_stroffs)))
389 
390 #define	DTRACE_HASHNEXT(hash, probe)	\
391 	(dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_nextoffs)
392 
393 #define	DTRACE_HASHPREV(hash, probe)	\
394 	(dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_prevoffs)
395 
396 #define	DTRACE_HASHEQ(hash, lhs, rhs)	\
397 	(strcmp(*((char **)((uintptr_t)(lhs) + (hash)->dth_stroffs)), \
398 	    *((char **)((uintptr_t)(rhs) + (hash)->dth_stroffs))) == 0)
399 
400 #define	DTRACE_AGGHASHSIZE_SLEW		17
401 
402 #define	DTRACE_V4MAPPED_OFFSET		(sizeof (uint32_t) * 3)
403 
404 /*
405  * The key for a thread-local variable consists of the lower 61 bits of the
406  * t_did, plus the 3 bits of the highest active interrupt above LOCK_LEVEL.
407  * We add DIF_VARIABLE_MAX to t_did to assure that the thread key is never
408  * equal to a variable identifier.  This is necessary (but not sufficient) to
409  * assure that global associative arrays never collide with thread-local
410  * variables.  To guarantee that they cannot collide, we must also define the
411  * order for keying dynamic variables.  That order is:
412  *
413  *   [ key0 ] ... [ keyn ] [ variable-key ] [ tls-key ]
414  *
415  * Because the variable-key and the tls-key are in orthogonal spaces, there is
416  * no way for a global variable key signature to match a thread-local key
417  * signature.
418  */
419 #if defined(sun)
420 #define	DTRACE_TLS_THRKEY(where) { \
421 	uint_t intr = 0; \
422 	uint_t actv = CPU->cpu_intr_actv >> (LOCK_LEVEL + 1); \
423 	for (; actv; actv >>= 1) \
424 		intr++; \
425 	ASSERT(intr < (1 << 3)); \
426 	(where) = ((curthread->t_did + DIF_VARIABLE_MAX) & \
427 	    (((uint64_t)1 << 61) - 1)) | ((uint64_t)intr << 61); \
428 }
429 #else
430 #define	DTRACE_TLS_THRKEY(where) { \
431 	solaris_cpu_t *_c = &solaris_cpu[curcpu]; \
432 	uint_t intr = 0; \
433 	uint_t actv = _c->cpu_intr_actv; \
434 	for (; actv; actv >>= 1) \
435 		intr++; \
436 	ASSERT(intr < (1 << 3)); \
437 	(where) = ((curthread->td_tid + DIF_VARIABLE_MAX) & \
438 	    (((uint64_t)1 << 61) - 1)) | ((uint64_t)intr << 61); \
439 }
440 #endif
441 
442 #define	DT_BSWAP_8(x)	((x) & 0xff)
443 #define	DT_BSWAP_16(x)	((DT_BSWAP_8(x) << 8) | DT_BSWAP_8((x) >> 8))
444 #define	DT_BSWAP_32(x)	((DT_BSWAP_16(x) << 16) | DT_BSWAP_16((x) >> 16))
445 #define	DT_BSWAP_64(x)	((DT_BSWAP_32(x) << 32) | DT_BSWAP_32((x) >> 32))
446 
447 #define	DT_MASK_LO 0x00000000FFFFFFFFULL
448 
449 #define	DTRACE_STORE(type, tomax, offset, what) \
450 	*((type *)((uintptr_t)(tomax) + (uintptr_t)offset)) = (type)(what);
451 
452 #ifndef __x86
453 #define	DTRACE_ALIGNCHECK(addr, size, flags)				\
454 	if (addr & (size - 1)) {					\
455 		*flags |= CPU_DTRACE_BADALIGN;				\
456 		cpu_core[curcpu].cpuc_dtrace_illval = addr;	\
457 		return (0);						\
458 	}
459 #else
460 #define	DTRACE_ALIGNCHECK(addr, size, flags)
461 #endif
462 
463 /*
464  * Test whether a range of memory starting at testaddr of size testsz falls
465  * within the range of memory described by addr, sz.  We take care to avoid
466  * problems with overflow and underflow of the unsigned quantities, and
467  * disallow all negative sizes.  Ranges of size 0 are allowed.
468  */
469 #define	DTRACE_INRANGE(testaddr, testsz, baseaddr, basesz) \
470 	((testaddr) - (baseaddr) < (basesz) && \
471 	(testaddr) + (testsz) - (baseaddr) <= (basesz) && \
472 	(testaddr) + (testsz) >= (testaddr))
473 
474 /*
475  * Test whether alloc_sz bytes will fit in the scratch region.  We isolate
476  * alloc_sz on the righthand side of the comparison in order to avoid overflow
477  * or underflow in the comparison with it.  This is simpler than the INRANGE
478  * check above, because we know that the dtms_scratch_ptr is valid in the
479  * range.  Allocations of size zero are allowed.
480  */
481 #define	DTRACE_INSCRATCH(mstate, alloc_sz) \
482 	((mstate)->dtms_scratch_base + (mstate)->dtms_scratch_size - \
483 	(mstate)->dtms_scratch_ptr >= (alloc_sz))
484 
485 #define	DTRACE_LOADFUNC(bits)						\
486 /*CSTYLED*/								\
487 uint##bits##_t								\
488 dtrace_load##bits(uintptr_t addr)					\
489 {									\
490 	size_t size = bits / NBBY;					\
491 	/*CSTYLED*/							\
492 	uint##bits##_t rval;						\
493 	int i;								\
494 	volatile uint16_t *flags = (volatile uint16_t *)		\
495 	    &cpu_core[curcpu].cpuc_dtrace_flags;			\
496 									\
497 	DTRACE_ALIGNCHECK(addr, size, flags);				\
498 									\
499 	for (i = 0; i < dtrace_toxranges; i++) {			\
500 		if (addr >= dtrace_toxrange[i].dtt_limit)		\
501 			continue;					\
502 									\
503 		if (addr + size <= dtrace_toxrange[i].dtt_base)		\
504 			continue;					\
505 									\
506 		/*							\
507 		 * This address falls within a toxic region; return 0.	\
508 		 */							\
509 		*flags |= CPU_DTRACE_BADADDR;				\
510 		cpu_core[curcpu].cpuc_dtrace_illval = addr;		\
511 		return (0);						\
512 	}								\
513 									\
514 	*flags |= CPU_DTRACE_NOFAULT;					\
515 	/*CSTYLED*/							\
516 	rval = *((volatile uint##bits##_t *)addr);			\
517 	*flags &= ~CPU_DTRACE_NOFAULT;					\
518 									\
519 	return (!(*flags & CPU_DTRACE_FAULT) ? rval : 0);		\
520 }
521 
522 #ifdef _LP64
523 #define	dtrace_loadptr	dtrace_load64
524 #else
525 #define	dtrace_loadptr	dtrace_load32
526 #endif
527 
528 #define	DTRACE_DYNHASH_FREE	0
529 #define	DTRACE_DYNHASH_SINK	1
530 #define	DTRACE_DYNHASH_VALID	2
531 
532 #define	DTRACE_MATCH_NEXT	0
533 #define	DTRACE_MATCH_DONE	1
534 #define	DTRACE_ANCHORED(probe)	((probe)->dtpr_func[0] != '\0')
535 #define	DTRACE_STATE_ALIGN	64
536 
537 #define	DTRACE_FLAGS2FLT(flags)						\
538 	(((flags) & CPU_DTRACE_BADADDR) ? DTRACEFLT_BADADDR :		\
539 	((flags) & CPU_DTRACE_ILLOP) ? DTRACEFLT_ILLOP :		\
540 	((flags) & CPU_DTRACE_DIVZERO) ? DTRACEFLT_DIVZERO :		\
541 	((flags) & CPU_DTRACE_KPRIV) ? DTRACEFLT_KPRIV :		\
542 	((flags) & CPU_DTRACE_UPRIV) ? DTRACEFLT_UPRIV :		\
543 	((flags) & CPU_DTRACE_TUPOFLOW) ?  DTRACEFLT_TUPOFLOW :		\
544 	((flags) & CPU_DTRACE_BADALIGN) ?  DTRACEFLT_BADALIGN :		\
545 	((flags) & CPU_DTRACE_NOSCRATCH) ?  DTRACEFLT_NOSCRATCH :	\
546 	((flags) & CPU_DTRACE_BADSTACK) ?  DTRACEFLT_BADSTACK :		\
547 	DTRACEFLT_UNKNOWN)
548 
549 #define	DTRACEACT_ISSTRING(act)						\
550 	((act)->dta_kind == DTRACEACT_DIFEXPR &&			\
551 	(act)->dta_difo->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING)
552 
553 /* Function prototype definitions: */
554 static size_t dtrace_strlen(const char *, size_t);
555 static dtrace_probe_t *dtrace_probe_lookup_id(dtrace_id_t id);
556 static void dtrace_enabling_provide(dtrace_provider_t *);
557 static int dtrace_enabling_match(dtrace_enabling_t *, int *);
558 static void dtrace_enabling_matchall(void);
559 static void dtrace_enabling_reap(void);
560 static dtrace_state_t *dtrace_anon_grab(void);
561 static uint64_t dtrace_helper(int, dtrace_mstate_t *,
562     dtrace_state_t *, uint64_t, uint64_t);
563 static dtrace_helpers_t *dtrace_helpers_create(proc_t *);
564 static void dtrace_buffer_drop(dtrace_buffer_t *);
565 static int dtrace_buffer_consumed(dtrace_buffer_t *, hrtime_t when);
566 static intptr_t dtrace_buffer_reserve(dtrace_buffer_t *, size_t, size_t,
567     dtrace_state_t *, dtrace_mstate_t *);
568 static int dtrace_state_option(dtrace_state_t *, dtrace_optid_t,
569     dtrace_optval_t);
570 static int dtrace_ecb_create_enable(dtrace_probe_t *, void *);
571 static void dtrace_helper_provider_destroy(dtrace_helper_provider_t *);
572 uint16_t dtrace_load16(uintptr_t);
573 uint32_t dtrace_load32(uintptr_t);
574 uint64_t dtrace_load64(uintptr_t);
575 uint8_t dtrace_load8(uintptr_t);
576 void dtrace_dynvar_clean(dtrace_dstate_t *);
577 dtrace_dynvar_t *dtrace_dynvar(dtrace_dstate_t *, uint_t, dtrace_key_t *,
578     size_t, dtrace_dynvar_op_t, dtrace_mstate_t *, dtrace_vstate_t *);
579 uintptr_t dtrace_dif_varstr(uintptr_t, dtrace_state_t *, dtrace_mstate_t *);
580 
581 /*
582  * DTrace Probe Context Functions
583  *
584  * These functions are called from probe context.  Because probe context is
585  * any context in which C may be called, arbitrarily locks may be held,
586  * interrupts may be disabled, we may be in arbitrary dispatched state, etc.
587  * As a result, functions called from probe context may only call other DTrace
588  * support functions -- they may not interact at all with the system at large.
589  * (Note that the ASSERT macro is made probe-context safe by redefining it in
590  * terms of dtrace_assfail(), a probe-context safe function.) If arbitrary
591  * loads are to be performed from probe context, they _must_ be in terms of
592  * the safe dtrace_load*() variants.
593  *
594  * Some functions in this block are not actually called from probe context;
595  * for these functions, there will be a comment above the function reading
596  * "Note:  not called from probe context."
597  */
598 void
599 dtrace_panic(const char *format, ...)
600 {
601 	va_list alist;
602 
603 	va_start(alist, format);
604 	dtrace_vpanic(format, alist);
605 	va_end(alist);
606 }
607 
608 int
609 dtrace_assfail(const char *a, const char *f, int l)
610 {
611 	dtrace_panic("assertion failed: %s, file: %s, line: %d", a, f, l);
612 
613 	/*
614 	 * We just need something here that even the most clever compiler
615 	 * cannot optimize away.
616 	 */
617 	return (a[(uintptr_t)f]);
618 }
619 
620 /*
621  * Atomically increment a specified error counter from probe context.
622  */
623 static void
624 dtrace_error(uint32_t *counter)
625 {
626 	/*
627 	 * Most counters stored to in probe context are per-CPU counters.
628 	 * However, there are some error conditions that are sufficiently
629 	 * arcane that they don't merit per-CPU storage.  If these counters
630 	 * are incremented concurrently on different CPUs, scalability will be
631 	 * adversely affected -- but we don't expect them to be white-hot in a
632 	 * correctly constructed enabling...
633 	 */
634 	uint32_t oval, nval;
635 
636 	do {
637 		oval = *counter;
638 
639 		if ((nval = oval + 1) == 0) {
640 			/*
641 			 * If the counter would wrap, set it to 1 -- assuring
642 			 * that the counter is never zero when we have seen
643 			 * errors.  (The counter must be 32-bits because we
644 			 * aren't guaranteed a 64-bit compare&swap operation.)
645 			 * To save this code both the infamy of being fingered
646 			 * by a priggish news story and the indignity of being
647 			 * the target of a neo-puritan witch trial, we're
648 			 * carefully avoiding any colorful description of the
649 			 * likelihood of this condition -- but suffice it to
650 			 * say that it is only slightly more likely than the
651 			 * overflow of predicate cache IDs, as discussed in
652 			 * dtrace_predicate_create().
653 			 */
654 			nval = 1;
655 		}
656 	} while (dtrace_cas32(counter, oval, nval) != oval);
657 }
658 
659 /*
660  * Use the DTRACE_LOADFUNC macro to define functions for each of loading a
661  * uint8_t, a uint16_t, a uint32_t and a uint64_t.
662  */
663 DTRACE_LOADFUNC(8)
664 DTRACE_LOADFUNC(16)
665 DTRACE_LOADFUNC(32)
666 DTRACE_LOADFUNC(64)
667 
668 static int
669 dtrace_inscratch(uintptr_t dest, size_t size, dtrace_mstate_t *mstate)
670 {
671 	if (dest < mstate->dtms_scratch_base)
672 		return (0);
673 
674 	if (dest + size < dest)
675 		return (0);
676 
677 	if (dest + size > mstate->dtms_scratch_ptr)
678 		return (0);
679 
680 	return (1);
681 }
682 
683 static int
684 dtrace_canstore_statvar(uint64_t addr, size_t sz,
685     dtrace_statvar_t **svars, int nsvars)
686 {
687 	int i;
688 
689 	for (i = 0; i < nsvars; i++) {
690 		dtrace_statvar_t *svar = svars[i];
691 
692 		if (svar == NULL || svar->dtsv_size == 0)
693 			continue;
694 
695 		if (DTRACE_INRANGE(addr, sz, svar->dtsv_data, svar->dtsv_size))
696 			return (1);
697 	}
698 
699 	return (0);
700 }
701 
702 /*
703  * Check to see if the address is within a memory region to which a store may
704  * be issued.  This includes the DTrace scratch areas, and any DTrace variable
705  * region.  The caller of dtrace_canstore() is responsible for performing any
706  * alignment checks that are needed before stores are actually executed.
707  */
708 static int
709 dtrace_canstore(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
710     dtrace_vstate_t *vstate)
711 {
712 	/*
713 	 * First, check to see if the address is in scratch space...
714 	 */
715 	if (DTRACE_INRANGE(addr, sz, mstate->dtms_scratch_base,
716 	    mstate->dtms_scratch_size))
717 		return (1);
718 
719 	/*
720 	 * Now check to see if it's a dynamic variable.  This check will pick
721 	 * up both thread-local variables and any global dynamically-allocated
722 	 * variables.
723 	 */
724 	if (DTRACE_INRANGE(addr, sz, (uintptr_t)vstate->dtvs_dynvars.dtds_base,
725 	    vstate->dtvs_dynvars.dtds_size)) {
726 		dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
727 		uintptr_t base = (uintptr_t)dstate->dtds_base +
728 		    (dstate->dtds_hashsize * sizeof (dtrace_dynhash_t));
729 		uintptr_t chunkoffs;
730 
731 		/*
732 		 * Before we assume that we can store here, we need to make
733 		 * sure that it isn't in our metadata -- storing to our
734 		 * dynamic variable metadata would corrupt our state.  For
735 		 * the range to not include any dynamic variable metadata,
736 		 * it must:
737 		 *
738 		 *	(1) Start above the hash table that is at the base of
739 		 *	the dynamic variable space
740 		 *
741 		 *	(2) Have a starting chunk offset that is beyond the
742 		 *	dtrace_dynvar_t that is at the base of every chunk
743 		 *
744 		 *	(3) Not span a chunk boundary
745 		 *
746 		 */
747 		if (addr < base)
748 			return (0);
749 
750 		chunkoffs = (addr - base) % dstate->dtds_chunksize;
751 
752 		if (chunkoffs < sizeof (dtrace_dynvar_t))
753 			return (0);
754 
755 		if (chunkoffs + sz > dstate->dtds_chunksize)
756 			return (0);
757 
758 		return (1);
759 	}
760 
761 	/*
762 	 * Finally, check the static local and global variables.  These checks
763 	 * take the longest, so we perform them last.
764 	 */
765 	if (dtrace_canstore_statvar(addr, sz,
766 	    vstate->dtvs_locals, vstate->dtvs_nlocals))
767 		return (1);
768 
769 	if (dtrace_canstore_statvar(addr, sz,
770 	    vstate->dtvs_globals, vstate->dtvs_nglobals))
771 		return (1);
772 
773 	return (0);
774 }
775 
776 
777 /*
778  * Convenience routine to check to see if the address is within a memory
779  * region in which a load may be issued given the user's privilege level;
780  * if not, it sets the appropriate error flags and loads 'addr' into the
781  * illegal value slot.
782  *
783  * DTrace subroutines (DIF_SUBR_*) should use this helper to implement
784  * appropriate memory access protection.
785  */
786 static int
787 dtrace_canload(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
788     dtrace_vstate_t *vstate)
789 {
790 	volatile uintptr_t *illval = &cpu_core[curcpu].cpuc_dtrace_illval;
791 
792 	/*
793 	 * If we hold the privilege to read from kernel memory, then
794 	 * everything is readable.
795 	 */
796 	if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
797 		return (1);
798 
799 	/*
800 	 * You can obviously read that which you can store.
801 	 */
802 	if (dtrace_canstore(addr, sz, mstate, vstate))
803 		return (1);
804 
805 	/*
806 	 * We're allowed to read from our own string table.
807 	 */
808 	if (DTRACE_INRANGE(addr, sz, (uintptr_t)mstate->dtms_difo->dtdo_strtab,
809 	    mstate->dtms_difo->dtdo_strlen))
810 		return (1);
811 
812 	DTRACE_CPUFLAG_SET(CPU_DTRACE_KPRIV);
813 	*illval = addr;
814 	return (0);
815 }
816 
817 /*
818  * Convenience routine to check to see if a given string is within a memory
819  * region in which a load may be issued given the user's privilege level;
820  * this exists so that we don't need to issue unnecessary dtrace_strlen()
821  * calls in the event that the user has all privileges.
822  */
823 static int
824 dtrace_strcanload(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
825     dtrace_vstate_t *vstate)
826 {
827 	size_t strsz;
828 
829 	/*
830 	 * If we hold the privilege to read from kernel memory, then
831 	 * everything is readable.
832 	 */
833 	if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
834 		return (1);
835 
836 	strsz = 1 + dtrace_strlen((char *)(uintptr_t)addr, sz);
837 	if (dtrace_canload(addr, strsz, mstate, vstate))
838 		return (1);
839 
840 	return (0);
841 }
842 
843 /*
844  * Convenience routine to check to see if a given variable is within a memory
845  * region in which a load may be issued given the user's privilege level.
846  */
847 static int
848 dtrace_vcanload(void *src, dtrace_diftype_t *type, dtrace_mstate_t *mstate,
849     dtrace_vstate_t *vstate)
850 {
851 	size_t sz;
852 	ASSERT(type->dtdt_flags & DIF_TF_BYREF);
853 
854 	/*
855 	 * If we hold the privilege to read from kernel memory, then
856 	 * everything is readable.
857 	 */
858 	if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
859 		return (1);
860 
861 	if (type->dtdt_kind == DIF_TYPE_STRING)
862 		sz = dtrace_strlen(src,
863 		    vstate->dtvs_state->dts_options[DTRACEOPT_STRSIZE]) + 1;
864 	else
865 		sz = type->dtdt_size;
866 
867 	return (dtrace_canload((uintptr_t)src, sz, mstate, vstate));
868 }
869 
870 /*
871  * Compare two strings using safe loads.
872  */
873 static int
874 dtrace_strncmp(char *s1, char *s2, size_t limit)
875 {
876 	uint8_t c1, c2;
877 	volatile uint16_t *flags;
878 
879 	if (s1 == s2 || limit == 0)
880 		return (0);
881 
882 	flags = (volatile uint16_t *)&cpu_core[curcpu].cpuc_dtrace_flags;
883 
884 	do {
885 		if (s1 == NULL) {
886 			c1 = '\0';
887 		} else {
888 			c1 = dtrace_load8((uintptr_t)s1++);
889 		}
890 
891 		if (s2 == NULL) {
892 			c2 = '\0';
893 		} else {
894 			c2 = dtrace_load8((uintptr_t)s2++);
895 		}
896 
897 		if (c1 != c2)
898 			return (c1 - c2);
899 	} while (--limit && c1 != '\0' && !(*flags & CPU_DTRACE_FAULT));
900 
901 	return (0);
902 }
903 
904 /*
905  * Compute strlen(s) for a string using safe memory accesses.  The additional
906  * len parameter is used to specify a maximum length to ensure completion.
907  */
908 static size_t
909 dtrace_strlen(const char *s, size_t lim)
910 {
911 	uint_t len;
912 
913 	for (len = 0; len != lim; len++) {
914 		if (dtrace_load8((uintptr_t)s++) == '\0')
915 			break;
916 	}
917 
918 	return (len);
919 }
920 
921 /*
922  * Check if an address falls within a toxic region.
923  */
924 static int
925 dtrace_istoxic(uintptr_t kaddr, size_t size)
926 {
927 	uintptr_t taddr, tsize;
928 	int i;
929 
930 	for (i = 0; i < dtrace_toxranges; i++) {
931 		taddr = dtrace_toxrange[i].dtt_base;
932 		tsize = dtrace_toxrange[i].dtt_limit - taddr;
933 
934 		if (kaddr - taddr < tsize) {
935 			DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
936 			cpu_core[curcpu].cpuc_dtrace_illval = kaddr;
937 			return (1);
938 		}
939 
940 		if (taddr - kaddr < size) {
941 			DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
942 			cpu_core[curcpu].cpuc_dtrace_illval = taddr;
943 			return (1);
944 		}
945 	}
946 
947 	return (0);
948 }
949 
950 /*
951  * Copy src to dst using safe memory accesses.  The src is assumed to be unsafe
952  * memory specified by the DIF program.  The dst is assumed to be safe memory
953  * that we can store to directly because it is managed by DTrace.  As with
954  * standard bcopy, overlapping copies are handled properly.
955  */
956 static void
957 dtrace_bcopy(const void *src, void *dst, size_t len)
958 {
959 	if (len != 0) {
960 		uint8_t *s1 = dst;
961 		const uint8_t *s2 = src;
962 
963 		if (s1 <= s2) {
964 			do {
965 				*s1++ = dtrace_load8((uintptr_t)s2++);
966 			} while (--len != 0);
967 		} else {
968 			s2 += len;
969 			s1 += len;
970 
971 			do {
972 				*--s1 = dtrace_load8((uintptr_t)--s2);
973 			} while (--len != 0);
974 		}
975 	}
976 }
977 
978 /*
979  * Copy src to dst using safe memory accesses, up to either the specified
980  * length, or the point that a nul byte is encountered.  The src is assumed to
981  * be unsafe memory specified by the DIF program.  The dst is assumed to be
982  * safe memory that we can store to directly because it is managed by DTrace.
983  * Unlike dtrace_bcopy(), overlapping regions are not handled.
984  */
985 static void
986 dtrace_strcpy(const void *src, void *dst, size_t len)
987 {
988 	if (len != 0) {
989 		uint8_t *s1 = dst, c;
990 		const uint8_t *s2 = src;
991 
992 		do {
993 			*s1++ = c = dtrace_load8((uintptr_t)s2++);
994 		} while (--len != 0 && c != '\0');
995 	}
996 }
997 
998 /*
999  * Copy src to dst, deriving the size and type from the specified (BYREF)
1000  * variable type.  The src is assumed to be unsafe memory specified by the DIF
1001  * program.  The dst is assumed to be DTrace variable memory that is of the
1002  * specified type; we assume that we can store to directly.
1003  */
1004 static void
1005 dtrace_vcopy(void *src, void *dst, dtrace_diftype_t *type)
1006 {
1007 	ASSERT(type->dtdt_flags & DIF_TF_BYREF);
1008 
1009 	if (type->dtdt_kind == DIF_TYPE_STRING) {
1010 		dtrace_strcpy(src, dst, type->dtdt_size);
1011 	} else {
1012 		dtrace_bcopy(src, dst, type->dtdt_size);
1013 	}
1014 }
1015 
1016 /*
1017  * Compare s1 to s2 using safe memory accesses.  The s1 data is assumed to be
1018  * unsafe memory specified by the DIF program.  The s2 data is assumed to be
1019  * safe memory that we can access directly because it is managed by DTrace.
1020  */
1021 static int
1022 dtrace_bcmp(const void *s1, const void *s2, size_t len)
1023 {
1024 	volatile uint16_t *flags;
1025 
1026 	flags = (volatile uint16_t *)&cpu_core[curcpu].cpuc_dtrace_flags;
1027 
1028 	if (s1 == s2)
1029 		return (0);
1030 
1031 	if (s1 == NULL || s2 == NULL)
1032 		return (1);
1033 
1034 	if (s1 != s2 && len != 0) {
1035 		const uint8_t *ps1 = s1;
1036 		const uint8_t *ps2 = s2;
1037 
1038 		do {
1039 			if (dtrace_load8((uintptr_t)ps1++) != *ps2++)
1040 				return (1);
1041 		} while (--len != 0 && !(*flags & CPU_DTRACE_FAULT));
1042 	}
1043 	return (0);
1044 }
1045 
1046 /*
1047  * Zero the specified region using a simple byte-by-byte loop.  Note that this
1048  * is for safe DTrace-managed memory only.
1049  */
1050 static void
1051 dtrace_bzero(void *dst, size_t len)
1052 {
1053 	uchar_t *cp;
1054 
1055 	for (cp = dst; len != 0; len--)
1056 		*cp++ = 0;
1057 }
1058 
1059 static void
1060 dtrace_add_128(uint64_t *addend1, uint64_t *addend2, uint64_t *sum)
1061 {
1062 	uint64_t result[2];
1063 
1064 	result[0] = addend1[0] + addend2[0];
1065 	result[1] = addend1[1] + addend2[1] +
1066 	    (result[0] < addend1[0] || result[0] < addend2[0] ? 1 : 0);
1067 
1068 	sum[0] = result[0];
1069 	sum[1] = result[1];
1070 }
1071 
1072 /*
1073  * Shift the 128-bit value in a by b. If b is positive, shift left.
1074  * If b is negative, shift right.
1075  */
1076 static void
1077 dtrace_shift_128(uint64_t *a, int b)
1078 {
1079 	uint64_t mask;
1080 
1081 	if (b == 0)
1082 		return;
1083 
1084 	if (b < 0) {
1085 		b = -b;
1086 		if (b >= 64) {
1087 			a[0] = a[1] >> (b - 64);
1088 			a[1] = 0;
1089 		} else {
1090 			a[0] >>= b;
1091 			mask = 1LL << (64 - b);
1092 			mask -= 1;
1093 			a[0] |= ((a[1] & mask) << (64 - b));
1094 			a[1] >>= b;
1095 		}
1096 	} else {
1097 		if (b >= 64) {
1098 			a[1] = a[0] << (b - 64);
1099 			a[0] = 0;
1100 		} else {
1101 			a[1] <<= b;
1102 			mask = a[0] >> (64 - b);
1103 			a[1] |= mask;
1104 			a[0] <<= b;
1105 		}
1106 	}
1107 }
1108 
1109 /*
1110  * The basic idea is to break the 2 64-bit values into 4 32-bit values,
1111  * use native multiplication on those, and then re-combine into the
1112  * resulting 128-bit value.
1113  *
1114  * (hi1 << 32 + lo1) * (hi2 << 32 + lo2) =
1115  *     hi1 * hi2 << 64 +
1116  *     hi1 * lo2 << 32 +
1117  *     hi2 * lo1 << 32 +
1118  *     lo1 * lo2
1119  */
1120 static void
1121 dtrace_multiply_128(uint64_t factor1, uint64_t factor2, uint64_t *product)
1122 {
1123 	uint64_t hi1, hi2, lo1, lo2;
1124 	uint64_t tmp[2];
1125 
1126 	hi1 = factor1 >> 32;
1127 	hi2 = factor2 >> 32;
1128 
1129 	lo1 = factor1 & DT_MASK_LO;
1130 	lo2 = factor2 & DT_MASK_LO;
1131 
1132 	product[0] = lo1 * lo2;
1133 	product[1] = hi1 * hi2;
1134 
1135 	tmp[0] = hi1 * lo2;
1136 	tmp[1] = 0;
1137 	dtrace_shift_128(tmp, 32);
1138 	dtrace_add_128(product, tmp, product);
1139 
1140 	tmp[0] = hi2 * lo1;
1141 	tmp[1] = 0;
1142 	dtrace_shift_128(tmp, 32);
1143 	dtrace_add_128(product, tmp, product);
1144 }
1145 
1146 /*
1147  * This privilege check should be used by actions and subroutines to
1148  * verify that the user credentials of the process that enabled the
1149  * invoking ECB match the target credentials
1150  */
1151 static int
1152 dtrace_priv_proc_common_user(dtrace_state_t *state)
1153 {
1154 	cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1155 
1156 	/*
1157 	 * We should always have a non-NULL state cred here, since if cred
1158 	 * is null (anonymous tracing), we fast-path bypass this routine.
1159 	 */
1160 	ASSERT(s_cr != NULL);
1161 
1162 	if ((cr = CRED()) != NULL &&
1163 	    s_cr->cr_uid == cr->cr_uid &&
1164 	    s_cr->cr_uid == cr->cr_ruid &&
1165 	    s_cr->cr_uid == cr->cr_suid &&
1166 	    s_cr->cr_gid == cr->cr_gid &&
1167 	    s_cr->cr_gid == cr->cr_rgid &&
1168 	    s_cr->cr_gid == cr->cr_sgid)
1169 		return (1);
1170 
1171 	return (0);
1172 }
1173 
1174 /*
1175  * This privilege check should be used by actions and subroutines to
1176  * verify that the zone of the process that enabled the invoking ECB
1177  * matches the target credentials
1178  */
1179 static int
1180 dtrace_priv_proc_common_zone(dtrace_state_t *state)
1181 {
1182 #if defined(sun)
1183 	cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1184 
1185 	/*
1186 	 * We should always have a non-NULL state cred here, since if cred
1187 	 * is null (anonymous tracing), we fast-path bypass this routine.
1188 	 */
1189 	ASSERT(s_cr != NULL);
1190 
1191 	if ((cr = CRED()) != NULL &&
1192 	    s_cr->cr_zone == cr->cr_zone)
1193 		return (1);
1194 
1195 	return (0);
1196 #else
1197 	return (1);
1198 #endif
1199 }
1200 
1201 /*
1202  * This privilege check should be used by actions and subroutines to
1203  * verify that the process has not setuid or changed credentials.
1204  */
1205 static int
1206 dtrace_priv_proc_common_nocd(void)
1207 {
1208 	proc_t *proc;
1209 
1210 	if ((proc = ttoproc(curthread)) != NULL &&
1211 	    !(proc->p_flag & SNOCD))
1212 		return (1);
1213 
1214 	return (0);
1215 }
1216 
1217 static int
1218 dtrace_priv_proc_destructive(dtrace_state_t *state)
1219 {
1220 	int action = state->dts_cred.dcr_action;
1221 
1222 	if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE) == 0) &&
1223 	    dtrace_priv_proc_common_zone(state) == 0)
1224 		goto bad;
1225 
1226 	if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER) == 0) &&
1227 	    dtrace_priv_proc_common_user(state) == 0)
1228 		goto bad;
1229 
1230 	if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG) == 0) &&
1231 	    dtrace_priv_proc_common_nocd() == 0)
1232 		goto bad;
1233 
1234 	return (1);
1235 
1236 bad:
1237 	cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1238 
1239 	return (0);
1240 }
1241 
1242 static int
1243 dtrace_priv_proc_control(dtrace_state_t *state)
1244 {
1245 	if (state->dts_cred.dcr_action & DTRACE_CRA_PROC_CONTROL)
1246 		return (1);
1247 
1248 	if (dtrace_priv_proc_common_zone(state) &&
1249 	    dtrace_priv_proc_common_user(state) &&
1250 	    dtrace_priv_proc_common_nocd())
1251 		return (1);
1252 
1253 	cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1254 
1255 	return (0);
1256 }
1257 
1258 static int
1259 dtrace_priv_proc(dtrace_state_t *state)
1260 {
1261 	if (state->dts_cred.dcr_action & DTRACE_CRA_PROC)
1262 		return (1);
1263 
1264 	cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1265 
1266 	return (0);
1267 }
1268 
1269 static int
1270 dtrace_priv_kernel(dtrace_state_t *state)
1271 {
1272 	if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL)
1273 		return (1);
1274 
1275 	cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1276 
1277 	return (0);
1278 }
1279 
1280 static int
1281 dtrace_priv_kernel_destructive(dtrace_state_t *state)
1282 {
1283 	if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL_DESTRUCTIVE)
1284 		return (1);
1285 
1286 	cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1287 
1288 	return (0);
1289 }
1290 
1291 /*
1292  * Note:  not called from probe context.  This function is called
1293  * asynchronously (and at a regular interval) from outside of probe context to
1294  * clean the dirty dynamic variable lists on all CPUs.  Dynamic variable
1295  * cleaning is explained in detail in <sys/dtrace_impl.h>.
1296  */
1297 void
1298 dtrace_dynvar_clean(dtrace_dstate_t *dstate)
1299 {
1300 	dtrace_dynvar_t *dirty;
1301 	dtrace_dstate_percpu_t *dcpu;
1302 	int i, work = 0;
1303 
1304 	for (i = 0; i < NCPU; i++) {
1305 		dcpu = &dstate->dtds_percpu[i];
1306 
1307 		ASSERT(dcpu->dtdsc_rinsing == NULL);
1308 
1309 		/*
1310 		 * If the dirty list is NULL, there is no dirty work to do.
1311 		 */
1312 		if (dcpu->dtdsc_dirty == NULL)
1313 			continue;
1314 
1315 		/*
1316 		 * If the clean list is non-NULL, then we're not going to do
1317 		 * any work for this CPU -- it means that there has not been
1318 		 * a dtrace_dynvar() allocation on this CPU (or from this CPU)
1319 		 * since the last time we cleaned house.
1320 		 */
1321 		if (dcpu->dtdsc_clean != NULL)
1322 			continue;
1323 
1324 		work = 1;
1325 
1326 		/*
1327 		 * Atomically move the dirty list aside.
1328 		 */
1329 		do {
1330 			dirty = dcpu->dtdsc_dirty;
1331 
1332 			/*
1333 			 * Before we zap the dirty list, set the rinsing list.
1334 			 * (This allows for a potential assertion in
1335 			 * dtrace_dynvar():  if a free dynamic variable appears
1336 			 * on a hash chain, either the dirty list or the
1337 			 * rinsing list for some CPU must be non-NULL.)
1338 			 */
1339 			dcpu->dtdsc_rinsing = dirty;
1340 			dtrace_membar_producer();
1341 		} while (dtrace_casptr(&dcpu->dtdsc_dirty,
1342 		    dirty, NULL) != dirty);
1343 	}
1344 
1345 	if (!work) {
1346 		/*
1347 		 * We have no work to do; we can simply return.
1348 		 */
1349 		return;
1350 	}
1351 
1352 	dtrace_sync();
1353 
1354 	for (i = 0; i < NCPU; i++) {
1355 		dcpu = &dstate->dtds_percpu[i];
1356 
1357 		if (dcpu->dtdsc_rinsing == NULL)
1358 			continue;
1359 
1360 		/*
1361 		 * We are now guaranteed that no hash chain contains a pointer
1362 		 * into this dirty list; we can make it clean.
1363 		 */
1364 		ASSERT(dcpu->dtdsc_clean == NULL);
1365 		dcpu->dtdsc_clean = dcpu->dtdsc_rinsing;
1366 		dcpu->dtdsc_rinsing = NULL;
1367 	}
1368 
1369 	/*
1370 	 * Before we actually set the state to be DTRACE_DSTATE_CLEAN, make
1371 	 * sure that all CPUs have seen all of the dtdsc_clean pointers.
1372 	 * This prevents a race whereby a CPU incorrectly decides that
1373 	 * the state should be something other than DTRACE_DSTATE_CLEAN
1374 	 * after dtrace_dynvar_clean() has completed.
1375 	 */
1376 	dtrace_sync();
1377 
1378 	dstate->dtds_state = DTRACE_DSTATE_CLEAN;
1379 }
1380 
1381 /*
1382  * Depending on the value of the op parameter, this function looks-up,
1383  * allocates or deallocates an arbitrarily-keyed dynamic variable.  If an
1384  * allocation is requested, this function will return a pointer to a
1385  * dtrace_dynvar_t corresponding to the allocated variable -- or NULL if no
1386  * variable can be allocated.  If NULL is returned, the appropriate counter
1387  * will be incremented.
1388  */
1389 dtrace_dynvar_t *
1390 dtrace_dynvar(dtrace_dstate_t *dstate, uint_t nkeys,
1391     dtrace_key_t *key, size_t dsize, dtrace_dynvar_op_t op,
1392     dtrace_mstate_t *mstate, dtrace_vstate_t *vstate)
1393 {
1394 	uint64_t hashval = DTRACE_DYNHASH_VALID;
1395 	dtrace_dynhash_t *hash = dstate->dtds_hash;
1396 	dtrace_dynvar_t *free, *new_free, *next, *dvar, *start, *prev = NULL;
1397 	processorid_t me = curcpu, cpu = me;
1398 	dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[me];
1399 	size_t bucket, ksize;
1400 	size_t chunksize = dstate->dtds_chunksize;
1401 	uintptr_t kdata, lock, nstate;
1402 	uint_t i;
1403 
1404 	ASSERT(nkeys != 0);
1405 
1406 	/*
1407 	 * Hash the key.  As with aggregations, we use Jenkins' "One-at-a-time"
1408 	 * algorithm.  For the by-value portions, we perform the algorithm in
1409 	 * 16-bit chunks (as opposed to 8-bit chunks).  This speeds things up a
1410 	 * bit, and seems to have only a minute effect on distribution.  For
1411 	 * the by-reference data, we perform "One-at-a-time" iterating (safely)
1412 	 * over each referenced byte.  It's painful to do this, but it's much
1413 	 * better than pathological hash distribution.  The efficacy of the
1414 	 * hashing algorithm (and a comparison with other algorithms) may be
1415 	 * found by running the ::dtrace_dynstat MDB dcmd.
1416 	 */
1417 	for (i = 0; i < nkeys; i++) {
1418 		if (key[i].dttk_size == 0) {
1419 			uint64_t val = key[i].dttk_value;
1420 
1421 			hashval += (val >> 48) & 0xffff;
1422 			hashval += (hashval << 10);
1423 			hashval ^= (hashval >> 6);
1424 
1425 			hashval += (val >> 32) & 0xffff;
1426 			hashval += (hashval << 10);
1427 			hashval ^= (hashval >> 6);
1428 
1429 			hashval += (val >> 16) & 0xffff;
1430 			hashval += (hashval << 10);
1431 			hashval ^= (hashval >> 6);
1432 
1433 			hashval += val & 0xffff;
1434 			hashval += (hashval << 10);
1435 			hashval ^= (hashval >> 6);
1436 		} else {
1437 			/*
1438 			 * This is incredibly painful, but it beats the hell
1439 			 * out of the alternative.
1440 			 */
1441 			uint64_t j, size = key[i].dttk_size;
1442 			uintptr_t base = (uintptr_t)key[i].dttk_value;
1443 
1444 			if (!dtrace_canload(base, size, mstate, vstate))
1445 				break;
1446 
1447 			for (j = 0; j < size; j++) {
1448 				hashval += dtrace_load8(base + j);
1449 				hashval += (hashval << 10);
1450 				hashval ^= (hashval >> 6);
1451 			}
1452 		}
1453 	}
1454 
1455 	if (DTRACE_CPUFLAG_ISSET(CPU_DTRACE_FAULT))
1456 		return (NULL);
1457 
1458 	hashval += (hashval << 3);
1459 	hashval ^= (hashval >> 11);
1460 	hashval += (hashval << 15);
1461 
1462 	/*
1463 	 * There is a remote chance (ideally, 1 in 2^31) that our hashval
1464 	 * comes out to be one of our two sentinel hash values.  If this
1465 	 * actually happens, we set the hashval to be a value known to be a
1466 	 * non-sentinel value.
1467 	 */
1468 	if (hashval == DTRACE_DYNHASH_FREE || hashval == DTRACE_DYNHASH_SINK)
1469 		hashval = DTRACE_DYNHASH_VALID;
1470 
1471 	/*
1472 	 * Yes, it's painful to do a divide here.  If the cycle count becomes
1473 	 * important here, tricks can be pulled to reduce it.  (However, it's
1474 	 * critical that hash collisions be kept to an absolute minimum;
1475 	 * they're much more painful than a divide.)  It's better to have a
1476 	 * solution that generates few collisions and still keeps things
1477 	 * relatively simple.
1478 	 */
1479 	bucket = hashval % dstate->dtds_hashsize;
1480 
1481 	if (op == DTRACE_DYNVAR_DEALLOC) {
1482 		volatile uintptr_t *lockp = &hash[bucket].dtdh_lock;
1483 
1484 		for (;;) {
1485 			while ((lock = *lockp) & 1)
1486 				continue;
1487 
1488 			if (dtrace_casptr((volatile void *)lockp,
1489 			    (volatile void *)lock, (volatile void *)(lock + 1)) == (void *)lock)
1490 				break;
1491 		}
1492 
1493 		dtrace_membar_producer();
1494 	}
1495 
1496 top:
1497 	prev = NULL;
1498 	lock = hash[bucket].dtdh_lock;
1499 
1500 	dtrace_membar_consumer();
1501 
1502 	start = hash[bucket].dtdh_chain;
1503 	ASSERT(start != NULL && (start->dtdv_hashval == DTRACE_DYNHASH_SINK ||
1504 	    start->dtdv_hashval != DTRACE_DYNHASH_FREE ||
1505 	    op != DTRACE_DYNVAR_DEALLOC));
1506 
1507 	for (dvar = start; dvar != NULL; dvar = dvar->dtdv_next) {
1508 		dtrace_tuple_t *dtuple = &dvar->dtdv_tuple;
1509 		dtrace_key_t *dkey = &dtuple->dtt_key[0];
1510 
1511 		if (dvar->dtdv_hashval != hashval) {
1512 			if (dvar->dtdv_hashval == DTRACE_DYNHASH_SINK) {
1513 				/*
1514 				 * We've reached the sink, and therefore the
1515 				 * end of the hash chain; we can kick out of
1516 				 * the loop knowing that we have seen a valid
1517 				 * snapshot of state.
1518 				 */
1519 				ASSERT(dvar->dtdv_next == NULL);
1520 				ASSERT(dvar == &dtrace_dynhash_sink);
1521 				break;
1522 			}
1523 
1524 			if (dvar->dtdv_hashval == DTRACE_DYNHASH_FREE) {
1525 				/*
1526 				 * We've gone off the rails:  somewhere along
1527 				 * the line, one of the members of this hash
1528 				 * chain was deleted.  Note that we could also
1529 				 * detect this by simply letting this loop run
1530 				 * to completion, as we would eventually hit
1531 				 * the end of the dirty list.  However, we
1532 				 * want to avoid running the length of the
1533 				 * dirty list unnecessarily (it might be quite
1534 				 * long), so we catch this as early as
1535 				 * possible by detecting the hash marker.  In
1536 				 * this case, we simply set dvar to NULL and
1537 				 * break; the conditional after the loop will
1538 				 * send us back to top.
1539 				 */
1540 				dvar = NULL;
1541 				break;
1542 			}
1543 
1544 			goto next;
1545 		}
1546 
1547 		if (dtuple->dtt_nkeys != nkeys)
1548 			goto next;
1549 
1550 		for (i = 0; i < nkeys; i++, dkey++) {
1551 			if (dkey->dttk_size != key[i].dttk_size)
1552 				goto next; /* size or type mismatch */
1553 
1554 			if (dkey->dttk_size != 0) {
1555 				if (dtrace_bcmp(
1556 				    (void *)(uintptr_t)key[i].dttk_value,
1557 				    (void *)(uintptr_t)dkey->dttk_value,
1558 				    dkey->dttk_size))
1559 					goto next;
1560 			} else {
1561 				if (dkey->dttk_value != key[i].dttk_value)
1562 					goto next;
1563 			}
1564 		}
1565 
1566 		if (op != DTRACE_DYNVAR_DEALLOC)
1567 			return (dvar);
1568 
1569 		ASSERT(dvar->dtdv_next == NULL ||
1570 		    dvar->dtdv_next->dtdv_hashval != DTRACE_DYNHASH_FREE);
1571 
1572 		if (prev != NULL) {
1573 			ASSERT(hash[bucket].dtdh_chain != dvar);
1574 			ASSERT(start != dvar);
1575 			ASSERT(prev->dtdv_next == dvar);
1576 			prev->dtdv_next = dvar->dtdv_next;
1577 		} else {
1578 			if (dtrace_casptr(&hash[bucket].dtdh_chain,
1579 			    start, dvar->dtdv_next) != start) {
1580 				/*
1581 				 * We have failed to atomically swing the
1582 				 * hash table head pointer, presumably because
1583 				 * of a conflicting allocation on another CPU.
1584 				 * We need to reread the hash chain and try
1585 				 * again.
1586 				 */
1587 				goto top;
1588 			}
1589 		}
1590 
1591 		dtrace_membar_producer();
1592 
1593 		/*
1594 		 * Now set the hash value to indicate that it's free.
1595 		 */
1596 		ASSERT(hash[bucket].dtdh_chain != dvar);
1597 		dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
1598 
1599 		dtrace_membar_producer();
1600 
1601 		/*
1602 		 * Set the next pointer to point at the dirty list, and
1603 		 * atomically swing the dirty pointer to the newly freed dvar.
1604 		 */
1605 		do {
1606 			next = dcpu->dtdsc_dirty;
1607 			dvar->dtdv_next = next;
1608 		} while (dtrace_casptr(&dcpu->dtdsc_dirty, next, dvar) != next);
1609 
1610 		/*
1611 		 * Finally, unlock this hash bucket.
1612 		 */
1613 		ASSERT(hash[bucket].dtdh_lock == lock);
1614 		ASSERT(lock & 1);
1615 		hash[bucket].dtdh_lock++;
1616 
1617 		return (NULL);
1618 next:
1619 		prev = dvar;
1620 		continue;
1621 	}
1622 
1623 	if (dvar == NULL) {
1624 		/*
1625 		 * If dvar is NULL, it is because we went off the rails:
1626 		 * one of the elements that we traversed in the hash chain
1627 		 * was deleted while we were traversing it.  In this case,
1628 		 * we assert that we aren't doing a dealloc (deallocs lock
1629 		 * the hash bucket to prevent themselves from racing with
1630 		 * one another), and retry the hash chain traversal.
1631 		 */
1632 		ASSERT(op != DTRACE_DYNVAR_DEALLOC);
1633 		goto top;
1634 	}
1635 
1636 	if (op != DTRACE_DYNVAR_ALLOC) {
1637 		/*
1638 		 * If we are not to allocate a new variable, we want to
1639 		 * return NULL now.  Before we return, check that the value
1640 		 * of the lock word hasn't changed.  If it has, we may have
1641 		 * seen an inconsistent snapshot.
1642 		 */
1643 		if (op == DTRACE_DYNVAR_NOALLOC) {
1644 			if (hash[bucket].dtdh_lock != lock)
1645 				goto top;
1646 		} else {
1647 			ASSERT(op == DTRACE_DYNVAR_DEALLOC);
1648 			ASSERT(hash[bucket].dtdh_lock == lock);
1649 			ASSERT(lock & 1);
1650 			hash[bucket].dtdh_lock++;
1651 		}
1652 
1653 		return (NULL);
1654 	}
1655 
1656 	/*
1657 	 * We need to allocate a new dynamic variable.  The size we need is the
1658 	 * size of dtrace_dynvar plus the size of nkeys dtrace_key_t's plus the
1659 	 * size of any auxiliary key data (rounded up to 8-byte alignment) plus
1660 	 * the size of any referred-to data (dsize).  We then round the final
1661 	 * size up to the chunksize for allocation.
1662 	 */
1663 	for (ksize = 0, i = 0; i < nkeys; i++)
1664 		ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
1665 
1666 	/*
1667 	 * This should be pretty much impossible, but could happen if, say,
1668 	 * strange DIF specified the tuple.  Ideally, this should be an
1669 	 * assertion and not an error condition -- but that requires that the
1670 	 * chunksize calculation in dtrace_difo_chunksize() be absolutely
1671 	 * bullet-proof.  (That is, it must not be able to be fooled by
1672 	 * malicious DIF.)  Given the lack of backwards branches in DIF,
1673 	 * solving this would presumably not amount to solving the Halting
1674 	 * Problem -- but it still seems awfully hard.
1675 	 */
1676 	if (sizeof (dtrace_dynvar_t) + sizeof (dtrace_key_t) * (nkeys - 1) +
1677 	    ksize + dsize > chunksize) {
1678 		dcpu->dtdsc_drops++;
1679 		return (NULL);
1680 	}
1681 
1682 	nstate = DTRACE_DSTATE_EMPTY;
1683 
1684 	do {
1685 retry:
1686 		free = dcpu->dtdsc_free;
1687 
1688 		if (free == NULL) {
1689 			dtrace_dynvar_t *clean = dcpu->dtdsc_clean;
1690 			void *rval;
1691 
1692 			if (clean == NULL) {
1693 				/*
1694 				 * We're out of dynamic variable space on
1695 				 * this CPU.  Unless we have tried all CPUs,
1696 				 * we'll try to allocate from a different
1697 				 * CPU.
1698 				 */
1699 				switch (dstate->dtds_state) {
1700 				case DTRACE_DSTATE_CLEAN: {
1701 					void *sp = &dstate->dtds_state;
1702 
1703 					if (++cpu >= NCPU)
1704 						cpu = 0;
1705 
1706 					if (dcpu->dtdsc_dirty != NULL &&
1707 					    nstate == DTRACE_DSTATE_EMPTY)
1708 						nstate = DTRACE_DSTATE_DIRTY;
1709 
1710 					if (dcpu->dtdsc_rinsing != NULL)
1711 						nstate = DTRACE_DSTATE_RINSING;
1712 
1713 					dcpu = &dstate->dtds_percpu[cpu];
1714 
1715 					if (cpu != me)
1716 						goto retry;
1717 
1718 					(void) dtrace_cas32(sp,
1719 					    DTRACE_DSTATE_CLEAN, nstate);
1720 
1721 					/*
1722 					 * To increment the correct bean
1723 					 * counter, take another lap.
1724 					 */
1725 					goto retry;
1726 				}
1727 
1728 				case DTRACE_DSTATE_DIRTY:
1729 					dcpu->dtdsc_dirty_drops++;
1730 					break;
1731 
1732 				case DTRACE_DSTATE_RINSING:
1733 					dcpu->dtdsc_rinsing_drops++;
1734 					break;
1735 
1736 				case DTRACE_DSTATE_EMPTY:
1737 					dcpu->dtdsc_drops++;
1738 					break;
1739 				}
1740 
1741 				DTRACE_CPUFLAG_SET(CPU_DTRACE_DROP);
1742 				return (NULL);
1743 			}
1744 
1745 			/*
1746 			 * The clean list appears to be non-empty.  We want to
1747 			 * move the clean list to the free list; we start by
1748 			 * moving the clean pointer aside.
1749 			 */
1750 			if (dtrace_casptr(&dcpu->dtdsc_clean,
1751 			    clean, NULL) != clean) {
1752 				/*
1753 				 * We are in one of two situations:
1754 				 *
1755 				 *  (a)	The clean list was switched to the
1756 				 *	free list by another CPU.
1757 				 *
1758 				 *  (b)	The clean list was added to by the
1759 				 *	cleansing cyclic.
1760 				 *
1761 				 * In either of these situations, we can
1762 				 * just reattempt the free list allocation.
1763 				 */
1764 				goto retry;
1765 			}
1766 
1767 			ASSERT(clean->dtdv_hashval == DTRACE_DYNHASH_FREE);
1768 
1769 			/*
1770 			 * Now we'll move the clean list to the free list.
1771 			 * It's impossible for this to fail:  the only way
1772 			 * the free list can be updated is through this
1773 			 * code path, and only one CPU can own the clean list.
1774 			 * Thus, it would only be possible for this to fail if
1775 			 * this code were racing with dtrace_dynvar_clean().
1776 			 * (That is, if dtrace_dynvar_clean() updated the clean
1777 			 * list, and we ended up racing to update the free
1778 			 * list.)  This race is prevented by the dtrace_sync()
1779 			 * in dtrace_dynvar_clean() -- which flushes the
1780 			 * owners of the clean lists out before resetting
1781 			 * the clean lists.
1782 			 */
1783 			rval = dtrace_casptr(&dcpu->dtdsc_free, NULL, clean);
1784 			ASSERT(rval == NULL);
1785 			goto retry;
1786 		}
1787 
1788 		dvar = free;
1789 		new_free = dvar->dtdv_next;
1790 	} while (dtrace_casptr(&dcpu->dtdsc_free, free, new_free) != free);
1791 
1792 	/*
1793 	 * We have now allocated a new chunk.  We copy the tuple keys into the
1794 	 * tuple array and copy any referenced key data into the data space
1795 	 * following the tuple array.  As we do this, we relocate dttk_value
1796 	 * in the final tuple to point to the key data address in the chunk.
1797 	 */
1798 	kdata = (uintptr_t)&dvar->dtdv_tuple.dtt_key[nkeys];
1799 	dvar->dtdv_data = (void *)(kdata + ksize);
1800 	dvar->dtdv_tuple.dtt_nkeys = nkeys;
1801 
1802 	for (i = 0; i < nkeys; i++) {
1803 		dtrace_key_t *dkey = &dvar->dtdv_tuple.dtt_key[i];
1804 		size_t kesize = key[i].dttk_size;
1805 
1806 		if (kesize != 0) {
1807 			dtrace_bcopy(
1808 			    (const void *)(uintptr_t)key[i].dttk_value,
1809 			    (void *)kdata, kesize);
1810 			dkey->dttk_value = kdata;
1811 			kdata += P2ROUNDUP(kesize, sizeof (uint64_t));
1812 		} else {
1813 			dkey->dttk_value = key[i].dttk_value;
1814 		}
1815 
1816 		dkey->dttk_size = kesize;
1817 	}
1818 
1819 	ASSERT(dvar->dtdv_hashval == DTRACE_DYNHASH_FREE);
1820 	dvar->dtdv_hashval = hashval;
1821 	dvar->dtdv_next = start;
1822 
1823 	if (dtrace_casptr(&hash[bucket].dtdh_chain, start, dvar) == start)
1824 		return (dvar);
1825 
1826 	/*
1827 	 * The cas has failed.  Either another CPU is adding an element to
1828 	 * this hash chain, or another CPU is deleting an element from this
1829 	 * hash chain.  The simplest way to deal with both of these cases
1830 	 * (though not necessarily the most efficient) is to free our
1831 	 * allocated block and tail-call ourselves.  Note that the free is
1832 	 * to the dirty list and _not_ to the free list.  This is to prevent
1833 	 * races with allocators, above.
1834 	 */
1835 	dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
1836 
1837 	dtrace_membar_producer();
1838 
1839 	do {
1840 		free = dcpu->dtdsc_dirty;
1841 		dvar->dtdv_next = free;
1842 	} while (dtrace_casptr(&dcpu->dtdsc_dirty, free, dvar) != free);
1843 
1844 	return (dtrace_dynvar(dstate, nkeys, key, dsize, op, mstate, vstate));
1845 }
1846 
1847 /*ARGSUSED*/
1848 static void
1849 dtrace_aggregate_min(uint64_t *oval, uint64_t nval, uint64_t arg)
1850 {
1851 	if ((int64_t)nval < (int64_t)*oval)
1852 		*oval = nval;
1853 }
1854 
1855 /*ARGSUSED*/
1856 static void
1857 dtrace_aggregate_max(uint64_t *oval, uint64_t nval, uint64_t arg)
1858 {
1859 	if ((int64_t)nval > (int64_t)*oval)
1860 		*oval = nval;
1861 }
1862 
1863 static void
1864 dtrace_aggregate_quantize(uint64_t *quanta, uint64_t nval, uint64_t incr)
1865 {
1866 	int i, zero = DTRACE_QUANTIZE_ZEROBUCKET;
1867 	int64_t val = (int64_t)nval;
1868 
1869 	if (val < 0) {
1870 		for (i = 0; i < zero; i++) {
1871 			if (val <= DTRACE_QUANTIZE_BUCKETVAL(i)) {
1872 				quanta[i] += incr;
1873 				return;
1874 			}
1875 		}
1876 	} else {
1877 		for (i = zero + 1; i < DTRACE_QUANTIZE_NBUCKETS; i++) {
1878 			if (val < DTRACE_QUANTIZE_BUCKETVAL(i)) {
1879 				quanta[i - 1] += incr;
1880 				return;
1881 			}
1882 		}
1883 
1884 		quanta[DTRACE_QUANTIZE_NBUCKETS - 1] += incr;
1885 		return;
1886 	}
1887 
1888 	ASSERT(0);
1889 }
1890 
1891 static void
1892 dtrace_aggregate_lquantize(uint64_t *lquanta, uint64_t nval, uint64_t incr)
1893 {
1894 	uint64_t arg = *lquanta++;
1895 	int32_t base = DTRACE_LQUANTIZE_BASE(arg);
1896 	uint16_t step = DTRACE_LQUANTIZE_STEP(arg);
1897 	uint16_t levels = DTRACE_LQUANTIZE_LEVELS(arg);
1898 	int32_t val = (int32_t)nval, level;
1899 
1900 	ASSERT(step != 0);
1901 	ASSERT(levels != 0);
1902 
1903 	if (val < base) {
1904 		/*
1905 		 * This is an underflow.
1906 		 */
1907 		lquanta[0] += incr;
1908 		return;
1909 	}
1910 
1911 	level = (val - base) / step;
1912 
1913 	if (level < levels) {
1914 		lquanta[level + 1] += incr;
1915 		return;
1916 	}
1917 
1918 	/*
1919 	 * This is an overflow.
1920 	 */
1921 	lquanta[levels + 1] += incr;
1922 }
1923 
1924 static int
1925 dtrace_aggregate_llquantize_bucket(uint16_t factor, uint16_t low,
1926     uint16_t high, uint16_t nsteps, int64_t value)
1927 {
1928 	int64_t this = 1, last, next;
1929 	int base = 1, order;
1930 
1931 	ASSERT(factor <= nsteps);
1932 	ASSERT(nsteps % factor == 0);
1933 
1934 	for (order = 0; order < low; order++)
1935 		this *= factor;
1936 
1937 	/*
1938 	 * If our value is less than our factor taken to the power of the
1939 	 * low order of magnitude, it goes into the zeroth bucket.
1940 	 */
1941 	if (value < (last = this))
1942 		return (0);
1943 
1944 	for (this *= factor; order <= high; order++) {
1945 		int nbuckets = this > nsteps ? nsteps : this;
1946 
1947 		if ((next = this * factor) < this) {
1948 			/*
1949 			 * We should not generally get log/linear quantizations
1950 			 * with a high magnitude that allows 64-bits to
1951 			 * overflow, but we nonetheless protect against this
1952 			 * by explicitly checking for overflow, and clamping
1953 			 * our value accordingly.
1954 			 */
1955 			value = this - 1;
1956 		}
1957 
1958 		if (value < this) {
1959 			/*
1960 			 * If our value lies within this order of magnitude,
1961 			 * determine its position by taking the offset within
1962 			 * the order of magnitude, dividing by the bucket
1963 			 * width, and adding to our (accumulated) base.
1964 			 */
1965 			return (base + (value - last) / (this / nbuckets));
1966 		}
1967 
1968 		base += nbuckets - (nbuckets / factor);
1969 		last = this;
1970 		this = next;
1971 	}
1972 
1973 	/*
1974 	 * Our value is greater than or equal to our factor taken to the
1975 	 * power of one plus the high magnitude -- return the top bucket.
1976 	 */
1977 	return (base);
1978 }
1979 
1980 static void
1981 dtrace_aggregate_llquantize(uint64_t *llquanta, uint64_t nval, uint64_t incr)
1982 {
1983 	uint64_t arg = *llquanta++;
1984 	uint16_t factor = DTRACE_LLQUANTIZE_FACTOR(arg);
1985 	uint16_t low = DTRACE_LLQUANTIZE_LOW(arg);
1986 	uint16_t high = DTRACE_LLQUANTIZE_HIGH(arg);
1987 	uint16_t nsteps = DTRACE_LLQUANTIZE_NSTEP(arg);
1988 
1989 	llquanta[dtrace_aggregate_llquantize_bucket(factor,
1990 	    low, high, nsteps, nval)] += incr;
1991 }
1992 
1993 /*ARGSUSED*/
1994 static void
1995 dtrace_aggregate_avg(uint64_t *data, uint64_t nval, uint64_t arg)
1996 {
1997 	data[0]++;
1998 	data[1] += nval;
1999 }
2000 
2001 /*ARGSUSED*/
2002 static void
2003 dtrace_aggregate_stddev(uint64_t *data, uint64_t nval, uint64_t arg)
2004 {
2005 	int64_t snval = (int64_t)nval;
2006 	uint64_t tmp[2];
2007 
2008 	data[0]++;
2009 	data[1] += nval;
2010 
2011 	/*
2012 	 * What we want to say here is:
2013 	 *
2014 	 * data[2] += nval * nval;
2015 	 *
2016 	 * But given that nval is 64-bit, we could easily overflow, so
2017 	 * we do this as 128-bit arithmetic.
2018 	 */
2019 	if (snval < 0)
2020 		snval = -snval;
2021 
2022 	dtrace_multiply_128((uint64_t)snval, (uint64_t)snval, tmp);
2023 	dtrace_add_128(data + 2, tmp, data + 2);
2024 }
2025 
2026 /*ARGSUSED*/
2027 static void
2028 dtrace_aggregate_count(uint64_t *oval, uint64_t nval, uint64_t arg)
2029 {
2030 	*oval = *oval + 1;
2031 }
2032 
2033 /*ARGSUSED*/
2034 static void
2035 dtrace_aggregate_sum(uint64_t *oval, uint64_t nval, uint64_t arg)
2036 {
2037 	*oval += nval;
2038 }
2039 
2040 /*
2041  * Aggregate given the tuple in the principal data buffer, and the aggregating
2042  * action denoted by the specified dtrace_aggregation_t.  The aggregation
2043  * buffer is specified as the buf parameter.  This routine does not return
2044  * failure; if there is no space in the aggregation buffer, the data will be
2045  * dropped, and a corresponding counter incremented.
2046  */
2047 static void
2048 dtrace_aggregate(dtrace_aggregation_t *agg, dtrace_buffer_t *dbuf,
2049     intptr_t offset, dtrace_buffer_t *buf, uint64_t expr, uint64_t arg)
2050 {
2051 	dtrace_recdesc_t *rec = &agg->dtag_action.dta_rec;
2052 	uint32_t i, ndx, size, fsize;
2053 	uint32_t align = sizeof (uint64_t) - 1;
2054 	dtrace_aggbuffer_t *agb;
2055 	dtrace_aggkey_t *key;
2056 	uint32_t hashval = 0, limit, isstr;
2057 	caddr_t tomax, data, kdata;
2058 	dtrace_actkind_t action;
2059 	dtrace_action_t *act;
2060 	uintptr_t offs;
2061 
2062 	if (buf == NULL)
2063 		return;
2064 
2065 	if (!agg->dtag_hasarg) {
2066 		/*
2067 		 * Currently, only quantize() and lquantize() take additional
2068 		 * arguments, and they have the same semantics:  an increment
2069 		 * value that defaults to 1 when not present.  If additional
2070 		 * aggregating actions take arguments, the setting of the
2071 		 * default argument value will presumably have to become more
2072 		 * sophisticated...
2073 		 */
2074 		arg = 1;
2075 	}
2076 
2077 	action = agg->dtag_action.dta_kind - DTRACEACT_AGGREGATION;
2078 	size = rec->dtrd_offset - agg->dtag_base;
2079 	fsize = size + rec->dtrd_size;
2080 
2081 	ASSERT(dbuf->dtb_tomax != NULL);
2082 	data = dbuf->dtb_tomax + offset + agg->dtag_base;
2083 
2084 	if ((tomax = buf->dtb_tomax) == NULL) {
2085 		dtrace_buffer_drop(buf);
2086 		return;
2087 	}
2088 
2089 	/*
2090 	 * The metastructure is always at the bottom of the buffer.
2091 	 */
2092 	agb = (dtrace_aggbuffer_t *)(tomax + buf->dtb_size -
2093 	    sizeof (dtrace_aggbuffer_t));
2094 
2095 	if (buf->dtb_offset == 0) {
2096 		/*
2097 		 * We just kludge up approximately 1/8th of the size to be
2098 		 * buckets.  If this guess ends up being routinely
2099 		 * off-the-mark, we may need to dynamically readjust this
2100 		 * based on past performance.
2101 		 */
2102 		uintptr_t hashsize = (buf->dtb_size >> 3) / sizeof (uintptr_t);
2103 
2104 		if ((uintptr_t)agb - hashsize * sizeof (dtrace_aggkey_t *) <
2105 		    (uintptr_t)tomax || hashsize == 0) {
2106 			/*
2107 			 * We've been given a ludicrously small buffer;
2108 			 * increment our drop count and leave.
2109 			 */
2110 			dtrace_buffer_drop(buf);
2111 			return;
2112 		}
2113 
2114 		/*
2115 		 * And now, a pathetic attempt to try to get a an odd (or
2116 		 * perchance, a prime) hash size for better hash distribution.
2117 		 */
2118 		if (hashsize > (DTRACE_AGGHASHSIZE_SLEW << 3))
2119 			hashsize -= DTRACE_AGGHASHSIZE_SLEW;
2120 
2121 		agb->dtagb_hashsize = hashsize;
2122 		agb->dtagb_hash = (dtrace_aggkey_t **)((uintptr_t)agb -
2123 		    agb->dtagb_hashsize * sizeof (dtrace_aggkey_t *));
2124 		agb->dtagb_free = (uintptr_t)agb->dtagb_hash;
2125 
2126 		for (i = 0; i < agb->dtagb_hashsize; i++)
2127 			agb->dtagb_hash[i] = NULL;
2128 	}
2129 
2130 	ASSERT(agg->dtag_first != NULL);
2131 	ASSERT(agg->dtag_first->dta_intuple);
2132 
2133 	/*
2134 	 * Calculate the hash value based on the key.  Note that we _don't_
2135 	 * include the aggid in the hashing (but we will store it as part of
2136 	 * the key).  The hashing algorithm is Bob Jenkins' "One-at-a-time"
2137 	 * algorithm: a simple, quick algorithm that has no known funnels, and
2138 	 * gets good distribution in practice.  The efficacy of the hashing
2139 	 * algorithm (and a comparison with other algorithms) may be found by
2140 	 * running the ::dtrace_aggstat MDB dcmd.
2141 	 */
2142 	for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
2143 		i = act->dta_rec.dtrd_offset - agg->dtag_base;
2144 		limit = i + act->dta_rec.dtrd_size;
2145 		ASSERT(limit <= size);
2146 		isstr = DTRACEACT_ISSTRING(act);
2147 
2148 		for (; i < limit; i++) {
2149 			hashval += data[i];
2150 			hashval += (hashval << 10);
2151 			hashval ^= (hashval >> 6);
2152 
2153 			if (isstr && data[i] == '\0')
2154 				break;
2155 		}
2156 	}
2157 
2158 	hashval += (hashval << 3);
2159 	hashval ^= (hashval >> 11);
2160 	hashval += (hashval << 15);
2161 
2162 	/*
2163 	 * Yes, the divide here is expensive -- but it's generally the least
2164 	 * of the performance issues given the amount of data that we iterate
2165 	 * over to compute hash values, compare data, etc.
2166 	 */
2167 	ndx = hashval % agb->dtagb_hashsize;
2168 
2169 	for (key = agb->dtagb_hash[ndx]; key != NULL; key = key->dtak_next) {
2170 		ASSERT((caddr_t)key >= tomax);
2171 		ASSERT((caddr_t)key < tomax + buf->dtb_size);
2172 
2173 		if (hashval != key->dtak_hashval || key->dtak_size != size)
2174 			continue;
2175 
2176 		kdata = key->dtak_data;
2177 		ASSERT(kdata >= tomax && kdata < tomax + buf->dtb_size);
2178 
2179 		for (act = agg->dtag_first; act->dta_intuple;
2180 		    act = act->dta_next) {
2181 			i = act->dta_rec.dtrd_offset - agg->dtag_base;
2182 			limit = i + act->dta_rec.dtrd_size;
2183 			ASSERT(limit <= size);
2184 			isstr = DTRACEACT_ISSTRING(act);
2185 
2186 			for (; i < limit; i++) {
2187 				if (kdata[i] != data[i])
2188 					goto next;
2189 
2190 				if (isstr && data[i] == '\0')
2191 					break;
2192 			}
2193 		}
2194 
2195 		if (action != key->dtak_action) {
2196 			/*
2197 			 * We are aggregating on the same value in the same
2198 			 * aggregation with two different aggregating actions.
2199 			 * (This should have been picked up in the compiler,
2200 			 * so we may be dealing with errant or devious DIF.)
2201 			 * This is an error condition; we indicate as much,
2202 			 * and return.
2203 			 */
2204 			DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
2205 			return;
2206 		}
2207 
2208 		/*
2209 		 * This is a hit:  we need to apply the aggregator to
2210 		 * the value at this key.
2211 		 */
2212 		agg->dtag_aggregate((uint64_t *)(kdata + size), expr, arg);
2213 		return;
2214 next:
2215 		continue;
2216 	}
2217 
2218 	/*
2219 	 * We didn't find it.  We need to allocate some zero-filled space,
2220 	 * link it into the hash table appropriately, and apply the aggregator
2221 	 * to the (zero-filled) value.
2222 	 */
2223 	offs = buf->dtb_offset;
2224 	while (offs & (align - 1))
2225 		offs += sizeof (uint32_t);
2226 
2227 	/*
2228 	 * If we don't have enough room to both allocate a new key _and_
2229 	 * its associated data, increment the drop count and return.
2230 	 */
2231 	if ((uintptr_t)tomax + offs + fsize >
2232 	    agb->dtagb_free - sizeof (dtrace_aggkey_t)) {
2233 		dtrace_buffer_drop(buf);
2234 		return;
2235 	}
2236 
2237 	/*CONSTCOND*/
2238 	ASSERT(!(sizeof (dtrace_aggkey_t) & (sizeof (uintptr_t) - 1)));
2239 	key = (dtrace_aggkey_t *)(agb->dtagb_free - sizeof (dtrace_aggkey_t));
2240 	agb->dtagb_free -= sizeof (dtrace_aggkey_t);
2241 
2242 	key->dtak_data = kdata = tomax + offs;
2243 	buf->dtb_offset = offs + fsize;
2244 
2245 	/*
2246 	 * Now copy the data across.
2247 	 */
2248 	*((dtrace_aggid_t *)kdata) = agg->dtag_id;
2249 
2250 	for (i = sizeof (dtrace_aggid_t); i < size; i++)
2251 		kdata[i] = data[i];
2252 
2253 	/*
2254 	 * Because strings are not zeroed out by default, we need to iterate
2255 	 * looking for actions that store strings, and we need to explicitly
2256 	 * pad these strings out with zeroes.
2257 	 */
2258 	for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
2259 		int nul;
2260 
2261 		if (!DTRACEACT_ISSTRING(act))
2262 			continue;
2263 
2264 		i = act->dta_rec.dtrd_offset - agg->dtag_base;
2265 		limit = i + act->dta_rec.dtrd_size;
2266 		ASSERT(limit <= size);
2267 
2268 		for (nul = 0; i < limit; i++) {
2269 			if (nul) {
2270 				kdata[i] = '\0';
2271 				continue;
2272 			}
2273 
2274 			if (data[i] != '\0')
2275 				continue;
2276 
2277 			nul = 1;
2278 		}
2279 	}
2280 
2281 	for (i = size; i < fsize; i++)
2282 		kdata[i] = 0;
2283 
2284 	key->dtak_hashval = hashval;
2285 	key->dtak_size = size;
2286 	key->dtak_action = action;
2287 	key->dtak_next = agb->dtagb_hash[ndx];
2288 	agb->dtagb_hash[ndx] = key;
2289 
2290 	/*
2291 	 * Finally, apply the aggregator.
2292 	 */
2293 	*((uint64_t *)(key->dtak_data + size)) = agg->dtag_initial;
2294 	agg->dtag_aggregate((uint64_t *)(key->dtak_data + size), expr, arg);
2295 }
2296 
2297 /*
2298  * Given consumer state, this routine finds a speculation in the INACTIVE
2299  * state and transitions it into the ACTIVE state.  If there is no speculation
2300  * in the INACTIVE state, 0 is returned.  In this case, no error counter is
2301  * incremented -- it is up to the caller to take appropriate action.
2302  */
2303 static int
2304 dtrace_speculation(dtrace_state_t *state)
2305 {
2306 	int i = 0;
2307 	dtrace_speculation_state_t current;
2308 	uint32_t *stat = &state->dts_speculations_unavail, count;
2309 
2310 	while (i < state->dts_nspeculations) {
2311 		dtrace_speculation_t *spec = &state->dts_speculations[i];
2312 
2313 		current = spec->dtsp_state;
2314 
2315 		if (current != DTRACESPEC_INACTIVE) {
2316 			if (current == DTRACESPEC_COMMITTINGMANY ||
2317 			    current == DTRACESPEC_COMMITTING ||
2318 			    current == DTRACESPEC_DISCARDING)
2319 				stat = &state->dts_speculations_busy;
2320 			i++;
2321 			continue;
2322 		}
2323 
2324 		if (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2325 		    current, DTRACESPEC_ACTIVE) == current)
2326 			return (i + 1);
2327 	}
2328 
2329 	/*
2330 	 * We couldn't find a speculation.  If we found as much as a single
2331 	 * busy speculation buffer, we'll attribute this failure as "busy"
2332 	 * instead of "unavail".
2333 	 */
2334 	do {
2335 		count = *stat;
2336 	} while (dtrace_cas32(stat, count, count + 1) != count);
2337 
2338 	return (0);
2339 }
2340 
2341 /*
2342  * This routine commits an active speculation.  If the specified speculation
2343  * is not in a valid state to perform a commit(), this routine will silently do
2344  * nothing.  The state of the specified speculation is transitioned according
2345  * to the state transition diagram outlined in <sys/dtrace_impl.h>
2346  */
2347 static void
2348 dtrace_speculation_commit(dtrace_state_t *state, processorid_t cpu,
2349     dtrace_specid_t which)
2350 {
2351 	dtrace_speculation_t *spec;
2352 	dtrace_buffer_t *src, *dest;
2353 	uintptr_t daddr, saddr, dlimit, slimit;
2354 	dtrace_speculation_state_t current, new = 0;
2355 	intptr_t offs;
2356 	uint64_t timestamp;
2357 
2358 	if (which == 0)
2359 		return;
2360 
2361 	if (which > state->dts_nspeculations) {
2362 		cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2363 		return;
2364 	}
2365 
2366 	spec = &state->dts_speculations[which - 1];
2367 	src = &spec->dtsp_buffer[cpu];
2368 	dest = &state->dts_buffer[cpu];
2369 
2370 	do {
2371 		current = spec->dtsp_state;
2372 
2373 		if (current == DTRACESPEC_COMMITTINGMANY)
2374 			break;
2375 
2376 		switch (current) {
2377 		case DTRACESPEC_INACTIVE:
2378 		case DTRACESPEC_DISCARDING:
2379 			return;
2380 
2381 		case DTRACESPEC_COMMITTING:
2382 			/*
2383 			 * This is only possible if we are (a) commit()'ing
2384 			 * without having done a prior speculate() on this CPU
2385 			 * and (b) racing with another commit() on a different
2386 			 * CPU.  There's nothing to do -- we just assert that
2387 			 * our offset is 0.
2388 			 */
2389 			ASSERT(src->dtb_offset == 0);
2390 			return;
2391 
2392 		case DTRACESPEC_ACTIVE:
2393 			new = DTRACESPEC_COMMITTING;
2394 			break;
2395 
2396 		case DTRACESPEC_ACTIVEONE:
2397 			/*
2398 			 * This speculation is active on one CPU.  If our
2399 			 * buffer offset is non-zero, we know that the one CPU
2400 			 * must be us.  Otherwise, we are committing on a
2401 			 * different CPU from the speculate(), and we must
2402 			 * rely on being asynchronously cleaned.
2403 			 */
2404 			if (src->dtb_offset != 0) {
2405 				new = DTRACESPEC_COMMITTING;
2406 				break;
2407 			}
2408 			/*FALLTHROUGH*/
2409 
2410 		case DTRACESPEC_ACTIVEMANY:
2411 			new = DTRACESPEC_COMMITTINGMANY;
2412 			break;
2413 
2414 		default:
2415 			ASSERT(0);
2416 		}
2417 	} while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2418 	    current, new) != current);
2419 
2420 	/*
2421 	 * We have set the state to indicate that we are committing this
2422 	 * speculation.  Now reserve the necessary space in the destination
2423 	 * buffer.
2424 	 */
2425 	if ((offs = dtrace_buffer_reserve(dest, src->dtb_offset,
2426 	    sizeof (uint64_t), state, NULL)) < 0) {
2427 		dtrace_buffer_drop(dest);
2428 		goto out;
2429 	}
2430 
2431 	/*
2432 	 * We have sufficient space to copy the speculative buffer into the
2433 	 * primary buffer.  First, modify the speculative buffer, filling
2434 	 * in the timestamp of all entries with the current time.  The data
2435 	 * must have the commit() time rather than the time it was traced,
2436 	 * so that all entries in the primary buffer are in timestamp order.
