xref: /freebsd/sys/cddl/contrib/opensolaris/uts/common/dtrace/dtrace.c (revision 656f49f8e2b0656824a5f10aeb760a00fdd3753f)
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 (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
26  * Copyright (c) 2015, Joyent, Inc. All rights reserved.
27  * Copyright (c) 2012, 2014 by Delphix. All rights reserved.
28  */
29 
30 /*
31  * DTrace - Dynamic Tracing for Solaris
32  *
33  * This is the implementation of the Solaris Dynamic Tracing framework
34  * (DTrace).  The user-visible interface to DTrace is described at length in
35  * the "Solaris Dynamic Tracing Guide".  The interfaces between the libdtrace
36  * library, the in-kernel DTrace framework, and the DTrace providers are
37  * described in the block comments in the <sys/dtrace.h> header file.  The
38  * internal architecture of DTrace is described in the block comments in the
39  * <sys/dtrace_impl.h> header file.  The comments contained within the DTrace
40  * implementation very much assume mastery of all of these sources; if one has
41  * an unanswered question about the implementation, one should consult them
42  * first.
43  *
44  * The functions here are ordered roughly as follows:
45  *
46  *   - Probe context functions
47  *   - Probe hashing functions
48  *   - Non-probe context utility functions
49  *   - Matching functions
50  *   - Provider-to-Framework API functions
51  *   - Probe management functions
52  *   - DIF object functions
53  *   - Format functions
54  *   - Predicate functions
55  *   - ECB functions
56  *   - Buffer functions
57  *   - Enabling functions
58  *   - DOF functions
59  *   - Anonymous enabling functions
60  *   - Consumer state functions
61  *   - Helper functions
62  *   - Hook functions
63  *   - Driver cookbook functions
64  *
65  * Each group of functions begins with a block comment labelled the "DTrace
66  * [Group] Functions", allowing one to find each block by searching forward
67  * on capital-f functions.
68  */
69 #include <sys/errno.h>
70 #ifndef illumos
71 #include <sys/time.h>
72 #endif
73 #include <sys/stat.h>
74 #include <sys/modctl.h>
75 #include <sys/conf.h>
76 #include <sys/systm.h>
77 #ifdef illumos
78 #include <sys/ddi.h>
79 #include <sys/sunddi.h>
80 #endif
81 #include <sys/cpuvar.h>
82 #include <sys/kmem.h>
83 #ifdef illumos
84 #include <sys/strsubr.h>
85 #endif
86 #include <sys/sysmacros.h>
87 #include <sys/dtrace_impl.h>
88 #include <sys/atomic.h>
89 #include <sys/cmn_err.h>
90 #ifdef illumos
91 #include <sys/mutex_impl.h>
92 #include <sys/rwlock_impl.h>
93 #endif
94 #include <sys/ctf_api.h>
95 #ifdef illumos
96 #include <sys/panic.h>
97 #include <sys/priv_impl.h>
98 #endif
99 #include <sys/policy.h>
100 #ifdef illumos
101 #include <sys/cred_impl.h>
102 #include <sys/procfs_isa.h>
103 #endif
104 #include <sys/taskq.h>
105 #ifdef illumos
106 #include <sys/mkdev.h>
107 #include <sys/kdi.h>
108 #endif
109 #include <sys/zone.h>
110 #include <sys/socket.h>
111 #include <netinet/in.h>
112 #include "strtolctype.h"
113 
114 /* FreeBSD includes: */
115 #ifndef illumos
116 #include <sys/callout.h>
117 #include <sys/ctype.h>
118 #include <sys/eventhandler.h>
119 #include <sys/limits.h>
120 #include <sys/kdb.h>
121 #include <sys/kernel.h>
122 #include <sys/malloc.h>
123 #include <sys/sysctl.h>
124 #include <sys/lock.h>
125 #include <sys/mutex.h>
126 #include <sys/rwlock.h>
127 #include <sys/sx.h>
128 #include <sys/dtrace_bsd.h>
129 #include <netinet/in.h>
130 #include "dtrace_cddl.h"
131 #include "dtrace_debug.c"
132 #endif
133 
134 /*
135  * DTrace Tunable Variables
136  *
137  * The following variables may be tuned by adding a line to /etc/system that
138  * includes both the name of the DTrace module ("dtrace") and the name of the
139  * variable.  For example:
140  *
141  *   set dtrace:dtrace_destructive_disallow = 1
142  *
143  * In general, the only variables that one should be tuning this way are those
144  * that affect system-wide DTrace behavior, and for which the default behavior
145  * is undesirable.  Most of these variables are tunable on a per-consumer
146  * basis using DTrace options, and need not be tuned on a system-wide basis.
147  * When tuning these variables, avoid pathological values; while some attempt
148  * is made to verify the integrity of these variables, they are not considered
149  * part of the supported interface to DTrace, and they are therefore not
150  * checked comprehensively.  Further, these variables should not be tuned
151  * dynamically via "mdb -kw" or other means; they should only be tuned via
152  * /etc/system.
153  */
154 int		dtrace_destructive_disallow = 0;
155 dtrace_optval_t	dtrace_nonroot_maxsize = (16 * 1024 * 1024);
156 size_t		dtrace_difo_maxsize = (256 * 1024);
157 dtrace_optval_t	dtrace_dof_maxsize = (8 * 1024 * 1024);
158 size_t		dtrace_statvar_maxsize = (16 * 1024);
159 size_t		dtrace_actions_max = (16 * 1024);
160 size_t		dtrace_retain_max = 1024;
161 dtrace_optval_t	dtrace_helper_actions_max = 128;
162 dtrace_optval_t	dtrace_helper_providers_max = 32;
163 dtrace_optval_t	dtrace_dstate_defsize = (1 * 1024 * 1024);
164 size_t		dtrace_strsize_default = 256;
165 dtrace_optval_t	dtrace_cleanrate_default = 9900990;		/* 101 hz */
166 dtrace_optval_t	dtrace_cleanrate_min = 200000;			/* 5000 hz */
167 dtrace_optval_t	dtrace_cleanrate_max = (uint64_t)60 * NANOSEC;	/* 1/minute */
168 dtrace_optval_t	dtrace_aggrate_default = NANOSEC;		/* 1 hz */
169 dtrace_optval_t	dtrace_statusrate_default = NANOSEC;		/* 1 hz */
170 dtrace_optval_t dtrace_statusrate_max = (hrtime_t)10 * NANOSEC;	 /* 6/minute */
171 dtrace_optval_t	dtrace_switchrate_default = NANOSEC;		/* 1 hz */
172 dtrace_optval_t	dtrace_nspec_default = 1;
173 dtrace_optval_t	dtrace_specsize_default = 32 * 1024;
174 dtrace_optval_t dtrace_stackframes_default = 20;
175 dtrace_optval_t dtrace_ustackframes_default = 20;
176 dtrace_optval_t dtrace_jstackframes_default = 50;
177 dtrace_optval_t dtrace_jstackstrsize_default = 512;
178 int		dtrace_msgdsize_max = 128;
179 hrtime_t	dtrace_chill_max = MSEC2NSEC(500);		/* 500 ms */
180 hrtime_t	dtrace_chill_interval = NANOSEC;		/* 1000 ms */
181 int		dtrace_devdepth_max = 32;
182 int		dtrace_err_verbose;
183 hrtime_t	dtrace_deadman_interval = NANOSEC;
184 hrtime_t	dtrace_deadman_timeout = (hrtime_t)10 * NANOSEC;
185 hrtime_t	dtrace_deadman_user = (hrtime_t)30 * NANOSEC;
186 hrtime_t	dtrace_unregister_defunct_reap = (hrtime_t)60 * NANOSEC;
187 #ifndef illumos
188 int		dtrace_memstr_max = 4096;
189 #endif
190 
191 /*
192  * DTrace External Variables
193  *
194  * As dtrace(7D) is a kernel module, any DTrace variables are obviously
195  * available to DTrace consumers via the backtick (`) syntax.  One of these,
196  * dtrace_zero, is made deliberately so:  it is provided as a source of
197  * well-known, zero-filled memory.  While this variable is not documented,
198  * it is used by some translators as an implementation detail.
199  */
200 const char	dtrace_zero[256] = { 0 };	/* zero-filled memory */
201 
202 /*
203  * DTrace Internal Variables
204  */
205 #ifdef illumos
206 static dev_info_t	*dtrace_devi;		/* device info */
207 #endif
208 #ifdef illumos
209 static vmem_t		*dtrace_arena;		/* probe ID arena */
210 static vmem_t		*dtrace_minor;		/* minor number arena */
211 #else
212 static taskq_t		*dtrace_taskq;		/* task queue */
213 static struct unrhdr	*dtrace_arena;		/* Probe ID number.     */
214 #endif
215 static dtrace_probe_t	**dtrace_probes;	/* array of all probes */
216 static int		dtrace_nprobes;		/* number of probes */
217 static dtrace_provider_t *dtrace_provider;	/* provider list */
218 static dtrace_meta_t	*dtrace_meta_pid;	/* user-land meta provider */
219 static int		dtrace_opens;		/* number of opens */
220 static int		dtrace_helpers;		/* number of helpers */
221 static int		dtrace_getf;		/* number of unpriv getf()s */
222 #ifdef illumos
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_genid_t	dtrace_retained_gen;	/* current retained enab gen */
240 static dtrace_dynvar_t	dtrace_dynhash_sink;	/* end of dynamic hash chains */
241 static int		dtrace_dynvar_failclean; /* dynvars failed to clean */
242 #ifndef illumos
243 static struct mtx	dtrace_unr_mtx;
244 MTX_SYSINIT(dtrace_unr_mtx, &dtrace_unr_mtx, "Unique resource identifier", MTX_DEF);
245 int		dtrace_in_probe;	/* non-zero if executing a probe */
246 #if defined(__i386__) || defined(__amd64__) || defined(__mips__) || defined(__powerpc__)
247 uintptr_t	dtrace_in_probe_addr;	/* Address of invop when already in probe */
248 #endif
249 static eventhandler_tag	dtrace_kld_load_tag;
250 static eventhandler_tag	dtrace_kld_unload_try_tag;
251 #endif
252 
253 /*
254  * DTrace Locking
255  * DTrace is protected by three (relatively coarse-grained) locks:
256  *
257  * (1) dtrace_lock is required to manipulate essentially any DTrace state,
258  *     including enabling state, probes, ECBs, consumer state, helper state,
259  *     etc.  Importantly, dtrace_lock is _not_ required when in probe context;
260  *     probe context is lock-free -- synchronization is handled via the
261  *     dtrace_sync() cross call mechanism.
262  *
263  * (2) dtrace_provider_lock is required when manipulating provider state, or
264  *     when provider state must be held constant.
265  *
266  * (3) dtrace_meta_lock is required when manipulating meta provider state, or
267  *     when meta provider state must be held constant.
268  *
269  * The lock ordering between these three locks is dtrace_meta_lock before
270  * dtrace_provider_lock before dtrace_lock.  (In particular, there are
271  * several places where dtrace_provider_lock is held by the framework as it
272  * calls into the providers -- which then call back into the framework,
273  * grabbing dtrace_lock.)
274  *
275  * There are two other locks in the mix:  mod_lock and cpu_lock.  With respect
276  * to dtrace_provider_lock and dtrace_lock, cpu_lock continues its historical
277  * role as a coarse-grained lock; it is acquired before both of these locks.
278  * With respect to dtrace_meta_lock, its behavior is stranger:  cpu_lock must
279  * be acquired _between_ dtrace_meta_lock and any other DTrace locks.
280  * mod_lock is similar with respect to dtrace_provider_lock in that it must be
281  * acquired _between_ dtrace_provider_lock and dtrace_lock.
282  */
283 static kmutex_t		dtrace_lock;		/* probe state lock */
284 static kmutex_t		dtrace_provider_lock;	/* provider state lock */
285 static kmutex_t		dtrace_meta_lock;	/* meta-provider state lock */
286 
287 #ifndef illumos
288 /* XXX FreeBSD hacks. */
289 #define cr_suid		cr_svuid
290 #define cr_sgid		cr_svgid
291 #define	ipaddr_t	in_addr_t
292 #define mod_modname	pathname
293 #define vuprintf	vprintf
294 #define ttoproc(_a)	((_a)->td_proc)
295 #define crgetzoneid(_a)	0
296 #define	NCPU		MAXCPU
297 #define SNOCD		0
298 #define CPU_ON_INTR(_a)	0
299 
300 #define PRIV_EFFECTIVE		(1 << 0)
301 #define PRIV_DTRACE_KERNEL	(1 << 1)
302 #define PRIV_DTRACE_PROC	(1 << 2)
303 #define PRIV_DTRACE_USER	(1 << 3)
304 #define PRIV_PROC_OWNER		(1 << 4)
305 #define PRIV_PROC_ZONE		(1 << 5)
306 #define PRIV_ALL		~0
307 
308 SYSCTL_DECL(_debug_dtrace);
309 SYSCTL_DECL(_kern_dtrace);
310 #endif
311 
312 #ifdef illumos
313 #define curcpu	CPU->cpu_id
314 #endif
315 
316 
317 /*
318  * DTrace Provider Variables
319  *
320  * These are the variables relating to DTrace as a provider (that is, the
321  * provider of the BEGIN, END, and ERROR probes).
322  */
323 static dtrace_pattr_t	dtrace_provider_attr = {
324 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
325 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
326 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
327 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
328 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
329 };
330 
331 static void
332 dtrace_nullop(void)
333 {}
334 
335 static dtrace_pops_t	dtrace_provider_ops = {
336 	(void (*)(void *, dtrace_probedesc_t *))dtrace_nullop,
337 	(void (*)(void *, modctl_t *))dtrace_nullop,
338 	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
339 	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
340 	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
341 	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
342 	NULL,
343 	NULL,
344 	NULL,
345 	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop
346 };
347 
348 static dtrace_id_t	dtrace_probeid_begin;	/* special BEGIN probe */
349 static dtrace_id_t	dtrace_probeid_end;	/* special END probe */
350 dtrace_id_t		dtrace_probeid_error;	/* special ERROR probe */
351 
352 /*
353  * DTrace Helper Tracing Variables
354  *
355  * These variables should be set dynamically to enable helper tracing.  The
356  * only variables that should be set are dtrace_helptrace_enable (which should
357  * be set to a non-zero value to allocate helper tracing buffers on the next
358  * open of /dev/dtrace) and dtrace_helptrace_disable (which should be set to a
359  * non-zero value to deallocate helper tracing buffers on the next close of
360  * /dev/dtrace).  When (and only when) helper tracing is disabled, the
361  * buffer size may also be set via dtrace_helptrace_bufsize.
362  */
363 int			dtrace_helptrace_enable = 0;
364 int			dtrace_helptrace_disable = 0;
365 int			dtrace_helptrace_bufsize = 16 * 1024 * 1024;
366 uint32_t		dtrace_helptrace_nlocals;
367 static dtrace_helptrace_t *dtrace_helptrace_buffer;
368 static uint32_t		dtrace_helptrace_next = 0;
369 static int		dtrace_helptrace_wrapped = 0;
370 
371 /*
372  * DTrace Error Hashing
373  *
374  * On DEBUG kernels, DTrace will track the errors that has seen in a hash
375  * table.  This is very useful for checking coverage of tests that are
376  * expected to induce DIF or DOF processing errors, and may be useful for
377  * debugging problems in the DIF code generator or in DOF generation .  The
378  * error hash may be examined with the ::dtrace_errhash MDB dcmd.
379  */
380 #ifdef DEBUG
381 static dtrace_errhash_t	dtrace_errhash[DTRACE_ERRHASHSZ];
382 static const char *dtrace_errlast;
383 static kthread_t *dtrace_errthread;
384 static kmutex_t dtrace_errlock;
385 #endif
386 
387 /*
388  * DTrace Macros and Constants
389  *
390  * These are various macros that are useful in various spots in the
391  * implementation, along with a few random constants that have no meaning
392  * outside of the implementation.  There is no real structure to this cpp
393  * mishmash -- but is there ever?
394  */
395 #define	DTRACE_HASHSTR(hash, probe)	\
396 	dtrace_hash_str(*((char **)((uintptr_t)(probe) + (hash)->dth_stroffs)))
397 
398 #define	DTRACE_HASHNEXT(hash, probe)	\
399 	(dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_nextoffs)
400 
401 #define	DTRACE_HASHPREV(hash, probe)	\
402 	(dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_prevoffs)
403 
404 #define	DTRACE_HASHEQ(hash, lhs, rhs)	\
405 	(strcmp(*((char **)((uintptr_t)(lhs) + (hash)->dth_stroffs)), \
406 	    *((char **)((uintptr_t)(rhs) + (hash)->dth_stroffs))) == 0)
407 
408 #define	DTRACE_AGGHASHSIZE_SLEW		17
409 
410 #define	DTRACE_V4MAPPED_OFFSET		(sizeof (uint32_t) * 3)
411 
412 /*
413  * The key for a thread-local variable consists of the lower 61 bits of the
414  * t_did, plus the 3 bits of the highest active interrupt above LOCK_LEVEL.
415  * We add DIF_VARIABLE_MAX to t_did to assure that the thread key is never
416  * equal to a variable identifier.  This is necessary (but not sufficient) to
417  * assure that global associative arrays never collide with thread-local
418  * variables.  To guarantee that they cannot collide, we must also define the
419  * order for keying dynamic variables.  That order is:
420  *
421  *   [ key0 ] ... [ keyn ] [ variable-key ] [ tls-key ]
422  *
423  * Because the variable-key and the tls-key are in orthogonal spaces, there is
424  * no way for a global variable key signature to match a thread-local key
425  * signature.
426  */
427 #ifdef illumos
428 #define	DTRACE_TLS_THRKEY(where) { \
429 	uint_t intr = 0; \
430 	uint_t actv = CPU->cpu_intr_actv >> (LOCK_LEVEL + 1); \
431 	for (; actv; actv >>= 1) \
432 		intr++; \
433 	ASSERT(intr < (1 << 3)); \
434 	(where) = ((curthread->t_did + DIF_VARIABLE_MAX) & \
435 	    (((uint64_t)1 << 61) - 1)) | ((uint64_t)intr << 61); \
436 }
437 #else
438 #define	DTRACE_TLS_THRKEY(where) { \
439 	solaris_cpu_t *_c = &solaris_cpu[curcpu]; \
440 	uint_t intr = 0; \
441 	uint_t actv = _c->cpu_intr_actv; \
442 	for (; actv; actv >>= 1) \
443 		intr++; \
444 	ASSERT(intr < (1 << 3)); \
445 	(where) = ((curthread->td_tid + DIF_VARIABLE_MAX) & \
446 	    (((uint64_t)1 << 61) - 1)) | ((uint64_t)intr << 61); \
447 }
448 #endif
449 
450 #define	DT_BSWAP_8(x)	((x) & 0xff)
451 #define	DT_BSWAP_16(x)	((DT_BSWAP_8(x) << 8) | DT_BSWAP_8((x) >> 8))
452 #define	DT_BSWAP_32(x)	((DT_BSWAP_16(x) << 16) | DT_BSWAP_16((x) >> 16))
453 #define	DT_BSWAP_64(x)	((DT_BSWAP_32(x) << 32) | DT_BSWAP_32((x) >> 32))
454 
455 #define	DT_MASK_LO 0x00000000FFFFFFFFULL
456 
457 #define	DTRACE_STORE(type, tomax, offset, what) \
458 	*((type *)((uintptr_t)(tomax) + (uintptr_t)offset)) = (type)(what);
459 
460 #ifndef __x86
461 #define	DTRACE_ALIGNCHECK(addr, size, flags)				\
462 	if (addr & (size - 1)) {					\
463 		*flags |= CPU_DTRACE_BADALIGN;				\
464 		cpu_core[curcpu].cpuc_dtrace_illval = addr;	\
465 		return (0);						\
466 	}
467 #else
468 #define	DTRACE_ALIGNCHECK(addr, size, flags)
469 #endif
470 
471 /*
472  * Test whether a range of memory starting at testaddr of size testsz falls
473  * within the range of memory described by addr, sz.  We take care to avoid
474  * problems with overflow and underflow of the unsigned quantities, and
475  * disallow all negative sizes.  Ranges of size 0 are allowed.
476  */
477 #define	DTRACE_INRANGE(testaddr, testsz, baseaddr, basesz) \
478 	((testaddr) - (uintptr_t)(baseaddr) < (basesz) && \
479 	(testaddr) + (testsz) - (uintptr_t)(baseaddr) <= (basesz) && \
480 	(testaddr) + (testsz) >= (testaddr))
481 
482 /*
483  * Test whether alloc_sz bytes will fit in the scratch region.  We isolate
484  * alloc_sz on the righthand side of the comparison in order to avoid overflow
485  * or underflow in the comparison with it.  This is simpler than the INRANGE
486  * check above, because we know that the dtms_scratch_ptr is valid in the
487  * range.  Allocations of size zero are allowed.
488  */
489 #define	DTRACE_INSCRATCH(mstate, alloc_sz) \
490 	((mstate)->dtms_scratch_base + (mstate)->dtms_scratch_size - \
491 	(mstate)->dtms_scratch_ptr >= (alloc_sz))
492 
493 #define	DTRACE_LOADFUNC(bits)						\
494 /*CSTYLED*/								\
495 uint##bits##_t								\
496 dtrace_load##bits(uintptr_t addr)					\
497 {									\
498 	size_t size = bits / NBBY;					\
499 	/*CSTYLED*/							\
500 	uint##bits##_t rval;						\
501 	int i;								\
502 	volatile uint16_t *flags = (volatile uint16_t *)		\
503 	    &cpu_core[curcpu].cpuc_dtrace_flags;			\
504 									\
505 	DTRACE_ALIGNCHECK(addr, size, flags);				\
506 									\
507 	for (i = 0; i < dtrace_toxranges; i++) {			\
508 		if (addr >= dtrace_toxrange[i].dtt_limit)		\
509 			continue;					\
510 									\
511 		if (addr + size <= dtrace_toxrange[i].dtt_base)		\
512 			continue;					\
513 									\
514 		/*							\
515 		 * This address falls within a toxic region; return 0.	\
516 		 */							\
517 		*flags |= CPU_DTRACE_BADADDR;				\
518 		cpu_core[curcpu].cpuc_dtrace_illval = addr;		\
519 		return (0);						\
520 	}								\
521 									\
522 	*flags |= CPU_DTRACE_NOFAULT;					\
523 	/*CSTYLED*/							\
524 	rval = *((volatile uint##bits##_t *)addr);			\
525 	*flags &= ~CPU_DTRACE_NOFAULT;					\
526 									\
527 	return (!(*flags & CPU_DTRACE_FAULT) ? rval : 0);		\
528 }
529 
530 #ifdef _LP64
531 #define	dtrace_loadptr	dtrace_load64
532 #else
533 #define	dtrace_loadptr	dtrace_load32
534 #endif
535 
536 #define	DTRACE_DYNHASH_FREE	0
537 #define	DTRACE_DYNHASH_SINK	1
538 #define	DTRACE_DYNHASH_VALID	2
539 
540 #define	DTRACE_MATCH_NEXT	0
541 #define	DTRACE_MATCH_DONE	1
542 #define	DTRACE_ANCHORED(probe)	((probe)->dtpr_func[0] != '\0')
543 #define	DTRACE_STATE_ALIGN	64
544 
545 #define	DTRACE_FLAGS2FLT(flags)						\
546 	(((flags) & CPU_DTRACE_BADADDR) ? DTRACEFLT_BADADDR :		\
547 	((flags) & CPU_DTRACE_ILLOP) ? DTRACEFLT_ILLOP :		\
548 	((flags) & CPU_DTRACE_DIVZERO) ? DTRACEFLT_DIVZERO :		\
549 	((flags) & CPU_DTRACE_KPRIV) ? DTRACEFLT_KPRIV :		\
550 	((flags) & CPU_DTRACE_UPRIV) ? DTRACEFLT_UPRIV :		\
551 	((flags) & CPU_DTRACE_TUPOFLOW) ?  DTRACEFLT_TUPOFLOW :		\
552 	((flags) & CPU_DTRACE_BADALIGN) ?  DTRACEFLT_BADALIGN :		\
553 	((flags) & CPU_DTRACE_NOSCRATCH) ?  DTRACEFLT_NOSCRATCH :	\
554 	((flags) & CPU_DTRACE_BADSTACK) ?  DTRACEFLT_BADSTACK :		\
555 	DTRACEFLT_UNKNOWN)
556 
557 #define	DTRACEACT_ISSTRING(act)						\
558 	((act)->dta_kind == DTRACEACT_DIFEXPR &&			\
559 	(act)->dta_difo->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING)
560 
561 /* Function prototype definitions: */
562 static size_t dtrace_strlen(const char *, size_t);
563 static dtrace_probe_t *dtrace_probe_lookup_id(dtrace_id_t id);
564 static void dtrace_enabling_provide(dtrace_provider_t *);
565 static int dtrace_enabling_match(dtrace_enabling_t *, int *);
566 static void dtrace_enabling_matchall(void);
567 static void dtrace_enabling_reap(void);
568 static dtrace_state_t *dtrace_anon_grab(void);
569 static uint64_t dtrace_helper(int, dtrace_mstate_t *,
570     dtrace_state_t *, uint64_t, uint64_t);
571 static dtrace_helpers_t *dtrace_helpers_create(proc_t *);
572 static void dtrace_buffer_drop(dtrace_buffer_t *);
573 static int dtrace_buffer_consumed(dtrace_buffer_t *, hrtime_t when);
574 static intptr_t dtrace_buffer_reserve(dtrace_buffer_t *, size_t, size_t,
575     dtrace_state_t *, dtrace_mstate_t *);
576 static int dtrace_state_option(dtrace_state_t *, dtrace_optid_t,
577     dtrace_optval_t);
578 static int dtrace_ecb_create_enable(dtrace_probe_t *, void *);
579 static void dtrace_helper_provider_destroy(dtrace_helper_provider_t *);
580 uint16_t dtrace_load16(uintptr_t);
581 uint32_t dtrace_load32(uintptr_t);
582 uint64_t dtrace_load64(uintptr_t);
583 uint8_t dtrace_load8(uintptr_t);
584 void dtrace_dynvar_clean(dtrace_dstate_t *);
585 dtrace_dynvar_t *dtrace_dynvar(dtrace_dstate_t *, uint_t, dtrace_key_t *,
586     size_t, dtrace_dynvar_op_t, dtrace_mstate_t *, dtrace_vstate_t *);
587 uintptr_t dtrace_dif_varstr(uintptr_t, dtrace_state_t *, dtrace_mstate_t *);
588 static int dtrace_priv_proc(dtrace_state_t *);
589 static void dtrace_getf_barrier(void);
590 
591 /*
592  * DTrace Probe Context Functions
593  *
594  * These functions are called from probe context.  Because probe context is
595  * any context in which C may be called, arbitrarily locks may be held,
596  * interrupts may be disabled, we may be in arbitrary dispatched state, etc.
597  * As a result, functions called from probe context may only call other DTrace
598  * support functions -- they may not interact at all with the system at large.
599  * (Note that the ASSERT macro is made probe-context safe by redefining it in
600  * terms of dtrace_assfail(), a probe-context safe function.) If arbitrary
601  * loads are to be performed from probe context, they _must_ be in terms of
602  * the safe dtrace_load*() variants.
603  *
604  * Some functions in this block are not actually called from probe context;
605  * for these functions, there will be a comment above the function reading
606  * "Note:  not called from probe context."
607  */
608 void
609 dtrace_panic(const char *format, ...)
610 {
611 	va_list alist;
612 
613 	va_start(alist, format);
614 #ifdef __FreeBSD__
615 	vpanic(format, alist);
616 #else
617 	dtrace_vpanic(format, alist);
618 #endif
619 	va_end(alist);
620 }
621 
622 int
623 dtrace_assfail(const char *a, const char *f, int l)
624 {
625 	dtrace_panic("assertion failed: %s, file: %s, line: %d", a, f, l);
626 
627 	/*
628 	 * We just need something here that even the most clever compiler
629 	 * cannot optimize away.
630 	 */
631 	return (a[(uintptr_t)f]);
632 }
633 
634 /*
635  * Atomically increment a specified error counter from probe context.
636  */
637 static void
638 dtrace_error(uint32_t *counter)
639 {
640 	/*
641 	 * Most counters stored to in probe context are per-CPU counters.
642 	 * However, there are some error conditions that are sufficiently
643 	 * arcane that they don't merit per-CPU storage.  If these counters
644 	 * are incremented concurrently on different CPUs, scalability will be
645 	 * adversely affected -- but we don't expect them to be white-hot in a
646 	 * correctly constructed enabling...
647 	 */
648 	uint32_t oval, nval;
649 
650 	do {
651 		oval = *counter;
652 
653 		if ((nval = oval + 1) == 0) {
654 			/*
655 			 * If the counter would wrap, set it to 1 -- assuring
656 			 * that the counter is never zero when we have seen
657 			 * errors.  (The counter must be 32-bits because we
658 			 * aren't guaranteed a 64-bit compare&swap operation.)
659 			 * To save this code both the infamy of being fingered
660 			 * by a priggish news story and the indignity of being
661 			 * the target of a neo-puritan witch trial, we're
662 			 * carefully avoiding any colorful description of the
663 			 * likelihood of this condition -- but suffice it to
664 			 * say that it is only slightly more likely than the
665 			 * overflow of predicate cache IDs, as discussed in
666 			 * dtrace_predicate_create().
667 			 */
668 			nval = 1;
669 		}
670 	} while (dtrace_cas32(counter, oval, nval) != oval);
671 }
672 
673 /*
674  * Use the DTRACE_LOADFUNC macro to define functions for each of loading a
675  * uint8_t, a uint16_t, a uint32_t and a uint64_t.
676  */
677 DTRACE_LOADFUNC(8)
678 DTRACE_LOADFUNC(16)
679 DTRACE_LOADFUNC(32)
680 DTRACE_LOADFUNC(64)
681 
682 static int
683 dtrace_inscratch(uintptr_t dest, size_t size, dtrace_mstate_t *mstate)
684 {
685 	if (dest < mstate->dtms_scratch_base)
686 		return (0);
687 
688 	if (dest + size < dest)
689 		return (0);
690 
691 	if (dest + size > mstate->dtms_scratch_ptr)
692 		return (0);
693 
694 	return (1);
695 }
696 
697 static int
698 dtrace_canstore_statvar(uint64_t addr, size_t sz,
699     dtrace_statvar_t **svars, int nsvars)
700 {
701 	int i;
702 	size_t maxglobalsize, maxlocalsize;
703 
704 	if (nsvars == 0)
705 		return (0);
706 
707 	maxglobalsize = dtrace_statvar_maxsize;
708 	maxlocalsize = (maxglobalsize + sizeof (uint64_t)) * NCPU;
709 
710 	for (i = 0; i < nsvars; i++) {
711 		dtrace_statvar_t *svar = svars[i];
712 		uint8_t scope;
713 		size_t size;
714 
715 		if (svar == NULL || (size = svar->dtsv_size) == 0)
716 			continue;
717 
718 		scope = svar->dtsv_var.dtdv_scope;
719 
720 		/*
721 		 * We verify that our size is valid in the spirit of providing
722 		 * defense in depth:  we want to prevent attackers from using
723 		 * DTrace to escalate an orthogonal kernel heap corruption bug
724 		 * into the ability to store to arbitrary locations in memory.
725 		 */
726 		VERIFY((scope == DIFV_SCOPE_GLOBAL && size < maxglobalsize) ||
727 		    (scope == DIFV_SCOPE_LOCAL && size < maxlocalsize));
728 
729 		if (DTRACE_INRANGE(addr, sz, svar->dtsv_data, svar->dtsv_size))
730 			return (1);
731 	}
732 
733 	return (0);
734 }
735 
736 /*
737  * Check to see if the address is within a memory region to which a store may
738  * be issued.  This includes the DTrace scratch areas, and any DTrace variable
739  * region.  The caller of dtrace_canstore() is responsible for performing any
740  * alignment checks that are needed before stores are actually executed.
741  */
742 static int
743 dtrace_canstore(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
744     dtrace_vstate_t *vstate)
745 {
746 	/*
747 	 * First, check to see if the address is in scratch space...
748 	 */
749 	if (DTRACE_INRANGE(addr, sz, mstate->dtms_scratch_base,
750 	    mstate->dtms_scratch_size))
751 		return (1);
752 
753 	/*
754 	 * Now check to see if it's a dynamic variable.  This check will pick
755 	 * up both thread-local variables and any global dynamically-allocated
756 	 * variables.