2437 	 */
2438 	timestamp = dtrace_gethrtime();
2439 	saddr = (uintptr_t)src->dtb_tomax;
2440 	slimit = saddr + src->dtb_offset;
2441 	while (saddr < slimit) {
2442 		size_t size;
2443 		dtrace_rechdr_t *dtrh = (dtrace_rechdr_t *)saddr;
2444 
2445 		if (dtrh->dtrh_epid == DTRACE_EPIDNONE) {
2446 			saddr += sizeof (dtrace_epid_t);
2447 			continue;
2448 		}
2449 		ASSERT3U(dtrh->dtrh_epid, <=, state->dts_necbs);
2450 		size = state->dts_ecbs[dtrh->dtrh_epid - 1]->dte_size;
2451 
2452 		ASSERT3U(saddr + size, <=, slimit);
2453 		ASSERT3U(size, >=, sizeof (dtrace_rechdr_t));
2454 		ASSERT3U(DTRACE_RECORD_LOAD_TIMESTAMP(dtrh), ==, UINT64_MAX);
2455 
2456 		DTRACE_RECORD_STORE_TIMESTAMP(dtrh, timestamp);
2457 
2458 		saddr += size;
2459 	}
2460 
2461 	/*
2462 	 * Copy the buffer across.  (Note that this is a
2463 	 * highly subobtimal bcopy(); in the unlikely event that this becomes
2464 	 * a serious performance issue, a high-performance DTrace-specific
2465 	 * bcopy() should obviously be invented.)
2466 	 */
2467 	daddr = (uintptr_t)dest->dtb_tomax + offs;
2468 	dlimit = daddr + src->dtb_offset;
2469 	saddr = (uintptr_t)src->dtb_tomax;
2470 
2471 	/*
2472 	 * First, the aligned portion.
2473 	 */
2474 	while (dlimit - daddr >= sizeof (uint64_t)) {
2475 		*((uint64_t *)daddr) = *((uint64_t *)saddr);
2476 
2477 		daddr += sizeof (uint64_t);
2478 		saddr += sizeof (uint64_t);
2479 	}
2480 
2481 	/*
2482 	 * Now any left-over bit...
2483 	 */
2484 	while (dlimit - daddr)
2485 		*((uint8_t *)daddr++) = *((uint8_t *)saddr++);
2486 
2487 	/*
2488 	 * Finally, commit the reserved space in the destination buffer.
2489 	 */
2490 	dest->dtb_offset = offs + src->dtb_offset;
2491 
2492 out:
2493 	/*
2494 	 * If we're lucky enough to be the only active CPU on this speculation
2495 	 * buffer, we can just set the state back to DTRACESPEC_INACTIVE.
2496 	 */
2497 	if (current == DTRACESPEC_ACTIVE ||
2498 	    (current == DTRACESPEC_ACTIVEONE && new == DTRACESPEC_COMMITTING)) {
2499 		uint32_t rval = dtrace_cas32((uint32_t *)&spec->dtsp_state,
2500 		    DTRACESPEC_COMMITTING, DTRACESPEC_INACTIVE);
2501 
2502 		ASSERT(rval == DTRACESPEC_COMMITTING);
2503 	}
2504 
2505 	src->dtb_offset = 0;
2506 	src->dtb_xamot_drops += src->dtb_drops;
2507 	src->dtb_drops = 0;
2508 }
2509 
2510 /*
2511  * This routine discards an active speculation.  If the specified speculation
2512  * is not in a valid state to perform a discard(), this routine will silently
2513  * do nothing.  The state of the specified speculation is transitioned
2514  * according to the state transition diagram outlined in <sys/dtrace_impl.h>
2515  */
2516 static void
2517 dtrace_speculation_discard(dtrace_state_t *state, processorid_t cpu,
2518     dtrace_specid_t which)
2519 {
2520 	dtrace_speculation_t *spec;
2521 	dtrace_speculation_state_t current, new = 0;
2522 	dtrace_buffer_t *buf;
2523 
2524 	if (which == 0)
2525 		return;
2526 
2527 	if (which > state->dts_nspeculations) {
2528 		cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2529 		return;
2530 	}
2531 
2532 	spec = &state->dts_speculations[which - 1];
2533 	buf = &spec->dtsp_buffer[cpu];
2534 
2535 	do {
2536 		current = spec->dtsp_state;
2537 
2538 		switch (current) {
2539 		case DTRACESPEC_INACTIVE:
2540 		case DTRACESPEC_COMMITTINGMANY:
2541 		case DTRACESPEC_COMMITTING:
2542 		case DTRACESPEC_DISCARDING:
2543 			return;
2544 
2545 		case DTRACESPEC_ACTIVE:
2546 		case DTRACESPEC_ACTIVEMANY:
2547 			new = DTRACESPEC_DISCARDING;
2548 			break;
2549 
2550 		case DTRACESPEC_ACTIVEONE:
2551 			if (buf->dtb_offset != 0) {
2552 				new = DTRACESPEC_INACTIVE;
2553 			} else {
2554 				new = DTRACESPEC_DISCARDING;
2555 			}
2556 			break;
2557 
2558 		default:
2559 			ASSERT(0);
2560 		}
2561 	} while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2562 	    current, new) != current);
2563 
2564 	buf->dtb_offset = 0;
2565 	buf->dtb_drops = 0;
2566 }
2567 
2568 /*
2569  * Note:  not called from probe context.  This function is called
2570  * asynchronously from cross call context to clean any speculations that are
2571  * in the COMMITTINGMANY or DISCARDING states.  These speculations may not be
2572  * transitioned back to the INACTIVE state until all CPUs have cleaned the
2573  * speculation.
2574  */
2575 static void
2576 dtrace_speculation_clean_here(dtrace_state_t *state)
2577 {
2578 	dtrace_icookie_t cookie;
2579 	processorid_t cpu = curcpu;
2580 	dtrace_buffer_t *dest = &state->dts_buffer[cpu];
2581 	dtrace_specid_t i;
2582 
2583 	cookie = dtrace_interrupt_disable();
2584 
2585 	if (dest->dtb_tomax == NULL) {
2586 		dtrace_interrupt_enable(cookie);
2587 		return;
2588 	}
2589 
2590 	for (i = 0; i < state->dts_nspeculations; i++) {
2591 		dtrace_speculation_t *spec = &state->dts_speculations[i];
2592 		dtrace_buffer_t *src = &spec->dtsp_buffer[cpu];
2593 
2594 		if (src->dtb_tomax == NULL)
2595 			continue;
2596 
2597 		if (spec->dtsp_state == DTRACESPEC_DISCARDING) {
2598 			src->dtb_offset = 0;
2599 			continue;
2600 		}
2601 
2602 		if (spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
2603 			continue;
2604 
2605 		if (src->dtb_offset == 0)
2606 			continue;
2607 
2608 		dtrace_speculation_commit(state, cpu, i + 1);
2609 	}
2610 
2611 	dtrace_interrupt_enable(cookie);
2612 }
2613 
2614 /*
2615  * Note:  not called from probe context.  This function is called
2616  * asynchronously (and at a regular interval) to clean any speculations that
2617  * are in the COMMITTINGMANY or DISCARDING states.  If it discovers that there
2618  * is work to be done, it cross calls all CPUs to perform that work;
2619  * COMMITMANY and DISCARDING speculations may not be transitioned back to the
2620  * INACTIVE state until they have been cleaned by all CPUs.
2621  */
2622 static void
2623 dtrace_speculation_clean(dtrace_state_t *state)
2624 {
2625 	int work = 0, rv;
2626 	dtrace_specid_t i;
2627 
2628 	for (i = 0; i < state->dts_nspeculations; i++) {
2629 		dtrace_speculation_t *spec = &state->dts_speculations[i];
2630 
2631 		ASSERT(!spec->dtsp_cleaning);
2632 
2633 		if (spec->dtsp_state != DTRACESPEC_DISCARDING &&
2634 		    spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
2635 			continue;
2636 
2637 		work++;
2638 		spec->dtsp_cleaning = 1;
2639 	}
2640 
2641 	if (!work)
2642 		return;
2643 
2644 	dtrace_xcall(DTRACE_CPUALL,
2645 	    (dtrace_xcall_t)dtrace_speculation_clean_here, state);
2646 
2647 	/*
2648 	 * We now know that all CPUs have committed or discarded their
2649 	 * speculation buffers, as appropriate.  We can now set the state
2650 	 * to inactive.
2651 	 */
2652 	for (i = 0; i < state->dts_nspeculations; i++) {
2653 		dtrace_speculation_t *spec = &state->dts_speculations[i];
2654 		dtrace_speculation_state_t current, new;
2655 
2656 		if (!spec->dtsp_cleaning)
2657 			continue;
2658 
2659 		current = spec->dtsp_state;
2660 		ASSERT(current == DTRACESPEC_DISCARDING ||
2661 		    current == DTRACESPEC_COMMITTINGMANY);
2662 
2663 		new = DTRACESPEC_INACTIVE;
2664 
2665 		rv = dtrace_cas32((uint32_t *)&spec->dtsp_state, current, new);
2666 		ASSERT(rv == current);
2667 		spec->dtsp_cleaning = 0;
2668 	}
2669 }
2670 
2671 /*
2672  * Called as part of a speculate() to get the speculative buffer associated
2673  * with a given speculation.  Returns NULL if the specified speculation is not
2674  * in an ACTIVE state.  If the speculation is in the ACTIVEONE state -- and
2675  * the active CPU is not the specified CPU -- the speculation will be
2676  * atomically transitioned into the ACTIVEMANY state.
2677  */
2678 static dtrace_buffer_t *
2679 dtrace_speculation_buffer(dtrace_state_t *state, processorid_t cpuid,
2680     dtrace_specid_t which)
2681 {
2682 	dtrace_speculation_t *spec;
2683 	dtrace_speculation_state_t current, new = 0;
2684 	dtrace_buffer_t *buf;
2685 
2686 	if (which == 0)
2687 		return (NULL);
2688 
2689 	if (which > state->dts_nspeculations) {
2690 		cpu_core[cpuid].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2691 		return (NULL);
2692 	}
2693 
2694 	spec = &state->dts_speculations[which - 1];
2695 	buf = &spec->dtsp_buffer[cpuid];
2696 
2697 	do {
2698 		current = spec->dtsp_state;
2699 
2700 		switch (current) {
2701 		case DTRACESPEC_INACTIVE:
2702 		case DTRACESPEC_COMMITTINGMANY:
2703 		case DTRACESPEC_DISCARDING:
2704 			return (NULL);
2705 
2706 		case DTRACESPEC_COMMITTING:
2707 			ASSERT(buf->dtb_offset == 0);
2708 			return (NULL);
2709 
2710 		case DTRACESPEC_ACTIVEONE:
2711 			/*
2712 			 * This speculation is currently active on one CPU.
2713 			 * Check the offset in the buffer; if it's non-zero,
2714 			 * that CPU must be us (and we leave the state alone).
2715 			 * If it's zero, assume that we're starting on a new
2716 			 * CPU -- and change the state to indicate that the
2717 			 * speculation is active on more than one CPU.
2718 			 */
2719 			if (buf->dtb_offset != 0)
2720 				return (buf);
2721 
2722 			new = DTRACESPEC_ACTIVEMANY;
2723 			break;
2724 
2725 		case DTRACESPEC_ACTIVEMANY:
2726 			return (buf);
2727 
2728 		case DTRACESPEC_ACTIVE:
2729 			new = DTRACESPEC_ACTIVEONE;
2730 			break;
2731 
2732 		default:
2733 			ASSERT(0);
2734 		}
2735 	} while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2736 	    current, new) != current);
2737 
2738 	ASSERT(new == DTRACESPEC_ACTIVEONE || new == DTRACESPEC_ACTIVEMANY);
2739 	return (buf);
2740 }
2741 
2742 /*
2743  * Return a string.  In the event that the user lacks the privilege to access
2744  * arbitrary kernel memory, we copy the string out to scratch memory so that we
2745  * don't fail access checking.
2746  *
2747  * dtrace_dif_variable() uses this routine as a helper for various
2748  * builtin values such as 'execname' and 'probefunc.'
2749  */
2750 uintptr_t
2751 dtrace_dif_varstr(uintptr_t addr, dtrace_state_t *state,
2752     dtrace_mstate_t *mstate)
2753 {
2754 	uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
2755 	uintptr_t ret;
2756 	size_t strsz;
2757 
2758 	/*
2759 	 * The easy case: this probe is allowed to read all of memory, so
2760 	 * we can just return this as a vanilla pointer.
2761 	 */
2762 	if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
2763 		return (addr);
2764 
2765 	/*
2766 	 * This is the tougher case: we copy the string in question from
2767 	 * kernel memory into scratch memory and return it that way: this
2768 	 * ensures that we won't trip up when access checking tests the
2769 	 * BYREF return value.
2770 	 */
2771 	strsz = dtrace_strlen((char *)addr, size) + 1;
2772 
2773 	if (mstate->dtms_scratch_ptr + strsz >
2774 	    mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
2775 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
2776 		return (0);
2777 	}
2778 
2779 	dtrace_strcpy((const void *)addr, (void *)mstate->dtms_scratch_ptr,
2780 	    strsz);
2781 	ret = mstate->dtms_scratch_ptr;
2782 	mstate->dtms_scratch_ptr += strsz;
2783 	return (ret);
2784 }
2785 
2786 /*
2787  * Return a string from a memoy address which is known to have one or
2788  * more concatenated, individually zero terminated, sub-strings.
2789  * In the event that the user lacks the privilege to access
2790  * arbitrary kernel memory, we copy the string out to scratch memory so that we
2791  * don't fail access checking.
2792  *
2793  * dtrace_dif_variable() uses this routine as a helper for various
2794  * builtin values such as 'execargs'.
2795  */
2796 static uintptr_t
2797 dtrace_dif_varstrz(uintptr_t addr, size_t strsz, dtrace_state_t *state,
2798     dtrace_mstate_t *mstate)
2799 {
2800 	char *p;
2801 	size_t i;
2802 	uintptr_t ret;
2803 
2804 	if (mstate->dtms_scratch_ptr + strsz >
2805 	    mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
2806 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
2807 		return (0);
2808 	}
2809 
2810 	dtrace_bcopy((const void *)addr, (void *)mstate->dtms_scratch_ptr,
2811 	    strsz);
2812 
2813 	/* Replace sub-string termination characters with a space. */
2814 	for (p = (char *) mstate->dtms_scratch_ptr, i = 0; i < strsz - 1;
2815 	    p++, i++)
2816 		if (*p == '\0')
2817 			*p = ' ';
2818 
2819 	ret = mstate->dtms_scratch_ptr;
2820 	mstate->dtms_scratch_ptr += strsz;
2821 	return (ret);
2822 }
2823 
2824 /*
2825  * This function implements the DIF emulator's variable lookups.  The emulator
2826  * passes a reserved variable identifier and optional built-in array index.
2827  */
2828 static uint64_t
2829 dtrace_dif_variable(dtrace_mstate_t *mstate, dtrace_state_t *state, uint64_t v,
2830     uint64_t ndx)
2831 {
2832 	/*
2833 	 * If we're accessing one of the uncached arguments, we'll turn this
2834 	 * into a reference in the args array.
2835 	 */
2836 	if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9) {
2837 		ndx = v - DIF_VAR_ARG0;
2838 		v = DIF_VAR_ARGS;
2839 	}
2840 
2841 	switch (v) {
2842 	case DIF_VAR_ARGS:
2843 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_ARGS);
2844 		if (ndx >= sizeof (mstate->dtms_arg) /
2845 		    sizeof (mstate->dtms_arg[0])) {
2846 			int aframes = mstate->dtms_probe->dtpr_aframes + 2;
2847 			dtrace_provider_t *pv;
2848 			uint64_t val;
2849 
2850 			pv = mstate->dtms_probe->dtpr_provider;
2851 			if (pv->dtpv_pops.dtps_getargval != NULL)
2852 				val = pv->dtpv_pops.dtps_getargval(pv->dtpv_arg,
2853 				    mstate->dtms_probe->dtpr_id,
2854 				    mstate->dtms_probe->dtpr_arg, ndx, aframes);
2855 			else
2856 				val = dtrace_getarg(ndx, aframes);
2857 
2858 			/*
2859 			 * This is regrettably required to keep the compiler
2860 			 * from tail-optimizing the call to dtrace_getarg().
2861 			 * The condition always evaluates to true, but the
2862 			 * compiler has no way of figuring that out a priori.
2863 			 * (None of this would be necessary if the compiler
2864 			 * could be relied upon to _always_ tail-optimize
2865 			 * the call to dtrace_getarg() -- but it can't.)
2866 			 */
2867 			if (mstate->dtms_probe != NULL)
2868 				return (val);
2869 
2870 			ASSERT(0);
2871 		}
2872 
2873 		return (mstate->dtms_arg[ndx]);
2874 
2875 #if defined(sun)
2876 	case DIF_VAR_UREGS: {
2877 		klwp_t *lwp;
2878 
2879 		if (!dtrace_priv_proc(state))
2880 			return (0);
2881 
2882 		if ((lwp = curthread->t_lwp) == NULL) {
2883 			DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
2884 			cpu_core[curcpu].cpuc_dtrace_illval = NULL;
2885 			return (0);
2886 		}
2887 
2888 		return (dtrace_getreg(lwp->lwp_regs, ndx));
2889 		return (0);
2890 	}
2891 #else
2892 	case DIF_VAR_UREGS: {
2893 		struct trapframe *tframe;
2894 
2895 		if (!dtrace_priv_proc(state))
2896 			return (0);
2897 
2898 		if ((tframe = curthread->td_frame) == NULL) {
2899 			DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
2900 			cpu_core[curcpu].cpuc_dtrace_illval = 0;
2901 			return (0);
2902 		}
2903 
2904 		return (dtrace_getreg(tframe, ndx));
2905 	}
2906 #endif
2907 
2908 	case DIF_VAR_CURTHREAD:
2909 		if (!dtrace_priv_kernel(state))
2910 			return (0);
2911 		return ((uint64_t)(uintptr_t)curthread);
2912 
2913 	case DIF_VAR_TIMESTAMP:
2914 		if (!(mstate->dtms_present & DTRACE_MSTATE_TIMESTAMP)) {
2915 			mstate->dtms_timestamp = dtrace_gethrtime();
2916 			mstate->dtms_present |= DTRACE_MSTATE_TIMESTAMP;
2917 		}
2918 		return (mstate->dtms_timestamp);
2919 
2920 	case DIF_VAR_VTIMESTAMP:
2921 		ASSERT(dtrace_vtime_references != 0);
2922 		return (curthread->t_dtrace_vtime);
2923 
2924 	case DIF_VAR_WALLTIMESTAMP:
2925 		if (!(mstate->dtms_present & DTRACE_MSTATE_WALLTIMESTAMP)) {
2926 			mstate->dtms_walltimestamp = dtrace_gethrestime();
2927 			mstate->dtms_present |= DTRACE_MSTATE_WALLTIMESTAMP;
2928 		}
2929 		return (mstate->dtms_walltimestamp);
2930 
2931 #if defined(sun)
2932 	case DIF_VAR_IPL:
2933 		if (!dtrace_priv_kernel(state))
2934 			return (0);
2935 		if (!(mstate->dtms_present & DTRACE_MSTATE_IPL)) {
2936 			mstate->dtms_ipl = dtrace_getipl();
2937 			mstate->dtms_present |= DTRACE_MSTATE_IPL;
2938 		}
2939 		return (mstate->dtms_ipl);
2940 #endif
2941 
2942 	case DIF_VAR_EPID:
2943 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_EPID);
2944 		return (mstate->dtms_epid);
2945 
2946 	case DIF_VAR_ID:
2947 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
2948 		return (mstate->dtms_probe->dtpr_id);
2949 
2950 	case DIF_VAR_STACKDEPTH:
2951 		if (!dtrace_priv_kernel(state))
2952 			return (0);
2953 		if (!(mstate->dtms_present & DTRACE_MSTATE_STACKDEPTH)) {
2954 			int aframes = mstate->dtms_probe->dtpr_aframes + 2;
2955 
2956 			mstate->dtms_stackdepth = dtrace_getstackdepth(aframes);
2957 			mstate->dtms_present |= DTRACE_MSTATE_STACKDEPTH;
2958 		}
2959 		return (mstate->dtms_stackdepth);
2960 
2961 	case DIF_VAR_USTACKDEPTH:
2962 		if (!dtrace_priv_proc(state))
2963 			return (0);
2964 		if (!(mstate->dtms_present & DTRACE_MSTATE_USTACKDEPTH)) {
2965 			/*
2966 			 * See comment in DIF_VAR_PID.
2967 			 */
2968 			if (DTRACE_ANCHORED(mstate->dtms_probe) &&
2969 			    CPU_ON_INTR(CPU)) {
2970 				mstate->dtms_ustackdepth = 0;
2971 			} else {
2972 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
2973 				mstate->dtms_ustackdepth =
2974 				    dtrace_getustackdepth();
2975 				DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
2976 			}
2977 			mstate->dtms_present |= DTRACE_MSTATE_USTACKDEPTH;
2978 		}
2979 		return (mstate->dtms_ustackdepth);
2980 
2981 	case DIF_VAR_CALLER:
2982 		if (!dtrace_priv_kernel(state))
2983 			return (0);
2984 		if (!(mstate->dtms_present & DTRACE_MSTATE_CALLER)) {
2985 			int aframes = mstate->dtms_probe->dtpr_aframes + 2;
2986 
2987 			if (!DTRACE_ANCHORED(mstate->dtms_probe)) {
2988 				/*
2989 				 * If this is an unanchored probe, we are
2990 				 * required to go through the slow path:
2991 				 * dtrace_caller() only guarantees correct
2992 				 * results for anchored probes.
2993 				 */
2994 				pc_t caller[2] = {0, 0};
2995 
2996 				dtrace_getpcstack(caller, 2, aframes,
2997 				    (uint32_t *)(uintptr_t)mstate->dtms_arg[0]);
2998 				mstate->dtms_caller = caller[1];
2999 			} else if ((mstate->dtms_caller =
3000 			    dtrace_caller(aframes)) == -1) {
3001 				/*
3002 				 * We have failed to do this the quick way;
3003 				 * we must resort to the slower approach of
3004 				 * calling dtrace_getpcstack().
3005 				 */
3006 				pc_t caller = 0;
3007 
3008 				dtrace_getpcstack(&caller, 1, aframes, NULL);
3009 				mstate->dtms_caller = caller;
3010 			}
3011 
3012 			mstate->dtms_present |= DTRACE_MSTATE_CALLER;
3013 		}
3014 		return (mstate->dtms_caller);
3015 
3016 	case DIF_VAR_UCALLER:
3017 		if (!dtrace_priv_proc(state))
3018 			return (0);
3019 
3020 		if (!(mstate->dtms_present & DTRACE_MSTATE_UCALLER)) {
3021 			uint64_t ustack[3];
3022 
3023 			/*
3024 			 * dtrace_getupcstack() fills in the first uint64_t
3025 			 * with the current PID.  The second uint64_t will
3026 			 * be the program counter at user-level.  The third
3027 			 * uint64_t will contain the caller, which is what
3028 			 * we're after.
3029 			 */
3030 			ustack[2] = 0;
3031 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3032 			dtrace_getupcstack(ustack, 3);
3033 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3034 			mstate->dtms_ucaller = ustack[2];
3035 			mstate->dtms_present |= DTRACE_MSTATE_UCALLER;
3036 		}
3037 
3038 		return (mstate->dtms_ucaller);
3039 
3040 	case DIF_VAR_PROBEPROV:
3041 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3042 		return (dtrace_dif_varstr(
3043 		    (uintptr_t)mstate->dtms_probe->dtpr_provider->dtpv_name,
3044 		    state, mstate));
3045 
3046 	case DIF_VAR_PROBEMOD:
3047 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3048 		return (dtrace_dif_varstr(
3049 		    (uintptr_t)mstate->dtms_probe->dtpr_mod,
3050 		    state, mstate));
3051 
3052 	case DIF_VAR_PROBEFUNC:
3053 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3054 		return (dtrace_dif_varstr(
3055 		    (uintptr_t)mstate->dtms_probe->dtpr_func,
3056 		    state, mstate));
3057 
3058 	case DIF_VAR_PROBENAME:
3059 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3060 		return (dtrace_dif_varstr(
3061 		    (uintptr_t)mstate->dtms_probe->dtpr_name,
3062 		    state, mstate));
3063 
3064 	case DIF_VAR_PID:
3065 		if (!dtrace_priv_proc(state))
3066 			return (0);
3067 
3068 #if defined(sun)
3069 		/*
3070 		 * Note that we are assuming that an unanchored probe is
3071 		 * always due to a high-level interrupt.  (And we're assuming
3072 		 * that there is only a single high level interrupt.)
3073 		 */
3074 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3075 			return (pid0.pid_id);
3076 
3077 		/*
3078 		 * It is always safe to dereference one's own t_procp pointer:
3079 		 * it always points to a valid, allocated proc structure.
3080 		 * Further, it is always safe to dereference the p_pidp member
3081 		 * of one's own proc structure.  (These are truisms becuase
3082 		 * threads and processes don't clean up their own state --
3083 		 * they leave that task to whomever reaps them.)
3084 		 */
3085 		return ((uint64_t)curthread->t_procp->p_pidp->pid_id);
3086 #else
3087 		return ((uint64_t)curproc->p_pid);
3088 #endif
3089 
3090 	case DIF_VAR_PPID:
3091 		if (!dtrace_priv_proc(state))
3092 			return (0);
3093 
3094 #if defined(sun)
3095 		/*
3096 		 * See comment in DIF_VAR_PID.
3097 		 */
3098 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3099 			return (pid0.pid_id);
3100 
3101 		/*
3102 		 * It is always safe to dereference one's own t_procp pointer:
3103 		 * it always points to a valid, allocated proc structure.
3104 		 * (This is true because threads don't clean up their own
3105 		 * state -- they leave that task to whomever reaps them.)
3106 		 */
3107 		return ((uint64_t)curthread->t_procp->p_ppid);
3108 #else
3109 		return ((uint64_t)curproc->p_pptr->p_pid);
3110 #endif
3111 
3112 	case DIF_VAR_TID:
3113 #if defined(sun)
3114 		/*
3115 		 * See comment in DIF_VAR_PID.
3116 		 */
3117 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3118 			return (0);
3119 #endif
3120 
3121 		return ((uint64_t)curthread->t_tid);
3122 
3123 	case DIF_VAR_EXECARGS: {
3124 		struct pargs *p_args = curthread->td_proc->p_args;
3125 
3126 		if (p_args == NULL)
3127 			return(0);
3128 
3129 		return (dtrace_dif_varstrz(
3130 		    (uintptr_t) p_args->ar_args, p_args->ar_length, state, mstate));
3131 	}
3132 
3133 	case DIF_VAR_EXECNAME:
3134 #if defined(sun)
3135 		if (!dtrace_priv_proc(state))
3136 			return (0);
3137 
3138 		/*
3139 		 * See comment in DIF_VAR_PID.
3140 		 */
3141 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3142 			return ((uint64_t)(uintptr_t)p0.p_user.u_comm);
3143 
3144 		/*
3145 		 * It is always safe to dereference one's own t_procp pointer:
3146 		 * it always points to a valid, allocated proc structure.
3147 		 * (This is true because threads don't clean up their own
3148 		 * state -- they leave that task to whomever reaps them.)
3149 		 */
3150 		return (dtrace_dif_varstr(
3151 		    (uintptr_t)curthread->t_procp->p_user.u_comm,
3152 		    state, mstate));
3153 #else
3154 		return (dtrace_dif_varstr(
3155 		    (uintptr_t) curthread->td_proc->p_comm, state, mstate));
3156 #endif
3157 
3158 	case DIF_VAR_ZONENAME:
3159 #if defined(sun)
3160 		if (!dtrace_priv_proc(state))
3161 			return (0);
3162 
3163 		/*
3164 		 * See comment in DIF_VAR_PID.
3165 		 */
3166 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3167 			return ((uint64_t)(uintptr_t)p0.p_zone->zone_name);
3168 
3169 		/*
3170 		 * It is always safe to dereference one's own t_procp pointer:
3171 		 * it always points to a valid, allocated proc structure.
3172 		 * (This is true because threads don't clean up their own
3173 		 * state -- they leave that task to whomever reaps them.)
3174 		 */
3175 		return (dtrace_dif_varstr(
3176 		    (uintptr_t)curthread->t_procp->p_zone->zone_name,
3177 		    state, mstate));
3178 #else
3179 		return (0);
3180 #endif
3181 
3182 	case DIF_VAR_UID:
3183 		if (!dtrace_priv_proc(state))
3184 			return (0);
3185 
3186 #if defined(sun)
3187 		/*
3188 		 * See comment in DIF_VAR_PID.
3189 		 */
3190 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3191 			return ((uint64_t)p0.p_cred->cr_uid);
3192 #endif
3193 
3194 		/*
3195 		 * It is always safe to dereference one's own t_procp pointer:
3196 		 * it always points to a valid, allocated proc structure.
3197 		 * (This is true because threads don't clean up their own
3198 		 * state -- they leave that task to whomever reaps them.)
3199 		 *
3200 		 * Additionally, it is safe to dereference one's own process
3201 		 * credential, since this is never NULL after process birth.
3202 		 */
3203 		return ((uint64_t)curthread->t_procp->p_cred->cr_uid);
3204 
3205 	case DIF_VAR_GID:
3206 		if (!dtrace_priv_proc(state))
3207 			return (0);
3208 
3209 #if defined(sun)
3210 		/*
3211 		 * See comment in DIF_VAR_PID.
3212 		 */
3213 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3214 			return ((uint64_t)p0.p_cred->cr_gid);
3215 #endif
3216 
3217 		/*
3218 		 * It is always safe to dereference one's own t_procp pointer:
3219 		 * it always points to a valid, allocated proc structure.
3220 		 * (This is true because threads don't clean up their own
3221 		 * state -- they leave that task to whomever reaps them.)
3222 		 *
3223 		 * Additionally, it is safe to dereference one's own process
3224 		 * credential, since this is never NULL after process birth.
3225 		 */
3226 		return ((uint64_t)curthread->t_procp->p_cred->cr_gid);
3227 
3228 	case DIF_VAR_ERRNO: {
3229 #if defined(sun)
3230 		klwp_t *lwp;
3231 		if (!dtrace_priv_proc(state))
3232 			return (0);
3233 
3234 		/*
3235 		 * See comment in DIF_VAR_PID.
3236 		 */
3237 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3238 			return (0);
3239 
3240 		/*
3241 		 * It is always safe to dereference one's own t_lwp pointer in
3242 		 * the event that this pointer is non-NULL.  (This is true
3243 		 * because threads and lwps don't clean up their own state --
3244 		 * they leave that task to whomever reaps them.)
3245 		 */
3246 		if ((lwp = curthread->t_lwp) == NULL)
3247 			return (0);
3248 
3249 		return ((uint64_t)lwp->lwp_errno);
3250 #else
3251 		return (curthread->td_errno);
3252 #endif
3253 	}
3254 #if !defined(sun)
3255 	case DIF_VAR_CPU: {
3256 		return curcpu;
3257 	}
3258 #endif
3259 	default:
3260 		DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
3261 		return (0);
3262 	}
3263 }
3264 
3265 /*
3266  * Emulate the execution of DTrace ID subroutines invoked by the call opcode.
3267  * Notice that we don't bother validating the proper number of arguments or
3268  * their types in the tuple stack.  This isn't needed because all argument
3269  * interpretation is safe because of our load safety -- the worst that can
3270  * happen is that a bogus program can obtain bogus results.
3271  */
3272 static void
3273 dtrace_dif_subr(uint_t subr, uint_t rd, uint64_t *regs,
3274     dtrace_key_t *tupregs, int nargs,
3275     dtrace_mstate_t *mstate, dtrace_state_t *state)
3276 {
3277 	volatile uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
3278 	volatile uintptr_t *illval = &cpu_core[curcpu].cpuc_dtrace_illval;
3279 	dtrace_vstate_t *vstate = &state->dts_vstate;
3280 
3281 #if defined(sun)
3282 	union {
3283 		mutex_impl_t mi;
3284 		uint64_t mx;
3285 	} m;
3286 
3287 	union {
3288 		krwlock_t ri;
3289 		uintptr_t rw;
3290 	} r;
3291 #else
3292 	struct thread *lowner;
3293 	union {
3294 		struct lock_object *li;
3295 		uintptr_t lx;
3296 	} l;
3297 #endif
3298 
3299 	switch (subr) {
3300 	case DIF_SUBR_RAND:
3301 		regs[rd] = (dtrace_gethrtime() * 2416 + 374441) % 1771875;
3302 		break;
3303 
3304 #if defined(sun)
3305 	case DIF_SUBR_MUTEX_OWNED:
3306 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3307 		    mstate, vstate)) {
3308 			regs[rd] = 0;
3309 			break;
3310 		}
3311 
3312 		m.mx = dtrace_load64(tupregs[0].dttk_value);
3313 		if (MUTEX_TYPE_ADAPTIVE(&m.mi))
3314 			regs[rd] = MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER;
3315 		else
3316 			regs[rd] = LOCK_HELD(&m.mi.m_spin.m_spinlock);
3317 		break;
3318 
3319 	case DIF_SUBR_MUTEX_OWNER:
3320 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3321 		    mstate, vstate)) {
3322 			regs[rd] = 0;
3323 			break;
3324 		}
3325 
3326 		m.mx = dtrace_load64(tupregs[0].dttk_value);
3327 		if (MUTEX_TYPE_ADAPTIVE(&m.mi) &&
3328 		    MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER)
3329 			regs[rd] = (uintptr_t)MUTEX_OWNER(&m.mi);
3330 		else
3331 			regs[rd] = 0;
3332 		break;
3333 
3334 	case DIF_SUBR_MUTEX_TYPE_ADAPTIVE:
3335 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3336 		    mstate, vstate)) {
3337 			regs[rd] = 0;
3338 			break;
3339 		}
3340 
3341 		m.mx = dtrace_load64(tupregs[0].dttk_value);
3342 		regs[rd] = MUTEX_TYPE_ADAPTIVE(&m.mi);
3343 		break;
3344 
3345 	case DIF_SUBR_MUTEX_TYPE_SPIN:
3346 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3347 		    mstate, vstate)) {
3348 			regs[rd] = 0;
3349 			break;
3350 		}
3351 
3352 		m.mx = dtrace_load64(tupregs[0].dttk_value);
3353 		regs[rd] = MUTEX_TYPE_SPIN(&m.mi);
3354 		break;
3355 
3356 	case DIF_SUBR_RW_READ_HELD: {
3357 		uintptr_t tmp;
3358 
3359 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
3360 		    mstate, vstate)) {
3361 			regs[rd] = 0;
3362 			break;
3363 		}
3364 
3365 		r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3366 		regs[rd] = _RW_READ_HELD(&r.ri, tmp);
3367 		break;
3368 	}
3369 
3370 	case DIF_SUBR_RW_WRITE_HELD:
3371 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
3372 		    mstate, vstate)) {
3373 			regs[rd] = 0;
3374 			break;
3375 		}
3376 
3377 		r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3378 		regs[rd] = _RW_WRITE_HELD(&r.ri);
3379 		break;
3380 
3381 	case DIF_SUBR_RW_ISWRITER:
3382 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
3383 		    mstate, vstate)) {
3384 			regs[rd] = 0;
3385 			break;
3386 		}
3387 
3388 		r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3389 		regs[rd] = _RW_ISWRITER(&r.ri);
3390 		break;
3391 
3392 #else
3393 	case DIF_SUBR_MUTEX_OWNED:
3394 		if (!dtrace_canload(tupregs[0].dttk_value,
3395 			sizeof (struct lock_object), mstate, vstate)) {
3396 			regs[rd] = 0;
3397 			break;
3398 		}
3399 		l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3400 		regs[rd] = LOCK_CLASS(l.li)->lc_owner(l.li, &lowner);
3401 		break;
3402 
3403 	case DIF_SUBR_MUTEX_OWNER:
3404 		if (!dtrace_canload(tupregs[0].dttk_value,
3405 			sizeof (struct lock_object), mstate, vstate)) {
3406 			regs[rd] = 0;
3407 			break;
3408 		}
3409 		l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3410 		LOCK_CLASS(l.li)->lc_owner(l.li, &lowner);
3411 		regs[rd] = (uintptr_t)lowner;
3412 		break;
3413 
3414 	case DIF_SUBR_MUTEX_TYPE_ADAPTIVE:
3415 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (struct mtx),
3416 		    mstate, vstate)) {
3417 			regs[rd] = 0;
3418 			break;
3419 		}
3420 		l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3421 		/* XXX - should be only LC_SLEEPABLE? */
3422 		regs[rd] = (LOCK_CLASS(l.li)->lc_flags &
3423 		    (LC_SLEEPLOCK | LC_SLEEPABLE)) != 0;
3424 		break;
3425 
3426 	case DIF_SUBR_MUTEX_TYPE_SPIN:
3427 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (struct mtx),
3428 		    mstate, vstate)) {
3429 			regs[rd] = 0;
3430 			break;
3431 		}
3432 		l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3433 		regs[rd] = (LOCK_CLASS(l.li)->lc_flags & LC_SPINLOCK) != 0;
3434 		break;
3435 
3436 	case DIF_SUBR_RW_READ_HELD:
3437 	case DIF_SUBR_SX_SHARED_HELD:
3438 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
3439 		    mstate, vstate)) {
3440 			regs[rd] = 0;
3441 			break;
3442 		}
3443 		l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3444 		regs[rd] = LOCK_CLASS(l.li)->lc_owner(l.li, &lowner) &&
3445 		    lowner == NULL;
3446 		break;
3447 
3448 	case DIF_SUBR_RW_WRITE_HELD:
3449 	case DIF_SUBR_SX_EXCLUSIVE_HELD:
3450 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
3451 		    mstate, vstate)) {
3452 			regs[rd] = 0;
3453 			break;
3454 		}
3455 		l.lx = dtrace_loadptr(tupregs[0].dttk_value);
3456 		LOCK_CLASS(l.li)->lc_owner(l.li, &lowner);
3457 		regs[rd] = (lowner == curthread);
3458 		break;
3459 
3460 	case DIF_SUBR_RW_ISWRITER:
3461 	case DIF_SUBR_SX_ISEXCLUSIVE:
3462 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
3463 		    mstate, vstate)) {
3464 			regs[rd] = 0;
3465 			break;
3466 		}
3467 		l.lx = dtrace_loadptr(tupregs[0].dttk_value);
3468 		regs[rd] = LOCK_CLASS(l.li)->lc_owner(l.li, &lowner) &&
3469 		    lowner != NULL;
3470 		break;
3471 #endif /* ! defined(sun) */
3472 
3473 	case DIF_SUBR_BCOPY: {
3474 		/*
3475 		 * We need to be sure that the destination is in the scratch
3476 		 * region -- no other region is allowed.
3477 		 */
3478 		uintptr_t src = tupregs[0].dttk_value;
3479 		uintptr_t dest = tupregs[1].dttk_value;
3480 		size_t size = tupregs[2].dttk_value;
3481 
3482 		if (!dtrace_inscratch(dest, size, mstate)) {
3483 			*flags |= CPU_DTRACE_BADADDR;
3484 			*illval = regs[rd];
3485 			break;
3486 		}
3487 
3488 		if (!dtrace_canload(src, size, mstate, vstate)) {
3489 			regs[rd] = 0;
3490 			break;
3491 		}
3492 
3493 		dtrace_bcopy((void *)src, (void *)dest, size);
3494 		break;
3495 	}
3496 
3497 	case DIF_SUBR_ALLOCA:
3498 	case DIF_SUBR_COPYIN: {
3499 		uintptr_t dest = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
3500 		uint64_t size =
3501 		    tupregs[subr == DIF_SUBR_ALLOCA ? 0 : 1].dttk_value;
3502 		size_t scratch_size = (dest - mstate->dtms_scratch_ptr) + size;
3503 
3504 		/*
3505 		 * This action doesn't require any credential checks since
3506 		 * probes will not activate in user contexts to which the
3507 		 * enabling user does not have permissions.
3508 		 */
3509 
3510 		/*
3511 		 * Rounding up the user allocation size could have overflowed
3512 		 * a large, bogus allocation (like -1ULL) to 0.
3513 		 */
3514 		if (scratch_size < size ||
3515 		    !DTRACE_INSCRATCH(mstate, scratch_size)) {
3516 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3517 			regs[rd] = 0;
3518 			break;
3519 		}
3520 
3521 		if (subr == DIF_SUBR_COPYIN) {
3522 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3523 			dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
3524 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3525 		}
3526 
3527 		mstate->dtms_scratch_ptr += scratch_size;
3528 		regs[rd] = dest;
3529 		break;
3530 	}
3531 
3532 	case DIF_SUBR_COPYINTO: {
3533 		uint64_t size = tupregs[1].dttk_value;
3534 		uintptr_t dest = tupregs[2].dttk_value;
3535 
3536 		/*
3537 		 * This action doesn't require any credential checks since
3538 		 * probes will not activate in user contexts to which the
3539 		 * enabling user does not have permissions.
3540 		 */
3541 		if (!dtrace_inscratch(dest, size, mstate)) {
3542 			*flags |= CPU_DTRACE_BADADDR;
3543 			*illval = regs[rd];
3544 			break;
3545 		}
3546 
3547 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3548 		dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
3549 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3550 		break;
3551 	}
3552 
3553 	case DIF_SUBR_COPYINSTR: {
3554 		uintptr_t dest = mstate->dtms_scratch_ptr;
3555 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3556 
3557 		if (nargs > 1 && tupregs[1].dttk_value < size)
3558 			size = tupregs[1].dttk_value + 1;
3559 
3560 		/*
3561 		 * This action doesn't require any credential checks since
3562 		 * probes will not activate in user contexts to which the
3563 		 * enabling user does not have permissions.
3564 		 */
3565 		if (!DTRACE_INSCRATCH(mstate, size)) {
3566 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3567 			regs[rd] = 0;
3568 			break;
3569 		}
3570 
3571 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3572 		dtrace_copyinstr(tupregs[0].dttk_value, dest, size, flags);
3573 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3574 
3575 		((char *)dest)[size - 1] = '\0';
3576 		mstate->dtms_scratch_ptr += size;
3577 		regs[rd] = dest;
3578 		break;
3579 	}
3580 
3581 #if defined(sun)
3582 	case DIF_SUBR_MSGSIZE:
3583 	case DIF_SUBR_MSGDSIZE: {
3584 		uintptr_t baddr = tupregs[0].dttk_value, daddr;
3585 		uintptr_t wptr, rptr;
3586 		size_t count = 0;
3587 		int cont = 0;
3588 
3589 		while (baddr != 0 && !(*flags & CPU_DTRACE_FAULT)) {
3590 
3591 			if (!dtrace_canload(baddr, sizeof (mblk_t), mstate,
3592 			    vstate)) {
3593 				regs[rd] = 0;
3594 				break;
3595 			}
3596 
3597 			wptr = dtrace_loadptr(baddr +
3598 			    offsetof(mblk_t, b_wptr));
3599 
3600 			rptr = dtrace_loadptr(baddr +
3601 			    offsetof(mblk_t, b_rptr));
3602 
3603 			if (wptr < rptr) {
3604 				*flags |= CPU_DTRACE_BADADDR;
3605 				*illval = tupregs[0].dttk_value;
3606 				break;
3607 			}
3608 
3609 			daddr = dtrace_loadptr(baddr +
3610 			    offsetof(mblk_t, b_datap));
3611 
3612 			baddr = dtrace_loadptr(baddr +
3613 			    offsetof(mblk_t, b_cont));
3614 
3615 			/*
3616 			 * We want to prevent against denial-of-service here,
3617 			 * so we're only going to search the list for
3618 			 * dtrace_msgdsize_max mblks.
3619 			 */
3620 			if (cont++ > dtrace_msgdsize_max) {
3621 				*flags |= CPU_DTRACE_ILLOP;
3622 				break;
3623 			}
3624 
3625 			if (subr == DIF_SUBR_MSGDSIZE) {
3626 				if (dtrace_load8(daddr +
3627 				    offsetof(dblk_t, db_type)) != M_DATA)
3628 					continue;
3629 			}
3630 
3631 			count += wptr - rptr;
3632 		}
3633 
3634 		if (!(*flags & CPU_DTRACE_FAULT))
3635 			regs[rd] = count;
3636 
3637 		break;
3638 	}
3639 #endif
3640 
3641 	case DIF_SUBR_PROGENYOF: {
3642 		pid_t pid = tupregs[0].dttk_value;
3643 		proc_t *p;
3644 		int rval = 0;
3645 
3646 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3647 
3648 		for (p = curthread->t_procp; p != NULL; p = p->p_parent) {
3649 #if defined(sun)
3650 			if (p->p_pidp->pid_id == pid) {
3651 #else
3652 			if (p->p_pid == pid) {
3653 #endif
3654 				rval = 1;
3655 				break;
3656 			}
3657 		}
3658 
3659 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3660 
3661 		regs[rd] = rval;
3662 		break;
3663 	}
3664 
3665 	case DIF_SUBR_SPECULATION:
3666 		regs[rd] = dtrace_speculation(state);
3667 		break;
3668 
3669 	case DIF_SUBR_COPYOUT: {
3670 		uintptr_t kaddr = tupregs[0].dttk_value;
3671 		uintptr_t uaddr = tupregs[1].dttk_value;
3672 		uint64_t size = tupregs[2].dttk_value;
3673 
3674 		if (!dtrace_destructive_disallow &&
3675 		    dtrace_priv_proc_control(state) &&
3676 		    !dtrace_istoxic(kaddr, size)) {
3677 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3678 			dtrace_copyout(kaddr, uaddr, size, flags);
3679 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3680 		}
3681 		break;
3682 	}
3683 
3684 	case DIF_SUBR_COPYOUTSTR: {
3685 		uintptr_t kaddr = tupregs[0].dttk_value;
3686 		uintptr_t uaddr = tupregs[1].dttk_value;
3687 		uint64_t size = tupregs[2].dttk_value;
3688 
3689 		if (!dtrace_destructive_disallow &&
3690 		    dtrace_priv_proc_control(state) &&
3691 		    !dtrace_istoxic(kaddr, size)) {
3692 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3693 			dtrace_copyoutstr(kaddr, uaddr, size, flags);
3694 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3695 		}
3696 		break;
3697 	}
3698 
3699 	case DIF_SUBR_STRLEN: {
3700 		size_t sz;
3701 		uintptr_t addr = (uintptr_t)tupregs[0].dttk_value;
3702 		sz = dtrace_strlen((char *)addr,
3703 		    state->dts_options[DTRACEOPT_STRSIZE]);
3704 
3705 		if (!dtrace_canload(addr, sz + 1, mstate, vstate)) {
3706 			regs[rd] = 0;
3707 			break;
3708 		}
3709 
3710 		regs[rd] = sz;
3711 
3712 		break;
3713 	}
3714 
3715 	case DIF_SUBR_STRCHR:
3716 	case DIF_SUBR_STRRCHR: {
3717 		/*
3718 		 * We're going to iterate over the string looking for the
3719 		 * specified character.  We will iterate until we have reached
3720 		 * the string length or we have found the character.  If this
3721 		 * is DIF_SUBR_STRRCHR, we will look for the last occurrence
3722 		 * of the specified character instead of the first.
3723 		 */
3724 		uintptr_t saddr = tupregs[0].dttk_value;
3725 		uintptr_t addr = tupregs[0].dttk_value;
3726 		uintptr_t limit = addr + state->dts_options[DTRACEOPT_STRSIZE];
3727 		char c, target = (char)tupregs[1].dttk_value;
3728 
3729 		for (regs[rd] = 0; addr < limit; addr++) {
3730 			if ((c = dtrace_load8(addr)) == target) {
3731 				regs[rd] = addr;
3732 
3733 				if (subr == DIF_SUBR_STRCHR)
3734 					break;
3735 			}
3736 
3737 			if (c == '\0')
3738 				break;
3739 		}
3740 
3741 		if (!dtrace_canload(saddr, addr - saddr, mstate, vstate)) {
3742 			regs[rd] = 0;
3743 			break;
3744 		}
3745 
3746 		break;
3747 	}
3748 
3749 	case DIF_SUBR_STRSTR:
3750 	case DIF_SUBR_INDEX:
3751 	case DIF_SUBR_RINDEX: {
3752 		/*
3753 		 * We're going to iterate over the string looking for the
3754 		 * specified string.  We will iterate until we have reached
3755 		 * the string length or we have found the string.  (Yes, this
3756 		 * is done in the most naive way possible -- but considering
3757 		 * that the string we're searching for is likely to be
3758 		 * relatively short, the complexity of Rabin-Karp or similar
3759 		 * hardly seems merited.)
3760 		 */
3761 		char *addr = (char *)(uintptr_t)tupregs[0].dttk_value;
3762 		char *substr = (char *)(uintptr_t)tupregs[1].dttk_value;
3763 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3764 		size_t len = dtrace_strlen(addr, size);
3765 		size_t sublen = dtrace_strlen(substr, size);
3766 		char *limit = addr + len, *orig = addr;
3767 		int notfound = subr == DIF_SUBR_STRSTR ? 0 : -1;
3768 		int inc = 1;
3769 
3770 		regs[rd] = notfound;
3771 
3772 		if (!dtrace_canload((uintptr_t)addr, len + 1, mstate, vstate)) {
3773 			regs[rd] = 0;
3774 			break;
3775 		}
3776 
3777 		if (!dtrace_canload((uintptr_t)substr, sublen + 1, mstate,
3778 		    vstate)) {
3779 			regs[rd] = 0;
3780 			break;
3781 		}
3782 
3783 		/*
3784 		 * strstr() and index()/rindex() have similar semantics if
3785 		 * both strings are the empty string: strstr() returns a
3786 		 * pointer to the (empty) string, and index() and rindex()
3787 		 * both return index 0 (regardless of any position argument).
3788 		 */
3789 		if (sublen == 0 && len == 0) {
3790 			if (subr == DIF_SUBR_STRSTR)
3791 				regs[rd] = (uintptr_t)addr;
3792 			else
3793 				regs[rd] = 0;
3794 			break;
3795 		}
3796 
3797 		if (subr != DIF_SUBR_STRSTR) {
3798 			if (subr == DIF_SUBR_RINDEX) {
3799 				limit = orig - 1;
3800 				addr += len;
3801 				inc = -1;
3802 			}
3803 
3804 			/*
3805 			 * Both index() and rindex() take an optional position
3806 			 * argument that denotes the starting position.
3807 			 */
3808 			if (nargs == 3) {
3809 				int64_t pos = (int64_t)tupregs[2].dttk_value;
3810 
3811 				/*
3812 				 * If the position argument to index() is
3813 				 * negative, Perl implicitly clamps it at
3814 				 * zero.  This semantic is a little surprising
3815 				 * given the special meaning of negative
3816 				 * positions to similar Perl functions like
3817 				 * substr(), but it appears to reflect a
3818 				 * notion that index() can start from a
3819 				 * negative index and increment its way up to
3820 				 * the string.  Given this notion, Perl's
3821 				 * rindex() is at least self-consistent in
3822 				 * that it implicitly clamps positions greater
3823 				 * than the string length to be the string
3824 				 * length.  Where Perl completely loses
3825 				 * coherence, however, is when the specified
3826 				 * substring is the empty string ("").  In
3827 				 * this case, even if the position is
3828 				 * negative, rindex() returns 0 -- and even if
3829 				 * the position is greater than the length,
3830 				 * index() returns the string length.  These
3831 				 * semantics violate the notion that index()
3832 				 * should never return a value less than the
3833 				 * specified position and that rindex() should
3834 				 * never return a value greater than the
3835 				 * specified position.  (One assumes that
3836 				 * these semantics are artifacts of Perl's
3837 				 * implementation and not the results of
3838 				 * deliberate design -- it beggars belief that
3839 				 * even Larry Wall could desire such oddness.)
3840 				 * While in the abstract one would wish for
3841 				 * consistent position semantics across
3842 				 * substr(), index() and rindex() -- or at the
3843 				 * very least self-consistent position
3844 				 * semantics for index() and rindex() -- we
3845 				 * instead opt to keep with the extant Perl
3846 				 * semantics, in all their broken glory.  (Do
3847 				 * we have more desire to maintain Perl's
3848 				 * semantics than Perl does?  Probably.)
3849 				 */
3850 				if (subr == DIF_SUBR_RINDEX) {
3851 					if (pos < 0) {
3852 						if (sublen == 0)
3853 							regs[rd] = 0;
3854 						break;
3855 					}
3856 
3857 					if (pos > len)
3858 						pos = len;
3859 				} else {
3860 					if (pos < 0)
3861 						pos = 0;
3862 
3863 					if (pos >= len) {
3864 						if (sublen == 0)
3865 							regs[rd] = len;
3866 						break;
3867 					}
3868 				}
3869 
3870 				addr = orig + pos;
3871 			}
3872 		}
3873 
3874 		for (regs[rd] = notfound; addr != limit; addr += inc) {
3875 			if (dtrace_strncmp(addr, substr, sublen) == 0) {
3876 				if (subr != DIF_SUBR_STRSTR) {
3877 					/*
3878 					 * As D index() and rindex() are
3879 					 * modeled on Perl (and not on awk),
3880 					 * we return a zero-based (and not a
3881 					 * one-based) index.  (For you Perl
3882 					 * weenies: no, we're not going to add
3883 					 * $[ -- and shouldn't you be at a con
3884 					 * or something?)
3885 					 */
3886 					regs[rd] = (uintptr_t)(addr - orig);
3887 					break;
3888 				}
3889 
3890 				ASSERT(subr == DIF_SUBR_STRSTR);
3891 				regs[rd] = (uintptr_t)addr;
3892 				break;
3893 			}
3894 		}
3895 
3896 		break;
3897 	}
3898 
3899 	case DIF_SUBR_STRTOK: {
3900 		uintptr_t addr = tupregs[0].dttk_value;
3901 		uintptr_t tokaddr = tupregs[1].dttk_value;
3902 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3903 		uintptr_t limit, toklimit = tokaddr + size;
3904 		uint8_t c = 0, tokmap[32];	 /* 256 / 8 */
3905 		char *dest = (char *)mstate->dtms_scratch_ptr;
3906 		int i;
3907 
3908 		/*
3909 		 * Check both the token buffer and (later) the input buffer,
3910 		 * since both could be non-scratch addresses.
3911 		 */
3912 		if (!dtrace_strcanload(tokaddr, size, mstate, vstate)) {
3913 			regs[rd] = 0;
3914 			break;
3915 		}
3916 
3917 		if (!DTRACE_INSCRATCH(mstate, size)) {
3918 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3919 			regs[rd] = 0;
3920 			break;
3921 		}
3922 
3923 		if (addr == 0) {
3924 			/*
3925 			 * If the address specified is NULL, we use our saved
3926 			 * strtok pointer from the mstate.  Note that this
3927 			 * means that the saved strtok pointer is _only_
3928 			 * valid within multiple enablings of the same probe --
3929 			 * it behaves like an implicit clause-local variable.
3930 			 */
3931 			addr = mstate->dtms_strtok;
3932 		} else {
3933 			/*
3934 			 * If the user-specified address is non-NULL we must
3935 			 * access check it.  This is the only time we have
3936 			 * a chance to do so, since this address may reside
3937 			 * in the string table of this clause-- future calls
3938 			 * (when we fetch addr from mstate->dtms_strtok)
3939 			 * would fail this access check.
3940 			 */
3941 			if (!dtrace_strcanload(addr, size, mstate, vstate)) {
3942 				regs[rd] = 0;
3943 				break;
3944 			}
3945 		}
3946 
3947 		/*
3948 		 * First, zero the token map, and then process the token
3949 		 * string -- setting a bit in the map for every character
3950 		 * found in the token string.
3951 		 */
3952 		for (i = 0; i < sizeof (tokmap); i++)
3953 			tokmap[i] = 0;
3954 
3955 		for (; tokaddr < toklimit; tokaddr++) {
3956 			if ((c = dtrace_load8(tokaddr)) == '\0')
3957 				break;
3958 
3959 			ASSERT((c >> 3) < sizeof (tokmap));
3960 			tokmap[c >> 3] |= (1 << (c & 0x7));
3961 		}
3962 
3963 		for (limit = addr + size; addr < limit; addr++) {
3964 			/*
3965 			 * We're looking for a character that is _not_ contained
3966 			 * in the token string.
3967 			 */
3968 			if ((c = dtrace_load8(addr)) == '\0')
3969 				break;
3970 
3971 			if (!(tokmap[c >> 3] & (1 << (c & 0x7))))
3972 				break;
3973 		}
3974 
3975 		if (c == '\0') {
3976 			/*
3977 			 * We reached the end of the string without finding
3978 			 * any character that was not in the token string.
3979 			 * We return NULL in this case, and we set the saved
3980 			 * address to NULL as well.
3981 			 */
3982 			regs[rd] = 0;
3983 			mstate->dtms_strtok = 0;
3984 			break;
3985 		}
3986 
3987 		/*
3988 		 * From here on, we're copying into the destination string.
3989 		 */
3990 		for (i = 0; addr < limit && i < size - 1; addr++) {
3991 			if ((c = dtrace_load8(addr)) == '\0')
3992 				break;
3993 
3994 			if (tokmap[c >> 3] & (1 << (c & 0x7)))
3995 				break;
3996 
3997 			ASSERT(i < size);
3998 			dest[i++] = c;
3999 		}
4000 
4001 		ASSERT(i < size);
4002 		dest[i] = '\0';
4003 		regs[rd] = (uintptr_t)dest;
4004 		mstate->dtms_scratch_ptr += size;
4005 		mstate->dtms_strtok = addr;
4006 		break;
4007 	}
4008 
4009 	case DIF_SUBR_SUBSTR: {
4010 		uintptr_t s = tupregs[0].dttk_value;
4011 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4012 		char *d = (char *)mstate->dtms_scratch_ptr;
4013 		int64_t index = (int64_t)tupregs[1].dttk_value;
4014 		int64_t remaining = (int64_t)tupregs[2].dttk_value;
4015 		size_t len = dtrace_strlen((char *)s, size);
4016 		int64_t i = 0;
4017 
4018 		if (!dtrace_canload(s, len + 1, mstate, vstate)) {
4019 			regs[rd] = 0;
4020 			break;
4021 		}
4022 
4023 		if (!DTRACE_INSCRATCH(mstate, size)) {
4024 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4025 			regs[rd] = 0;
4026 			break;
4027 		}
4028 
4029 		if (nargs <= 2)
4030 			remaining = (int64_t)size;
4031 
4032 		if (index < 0) {
4033 			index += len;
4034 
4035 			if (index < 0 && index + remaining > 0) {
4036 				remaining += index;
4037 				index = 0;
4038 			}
4039 		}
4040 
4041 		if (index >= len || index < 0) {
4042 			remaining = 0;
4043 		} else if (remaining < 0) {
4044 			remaining += len - index;
4045 		} else if (index + remaining > size) {
4046 			remaining = size - index;
4047 		}
4048 
4049 		for (i = 0; i < remaining; i++) {
4050 			if ((d[i] = dtrace_load8(s + index + i)) == '\0')
4051 				break;
4052 		}
4053 
4054 		d[i] = '\0';
4055 
4056 		mstate->dtms_scratch_ptr += size;
4057 		regs[rd] = (uintptr_t)d;
4058 		break;
4059 	}
4060 
4061 	case DIF_SUBR_TOUPPER:
4062 	case DIF_SUBR_TOLOWER: {
4063 		uintptr_t s = tupregs[0].dttk_value;
4064 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4065 		char *dest = (char *)mstate->dtms_scratch_ptr, c;
4066 		size_t len = dtrace_strlen((char *)s, size);
4067 		char lower, upper, convert;
4068 		int64_t i;
4069 
4070 		if (subr == DIF_SUBR_TOUPPER) {
4071 			lower = 'a';
4072 			upper = 'z';
4073 			convert = 'A';
4074 		} else {
4075 			lower = 'A';
4076 			upper = 'Z';
4077 			convert = 'a';
4078 		}
4079 
4080 		if (!dtrace_canload(s, len + 1, mstate, vstate)) {
4081 			regs[rd] = 0;
4082 			break;
4083 		}
4084 
4085 		if (!DTRACE_INSCRATCH(mstate, size)) {
4086 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4087 			regs[rd] = 0;
4088 			break;
4089 		}
4090 
4091 		for (i = 0; i < size - 1; i++) {
4092 			if ((c = dtrace_load8(s + i)) == '\0')
4093 				break;
4094 
4095 			if (c >= lower && c <= upper)
4096 				c = convert + (c - lower);
4097 
4098 			dest[i] = c;
4099 		}
4100 
4101 		ASSERT(i < size);
4102 		dest[i] = '\0';
4103 		regs[rd] = (uintptr_t)dest;
4104 		mstate->dtms_scratch_ptr += size;
4105 		break;
4106 	}
4107 
4108 #if defined(sun)
4109 	case DIF_SUBR_GETMAJOR:
4110 #ifdef _LP64
4111 		regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR64) & MAXMAJ64;
4112 #else
4113 		regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR) & MAXMAJ;
4114 #endif
4115 		break;
4116 
4117 	case DIF_SUBR_GETMINOR:
4118 #ifdef _LP64
4119 		regs[rd] = tupregs[0].dttk_value & MAXMIN64;
4120 #else
4121 		regs[rd] = tupregs[0].dttk_value & MAXMIN;
4122 #endif
4123 		break;
4124 
4125 	case DIF_SUBR_DDI_PATHNAME: {
4126 		/*
4127 		 * This one is a galactic mess.  We are going to roughly
4128 		 * emulate ddi_pathname(), but it's made more complicated
4129 		 * by the fact that we (a) want to include the minor name and
4130 		 * (b) must proceed iteratively instead of recursively.
4131 		 */
4132 		uintptr_t dest = mstate->dtms_scratch_ptr;
4133 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4134 		char *start = (char *)dest, *end = start + size - 1;
4135 		uintptr_t daddr = tupregs[0].dttk_value;
4136 		int64_t minor = (int64_t)tupregs[1].dttk_value;
4137 		char *s;
4138 		int i, len, depth = 0;
4139 
4140 		/*
4141 		 * Due to all the pointer jumping we do and context we must
4142 		 * rely upon, we just mandate that the user must have kernel
4143 		 * read privileges to use this routine.
4144 		 */
4145 		if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) == 0) {
4146 			*flags |= CPU_DTRACE_KPRIV;
4147 			*illval = daddr;
4148 			regs[rd] = 0;
4149 		}
4150 
4151 		if (!DTRACE_INSCRATCH(mstate, size)) {
4152 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4153 			regs[rd] = 0;
4154 			break;
4155 		}
4156 
4157 		*end = '\0';
4158 
4159 		/*
4160 		 * We want to have a name for the minor.  In order to do this,
4161 		 * we need to walk the minor list from the devinfo.  We want
4162 		 * to be sure that we don't infinitely walk a circular list,
4163 		 * so we check for circularity by sending a scout pointer
4164 		 * ahead two elements for every element that we iterate over;
4165 		 * if the list is circular, these will ultimately point to the
4166 		 * same element.  You may recognize this little trick as the
4167 		 * answer to a stupid interview question -- one that always
4168 		 * seems to be asked by those who had to have it laboriously
4169 		 * explained to them, and who can't even concisely describe
4170 		 * the conditions under which one would be forced to resort to
4171 		 * this technique.  Needless to say, those conditions are
4172 		 * found here -- and probably only here.  Is this the only use
4173 		 * of this infamous trick in shipping, production code?  If it
4174 		 * isn't, it probably should be...
4175 		 */
4176 		if (minor != -1) {
4177 			uintptr_t maddr = dtrace_loadptr(daddr +
4178 			    offsetof(struct dev_info, devi_minor));
4179 
4180 			uintptr_t next = offsetof(struct ddi_minor_data, next);
4181 			uintptr_t name = offsetof(struct ddi_minor_data,
4182 			    d_minor) + offsetof(struct ddi_minor, name);
4183 			uintptr_t dev = offsetof(struct ddi_minor_data,
4184 			    d_minor) + offsetof(struct ddi_minor, dev);
4185 			uintptr_t scout;
4186 
4187 			if (maddr != NULL)
4188 				scout = dtrace_loadptr(maddr + next);
4189 
4190 			while (maddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
4191 				uint64_t m;
4192 #ifdef _LP64
4193 				m = dtrace_load64(maddr + dev) & MAXMIN64;
4194 #else
4195 				m = dtrace_load32(maddr + dev) & MAXMIN;
4196 #endif
4197 				if (m != minor) {
4198 					maddr = dtrace_loadptr(maddr + next);
4199 
4200 					if (scout == NULL)
4201 						continue;
4202 
4203 					scout = dtrace_loadptr(scout + next);
4204 
4205 					if (scout == NULL)
4206 						continue;
4207 
4208 					scout = dtrace_loadptr(scout + next);
4209 
4210 					if (scout == NULL)
4211 						continue;
4212 
4213 					if (scout == maddr) {
4214 						*flags |= CPU_DTRACE_ILLOP;
4215 						break;
4216 					}
4217 
4218 					continue;
4219 				}
4220 
4221 				/*
4222 				 * We have the minor data.  Now we need to
4223 				 * copy the minor's name into the end of the
4224 				 * pathname.
4225 				 */
4226 				s = (char *)dtrace_loadptr(maddr + name);
4227 				len = dtrace_strlen(s, size);
4228 
4229 				if (*flags & CPU_DTRACE_FAULT)
4230 					break;
4231 
4232 				if (len != 0) {
4233 					if ((end -= (len + 1)) < start)
4234 						break;
4235 
4236 					*end = ':';
4237 				}
4238 
4239 				for (i = 1; i <= len; i++)
4240 					end[i] = dtrace_load8((uintptr_t)s++);
4241 				break;
4242 			}
4243 		}
4244 
4245 		while (daddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
4246 			ddi_node_state_t devi_state;
4247 
4248 			devi_state = dtrace_load32(daddr +
4249 			    offsetof(struct dev_info, devi_node_state));
4250 
4251 			if (*flags & CPU_DTRACE_FAULT)
4252 				break;
4253 
4254 			if (devi_state >= DS_INITIALIZED) {
4255 				s = (char *)dtrace_loadptr(daddr +
4256 				    offsetof(struct dev_info, devi_addr));
4257 				len = dtrace_strlen(s, size);
4258 
4259 				if (*flags & CPU_DTRACE_FAULT)
4260 					break;
4261 
4262 				if (len != 0) {
4263 					if ((end -= (len + 1)) < start)
4264 						break;
4265 
4266 					*end = '@';
4267 				}
4268 
4269 				for (i = 1; i <= len; i++)
4270 					end[i] = dtrace_load8((uintptr_t)s++);
4271 			}
4272 
4273 			/*
4274 			 * Now for the node name...
4275 			 */
4276 			s = (char *)dtrace_loadptr(daddr +
4277 			    offsetof(struct dev_info, devi_node_name));
4278 
4279 			daddr = dtrace_loadptr(daddr +
4280 			    offsetof(struct dev_info, devi_parent));
4281 
4282 			/*
4283 			 * If our parent is NULL (that is, if we're the root
4284 			 * node), we're going to use the special path
4285 			 * "devices".
4286 			 */
4287 			if (daddr == 0)
4288 				s = "devices";
4289 
4290 			len = dtrace_strlen(s, size);
4291 			if (*flags & CPU_DTRACE_FAULT)
4292 				break;
4293 
4294 			if ((end -= (len + 1)) < start)
4295 				break;
4296 
4297 			for (i = 1; i <= len; i++)
4298 				end[i] = dtrace_load8((uintptr_t)s++);
4299 			*end = '/';
4300 
4301 			if (depth++ > dtrace_devdepth_max) {
4302 				*flags |= CPU_DTRACE_ILLOP;
4303 				break;
4304 			}
4305 		}
4306 
4307 		if (end < start)
4308 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4309 
4310 		if (daddr == 0) {
4311 			regs[rd] = (uintptr_t)end;
4312 			mstate->dtms_scratch_ptr += size;
4313 		}
4314 
4315 		break;
4316 	}
4317 #endif
4318 
4319 	case DIF_SUBR_STRJOIN: {
4320 		char *d = (char *)mstate->dtms_scratch_ptr;
4321 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4322 		uintptr_t s1 = tupregs[0].dttk_value;
4323 		uintptr_t s2 = tupregs[1].dttk_value;
4324 		int i = 0;
4325 
4326 		if (!dtrace_strcanload(s1, size, mstate, vstate) ||
4327 		    !dtrace_strcanload(s2, size, mstate, vstate)) {
4328 			regs[rd] = 0;
4329 			break;
4330 		}
4331 
4332 		if (!DTRACE_INSCRATCH(mstate, size)) {
4333 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4334 			regs[rd] = 0;
4335 			break;
4336 		}
4337 
4338 		for (;;) {
4339 			if (i >= size) {
4340 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4341 				regs[rd] = 0;
4342 				break;
4343 			}
4344 
4345 			if ((d[i++] = dtrace_load8(s1++)) == '\0') {
4346 				i--;
4347 				break;
4348 			}
4349 		}
4350 
4351 		for (;;) {
4352 			if (i >= size) {
4353 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4354 				regs[rd] = 0;
4355 				break;
4356 			}
4357 
4358 			if ((d[i++] = dtrace_load8(s2++)) == '\0')
4359 				break;
4360 		}
4361 
4362 		if (i < size) {
4363 			mstate->dtms_scratch_ptr += i;
4364 			regs[rd] = (uintptr_t)d;
4365 		}
4366 
4367 		break;
4368 	}
4369 
4370 	case DIF_SUBR_LLTOSTR: {
4371 		int64_t i = (int64_t)tupregs[0].dttk_value;
4372 		uint64_t val, digit;
4373 		uint64_t size = 65;	/* enough room for 2^64 in binary */
4374 		char *end = (char *)mstate->dtms_scratch_ptr + size - 1;
4375 		int base = 10;
4376 
4377 		if (nargs > 1) {
4378 			if ((base = tupregs[1].dttk_value) <= 1 ||
4379 			    base > ('z' - 'a' + 1) + ('9' - '0' + 1)) {
4380 				*flags |= CPU_DTRACE_ILLOP;
4381 				break;
4382 			}
4383 		}
4384 
4385 		val = (base == 10 && i < 0) ? i * -1 : i;
4386 
4387 		if (!DTRACE_INSCRATCH(mstate, size)) {
4388 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4389 			regs[rd] = 0;
4390 			break;
4391 		}
4392 
4393 		for (*end-- = '\0'; val; val /= base) {
4394 			if ((digit = val % base) <= '9' - '0') {
4395 				*end-- = '0' + digit;
4396 			} else {
4397 				*end-- = 'a' + (digit - ('9' - '0') - 1);
4398 			}
4399 		}
4400 
4401 		if (i == 0 && base == 16)
4402 			*end-- = '0';
4403 
4404 		if (base == 16)
4405 			*end-- = 'x';
4406 
4407 		if (i == 0 || base == 8 || base == 16)
4408 			*end-- = '0';
4409 
4410 		if (i < 0 && base == 10)
4411 			*end-- = '-';
4412 
4413 		regs[rd] = (uintptr_t)end + 1;
4414 		mstate->dtms_scratch_ptr += size;
4415 		break;
4416 	}
4417 
4418 	case DIF_SUBR_HTONS:
4419 	case DIF_SUBR_NTOHS:
4420 #if BYTE_ORDER == BIG_ENDIAN
4421 		regs[rd] = (uint16_t)tupregs[0].dttk_value;
4422 #else
4423 		regs[rd] = DT_BSWAP_16((uint16_t)tupregs[0].dttk_value);
4424 #endif
4425 		break;
4426 
4427 
4428 	case DIF_SUBR_HTONL:
4429 	case DIF_SUBR_NTOHL:
4430 #if BYTE_ORDER == BIG_ENDIAN
4431 		regs[rd] = (uint32_t)tupregs[0].dttk_value;
4432 #else
4433 		regs[rd] = DT_BSWAP_32((uint32_t)tupregs[0].dttk_value);
4434 #endif
4435 		break;
4436 
4437 
4438 	case DIF_SUBR_HTONLL:
4439 	case DIF_SUBR_NTOHLL:
4440 #if BYTE_ORDER == BIG_ENDIAN
4441 		regs[rd] = (uint64_t)tupregs[0].dttk_value;
4442 #else
4443 		regs[rd] = DT_BSWAP_64((uint64_t)tupregs[0].dttk_value);
4444 #endif
4445 		break;
4446 
4447 
4448 	case DIF_SUBR_DIRNAME:
4449 	case DIF_SUBR_BASENAME: {
4450 		char *dest = (char *)mstate->dtms_scratch_ptr;
4451 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4452 		uintptr_t src = tupregs[0].dttk_value;
4453 		int i, j, len = dtrace_strlen((char *)src, size);
4454 		int lastbase = -1, firstbase = -1, lastdir = -1;
4455 		int start, end;
4456 
4457 		if (!dtrace_canload(src, len + 1, mstate, vstate)) {
4458 			regs[rd] = 0;
4459 			break;
4460 		}
4461 
4462 		if (!DTRACE_INSCRATCH(mstate, size)) {
4463 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4464 			regs[rd] = 0;
4465 			break;
4466 		}
4467 
4468 		/*
4469 		 * The basename and dirname for a zero-length string is
4470 		 * defined to be "."
4471 		 */
4472 		if (len == 0) {
4473 			len = 1;
4474 			src = (uintptr_t)".";
4475 		}
4476 
4477 		/*
4478 		 * Start from the back of the string, moving back toward the
4479 		 * front until we see a character that isn't a slash.  That
4480 		 * character is the last character in the basename.
4481 		 */
4482 		for (i = len - 1; i >= 0; i--) {
4483 			if (dtrace_load8(src + i) != '/')
4484 				break;
4485 		}
4486 
4487 		if (i >= 0)
4488 			lastbase = i;
4489 
4490 		/*
4491 		 * Starting from the last character in the basename, move
4492 		 * towards the front until we find a slash.  The character
4493 		 * that we processed immediately before that is the first
4494 		 * character in the basename.
4495 		 */
4496 		for (; i >= 0; i--) {
4497 			if (dtrace_load8(src + i) == '/')
4498 				break;
4499 		}
4500 
4501 		if (i >= 0)
4502 			firstbase = i + 1;
4503 
4504 		/*
4505 		 * Now keep going until we find a non-slash character.  That
4506 		 * character is the last character in the dirname.
4507 		 */
4508 		for (; i >= 0; i--) {
4509 			if (dtrace_load8(src + i) != '/')
4510 				break;
4511 		}
4512 
4513 		if (i >= 0)
4514 			lastdir = i;
4515 
4516 		ASSERT(!(lastbase == -1 && firstbase != -1));
4517 		ASSERT(!(firstbase == -1 && lastdir != -1));
4518 
4519 		if (lastbase == -1) {
4520 			/*
4521 			 * We didn't find a non-slash character.  We know that
4522 			 * the length is non-zero, so the whole string must be
4523 			 * slashes.  In either the dirname or the basename
4524 			 * case, we return '/'.
4525 			 */
4526 			ASSERT(firstbase == -1);
4527 			firstbase = lastbase = lastdir = 0;
4528 		}
4529 
4530 		if (firstbase == -1) {
4531 			/*
4532 			 * The entire string consists only of a basename
4533 			 * component.  If we're looking for dirname, we need
4534 			 * to change our string to be just "."; if we're
4535 			 * looking for a basename, we'll just set the first
4536 			 * character of the basename to be 0.
4537 			 */
4538 			if (subr == DIF_SUBR_DIRNAME) {
4539 				ASSERT(lastdir == -1);
4540 				src = (uintptr_t)".";
4541 				lastdir = 0;
4542 			} else {
4543 				firstbase = 0;
4544 			}
4545 		}
4546 
4547 		if (subr == DIF_SUBR_DIRNAME) {
4548 			if (lastdir == -1) {
4549 				/*
4550 				 * We know that we have a slash in the name --
4551 				 * or lastdir would be set to 0, above.  And
4552 				 * because lastdir is -1, we know that this
4553 				 * slash must be the first character.  (That
4554 				 * is, the full string must be of the form
4555 				 * "/basename".)  In this case, the last
4556 				 * character of the directory name is 0.
4557 				 */
4558 				lastdir = 0;
4559 			}
4560 
4561 			start = 0;
4562 			end = lastdir;
4563 		} else {
4564 			ASSERT(subr == DIF_SUBR_BASENAME);
4565 			ASSERT(firstbase != -1 && lastbase != -1);
4566 			start = firstbase;
4567 			end = lastbase;
4568 		}
4569 
4570 		for (i = start, j = 0; i <= end && j < size - 1; i++, j++)
4571 			dest[j] = dtrace_load8(src + i);
4572 
4573 		dest[j] = '\0';
4574 		regs[rd] = (uintptr_t)dest;
4575 		mstate->dtms_scratch_ptr += size;
4576 		break;
4577 	}
4578 
4579 	case DIF_SUBR_CLEANPATH: {
4580 		char *dest = (char *)mstate->dtms_scratch_ptr, c;
4581 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4582 		uintptr_t src = tupregs[0].dttk_value;
4583 		int i = 0, j = 0;
4584 
4585 		if (!dtrace_strcanload(src, size, mstate, vstate)) {
4586 			regs[rd] = 0;
4587 			break;
4588 		}
4589 
4590 		if (!DTRACE_INSCRATCH(mstate, size)) {
4591 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4592 			regs[rd] = 0;
4593 			break;
4594 		}
4595 
4596 		/*
4597 		 * Move forward, loading each character.
4598 		 */
4599 		do {
4600 			c = dtrace_load8(src + i++);
4601 next:
4602 			if (j + 5 >= size)	/* 5 = strlen("/..c\0") */
4603 				break;
4604 
4605 			if (c != '/') {
4606 				dest[j++] = c;
4607 				continue;
4608 			}
4609 
4610 			c = dtrace_load8(src + i++);
4611 
4612 			if (c == '/') {
4613 				/*
4614 				 * We have two slashes -- we can just advance
4615 				 * to the next character.
4616 				 */
4617 				goto next;
4618 			}
4619 
4620 			if (c != '.') {
4621 				/*
4622 				 * This is not "." and it's not ".." -- we can
4623 				 * just store the "/" and this character and
4624 				 * drive on.
4625 				 */
4626 				dest[j++] = '/';
4627 				dest[j++] = c;
4628 				continue;
4629 			}
4630 
4631 			c = dtrace_load8(src + i++);
4632 
4633 			if (c == '/') {
4634 				/*
4635 				 * This is a "/./" component.  We're not going
4636 				 * to store anything in the destination buffer;
4637 				 * we're just going to go to the next component.
4638 				 */
4639 				goto next;
4640 			}
4641 
4642 			if (c != '.') {
4643 				/*
4644 				 * This is not ".." -- we can just store the
4645 				 * "/." and this character and continue
4646 				 * processing.
4647 				 */
4648 				dest[j++] = '/';
4649 				dest[j++] = '.';
4650 				dest[j++] = c;
4651 				continue;
4652 			}
4653 
4654 			c = dtrace_load8(src + i++);
4655 
4656 			if (c != '/' && c != '\0') {
4657 				/*
4658 				 * This is not ".." -- it's "..[mumble]".
4659 				 * We'll store the "/.." and this character
4660 				 * and continue processing.
4661 				 */
4662 				dest[j++] = '/';
4663 				dest[j++] = '.';
4664 				dest[j++] = '.';
4665 				dest[j++] = c;
4666 				continue;
4667 			}
4668 
4669 			/*
4670 			 * This is "/../" or "/..\0".  We need to back up
4671 			 * our destination pointer until we find a "/".
4672 			 */
4673 			i--;
4674 			while (j != 0 && dest[--j] != '/')
4675 				continue;
4676 
4677 			if (c == '\0')
4678 				dest[++j] = '/';
4679 		} while (c != '\0');
4680 
4681 		dest[j] = '\0';
4682 		regs[rd] = (uintptr_t)dest;
4683 		mstate->dtms_scratch_ptr += size;
4684 		break;
4685 	}
4686 
4687 	case DIF_SUBR_INET_NTOA:
4688 	case DIF_SUBR_INET_NTOA6:
4689 	case DIF_SUBR_INET_NTOP: {
4690 		size_t size;
4691 		int af, argi, i;
4692 		char *base, *end;
4693 
4694 		if (subr == DIF_SUBR_INET_NTOP) {
4695 			af = (int)tupregs[0].dttk_value;
4696 			argi = 1;
4697 		} else {
4698 			af = subr == DIF_SUBR_INET_NTOA ? AF_INET: AF_INET6;
4699 			argi = 0;
4700 		}
4701 
4702 		if (af == AF_INET) {
4703 			ipaddr_t ip4;
4704 			uint8_t *ptr8, val;
4705 
4706 			/*
4707 			 * Safely load the IPv4 address.
4708 			 */
4709 			ip4 = dtrace_load32(tupregs[argi].dttk_value);
4710 
4711 			/*
4712 			 * Check an IPv4 string will fit in scratch.
4713 			 */
4714 			size = INET_ADDRSTRLEN;
4715 			if (!DTRACE_INSCRATCH(mstate, size)) {
4716 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4717 				regs[rd] = 0;
4718 				break;
4719 			}
4720 			base = (char *)mstate->dtms_scratch_ptr;
4721 			end = (char *)mstate->dtms_scratch_ptr + size - 1;
4722 
4723 			/*
4724 			 * Stringify as a dotted decimal quad.
4725 			 */
4726 			*end-- = '\0';
4727 			ptr8 = (uint8_t *)&ip4;
4728 			for (i = 3; i >= 0; i--) {
4729 				val = ptr8[i];
4730 
4731 				if (val == 0) {
4732 					*end-- = '0';
4733 				} else {
4734 					for (; val; val /= 10) {
4735 						*end-- = '0' + (val % 10);
4736 					}
4737 				}
4738 
4739 				if (i > 0)
4740 					*end-- = '.';
4741 			}
4742 			ASSERT(end + 1 >= base);
4743 
4744 		} else if (af == AF_INET6) {
4745 			struct in6_addr ip6;
4746 			int firstzero, tryzero, numzero, v6end;
4747 			uint16_t val;
4748 			const char digits[] = "0123456789abcdef";
4749 
4750 			/*
4751 			 * Stringify using RFC 1884 convention 2 - 16 bit
4752 			 * hexadecimal values with a zero-run compression.
4753 			 * Lower case hexadecimal digits are used.
4754 			 * 	eg, fe80::214:4fff:fe0b:76c8.
4755 			 * The IPv4 embedded form is returned for inet_ntop,
4756 			 * just the IPv4 string is returned for inet_ntoa6.
4757 			 */
4758 
4759 			/*
4760 			 * Safely load the IPv6 address.
4761 			 */
4762 			dtrace_bcopy(
4763 			    (void *)(uintptr_t)tupregs[argi].dttk_value,
4764 			    (void *)(uintptr_t)&ip6, sizeof (struct in6_addr));
4765 
4766 			/*
4767 			 * Check an IPv6 string will fit in scratch.
4768 			 */
4769 			size = INET6_ADDRSTRLEN;
4770 			if (!DTRACE_INSCRATCH(mstate, size)) {
4771 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4772 				regs[rd] = 0;
4773 				break;
4774 			}
4775 			base = (char *)mstate->dtms_scratch_ptr;
4776 			end = (char *)mstate->dtms_scratch_ptr + size - 1;
4777 			*end-- = '\0';
4778 
4779 			/*
4780 			 * Find the longest run of 16 bit zero values
4781 			 * for the single allowed zero compression - "::".
4782 			 */
4783 			firstzero = -1;
4784 			tryzero = -1;
4785 			numzero = 1;
4786 			for (i = 0; i < sizeof (struct in6_addr); i++) {
4787 #if defined(sun)
4788 				if (ip6._S6_un._S6_u8[i] == 0 &&
4789 #else
4790 				if (ip6.__u6_addr.__u6_addr8[i] == 0 &&
4791 #endif
4792 				    tryzero == -1 && i % 2 == 0) {
4793 					tryzero = i;
4794 					continue;
4795 				}
4796 
4797 				if (tryzero != -1 &&
4798 #if defined(sun)
4799 				    (ip6._S6_un._S6_u8[i] != 0 ||
4800 #else
4801 				    (ip6.__u6_addr.__u6_addr8[i] != 0 ||
4802 #endif
4803 				    i == sizeof (struct in6_addr) - 1)) {
4804 
4805 					if (i - tryzero <= numzero) {
4806 						tryzero = -1;
4807 						continue;
4808 					}
4809 
4810 					firstzero = tryzero;
4811 					numzero = i - i % 2 - tryzero;
4812 					tryzero = -1;
4813 
4814 #if defined(sun)
4815 					if (ip6._S6_un._S6_u8[i] == 0 &&
4816 #else
4817 					if (ip6.__u6_addr.__u6_addr8[i] == 0 &&
4818 #endif
4819 					    i == sizeof (struct in6_addr) - 1)
4820 						numzero += 2;
4821 				}
4822 			}
4823 			ASSERT(firstzero + numzero <= sizeof (struct in6_addr));
4824 
4825 			/*
4826 			 * Check for an IPv4 embedded address.
4827 			 */
4828 			v6end = sizeof (struct in6_addr) - 2;
4829 			if (IN6_IS_ADDR_V4MAPPED(&ip6) ||
4830 			    IN6_IS_ADDR_V4COMPAT(&ip6)) {
4831 				for (i = sizeof (struct in6_addr) - 1;
4832 				    i >= DTRACE_V4MAPPED_OFFSET; i--) {
4833 					ASSERT(end >= base);
4834 
4835 #if defined(sun)
4836 					val = ip6._S6_un._S6_u8[i];
4837 #else
4838 					val = ip6.__u6_addr.__u6_addr8[i];
4839 #endif
4840 
4841 					if (val == 0) {
4842 						*end-- = '0';
4843 					} else {
4844 						for (; val; val /= 10) {
4845 							*end-- = '0' + val % 10;
4846 						}
4847 					}
4848 
4849 					if (i > DTRACE_V4MAPPED_OFFSET)
4850 						*end-- = '.';
4851 				}
4852 
4853 				if (subr == DIF_SUBR_INET_NTOA6)
4854 					goto inetout;
4855 
4856 				/*
4857 				 * Set v6end to skip the IPv4 address that
4858 				 * we have already stringified.
4859 				 */
4860 				v6end = 10;
4861 			}
4862 
4863 			/*
4864 			 * Build the IPv6 string by working through the
4865 			 * address in reverse.
4866 			 */
4867 			for (i = v6end; i >= 0; i -= 2) {
4868 				ASSERT(end >= base);
4869 
4870 				if (i == firstzero + numzero - 2) {
4871 					*end-- = ':';
4872 					*end-- = ':';
4873 					i -= numzero - 2;
4874 					continue;
4875 				}
4876 
4877 				if (i < 14 && i != firstzero - 2)
4878 					*end-- = ':';
4879 
4880 #if defined(sun)
4881 				val = (ip6._S6_un._S6_u8[i] << 8) +
4882 				    ip6._S6_un._S6_u8[i + 1];
4883 #else
4884 				val = (ip6.__u6_addr.__u6_addr8[i] << 8) +
4885 				    ip6.__u6_addr.__u6_addr8[i + 1];
4886 #endif
4887 
4888 				if (val == 0) {
4889 					*end-- = '0';
4890 				} else {
4891 					for (; val; val /= 16) {
4892 						*end-- = digits[val % 16];
4893 					}
4894 				}
4895 			}
4896 			ASSERT(end + 1 >= base);
4897 
4898 		} else {
4899 			/*
4900 			 * The user didn't use AH_INET or AH_INET6.
4901 			 */
4902 			DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
4903 			regs[rd] = 0;
4904 			break;
4905 		}
4906 
4907 inetout:	regs[rd] = (uintptr_t)end + 1;
4908 		mstate->dtms_scratch_ptr += size;
4909 		break;
4910 	}
4911 
4912 	case DIF_SUBR_MEMREF: {
4913 		uintptr_t size = 2 * sizeof(uintptr_t);
4914 		uintptr_t *memref = (uintptr_t *) P2ROUNDUP(mstate->dtms_scratch_ptr, sizeof(uintptr_t));
4915 		size_t scratch_size = ((uintptr_t) memref - mstate->dtms_scratch_ptr) + size;
4916 
4917 		/* address and length */
4918 		memref[0] = tupregs[0].dttk_value;
4919 		memref[1] = tupregs[1].dttk_value;
4920 
4921 		regs[rd] = (uintptr_t) memref;
4922 		mstate->dtms_scratch_ptr += scratch_size;
4923 		break;
4924 	}
4925 
4926 #if !defined(sun)
4927 	case DIF_SUBR_MEMSTR: {
4928 		char *str = (char *)mstate->dtms_scratch_ptr;
4929 		uintptr_t mem = tupregs[0].dttk_value;
4930 		char c = tupregs[1].dttk_value;
4931 		size_t size = tupregs[2].dttk_value;
4932 		uint8_t n;
4933 		int i;
4934 
4935 		regs[rd] = 0;
4936 
4937 		if (size == 0)
4938 			break;
4939 
4940 		if (!dtrace_canload(mem, size - 1, mstate, vstate))
4941 			break;
4942 
4943 		if (!DTRACE_INSCRATCH(mstate, size)) {
4944 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4945 			break;
4946 		}
4947 
4948 		if (dtrace_memstr_max != 0 && size > dtrace_memstr_max) {
4949 			*flags |= CPU_DTRACE_ILLOP;
4950 			break;
4951 		}
4952 
4953 		for (i = 0; i < size - 1; i++) {
4954 			n = dtrace_load8(mem++);
4955 			str[i] = (n == 0) ? c : n;
4956 		}
4957 		str[size - 1] = 0;
4958 
4959 		regs[rd] = (uintptr_t)str;
4960 		mstate->dtms_scratch_ptr += size;
4961 		break;
4962 	}
4963 #endif
4964 
4965 	case DIF_SUBR_TYPEREF: {
4966 		uintptr_t size = 4 * sizeof(uintptr_t);
4967 		uintptr_t *typeref = (uintptr_t *) P2ROUNDUP(mstate->dtms_scratch_ptr, sizeof(uintptr_t));
4968 		size_t scratch_size = ((uintptr_t) typeref - mstate->dtms_scratch_ptr) + size;
4969 
4970 		/* address, num_elements, type_str, type_len */
4971 		typeref[0] = tupregs[0].dttk_value;
4972 		typeref[1] = tupregs[1].dttk_value;
4973 		typeref[2] = tupregs[2].dttk_value;
4974 		typeref[3] = tupregs[3].dttk_value;
4975 
4976 		regs[rd] = (uintptr_t) typeref;
4977 		mstate->dtms_scratch_ptr += scratch_size;
4978 		break;
4979 	}
4980 	}
4981 }
4982 
4983 /*
4984  * Emulate the execution of DTrace IR instructions specified by the given
4985  * DIF object.  This function is deliberately void of assertions as all of
4986  * the necessary checks are handled by a call to dtrace_difo_validate().
4987  */
4988 static uint64_t
4989 dtrace_dif_emulate(dtrace_difo_t *difo, dtrace_mstate_t *mstate,
4990     dtrace_vstate_t *vstate, dtrace_state_t *state)
4991 {
4992 	const dif_instr_t *text = difo->dtdo_buf;
4993 	const uint_t textlen = difo->dtdo_len;
4994 	const char *strtab = difo->dtdo_strtab;
4995 	const uint64_t *inttab = difo->dtdo_inttab;
4996 
4997 	uint64_t rval = 0;
4998 	dtrace_statvar_t *svar;
4999 	dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
5000 	dtrace_difv_t *v;
5001 	volatile uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
5002 	volatile uintptr_t *illval = &cpu_core[curcpu].cpuc_dtrace_illval;
5003 
5004 	dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
5005 	uint64_t regs[DIF_DIR_NREGS];
5006 	uint64_t *tmp;
5007 
5008 	uint8_t cc_n = 0, cc_z = 0, cc_v = 0, cc_c = 0;
5009 	int64_t cc_r;
5010 	uint_t pc = 0, id, opc = 0;
5011 	uint8_t ttop = 0;
5012 	dif_instr_t instr;
5013 	uint_t r1, r2, rd;
5014 
5015 	/*
5016 	 * We stash the current DIF object into the machine state: we need it
5017 	 * for subsequent access checking.
5018 	 */
5019 	mstate->dtms_difo = difo;
5020 
5021 	regs[DIF_REG_R0] = 0; 		/* %r0 is fixed at zero */
5022 
5023 	while (pc < textlen && !(*flags & CPU_DTRACE_FAULT)) {
5024 		opc = pc;
5025 
5026 		instr = text[pc++];
5027 		r1 = DIF_INSTR_R1(instr);
5028 		r2 = DIF_INSTR_R2(instr);
5029 		rd = DIF_INSTR_RD(instr);
5030 
5031 		switch (DIF_INSTR_OP(instr)) {
5032 		case DIF_OP_OR:
5033 			regs[rd] = regs[r1] | regs[r2];
5034 			break;
5035 		case DIF_OP_XOR:
5036 			regs[rd] = regs[r1] ^ regs[r2];
5037 			break;
5038 		case DIF_OP_AND:
5039 			regs[rd] = regs[r1] & regs[r2];
5040 			break;
5041 		case DIF_OP_SLL:
5042 			regs[rd] = regs[r1] << regs[r2];
5043 			break;
5044 		case DIF_OP_SRL:
5045 			regs[rd] = regs[r1] >> regs[r2];
5046 			break;
5047 		case DIF_OP_SUB:
5048 			regs[rd] = regs[r1] - regs[r2];
5049 			break;
5050 		case DIF_OP_ADD:
5051 			regs[rd] = regs[r1] + regs[r2];
5052 			break;
5053 		case DIF_OP_MUL:
5054 			regs[rd] = regs[r1] * regs[r2];
5055 			break;
5056 		case DIF_OP_SDIV:
5057 			if (regs[r2] == 0) {
5058 				regs[rd] = 0;
5059 				*flags |= CPU_DTRACE_DIVZERO;
5060 			} else {
5061 				regs[rd] = (int64_t)regs[r1] /
5062 				    (int64_t)regs[r2];
5063 			}
5064 			break;
5065 
5066 		case DIF_OP_UDIV:
5067 			if (regs[r2] == 0) {
5068 				regs[rd] = 0;
5069 				*flags |= CPU_DTRACE_DIVZERO;
5070 			} else {
5071 				regs[rd] = regs[r1] / regs[r2];
5072 			}
5073 			break;
5074 
5075 		case DIF_OP_SREM:
5076 			if (regs[r2] == 0) {
5077 				regs[rd] = 0;
5078 				*flags |= CPU_DTRACE_DIVZERO;
5079 			} else {
5080 				regs[rd] = (int64_t)regs[r1] %
5081 				    (int64_t)regs[r2];
5082 			}
5083 			break;
5084 
5085 		case DIF_OP_UREM:
5086 			if (regs[r2] == 0) {
5087 				regs[rd] = 0;
5088 				*flags |= CPU_DTRACE_DIVZERO;
5089 			} else {
5090 				regs[rd] = regs[r1] % regs[r2];
5091 			}
5092 			break;
5093 
5094 		case DIF_OP_NOT:
5095 			regs[rd] = ~regs[r1];
5096 			break;
5097 		case DIF_OP_MOV:
5098 			regs[rd] = regs[r1];
5099 			break;
5100 		case DIF_OP_CMP:
5101 			cc_r = regs[r1] - regs[r2];
5102 			cc_n = cc_r < 0;
5103 			cc_z = cc_r == 0;
5104 			cc_v = 0;
5105 			cc_c = regs[r1] < regs[r2];
5106 			break;
5107 		case DIF_OP_TST:
5108 			cc_n = cc_v = cc_c = 0;
5109 			cc_z = regs[r1] == 0;
5110 			break;
5111 		case DIF_OP_BA:
5112 			pc = DIF_INSTR_LABEL(instr);
5113 			break;
5114 		case DIF_OP_BE:
5115 			if (cc_z)
5116 				pc = DIF_INSTR_LABEL(instr);
5117 			break;
5118 		case DIF_OP_BNE:
5119 			if (cc_z == 0)
5120 				pc = DIF_INSTR_LABEL(instr);
5121 			break;
5122 		case DIF_OP_BG:
5123 			if ((cc_z | (cc_n ^ cc_v)) == 0)
5124 				pc = DIF_INSTR_LABEL(instr);
5125 			break;
5126 		case DIF_OP_BGU:
5127 			if ((cc_c | cc_z) == 0)
5128 				pc = DIF_INSTR_LABEL(instr);
5129 			break;
5130 		case DIF_OP_BGE:
5131 			if ((cc_n ^ cc_v) == 0)
5132 				pc = DIF_INSTR_LABEL(instr);
5133 			break;
5134 		case DIF_OP_BGEU:
5135 			if (cc_c == 0)
5136 				pc = DIF_INSTR_LABEL(instr);
5137 			break;
5138 		case DIF_OP_BL:
5139 			if (cc_n ^ cc_v)
5140 				pc = DIF_INSTR_LABEL(instr);
5141 			break;
5142 		case DIF_OP_BLU:
5143 			if (cc_c)
5144 				pc = DIF_INSTR_LABEL(instr);
5145 			break;
5146 		case DIF_OP_BLE:
5147 			if (cc_z | (cc_n ^ cc_v))
5148 				pc = DIF_INSTR_LABEL(instr);
5149 			break;
5150 		case DIF_OP_BLEU:
5151 			if (cc_c | cc_z)
5152 				pc = DIF_INSTR_LABEL(instr);
5153 			break;
5154 		case DIF_OP_RLDSB:
5155 			if (!dtrace_canstore(regs[r1], 1, mstate, vstate)) {
5156 				*flags |= CPU_DTRACE_KPRIV;
5157 				*illval = regs[r1];
5158 				break;
5159 			}
5160 			/*FALLTHROUGH*/
5161 		case DIF_OP_LDSB:
5162 			regs[rd] = (int8_t)dtrace_load8(regs[r1]);
5163 			break;
5164 		case DIF_OP_RLDSH:
5165 			if (!dtrace_canstore(regs[r1], 2, mstate, vstate)) {
5166 				*flags |= CPU_DTRACE_KPRIV;
5167 				*illval = regs[r1];
5168 				break;
5169 			}
5170 			/*FALLTHROUGH*/
5171 		case DIF_OP_LDSH:
5172 			regs[rd] = (int16_t)dtrace_load16(regs[r1]);
5173 			break;
5174 		case DIF_OP_RLDSW:
5175 			if (!dtrace_canstore(regs[r1], 4, mstate, vstate)) {
5176 				*flags |= CPU_DTRACE_KPRIV;
5177 				*illval = regs[r1];
5178 				break;
5179 			}
5180 			/*FALLTHROUGH*/
5181 		case DIF_OP_LDSW:
5182 			regs[rd] = (int32_t)dtrace_load32(regs[r1]);
5183 			break;
5184 		case DIF_OP_RLDUB:
5185 			if (!dtrace_canstore(regs[r1], 1, mstate, vstate)) {
5186 				*flags |= CPU_DTRACE_KPRIV;
5187 				*illval = regs[r1];
5188 				break;
5189 			}
5190 			/*FALLTHROUGH*/
5191 		case DIF_OP_LDUB:
5192 			regs[rd] = dtrace_load8(regs[r1]);
5193 			break;
5194 		case DIF_OP_RLDUH:
5195 			if (!dtrace_canstore(regs[r1], 2, mstate, vstate)) {
5196 				*flags |= CPU_DTRACE_KPRIV;
5197 				*illval = regs[r1];
5198 				break;
5199 			}
5200 			/*FALLTHROUGH*/
5201 		case DIF_OP_LDUH:
5202 			regs[rd] = dtrace_load16(regs[r1]);
5203 			break;
5204 		case DIF_OP_RLDUW:
5205 			if (!dtrace_canstore(regs[r1], 4, mstate, vstate)) {
5206 				*flags |= CPU_DTRACE_KPRIV;
5207 				*illval = regs[r1];
5208 				break;
5209 			}
5210 			/*FALLTHROUGH*/
5211 		case DIF_OP_LDUW:
5212 			regs[rd] = dtrace_load32(regs[r1]);
5213 			break;
5214 		case DIF_OP_RLDX:
5215 			if (!dtrace_canstore(regs[r1], 8, mstate, vstate)) {
5216 				*flags |= CPU_DTRACE_KPRIV;
5217 				*illval = regs[r1];
5218 				break;
5219 			}
5220 			/*FALLTHROUGH*/
5221 		case DIF_OP_LDX:
5222 			regs[rd] = dtrace_load64(regs[r1]);
5223 			break;
5224 		case DIF_OP_ULDSB:
5225 			regs[rd] = (int8_t)
5226 			    dtrace_fuword8((void *)(uintptr_t)regs[r1]);
5227 			break;
5228 		case DIF_OP_ULDSH:
5229 			regs[rd] = (int16_t)
5230 			    dtrace_fuword16((void *)(uintptr_t)regs[r1]);
5231 			break;
5232 		case DIF_OP_ULDSW:
5233 			regs[rd] = (int32_t)
5234 			    dtrace_fuword32((void *)(uintptr_t)regs[r1]);
5235 			break;
5236 		case DIF_OP_ULDUB:
5237 			regs[rd] =
5238 			    dtrace_fuword8((void *)(uintptr_t)regs[r1]);
5239 			break;
5240 		case DIF_OP_ULDUH:
5241 			regs[rd] =
5242 			    dtrace_fuword16((void *)(uintptr_t)regs[r1]);
5243 			break;
5244 		case DIF_OP_ULDUW:
5245 			regs[rd] =
5246 			    dtrace_fuword32((void *)(uintptr_t)regs[r1]);
5247 			break;
5248 		case DIF_OP_ULDX:
5249 			regs[rd] =
5250 			    dtrace_fuword64((void *)(uintptr_t)regs[r1]);
5251 			break;
5252 		case DIF_OP_RET:
5253 			rval = regs[rd];
5254 			pc = textlen;
5255 			break;
5256 		case DIF_OP_NOP:
5257 			break;
5258 		case DIF_OP_SETX:
5259 			regs[rd] = inttab[DIF_INSTR_INTEGER(instr)];
5260 			break;
5261 		case DIF_OP_SETS:
5262 			regs[rd] = (uint64_t)(uintptr_t)
5263 			    (strtab + DIF_INSTR_STRING(instr));
5264 			break;
5265 		case DIF_OP_SCMP: {
5266 			size_t sz = state->dts_options[DTRACEOPT_STRSIZE];
5267 			uintptr_t s1 = regs[r1];
5268 			uintptr_t s2 = regs[r2];
5269 
5270 			if (s1 != 0 &&
5271 			    !dtrace_strcanload(s1, sz, mstate, vstate))
5272 				break;
5273 			if (s2 != 0 &&
5274 			    !dtrace_strcanload(s2, sz, mstate, vstate))
5275 				break;
5276 
5277 			cc_r = dtrace_strncmp((char *)s1, (char *)s2, sz);
5278 
5279 			cc_n = cc_r < 0;
5280 			cc_z = cc_r == 0;
5281 			cc_v = cc_c = 0;
5282 			break;
5283 		}
5284 		case DIF_OP_LDGA:
5285 			regs[rd] = dtrace_dif_variable(mstate, state,
5286 			    r1, regs[r2]);
5287 			break;
5288 		case DIF_OP_LDGS:
5289 			id = DIF_INSTR_VAR(instr);
5290 
5291 			if (id >= DIF_VAR_OTHER_UBASE) {
5292 				uintptr_t a;
5293 
5294 				id -= DIF_VAR_OTHER_UBASE;
5295 				svar = vstate->dtvs_globals[id];
5296 				ASSERT(svar != NULL);
5297 				v = &svar->dtsv_var;
5298 
5299 				if (!(v->dtdv_type.dtdt_flags & DIF_TF_BYREF)) {
5300 					regs[rd] = svar->dtsv_data;
5301 					break;
5302 				}
5303 
5304 				a = (uintptr_t)svar->dtsv_data;
5305 
5306 				if (*(uint8_t *)a == UINT8_MAX) {
5307 					/*
5308 					 * If the 0th byte is set to UINT8_MAX
5309 					 * then this is to be treated as a
5310 					 * reference to a NULL variable.
5311 					 */
5312 					regs[rd] = 0;
5313 				} else {
5314 					regs[rd] = a + sizeof (uint64_t);
5315 				}
5316 
5317 				break;
5318 			}
5319 
5320 			regs[rd] = dtrace_dif_variable(mstate, state, id, 0);
5321 			break;
5322 
5323 		case DIF_OP_STGS:
5324 			id = DIF_INSTR_VAR(instr);
5325 
5326 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
5327 			id -= DIF_VAR_OTHER_UBASE;
5328 
5329 			svar = vstate->dtvs_globals[id];
5330 			ASSERT(svar != NULL);
5331 			v = &svar->dtsv_var;
5332 
5333 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5334 				uintptr_t a = (uintptr_t)svar->dtsv_data;
5335 
5336 				ASSERT(a != 0);
5337 				ASSERT(svar->dtsv_size != 0);
5338 
5339 				if (regs[rd] == 0) {
5340 					*(uint8_t *)a = UINT8_MAX;
5341 					break;
5342 				} else {
5343 					*(uint8_t *)a = 0;
5344 					a += sizeof (uint64_t);
5345 				}
5346 				if (!dtrace_vcanload(
5347 				    (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5348 				    mstate, vstate))
5349 					break;
5350 
5351 				dtrace_vcopy((void *)(uintptr_t)regs[rd],
5352 				    (void *)a, &v->dtdv_type);
5353 				break;
5354 			}
5355 
5356 			svar->dtsv_data = regs[rd];
5357 			break;
5358 
5359 		case DIF_OP_LDTA:
5360 			/*
5361 			 * There are no DTrace built-in thread-local arrays at
5362 			 * present.  This opcode is saved for future work.
5363 			 */
5364 			*flags |= CPU_DTRACE_ILLOP;
5365 			regs[rd] = 0;
5366 			break;
5367 
5368 		case DIF_OP_LDLS:
5369 			id = DIF_INSTR_VAR(instr);
5370 
5371 			if (id < DIF_VAR_OTHER_UBASE) {
5372 				/*
5373 				 * For now, this has no meaning.
5374 				 */
5375 				regs[rd] = 0;
5376 				break;
5377 			}
5378 
5379 			id -= DIF_VAR_OTHER_UBASE;
5380 
5381 			ASSERT(id < vstate->dtvs_nlocals);
5382 			ASSERT(vstate->dtvs_locals != NULL);
5383 
5384 			svar = vstate->dtvs_locals[id];
5385 			ASSERT(svar != NULL);
5386 			v = &svar->dtsv_var;
5387 
5388 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5389 				uintptr_t a = (uintptr_t)svar->dtsv_data;
5390 				size_t sz = v->dtdv_type.dtdt_size;
5391 
5392 				sz += sizeof (uint64_t);
5393 				ASSERT(svar->dtsv_size == NCPU * sz);
5394 				a += curcpu * sz;
5395 
5396 				if (*(uint8_t *)a == UINT8_MAX) {
5397 					/*
5398 					 * If the 0th byte is set to UINT8_MAX
5399 					 * then this is to be treated as a
5400 					 * reference to a NULL variable.
5401 					 */
5402 					regs[rd] = 0;
5403 				} else {
5404 					regs[rd] = a + sizeof (uint64_t);
5405 				}
5406 
5407 				break;
5408 			}
5409 
5410 			ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
5411 			tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
5412 			regs[rd] = tmp[curcpu];
5413 			break;
5414 
5415 		case DIF_OP_STLS:
5416 			id = DIF_INSTR_VAR(instr);
5417 
5418 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
5419 			id -= DIF_VAR_OTHER_UBASE;
5420 			ASSERT(id < vstate->dtvs_nlocals);
5421 
5422 			ASSERT(vstate->dtvs_locals != NULL);
5423 			svar = vstate->dtvs_locals[id];
5424 			ASSERT(svar != NULL);
5425 			v = &svar->dtsv_var;
5426 
5427 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5428 				uintptr_t a = (uintptr_t)svar->dtsv_data;
5429 				size_t sz = v->dtdv_type.dtdt_size;
5430 
5431 				sz += sizeof (uint64_t);
5432 				ASSERT(svar->dtsv_size == NCPU * sz);
5433 				a += curcpu * sz;
5434 
5435 				if (regs[rd] == 0) {
5436 					*(uint8_t *)a = UINT8_MAX;
5437 					break;
5438 				} else {
5439 					*(uint8_t *)a = 0;
5440 					a += sizeof (uint64_t);
5441 				}
5442 
5443 				if (!dtrace_vcanload(
5444 				    (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5445 				    mstate, vstate))
5446 					break;
5447 
5448 				dtrace_vcopy((void *)(uintptr_t)regs[rd],
5449 				    (void *)a, &v->dtdv_type);
5450 				break;
5451 			}
5452 
5453 			ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
5454 			tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
5455 			tmp[curcpu] = regs[rd];
5456 			break;
5457 
5458 		case DIF_OP_LDTS: {
5459 			dtrace_dynvar_t *dvar;
5460 			dtrace_key_t *key;
5461 
5462 			id = DIF_INSTR_VAR(instr);
5463 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
5464 			id -= DIF_VAR_OTHER_UBASE;
5465 			v = &vstate->dtvs_tlocals[id];
5466 
5467 			key = &tupregs[DIF_DTR_NREGS];
5468 			key[0].dttk_value = (uint64_t)id;
5469 			key[0].dttk_size = 0;
5470 			DTRACE_TLS_THRKEY(key[1].dttk_value);
5471 			key[1].dttk_size = 0;
5472 
5473 			dvar = dtrace_dynvar(dstate, 2, key,
5474 			    sizeof (uint64_t), DTRACE_DYNVAR_NOALLOC,
5475 			    mstate, vstate);
5476 
5477 			if (dvar == NULL) {
5478 				regs[rd] = 0;
5479 				break;
5480 			}
5481 
5482 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5483 				regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
5484 			} else {
5485 				regs[rd] = *((uint64_t *)dvar->dtdv_data);
5486 			}
5487 
5488 			break;
5489 		}
5490 
5491 		case DIF_OP_STTS: {
5492 			dtrace_dynvar_t *dvar;
5493 			dtrace_key_t *key;
5494 
5495 			id = DIF_INSTR_VAR(instr);
5496 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
5497 			id -= DIF_VAR_OTHER_UBASE;
5498 
5499 			key = &tupregs[DIF_DTR_NREGS];
5500 			key[0].dttk_value = (uint64_t)id;
5501 			key[0].dttk_size = 0;
5502 			DTRACE_TLS_THRKEY(key[1].dttk_value);
5503 			key[1].dttk_size = 0;
5504 			v = &vstate->dtvs_tlocals[id];
5505 
5506 			dvar = dtrace_dynvar(dstate, 2, key,
5507 			    v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
5508 			    v->dtdv_type.dtdt_size : sizeof (uint64_t),
5509 			    regs[rd] ? DTRACE_DYNVAR_ALLOC :
5510 			    DTRACE_DYNVAR_DEALLOC, mstate, vstate);
5511 
5512 			/*
5513 			 * Given that we're storing to thread-local data,
5514 			 * we need to flush our predicate cache.
5515 			 */
5516 			curthread->t_predcache = 0;
5517 
5518 			if (dvar == NULL)
5519 				break;
5520 
5521 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5522 				if (!dtrace_vcanload(
5523 				    (void *)(uintptr_t)regs[rd],
5524 				    &v->dtdv_type, mstate, vstate))
5525 					break;
5526 
5527 				dtrace_vcopy((void *)(uintptr_t)regs[rd],
5528 				    dvar->dtdv_data, &v->dtdv_type);
5529 			} else {
5530 				*((uint64_t *)dvar->dtdv_data) = regs[rd];
5531 			}
5532 
5533 			break;
5534 		}
5535 
5536 		case DIF_OP_SRA:
5537 			regs[rd] = (int64_t)regs[r1] >> regs[r2];
5538 			break;
5539 
5540 		case DIF_OP_CALL:
5541 			dtrace_dif_subr(DIF_INSTR_SUBR(instr), rd,
5542 			    regs, tupregs, ttop, mstate, state);
5543 			break;
5544 
5545 		case DIF_OP_PUSHTR:
5546 			if (ttop == DIF_DTR_NREGS) {
5547 				*flags |= CPU_DTRACE_TUPOFLOW;
5548 				break;
5549 			}
5550 
5551 			if (r1 == DIF_TYPE_STRING) {
5552 				/*
5553 				 * If this is a string type and the size is 0,
5554 				 * we'll use the system-wide default string
5555 				 * size.  Note that we are _not_ looking at
5556 				 * the value of the DTRACEOPT_STRSIZE option;
5557 				 * had this been set, we would expect to have
5558 				 * a non-zero size value in the "pushtr".
5559 				 */
5560 				tupregs[ttop].dttk_size =
5561 				    dtrace_strlen((char *)(uintptr_t)regs[rd],
5562 				    regs[r2] ? regs[r2] :
5563 				    dtrace_strsize_default) + 1;
5564 			} else {
5565 				tupregs[ttop].dttk_size = regs[r2];
5566 			}
5567 
5568 			tupregs[ttop++].dttk_value = regs[rd];
5569 			break;
5570 
5571 		case DIF_OP_PUSHTV:
5572 			if (ttop == DIF_DTR_NREGS) {
5573 				*flags |= CPU_DTRACE_TUPOFLOW;
5574 				break;
5575 			}
5576 
5577 			tupregs[ttop].dttk_value = regs[rd];
5578 			tupregs[ttop++].dttk_size = 0;
5579 			break;
5580 
5581 		case DIF_OP_POPTS:
5582 			if (ttop != 0)
5583 				ttop--;
5584 			break;
5585 
5586 		case DIF_OP_FLUSHTS:
5587 			ttop = 0;
5588 			break;
5589 
5590 		case DIF_OP_LDGAA:
5591 		case DIF_OP_LDTAA: {
5592 			dtrace_dynvar_t *dvar;
5593 			dtrace_key_t *key = tupregs;
5594 			uint_t nkeys = ttop;
5595 
5596 			id = DIF_INSTR_VAR(instr);
5597 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
5598 			id -= DIF_VAR_OTHER_UBASE;
5599 
5600 			key[nkeys].dttk_value = (uint64_t)id;
5601 			key[nkeys++].dttk_size = 0;
5602 
5603 			if (DIF_INSTR_OP(instr) == DIF_OP_LDTAA) {
5604 				DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
5605 				key[nkeys++].dttk_size = 0;
5606 				v = &vstate->dtvs_tlocals[id];
5607 			} else {
5608 				v = &vstate->dtvs_globals[id]->dtsv_var;
5609 			}
5610 
5611 			dvar = dtrace_dynvar(dstate, nkeys, key,
5612 			    v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
5613 			    v->dtdv_type.dtdt_size : sizeof (uint64_t),
5614 			    DTRACE_DYNVAR_NOALLOC, mstate, vstate);
5615 
5616 			if (dvar == NULL) {
5617 				regs[rd] = 0;
5618 				break;
5619 			}
5620 
5621 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5622 				regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
5623 			} else {
5624 				regs[rd] = *((uint64_t *)dvar->dtdv_data);
5625 			}
5626 
5627 			break;
5628 		}
5629 
5630 		case DIF_OP_STGAA:
5631 		case DIF_OP_STTAA: {
5632 			dtrace_dynvar_t *dvar;
5633 			dtrace_key_t *key = tupregs;
5634 			uint_t nkeys = ttop;
5635 
5636 			id = DIF_INSTR_VAR(instr);
5637 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
5638 			id -= DIF_VAR_OTHER_UBASE;
5639 
5640 			key[nkeys].dttk_value = (uint64_t)id;
5641 			key[nkeys++].dttk_size = 0;
5642 
5643 			if (DIF_INSTR_OP(instr) == DIF_OP_STTAA) {
5644 				DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
5645 				key[nkeys++].dttk_size = 0;
5646 				v = &vstate->dtvs_tlocals[id];
5647 			} else {
5648 				v = &vstate->dtvs_globals[id]->dtsv_var;
5649 			}
5650 
5651 			dvar = dtrace_dynvar(dstate, nkeys, key,
5652 			    v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
5653 			    v->dtdv_type.dtdt_size : sizeof (uint64_t),
5654 			    regs[rd] ? DTRACE_DYNVAR_ALLOC :
5655 			    DTRACE_DYNVAR_DEALLOC, mstate, vstate);
5656 
5657 			if (dvar == NULL)
5658 				break;
5659 
5660 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5661 				if (!dtrace_vcanload(
5662 				    (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5663 				    mstate, vstate))
5664 					break;
5665 
5666 				dtrace_vcopy((void *)(uintptr_t)regs[rd],
5667 				    dvar->dtdv_data, &v->dtdv_type);
5668 			} else {
5669 				*((uint64_t *)dvar->dtdv_data) = regs[rd];
5670 			}
5671 
5672 			break;
5673 		}
5674 
5675 		case DIF_OP_ALLOCS: {
5676 			uintptr_t ptr = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
5677 			size_t size = ptr - mstate->dtms_scratch_ptr + regs[r1];
5678 
5679 			/*
5680 			 * Rounding up the user allocation size could have
5681 			 * overflowed large, bogus allocations (like -1ULL) to
5682 			 * 0.
5683 			 */
5684 			if (size < regs[r1] ||
5685 			    !DTRACE_INSCRATCH(mstate, size)) {
5686 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5687 				regs[rd] = 0;
5688 				break;
5689 			}
5690 
5691 			dtrace_bzero((void *) mstate->dtms_scratch_ptr, size);
5692 			mstate->dtms_scratch_ptr += size;
5693 			regs[rd] = ptr;
5694 			break;
5695 		}
5696 
5697 		case DIF_OP_COPYS:
5698 			if (!dtrace_canstore(regs[rd], regs[r2],
5699 			    mstate, vstate)) {
5700 				*flags |= CPU_DTRACE_BADADDR;
5701 				*illval = regs[rd];
5702 				break;
5703 			}
5704 
5705 			if (!dtrace_canload(regs[r1], regs[r2], mstate, vstate))
5706 				break;
5707 
5708 			dtrace_bcopy((void *)(uintptr_t)regs[r1],
5709 			    (void *)(uintptr_t)regs[rd], (size_t)regs[r2]);
5710 			break;
5711 
5712 		case DIF_OP_STB:
5713 			if (!dtrace_canstore(regs[rd], 1, mstate, vstate)) {
5714 				*flags |= CPU_DTRACE_BADADDR;
5715 				*illval = regs[rd];
5716 				break;
5717 			}
5718 			*((uint8_t *)(uintptr_t)regs[rd]) = (uint8_t)regs[r1];
5719 			break;
5720 
5721 		case DIF_OP_STH:
5722 			if (!dtrace_canstore(regs[rd], 2, mstate, vstate)) {
5723 				*flags |= CPU_DTRACE_BADADDR;
5724 				*illval = regs[rd];
5725 				break;
5726 			}
5727 			if (regs[rd] & 1) {
5728 				*flags |= CPU_DTRACE_BADALIGN;
5729 				*illval = regs[rd];
5730 				break;
5731 			}
5732 			*((uint16_t *)(uintptr_t)regs[rd]) = (uint16_t)regs[r1];
5733 			break;
5734 
5735 		case DIF_OP_STW:
5736 			if (!dtrace_canstore(regs[rd], 4, mstate, vstate)) {
5737 				*flags |= CPU_DTRACE_BADADDR;
5738 				*illval = regs[rd];
5739 				break;
5740 			}
5741 			if (regs[rd] & 3) {
5742 				*flags |= CPU_DTRACE_BADALIGN;
5743 				*illval = regs[rd];
5744 				break;
5745 			}
5746 			*((uint32_t *)(uintptr_t)regs[rd]) = (uint32_t)regs[r1];
5747 			break;
5748 
5749 		case DIF_OP_STX:
5750 			if (!dtrace_canstore(regs[rd], 8, mstate, vstate)) {
5751 				*flags |= CPU_DTRACE_BADADDR;
5752 				*illval = regs[rd];
5753 				break;
5754 			}
5755 			if (regs[rd] & 7) {
5756 				*flags |= CPU_DTRACE_BADALIGN;
5757 				*illval = regs[rd];
5758 				break;
5759 			}
5760 			*((uint64_t *)(uintptr_t)regs[rd]) = regs[r1];
5761 			break;
5762 		}
5763 	}
5764 
5765 	if (!(*flags & CPU_DTRACE_FAULT))
5766 		return (rval);
5767 
5768 	mstate->dtms_fltoffs = opc * sizeof (dif_instr_t);
5769 	mstate->dtms_present |= DTRACE_MSTATE_FLTOFFS;
5770 
5771 	return (0);
5772 }
5773 
5774 static void
5775 dtrace_action_breakpoint(dtrace_ecb_t *ecb)
5776 {
5777 	dtrace_probe_t *probe = ecb->dte_probe;
5778 	dtrace_provider_t *prov = probe->dtpr_provider;
5779 	char c[DTRACE_FULLNAMELEN + 80], *str;
5780 	char *msg = "dtrace: breakpoint action at probe ";
5781 	char *ecbmsg = " (ecb ";
5782 	uintptr_t mask = (0xf << (sizeof (uintptr_t) * NBBY / 4));
5783 	uintptr_t val = (uintptr_t)ecb;
5784 	int shift = (sizeof (uintptr_t) * NBBY) - 4, i = 0;
5785 
5786 	if (dtrace_destructive_disallow)
5787 		return;
5788 
5789 	/*
5790 	 * It's impossible to be taking action on the NULL probe.
5791 	 */
5792 	ASSERT(probe != NULL);
5793 
5794 	/*
5795 	 * This is a poor man's (destitute man's?) sprintf():  we want to
5796 	 * print the provider name, module name, function name and name of
5797 	 * the probe, along with the hex address of the ECB with the breakpoint
5798 	 * action -- all of which we must place in the character buffer by
5799 	 * hand.
5800 	 */
5801 	while (*msg != '\0')
5802 		c[i++] = *msg++;
5803 
5804 	for (str = prov->dtpv_name; *str != '\0'; str++)
5805 		c[i++] = *str;
5806 	c[i++] = ':';
5807 
5808 	for (str = probe->dtpr_mod; *str != '\0'; str++)
5809 		c[i++] = *str;
5810 	c[i++] = ':';
5811 
5812 	for (str = probe->dtpr_func; *str != '\0'; str++)
5813 		c[i++] = *str;
5814 	c[i++] = ':';
5815 
5816 	for (str = probe->dtpr_name; *str != '\0'; str++)
5817 		c[i++] = *str;
5818 
5819 	while (*ecbmsg != '\0')
5820 		c[i++] = *ecbmsg++;
5821 
5822 	while (shift >= 0) {
5823 		mask = (uintptr_t)0xf << shift;
5824 
5825 		if (val >= ((uintptr_t)1 << shift))
5826 			c[i++] = "0123456789abcdef"[(val & mask) >> shift];
5827 		shift -= 4;
5828 	}
5829 
5830 	c[i++] = ')';
5831 	c[i] = '\0';
5832 
5833 #if defined(sun)
5834 	debug_enter(c);
5835 #else
5836 	kdb_enter(KDB_WHY_DTRACE, "breakpoint action");
5837 #endif
5838 }
5839 
5840 static void
5841 dtrace_action_panic(dtrace_ecb_t *ecb)
5842 {
5843 	dtrace_probe_t *probe = ecb->dte_probe;
5844 
5845 	/*
5846 	 * It's impossible to be taking action on the NULL probe.
5847 	 */
5848 	ASSERT(probe != NULL);
5849 
5850 	if (dtrace_destructive_disallow)
5851 		return;
5852 
5853 	if (dtrace_panicked != NULL)
5854 		return;
5855 
5856 	if (dtrace_casptr(&dtrace_panicked, NULL, curthread) != NULL)
5857 		return;
5858 
5859 	/*
5860 	 * We won the right to panic.  (We want to be sure that only one
5861 	 * thread calls panic() from dtrace_probe(), and that panic() is
5862 	 * called exactly once.)
5863 	 */
5864 	dtrace_panic("dtrace: panic action at probe %s:%s:%s:%s (ecb %p)",
5865 	    probe->dtpr_provider->dtpv_name, probe->dtpr_mod,
5866 	    probe->dtpr_func, probe->dtpr_name, (void *)ecb);
5867 }
5868 
5869 static void
5870 dtrace_action_raise(uint64_t sig)
5871 {
5872 	if (dtrace_destructive_disallow)
5873 		return;
5874 
5875 	if (sig >= NSIG) {
5876 		DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
5877 		return;
5878 	}
5879 
5880 #if defined(sun)
5881 	/*
5882 	 * raise() has a queue depth of 1 -- we ignore all subsequent
5883 	 * invocations of the raise() action.
5884 	 */
5885 	if (curthread->t_dtrace_sig == 0)
5886 		curthread->t_dtrace_sig = (uint8_t)sig;
5887 
5888 	curthread->t_sig_check = 1;
5889 	aston(curthread);
5890 #else
5891 	struct proc *p = curproc;
5892 	PROC_LOCK(p);
5893 	kern_psignal(p, sig);
5894 	PROC_UNLOCK(p);
5895 #endif
5896 }
5897 
5898 static void
5899 dtrace_action_stop(void)
5900 {
5901 	if (dtrace_destructive_disallow)
5902 		return;
5903 
5904 #if defined(sun)
5905 	if (!curthread->t_dtrace_stop) {
5906 		curthread->t_dtrace_stop = 1;
5907 		curthread->t_sig_check = 1;
5908 		aston(curthread);
5909 	}
5910 #else
5911 	struct proc *p = curproc;
5912 	PROC_LOCK(p);
5913 	kern_psignal(p, SIGSTOP);
5914 	PROC_UNLOCK(p);
5915 #endif
5916 }
5917 
5918 static void
5919 dtrace_action_chill(dtrace_mstate_t *mstate, hrtime_t val)
5920 {
5921 	hrtime_t now;
5922 	volatile uint16_t *flags;
5923 #if defined(sun)
5924 	cpu_t *cpu = CPU;
5925 #else
5926 	cpu_t *cpu = &solaris_cpu[curcpu];
5927 #endif
5928 
5929 	if (dtrace_destructive_disallow)
5930 		return;
5931 
5932 	flags = (volatile uint16_t *)&cpu_core[cpu->cpu_id].cpuc_dtrace_flags;
5933 
5934 	now = dtrace_gethrtime();
5935 
5936 	if (now - cpu->cpu_dtrace_chillmark > dtrace_chill_interval) {
5937 		/*
5938 		 * We need to advance the mark to the current time.
5939 		 */
5940 		cpu->cpu_dtrace_chillmark = now;
5941 		cpu->cpu_dtrace_chilled = 0;
5942 	}
5943 
5944 	/*
5945 	 * Now check to see if the requested chill time would take us over
5946 	 * the maximum amount of time allowed in the chill interval.  (Or
5947 	 * worse, if the calculation itself induces overflow.)
5948 	 */
5949 	if (cpu->cpu_dtrace_chilled + val > dtrace_chill_max ||
5950 	    cpu->cpu_dtrace_chilled + val < cpu->cpu_dtrace_chilled) {
5951 		*flags |= CPU_DTRACE_ILLOP;
5952 		return;
5953 	}
5954 
5955 	while (dtrace_gethrtime() - now < val)
5956 		continue;
5957 
5958 	/*
5959 	 * Normally, we assure that the value of the variable "timestamp" does
5960 	 * not change within an ECB.  The presence of chill() represents an
5961 	 * exception to this rule, however.
5962 	 */
5963 	mstate->dtms_present &= ~DTRACE_MSTATE_TIMESTAMP;
5964 	cpu->cpu_dtrace_chilled += val;
5965 }
5966 
5967 static void
5968 dtrace_action_ustack(dtrace_mstate_t *mstate, dtrace_state_t *state,
5969     uint64_t *buf, uint64_t arg)
5970 {
5971 	int nframes = DTRACE_USTACK_NFRAMES(arg);
5972 	int strsize = DTRACE_USTACK_STRSIZE(arg);
5973 	uint64_t *pcs = &buf[1], *fps;
5974 	char *str = (char *)&pcs[nframes];
5975 	int size, offs = 0, i, j;
5976 	uintptr_t old = mstate->dtms_scratch_ptr, saved;
5977 	uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
5978 	char *sym;
5979 
5980 	/*
5981 	 * Should be taking a faster path if string space has not been
5982 	 * allocated.
5983 	 */
5984 	ASSERT(strsize != 0);
5985 
5986 	/*
5987 	 * We will first allocate some temporary space for the frame pointers.
5988 	 */
5989 	fps = (uint64_t *)P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
5990 	size = (uintptr_t)fps - mstate->dtms_scratch_ptr +
5991 	    (nframes * sizeof (uint64_t));
5992 
5993 	if (!DTRACE_INSCRATCH(mstate, size)) {
5994 		/*
5995 		 * Not enough room for our frame pointers -- need to indicate
5996 		 * that we ran out of scratch space.
5997 		 */
5998 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5999 		return;
6000 	}
6001 
6002 	mstate->dtms_scratch_ptr += size;
6003 	saved = mstate->dtms_scratch_ptr;
6004 
6005 	/*
6006 	 * Now get a stack with both program counters and frame pointers.
6007 	 */
6008 	DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6009 	dtrace_getufpstack(buf, fps, nframes + 1);
6010 	DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6011 
6012 	/*
6013 	 * If that faulted, we're cooked.
6014 	 */
6015 	if (*flags & CPU_DTRACE_FAULT)
6016 		goto out;
6017 
6018 	/*
6019 	 * Now we want to walk up the stack, calling the USTACK helper.  For
6020 	 * each iteration, we restore the scratch pointer.
6021 	 */
6022 	for (i = 0; i < nframes; i++) {
6023 		mstate->dtms_scratch_ptr = saved;
6024 
6025 		if (offs >= strsize)
6026 			break;
6027 
6028 		sym = (char *)(uintptr_t)dtrace_helper(
6029 		    DTRACE_HELPER_ACTION_USTACK,
6030 		    mstate, state, pcs[i], fps[i]);
6031 
6032 		/*
6033 		 * If we faulted while running the helper, we're going to
6034 		 * clear the fault and null out the corresponding string.
6035 		 */
6036 		if (*flags & CPU_DTRACE_FAULT) {
6037 			*flags &= ~CPU_DTRACE_FAULT;
6038 			str[offs++] = '\0';
6039 			continue;
6040 		}
6041 
6042 		if (sym == NULL) {
6043 			str[offs++] = '\0';
6044 			continue;
6045 		}
6046 
6047 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6048 
6049 		/*
6050 		 * Now copy in the string that the helper returned to us.
6051 		 */
6052 		for (j = 0; offs + j < strsize; j++) {
6053 			if ((str[offs + j] = sym[j]) == '\0')
6054 				break;
6055 		}
6056 
6057 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6058 
6059 		offs += j + 1;
6060 	}
6061 
6062 	if (offs >= strsize) {
6063 		/*
6064 		 * If we didn't have room for all of the strings, we don't
6065 		 * abort processing -- this needn't be a fatal error -- but we
6066 		 * still want to increment a counter (dts_stkstroverflows) to
6067 		 * allow this condition to be warned about.  (If this is from
6068 		 * a jstack() action, it is easily tuned via jstackstrsize.)
6069 		 */
6070 		dtrace_error(&state->dts_stkstroverflows);
6071 	}
6072 
6073 	while (offs < strsize)
6074 		str[offs++] = '\0';
6075 
6076 out:
6077 	mstate->dtms_scratch_ptr = old;
6078 }
6079 
6080 /*
6081  * If you're looking for the epicenter of DTrace, you just found it.  This
6082  * is the function called by the provider to fire a probe -- from which all
6083  * subsequent probe-context DTrace activity emanates.
6084  */
6085 void
6086 dtrace_probe(dtrace_id_t id, uintptr_t arg0, uintptr_t arg1,
6087     uintptr_t arg2, uintptr_t arg3, uintptr_t arg4)
6088 {
6089 	processorid_t cpuid;
6090 	dtrace_icookie_t cookie;
6091 	dtrace_probe_t *probe;
6092 	dtrace_mstate_t mstate;
6093 	dtrace_ecb_t *ecb;
6094 	dtrace_action_t *act;
6095 	intptr_t offs;
6096 	size_t size;
6097 	int vtime, onintr;
6098 	volatile uint16_t *flags;
6099 	hrtime_t now;
6100 
6101 	if (panicstr != NULL)
6102 		return;
6103 
6104 #if defined(sun)
6105 	/*
6106 	 * Kick out immediately if this CPU is still being born (in which case
6107 	 * curthread will be set to -1) or the current thread can't allow
6108 	 * probes in its current context.
6109 	 */
6110 	if (((uintptr_t)curthread & 1) || (curthread->t_flag & T_DONTDTRACE))
6111 		return;
6112 #endif
6113 
6114 	cookie = dtrace_interrupt_disable();
6115 	probe = dtrace_probes[id - 1];
6116 	cpuid = curcpu;
6117 	onintr = CPU_ON_INTR(CPU);
6118 
6119 	if (!onintr && probe->dtpr_predcache != DTRACE_CACHEIDNONE &&
6120 	    probe->dtpr_predcache == curthread->t_predcache) {
6121 		/*
6122 		 * We have hit in the predicate cache; we know that
6123 		 * this predicate would evaluate to be false.
6124 		 */
6125 		dtrace_interrupt_enable(cookie);
6126 		return;
6127 	}
6128 
6129 #if defined(sun)
6130 	if (panic_quiesce) {
6131 #else
6132 	if (panicstr != NULL) {
6133 #endif
6134 		/*
6135 		 * We don't trace anything if we're panicking.
6136 		 */
6137 		dtrace_interrupt_enable(cookie);
6138 		return;
6139 	}
6140 
6141 	now = dtrace_gethrtime();
6142 	vtime = dtrace_vtime_references != 0;
6143 
6144 	if (vtime && curthread->t_dtrace_start)
6145 		curthread->t_dtrace_vtime += now - curthread->t_dtrace_start;
6146 
6147 	mstate.dtms_difo = NULL;
6148 	mstate.dtms_probe = probe;
6149 	mstate.dtms_strtok = 0;
6150 	mstate.dtms_arg[0] = arg0;
6151 	mstate.dtms_arg[1] = arg1;
6152 	mstate.dtms_arg[2] = arg2;
6153 	mstate.dtms_arg[3] = arg3;
6154 	mstate.dtms_arg[4] = arg4;
6155 
6156 	flags = (volatile uint16_t *)&cpu_core[cpuid].cpuc_dtrace_flags;
6157 
6158 	for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
6159 		dtrace_predicate_t *pred = ecb->dte_predicate;
6160 		dtrace_state_t *state = ecb->dte_state;
6161 		dtrace_buffer_t *buf = &state->dts_buffer[cpuid];
6162 		dtrace_buffer_t *aggbuf = &state->dts_aggbuffer[cpuid];
6163 		dtrace_vstate_t *vstate = &state->dts_vstate;
6164 		dtrace_provider_t *prov = probe->dtpr_provider;
6165 		uint64_t tracememsize = 0;
6166 		int committed = 0;
6167 		caddr_t tomax;
6168 
6169 		/*
6170 		 * A little subtlety with the following (seemingly innocuous)
6171 		 * declaration of the automatic 'val':  by looking at the
6172 		 * code, you might think that it could be declared in the
6173 		 * action processing loop, below.  (That is, it's only used in
6174 		 * the action processing loop.)  However, it must be declared
6175 		 * out of that scope because in the case of DIF expression
6176 		 * arguments to aggregating actions, one iteration of the
6177 		 * action loop will use the last iteration's value.
6178 		 */
6179 		uint64_t val = 0;
6180 
6181 		mstate.dtms_present = DTRACE_MSTATE_ARGS | DTRACE_MSTATE_PROBE;
6182 		*flags &= ~CPU_DTRACE_ERROR;
6183 
6184 		if (prov == dtrace_provider) {
6185 			/*
6186 			 * If dtrace itself is the provider of this probe,
6187 			 * we're only going to continue processing the ECB if
6188 			 * arg0 (the dtrace_state_t) is equal to the ECB's
6189 			 * creating state.  (This prevents disjoint consumers
6190 			 * from seeing one another's metaprobes.)
6191 			 */
6192 			if (arg0 != (uint64_t)(uintptr_t)state)
6193 				continue;
6194 		}
6195 
6196 		if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE) {
6197 			/*
6198 			 * We're not currently active.  If our provider isn't
6199 			 * the dtrace pseudo provider, we're not interested.
6200 			 */
6201 			if (prov != dtrace_provider)
6202 				continue;
6203 
6204 			/*
6205 			 * Now we must further check if we are in the BEGIN
6206 			 * probe.  If we are, we will only continue processing
6207 			 * if we're still in WARMUP -- if one BEGIN enabling
6208 			 * has invoked the exit() action, we don't want to
6209 			 * evaluate subsequent BEGIN enablings.
6210 			 */
6211 			if (probe->dtpr_id == dtrace_probeid_begin &&
6212 			    state->dts_activity != DTRACE_ACTIVITY_WARMUP) {
6213 				ASSERT(state->dts_activity ==
6214 				    DTRACE_ACTIVITY_DRAINING);
6215 				continue;
6216 			}
6217 		}
6218 
6219 		if (ecb->dte_cond) {
6220 			/*
6221 			 * If the dte_cond bits indicate that this
6222 			 * consumer is only allowed to see user-mode firings
6223 			 * of this probe, call the provider's dtps_usermode()
6224 			 * entry point to check that the probe was fired
6225 			 * while in a user context. Skip this ECB if that's
6226 			 * not the case.
6227 			 */
6228 			if ((ecb->dte_cond & DTRACE_COND_USERMODE) &&
6229 			    prov->dtpv_pops.dtps_usermode(prov->dtpv_arg,
6230 			    probe->dtpr_id, probe->dtpr_arg) == 0)
6231 				continue;
6232 
6233 #if defined(sun)
6234 			/*
6235 			 * This is more subtle than it looks. We have to be
6236 			 * absolutely certain that CRED() isn't going to
6237 			 * change out from under us so it's only legit to
6238 			 * examine that structure if we're in constrained
6239 			 * situations. Currently, the only times we'll this
6240 			 * check is if a non-super-user has enabled the
6241 			 * profile or syscall providers -- providers that
6242 			 * allow visibility of all processes. For the
6243 			 * profile case, the check above will ensure that
6244 			 * we're examining a user context.
6245 			 */
6246 			if (ecb->dte_cond & DTRACE_COND_OWNER) {
6247 				cred_t *cr;
6248 				cred_t *s_cr =
6249 				    ecb->dte_state->dts_cred.dcr_cred;
6250 				proc_t *proc;
6251 
6252 				ASSERT(s_cr != NULL);
6253 
6254 				if ((cr = CRED()) == NULL ||
6255 				    s_cr->cr_uid != cr->cr_uid ||
6256 				    s_cr->cr_uid != cr->cr_ruid ||
6257 				    s_cr->cr_uid != cr->cr_suid ||
6258 				    s_cr->cr_gid != cr->cr_gid ||
6259 				    s_cr->cr_gid != cr->cr_rgid ||
6260 				    s_cr->cr_gid != cr->cr_sgid ||
6261 				    (proc = ttoproc(curthread)) == NULL ||
6262 				    (proc->p_flag & SNOCD))
6263 					continue;
6264 			}
6265 
6266 			if (ecb->dte_cond & DTRACE_COND_ZONEOWNER) {
6267 				cred_t *cr;
6268 				cred_t *s_cr =
6269 				    ecb->dte_state->dts_cred.dcr_cred;
6270 
6271 				ASSERT(s_cr != NULL);
6272 
6273 				if ((cr = CRED()) == NULL ||
6274 				    s_cr->cr_zone->zone_id !=
6275 				    cr->cr_zone->zone_id)
6276 					continue;
6277 			}
6278 #endif
6279 		}
6280 
6281 		if (now - state->dts_alive > dtrace_deadman_timeout) {
6282 			/*
6283 			 * We seem to be dead.  Unless we (a) have kernel
6284 			 * destructive permissions (b) have explicitly enabled
6285 			 * destructive actions and (c) destructive actions have
6286 			 * not been disabled, we're going to transition into
6287 			 * the KILLED state, from which no further processing
6288 			 * on this state will be performed.
6289 			 */
6290 			if (!dtrace_priv_kernel_destructive(state) ||
6291 			    !state->dts_cred.dcr_destructive ||
6292 			    dtrace_destructive_disallow) {
6293 				void *activity = &state->dts_activity;
6294 				dtrace_activity_t current;
6295 
6296 				do {
6297 					current = state->dts_activity;
6298 				} while (dtrace_cas32(activity, current,
6299 				    DTRACE_ACTIVITY_KILLED) != current);
6300 
6301 				continue;
6302 			}
6303 		}
6304 
6305 		if ((offs = dtrace_buffer_reserve(buf, ecb->dte_needed,
6306 		    ecb->dte_alignment, state, &mstate)) < 0)
6307 			continue;
6308 
6309 		tomax = buf->dtb_tomax;
6310 		ASSERT(tomax != NULL);
6311 
6312 		if (ecb->dte_size != 0) {
6313 			dtrace_rechdr_t dtrh;
6314 			if (!(mstate.dtms_present & DTRACE_MSTATE_TIMESTAMP)) {
6315 				mstate.dtms_timestamp = dtrace_gethrtime();
6316 				mstate.dtms_present |= DTRACE_MSTATE_TIMESTAMP;
6317 			}
6318 			ASSERT3U(ecb->dte_size, >=, sizeof (dtrace_rechdr_t));
6319 			dtrh.dtrh_epid = ecb->dte_epid;
6320 			DTRACE_RECORD_STORE_TIMESTAMP(&dtrh,
6321 			    mstate.dtms_timestamp);
6322 			*((dtrace_rechdr_t *)(tomax + offs)) = dtrh;
6323 		}
6324 
6325 		mstate.dtms_epid = ecb->dte_epid;
6326 		mstate.dtms_present |= DTRACE_MSTATE_EPID;
6327 
6328 		if (state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)
6329 			mstate.dtms_access = DTRACE_ACCESS_KERNEL;
6330 		else
6331 			mstate.dtms_access = 0;
6332 
6333 		if (pred != NULL) {
6334 			dtrace_difo_t *dp = pred->dtp_difo;
6335 			int rval;
6336 
6337 			rval = dtrace_dif_emulate(dp, &mstate, vstate, state);
6338 
6339 			if (!(*flags & CPU_DTRACE_ERROR) && !rval) {
6340 				dtrace_cacheid_t cid = probe->dtpr_predcache;
6341 
6342 				if (cid != DTRACE_CACHEIDNONE && !onintr) {
6343 					/*
6344 					 * Update the predicate cache...
6345 					 */
6346 					ASSERT(cid == pred->dtp_cacheid);
6347 					curthread->t_predcache = cid;
6348 				}
6349 
6350 				continue;
6351 			}
6352 		}
6353 
6354 		for (act = ecb->dte_action; !(*flags & CPU_DTRACE_ERROR) &&
6355 		    act != NULL; act = act->dta_next) {
6356 			size_t valoffs;
6357 			dtrace_difo_t *dp;
6358 			dtrace_recdesc_t *rec = &act->dta_rec;
6359 
6360 			size = rec->dtrd_size;
6361 			valoffs = offs + rec->dtrd_offset;
6362 
6363 			if (DTRACEACT_ISAGG(act->dta_kind)) {
6364 				uint64_t v = 0xbad;
6365 				dtrace_aggregation_t *agg;
6366 
6367 				agg = (dtrace_aggregation_t *)act;
6368 
6369 				if ((dp = act->dta_difo) != NULL)
6370 					v = dtrace_dif_emulate(dp,
6371 					    &mstate, vstate, state);
6372 
6373 				if (*flags & CPU_DTRACE_ERROR)
6374 					continue;
6375 
6376 				/*
6377 				 * Note that we always pass the expression
6378 				 * value from the previous iteration of the
6379 				 * action loop.  This value will only be used
6380 				 * if there is an expression argument to the
6381 				 * aggregating action, denoted by the
6382 				 * dtag_hasarg field.
6383 				 */
6384 				dtrace_aggregate(agg, buf,
6385 				    offs, aggbuf, v, val);
6386 				continue;
6387 			}
6388 
6389 			switch (act->dta_kind) {
6390 			case DTRACEACT_STOP:
6391 				if (dtrace_priv_proc_destructive(state))
6392 					dtrace_action_stop();
6393 				continue;
6394 
6395 			case DTRACEACT_BREAKPOINT:
6396 				if (dtrace_priv_kernel_destructive(state))
6397 					dtrace_action_breakpoint(ecb);
6398 				continue;
6399 
6400 			case DTRACEACT_PANIC:
6401 				if (dtrace_priv_kernel_destructive(state))
6402 					dtrace_action_panic(ecb);
6403 				continue;
6404 
6405 			case DTRACEACT_STACK:
6406 				if (!dtrace_priv_kernel(state))
6407 					continue;
6408 
6409 				dtrace_getpcstack((pc_t *)(tomax + valoffs),
6410 				    size / sizeof (pc_t), probe->dtpr_aframes,
6411 				    DTRACE_ANCHORED(probe) ? NULL :
6412 				    (uint32_t *)arg0);
6413 				continue;
6414 
6415 			case DTRACEACT_JSTACK:
6416 			case DTRACEACT_USTACK:
6417 				if (!dtrace_priv_proc(state))
6418 					continue;
6419 
6420 				/*
6421 				 * See comment in DIF_VAR_PID.
6422 				 */
6423 				if (DTRACE_ANCHORED(mstate.dtms_probe) &&
6424 				    CPU_ON_INTR(CPU)) {
6425 					int depth = DTRACE_USTACK_NFRAMES(
6426 					    rec->dtrd_arg) + 1;
6427 
6428 					dtrace_bzero((void *)(tomax + valoffs),
6429 					    DTRACE_USTACK_STRSIZE(rec->dtrd_arg)
6430 					    + depth * sizeof (uint64_t));
6431 
6432 					continue;
6433 				}
6434 
6435 				if (DTRACE_USTACK_STRSIZE(rec->dtrd_arg) != 0 &&
6436 				    curproc->p_dtrace_helpers != NULL) {
6437 					/*
6438 					 * This is the slow path -- we have
6439 					 * allocated string space, and we're
6440 					 * getting the stack of a process that
6441 					 * has helpers.  Call into a separate
6442 					 * routine to perform this processing.
6443 					 */
6444 					dtrace_action_ustack(&mstate, state,
6445 					    (uint64_t *)(tomax + valoffs),
6446 					    rec->dtrd_arg);
6447 					continue;
6448 				}
6449 
6450 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6451 				dtrace_getupcstack((uint64_t *)
6452 				    (tomax + valoffs),
6453 				    DTRACE_USTACK_NFRAMES(rec->dtrd_arg) + 1);
6454 				DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6455 				continue;
6456 
6457 			default:
6458 				break;
6459 			}
6460 
6461 			dp = act->dta_difo;
6462 			ASSERT(dp != NULL);
6463 
6464 			val = dtrace_dif_emulate(dp, &mstate, vstate, state);
6465 
6466 			if (*flags & CPU_DTRACE_ERROR)
6467 				continue;
6468 
6469 			switch (act->dta_kind) {
6470 			case DTRACEACT_SPECULATE: {
6471 				dtrace_rechdr_t *dtrh;
6472 
6473 				ASSERT(buf == &state->dts_buffer[cpuid]);
6474 				buf = dtrace_speculation_buffer(state,
6475 				    cpuid, val);
6476 
6477 				if (buf == NULL) {
6478 					*flags |= CPU_DTRACE_DROP;
6479 					continue;
6480 				}
6481 
6482 				offs = dtrace_buffer_reserve(buf,
6483 				    ecb->dte_needed, ecb->dte_alignment,
6484 				    state, NULL);
6485 
6486 				if (offs < 0) {
6487 					*flags |= CPU_DTRACE_DROP;
6488 					continue;
6489 				}
6490 
6491 				tomax = buf->dtb_tomax;
6492 				ASSERT(tomax != NULL);
6493 
6494 				if (ecb->dte_size == 0)
6495 					continue;
6496 
6497 				ASSERT3U(ecb->dte_size, >=,
6498 				    sizeof (dtrace_rechdr_t));
6499 				dtrh = ((void *)(tomax + offs));
6500 				dtrh->dtrh_epid = ecb->dte_epid;
6501 				/*
6502 				 * When the speculation is committed, all of
6503 				 * the records in the speculative buffer will
6504 				 * have their timestamps set to the commit
6505 				 * time.  Until then, it is set to a sentinel
6506 				 * value, for debugability.
6507 				 */
6508 				DTRACE_RECORD_STORE_TIMESTAMP(dtrh, UINT64_MAX);
6509 				continue;
6510 			}
6511 
6512 			case DTRACEACT_PRINTM: {
6513 				/* The DIF returns a 'memref'. */
6514 				uintptr_t *memref = (uintptr_t *)(uintptr_t) val;
6515 
6516 				/* Get the size from the memref. */
6517 				size = memref[1];
6518 
6519 				/*
6520 				 * Check if the size exceeds the allocated
6521 				 * buffer size.
6522 				 */
6523 				if (size + sizeof(uintptr_t) > dp->dtdo_rtype.dtdt_size) {
6524 					/* Flag a drop! */
6525 					*flags |= CPU_DTRACE_DROP;
6526 					continue;
6527 				}
6528 
6529 				/* Store the size in the buffer first. */
6530 				DTRACE_STORE(uintptr_t, tomax,
6531 				    valoffs, size);
6532 
6533 				/*
6534 				 * Offset the buffer address to the start
6535 				 * of the data.
6536 				 */
6537 				valoffs += sizeof(uintptr_t);
6538 
6539 				/*
6540 				 * Reset to the memory address rather than
6541 				 * the memref array, then let the BYREF
6542 				 * code below do the work to store the
6543 				 * memory data in the buffer.
6544 				 */
6545 				val = memref[0];
6546 				break;
6547 			}
6548 
6549 			case DTRACEACT_PRINTT: {
6550 				/* The DIF returns a 'typeref'. */
6551 				uintptr_t *typeref = (uintptr_t *)(uintptr_t) val;
6552 				char c = '\0' + 1;
6553 				size_t s;
6554 
6555 				/*
6556 				 * Get the type string length and round it
6557 				 * up so that the data that follows is
6558 				 * aligned for easy access.
6559 				 */
6560 				size_t typs = strlen((char *) typeref[2]) + 1;
6561 				typs = roundup(typs,  sizeof(uintptr_t));
6562 
6563 				/*
6564 				 *Get the size from the typeref using the
6565 				 * number of elements and the type size.
6566 				 */
6567 				size = typeref[1] * typeref[3];
6568 
6569 				/*
6570 				 * Check if the size exceeds the allocated
6571 				 * buffer size.
6572 				 */
6573 				if (size + typs + 2 * sizeof(uintptr_t) > dp->dtdo_rtype.dtdt_size) {
6574 					/* Flag a drop! */
6575 					*flags |= CPU_DTRACE_DROP;
6576 
6577 				}
6578 
6579 				/* Store the size in the buffer first. */
6580 				DTRACE_STORE(uintptr_t, tomax,
6581 				    valoffs, size);
6582 				valoffs += sizeof(uintptr_t);
6583 
6584 				/* Store the type size in the buffer. */
6585 				DTRACE_STORE(uintptr_t, tomax,
6586 				    valoffs, typeref[3]);
6587 				valoffs += sizeof(uintptr_t);
6588 
6589 				val = typeref[2];
6590 
6591 				for (s = 0; s < typs; s++) {
6592 					if (c != '\0')
6593 						c = dtrace_load8(val++);
6594 
6595 					DTRACE_STORE(uint8_t, tomax,
6596 					    valoffs++, c);
6597 				}
6598 
6599 				/*
6600 				 * Reset to the memory address rather than
6601 				 * the typeref array, then let the BYREF
6602 				 * code below do the work to store the
6603 				 * memory data in the buffer.
6604 				 */
6605 				val = typeref[0];
6606 				break;
6607 			}
6608 
6609 			case DTRACEACT_CHILL:
6610 				if (dtrace_priv_kernel_destructive(state))
6611 					dtrace_action_chill(&mstate, val);
6612 				continue;
6613 
6614 			case DTRACEACT_RAISE:
6615 				if (dtrace_priv_proc_destructive(state))
6616 					dtrace_action_raise(val);
6617 				continue;
6618 
6619 			case DTRACEACT_COMMIT:
6620 				ASSERT(!committed);
6621 
6622 				/*
6623 				 * We need to commit our buffer state.
6624 				 */
6625 				if (ecb->dte_size)
6626 					buf->dtb_offset = offs + ecb->dte_size;
6627 				buf = &state->dts_buffer[cpuid];
6628 				dtrace_speculation_commit(state, cpuid, val);
6629 				committed = 1;
6630 				continue;
6631 
6632 			case DTRACEACT_DISCARD:
6633 				dtrace_speculation_discard(state, cpuid, val);
6634 				continue;
6635 
6636 			case DTRACEACT_DIFEXPR:
6637 			case DTRACEACT_LIBACT:
6638 			case DTRACEACT_PRINTF:
6639 			case DTRACEACT_PRINTA:
6640 			case DTRACEACT_SYSTEM:
6641 			case DTRACEACT_FREOPEN:
6642 			case DTRACEACT_TRACEMEM:
6643 				break;
6644 
6645 			case DTRACEACT_TRACEMEM_DYNSIZE:
6646 				tracememsize = val;
6647 				break;
6648 
6649 			case DTRACEACT_SYM:
6650 			case DTRACEACT_MOD:
6651 				if (!dtrace_priv_kernel(state))
6652 					continue;
6653 				break;
6654 
6655 			case DTRACEACT_USYM:
6656 			case DTRACEACT_UMOD:
6657 			case DTRACEACT_UADDR: {
6658 #if defined(sun)
6659 				struct pid *pid = curthread->t_procp->p_pidp;
6660 #endif
6661 
6662 				if (!dtrace_priv_proc(state))
6663 					continue;
6664 
6665 				DTRACE_STORE(uint64_t, tomax,
6666 #if defined(sun)
6667 				    valoffs, (uint64_t)pid->pid_id);
6668 #else
6669 				    valoffs, (uint64_t) curproc->p_pid);
6670 #endif
6671 				DTRACE_STORE(uint64_t, tomax,
6672 				    valoffs + sizeof (uint64_t), val);
6673 
6674 				continue;
6675 			}
6676 
6677 			case DTRACEACT_EXIT: {
6678 				/*
6679 				 * For the exit action, we are going to attempt
6680 				 * to atomically set our activity to be
6681 				 * draining.  If this fails (either because
6682 				 * another CPU has beat us to the exit action,
6683 				 * or because our current activity is something
6684 				 * other than ACTIVE or WARMUP), we will
6685 				 * continue.  This assures that the exit action
6686 				 * can be successfully recorded at most once
6687 				 * when we're in the ACTIVE state.  If we're
6688 				 * encountering the exit() action while in
6689 				 * COOLDOWN, however, we want to honor the new
6690 				 * status code.  (We know that we're the only
6691 				 * thread in COOLDOWN, so there is no race.)
6692 				 */
6693 				void *activity = &state->dts_activity;
6694 				dtrace_activity_t current = state->dts_activity;
6695 
6696 				if (current == DTRACE_ACTIVITY_COOLDOWN)
6697 					break;
6698 
6699 				if (current != DTRACE_ACTIVITY_WARMUP)
6700 					current = DTRACE_ACTIVITY_ACTIVE;
6701 
6702 				if (dtrace_cas32(activity, current,
6703 				    DTRACE_ACTIVITY_DRAINING) != current) {
6704 					*flags |= CPU_DTRACE_DROP;
6705 					continue;
6706 				}
6707 
6708 				break;
6709 			}
6710 
6711 			default:
6712 				ASSERT(0);
6713 			}
6714 
6715 			if (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF) {
6716 				uintptr_t end = valoffs + size;
6717 
6718 				if (tracememsize != 0 &&
6719 				    valoffs + tracememsize < end) {
6720 					end = valoffs + tracememsize;
6721 					tracememsize = 0;
6722 				}
6723 
6724 				if (!dtrace_vcanload((void *)(uintptr_t)val,
6725 				    &dp->dtdo_rtype, &mstate, vstate))
6726 					continue;
6727 
6728 				/*
6729 				 * If this is a string, we're going to only
6730 				 * load until we find the zero byte -- after
6731 				 * which we'll store zero bytes.
6732 				 */
6733 				if (dp->dtdo_rtype.dtdt_kind ==
6734 				    DIF_TYPE_STRING) {
6735 					char c = '\0' + 1;
6736 					int intuple = act->dta_intuple;
6737 					size_t s;
6738 
6739 					for (s = 0; s < size; s++) {
6740 						if (c != '\0')
6741 							c = dtrace_load8(val++);
6742 
6743 						DTRACE_STORE(uint8_t, tomax,
6744 						    valoffs++, c);
6745 
6746 						if (c == '\0' && intuple)
6747 							break;
6748 					}
6749 
6750 					continue;
6751 				}
6752 
6753 				while (valoffs < end) {
6754 					DTRACE_STORE(uint8_t, tomax, valoffs++,
6755 					    dtrace_load8(val++));
6756 				}
6757 
6758 				continue;
6759 			}
6760 
6761 			switch (size) {
6762 			case 0:
6763 				break;
6764 
6765 			case sizeof (uint8_t):
6766 				DTRACE_STORE(uint8_t, tomax, valoffs, val);
6767 				break;
6768 			case sizeof (uint16_t):
6769 				DTRACE_STORE(uint16_t, tomax, valoffs, val);
6770 				break;
6771 			case sizeof (uint32_t):
6772 				DTRACE_STORE(uint32_t, tomax, valoffs, val);
6773 				break;
6774 			case sizeof (uint64_t):
6775 				DTRACE_STORE(uint64_t, tomax, valoffs, val);
6776 				break;
6777 			default:
6778 				/*
6779 				 * Any other size should have been returned by
6780 				 * reference, not by value.