757 	 */
758 	if (DTRACE_INRANGE(addr, sz, vstate->dtvs_dynvars.dtds_base,
759 	    vstate->dtvs_dynvars.dtds_size)) {
760 		dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
761 		uintptr_t base = (uintptr_t)dstate->dtds_base +
762 		    (dstate->dtds_hashsize * sizeof (dtrace_dynhash_t));
763 		uintptr_t chunkoffs;
764 
765 		/*
766 		 * Before we assume that we can store here, we need to make
767 		 * sure that it isn't in our metadata -- storing to our
768 		 * dynamic variable metadata would corrupt our state.  For
769 		 * the range to not include any dynamic variable metadata,
770 		 * it must:
771 		 *
772 		 *	(1) Start above the hash table that is at the base of
773 		 *	the dynamic variable space
774 		 *
775 		 *	(2) Have a starting chunk offset that is beyond the
776 		 *	dtrace_dynvar_t that is at the base of every chunk
777 		 *
778 		 *	(3) Not span a chunk boundary
779 		 *
780 		 */
781 		if (addr < base)
782 			return (0);
783 
784 		chunkoffs = (addr - base) % dstate->dtds_chunksize;
785 
786 		if (chunkoffs < sizeof (dtrace_dynvar_t))
787 			return (0);
788 
789 		if (chunkoffs + sz > dstate->dtds_chunksize)
790 			return (0);
791 
792 		return (1);
793 	}
794 
795 	/*
796 	 * Finally, check the static local and global variables.  These checks
797 	 * take the longest, so we perform them last.
798 	 */
799 	if (dtrace_canstore_statvar(addr, sz,
800 	    vstate->dtvs_locals, vstate->dtvs_nlocals))
801 		return (1);
802 
803 	if (dtrace_canstore_statvar(addr, sz,
804 	    vstate->dtvs_globals, vstate->dtvs_nglobals))
805 		return (1);
806 
807 	return (0);
808 }
809 
810 
811 /*
812  * Convenience routine to check to see if the address is within a memory
813  * region in which a load may be issued given the user's privilege level;
814  * if not, it sets the appropriate error flags and loads 'addr' into the
815  * illegal value slot.
816  *
817  * DTrace subroutines (DIF_SUBR_*) should use this helper to implement
818  * appropriate memory access protection.
819  */
820 static int
821 dtrace_canload(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
822     dtrace_vstate_t *vstate)
823 {
824 	volatile uintptr_t *illval = &cpu_core[curcpu].cpuc_dtrace_illval;
825 	file_t *fp;
826 
827 	/*
828 	 * If we hold the privilege to read from kernel memory, then
829 	 * everything is readable.
830 	 */
831 	if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
832 		return (1);
833 
834 	/*
835 	 * You can obviously read that which you can store.
836 	 */
837 	if (dtrace_canstore(addr, sz, mstate, vstate))
838 		return (1);
839 
840 	/*
841 	 * We're allowed to read from our own string table.
842 	 */
843 	if (DTRACE_INRANGE(addr, sz, mstate->dtms_difo->dtdo_strtab,
844 	    mstate->dtms_difo->dtdo_strlen))
845 		return (1);
846 
847 	if (vstate->dtvs_state != NULL &&
848 	    dtrace_priv_proc(vstate->dtvs_state)) {
849 		proc_t *p;
850 
851 		/*
852 		 * When we have privileges to the current process, there are
853 		 * several context-related kernel structures that are safe to
854 		 * read, even absent the privilege to read from kernel memory.
855 		 * These reads are safe because these structures contain only
856 		 * state that (1) we're permitted to read, (2) is harmless or
857 		 * (3) contains pointers to additional kernel state that we're
858 		 * not permitted to read (and as such, do not present an
859 		 * opportunity for privilege escalation).  Finally (and
860 		 * critically), because of the nature of their relation with
861 		 * the current thread context, the memory associated with these
862 		 * structures cannot change over the duration of probe context,
863 		 * and it is therefore impossible for this memory to be
864 		 * deallocated and reallocated as something else while it's
865 		 * being operated upon.
866 		 */
867 		if (DTRACE_INRANGE(addr, sz, curthread, sizeof (kthread_t)))
868 			return (1);
869 
870 		if ((p = curthread->t_procp) != NULL && DTRACE_INRANGE(addr,
871 		    sz, curthread->t_procp, sizeof (proc_t))) {
872 			return (1);
873 		}
874 
875 		if (curthread->t_cred != NULL && DTRACE_INRANGE(addr, sz,
876 		    curthread->t_cred, sizeof (cred_t))) {
877 			return (1);
878 		}
879 
880 #ifdef illumos
881 		if (p != NULL && p->p_pidp != NULL && DTRACE_INRANGE(addr, sz,
882 		    &(p->p_pidp->pid_id), sizeof (pid_t))) {
883 			return (1);
884 		}
885 
886 		if (curthread->t_cpu != NULL && DTRACE_INRANGE(addr, sz,
887 		    curthread->t_cpu, offsetof(cpu_t, cpu_pause_thread))) {
888 			return (1);
889 		}
890 #endif
891 	}
892 
893 	if ((fp = mstate->dtms_getf) != NULL) {
894 		uintptr_t psz = sizeof (void *);
895 		vnode_t *vp;
896 		vnodeops_t *op;
897 
898 		/*
899 		 * When getf() returns a file_t, the enabling is implicitly
900 		 * granted the (transient) right to read the returned file_t
901 		 * as well as the v_path and v_op->vnop_name of the underlying
902 		 * vnode.  These accesses are allowed after a successful
903 		 * getf() because the members that they refer to cannot change
904 		 * once set -- and the barrier logic in the kernel's closef()
905 		 * path assures that the file_t and its referenced vode_t
906 		 * cannot themselves be stale (that is, it impossible for
907 		 * either dtms_getf itself or its f_vnode member to reference
908 		 * freed memory).
909 		 */
910 		if (DTRACE_INRANGE(addr, sz, fp, sizeof (file_t)))
911 			return (1);
912 
913 		if ((vp = fp->f_vnode) != NULL) {
914 #ifdef illumos
915 			if (DTRACE_INRANGE(addr, sz, &vp->v_path, psz))
916 				return (1);
917 			if (vp->v_path != NULL && DTRACE_INRANGE(addr, sz,
918 			    vp->v_path, strlen(vp->v_path) + 1)) {
919 				return (1);
920 			}
921 #endif
922 
923 			if (DTRACE_INRANGE(addr, sz, &vp->v_op, psz))
924 				return (1);
925 
926 #ifdef illumos
927 			if ((op = vp->v_op) != NULL &&
928 			    DTRACE_INRANGE(addr, sz, &op->vnop_name, psz)) {
929 				return (1);
930 			}
931 
932 			if (op != NULL && op->vnop_name != NULL &&
933 			    DTRACE_INRANGE(addr, sz, op->vnop_name,
934 			    strlen(op->vnop_name) + 1)) {
935 				return (1);
936 			}
937 #endif
938 		}
939 	}
940 
941 	DTRACE_CPUFLAG_SET(CPU_DTRACE_KPRIV);
942 	*illval = addr;
943 	return (0);
944 }
945 
946 /*
947  * Convenience routine to check to see if a given string is within a memory
948  * region in which a load may be issued given the user's privilege level;
949  * this exists so that we don't need to issue unnecessary dtrace_strlen()
950  * calls in the event that the user has all privileges.
951  */
952 static int
953 dtrace_strcanload(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
954     dtrace_vstate_t *vstate)
955 {
956 	size_t strsz;
957 
958 	/*
959 	 * If we hold the privilege to read from kernel memory, then
960 	 * everything is readable.
961 	 */
962 	if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
963 		return (1);
964 
965 	strsz = 1 + dtrace_strlen((char *)(uintptr_t)addr, sz);
966 	if (dtrace_canload(addr, strsz, mstate, vstate))
967 		return (1);
968 
969 	return (0);
970 }
971 
972 /*
973  * Convenience routine to check to see if a given variable is within a memory
974  * region in which a load may be issued given the user's privilege level.
975  */
976 static int
977 dtrace_vcanload(void *src, dtrace_diftype_t *type, dtrace_mstate_t *mstate,
978     dtrace_vstate_t *vstate)
979 {
980 	size_t sz;
981 	ASSERT(type->dtdt_flags & DIF_TF_BYREF);
982 
983 	/*
984 	 * If we hold the privilege to read from kernel memory, then
985 	 * everything is readable.
986 	 */
987 	if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
988 		return (1);
989 
990 	if (type->dtdt_kind == DIF_TYPE_STRING)
991 		sz = dtrace_strlen(src,
992 		    vstate->dtvs_state->dts_options[DTRACEOPT_STRSIZE]) + 1;
993 	else
994 		sz = type->dtdt_size;
995 
996 	return (dtrace_canload((uintptr_t)src, sz, mstate, vstate));
997 }
998 
999 /*
1000  * Convert a string to a signed integer using safe loads.
1001  *
1002  * NOTE: This function uses various macros from strtolctype.h to manipulate
1003  * digit values, etc -- these have all been checked to ensure they make
1004  * no additional function calls.
1005  */
1006 static int64_t
1007 dtrace_strtoll(char *input, int base, size_t limit)
1008 {
1009 	uintptr_t pos = (uintptr_t)input;
1010 	int64_t val = 0;
1011 	int x;
1012 	boolean_t neg = B_FALSE;
1013 	char c, cc, ccc;
1014 	uintptr_t end = pos + limit;
1015 
1016 	/*
1017 	 * Consume any whitespace preceding digits.
1018 	 */
1019 	while ((c = dtrace_load8(pos)) == ' ' || c == '\t')
1020 		pos++;
1021 
1022 	/*
1023 	 * Handle an explicit sign if one is present.
1024 	 */
1025 	if (c == '-' || c == '+') {
1026 		if (c == '-')
1027 			neg = B_TRUE;
1028 		c = dtrace_load8(++pos);
1029 	}
1030 
1031 	/*
1032 	 * Check for an explicit hexadecimal prefix ("0x" or "0X") and skip it
1033 	 * if present.
1034 	 */
1035 	if (base == 16 && c == '0' && ((cc = dtrace_load8(pos + 1)) == 'x' ||
1036 	    cc == 'X') && isxdigit(ccc = dtrace_load8(pos + 2))) {
1037 		pos += 2;
1038 		c = ccc;
1039 	}
1040 
1041 	/*
1042 	 * Read in contiguous digits until the first non-digit character.
1043 	 */
1044 	for (; pos < end && c != '\0' && lisalnum(c) && (x = DIGIT(c)) < base;
1045 	    c = dtrace_load8(++pos))
1046 		val = val * base + x;
1047 
1048 	return (neg ? -val : val);
1049 }
1050 
1051 /*
1052  * Compare two strings using safe loads.
1053  */
1054 static int
1055 dtrace_strncmp(char *s1, char *s2, size_t limit)
1056 {
1057 	uint8_t c1, c2;
1058 	volatile uint16_t *flags;
1059 
1060 	if (s1 == s2 || limit == 0)
1061 		return (0);
1062 
1063 	flags = (volatile uint16_t *)&cpu_core[curcpu].cpuc_dtrace_flags;
1064 
1065 	do {
1066 		if (s1 == NULL) {
1067 			c1 = '\0';
1068 		} else {
1069 			c1 = dtrace_load8((uintptr_t)s1++);
1070 		}
1071 
1072 		if (s2 == NULL) {
1073 			c2 = '\0';
1074 		} else {
1075 			c2 = dtrace_load8((uintptr_t)s2++);
1076 		}
1077 
1078 		if (c1 != c2)
1079 			return (c1 - c2);
1080 	} while (--limit && c1 != '\0' && !(*flags & CPU_DTRACE_FAULT));
1081 
1082 	return (0);
1083 }
1084 
1085 /*
1086  * Compute strlen(s) for a string using safe memory accesses.  The additional
1087  * len parameter is used to specify a maximum length to ensure completion.
1088  */
1089 static size_t
1090 dtrace_strlen(const char *s, size_t lim)
1091 {
1092 	uint_t len;
1093 
1094 	for (len = 0; len != lim; len++) {
1095 		if (dtrace_load8((uintptr_t)s++) == '\0')
1096 			break;
1097 	}
1098 
1099 	return (len);
1100 }
1101 
1102 /*
1103  * Check if an address falls within a toxic region.
1104  */
1105 static int
1106 dtrace_istoxic(uintptr_t kaddr, size_t size)
1107 {
1108 	uintptr_t taddr, tsize;
1109 	int i;
1110 
1111 	for (i = 0; i < dtrace_toxranges; i++) {
1112 		taddr = dtrace_toxrange[i].dtt_base;
1113 		tsize = dtrace_toxrange[i].dtt_limit - taddr;
1114 
1115 		if (kaddr - taddr < tsize) {
1116 			DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
1117 			cpu_core[curcpu].cpuc_dtrace_illval = kaddr;
1118 			return (1);
1119 		}
1120 
1121 		if (taddr - kaddr < size) {
1122 			DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
1123 			cpu_core[curcpu].cpuc_dtrace_illval = taddr;
1124 			return (1);
1125 		}
1126 	}
1127 
1128 	return (0);
1129 }
1130 
1131 /*
1132  * Copy src to dst using safe memory accesses.  The src is assumed to be unsafe
1133  * memory specified by the DIF program.  The dst is assumed to be safe memory
1134  * that we can store to directly because it is managed by DTrace.  As with
1135  * standard bcopy, overlapping copies are handled properly.
1136  */
1137 static void
1138 dtrace_bcopy(const void *src, void *dst, size_t len)
1139 {
1140 	if (len != 0) {
1141 		uint8_t *s1 = dst;
1142 		const uint8_t *s2 = src;
1143 
1144 		if (s1 <= s2) {
1145 			do {
1146 				*s1++ = dtrace_load8((uintptr_t)s2++);
1147 			} while (--len != 0);
1148 		} else {
1149 			s2 += len;
1150 			s1 += len;
1151 
1152 			do {
1153 				*--s1 = dtrace_load8((uintptr_t)--s2);
1154 			} while (--len != 0);
1155 		}
1156 	}
1157 }
1158 
1159 /*
1160  * Copy src to dst using safe memory accesses, up to either the specified
1161  * length, or the point that a nul byte is encountered.  The src is assumed to
1162  * be unsafe memory specified by the DIF program.  The dst is assumed to be
1163  * safe memory that we can store to directly because it is managed by DTrace.
1164  * Unlike dtrace_bcopy(), overlapping regions are not handled.
1165  */
1166 static void
1167 dtrace_strcpy(const void *src, void *dst, size_t len)
1168 {
1169 	if (len != 0) {
1170 		uint8_t *s1 = dst, c;
1171 		const uint8_t *s2 = src;
1172 
1173 		do {
1174 			*s1++ = c = dtrace_load8((uintptr_t)s2++);
1175 		} while (--len != 0 && c != '\0');
1176 	}
1177 }
1178 
1179 /*
1180  * Copy src to dst, deriving the size and type from the specified (BYREF)
1181  * variable type.  The src is assumed to be unsafe memory specified by the DIF
1182  * program.  The dst is assumed to be DTrace variable memory that is of the
1183  * specified type; we assume that we can store to directly.
1184  */
1185 static void
1186 dtrace_vcopy(void *src, void *dst, dtrace_diftype_t *type)
1187 {
1188 	ASSERT(type->dtdt_flags & DIF_TF_BYREF);
1189 
1190 	if (type->dtdt_kind == DIF_TYPE_STRING) {
1191 		dtrace_strcpy(src, dst, type->dtdt_size);
1192 	} else {
1193 		dtrace_bcopy(src, dst, type->dtdt_size);
1194 	}
1195 }
1196 
1197 /*
1198  * Compare s1 to s2 using safe memory accesses.  The s1 data is assumed to be
1199  * unsafe memory specified by the DIF program.  The s2 data is assumed to be
1200  * safe memory that we can access directly because it is managed by DTrace.