6781 				 */
6782 				ASSERT(0);
6783 				break;
6784 			}
6785 		}
6786 
6787 		if (*flags & CPU_DTRACE_DROP)
6788 			continue;
6789 
6790 		if (*flags & CPU_DTRACE_FAULT) {
6791 			int ndx;
6792 			dtrace_action_t *err;
6793 
6794 			buf->dtb_errors++;
6795 
6796 			if (probe->dtpr_id == dtrace_probeid_error) {
6797 				/*
6798 				 * There's nothing we can do -- we had an
6799 				 * error on the error probe.  We bump an
6800 				 * error counter to at least indicate that
6801 				 * this condition happened.
6802 				 */
6803 				dtrace_error(&state->dts_dblerrors);
6804 				continue;
6805 			}
6806 
6807 			if (vtime) {
6808 				/*
6809 				 * Before recursing on dtrace_probe(), we
6810 				 * need to explicitly clear out our start
6811 				 * time to prevent it from being accumulated
6812 				 * into t_dtrace_vtime.
6813 				 */
6814 				curthread->t_dtrace_start = 0;
6815 			}
6816 
6817 			/*
6818 			 * Iterate over the actions to figure out which action
6819 			 * we were processing when we experienced the error.
6820 			 * Note that act points _past_ the faulting action; if
6821 			 * act is ecb->dte_action, the fault was in the
6822 			 * predicate, if it's ecb->dte_action->dta_next it's
6823 			 * in action #1, and so on.
6824 			 */
6825 			for (err = ecb->dte_action, ndx = 0;
6826 			    err != act; err = err->dta_next, ndx++)
6827 				continue;
6828 
6829 			dtrace_probe_error(state, ecb->dte_epid, ndx,
6830 			    (mstate.dtms_present & DTRACE_MSTATE_FLTOFFS) ?
6831 			    mstate.dtms_fltoffs : -1, DTRACE_FLAGS2FLT(*flags),
6832 			    cpu_core[cpuid].cpuc_dtrace_illval);
6833 
6834 			continue;
6835 		}
6836 
6837 		if (!committed)
6838 			buf->dtb_offset = offs + ecb->dte_size;
6839 	}
6840 
6841 	if (vtime)
6842 		curthread->t_dtrace_start = dtrace_gethrtime();
6843 
6844 	dtrace_interrupt_enable(cookie);
6845 }
6846 
6847 /*
6848  * DTrace Probe Hashing Functions
6849  *
6850  * The functions in this section (and indeed, the functions in remaining
6851  * sections) are not _called_ from probe context.  (Any exceptions to this are
6852  * marked with a "Note:".)  Rather, they are called from elsewhere in the
6853  * DTrace framework to look-up probes in, add probes to and remove probes from
6854  * the DTrace probe hashes.  (Each probe is hashed by each element of the
6855  * probe tuple -- allowing for fast lookups, regardless of what was
6856  * specified.)
6857  */
6858 static uint_t
6859 dtrace_hash_str(const char *p)
6860 {
6861 	unsigned int g;
6862 	uint_t hval = 0;
6863 
6864 	while (*p) {
6865 		hval = (hval << 4) + *p++;
6866 		if ((g = (hval & 0xf0000000)) != 0)
6867 			hval ^= g >> 24;
6868 		hval &= ~g;
6869 	}
6870 	return (hval);
6871 }
6872 
6873 static dtrace_hash_t *
6874 dtrace_hash_create(uintptr_t stroffs, uintptr_t nextoffs, uintptr_t prevoffs)
6875 {
6876 	dtrace_hash_t *hash = kmem_zalloc(sizeof (dtrace_hash_t), KM_SLEEP);
6877 
6878 	hash->dth_stroffs = stroffs;
6879 	hash->dth_nextoffs = nextoffs;
6880 	hash->dth_prevoffs = prevoffs;
6881 
6882 	hash->dth_size = 1;
6883 	hash->dth_mask = hash->dth_size - 1;
6884 
6885 	hash->dth_tab = kmem_zalloc(hash->dth_size *
6886 	    sizeof (dtrace_hashbucket_t *), KM_SLEEP);
6887 
6888 	return (hash);
6889 }
6890 
6891 static void
6892 dtrace_hash_destroy(dtrace_hash_t *hash)
6893 {
6894 #ifdef DEBUG
6895 	int i;
6896 
6897 	for (i = 0; i < hash->dth_size; i++)
6898 		ASSERT(hash->dth_tab[i] == NULL);
6899 #endif
6900 
6901 	kmem_free(hash->dth_tab,
6902 	    hash->dth_size * sizeof (dtrace_hashbucket_t *));
6903 	kmem_free(hash, sizeof (dtrace_hash_t));
6904 }
6905 
6906 static void
6907 dtrace_hash_resize(dtrace_hash_t *hash)
6908 {
6909 	int size = hash->dth_size, i, ndx;
6910 	int new_size = hash->dth_size << 1;
6911 	int new_mask = new_size - 1;
6912 	dtrace_hashbucket_t **new_tab, *bucket, *next;
6913 
6914 	ASSERT((new_size & new_mask) == 0);
6915 
6916 	new_tab = kmem_zalloc(new_size * sizeof (void *), KM_SLEEP);
6917 
6918 	for (i = 0; i < size; i++) {
6919 		for (bucket = hash->dth_tab[i]; bucket != NULL; bucket = next) {
6920 			dtrace_probe_t *probe = bucket->dthb_chain;
6921 
6922 			ASSERT(probe != NULL);
6923 			ndx = DTRACE_HASHSTR(hash, probe) & new_mask;
6924 
6925 			next = bucket->dthb_next;
6926 			bucket->dthb_next = new_tab[ndx];
6927 			new_tab[ndx] = bucket;
6928 		}
6929 	}
6930 
6931 	kmem_free(hash->dth_tab, hash->dth_size * sizeof (void *));
6932 	hash->dth_tab = new_tab;
6933 	hash->dth_size = new_size;
6934 	hash->dth_mask = new_mask;
6935 }
6936 
6937 static void
6938 dtrace_hash_add(dtrace_hash_t *hash, dtrace_probe_t *new)
6939 {
6940 	int hashval = DTRACE_HASHSTR(hash, new);
6941 	int ndx = hashval & hash->dth_mask;
6942 	dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6943 	dtrace_probe_t **nextp, **prevp;
6944 
6945 	for (; bucket != NULL; bucket = bucket->dthb_next) {
6946 		if (DTRACE_HASHEQ(hash, bucket->dthb_chain, new))
6947 			goto add;
6948 	}
6949 
6950 	if ((hash->dth_nbuckets >> 1) > hash->dth_size) {
6951 		dtrace_hash_resize(hash);
6952 		dtrace_hash_add(hash, new);
6953 		return;
6954 	}
6955 
6956 	bucket = kmem_zalloc(sizeof (dtrace_hashbucket_t), KM_SLEEP);
6957 	bucket->dthb_next = hash->dth_tab[ndx];
6958 	hash->dth_tab[ndx] = bucket;
6959 	hash->dth_nbuckets++;
6960 
6961 add:
6962 	nextp = DTRACE_HASHNEXT(hash, new);
6963 	ASSERT(*nextp == NULL && *(DTRACE_HASHPREV(hash, new)) == NULL);
6964 	*nextp = bucket->dthb_chain;
6965 
6966 	if (bucket->dthb_chain != NULL) {
6967 		prevp = DTRACE_HASHPREV(hash, bucket->dthb_chain);
6968 		ASSERT(*prevp == NULL);
6969 		*prevp = new;
6970 	}
6971 
6972 	bucket->dthb_chain = new;
6973 	bucket->dthb_len++;
6974 }
6975 
6976 static dtrace_probe_t *
6977 dtrace_hash_lookup(dtrace_hash_t *hash, dtrace_probe_t *template)
6978 {
6979 	int hashval = DTRACE_HASHSTR(hash, template);
6980 	int ndx = hashval & hash->dth_mask;
6981 	dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6982 
6983 	for (; bucket != NULL; bucket = bucket->dthb_next) {
6984 		if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
6985 			return (bucket->dthb_chain);
6986 	}
6987 
6988 	return (NULL);
6989 }
6990 
6991 static int
6992 dtrace_hash_collisions(dtrace_hash_t *hash, dtrace_probe_t *template)
6993 {
6994 	int hashval = DTRACE_HASHSTR(hash, template);
6995 	int ndx = hashval & hash->dth_mask;
6996 	dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6997 
6998 	for (; bucket != NULL; bucket = bucket->dthb_next) {
6999 		if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
7000 			return (bucket->dthb_len);
7001 	}
7002 
7003 	return (0);
7004 }
7005 
7006 static void
7007 dtrace_hash_remove(dtrace_hash_t *hash, dtrace_probe_t *probe)
7008 {
7009 	int ndx = DTRACE_HASHSTR(hash, probe) & hash->dth_mask;
7010 	dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
7011 
7012 	dtrace_probe_t **prevp = DTRACE_HASHPREV(hash, probe);
7013 	dtrace_probe_t **nextp = DTRACE_HASHNEXT(hash, probe);
7014 
7015 	/*
7016 	 * Find the bucket that we're removing this probe from.
7017 	 */
7018 	for (; bucket != NULL; bucket = bucket->dthb_next) {
7019 		if (DTRACE_HASHEQ(hash, bucket->dthb_chain, probe))
7020 			break;
7021 	}
7022 
7023 	ASSERT(bucket != NULL);
7024 
7025 	if (*prevp == NULL) {
7026 		if (*nextp == NULL) {
7027 			/*
7028 			 * The removed probe was the only probe on this
7029 			 * bucket; we need to remove the bucket.
7030 			 */
7031 			dtrace_hashbucket_t *b = hash->dth_tab[ndx];
7032 
7033 			ASSERT(bucket->dthb_chain == probe);
7034 			ASSERT(b != NULL);
7035 
7036 			if (b == bucket) {
7037 				hash->dth_tab[ndx] = bucket->dthb_next;
7038 			} else {
7039 				while (b->dthb_next != bucket)
7040 					b = b->dthb_next;
7041 				b->dthb_next = bucket->dthb_next;
7042 			}
7043 
7044 			ASSERT(hash->dth_nbuckets > 0);
7045 			hash->dth_nbuckets--;
7046 			kmem_free(bucket, sizeof (dtrace_hashbucket_t));
7047 			return;
7048 		}
7049 
7050 		bucket->dthb_chain = *nextp;
7051 	} else {
7052 		*(DTRACE_HASHNEXT(hash, *prevp)) = *nextp;
7053 	}
7054 
7055 	if (*nextp != NULL)
7056 		*(DTRACE_HASHPREV(hash, *nextp)) = *prevp;
7057 }
7058 
7059 /*
7060  * DTrace Utility Functions
7061  *
7062  * These are random utility functions that are _not_ called from probe context.
7063  */
7064 static int
7065 dtrace_badattr(const dtrace_attribute_t *a)
7066 {
7067 	return (a->dtat_name > DTRACE_STABILITY_MAX ||
7068 	    a->dtat_data > DTRACE_STABILITY_MAX ||
7069 	    a->dtat_class > DTRACE_CLASS_MAX);
7070 }
7071 
7072 /*
7073  * Return a duplicate copy of a string.  If the specified string is NULL,
7074  * this function returns a zero-length string.
7075  */
7076 static char *
7077 dtrace_strdup(const char *str)
7078 {
7079 	char *new = kmem_zalloc((str != NULL ? strlen(str) : 0) + 1, KM_SLEEP);
7080 
7081 	if (str != NULL)
7082 		(void) strcpy(new, str);
7083 
7084 	return (new);
7085 }
7086 
7087 #define	DTRACE_ISALPHA(c)	\
7088 	(((c) >= 'a' && (c) <= 'z') || ((c) >= 'A' && (c) <= 'Z'))
7089 
7090 static int
7091 dtrace_badname(const char *s)
7092 {
7093 	char c;
7094 
7095 	if (s == NULL || (c = *s++) == '\0')
7096 		return (0);
7097 
7098 	if (!DTRACE_ISALPHA(c) && c != '-' && c != '_' && c != '.')
7099 		return (1);
7100 
7101 	while ((c = *s++) != '\0') {
7102 		if (!DTRACE_ISALPHA(c) && (c < '0' || c > '9') &&
7103 		    c != '-' && c != '_' && c != '.' && c != '`')
7104 			return (1);
7105 	}
7106 
7107 	return (0);
7108 }
7109 
7110 static void
7111 dtrace_cred2priv(cred_t *cr, uint32_t *privp, uid_t *uidp, zoneid_t *zoneidp)
7112 {
7113 	uint32_t priv;
7114 
7115 #if defined(sun)
7116 	if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
7117 		/*
7118 		 * For DTRACE_PRIV_ALL, the uid and zoneid don't matter.
7119 		 */
7120 		priv = DTRACE_PRIV_ALL;
7121 	} else {
7122 		*uidp = crgetuid(cr);
7123 		*zoneidp = crgetzoneid(cr);
7124 
7125 		priv = 0;
7126 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE))
7127 			priv |= DTRACE_PRIV_KERNEL | DTRACE_PRIV_USER;
7128 		else if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE))
7129 			priv |= DTRACE_PRIV_USER;
7130 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE))
7131 			priv |= DTRACE_PRIV_PROC;
7132 		if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
7133 			priv |= DTRACE_PRIV_OWNER;
7134 		if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
7135 			priv |= DTRACE_PRIV_ZONEOWNER;
7136 	}
7137 #else
7138 	priv = DTRACE_PRIV_ALL;
7139 #endif
7140 
7141 	*privp = priv;
7142 }
7143 
7144 #ifdef DTRACE_ERRDEBUG
7145 static void
7146 dtrace_errdebug(const char *str)
7147 {
7148 	int hval = dtrace_hash_str(str) % DTRACE_ERRHASHSZ;
7149 	int occupied = 0;
7150 
7151 	mutex_enter(&dtrace_errlock);
7152 	dtrace_errlast = str;
7153 	dtrace_errthread = curthread;
7154 
7155 	while (occupied++ < DTRACE_ERRHASHSZ) {
7156 		if (dtrace_errhash[hval].dter_msg == str) {
7157 			dtrace_errhash[hval].dter_count++;
7158 			goto out;
7159 		}
7160 
7161 		if (dtrace_errhash[hval].dter_msg != NULL) {
7162 			hval = (hval + 1) % DTRACE_ERRHASHSZ;
7163 			continue;
7164 		}
7165 
7166 		dtrace_errhash[hval].dter_msg = str;
7167 		dtrace_errhash[hval].dter_count = 1;
7168 		goto out;
7169 	}
7170 
7171 	panic("dtrace: undersized error hash");
7172 out:
7173 	mutex_exit(&dtrace_errlock);
7174 }
7175 #endif
7176 
7177 /*
7178  * DTrace Matching Functions
7179  *
7180  * These functions are used to match groups of probes, given some elements of
7181  * a probe tuple, or some globbed expressions for elements of a probe tuple.
7182  */
7183 static int
7184 dtrace_match_priv(const dtrace_probe_t *prp, uint32_t priv, uid_t uid,
7185     zoneid_t zoneid)
7186 {
7187 	if (priv != DTRACE_PRIV_ALL) {
7188 		uint32_t ppriv = prp->dtpr_provider->dtpv_priv.dtpp_flags;
7189 		uint32_t match = priv & ppriv;
7190 
7191 		/*
7192 		 * No PRIV_DTRACE_* privileges...
7193 		 */
7194 		if ((priv & (DTRACE_PRIV_PROC | DTRACE_PRIV_USER |
7195 		    DTRACE_PRIV_KERNEL)) == 0)
7196 			return (0);
7197 
7198 		/*
7199 		 * No matching bits, but there were bits to match...
7200 		 */
7201 		if (match == 0 && ppriv != 0)
7202 			return (0);
7203 
7204 		/*
7205 		 * Need to have permissions to the process, but don't...
7206 		 */
7207 		if (((ppriv & ~match) & DTRACE_PRIV_OWNER) != 0 &&
7208 		    uid != prp->dtpr_provider->dtpv_priv.dtpp_uid) {
7209 			return (0);
7210 		}
7211 
7212 		/*
7213 		 * Need to be in the same zone unless we possess the
7214 		 * privilege to examine all zones.
7215 		 */
7216 		if (((ppriv & ~match) & DTRACE_PRIV_ZONEOWNER) != 0 &&
7217 		    zoneid != prp->dtpr_provider->dtpv_priv.dtpp_zoneid) {
7218 			return (0);
7219 		}
7220 	}
7221 
7222 	return (1);
7223 }
7224 
7225 /*
7226  * dtrace_match_probe compares a dtrace_probe_t to a pre-compiled key, which
7227  * consists of input pattern strings and an ops-vector to evaluate them.
7228  * This function returns >0 for match, 0 for no match, and <0 for error.
7229  */
7230 static int
7231 dtrace_match_probe(const dtrace_probe_t *prp, const dtrace_probekey_t *pkp,
7232     uint32_t priv, uid_t uid, zoneid_t zoneid)
7233 {
7234 	dtrace_provider_t *pvp = prp->dtpr_provider;
7235 	int rv;
7236 
7237 	if (pvp->dtpv_defunct)
7238 		return (0);
7239 
7240 	if ((rv = pkp->dtpk_pmatch(pvp->dtpv_name, pkp->dtpk_prov, 0)) <= 0)
7241 		return (rv);
7242 
7243 	if ((rv = pkp->dtpk_mmatch(prp->dtpr_mod, pkp->dtpk_mod, 0)) <= 0)
7244 		return (rv);
7245 
7246 	if ((rv = pkp->dtpk_fmatch(prp->dtpr_func, pkp->dtpk_func, 0)) <= 0)
7247 		return (rv);
7248 
7249 	if ((rv = pkp->dtpk_nmatch(prp->dtpr_name, pkp->dtpk_name, 0)) <= 0)
7250 		return (rv);
7251 
7252 	if (dtrace_match_priv(prp, priv, uid, zoneid) == 0)
7253 		return (0);
7254 
7255 	return (rv);
7256 }
7257 
7258 /*
7259  * dtrace_match_glob() is a safe kernel implementation of the gmatch(3GEN)
7260  * interface for matching a glob pattern 'p' to an input string 's'.  Unlike
7261  * libc's version, the kernel version only applies to 8-bit ASCII strings.
7262  * In addition, all of the recursion cases except for '*' matching have been
7263  * unwound.  For '*', we still implement recursive evaluation, but a depth
7264  * counter is maintained and matching is aborted if we recurse too deep.
7265  * The function returns 0 if no match, >0 if match, and <0 if recursion error.
7266  */
7267 static int
7268 dtrace_match_glob(const char *s, const char *p, int depth)
7269 {
7270 	const char *olds;
7271 	char s1, c;
7272 	int gs;
7273 
7274 	if (depth > DTRACE_PROBEKEY_MAXDEPTH)
7275 		return (-1);
7276 
7277 	if (s == NULL)
7278 		s = ""; /* treat NULL as empty string */
7279 
7280 top:
7281 	olds = s;
7282 	s1 = *s++;
7283 
7284 	if (p == NULL)
7285 		return (0);
7286 
7287 	if ((c = *p++) == '\0')
7288 		return (s1 == '\0');
7289 
7290 	switch (c) {
7291 	case '[': {
7292 		int ok = 0, notflag = 0;
7293 		char lc = '\0';
7294 
7295 		if (s1 == '\0')
7296 			return (0);
7297 
7298 		if (*p == '!') {
7299 			notflag = 1;
7300 			p++;
7301 		}
7302 
7303 		if ((c = *p++) == '\0')
7304 			return (0);
7305 
7306 		do {
7307 			if (c == '-' && lc != '\0' && *p != ']') {
7308 				if ((c = *p++) == '\0')
7309 					return (0);
7310 				if (c == '\\' && (c = *p++) == '\0')
7311 					return (0);
7312 
7313 				if (notflag) {
7314 					if (s1 < lc || s1 > c)
7315 						ok++;
7316 					else
7317 						return (0);
7318 				} else if (lc <= s1 && s1 <= c)
7319 					ok++;
7320 
7321 			} else if (c == '\\' && (c = *p++) == '\0')
7322 				return (0);
7323 
7324 			lc = c; /* save left-hand 'c' for next iteration */
7325 
7326 			if (notflag) {
7327 				if (s1 != c)
7328 					ok++;
7329 				else
7330 					return (0);
7331 			} else if (s1 == c)
7332 				ok++;
7333 
7334 			if ((c = *p++) == '\0')
7335 				return (0);
7336 
7337 		} while (c != ']');
7338 
7339 		if (ok)
7340 			goto top;
7341 
7342 		return (0);
7343 	}
7344 
7345 	case '\\':
7346 		if ((c = *p++) == '\0')
7347 			return (0);
7348 		/*FALLTHRU*/
7349 
7350 	default:
7351 		if (c != s1)
7352 			return (0);
7353 		/*FALLTHRU*/
7354 
7355 	case '?':
7356 		if (s1 != '\0')
7357 			goto top;
7358 		return (0);
7359 
7360 	case '*':
7361 		while (*p == '*')
7362 			p++; /* consecutive *'s are identical to a single one */
7363 
7364 		if (*p == '\0')
7365 			return (1);
7366 
7367 		for (s = olds; *s != '\0'; s++) {
7368 			if ((gs = dtrace_match_glob(s, p, depth + 1)) != 0)
7369 				return (gs);
7370 		}
7371 
7372 		return (0);
7373 	}
7374 }
7375 
7376 /*ARGSUSED*/
7377 static int
7378 dtrace_match_string(const char *s, const char *p, int depth)
7379 {
7380 	return (s != NULL && strcmp(s, p) == 0);
7381 }
7382 
7383 /*ARGSUSED*/
7384 static int
7385 dtrace_match_nul(const char *s, const char *p, int depth)
7386 {
7387 	return (1); /* always match the empty pattern */
7388 }
7389 
7390 /*ARGSUSED*/
7391 static int
7392 dtrace_match_nonzero(const char *s, const char *p, int depth)
7393 {
7394 	return (s != NULL && s[0] != '\0');
7395 }
7396 
7397 static int
7398 dtrace_match(const dtrace_probekey_t *pkp, uint32_t priv, uid_t uid,
7399     zoneid_t zoneid, int (*matched)(dtrace_probe_t *, void *), void *arg)
7400 {
7401 	dtrace_probe_t template, *probe;
7402 	dtrace_hash_t *hash = NULL;
7403 	int len, best = INT_MAX, nmatched = 0;
7404 	dtrace_id_t i;
7405 
7406 	ASSERT(MUTEX_HELD(&dtrace_lock));
7407 
7408 	/*
7409 	 * If the probe ID is specified in the key, just lookup by ID and
7410 	 * invoke the match callback once if a matching probe is found.
7411 	 */
7412 	if (pkp->dtpk_id != DTRACE_IDNONE) {
7413 		if ((probe = dtrace_probe_lookup_id(pkp->dtpk_id)) != NULL &&
7414 		    dtrace_match_probe(probe, pkp, priv, uid, zoneid) > 0) {
7415 			(void) (*matched)(probe, arg);
7416 			nmatched++;
7417 		}
7418 		return (nmatched);
7419 	}
7420 
7421 	template.dtpr_mod = (char *)pkp->dtpk_mod;
7422 	template.dtpr_func = (char *)pkp->dtpk_func;
7423 	template.dtpr_name = (char *)pkp->dtpk_name;
7424 
7425 	/*
7426 	 * We want to find the most distinct of the module name, function
7427 	 * name, and name.  So for each one that is not a glob pattern or
7428 	 * empty string, we perform a lookup in the corresponding hash and
7429 	 * use the hash table with the fewest collisions to do our search.
7430 	 */
7431 	if (pkp->dtpk_mmatch == &dtrace_match_string &&
7432 	    (len = dtrace_hash_collisions(dtrace_bymod, &template)) < best) {
7433 		best = len;
7434 		hash = dtrace_bymod;
7435 	}
7436 
7437 	if (pkp->dtpk_fmatch == &dtrace_match_string &&
7438 	    (len = dtrace_hash_collisions(dtrace_byfunc, &template)) < best) {
7439 		best = len;
7440 		hash = dtrace_byfunc;
7441 	}
7442 
7443 	if (pkp->dtpk_nmatch == &dtrace_match_string &&
7444 	    (len = dtrace_hash_collisions(dtrace_byname, &template)) < best) {
7445 		best = len;
7446 		hash = dtrace_byname;
7447 	}
7448 
7449 	/*
7450 	 * If we did not select a hash table, iterate over every probe and
7451 	 * invoke our callback for each one that matches our input probe key.
7452 	 */
7453 	if (hash == NULL) {
7454 		for (i = 0; i < dtrace_nprobes; i++) {
7455 			if ((probe = dtrace_probes[i]) == NULL ||
7456 			    dtrace_match_probe(probe, pkp, priv, uid,
7457 			    zoneid) <= 0)
7458 				continue;
7459 
7460 			nmatched++;
7461 
7462 			if ((*matched)(probe, arg) != DTRACE_MATCH_NEXT)
7463 				break;
7464 		}
7465 
7466 		return (nmatched);
7467 	}
7468 
7469 	/*
7470 	 * If we selected a hash table, iterate over each probe of the same key
7471 	 * name and invoke the callback for every probe that matches the other
7472 	 * attributes of our input probe key.
7473 	 */
7474 	for (probe = dtrace_hash_lookup(hash, &template); probe != NULL;
7475 	    probe = *(DTRACE_HASHNEXT(hash, probe))) {
7476 
7477 		if (dtrace_match_probe(probe, pkp, priv, uid, zoneid) <= 0)
7478 			continue;
7479 
7480 		nmatched++;
7481 
7482 		if ((*matched)(probe, arg) != DTRACE_MATCH_NEXT)
7483 			break;
7484 	}
7485 
7486 	return (nmatched);
7487 }
7488 
7489 /*
7490  * Return the function pointer dtrace_probecmp() should use to compare the
7491  * specified pattern with a string.  For NULL or empty patterns, we select
7492  * dtrace_match_nul().  For glob pattern strings, we use dtrace_match_glob().
7493  * For non-empty non-glob strings, we use dtrace_match_string().
7494  */
7495 static dtrace_probekey_f *
7496 dtrace_probekey_func(const char *p)
7497 {
7498 	char c;
7499 
7500 	if (p == NULL || *p == '\0')
7501 		return (&dtrace_match_nul);
7502 
7503 	while ((c = *p++) != '\0') {
7504 		if (c == '[' || c == '?' || c == '*' || c == '\\')
7505 			return (&dtrace_match_glob);
7506 	}
7507 
7508 	return (&dtrace_match_string);
7509 }
7510 
7511 /*
7512  * Build a probe comparison key for use with dtrace_match_probe() from the
7513  * given probe description.  By convention, a null key only matches anchored
7514  * probes: if each field is the empty string, reset dtpk_fmatch to
7515  * dtrace_match_nonzero().
7516  */
7517 static void
7518 dtrace_probekey(dtrace_probedesc_t *pdp, dtrace_probekey_t *pkp)
7519 {
7520 	pkp->dtpk_prov = pdp->dtpd_provider;
7521 	pkp->dtpk_pmatch = dtrace_probekey_func(pdp->dtpd_provider);
7522 
7523 	pkp->dtpk_mod = pdp->dtpd_mod;
7524 	pkp->dtpk_mmatch = dtrace_probekey_func(pdp->dtpd_mod);
7525 
7526 	pkp->dtpk_func = pdp->dtpd_func;
7527 	pkp->dtpk_fmatch = dtrace_probekey_func(pdp->dtpd_func);
7528 
7529 	pkp->dtpk_name = pdp->dtpd_name;
7530 	pkp->dtpk_nmatch = dtrace_probekey_func(pdp->dtpd_name);
7531 
7532 	pkp->dtpk_id = pdp->dtpd_id;
7533 
7534 	if (pkp->dtpk_id == DTRACE_IDNONE &&
7535 	    pkp->dtpk_pmatch == &dtrace_match_nul &&
7536 	    pkp->dtpk_mmatch == &dtrace_match_nul &&
7537 	    pkp->dtpk_fmatch == &dtrace_match_nul &&
7538 	    pkp->dtpk_nmatch == &dtrace_match_nul)
7539 		pkp->dtpk_fmatch = &dtrace_match_nonzero;
7540 }
7541 
7542 /*
7543  * DTrace Provider-to-Framework API Functions
7544  *
7545  * These functions implement much of the Provider-to-Framework API, as
7546  * described in <sys/dtrace.h>.  The parts of the API not in this section are
7547  * the functions in the API for probe management (found below), and
7548  * dtrace_probe() itself (found above).
7549  */
7550 
7551 /*
7552  * Register the calling provider with the DTrace framework.  This should
7553  * generally be called by DTrace providers in their attach(9E) entry point.
7554  */
7555 int
7556 dtrace_register(const char *name, const dtrace_pattr_t *pap, uint32_t priv,
7557     cred_t *cr, const dtrace_pops_t *pops, void *arg, dtrace_provider_id_t *idp)
7558 {
7559 	dtrace_provider_t *provider;
7560 
7561 	if (name == NULL || pap == NULL || pops == NULL || idp == NULL) {
7562 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7563 		    "arguments", name ? name : "<NULL>");
7564 		return (EINVAL);
7565 	}
7566 
7567 	if (name[0] == '\0' || dtrace_badname(name)) {
7568 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7569 		    "provider name", name);
7570 		return (EINVAL);
7571 	}
7572 
7573 	if ((pops->dtps_provide == NULL && pops->dtps_provide_module == NULL) ||
7574 	    pops->dtps_enable == NULL || pops->dtps_disable == NULL ||
7575 	    pops->dtps_destroy == NULL ||
7576 	    ((pops->dtps_resume == NULL) != (pops->dtps_suspend == NULL))) {
7577 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7578 		    "provider ops", name);
7579 		return (EINVAL);
7580 	}
7581 
7582 	if (dtrace_badattr(&pap->dtpa_provider) ||
7583 	    dtrace_badattr(&pap->dtpa_mod) ||
7584 	    dtrace_badattr(&pap->dtpa_func) ||
7585 	    dtrace_badattr(&pap->dtpa_name) ||
7586 	    dtrace_badattr(&pap->dtpa_args)) {
7587 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7588 		    "provider attributes", name);
7589 		return (EINVAL);
7590 	}
7591 
7592 	if (priv & ~DTRACE_PRIV_ALL) {
7593 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7594 		    "privilege attributes", name);
7595 		return (EINVAL);
7596 	}
7597 
7598 	if ((priv & DTRACE_PRIV_KERNEL) &&
7599 	    (priv & (DTRACE_PRIV_USER | DTRACE_PRIV_OWNER)) &&
7600 	    pops->dtps_usermode == NULL) {
7601 		cmn_err(CE_WARN, "failed to register provider '%s': need "
7602 		    "dtps_usermode() op for given privilege attributes", name);
7603 		return (EINVAL);
7604 	}
7605 
7606 	provider = kmem_zalloc(sizeof (dtrace_provider_t), KM_SLEEP);
7607 	provider->dtpv_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
7608 	(void) strcpy(provider->dtpv_name, name);
7609 
7610 	provider->dtpv_attr = *pap;
7611 	provider->dtpv_priv.dtpp_flags = priv;
7612 	if (cr != NULL) {
7613 		provider->dtpv_priv.dtpp_uid = crgetuid(cr);
7614 		provider->dtpv_priv.dtpp_zoneid = crgetzoneid(cr);
7615 	}
7616 	provider->dtpv_pops = *pops;
7617 
7618 	if (pops->dtps_provide == NULL) {
7619 		ASSERT(pops->dtps_provide_module != NULL);
7620 		provider->dtpv_pops.dtps_provide =
7621 		    (void (*)(void *, dtrace_probedesc_t *))dtrace_nullop;
7622 	}
7623 
7624 	if (pops->dtps_provide_module == NULL) {
7625 		ASSERT(pops->dtps_provide != NULL);
7626 		provider->dtpv_pops.dtps_provide_module =
7627 		    (void (*)(void *, modctl_t *))dtrace_nullop;
7628 	}
7629 
7630 	if (pops->dtps_suspend == NULL) {
7631 		ASSERT(pops->dtps_resume == NULL);
7632 		provider->dtpv_pops.dtps_suspend =
7633 		    (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
7634 		provider->dtpv_pops.dtps_resume =
7635 		    (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
7636 	}
7637 
7638 	provider->dtpv_arg = arg;
7639 	*idp = (dtrace_provider_id_t)provider;
7640 
7641 	if (pops == &dtrace_provider_ops) {
7642 		ASSERT(MUTEX_HELD(&dtrace_provider_lock));
7643 		ASSERT(MUTEX_HELD(&dtrace_lock));
7644 		ASSERT(dtrace_anon.dta_enabling == NULL);
7645 
7646 		/*
7647 		 * We make sure that the DTrace provider is at the head of
7648 		 * the provider chain.
7649 		 */
7650 		provider->dtpv_next = dtrace_provider;
7651 		dtrace_provider = provider;
7652 		return (0);
7653 	}
7654 
7655 	mutex_enter(&dtrace_provider_lock);
7656 	mutex_enter(&dtrace_lock);
7657 
7658 	/*
7659 	 * If there is at least one provider registered, we'll add this
7660 	 * provider after the first provider.
7661 	 */
7662 	if (dtrace_provider != NULL) {
7663 		provider->dtpv_next = dtrace_provider->dtpv_next;
7664 		dtrace_provider->dtpv_next = provider;
7665 	} else {
7666 		dtrace_provider = provider;
7667 	}
7668 
7669 	if (dtrace_retained != NULL) {
7670 		dtrace_enabling_provide(provider);
7671 
7672 		/*
7673 		 * Now we need to call dtrace_enabling_matchall() -- which
7674 		 * will acquire cpu_lock and dtrace_lock.  We therefore need
7675 		 * to drop all of our locks before calling into it...
7676 		 */
7677 		mutex_exit(&dtrace_lock);
7678 		mutex_exit(&dtrace_provider_lock);
7679 		dtrace_enabling_matchall();
7680 
7681 		return (0);
7682 	}
7683 
7684 	mutex_exit(&dtrace_lock);
7685 	mutex_exit(&dtrace_provider_lock);
7686 
7687 	return (0);
7688 }
7689 
7690 /*
7691  * Unregister the specified provider from the DTrace framework.  This should
7692  * generally be called by DTrace providers in their detach(9E) entry point.
7693  */
7694 int
7695 dtrace_unregister(dtrace_provider_id_t id)
7696 {
7697 	dtrace_provider_t *old = (dtrace_provider_t *)id;
7698 	dtrace_provider_t *prev = NULL;
7699 	int i, self = 0, noreap = 0;
7700 	dtrace_probe_t *probe, *first = NULL;
7701 
7702 	if (old->dtpv_pops.dtps_enable ==
7703 	    (void (*)(void *, dtrace_id_t, void *))dtrace_nullop) {
7704 		/*
7705 		 * If DTrace itself is the provider, we're called with locks
7706 		 * already held.
7707 		 */
7708 		ASSERT(old == dtrace_provider);
7709 #if defined(sun)
7710 		ASSERT(dtrace_devi != NULL);
7711 #endif
7712 		ASSERT(MUTEX_HELD(&dtrace_provider_lock));
7713 		ASSERT(MUTEX_HELD(&dtrace_lock));
7714 		self = 1;
7715 
7716 		if (dtrace_provider->dtpv_next != NULL) {
7717 			/*
7718 			 * There's another provider here; return failure.
7719 			 */
7720 			return (EBUSY);
7721 		}
7722 	} else {
7723 		mutex_enter(&dtrace_provider_lock);
7724 #if defined(sun)
7725 		mutex_enter(&mod_lock);
7726 #endif
7727 		mutex_enter(&dtrace_lock);
7728 	}
7729 
7730 	/*
7731 	 * If anyone has /dev/dtrace open, or if there are anonymous enabled
7732 	 * probes, we refuse to let providers slither away, unless this
7733 	 * provider has already been explicitly invalidated.
7734 	 */
7735 	if (!old->dtpv_defunct &&
7736 	    (dtrace_opens || (dtrace_anon.dta_state != NULL &&
7737 	    dtrace_anon.dta_state->dts_necbs > 0))) {
7738 		if (!self) {
7739 			mutex_exit(&dtrace_lock);
7740 #if defined(sun)
7741 			mutex_exit(&mod_lock);
7742 #endif
7743 			mutex_exit(&dtrace_provider_lock);
7744 		}
7745 		return (EBUSY);
7746 	}
7747 
7748 	/*
7749 	 * Attempt to destroy the probes associated with this provider.
7750 	 */
7751 	for (i = 0; i < dtrace_nprobes; i++) {
7752 		if ((probe = dtrace_probes[i]) == NULL)
7753 			continue;
7754 
7755 		if (probe->dtpr_provider != old)
7756 			continue;
7757 
7758 		if (probe->dtpr_ecb == NULL)
7759 			continue;
7760 
7761 		/*
7762 		 * If we are trying to unregister a defunct provider, and the
7763 		 * provider was made defunct within the interval dictated by
7764 		 * dtrace_unregister_defunct_reap, we'll (asynchronously)
7765 		 * attempt to reap our enablings.  To denote that the provider
7766 		 * should reattempt to unregister itself at some point in the
7767 		 * future, we will return a differentiable error code (EAGAIN
7768 		 * instead of EBUSY) in this case.
7769 		 */
7770 		if (dtrace_gethrtime() - old->dtpv_defunct >
7771 		    dtrace_unregister_defunct_reap)
7772 			noreap = 1;
7773 
7774 		if (!self) {
7775 			mutex_exit(&dtrace_lock);
7776 #if defined(sun)
7777 			mutex_exit(&mod_lock);
7778 #endif
7779 			mutex_exit(&dtrace_provider_lock);
7780 		}
7781 
7782 		if (noreap)
7783 			return (EBUSY);
7784 
7785 		(void) taskq_dispatch(dtrace_taskq,
7786 		    (task_func_t *)dtrace_enabling_reap, NULL, TQ_SLEEP);
7787 
7788 		return (EAGAIN);
7789 	}
7790 
7791 	/*
7792 	 * All of the probes for this provider are disabled; we can safely
7793 	 * remove all of them from their hash chains and from the probe array.
7794 	 */
7795 	for (i = 0; i < dtrace_nprobes; i++) {
7796 		if ((probe = dtrace_probes[i]) == NULL)
7797 			continue;
7798 
7799 		if (probe->dtpr_provider != old)
7800 			continue;
7801 
7802 		dtrace_probes[i] = NULL;
7803 
7804 		dtrace_hash_remove(dtrace_bymod, probe);
7805 		dtrace_hash_remove(dtrace_byfunc, probe);
7806 		dtrace_hash_remove(dtrace_byname, probe);
7807 
7808 		if (first == NULL) {
7809 			first = probe;
7810 			probe->dtpr_nextmod = NULL;
7811 		} else {
7812 			probe->dtpr_nextmod = first;
7813 			first = probe;
7814 		}
7815 	}
7816 
7817 	/*
7818 	 * The provider's probes have been removed from the hash chains and
7819 	 * from the probe array.  Now issue a dtrace_sync() to be sure that
7820 	 * everyone has cleared out from any probe array processing.
7821 	 */
7822 	dtrace_sync();
7823 
7824 	for (probe = first; probe != NULL; probe = first) {
7825 		first = probe->dtpr_nextmod;
7826 
7827 		old->dtpv_pops.dtps_destroy(old->dtpv_arg, probe->dtpr_id,
7828 		    probe->dtpr_arg);
7829 		kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
7830 		kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
7831 		kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
7832 #if defined(sun)
7833 		vmem_free(dtrace_arena, (void *)(uintptr_t)(probe->dtpr_id), 1);
7834 #else
7835 		free_unr(dtrace_arena, probe->dtpr_id);
7836 #endif
7837 		kmem_free(probe, sizeof (dtrace_probe_t));
7838 	}
7839 
7840 	if ((prev = dtrace_provider) == old) {
7841 #if defined(sun)
7842 		ASSERT(self || dtrace_devi == NULL);
7843 		ASSERT(old->dtpv_next == NULL || dtrace_devi == NULL);
7844 #endif
7845 		dtrace_provider = old->dtpv_next;
7846 	} else {
7847 		while (prev != NULL && prev->dtpv_next != old)
7848 			prev = prev->dtpv_next;
7849 
7850 		if (prev == NULL) {
7851 			panic("attempt to unregister non-existent "
7852 			    "dtrace provider %p\n", (void *)id);
7853 		}
7854 
7855 		prev->dtpv_next = old->dtpv_next;
7856 	}
7857 
7858 	if (!self) {
7859 		mutex_exit(&dtrace_lock);
7860 #if defined(sun)
7861 		mutex_exit(&mod_lock);
7862 #endif
7863 		mutex_exit(&dtrace_provider_lock);
7864 	}
7865 
7866 	kmem_free(old->dtpv_name, strlen(old->dtpv_name) + 1);
7867 	kmem_free(old, sizeof (dtrace_provider_t));
7868 
7869 	return (0);
7870 }
7871 
7872 /*
7873  * Invalidate the specified provider.  All subsequent probe lookups for the
7874  * specified provider will fail, but its probes will not be removed.
7875  */
7876 void
7877 dtrace_invalidate(dtrace_provider_id_t id)
7878 {
7879 	dtrace_provider_t *pvp = (dtrace_provider_t *)id;
7880 
7881 	ASSERT(pvp->dtpv_pops.dtps_enable !=
7882 	    (void (*)(void *, dtrace_id_t, void *))dtrace_nullop);
7883 
7884 	mutex_enter(&dtrace_provider_lock);
7885 	mutex_enter(&dtrace_lock);
7886 
7887 	pvp->dtpv_defunct = dtrace_gethrtime();
7888 
7889 	mutex_exit(&dtrace_lock);
7890 	mutex_exit(&dtrace_provider_lock);
7891 }
7892 
7893 /*
7894  * Indicate whether or not DTrace has attached.
7895  */
7896 int
7897 dtrace_attached(void)
7898 {
7899 	/*
7900 	 * dtrace_provider will be non-NULL iff the DTrace driver has
7901 	 * attached.  (It's non-NULL because DTrace is always itself a
7902 	 * provider.)
7903 	 */
7904 	return (dtrace_provider != NULL);
7905 }
7906 
7907 /*
7908  * Remove all the unenabled probes for the given provider.  This function is
7909  * not unlike dtrace_unregister(), except that it doesn't remove the provider
7910  * -- just as many of its associated probes as it can.
7911  */
7912 int
7913 dtrace_condense(dtrace_provider_id_t id)
7914 {
7915 	dtrace_provider_t *prov = (dtrace_provider_t *)id;
7916 	int i;
7917 	dtrace_probe_t *probe;
7918 
7919 	/*
7920 	 * Make sure this isn't the dtrace provider itself.
7921 	 */
7922 	ASSERT(prov->dtpv_pops.dtps_enable !=
7923 	    (void (*)(void *, dtrace_id_t, void *))dtrace_nullop);
7924 
7925 	mutex_enter(&dtrace_provider_lock);
7926 	mutex_enter(&dtrace_lock);
7927 
7928 	/*
7929 	 * Attempt to destroy the probes associated with this provider.
7930 	 */
7931 	for (i = 0; i < dtrace_nprobes; i++) {
7932 		if ((probe = dtrace_probes[i]) == NULL)
7933 			continue;
7934 
7935 		if (probe->dtpr_provider != prov)
7936 			continue;
7937 
7938 		if (probe->dtpr_ecb != NULL)
7939 			continue;
7940 
7941 		dtrace_probes[i] = NULL;
7942 
7943 		dtrace_hash_remove(dtrace_bymod, probe);
7944 		dtrace_hash_remove(dtrace_byfunc, probe);
7945 		dtrace_hash_remove(dtrace_byname, probe);
7946 
7947 		prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, i + 1,
7948 		    probe->dtpr_arg);
7949 		kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
7950 		kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
7951 		kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
7952 		kmem_free(probe, sizeof (dtrace_probe_t));
7953 #if defined(sun)
7954 		vmem_free(dtrace_arena, (void *)((uintptr_t)i + 1), 1);
7955 #else
7956 		free_unr(dtrace_arena, i + 1);
7957 #endif
7958 	}
7959 
7960 	mutex_exit(&dtrace_lock);
7961 	mutex_exit(&dtrace_provider_lock);
7962 
7963 	return (0);
7964 }
7965 
7966 /*
7967  * DTrace Probe Management Functions
7968  *
7969  * The functions in this section perform the DTrace probe management,
7970  * including functions to create probes, look-up probes, and call into the
7971  * providers to request that probes be provided.  Some of these functions are
7972  * in the Provider-to-Framework API; these functions can be identified by the
7973  * fact that they are not declared "static".
7974  */
7975 
7976 /*
7977  * Create a probe with the specified module name, function name, and name.
7978  */
7979 dtrace_id_t
7980 dtrace_probe_create(dtrace_provider_id_t prov, const char *mod,
7981     const char *func, const char *name, int aframes, void *arg)
7982 {
7983 	dtrace_probe_t *probe, **probes;
7984 	dtrace_provider_t *provider = (dtrace_provider_t *)prov;
7985 	dtrace_id_t id;
7986 
7987 	if (provider == dtrace_provider) {
7988 		ASSERT(MUTEX_HELD(&dtrace_lock));
7989 	} else {
7990 		mutex_enter(&dtrace_lock);
7991 	}
7992 
7993 #if defined(sun)
7994 	id = (dtrace_id_t)(uintptr_t)vmem_alloc(dtrace_arena, 1,
7995 	    VM_BESTFIT | VM_SLEEP);
7996 #else
7997 	id = alloc_unr(dtrace_arena);
7998 #endif
7999 	probe = kmem_zalloc(sizeof (dtrace_probe_t), KM_SLEEP);
8000 
8001 	probe->dtpr_id = id;
8002 	probe->dtpr_gen = dtrace_probegen++;
8003 	probe->dtpr_mod = dtrace_strdup(mod);
8004 	probe->dtpr_func = dtrace_strdup(func);
8005 	probe->dtpr_name = dtrace_strdup(name);
8006 	probe->dtpr_arg = arg;
8007 	probe->dtpr_aframes = aframes;
8008 	probe->dtpr_provider = provider;
8009 
8010 	dtrace_hash_add(dtrace_bymod, probe);
8011 	dtrace_hash_add(dtrace_byfunc, probe);
8012 	dtrace_hash_add(dtrace_byname, probe);
8013 
8014 	if (id - 1 >= dtrace_nprobes) {
8015 		size_t osize = dtrace_nprobes * sizeof (dtrace_probe_t *);
8016 		size_t nsize = osize << 1;
8017 
8018 		if (nsize == 0) {
8019 			ASSERT(osize == 0);
8020 			ASSERT(dtrace_probes == NULL);
8021 			nsize = sizeof (dtrace_probe_t *);
8022 		}
8023 
8024 		probes = kmem_zalloc(nsize, KM_SLEEP);
8025 
8026 		if (dtrace_probes == NULL) {
8027 			ASSERT(osize == 0);
8028 			dtrace_probes = probes;
8029 			dtrace_nprobes = 1;
8030 		} else {
8031 			dtrace_probe_t **oprobes = dtrace_probes;
8032 
8033 			bcopy(oprobes, probes, osize);
8034 			dtrace_membar_producer();
8035 			dtrace_probes = probes;
8036 
8037 			dtrace_sync();
8038 
8039 			/*
8040 			 * All CPUs are now seeing the new probes array; we can
8041 			 * safely free the old array.
8042 			 */
8043 			kmem_free(oprobes, osize);
8044 			dtrace_nprobes <<= 1;
8045 		}
8046 
8047 		ASSERT(id - 1 < dtrace_nprobes);
8048 	}
8049 
8050 	ASSERT(dtrace_probes[id - 1] == NULL);
8051 	dtrace_probes[id - 1] = probe;
8052 
8053 	if (provider != dtrace_provider)
8054 		mutex_exit(&dtrace_lock);
8055 
8056 	return (id);
8057 }
8058 
8059 static dtrace_probe_t *
8060 dtrace_probe_lookup_id(dtrace_id_t id)
8061 {
8062 	ASSERT(MUTEX_HELD(&dtrace_lock));
8063 
8064 	if (id == 0 || id > dtrace_nprobes)
8065 		return (NULL);
8066 
8067 	return (dtrace_probes[id - 1]);
8068 }
8069 
8070 static int
8071 dtrace_probe_lookup_match(dtrace_probe_t *probe, void *arg)
8072 {
8073 	*((dtrace_id_t *)arg) = probe->dtpr_id;
8074 
8075 	return (DTRACE_MATCH_DONE);
8076 }
8077 
8078 /*
8079  * Look up a probe based on provider and one or more of module name, function
8080  * name and probe name.
8081  */
8082 dtrace_id_t
8083 dtrace_probe_lookup(dtrace_provider_id_t prid, char *mod,
8084     char *func, char *name)
8085 {
8086 	dtrace_probekey_t pkey;
8087 	dtrace_id_t id;
8088 	int match;
8089 
8090 	pkey.dtpk_prov = ((dtrace_provider_t *)prid)->dtpv_name;
8091 	pkey.dtpk_pmatch = &dtrace_match_string;
8092 	pkey.dtpk_mod = mod;
8093 	pkey.dtpk_mmatch = mod ? &dtrace_match_string : &dtrace_match_nul;
8094 	pkey.dtpk_func = func;
8095 	pkey.dtpk_fmatch = func ? &dtrace_match_string : &dtrace_match_nul;
8096 	pkey.dtpk_name = name;
8097 	pkey.dtpk_nmatch = name ? &dtrace_match_string : &dtrace_match_nul;
8098 	pkey.dtpk_id = DTRACE_IDNONE;
8099 
8100 	mutex_enter(&dtrace_lock);
8101 	match = dtrace_match(&pkey, DTRACE_PRIV_ALL, 0, 0,
8102 	    dtrace_probe_lookup_match, &id);
8103 	mutex_exit(&dtrace_lock);
8104 
8105 	ASSERT(match == 1 || match == 0);
8106 	return (match ? id : 0);
8107 }
8108 
8109 /*
8110  * Returns the probe argument associated with the specified probe.
8111  */
8112 void *
8113 dtrace_probe_arg(dtrace_provider_id_t id, dtrace_id_t pid)
8114 {
8115 	dtrace_probe_t *probe;
8116 	void *rval = NULL;
8117 
8118 	mutex_enter(&dtrace_lock);
8119 
8120 	if ((probe = dtrace_probe_lookup_id(pid)) != NULL &&
8121 	    probe->dtpr_provider == (dtrace_provider_t *)id)
8122 		rval = probe->dtpr_arg;
8123 
8124 	mutex_exit(&dtrace_lock);
8125 
8126 	return (rval);
8127 }
8128 
8129 /*
8130  * Copy a probe into a probe description.
8131  */
8132 static void
8133 dtrace_probe_description(const dtrace_probe_t *prp, dtrace_probedesc_t *pdp)
8134 {
8135 	bzero(pdp, sizeof (dtrace_probedesc_t));
8136 	pdp->dtpd_id = prp->dtpr_id;
8137 
8138 	(void) strncpy(pdp->dtpd_provider,
8139 	    prp->dtpr_provider->dtpv_name, DTRACE_PROVNAMELEN - 1);
8140 
8141 	(void) strncpy(pdp->dtpd_mod, prp->dtpr_mod, DTRACE_MODNAMELEN - 1);
8142 	(void) strncpy(pdp->dtpd_func, prp->dtpr_func, DTRACE_FUNCNAMELEN - 1);
8143 	(void) strncpy(pdp->dtpd_name, prp->dtpr_name, DTRACE_NAMELEN - 1);
8144 }
8145 
8146 /*
8147  * Called to indicate that a probe -- or probes -- should be provided by a
8148  * specfied provider.  If the specified description is NULL, the provider will
8149  * be told to provide all of its probes.  (This is done whenever a new
8150  * consumer comes along, or whenever a retained enabling is to be matched.) If
8151  * the specified description is non-NULL, the provider is given the
8152  * opportunity to dynamically provide the specified probe, allowing providers
8153  * to support the creation of probes on-the-fly.  (So-called _autocreated_
8154  * probes.)  If the provider is NULL, the operations will be applied to all
8155  * providers; if the provider is non-NULL the operations will only be applied
8156  * to the specified provider.  The dtrace_provider_lock must be held, and the
8157  * dtrace_lock must _not_ be held -- the provider's dtps_provide() operation
8158  * will need to grab the dtrace_lock when it reenters the framework through
8159  * dtrace_probe_lookup(), dtrace_probe_create(), etc.
8160  */
8161 static void
8162 dtrace_probe_provide(dtrace_probedesc_t *desc, dtrace_provider_t *prv)
8163 {
8164 #if defined(sun)
8165 	modctl_t *ctl;
8166 #endif
8167 	int all = 0;
8168 
8169 	ASSERT(MUTEX_HELD(&dtrace_provider_lock));
8170 
8171 	if (prv == NULL) {
8172 		all = 1;
8173 		prv = dtrace_provider;
8174 	}
8175 
8176 	do {
8177 		/*
8178 		 * First, call the blanket provide operation.
8179 		 */
8180 		prv->dtpv_pops.dtps_provide(prv->dtpv_arg, desc);
8181 
8182 #if defined(sun)
8183 		/*
8184 		 * Now call the per-module provide operation.  We will grab
8185 		 * mod_lock to prevent the list from being modified.  Note
8186 		 * that this also prevents the mod_busy bits from changing.
8187 		 * (mod_busy can only be changed with mod_lock held.)
8188 		 */
8189 		mutex_enter(&mod_lock);
8190 
8191 		ctl = &modules;
8192 		do {
8193 			if (ctl->mod_busy || ctl->mod_mp == NULL)
8194 				continue;
8195 
8196 			prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
8197 
8198 		} while ((ctl = ctl->mod_next) != &modules);
8199 
8200 		mutex_exit(&mod_lock);
8201 #endif
8202 	} while (all && (prv = prv->dtpv_next) != NULL);
8203 }
8204 
8205 #if defined(sun)
8206 /*
8207  * Iterate over each probe, and call the Framework-to-Provider API function
8208  * denoted by offs.
8209  */
8210 static void
8211 dtrace_probe_foreach(uintptr_t offs)
8212 {
8213 	dtrace_provider_t *prov;
8214 	void (*func)(void *, dtrace_id_t, void *);
8215 	dtrace_probe_t *probe;
8216 	dtrace_icookie_t cookie;
8217 	int i;
8218 
8219 	/*
8220 	 * We disable interrupts to walk through the probe array.  This is
8221 	 * safe -- the dtrace_sync() in dtrace_unregister() assures that we
8222 	 * won't see stale data.
8223 	 */
8224 	cookie = dtrace_interrupt_disable();
8225 
8226 	for (i = 0; i < dtrace_nprobes; i++) {
8227 		if ((probe = dtrace_probes[i]) == NULL)
8228 			continue;
8229 
8230 		if (probe->dtpr_ecb == NULL) {
8231 			/*
8232 			 * This probe isn't enabled -- don't call the function.
8233 			 */
8234 			continue;
8235 		}
8236 
8237 		prov = probe->dtpr_provider;
8238 		func = *((void(**)(void *, dtrace_id_t, void *))
8239 		    ((uintptr_t)&prov->dtpv_pops + offs));
8240 
8241 		func(prov->dtpv_arg, i + 1, probe->dtpr_arg);
8242 	}
8243 
8244 	dtrace_interrupt_enable(cookie);
8245 }
8246 #endif
8247 
8248 static int
8249 dtrace_probe_enable(dtrace_probedesc_t *desc, dtrace_enabling_t *enab)
8250 {
8251 	dtrace_probekey_t pkey;
8252 	uint32_t priv;
8253 	uid_t uid;
8254 	zoneid_t zoneid;
8255 
8256 	ASSERT(MUTEX_HELD(&dtrace_lock));
8257 	dtrace_ecb_create_cache = NULL;
8258 
8259 	if (desc == NULL) {
8260 		/*
8261 		 * If we're passed a NULL description, we're being asked to
8262 		 * create an ECB with a NULL probe.
8263 		 */
8264 		(void) dtrace_ecb_create_enable(NULL, enab);
8265 		return (0);
8266 	}
8267 
8268 	dtrace_probekey(desc, &pkey);
8269 	dtrace_cred2priv(enab->dten_vstate->dtvs_state->dts_cred.dcr_cred,
8270 	    &priv, &uid, &zoneid);
8271 
8272 	return (dtrace_match(&pkey, priv, uid, zoneid, dtrace_ecb_create_enable,
8273 	    enab));
8274 }
8275 
8276 /*
8277  * DTrace Helper Provider Functions
8278  */
8279 static void
8280 dtrace_dofattr2attr(dtrace_attribute_t *attr, const dof_attr_t dofattr)
8281 {
8282 	attr->dtat_name = DOF_ATTR_NAME(dofattr);
8283 	attr->dtat_data = DOF_ATTR_DATA(dofattr);
8284 	attr->dtat_class = DOF_ATTR_CLASS(dofattr);
8285 }
8286 
8287 static void
8288 dtrace_dofprov2hprov(dtrace_helper_provdesc_t *hprov,
8289     const dof_provider_t *dofprov, char *strtab)
8290 {
8291 	hprov->dthpv_provname = strtab + dofprov->dofpv_name;
8292 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_provider,
8293 	    dofprov->dofpv_provattr);
8294 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_mod,
8295 	    dofprov->dofpv_modattr);
8296 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_func,
8297 	    dofprov->dofpv_funcattr);
8298 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_name,
8299 	    dofprov->dofpv_nameattr);
8300 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_args,
8301 	    dofprov->dofpv_argsattr);
8302 }
8303 
8304 static void
8305 dtrace_helper_provide_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
8306 {
8307 	uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8308 	dof_hdr_t *dof = (dof_hdr_t *)daddr;
8309 	dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
8310 	dof_provider_t *provider;
8311 	dof_probe_t *probe;
8312 	uint32_t *off, *enoff;
8313 	uint8_t *arg;
8314 	char *strtab;
8315 	uint_t i, nprobes;
8316 	dtrace_helper_provdesc_t dhpv;
8317 	dtrace_helper_probedesc_t dhpb;
8318 	dtrace_meta_t *meta = dtrace_meta_pid;
8319 	dtrace_mops_t *mops = &meta->dtm_mops;
8320 	void *parg;
8321 
8322 	provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
8323 	str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8324 	    provider->dofpv_strtab * dof->dofh_secsize);
8325 	prb_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8326 	    provider->dofpv_probes * dof->dofh_secsize);
8327 	arg_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8328 	    provider->dofpv_prargs * dof->dofh_secsize);
8329 	off_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8330 	    provider->dofpv_proffs * dof->dofh_secsize);
8331 
8332 	strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
8333 	off = (uint32_t *)(uintptr_t)(daddr + off_sec->dofs_offset);
8334 	arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
8335 	enoff = NULL;
8336 
8337 	/*
8338 	 * See dtrace_helper_provider_validate().
8339 	 */
8340 	if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
8341 	    provider->dofpv_prenoffs != DOF_SECT_NONE) {
8342 		enoff_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8343 		    provider->dofpv_prenoffs * dof->dofh_secsize);
8344 		enoff = (uint32_t *)(uintptr_t)(daddr + enoff_sec->dofs_offset);
8345 	}
8346 
8347 	nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
8348 
8349 	/*
8350 	 * Create the provider.
8351 	 */
8352 	dtrace_dofprov2hprov(&dhpv, provider, strtab);
8353 
8354 	if ((parg = mops->dtms_provide_pid(meta->dtm_arg, &dhpv, pid)) == NULL)
8355 		return;
8356 
8357 	meta->dtm_count++;
8358 
8359 	/*
8360 	 * Create the probes.
8361 	 */
8362 	for (i = 0; i < nprobes; i++) {
8363 		probe = (dof_probe_t *)(uintptr_t)(daddr +
8364 		    prb_sec->dofs_offset + i * prb_sec->dofs_entsize);
8365 
8366 		dhpb.dthpb_mod = dhp->dofhp_mod;
8367 		dhpb.dthpb_func = strtab + probe->dofpr_func;
8368 		dhpb.dthpb_name = strtab + probe->dofpr_name;
8369 		dhpb.dthpb_base = probe->dofpr_addr;
8370 		dhpb.dthpb_offs = off + probe->dofpr_offidx;
8371 		dhpb.dthpb_noffs = probe->dofpr_noffs;
8372 		if (enoff != NULL) {
8373 			dhpb.dthpb_enoffs = enoff + probe->dofpr_enoffidx;
8374 			dhpb.dthpb_nenoffs = probe->dofpr_nenoffs;
8375 		} else {
8376 			dhpb.dthpb_enoffs = NULL;
8377 			dhpb.dthpb_nenoffs = 0;
8378 		}
8379 		dhpb.dthpb_args = arg + probe->dofpr_argidx;
8380 		dhpb.dthpb_nargc = probe->dofpr_nargc;
8381 		dhpb.dthpb_xargc = probe->dofpr_xargc;
8382 		dhpb.dthpb_ntypes = strtab + probe->dofpr_nargv;
8383 		dhpb.dthpb_xtypes = strtab + probe->dofpr_xargv;
8384 
8385 		mops->dtms_create_probe(meta->dtm_arg, parg, &dhpb);
8386 	}
8387 }
8388 
8389 static void
8390 dtrace_helper_provide(dof_helper_t *dhp, pid_t pid)
8391 {
8392 	uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8393 	dof_hdr_t *dof = (dof_hdr_t *)daddr;
8394 	int i;
8395 
8396 	ASSERT(MUTEX_HELD(&dtrace_meta_lock));
8397 
8398 	for (i = 0; i < dof->dofh_secnum; i++) {
8399 		dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
8400 		    dof->dofh_secoff + i * dof->dofh_secsize);
8401 
8402 		if (sec->dofs_type != DOF_SECT_PROVIDER)
8403 			continue;
8404 
8405 		dtrace_helper_provide_one(dhp, sec, pid);
8406 	}
8407 
8408 	/*
8409 	 * We may have just created probes, so we must now rematch against
8410 	 * any retained enablings.  Note that this call will acquire both
8411 	 * cpu_lock and dtrace_lock; the fact that we are holding
8412 	 * dtrace_meta_lock now is what defines the ordering with respect to
8413 	 * these three locks.
8414 	 */
8415 	dtrace_enabling_matchall();
8416 }
8417 
8418 static void
8419 dtrace_helper_provider_remove_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
8420 {
8421 	uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8422 	dof_hdr_t *dof = (dof_hdr_t *)daddr;
8423 	dof_sec_t *str_sec;
8424 	dof_provider_t *provider;
8425 	char *strtab;
8426 	dtrace_helper_provdesc_t dhpv;
8427 	dtrace_meta_t *meta = dtrace_meta_pid;
8428 	dtrace_mops_t *mops = &meta->dtm_mops;
8429 
8430 	provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
8431 	str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8432 	    provider->dofpv_strtab * dof->dofh_secsize);
8433 
8434 	strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
8435 
8436 	/*
8437 	 * Create the provider.
8438 	 */
8439 	dtrace_dofprov2hprov(&dhpv, provider, strtab);
8440 
8441 	mops->dtms_remove_pid(meta->dtm_arg, &dhpv, pid);
8442 
8443 	meta->dtm_count--;
8444 }
8445 
8446 static void
8447 dtrace_helper_provider_remove(dof_helper_t *dhp, pid_t pid)
8448 {
8449 	uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8450 	dof_hdr_t *dof = (dof_hdr_t *)daddr;
8451 	int i;
8452 
8453 	ASSERT(MUTEX_HELD(&dtrace_meta_lock));
8454 
8455 	for (i = 0; i < dof->dofh_secnum; i++) {
8456 		dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
8457 		    dof->dofh_secoff + i * dof->dofh_secsize);
8458 
8459 		if (sec->dofs_type != DOF_SECT_PROVIDER)
8460 			continue;
8461 
8462 		dtrace_helper_provider_remove_one(dhp, sec, pid);
8463 	}
8464 }
8465 
8466 /*
8467  * DTrace Meta Provider-to-Framework API Functions
8468  *
8469  * These functions implement the Meta Provider-to-Framework API, as described
8470  * in <sys/dtrace.h>.
8471  */
8472 int
8473 dtrace_meta_register(const char *name, const dtrace_mops_t *mops, void *arg,
8474     dtrace_meta_provider_id_t *idp)
8475 {
8476 	dtrace_meta_t *meta;
8477 	dtrace_helpers_t *help, *next;
8478 	int i;
8479 
8480 	*idp = DTRACE_METAPROVNONE;
8481 
8482 	/*
8483 	 * We strictly don't need the name, but we hold onto it for
8484 	 * debuggability. All hail error queues!
8485 	 */
8486 	if (name == NULL) {
8487 		cmn_err(CE_WARN, "failed to register meta-provider: "
8488 		    "invalid name");
8489 		return (EINVAL);
8490 	}
8491 
8492 	if (mops == NULL ||
8493 	    mops->dtms_create_probe == NULL ||
8494 	    mops->dtms_provide_pid == NULL ||
8495 	    mops->dtms_remove_pid == NULL) {
8496 		cmn_err(CE_WARN, "failed to register meta-register %s: "
8497 		    "invalid ops", name);
8498 		return (EINVAL);
8499 	}
8500 
8501 	meta = kmem_zalloc(sizeof (dtrace_meta_t), KM_SLEEP);
8502 	meta->dtm_mops = *mops;
8503 	meta->dtm_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
8504 	(void) strcpy(meta->dtm_name, name);
8505 	meta->dtm_arg = arg;
8506 
8507 	mutex_enter(&dtrace_meta_lock);
8508 	mutex_enter(&dtrace_lock);
8509 
8510 	if (dtrace_meta_pid != NULL) {
8511 		mutex_exit(&dtrace_lock);
8512 		mutex_exit(&dtrace_meta_lock);
8513 		cmn_err(CE_WARN, "failed to register meta-register %s: "
8514 		    "user-land meta-provider exists", name);
8515 		kmem_free(meta->dtm_name, strlen(meta->dtm_name) + 1);
8516 		kmem_free(meta, sizeof (dtrace_meta_t));
8517 		return (EINVAL);
8518 	}
8519 
8520 	dtrace_meta_pid = meta;
8521 	*idp = (dtrace_meta_provider_id_t)meta;
8522 
8523 	/*
8524 	 * If there are providers and probes ready to go, pass them
8525 	 * off to the new meta provider now.
8526 	 */
8527 
8528 	help = dtrace_deferred_pid;
8529 	dtrace_deferred_pid = NULL;
8530 
8531 	mutex_exit(&dtrace_lock);
8532 
8533 	while (help != NULL) {
8534 		for (i = 0; i < help->dthps_nprovs; i++) {
8535 			dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
8536 			    help->dthps_pid);
8537 		}
8538 
8539 		next = help->dthps_next;
8540 		help->dthps_next = NULL;
8541 		help->dthps_prev = NULL;
8542 		help->dthps_deferred = 0;
8543 		help = next;
8544 	}
8545 
8546 	mutex_exit(&dtrace_meta_lock);
8547 
8548 	return (0);
8549 }
8550 
8551 int
8552 dtrace_meta_unregister(dtrace_meta_provider_id_t id)
8553 {
8554 	dtrace_meta_t **pp, *old = (dtrace_meta_t *)id;
8555 
8556 	mutex_enter(&dtrace_meta_lock);
8557 	mutex_enter(&dtrace_lock);
8558 
8559 	if (old == dtrace_meta_pid) {
8560 		pp = &dtrace_meta_pid;
8561 	} else {
8562 		panic("attempt to unregister non-existent "
8563 		    "dtrace meta-provider %p\n", (void *)old);
8564 	}
8565 
8566 	if (old->dtm_count != 0) {
8567 		mutex_exit(&dtrace_lock);
8568 		mutex_exit(&dtrace_meta_lock);
8569 		return (EBUSY);
8570 	}
8571 
8572 	*pp = NULL;
8573 
8574 	mutex_exit(&dtrace_lock);
8575 	mutex_exit(&dtrace_meta_lock);
8576 
8577 	kmem_free(old->dtm_name, strlen(old->dtm_name) + 1);
8578 	kmem_free(old, sizeof (dtrace_meta_t));
8579 
8580 	return (0);
8581 }
8582 
8583 
8584 /*
8585  * DTrace DIF Object Functions
8586  */
8587 static int
8588 dtrace_difo_err(uint_t pc, const char *format, ...)
8589 {
8590 	if (dtrace_err_verbose) {
8591 		va_list alist;
8592 
8593 		(void) uprintf("dtrace DIF object error: [%u]: ", pc);
8594 		va_start(alist, format);
8595 		(void) vuprintf(format, alist);
8596 		va_end(alist);
8597 	}
8598 
8599 #ifdef DTRACE_ERRDEBUG
8600 	dtrace_errdebug(format);
8601 #endif
8602 	return (1);
8603 }
8604 
8605 /*
8606  * Validate a DTrace DIF object by checking the IR instructions.  The following
8607  * rules are currently enforced by dtrace_difo_validate():
8608  *
8609  * 1. Each instruction must have a valid opcode
8610  * 2. Each register, string, variable, or subroutine reference must be valid
8611  * 3. No instruction can modify register %r0 (must be zero)
8612  * 4. All instruction reserved bits must be set to zero
8613  * 5. The last instruction must be a "ret" instruction
8614  * 6. All branch targets must reference a valid instruction _after_ the branch
8615  */
8616 static int
8617 dtrace_difo_validate(dtrace_difo_t *dp, dtrace_vstate_t *vstate, uint_t nregs,
8618     cred_t *cr)
8619 {
8620 	int err = 0, i;
8621 	int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
8622 	int kcheckload;
8623 	uint_t pc;
8624 
8625 	kcheckload = cr == NULL ||
8626 	    (vstate->dtvs_state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) == 0;
8627 
8628 	dp->dtdo_destructive = 0;
8629 
8630 	for (pc = 0; pc < dp->dtdo_len && err == 0; pc++) {
8631 		dif_instr_t instr = dp->dtdo_buf[pc];
8632 
8633 		uint_t r1 = DIF_INSTR_R1(instr);
8634 		uint_t r2 = DIF_INSTR_R2(instr);
8635 		uint_t rd = DIF_INSTR_RD(instr);
8636 		uint_t rs = DIF_INSTR_RS(instr);
8637 		uint_t label = DIF_INSTR_LABEL(instr);
8638 		uint_t v = DIF_INSTR_VAR(instr);
8639 		uint_t subr = DIF_INSTR_SUBR(instr);
8640 		uint_t type = DIF_INSTR_TYPE(instr);
8641 		uint_t op = DIF_INSTR_OP(instr);
8642 
8643 		switch (op) {
8644 		case DIF_OP_OR:
8645 		case DIF_OP_XOR:
8646 		case DIF_OP_AND:
8647 		case DIF_OP_SLL:
8648 		case DIF_OP_SRL:
8649 		case DIF_OP_SRA:
8650 		case DIF_OP_SUB:
8651 		case DIF_OP_ADD:
8652 		case DIF_OP_MUL:
8653 		case DIF_OP_SDIV:
8654 		case DIF_OP_UDIV:
8655 		case DIF_OP_SREM:
8656 		case DIF_OP_UREM:
8657 		case DIF_OP_COPYS:
8658 			if (r1 >= nregs)
8659 				err += efunc(pc, "invalid register %u\n", r1);
8660 			if (r2 >= nregs)
8661 				err += efunc(pc, "invalid register %u\n", r2);
8662 			if (rd >= nregs)
8663 				err += efunc(pc, "invalid register %u\n", rd);
8664 			if (rd == 0)
8665 				err += efunc(pc, "cannot write to %r0\n");
8666 			break;
8667 		case DIF_OP_NOT:
8668 		case DIF_OP_MOV:
8669 		case DIF_OP_ALLOCS:
8670 			if (r1 >= nregs)
8671 				err += efunc(pc, "invalid register %u\n", r1);
8672 			if (r2 != 0)
8673 				err += efunc(pc, "non-zero reserved bits\n");
8674 			if (rd >= nregs)
8675 				err += efunc(pc, "invalid register %u\n", rd);
8676 			if (rd == 0)
8677 				err += efunc(pc, "cannot write to %r0\n");
8678 			break;
8679 		case DIF_OP_LDSB:
8680 		case DIF_OP_LDSH:
8681 		case DIF_OP_LDSW:
8682 		case DIF_OP_LDUB:
8683 		case DIF_OP_LDUH:
8684 		case DIF_OP_LDUW:
8685 		case DIF_OP_LDX:
8686 			if (r1 >= nregs)
8687 				err += efunc(pc, "invalid register %u\n", r1);
8688 			if (r2 != 0)
8689 				err += efunc(pc, "non-zero reserved bits\n");
8690 			if (rd >= nregs)
8691 				err += efunc(pc, "invalid register %u\n", rd);
8692 			if (rd == 0)
8693 				err += efunc(pc, "cannot write to %r0\n");
8694 			if (kcheckload)
8695 				dp->dtdo_buf[pc] = DIF_INSTR_LOAD(op +
8696 				    DIF_OP_RLDSB - DIF_OP_LDSB, r1, rd);
8697 			break;
8698 		case DIF_OP_RLDSB:
8699 		case DIF_OP_RLDSH:
8700 		case DIF_OP_RLDSW:
8701 		case DIF_OP_RLDUB:
8702 		case DIF_OP_RLDUH:
8703 		case DIF_OP_RLDUW:
8704 		case DIF_OP_RLDX:
8705 			if (r1 >= nregs)
8706 				err += efunc(pc, "invalid register %u\n", r1);
8707 			if (r2 != 0)
8708 				err += efunc(pc, "non-zero reserved bits\n");
8709 			if (rd >= nregs)
8710 				err += efunc(pc, "invalid register %u\n", rd);
8711 			if (rd == 0)
8712 				err += efunc(pc, "cannot write to %r0\n");
8713 			break;
8714 		case DIF_OP_ULDSB:
8715 		case DIF_OP_ULDSH:
8716 		case DIF_OP_ULDSW:
8717 		case DIF_OP_ULDUB:
8718 		case DIF_OP_ULDUH:
8719 		case DIF_OP_ULDUW:
8720 		case DIF_OP_ULDX:
8721 			if (r1 >= nregs)
8722 				err += efunc(pc, "invalid register %u\n", r1);
8723 			if (r2 != 0)
8724 				err += efunc(pc, "non-zero reserved bits\n");
8725 			if (rd >= nregs)
8726 				err += efunc(pc, "invalid register %u\n", rd);
8727 			if (rd == 0)
8728 				err += efunc(pc, "cannot write to %r0\n");
8729 			break;
8730 		case DIF_OP_STB:
8731 		case DIF_OP_STH:
8732 		case DIF_OP_STW:
8733 		case DIF_OP_STX:
8734 			if (r1 >= nregs)
8735 				err += efunc(pc, "invalid register %u\n", r1);
8736 			if (r2 != 0)
8737 				err += efunc(pc, "non-zero reserved bits\n");
8738 			if (rd >= nregs)
8739 				err += efunc(pc, "invalid register %u\n", rd);
8740 			if (rd == 0)
8741 				err += efunc(pc, "cannot write to 0 address\n");
8742 			break;
8743 		case DIF_OP_CMP:
8744 		case DIF_OP_SCMP:
8745 			if (r1 >= nregs)
8746 				err += efunc(pc, "invalid register %u\n", r1);
8747 			if (r2 >= nregs)
8748 				err += efunc(pc, "invalid register %u\n", r2);
8749 			if (rd != 0)
8750 				err += efunc(pc, "non-zero reserved bits\n");
8751 			break;
8752 		case DIF_OP_TST:
8753 			if (r1 >= nregs)
8754 				err += efunc(pc, "invalid register %u\n", r1);
8755 			if (r2 != 0 || rd != 0)
8756 				err += efunc(pc, "non-zero reserved bits\n");
8757 			break;
8758 		case DIF_OP_BA:
8759 		case DIF_OP_BE:
8760 		case DIF_OP_BNE:
8761 		case DIF_OP_BG:
8762 		case DIF_OP_BGU:
8763 		case DIF_OP_BGE:
8764 		case DIF_OP_BGEU:
8765 		case DIF_OP_BL:
8766 		case DIF_OP_BLU:
8767 		case DIF_OP_BLE:
8768 		case DIF_OP_BLEU:
8769 			if (label >= dp->dtdo_len) {
8770 				err += efunc(pc, "invalid branch target %u\n",
8771 				    label);
8772 			}
8773 			if (label <= pc) {
8774 				err += efunc(pc, "backward branch to %u\n",
8775 				    label);
8776 			}
8777 			break;
8778 		case DIF_OP_RET:
8779 			if (r1 != 0 || r2 != 0)
8780 				err += efunc(pc, "non-zero reserved bits\n");
8781 			if (rd >= nregs)
8782 				err += efunc(pc, "invalid register %u\n", rd);
8783 			break;
8784 		case DIF_OP_NOP:
8785 		case DIF_OP_POPTS:
8786 		case DIF_OP_FLUSHTS:
8787 			if (r1 != 0 || r2 != 0 || rd != 0)
8788 				err += efunc(pc, "non-zero reserved bits\n");
8789 			break;
8790 		case DIF_OP_SETX:
8791 			if (DIF_INSTR_INTEGER(instr) >= dp->dtdo_intlen) {
8792 				err += efunc(pc, "invalid integer ref %u\n",
8793 				    DIF_INSTR_INTEGER(instr));
8794 			}
8795 			if (rd >= nregs)
8796 				err += efunc(pc, "invalid register %u\n", rd);
8797 			if (rd == 0)
8798 				err += efunc(pc, "cannot write to %r0\n");
8799 			break;
8800 		case DIF_OP_SETS:
8801 			if (DIF_INSTR_STRING(instr) >= dp->dtdo_strlen) {
8802 				err += efunc(pc, "invalid string ref %u\n",
8803 				    DIF_INSTR_STRING(instr));
8804 			}
8805 			if (rd >= nregs)
8806 				err += efunc(pc, "invalid register %u\n", rd);
8807 			if (rd == 0)
8808 				err += efunc(pc, "cannot write to %r0\n");
8809 			break;
8810 		case DIF_OP_LDGA:
8811 		case DIF_OP_LDTA:
8812 			if (r1 > DIF_VAR_ARRAY_MAX)
8813 				err += efunc(pc, "invalid array %u\n", r1);
8814 			if (r2 >= nregs)
8815 				err += efunc(pc, "invalid register %u\n", r2);
8816 			if (rd >= nregs)
8817 				err += efunc(pc, "invalid register %u\n", rd);
8818 			if (rd == 0)
8819 				err += efunc(pc, "cannot write to %r0\n");
8820 			break;
8821 		case DIF_OP_LDGS:
8822 		case DIF_OP_LDTS:
8823 		case DIF_OP_LDLS:
8824 		case DIF_OP_LDGAA:
8825 		case DIF_OP_LDTAA:
8826 			if (v < DIF_VAR_OTHER_MIN || v > DIF_VAR_OTHER_MAX)
8827 				err += efunc(pc, "invalid variable %u\n", v);
8828 			if (rd >= nregs)
8829 				err += efunc(pc, "invalid register %u\n", rd);
8830 			if (rd == 0)
8831 				err += efunc(pc, "cannot write to %r0\n");
8832 			break;
8833 		case DIF_OP_STGS:
8834 		case DIF_OP_STTS:
8835 		case DIF_OP_STLS:
8836 		case DIF_OP_STGAA:
8837 		case DIF_OP_STTAA:
8838 			if (v < DIF_VAR_OTHER_UBASE || v > DIF_VAR_OTHER_MAX)
8839 				err += efunc(pc, "invalid variable %u\n", v);
8840 			if (rs >= nregs)
8841 				err += efunc(pc, "invalid register %u\n", rd);
8842 			break;
8843 		case DIF_OP_CALL:
8844 			if (subr > DIF_SUBR_MAX)
8845 				err += efunc(pc, "invalid subr %u\n", subr);
8846 			if (rd >= nregs)
8847 				err += efunc(pc, "invalid register %u\n", rd);
8848 			if (rd == 0)
8849 				err += efunc(pc, "cannot write to %r0\n");
8850 
8851 			if (subr == DIF_SUBR_COPYOUT ||
8852 			    subr == DIF_SUBR_COPYOUTSTR) {
8853 				dp->dtdo_destructive = 1;
8854 			}
8855 			break;
8856 		case DIF_OP_PUSHTR:
8857 			if (type != DIF_TYPE_STRING && type != DIF_TYPE_CTF)
8858 				err += efunc(pc, "invalid ref type %u\n", type);
8859 			if (r2 >= nregs)
8860 				err += efunc(pc, "invalid register %u\n", r2);
8861 			if (rs >= nregs)
8862 				err += efunc(pc, "invalid register %u\n", rs);
8863 			break;
8864 		case DIF_OP_PUSHTV:
8865 			if (type != DIF_TYPE_CTF)
8866 				err += efunc(pc, "invalid val type %u\n", type);
8867 			if (r2 >= nregs)
8868 				err += efunc(pc, "invalid register %u\n", r2);
8869 			if (rs >= nregs)
8870 				err += efunc(pc, "invalid register %u\n", rs);
8871 			break;
8872 		default:
8873 			err += efunc(pc, "invalid opcode %u\n",
8874 			    DIF_INSTR_OP(instr));
8875 		}
8876 	}
8877 
8878 	if (dp->dtdo_len != 0 &&
8879 	    DIF_INSTR_OP(dp->dtdo_buf[dp->dtdo_len - 1]) != DIF_OP_RET) {
8880 		err += efunc(dp->dtdo_len - 1,
8881 		    "expected 'ret' as last DIF instruction\n");
8882 	}
8883 
8884 	if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF)) {
8885 		/*
8886 		 * If we're not returning by reference, the size must be either
8887 		 * 0 or the size of one of the base types.