1201  */
1202 static int
1203 dtrace_bcmp(const void *s1, const void *s2, size_t len)
1204 {
1205 	volatile uint16_t *flags;
1206 
1207 	flags = (volatile uint16_t *)&cpu_core[curcpu].cpuc_dtrace_flags;
1208 
1209 	if (s1 == s2)
1210 		return (0);
1211 
1212 	if (s1 == NULL || s2 == NULL)
1213 		return (1);
1214 
1215 	if (s1 != s2 && len != 0) {
1216 		const uint8_t *ps1 = s1;
1217 		const uint8_t *ps2 = s2;
1218 
1219 		do {
1220 			if (dtrace_load8((uintptr_t)ps1++) != *ps2++)
1221 				return (1);
1222 		} while (--len != 0 && !(*flags & CPU_DTRACE_FAULT));
1223 	}
1224 	return (0);
1225 }
1226 
1227 /*
1228  * Zero the specified region using a simple byte-by-byte loop.  Note that this
1229  * is for safe DTrace-managed memory only.
1230  */
1231 static void
1232 dtrace_bzero(void *dst, size_t len)
1233 {
1234 	uchar_t *cp;
1235 
1236 	for (cp = dst; len != 0; len--)
1237 		*cp++ = 0;
1238 }
1239 
1240 static void
1241 dtrace_add_128(uint64_t *addend1, uint64_t *addend2, uint64_t *sum)
1242 {
1243 	uint64_t result[2];
1244 
1245 	result[0] = addend1[0] + addend2[0];
1246 	result[1] = addend1[1] + addend2[1] +
1247 	    (result[0] < addend1[0] || result[0] < addend2[0] ? 1 : 0);
1248 
1249 	sum[0] = result[0];
1250 	sum[1] = result[1];
1251 }
1252 
1253 /*
1254  * Shift the 128-bit value in a by b. If b is positive, shift left.
1255  * If b is negative, shift right.
1256  */
1257 static void
1258 dtrace_shift_128(uint64_t *a, int b)
1259 {
1260 	uint64_t mask;
1261 
1262 	if (b == 0)
1263 		return;
1264 
1265 	if (b < 0) {
1266 		b = -b;
1267 		if (b >= 64) {
1268 			a[0] = a[1] >> (b - 64);
1269 			a[1] = 0;
1270 		} else {
1271 			a[0] >>= b;
1272 			mask = 1LL << (64 - b);
1273 			mask -= 1;
1274 			a[0] |= ((a[1] & mask) << (64 - b));
1275 			a[1] >>= b;
1276 		}
1277 	} else {
1278 		if (b >= 64) {
1279 			a[1] = a[0] << (b - 64);
1280 			a[0] = 0;
1281 		} else {
1282 			a[1] <<= b;
1283 			mask = a[0] >> (64 - b);
1284 			a[1] |= mask;
1285 			a[0] <<= b;
1286 		}
1287 	}
1288 }
1289 
1290 /*
1291  * The basic idea is to break the 2 64-bit values into 4 32-bit values,
1292  * use native multiplication on those, and then re-combine into the
1293  * resulting 128-bit value.
1294  *
1295  * (hi1 << 32 + lo1) * (hi2 << 32 + lo2) =
1296  *     hi1 * hi2 << 64 +
1297  *     hi1 * lo2 << 32 +
1298  *     hi2 * lo1 << 32 +
1299  *     lo1 * lo2
1300  */
1301 static void
1302 dtrace_multiply_128(uint64_t factor1, uint64_t factor2, uint64_t *product)
1303 {
1304 	uint64_t hi1, hi2, lo1, lo2;
1305 	uint64_t tmp[2];
1306 
1307 	hi1 = factor1 >> 32;
1308 	hi2 = factor2 >> 32;
1309 
1310 	lo1 = factor1 & DT_MASK_LO;
1311 	lo2 = factor2 & DT_MASK_LO;
1312 
1313 	product[0] = lo1 * lo2;
1314 	product[1] = hi1 * hi2;
1315 
1316 	tmp[0] = hi1 * lo2;
1317 	tmp[1] = 0;
1318 	dtrace_shift_128(tmp, 32);
1319 	dtrace_add_128(product, tmp, product);
1320 
1321 	tmp[0] = hi2 * lo1;
1322 	tmp[1] = 0;
1323 	dtrace_shift_128(tmp, 32);
1324 	dtrace_add_128(product, tmp, product);
1325 }
1326 
1327 /*
1328  * This privilege check should be used by actions and subroutines to
1329  * verify that the user credentials of the process that enabled the
1330  * invoking ECB match the target credentials
1331  */
1332 static int
1333 dtrace_priv_proc_common_user(dtrace_state_t *state)
1334 {
1335 	cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1336 
1337 	/*
1338 	 * We should always have a non-NULL state cred here, since if cred
1339 	 * is null (anonymous tracing), we fast-path bypass this routine.
1340 	 */
1341 	ASSERT(s_cr != NULL);
1342 
1343 	if ((cr = CRED()) != NULL &&
1344 	    s_cr->cr_uid == cr->cr_uid &&
1345 	    s_cr->cr_uid == cr->cr_ruid &&
1346 	    s_cr->cr_uid == cr->cr_suid &&
1347 	    s_cr->cr_gid == cr->cr_gid &&
1348 	    s_cr->cr_gid == cr->cr_rgid &&
1349 	    s_cr->cr_gid == cr->cr_sgid)
1350 		return (1);
1351 
1352 	return (0);
1353 }
1354 
1355 /*
1356  * This privilege check should be used by actions and subroutines to
1357  * verify that the zone of the process that enabled the invoking ECB
1358  * matches the target credentials
1359  */
1360 static int
1361 dtrace_priv_proc_common_zone(dtrace_state_t *state)
1362 {
1363 #ifdef illumos
1364 	cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1365 
1366 	/*
1367 	 * We should always have a non-NULL state cred here, since if cred
1368 	 * is null (anonymous tracing), we fast-path bypass this routine.
1369 	 */
1370 	ASSERT(s_cr != NULL);
1371 
1372 	if ((cr = CRED()) != NULL && s_cr->cr_zone == cr->cr_zone)
1373 		return (1);
1374 
1375 	return (0);
1376 #else
1377 	return (1);
1378 #endif
1379 }
1380 
1381 /*
1382  * This privilege check should be used by actions and subroutines to
1383  * verify that the process has not setuid or changed credentials.
1384  */
1385 static int
1386 dtrace_priv_proc_common_nocd(void)
1387 {
1388 	proc_t *proc;
1389 
1390 	if ((proc = ttoproc(curthread)) != NULL &&
1391 	    !(proc->p_flag & SNOCD))
1392 		return (1);
1393 
1394 	return (0);
1395 }
1396 
1397 static int
1398 dtrace_priv_proc_destructive(dtrace_state_t *state)
1399 {
1400 	int action = state->dts_cred.dcr_action;
1401 
1402 	if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE) == 0) &&
1403 	    dtrace_priv_proc_common_zone(state) == 0)
1404 		goto bad;
1405 
1406 	if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER) == 0) &&
1407 	    dtrace_priv_proc_common_user(state) == 0)
1408 		goto bad;
1409 
1410 	if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG) == 0) &&
1411 	    dtrace_priv_proc_common_nocd() == 0)
1412 		goto bad;
1413 
1414 	return (1);
1415 
1416 bad:
1417 	cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1418 
1419 	return (0);
1420 }
1421 
1422 static int
1423 dtrace_priv_proc_control(dtrace_state_t *state)
1424 {
1425 	if (state->dts_cred.dcr_action & DTRACE_CRA_PROC_CONTROL)
1426 		return (1);
1427 
1428 	if (dtrace_priv_proc_common_zone(state) &&
1429 	    dtrace_priv_proc_common_user(state) &&
1430 	    dtrace_priv_proc_common_nocd())
1431 		return (1);
1432 
1433 	cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1434 
1435 	return (0);
1436 }
1437 
1438 static int
1439 dtrace_priv_proc(dtrace_state_t *state)
1440 {
1441 	if (state->dts_cred.dcr_action & DTRACE_CRA_PROC)
1442 		return (1);
1443 
1444 	cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1445 
1446 	return (0);
1447 }
1448 
1449 static int
1450 dtrace_priv_kernel(dtrace_state_t *state)
1451 {
1452 	if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL)
1453 		return (1);
1454 
1455 	cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1456 
1457 	return (0);
1458 }
1459 
1460 static int
1461 dtrace_priv_kernel_destructive(dtrace_state_t *state)
1462 {
1463 	if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL_DESTRUCTIVE)
1464 		return (1);
1465 
1466 	cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1467 
1468 	return (0);
1469 }
1470 
1471 /*
1472  * Determine if the dte_cond of the specified ECB allows for processing of
1473  * the current probe to continue.  Note that this routine may allow continued
1474  * processing, but with access(es) stripped from the mstate's dtms_access
1475  * field.
1476  */
1477 static int
1478 dtrace_priv_probe(dtrace_state_t *state, dtrace_mstate_t *mstate,
1479     dtrace_ecb_t *ecb)
1480 {
1481 	dtrace_probe_t *probe = ecb->dte_probe;
1482 	dtrace_provider_t *prov = probe->dtpr_provider;
1483 	dtrace_pops_t *pops = &prov->dtpv_pops;
1484 	int mode = DTRACE_MODE_NOPRIV_DROP;
1485 
1486 	ASSERT(ecb->dte_cond);
1487 
1488 #ifdef illumos
1489 	if (pops->dtps_mode != NULL) {
1490 		mode = pops->dtps_mode(prov->dtpv_arg,
1491 		    probe->dtpr_id, probe->dtpr_arg);
1492 
1493 		ASSERT((mode & DTRACE_MODE_USER) ||
1494 		    (mode & DTRACE_MODE_KERNEL));
1495 		ASSERT((mode & DTRACE_MODE_NOPRIV_RESTRICT) ||
1496 		    (mode & DTRACE_MODE_NOPRIV_DROP));
1497 	}
1498 
1499 	/*
1500 	 * If the dte_cond bits indicate that this consumer is only allowed to
1501 	 * see user-mode firings of this probe, call the provider's dtps_mode()
1502 	 * entry point to check that the probe was fired while in a user
1503 	 * context.  If that's not the case, use the policy specified by the
1504 	 * provider to determine if we drop the probe or merely restrict
1505 	 * operation.
1506 	 */
1507 	if (ecb->dte_cond & DTRACE_COND_USERMODE) {
1508 		ASSERT(mode != DTRACE_MODE_NOPRIV_DROP);
1509 
1510 		if (!(mode & DTRACE_MODE_USER)) {
1511 			if (mode & DTRACE_MODE_NOPRIV_DROP)
1512 				return (0);
1513 
1514 			mstate->dtms_access &= ~DTRACE_ACCESS_ARGS;
1515 		}
1516 	}
1517 #endif
1518 
1519 	/*
1520 	 * This is more subtle than it looks. We have to be absolutely certain
1521 	 * that CRED() isn't going to change out from under us so it's only
1522 	 * legit to examine that structure if we're in constrained situations.
1523 	 * Currently, the only times we'll this check is if a non-super-user
1524 	 * has enabled the profile or syscall providers -- providers that
1525 	 * allow visibility of all processes. For the profile case, the check
1526 	 * above will ensure that we're examining a user context.
1527 	 */
1528 	if (ecb->dte_cond & DTRACE_COND_OWNER) {
1529 		cred_t *cr;
1530 		cred_t *s_cr = state->dts_cred.dcr_cred;
1531 		proc_t *proc;
1532 
1533 		ASSERT(s_cr != NULL);
1534 
1535 		if ((cr = CRED()) == NULL ||
1536 		    s_cr->cr_uid != cr->cr_uid ||
1537 		    s_cr->cr_uid != cr->cr_ruid ||
1538 		    s_cr->cr_uid != cr->cr_suid ||
1539 		    s_cr->cr_gid != cr->cr_gid ||
1540 		    s_cr->cr_gid != cr->cr_rgid ||
1541 		    s_cr->cr_gid != cr->cr_sgid ||
1542 		    (proc = ttoproc(curthread)) == NULL ||
1543 		    (proc->p_flag & SNOCD)) {
1544 			if (mode & DTRACE_MODE_NOPRIV_DROP)
1545 				return (0);
1546 
1547 #ifdef illumos
1548 			mstate->dtms_access &= ~DTRACE_ACCESS_PROC;
1549 #endif
1550 		}
1551 	}
1552 
1553 #ifdef illumos
1554 	/*
1555 	 * If our dte_cond is set to DTRACE_COND_ZONEOWNER and we are not
1556 	 * in our zone, check to see if our mode policy is to restrict rather
1557 	 * than to drop; if to restrict, strip away both DTRACE_ACCESS_PROC
1558 	 * and DTRACE_ACCESS_ARGS
1559 	 */
1560 	if (ecb->dte_cond & DTRACE_COND_ZONEOWNER) {
1561 		cred_t *cr;
1562 		cred_t *s_cr = state->dts_cred.dcr_cred;
1563 
1564 		ASSERT(s_cr != NULL);
1565 
1566 		if ((cr = CRED()) == NULL ||
1567 		    s_cr->cr_zone->zone_id != cr->cr_zone->zone_id) {
1568 			if (mode & DTRACE_MODE_NOPRIV_DROP)
1569 				return (0);
1570 
1571 			mstate->dtms_access &=
1572 			    ~(DTRACE_ACCESS_PROC | DTRACE_ACCESS_ARGS);
1573 		}
1574 	}
1575 #endif
1576 
1577 	return (1);
1578 }
1579 
1580 /*
1581  * Note:  not called from probe context.  This function is called
1582  * asynchronously (and at a regular interval) from outside of probe context to
1583  * clean the dirty dynamic variable lists on all CPUs.  Dynamic variable
1584  * cleaning is explained in detail in <sys/dtrace_impl.h>.
1585  */
1586 void
1587 dtrace_dynvar_clean(dtrace_dstate_t *dstate)
1588 {
1589 	dtrace_dynvar_t *dirty;
1590 	dtrace_dstate_percpu_t *dcpu;
1591 	dtrace_dynvar_t **rinsep;
1592 	int i, j, work = 0;
1593 
1594 	for (i = 0; i < NCPU; i++) {
1595 		dcpu = &dstate->dtds_percpu[i];
1596 		rinsep = &dcpu->dtdsc_rinsing;
1597 
1598 		/*
1599 		 * If the dirty list is NULL, there is no dirty work to do.
1600 		 */
1601 		if (dcpu->dtdsc_dirty == NULL)
1602 			continue;
1603 
1604 		if (dcpu->dtdsc_rinsing != NULL) {
1605 			/*
1606 			 * If the rinsing list is non-NULL, then it is because
1607 			 * this CPU was selected to accept another CPU's
1608 			 * dirty list -- and since that time, dirty buffers
1609 			 * have accumulated.  This is a highly unlikely
1610 			 * condition, but we choose to ignore the dirty
1611 			 * buffers -- they'll be picked up a future cleanse.
1612 			 */
1613 			continue;
1614 		}
1615 
1616 		if (dcpu->dtdsc_clean != NULL) {
1617 			/*
1618 			 * If the clean list is non-NULL, then we're in a
1619 			 * situation where a CPU has done deallocations (we
1620 			 * have a non-NULL dirty list) but no allocations (we
1621 			 * also have a non-NULL clean list).  We can't simply
1622 			 * move the dirty list into the clean list on this
1623 			 * CPU, yet we also don't want to allow this condition
1624 			 * to persist, lest a short clean list prevent a
1625 			 * massive dirty list from being cleaned (which in
1626 			 * turn could lead to otherwise avoidable dynamic
1627 			 * drops).  To deal with this, we look for some CPU
1628 			 * with a NULL clean list, NULL dirty list, and NULL
1629 			 * rinsing list -- and then we borrow this CPU to
1630 			 * rinse our dirty list.
1631 			 */
1632 			for (j = 0; j < NCPU; j++) {
1633 				dtrace_dstate_percpu_t *rinser;
1634 
1635 				rinser = &dstate->dtds_percpu[j];
1636 
1637 				if (rinser->dtdsc_rinsing != NULL)
1638 					continue;
1639 
1640 				if (rinser->dtdsc_dirty != NULL)
1641 					continue;
1642 
1643 				if (rinser->dtdsc_clean != NULL)
1644 					continue;
1645 
1646 				rinsep = &rinser->dtdsc_rinsing;
1647 				break;
1648 			}
1649 
1650 			if (j == NCPU) {
1651 				/*
1652 				 * We were unable to find another CPU that
1653 				 * could accept this dirty list -- we are
1654 				 * therefore unable to clean it now.
1655 				 */
1656 				dtrace_dynvar_failclean++;
1657 				continue;
1658 			}
1659 		}
1660 
1661 		work = 1;
1662 
1663 		/*
1664 		 * Atomically move the dirty list aside.