8888 		 */
8889 		switch (dp->dtdo_rtype.dtdt_size) {
8890 		case 0:
8891 		case sizeof (uint8_t):
8892 		case sizeof (uint16_t):
8893 		case sizeof (uint32_t):
8894 		case sizeof (uint64_t):
8895 			break;
8896 
8897 		default:
8898 			err += efunc(dp->dtdo_len - 1, "bad return size");
8899 		}
8900 	}
8901 
8902 	for (i = 0; i < dp->dtdo_varlen && err == 0; i++) {
8903 		dtrace_difv_t *v = &dp->dtdo_vartab[i], *existing = NULL;
8904 		dtrace_diftype_t *vt, *et;
8905 		uint_t id, ndx;
8906 
8907 		if (v->dtdv_scope != DIFV_SCOPE_GLOBAL &&
8908 		    v->dtdv_scope != DIFV_SCOPE_THREAD &&
8909 		    v->dtdv_scope != DIFV_SCOPE_LOCAL) {
8910 			err += efunc(i, "unrecognized variable scope %d\n",
8911 			    v->dtdv_scope);
8912 			break;
8913 		}
8914 
8915 		if (v->dtdv_kind != DIFV_KIND_ARRAY &&
8916 		    v->dtdv_kind != DIFV_KIND_SCALAR) {
8917 			err += efunc(i, "unrecognized variable type %d\n",
8918 			    v->dtdv_kind);
8919 			break;
8920 		}
8921 
8922 		if ((id = v->dtdv_id) > DIF_VARIABLE_MAX) {
8923 			err += efunc(i, "%d exceeds variable id limit\n", id);
8924 			break;
8925 		}
8926 
8927 		if (id < DIF_VAR_OTHER_UBASE)
8928 			continue;
8929 
8930 		/*
8931 		 * For user-defined variables, we need to check that this
8932 		 * definition is identical to any previous definition that we
8933 		 * encountered.
8934 		 */
8935 		ndx = id - DIF_VAR_OTHER_UBASE;
8936 
8937 		switch (v->dtdv_scope) {
8938 		case DIFV_SCOPE_GLOBAL:
8939 			if (ndx < vstate->dtvs_nglobals) {
8940 				dtrace_statvar_t *svar;
8941 
8942 				if ((svar = vstate->dtvs_globals[ndx]) != NULL)
8943 					existing = &svar->dtsv_var;
8944 			}
8945 
8946 			break;
8947 
8948 		case DIFV_SCOPE_THREAD:
8949 			if (ndx < vstate->dtvs_ntlocals)
8950 				existing = &vstate->dtvs_tlocals[ndx];
8951 			break;
8952 
8953 		case DIFV_SCOPE_LOCAL:
8954 			if (ndx < vstate->dtvs_nlocals) {
8955 				dtrace_statvar_t *svar;
8956 
8957 				if ((svar = vstate->dtvs_locals[ndx]) != NULL)
8958 					existing = &svar->dtsv_var;
8959 			}
8960 
8961 			break;
8962 		}
8963 
8964 		vt = &v->dtdv_type;
8965 
8966 		if (vt->dtdt_flags & DIF_TF_BYREF) {
8967 			if (vt->dtdt_size == 0) {
8968 				err += efunc(i, "zero-sized variable\n");
8969 				break;
8970 			}
8971 
8972 			if (v->dtdv_scope == DIFV_SCOPE_GLOBAL &&
8973 			    vt->dtdt_size > dtrace_global_maxsize) {
8974 				err += efunc(i, "oversized by-ref global\n");
8975 				break;
8976 			}
8977 		}
8978 
8979 		if (existing == NULL || existing->dtdv_id == 0)
8980 			continue;
8981 
8982 		ASSERT(existing->dtdv_id == v->dtdv_id);
8983 		ASSERT(existing->dtdv_scope == v->dtdv_scope);
8984 
8985 		if (existing->dtdv_kind != v->dtdv_kind)
8986 			err += efunc(i, "%d changed variable kind\n", id);
8987 
8988 		et = &existing->dtdv_type;
8989 
8990 		if (vt->dtdt_flags != et->dtdt_flags) {
8991 			err += efunc(i, "%d changed variable type flags\n", id);
8992 			break;
8993 		}
8994 
8995 		if (vt->dtdt_size != 0 && vt->dtdt_size != et->dtdt_size) {
8996 			err += efunc(i, "%d changed variable type size\n", id);
8997 			break;
8998 		}
8999 	}
9000 
9001 	return (err);
9002 }
9003 
9004 /*
9005  * Validate a DTrace DIF object that it is to be used as a helper.  Helpers
9006  * are much more constrained than normal DIFOs.  Specifically, they may
9007  * not:
9008  *
9009  * 1. Make calls to subroutines other than copyin(), copyinstr() or
9010  *    miscellaneous string routines
9011  * 2. Access DTrace variables other than the args[] array, and the
9012  *    curthread, pid, ppid, tid, execname, zonename, uid and gid variables.
9013  * 3. Have thread-local variables.
9014  * 4. Have dynamic variables.
9015  */
9016 static int
9017 dtrace_difo_validate_helper(dtrace_difo_t *dp)
9018 {
9019 	int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
9020 	int err = 0;
9021 	uint_t pc;
9022 
9023 	for (pc = 0; pc < dp->dtdo_len; pc++) {
9024 		dif_instr_t instr = dp->dtdo_buf[pc];
9025 
9026 		uint_t v = DIF_INSTR_VAR(instr);
9027 		uint_t subr = DIF_INSTR_SUBR(instr);
9028 		uint_t op = DIF_INSTR_OP(instr);
9029 
9030 		switch (op) {
9031 		case DIF_OP_OR:
9032 		case DIF_OP_XOR:
9033 		case DIF_OP_AND:
9034 		case DIF_OP_SLL:
9035 		case DIF_OP_SRL:
9036 		case DIF_OP_SRA:
9037 		case DIF_OP_SUB:
9038 		case DIF_OP_ADD:
9039 		case DIF_OP_MUL:
9040 		case DIF_OP_SDIV:
9041 		case DIF_OP_UDIV:
9042 		case DIF_OP_SREM:
9043 		case DIF_OP_UREM:
9044 		case DIF_OP_COPYS:
9045 		case DIF_OP_NOT:
9046 		case DIF_OP_MOV:
9047 		case DIF_OP_RLDSB:
9048 		case DIF_OP_RLDSH:
9049 		case DIF_OP_RLDSW:
9050 		case DIF_OP_RLDUB:
9051 		case DIF_OP_RLDUH:
9052 		case DIF_OP_RLDUW:
9053 		case DIF_OP_RLDX:
9054 		case DIF_OP_ULDSB:
9055 		case DIF_OP_ULDSH:
9056 		case DIF_OP_ULDSW:
9057 		case DIF_OP_ULDUB:
9058 		case DIF_OP_ULDUH:
9059 		case DIF_OP_ULDUW:
9060 		case DIF_OP_ULDX:
9061 		case DIF_OP_STB:
9062 		case DIF_OP_STH:
9063 		case DIF_OP_STW:
9064 		case DIF_OP_STX:
9065 		case DIF_OP_ALLOCS:
9066 		case DIF_OP_CMP:
9067 		case DIF_OP_SCMP:
9068 		case DIF_OP_TST:
9069 		case DIF_OP_BA:
9070 		case DIF_OP_BE:
9071 		case DIF_OP_BNE:
9072 		case DIF_OP_BG:
9073 		case DIF_OP_BGU:
9074 		case DIF_OP_BGE:
9075 		case DIF_OP_BGEU:
9076 		case DIF_OP_BL:
9077 		case DIF_OP_BLU:
9078 		case DIF_OP_BLE:
9079 		case DIF_OP_BLEU:
9080 		case DIF_OP_RET:
9081 		case DIF_OP_NOP:
9082 		case DIF_OP_POPTS:
9083 		case DIF_OP_FLUSHTS:
9084 		case DIF_OP_SETX:
9085 		case DIF_OP_SETS:
9086 		case DIF_OP_LDGA:
9087 		case DIF_OP_LDLS:
9088 		case DIF_OP_STGS:
9089 		case DIF_OP_STLS:
9090 		case DIF_OP_PUSHTR:
9091 		case DIF_OP_PUSHTV:
9092 			break;
9093 
9094 		case DIF_OP_LDGS:
9095 			if (v >= DIF_VAR_OTHER_UBASE)
9096 				break;
9097 
9098 			if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9)
9099 				break;
9100 
9101 			if (v == DIF_VAR_CURTHREAD || v == DIF_VAR_PID ||
9102 			    v == DIF_VAR_PPID || v == DIF_VAR_TID ||
9103 			    v == DIF_VAR_EXECARGS ||
9104 			    v == DIF_VAR_EXECNAME || v == DIF_VAR_ZONENAME ||
9105 			    v == DIF_VAR_UID || v == DIF_VAR_GID)
9106 				break;
9107 
9108 			err += efunc(pc, "illegal variable %u\n", v);
9109 			break;
9110 
9111 		case DIF_OP_LDTA:
9112 		case DIF_OP_LDTS:
9113 		case DIF_OP_LDGAA:
9114 		case DIF_OP_LDTAA:
9115 			err += efunc(pc, "illegal dynamic variable load\n");
9116 			break;
9117 
9118 		case DIF_OP_STTS:
9119 		case DIF_OP_STGAA:
9120 		case DIF_OP_STTAA:
9121 			err += efunc(pc, "illegal dynamic variable store\n");
9122 			break;
9123 
9124 		case DIF_OP_CALL:
9125 			if (subr == DIF_SUBR_ALLOCA ||
9126 			    subr == DIF_SUBR_BCOPY ||
9127 			    subr == DIF_SUBR_COPYIN ||
9128 			    subr == DIF_SUBR_COPYINTO ||
9129 			    subr == DIF_SUBR_COPYINSTR ||
9130 			    subr == DIF_SUBR_INDEX ||
9131 			    subr == DIF_SUBR_INET_NTOA ||
9132 			    subr == DIF_SUBR_INET_NTOA6 ||
9133 			    subr == DIF_SUBR_INET_NTOP ||
9134 			    subr == DIF_SUBR_LLTOSTR ||
9135 			    subr == DIF_SUBR_RINDEX ||
9136 			    subr == DIF_SUBR_STRCHR ||
9137 			    subr == DIF_SUBR_STRJOIN ||
9138 			    subr == DIF_SUBR_STRRCHR ||
9139 			    subr == DIF_SUBR_STRSTR ||
9140 			    subr == DIF_SUBR_HTONS ||
9141 			    subr == DIF_SUBR_HTONL ||
9142 			    subr == DIF_SUBR_HTONLL ||
9143 			    subr == DIF_SUBR_NTOHS ||
9144 			    subr == DIF_SUBR_NTOHL ||
9145 			    subr == DIF_SUBR_NTOHLL ||
9146 			    subr == DIF_SUBR_MEMREF ||
9147 #if !defined(sun)
9148 			    subr == DIF_SUBR_MEMSTR ||
9149 #endif
9150 			    subr == DIF_SUBR_TYPEREF)
9151 				break;
9152 
9153 			err += efunc(pc, "invalid subr %u\n", subr);
9154 			break;
9155 
9156 		default:
9157 			err += efunc(pc, "invalid opcode %u\n",
9158 			    DIF_INSTR_OP(instr));
9159 		}
9160 	}
9161 
9162 	return (err);
9163 }
9164 
9165 /*
9166  * Returns 1 if the expression in the DIF object can be cached on a per-thread
9167  * basis; 0 if not.
9168  */
9169 static int
9170 dtrace_difo_cacheable(dtrace_difo_t *dp)
9171 {
9172 	int i;
9173 
9174 	if (dp == NULL)
9175 		return (0);
9176 
9177 	for (i = 0; i < dp->dtdo_varlen; i++) {
9178 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
9179 
9180 		if (v->dtdv_scope != DIFV_SCOPE_GLOBAL)
9181 			continue;
9182 
9183 		switch (v->dtdv_id) {
9184 		case DIF_VAR_CURTHREAD:
9185 		case DIF_VAR_PID:
9186 		case DIF_VAR_TID:
9187 		case DIF_VAR_EXECARGS:
9188 		case DIF_VAR_EXECNAME:
9189 		case DIF_VAR_ZONENAME:
9190 			break;
9191 
9192 		default:
9193 			return (0);
9194 		}
9195 	}
9196 
9197 	/*
9198 	 * This DIF object may be cacheable.  Now we need to look for any
9199 	 * array loading instructions, any memory loading instructions, or
9200 	 * any stores to thread-local variables.
9201 	 */
9202 	for (i = 0; i < dp->dtdo_len; i++) {
9203 		uint_t op = DIF_INSTR_OP(dp->dtdo_buf[i]);
9204 
9205 		if ((op >= DIF_OP_LDSB && op <= DIF_OP_LDX) ||
9206 		    (op >= DIF_OP_ULDSB && op <= DIF_OP_ULDX) ||
9207 		    (op >= DIF_OP_RLDSB && op <= DIF_OP_RLDX) ||
9208 		    op == DIF_OP_LDGA || op == DIF_OP_STTS)
9209 			return (0);
9210 	}
9211 
9212 	return (1);
9213 }
9214 
9215 static void
9216 dtrace_difo_hold(dtrace_difo_t *dp)
9217 {
9218 	int i;
9219 
9220 	ASSERT(MUTEX_HELD(&dtrace_lock));
9221 
9222 	dp->dtdo_refcnt++;
9223 	ASSERT(dp->dtdo_refcnt != 0);
9224 
9225 	/*
9226 	 * We need to check this DIF object for references to the variable
9227 	 * DIF_VAR_VTIMESTAMP.
9228 	 */
9229 	for (i = 0; i < dp->dtdo_varlen; i++) {
9230 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
9231 
9232 		if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
9233 			continue;
9234 
9235 		if (dtrace_vtime_references++ == 0)
9236 			dtrace_vtime_enable();
9237 	}
9238 }
9239 
9240 /*
9241  * This routine calculates the dynamic variable chunksize for a given DIF
9242  * object.  The calculation is not fool-proof, and can probably be tricked by
9243  * malicious DIF -- but it works for all compiler-generated DIF.  Because this
9244  * calculation is likely imperfect, dtrace_dynvar() is able to gracefully fail
9245  * if a dynamic variable size exceeds the chunksize.
9246  */
9247 static void
9248 dtrace_difo_chunksize(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9249 {
9250 	uint64_t sval = 0;
9251 	dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
9252 	const dif_instr_t *text = dp->dtdo_buf;
9253 	uint_t pc, srd = 0;
9254 	uint_t ttop = 0;
9255 	size_t size, ksize;
9256 	uint_t id, i;
9257 
9258 	for (pc = 0; pc < dp->dtdo_len; pc++) {
9259 		dif_instr_t instr = text[pc];
9260 		uint_t op = DIF_INSTR_OP(instr);
9261 		uint_t rd = DIF_INSTR_RD(instr);
9262 		uint_t r1 = DIF_INSTR_R1(instr);
9263 		uint_t nkeys = 0;
9264 		uchar_t scope = 0;
9265 
9266 		dtrace_key_t *key = tupregs;
9267 
9268 		switch (op) {
9269 		case DIF_OP_SETX:
9270 			sval = dp->dtdo_inttab[DIF_INSTR_INTEGER(instr)];
9271 			srd = rd;
9272 			continue;
9273 
9274 		case DIF_OP_STTS:
9275 			key = &tupregs[DIF_DTR_NREGS];
9276 			key[0].dttk_size = 0;
9277 			key[1].dttk_size = 0;
9278 			nkeys = 2;
9279 			scope = DIFV_SCOPE_THREAD;
9280 			break;
9281 
9282 		case DIF_OP_STGAA:
9283 		case DIF_OP_STTAA:
9284 			nkeys = ttop;
9285 
9286 			if (DIF_INSTR_OP(instr) == DIF_OP_STTAA)
9287 				key[nkeys++].dttk_size = 0;
9288 
9289 			key[nkeys++].dttk_size = 0;
9290 
9291 			if (op == DIF_OP_STTAA) {
9292 				scope = DIFV_SCOPE_THREAD;
9293 			} else {
9294 				scope = DIFV_SCOPE_GLOBAL;
9295 			}
9296 
9297 			break;
9298 
9299 		case DIF_OP_PUSHTR:
9300 			if (ttop == DIF_DTR_NREGS)
9301 				return;
9302 
9303 			if ((srd == 0 || sval == 0) && r1 == DIF_TYPE_STRING) {
9304 				/*
9305 				 * If the register for the size of the "pushtr"
9306 				 * is %r0 (or the value is 0) and the type is
9307 				 * a string, we'll use the system-wide default
9308 				 * string size.
9309 				 */
9310 				tupregs[ttop++].dttk_size =
9311 				    dtrace_strsize_default;
9312 			} else {
9313 				if (srd == 0)
9314 					return;
9315 
9316 				tupregs[ttop++].dttk_size = sval;
9317 			}
9318 
9319 			break;
9320 
9321 		case DIF_OP_PUSHTV:
9322 			if (ttop == DIF_DTR_NREGS)
9323 				return;
9324 
9325 			tupregs[ttop++].dttk_size = 0;
9326 			break;
9327 
9328 		case DIF_OP_FLUSHTS:
9329 			ttop = 0;
9330 			break;
9331 
9332 		case DIF_OP_POPTS:
9333 			if (ttop != 0)
9334 				ttop--;
9335 			break;
9336 		}
9337 
9338 		sval = 0;
9339 		srd = 0;
9340 
9341 		if (nkeys == 0)
9342 			continue;
9343 
9344 		/*
9345 		 * We have a dynamic variable allocation; calculate its size.
9346 		 */
9347 		for (ksize = 0, i = 0; i < nkeys; i++)
9348 			ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
9349 
9350 		size = sizeof (dtrace_dynvar_t);
9351 		size += sizeof (dtrace_key_t) * (nkeys - 1);
9352 		size += ksize;
9353 
9354 		/*
9355 		 * Now we need to determine the size of the stored data.
9356 		 */
9357 		id = DIF_INSTR_VAR(instr);
9358 
9359 		for (i = 0; i < dp->dtdo_varlen; i++) {
9360 			dtrace_difv_t *v = &dp->dtdo_vartab[i];
9361 
9362 			if (v->dtdv_id == id && v->dtdv_scope == scope) {
9363 				size += v->dtdv_type.dtdt_size;
9364 				break;
9365 			}
9366 		}
9367 
9368 		if (i == dp->dtdo_varlen)
9369 			return;
9370 
9371 		/*
9372 		 * We have the size.  If this is larger than the chunk size
9373 		 * for our dynamic variable state, reset the chunk size.
9374 		 */
9375 		size = P2ROUNDUP(size, sizeof (uint64_t));
9376 
9377 		if (size > vstate->dtvs_dynvars.dtds_chunksize)
9378 			vstate->dtvs_dynvars.dtds_chunksize = size;
9379 	}
9380 }
9381 
9382 static void
9383 dtrace_difo_init(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9384 {
9385 	int i, oldsvars, osz, nsz, otlocals, ntlocals;
9386 	uint_t id;
9387 
9388 	ASSERT(MUTEX_HELD(&dtrace_lock));
9389 	ASSERT(dp->dtdo_buf != NULL && dp->dtdo_len != 0);
9390 
9391 	for (i = 0; i < dp->dtdo_varlen; i++) {
9392 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
9393 		dtrace_statvar_t *svar, ***svarp = NULL;
9394 		size_t dsize = 0;
9395 		uint8_t scope = v->dtdv_scope;
9396 		int *np = NULL;
9397 
9398 		if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
9399 			continue;
9400 
9401 		id -= DIF_VAR_OTHER_UBASE;
9402 
9403 		switch (scope) {
9404 		case DIFV_SCOPE_THREAD:
9405 			while (id >= (otlocals = vstate->dtvs_ntlocals)) {
9406 				dtrace_difv_t *tlocals;
9407 
9408 				if ((ntlocals = (otlocals << 1)) == 0)
9409 					ntlocals = 1;
9410 
9411 				osz = otlocals * sizeof (dtrace_difv_t);
9412 				nsz = ntlocals * sizeof (dtrace_difv_t);
9413 
9414 				tlocals = kmem_zalloc(nsz, KM_SLEEP);
9415 
9416 				if (osz != 0) {
9417 					bcopy(vstate->dtvs_tlocals,
9418 					    tlocals, osz);
9419 					kmem_free(vstate->dtvs_tlocals, osz);
9420 				}
9421 
9422 				vstate->dtvs_tlocals = tlocals;
9423 				vstate->dtvs_ntlocals = ntlocals;
9424 			}
9425 
9426 			vstate->dtvs_tlocals[id] = *v;
9427 			continue;
9428 
9429 		case DIFV_SCOPE_LOCAL:
9430 			np = &vstate->dtvs_nlocals;
9431 			svarp = &vstate->dtvs_locals;
9432 
9433 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
9434 				dsize = NCPU * (v->dtdv_type.dtdt_size +
9435 				    sizeof (uint64_t));
9436 			else
9437 				dsize = NCPU * sizeof (uint64_t);
9438 
9439 			break;
9440 
9441 		case DIFV_SCOPE_GLOBAL:
9442 			np = &vstate->dtvs_nglobals;
9443 			svarp = &vstate->dtvs_globals;
9444 
9445 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
9446 				dsize = v->dtdv_type.dtdt_size +
9447 				    sizeof (uint64_t);
9448 
9449 			break;
9450 
9451 		default:
9452 			ASSERT(0);
9453 		}
9454 
9455 		while (id >= (oldsvars = *np)) {
9456 			dtrace_statvar_t **statics;
9457 			int newsvars, oldsize, newsize;
9458 
9459 			if ((newsvars = (oldsvars << 1)) == 0)
9460 				newsvars = 1;
9461 
9462 			oldsize = oldsvars * sizeof (dtrace_statvar_t *);
9463 			newsize = newsvars * sizeof (dtrace_statvar_t *);
9464 
9465 			statics = kmem_zalloc(newsize, KM_SLEEP);
9466 
9467 			if (oldsize != 0) {
9468 				bcopy(*svarp, statics, oldsize);
9469 				kmem_free(*svarp, oldsize);
9470 			}
9471 
9472 			*svarp = statics;
9473 			*np = newsvars;
9474 		}
9475 
9476 		if ((svar = (*svarp)[id]) == NULL) {
9477 			svar = kmem_zalloc(sizeof (dtrace_statvar_t), KM_SLEEP);
9478 			svar->dtsv_var = *v;
9479 
9480 			if ((svar->dtsv_size = dsize) != 0) {
9481 				svar->dtsv_data = (uint64_t)(uintptr_t)
9482 				    kmem_zalloc(dsize, KM_SLEEP);
9483 			}
9484 
9485 			(*svarp)[id] = svar;
9486 		}
9487 
9488 		svar->dtsv_refcnt++;
9489 	}
9490 
9491 	dtrace_difo_chunksize(dp, vstate);
9492 	dtrace_difo_hold(dp);
9493 }
9494 
9495 static dtrace_difo_t *
9496 dtrace_difo_duplicate(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9497 {
9498 	dtrace_difo_t *new;
9499 	size_t sz;
9500 
9501 	ASSERT(dp->dtdo_buf != NULL);
9502 	ASSERT(dp->dtdo_refcnt != 0);
9503 
9504 	new = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
9505 
9506 	ASSERT(dp->dtdo_buf != NULL);
9507 	sz = dp->dtdo_len * sizeof (dif_instr_t);
9508 	new->dtdo_buf = kmem_alloc(sz, KM_SLEEP);
9509 	bcopy(dp->dtdo_buf, new->dtdo_buf, sz);
9510 	new->dtdo_len = dp->dtdo_len;
9511 
9512 	if (dp->dtdo_strtab != NULL) {
9513 		ASSERT(dp->dtdo_strlen != 0);
9514 		new->dtdo_strtab = kmem_alloc(dp->dtdo_strlen, KM_SLEEP);
9515 		bcopy(dp->dtdo_strtab, new->dtdo_strtab, dp->dtdo_strlen);
9516 		new->dtdo_strlen = dp->dtdo_strlen;
9517 	}
9518 
9519 	if (dp->dtdo_inttab != NULL) {
9520 		ASSERT(dp->dtdo_intlen != 0);
9521 		sz = dp->dtdo_intlen * sizeof (uint64_t);
9522 		new->dtdo_inttab = kmem_alloc(sz, KM_SLEEP);
9523 		bcopy(dp->dtdo_inttab, new->dtdo_inttab, sz);
9524 		new->dtdo_intlen = dp->dtdo_intlen;
9525 	}
9526 
9527 	if (dp->dtdo_vartab != NULL) {
9528 		ASSERT(dp->dtdo_varlen != 0);
9529 		sz = dp->dtdo_varlen * sizeof (dtrace_difv_t);
9530 		new->dtdo_vartab = kmem_alloc(sz, KM_SLEEP);
9531 		bcopy(dp->dtdo_vartab, new->dtdo_vartab, sz);
9532 		new->dtdo_varlen = dp->dtdo_varlen;
9533 	}
9534 
9535 	dtrace_difo_init(new, vstate);
9536 	return (new);
9537 }
9538 
9539 static void
9540 dtrace_difo_destroy(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9541 {
9542 	int i;
9543 
9544 	ASSERT(dp->dtdo_refcnt == 0);
9545 
9546 	for (i = 0; i < dp->dtdo_varlen; i++) {
9547 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
9548 		dtrace_statvar_t *svar, **svarp = NULL;
9549 		uint_t id;
9550 		uint8_t scope = v->dtdv_scope;
9551 		int *np = NULL;
9552 
9553 		switch (scope) {
9554 		case DIFV_SCOPE_THREAD:
9555 			continue;
9556 
9557 		case DIFV_SCOPE_LOCAL:
9558 			np = &vstate->dtvs_nlocals;
9559 			svarp = vstate->dtvs_locals;
9560 			break;
9561 
9562 		case DIFV_SCOPE_GLOBAL:
9563 			np = &vstate->dtvs_nglobals;
9564 			svarp = vstate->dtvs_globals;
9565 			break;
9566 
9567 		default:
9568 			ASSERT(0);
9569 		}
9570 
9571 		if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
9572 			continue;
9573 
9574 		id -= DIF_VAR_OTHER_UBASE;
9575 		ASSERT(id < *np);
9576 
9577 		svar = svarp[id];
9578 		ASSERT(svar != NULL);
9579 		ASSERT(svar->dtsv_refcnt > 0);
9580 
9581 		if (--svar->dtsv_refcnt > 0)
9582 			continue;
9583 
9584 		if (svar->dtsv_size != 0) {
9585 			ASSERT(svar->dtsv_data != 0);
9586 			kmem_free((void *)(uintptr_t)svar->dtsv_data,
9587 			    svar->dtsv_size);
9588 		}
9589 
9590 		kmem_free(svar, sizeof (dtrace_statvar_t));
9591 		svarp[id] = NULL;
9592 	}
9593 
9594 	if (dp->dtdo_buf != NULL)
9595 		kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
9596 	if (dp->dtdo_inttab != NULL)
9597 		kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
9598 	if (dp->dtdo_strtab != NULL)
9599 		kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
9600 	if (dp->dtdo_vartab != NULL)
9601 		kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
9602 
9603 	kmem_free(dp, sizeof (dtrace_difo_t));
9604 }
9605 
9606 static void
9607 dtrace_difo_release(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9608 {
9609 	int i;
9610 
9611 	ASSERT(MUTEX_HELD(&dtrace_lock));
9612 	ASSERT(dp->dtdo_refcnt != 0);
9613 
9614 	for (i = 0; i < dp->dtdo_varlen; i++) {
9615 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
9616 
9617 		if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
9618 			continue;
9619 
9620 		ASSERT(dtrace_vtime_references > 0);
9621 		if (--dtrace_vtime_references == 0)
9622 			dtrace_vtime_disable();
9623 	}
9624 
9625 	if (--dp->dtdo_refcnt == 0)
9626 		dtrace_difo_destroy(dp, vstate);
9627 }
9628 
9629 /*
9630  * DTrace Format Functions
9631  */
9632 static uint16_t
9633 dtrace_format_add(dtrace_state_t *state, char *str)
9634 {
9635 	char *fmt, **new;
9636 	uint16_t ndx, len = strlen(str) + 1;
9637 
9638 	fmt = kmem_zalloc(len, KM_SLEEP);
9639 	bcopy(str, fmt, len);
9640 
9641 	for (ndx = 0; ndx < state->dts_nformats; ndx++) {
9642 		if (state->dts_formats[ndx] == NULL) {
9643 			state->dts_formats[ndx] = fmt;
9644 			return (ndx + 1);
9645 		}
9646 	}
9647 
9648 	if (state->dts_nformats == USHRT_MAX) {
9649 		/*
9650 		 * This is only likely if a denial-of-service attack is being
9651 		 * attempted.  As such, it's okay to fail silently here.
9652 		 */
9653 		kmem_free(fmt, len);
9654 		return (0);
9655 	}
9656 
9657 	/*
9658 	 * For simplicity, we always resize the formats array to be exactly the
9659 	 * number of formats.
9660 	 */
9661 	ndx = state->dts_nformats++;
9662 	new = kmem_alloc((ndx + 1) * sizeof (char *), KM_SLEEP);
9663 
9664 	if (state->dts_formats != NULL) {
9665 		ASSERT(ndx != 0);
9666 		bcopy(state->dts_formats, new, ndx * sizeof (char *));
9667 		kmem_free(state->dts_formats, ndx * sizeof (char *));
9668 	}
9669 
9670 	state->dts_formats = new;
9671 	state->dts_formats[ndx] = fmt;
9672 
9673 	return (ndx + 1);
9674 }
9675 
9676 static void
9677 dtrace_format_remove(dtrace_state_t *state, uint16_t format)
9678 {
9679 	char *fmt;
9680 
9681 	ASSERT(state->dts_formats != NULL);
9682 	ASSERT(format <= state->dts_nformats);
9683 	ASSERT(state->dts_formats[format - 1] != NULL);
9684 
9685 	fmt = state->dts_formats[format - 1];
9686 	kmem_free(fmt, strlen(fmt) + 1);
9687 	state->dts_formats[format - 1] = NULL;
9688 }
9689 
9690 static void
9691 dtrace_format_destroy(dtrace_state_t *state)
9692 {
9693 	int i;
9694 
9695 	if (state->dts_nformats == 0) {
9696 		ASSERT(state->dts_formats == NULL);
9697 		return;
9698 	}
9699 
9700 	ASSERT(state->dts_formats != NULL);
9701 
9702 	for (i = 0; i < state->dts_nformats; i++) {
9703 		char *fmt = state->dts_formats[i];
9704 
9705 		if (fmt == NULL)
9706 			continue;
9707 
9708 		kmem_free(fmt, strlen(fmt) + 1);
9709 	}
9710 
9711 	kmem_free(state->dts_formats, state->dts_nformats * sizeof (char *));
9712 	state->dts_nformats = 0;
9713 	state->dts_formats = NULL;
9714 }
9715 
9716 /*
9717  * DTrace Predicate Functions
9718  */
9719 static dtrace_predicate_t *
9720 dtrace_predicate_create(dtrace_difo_t *dp)
9721 {
9722 	dtrace_predicate_t *pred;
9723 
9724 	ASSERT(MUTEX_HELD(&dtrace_lock));
9725 	ASSERT(dp->dtdo_refcnt != 0);
9726 
9727 	pred = kmem_zalloc(sizeof (dtrace_predicate_t), KM_SLEEP);
9728 	pred->dtp_difo = dp;
9729 	pred->dtp_refcnt = 1;
9730 
9731 	if (!dtrace_difo_cacheable(dp))
9732 		return (pred);
9733 
9734 	if (dtrace_predcache_id == DTRACE_CACHEIDNONE) {
9735 		/*
9736 		 * This is only theoretically possible -- we have had 2^32
9737 		 * cacheable predicates on this machine.  We cannot allow any
9738 		 * more predicates to become cacheable:  as unlikely as it is,
9739 		 * there may be a thread caching a (now stale) predicate cache
9740 		 * ID. (N.B.: the temptation is being successfully resisted to
9741 		 * have this cmn_err() "Holy shit -- we executed this code!")
9742 		 */
9743 		return (pred);
9744 	}
9745 
9746 	pred->dtp_cacheid = dtrace_predcache_id++;
9747 
9748 	return (pred);
9749 }
9750 
9751 static void
9752 dtrace_predicate_hold(dtrace_predicate_t *pred)
9753 {
9754 	ASSERT(MUTEX_HELD(&dtrace_lock));
9755 	ASSERT(pred->dtp_difo != NULL && pred->dtp_difo->dtdo_refcnt != 0);
9756 	ASSERT(pred->dtp_refcnt > 0);
9757 
9758 	pred->dtp_refcnt++;
9759 }
9760 
9761 static void
9762 dtrace_predicate_release(dtrace_predicate_t *pred, dtrace_vstate_t *vstate)
9763 {
9764 	dtrace_difo_t *dp = pred->dtp_difo;
9765 
9766 	ASSERT(MUTEX_HELD(&dtrace_lock));
9767 	ASSERT(dp != NULL && dp->dtdo_refcnt != 0);
9768 	ASSERT(pred->dtp_refcnt > 0);
9769 
9770 	if (--pred->dtp_refcnt == 0) {
9771 		dtrace_difo_release(pred->dtp_difo, vstate);
9772 		kmem_free(pred, sizeof (dtrace_predicate_t));
9773 	}
9774 }
9775 
9776 /*
9777  * DTrace Action Description Functions
9778  */
9779 static dtrace_actdesc_t *
9780 dtrace_actdesc_create(dtrace_actkind_t kind, uint32_t ntuple,
9781     uint64_t uarg, uint64_t arg)
9782 {
9783 	dtrace_actdesc_t *act;
9784 
9785 #if defined(sun)
9786 	ASSERT(!DTRACEACT_ISPRINTFLIKE(kind) || (arg != NULL &&
9787 	    arg >= KERNELBASE) || (arg == NULL && kind == DTRACEACT_PRINTA));
9788 #endif
9789 
9790 	act = kmem_zalloc(sizeof (dtrace_actdesc_t), KM_SLEEP);
9791 	act->dtad_kind = kind;
9792 	act->dtad_ntuple = ntuple;
9793 	act->dtad_uarg = uarg;
9794 	act->dtad_arg = arg;
9795 	act->dtad_refcnt = 1;
9796 
9797 	return (act);
9798 }
9799 
9800 static void
9801 dtrace_actdesc_hold(dtrace_actdesc_t *act)
9802 {
9803 	ASSERT(act->dtad_refcnt >= 1);
9804 	act->dtad_refcnt++;
9805 }
9806 
9807 static void
9808 dtrace_actdesc_release(dtrace_actdesc_t *act, dtrace_vstate_t *vstate)
9809 {
9810 	dtrace_actkind_t kind = act->dtad_kind;
9811 	dtrace_difo_t *dp;
9812 
9813 	ASSERT(act->dtad_refcnt >= 1);
9814 
9815 	if (--act->dtad_refcnt != 0)
9816 		return;
9817 
9818 	if ((dp = act->dtad_difo) != NULL)
9819 		dtrace_difo_release(dp, vstate);
9820 
9821 	if (DTRACEACT_ISPRINTFLIKE(kind)) {
9822 		char *str = (char *)(uintptr_t)act->dtad_arg;
9823 
9824 #if defined(sun)
9825 		ASSERT((str != NULL && (uintptr_t)str >= KERNELBASE) ||
9826 		    (str == NULL && act->dtad_kind == DTRACEACT_PRINTA));
9827 #endif
9828 
9829 		if (str != NULL)
9830 			kmem_free(str, strlen(str) + 1);
9831 	}
9832 
9833 	kmem_free(act, sizeof (dtrace_actdesc_t));
9834 }
9835 
9836 /*
9837  * DTrace ECB Functions
9838  */
9839 static dtrace_ecb_t *
9840 dtrace_ecb_add(dtrace_state_t *state, dtrace_probe_t *probe)
9841 {
9842 	dtrace_ecb_t *ecb;
9843 	dtrace_epid_t epid;
9844 
9845 	ASSERT(MUTEX_HELD(&dtrace_lock));
9846 
9847 	ecb = kmem_zalloc(sizeof (dtrace_ecb_t), KM_SLEEP);
9848 	ecb->dte_predicate = NULL;
9849 	ecb->dte_probe = probe;
9850 
9851 	/*
9852 	 * The default size is the size of the default action: recording
9853 	 * the header.
9854 	 */
9855 	ecb->dte_size = ecb->dte_needed = sizeof (dtrace_rechdr_t);
9856 	ecb->dte_alignment = sizeof (dtrace_epid_t);
9857 
9858 	epid = state->dts_epid++;
9859 
9860 	if (epid - 1 >= state->dts_necbs) {
9861 		dtrace_ecb_t **oecbs = state->dts_ecbs, **ecbs;
9862 		int necbs = state->dts_necbs << 1;
9863 
9864 		ASSERT(epid == state->dts_necbs + 1);
9865 
9866 		if (necbs == 0) {
9867 			ASSERT(oecbs == NULL);
9868 			necbs = 1;
9869 		}
9870 
9871 		ecbs = kmem_zalloc(necbs * sizeof (*ecbs), KM_SLEEP);
9872 
9873 		if (oecbs != NULL)
9874 			bcopy(oecbs, ecbs, state->dts_necbs * sizeof (*ecbs));
9875 
9876 		dtrace_membar_producer();
9877 		state->dts_ecbs = ecbs;
9878 
9879 		if (oecbs != NULL) {
9880 			/*
9881 			 * If this state is active, we must dtrace_sync()
9882 			 * before we can free the old dts_ecbs array:  we're
9883 			 * coming in hot, and there may be active ring
9884 			 * buffer processing (which indexes into the dts_ecbs
9885 			 * array) on another CPU.
9886 			 */
9887 			if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
9888 				dtrace_sync();
9889 
9890 			kmem_free(oecbs, state->dts_necbs * sizeof (*ecbs));
9891 		}
9892 
9893 		dtrace_membar_producer();
9894 		state->dts_necbs = necbs;
9895 	}
9896 
9897 	ecb->dte_state = state;
9898 
9899 	ASSERT(state->dts_ecbs[epid - 1] == NULL);
9900 	dtrace_membar_producer();
9901 	state->dts_ecbs[(ecb->dte_epid = epid) - 1] = ecb;
9902 
9903 	return (ecb);
9904 }
9905 
9906 static void
9907 dtrace_ecb_enable(dtrace_ecb_t *ecb)
9908 {
9909 	dtrace_probe_t *probe = ecb->dte_probe;
9910 
9911 	ASSERT(MUTEX_HELD(&cpu_lock));
9912 	ASSERT(MUTEX_HELD(&dtrace_lock));
9913 	ASSERT(ecb->dte_next == NULL);
9914 
9915 	if (probe == NULL) {
9916 		/*
9917 		 * This is the NULL probe -- there's nothing to do.
9918 		 */
9919 		return;
9920 	}
9921 
9922 	if (probe->dtpr_ecb == NULL) {
9923 		dtrace_provider_t *prov = probe->dtpr_provider;
9924 
9925 		/*
9926 		 * We're the first ECB on this probe.
9927 		 */
9928 		probe->dtpr_ecb = probe->dtpr_ecb_last = ecb;
9929 
9930 		if (ecb->dte_predicate != NULL)
9931 			probe->dtpr_predcache = ecb->dte_predicate->dtp_cacheid;
9932 
9933 		prov->dtpv_pops.dtps_enable(prov->dtpv_arg,
9934 		    probe->dtpr_id, probe->dtpr_arg);
9935 	} else {
9936 		/*
9937 		 * This probe is already active.  Swing the last pointer to
9938 		 * point to the new ECB, and issue a dtrace_sync() to assure
9939 		 * that all CPUs have seen the change.
9940 		 */
9941 		ASSERT(probe->dtpr_ecb_last != NULL);
9942 		probe->dtpr_ecb_last->dte_next = ecb;
9943 		probe->dtpr_ecb_last = ecb;
9944 		probe->dtpr_predcache = 0;
9945 
9946 		dtrace_sync();
9947 	}
9948 }
9949 
9950 static void
9951 dtrace_ecb_resize(dtrace_ecb_t *ecb)
9952 {
9953 	dtrace_action_t *act;
9954 	uint32_t curneeded = UINT32_MAX;
9955 	uint32_t aggbase = UINT32_MAX;
9956 
9957 	/*
9958 	 * If we record anything, we always record the dtrace_rechdr_t.  (And
9959 	 * we always record it first.)
9960 	 */
9961 	ecb->dte_size = sizeof (dtrace_rechdr_t);
9962 	ecb->dte_alignment = sizeof (dtrace_epid_t);
9963 
9964 	for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
9965 		dtrace_recdesc_t *rec = &act->dta_rec;
9966 		ASSERT(rec->dtrd_size > 0 || rec->dtrd_alignment == 1);
9967 
9968 		ecb->dte_alignment = MAX(ecb->dte_alignment,
9969 		    rec->dtrd_alignment);
9970 
9971 		if (DTRACEACT_ISAGG(act->dta_kind)) {
9972 			dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
9973 
9974 			ASSERT(rec->dtrd_size != 0);
9975 			ASSERT(agg->dtag_first != NULL);
9976 			ASSERT(act->dta_prev->dta_intuple);
9977 			ASSERT(aggbase != UINT32_MAX);
9978 			ASSERT(curneeded != UINT32_MAX);
9979 
9980 			agg->dtag_base = aggbase;
9981 
9982 			curneeded = P2ROUNDUP(curneeded, rec->dtrd_alignment);
9983 			rec->dtrd_offset = curneeded;
9984 			curneeded += rec->dtrd_size;
9985 			ecb->dte_needed = MAX(ecb->dte_needed, curneeded);
9986 
9987 			aggbase = UINT32_MAX;
9988 			curneeded = UINT32_MAX;
9989 		} else if (act->dta_intuple) {
9990 			if (curneeded == UINT32_MAX) {
9991 				/*
9992 				 * This is the first record in a tuple.  Align
9993 				 * curneeded to be at offset 4 in an 8-byte
9994 				 * aligned block.
9995 				 */
9996 				ASSERT(act->dta_prev == NULL ||
9997 				    !act->dta_prev->dta_intuple);
9998 				ASSERT3U(aggbase, ==, UINT32_MAX);
9999 				curneeded = P2PHASEUP(ecb->dte_size,
10000 				    sizeof (uint64_t), sizeof (dtrace_aggid_t));
10001 
10002 				aggbase = curneeded - sizeof (dtrace_aggid_t);
10003 				ASSERT(IS_P2ALIGNED(aggbase,
10004 				    sizeof (uint64_t)));
10005 			}
10006 			curneeded = P2ROUNDUP(curneeded, rec->dtrd_alignment);
10007 			rec->dtrd_offset = curneeded;
10008 			curneeded += rec->dtrd_size;
10009 		} else {
10010 			/* tuples must be followed by an aggregation */
10011 			ASSERT(act->dta_prev == NULL ||
10012 			    !act->dta_prev->dta_intuple);
10013 
10014 			ecb->dte_size = P2ROUNDUP(ecb->dte_size,
10015 			    rec->dtrd_alignment);
10016 			rec->dtrd_offset = ecb->dte_size;
10017 			ecb->dte_size += rec->dtrd_size;
10018 			ecb->dte_needed = MAX(ecb->dte_needed, ecb->dte_size);
10019 		}
10020 	}
10021 
10022 	if ((act = ecb->dte_action) != NULL &&
10023 	    !(act->dta_kind == DTRACEACT_SPECULATE && act->dta_next == NULL) &&
10024 	    ecb->dte_size == sizeof (dtrace_rechdr_t)) {
10025 		/*
10026 		 * If the size is still sizeof (dtrace_rechdr_t), then all
10027 		 * actions store no data; set the size to 0.
10028 		 */
10029 		ecb->dte_size = 0;
10030 	}
10031 
10032 	ecb->dte_size = P2ROUNDUP(ecb->dte_size, sizeof (dtrace_epid_t));
10033 	ecb->dte_needed = P2ROUNDUP(ecb->dte_needed, (sizeof (dtrace_epid_t)));
10034 	ecb->dte_state->dts_needed = MAX(ecb->dte_state->dts_needed,
10035 	    ecb->dte_needed);
10036 }
10037 
10038 static dtrace_action_t *
10039 dtrace_ecb_aggregation_create(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
10040 {
10041 	dtrace_aggregation_t *agg;
10042 	size_t size = sizeof (uint64_t);
10043 	int ntuple = desc->dtad_ntuple;
10044 	dtrace_action_t *act;
10045 	dtrace_recdesc_t *frec;
10046 	dtrace_aggid_t aggid;
10047 	dtrace_state_t *state = ecb->dte_state;
10048 
10049 	agg = kmem_zalloc(sizeof (dtrace_aggregation_t), KM_SLEEP);
10050 	agg->dtag_ecb = ecb;
10051 
10052 	ASSERT(DTRACEACT_ISAGG(desc->dtad_kind));
10053 
10054 	switch (desc->dtad_kind) {
10055 	case DTRACEAGG_MIN:
10056 		agg->dtag_initial = INT64_MAX;
10057 		agg->dtag_aggregate = dtrace_aggregate_min;
10058 		break;
10059 
10060 	case DTRACEAGG_MAX:
10061 		agg->dtag_initial = INT64_MIN;
10062 		agg->dtag_aggregate = dtrace_aggregate_max;
10063 		break;
10064 
10065 	case DTRACEAGG_COUNT:
10066 		agg->dtag_aggregate = dtrace_aggregate_count;
10067 		break;
10068 
10069 	case DTRACEAGG_QUANTIZE:
10070 		agg->dtag_aggregate = dtrace_aggregate_quantize;
10071 		size = (((sizeof (uint64_t) * NBBY) - 1) * 2 + 1) *
10072 		    sizeof (uint64_t);
10073 		break;
10074 
10075 	case DTRACEAGG_LQUANTIZE: {
10076 		uint16_t step = DTRACE_LQUANTIZE_STEP(desc->dtad_arg);
10077 		uint16_t levels = DTRACE_LQUANTIZE_LEVELS(desc->dtad_arg);
10078 
10079 		agg->dtag_initial = desc->dtad_arg;
10080 		agg->dtag_aggregate = dtrace_aggregate_lquantize;
10081 
10082 		if (step == 0 || levels == 0)
10083 			goto err;
10084 
10085 		size = levels * sizeof (uint64_t) + 3 * sizeof (uint64_t);
10086 		break;
10087 	}
10088 
10089 	case DTRACEAGG_LLQUANTIZE: {
10090 		uint16_t factor = DTRACE_LLQUANTIZE_FACTOR(desc->dtad_arg);
10091 		uint16_t low = DTRACE_LLQUANTIZE_LOW(desc->dtad_arg);
10092 		uint16_t high = DTRACE_LLQUANTIZE_HIGH(desc->dtad_arg);
10093 		uint16_t nsteps = DTRACE_LLQUANTIZE_NSTEP(desc->dtad_arg);
10094 		int64_t v;
10095 
10096 		agg->dtag_initial = desc->dtad_arg;
10097 		agg->dtag_aggregate = dtrace_aggregate_llquantize;
10098 
10099 		if (factor < 2 || low >= high || nsteps < factor)
10100 			goto err;
10101 
10102 		/*
10103 		 * Now check that the number of steps evenly divides a power
10104 		 * of the factor.  (This assures both integer bucket size and
10105 		 * linearity within each magnitude.)
10106 		 */
10107 		for (v = factor; v < nsteps; v *= factor)
10108 			continue;
10109 
10110 		if ((v % nsteps) || (nsteps % factor))
10111 			goto err;
10112 
10113 		size = (dtrace_aggregate_llquantize_bucket(factor,
10114 		    low, high, nsteps, INT64_MAX) + 2) * sizeof (uint64_t);
10115 		break;
10116 	}
10117 
10118 	case DTRACEAGG_AVG:
10119 		agg->dtag_aggregate = dtrace_aggregate_avg;
10120 		size = sizeof (uint64_t) * 2;
10121 		break;
10122 
10123 	case DTRACEAGG_STDDEV:
10124 		agg->dtag_aggregate = dtrace_aggregate_stddev;
10125 		size = sizeof (uint64_t) * 4;
10126 		break;
10127 
10128 	case DTRACEAGG_SUM:
10129 		agg->dtag_aggregate = dtrace_aggregate_sum;
10130 		break;
10131 
10132 	default:
10133 		goto err;
10134 	}
10135 
10136 	agg->dtag_action.dta_rec.dtrd_size = size;
10137 
10138 	if (ntuple == 0)
10139 		goto err;
10140 
10141 	/*
10142 	 * We must make sure that we have enough actions for the n-tuple.
10143 	 */
10144 	for (act = ecb->dte_action_last; act != NULL; act = act->dta_prev) {
10145 		if (DTRACEACT_ISAGG(act->dta_kind))
10146 			break;
10147 
10148 		if (--ntuple == 0) {
10149 			/*
10150 			 * This is the action with which our n-tuple begins.
10151 			 */
10152 			agg->dtag_first = act;
10153 			goto success;
10154 		}
10155 	}
10156 
10157 	/*
10158 	 * This n-tuple is short by ntuple elements.  Return failure.
10159 	 */
10160 	ASSERT(ntuple != 0);
10161 err:
10162 	kmem_free(agg, sizeof (dtrace_aggregation_t));
10163 	return (NULL);
10164 
10165 success:
10166 	/*
10167 	 * If the last action in the tuple has a size of zero, it's actually
10168 	 * an expression argument for the aggregating action.
10169 	 */
10170 	ASSERT(ecb->dte_action_last != NULL);
10171 	act = ecb->dte_action_last;
10172 
10173 	if (act->dta_kind == DTRACEACT_DIFEXPR) {
10174 		ASSERT(act->dta_difo != NULL);
10175 
10176 		if (act->dta_difo->dtdo_rtype.dtdt_size == 0)
10177 			agg->dtag_hasarg = 1;
10178 	}
10179 
10180 	/*
10181 	 * We need to allocate an id for this aggregation.
10182 	 */
10183 #if defined(sun)
10184 	aggid = (dtrace_aggid_t)(uintptr_t)vmem_alloc(state->dts_aggid_arena, 1,
10185 	    VM_BESTFIT | VM_SLEEP);
10186 #else
10187 	aggid = alloc_unr(state->dts_aggid_arena);
10188 #endif
10189 
10190 	if (aggid - 1 >= state->dts_naggregations) {
10191 		dtrace_aggregation_t **oaggs = state->dts_aggregations;
10192 		dtrace_aggregation_t **aggs;
10193 		int naggs = state->dts_naggregations << 1;
10194 		int onaggs = state->dts_naggregations;
10195 
10196 		ASSERT(aggid == state->dts_naggregations + 1);
10197 
10198 		if (naggs == 0) {
10199 			ASSERT(oaggs == NULL);
10200 			naggs = 1;
10201 		}
10202 
10203 		aggs = kmem_zalloc(naggs * sizeof (*aggs), KM_SLEEP);
10204 
10205 		if (oaggs != NULL) {
10206 			bcopy(oaggs, aggs, onaggs * sizeof (*aggs));
10207 			kmem_free(oaggs, onaggs * sizeof (*aggs));
10208 		}
10209 
10210 		state->dts_aggregations = aggs;
10211 		state->dts_naggregations = naggs;
10212 	}
10213 
10214 	ASSERT(state->dts_aggregations[aggid - 1] == NULL);
10215 	state->dts_aggregations[(agg->dtag_id = aggid) - 1] = agg;
10216 
10217 	frec = &agg->dtag_first->dta_rec;
10218 	if (frec->dtrd_alignment < sizeof (dtrace_aggid_t))
10219 		frec->dtrd_alignment = sizeof (dtrace_aggid_t);
10220 
10221 	for (act = agg->dtag_first; act != NULL; act = act->dta_next) {
10222 		ASSERT(!act->dta_intuple);
10223 		act->dta_intuple = 1;
10224 	}
10225 
10226 	return (&agg->dtag_action);
10227 }
10228 
10229 static void
10230 dtrace_ecb_aggregation_destroy(dtrace_ecb_t *ecb, dtrace_action_t *act)
10231 {
10232 	dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
10233 	dtrace_state_t *state = ecb->dte_state;
10234 	dtrace_aggid_t aggid = agg->dtag_id;
10235 
10236 	ASSERT(DTRACEACT_ISAGG(act->dta_kind));
10237 #if defined(sun)
10238 	vmem_free(state->dts_aggid_arena, (void *)(uintptr_t)aggid, 1);
10239 #else
10240 	free_unr(state->dts_aggid_arena, aggid);
10241 #endif
10242 
10243 	ASSERT(state->dts_aggregations[aggid - 1] == agg);
10244 	state->dts_aggregations[aggid - 1] = NULL;
10245 
10246 	kmem_free(agg, sizeof (dtrace_aggregation_t));
10247 }
10248 
10249 static int
10250 dtrace_ecb_action_add(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
10251 {
10252 	dtrace_action_t *action, *last;
10253 	dtrace_difo_t *dp = desc->dtad_difo;
10254 	uint32_t size = 0, align = sizeof (uint8_t), mask;
10255 	uint16_t format = 0;
10256 	dtrace_recdesc_t *rec;
10257 	dtrace_state_t *state = ecb->dte_state;
10258 	dtrace_optval_t *opt = state->dts_options, nframes = 0, strsize;
10259 	uint64_t arg = desc->dtad_arg;
10260 
10261 	ASSERT(MUTEX_HELD(&dtrace_lock));
10262 	ASSERT(ecb->dte_action == NULL || ecb->dte_action->dta_refcnt == 1);
10263 
10264 	if (DTRACEACT_ISAGG(desc->dtad_kind)) {
10265 		/*
10266 		 * If this is an aggregating action, there must be neither
10267 		 * a speculate nor a commit on the action chain.
10268 		 */
10269 		dtrace_action_t *act;
10270 
10271 		for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
10272 			if (act->dta_kind == DTRACEACT_COMMIT)
10273 				return (EINVAL);
10274 
10275 			if (act->dta_kind == DTRACEACT_SPECULATE)
10276 				return (EINVAL);
10277 		}
10278 
10279 		action = dtrace_ecb_aggregation_create(ecb, desc);
10280 
10281 		if (action == NULL)
10282 			return (EINVAL);
10283 	} else {
10284 		if (DTRACEACT_ISDESTRUCTIVE(desc->dtad_kind) ||
10285 		    (desc->dtad_kind == DTRACEACT_DIFEXPR &&
10286 		    dp != NULL && dp->dtdo_destructive)) {
10287 			state->dts_destructive = 1;
10288 		}
10289 
10290 		switch (desc->dtad_kind) {
10291 		case DTRACEACT_PRINTF:
10292 		case DTRACEACT_PRINTA:
10293 		case DTRACEACT_SYSTEM:
10294 		case DTRACEACT_FREOPEN:
10295 		case DTRACEACT_DIFEXPR:
10296 			/*
10297 			 * We know that our arg is a string -- turn it into a
10298 			 * format.
10299 			 */
10300 			if (arg == 0) {
10301 				ASSERT(desc->dtad_kind == DTRACEACT_PRINTA ||
10302 				    desc->dtad_kind == DTRACEACT_DIFEXPR);
10303 				format = 0;
10304 			} else {
10305 				ASSERT(arg != 0);
10306 #if defined(sun)
10307 				ASSERT(arg > KERNELBASE);
10308 #endif
10309 				format = dtrace_format_add(state,
10310 				    (char *)(uintptr_t)arg);
10311 			}
10312 
10313 			/*FALLTHROUGH*/
10314 		case DTRACEACT_LIBACT:
10315 		case DTRACEACT_TRACEMEM:
10316 		case DTRACEACT_TRACEMEM_DYNSIZE:
10317 			if (dp == NULL)
10318 				return (EINVAL);
10319 
10320 			if ((size = dp->dtdo_rtype.dtdt_size) != 0)
10321 				break;
10322 
10323 			if (dp->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING) {
10324 				if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10325 					return (EINVAL);
10326 
10327 				size = opt[DTRACEOPT_STRSIZE];
10328 			}
10329 
10330 			break;
10331 
10332 		case DTRACEACT_STACK:
10333 			if ((nframes = arg) == 0) {
10334 				nframes = opt[DTRACEOPT_STACKFRAMES];
10335 				ASSERT(nframes > 0);
10336 				arg = nframes;
10337 			}
10338 
10339 			size = nframes * sizeof (pc_t);
10340 			break;
10341 
10342 		case DTRACEACT_JSTACK:
10343 			if ((strsize = DTRACE_USTACK_STRSIZE(arg)) == 0)
10344 				strsize = opt[DTRACEOPT_JSTACKSTRSIZE];
10345 
10346 			if ((nframes = DTRACE_USTACK_NFRAMES(arg)) == 0)
10347 				nframes = opt[DTRACEOPT_JSTACKFRAMES];
10348 
10349 			arg = DTRACE_USTACK_ARG(nframes, strsize);
10350 
10351 			/*FALLTHROUGH*/
10352 		case DTRACEACT_USTACK:
10353 			if (desc->dtad_kind != DTRACEACT_JSTACK &&
10354 			    (nframes = DTRACE_USTACK_NFRAMES(arg)) == 0) {
10355 				strsize = DTRACE_USTACK_STRSIZE(arg);
10356 				nframes = opt[DTRACEOPT_USTACKFRAMES];
10357 				ASSERT(nframes > 0);
10358 				arg = DTRACE_USTACK_ARG(nframes, strsize);
10359 			}
10360 
10361 			/*
10362 			 * Save a slot for the pid.
10363 			 */
10364 			size = (nframes + 1) * sizeof (uint64_t);
10365 			size += DTRACE_USTACK_STRSIZE(arg);
10366 			size = P2ROUNDUP(size, (uint32_t)(sizeof (uintptr_t)));
10367 
10368 			break;
10369 
10370 		case DTRACEACT_SYM:
10371 		case DTRACEACT_MOD:
10372 			if (dp == NULL || ((size = dp->dtdo_rtype.dtdt_size) !=
10373 			    sizeof (uint64_t)) ||
10374 			    (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10375 				return (EINVAL);
10376 			break;
10377 
10378 		case DTRACEACT_USYM:
10379 		case DTRACEACT_UMOD:
10380 		case DTRACEACT_UADDR:
10381 			if (dp == NULL ||
10382 			    (dp->dtdo_rtype.dtdt_size != sizeof (uint64_t)) ||
10383 			    (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10384 				return (EINVAL);
10385 
10386 			/*
10387 			 * We have a slot for the pid, plus a slot for the
10388 			 * argument.  To keep things simple (aligned with
10389 			 * bitness-neutral sizing), we store each as a 64-bit
10390 			 * quantity.
10391 			 */
10392 			size = 2 * sizeof (uint64_t);
10393 			break;
10394 
10395 		case DTRACEACT_STOP:
10396 		case DTRACEACT_BREAKPOINT:
10397 		case DTRACEACT_PANIC:
10398 			break;
10399 
10400 		case DTRACEACT_CHILL:
10401 		case DTRACEACT_DISCARD:
10402 		case DTRACEACT_RAISE:
10403 			if (dp == NULL)
10404 				return (EINVAL);
10405 			break;
10406 
10407 		case DTRACEACT_EXIT:
10408 			if (dp == NULL ||
10409 			    (size = dp->dtdo_rtype.dtdt_size) != sizeof (int) ||
10410 			    (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10411 				return (EINVAL);
10412 			break;
10413 
10414 		case DTRACEACT_SPECULATE:
10415 			if (ecb->dte_size > sizeof (dtrace_rechdr_t))
10416 				return (EINVAL);
10417 
10418 			if (dp == NULL)
10419 				return (EINVAL);
10420 
10421 			state->dts_speculates = 1;
10422 			break;
10423 
10424 		case DTRACEACT_PRINTM:
10425 		    	size = dp->dtdo_rtype.dtdt_size;
10426 			break;
10427 
10428 		case DTRACEACT_PRINTT:
10429 		    	size = dp->dtdo_rtype.dtdt_size;
10430 			break;
10431 
10432 		case DTRACEACT_COMMIT: {
10433 			dtrace_action_t *act = ecb->dte_action;
10434 
10435 			for (; act != NULL; act = act->dta_next) {
10436 				if (act->dta_kind == DTRACEACT_COMMIT)
10437 					return (EINVAL);
10438 			}
10439 
10440 			if (dp == NULL)
10441 				return (EINVAL);
10442 			break;
10443 		}
10444 
10445 		default:
10446 			return (EINVAL);
10447 		}
10448 
10449 		if (size != 0 || desc->dtad_kind == DTRACEACT_SPECULATE) {
10450 			/*
10451 			 * If this is a data-storing action or a speculate,
10452 			 * we must be sure that there isn't a commit on the
10453 			 * action chain.
10454 			 */
10455 			dtrace_action_t *act = ecb->dte_action;
10456 
10457 			for (; act != NULL; act = act->dta_next) {
10458 				if (act->dta_kind == DTRACEACT_COMMIT)
10459 					return (EINVAL);
10460 			}
10461 		}
10462 
10463 		action = kmem_zalloc(sizeof (dtrace_action_t), KM_SLEEP);
10464 		action->dta_rec.dtrd_size = size;
10465 	}
10466 
10467 	action->dta_refcnt = 1;
10468 	rec = &action->dta_rec;
10469 	size = rec->dtrd_size;
10470 
10471 	for (mask = sizeof (uint64_t) - 1; size != 0 && mask > 0; mask >>= 1) {
10472 		if (!(size & mask)) {
10473 			align = mask + 1;
10474 			break;
10475 		}
10476 	}
10477 
10478 	action->dta_kind = desc->dtad_kind;
10479 
10480 	if ((action->dta_difo = dp) != NULL)
10481 		dtrace_difo_hold(dp);
10482 
10483 	rec->dtrd_action = action->dta_kind;
10484 	rec->dtrd_arg = arg;
10485 	rec->dtrd_uarg = desc->dtad_uarg;
10486 	rec->dtrd_alignment = (uint16_t)align;
10487 	rec->dtrd_format = format;
10488 
10489 	if ((last = ecb->dte_action_last) != NULL) {
10490 		ASSERT(ecb->dte_action != NULL);
10491 		action->dta_prev = last;
10492 		last->dta_next = action;
10493 	} else {
10494 		ASSERT(ecb->dte_action == NULL);
10495 		ecb->dte_action = action;
10496 	}
10497 
10498 	ecb->dte_action_last = action;
10499 
10500 	return (0);
10501 }
10502 
10503 static void
10504 dtrace_ecb_action_remove(dtrace_ecb_t *ecb)
10505 {
10506 	dtrace_action_t *act = ecb->dte_action, *next;
10507 	dtrace_vstate_t *vstate = &ecb->dte_state->dts_vstate;
10508 	dtrace_difo_t *dp;
10509 	uint16_t format;
10510 
10511 	if (act != NULL && act->dta_refcnt > 1) {
10512 		ASSERT(act->dta_next == NULL || act->dta_next->dta_refcnt == 1);
10513 		act->dta_refcnt--;
10514 	} else {
10515 		for (; act != NULL; act = next) {
10516 			next = act->dta_next;
10517 			ASSERT(next != NULL || act == ecb->dte_action_last);
10518 			ASSERT(act->dta_refcnt == 1);
10519 
10520 			if ((format = act->dta_rec.dtrd_format) != 0)
10521 				dtrace_format_remove(ecb->dte_state, format);
10522 
10523 			if ((dp = act->dta_difo) != NULL)
10524 				dtrace_difo_release(dp, vstate);
10525 
10526 			if (DTRACEACT_ISAGG(act->dta_kind)) {
10527 				dtrace_ecb_aggregation_destroy(ecb, act);
10528 			} else {
10529 				kmem_free(act, sizeof (dtrace_action_t));
10530 			}
10531 		}
10532 	}
10533 
10534 	ecb->dte_action = NULL;
10535 	ecb->dte_action_last = NULL;
10536 	ecb->dte_size = 0;
10537 }
10538 
10539 static void
10540 dtrace_ecb_disable(dtrace_ecb_t *ecb)
10541 {
10542 	/*
10543 	 * We disable the ECB by removing it from its probe.
10544 	 */
10545 	dtrace_ecb_t *pecb, *prev = NULL;
10546 	dtrace_probe_t *probe = ecb->dte_probe;
10547 
10548 	ASSERT(MUTEX_HELD(&dtrace_lock));
10549 
10550 	if (probe == NULL) {
10551 		/*
10552 		 * This is the NULL probe; there is nothing to disable.
10553 		 */
10554 		return;
10555 	}
10556 
10557 	for (pecb = probe->dtpr_ecb; pecb != NULL; pecb = pecb->dte_next) {
10558 		if (pecb == ecb)
10559 			break;
10560 		prev = pecb;
10561 	}
10562 
10563 	ASSERT(pecb != NULL);
10564 
10565 	if (prev == NULL) {
10566 		probe->dtpr_ecb = ecb->dte_next;
10567 	} else {
10568 		prev->dte_next = ecb->dte_next;
10569 	}
10570 
10571 	if (ecb == probe->dtpr_ecb_last) {
10572 		ASSERT(ecb->dte_next == NULL);
10573 		probe->dtpr_ecb_last = prev;
10574 	}
10575 
10576 	/*
10577 	 * The ECB has been disconnected from the probe; now sync to assure
10578 	 * that all CPUs have seen the change before returning.
10579 	 */
10580 	dtrace_sync();
10581 
10582 	if (probe->dtpr_ecb == NULL) {
10583 		/*
10584 		 * That was the last ECB on the probe; clear the predicate
10585 		 * cache ID for the probe, disable it and sync one more time
10586 		 * to assure that we'll never hit it again.
10587 		 */
10588 		dtrace_provider_t *prov = probe->dtpr_provider;
10589 
10590 		ASSERT(ecb->dte_next == NULL);
10591 		ASSERT(probe->dtpr_ecb_last == NULL);
10592 		probe->dtpr_predcache = DTRACE_CACHEIDNONE;
10593 		prov->dtpv_pops.dtps_disable(prov->dtpv_arg,
10594 		    probe->dtpr_id, probe->dtpr_arg);
10595 		dtrace_sync();
10596 	} else {
10597 		/*
10598 		 * There is at least one ECB remaining on the probe.  If there
10599 		 * is _exactly_ one, set the probe's predicate cache ID to be
10600 		 * the predicate cache ID of the remaining ECB.
10601 		 */
10602 		ASSERT(probe->dtpr_ecb_last != NULL);
10603 		ASSERT(probe->dtpr_predcache == DTRACE_CACHEIDNONE);
10604 
10605 		if (probe->dtpr_ecb == probe->dtpr_ecb_last) {
10606 			dtrace_predicate_t *p = probe->dtpr_ecb->dte_predicate;
10607 
10608 			ASSERT(probe->dtpr_ecb->dte_next == NULL);
10609 
10610 			if (p != NULL)
10611 				probe->dtpr_predcache = p->dtp_cacheid;
10612 		}
10613 
10614 		ecb->dte_next = NULL;
10615 	}
10616 }
10617 
10618 static void
10619 dtrace_ecb_destroy(dtrace_ecb_t *ecb)
10620 {
10621 	dtrace_state_t *state = ecb->dte_state;
10622 	dtrace_vstate_t *vstate = &state->dts_vstate;
10623 	dtrace_predicate_t *pred;
10624 	dtrace_epid_t epid = ecb->dte_epid;
10625 
10626 	ASSERT(MUTEX_HELD(&dtrace_lock));
10627 	ASSERT(ecb->dte_next == NULL);
10628 	ASSERT(ecb->dte_probe == NULL || ecb->dte_probe->dtpr_ecb != ecb);
10629 
10630 	if ((pred = ecb->dte_predicate) != NULL)
10631 		dtrace_predicate_release(pred, vstate);
10632 
10633 	dtrace_ecb_action_remove(ecb);
10634 
10635 	ASSERT(state->dts_ecbs[epid - 1] == ecb);
10636 	state->dts_ecbs[epid - 1] = NULL;
10637 
10638 	kmem_free(ecb, sizeof (dtrace_ecb_t));
10639 }
10640 
10641 static dtrace_ecb_t *
10642 dtrace_ecb_create(dtrace_state_t *state, dtrace_probe_t *probe,
10643     dtrace_enabling_t *enab)
10644 {
10645 	dtrace_ecb_t *ecb;
10646 	dtrace_predicate_t *pred;
10647 	dtrace_actdesc_t *act;
10648 	dtrace_provider_t *prov;
10649 	dtrace_ecbdesc_t *desc = enab->dten_current;
10650 
10651 	ASSERT(MUTEX_HELD(&dtrace_lock));
10652 	ASSERT(state != NULL);
10653 
10654 	ecb = dtrace_ecb_add(state, probe);
10655 	ecb->dte_uarg = desc->dted_uarg;
10656 
10657 	if ((pred = desc->dted_pred.dtpdd_predicate) != NULL) {
10658 		dtrace_predicate_hold(pred);
10659 		ecb->dte_predicate = pred;
10660 	}
10661 
10662 	if (probe != NULL) {
10663 		/*
10664 		 * If the provider shows more leg than the consumer is old
10665 		 * enough to see, we need to enable the appropriate implicit
10666 		 * predicate bits to prevent the ecb from activating at
10667 		 * revealing times.
10668 		 *
10669 		 * Providers specifying DTRACE_PRIV_USER at register time
10670 		 * are stating that they need the /proc-style privilege
10671 		 * model to be enforced, and this is what DTRACE_COND_OWNER
10672 		 * and DTRACE_COND_ZONEOWNER will then do at probe time.
10673 		 */
10674 		prov = probe->dtpr_provider;
10675 		if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLPROC) &&
10676 		    (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
10677 			ecb->dte_cond |= DTRACE_COND_OWNER;
10678 
10679 		if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLZONE) &&
10680 		    (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
10681 			ecb->dte_cond |= DTRACE_COND_ZONEOWNER;
10682 
10683 		/*
10684 		 * If the provider shows us kernel innards and the user
10685 		 * is lacking sufficient privilege, enable the
10686 		 * DTRACE_COND_USERMODE implicit predicate.
10687 		 */
10688 		if (!(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) &&
10689 		    (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_KERNEL))
10690 			ecb->dte_cond |= DTRACE_COND_USERMODE;
10691 	}
10692 
10693 	if (dtrace_ecb_create_cache != NULL) {
10694 		/*
10695 		 * If we have a cached ecb, we'll use its action list instead
10696 		 * of creating our own (saving both time and space).
10697 		 */
10698 		dtrace_ecb_t *cached = dtrace_ecb_create_cache;
10699 		dtrace_action_t *act = cached->dte_action;
10700 
10701 		if (act != NULL) {
10702 			ASSERT(act->dta_refcnt > 0);
10703 			act->dta_refcnt++;
10704 			ecb->dte_action = act;
10705 			ecb->dte_action_last = cached->dte_action_last;
10706 			ecb->dte_needed = cached->dte_needed;
10707 			ecb->dte_size = cached->dte_size;
10708 			ecb->dte_alignment = cached->dte_alignment;
10709 		}
10710 
10711 		return (ecb);
10712 	}
10713 
10714 	for (act = desc->dted_action; act != NULL; act = act->dtad_next) {
10715 		if ((enab->dten_error = dtrace_ecb_action_add(ecb, act)) != 0) {
10716 			dtrace_ecb_destroy(ecb);
10717 			return (NULL);
10718 		}
10719 	}
10720 
10721 	dtrace_ecb_resize(ecb);
10722 
10723 	return (dtrace_ecb_create_cache = ecb);
10724 }
10725 
10726 static int
10727 dtrace_ecb_create_enable(dtrace_probe_t *probe, void *arg)
10728 {
10729 	dtrace_ecb_t *ecb;
10730 	dtrace_enabling_t *enab = arg;
10731 	dtrace_state_t *state = enab->dten_vstate->dtvs_state;
10732 
10733 	ASSERT(state != NULL);
10734 
10735 	if (probe != NULL && probe->dtpr_gen < enab->dten_probegen) {
10736 		/*
10737 		 * This probe was created in a generation for which this
10738 		 * enabling has previously created ECBs; we don't want to
10739 		 * enable it again, so just kick out.
10740 		 */
10741 		return (DTRACE_MATCH_NEXT);
10742 	}
10743 
10744 	if ((ecb = dtrace_ecb_create(state, probe, enab)) == NULL)
10745 		return (DTRACE_MATCH_DONE);
10746 
10747 	dtrace_ecb_enable(ecb);
10748 	return (DTRACE_MATCH_NEXT);
10749 }
10750 
10751 static dtrace_ecb_t *
10752 dtrace_epid2ecb(dtrace_state_t *state, dtrace_epid_t id)
10753 {
10754 	dtrace_ecb_t *ecb;
10755 
10756 	ASSERT(MUTEX_HELD(&dtrace_lock));
10757 
10758 	if (id == 0 || id > state->dts_necbs)
10759 		return (NULL);
10760 
10761 	ASSERT(state->dts_necbs > 0 && state->dts_ecbs != NULL);
10762 	ASSERT((ecb = state->dts_ecbs[id - 1]) == NULL || ecb->dte_epid == id);
10763 
10764 	return (state->dts_ecbs[id - 1]);
10765 }
10766 
10767 static dtrace_aggregation_t *
10768 dtrace_aggid2agg(dtrace_state_t *state, dtrace_aggid_t id)
10769 {
10770 	dtrace_aggregation_t *agg;
10771 
10772 	ASSERT(MUTEX_HELD(&dtrace_lock));
10773 
10774 	if (id == 0 || id > state->dts_naggregations)
10775 		return (NULL);
10776 
10777 	ASSERT(state->dts_naggregations > 0 && state->dts_aggregations != NULL);
10778 	ASSERT((agg = state->dts_aggregations[id - 1]) == NULL ||
10779 	    agg->dtag_id == id);
10780 
10781 	return (state->dts_aggregations[id - 1]);
10782 }
10783 
10784 /*
10785  * DTrace Buffer Functions
10786  *
10787  * The following functions manipulate DTrace buffers.  Most of these functions
10788  * are called in the context of establishing or processing consumer state;
10789  * exceptions are explicitly noted.
10790  */
10791 
10792 /*
10793  * Note:  called from cross call context.  This function switches the two
10794  * buffers on a given CPU.  The atomicity of this operation is assured by
10795  * disabling interrupts while the actual switch takes place; the disabling of
10796  * interrupts serializes the execution with any execution of dtrace_probe() on
10797  * the same CPU.
10798  */
10799 static void
10800 dtrace_buffer_switch(dtrace_buffer_t *buf)
10801 {
10802 	caddr_t tomax = buf->dtb_tomax;
10803 	caddr_t xamot = buf->dtb_xamot;
10804 	dtrace_icookie_t cookie;
10805 	hrtime_t now;
10806 
10807 	ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
10808 	ASSERT(!(buf->dtb_flags & DTRACEBUF_RING));
10809 
10810 	cookie = dtrace_interrupt_disable();
10811 	now = dtrace_gethrtime();
10812 	buf->dtb_tomax = xamot;
10813 	buf->dtb_xamot = tomax;
10814 	buf->dtb_xamot_drops = buf->dtb_drops;
10815 	buf->dtb_xamot_offset = buf->dtb_offset;
10816 	buf->dtb_xamot_errors = buf->dtb_errors;
10817 	buf->dtb_xamot_flags = buf->dtb_flags;
10818 	buf->dtb_offset = 0;
10819 	buf->dtb_drops = 0;
10820 	buf->dtb_errors = 0;
10821 	buf->dtb_flags &= ~(DTRACEBUF_ERROR | DTRACEBUF_DROPPED);
10822 	buf->dtb_interval = now - buf->dtb_switched;
10823 	buf->dtb_switched = now;
10824 	dtrace_interrupt_enable(cookie);
10825 }
10826 
10827 /*
10828  * Note:  called from cross call context.  This function activates a buffer
10829  * on a CPU.  As with dtrace_buffer_switch(), the atomicity of the operation
10830  * is guaranteed by the disabling of interrupts.