1665 		 */
1666 		do {
1667 			dirty = dcpu->dtdsc_dirty;
1668 
1669 			/*
1670 			 * Before we zap the dirty list, set the rinsing list.
1671 			 * (This allows for a potential assertion in
1672 			 * dtrace_dynvar():  if a free dynamic variable appears
1673 			 * on a hash chain, either the dirty list or the
1674 			 * rinsing list for some CPU must be non-NULL.)
1675 			 */
1676 			*rinsep = dirty;
1677 			dtrace_membar_producer();
1678 		} while (dtrace_casptr(&dcpu->dtdsc_dirty,
1679 		    dirty, NULL) != dirty);
1680 	}
1681 
1682 	if (!work) {
1683 		/*
1684 		 * We have no work to do; we can simply return.
1685 		 */
1686 		return;
1687 	}
1688 
1689 	dtrace_sync();
1690 
1691 	for (i = 0; i < NCPU; i++) {
1692 		dcpu = &dstate->dtds_percpu[i];
1693 
1694 		if (dcpu->dtdsc_rinsing == NULL)
1695 			continue;
1696 
1697 		/*
1698 		 * We are now guaranteed that no hash chain contains a pointer
1699 		 * into this dirty list; we can make it clean.
1700 		 */
1701 		ASSERT(dcpu->dtdsc_clean == NULL);
1702 		dcpu->dtdsc_clean = dcpu->dtdsc_rinsing;
1703 		dcpu->dtdsc_rinsing = NULL;
1704 	}
1705 
1706 	/*
1707 	 * Before we actually set the state to be DTRACE_DSTATE_CLEAN, make
1708 	 * sure that all CPUs have seen all of the dtdsc_clean pointers.
1709 	 * This prevents a race whereby a CPU incorrectly decides that
1710 	 * the state should be something other than DTRACE_DSTATE_CLEAN
1711 	 * after dtrace_dynvar_clean() has completed.
1712 	 */
1713 	dtrace_sync();
1714 
1715 	dstate->dtds_state = DTRACE_DSTATE_CLEAN;
1716 }
1717 
1718 /*
1719  * Depending on the value of the op parameter, this function looks-up,
1720  * allocates or deallocates an arbitrarily-keyed dynamic variable.  If an
1721  * allocation is requested, this function will return a pointer to a
1722  * dtrace_dynvar_t corresponding to the allocated variable -- or NULL if no
1723  * variable can be allocated.  If NULL is returned, the appropriate counter
1724  * will be incremented.
1725  */
1726 dtrace_dynvar_t *
1727 dtrace_dynvar(dtrace_dstate_t *dstate, uint_t nkeys,
1728     dtrace_key_t *key, size_t dsize, dtrace_dynvar_op_t op,
1729     dtrace_mstate_t *mstate, dtrace_vstate_t *vstate)
1730 {
1731 	uint64_t hashval = DTRACE_DYNHASH_VALID;
1732 	dtrace_dynhash_t *hash = dstate->dtds_hash;
1733 	dtrace_dynvar_t *free, *new_free, *next, *dvar, *start, *prev = NULL;
1734 	processorid_t me = curcpu, cpu = me;
1735 	dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[me];
1736 	size_t bucket, ksize;
1737 	size_t chunksize = dstate->dtds_chunksize;
1738 	uintptr_t kdata, lock, nstate;
1739 	uint_t i;
1740 
1741 	ASSERT(nkeys != 0);
1742 
1743 	/*
1744 	 * Hash the key.  As with aggregations, we use Jenkins' "One-at-a-time"
1745 	 * algorithm.  For the by-value portions, we perform the algorithm in
1746 	 * 16-bit chunks (as opposed to 8-bit chunks).  This speeds things up a
1747 	 * bit, and seems to have only a minute effect on distribution.  For
1748 	 * the by-reference data, we perform "One-at-a-time" iterating (safely)
1749 	 * over each referenced byte.  It's painful to do this, but it's much
1750 	 * better than pathological hash distribution.  The efficacy of the
1751 	 * hashing algorithm (and a comparison with other algorithms) may be
1752 	 * found by running the ::dtrace_dynstat MDB dcmd.
1753 	 */
1754 	for (i = 0; i < nkeys; i++) {
1755 		if (key[i].dttk_size == 0) {
1756 			uint64_t val = key[i].dttk_value;
1757 
1758 			hashval += (val >> 48) & 0xffff;
1759 			hashval += (hashval << 10);
1760 			hashval ^= (hashval >> 6);
1761 
1762 			hashval += (val >> 32) & 0xffff;
1763 			hashval += (hashval << 10);
1764 			hashval ^= (hashval >> 6);
1765 
1766 			hashval += (val >> 16) & 0xffff;
1767 			hashval += (hashval << 10);
1768 			hashval ^= (hashval >> 6);
1769 
1770 			hashval += val & 0xffff;
1771 			hashval += (hashval << 10);
1772 			hashval ^= (hashval >> 6);
1773 		} else {
1774 			/*
1775 			 * This is incredibly painful, but it beats the hell
1776 			 * out of the alternative.
1777 			 */
1778 			uint64_t j, size = key[i].dttk_size;
1779 			uintptr_t base = (uintptr_t)key[i].dttk_value;
1780 
1781 			if (!dtrace_canload(base, size, mstate, vstate))
1782 				break;
1783 
1784 			for (j = 0; j < size; j++) {
1785 				hashval += dtrace_load8(base + j);
1786 				hashval += (hashval << 10);
1787 				hashval ^= (hashval >> 6);
1788 			}
1789 		}
1790 	}
1791 
1792 	if (DTRACE_CPUFLAG_ISSET(CPU_DTRACE_FAULT))
1793 		return (NULL);
1794 
1795 	hashval += (hashval << 3);
1796 	hashval ^= (hashval >> 11);
1797 	hashval += (hashval << 15);
1798 
1799 	/*
1800 	 * There is a remote chance (ideally, 1 in 2^31) that our hashval
1801 	 * comes out to be one of our two sentinel hash values.  If this
1802 	 * actually happens, we set the hashval to be a value known to be a
1803 	 * non-sentinel value.
1804 	 */
1805 	if (hashval == DTRACE_DYNHASH_FREE || hashval == DTRACE_DYNHASH_SINK)
1806 		hashval = DTRACE_DYNHASH_VALID;
1807 
1808 	/*
1809 	 * Yes, it's painful to do a divide here.  If the cycle count becomes
1810 	 * important here, tricks can be pulled to reduce it.  (However, it's
1811 	 * critical that hash collisions be kept to an absolute minimum;
1812 	 * they're much more painful than a divide.)  It's better to have a
1813 	 * solution that generates few collisions and still keeps things
1814 	 * relatively simple.
1815 	 */
1816 	bucket = hashval % dstate->dtds_hashsize;
1817 
1818 	if (op == DTRACE_DYNVAR_DEALLOC) {
1819 		volatile uintptr_t *lockp = &hash[bucket].dtdh_lock;
1820 
1821 		for (;;) {
1822 			while ((lock = *lockp) & 1)
1823 				continue;
1824 
1825 			if (dtrace_casptr((volatile void *)lockp,
1826 			    (volatile void *)lock, (volatile void *)(lock + 1)) == (void *)lock)
1827 				break;
1828 		}
1829 
1830 		dtrace_membar_producer();
1831 	}
1832 
1833 top:
1834 	prev = NULL;
1835 	lock = hash[bucket].dtdh_lock;
1836 
1837 	dtrace_membar_consumer();
1838 
1839 	start = hash[bucket].dtdh_chain;
1840 	ASSERT(start != NULL && (start->dtdv_hashval == DTRACE_DYNHASH_SINK ||
1841 	    start->dtdv_hashval != DTRACE_DYNHASH_FREE ||
1842 	    op != DTRACE_DYNVAR_DEALLOC));
1843 
1844 	for (dvar = start; dvar != NULL; dvar = dvar->dtdv_next) {
1845 		dtrace_tuple_t *dtuple = &dvar->dtdv_tuple;
1846 		dtrace_key_t *dkey = &dtuple->dtt_key[0];
1847 
1848 		if (dvar->dtdv_hashval != hashval) {
1849 			if (dvar->dtdv_hashval == DTRACE_DYNHASH_SINK) {
1850 				/*
1851 				 * We've reached the sink, and therefore the
1852 				 * end of the hash chain; we can kick out of
1853 				 * the loop knowing that we have seen a valid
1854 				 * snapshot of state.
1855 				 */
1856 				ASSERT(dvar->dtdv_next == NULL);
1857 				ASSERT(dvar == &dtrace_dynhash_sink);
1858 				break;
1859 			}
1860 
1861 			if (dvar->dtdv_hashval == DTRACE_DYNHASH_FREE) {
1862 				/*
1863 				 * We've gone off the rails:  somewhere along
1864 				 * the line, one of the members of this hash
1865 				 * chain was deleted.  Note that we could also
1866 				 * detect this by simply letting this loop run
1867 				 * to completion, as we would eventually hit
1868 				 * the end of the dirty list.  However, we
1869 				 * want to avoid running the length of the
1870 				 * dirty list unnecessarily (it might be quite
1871 				 * long), so we catch this as early as
1872 				 * possible by detecting the hash marker.  In
1873 				 * this case, we simply set dvar to NULL and
1874 				 * break; the conditional after the loop will
1875 				 * send us back to top.
1876 				 */
1877 				dvar = NULL;
1878 				break;
1879 			}
1880 
1881 			goto next;
1882 		}
1883 
1884 		if (dtuple->dtt_nkeys != nkeys)
1885 			goto next;
1886 
1887 		for (i = 0; i < nkeys; i++, dkey++) {
1888 			if (dkey->dttk_size != key[i].dttk_size)
1889 				goto next; /* size or type mismatch */
1890 
1891 			if (dkey->dttk_size != 0) {
1892 				if (dtrace_bcmp(
1893 				    (void *)(uintptr_t)key[i].dttk_value,
1894 				    (void *)(uintptr_t)dkey->dttk_value,
1895 				    dkey->dttk_size))
1896 					goto next;
1897 			} else {
1898 				if (dkey->dttk_value != key[i].dttk_value)
1899 					goto next;
1900 			}
1901 		}
1902 
1903 		if (op != DTRACE_DYNVAR_DEALLOC)
1904 			return (dvar);
1905 
1906 		ASSERT(dvar->dtdv_next == NULL ||
1907 		    dvar->dtdv_next->dtdv_hashval != DTRACE_DYNHASH_FREE);
1908 
1909 		if (prev != NULL) {
1910 			ASSERT(hash[bucket].dtdh_chain != dvar);
1911 			ASSERT(start != dvar);
1912 			ASSERT(prev->dtdv_next == dvar);
1913 			prev->dtdv_next = dvar->dtdv_next;
1914 		} else {
1915 			if (dtrace_casptr(&hash[bucket].dtdh_chain,
1916 			    start, dvar->dtdv_next) != start) {
1917 				/*
1918 				 * We have failed to atomically swing the
1919 				 * hash table head pointer, presumably because
1920 				 * of a conflicting allocation on another CPU.
1921 				 * We need to reread the hash chain and try
1922 				 * again.
1923 				 */
1924 				goto top;
1925 			}
1926 		}
1927 
1928 		dtrace_membar_producer();
1929 
1930 		/*
1931 		 * Now set the hash value to indicate that it's free.
1932 		 */
1933 		ASSERT(hash[bucket].dtdh_chain != dvar);
1934 		dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
1935 
1936 		dtrace_membar_producer();
1937 
1938 		/*
1939 		 * Set the next pointer to point at the dirty list, and
1940 		 * atomically swing the dirty pointer to the newly freed dvar.
1941 		 */
1942 		do {
1943 			next = dcpu->dtdsc_dirty;
1944 			dvar->dtdv_next = next;
1945 		} while (dtrace_casptr(&dcpu->dtdsc_dirty, next, dvar) != next);
1946 
1947 		/*
1948 		 * Finally, unlock this hash bucket.
1949 		 */
1950 		ASSERT(hash[bucket].dtdh_lock == lock);
1951 		ASSERT(lock & 1);
1952 		hash[bucket].dtdh_lock++;
1953 
1954 		return (NULL);
1955 next:
1956 		prev = dvar;
1957 		continue;
1958 	}
1959 
1960 	if (dvar == NULL) {
1961 		/*
1962 		 * If dvar is NULL, it is because we went off the rails:
1963 		 * one of the elements that we traversed in the hash chain
1964 		 * was deleted while we were traversing it.  In this case,
1965 		 * we assert that we aren't doing a dealloc (deallocs lock
1966 		 * the hash bucket to prevent themselves from racing with
1967 		 * one another), and retry the hash chain traversal.
1968 		 */
1969 		ASSERT(op != DTRACE_DYNVAR_DEALLOC);
1970 		goto top;
1971 	}
1972 
1973 	if (op != DTRACE_DYNVAR_ALLOC) {
1974 		/*
1975 		 * If we are not to allocate a new variable, we want to
1976 		 * return NULL now.  Before we return, check that the value
1977 		 * of the lock word hasn't changed.  If it has, we may have
1978 		 * seen an inconsistent snapshot.
1979 		 */
1980 		if (op == DTRACE_DYNVAR_NOALLOC) {
1981 			if (hash[bucket].dtdh_lock != lock)
1982 				goto top;
1983 		} else {
1984 			ASSERT(op == DTRACE_DYNVAR_DEALLOC);
1985 			ASSERT(hash[bucket].dtdh_lock == lock);
1986 			ASSERT(lock & 1);
1987 			hash[bucket].dtdh_lock++;
1988 		}
1989 
1990 		return (NULL);
1991 	}
1992 
1993 	/*
1994 	 * We need to allocate a new dynamic variable.  The size we need is the
1995 	 * size of dtrace_dynvar plus the size of nkeys dtrace_key_t's plus the
1996 	 * size of any auxiliary key data (rounded up to 8-byte alignment) plus
1997 	 * the size of any referred-to data (dsize).  We then round the final
1998 	 * size up to the chunksize for allocation.
1999 	 */
2000 	for (ksize = 0, i = 0; i < nkeys; i++)
2001 		ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
2002 
2003 	/*
2004 	 * This should be pretty much impossible, but could happen if, say,
2005 	 * strange DIF specified the tuple.  Ideally, this should be an
2006 	 * assertion and not an error condition -- but that requires that the
2007 	 * chunksize calculation in dtrace_difo_chunksize() be absolutely
2008 	 * bullet-proof.  (That is, it must not be able to be fooled by
2009 	 * malicious DIF.)  Given the lack of backwards branches in DIF,
2010 	 * solving this would presumably not amount to solving the Halting
2011 	 * Problem -- but it still seems awfully hard.
2012 	 */
2013 	if (sizeof (dtrace_dynvar_t) + sizeof (dtrace_key_t) * (nkeys - 1) +
2014 	    ksize + dsize > chunksize) {
2015 		dcpu->dtdsc_drops++;
2016 		return (NULL);
2017 	}
2018 
2019 	nstate = DTRACE_DSTATE_EMPTY;
2020 
2021 	do {
2022 retry:
2023 		free = dcpu->dtdsc_free;
2024 
2025 		if (free == NULL) {
2026 			dtrace_dynvar_t *clean = dcpu->dtdsc_clean;
2027 			void *rval;
2028 
2029 			if (clean == NULL) {
2030 				/*
2031 				 * We're out of dynamic variable space on
2032 				 * this CPU.  Unless we have tried all CPUs,
2033 				 * we'll try to allocate from a different
2034 				 * CPU.
2035 				 */
2036 				switch (dstate->dtds_state) {
2037 				case DTRACE_DSTATE_CLEAN: {
2038 					void *sp = &dstate->dtds_state;
2039 
2040 					if (++cpu >= NCPU)
2041 						cpu = 0;
2042 
2043 					if (dcpu->dtdsc_dirty != NULL &&
2044 					    nstate == DTRACE_DSTATE_EMPTY)
2045 						nstate = DTRACE_DSTATE_DIRTY;
2046 
2047 					if (dcpu->dtdsc_rinsing != NULL)
2048 						nstate = DTRACE_DSTATE_RINSING;
2049 
2050 					dcpu = &dstate->dtds_percpu[cpu];
2051 
2052 					if (cpu != me)
2053 						goto retry;
2054 
2055 					(void) dtrace_cas32(sp,
2056 					    DTRACE_DSTATE_CLEAN, nstate);
2057 
2058 					/*
2059 					 * To increment the correct bean
2060 					 * counter, take another lap.