10831  */
10832 static void
10833 dtrace_buffer_activate(dtrace_state_t *state)
10834 {
10835 	dtrace_buffer_t *buf;
10836 	dtrace_icookie_t cookie = dtrace_interrupt_disable();
10837 
10838 	buf = &state->dts_buffer[curcpu];
10839 
10840 	if (buf->dtb_tomax != NULL) {
10841 		/*
10842 		 * We might like to assert that the buffer is marked inactive,
10843 		 * but this isn't necessarily true:  the buffer for the CPU
10844 		 * that processes the BEGIN probe has its buffer activated
10845 		 * manually.  In this case, we take the (harmless) action
10846 		 * re-clearing the bit INACTIVE bit.
10847 		 */
10848 		buf->dtb_flags &= ~DTRACEBUF_INACTIVE;
10849 	}
10850 
10851 	dtrace_interrupt_enable(cookie);
10852 }
10853 
10854 static int
10855 dtrace_buffer_alloc(dtrace_buffer_t *bufs, size_t size, int flags,
10856     processorid_t cpu)
10857 {
10858 #if defined(sun)
10859 	cpu_t *cp;
10860 #endif
10861 	dtrace_buffer_t *buf;
10862 
10863 #if defined(sun)
10864 	ASSERT(MUTEX_HELD(&cpu_lock));
10865 	ASSERT(MUTEX_HELD(&dtrace_lock));
10866 
10867 	if (size > dtrace_nonroot_maxsize &&
10868 	    !PRIV_POLICY_CHOICE(CRED(), PRIV_ALL, B_FALSE))
10869 		return (EFBIG);
10870 
10871 	cp = cpu_list;
10872 
10873 	do {
10874 		if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
10875 			continue;
10876 
10877 		buf = &bufs[cp->cpu_id];
10878 
10879 		/*
10880 		 * If there is already a buffer allocated for this CPU, it
10881 		 * is only possible that this is a DR event.  In this case,
10882 		 */
10883 		if (buf->dtb_tomax != NULL) {
10884 			ASSERT(buf->dtb_size == size);
10885 			continue;
10886 		}
10887 
10888 		ASSERT(buf->dtb_xamot == NULL);
10889 
10890 		if ((buf->dtb_tomax = kmem_zalloc(size, KM_NOSLEEP)) == NULL)
10891 			goto err;
10892 
10893 		buf->dtb_size = size;
10894 		buf->dtb_flags = flags;
10895 		buf->dtb_offset = 0;
10896 		buf->dtb_drops = 0;
10897 
10898 		if (flags & DTRACEBUF_NOSWITCH)
10899 			continue;
10900 
10901 		if ((buf->dtb_xamot = kmem_zalloc(size, KM_NOSLEEP)) == NULL)
10902 			goto err;
10903 	} while ((cp = cp->cpu_next) != cpu_list);
10904 
10905 	return (0);
10906 
10907 err:
10908 	cp = cpu_list;
10909 
10910 	do {
10911 		if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
10912 			continue;
10913 
10914 		buf = &bufs[cp->cpu_id];
10915 
10916 		if (buf->dtb_xamot != NULL) {
10917 			ASSERT(buf->dtb_tomax != NULL);
10918 			ASSERT(buf->dtb_size == size);
10919 			kmem_free(buf->dtb_xamot, size);
10920 		}
10921 
10922 		if (buf->dtb_tomax != NULL) {
10923 			ASSERT(buf->dtb_size == size);
10924 			kmem_free(buf->dtb_tomax, size);
10925 		}
10926 
10927 		buf->dtb_tomax = NULL;
10928 		buf->dtb_xamot = NULL;
10929 		buf->dtb_size = 0;
10930 	} while ((cp = cp->cpu_next) != cpu_list);
10931 
10932 	return (ENOMEM);
10933 #else
10934 	int i;
10935 
10936 #if defined(__amd64__) || defined(__mips__) || defined(__powerpc__)
10937 	/*
10938 	 * FreeBSD isn't good at limiting the amount of memory we
10939 	 * ask to malloc, so let's place a limit here before trying
10940 	 * to do something that might well end in tears at bedtime.
10941 	 */
10942 	if (size > physmem * PAGE_SIZE / (128 * (mp_maxid + 1)))
10943 		return(ENOMEM);
10944 #endif
10945 
10946 	ASSERT(MUTEX_HELD(&dtrace_lock));
10947 	CPU_FOREACH(i) {
10948 		if (cpu != DTRACE_CPUALL && cpu != i)
10949 			continue;
10950 
10951 		buf = &bufs[i];
10952 
10953 		/*
10954 		 * If there is already a buffer allocated for this CPU, it
10955 		 * is only possible that this is a DR event.  In this case,
10956 		 * the buffer size must match our specified size.
10957 		 */
10958 		if (buf->dtb_tomax != NULL) {
10959 			ASSERT(buf->dtb_size == size);
10960 			continue;
10961 		}
10962 
10963 		ASSERT(buf->dtb_xamot == NULL);
10964 
10965 		if ((buf->dtb_tomax = kmem_zalloc(size, KM_NOSLEEP)) == NULL)
10966 			goto err;
10967 
10968 		buf->dtb_size = size;
10969 		buf->dtb_flags = flags;
10970 		buf->dtb_offset = 0;
10971 		buf->dtb_drops = 0;
10972 
10973 		if (flags & DTRACEBUF_NOSWITCH)
10974 			continue;
10975 
10976 		if ((buf->dtb_xamot = kmem_zalloc(size, KM_NOSLEEP)) == NULL)
10977 			goto err;
10978 	}
10979 
10980 	return (0);
10981 
10982 err:
10983 	/*
10984 	 * Error allocating memory, so free the buffers that were
10985 	 * allocated before the failed allocation.
10986 	 */
10987 	CPU_FOREACH(i) {
10988 		if (cpu != DTRACE_CPUALL && cpu != i)
10989 			continue;
10990 
10991 		buf = &bufs[i];
10992 
10993 		if (buf->dtb_xamot != NULL) {
10994 			ASSERT(buf->dtb_tomax != NULL);
10995 			ASSERT(buf->dtb_size == size);
10996 			kmem_free(buf->dtb_xamot, size);
10997 		}
10998 
10999 		if (buf->dtb_tomax != NULL) {
11000 			ASSERT(buf->dtb_size == size);
11001 			kmem_free(buf->dtb_tomax, size);
11002 		}
11003 
11004 		buf->dtb_tomax = NULL;
11005 		buf->dtb_xamot = NULL;
11006 		buf->dtb_size = 0;
11007 
11008 	}
11009 
11010 	return (ENOMEM);
11011 #endif
11012 }
11013 
11014 /*
11015  * Note:  called from probe context.  This function just increments the drop
11016  * count on a buffer.  It has been made a function to allow for the
11017  * possibility of understanding the source of mysterious drop counts.  (A
11018  * problem for which one may be particularly disappointed that DTrace cannot
11019  * be used to understand DTrace.)
11020  */
11021 static void
11022 dtrace_buffer_drop(dtrace_buffer_t *buf)
11023 {
11024 	buf->dtb_drops++;
11025 }
11026 
11027 /*
11028  * Note:  called from probe context.  This function is called to reserve space
11029  * in a buffer.  If mstate is non-NULL, sets the scratch base and size in the
11030  * mstate.  Returns the new offset in the buffer, or a negative value if an
11031  * error has occurred.
11032  */
11033 static intptr_t
11034 dtrace_buffer_reserve(dtrace_buffer_t *buf, size_t needed, size_t align,
11035     dtrace_state_t *state, dtrace_mstate_t *mstate)
11036 {
11037 	intptr_t offs = buf->dtb_offset, soffs;
11038 	intptr_t woffs;
11039 	caddr_t tomax;
11040 	size_t total;
11041 
11042 	if (buf->dtb_flags & DTRACEBUF_INACTIVE)
11043 		return (-1);
11044 
11045 	if ((tomax = buf->dtb_tomax) == NULL) {
11046 		dtrace_buffer_drop(buf);
11047 		return (-1);
11048 	}
11049 
11050 	if (!(buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL))) {
11051 		while (offs & (align - 1)) {
11052 			/*
11053 			 * Assert that our alignment is off by a number which
11054 			 * is itself sizeof (uint32_t) aligned.
11055 			 */
11056 			ASSERT(!((align - (offs & (align - 1))) &
11057 			    (sizeof (uint32_t) - 1)));
11058 			DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
11059 			offs += sizeof (uint32_t);
11060 		}
11061 
11062 		if ((soffs = offs + needed) > buf->dtb_size) {
11063 			dtrace_buffer_drop(buf);
11064 			return (-1);
11065 		}
11066 
11067 		if (mstate == NULL)
11068 			return (offs);
11069 
11070 		mstate->dtms_scratch_base = (uintptr_t)tomax + soffs;
11071 		mstate->dtms_scratch_size = buf->dtb_size - soffs;
11072 		mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
11073 
11074 		return (offs);
11075 	}
11076 
11077 	if (buf->dtb_flags & DTRACEBUF_FILL) {
11078 		if (state->dts_activity != DTRACE_ACTIVITY_COOLDOWN &&
11079 		    (buf->dtb_flags & DTRACEBUF_FULL))
11080 			return (-1);
11081 		goto out;
11082 	}
11083 
11084 	total = needed + (offs & (align - 1));
11085 
11086 	/*
11087 	 * For a ring buffer, life is quite a bit more complicated.  Before
11088 	 * we can store any padding, we need to adjust our wrapping offset.
11089 	 * (If we've never before wrapped or we're not about to, no adjustment
11090 	 * is required.)
11091 	 */
11092 	if ((buf->dtb_flags & DTRACEBUF_WRAPPED) ||
11093 	    offs + total > buf->dtb_size) {
11094 		woffs = buf->dtb_xamot_offset;
11095 
11096 		if (offs + total > buf->dtb_size) {
11097 			/*
11098 			 * We can't fit in the end of the buffer.  First, a
11099 			 * sanity check that we can fit in the buffer at all.
11100 			 */
11101 			if (total > buf->dtb_size) {
11102 				dtrace_buffer_drop(buf);
11103 				return (-1);
11104 			}
11105 
11106 			/*
11107 			 * We're going to be storing at the top of the buffer,
11108 			 * so now we need to deal with the wrapped offset.  We
11109 			 * only reset our wrapped offset to 0 if it is
11110 			 * currently greater than the current offset.  If it
11111 			 * is less than the current offset, it is because a
11112 			 * previous allocation induced a wrap -- but the
11113 			 * allocation didn't subsequently take the space due
11114 			 * to an error or false predicate evaluation.  In this
11115 			 * case, we'll just leave the wrapped offset alone: if
11116 			 * the wrapped offset hasn't been advanced far enough
11117 			 * for this allocation, it will be adjusted in the
11118 			 * lower loop.
11119 			 */
11120 			if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
11121 				if (woffs >= offs)
11122 					woffs = 0;
11123 			} else {
11124 				woffs = 0;
11125 			}
11126 
11127 			/*
11128 			 * Now we know that we're going to be storing to the
11129 			 * top of the buffer and that there is room for us
11130 			 * there.  We need to clear the buffer from the current
11131 			 * offset to the end (there may be old gunk there).
11132 			 */
11133 			while (offs < buf->dtb_size)
11134 				tomax[offs++] = 0;
11135 
11136 			/*
11137 			 * We need to set our offset to zero.  And because we
11138 			 * are wrapping, we need to set the bit indicating as
11139 			 * much.  We can also adjust our needed space back
11140 			 * down to the space required by the ECB -- we know
11141 			 * that the top of the buffer is aligned.
11142 			 */
11143 			offs = 0;
11144 			total = needed;
11145 			buf->dtb_flags |= DTRACEBUF_WRAPPED;
11146 		} else {
11147 			/*
11148 			 * There is room for us in the buffer, so we simply
11149 			 * need to check the wrapped offset.
11150 			 */
11151 			if (woffs < offs) {
11152 				/*
11153 				 * The wrapped offset is less than the offset.
11154 				 * This can happen if we allocated buffer space
11155 				 * that induced a wrap, but then we didn't
11156 				 * subsequently take the space due to an error
11157 				 * or false predicate evaluation.  This is
11158 				 * okay; we know that _this_ allocation isn't
11159 				 * going to induce a wrap.  We still can't
11160 				 * reset the wrapped offset to be zero,
11161 				 * however: the space may have been trashed in
11162 				 * the previous failed probe attempt.  But at
11163 				 * least the wrapped offset doesn't need to
11164 				 * be adjusted at all...
11165 				 */
11166 				goto out;
11167 			}
11168 		}
11169 
11170 		while (offs + total > woffs) {
11171 			dtrace_epid_t epid = *(uint32_t *)(tomax + woffs);
11172 			size_t size;
11173 
11174 			if (epid == DTRACE_EPIDNONE) {
11175 				size = sizeof (uint32_t);
11176 			} else {
11177 				ASSERT3U(epid, <=, state->dts_necbs);
11178 				ASSERT(state->dts_ecbs[epid - 1] != NULL);
11179 
11180 				size = state->dts_ecbs[epid - 1]->dte_size;
11181 			}
11182 
11183 			ASSERT(woffs + size <= buf->dtb_size);
11184 			ASSERT(size != 0);
11185 
11186 			if (woffs + size == buf->dtb_size) {
11187 				/*
11188 				 * We've reached the end of the buffer; we want
11189 				 * to set the wrapped offset to 0 and break
11190 				 * out.  However, if the offs is 0, then we're
11191 				 * in a strange edge-condition:  the amount of
11192 				 * space that we want to reserve plus the size
11193 				 * of the record that we're overwriting is
11194 				 * greater than the size of the buffer.  This
11195 				 * is problematic because if we reserve the
11196 				 * space but subsequently don't consume it (due
11197 				 * to a failed predicate or error) the wrapped
11198 				 * offset will be 0 -- yet the EPID at offset 0
11199 				 * will not be committed.  This situation is
11200 				 * relatively easy to deal with:  if we're in
11201 				 * this case, the buffer is indistinguishable
11202 				 * from one that hasn't wrapped; we need only
11203 				 * finish the job by clearing the wrapped bit,
11204 				 * explicitly setting the offset to be 0, and
11205 				 * zero'ing out the old data in the buffer.
11206 				 */
11207 				if (offs == 0) {
11208 					buf->dtb_flags &= ~DTRACEBUF_WRAPPED;
11209 					buf->dtb_offset = 0;
11210 					woffs = total;
11211 
11212 					while (woffs < buf->dtb_size)
11213 						tomax[woffs++] = 0;
11214 				}
11215 
11216 				woffs = 0;
11217 				break;
11218 			}
11219 
11220 			woffs += size;
11221 		}
11222 
11223 		/*
11224 		 * We have a wrapped offset.  It may be that the wrapped offset
11225 		 * has become zero -- that's okay.
11226 		 */
11227 		buf->dtb_xamot_offset = woffs;
11228 	}
11229 
11230 out:
11231 	/*
11232 	 * Now we can plow the buffer with any necessary padding.
11233 	 */
11234 	while (offs & (align - 1)) {
11235 		/*
11236 		 * Assert that our alignment is off by a number which
11237 		 * is itself sizeof (uint32_t) aligned.
11238 		 */
11239 		ASSERT(!((align - (offs & (align - 1))) &
11240 		    (sizeof (uint32_t) - 1)));
11241 		DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
11242 		offs += sizeof (uint32_t);
11243 	}
11244 
11245 	if (buf->dtb_flags & DTRACEBUF_FILL) {
11246 		if (offs + needed > buf->dtb_size - state->dts_reserve) {
11247 			buf->dtb_flags |= DTRACEBUF_FULL;
11248 			return (-1);
11249 		}
11250 	}
11251 
11252 	if (mstate == NULL)
11253 		return (offs);
11254 
11255 	/*
11256 	 * For ring buffers and fill buffers, the scratch space is always
11257 	 * the inactive buffer.
11258 	 */
11259 	mstate->dtms_scratch_base = (uintptr_t)buf->dtb_xamot;
11260 	mstate->dtms_scratch_size = buf->dtb_size;
11261 	mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
11262 
11263 	return (offs);
11264 }
11265 
11266 static void
11267 dtrace_buffer_polish(dtrace_buffer_t *buf)
11268 {
11269 	ASSERT(buf->dtb_flags & DTRACEBUF_RING);
11270 	ASSERT(MUTEX_HELD(&dtrace_lock));
11271 
11272 	if (!(buf->dtb_flags & DTRACEBUF_WRAPPED))
11273 		return;
11274 
11275 	/*
11276 	 * We need to polish the ring buffer.  There are three cases:
11277 	 *
11278 	 * - The first (and presumably most common) is that there is no gap
11279 	 *   between the buffer offset and the wrapped offset.  In this case,
11280 	 *   there is nothing in the buffer that isn't valid data; we can
11281 	 *   mark the buffer as polished and return.
11282 	 *
11283 	 * - The second (less common than the first but still more common
11284 	 *   than the third) is that there is a gap between the buffer offset
11285 	 *   and the wrapped offset, and the wrapped offset is larger than the
11286 	 *   buffer offset.  This can happen because of an alignment issue, or
11287 	 *   can happen because of a call to dtrace_buffer_reserve() that
11288 	 *   didn't subsequently consume the buffer space.  In this case,
11289 	 *   we need to zero the data from the buffer offset to the wrapped
11290 	 *   offset.
11291 	 *
11292 	 * - The third (and least common) is that there is a gap between the
11293 	 *   buffer offset and the wrapped offset, but the wrapped offset is
11294 	 *   _less_ than the buffer offset.  This can only happen because a
11295 	 *   call to dtrace_buffer_reserve() induced a wrap, but the space
11296 	 *   was not subsequently consumed.  In this case, we need to zero the
11297 	 *   space from the offset to the end of the buffer _and_ from the
11298 	 *   top of the buffer to the wrapped offset.
11299 	 */
11300 	if (buf->dtb_offset < buf->dtb_xamot_offset) {
11301 		bzero(buf->dtb_tomax + buf->dtb_offset,
11302 		    buf->dtb_xamot_offset - buf->dtb_offset);
11303 	}
11304 
11305 	if (buf->dtb_offset > buf->dtb_xamot_offset) {
11306 		bzero(buf->dtb_tomax + buf->dtb_offset,
11307 		    buf->dtb_size - buf->dtb_offset);
11308 		bzero(buf->dtb_tomax, buf->dtb_xamot_offset);
11309 	}
11310 }
11311 
11312 /*
11313  * This routine determines if data generated at the specified time has likely
11314  * been entirely consumed at user-level.  This routine is called to determine
11315  * if an ECB on a defunct probe (but for an active enabling) can be safely
11316  * disabled and destroyed.
11317  */
11318 static int
11319 dtrace_buffer_consumed(dtrace_buffer_t *bufs, hrtime_t when)
11320 {
11321 	int i;
11322 
11323 	for (i = 0; i < NCPU; i++) {
11324 		dtrace_buffer_t *buf = &bufs[i];
11325 
11326 		if (buf->dtb_size == 0)
11327 			continue;
11328 
11329 		if (buf->dtb_flags & DTRACEBUF_RING)
11330 			return (0);
11331 
11332 		if (!buf->dtb_switched && buf->dtb_offset != 0)
11333 			return (0);
11334 
11335 		if (buf->dtb_switched - buf->dtb_interval < when)
11336 			return (0);
11337 	}
11338 
11339 	return (1);
11340 }
11341 
11342 static void
11343 dtrace_buffer_free(dtrace_buffer_t *bufs)
11344 {
11345 	int i;
11346 
11347 	for (i = 0; i < NCPU; i++) {
11348 		dtrace_buffer_t *buf = &bufs[i];
11349 
11350 		if (buf->dtb_tomax == NULL) {
11351 			ASSERT(buf->dtb_xamot == NULL);
11352 			ASSERT(buf->dtb_size == 0);
11353 			continue;
11354 		}
11355 
11356 		if (buf->dtb_xamot != NULL) {
11357 			ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
11358 			kmem_free(buf->dtb_xamot, buf->dtb_size);
11359 		}
11360 
11361 		kmem_free(buf->dtb_tomax, buf->dtb_size);
11362 		buf->dtb_size = 0;
11363 		buf->dtb_tomax = NULL;
11364 		buf->dtb_xamot = NULL;
11365 	}
11366 }
11367 
11368 /*
11369  * DTrace Enabling Functions
11370  */
11371 static dtrace_enabling_t *
11372 dtrace_enabling_create(dtrace_vstate_t *vstate)
11373 {
11374 	dtrace_enabling_t *enab;
11375 
11376 	enab = kmem_zalloc(sizeof (dtrace_enabling_t), KM_SLEEP);
11377 	enab->dten_vstate = vstate;
11378 
11379 	return (enab);
11380 }
11381 
11382 static void
11383 dtrace_enabling_add(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb)
11384 {
11385 	dtrace_ecbdesc_t **ndesc;
11386 	size_t osize, nsize;
11387 
11388 	/*
11389 	 * We can't add to enablings after we've enabled them, or after we've
11390 	 * retained them.
11391 	 */
11392 	ASSERT(enab->dten_probegen == 0);
11393 	ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
11394 
11395 	if (enab->dten_ndesc < enab->dten_maxdesc) {
11396 		enab->dten_desc[enab->dten_ndesc++] = ecb;
11397 		return;
11398 	}
11399 
11400 	osize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
11401 
11402 	if (enab->dten_maxdesc == 0) {
11403 		enab->dten_maxdesc = 1;
11404 	} else {
11405 		enab->dten_maxdesc <<= 1;
11406 	}
11407 
11408 	ASSERT(enab->dten_ndesc < enab->dten_maxdesc);
11409 
11410 	nsize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
11411 	ndesc = kmem_zalloc(nsize, KM_SLEEP);
11412 	bcopy(enab->dten_desc, ndesc, osize);
11413 	if (enab->dten_desc != NULL)
11414 		kmem_free(enab->dten_desc, osize);
11415 
11416 	enab->dten_desc = ndesc;
11417 	enab->dten_desc[enab->dten_ndesc++] = ecb;
11418 }
11419 
11420 static void
11421 dtrace_enabling_addlike(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb,
11422     dtrace_probedesc_t *pd)
11423 {
11424 	dtrace_ecbdesc_t *new;
11425 	dtrace_predicate_t *pred;
11426 	dtrace_actdesc_t *act;
11427 
11428 	/*
11429 	 * We're going to create a new ECB description that matches the
11430 	 * specified ECB in every way, but has the specified probe description.
11431 	 */
11432 	new = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
11433 
11434 	if ((pred = ecb->dted_pred.dtpdd_predicate) != NULL)
11435 		dtrace_predicate_hold(pred);
11436 
11437 	for (act = ecb->dted_action; act != NULL; act = act->dtad_next)
11438 		dtrace_actdesc_hold(act);
11439 
11440 	new->dted_action = ecb->dted_action;
11441 	new->dted_pred = ecb->dted_pred;
11442 	new->dted_probe = *pd;
11443 	new->dted_uarg = ecb->dted_uarg;
11444 
11445 	dtrace_enabling_add(enab, new);
11446 }
11447 
11448 static void
11449 dtrace_enabling_dump(dtrace_enabling_t *enab)
11450 {
11451 	int i;
11452 
11453 	for (i = 0; i < enab->dten_ndesc; i++) {
11454 		dtrace_probedesc_t *desc = &enab->dten_desc[i]->dted_probe;
11455 
11456 		cmn_err(CE_NOTE, "enabling probe %d (%s:%s:%s:%s)", i,
11457 		    desc->dtpd_provider, desc->dtpd_mod,
11458 		    desc->dtpd_func, desc->dtpd_name);
11459 	}
11460 }
11461 
11462 static void
11463 dtrace_enabling_destroy(dtrace_enabling_t *enab)
11464 {
11465 	int i;
11466 	dtrace_ecbdesc_t *ep;
11467 	dtrace_vstate_t *vstate = enab->dten_vstate;
11468 
11469 	ASSERT(MUTEX_HELD(&dtrace_lock));
11470 
11471 	for (i = 0; i < enab->dten_ndesc; i++) {
11472 		dtrace_actdesc_t *act, *next;
11473 		dtrace_predicate_t *pred;
11474 
11475 		ep = enab->dten_desc[i];
11476 
11477 		if ((pred = ep->dted_pred.dtpdd_predicate) != NULL)
11478 			dtrace_predicate_release(pred, vstate);
11479 
11480 		for (act = ep->dted_action; act != NULL; act = next) {
11481 			next = act->dtad_next;
11482 			dtrace_actdesc_release(act, vstate);
11483 		}
11484 
11485 		kmem_free(ep, sizeof (dtrace_ecbdesc_t));
11486 	}
11487 
11488 	if (enab->dten_desc != NULL)
11489 		kmem_free(enab->dten_desc,
11490 		    enab->dten_maxdesc * sizeof (dtrace_enabling_t *));
11491 
11492 	/*
11493 	 * If this was a retained enabling, decrement the dts_nretained count
11494 	 * and take it off of the dtrace_retained list.
11495 	 */
11496 	if (enab->dten_prev != NULL || enab->dten_next != NULL ||
11497 	    dtrace_retained == enab) {
11498 		ASSERT(enab->dten_vstate->dtvs_state != NULL);
11499 		ASSERT(enab->dten_vstate->dtvs_state->dts_nretained > 0);
11500 		enab->dten_vstate->dtvs_state->dts_nretained--;
11501 	}
11502 
11503 	if (enab->dten_prev == NULL) {
11504 		if (dtrace_retained == enab) {
11505 			dtrace_retained = enab->dten_next;
11506 
11507 			if (dtrace_retained != NULL)
11508 				dtrace_retained->dten_prev = NULL;
11509 		}
11510 	} else {
11511 		ASSERT(enab != dtrace_retained);
11512 		ASSERT(dtrace_retained != NULL);
11513 		enab->dten_prev->dten_next = enab->dten_next;
11514 	}
11515 
11516 	if (enab->dten_next != NULL) {
11517 		ASSERT(dtrace_retained != NULL);
11518 		enab->dten_next->dten_prev = enab->dten_prev;
11519 	}
11520 
11521 	kmem_free(enab, sizeof (dtrace_enabling_t));
11522 }
11523 
11524 static int
11525 dtrace_enabling_retain(dtrace_enabling_t *enab)
11526 {
11527 	dtrace_state_t *state;
11528 
11529 	ASSERT(MUTEX_HELD(&dtrace_lock));
11530 	ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
11531 	ASSERT(enab->dten_vstate != NULL);
11532 
11533 	state = enab->dten_vstate->dtvs_state;
11534 	ASSERT(state != NULL);
11535 
11536 	/*
11537 	 * We only allow each state to retain dtrace_retain_max enablings.
11538 	 */
11539 	if (state->dts_nretained >= dtrace_retain_max)
11540 		return (ENOSPC);
11541 
11542 	state->dts_nretained++;
11543 
11544 	if (dtrace_retained == NULL) {
11545 		dtrace_retained = enab;
11546 		return (0);
11547 	}
11548 
11549 	enab->dten_next = dtrace_retained;
11550 	dtrace_retained->dten_prev = enab;
11551 	dtrace_retained = enab;
11552 
11553 	return (0);
11554 }
11555 
11556 static int
11557 dtrace_enabling_replicate(dtrace_state_t *state, dtrace_probedesc_t *match,
11558     dtrace_probedesc_t *create)
11559 {
11560 	dtrace_enabling_t *new, *enab;
11561 	int found = 0, err = ENOENT;
11562 
11563 	ASSERT(MUTEX_HELD(&dtrace_lock));
11564 	ASSERT(strlen(match->dtpd_provider) < DTRACE_PROVNAMELEN);
11565 	ASSERT(strlen(match->dtpd_mod) < DTRACE_MODNAMELEN);
11566 	ASSERT(strlen(match->dtpd_func) < DTRACE_FUNCNAMELEN);
11567 	ASSERT(strlen(match->dtpd_name) < DTRACE_NAMELEN);
11568 
11569 	new = dtrace_enabling_create(&state->dts_vstate);
11570 
11571 	/*
11572 	 * Iterate over all retained enablings, looking for enablings that
11573 	 * match the specified state.
11574 	 */
11575 	for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11576 		int i;
11577 
11578 		/*
11579 		 * dtvs_state can only be NULL for helper enablings -- and
11580 		 * helper enablings can't be retained.
11581 		 */
11582 		ASSERT(enab->dten_vstate->dtvs_state != NULL);
11583 
11584 		if (enab->dten_vstate->dtvs_state != state)
11585 			continue;
11586 
11587 		/*
11588 		 * Now iterate over each probe description; we're looking for
11589 		 * an exact match to the specified probe description.
11590 		 */
11591 		for (i = 0; i < enab->dten_ndesc; i++) {
11592 			dtrace_ecbdesc_t *ep = enab->dten_desc[i];
11593 			dtrace_probedesc_t *pd = &ep->dted_probe;
11594 
11595 			if (strcmp(pd->dtpd_provider, match->dtpd_provider))
11596 				continue;
11597 
11598 			if (strcmp(pd->dtpd_mod, match->dtpd_mod))
11599 				continue;
11600 
11601 			if (strcmp(pd->dtpd_func, match->dtpd_func))
11602 				continue;
11603 
11604 			if (strcmp(pd->dtpd_name, match->dtpd_name))
11605 				continue;
11606 
11607 			/*
11608 			 * We have a winning probe!  Add it to our growing
11609 			 * enabling.
11610 			 */
11611 			found = 1;
11612 			dtrace_enabling_addlike(new, ep, create);
11613 		}
11614 	}
11615 
11616 	if (!found || (err = dtrace_enabling_retain(new)) != 0) {
11617 		dtrace_enabling_destroy(new);
11618 		return (err);
11619 	}
11620 
11621 	return (0);
11622 }
11623 
11624 static void
11625 dtrace_enabling_retract(dtrace_state_t *state)
11626 {
11627 	dtrace_enabling_t *enab, *next;
11628 
11629 	ASSERT(MUTEX_HELD(&dtrace_lock));
11630 
11631 	/*
11632 	 * Iterate over all retained enablings, destroy the enablings retained
11633 	 * for the specified state.
11634 	 */
11635 	for (enab = dtrace_retained; enab != NULL; enab = next) {
11636 		next = enab->dten_next;
11637 
11638 		/*
11639 		 * dtvs_state can only be NULL for helper enablings -- and
11640 		 * helper enablings can't be retained.
11641 		 */
11642 		ASSERT(enab->dten_vstate->dtvs_state != NULL);
11643 
11644 		if (enab->dten_vstate->dtvs_state == state) {
11645 			ASSERT(state->dts_nretained > 0);
11646 			dtrace_enabling_destroy(enab);
11647 		}
11648 	}
11649 
11650 	ASSERT(state->dts_nretained == 0);
11651 }
11652 
11653 static int
11654 dtrace_enabling_match(dtrace_enabling_t *enab, int *nmatched)
11655 {
11656 	int i = 0;
11657 	int matched = 0;
11658 
11659 	ASSERT(MUTEX_HELD(&cpu_lock));
11660 	ASSERT(MUTEX_HELD(&dtrace_lock));
11661 
11662 	for (i = 0; i < enab->dten_ndesc; i++) {
11663 		dtrace_ecbdesc_t *ep = enab->dten_desc[i];
11664 
11665 		enab->dten_current = ep;
11666 		enab->dten_error = 0;
11667 
11668 		matched += dtrace_probe_enable(&ep->dted_probe, enab);
11669 
11670 		if (enab->dten_error != 0) {
11671 			/*
11672 			 * If we get an error half-way through enabling the
11673 			 * probes, we kick out -- perhaps with some number of
11674 			 * them enabled.  Leaving enabled probes enabled may
11675 			 * be slightly confusing for user-level, but we expect
11676 			 * that no one will attempt to actually drive on in
11677 			 * the face of such errors.  If this is an anonymous
11678 			 * enabling (indicated with a NULL nmatched pointer),
11679 			 * we cmn_err() a message.  We aren't expecting to
11680 			 * get such an error -- such as it can exist at all,
11681 			 * it would be a result of corrupted DOF in the driver
11682 			 * properties.
11683 			 */
11684 			if (nmatched == NULL) {
11685 				cmn_err(CE_WARN, "dtrace_enabling_match() "
11686 				    "error on %p: %d", (void *)ep,
11687 				    enab->dten_error);
11688 			}
11689 
11690 			return (enab->dten_error);
11691 		}
11692 	}
11693 
11694 	enab->dten_probegen = dtrace_probegen;
11695 	if (nmatched != NULL)
11696 		*nmatched = matched;
11697 
11698 	return (0);
11699 }
11700 
11701 static void
11702 dtrace_enabling_matchall(void)
11703 {
11704 	dtrace_enabling_t *enab;
11705 
11706 	mutex_enter(&cpu_lock);
11707 	mutex_enter(&dtrace_lock);
11708 
11709 	/*
11710 	 * Iterate over all retained enablings to see if any probes match
11711 	 * against them.  We only perform this operation on enablings for which
11712 	 * we have sufficient permissions by virtue of being in the global zone
11713 	 * or in the same zone as the DTrace client.  Because we can be called
11714 	 * after dtrace_detach() has been called, we cannot assert that there
11715 	 * are retained enablings.  We can safely load from dtrace_retained,
11716 	 * however:  the taskq_destroy() at the end of dtrace_detach() will
11717 	 * block pending our completion.
11718 	 */
11719 	for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11720 #if defined(sun)
11721 		cred_t *cr = enab->dten_vstate->dtvs_state->dts_cred.dcr_cred;
11722 
11723 		if (INGLOBALZONE(curproc) || getzoneid() == crgetzoneid(cr))
11724 #endif
11725 			(void) dtrace_enabling_match(enab, NULL);
11726 	}
11727 
11728 	mutex_exit(&dtrace_lock);
11729 	mutex_exit(&cpu_lock);
11730 }
11731 
11732 /*
11733  * If an enabling is to be enabled without having matched probes (that is, if
11734  * dtrace_state_go() is to be called on the underlying dtrace_state_t), the
11735  * enabling must be _primed_ by creating an ECB for every ECB description.
11736  * This must be done to assure that we know the number of speculations, the
11737  * number of aggregations, the minimum buffer size needed, etc. before we
11738  * transition out of DTRACE_ACTIVITY_INACTIVE.  To do this without actually
11739  * enabling any probes, we create ECBs for every ECB decription, but with a
11740  * NULL probe -- which is exactly what this function does.
11741  */
11742 static void
11743 dtrace_enabling_prime(dtrace_state_t *state)
11744 {
11745 	dtrace_enabling_t *enab;
11746 	int i;
11747 
11748 	for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11749 		ASSERT(enab->dten_vstate->dtvs_state != NULL);
11750 
11751 		if (enab->dten_vstate->dtvs_state != state)
11752 			continue;
11753 
11754 		/*
11755 		 * We don't want to prime an enabling more than once, lest
11756 		 * we allow a malicious user to induce resource exhaustion.
11757 		 * (The ECBs that result from priming an enabling aren't
11758 		 * leaked -- but they also aren't deallocated until the
11759 		 * consumer state is destroyed.)
11760 		 */
11761 		if (enab->dten_primed)
11762 			continue;
11763 
11764 		for (i = 0; i < enab->dten_ndesc; i++) {
11765 			enab->dten_current = enab->dten_desc[i];
11766 			(void) dtrace_probe_enable(NULL, enab);
11767 		}
11768 
11769 		enab->dten_primed = 1;
11770 	}
11771 }
11772 
11773 /*
11774  * Called to indicate that probes should be provided due to retained
11775  * enablings.  This is implemented in terms of dtrace_probe_provide(), but it
11776  * must take an initial lap through the enabling calling the dtps_provide()
11777  * entry point explicitly to allow for autocreated probes.
11778  */
11779 static void
11780 dtrace_enabling_provide(dtrace_provider_t *prv)
11781 {
11782 	int i, all = 0;
11783 	dtrace_probedesc_t desc;
11784 
11785 	ASSERT(MUTEX_HELD(&dtrace_lock));
11786 	ASSERT(MUTEX_HELD(&dtrace_provider_lock));
11787 
11788 	if (prv == NULL) {
11789 		all = 1;
11790 		prv = dtrace_provider;
11791 	}
11792 
11793 	do {
11794 		dtrace_enabling_t *enab = dtrace_retained;
11795 		void *parg = prv->dtpv_arg;
11796 
11797 		for (; enab != NULL; enab = enab->dten_next) {
11798 			for (i = 0; i < enab->dten_ndesc; i++) {
11799 				desc = enab->dten_desc[i]->dted_probe;
11800 				mutex_exit(&dtrace_lock);
11801 				prv->dtpv_pops.dtps_provide(parg, &desc);
11802 				mutex_enter(&dtrace_lock);
11803 			}
11804 		}
11805 	} while (all && (prv = prv->dtpv_next) != NULL);
11806 
11807 	mutex_exit(&dtrace_lock);
11808 	dtrace_probe_provide(NULL, all ? NULL : prv);
11809 	mutex_enter(&dtrace_lock);
11810 }
11811 
11812 /*
11813  * Called to reap ECBs that are attached to probes from defunct providers.
11814  */
11815 static void
11816 dtrace_enabling_reap(void)
11817 {
11818 	dtrace_provider_t *prov;
11819 	dtrace_probe_t *probe;
11820 	dtrace_ecb_t *ecb;
11821 	hrtime_t when;
11822 	int i;
11823 
11824 	mutex_enter(&cpu_lock);
11825 	mutex_enter(&dtrace_lock);
11826 
11827 	for (i = 0; i < dtrace_nprobes; i++) {
11828 		if ((probe = dtrace_probes[i]) == NULL)
11829 			continue;
11830 
11831 		if (probe->dtpr_ecb == NULL)
11832 			continue;
11833 
11834 		prov = probe->dtpr_provider;
11835 
11836 		if ((when = prov->dtpv_defunct) == 0)
11837 			continue;
11838 
11839 		/*
11840 		 * We have ECBs on a defunct provider:  we want to reap these
11841 		 * ECBs to allow the provider to unregister.  The destruction
11842 		 * of these ECBs must be done carefully:  if we destroy the ECB
11843 		 * and the consumer later wishes to consume an EPID that
11844 		 * corresponds to the destroyed ECB (and if the EPID metadata
11845 		 * has not been previously consumed), the consumer will abort
11846 		 * processing on the unknown EPID.  To reduce (but not, sadly,
11847 		 * eliminate) the possibility of this, we will only destroy an
11848 		 * ECB for a defunct provider if, for the state that
11849 		 * corresponds to the ECB:
11850 		 *
11851 		 *  (a)	There is no speculative tracing (which can effectively
11852 		 *	cache an EPID for an arbitrary amount of time).
11853 		 *
11854 		 *  (b)	The principal buffers have been switched twice since the
11855 		 *	provider became defunct.
11856 		 *
11857 		 *  (c)	The aggregation buffers are of zero size or have been
11858 		 *	switched twice since the provider became defunct.
11859 		 *
11860 		 * We use dts_speculates to determine (a) and call a function
11861 		 * (dtrace_buffer_consumed()) to determine (b) and (c).  Note
11862 		 * that as soon as we've been unable to destroy one of the ECBs
11863 		 * associated with the probe, we quit trying -- reaping is only
11864 		 * fruitful in as much as we can destroy all ECBs associated
11865 		 * with the defunct provider's probes.
11866 		 */
11867 		while ((ecb = probe->dtpr_ecb) != NULL) {
11868 			dtrace_state_t *state = ecb->dte_state;
11869 			dtrace_buffer_t *buf = state->dts_buffer;
11870 			dtrace_buffer_t *aggbuf = state->dts_aggbuffer;
11871 
11872 			if (state->dts_speculates)
11873 				break;
11874 
11875 			if (!dtrace_buffer_consumed(buf, when))
11876 				break;
11877 
11878 			if (!dtrace_buffer_consumed(aggbuf, when))
11879 				break;
11880 
11881 			dtrace_ecb_disable(ecb);
11882 			ASSERT(probe->dtpr_ecb != ecb);
11883 			dtrace_ecb_destroy(ecb);
11884 		}
11885 	}
11886 
11887 	mutex_exit(&dtrace_lock);
11888 	mutex_exit(&cpu_lock);
11889 }
11890 
11891 /*
11892  * DTrace DOF Functions
11893  */
11894 /*ARGSUSED*/
11895 static void
11896 dtrace_dof_error(dof_hdr_t *dof, const char *str)
11897 {
11898 	if (dtrace_err_verbose)
11899 		cmn_err(CE_WARN, "failed to process DOF: %s", str);
11900 
11901 #ifdef DTRACE_ERRDEBUG
11902 	dtrace_errdebug(str);
11903 #endif
11904 }
11905 
11906 /*
11907  * Create DOF out of a currently enabled state.  Right now, we only create
11908  * DOF containing the run-time options -- but this could be expanded to create
11909  * complete DOF representing the enabled state.
11910  */
11911 static dof_hdr_t *
11912 dtrace_dof_create(dtrace_state_t *state)
11913 {
11914 	dof_hdr_t *dof;
11915 	dof_sec_t *sec;
11916 	dof_optdesc_t *opt;
11917 	int i, len = sizeof (dof_hdr_t) +
11918 	    roundup(sizeof (dof_sec_t), sizeof (uint64_t)) +
11919 	    sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
11920 
11921 	ASSERT(MUTEX_HELD(&dtrace_lock));
11922 
11923 	dof = kmem_zalloc(len, KM_SLEEP);
11924 	dof->dofh_ident[DOF_ID_MAG0] = DOF_MAG_MAG0;
11925 	dof->dofh_ident[DOF_ID_MAG1] = DOF_MAG_MAG1;
11926 	dof->dofh_ident[DOF_ID_MAG2] = DOF_MAG_MAG2;
11927 	dof->dofh_ident[DOF_ID_MAG3] = DOF_MAG_MAG3;
11928 
11929 	dof->dofh_ident[DOF_ID_MODEL] = DOF_MODEL_NATIVE;
11930 	dof->dofh_ident[DOF_ID_ENCODING] = DOF_ENCODE_NATIVE;
11931 	dof->dofh_ident[DOF_ID_VERSION] = DOF_VERSION;
11932 	dof->dofh_ident[DOF_ID_DIFVERS] = DIF_VERSION;
11933 	dof->dofh_ident[DOF_ID_DIFIREG] = DIF_DIR_NREGS;
11934 	dof->dofh_ident[DOF_ID_DIFTREG] = DIF_DTR_NREGS;
11935 
11936 	dof->dofh_flags = 0;
11937 	dof->dofh_hdrsize = sizeof (dof_hdr_t);
11938 	dof->dofh_secsize = sizeof (dof_sec_t);
11939 	dof->dofh_secnum = 1;	/* only DOF_SECT_OPTDESC */
11940 	dof->dofh_secoff = sizeof (dof_hdr_t);
11941 	dof->dofh_loadsz = len;
11942 	dof->dofh_filesz = len;
11943 	dof->dofh_pad = 0;
11944 
11945 	/*
11946 	 * Fill in the option section header...
11947 	 */
11948 	sec = (dof_sec_t *)((uintptr_t)dof + sizeof (dof_hdr_t));
11949 	sec->dofs_type = DOF_SECT_OPTDESC;
11950 	sec->dofs_align = sizeof (uint64_t);
11951 	sec->dofs_flags = DOF_SECF_LOAD;
11952 	sec->dofs_entsize = sizeof (dof_optdesc_t);
11953 
11954 	opt = (dof_optdesc_t *)((uintptr_t)sec +
11955 	    roundup(sizeof (dof_sec_t), sizeof (uint64_t)));
11956 
11957 	sec->dofs_offset = (uintptr_t)opt - (uintptr_t)dof;
11958 	sec->dofs_size = sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
11959 
11960 	for (i = 0; i < DTRACEOPT_MAX; i++) {
11961 		opt[i].dofo_option = i;
11962 		opt[i].dofo_strtab = DOF_SECIDX_NONE;
11963 		opt[i].dofo_value = state->dts_options[i];
11964 	}
11965 
11966 	return (dof);
11967 }
11968 
11969 static dof_hdr_t *
11970 dtrace_dof_copyin(uintptr_t uarg, int *errp)
11971 {
11972 	dof_hdr_t hdr, *dof;
11973 
11974 	ASSERT(!MUTEX_HELD(&dtrace_lock));
11975 
11976 	/*
11977 	 * First, we're going to copyin() the sizeof (dof_hdr_t).
11978 	 */
11979 	if (copyin((void *)uarg, &hdr, sizeof (hdr)) != 0) {
11980 		dtrace_dof_error(NULL, "failed to copyin DOF header");
11981 		*errp = EFAULT;
11982 		return (NULL);
11983 	}
11984 
11985 	/*
11986 	 * Now we'll allocate the entire DOF and copy it in -- provided
11987 	 * that the length isn't outrageous.
11988 	 */
11989 	if (hdr.dofh_loadsz >= dtrace_dof_maxsize) {
11990 		dtrace_dof_error(&hdr, "load size exceeds maximum");
11991 		*errp = E2BIG;
11992 		return (NULL);
11993 	}
11994 
11995 	if (hdr.dofh_loadsz < sizeof (hdr)) {
11996 		dtrace_dof_error(&hdr, "invalid load size");
11997 		*errp = EINVAL;
11998 		return (NULL);
11999 	}
12000 
12001 	dof = kmem_alloc(hdr.dofh_loadsz, KM_SLEEP);
12002 
12003 	if (copyin((void *)uarg, dof, hdr.dofh_loadsz) != 0) {
12004 		kmem_free(dof, hdr.dofh_loadsz);
12005 		*errp = EFAULT;
12006 		return (NULL);
12007 	}
12008 
12009 	return (dof);
12010 }
12011 
12012 #if !defined(sun)
12013 static __inline uchar_t
12014 dtrace_dof_char(char c) {
12015 	switch (c) {
12016 	case '0':
12017 	case '1':
12018 	case '2':
12019 	case '3':
12020 	case '4':
12021 	case '5':
12022 	case '6':
12023 	case '7':
12024 	case '8':
12025 	case '9':
12026 		return (c - '0');
12027 	case 'A':
12028 	case 'B':
12029 	case 'C':
12030 	case 'D':
12031 	case 'E':
12032 	case 'F':
12033 		return (c - 'A' + 10);
12034 	case 'a':
12035 	case 'b':
12036 	case 'c':
12037 	case 'd':
12038 	case 'e':
12039 	case 'f':
12040 		return (c - 'a' + 10);
12041 	}
12042 	/* Should not reach here. */
12043 	return (0);
12044 }
12045 #endif
12046 
12047 static dof_hdr_t *
12048 dtrace_dof_property(const char *name)
12049 {
12050 	uchar_t *buf;
12051 	uint64_t loadsz;
12052 	unsigned int len, i;
12053 	dof_hdr_t *dof;
12054 
12055 #if defined(sun)
12056 	/*
12057 	 * Unfortunately, array of values in .conf files are always (and
12058 	 * only) interpreted to be integer arrays.  We must read our DOF
12059 	 * as an integer array, and then squeeze it into a byte array.
12060 	 */
12061 	if (ddi_prop_lookup_int_array(DDI_DEV_T_ANY, dtrace_devi, 0,
12062 	    (char *)name, (int **)&buf, &len) != DDI_PROP_SUCCESS)
12063 		return (NULL);
12064 
12065 	for (i = 0; i < len; i++)
12066 		buf[i] = (uchar_t)(((int *)buf)[i]);
12067 
12068 	if (len < sizeof (dof_hdr_t)) {
12069 		ddi_prop_free(buf);
12070 		dtrace_dof_error(NULL, "truncated header");
12071 		return (NULL);
12072 	}
12073 
12074 	if (len < (loadsz = ((dof_hdr_t *)buf)->dofh_loadsz)) {
12075 		ddi_prop_free(buf);
12076 		dtrace_dof_error(NULL, "truncated DOF");
12077 		return (NULL);
12078 	}
12079 
12080 	if (loadsz >= dtrace_dof_maxsize) {
12081 		ddi_prop_free(buf);
12082 		dtrace_dof_error(NULL, "oversized DOF");
12083 		return (NULL);
12084 	}
12085 
12086 	dof = kmem_alloc(loadsz, KM_SLEEP);
12087 	bcopy(buf, dof, loadsz);
12088 	ddi_prop_free(buf);
12089 #else
12090 	char *p;
12091 	char *p_env;
12092 
12093 	if ((p_env = getenv(name)) == NULL)
12094 		return (NULL);
12095 
12096 	len = strlen(p_env) / 2;
12097 
12098 	buf = kmem_alloc(len, KM_SLEEP);
12099 
12100 	dof = (dof_hdr_t *) buf;
12101 
12102 	p = p_env;
12103 
12104 	for (i = 0; i < len; i++) {
12105 		buf[i] = (dtrace_dof_char(p[0]) << 4) |
12106 		     dtrace_dof_char(p[1]);
12107 		p += 2;
12108 	}
12109 
12110 	freeenv(p_env);
12111 
12112 	if (len < sizeof (dof_hdr_t)) {
12113 		kmem_free(buf, 0);
12114 		dtrace_dof_error(NULL, "truncated header");
12115 		return (NULL);
12116 	}
12117 
12118 	if (len < (loadsz = dof->dofh_loadsz)) {
12119 		kmem_free(buf, 0);
12120 		dtrace_dof_error(NULL, "truncated DOF");
12121 		return (NULL);
12122 	}
12123 
12124 	if (loadsz >= dtrace_dof_maxsize) {
12125 		kmem_free(buf, 0);
12126 		dtrace_dof_error(NULL, "oversized DOF");
12127 		return (NULL);
12128 	}
12129 #endif
12130 
12131 	return (dof);
12132 }
12133 
12134 static void
12135 dtrace_dof_destroy(dof_hdr_t *dof)
12136 {
12137 	kmem_free(dof, dof->dofh_loadsz);
12138 }
12139 
12140 /*
12141  * Return the dof_sec_t pointer corresponding to a given section index.  If the
12142  * index is not valid, dtrace_dof_error() is called and NULL is returned.  If
12143  * a type other than DOF_SECT_NONE is specified, the header is checked against
12144  * this type and NULL is returned if the types do not match.
12145  */
12146 static dof_sec_t *
12147 dtrace_dof_sect(dof_hdr_t *dof, uint32_t type, dof_secidx_t i)
12148 {
12149 	dof_sec_t *sec = (dof_sec_t *)(uintptr_t)
12150 	    ((uintptr_t)dof + dof->dofh_secoff + i * dof->dofh_secsize);
12151 
12152 	if (i >= dof->dofh_secnum) {
12153 		dtrace_dof_error(dof, "referenced section index is invalid");
12154 		return (NULL);
12155 	}
12156 
12157 	if (!(sec->dofs_flags & DOF_SECF_LOAD)) {
12158 		dtrace_dof_error(dof, "referenced section is not loadable");
12159 		return (NULL);
12160 	}
12161 
12162 	if (type != DOF_SECT_NONE && type != sec->dofs_type) {
12163 		dtrace_dof_error(dof, "referenced section is the wrong type");
12164 		return (NULL);
12165 	}
12166 
12167 	return (sec);
12168 }
12169 
12170 static dtrace_probedesc_t *
12171 dtrace_dof_probedesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_probedesc_t *desc)
12172 {
12173 	dof_probedesc_t *probe;
12174 	dof_sec_t *strtab;
12175 	uintptr_t daddr = (uintptr_t)dof;
12176 	uintptr_t str;
12177 	size_t size;
12178 
12179 	if (sec->dofs_type != DOF_SECT_PROBEDESC) {
12180 		dtrace_dof_error(dof, "invalid probe section");
12181 		return (NULL);
12182 	}
12183 
12184 	if (sec->dofs_align != sizeof (dof_secidx_t)) {
12185 		dtrace_dof_error(dof, "bad alignment in probe description");
12186 		return (NULL);
12187 	}
12188 
12189 	if (sec->dofs_offset + sizeof (dof_probedesc_t) > dof->dofh_loadsz) {
12190 		dtrace_dof_error(dof, "truncated probe description");
12191 		return (NULL);
12192 	}
12193 
12194 	probe = (dof_probedesc_t *)(uintptr_t)(daddr + sec->dofs_offset);
12195 	strtab = dtrace_dof_sect(dof, DOF_SECT_STRTAB, probe->dofp_strtab);
12196 
12197 	if (strtab == NULL)
12198 		return (NULL);
12199 
12200 	str = daddr + strtab->dofs_offset;
12201 	size = strtab->dofs_size;
12202 
12203 	if (probe->dofp_provider >= strtab->dofs_size) {
12204 		dtrace_dof_error(dof, "corrupt probe provider");
12205 		return (NULL);
12206 	}
12207 
12208 	(void) strncpy(desc->dtpd_provider,
12209 	    (char *)(str + probe->dofp_provider),
12210 	    MIN(DTRACE_PROVNAMELEN - 1, size - probe->dofp_provider));
12211 
12212 	if (probe->dofp_mod >= strtab->dofs_size) {
12213 		dtrace_dof_error(dof, "corrupt probe module");
12214 		return (NULL);
12215 	}
12216 
12217 	(void) strncpy(desc->dtpd_mod, (char *)(str + probe->dofp_mod),
12218 	    MIN(DTRACE_MODNAMELEN - 1, size - probe->dofp_mod));
12219 
12220 	if (probe->dofp_func >= strtab->dofs_size) {
12221 		dtrace_dof_error(dof, "corrupt probe function");
12222 		return (NULL);
12223 	}
12224 
12225 	(void) strncpy(desc->dtpd_func, (char *)(str + probe->dofp_func),
12226 	    MIN(DTRACE_FUNCNAMELEN - 1, size - probe->dofp_func));
12227 
12228 	if (probe->dofp_name >= strtab->dofs_size) {
12229 		dtrace_dof_error(dof, "corrupt probe name");
12230 		return (NULL);
12231 	}
12232 
12233 	(void) strncpy(desc->dtpd_name, (char *)(str + probe->dofp_name),
12234 	    MIN(DTRACE_NAMELEN - 1, size - probe->dofp_name));
12235 
12236 	return (desc);
12237 }
12238 
12239 static dtrace_difo_t *
12240 dtrace_dof_difo(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12241     cred_t *cr)
12242 {
12243 	dtrace_difo_t *dp;
12244 	size_t ttl = 0;
12245 	dof_difohdr_t *dofd;
12246 	uintptr_t daddr = (uintptr_t)dof;
12247 	size_t max = dtrace_difo_maxsize;
12248 	int i, l, n;
12249 
12250 	static const struct {
12251 		int section;
12252 		int bufoffs;
12253 		int lenoffs;
12254 		int entsize;
12255 		int align;
12256 		const char *msg;
12257 	} difo[] = {
12258 		{ DOF_SECT_DIF, offsetof(dtrace_difo_t, dtdo_buf),
12259 		offsetof(dtrace_difo_t, dtdo_len), sizeof (dif_instr_t),
12260 		sizeof (dif_instr_t), "multiple DIF sections" },
12261 
12262 		{ DOF_SECT_INTTAB, offsetof(dtrace_difo_t, dtdo_inttab),
12263 		offsetof(dtrace_difo_t, dtdo_intlen), sizeof (uint64_t),
12264 		sizeof (uint64_t), "multiple integer tables" },
12265 
12266 		{ DOF_SECT_STRTAB, offsetof(dtrace_difo_t, dtdo_strtab),
12267 		offsetof(dtrace_difo_t, dtdo_strlen), 0,
12268 		sizeof (char), "multiple string tables" },
12269 
12270 		{ DOF_SECT_VARTAB, offsetof(dtrace_difo_t, dtdo_vartab),
12271 		offsetof(dtrace_difo_t, dtdo_varlen), sizeof (dtrace_difv_t),
12272 		sizeof (uint_t), "multiple variable tables" },
12273 
12274 		{ DOF_SECT_NONE, 0, 0, 0, 0, NULL }
12275 	};
12276 
12277 	if (sec->dofs_type != DOF_SECT_DIFOHDR) {
12278 		dtrace_dof_error(dof, "invalid DIFO header section");
12279 		return (NULL);
12280 	}
12281 
12282 	if (sec->dofs_align != sizeof (dof_secidx_t)) {
12283 		dtrace_dof_error(dof, "bad alignment in DIFO header");
12284 		return (NULL);
12285 	}
12286 
12287 	if (sec->dofs_size < sizeof (dof_difohdr_t) ||
12288 	    sec->dofs_size % sizeof (dof_secidx_t)) {
12289 		dtrace_dof_error(dof, "bad size in DIFO header");
12290 		return (NULL);
12291 	}
12292 
12293 	dofd = (dof_difohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
12294 	n = (sec->dofs_size - sizeof (*dofd)) / sizeof (dof_secidx_t) + 1;
12295 
12296 	dp = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
12297 	dp->dtdo_rtype = dofd->dofd_rtype;
12298 
12299 	for (l = 0; l < n; l++) {
12300 		dof_sec_t *subsec;
12301 		void **bufp;
12302 		uint32_t *lenp;
12303 
12304 		if ((subsec = dtrace_dof_sect(dof, DOF_SECT_NONE,
12305 		    dofd->dofd_links[l])) == NULL)
12306 			goto err; /* invalid section link */
12307 
12308 		if (ttl + subsec->dofs_size > max) {
12309 			dtrace_dof_error(dof, "exceeds maximum size");
12310 			goto err;
12311 		}
12312 
12313 		ttl += subsec->dofs_size;
12314 
12315 		for (i = 0; difo[i].section != DOF_SECT_NONE; i++) {
12316 			if (subsec->dofs_type != difo[i].section)
12317 				continue;
12318 
12319 			if (!(subsec->dofs_flags & DOF_SECF_LOAD)) {
12320 				dtrace_dof_error(dof, "section not loaded");
12321 				goto err;
12322 			}
12323 
12324 			if (subsec->dofs_align != difo[i].align) {
12325 				dtrace_dof_error(dof, "bad alignment");
12326 				goto err;
12327 			}
12328 
12329 			bufp = (void **)((uintptr_t)dp + difo[i].bufoffs);
12330 			lenp = (uint32_t *)((uintptr_t)dp + difo[i].lenoffs);
12331 
12332 			if (*bufp != NULL) {
12333 				dtrace_dof_error(dof, difo[i].msg);
12334 				goto err;
12335 			}
12336 
12337 			if (difo[i].entsize != subsec->dofs_entsize) {
12338 				dtrace_dof_error(dof, "entry size mismatch");
12339 				goto err;
12340 			}
12341 
12342 			if (subsec->dofs_entsize != 0 &&
12343 			    (subsec->dofs_size % subsec->dofs_entsize) != 0) {
12344 				dtrace_dof_error(dof, "corrupt entry size");
12345 				goto err;
12346 			}
12347 
12348 			*lenp = subsec->dofs_size;
12349 			*bufp = kmem_alloc(subsec->dofs_size, KM_SLEEP);
12350 			bcopy((char *)(uintptr_t)(daddr + subsec->dofs_offset),
12351 			    *bufp, subsec->dofs_size);
12352 
12353 			if (subsec->dofs_entsize != 0)
12354 				*lenp /= subsec->dofs_entsize;
12355 
12356 			break;
12357 		}
12358 
12359 		/*
12360 		 * If we encounter a loadable DIFO sub-section that is not
12361 		 * known to us, assume this is a broken program and fail.
12362 		 */
12363 		if (difo[i].section == DOF_SECT_NONE &&
12364 		    (subsec->dofs_flags & DOF_SECF_LOAD)) {
12365 			dtrace_dof_error(dof, "unrecognized DIFO subsection");
12366 			goto err;
12367 		}
12368 	}
12369 
12370 	if (dp->dtdo_buf == NULL) {
12371 		/*
12372 		 * We can't have a DIF object without DIF text.
12373 		 */
12374 		dtrace_dof_error(dof, "missing DIF text");
12375 		goto err;
12376 	}
12377 
12378 	/*
12379 	 * Before we validate the DIF object, run through the variable table
12380 	 * looking for the strings -- if any of their size are under, we'll set
12381 	 * their size to be the system-wide default string size.  Note that
12382 	 * this should _not_ happen if the "strsize" option has been set --
12383 	 * in this case, the compiler should have set the size to reflect the
12384 	 * setting of the option.
12385 	 */
12386 	for (i = 0; i < dp->dtdo_varlen; i++) {
12387 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
12388 		dtrace_diftype_t *t = &v->dtdv_type;
12389 
12390 		if (v->dtdv_id < DIF_VAR_OTHER_UBASE)
12391 			continue;
12392 
12393 		if (t->dtdt_kind == DIF_TYPE_STRING && t->dtdt_size == 0)
12394 			t->dtdt_size = dtrace_strsize_default;
12395 	}
12396 
12397 	if (dtrace_difo_validate(dp, vstate, DIF_DIR_NREGS, cr) != 0)
12398 		goto err;
12399 
12400 	dtrace_difo_init(dp, vstate);
12401 	return (dp);
12402 
12403 err:
12404 	kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
12405 	kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
12406 	kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
12407 	kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
12408 
12409 	kmem_free(dp, sizeof (dtrace_difo_t));
12410 	return (NULL);
12411 }
12412 
12413 static dtrace_predicate_t *
12414 dtrace_dof_predicate(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12415     cred_t *cr)
12416 {
12417 	dtrace_difo_t *dp;
12418 
12419 	if ((dp = dtrace_dof_difo(dof, sec, vstate, cr)) == NULL)
12420 		return (NULL);
12421 
12422 	return (dtrace_predicate_create(dp));
12423 }
12424 
12425 static dtrace_actdesc_t *
12426 dtrace_dof_actdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12427     cred_t *cr)
12428 {
12429 	dtrace_actdesc_t *act, *first = NULL, *last = NULL, *next;
12430 	dof_actdesc_t *desc;
12431 	dof_sec_t *difosec;
12432 	size_t offs;
12433 	uintptr_t daddr = (uintptr_t)dof;
12434 	uint64_t arg;
12435 	dtrace_actkind_t kind;
12436 
12437 	if (sec->dofs_type != DOF_SECT_ACTDESC) {
12438 		dtrace_dof_error(dof, "invalid action section");
12439 		return (NULL);
12440 	}
12441 
12442 	if (sec->dofs_offset + sizeof (dof_actdesc_t) > dof->dofh_loadsz) {
12443 		dtrace_dof_error(dof, "truncated action description");
12444 		return (NULL);
12445 	}
12446 
12447 	if (sec->dofs_align != sizeof (uint64_t)) {
12448 		dtrace_dof_error(dof, "bad alignment in action description");
12449 		return (NULL);
12450 	}
12451 
12452 	if (sec->dofs_size < sec->dofs_entsize) {
12453 		dtrace_dof_error(dof, "section entry size exceeds total size");
12454 		return (NULL);
12455 	}
12456 
12457 	if (sec->dofs_entsize != sizeof (dof_actdesc_t)) {
12458 		dtrace_dof_error(dof, "bad entry size in action description");
12459 		return (NULL);
12460 	}
12461 
12462 	if (sec->dofs_size / sec->dofs_entsize > dtrace_actions_max) {
12463 		dtrace_dof_error(dof, "actions exceed dtrace_actions_max");
12464 		return (NULL);
12465 	}
12466 
12467 	for (offs = 0; offs < sec->dofs_size; offs += sec->dofs_entsize) {
12468 		desc = (dof_actdesc_t *)(daddr +
12469 		    (uintptr_t)sec->dofs_offset + offs);
12470 		kind = (dtrace_actkind_t)desc->dofa_kind;
12471 
12472 		if ((DTRACEACT_ISPRINTFLIKE(kind) &&
12473 		    (kind != DTRACEACT_PRINTA ||
12474 		    desc->dofa_strtab != DOF_SECIDX_NONE)) ||
12475 		    (kind == DTRACEACT_DIFEXPR &&
12476 		    desc->dofa_strtab != DOF_SECIDX_NONE)) {
12477 			dof_sec_t *strtab;
12478 			char *str, *fmt;
12479 			uint64_t i;
12480 
12481 			/*
12482 			 * The argument to these actions is an index into the
12483 			 * DOF string table.  For printf()-like actions, this
12484 			 * is the format string.  For print(), this is the
12485 			 * CTF type of the expression result.
12486 			 */
12487 			if ((strtab = dtrace_dof_sect(dof,
12488 			    DOF_SECT_STRTAB, desc->dofa_strtab)) == NULL)
12489 				goto err;
12490 
12491 			str = (char *)((uintptr_t)dof +
12492 			    (uintptr_t)strtab->dofs_offset);
12493 
12494 			for (i = desc->dofa_arg; i < strtab->dofs_size; i++) {
12495 				if (str[i] == '\0')
12496 					break;
12497 			}
12498 
12499 			if (i >= strtab->dofs_size) {
12500 				dtrace_dof_error(dof, "bogus format string");
12501 				goto err;
12502 			}
12503 
12504 			if (i == desc->dofa_arg) {
12505 				dtrace_dof_error(dof, "empty format string");
12506 				goto err;
12507 			}
12508 
12509 			i -= desc->dofa_arg;
12510 			fmt = kmem_alloc(i + 1, KM_SLEEP);
12511 			bcopy(&str[desc->dofa_arg], fmt, i + 1);
12512 			arg = (uint64_t)(uintptr_t)fmt;
12513 		} else {
12514 			if (kind == DTRACEACT_PRINTA) {
12515 				ASSERT(desc->dofa_strtab == DOF_SECIDX_NONE);
12516 				arg = 0;
12517 			} else {
12518 				arg = desc->dofa_arg;
12519 			}
12520 		}
12521 
12522 		act = dtrace_actdesc_create(kind, desc->dofa_ntuple,
12523 		    desc->dofa_uarg, arg);
12524 
12525 		if (last != NULL) {
12526 			last->dtad_next = act;
12527 		} else {
12528 			first = act;
12529 		}
12530 
12531 		last = act;
12532 
12533 		if (desc->dofa_difo == DOF_SECIDX_NONE)
12534 			continue;
12535 
12536 		if ((difosec = dtrace_dof_sect(dof,
12537 		    DOF_SECT_DIFOHDR, desc->dofa_difo)) == NULL)
12538 			goto err;
12539 
12540 		act->dtad_difo = dtrace_dof_difo(dof, difosec, vstate, cr);
12541 
12542 		if (act->dtad_difo == NULL)
12543 			goto err;
12544 	}
12545 
12546 	ASSERT(first != NULL);
12547 	return (first);
12548 
12549 err:
12550 	for (act = first; act != NULL; act = next) {
12551 		next = act->dtad_next;
12552 		dtrace_actdesc_release(act, vstate);
12553 	}
12554 
12555 	return (NULL);
12556 }
12557 
12558 static dtrace_ecbdesc_t *
12559 dtrace_dof_ecbdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12560     cred_t *cr)
12561 {
12562 	dtrace_ecbdesc_t *ep;
12563 	dof_ecbdesc_t *ecb;
12564 	dtrace_probedesc_t *desc;
12565 	dtrace_predicate_t *pred = NULL;
12566 
12567 	if (sec->dofs_size < sizeof (dof_ecbdesc_t)) {
12568 		dtrace_dof_error(dof, "truncated ECB description");
12569 		return (NULL);
12570 	}
12571 
12572 	if (sec->dofs_align != sizeof (uint64_t)) {
12573 		dtrace_dof_error(dof, "bad alignment in ECB description");
12574 		return (NULL);
12575 	}
12576 
12577 	ecb = (dof_ecbdesc_t *)((uintptr_t)dof + (uintptr_t)sec->dofs_offset);
12578 	sec = dtrace_dof_sect(dof, DOF_SECT_PROBEDESC, ecb->dofe_probes);
12579 
12580 	if (sec == NULL)
12581 		return (NULL);
12582 
12583 	ep = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
12584 	ep->dted_uarg = ecb->dofe_uarg;
12585 	desc = &ep->dted_probe;
12586 
12587 	if (dtrace_dof_probedesc(dof, sec, desc) == NULL)
12588 		goto err;
12589 
12590 	if (ecb->dofe_pred != DOF_SECIDX_NONE) {
12591 		if ((sec = dtrace_dof_sect(dof,
12592 		    DOF_SECT_DIFOHDR, ecb->dofe_pred)) == NULL)
12593 			goto err;
12594 
12595 		if ((pred = dtrace_dof_predicate(dof, sec, vstate, cr)) == NULL)
12596 			goto err;
12597 
12598 		ep->dted_pred.dtpdd_predicate = pred;
12599 	}
12600 
12601 	if (ecb->dofe_actions != DOF_SECIDX_NONE) {
12602 		if ((sec = dtrace_dof_sect(dof,
12603 		    DOF_SECT_ACTDESC, ecb->dofe_actions)) == NULL)
12604 			goto err;
12605 
12606 		ep->dted_action = dtrace_dof_actdesc(dof, sec, vstate, cr);
12607 
12608 		if (ep->dted_action == NULL)
12609 			goto err;
12610 	}
12611 
12612 	return (ep);
12613 
12614 err:
12615 	if (pred != NULL)
12616 		dtrace_predicate_release(pred, vstate);
12617 	kmem_free(ep, sizeof (dtrace_ecbdesc_t));
12618 	return (NULL);
12619 }
12620 
12621 /*
12622  * Apply the relocations from the specified 'sec' (a DOF_SECT_URELHDR) to the
12623  * specified DOF.  At present, this amounts to simply adding 'ubase' to the
12624  * site of any user SETX relocations to account for load object base address.
12625  * In the future, if we need other relocations, this function can be extended.
12626  */
12627 static int
12628 dtrace_dof_relocate(dof_hdr_t *dof, dof_sec_t *sec, uint64_t ubase)
12629 {
12630 	uintptr_t daddr = (uintptr_t)dof;
12631 	dof_relohdr_t *dofr =
12632 	    (dof_relohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
12633 	dof_sec_t *ss, *rs, *ts;
12634 	dof_relodesc_t *r;
12635 	uint_t i, n;
12636 
12637 	if (sec->dofs_size < sizeof (dof_relohdr_t) ||
12638 	    sec->dofs_align != sizeof (dof_secidx_t)) {
12639 		dtrace_dof_error(dof, "invalid relocation header");
12640 		return (-1);
12641 	}
12642 
12643 	ss = dtrace_dof_sect(dof, DOF_SECT_STRTAB, dofr->dofr_strtab);
12644 	rs = dtrace_dof_sect(dof, DOF_SECT_RELTAB, dofr->dofr_relsec);
12645 	ts = dtrace_dof_sect(dof, DOF_SECT_NONE, dofr->dofr_tgtsec);
12646 
12647 	if (ss == NULL || rs == NULL || ts == NULL)
12648 		return (-1); /* dtrace_dof_error() has been called already */
12649 
12650 	if (rs->dofs_entsize < sizeof (dof_relodesc_t) ||
12651 	    rs->dofs_align != sizeof (uint64_t)) {
12652 		dtrace_dof_error(dof, "invalid relocation section");
12653 		return (-1);
12654 	}
12655 
12656 	r = (dof_relodesc_t *)(uintptr_t)(daddr + rs->dofs_offset);
12657 	n = rs->dofs_size / rs->dofs_entsize;
12658 
12659 	for (i = 0; i < n; i++) {
12660 		uintptr_t taddr = daddr + ts->dofs_offset + r->dofr_offset;
12661 
12662 		switch (r->dofr_type) {
12663 		case DOF_RELO_NONE:
12664 			break;
12665 		case DOF_RELO_SETX:
12666 			if (r->dofr_offset >= ts->dofs_size || r->dofr_offset +
12667 			    sizeof (uint64_t) > ts->dofs_size) {
12668 				dtrace_dof_error(dof, "bad relocation offset");
12669 				return (-1);
12670 			}
12671 
12672 			if (!IS_P2ALIGNED(taddr, sizeof (uint64_t))) {
12673 				dtrace_dof_error(dof, "misaligned setx relo");
12674 				return (-1);
12675 			}
12676 
12677 			*(uint64_t *)taddr += ubase;
12678 			break;
12679 		default:
12680 			dtrace_dof_error(dof, "invalid relocation type");
12681 			return (-1);
12682 		}
12683 
12684 		r = (dof_relodesc_t *)((uintptr_t)r + rs->dofs_entsize);
12685 	}
12686 
12687 	return (0);
12688 }
12689 
12690 /*
12691  * The dof_hdr_t passed to dtrace_dof_slurp() should be a partially validated
12692  * header:  it should be at the front of a memory region that is at least
12693  * sizeof (dof_hdr_t) in size -- and then at least dof_hdr.dofh_loadsz in
12694  * size.  It need not be validated in any other way.
12695  */
12696 static int
12697 dtrace_dof_slurp(dof_hdr_t *dof, dtrace_vstate_t *vstate, cred_t *cr,
12698     dtrace_enabling_t **enabp, uint64_t ubase, int noprobes)
12699 {
12700 	uint64_t len = dof->dofh_loadsz, seclen;
12701 	uintptr_t daddr = (uintptr_t)dof;
12702 	dtrace_ecbdesc_t *ep;
12703 	dtrace_enabling_t *enab;
12704 	uint_t i;
12705 
12706 	ASSERT(MUTEX_HELD(&dtrace_lock));
12707 	ASSERT(dof->dofh_loadsz >= sizeof (dof_hdr_t));
12708 
12709 	/*
12710 	 * Check the DOF header identification bytes.  In addition to checking
12711 	 * valid settings, we also verify that unused bits/bytes are zeroed so
12712 	 * we can use them later without fear of regressing existing binaries.
12713 	 */
12714 	if (bcmp(&dof->dofh_ident[DOF_ID_MAG0],
12715 	    DOF_MAG_STRING, DOF_MAG_STRLEN) != 0) {
12716 		dtrace_dof_error(dof, "DOF magic string mismatch");
12717 		return (-1);
12718 	}
12719 
12720 	if (dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_ILP32 &&
12721 	    dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_LP64) {
12722 		dtrace_dof_error(dof, "DOF has invalid data model");
12723 		return (-1);
12724 	}
12725 
12726 	if (dof->dofh_ident[DOF_ID_ENCODING] != DOF_ENCODE_NATIVE) {
12727 		dtrace_dof_error(dof, "DOF encoding mismatch");
12728 		return (-1);
12729 	}
12730 
12731 	if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
12732 	    dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_2) {
12733 		dtrace_dof_error(dof, "DOF version mismatch");
12734 		return (-1);
12735 	}
12736 
12737 	if (dof->dofh_ident[DOF_ID_DIFVERS] != DIF_VERSION_2) {
12738 		dtrace_dof_error(dof, "DOF uses unsupported instruction set");
12739 		return (-1);
12740 	}
12741 
12742 	if (dof->dofh_ident[DOF_ID_DIFIREG] > DIF_DIR_NREGS) {
12743 		dtrace_dof_error(dof, "DOF uses too many integer registers");
12744 		return (-1);
12745 	}
12746 
12747 	if (dof->dofh_ident[DOF_ID_DIFTREG] > DIF_DTR_NREGS) {
12748 		dtrace_dof_error(dof, "DOF uses too many tuple registers");
12749 		return (-1);
12750 	}
12751 
12752 	for (i = DOF_ID_PAD; i < DOF_ID_SIZE; i++) {
12753 		if (dof->dofh_ident[i] != 0) {
12754 			dtrace_dof_error(dof, "DOF has invalid ident byte set");
12755 			return (-1);
12756 		}
12757 	}
12758 
12759 	if (dof->dofh_flags & ~DOF_FL_VALID) {
12760 		dtrace_dof_error(dof, "DOF has invalid flag bits set");
12761 		return (-1);
12762 	}
12763 
12764 	if (dof->dofh_secsize == 0) {
12765 		dtrace_dof_error(dof, "zero section header size");
12766 		return (-1);
12767 	}
12768 
12769 	/*
12770 	 * Check that the section headers don't exceed the amount of DOF
12771 	 * data.  Note that we cast the section size and number of sections
12772 	 * to uint64_t's to prevent possible overflow in the multiplication.
12773 	 */
12774 	seclen = (uint64_t)dof->dofh_secnum * (uint64_t)dof->dofh_secsize;
12775 
12776 	if (dof->dofh_secoff > len || seclen > len ||
12777 	    dof->dofh_secoff + seclen > len) {
12778 		dtrace_dof_error(dof, "truncated section headers");
12779 		return (-1);
12780 	}
12781 
12782 	if (!IS_P2ALIGNED(dof->dofh_secoff, sizeof (uint64_t))) {
12783 		dtrace_dof_error(dof, "misaligned section headers");
12784 		return (-1);
12785 	}
12786 
12787 	if (!IS_P2ALIGNED(dof->dofh_secsize, sizeof (uint64_t))) {
12788 		dtrace_dof_error(dof, "misaligned section size");
12789 		return (-1);
12790 	}
12791 
12792 	/*
12793 	 * Take an initial pass through the section headers to be sure that
12794 	 * the headers don't have stray offsets.  If the 'noprobes' flag is
12795 	 * set, do not permit sections relating to providers, probes, or args.
12796 	 */
12797 	for (i = 0; i < dof->dofh_secnum; i++) {
12798 		dof_sec_t *sec = (dof_sec_t *)(daddr +
12799 		    (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12800 
12801 		if (noprobes) {
12802 			switch (sec->dofs_type) {
12803 			case DOF_SECT_PROVIDER:
12804 			case DOF_SECT_PROBES:
12805 			case DOF_SECT_PRARGS:
12806 			case DOF_SECT_PROFFS:
12807 				dtrace_dof_error(dof, "illegal sections "
12808 				    "for enabling");
12809 				return (-1);
12810 			}
12811 		}
12812 
12813 		if (!(sec->dofs_flags & DOF_SECF_LOAD))
12814 			continue; /* just ignore non-loadable sections */
12815 
12816 		if (sec->dofs_align & (sec->dofs_align - 1)) {
12817 			dtrace_dof_error(dof, "bad section alignment");
12818 			return (-1);
12819 		}
12820 
12821 		if (sec->dofs_offset & (sec->dofs_align - 1)) {
12822 			dtrace_dof_error(dof, "misaligned section");
12823 			return (-1);
12824 		}
12825 
12826 		if (sec->dofs_offset > len || sec->dofs_size > len ||
12827 		    sec->dofs_offset + sec->dofs_size > len) {
12828 			dtrace_dof_error(dof, "corrupt section header");
12829 			return (-1);
12830 		}
12831 
12832 		if (sec->dofs_type == DOF_SECT_STRTAB && *((char *)daddr +
12833 		    sec->dofs_offset + sec->dofs_size - 1) != '\0') {
12834 			dtrace_dof_error(dof, "non-terminating string table");
12835 			return (-1);
12836 		}
12837 	}
12838 
12839 	/*
12840 	 * Take a second pass through the sections and locate and perform any
12841 	 * relocations that are present.  We do this after the first pass to
12842 	 * be sure that all sections have had their headers validated.
12843 	 */
12844 	for (i = 0; i < dof->dofh_secnum; i++) {
12845 		dof_sec_t *sec = (dof_sec_t *)(daddr +
12846 		    (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12847 
12848 		if (!(sec->dofs_flags & DOF_SECF_LOAD))
12849 			continue; /* skip sections that are not loadable */
12850 
12851 		switch (sec->dofs_type) {
12852 		case DOF_SECT_URELHDR:
12853 			if (dtrace_dof_relocate(dof, sec, ubase) != 0)
12854 				return (-1);
12855 			break;
12856 		}
12857 	}
12858 
12859 	if ((enab = *enabp) == NULL)
12860 		enab = *enabp = dtrace_enabling_create(vstate);
12861 
12862 	for (i = 0; i < dof->dofh_secnum; i++) {
12863 		dof_sec_t *sec = (dof_sec_t *)(daddr +
12864 		    (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12865 
12866 		if (sec->dofs_type != DOF_SECT_ECBDESC)
12867 			continue;
12868 
12869 		if ((ep = dtrace_dof_ecbdesc(dof, sec, vstate, cr)) == NULL) {
12870 			dtrace_enabling_destroy(enab);
12871 			*enabp = NULL;
12872 			return (-1);
12873 		}
12874 
12875 		dtrace_enabling_add(enab, ep);
12876 	}
12877 
12878 	return (0);
12879 }
12880 
12881 /*
12882  * Process DOF for any options.  This routine assumes that the DOF has been
12883  * at least processed by dtrace_dof_slurp().
12884  */
12885 static int
12886 dtrace_dof_options(dof_hdr_t *dof, dtrace_state_t *state)
12887 {
12888 	int i, rval;
12889 	uint32_t entsize;
12890 	size_t offs;
12891 	dof_optdesc_t *desc;
12892 
12893 	for (i = 0; i < dof->dofh_secnum; i++) {
12894 		dof_sec_t *sec = (dof_sec_t *)((uintptr_t)dof +
12895 		    (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12896 
12897 		if (sec->dofs_type != DOF_SECT_OPTDESC)
12898 			continue;
12899 
12900 		if (sec->dofs_align != sizeof (uint64_t)) {
12901 			dtrace_dof_error(dof, "bad alignment in "
12902 			    "option description");
12903 			return (EINVAL);
12904 		}
12905 
12906 		if ((entsize = sec->dofs_entsize) == 0) {
12907 			dtrace_dof_error(dof, "zeroed option entry size");
12908 			return (EINVAL);
12909 		}
12910 
12911 		if (entsize < sizeof (dof_optdesc_t)) {
12912 			dtrace_dof_error(dof, "bad option entry size");
12913 			return (EINVAL);
12914 		}
12915 
12916 		for (offs = 0; offs < sec->dofs_size; offs += entsize) {
12917 			desc = (dof_optdesc_t *)((uintptr_t)dof +
12918 			    (uintptr_t)sec->dofs_offset + offs);
12919 
12920 			if (desc->dofo_strtab != DOF_SECIDX_NONE) {
12921 				dtrace_dof_error(dof, "non-zero option string");
12922 				return (EINVAL);
12923 			}
12924 
12925 			if (desc->dofo_value == DTRACEOPT_UNSET) {
12926 				dtrace_dof_error(dof, "unset option");
12927 				return (EINVAL);
12928 			}
12929 
12930 			if ((rval = dtrace_state_option(state,
12931 			    desc->dofo_option, desc->dofo_value)) != 0) {
12932 				dtrace_dof_error(dof, "rejected option");
12933 				return (rval);
12934 			}
12935 		}
12936 	}
12937 
12938 	return (0);
12939 }
12940 
12941 /*
12942  * DTrace Consumer State Functions
12943  */
12944 static int
12945 dtrace_dstate_init(dtrace_dstate_t *dstate, size_t size)
12946 {
12947 	size_t hashsize, maxper, min, chunksize = dstate->dtds_chunksize;
12948 	void *base;
12949 	uintptr_t limit;
12950 	dtrace_dynvar_t *dvar, *next, *start;
12951 	int i;
12952 
12953 	ASSERT(MUTEX_HELD(&dtrace_lock));
12954 	ASSERT(dstate->dtds_base == NULL && dstate->dtds_percpu == NULL);
12955 
12956 	bzero(dstate, sizeof (dtrace_dstate_t));
12957 
12958 	if ((dstate->dtds_chunksize = chunksize) == 0)
12959 		dstate->dtds_chunksize = DTRACE_DYNVAR_CHUNKSIZE;
12960 
12961 	if (size < (min = dstate->dtds_chunksize + sizeof (dtrace_dynhash_t)))
12962 		size = min;
12963 
12964 	if ((base = kmem_zalloc(size, KM_NOSLEEP)) == NULL)
12965 		return (ENOMEM);
12966 
12967 	dstate->dtds_size = size;
12968 	dstate->dtds_base = base;
12969 	dstate->dtds_percpu = kmem_cache_alloc(dtrace_state_cache, KM_SLEEP);
12970 	bzero(dstate->dtds_percpu, NCPU * sizeof (dtrace_dstate_percpu_t));
12971 
12972 	hashsize = size / (dstate->dtds_chunksize + sizeof (dtrace_dynhash_t));
12973 
12974 	if (hashsize != 1 && (hashsize & 1))
12975 		hashsize--;
12976 
12977 	dstate->dtds_hashsize = hashsize;
12978 	dstate->dtds_hash = dstate->dtds_base;
12979 
12980 	/*
12981 	 * Set all of our hash buckets to point to the single sink, and (if
12982 	 * it hasn't already been set), set the sink's hash value to be the
12983 	 * sink sentinel value.  The sink is needed for dynamic variable
12984 	 * lookups to know that they have iterated over an entire, valid hash
12985 	 * chain.
12986 	 */
12987 	for (i = 0; i < hashsize; i++)
12988 		dstate->dtds_hash[i].dtdh_chain = &dtrace_dynhash_sink;
12989 
12990 	if (dtrace_dynhash_sink.dtdv_hashval != DTRACE_DYNHASH_SINK)
12991 		dtrace_dynhash_sink.dtdv_hashval = DTRACE_DYNHASH_SINK;
12992 
12993 	/*
12994 	 * Determine number of active CPUs.  Divide free list evenly among
12995 	 * active CPUs.
12996 	 */
12997 	start = (dtrace_dynvar_t *)
12998 	    ((uintptr_t)base + hashsize * sizeof (dtrace_dynhash_t));
12999 	limit = (uintptr_t)base + size;
13000 
13001 	maxper = (limit - (uintptr_t)start) / NCPU;
13002 	maxper = (maxper / dstate->dtds_chunksize) * dstate->dtds_chunksize;
13003 
13004 #if !defined(sun)
13005 	CPU_FOREACH(i) {
13006 #else
13007 	for (i = 0; i < NCPU; i++) {
13008 #endif
13009 		dstate->dtds_percpu[i].dtdsc_free = dvar = start;
13010 
13011 		/*
13012 		 * If we don't even have enough chunks to make it once through
13013 		 * NCPUs, we're just going to allocate everything to the first
13014 		 * CPU.  And if we're on the last CPU, we're going to allocate
13015 		 * whatever is left over.  In either case, we set the limit to
13016 		 * be the limit of the dynamic variable space.
13017 		 */
13018 		if (maxper == 0 || i == NCPU - 1) {
13019 			limit = (uintptr_t)base + size;
13020 			start = NULL;
13021 		} else {
13022 			limit = (uintptr_t)start + maxper;
13023 			start = (dtrace_dynvar_t *)limit;
13024 		}
13025 
13026 		ASSERT(limit <= (uintptr_t)base + size);
13027 
13028 		for (;;) {
13029 			next = (dtrace_dynvar_t *)((uintptr_t)dvar +
13030 			    dstate->dtds_chunksize);
13031 
13032 			if ((uintptr_t)next + dstate->dtds_chunksize >= limit)
13033 				break;
13034 
13035 			dvar->dtdv_next = next;
13036 			dvar = next;
13037 		}
13038 
13039 		if (maxper == 0)
13040 			break;
13041 	}
13042 
13043 	return (0);
13044 }
13045 
13046 static void
13047 dtrace_dstate_fini(dtrace_dstate_t *dstate)
13048 {
13049 	ASSERT(MUTEX_HELD(&cpu_lock));
13050 
13051 	if (dstate->dtds_base == NULL)
13052 		return;
13053 
13054 	kmem_free(dstate->dtds_base, dstate->dtds_size);
13055 	kmem_cache_free(dtrace_state_cache, dstate->dtds_percpu);
13056 }
13057 
13058 static void
13059 dtrace_vstate_fini(dtrace_vstate_t *vstate)
13060 {
13061 	/*
13062 	 * Logical XOR, where are you?
13063 	 */
13064 	ASSERT((vstate->dtvs_nglobals == 0) ^ (vstate->dtvs_globals != NULL));
13065 
13066 	if (vstate->dtvs_nglobals > 0) {
13067 		kmem_free(vstate->dtvs_globals, vstate->dtvs_nglobals *
13068 		    sizeof (dtrace_statvar_t *));
13069 	}
13070 
13071 	if (vstate->dtvs_ntlocals > 0) {
13072 		kmem_free(vstate->dtvs_tlocals, vstate->dtvs_ntlocals *
13073 		    sizeof (dtrace_difv_t));
13074 	}
13075 
13076 	ASSERT((vstate->dtvs_nlocals == 0) ^ (vstate->dtvs_locals != NULL));
13077 
13078 	if (vstate->dtvs_nlocals > 0) {
13079 		kmem_free(vstate->dtvs_locals, vstate->dtvs_nlocals *
13080 		    sizeof (dtrace_statvar_t *));
13081 	}
13082 }
13083 
13084 #if defined(sun)
13085 static void
13086 dtrace_state_clean(dtrace_state_t *state)
13087 {
13088 	if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE)
13089 		return;
13090 
13091 	dtrace_dynvar_clean(&state->dts_vstate.dtvs_dynvars);
13092 	dtrace_speculation_clean(state);
13093 }
13094 
13095 static void
13096 dtrace_state_deadman(dtrace_state_t *state)
13097 {
13098 	hrtime_t now;
13099 
13100 	dtrace_sync();
13101 
13102 	now = dtrace_gethrtime();
13103 
13104 	if (state != dtrace_anon.dta_state &&
13105 	    now - state->dts_laststatus >= dtrace_deadman_user)
13106 		return;
13107 
13108 	/*
13109 	 * We must be sure that dts_alive never appears to be less than the
13110 	 * value upon entry to dtrace_state_deadman(), and because we lack a
13111 	 * dtrace_cas64(), we cannot store to it atomically.  We thus instead
13112 	 * store INT64_MAX to it, followed by a memory barrier, followed by
13113 	 * the new value.  This assures that dts_alive never appears to be
13114 	 * less than its true value, regardless of the order in which the
13115 	 * stores to the underlying storage are issued.
13116 	 */
13117 	state->dts_alive = INT64_MAX;
13118 	dtrace_membar_producer();
13119 	state->dts_alive = now;
13120 }
13121 #else
13122 static void
13123 dtrace_state_clean(void *arg)
13124 {
13125 	dtrace_state_t *state = arg;
13126 	dtrace_optval_t *opt = state->dts_options;
13127 
13128 	if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE)
13129 		return;
13130 
13131 	dtrace_dynvar_clean(&state->dts_vstate.dtvs_dynvars);
13132 	dtrace_speculation_clean(state);
13133 
13134 	callout_reset(&state->dts_cleaner, hz * opt[DTRACEOPT_CLEANRATE] / NANOSEC,
13135 	    dtrace_state_clean, state);
13136 }
13137 
13138 static void
13139 dtrace_state_deadman(void *arg)
13140 {
13141 	dtrace_state_t *state = arg;
13142 	hrtime_t now;
13143 
13144 	dtrace_sync();
13145 
13146 	dtrace_debug_output();
13147 
13148 	now = dtrace_gethrtime();
13149 
13150 	if (state != dtrace_anon.dta_state &&
13151 	    now - state->dts_laststatus >= dtrace_deadman_user)
13152 		return;
13153 
13154 	/*
13155 	 * We must be sure that dts_alive never appears to be less than the
13156 	 * value upon entry to dtrace_state_deadman(), and because we lack a
13157 	 * dtrace_cas64(), we cannot store to it atomically.  We thus instead
13158 	 * store INT64_MAX to it, followed by a memory barrier, followed by
13159 	 * the new value.  This assures that dts_alive never appears to be
13160 	 * less than its true value, regardless of the order in which the
13161 	 * stores to the underlying storage are issued.
13162 	 */
13163 	state->dts_alive = INT64_MAX;
13164 	dtrace_membar_producer();
13165 	state->dts_alive = now;
13166 
13167 	callout_reset(&state->dts_deadman, hz * dtrace_deadman_interval / NANOSEC,
13168 	    dtrace_state_deadman, state);
13169 }
13170 #endif
13171 
13172 static dtrace_state_t *
13173 #if defined(sun)
13174 dtrace_state_create(dev_t *devp, cred_t *cr)
13175 #else
13176 dtrace_state_create(struct cdev *dev)
13177 #endif
13178 {
13179 #if defined(sun)
13180 	minor_t minor;
13181 	major_t major;
13182 #else
13183 	cred_t *cr = NULL;
13184 	int m = 0;
13185 #endif
13186 	char c[30];
13187 	dtrace_state_t *state;
13188 	dtrace_optval_t *opt;
13189 	int bufsize = NCPU * sizeof (dtrace_buffer_t), i;
13190 
13191 	ASSERT(MUTEX_HELD(&dtrace_lock));
13192 	ASSERT(MUTEX_HELD(&cpu_lock));
13193 
13194 #if defined(sun)
13195 	minor = (minor_t)(uintptr_t)vmem_alloc(dtrace_minor, 1,
13196 	    VM_BESTFIT | VM_SLEEP);
13197 
13198 	if (ddi_soft_state_zalloc(dtrace_softstate, minor) != DDI_SUCCESS) {
13199 		vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
13200 		return (NULL);
13201 	}
13202 
13203 	state = ddi_get_soft_state(dtrace_softstate, minor);
13204 #else
13205 	if (dev != NULL) {
13206 		cr = dev->si_cred;
13207 		m = dev2unit(dev);
13208 		}
13209 
13210 	/* Allocate memory for the state. */
13211 	state = kmem_zalloc(sizeof(dtrace_state_t), KM_SLEEP);
13212 #endif
13213 
13214 	state->dts_epid = DTRACE_EPIDNONE + 1;
13215 
13216 	(void) snprintf(c, sizeof (c), "dtrace_aggid_%d", m);
13217 #if defined(sun)
13218 	state->dts_aggid_arena = vmem_create(c, (void *)1, UINT32_MAX, 1,
13219 	    NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
13220 
13221 	if (devp != NULL) {
13222 		major = getemajor(*devp);
13223 	} else {
13224 		major = ddi_driver_major(dtrace_devi);
13225 	}
13226 
13227 	state->dts_dev = makedevice(major, minor);
13228 
13229 	if (devp != NULL)
13230 		*devp = state->dts_dev;
13231 #else
13232 	state->dts_aggid_arena = new_unrhdr(1, INT_MAX, &dtrace_unr_mtx);
13233 	state->dts_dev = dev;
13234 #endif
13235 
13236 	/*
13237 	 * We allocate NCPU buffers.  On the one hand, this can be quite
13238 	 * a bit of memory per instance (nearly 36K on a Starcat).  On the
13239 	 * other hand, it saves an additional memory reference in the probe
13240 	 * path.
13241 	 */
13242 	state->dts_buffer = kmem_zalloc(bufsize, KM_SLEEP);
13243 	state->dts_aggbuffer = kmem_zalloc(bufsize, KM_SLEEP);
13244 
13245 #if defined(sun)
13246 	state->dts_cleaner = CYCLIC_NONE;
13247 	state->dts_deadman = CYCLIC_NONE;
13248 #else
13249 	callout_init(&state->dts_cleaner, CALLOUT_MPSAFE);
13250 	callout_init(&state->dts_deadman, CALLOUT_MPSAFE);
13251 #endif
13252 	state->dts_vstate.dtvs_state = state;
13253 
13254 	for (i = 0; i < DTRACEOPT_MAX; i++)
13255 		state->dts_options[i] = DTRACEOPT_UNSET;
13256 
13257 	/*
13258 	 * Set the default options.
13259 	 */
13260 	opt = state->dts_options;
13261 	opt[DTRACEOPT_BUFPOLICY] = DTRACEOPT_BUFPOLICY_SWITCH;
13262 	opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_AUTO;
13263 	opt[DTRACEOPT_NSPEC] = dtrace_nspec_default;
13264 	opt[DTRACEOPT_SPECSIZE] = dtrace_specsize_default;
13265 	opt[DTRACEOPT_CPU] = (dtrace_optval_t)DTRACE_CPUALL;
13266 	opt[DTRACEOPT_STRSIZE] = dtrace_strsize_default;
13267 	opt[DTRACEOPT_STACKFRAMES] = dtrace_stackframes_default;
13268 	opt[DTRACEOPT_USTACKFRAMES] = dtrace_ustackframes_default;
13269 	opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_default;
13270 	opt[DTRACEOPT_AGGRATE] = dtrace_aggrate_default;
13271 	opt[DTRACEOPT_SWITCHRATE] = dtrace_switchrate_default;
13272 	opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_default;
13273 	opt[DTRACEOPT_JSTACKFRAMES] = dtrace_jstackframes_default;
13274 	opt[DTRACEOPT_JSTACKSTRSIZE] = dtrace_jstackstrsize_default;
13275 
13276 	state->dts_activity = DTRACE_ACTIVITY_INACTIVE;
13277 
13278 	/*
13279 	 * Depending on the user credentials, we set flag bits which alter probe
13280 	 * visibility or the amount of destructiveness allowed.  In the case of
13281 	 * actual anonymous tracing, or the possession of all privileges, all of
13282 	 * the normal checks are bypassed.
13283 	 */
13284 	if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
13285 		state->dts_cred.dcr_visible = DTRACE_CRV_ALL;
13286 		state->dts_cred.dcr_action = DTRACE_CRA_ALL;
13287 	} else {
13288 		/*
13289 		 * Set up the credentials for this instantiation.  We take a
13290 		 * hold on the credential to prevent it from disappearing on
13291 		 * us; this in turn prevents the zone_t referenced by this
13292 		 * credential from disappearing.  This means that we can
13293 		 * examine the credential and the zone from probe context.
13294 		 */
13295 		crhold(cr);
13296 		state->dts_cred.dcr_cred = cr;
13297 
13298 		/*
13299 		 * CRA_PROC means "we have *some* privilege for dtrace" and
13300 		 * unlocks the use of variables like pid, zonename, etc.
13301 		 */
13302 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE) ||
13303 		    PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
13304 			state->dts_cred.dcr_action |= DTRACE_CRA_PROC;
13305 		}
13306 
13307 		/*
13308 		 * dtrace_user allows use of syscall and profile providers.
13309 		 * If the user also has proc_owner and/or proc_zone, we
13310 		 * extend the scope to include additional visibility and
13311 		 * destructive power.
13312 		 */
13313 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE)) {
13314 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE)) {
13315 				state->dts_cred.dcr_visible |=
13316 				    DTRACE_CRV_ALLPROC;
13317 
13318 				state->dts_cred.dcr_action |=
13319 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
13320 			}
13321 
13322 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE)) {
13323 				state->dts_cred.dcr_visible |=
13324 				    DTRACE_CRV_ALLZONE;
13325 
13326 				state->dts_cred.dcr_action |=
13327 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
13328 			}
13329 
13330 			/*
13331 			 * If we have all privs in whatever zone this is,
13332 			 * we can do destructive things to processes which
13333 			 * have altered credentials.
13334 			 */
13335 #if defined(sun)
13336 			if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
13337 			    cr->cr_zone->zone_privset)) {
13338 				state->dts_cred.dcr_action |=
13339 				    DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
13340 			}
13341 #endif
13342 		}
13343 
13344 		/*
13345 		 * Holding the dtrace_kernel privilege also implies that
13346 		 * the user has the dtrace_user privilege from a visibility
13347 		 * perspective.  But without further privileges, some
13348 		 * destructive actions are not available.
13349 		 */
13350 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE)) {
13351 			/*
13352 			 * Make all probes in all zones visible.  However,
13353 			 * this doesn't mean that all actions become available
13354 			 * to all zones.
13355 			 */
13356 			state->dts_cred.dcr_visible |= DTRACE_CRV_KERNEL |
13357 			    DTRACE_CRV_ALLPROC | DTRACE_CRV_ALLZONE;
13358 
13359 			state->dts_cred.dcr_action |= DTRACE_CRA_KERNEL |
13360 			    DTRACE_CRA_PROC;
13361 			/*
13362 			 * Holding proc_owner means that destructive actions
13363 			 * for *this* zone are allowed.
13364 			 */
13365 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
13366 				state->dts_cred.dcr_action |=
13367 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
13368 
13369 			/*
13370 			 * Holding proc_zone means that destructive actions
13371 			 * for this user/group ID in all zones is allowed.
13372 			 */
13373 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
13374 				state->dts_cred.dcr_action |=
13375 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
13376 
13377 #if defined(sun)
13378 			/*
13379 			 * If we have all privs in whatever zone this is,
13380 			 * we can do destructive things to processes which
13381 			 * have altered credentials.
13382 			 */
13383 			if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
13384 			    cr->cr_zone->zone_privset)) {
13385 				state->dts_cred.dcr_action |=
13386 				    DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
13387 			}
13388 #endif
13389 		}
13390 
13391 		/*
13392 		 * Holding the dtrace_proc privilege gives control over fasttrap
13393 		 * and pid providers.  We need to grant wider destructive
13394 		 * privileges in the event that the user has proc_owner and/or
13395 		 * proc_zone.
13396 		 */
13397 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
13398 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
13399 				state->dts_cred.dcr_action |=
13400 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
13401 
13402 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
13403 				state->dts_cred.dcr_action |=
13404 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
13405 		}
13406 	}
13407 
13408 	return (state);
13409 }
13410 
13411 static int
13412 dtrace_state_buffer(dtrace_state_t *state, dtrace_buffer_t *buf, int which)
13413 {
13414 	dtrace_optval_t *opt = state->dts_options, size;
13415 	processorid_t cpu = 0;;
13416 	int flags = 0, rval;
13417 
13418 	ASSERT(MUTEX_HELD(&dtrace_lock));
13419 	ASSERT(MUTEX_HELD(&cpu_lock));
13420 	ASSERT(which < DTRACEOPT_MAX);
13421 	ASSERT(state->dts_activity == DTRACE_ACTIVITY_INACTIVE ||
13422 	    (state == dtrace_anon.dta_state &&
13423 	    state->dts_activity == DTRACE_ACTIVITY_ACTIVE));
13424 
13425 	if (opt[which] == DTRACEOPT_UNSET || opt[which] == 0)
13426 		return (0);
13427 
13428 	if (opt[DTRACEOPT_CPU] != DTRACEOPT_UNSET)
13429 		cpu = opt[DTRACEOPT_CPU];
13430 
13431 	if (which == DTRACEOPT_SPECSIZE)
13432 		flags |= DTRACEBUF_NOSWITCH;
13433 
13434 	if (which == DTRACEOPT_BUFSIZE) {
13435 		if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_RING)
13436 			flags |= DTRACEBUF_RING;
13437 
13438 		if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_FILL)
13439 			flags |= DTRACEBUF_FILL;
13440 
13441 		if (state != dtrace_anon.dta_state ||
13442 		    state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
13443 			flags |= DTRACEBUF_INACTIVE;
13444 	}
13445 
13446 	for (size = opt[which]; size >= sizeof (uint64_t); size >>= 1) {
13447 		/*
13448 		 * The size must be 8-byte aligned.  If the size is not 8-byte
13449 		 * aligned, drop it down by the difference.
13450 		 */
13451 		if (size & (sizeof (uint64_t) - 1))
13452 			size -= size & (sizeof (uint64_t) - 1);
13453 
13454 		if (size < state->dts_reserve) {
13455 			/*
13456 			 * Buffers always must be large enough to accommodate
13457 			 * their prereserved space.  We return E2BIG instead
13458 			 * of ENOMEM in this case to allow for user-level
13459 			 * software to differentiate the cases.
13460 			 */
13461 			return (E2BIG);
13462 		}
13463 
13464 		rval = dtrace_buffer_alloc(buf, size, flags, cpu);
13465 
13466 		if (rval != ENOMEM) {
13467 			opt[which] = size;
13468 			return (rval);
13469 		}
13470 
13471 		if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
13472 			return (rval);
13473 	}
13474 
13475 	return (ENOMEM);
13476 }
13477 
13478 static int
13479 dtrace_state_buffers(dtrace_state_t *state)
13480 {
13481 	dtrace_speculation_t *spec = state->dts_speculations;
13482 	int rval, i;
13483 
13484 	if ((rval = dtrace_state_buffer(state, state->dts_buffer,
13485 	    DTRACEOPT_BUFSIZE)) != 0)
13486 		return (rval);
13487 
13488 	if ((rval = dtrace_state_buffer(state, state->dts_aggbuffer,
13489 	    DTRACEOPT_AGGSIZE)) != 0)
13490 		return (rval);
13491 
13492 	for (i = 0; i < state->dts_nspeculations; i++) {
13493 		if ((rval = dtrace_state_buffer(state,
13494 		    spec[i].dtsp_buffer, DTRACEOPT_SPECSIZE)) != 0)
13495 			return (rval);
13496 	}
13497 
13498 	return (0);
13499 }
13500 
13501 static void
13502 dtrace_state_prereserve(dtrace_state_t *state)
13503 {
13504 	dtrace_ecb_t *ecb;
13505 	dtrace_probe_t *probe;
13506 
13507 	state->dts_reserve = 0;
13508 
13509 	if (state->dts_options[DTRACEOPT_BUFPOLICY] != DTRACEOPT_BUFPOLICY_FILL)
13510 		return;
13511 
13512 	/*
13513 	 * If our buffer policy is a "fill" buffer policy, we need to set the
13514 	 * prereserved space to be the space required by the END probes.
13515 	 */
13516 	probe = dtrace_probes[dtrace_probeid_end - 1];
13517 	ASSERT(probe != NULL);
13518 
13519 	for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
13520 		if (ecb->dte_state != state)
13521 			continue;
13522 
13523 		state->dts_reserve += ecb->dte_needed + ecb->dte_alignment;
13524 	}
13525 }
13526 
13527 static int
13528 dtrace_state_go(dtrace_state_t *state, processorid_t *cpu)
13529 {
13530 	dtrace_optval_t *opt = state->dts_options, sz, nspec;
13531 	dtrace_speculation_t *spec;
13532 	dtrace_buffer_t *buf;
13533 #if defined(sun)
13534 	cyc_handler_t hdlr;
13535 	cyc_time_t when;
13536 #endif
13537 	int rval = 0, i, bufsize = NCPU * sizeof (dtrace_buffer_t);
13538 	dtrace_icookie_t cookie;
13539 
13540 	mutex_enter(&cpu_lock);
13541 	mutex_enter(&dtrace_lock);
13542 
13543 	if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
13544 		rval = EBUSY;
13545 		goto out;
13546 	}
13547 
13548 	/*
13549 	 * Before we can perform any checks, we must prime all of the
13550 	 * retained enablings that correspond to this state.
13551 	 */
13552 	dtrace_enabling_prime(state);
13553 
13554 	if (state->dts_destructive && !state->dts_cred.dcr_destructive) {
13555 		rval = EACCES;
13556 		goto out;
13557 	}
13558 
13559 	dtrace_state_prereserve(state);
13560 
13561 	/*
13562 	 * Now we want to do is try to allocate our speculations.
13563 	 * We do not automatically resize the number of speculations; if
13564 	 * this fails, we will fail the operation.
13565 	 */
13566 	nspec = opt[DTRACEOPT_NSPEC];
13567 	ASSERT(nspec != DTRACEOPT_UNSET);
13568 
13569 	if (nspec > INT_MAX) {
13570 		rval = ENOMEM;
13571 		goto out;
13572 	}
13573 
13574 	spec = kmem_zalloc(nspec * sizeof (dtrace_speculation_t), KM_NOSLEEP);
13575 
13576 	if (spec == NULL) {
13577 		rval = ENOMEM;
13578 		goto out;
13579 	}
13580 
13581 	state->dts_speculations = spec;
13582 	state->dts_nspeculations = (int)nspec;
13583 
13584 	for (i = 0; i < nspec; i++) {
13585 		if ((buf = kmem_zalloc(bufsize, KM_NOSLEEP)) == NULL) {
13586 			rval = ENOMEM;
13587 			goto err;
13588 		}
13589 
13590 		spec[i].dtsp_buffer = buf;
13591 	}
13592 
13593 	if (opt[DTRACEOPT_GRABANON] != DTRACEOPT_UNSET) {
13594 		if (dtrace_anon.dta_state == NULL) {
13595 			rval = ENOENT;
13596 			goto out;
13597 		}
13598 
13599 		if (state->dts_necbs != 0) {
13600 			rval = EALREADY;
13601 			goto out;
13602 		}
13603 
13604 		state->dts_anon = dtrace_anon_grab();
13605 		ASSERT(state->dts_anon != NULL);
13606 		state = state->dts_anon;
13607 
13608 		/*
13609 		 * We want "grabanon" to be set in the grabbed state, so we'll
13610 		 * copy that option value from the grabbing state into the
13611 		 * grabbed state.
13612 		 */
13613 		state->dts_options[DTRACEOPT_GRABANON] =
13614 		    opt[DTRACEOPT_GRABANON];
13615 
13616 		*cpu = dtrace_anon.dta_beganon;
13617 
13618 		/*
13619 		 * If the anonymous state is active (as it almost certainly
13620 		 * is if the anonymous enabling ultimately matched anything),
13621 		 * we don't allow any further option processing -- but we
13622 		 * don't return failure.
13623 		 */
13624 		if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
13625 			goto out;
13626 	}
13627 
13628 	if (opt[DTRACEOPT_AGGSIZE] != DTRACEOPT_UNSET &&
13629 	    opt[DTRACEOPT_AGGSIZE] != 0) {
13630 		if (state->dts_aggregations == NULL) {
13631 			/*
13632 			 * We're not going to create an aggregation buffer
13633 			 * because we don't have any ECBs that contain
13634 			 * aggregations -- set this option to 0.
13635 			 */
13636 			opt[DTRACEOPT_AGGSIZE] = 0;
13637 		} else {
13638 			/*
13639 			 * If we have an aggregation buffer, we must also have
13640 			 * a buffer to use as scratch.
13641 			 */
13642 			if (opt[DTRACEOPT_BUFSIZE] == DTRACEOPT_UNSET ||
13643 			    opt[DTRACEOPT_BUFSIZE] < state->dts_needed) {
13644 				opt[DTRACEOPT_BUFSIZE] = state->dts_needed;
13645 			}
13646 		}
13647 	}
13648 
13649 	if (opt[DTRACEOPT_SPECSIZE] != DTRACEOPT_UNSET &&
13650 	    opt[DTRACEOPT_SPECSIZE] != 0) {
13651 		if (!state->dts_speculates) {
13652 			/*
13653 			 * We're not going to create speculation buffers
13654 			 * because we don't have any ECBs that actually
13655 			 * speculate -- set the speculation size to 0.
13656 			 */
13657 			opt[DTRACEOPT_SPECSIZE] = 0;
13658 		}
13659 	}
13660 
13661 	/*
13662 	 * The bare minimum size for any buffer that we're actually going to
13663 	 * do anything to is sizeof (uint64_t).
13664 	 */
13665 	sz = sizeof (uint64_t);
13666 
13667 	if ((state->dts_needed != 0 && opt[DTRACEOPT_BUFSIZE] < sz) ||
13668 	    (state->dts_speculates && opt[DTRACEOPT_SPECSIZE] < sz) ||
13669 	    (state->dts_aggregations != NULL && opt[DTRACEOPT_AGGSIZE] < sz)) {
13670 		/*
13671 		 * A buffer size has been explicitly set to 0 (or to a size
13672 		 * that will be adjusted to 0) and we need the space -- we
13673 		 * need to return failure.  We return ENOSPC to differentiate
13674 		 * it from failing to allocate a buffer due to failure to meet
13675 		 * the reserve (for which we return E2BIG).
13676 		 */
13677 		rval = ENOSPC;
13678 		goto out;
13679 	}
13680 
13681 	if ((rval = dtrace_state_buffers(state)) != 0)
13682 		goto err;
13683 
13684 	if ((sz = opt[DTRACEOPT_DYNVARSIZE]) == DTRACEOPT_UNSET)
13685 		sz = dtrace_dstate_defsize;
13686 
13687 	do {
13688 		rval = dtrace_dstate_init(&state->dts_vstate.dtvs_dynvars, sz);
13689 
13690 		if (rval == 0)
13691 			break;
13692 
13693 		if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
13694 			goto err;
13695 	} while (sz >>= 1);
13696 
13697 	opt[DTRACEOPT_DYNVARSIZE] = sz;
13698 
13699 	if (rval != 0)
13700 		goto err;
13701 
13702 	if (opt[DTRACEOPT_STATUSRATE] > dtrace_statusrate_max)
13703 		opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_max;
13704 
13705 	if (opt[DTRACEOPT_CLEANRATE] == 0)
13706 		opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
13707 
13708 	if (opt[DTRACEOPT_CLEANRATE] < dtrace_cleanrate_min)
13709 		opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_min;
13710 
13711 	if (opt[DTRACEOPT_CLEANRATE] > dtrace_cleanrate_max)
13712 		opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
13713 
13714 	state->dts_alive = state->dts_laststatus = dtrace_gethrtime();
13715 #if defined(sun)
13716 	hdlr.cyh_func = (cyc_func_t)dtrace_state_clean;
13717 	hdlr.cyh_arg = state;
13718 	hdlr.cyh_level = CY_LOW_LEVEL;
13719 
13720 	when.cyt_when = 0;
13721 	when.cyt_interval = opt[DTRACEOPT_CLEANRATE];
13722 
13723 	state->dts_cleaner = cyclic_add(&hdlr, &when);
13724 
13725 	hdlr.cyh_func = (cyc_func_t)dtrace_state_deadman;
13726 	hdlr.cyh_arg = state;
13727 	hdlr.cyh_level = CY_LOW_LEVEL;
13728 
13729 	when.cyt_when = 0;
13730 	when.cyt_interval = dtrace_deadman_interval;
13731 
13732 	state->dts_deadman = cyclic_add(&hdlr, &when);
13733 #else
13734 	callout_reset(&state->dts_cleaner, hz * opt[DTRACEOPT_CLEANRATE] / NANOSEC,
13735 	    dtrace_state_clean, state);
13736 	callout_reset(&state->dts_deadman, hz * dtrace_deadman_interval / NANOSEC,
13737 	    dtrace_state_deadman, state);
13738 #endif
13739 
13740 	state->dts_activity = DTRACE_ACTIVITY_WARMUP;
13741 
13742 	/*
13743 	 * Now it's time to actually fire the BEGIN probe.  We need to disable
13744 	 * interrupts here both to record the CPU on which we fired the BEGIN
13745 	 * probe (the data from this CPU will be processed first at user
13746 	 * level) and to manually activate the buffer for this CPU.
13747 	 */
13748 	cookie = dtrace_interrupt_disable();
13749 	*cpu = curcpu;
13750 	ASSERT(state->dts_buffer[*cpu].dtb_flags & DTRACEBUF_INACTIVE);
13751 	state->dts_buffer[*cpu].dtb_flags &= ~DTRACEBUF_INACTIVE;
13752 
13753 	dtrace_probe(dtrace_probeid_begin,
13754 	    (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
13755 	dtrace_interrupt_enable(cookie);
13756 	/*
13757 	 * We may have had an exit action from a BEGIN probe; only change our
13758 	 * state to ACTIVE if we're still in WARMUP.
13759 	 */
13760 	ASSERT(state->dts_activity == DTRACE_ACTIVITY_WARMUP ||
13761 	    state->dts_activity == DTRACE_ACTIVITY_DRAINING);
13762 
13763 	if (state->dts_activity == DTRACE_ACTIVITY_WARMUP)
13764 		state->dts_activity = DTRACE_ACTIVITY_ACTIVE;
13765 
13766 	/*
13767 	 * Regardless of whether or not now we're in ACTIVE or DRAINING, we
13768 	 * want each CPU to transition its principal buffer out of the
13769 	 * INACTIVE state.  Doing this assures that no CPU will suddenly begin
13770 	 * processing an ECB halfway down a probe's ECB chain; all CPUs will
13771 	 * atomically transition from processing none of a state's ECBs to
13772 	 * processing all of them.
13773 	 */
13774 	dtrace_xcall(DTRACE_CPUALL,
13775 	    (dtrace_xcall_t)dtrace_buffer_activate, state);
13776 	goto out;
13777 
13778 err:
13779 	dtrace_buffer_free(state->dts_buffer);
13780 	dtrace_buffer_free(state->dts_aggbuffer);
13781 
13782 	if ((nspec = state->dts_nspeculations) == 0) {
13783 		ASSERT(state->dts_speculations == NULL);
13784 		goto out;
13785 	}
13786 
13787 	spec = state->dts_speculations;
13788 	ASSERT(spec != NULL);
13789 
13790 	for (i = 0; i < state->dts_nspeculations; i++) {
13791 		if ((buf = spec[i].dtsp_buffer) == NULL)
13792 			break;
13793 
13794 		dtrace_buffer_free(buf);
13795 		kmem_free(buf, bufsize);
13796 	}
13797 
13798 	kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
13799 	state->dts_nspeculations = 0;
13800 	state->dts_speculations = NULL;
13801 
13802 out:
13803 	mutex_exit(&dtrace_lock);
13804 	mutex_exit(&cpu_lock);
13805 
13806 	return (rval);
13807 }
13808 
13809 static int
13810 dtrace_state_stop(dtrace_state_t *state, processorid_t *cpu)
13811 {
13812 	dtrace_icookie_t cookie;
13813 
13814 	ASSERT(MUTEX_HELD(&dtrace_lock));
13815 
13816 	if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE &&
13817 	    state->dts_activity != DTRACE_ACTIVITY_DRAINING)
13818 		return (EINVAL);
13819 
13820 	/*
13821 	 * We'll set the activity to DTRACE_ACTIVITY_DRAINING, and issue a sync
13822 	 * to be sure that every CPU has seen it.  See below for the details
13823 	 * on why this is done.
13824 	 */
13825 	state->dts_activity = DTRACE_ACTIVITY_DRAINING;
13826 	dtrace_sync();
13827 
13828 	/*
13829 	 * By this point, it is impossible for any CPU to be still processing
13830 	 * with DTRACE_ACTIVITY_ACTIVE.  We can thus set our activity to
13831 	 * DTRACE_ACTIVITY_COOLDOWN and know that we're not racing with any
13832 	 * other CPU in dtrace_buffer_reserve().  This allows dtrace_probe()
13833 	 * and callees to know that the activity is DTRACE_ACTIVITY_COOLDOWN
13834 	 * iff we're in the END probe.
13835 	 */
13836 	state->dts_activity = DTRACE_ACTIVITY_COOLDOWN;
13837 	dtrace_sync();
13838 	ASSERT(state->dts_activity == DTRACE_ACTIVITY_COOLDOWN);
13839 
13840 	/*
13841 	 * Finally, we can release the reserve and call the END probe.  We
13842 	 * disable interrupts across calling the END probe to allow us to
13843 	 * return the CPU on which we actually called the END probe.  This
13844 	 * allows user-land to be sure that this CPU's principal buffer is
13845 	 * processed last.
13846 	 */
13847 	state->dts_reserve = 0;
13848 
13849 	cookie = dtrace_interrupt_disable();
13850 	*cpu = curcpu;
13851 	dtrace_probe(dtrace_probeid_end,
13852 	    (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
13853 	dtrace_interrupt_enable(cookie);
13854 
13855 	state->dts_activity = DTRACE_ACTIVITY_STOPPED;
13856 	dtrace_sync();
13857 
13858 	return (0);
13859 }
13860 
13861 static int
13862 dtrace_state_option(dtrace_state_t *state, dtrace_optid_t option,
13863     dtrace_optval_t val)
13864 {
13865 	ASSERT(MUTEX_HELD(&dtrace_lock));
13866 
13867 	if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
13868 		return (EBUSY);
13869 
13870 	if (option >= DTRACEOPT_MAX)
13871 		return (EINVAL);
13872 
13873 	if (option != DTRACEOPT_CPU && val < 0)
13874 		return (EINVAL);
13875 
13876 	switch (option) {
13877 	case DTRACEOPT_DESTRUCTIVE:
13878 		if (dtrace_destructive_disallow)
13879 			return (EACCES);
13880 
13881 		state->dts_cred.dcr_destructive = 1;
13882 		break;
13883 
13884 	case DTRACEOPT_BUFSIZE:
13885 	case DTRACEOPT_DYNVARSIZE:
13886 	case DTRACEOPT_AGGSIZE:
13887 	case DTRACEOPT_SPECSIZE:
13888 	case DTRACEOPT_STRSIZE:
13889 		if (val < 0)
13890 			return (EINVAL);
13891 
13892 		if (val >= LONG_MAX) {
13893 			/*
13894 			 * If this is an otherwise negative value, set it to
13895 			 * the highest multiple of 128m less than LONG_MAX.
13896 			 * Technically, we're adjusting the size without
13897 			 * regard to the buffer resizing policy, but in fact,
13898 			 * this has no effect -- if we set the buffer size to
13899 			 * ~LONG_MAX and the buffer policy is ultimately set to
13900 			 * be "manual", the buffer allocation is guaranteed to
13901 			 * fail, if only because the allocation requires two
13902 			 * buffers.  (We set the the size to the highest
13903 			 * multiple of 128m because it ensures that the size
13904 			 * will remain a multiple of a megabyte when
13905 			 * repeatedly halved -- all the way down to 15m.)
13906 			 */
13907 			val = LONG_MAX - (1 << 27) + 1;
13908 		}
13909 	}
13910 
13911 	state->dts_options[option] = val;
13912 
13913 	return (0);
13914 }
13915 
13916 static void
13917 dtrace_state_destroy(dtrace_state_t *state)
13918 {
13919 	dtrace_ecb_t *ecb;
13920 	dtrace_vstate_t *vstate = &state->dts_vstate;
13921 #if defined(sun)
13922 	minor_t minor = getminor(state->dts_dev);
13923 #endif
13924 	int i, bufsize = NCPU * sizeof (dtrace_buffer_t);
13925 	dtrace_speculation_t *spec = state->dts_speculations;
13926 	int nspec = state->dts_nspeculations;
13927 	uint32_t match;
13928 
13929 	ASSERT(MUTEX_HELD(&dtrace_lock));
13930 	ASSERT(MUTEX_HELD(&cpu_lock));
13931 
13932 	/*
13933 	 * First, retract any retained enablings for this state.
13934 	 */
13935 	dtrace_enabling_retract(state);
13936 	ASSERT(state->dts_nretained == 0);
13937 
13938 	if (state->dts_activity == DTRACE_ACTIVITY_ACTIVE ||
13939 	    state->dts_activity == DTRACE_ACTIVITY_DRAINING) {
13940 		/*
13941 		 * We have managed to come into dtrace_state_destroy() on a
13942 		 * hot enabling -- almost certainly because of a disorderly
13943 		 * shutdown of a consumer.  (That is, a consumer that is
13944 		 * exiting without having called dtrace_stop().) In this case,
13945 		 * we're going to set our activity to be KILLED, and then
13946 		 * issue a sync to be sure that everyone is out of probe
13947 		 * context before we start blowing away ECBs.
13948 		 */
13949 		state->dts_activity = DTRACE_ACTIVITY_KILLED;
13950 		dtrace_sync();
13951 	}
13952 
13953 	/*
13954 	 * Release the credential hold we took in dtrace_state_create().
13955 	 */
13956 	if (state->dts_cred.dcr_cred != NULL)
13957 		crfree(state->dts_cred.dcr_cred);
13958 
13959 	/*
13960 	 * Now we can safely disable and destroy any enabled probes.  Because
13961 	 * any DTRACE_PRIV_KERNEL probes may actually be slowing our progress
13962 	 * (especially if they're all enabled), we take two passes through the
13963 	 * ECBs:  in the first, we disable just DTRACE_PRIV_KERNEL probes, and
13964 	 * in the second we disable whatever is left over.
13965 	 */
13966 	for (match = DTRACE_PRIV_KERNEL; ; match = 0) {
13967 		for (i = 0; i < state->dts_necbs; i++) {
13968 			if ((ecb = state->dts_ecbs[i]) == NULL)
13969 				continue;
13970 
13971 			if (match && ecb->dte_probe != NULL) {
13972 				dtrace_probe_t *probe = ecb->dte_probe;
13973 				dtrace_provider_t *prov = probe->dtpr_provider;
13974 
13975 				if (!(prov->dtpv_priv.dtpp_flags & match))
13976 					continue;
13977 			}
13978 
13979 			dtrace_ecb_disable(ecb);
13980 			dtrace_ecb_destroy(ecb);
13981 		}
13982 
13983 		if (!match)
13984 			break;
13985 	}
13986 
13987 	/*
13988 	 * Before we free the buffers, perform one more sync to assure that
13989 	 * every CPU is out of probe context.
13990 	 */
13991 	dtrace_sync();
13992 
13993 	dtrace_buffer_free(state->dts_buffer);
13994 	dtrace_buffer_free(state->dts_aggbuffer);
13995 
13996 	for (i = 0; i < nspec; i++)
13997 		dtrace_buffer_free(spec[i].dtsp_buffer);
13998 
13999 #if defined(sun)
14000 	if (state->dts_cleaner != CYCLIC_NONE)
14001 		cyclic_remove(state->dts_cleaner);
14002 
14003 	if (state->dts_deadman != CYCLIC_NONE)
14004 		cyclic_remove(state->dts_deadman);
14005 #else
14006 	callout_stop(&state->dts_cleaner);
14007 	callout_drain(&state->dts_cleaner);
14008 	callout_stop(&state->dts_deadman);
14009 	callout_drain(&state->dts_deadman);
14010 #endif
14011 
14012 	dtrace_dstate_fini(&vstate->dtvs_dynvars);
14013 	dtrace_vstate_fini(vstate);
14014 	if (state->dts_ecbs != NULL)
14015 		kmem_free(state->dts_ecbs, state->dts_necbs * sizeof (dtrace_ecb_t *));
14016 
14017 	if (state->dts_aggregations != NULL) {
14018 #ifdef DEBUG
14019 		for (i = 0; i < state->dts_naggregations; i++)
14020 			ASSERT(state->dts_aggregations[i] == NULL);
14021 #endif
14022 		ASSERT(state->dts_naggregations > 0);
14023 		kmem_free(state->dts_aggregations,
14024 		    state->dts_naggregations * sizeof (dtrace_aggregation_t *));
14025 	}
14026 
14027 	kmem_free(state->dts_buffer, bufsize);
14028 	kmem_free(state->dts_aggbuffer, bufsize);
14029 
14030 	for (i = 0; i < nspec; i++)
14031 		kmem_free(spec[i].dtsp_buffer, bufsize);
14032 
14033 	if (spec != NULL)
14034 		kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
14035 
14036 	dtrace_format_destroy(state);
14037 
14038 	if (state->dts_aggid_arena != NULL) {
14039 #if defined(sun)
14040 		vmem_destroy(state->dts_aggid_arena);
14041 #else
14042 		delete_unrhdr(state->dts_aggid_arena);
14043 #endif
14044 		state->dts_aggid_arena = NULL;
14045 	}
14046 #if defined(sun)
14047 	ddi_soft_state_free(dtrace_softstate, minor);
14048 	vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
14049 #endif
14050 }
14051 
14052 /*
14053  * DTrace Anonymous Enabling Functions
14054  */
14055 static dtrace_state_t *
14056 dtrace_anon_grab(void)
14057 {
14058 	dtrace_state_t *state;
14059 
14060 	ASSERT(MUTEX_HELD(&dtrace_lock));
14061 
14062 	if ((state = dtrace_anon.dta_state) == NULL) {
14063 		ASSERT(dtrace_anon.dta_enabling == NULL);
14064 		return (NULL);
14065 	}
14066 
14067 	ASSERT(dtrace_anon.dta_enabling != NULL);
14068 	ASSERT(dtrace_retained != NULL);
14069 
14070 	dtrace_enabling_destroy(dtrace_anon.dta_enabling);
14071 	dtrace_anon.dta_enabling = NULL;
14072 	dtrace_anon.dta_state = NULL;
14073 
14074 	return (state);
14075 }
14076 
14077 static void
14078 dtrace_anon_property(void)
14079 {
14080 	int i, rv;
14081 	dtrace_state_t *state;
14082 	dof_hdr_t *dof;
14083 	char c[32];		/* enough for "dof-data-" + digits */
14084 
14085 	ASSERT(MUTEX_HELD(&dtrace_lock));
14086 	ASSERT(MUTEX_HELD(&cpu_lock));
14087 
14088 	for (i = 0; ; i++) {
14089 		(void) snprintf(c, sizeof (c), "dof-data-%d", i);
14090 
14091 		dtrace_err_verbose = 1;
14092 
14093 		if ((dof = dtrace_dof_property(c)) == NULL) {
14094 			dtrace_err_verbose = 0;
14095 			break;
14096 		}
14097 
14098 #if defined(sun)
14099 		/*
14100 		 * We want to create anonymous state, so we need to transition
14101 		 * the kernel debugger to indicate that DTrace is active.  If
14102 		 * this fails (e.g. because the debugger has modified text in
14103 		 * some way), we won't continue with the processing.
14104 		 */
14105 		if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
14106 			cmn_err(CE_NOTE, "kernel debugger active; anonymous "
14107 			    "enabling ignored.");
14108 			dtrace_dof_destroy(dof);
14109 			break;
14110 		}
14111 #endif
14112 
14113 		/*
14114 		 * If we haven't allocated an anonymous state, we'll do so now.
14115 		 */
14116 		if ((state = dtrace_anon.dta_state) == NULL) {
14117 #if defined(sun)
14118 			state = dtrace_state_create(NULL, NULL);
14119 #else
14120 			state = dtrace_state_create(NULL);
14121 #endif
14122 			dtrace_anon.dta_state = state;
14123 
14124 			if (state == NULL) {
14125 				/*
14126 				 * This basically shouldn't happen:  the only
14127 				 * failure mode from dtrace_state_create() is a
14128 				 * failure of ddi_soft_state_zalloc() that
14129 				 * itself should never happen.  Still, the
14130 				 * interface allows for a failure mode, and
14131 				 * we want to fail as gracefully as possible:
14132 				 * we'll emit an error message and cease
14133 				 * processing anonymous state in this case.
14134 				 */
14135 				cmn_err(CE_WARN, "failed to create "
14136 				    "anonymous state");
14137 				dtrace_dof_destroy(dof);
14138 				break;
14139 			}
14140 		}
14141 
14142 		rv = dtrace_dof_slurp(dof, &state->dts_vstate, CRED(),
14143 		    &dtrace_anon.dta_enabling, 0, B_TRUE);
14144 
14145 		if (rv == 0)
14146 			rv = dtrace_dof_options(dof, state);
14147 
14148 		dtrace_err_verbose = 0;
14149 		dtrace_dof_destroy(dof);
14150 
14151 		if (rv != 0) {
14152 			/*
14153 			 * This is malformed DOF; chuck any anonymous state
14154 			 * that we created.
14155 			 */
14156 			ASSERT(dtrace_anon.dta_enabling == NULL);
14157 			dtrace_state_destroy(state);
14158 			dtrace_anon.dta_state = NULL;
14159 			break;
14160 		}
14161 
14162 		ASSERT(dtrace_anon.dta_enabling != NULL);
14163 	}
14164 
14165 	if (dtrace_anon.dta_enabling != NULL) {
14166 		int rval;
14167 
14168 		/*
14169 		 * dtrace_enabling_retain() can only fail because we are
14170 		 * trying to retain more enablings than are allowed -- but
14171 		 * we only have one anonymous enabling, and we are guaranteed
14172 		 * to be allowed at least one retained enabling; we assert
14173 		 * that dtrace_enabling_retain() returns success.
14174 		 */
14175 		rval = dtrace_enabling_retain(dtrace_anon.dta_enabling);
14176 		ASSERT(rval == 0);
14177 
14178 		dtrace_enabling_dump(dtrace_anon.dta_enabling);
14179 	}
14180 }
14181 
14182 /*
14183  * DTrace Helper Functions
14184  */
14185 static void
14186 dtrace_helper_trace(dtrace_helper_action_t *helper,
14187     dtrace_mstate_t *mstate, dtrace_vstate_t *vstate, int where)
14188 {
14189 	uint32_t size, next, nnext, i;
14190 	dtrace_helptrace_t *ent;
14191 	uint16_t flags = cpu_core[curcpu].cpuc_dtrace_flags;
14192 
14193 	if (!dtrace_helptrace_enabled)
14194 		return;
14195 
14196 	ASSERT(vstate->dtvs_nlocals <= dtrace_helptrace_nlocals);
14197 
14198 	/*
14199 	 * What would a tracing framework be without its own tracing
14200 	 * framework?  (Well, a hell of a lot simpler, for starters...)
14201 	 */
14202 	size = sizeof (dtrace_helptrace_t) + dtrace_helptrace_nlocals *
14203 	    sizeof (uint64_t) - sizeof (uint64_t);
14204 
14205 	/*
14206 	 * Iterate until we can allocate a slot in the trace buffer.
14207 	 */
14208 	do {
14209 		next = dtrace_helptrace_next;
14210 
14211 		if (next + size < dtrace_helptrace_bufsize) {
14212 			nnext = next + size;
14213 		} else {
14214 			nnext = size;
14215 		}
14216 	} while (dtrace_cas32(&dtrace_helptrace_next, next, nnext) != next);
14217 
14218 	/*
14219 	 * We have our slot; fill it in.
14220 	 */
14221 	if (nnext == size)
14222 		next = 0;
14223 
14224 	ent = (dtrace_helptrace_t *)&dtrace_helptrace_buffer[next];
14225 	ent->dtht_helper = helper;
14226 	ent->dtht_where = where;
14227 	ent->dtht_nlocals = vstate->dtvs_nlocals;
14228 
14229 	ent->dtht_fltoffs = (mstate->dtms_present & DTRACE_MSTATE_FLTOFFS) ?
14230 	    mstate->dtms_fltoffs : -1;
14231 	ent->dtht_fault = DTRACE_FLAGS2FLT(flags);
14232 	ent->dtht_illval = cpu_core[curcpu].cpuc_dtrace_illval;
14233 
14234 	for (i = 0; i < vstate->dtvs_nlocals; i++) {
14235 		dtrace_statvar_t *svar;
14236 
14237 		if ((svar = vstate->dtvs_locals[i]) == NULL)
14238 			continue;
14239 
14240 		ASSERT(svar->dtsv_size >= NCPU * sizeof (uint64_t));
14241 		ent->dtht_locals[i] =
14242 		    ((uint64_t *)(uintptr_t)svar->dtsv_data)[curcpu];
14243 	}
14244 }
14245 
14246 static uint64_t
14247 dtrace_helper(int which, dtrace_mstate_t *mstate,
14248     dtrace_state_t *state, uint64_t arg0, uint64_t arg1)
14249 {
14250 	uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
14251 	uint64_t sarg0 = mstate->dtms_arg[0];
14252 	uint64_t sarg1 = mstate->dtms_arg[1];
14253 	uint64_t rval = 0;
14254 	dtrace_helpers_t *helpers = curproc->p_dtrace_helpers;
14255 	dtrace_helper_action_t *helper;
14256 	dtrace_vstate_t *vstate;
14257 	dtrace_difo_t *pred;
14258 	int i, trace = dtrace_helptrace_enabled;
14259 
14260 	ASSERT(which >= 0 && which < DTRACE_NHELPER_ACTIONS);
14261 
14262 	if (helpers == NULL)
14263 		return (0);
14264 
14265 	if ((helper = helpers->dthps_actions[which]) == NULL)
14266 		return (0);
14267 
14268 	vstate = &helpers->dthps_vstate;
14269 	mstate->dtms_arg[0] = arg0;
14270 	mstate->dtms_arg[1] = arg1;
14271 
14272 	/*
14273 	 * Now iterate over each helper.  If its predicate evaluates to 'true',
14274 	 * we'll call the corresponding actions.  Note that the below calls
14275 	 * to dtrace_dif_emulate() may set faults in machine state.  This is
14276 	 * okay:  our caller (the outer dtrace_dif_emulate()) will simply plow
14277 	 * the stored DIF offset with its own (which is the desired behavior).
14278 	 * Also, note the calls to dtrace_dif_emulate() may allocate scratch
14279 	 * from machine state; this is okay, too.
14280 	 */
14281 	for (; helper != NULL; helper = helper->dtha_next) {
14282 		if ((pred = helper->dtha_predicate) != NULL) {
14283 			if (trace)
14284 				dtrace_helper_trace(helper, mstate, vstate, 0);
14285 
14286 			if (!dtrace_dif_emulate(pred, mstate, vstate, state))
14287 				goto next;
14288 
14289 			if (*flags & CPU_DTRACE_FAULT)
14290 				goto err;
14291 		}
14292 
14293 		for (i = 0; i < helper->dtha_nactions; i++) {
14294 			if (trace)
14295 				dtrace_helper_trace(helper,
14296 				    mstate, vstate, i + 1);
14297 
14298 			rval = dtrace_dif_emulate(helper->dtha_actions[i],
14299 			    mstate, vstate, state);
14300 
14301 			if (*flags & CPU_DTRACE_FAULT)
14302 				goto err;
14303 		}
14304 
14305 next:
14306 		if (trace)
14307 			dtrace_helper_trace(helper, mstate, vstate,
14308 			    DTRACE_HELPTRACE_NEXT);
14309 	}
14310 
14311 	if (trace)
14312 		dtrace_helper_trace(helper, mstate, vstate,
14313 		    DTRACE_HELPTRACE_DONE);
14314 
14315 	/*
14316 	 * Restore the arg0 that we saved upon entry.
14317 	 */
14318 	mstate->dtms_arg[0] = sarg0;
14319 	mstate->dtms_arg[1] = sarg1;
14320 
14321 	return (rval);
14322 
14323 err:
14324 	if (trace)
14325 		dtrace_helper_trace(helper, mstate, vstate,
14326 		    DTRACE_HELPTRACE_ERR);
14327 
14328 	/*
14329 	 * Restore the arg0 that we saved upon entry.
14330 	 */
14331 	mstate->dtms_arg[0] = sarg0;
14332 	mstate->dtms_arg[1] = sarg1;
14333 
14334 	return (0);
14335 }
14336 
14337 static void
14338 dtrace_helper_action_destroy(dtrace_helper_action_t *helper,
14339     dtrace_vstate_t *vstate)
14340 {
14341 	int i;
14342 
14343 	if (helper->dtha_predicate != NULL)
14344 		dtrace_difo_release(helper->dtha_predicate, vstate);
14345 
14346 	for (i = 0; i < helper->dtha_nactions; i++) {
14347 		ASSERT(helper->dtha_actions[i] != NULL);
14348 		dtrace_difo_release(helper->dtha_actions[i], vstate);
14349 	}
14350 
14351 	kmem_free(helper->dtha_actions,
14352 	    helper->dtha_nactions * sizeof (dtrace_difo_t *));
14353 	kmem_free(helper, sizeof (dtrace_helper_action_t));
14354 }
14355 
14356 static int
14357 dtrace_helper_destroygen(int gen)
14358 {
14359 	proc_t *p = curproc;
14360 	dtrace_helpers_t *help = p->p_dtrace_helpers;
14361 	dtrace_vstate_t *vstate;
14362 	int i;
14363 
14364 	ASSERT(MUTEX_HELD(&dtrace_lock));
14365 
14366 	if (help == NULL || gen > help->dthps_generation)
14367 		return (EINVAL);
14368 
14369 	vstate = &help->dthps_vstate;
14370 
14371 	for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
14372 		dtrace_helper_action_t *last = NULL, *h, *next;
14373 
14374 		for (h = help->dthps_actions[i]; h != NULL; h = next) {
14375 			next = h->dtha_next;
14376 
14377 			if (h->dtha_generation == gen) {
14378 				if (last != NULL) {
14379 					last->dtha_next = next;
14380 				} else {
14381 					help->dthps_actions[i] = next;
14382 				}
14383 
14384 				dtrace_helper_action_destroy(h, vstate);
14385 			} else {
14386 				last = h;
14387 			}
14388 		}
14389 	}
14390 
14391 	/*
14392 	 * Interate until we've cleared out all helper providers with the
14393 	 * given generation number.
14394 	 */
14395 	for (;;) {
14396 		dtrace_helper_provider_t *prov;
14397 
14398 		/*
14399 		 * Look for a helper provider with the right generation. We
14400 		 * have to start back at the beginning of the list each time
14401 		 * because we drop dtrace_lock. It's unlikely that we'll make
14402 		 * more than two passes.
14403 		 */
14404 		for (i = 0; i < help->dthps_nprovs; i++) {
14405 			prov = help->dthps_provs[i];
14406 
14407 			if (prov->dthp_generation == gen)
14408 				break;
14409 		}
14410 
14411 		/*
14412 		 * If there were no matches, we're done.
14413 		 */
14414 		if (i == help->dthps_nprovs)
14415 			break;
14416 
14417 		/*
14418 		 * Move the last helper provider into this slot.
14419 		 */
14420 		help->dthps_nprovs--;
14421 		help->dthps_provs[i] = help->dthps_provs[help->dthps_nprovs];
14422 		help->dthps_provs[help->dthps_nprovs] = NULL;
14423 
14424 		mutex_exit(&dtrace_lock);
14425 
14426 		/*
14427 		 * If we have a meta provider, remove this helper provider.
14428 		 */
14429 		mutex_enter(&dtrace_meta_lock);
14430 		if (dtrace_meta_pid != NULL) {
14431 			ASSERT(dtrace_deferred_pid == NULL);
14432 			dtrace_helper_provider_remove(&prov->dthp_prov,
14433 			    p->p_pid);
14434 		}
14435 		mutex_exit(&dtrace_meta_lock);
14436 
14437 		dtrace_helper_provider_destroy(prov);
14438 
14439 		mutex_enter(&dtrace_lock);
14440 	}
14441 
14442 	return (0);
14443 }
14444 
14445 static int
14446 dtrace_helper_validate(dtrace_helper_action_t *helper)
14447 {
14448 	int err = 0, i;
14449 	dtrace_difo_t *dp;
14450 
14451 	if ((dp = helper->dtha_predicate) != NULL)
14452 		err += dtrace_difo_validate_helper(dp);
14453 
14454 	for (i = 0; i < helper->dtha_nactions; i++)
14455 		err += dtrace_difo_validate_helper(helper->dtha_actions[i]);
14456 
14457 	return (err == 0);
14458 }
14459 
14460 static int
14461 dtrace_helper_action_add(int which, dtrace_ecbdesc_t *ep)
14462 {
14463 	dtrace_helpers_t *help;
14464 	dtrace_helper_action_t *helper, *last;
14465 	dtrace_actdesc_t *act;
14466 	dtrace_vstate_t *vstate;
14467 	dtrace_predicate_t *pred;
14468 	int count = 0, nactions = 0, i;
14469 
14470 	if (which < 0 || which >= DTRACE_NHELPER_ACTIONS)
14471 		return (EINVAL);
14472 
14473 	help = curproc->p_dtrace_helpers;
14474 	last = help->dthps_actions[which];
14475 	vstate = &help->dthps_vstate;
14476 
14477 	for (count = 0; last != NULL; last = last->dtha_next) {
14478 		count++;
14479 		if (last->dtha_next == NULL)
14480 			break;
14481 	}
14482 
14483 	/*
14484 	 * If we already have dtrace_helper_actions_max helper actions for this
14485 	 * helper action type, we'll refuse to add a new one.
14486 	 */
14487 	if (count >= dtrace_helper_actions_max)
14488 		return (ENOSPC);
14489 
14490 	helper = kmem_zalloc(sizeof (dtrace_helper_action_t), KM_SLEEP);
14491 	helper->dtha_generation = help->dthps_generation;
14492 
14493 	if ((pred = ep->dted_pred.dtpdd_predicate) != NULL) {
14494 		ASSERT(pred->dtp_difo != NULL);
14495 		dtrace_difo_hold(pred->dtp_difo);
14496 		helper->dtha_predicate = pred->dtp_difo;
14497 	}
14498 
14499 	for (act = ep->dted_action; act != NULL; act = act->dtad_next) {
14500 		if (act->dtad_kind != DTRACEACT_DIFEXPR)
14501 			goto err;
14502 
14503 		if (act->dtad_difo == NULL)
14504 			goto err;
14505 
14506 		nactions++;
14507 	}
14508 
14509 	helper->dtha_actions = kmem_zalloc(sizeof (dtrace_difo_t *) *
14510 	    (helper->dtha_nactions = nactions), KM_SLEEP);
14511 
14512 	for (act = ep->dted_action, i = 0; act != NULL; act = act->dtad_next) {
14513 		dtrace_difo_hold(act->dtad_difo);
14514 		helper->dtha_actions[i++] = act->dtad_difo;
14515 	}
14516 
14517 	if (!dtrace_helper_validate(helper))
14518 		goto err;
14519 
14520 	if (last == NULL) {
14521 		help->dthps_actions[which] = helper;
14522 	} else {
14523 		last->dtha_next = helper;
14524 	}
14525 
14526 	if (vstate->dtvs_nlocals > dtrace_helptrace_nlocals) {
14527 		dtrace_helptrace_nlocals = vstate->dtvs_nlocals;
14528 		dtrace_helptrace_next = 0;
14529 	}
14530 
14531 	return (0);
14532 err:
14533 	dtrace_helper_action_destroy(helper, vstate);
14534 	return (EINVAL);
14535 }
14536 
14537 static void
14538 dtrace_helper_provider_register(proc_t *p, dtrace_helpers_t *help,
14539     dof_helper_t *dofhp)
14540 {
14541 	ASSERT(MUTEX_NOT_HELD(&dtrace_lock));
14542 
14543 	mutex_enter(&dtrace_meta_lock);
14544 	mutex_enter(&dtrace_lock);
14545 
14546 	if (!dtrace_attached() || dtrace_meta_pid == NULL) {
14547 		/*
14548 		 * If the dtrace module is loaded but not attached, or if
14549 		 * there aren't isn't a meta provider registered to deal with
14550 		 * these provider descriptions, we need to postpone creating
14551 		 * the actual providers until later.
14552 		 */
14553 
14554 		if (help->dthps_next == NULL && help->dthps_prev == NULL &&
14555 		    dtrace_deferred_pid != help) {
14556 			help->dthps_deferred = 1;
14557 			help->dthps_pid = p->p_pid;
14558 			help->dthps_next = dtrace_deferred_pid;
14559 			help->dthps_prev = NULL;
14560 			if (dtrace_deferred_pid != NULL)
14561 				dtrace_deferred_pid->dthps_prev = help;
14562 			dtrace_deferred_pid = help;
14563 		}
14564 
14565 		mutex_exit(&dtrace_lock);
14566 
14567 	} else if (dofhp != NULL) {
14568 		/*
14569 		 * If the dtrace module is loaded and we have a particular
14570 		 * helper provider description, pass that off to the
14571 		 * meta provider.
14572 		 */
14573 
14574 		mutex_exit(&dtrace_lock);
14575 
14576 		dtrace_helper_provide(dofhp, p->p_pid);
14577 
14578 	} else {
14579 		/*
14580 		 * Otherwise, just pass all the helper provider descriptions
14581 		 * off to the meta provider.
14582 		 */
14583 
14584 		int i;
14585 		mutex_exit(&dtrace_lock);
14586 
14587 		for (i = 0; i < help->dthps_nprovs; i++) {
14588 			dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
14589 			    p->p_pid);
14590 		}
14591 	}
14592 
14593 	mutex_exit(&dtrace_meta_lock);
14594 }
14595 
14596 static int
14597 dtrace_helper_provider_add(dof_helper_t *dofhp, int gen)
14598 {
14599 	dtrace_helpers_t *help;
14600 	dtrace_helper_provider_t *hprov, **tmp_provs;
14601 	uint_t tmp_maxprovs, i;
14602 
14603 	ASSERT(MUTEX_HELD(&dtrace_lock));
14604 
14605 	help = curproc->p_dtrace_helpers;
14606 	ASSERT(help != NULL);
14607 
14608 	/*
14609 	 * If we already have dtrace_helper_providers_max helper providers,
14610 	 * we're refuse to add a new one.
14611 	 */
14612 	if (help->dthps_nprovs >= dtrace_helper_providers_max)
14613 		return (ENOSPC);
14614 
14615 	/*
14616 	 * Check to make sure this isn't a duplicate.
14617 	 */
14618 	for (i = 0; i < help->dthps_nprovs; i++) {
14619 		if (dofhp->dofhp_addr ==
14620 		    help->dthps_provs[i]->dthp_prov.dofhp_addr)
14621 			return (EALREADY);
14622 	}
14623 
14624 	hprov = kmem_zalloc(sizeof (dtrace_helper_provider_t), KM_SLEEP);
14625 	hprov->dthp_prov = *dofhp;
14626 	hprov->dthp_ref = 1;
14627 	hprov->dthp_generation = gen;
14628 
14629 	/*
14630 	 * Allocate a bigger table for helper providers if it's already full.
14631 	 */
14632 	if (help->dthps_maxprovs == help->dthps_nprovs) {
14633 		tmp_maxprovs = help->dthps_maxprovs;
14634 		tmp_provs = help->dthps_provs;
14635 
14636 		if (help->dthps_maxprovs == 0)
14637 			help->dthps_maxprovs = 2;
14638 		else
14639 			help->dthps_maxprovs *= 2;
14640 		if (help->dthps_maxprovs > dtrace_helper_providers_max)
14641 			help->dthps_maxprovs = dtrace_helper_providers_max;
14642 
14643 		ASSERT(tmp_maxprovs < help->dthps_maxprovs);
14644 
14645 		help->dthps_provs = kmem_zalloc(help->dthps_maxprovs *
14646 		    sizeof (dtrace_helper_provider_t *), KM_SLEEP);
14647 
14648 		if (tmp_provs != NULL) {
14649 			bcopy(tmp_provs, help->dthps_provs, tmp_maxprovs *
14650 			    sizeof (dtrace_helper_provider_t *));
14651 			kmem_free(tmp_provs, tmp_maxprovs *
14652 			    sizeof (dtrace_helper_provider_t *));
14653 		}
14654 	}
14655 
14656 	help->dthps_provs[help->dthps_nprovs] = hprov;
14657 	help->dthps_nprovs++;
14658 
14659 	return (0);
14660 }
14661 
14662 static void
14663 dtrace_helper_provider_destroy(dtrace_helper_provider_t *hprov)
14664 {
14665 	mutex_enter(&dtrace_lock);
14666 
14667 	if (--hprov->dthp_ref == 0) {
14668 		dof_hdr_t *dof;
14669 		mutex_exit(&dtrace_lock);
14670 		dof = (dof_hdr_t *)(uintptr_t)hprov->dthp_prov.dofhp_dof;
14671 		dtrace_dof_destroy(dof);
14672 		kmem_free(hprov, sizeof (dtrace_helper_provider_t));
14673 	} else {
14674 		mutex_exit(&dtrace_lock);
14675 	}
14676 }
14677 
14678 static int
14679 dtrace_helper_provider_validate(dof_hdr_t *dof, dof_sec_t *sec)
14680 {
14681 	uintptr_t daddr = (uintptr_t)dof;
14682 	dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
14683 	dof_provider_t *provider;
14684 	dof_probe_t *probe;
14685 	uint8_t *arg;
14686 	char *strtab, *typestr;
14687 	dof_stridx_t typeidx;
14688 	size_t typesz;
14689 	uint_t nprobes, j, k;
14690 
14691 	ASSERT(sec->dofs_type == DOF_SECT_PROVIDER);
14692 
14693 	if (sec->dofs_offset & (sizeof (uint_t) - 1)) {
14694 		dtrace_dof_error(dof, "misaligned section offset");
14695 		return (-1);
14696 	}
14697 
14698 	/*
14699 	 * The section needs to be large enough to contain the DOF provider
14700 	 * structure appropriate for the given version.
14701 	 */
14702 	if (sec->dofs_size <
14703 	    ((dof->dofh_ident[DOF_ID_VERSION] == DOF_VERSION_1) ?
14704 	    offsetof(dof_provider_t, dofpv_prenoffs) :
14705 	    sizeof (dof_provider_t))) {
14706 		dtrace_dof_error(dof, "provider section too small");
14707 		return (-1);
14708 	}
14709 
14710 	provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
14711 	str_sec = dtrace_dof_sect(dof, DOF_SECT_STRTAB, provider->dofpv_strtab);
14712 	prb_sec = dtrace_dof_sect(dof, DOF_SECT_PROBES, provider->dofpv_probes);
14713 	arg_sec = dtrace_dof_sect(dof, DOF_SECT_PRARGS, provider->dofpv_prargs);
14714 	off_sec = dtrace_dof_sect(dof, DOF_SECT_PROFFS, provider->dofpv_proffs);
14715 
14716 	if (str_sec == NULL || prb_sec == NULL ||
14717 	    arg_sec == NULL || off_sec == NULL)
14718 		return (-1);
14719 
14720 	enoff_sec = NULL;
14721 
14722 	if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
14723 	    provider->dofpv_prenoffs != DOF_SECT_NONE &&
14724 	    (enoff_sec = dtrace_dof_sect(dof, DOF_SECT_PRENOFFS,
14725 	    provider->dofpv_prenoffs)) == NULL)
14726 		return (-1);
14727 
14728 	strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
14729 
14730 	if (provider->dofpv_name >= str_sec->dofs_size ||
14731 	    strlen(strtab + provider->dofpv_name) >= DTRACE_PROVNAMELEN) {
14732 		dtrace_dof_error(dof, "invalid provider name");
14733 		return (-1);
14734 	}
14735 
14736 	if (prb_sec->dofs_entsize == 0 ||
14737 	    prb_sec->dofs_entsize > prb_sec->dofs_size) {
14738 		dtrace_dof_error(dof, "invalid entry size");
14739 		return (-1);
14740 	}
14741 
14742 	if (prb_sec->dofs_entsize & (sizeof (uintptr_t) - 1)) {
14743 		dtrace_dof_error(dof, "misaligned entry size");
14744 		return (-1);
14745 	}
14746 
14747 	if (off_sec->dofs_entsize != sizeof (uint32_t)) {
14748 		dtrace_dof_error(dof, "invalid entry size");
14749 		return (-1);
14750 	}
14751 
14752 	if (off_sec->dofs_offset & (sizeof (uint32_t) - 1)) {
14753 		dtrace_dof_error(dof, "misaligned section offset");
14754 		return (-1);
14755 	}
14756 
14757 	if (arg_sec->dofs_entsize != sizeof (uint8_t)) {
14758 		dtrace_dof_error(dof, "invalid entry size");
14759 		return (-1);
14760 	}
14761 
14762 	arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
14763 
14764 	nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
14765 
14766 	/*
14767 	 * Take a pass through the probes to check for errors.
14768 	 */
14769 	for (j = 0; j < nprobes; j++) {
14770 		probe = (dof_probe_t *)(uintptr_t)(daddr +
14771 		    prb_sec->dofs_offset + j * prb_sec->dofs_entsize);
14772 
14773 		if (probe->dofpr_func >= str_sec->dofs_size) {
14774 			dtrace_dof_error(dof, "invalid function name");
14775 			return (-1);
14776 		}
14777 
14778 		if (strlen(strtab + probe->dofpr_func) >= DTRACE_FUNCNAMELEN) {
14779 			dtrace_dof_error(dof, "function name too long");
14780 			return (-1);
14781 		}
14782 
14783 		if (probe->dofpr_name >= str_sec->dofs_size ||
14784 		    strlen(strtab + probe->dofpr_name) >= DTRACE_NAMELEN) {
14785 			dtrace_dof_error(dof, "invalid probe name");
14786 			return (-1);
14787 		}
14788 
14789 		/*
14790 		 * The offset count must not wrap the index, and the offsets
14791 		 * must also not overflow the section's data.
14792 		 */
14793 		if (probe->dofpr_offidx + probe->dofpr_noffs <
14794 		    probe->dofpr_offidx ||
14795 		    (probe->dofpr_offidx + probe->dofpr_noffs) *
14796 		    off_sec->dofs_entsize > off_sec->dofs_size) {
14797 			dtrace_dof_error(dof, "invalid probe offset");
14798 			return (-1);
14799 		}
14800 
14801 		if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1) {
14802 			/*
14803 			 * If there's no is-enabled offset section, make sure
14804 			 * there aren't any is-enabled offsets. Otherwise
14805 			 * perform the same checks as for probe offsets
14806 			 * (immediately above).
14807 			 */
14808 			if (enoff_sec == NULL) {
14809 				if (probe->dofpr_enoffidx != 0 ||
14810 				    probe->dofpr_nenoffs != 0) {
14811 					dtrace_dof_error(dof, "is-enabled "
14812 					    "offsets with null section");
14813 					return (-1);
14814 				}
14815 			} else if (probe->dofpr_enoffidx +
14816 			    probe->dofpr_nenoffs < probe->dofpr_enoffidx ||
14817 			    (probe->dofpr_enoffidx + probe->dofpr_nenoffs) *
14818 			    enoff_sec->dofs_entsize > enoff_sec->dofs_size) {
14819 				dtrace_dof_error(dof, "invalid is-enabled "
14820 				    "offset");
14821 				return (-1);
14822 			}
14823 
14824 			if (probe->dofpr_noffs + probe->dofpr_nenoffs == 0) {
14825 				dtrace_dof_error(dof, "zero probe and "
14826 				    "is-enabled offsets");
14827 				return (-1);
14828 			}
14829 		} else if (probe->dofpr_noffs == 0) {
14830 			dtrace_dof_error(dof, "zero probe offsets");
14831 			return (-1);
14832 		}
14833 
14834 		if (probe->dofpr_argidx + probe->dofpr_xargc <
14835 		    probe->dofpr_argidx ||
14836 		    (probe->dofpr_argidx + probe->dofpr_xargc) *
14837 		    arg_sec->dofs_entsize > arg_sec->dofs_size) {
14838 			dtrace_dof_error(dof, "invalid args");
14839 			return (-1);
14840 		}
14841 
14842 		typeidx = probe->dofpr_nargv;
14843 		typestr = strtab + probe->dofpr_nargv;
14844 		for (k = 0; k < probe->dofpr_nargc; k++) {
14845 			if (typeidx >= str_sec->dofs_size) {
14846 				dtrace_dof_error(dof, "bad "
14847 				    "native argument type");
14848 				return (-1);
14849 			}
14850 
14851 			typesz = strlen(typestr) + 1;
14852 			if (typesz > DTRACE_ARGTYPELEN) {
14853 				dtrace_dof_error(dof, "native "
14854 				    "argument type too long");
14855 				return (-1);
14856 			}
14857 			typeidx += typesz;
14858 			typestr += typesz;
14859 		}
14860 
14861 		typeidx = probe->dofpr_xargv;
14862 		typestr = strtab + probe->dofpr_xargv;
14863 		for (k = 0; k < probe->dofpr_xargc; k++) {
14864 			if (arg[probe->dofpr_argidx + k] > probe->dofpr_nargc) {
14865 				dtrace_dof_error(dof, "bad "
14866 				    "native argument index");
14867 				return (-1);
14868 			}
14869 
14870 			if (typeidx >= str_sec->dofs_size) {
14871 				dtrace_dof_error(dof, "bad "
14872 				    "translated argument type");
14873 				return (-1);
14874 			}
14875 
14876 			typesz = strlen(typestr) + 1;
14877 			if (typesz > DTRACE_ARGTYPELEN) {
14878 				dtrace_dof_error(dof, "translated argument "
14879 				    "type too long");
14880 				return (-1);
14881 			}
14882 
14883 			typeidx += typesz;
14884 			typestr += typesz;
14885 		}
14886 	}
14887 
14888 	return (0);
14889 }
14890 
14891 static int
14892 dtrace_helper_slurp(dof_hdr_t *dof, dof_helper_t *dhp)
14893 {
14894 	dtrace_helpers_t *help;
14895 	dtrace_vstate_t *vstate;
14896 	dtrace_enabling_t *enab = NULL;
14897 	int i, gen, rv, nhelpers = 0, nprovs = 0, destroy = 1;
14898 	uintptr_t daddr = (uintptr_t)dof;
14899 
14900 	ASSERT(MUTEX_HELD(&dtrace_lock));
14901 
14902 	if ((help = curproc->p_dtrace_helpers) == NULL)
14903 		help = dtrace_helpers_create(curproc);
14904 
14905 	vstate = &help->dthps_vstate;
14906 
14907 	if ((rv = dtrace_dof_slurp(dof, vstate, NULL, &enab,
14908 	    dhp != NULL ? dhp->dofhp_addr : 0, B_FALSE)) != 0) {
14909 		dtrace_dof_destroy(dof);
14910 		return (rv);
14911 	}
14912 
14913 	/*
14914 	 * Look for helper providers and validate their descriptions.