2061 					 */
2062 					goto retry;
2063 				}
2064 
2065 				case DTRACE_DSTATE_DIRTY:
2066 					dcpu->dtdsc_dirty_drops++;
2067 					break;
2068 
2069 				case DTRACE_DSTATE_RINSING:
2070 					dcpu->dtdsc_rinsing_drops++;
2071 					break;
2072 
2073 				case DTRACE_DSTATE_EMPTY:
2074 					dcpu->dtdsc_drops++;
2075 					break;
2076 				}
2077 
2078 				DTRACE_CPUFLAG_SET(CPU_DTRACE_DROP);
2079 				return (NULL);
2080 			}
2081 
2082 			/*
2083 			 * The clean list appears to be non-empty.  We want to
2084 			 * move the clean list to the free list; we start by
2085 			 * moving the clean pointer aside.
2086 			 */
2087 			if (dtrace_casptr(&dcpu->dtdsc_clean,
2088 			    clean, NULL) != clean) {
2089 				/*
2090 				 * We are in one of two situations:
2091 				 *
2092 				 *  (a)	The clean list was switched to the
2093 				 *	free list by another CPU.
2094 				 *
2095 				 *  (b)	The clean list was added to by the
2096 				 *	cleansing cyclic.
2097 				 *
2098 				 * In either of these situations, we can
2099 				 * just reattempt the free list allocation.
2100 				 */
2101 				goto retry;
2102 			}
2103 
2104 			ASSERT(clean->dtdv_hashval == DTRACE_DYNHASH_FREE);
2105 
2106 			/*
2107 			 * Now we'll move the clean list to our free list.
2108 			 * It's impossible for this to fail:  the only way
2109 			 * the free list can be updated is through this
2110 			 * code path, and only one CPU can own the clean list.
2111 			 * Thus, it would only be possible for this to fail if
2112 			 * this code were racing with dtrace_dynvar_clean().
2113 			 * (That is, if dtrace_dynvar_clean() updated the clean
2114 			 * list, and we ended up racing to update the free
2115 			 * list.)  This race is prevented by the dtrace_sync()
2116 			 * in dtrace_dynvar_clean() -- which flushes the
2117 			 * owners of the clean lists out before resetting
2118 			 * the clean lists.
2119 			 */
2120 			dcpu = &dstate->dtds_percpu[me];
2121 			rval = dtrace_casptr(&dcpu->dtdsc_free, NULL, clean);
2122 			ASSERT(rval == NULL);
2123 			goto retry;
2124 		}
2125 
2126 		dvar = free;
2127 		new_free = dvar->dtdv_next;
2128 	} while (dtrace_casptr(&dcpu->dtdsc_free, free, new_free) != free);
2129 
2130 	/*
2131 	 * We have now allocated a new chunk.  We copy the tuple keys into the
2132 	 * tuple array and copy any referenced key data into the data space
2133 	 * following the tuple array.  As we do this, we relocate dttk_value
2134 	 * in the final tuple to point to the key data address in the chunk.
2135 	 */
2136 	kdata = (uintptr_t)&dvar->dtdv_tuple.dtt_key[nkeys];
2137 	dvar->dtdv_data = (void *)(kdata + ksize);
2138 	dvar->dtdv_tuple.dtt_nkeys = nkeys;
2139 
2140 	for (i = 0; i < nkeys; i++) {
2141 		dtrace_key_t *dkey = &dvar->dtdv_tuple.dtt_key[i];
2142 		size_t kesize = key[i].dttk_size;
2143 
2144 		if (kesize != 0) {
2145 			dtrace_bcopy(
2146 			    (const void *)(uintptr_t)key[i].dttk_value,
2147 			    (void *)kdata, kesize);
2148 			dkey->dttk_value = kdata;
2149 			kdata += P2ROUNDUP(kesize, sizeof (uint64_t));
2150 		} else {
2151 			dkey->dttk_value = key[i].dttk_value;
2152 		}
2153 
2154 		dkey->dttk_size = kesize;
2155 	}
2156 
2157 	ASSERT(dvar->dtdv_hashval == DTRACE_DYNHASH_FREE);
2158 	dvar->dtdv_hashval = hashval;
2159 	dvar->dtdv_next = start;
2160 
2161 	if (dtrace_casptr(&hash[bucket].dtdh_chain, start, dvar) == start)
2162 		return (dvar);
2163 
2164 	/*
2165 	 * The cas has failed.  Either another CPU is adding an element to
2166 	 * this hash chain, or another CPU is deleting an element from this
2167 	 * hash chain.  The simplest way to deal with both of these cases
2168 	 * (though not necessarily the most efficient) is to free our
2169 	 * allocated block and re-attempt it all.  Note that the free is
2170 	 * to the dirty list and _not_ to the free list.  This is to prevent
2171 	 * races with allocators, above.
2172 	 */
2173 	dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
2174 
2175 	dtrace_membar_producer();
2176 
2177 	do {
2178 		free = dcpu->dtdsc_dirty;
2179 		dvar->dtdv_next = free;
2180 	} while (dtrace_casptr(&dcpu->dtdsc_dirty, free, dvar) != free);
2181 
2182 	goto top;
2183 }
2184 
2185 /*ARGSUSED*/
2186 static void
2187 dtrace_aggregate_min(uint64_t *oval, uint64_t nval, uint64_t arg)
2188 {
2189 	if ((int64_t)nval < (int64_t)*oval)
2190 		*oval = nval;
2191 }
2192 
2193 /*ARGSUSED*/
2194 static void
2195 dtrace_aggregate_max(uint64_t *oval, uint64_t nval, uint64_t arg)
2196 {
2197 	if ((int64_t)nval > (int64_t)*oval)
2198 		*oval = nval;
2199 }
2200 
2201 static void
2202 dtrace_aggregate_quantize(uint64_t *quanta, uint64_t nval, uint64_t incr)
2203 {
2204 	int i, zero = DTRACE_QUANTIZE_ZEROBUCKET;
2205 	int64_t val = (int64_t)nval;
2206 
2207 	if (val < 0) {
2208 		for (i = 0; i < zero; i++) {
2209 			if (val <= DTRACE_QUANTIZE_BUCKETVAL(i)) {
2210 				quanta[i] += incr;
2211 				return;
2212 			}
2213 		}
2214 	} else {
2215 		for (i = zero + 1; i < DTRACE_QUANTIZE_NBUCKETS; i++) {
2216 			if (val < DTRACE_QUANTIZE_BUCKETVAL(i)) {
2217 				quanta[i - 1] += incr;
2218 				return;
2219 			}
2220 		}
2221 
2222 		quanta[DTRACE_QUANTIZE_NBUCKETS - 1] += incr;
2223 		return;
2224 	}
2225 
2226 	ASSERT(0);
2227 }
2228 
2229 static void
2230 dtrace_aggregate_lquantize(uint64_t *lquanta, uint64_t nval, uint64_t incr)
2231 {
2232 	uint64_t arg = *lquanta++;
2233 	int32_t base = DTRACE_LQUANTIZE_BASE(arg);
2234 	uint16_t step = DTRACE_LQUANTIZE_STEP(arg);
2235 	uint16_t levels = DTRACE_LQUANTIZE_LEVELS(arg);
2236 	int32_t val = (int32_t)nval, level;
2237 
2238 	ASSERT(step != 0);
2239 	ASSERT(levels != 0);
2240 
2241 	if (val < base) {
2242 		/*
2243 		 * This is an underflow.
2244 		 */
2245 		lquanta[0] += incr;
2246 		return;
2247 	}
2248 
2249 	level = (val - base) / step;
2250 
2251 	if (level < levels) {
2252 		lquanta[level + 1] += incr;
2253 		return;
2254 	}
2255 
2256 	/*
2257 	 * This is an overflow.
2258 	 */
2259 	lquanta[levels + 1] += incr;
2260 }
2261 
2262 static int
2263 dtrace_aggregate_llquantize_bucket(uint16_t factor, uint16_t low,
2264     uint16_t high, uint16_t nsteps, int64_t value)
2265 {
2266 	int64_t this = 1, last, next;
2267 	int base = 1, order;
2268 
2269 	ASSERT(factor <= nsteps);
2270 	ASSERT(nsteps % factor == 0);
2271 
2272 	for (order = 0; order < low; order++)
2273 		this *= factor;
2274 
2275 	/*
2276 	 * If our value is less than our factor taken to the power of the
2277 	 * low order of magnitude, it goes into the zeroth bucket.
2278 	 */
2279 	if (value < (last = this))
2280 		return (0);
2281 
2282 	for (this *= factor; order <= high; order++) {
2283 		int nbuckets = this > nsteps ? nsteps : this;
2284 
2285 		if ((next = this * factor) < this) {
2286 			/*
2287 			 * We should not generally get log/linear quantizations
2288 			 * with a high magnitude that allows 64-bits to
2289 			 * overflow, but we nonetheless protect against this
2290 			 * by explicitly checking for overflow, and clamping
2291 			 * our value accordingly.
2292 			 */
2293 			value = this - 1;
2294 		}
2295 
2296 		if (value < this) {
2297 			/*
2298 			 * If our value lies within this order of magnitude,
2299 			 * determine its position by taking the offset within
2300 			 * the order of magnitude, dividing by the bucket
2301 			 * width, and adding to our (accumulated) base.
2302 			 */
2303 			return (base + (value - last) / (this / nbuckets));
2304 		}
2305 
2306 		base += nbuckets - (nbuckets / factor);
2307 		last = this;
2308 		this = next;
2309 	}
2310 
2311 	/*
2312 	 * Our value is greater than or equal to our factor taken to the
2313 	 * power of one plus the high magnitude -- return the top bucket.
2314 	 */
2315 	return (base);
2316 }
2317 
2318 static void
2319 dtrace_aggregate_llquantize(uint64_t *llquanta, uint64_t nval, uint64_t incr)
2320 {
2321 	uint64_t arg = *llquanta++;
2322 	uint16_t factor = DTRACE_LLQUANTIZE_FACTOR(arg);
2323 	uint16_t low = DTRACE_LLQUANTIZE_LOW(arg);
2324 	uint16_t high = DTRACE_LLQUANTIZE_HIGH(arg);
2325 	uint16_t nsteps = DTRACE_LLQUANTIZE_NSTEP(arg);
2326 
2327 	llquanta[dtrace_aggregate_llquantize_bucket(factor,
2328 	    low, high, nsteps, nval)] += incr;
2329 }
2330 
2331 /*ARGSUSED*/
2332 static void
2333 dtrace_aggregate_avg(uint64_t *data, uint64_t nval, uint64_t arg)
2334 {
2335 	data[0]++;
2336 	data[1] += nval;
2337 }
2338 
2339 /*ARGSUSED*/
2340 static void
2341 dtrace_aggregate_stddev(uint64_t *data, uint64_t nval, uint64_t arg)
2342 {
2343 	int64_t snval = (int64_t)nval;
2344 	uint64_t tmp[2];
2345 
2346 	data[0]++;
2347 	data[1] += nval;
2348 
2349 	/*
2350 	 * What we want to say here is:
2351 	 *
2352 	 * data[2] += nval * nval;
2353 	 *
2354 	 * But given that nval is 64-bit, we could easily overflow, so
2355 	 * we do this as 128-bit arithmetic.
2356 	 */
2357 	if (snval < 0)
2358 		snval = -snval;
2359 
2360 	dtrace_multiply_128((uint64_t)snval, (uint64_t)snval, tmp);
2361 	dtrace_add_128(data + 2, tmp, data + 2);
2362 }
2363 
2364 /*ARGSUSED*/
2365 static void
2366 dtrace_aggregate_count(uint64_t *oval, uint64_t nval, uint64_t arg)
2367 {
2368 	*oval = *oval + 1;
2369 }
2370 
2371 /*ARGSUSED*/
2372 static void
2373 dtrace_aggregate_sum(uint64_t *oval, uint64_t nval, uint64_t arg)
2374 {
2375 	*oval += nval;
2376 }
2377 
2378 /*
2379  * Aggregate given the tuple in the principal data buffer, and the aggregating
2380  * action denoted by the specified dtrace_aggregation_t.  The aggregation
2381  * buffer is specified as the buf parameter.  This routine does not return
2382  * failure; if there is no space in the aggregation buffer, the data will be
2383  * dropped, and a corresponding counter incremented.
2384  */
2385 static void
2386 dtrace_aggregate(dtrace_aggregation_t *agg, dtrace_buffer_t *dbuf,
2387     intptr_t offset, dtrace_buffer_t *buf, uint64_t expr, uint64_t arg)
2388 {
2389 	dtrace_recdesc_t *rec = &agg->dtag_action.dta_rec;
2390 	uint32_t i, ndx, size, fsize;
2391 	uint32_t align = sizeof (uint64_t) - 1;
2392 	dtrace_aggbuffer_t *agb;
2393 	dtrace_aggkey_t *key;
2394 	uint32_t hashval = 0, limit, isstr;
2395 	caddr_t tomax, data, kdata;
2396 	dtrace_actkind_t action;
2397 	dtrace_action_t *act;
2398 	uintptr_t offs;
2399 
2400 	if (buf == NULL)
2401 		return;
2402 
2403 	if (!agg->dtag_hasarg) {
2404 		/*
2405 		 * Currently, only quantize() and lquantize() take additional
2406 		 * arguments, and they have the same semantics:  an increment
2407 		 * value that defaults to 1 when not present.  If additional
2408 		 * aggregating actions take arguments, the setting of the
2409 		 * default argument value will presumably have to become more
2410 		 * sophisticated...
2411 		 */
2412 		arg = 1;
2413 	}
2414 
2415 	action = agg->dtag_action.dta_kind - DTRACEACT_AGGREGATION;
2416 	size = rec->dtrd_offset - agg->dtag_base;
2417 	fsize = size + rec->dtrd_size;
2418 
2419 	ASSERT(dbuf->dtb_tomax != NULL);
2420 	data = dbuf->dtb_tomax + offset + agg->dtag_base;
2421 
2422 	if ((tomax = buf->dtb_tomax) == NULL) {
2423 		dtrace_buffer_drop(buf);
2424 		return;
2425 	}
2426 
2427 	/*
2428 	 * The metastructure is always at the bottom of the buffer.
2429 	 */
2430 	agb = (dtrace_aggbuffer_t *)(tomax + buf->dtb_size -
2431 	    sizeof (dtrace_aggbuffer_t));
2432 
2433 	if (buf->dtb_offset == 0) {
2434 		/*
2435 		 * We just kludge up approximately 1/8th of the size to be
2436 		 * buckets.  If this guess ends up being routinely
2437 		 * off-the-mark, we may need to dynamically readjust this
2438 		 * based on past performance.
2439 		 */
2440 		uintptr_t hashsize = (buf->dtb_size >> 3) / sizeof (uintptr_t);
2441 
2442 		if ((uintptr_t)agb - hashsize * sizeof (dtrace_aggkey_t *) <
2443 		    (uintptr_t)tomax || hashsize == 0) {
2444 			/*
2445 			 * We've been given a ludicrously small buffer;
2446 			 * increment our drop count and leave.
2447 			 */
2448 			dtrace_buffer_drop(buf);
2449 			return;
2450 		}
2451 
2452 		/*
2453 		 * And now, a pathetic attempt to try to get a an odd (or
2454 		 * perchance, a prime) hash size for better hash distribution.
2455 		 */
2456 		if (hashsize > (DTRACE_AGGHASHSIZE_SLEW << 3))
2457 			hashsize -= DTRACE_AGGHASHSIZE_SLEW;
2458 
2459 		agb->dtagb_hashsize = hashsize;
2460 		agb->dtagb_hash = (dtrace_aggkey_t **)((uintptr_t)agb -
2461 		    agb->dtagb_hashsize * sizeof (dtrace_aggkey_t *));
2462 		agb->dtagb_free = (uintptr_t)agb->dtagb_hash;
2463 
2464 		for (i = 0; i < agb->dtagb_hashsize; i++)
2465 			agb->dtagb_hash[i] = NULL;
2466 	}
2467 
2468 	ASSERT(agg->dtag_first != NULL);
2469 	ASSERT(agg->dtag_first->dta_intuple);
2470 
2471 	/*
2472 	 * Calculate the hash value based on the key.  Note that we _don't_
2473 	 * include the aggid in the hashing (but we will store it as part of
2474 	 * the key).  The hashing algorithm is Bob Jenkins' "One-at-a-time"
2475 	 * algorithm: a simple, quick algorithm that has no known funnels, and
2476 	 * gets good distribution in practice.  The efficacy of the hashing
2477 	 * algorithm (and a comparison with other algorithms) may be found by
2478 	 * running the ::dtrace_aggstat MDB dcmd.
2479 	 */
2480 	for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
2481 		i = act->dta_rec.dtrd_offset - agg->dtag_base;
2482 		limit = i + act->dta_rec.dtrd_size;
2483 		ASSERT(limit <= size);
2484 		isstr = DTRACEACT_ISSTRING(act);
2485 
2486 		for (; i < limit; i++) {
2487 			hashval += data[i];
2488 			hashval += (hashval << 10);
2489 			hashval ^= (hashval >> 6);
2490 
2491 			if (isstr && data[i] == '\0')
2492 				break;
2493 		}
2494 	}
2495 
2496 	hashval += (hashval << 3);
2497 	hashval ^= (hashval >> 11);
2498 	hashval += (hashval << 15);
2499 
2500 	/*
2501 	 * Yes, the divide here is expensive -- but it's generally the least
2502 	 * of the performance issues given the amount of data that we iterate
2503 	 * over to compute hash values, compare data, etc.
2504 	 */
2505 	ndx = hashval % agb->dtagb_hashsize;
2506 
2507 	for (key = agb->dtagb_hash[ndx]; key != NULL; key = key->dtak_next) {
2508 		ASSERT((caddr_t)key >= tomax);
2509 		ASSERT((caddr_t)key < tomax + buf->dtb_size);
2510 
2511 		if (hashval != key->dtak_hashval || key->dtak_size != size)
2512 			continue;
2513 
2514 		kdata = key->dtak_data;
2515 		ASSERT(kdata >= tomax && kdata < tomax + buf->dtb_size);
2516 
2517 		for (act = agg->dtag_first; act->dta_intuple;
2518 		    act = act->dta_next) {
2519 			i = act->dta_rec.dtrd_offset - agg->dtag_base;
2520 			limit = i + act->dta_rec.dtrd_size;
2521 			ASSERT(limit <= size);
2522 			isstr = DTRACEACT_ISSTRING(act);
2523 
2524 			for (; i < limit; i++) {
2525 				if (kdata[i] != data[i])
2526 					goto next;
2527 
2528 				if (isstr && data[i] == '\0')
2529 					break;
2530 			}
2531 		}
2532 
2533 		if (action != key->dtak_action) {
2534 			/*
2535 			 * We are aggregating on the same value in the same
2536 			 * aggregation with two different aggregating actions.
2537 			 * (This should have been picked up in the compiler,
2538 			 * so we may be dealing with errant or devious DIF.)
2539 			 * This is an error condition; we indicate as much,
2540 			 * and return.
2541 			 */
2542 			DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
2543 			return;
2544 		}
2545 
2546 		/*
2547 		 * This is a hit:  we need to apply the aggregator to
2548 		 * the value at this key.
2549 		 */
2550 		agg->dtag_aggregate((uint64_t *)(kdata + size), expr, arg);
2551 		return;
2552 next:
2553 		continue;
2554 	}
2555 
2556 	/*
2557 	 * We didn't find it.  We need to allocate some zero-filled space,
2558 	 * link it into the hash table appropriately, and apply the aggregator
2559 	 * to the (zero-filled) value.
2560 	 */
2561 	offs = buf->dtb_offset;
2562 	while (offs & (align - 1))
2563 		offs += sizeof (uint32_t);
2564 
2565 	/*
2566 	 * If we don't have enough room to both allocate a new key _and_
2567 	 * its associated data, increment the drop count and return.
2568 	 */
2569 	if ((uintptr_t)tomax + offs + fsize >
2570 	    agb->dtagb_free - sizeof (dtrace_aggkey_t)) {
2571 		dtrace_buffer_drop(buf);
2572 		return;
2573 	}
2574 
2575 	/*CONSTCOND*/
2576 	ASSERT(!(sizeof (dtrace_aggkey_t) & (sizeof (uintptr_t) - 1)));
2577 	key = (dtrace_aggkey_t *)(agb->dtagb_free - sizeof (dtrace_aggkey_t));
2578 	agb->dtagb_free -= sizeof (dtrace_aggkey_t);
2579 
2580 	key->dtak_data = kdata = tomax + offs;
2581 	buf->dtb_offset = offs + fsize;
2582 
2583 	/*
2584 	 * Now copy the data across.
2585 	 */
2586 	*((dtrace_aggid_t *)kdata) = agg->dtag_id;
2587 
2588 	for (i = sizeof (dtrace_aggid_t); i < size; i++)
2589 		kdata[i] = data[i];
2590 
2591 	/*
2592 	 * Because strings are not zeroed out by default, we need to iterate
2593 	 * looking for actions that store strings, and we need to explicitly
2594 	 * pad these strings out with zeroes.
2595 	 */
2596 	for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
2597 		int nul;
2598 
2599 		if (!DTRACEACT_ISSTRING(act))
2600 			continue;
2601 
2602 		i = act->dta_rec.dtrd_offset - agg->dtag_base;
2603 		limit = i + act->dta_rec.dtrd_size;
2604 		ASSERT(limit <= size);
2605 
2606 		for (nul = 0; i < limit; i++) {
2607 			if (nul) {
2608 				kdata[i] = '\0';
2609 				continue;
2610 			}
2611 
2612 			if (data[i] != '\0')
2613 				continue;
2614 
2615 			nul = 1;
2616 		}
2617 	}
2618 
2619 	for (i = size; i < fsize; i++)
2620 		kdata[i] = 0;
2621 
2622 	key->dtak_hashval = hashval;
2623 	key->dtak_size = size;
2624 	key->dtak_action = action;
2625 	key->dtak_next = agb->dtagb_hash[ndx];
2626 	agb->dtagb_hash[ndx] = key;
2627 
2628 	/*
2629 	 * Finally, apply the aggregator.
2630 	 */
2631 	*((uint64_t *)(key->dtak_data + size)) = agg->dtag_initial;
2632 	agg->dtag_aggregate((uint64_t *)(key->dtak_data + size), expr, arg);
2633 }
2634 
2635 /*
2636  * Given consumer state, this routine finds a speculation in the INACTIVE
2637  * state and transitions it into the ACTIVE state.  If there is no speculation
2638  * in the INACTIVE state, 0 is returned.  In this case, no error counter is
2639  * incremented -- it is up to the caller to take appropriate action.
2640  */
2641 static int
2642 dtrace_speculation(dtrace_state_t *state)
2643 {
2644 	int i = 0;
2645 	dtrace_speculation_state_t current;
2646 	uint32_t *stat = &state->dts_speculations_unavail, count;
2647 
2648 	while (i < state->dts_nspeculations) {
2649 		dtrace_speculation_t *spec = &state->dts_speculations[i];
2650 
2651 		current = spec->dtsp_state;
2652 
2653 		if (current != DTRACESPEC_INACTIVE) {
2654 			if (current == DTRACESPEC_COMMITTINGMANY ||
2655 			    current == DTRACESPEC_COMMITTING ||
2656 			    current == DTRACESPEC_DISCARDING)
2657 				stat = &state->dts_speculations_busy;
2658 			i++;
2659 			continue;
2660 		}
2661 
2662 		if (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2663 		    current, DTRACESPEC_ACTIVE) == current)
2664 			return (i + 1);
2665 	}
2666 
2667 	/*
2668 	 * We couldn't find a speculation.  If we found as much as a single
2669 	 * busy speculation buffer, we'll attribute this failure as "busy"
2670 	 * instead of "unavail".
2671 	 */
2672 	do {
2673 		count = *stat;
2674 	} while (dtrace_cas32(stat, count, count + 1) != count);
2675 
2676 	return (0);
2677 }
2678 
2679 /*
2680  * This routine commits an active speculation.  If the specified speculation
2681  * is not in a valid state to perform a commit(), this routine will silently do
2682  * nothing.  The state of the specified speculation is transitioned according
2683  * to the state transition diagram outlined in <sys/dtrace_impl.h>
2684  */
2685 static void
2686 dtrace_speculation_commit(dtrace_state_t *state, processorid_t cpu,
2687     dtrace_specid_t which)
2688 {
2689 	dtrace_speculation_t *spec;
2690 	dtrace_buffer_t *src, *dest;
2691 	uintptr_t daddr, saddr, dlimit, slimit;
2692 	dtrace_speculation_state_t current, new = 0;
2693 	intptr_t offs;
2694 	uint64_t timestamp;
2695 
2696 	if (which == 0)
2697 		return;
2698 
2699 	if (which > state->dts_nspeculations) {
2700 		cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2701 		return;
2702 	}
2703 
2704 	spec = &state->dts_speculations[which - 1];
2705 	src = &spec->dtsp_buffer[cpu];
2706 	dest = &state->dts_buffer[cpu];
2707 
2708 	do {
2709 		current = spec->dtsp_state;
2710 
2711 		if (current == DTRACESPEC_COMMITTINGMANY)
2712 			break;
2713 
2714 		switch (current) {
2715 		case DTRACESPEC_INACTIVE:
2716 		case DTRACESPEC_DISCARDING:
2717 			return;
2718 
2719 		case DTRACESPEC_COMMITTING:
2720 			/*
2721 			 * This is only possible if we are (a) commit()'ing
2722 			 * without having done a prior speculate() on this CPU
2723 			 * and (b) racing with another commit() on a different
2724 			 * CPU.  There's nothing to do -- we just assert that
2725 			 * our offset is 0.
2726 			 */
2727 			ASSERT(src->dtb_offset == 0);
2728 			return;
2729 
2730 		case DTRACESPEC_ACTIVE:
2731 			new = DTRACESPEC_COMMITTING;
2732 			break;
2733 
2734 		case DTRACESPEC_ACTIVEONE:
2735 			/*
2736 			 * This speculation is active on one CPU.  If our
2737 			 * buffer offset is non-zero, we know that the one CPU
2738 			 * must be us.  Otherwise, we are committing on a
2739 			 * different CPU from the speculate(), and we must
2740 			 * rely on being asynchronously cleaned.
2741 			 */
2742 			if (src->dtb_offset != 0) {
2743 				new = DTRACESPEC_COMMITTING;
2744 				break;
2745 			}
2746 			/*FALLTHROUGH*/
2747 
2748 		case DTRACESPEC_ACTIVEMANY:
2749 			new = DTRACESPEC_COMMITTINGMANY;
2750 			break;
2751 
2752 		default:
2753 			ASSERT(0);
2754 		}
2755 	} while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2756 	    current, new) != current);
2757 
2758 	/*
2759 	 * We have set the state to indicate that we are committing this
2760 	 * speculation.  Now reserve the necessary space in the destination
2761 	 * buffer.
2762 	 */
2763 	if ((offs = dtrace_buffer_reserve(dest, src->dtb_offset,
2764 	    sizeof (uint64_t), state, NULL)) < 0) {
2765 		dtrace_buffer_drop(dest);
2766 		goto out;
2767 	}
2768 
2769 	/*
2770 	 * We have sufficient space to copy the speculative buffer into the
2771 	 * primary buffer.  First, modify the speculative buffer, filling
2772 	 * in the timestamp of all entries with the current time.  The data
2773 	 * must have the commit() time rather than the time it was traced,
2774 	 * so that all entries in the primary buffer are in timestamp order.
2775 	 */
2776 	timestamp = dtrace_gethrtime();
2777 	saddr = (uintptr_t)src->dtb_tomax;
2778 	slimit = saddr + src->dtb_offset;
2779 	while (saddr < slimit) {
2780 		size_t size;
2781 		dtrace_rechdr_t *dtrh = (dtrace_rechdr_t *)saddr;
2782 
2783 		if (dtrh->dtrh_epid == DTRACE_EPIDNONE) {
2784 			saddr += sizeof (dtrace_epid_t);
2785 			continue;
2786 		}
2787 		ASSERT3U(dtrh->dtrh_epid, <=, state->dts_necbs);
2788 		size = state->dts_ecbs[dtrh->dtrh_epid - 1]->dte_size;
2789 
2790 		ASSERT3U(saddr + size, <=, slimit);
2791 		ASSERT3U(size, >=, sizeof (dtrace_rechdr_t));
2792 		ASSERT3U(DTRACE_RECORD_LOAD_TIMESTAMP(dtrh), ==, UINT64_MAX);
2793 
2794 		DTRACE_RECORD_STORE_TIMESTAMP(dtrh, timestamp);
2795 
2796 		saddr += size;
2797 	}
2798 
2799 	/*
2800 	 * Copy the buffer across.  (Note that this is a
2801 	 * highly subobtimal bcopy(); in the unlikely event that this becomes
2802 	 * a serious performance issue, a high-performance DTrace-specific
2803 	 * bcopy() should obviously be invented.)
2804 	 */
2805 	daddr = (uintptr_t)dest->dtb_tomax + offs;
2806 	dlimit = daddr + src->dtb_offset;
2807 	saddr = (uintptr_t)src->dtb_tomax;
2808 
2809 	/*
2810 	 * First, the aligned portion.
2811 	 */
2812 	while (dlimit - daddr >= sizeof (uint64_t)) {
2813 		*((uint64_t *)daddr) = *((uint64_t *)saddr);
2814 
2815 		daddr += sizeof (uint64_t);
2816 		saddr += sizeof (uint64_t);
2817 	}
2818 
2819 	/*
2820 	 * Now any left-over bit...
2821 	 */
2822 	while (dlimit - daddr)
2823 		*((uint8_t *)daddr++) = *((uint8_t *)saddr++);
2824 
2825 	/*
2826 	 * Finally, commit the reserved space in the destination buffer.
2827 	 */
2828 	dest->dtb_offset = offs + src->dtb_offset;
2829 
2830 out:
2831 	/*
2832 	 * If we're lucky enough to be the only active CPU on this speculation
2833 	 * buffer, we can just set the state back to DTRACESPEC_INACTIVE.
2834 	 */
2835 	if (current == DTRACESPEC_ACTIVE ||
2836 	    (current == DTRACESPEC_ACTIVEONE && new == DTRACESPEC_COMMITTING)) {
2837 		uint32_t rval = dtrace_cas32((uint32_t *)&spec->dtsp_state,
2838 		    DTRACESPEC_COMMITTING, DTRACESPEC_INACTIVE);
2839 
2840 		ASSERT(rval == DTRACESPEC_COMMITTING);
2841 	}
2842 
2843 	src->dtb_offset = 0;
2844 	src->dtb_xamot_drops += src->dtb_drops;
2845 	src->dtb_drops = 0;
2846 }
2847 
2848 /*
2849  * This routine discards an active speculation.  If the specified speculation
2850  * is not in a valid state to perform a discard(), this routine will silently
2851  * do nothing.  The state of the specified speculation is transitioned
2852  * according to the state transition diagram outlined in <sys/dtrace_impl.h>
2853  */
2854 static void
2855 dtrace_speculation_discard(dtrace_state_t *state, processorid_t cpu,
2856     dtrace_specid_t which)
2857 {
2858 	dtrace_speculation_t *spec;
2859 	dtrace_speculation_state_t current, new = 0;
2860 	dtrace_buffer_t *buf;
2861 
2862 	if (which == 0)
2863 		return;
2864 
2865 	if (which > state->dts_nspeculations) {
2866 		cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2867 		return;
2868 	}
2869 
2870 	spec = &state->dts_speculations[which - 1];
2871 	buf = &spec->dtsp_buffer[cpu];
2872 
2873 	do {
2874 		current = spec->dtsp_state;
2875 
2876 		switch (current) {
2877 		case DTRACESPEC_INACTIVE:
2878 		case DTRACESPEC_COMMITTINGMANY:
2879 		case DTRACESPEC_COMMITTING:
2880 		case DTRACESPEC_DISCARDING:
2881 			return;
2882 
2883 		case DTRACESPEC_ACTIVE:
2884 		case DTRACESPEC_ACTIVEMANY:
2885 			new = DTRACESPEC_DISCARDING;
2886 			break;
2887 
2888 		case DTRACESPEC_ACTIVEONE:
2889 			if (buf->dtb_offset != 0) {
2890 				new = DTRACESPEC_INACTIVE;
2891 			} else {
2892 				new = DTRACESPEC_DISCARDING;
2893 			}
2894 			break;
2895 
2896 		default:
2897 			ASSERT(0);
2898 		}
2899 	} while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2900 	    current, new) != current);
2901 
2902 	buf->dtb_offset = 0;
2903 	buf->dtb_drops = 0;
2904 }
2905 
2906 /*
2907  * Note:  not called from probe context.  This function is called
2908  * asynchronously from cross call context to clean any speculations that are
2909  * in the COMMITTINGMANY or DISCARDING states.  These speculations may not be
2910  * transitioned back to the INACTIVE state until all CPUs have cleaned the
2911  * speculation.