14915 	 */
14916 	if (dhp != NULL) {
14917 		for (i = 0; i < dof->dofh_secnum; i++) {
14918 			dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
14919 			    dof->dofh_secoff + i * dof->dofh_secsize);
14920 
14921 			if (sec->dofs_type != DOF_SECT_PROVIDER)
14922 				continue;
14923 
14924 			if (dtrace_helper_provider_validate(dof, sec) != 0) {
14925 				dtrace_enabling_destroy(enab);
14926 				dtrace_dof_destroy(dof);
14927 				return (-1);
14928 			}
14929 
14930 			nprovs++;
14931 		}
14932 	}
14933 
14934 	/*
14935 	 * Now we need to walk through the ECB descriptions in the enabling.
14936 	 */
14937 	for (i = 0; i < enab->dten_ndesc; i++) {
14938 		dtrace_ecbdesc_t *ep = enab->dten_desc[i];
14939 		dtrace_probedesc_t *desc = &ep->dted_probe;
14940 
14941 		if (strcmp(desc->dtpd_provider, "dtrace") != 0)
14942 			continue;
14943 
14944 		if (strcmp(desc->dtpd_mod, "helper") != 0)
14945 			continue;
14946 
14947 		if (strcmp(desc->dtpd_func, "ustack") != 0)
14948 			continue;
14949 
14950 		if ((rv = dtrace_helper_action_add(DTRACE_HELPER_ACTION_USTACK,
14951 		    ep)) != 0) {
14952 			/*
14953 			 * Adding this helper action failed -- we are now going
14954 			 * to rip out the entire generation and return failure.
14955 			 */
14956 			(void) dtrace_helper_destroygen(help->dthps_generation);
14957 			dtrace_enabling_destroy(enab);
14958 			dtrace_dof_destroy(dof);
14959 			return (-1);
14960 		}
14961 
14962 		nhelpers++;
14963 	}
14964 
14965 	if (nhelpers < enab->dten_ndesc)
14966 		dtrace_dof_error(dof, "unmatched helpers");
14967 
14968 	gen = help->dthps_generation++;
14969 	dtrace_enabling_destroy(enab);
14970 
14971 	if (dhp != NULL && nprovs > 0) {
14972 		dhp->dofhp_dof = (uint64_t)(uintptr_t)dof;
14973 		if (dtrace_helper_provider_add(dhp, gen) == 0) {
14974 			mutex_exit(&dtrace_lock);
14975 			dtrace_helper_provider_register(curproc, help, dhp);
14976 			mutex_enter(&dtrace_lock);
14977 
14978 			destroy = 0;
14979 		}
14980 	}
14981 
14982 	if (destroy)
14983 		dtrace_dof_destroy(dof);
14984 
14985 	return (gen);
14986 }
14987 
14988 static dtrace_helpers_t *
14989 dtrace_helpers_create(proc_t *p)
14990 {
14991 	dtrace_helpers_t *help;
14992 
14993 	ASSERT(MUTEX_HELD(&dtrace_lock));
14994 	ASSERT(p->p_dtrace_helpers == NULL);
14995 
14996 	help = kmem_zalloc(sizeof (dtrace_helpers_t), KM_SLEEP);
14997 	help->dthps_actions = kmem_zalloc(sizeof (dtrace_helper_action_t *) *
14998 	    DTRACE_NHELPER_ACTIONS, KM_SLEEP);
14999 
15000 	p->p_dtrace_helpers = help;
15001 	dtrace_helpers++;
15002 
15003 	return (help);
15004 }
15005 
15006 #if defined(sun)
15007 static
15008 #endif
15009 void
15010 dtrace_helpers_destroy(proc_t *p)
15011 {
15012 	dtrace_helpers_t *help;
15013 	dtrace_vstate_t *vstate;
15014 #if defined(sun)
15015 	proc_t *p = curproc;
15016 #endif
15017 	int i;
15018 
15019 	mutex_enter(&dtrace_lock);
15020 
15021 	ASSERT(p->p_dtrace_helpers != NULL);
15022 	ASSERT(dtrace_helpers > 0);
15023 
15024 	help = p->p_dtrace_helpers;
15025 	vstate = &help->dthps_vstate;
15026 
15027 	/*
15028 	 * We're now going to lose the help from this process.
15029 	 */
15030 	p->p_dtrace_helpers = NULL;
15031 	dtrace_sync();
15032 
15033 	/*
15034 	 * Destory the helper actions.
15035 	 */
15036 	for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
15037 		dtrace_helper_action_t *h, *next;
15038 
15039 		for (h = help->dthps_actions[i]; h != NULL; h = next) {
15040 			next = h->dtha_next;
15041 			dtrace_helper_action_destroy(h, vstate);
15042 			h = next;
15043 		}
15044 	}
15045 
15046 	mutex_exit(&dtrace_lock);
15047 
15048 	/*
15049 	 * Destroy the helper providers.
15050 	 */
15051 	if (help->dthps_maxprovs > 0) {
15052 		mutex_enter(&dtrace_meta_lock);
15053 		if (dtrace_meta_pid != NULL) {
15054 			ASSERT(dtrace_deferred_pid == NULL);
15055 
15056 			for (i = 0; i < help->dthps_nprovs; i++) {
15057 				dtrace_helper_provider_remove(
15058 				    &help->dthps_provs[i]->dthp_prov, p->p_pid);
15059 			}
15060 		} else {
15061 			mutex_enter(&dtrace_lock);
15062 			ASSERT(help->dthps_deferred == 0 ||
15063 			    help->dthps_next != NULL ||
15064 			    help->dthps_prev != NULL ||
15065 			    help == dtrace_deferred_pid);
15066 
15067 			/*
15068 			 * Remove the helper from the deferred list.
15069 			 */
15070 			if (help->dthps_next != NULL)
15071 				help->dthps_next->dthps_prev = help->dthps_prev;
15072 			if (help->dthps_prev != NULL)
15073 				help->dthps_prev->dthps_next = help->dthps_next;
15074 			if (dtrace_deferred_pid == help) {
15075 				dtrace_deferred_pid = help->dthps_next;
15076 				ASSERT(help->dthps_prev == NULL);
15077 			}
15078 
15079 			mutex_exit(&dtrace_lock);
15080 		}
15081 
15082 		mutex_exit(&dtrace_meta_lock);
15083 
15084 		for (i = 0; i < help->dthps_nprovs; i++) {
15085 			dtrace_helper_provider_destroy(help->dthps_provs[i]);
15086 		}
15087 
15088 		kmem_free(help->dthps_provs, help->dthps_maxprovs *
15089 		    sizeof (dtrace_helper_provider_t *));
15090 	}
15091 
15092 	mutex_enter(&dtrace_lock);
15093 
15094 	dtrace_vstate_fini(&help->dthps_vstate);
15095 	kmem_free(help->dthps_actions,
15096 	    sizeof (dtrace_helper_action_t *) * DTRACE_NHELPER_ACTIONS);
15097 	kmem_free(help, sizeof (dtrace_helpers_t));
15098 
15099 	--dtrace_helpers;
15100 	mutex_exit(&dtrace_lock);
15101 }
15102 
15103 #if defined(sun)
15104 static
15105 #endif
15106 void
15107 dtrace_helpers_duplicate(proc_t *from, proc_t *to)
15108 {
15109 	dtrace_helpers_t *help, *newhelp;
15110 	dtrace_helper_action_t *helper, *new, *last;
15111 	dtrace_difo_t *dp;
15112 	dtrace_vstate_t *vstate;
15113 	int i, j, sz, hasprovs = 0;
15114 
15115 	mutex_enter(&dtrace_lock);
15116 	ASSERT(from->p_dtrace_helpers != NULL);
15117 	ASSERT(dtrace_helpers > 0);
15118 
15119 	help = from->p_dtrace_helpers;
15120 	newhelp = dtrace_helpers_create(to);
15121 	ASSERT(to->p_dtrace_helpers != NULL);
15122 
15123 	newhelp->dthps_generation = help->dthps_generation;
15124 	vstate = &newhelp->dthps_vstate;
15125 
15126 	/*
15127 	 * Duplicate the helper actions.
15128 	 */
15129 	for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
15130 		if ((helper = help->dthps_actions[i]) == NULL)
15131 			continue;
15132 
15133 		for (last = NULL; helper != NULL; helper = helper->dtha_next) {
15134 			new = kmem_zalloc(sizeof (dtrace_helper_action_t),
15135 			    KM_SLEEP);
15136 			new->dtha_generation = helper->dtha_generation;
15137 
15138 			if ((dp = helper->dtha_predicate) != NULL) {
15139 				dp = dtrace_difo_duplicate(dp, vstate);
15140 				new->dtha_predicate = dp;
15141 			}
15142 
15143 			new->dtha_nactions = helper->dtha_nactions;
15144 			sz = sizeof (dtrace_difo_t *) * new->dtha_nactions;
15145 			new->dtha_actions = kmem_alloc(sz, KM_SLEEP);
15146 
15147 			for (j = 0; j < new->dtha_nactions; j++) {
15148 				dtrace_difo_t *dp = helper->dtha_actions[j];
15149 
15150 				ASSERT(dp != NULL);
15151 				dp = dtrace_difo_duplicate(dp, vstate);
15152 				new->dtha_actions[j] = dp;
15153 			}
15154 
15155 			if (last != NULL) {
15156 				last->dtha_next = new;
15157 			} else {
15158 				newhelp->dthps_actions[i] = new;
15159 			}
15160 
15161 			last = new;
15162 		}
15163 	}
15164 
15165 	/*
15166 	 * Duplicate the helper providers and register them with the
15167 	 * DTrace framework.
15168 	 */
15169 	if (help->dthps_nprovs > 0) {
15170 		newhelp->dthps_nprovs = help->dthps_nprovs;
15171 		newhelp->dthps_maxprovs = help->dthps_nprovs;
15172 		newhelp->dthps_provs = kmem_alloc(newhelp->dthps_nprovs *
15173 		    sizeof (dtrace_helper_provider_t *), KM_SLEEP);
15174 		for (i = 0; i < newhelp->dthps_nprovs; i++) {
15175 			newhelp->dthps_provs[i] = help->dthps_provs[i];
15176 			newhelp->dthps_provs[i]->dthp_ref++;
15177 		}
15178 
15179 		hasprovs = 1;
15180 	}
15181 
15182 	mutex_exit(&dtrace_lock);
15183 
15184 	if (hasprovs)
15185 		dtrace_helper_provider_register(to, newhelp, NULL);
15186 }
15187 
15188 /*
15189  * DTrace Hook Functions
15190  */
15191 static void
15192 dtrace_module_loaded(modctl_t *ctl)
15193 {
15194 	dtrace_provider_t *prv;
15195 
15196 	mutex_enter(&dtrace_provider_lock);
15197 #if defined(sun)
15198 	mutex_enter(&mod_lock);
15199 #endif
15200 
15201 #if defined(sun)
15202 	ASSERT(ctl->mod_busy);
15203 #endif
15204 
15205 	/*
15206 	 * We're going to call each providers per-module provide operation
15207 	 * specifying only this module.
15208 	 */
15209 	for (prv = dtrace_provider; prv != NULL; prv = prv->dtpv_next)
15210 		prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
15211 
15212 #if defined(sun)
15213 	mutex_exit(&mod_lock);
15214 #endif
15215 	mutex_exit(&dtrace_provider_lock);
15216 
15217 	/*
15218 	 * If we have any retained enablings, we need to match against them.
15219 	 * Enabling probes requires that cpu_lock be held, and we cannot hold
15220 	 * cpu_lock here -- it is legal for cpu_lock to be held when loading a
15221 	 * module.  (In particular, this happens when loading scheduling
15222 	 * classes.)  So if we have any retained enablings, we need to dispatch
15223 	 * our task queue to do the match for us.
15224 	 */
15225 	mutex_enter(&dtrace_lock);
15226 
15227 	if (dtrace_retained == NULL) {
15228 		mutex_exit(&dtrace_lock);
15229 		return;
15230 	}
15231 
15232 	(void) taskq_dispatch(dtrace_taskq,
15233 	    (task_func_t *)dtrace_enabling_matchall, NULL, TQ_SLEEP);
15234 
15235 	mutex_exit(&dtrace_lock);
15236 
15237 	/*
15238 	 * And now, for a little heuristic sleaze:  in general, we want to
15239 	 * match modules as soon as they load.  However, we cannot guarantee
15240 	 * this, because it would lead us to the lock ordering violation
15241 	 * outlined above.  The common case, of course, is that cpu_lock is
15242 	 * _not_ held -- so we delay here for a clock tick, hoping that that's
15243 	 * long enough for the task queue to do its work.  If it's not, it's
15244 	 * not a serious problem -- it just means that the module that we
15245 	 * just loaded may not be immediately instrumentable.
15246 	 */
15247 	delay(1);
15248 }
15249 
15250 static void
15251 #if defined(sun)
15252 dtrace_module_unloaded(modctl_t *ctl)
15253 #else
15254 dtrace_module_unloaded(modctl_t *ctl, int *error)
15255 #endif
15256 {
15257 	dtrace_probe_t template, *probe, *first, *next;
15258 	dtrace_provider_t *prov;
15259 #if !defined(sun)
15260 	char modname[DTRACE_MODNAMELEN];
15261 	size_t len;
15262 #endif
15263 
15264 #if defined(sun)
15265 	template.dtpr_mod = ctl->mod_modname;
15266 #else
15267 	/* Handle the fact that ctl->filename may end in ".ko". */
15268 	strlcpy(modname, ctl->filename, sizeof(modname));
15269 	len = strlen(ctl->filename);
15270 	if (len > 3 && strcmp(modname + len - 3, ".ko") == 0)
15271 		modname[len - 3] = '\0';
15272 	template.dtpr_mod = modname;
15273 #endif
15274 
15275 	mutex_enter(&dtrace_provider_lock);
15276 #if defined(sun)
15277 	mutex_enter(&mod_lock);
15278 #endif
15279 	mutex_enter(&dtrace_lock);
15280 
15281 #if !defined(sun)
15282 	if (ctl->nenabled > 0) {
15283 		/* Don't allow unloads if a probe is enabled. */
15284 		mutex_exit(&dtrace_provider_lock);
15285 		mutex_exit(&dtrace_lock);
15286 		*error = -1;
15287 		printf(
15288 	"kldunload: attempt to unload module that has DTrace probes enabled\n");
15289 		return;
15290 	}
15291 #endif
15292 
15293 	if (dtrace_bymod == NULL) {
15294 		/*
15295 		 * The DTrace module is loaded (obviously) but not attached;
15296 		 * we don't have any work to do.
15297 		 */
15298 		mutex_exit(&dtrace_provider_lock);
15299 #if defined(sun)
15300 		mutex_exit(&mod_lock);
15301 #endif
15302 		mutex_exit(&dtrace_lock);
15303 		return;
15304 	}
15305 
15306 	for (probe = first = dtrace_hash_lookup(dtrace_bymod, &template);
15307 	    probe != NULL; probe = probe->dtpr_nextmod) {
15308 		if (probe->dtpr_ecb != NULL) {
15309 			mutex_exit(&dtrace_provider_lock);
15310 #if defined(sun)
15311 			mutex_exit(&mod_lock);
15312 #endif
15313 			mutex_exit(&dtrace_lock);
15314 
15315 			/*
15316 			 * This shouldn't _actually_ be possible -- we're
15317 			 * unloading a module that has an enabled probe in it.
15318 			 * (It's normally up to the provider to make sure that
15319 			 * this can't happen.)  However, because dtps_enable()
15320 			 * doesn't have a failure mode, there can be an
15321 			 * enable/unload race.  Upshot:  we don't want to
15322 			 * assert, but we're not going to disable the
15323 			 * probe, either.
15324 			 */
15325 			if (dtrace_err_verbose) {
15326 #if defined(sun)
15327 				cmn_err(CE_WARN, "unloaded module '%s' had "
15328 				    "enabled probes", ctl->mod_modname);
15329 #else
15330 				cmn_err(CE_WARN, "unloaded module '%s' had "
15331 				    "enabled probes", modname);
15332 #endif
15333 			}
15334 
15335 			return;
15336 		}
15337 	}
15338 
15339 	probe = first;
15340 
15341 	for (first = NULL; probe != NULL; probe = next) {
15342 		ASSERT(dtrace_probes[probe->dtpr_id - 1] == probe);
15343 
15344 		dtrace_probes[probe->dtpr_id - 1] = NULL;
15345 
15346 		next = probe->dtpr_nextmod;
15347 		dtrace_hash_remove(dtrace_bymod, probe);
15348 		dtrace_hash_remove(dtrace_byfunc, probe);
15349 		dtrace_hash_remove(dtrace_byname, probe);
15350 
15351 		if (first == NULL) {
15352 			first = probe;
15353 			probe->dtpr_nextmod = NULL;
15354 		} else {
15355 			probe->dtpr_nextmod = first;
15356 			first = probe;
15357 		}
15358 	}
15359 
15360 	/*
15361 	 * We've removed all of the module's probes from the hash chains and
15362 	 * from the probe array.  Now issue a dtrace_sync() to be sure that
15363 	 * everyone has cleared out from any probe array processing.
15364 	 */
15365 	dtrace_sync();
15366 
15367 	for (probe = first; probe != NULL; probe = first) {
15368 		first = probe->dtpr_nextmod;
15369 		prov = probe->dtpr_provider;
15370 		prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, probe->dtpr_id,
15371 		    probe->dtpr_arg);
15372 		kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
15373 		kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
15374 		kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
15375 #if defined(sun)
15376 		vmem_free(dtrace_arena, (void *)(uintptr_t)probe->dtpr_id, 1);
15377 #else
15378 		free_unr(dtrace_arena, probe->dtpr_id);
15379 #endif
15380 		kmem_free(probe, sizeof (dtrace_probe_t));
15381 	}
15382 
15383 	mutex_exit(&dtrace_lock);
15384 #if defined(sun)
15385 	mutex_exit(&mod_lock);
15386 #endif
15387 	mutex_exit(&dtrace_provider_lock);
15388 }
15389 
15390 #if !defined(sun)
15391 static void
15392 dtrace_kld_load(void *arg __unused, linker_file_t lf)
15393 {
15394 
15395 	dtrace_module_loaded(lf);
15396 }
15397 
15398 static void
15399 dtrace_kld_unload_try(void *arg __unused, linker_file_t lf, int *error)
15400 {
15401 
15402 	if (*error != 0)
15403 		/* We already have an error, so don't do anything. */
15404 		return;
15405 	dtrace_module_unloaded(lf, error);
15406 }
15407 #endif
15408 
15409 #if defined(sun)
15410 static void
15411 dtrace_suspend(void)
15412 {
15413 	dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_suspend));
15414 }
15415 
15416 static void
15417 dtrace_resume(void)
15418 {
15419 	dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_resume));
15420 }
15421 #endif
15422 
15423 static int
15424 dtrace_cpu_setup(cpu_setup_t what, processorid_t cpu)
15425 {
15426 	ASSERT(MUTEX_HELD(&cpu_lock));
15427 	mutex_enter(&dtrace_lock);
15428 
15429 	switch (what) {
15430 	case CPU_CONFIG: {
15431 		dtrace_state_t *state;
15432 		dtrace_optval_t *opt, rs, c;
15433 
15434 		/*
15435 		 * For now, we only allocate a new buffer for anonymous state.
15436 		 */
15437 		if ((state = dtrace_anon.dta_state) == NULL)
15438 			break;
15439 
15440 		if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
15441 			break;
15442 
15443 		opt = state->dts_options;
15444 		c = opt[DTRACEOPT_CPU];
15445 
15446 		if (c != DTRACE_CPUALL && c != DTRACEOPT_UNSET && c != cpu)
15447 			break;
15448 
15449 		/*
15450 		 * Regardless of what the actual policy is, we're going to
15451 		 * temporarily set our resize policy to be manual.  We're
15452 		 * also going to temporarily set our CPU option to denote
15453 		 * the newly configured CPU.
15454 		 */
15455 		rs = opt[DTRACEOPT_BUFRESIZE];
15456 		opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_MANUAL;
15457 		opt[DTRACEOPT_CPU] = (dtrace_optval_t)cpu;
15458 
15459 		(void) dtrace_state_buffers(state);
15460 
15461 		opt[DTRACEOPT_BUFRESIZE] = rs;
15462 		opt[DTRACEOPT_CPU] = c;
15463 
15464 		break;
15465 	}
15466 
15467 	case CPU_UNCONFIG:
15468 		/*
15469 		 * We don't free the buffer in the CPU_UNCONFIG case.  (The
15470 		 * buffer will be freed when the consumer exits.)
15471 		 */
15472 		break;
15473 
15474 	default:
15475 		break;
15476 	}
15477 
15478 	mutex_exit(&dtrace_lock);
15479 	return (0);
15480 }
15481 
15482 #if defined(sun)
15483 static void
15484 dtrace_cpu_setup_initial(processorid_t cpu)
15485 {
15486 	(void) dtrace_cpu_setup(CPU_CONFIG, cpu);
15487 }
15488 #endif
15489 
15490 static void
15491 dtrace_toxrange_add(uintptr_t base, uintptr_t limit)
15492 {
15493 	if (dtrace_toxranges >= dtrace_toxranges_max) {
15494 		int osize, nsize;
15495 		dtrace_toxrange_t *range;
15496 
15497 		osize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
15498 
15499 		if (osize == 0) {
15500 			ASSERT(dtrace_toxrange == NULL);
15501 			ASSERT(dtrace_toxranges_max == 0);
15502 			dtrace_toxranges_max = 1;
15503 		} else {
15504 			dtrace_toxranges_max <<= 1;
15505 		}
15506 
15507 		nsize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
15508 		range = kmem_zalloc(nsize, KM_SLEEP);
15509 
15510 		if (dtrace_toxrange != NULL) {
15511 			ASSERT(osize != 0);
15512 			bcopy(dtrace_toxrange, range, osize);
15513 			kmem_free(dtrace_toxrange, osize);
15514 		}
15515 
15516 		dtrace_toxrange = range;
15517 	}
15518 
15519 	ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_base == 0);
15520 	ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_limit == 0);
15521 
15522 	dtrace_toxrange[dtrace_toxranges].dtt_base = base;
15523 	dtrace_toxrange[dtrace_toxranges].dtt_limit = limit;
15524 	dtrace_toxranges++;
15525 }
15526 
15527 /*
15528  * DTrace Driver Cookbook Functions
15529  */
15530 #if defined(sun)
15531 /*ARGSUSED*/
15532 static int
15533 dtrace_attach(dev_info_t *devi, ddi_attach_cmd_t cmd)
15534 {
15535 	dtrace_provider_id_t id;
15536 	dtrace_state_t *state = NULL;
15537 	dtrace_enabling_t *enab;
15538 
15539 	mutex_enter(&cpu_lock);
15540 	mutex_enter(&dtrace_provider_lock);
15541 	mutex_enter(&dtrace_lock);
15542 
15543 	if (ddi_soft_state_init(&dtrace_softstate,
15544 	    sizeof (dtrace_state_t), 0) != 0) {
15545 		cmn_err(CE_NOTE, "/dev/dtrace failed to initialize soft state");
15546 		mutex_exit(&cpu_lock);
15547 		mutex_exit(&dtrace_provider_lock);
15548 		mutex_exit(&dtrace_lock);
15549 		return (DDI_FAILURE);
15550 	}
15551 
15552 	if (ddi_create_minor_node(devi, DTRACEMNR_DTRACE, S_IFCHR,
15553 	    DTRACEMNRN_DTRACE, DDI_PSEUDO, NULL) == DDI_FAILURE ||
15554 	    ddi_create_minor_node(devi, DTRACEMNR_HELPER, S_IFCHR,
15555 	    DTRACEMNRN_HELPER, DDI_PSEUDO, NULL) == DDI_FAILURE) {
15556 		cmn_err(CE_NOTE, "/dev/dtrace couldn't create minor nodes");
15557 		ddi_remove_minor_node(devi, NULL);
15558 		ddi_soft_state_fini(&dtrace_softstate);
15559 		mutex_exit(&cpu_lock);
15560 		mutex_exit(&dtrace_provider_lock);
15561 		mutex_exit(&dtrace_lock);
15562 		return (DDI_FAILURE);
15563 	}
15564 
15565 	ddi_report_dev(devi);
15566 	dtrace_devi = devi;
15567 
15568 	dtrace_modload = dtrace_module_loaded;
15569 	dtrace_modunload = dtrace_module_unloaded;
15570 	dtrace_cpu_init = dtrace_cpu_setup_initial;
15571 	dtrace_helpers_cleanup = dtrace_helpers_destroy;
15572 	dtrace_helpers_fork = dtrace_helpers_duplicate;
15573 	dtrace_cpustart_init = dtrace_suspend;
15574 	dtrace_cpustart_fini = dtrace_resume;
15575 	dtrace_debugger_init = dtrace_suspend;
15576 	dtrace_debugger_fini = dtrace_resume;
15577 
15578 	register_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
15579 
15580 	ASSERT(MUTEX_HELD(&cpu_lock));
15581 
15582 	dtrace_arena = vmem_create("dtrace", (void *)1, UINT32_MAX, 1,
15583 	    NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
15584 	dtrace_minor = vmem_create("dtrace_minor", (void *)DTRACEMNRN_CLONE,
15585 	    UINT32_MAX - DTRACEMNRN_CLONE, 1, NULL, NULL, NULL, 0,
15586 	    VM_SLEEP | VMC_IDENTIFIER);
15587 	dtrace_taskq = taskq_create("dtrace_taskq", 1, maxclsyspri,
15588 	    1, INT_MAX, 0);
15589 
15590 	dtrace_state_cache = kmem_cache_create("dtrace_state_cache",
15591 	    sizeof (dtrace_dstate_percpu_t) * NCPU, DTRACE_STATE_ALIGN,
15592 	    NULL, NULL, NULL, NULL, NULL, 0);
15593 
15594 	ASSERT(MUTEX_HELD(&cpu_lock));
15595 	dtrace_bymod = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_mod),
15596 	    offsetof(dtrace_probe_t, dtpr_nextmod),
15597 	    offsetof(dtrace_probe_t, dtpr_prevmod));
15598 
15599 	dtrace_byfunc = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_func),
15600 	    offsetof(dtrace_probe_t, dtpr_nextfunc),
15601 	    offsetof(dtrace_probe_t, dtpr_prevfunc));
15602 
15603 	dtrace_byname = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_name),
15604 	    offsetof(dtrace_probe_t, dtpr_nextname),
15605 	    offsetof(dtrace_probe_t, dtpr_prevname));
15606 
15607 	if (dtrace_retain_max < 1) {
15608 		cmn_err(CE_WARN, "illegal value (%lu) for dtrace_retain_max; "
15609 		    "setting to 1", dtrace_retain_max);
15610 		dtrace_retain_max = 1;
15611 	}
15612 
15613 	/*
15614 	 * Now discover our toxic ranges.
15615 	 */
15616 	dtrace_toxic_ranges(dtrace_toxrange_add);
15617 
15618 	/*
15619 	 * Before we register ourselves as a provider to our own framework,
15620 	 * we would like to assert that dtrace_provider is NULL -- but that's
15621 	 * not true if we were loaded as a dependency of a DTrace provider.
15622 	 * Once we've registered, we can assert that dtrace_provider is our
15623 	 * pseudo provider.
15624 	 */
15625 	(void) dtrace_register("dtrace", &dtrace_provider_attr,
15626 	    DTRACE_PRIV_NONE, 0, &dtrace_provider_ops, NULL, &id);
15627 
15628 	ASSERT(dtrace_provider != NULL);
15629 	ASSERT((dtrace_provider_id_t)dtrace_provider == id);
15630 
15631 	dtrace_probeid_begin = dtrace_probe_create((dtrace_provider_id_t)
15632 	    dtrace_provider, NULL, NULL, "BEGIN", 0, NULL);
15633 	dtrace_probeid_end = dtrace_probe_create((dtrace_provider_id_t)
15634 	    dtrace_provider, NULL, NULL, "END", 0, NULL);
15635 	dtrace_probeid_error = dtrace_probe_create((dtrace_provider_id_t)
15636 	    dtrace_provider, NULL, NULL, "ERROR", 1, NULL);
15637 
15638 	dtrace_anon_property();
15639 	mutex_exit(&cpu_lock);
15640 
15641 	/*
15642 	 * If DTrace helper tracing is enabled, we need to allocate the
15643 	 * trace buffer and initialize the values.
15644 	 */
15645 	if (dtrace_helptrace_enabled) {
15646 		ASSERT(dtrace_helptrace_buffer == NULL);
15647 		dtrace_helptrace_buffer =
15648 		    kmem_zalloc(dtrace_helptrace_bufsize, KM_SLEEP);
15649 		dtrace_helptrace_next = 0;
15650 	}
15651 
15652 	/*
15653 	 * If there are already providers, we must ask them to provide their
15654 	 * probes, and then match any anonymous enabling against them.  Note
15655 	 * that there should be no other retained enablings at this time:
15656 	 * the only retained enablings at this time should be the anonymous
15657 	 * enabling.
15658 	 */
15659 	if (dtrace_anon.dta_enabling != NULL) {
15660 		ASSERT(dtrace_retained == dtrace_anon.dta_enabling);
15661 
15662 		dtrace_enabling_provide(NULL);
15663 		state = dtrace_anon.dta_state;
15664 
15665 		/*
15666 		 * We couldn't hold cpu_lock across the above call to
15667 		 * dtrace_enabling_provide(), but we must hold it to actually
15668 		 * enable the probes.  We have to drop all of our locks, pick
15669 		 * up cpu_lock, and regain our locks before matching the
15670 		 * retained anonymous enabling.
15671 		 */
15672 		mutex_exit(&dtrace_lock);
15673 		mutex_exit(&dtrace_provider_lock);
15674 
15675 		mutex_enter(&cpu_lock);
15676 		mutex_enter(&dtrace_provider_lock);
15677 		mutex_enter(&dtrace_lock);
15678 
15679 		if ((enab = dtrace_anon.dta_enabling) != NULL)
15680 			(void) dtrace_enabling_match(enab, NULL);
15681 
15682 		mutex_exit(&cpu_lock);
15683 	}
15684 
15685 	mutex_exit(&dtrace_lock);
15686 	mutex_exit(&dtrace_provider_lock);
15687 
15688 	if (state != NULL) {
15689 		/*
15690 		 * If we created any anonymous state, set it going now.
15691 		 */
15692 		(void) dtrace_state_go(state, &dtrace_anon.dta_beganon);
15693 	}
15694 
15695 	return (DDI_SUCCESS);
15696 }
15697 #endif
15698 
15699 #if !defined(sun)
15700 #if __FreeBSD_version >= 800039
15701 static void dtrace_dtr(void *);
15702 #endif
15703 #endif
15704 
15705 /*ARGSUSED*/
15706 static int
15707 #if defined(sun)
15708 dtrace_open(dev_t *devp, int flag, int otyp, cred_t *cred_p)
15709 #else
15710 dtrace_open(struct cdev *dev, int oflags, int devtype, struct thread *td)
15711 #endif
15712 {
15713 	dtrace_state_t *state;
15714 	uint32_t priv;
15715 	uid_t uid;
15716 	zoneid_t zoneid;
15717 
15718 #if defined(sun)
15719 	if (getminor(*devp) == DTRACEMNRN_HELPER)
15720 		return (0);
15721 
15722 	/*
15723 	 * If this wasn't an open with the "helper" minor, then it must be
15724 	 * the "dtrace" minor.
15725 	 */
15726 	ASSERT(getminor(*devp) == DTRACEMNRN_DTRACE);
15727 #else
15728 	cred_t *cred_p = NULL;
15729 
15730 #if __FreeBSD_version < 800039
15731 	/*
15732 	 * The first minor device is the one that is cloned so there is
15733 	 * nothing more to do here.
15734 	 */
15735 	if (dev2unit(dev) == 0)
15736 		return 0;
15737 
15738 	/*
15739 	 * Devices are cloned, so if the DTrace state has already
15740 	 * been allocated, that means this device belongs to a
15741 	 * different client. Each client should open '/dev/dtrace'
15742 	 * to get a cloned device.
15743 	 */
15744 	if (dev->si_drv1 != NULL)
15745 		return (EBUSY);
15746 #endif
15747 
15748 	cred_p = dev->si_cred;
15749 #endif
15750 
15751 	/*
15752 	 * If no DTRACE_PRIV_* bits are set in the credential, then the
15753 	 * caller lacks sufficient permission to do anything with DTrace.
15754 	 */
15755 	dtrace_cred2priv(cred_p, &priv, &uid, &zoneid);
15756 	if (priv == DTRACE_PRIV_NONE) {
15757 #if !defined(sun)
15758 #if __FreeBSD_version < 800039
15759 		/* Destroy the cloned device. */
15760                 destroy_dev(dev);
15761 #endif
15762 #endif
15763 
15764 		return (EACCES);
15765 	}
15766 
15767 	/*
15768 	 * Ask all providers to provide all their probes.
15769 	 */
15770 	mutex_enter(&dtrace_provider_lock);
15771 	dtrace_probe_provide(NULL, NULL);
15772 	mutex_exit(&dtrace_provider_lock);
15773 
15774 	mutex_enter(&cpu_lock);
15775 	mutex_enter(&dtrace_lock);
15776 	dtrace_opens++;
15777 	dtrace_membar_producer();
15778 
15779 #if defined(sun)
15780 	/*
15781 	 * If the kernel debugger is active (that is, if the kernel debugger
15782 	 * modified text in some way), we won't allow the open.
15783 	 */
15784 	if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
15785 		dtrace_opens--;
15786 		mutex_exit(&cpu_lock);
15787 		mutex_exit(&dtrace_lock);
15788 		return (EBUSY);
15789 	}
15790 
15791 	state = dtrace_state_create(devp, cred_p);
15792 #else
15793 	state = dtrace_state_create(dev);
15794 #if __FreeBSD_version < 800039
15795 	dev->si_drv1 = state;
15796 #else
15797 	devfs_set_cdevpriv(state, dtrace_dtr);
15798 #endif
15799 #endif
15800 
15801 	mutex_exit(&cpu_lock);
15802 
15803 	if (state == NULL) {
15804 #if defined(sun)
15805 		if (--dtrace_opens == 0)
15806 			(void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
15807 #else
15808 		--dtrace_opens;
15809 #endif
15810 		mutex_exit(&dtrace_lock);
15811 #if !defined(sun)
15812 #if __FreeBSD_version < 800039
15813 		/* Destroy the cloned device. */
15814                 destroy_dev(dev);
15815 #endif
15816 #endif
15817 		return (EAGAIN);
15818 	}
15819 
15820 	mutex_exit(&dtrace_lock);
15821 
15822 	return (0);
15823 }
15824 
15825 /*ARGSUSED*/
15826 #if defined(sun)
15827 static int
15828 dtrace_close(dev_t dev, int flag, int otyp, cred_t *cred_p)
15829 #elif __FreeBSD_version < 800039
15830 static int
15831 dtrace_close(struct cdev *dev, int flags, int fmt __unused, struct thread *td)
15832 #else
15833 static void
15834 dtrace_dtr(void *data)
15835 #endif
15836 {
15837 #if defined(sun)
15838 	minor_t minor = getminor(dev);
15839 	dtrace_state_t *state;
15840 
15841 	if (minor == DTRACEMNRN_HELPER)
15842 		return (0);
15843 
15844 	state = ddi_get_soft_state(dtrace_softstate, minor);
15845 #else
15846 #if __FreeBSD_version < 800039
15847 	dtrace_state_t *state = dev->si_drv1;
15848 
15849 	/* Check if this is not a cloned device. */
15850 	if (dev2unit(dev) == 0)
15851 		return (0);
15852 #else
15853 	dtrace_state_t *state = data;
15854 #endif
15855 
15856 #endif
15857 
15858 	mutex_enter(&cpu_lock);
15859 	mutex_enter(&dtrace_lock);
15860 
15861 	if (state != NULL) {
15862 		if (state->dts_anon) {
15863 			/*
15864 			 * There is anonymous state. Destroy that first.
15865 			 */
15866 			ASSERT(dtrace_anon.dta_state == NULL);
15867 			dtrace_state_destroy(state->dts_anon);
15868 		}
15869 
15870 		dtrace_state_destroy(state);
15871 
15872 #if !defined(sun)
15873 		kmem_free(state, 0);
15874 #if __FreeBSD_version < 800039
15875 		dev->si_drv1 = NULL;
15876 #endif
15877 #endif
15878 	}
15879 
15880 	ASSERT(dtrace_opens > 0);
15881 #if defined(sun)
15882 	if (--dtrace_opens == 0)
15883 		(void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
15884 #else
15885 	--dtrace_opens;
15886 #endif
15887 
15888 	mutex_exit(&dtrace_lock);
15889 	mutex_exit(&cpu_lock);
15890 
15891 #if __FreeBSD_version < 800039
15892 	/* Schedule this cloned device to be destroyed. */
15893 	destroy_dev_sched(dev);
15894 #endif
15895 
15896 #if defined(sun) || __FreeBSD_version < 800039
15897 	return (0);
15898 #endif
15899 }
15900 
15901 #if defined(sun)
15902 /*ARGSUSED*/
15903 static int
15904 dtrace_ioctl_helper(int cmd, intptr_t arg, int *rv)
15905 {
15906 	int rval;
15907 	dof_helper_t help, *dhp = NULL;
15908 
15909 	switch (cmd) {
15910 	case DTRACEHIOC_ADDDOF:
15911 		if (copyin((void *)arg, &help, sizeof (help)) != 0) {
15912 			dtrace_dof_error(NULL, "failed to copyin DOF helper");
15913 			return (EFAULT);
15914 		}
15915 
15916 		dhp = &help;
15917 		arg = (intptr_t)help.dofhp_dof;
15918 		/*FALLTHROUGH*/
15919 
15920 	case DTRACEHIOC_ADD: {
15921 		dof_hdr_t *dof = dtrace_dof_copyin(arg, &rval);
15922 
15923 		if (dof == NULL)
15924 			return (rval);
15925 
15926 		mutex_enter(&dtrace_lock);
15927 
15928 		/*
15929 		 * dtrace_helper_slurp() takes responsibility for the dof --
15930 		 * it may free it now or it may save it and free it later.
15931 		 */
15932 		if ((rval = dtrace_helper_slurp(dof, dhp)) != -1) {
15933 			*rv = rval;
15934 			rval = 0;
15935 		} else {
15936 			rval = EINVAL;
15937 		}
15938 
15939 		mutex_exit(&dtrace_lock);
15940 		return (rval);
15941 	}
15942 
15943 	case DTRACEHIOC_REMOVE: {
15944 		mutex_enter(&dtrace_lock);
15945 		rval = dtrace_helper_destroygen(arg);
15946 		mutex_exit(&dtrace_lock);
15947 
15948 		return (rval);
15949 	}
15950 
15951 	default:
15952 		break;
15953 	}
15954 
15955 	return (ENOTTY);
15956 }
15957 
15958 /*ARGSUSED*/
15959 static int
15960 dtrace_ioctl(dev_t dev, int cmd, intptr_t arg, int md, cred_t *cr, int *rv)
15961 {
15962 	minor_t minor = getminor(dev);
15963 	dtrace_state_t *state;
15964 	int rval;
15965 
15966 	if (minor == DTRACEMNRN_HELPER)
15967 		return (dtrace_ioctl_helper(cmd, arg, rv));
15968 
15969 	state = ddi_get_soft_state(dtrace_softstate, minor);
15970 
15971 	if (state->dts_anon) {
15972 		ASSERT(dtrace_anon.dta_state == NULL);
15973 		state = state->dts_anon;
15974 	}
15975 
15976 	switch (cmd) {
15977 	case DTRACEIOC_PROVIDER: {
15978 		dtrace_providerdesc_t pvd;
15979 		dtrace_provider_t *pvp;
15980 
15981 		if (copyin((void *)arg, &pvd, sizeof (pvd)) != 0)
15982 			return (EFAULT);
15983 
15984 		pvd.dtvd_name[DTRACE_PROVNAMELEN - 1] = '\0';
15985 		mutex_enter(&dtrace_provider_lock);
15986 
15987 		for (pvp = dtrace_provider; pvp != NULL; pvp = pvp->dtpv_next) {
15988 			if (strcmp(pvp->dtpv_name, pvd.dtvd_name) == 0)
15989 				break;
15990 		}
15991 
15992 		mutex_exit(&dtrace_provider_lock);
15993 
15994 		if (pvp == NULL)
15995 			return (ESRCH);
15996 
15997 		bcopy(&pvp->dtpv_priv, &pvd.dtvd_priv, sizeof (dtrace_ppriv_t));
15998 		bcopy(&pvp->dtpv_attr, &pvd.dtvd_attr, sizeof (dtrace_pattr_t));
15999 
16000 		if (copyout(&pvd, (void *)arg, sizeof (pvd)) != 0)
16001 			return (EFAULT);
16002 
16003 		return (0);
16004 	}
16005 
16006 	case DTRACEIOC_EPROBE: {
16007 		dtrace_eprobedesc_t epdesc;
16008 		dtrace_ecb_t *ecb;
16009 		dtrace_action_t *act;
16010 		void *buf;
16011 		size_t size;
16012 		uintptr_t dest;
16013 		int nrecs;
16014 
16015 		if (copyin((void *)arg, &epdesc, sizeof (epdesc)) != 0)
16016 			return (EFAULT);
16017 
16018 		mutex_enter(&dtrace_lock);
16019 
16020 		if ((ecb = dtrace_epid2ecb(state, epdesc.dtepd_epid)) == NULL) {
16021 			mutex_exit(&dtrace_lock);
16022 			return (EINVAL);
16023 		}
16024 
16025 		if (ecb->dte_probe == NULL) {
16026 			mutex_exit(&dtrace_lock);
16027 			return (EINVAL);
16028 		}
16029 
16030 		epdesc.dtepd_probeid = ecb->dte_probe->dtpr_id;
16031 		epdesc.dtepd_uarg = ecb->dte_uarg;
16032 		epdesc.dtepd_size = ecb->dte_size;
16033 
16034 		nrecs = epdesc.dtepd_nrecs;
16035 		epdesc.dtepd_nrecs = 0;
16036 		for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
16037 			if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
16038 				continue;
16039 
16040 			epdesc.dtepd_nrecs++;
16041 		}
16042 
16043 		/*
16044 		 * Now that we have the size, we need to allocate a temporary
16045 		 * buffer in which to store the complete description.  We need
16046 		 * the temporary buffer to be able to drop dtrace_lock()
16047 		 * across the copyout(), below.
16048 		 */
16049 		size = sizeof (dtrace_eprobedesc_t) +
16050 		    (epdesc.dtepd_nrecs * sizeof (dtrace_recdesc_t));
16051 
16052 		buf = kmem_alloc(size, KM_SLEEP);
16053 		dest = (uintptr_t)buf;
16054 
16055 		bcopy(&epdesc, (void *)dest, sizeof (epdesc));
16056 		dest += offsetof(dtrace_eprobedesc_t, dtepd_rec[0]);
16057 
16058 		for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
16059 			if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
16060 				continue;
16061 
16062 			if (nrecs-- == 0)
16063 				break;
16064 
16065 			bcopy(&act->dta_rec, (void *)dest,
16066 			    sizeof (dtrace_recdesc_t));
16067 			dest += sizeof (dtrace_recdesc_t);
16068 		}
16069 
16070 		mutex_exit(&dtrace_lock);
16071 
16072 		if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
16073 			kmem_free(buf, size);
16074 			return (EFAULT);
16075 		}
16076 
16077 		kmem_free(buf, size);
16078 		return (0);
16079 	}
16080 
16081 	case DTRACEIOC_AGGDESC: {
16082 		dtrace_aggdesc_t aggdesc;
16083 		dtrace_action_t *act;
16084 		dtrace_aggregation_t *agg;
16085 		int nrecs;
16086 		uint32_t offs;
16087 		dtrace_recdesc_t *lrec;
16088 		void *buf;
16089 		size_t size;
16090 		uintptr_t dest;
16091 
16092 		if (copyin((void *)arg, &aggdesc, sizeof (aggdesc)) != 0)
16093 			return (EFAULT);
16094 
16095 		mutex_enter(&dtrace_lock);
16096 
16097 		if ((agg = dtrace_aggid2agg(state, aggdesc.dtagd_id)) == NULL) {
16098 			mutex_exit(&dtrace_lock);
16099 			return (EINVAL);
16100 		}
16101 
16102 		aggdesc.dtagd_epid = agg->dtag_ecb->dte_epid;
16103 
16104 		nrecs = aggdesc.dtagd_nrecs;
16105 		aggdesc.dtagd_nrecs = 0;
16106 
16107 		offs = agg->dtag_base;
16108 		lrec = &agg->dtag_action.dta_rec;
16109 		aggdesc.dtagd_size = lrec->dtrd_offset + lrec->dtrd_size - offs;
16110 
16111 		for (act = agg->dtag_first; ; act = act->dta_next) {
16112 			ASSERT(act->dta_intuple ||
16113 			    DTRACEACT_ISAGG(act->dta_kind));
16114 
16115 			/*
16116 			 * If this action has a record size of zero, it
16117 			 * denotes an argument to the aggregating action.
16118 			 * Because the presence of this record doesn't (or
16119 			 * shouldn't) affect the way the data is interpreted,
16120 			 * we don't copy it out to save user-level the
16121 			 * confusion of dealing with a zero-length record.
16122 			 */
16123 			if (act->dta_rec.dtrd_size == 0) {
16124 				ASSERT(agg->dtag_hasarg);
16125 				continue;
16126 			}
16127 
16128 			aggdesc.dtagd_nrecs++;
16129 
16130 			if (act == &agg->dtag_action)
16131 				break;
16132 		}
16133 
16134 		/*
16135 		 * Now that we have the size, we need to allocate a temporary
16136 		 * buffer in which to store the complete description.  We need
16137 		 * the temporary buffer to be able to drop dtrace_lock()
16138 		 * across the copyout(), below.
16139 		 */
16140 		size = sizeof (dtrace_aggdesc_t) +
16141 		    (aggdesc.dtagd_nrecs * sizeof (dtrace_recdesc_t));
16142 
16143 		buf = kmem_alloc(size, KM_SLEEP);
16144 		dest = (uintptr_t)buf;
16145 
16146 		bcopy(&aggdesc, (void *)dest, sizeof (aggdesc));
16147 		dest += offsetof(dtrace_aggdesc_t, dtagd_rec[0]);
16148 
16149 		for (act = agg->dtag_first; ; act = act->dta_next) {
16150 			dtrace_recdesc_t rec = act->dta_rec;
16151 
16152 			/*
16153 			 * See the comment in the above loop for why we pass
16154 			 * over zero-length records.
16155 			 */
16156 			if (rec.dtrd_size == 0) {
16157 				ASSERT(agg->dtag_hasarg);
16158 				continue;
16159 			}
16160 
16161 			if (nrecs-- == 0)
16162 				break;
16163 
16164 			rec.dtrd_offset -= offs;
16165 			bcopy(&rec, (void *)dest, sizeof (rec));
16166 			dest += sizeof (dtrace_recdesc_t);
16167 
16168 			if (act == &agg->dtag_action)
16169 				break;
16170 		}
16171 
16172 		mutex_exit(&dtrace_lock);
16173 
16174 		if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
16175 			kmem_free(buf, size);
16176 			return (EFAULT);
16177 		}
16178 
16179 		kmem_free(buf, size);
16180 		return (0);
16181 	}
16182 
16183 	case DTRACEIOC_ENABLE: {
16184 		dof_hdr_t *dof;
16185 		dtrace_enabling_t *enab = NULL;
16186 		dtrace_vstate_t *vstate;
16187 		int err = 0;
16188 
16189 		*rv = 0;
16190 
16191 		/*
16192 		 * If a NULL argument has been passed, we take this as our
16193 		 * cue to reevaluate our enablings.
16194 		 */
16195 		if (arg == NULL) {
16196 			dtrace_enabling_matchall();
16197 
16198 			return (0);
16199 		}
16200 
16201 		if ((dof = dtrace_dof_copyin(arg, &rval)) == NULL)
16202 			return (rval);
16203 
16204 		mutex_enter(&cpu_lock);
16205 		mutex_enter(&dtrace_lock);
16206 		vstate = &state->dts_vstate;
16207 
16208 		if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
16209 			mutex_exit(&dtrace_lock);
16210 			mutex_exit(&cpu_lock);
16211 			dtrace_dof_destroy(dof);
16212 			return (EBUSY);
16213 		}
16214 
16215 		if (dtrace_dof_slurp(dof, vstate, cr, &enab, 0, B_TRUE) != 0) {
16216 			mutex_exit(&dtrace_lock);
16217 			mutex_exit(&cpu_lock);
16218 			dtrace_dof_destroy(dof);
16219 			return (EINVAL);
16220 		}
16221 
16222 		if ((rval = dtrace_dof_options(dof, state)) != 0) {
16223 			dtrace_enabling_destroy(enab);
16224 			mutex_exit(&dtrace_lock);
16225 			mutex_exit(&cpu_lock);
16226 			dtrace_dof_destroy(dof);
16227 			return (rval);
16228 		}
16229 
16230 		if ((err = dtrace_enabling_match(enab, rv)) == 0) {
16231 			err = dtrace_enabling_retain(enab);
16232 		} else {
16233 			dtrace_enabling_destroy(enab);
16234 		}
16235 
16236 		mutex_exit(&cpu_lock);
16237 		mutex_exit(&dtrace_lock);
16238 		dtrace_dof_destroy(dof);
16239 
16240 		return (err);
16241 	}
16242 
16243 	case DTRACEIOC_REPLICATE: {
16244 		dtrace_repldesc_t desc;
16245 		dtrace_probedesc_t *match = &desc.dtrpd_match;
16246 		dtrace_probedesc_t *create = &desc.dtrpd_create;
16247 		int err;
16248 
16249 		if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16250 			return (EFAULT);
16251 
16252 		match->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
16253 		match->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
16254 		match->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
16255 		match->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
16256 
16257 		create->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
16258 		create->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
16259 		create->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
16260 		create->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
16261 
16262 		mutex_enter(&dtrace_lock);
16263 		err = dtrace_enabling_replicate(state, match, create);
16264 		mutex_exit(&dtrace_lock);
16265 
16266 		return (err);
16267 	}
16268 
16269 	case DTRACEIOC_PROBEMATCH:
16270 	case DTRACEIOC_PROBES: {
16271 		dtrace_probe_t *probe = NULL;
16272 		dtrace_probedesc_t desc;
16273 		dtrace_probekey_t pkey;
16274 		dtrace_id_t i;
16275 		int m = 0;
16276 		uint32_t priv;
16277 		uid_t uid;
16278 		zoneid_t zoneid;
16279 
16280 		if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16281 			return (EFAULT);
16282 
16283 		desc.dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
16284 		desc.dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
16285 		desc.dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
16286 		desc.dtpd_name[DTRACE_NAMELEN - 1] = '\0';
16287 
16288 		/*
16289 		 * Before we attempt to match this probe, we want to give
16290 		 * all providers the opportunity to provide it.
16291 		 */
16292 		if (desc.dtpd_id == DTRACE_IDNONE) {
16293 			mutex_enter(&dtrace_provider_lock);
16294 			dtrace_probe_provide(&desc, NULL);
16295 			mutex_exit(&dtrace_provider_lock);
16296 			desc.dtpd_id++;
16297 		}
16298 
16299 		if (cmd == DTRACEIOC_PROBEMATCH)  {
16300 			dtrace_probekey(&desc, &pkey);
16301 			pkey.dtpk_id = DTRACE_IDNONE;
16302 		}
16303 
16304 		dtrace_cred2priv(cr, &priv, &uid, &zoneid);
16305 
16306 		mutex_enter(&dtrace_lock);
16307 
16308 		if (cmd == DTRACEIOC_PROBEMATCH) {
16309 			for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
16310 				if ((probe = dtrace_probes[i - 1]) != NULL &&
16311 				    (m = dtrace_match_probe(probe, &pkey,
16312 				    priv, uid, zoneid)) != 0)
16313 					break;
16314 			}
16315 
16316 			if (m < 0) {
16317 				mutex_exit(&dtrace_lock);
16318 				return (EINVAL);
16319 			}
16320 
16321 		} else {
16322 			for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
16323 				if ((probe = dtrace_probes[i - 1]) != NULL &&
16324 				    dtrace_match_priv(probe, priv, uid, zoneid))
16325 					break;
16326 			}
16327 		}
16328 
16329 		if (probe == NULL) {
16330 			mutex_exit(&dtrace_lock);
16331 			return (ESRCH);
16332 		}
16333 
16334 		dtrace_probe_description(probe, &desc);
16335 		mutex_exit(&dtrace_lock);
16336 
16337 		if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16338 			return (EFAULT);
16339 
16340 		return (0);
16341 	}
16342 
16343 	case DTRACEIOC_PROBEARG: {
16344 		dtrace_argdesc_t desc;
16345 		dtrace_probe_t *probe;
16346 		dtrace_provider_t *prov;
16347 
16348 		if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16349 			return (EFAULT);
16350 
16351 		if (desc.dtargd_id == DTRACE_IDNONE)
16352 			return (EINVAL);
16353 
16354 		if (desc.dtargd_ndx == DTRACE_ARGNONE)
16355 			return (EINVAL);
16356 
16357 		mutex_enter(&dtrace_provider_lock);
16358 		mutex_enter(&mod_lock);
16359 		mutex_enter(&dtrace_lock);
16360 
16361 		if (desc.dtargd_id > dtrace_nprobes) {
16362 			mutex_exit(&dtrace_lock);
16363 			mutex_exit(&mod_lock);
16364 			mutex_exit(&dtrace_provider_lock);
16365 			return (EINVAL);
16366 		}
16367 
16368 		if ((probe = dtrace_probes[desc.dtargd_id - 1]) == NULL) {
16369 			mutex_exit(&dtrace_lock);
16370 			mutex_exit(&mod_lock);
16371 			mutex_exit(&dtrace_provider_lock);
16372 			return (EINVAL);
16373 		}
16374 
16375 		mutex_exit(&dtrace_lock);
16376 
16377 		prov = probe->dtpr_provider;
16378 
16379 		if (prov->dtpv_pops.dtps_getargdesc == NULL) {
16380 			/*
16381 			 * There isn't any typed information for this probe.
16382 			 * Set the argument number to DTRACE_ARGNONE.
16383 			 */
16384 			desc.dtargd_ndx = DTRACE_ARGNONE;
16385 		} else {
16386 			desc.dtargd_native[0] = '\0';
16387 			desc.dtargd_xlate[0] = '\0';
16388 			desc.dtargd_mapping = desc.dtargd_ndx;
16389 
16390 			prov->dtpv_pops.dtps_getargdesc(prov->dtpv_arg,
16391 			    probe->dtpr_id, probe->dtpr_arg, &desc);
16392 		}
16393 
16394 		mutex_exit(&mod_lock);
16395 		mutex_exit(&dtrace_provider_lock);
16396 
16397 		if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16398 			return (EFAULT);
16399 
16400 		return (0);
16401 	}
16402 
16403 	case DTRACEIOC_GO: {
16404 		processorid_t cpuid;
16405 		rval = dtrace_state_go(state, &cpuid);
16406 
16407 		if (rval != 0)
16408 			return (rval);
16409 
16410 		if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
16411 			return (EFAULT);
16412 
16413 		return (0);
16414 	}
16415 
16416 	case DTRACEIOC_STOP: {
16417 		processorid_t cpuid;
16418 
16419 		mutex_enter(&dtrace_lock);
16420 		rval = dtrace_state_stop(state, &cpuid);
16421 		mutex_exit(&dtrace_lock);
16422 
16423 		if (rval != 0)
16424 			return (rval);
16425 
16426 		if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
16427 			return (EFAULT);
16428 
16429 		return (0);
16430 	}
16431 
16432 	case DTRACEIOC_DOFGET: {
16433 		dof_hdr_t hdr, *dof;
16434 		uint64_t len;
16435 
16436 		if (copyin((void *)arg, &hdr, sizeof (hdr)) != 0)
16437 			return (EFAULT);
16438 
16439 		mutex_enter(&dtrace_lock);
16440 		dof = dtrace_dof_create(state);
16441 		mutex_exit(&dtrace_lock);
16442 
16443 		len = MIN(hdr.dofh_loadsz, dof->dofh_loadsz);
16444 		rval = copyout(dof, (void *)arg, len);
16445 		dtrace_dof_destroy(dof);
16446 
16447 		return (rval == 0 ? 0 : EFAULT);
16448 	}
16449 
16450 	case DTRACEIOC_AGGSNAP:
16451 	case DTRACEIOC_BUFSNAP: {
16452 		dtrace_bufdesc_t desc;
16453 		caddr_t cached;
16454 		dtrace_buffer_t *buf;
16455 
16456 		if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16457 			return (EFAULT);
16458 
16459 		if (desc.dtbd_cpu < 0 || desc.dtbd_cpu >= NCPU)
16460 			return (EINVAL);
16461 
16462 		mutex_enter(&dtrace_lock);
16463 
16464 		if (cmd == DTRACEIOC_BUFSNAP) {
16465 			buf = &state->dts_buffer[desc.dtbd_cpu];
16466 		} else {
16467 			buf = &state->dts_aggbuffer[desc.dtbd_cpu];
16468 		}
16469 
16470 		if (buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL)) {
16471 			size_t sz = buf->dtb_offset;
16472 
16473 			if (state->dts_activity != DTRACE_ACTIVITY_STOPPED) {
16474 				mutex_exit(&dtrace_lock);
16475 				return (EBUSY);
16476 			}
16477 
16478 			/*
16479 			 * If this buffer has already been consumed, we're
16480 			 * going to indicate that there's nothing left here
16481 			 * to consume.
16482 			 */
16483 			if (buf->dtb_flags & DTRACEBUF_CONSUMED) {
16484 				mutex_exit(&dtrace_lock);
16485 
16486 				desc.dtbd_size = 0;
16487 				desc.dtbd_drops = 0;
16488 				desc.dtbd_errors = 0;
16489 				desc.dtbd_oldest = 0;
16490 				sz = sizeof (desc);
16491 
16492 				if (copyout(&desc, (void *)arg, sz) != 0)
16493 					return (EFAULT);
16494 
16495 				return (0);
16496 			}
16497 
16498 			/*
16499 			 * If this is a ring buffer that has wrapped, we want
16500 			 * to copy the whole thing out.
16501 			 */
16502 			if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
16503 				dtrace_buffer_polish(buf);
16504 				sz = buf->dtb_size;
16505 			}
16506 
16507 			if (copyout(buf->dtb_tomax, desc.dtbd_data, sz) != 0) {
16508 				mutex_exit(&dtrace_lock);
16509 				return (EFAULT);
16510 			}
16511 
16512 			desc.dtbd_size = sz;
16513 			desc.dtbd_drops = buf->dtb_drops;
16514 			desc.dtbd_errors = buf->dtb_errors;
16515 			desc.dtbd_oldest = buf->dtb_xamot_offset;
16516 			desc.dtbd_timestamp = dtrace_gethrtime();
16517 
16518 			mutex_exit(&dtrace_lock);
16519 
16520 			if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16521 				return (EFAULT);
16522 
16523 			buf->dtb_flags |= DTRACEBUF_CONSUMED;
16524 
16525 			return (0);
16526 		}
16527 
16528 		if (buf->dtb_tomax == NULL) {
16529 			ASSERT(buf->dtb_xamot == NULL);
16530 			mutex_exit(&dtrace_lock);
16531 			return (ENOENT);
16532 		}
16533 
16534 		cached = buf->dtb_tomax;
16535 		ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
16536 
16537 		dtrace_xcall(desc.dtbd_cpu,
16538 		    (dtrace_xcall_t)dtrace_buffer_switch, buf);
16539 
16540 		state->dts_errors += buf->dtb_xamot_errors;
16541 
16542 		/*
16543 		 * If the buffers did not actually switch, then the cross call
16544 		 * did not take place -- presumably because the given CPU is
16545 		 * not in the ready set.  If this is the case, we'll return
16546 		 * ENOENT.
16547 		 */
16548 		if (buf->dtb_tomax == cached) {
16549 			ASSERT(buf->dtb_xamot != cached);
16550 			mutex_exit(&dtrace_lock);
16551 			return (ENOENT);
16552 		}
16553 
16554 		ASSERT(cached == buf->dtb_xamot);
16555 
16556 		/*
16557 		 * We have our snapshot; now copy it out.
16558 		 */
16559 		if (copyout(buf->dtb_xamot, desc.dtbd_data,
16560 		    buf->dtb_xamot_offset) != 0) {
16561 			mutex_exit(&dtrace_lock);
16562 			return (EFAULT);
16563 		}
16564 
16565 		desc.dtbd_size = buf->dtb_xamot_offset;
16566 		desc.dtbd_drops = buf->dtb_xamot_drops;
16567 		desc.dtbd_errors = buf->dtb_xamot_errors;
16568 		desc.dtbd_oldest = 0;
16569 		desc.dtbd_timestamp = buf->dtb_switched;
16570 
16571 		mutex_exit(&dtrace_lock);
16572 
16573 		/*
16574 		 * Finally, copy out the buffer description.
16575 		 */
16576 		if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16577 			return (EFAULT);
16578 
16579 		return (0);
16580 	}
16581 
16582 	case DTRACEIOC_CONF: {
16583 		dtrace_conf_t conf;
16584 
16585 		bzero(&conf, sizeof (conf));
16586 		conf.dtc_difversion = DIF_VERSION;
16587 		conf.dtc_difintregs = DIF_DIR_NREGS;
16588 		conf.dtc_diftupregs = DIF_DTR_NREGS;
16589 		conf.dtc_ctfmodel = CTF_MODEL_NATIVE;
16590 
16591 		if (copyout(&conf, (void *)arg, sizeof (conf)) != 0)
16592 			return (EFAULT);
16593 
16594 		return (0);
16595 	}
16596 
16597 	case DTRACEIOC_STATUS: {
16598 		dtrace_status_t stat;
16599 		dtrace_dstate_t *dstate;
16600 		int i, j;
16601 		uint64_t nerrs;
16602 
16603 		/*
16604 		 * See the comment in dtrace_state_deadman() for the reason
16605 		 * for setting dts_laststatus to INT64_MAX before setting
16606 		 * it to the correct value.
16607 		 */
16608 		state->dts_laststatus = INT64_MAX;
16609 		dtrace_membar_producer();
16610 		state->dts_laststatus = dtrace_gethrtime();
16611 
16612 		bzero(&stat, sizeof (stat));
16613 
16614 		mutex_enter(&dtrace_lock);
16615 
16616 		if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE) {
16617 			mutex_exit(&dtrace_lock);
16618 			return (ENOENT);
16619 		}
16620 
16621 		if (state->dts_activity == DTRACE_ACTIVITY_DRAINING)
16622 			stat.dtst_exiting = 1;
16623 
16624 		nerrs = state->dts_errors;
16625 		dstate = &state->dts_vstate.dtvs_dynvars;
16626 
16627 		for (i = 0; i < NCPU; i++) {
16628 			dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[i];
16629 
16630 			stat.dtst_dyndrops += dcpu->dtdsc_drops;
16631 			stat.dtst_dyndrops_dirty += dcpu->dtdsc_dirty_drops;
16632 			stat.dtst_dyndrops_rinsing += dcpu->dtdsc_rinsing_drops;
16633 
16634 			if (state->dts_buffer[i].dtb_flags & DTRACEBUF_FULL)
16635 				stat.dtst_filled++;
16636 
16637 			nerrs += state->dts_buffer[i].dtb_errors;
16638 
16639 			for (j = 0; j < state->dts_nspeculations; j++) {
16640 				dtrace_speculation_t *spec;
16641 				dtrace_buffer_t *buf;
16642 
16643 				spec = &state->dts_speculations[j];
16644 				buf = &spec->dtsp_buffer[i];
16645 				stat.dtst_specdrops += buf->dtb_xamot_drops;
16646 			}
16647 		}
16648 
16649 		stat.dtst_specdrops_busy = state->dts_speculations_busy;
16650 		stat.dtst_specdrops_unavail = state->dts_speculations_unavail;
16651 		stat.dtst_stkstroverflows = state->dts_stkstroverflows;
16652 		stat.dtst_dblerrors = state->dts_dblerrors;
16653 		stat.dtst_killed =
16654 		    (state->dts_activity == DTRACE_ACTIVITY_KILLED);
16655 		stat.dtst_errors = nerrs;
16656 
16657 		mutex_exit(&dtrace_lock);
16658 
16659 		if (copyout(&stat, (void *)arg, sizeof (stat)) != 0)
16660 			return (EFAULT);
16661 
16662 		return (0);
16663 	}
16664 
16665 	case DTRACEIOC_FORMAT: {
16666 		dtrace_fmtdesc_t fmt;
16667 		char *str;
16668 		int len;
16669 
16670 		if (copyin((void *)arg, &fmt, sizeof (fmt)) != 0)
16671 			return (EFAULT);
16672 
16673 		mutex_enter(&dtrace_lock);
16674 
16675 		if (fmt.dtfd_format == 0 ||
16676 		    fmt.dtfd_format > state->dts_nformats) {
16677 			mutex_exit(&dtrace_lock);
16678 			return (EINVAL);
16679 		}
16680 
16681 		/*
16682 		 * Format strings are allocated contiguously and they are
16683 		 * never freed; if a format index is less than the number
16684 		 * of formats, we can assert that the format map is non-NULL
16685 		 * and that the format for the specified index is non-NULL.
16686 		 */
16687 		ASSERT(state->dts_formats != NULL);
16688 		str = state->dts_formats[fmt.dtfd_format - 1];
16689 		ASSERT(str != NULL);
16690 
16691 		len = strlen(str) + 1;
16692 
16693 		if (len > fmt.dtfd_length) {
16694 			fmt.dtfd_length = len;
16695 
16696 			if (copyout(&fmt, (void *)arg, sizeof (fmt)) != 0) {
16697 				mutex_exit(&dtrace_lock);
16698 				return (EINVAL);
16699 			}
16700 		} else {
16701 			if (copyout(str, fmt.dtfd_string, len) != 0) {
16702 				mutex_exit(&dtrace_lock);
16703 				return (EINVAL);
16704 			}
16705 		}
16706 
16707 		mutex_exit(&dtrace_lock);
16708 		return (0);
16709 	}
16710 
16711 	default:
16712 		break;
16713 	}
16714 
16715 	return (ENOTTY);
16716 }
16717 
16718 /*ARGSUSED*/
16719 static int
16720 dtrace_detach(dev_info_t *dip, ddi_detach_cmd_t cmd)
16721 {
16722 	dtrace_state_t *state;
16723 
16724 	switch (cmd) {
16725 	case DDI_DETACH:
16726 		break;
16727 
16728 	case DDI_SUSPEND:
16729 		return (DDI_SUCCESS);
16730 
16731 	default:
16732 		return (DDI_FAILURE);
16733 	}
16734 
16735 	mutex_enter(&cpu_lock);
16736 	mutex_enter(&dtrace_provider_lock);
16737 	mutex_enter(&dtrace_lock);
16738 
16739 	ASSERT(dtrace_opens == 0);
16740 
16741 	if (dtrace_helpers > 0) {
16742 		mutex_exit(&dtrace_provider_lock);
16743 		mutex_exit(&dtrace_lock);
16744 		mutex_exit(&cpu_lock);
16745 		return (DDI_FAILURE);
16746 	}
16747 
16748 	if (dtrace_unregister((dtrace_provider_id_t)dtrace_provider) != 0) {
16749 		mutex_exit(&dtrace_provider_lock);
16750 		mutex_exit(&dtrace_lock);
16751 		mutex_exit(&cpu_lock);
16752 		return (DDI_FAILURE);
16753 	}
16754 
16755 	dtrace_provider = NULL;
16756 
16757 	if ((state = dtrace_anon_grab()) != NULL) {
16758 		/*
16759 		 * If there were ECBs on this state, the provider should
16760 		 * have not been allowed to detach; assert that there is
16761 		 * none.
16762 		 */
16763 		ASSERT(state->dts_necbs == 0);
16764 		dtrace_state_destroy(state);
16765 
16766 		/*
16767 		 * If we're being detached with anonymous state, we need to
16768 		 * indicate to the kernel debugger that DTrace is now inactive.
16769 		 */
16770 		(void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
16771 	}
16772 
16773 	bzero(&dtrace_anon, sizeof (dtrace_anon_t));
16774 	unregister_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
16775 	dtrace_cpu_init = NULL;
16776 	dtrace_helpers_cleanup = NULL;
16777 	dtrace_helpers_fork = NULL;
16778 	dtrace_cpustart_init = NULL;
16779 	dtrace_cpustart_fini = NULL;
16780 	dtrace_debugger_init = NULL;
16781 	dtrace_debugger_fini = NULL;
16782 	dtrace_modload = NULL;
16783 	dtrace_modunload = NULL;
16784 
16785 	mutex_exit(&cpu_lock);
16786 
16787 	if (dtrace_helptrace_enabled) {
16788 		kmem_free(dtrace_helptrace_buffer, dtrace_helptrace_bufsize);
16789 		dtrace_helptrace_buffer = NULL;
16790 	}
16791 
16792 	kmem_free(dtrace_probes, dtrace_nprobes * sizeof (dtrace_probe_t *));
16793 	dtrace_probes = NULL;
16794 	dtrace_nprobes = 0;
16795 
16796 	dtrace_hash_destroy(dtrace_bymod);
16797 	dtrace_hash_destroy(dtrace_byfunc);
16798 	dtrace_hash_destroy(dtrace_byname);
16799 	dtrace_bymod = NULL;
16800 	dtrace_byfunc = NULL;
16801 	dtrace_byname = NULL;
16802 
16803 	kmem_cache_destroy(dtrace_state_cache);
16804 	vmem_destroy(dtrace_minor);
16805 	vmem_destroy(dtrace_arena);
16806 
16807 	if (dtrace_toxrange != NULL) {
16808 		kmem_free(dtrace_toxrange,
16809 		    dtrace_toxranges_max * sizeof (dtrace_toxrange_t));
16810 		dtrace_toxrange = NULL;
16811 		dtrace_toxranges = 0;
16812 		dtrace_toxranges_max = 0;
16813 	}
16814 
16815 	ddi_remove_minor_node(dtrace_devi, NULL);
16816 	dtrace_devi = NULL;
16817 
16818 	ddi_soft_state_fini(&dtrace_softstate);
16819 
16820 	ASSERT(dtrace_vtime_references == 0);
16821 	ASSERT(dtrace_opens == 0);
16822 	ASSERT(dtrace_retained == NULL);
16823 
16824 	mutex_exit(&dtrace_lock);
16825 	mutex_exit(&dtrace_provider_lock);
16826 
16827 	/*
16828 	 * We don't destroy the task queue until after we have dropped our
16829 	 * locks (taskq_destroy() may block on running tasks).  To prevent
16830 	 * attempting to do work after we have effectively detached but before
16831 	 * the task queue has been destroyed, all tasks dispatched via the
16832 	 * task queue must check that DTrace is still attached before
16833 	 * performing any operation.
16834 	 */
16835 	taskq_destroy(dtrace_taskq);
16836 	dtrace_taskq = NULL;
16837 
16838 	return (DDI_SUCCESS);
16839 }
16840 #endif
16841 
16842 #if defined(sun)
16843 /*ARGSUSED*/
16844 static int
16845 dtrace_info(dev_info_t *dip, ddi_info_cmd_t infocmd, void *arg, void **result)
16846 {
16847 	int error;
16848 
16849 	switch (infocmd) {
16850 	case DDI_INFO_DEVT2DEVINFO:
16851 		*result = (void *)dtrace_devi;
16852 		error = DDI_SUCCESS;
16853 		break;
16854 	case DDI_INFO_DEVT2INSTANCE:
16855 		*result = (void *)0;
16856 		error = DDI_SUCCESS;
16857 		break;
16858 	default:
16859 		error = DDI_FAILURE;
16860 	}
16861 	return (error);
16862 }
16863 #endif
16864 
16865 #if defined(sun)
16866 static struct cb_ops dtrace_cb_ops = {
16867 	dtrace_open,		/* open */
16868 	dtrace_close,		/* close */
16869 	nulldev,		/* strategy */
16870 	nulldev,		/* print */
16871 	nodev,			/* dump */
16872 	nodev,			/* read */
16873 	nodev,			/* write */
16874 	dtrace_ioctl,		/* ioctl */
16875 	nodev,			/* devmap */
16876 	nodev,			/* mmap */
16877 	nodev,			/* segmap */
16878 	nochpoll,		/* poll */
16879 	ddi_prop_op,		/* cb_prop_op */
16880 	0,			/* streamtab  */
16881 	D_NEW | D_MP		/* Driver compatibility flag */
16882 };
16883 
16884 static struct dev_ops dtrace_ops = {
16885 	DEVO_REV,		/* devo_rev */
16886 	0,			/* refcnt */
16887 	dtrace_info,		/* get_dev_info */
16888 	nulldev,		/* identify */
16889 	nulldev,		/* probe */
16890 	dtrace_attach,		/* attach */
16891 	dtrace_detach,		/* detach */
16892 	nodev,			/* reset */
16893 	&dtrace_cb_ops,		/* driver operations */
16894 	NULL,			/* bus operations */
16895 	nodev			/* dev power */
16896 };
16897 
16898 static struct modldrv modldrv = {
16899 	&mod_driverops,		/* module type (this is a pseudo driver) */
16900 	"Dynamic Tracing",	/* name of module */
16901 	&dtrace_ops,		/* driver ops */
16902 };
16903 
16904 static struct modlinkage modlinkage = {
16905 	MODREV_1,
16906 	(void *)&modldrv,
16907 	NULL
16908 };
16909 
16910 int
16911 _init(void)
16912 {
16913 	return (mod_install(&modlinkage));
16914 }
16915 
16916 int
16917 _info(struct modinfo *modinfop)
16918 {
16919 	return (mod_info(&modlinkage, modinfop));
16920 }
16921 
16922 int
16923 _fini(void)
16924 {
16925 	return (mod_remove(&modlinkage));
16926 }
16927 #else
16928 
16929 static d_ioctl_t	dtrace_ioctl;
16930 static d_ioctl_t	dtrace_ioctl_helper;
16931 static void		dtrace_load(void *);
16932 static int		dtrace_unload(void);
16933 #if __FreeBSD_version < 800039
16934 static void		dtrace_clone(void *, struct ucred *, char *, int , struct cdev **);
16935 static struct clonedevs	*dtrace_clones;		/* Ptr to the array of cloned devices. */
16936 static eventhandler_tag	eh_tag;			/* Event handler tag. */
16937 #else
16938 static struct cdev	*dtrace_dev;
16939 static struct cdev	*helper_dev;
16940 #endif
16941 
16942 void dtrace_invop_init(void);
16943 void dtrace_invop_uninit(void);
16944 
16945 static struct cdevsw dtrace_cdevsw = {
16946 	.d_version	= D_VERSION,
16947 #if __FreeBSD_version < 800039
16948 	.d_flags	= D_TRACKCLOSE | D_NEEDMINOR,
16949 	.d_close	= dtrace_close,
16950 #endif
16951 	.d_ioctl	= dtrace_ioctl,
16952 	.d_open		= dtrace_open,
16953 	.d_name		= "dtrace",
16954 };
16955 
16956 static struct cdevsw helper_cdevsw = {
16957 	.d_version	= D_VERSION,
16958 	.d_ioctl	= dtrace_ioctl_helper,
16959 	.d_name		= "helper",
16960 };
16961 
16962 #include <dtrace_anon.c>
16963 #if __FreeBSD_version < 800039
16964 #include <dtrace_clone.c>
16965 #endif
16966 #include <dtrace_ioctl.c>
16967 #include <dtrace_load.c>
16968 #include <dtrace_modevent.c>
16969 #include <dtrace_sysctl.c>
16970 #include <dtrace_unload.c>
16971 #include <dtrace_vtime.c>
16972 #include <dtrace_hacks.c>
16973 #include <dtrace_isa.c>
16974 
16975 SYSINIT(dtrace_load, SI_SUB_DTRACE, SI_ORDER_FIRST, dtrace_load, NULL);
16976 SYSUNINIT(dtrace_unload, SI_SUB_DTRACE, SI_ORDER_FIRST, dtrace_unload, NULL);
16977 SYSINIT(dtrace_anon_init, SI_SUB_DTRACE_ANON, SI_ORDER_FIRST, dtrace_anon_init, NULL);
16978 
16979 DEV_MODULE(dtrace, dtrace_modevent, NULL);
16980 MODULE_VERSION(dtrace, 1);
16981 MODULE_DEPEND(dtrace, cyclic, 1, 1, 1);
16982 MODULE_DEPEND(dtrace, opensolaris, 1, 1, 1);
16983 #endif
16984