2912  */
2913 static void
2914 dtrace_speculation_clean_here(dtrace_state_t *state)
2915 {
2916 	dtrace_icookie_t cookie;
2917 	processorid_t cpu = curcpu;
2918 	dtrace_buffer_t *dest = &state->dts_buffer[cpu];
2919 	dtrace_specid_t i;
2920 
2921 	cookie = dtrace_interrupt_disable();
2922 
2923 	if (dest->dtb_tomax == NULL) {
2924 		dtrace_interrupt_enable(cookie);
2925 		return;
2926 	}
2927 
2928 	for (i = 0; i < state->dts_nspeculations; i++) {
2929 		dtrace_speculation_t *spec = &state->dts_speculations[i];
2930 		dtrace_buffer_t *src = &spec->dtsp_buffer[cpu];
2931 
2932 		if (src->dtb_tomax == NULL)
2933 			continue;
2934 
2935 		if (spec->dtsp_state == DTRACESPEC_DISCARDING) {
2936 			src->dtb_offset = 0;
2937 			continue;
2938 		}
2939 
2940 		if (spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
2941 			continue;
2942 
2943 		if (src->dtb_offset == 0)
2944 			continue;
2945 
2946 		dtrace_speculation_commit(state, cpu, i + 1);
2947 	}
2948 
2949 	dtrace_interrupt_enable(cookie);
2950 }
2951 
2952 /*
2953  * Note:  not called from probe context.  This function is called
2954  * asynchronously (and at a regular interval) to clean any speculations that
2955  * are in the COMMITTINGMANY or DISCARDING states.  If it discovers that there
2956  * is work to be done, it cross calls all CPUs to perform that work;
2957  * COMMITMANY and DISCARDING speculations may not be transitioned back to the
2958  * INACTIVE state until they have been cleaned by all CPUs.
2959  */
2960 static void
2961 dtrace_speculation_clean(dtrace_state_t *state)
2962 {
2963 	int work = 0, rv;
2964 	dtrace_specid_t i;
2965 
2966 	for (i = 0; i < state->dts_nspeculations; i++) {
2967 		dtrace_speculation_t *spec = &state->dts_speculations[i];
2968 
2969 		ASSERT(!spec->dtsp_cleaning);
2970 
2971 		if (spec->dtsp_state != DTRACESPEC_DISCARDING &&
2972 		    spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
2973 			continue;
2974 
2975 		work++;
2976 		spec->dtsp_cleaning = 1;
2977 	}
2978 
2979 	if (!work)
2980 		return;
2981 
2982 	dtrace_xcall(DTRACE_CPUALL,
2983 	    (dtrace_xcall_t)dtrace_speculation_clean_here, state);
2984 
2985 	/*
2986 	 * We now know that all CPUs have committed or discarded their
2987 	 * speculation buffers, as appropriate.  We can now set the state
2988 	 * to inactive.
2989 	 */
2990 	for (i = 0; i < state->dts_nspeculations; i++) {
2991 		dtrace_speculation_t *spec = &state->dts_speculations[i];
2992 		dtrace_speculation_state_t current, new;
2993 
2994 		if (!spec->dtsp_cleaning)
2995 			continue;
2996 
2997 		current = spec->dtsp_state;
2998 		ASSERT(current == DTRACESPEC_DISCARDING ||
2999 		    current == DTRACESPEC_COMMITTINGMANY);
3000 
3001 		new = DTRACESPEC_INACTIVE;
3002 
3003 		rv = dtrace_cas32((uint32_t *)&spec->dtsp_state, current, new);
3004 		ASSERT(rv == current);
3005 		spec->dtsp_cleaning = 0;
3006 	}
3007 }
3008 
3009 /*
3010  * Called as part of a speculate() to get the speculative buffer associated
3011  * with a given speculation.  Returns NULL if the specified speculation is not
3012  * in an ACTIVE state.  If the speculation is in the ACTIVEONE state -- and
3013  * the active CPU is not the specified CPU -- the speculation will be
3014  * atomically transitioned into the ACTIVEMANY state.
3015  */
3016 static dtrace_buffer_t *
3017 dtrace_speculation_buffer(dtrace_state_t *state, processorid_t cpuid,
3018     dtrace_specid_t which)
3019 {
3020 	dtrace_speculation_t *spec;
3021 	dtrace_speculation_state_t current, new = 0;
3022 	dtrace_buffer_t *buf;
3023 
3024 	if (which == 0)
3025 		return (NULL);
3026 
3027 	if (which > state->dts_nspeculations) {
3028 		cpu_core[cpuid].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
3029 		return (NULL);
3030 	}
3031 
3032 	spec = &state->dts_speculations[which - 1];
3033 	buf = &spec->dtsp_buffer[cpuid];
3034 
3035 	do {
3036 		current = spec->dtsp_state;
3037 
3038 		switch (current) {
3039 		case DTRACESPEC_INACTIVE:
3040 		case DTRACESPEC_COMMITTINGMANY:
3041 		case DTRACESPEC_DISCARDING:
3042 			return (NULL);
3043 
3044 		case DTRACESPEC_COMMITTING:
3045 			ASSERT(buf->dtb_offset == 0);
3046 			return (NULL);
3047 
3048 		case DTRACESPEC_ACTIVEONE:
3049 			/*
3050 			 * This speculation is currently active on one CPU.
3051 			 * Check the offset in the buffer; if it's non-zero,
3052 			 * that CPU must be us (and we leave the state alone).
3053 			 * If it's zero, assume that we're starting on a new
3054 			 * CPU -- and change the state to indicate that the
3055 			 * speculation is active on more than one CPU.
3056 			 */
3057 			if (buf->dtb_offset != 0)
3058 				return (buf);
3059 
3060 			new = DTRACESPEC_ACTIVEMANY;
3061 			break;
3062 
3063 		case DTRACESPEC_ACTIVEMANY:
3064 			return (buf);
3065 
3066 		case DTRACESPEC_ACTIVE:
3067 			new = DTRACESPEC_ACTIVEONE;
3068 			break;
3069 
3070 		default:
3071 			ASSERT(0);
3072 		}
3073 	} while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
3074 	    current, new) != current);
3075 
3076 	ASSERT(new == DTRACESPEC_ACTIVEONE || new == DTRACESPEC_ACTIVEMANY);
3077 	return (buf);
3078 }
3079 
3080 /*
3081  * Return a string.  In the event that the user lacks the privilege to access
3082  * arbitrary kernel memory, we copy the string out to scratch memory so that we
3083  * don't fail access checking.
3084  *
3085  * dtrace_dif_variable() uses this routine as a helper for various
3086  * builtin values such as 'execname' and 'probefunc.'
3087  */
3088 uintptr_t
3089 dtrace_dif_varstr(uintptr_t addr, dtrace_state_t *state,
3090     dtrace_mstate_t *mstate)
3091 {
3092 	uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3093 	uintptr_t ret;
3094 	size_t strsz;
3095 
3096 	/*
3097 	 * The easy case: this probe is allowed to read all of memory, so
3098 	 * we can just return this as a vanilla pointer.
3099 	 */
3100 	if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
3101 		return (addr);
3102 
3103 	/*
3104 	 * This is the tougher case: we copy the string in question from
3105 	 * kernel memory into scratch memory and return it that way: this
3106 	 * ensures that we won't trip up when access checking tests the
3107 	 * BYREF return value.
3108 	 */
3109 	strsz = dtrace_strlen((char *)addr, size) + 1;
3110 
3111 	if (mstate->dtms_scratch_ptr + strsz >
3112 	    mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
3113 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3114 		return (0);
3115 	}
3116 
3117 	dtrace_strcpy((const void *)addr, (void *)mstate->dtms_scratch_ptr,
3118 	    strsz);
3119 	ret = mstate->dtms_scratch_ptr;
3120 	mstate->dtms_scratch_ptr += strsz;
3121 	return (ret);
3122 }
3123 
3124 /*
3125  * Return a string from a memoy address which is known to have one or
3126  * more concatenated, individually zero terminated, sub-strings.
3127  * In the event that the user lacks the privilege to access
3128  * arbitrary kernel memory, we copy the string out to scratch memory so that we
3129  * don't fail access checking.
3130  *
3131  * dtrace_dif_variable() uses this routine as a helper for various
3132  * builtin values such as 'execargs'.
3133  */
3134 static uintptr_t
3135 dtrace_dif_varstrz(uintptr_t addr, size_t strsz, dtrace_state_t *state,
3136     dtrace_mstate_t *mstate)
3137 {
3138 	char *p;
3139 	size_t i;
3140 	uintptr_t ret;
3141 
3142 	if (mstate->dtms_scratch_ptr + strsz >
3143 	    mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
3144 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3145 		return (0);
3146 	}
3147 
3148 	dtrace_bcopy((const void *)addr, (void *)mstate->dtms_scratch_ptr,
3149 	    strsz);
3150 
3151 	/* Replace sub-string termination characters with a space. */
3152 	for (p = (char *) mstate->dtms_scratch_ptr, i = 0; i < strsz - 1;
3153 	    p++, i++)
3154 		if (*p == '\0')
3155 			*p = ' ';
3156 
3157 	ret = mstate->dtms_scratch_ptr;
3158 	mstate->dtms_scratch_ptr += strsz;
3159 	return (ret);
3160 }
3161 
3162 /*
3163  * This function implements the DIF emulator's variable lookups.  The emulator
3164  * passes a reserved variable identifier and optional built-in array index.
3165  */
3166 static uint64_t
3167 dtrace_dif_variable(dtrace_mstate_t *mstate, dtrace_state_t *state, uint64_t v,
3168     uint64_t ndx)
3169 {
3170 	/*
3171 	 * If we're accessing one of the uncached arguments, we'll turn this
3172 	 * into a reference in the args array.
3173 	 */
3174 	if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9) {
3175 		ndx = v - DIF_VAR_ARG0;
3176 		v = DIF_VAR_ARGS;
3177 	}
3178 
3179 	switch (v) {
3180 	case DIF_VAR_ARGS:
3181 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_ARGS);
3182 		if (ndx >= sizeof (mstate->dtms_arg) /
3183 		    sizeof (mstate->dtms_arg[0])) {
3184 			int aframes = mstate->dtms_probe->dtpr_aframes + 2;
3185 			dtrace_provider_t *pv;
3186 			uint64_t val;
3187 
3188 			pv = mstate->dtms_probe->dtpr_provider;
3189 			if (pv->dtpv_pops.dtps_getargval != NULL)
3190 				val = pv->dtpv_pops.dtps_getargval(pv->dtpv_arg,
3191 				    mstate->dtms_probe->dtpr_id,
3192 				    mstate->dtms_probe->dtpr_arg, ndx, aframes);
3193 			else
3194 				val = dtrace_getarg(ndx, aframes);
3195 
3196 			/*
3197 			 * This is regrettably required to keep the compiler
3198 			 * from tail-optimizing the call to dtrace_getarg().
3199 			 * The condition always evaluates to true, but the
3200 			 * compiler has no way of figuring that out a priori.
3201 			 * (None of this would be necessary if the compiler
3202 			 * could be relied upon to _always_ tail-optimize
3203 			 * the call to dtrace_getarg() -- but it can't.)
3204 			 */
3205 			if (mstate->dtms_probe != NULL)
3206 				return (val);
3207 
3208 			ASSERT(0);
3209 		}
3210 
3211 		return (mstate->dtms_arg[ndx]);
3212 
3213 #ifdef illumos
3214 	case DIF_VAR_UREGS: {
3215 		klwp_t *lwp;
3216 
3217 		if (!dtrace_priv_proc(state))
3218 			return (0);
3219 
3220 		if ((lwp = curthread->t_lwp) == NULL) {
3221 			DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
3222 			cpu_core[curcpu].cpuc_dtrace_illval = NULL;
3223 			return (0);
3224 		}
3225 
3226 		return (dtrace_getreg(lwp->lwp_regs, ndx));
3227 		return (0);
3228 	}
3229 #else
3230 	case DIF_VAR_UREGS: {
3231 		struct trapframe *tframe;
3232 
3233 		if (!dtrace_priv_proc(state))
3234 			return (0);
3235 
3236 		if ((tframe = curthread->td_frame) == NULL) {
3237 			DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
3238 			cpu_core[curcpu].cpuc_dtrace_illval = 0;
3239 			return (0);
3240 		}
3241 
3242 		return (dtrace_getreg(tframe, ndx));
3243 	}
3244 #endif
3245 
3246 	case DIF_VAR_CURTHREAD:
3247 		if (!dtrace_priv_proc(state))
3248 			return (0);
3249 		return ((uint64_t)(uintptr_t)curthread);
3250 
3251 	case DIF_VAR_TIMESTAMP:
3252 		if (!(mstate->dtms_present & DTRACE_MSTATE_TIMESTAMP)) {
3253 			mstate->dtms_timestamp = dtrace_gethrtime();
3254 			mstate->dtms_present |= DTRACE_MSTATE_TIMESTAMP;
3255 		}
3256 		return (mstate->dtms_timestamp);
3257 
3258 	case DIF_VAR_VTIMESTAMP:
3259 		ASSERT(dtrace_vtime_references != 0);
3260 		return (curthread->t_dtrace_vtime);
3261 
3262 	case DIF_VAR_WALLTIMESTAMP:
3263 		if (!(mstate->dtms_present & DTRACE_MSTATE_WALLTIMESTAMP)) {
3264 			mstate->dtms_walltimestamp = dtrace_gethrestime();
3265 			mstate->dtms_present |= DTRACE_MSTATE_WALLTIMESTAMP;
3266 		}
3267 		return (mstate->dtms_walltimestamp);
3268 
3269 #ifdef illumos
3270 	case DIF_VAR_IPL:
3271 		if (!dtrace_priv_kernel(state))
3272 			return (0);
3273 		if (!(mstate->dtms_present & DTRACE_MSTATE_IPL)) {
3274 			mstate->dtms_ipl = dtrace_getipl();
3275 			mstate->dtms_present |= DTRACE_MSTATE_IPL;
3276 		}
3277 		return (mstate->dtms_ipl);
3278 #endif
3279 
3280 	case DIF_VAR_EPID:
3281 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_EPID);
3282 		return (mstate->dtms_epid);
3283 
3284 	case DIF_VAR_ID:
3285 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3286 		return (mstate->dtms_probe->dtpr_id);
3287 
3288 	case DIF_VAR_STACKDEPTH:
3289 		if (!dtrace_priv_kernel(state))
3290 			return (0);
3291 		if (!(mstate->dtms_present & DTRACE_MSTATE_STACKDEPTH)) {
3292 			int aframes = mstate->dtms_probe->dtpr_aframes + 2;
3293 
3294 			mstate->dtms_stackdepth = dtrace_getstackdepth(aframes);
3295 			mstate->dtms_present |= DTRACE_MSTATE_STACKDEPTH;
3296 		}
3297 		return (mstate->dtms_stackdepth);
3298 
3299 	case DIF_VAR_USTACKDEPTH:
3300 		if (!dtrace_priv_proc(state))
3301 			return (0);
3302 		if (!(mstate->dtms_present & DTRACE_MSTATE_USTACKDEPTH)) {
3303 			/*
3304 			 * See comment in DIF_VAR_PID.
3305 			 */
3306 			if (DTRACE_ANCHORED(mstate->dtms_probe) &&
3307 			    CPU_ON_INTR(CPU)) {
3308 				mstate->dtms_ustackdepth = 0;
3309 			} else {
3310 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3311 				mstate->dtms_ustackdepth =
3312 				    dtrace_getustackdepth();
3313 				DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3314 			}
3315 			mstate->dtms_present |= DTRACE_MSTATE_USTACKDEPTH;
3316 		}
3317 		return (mstate->dtms_ustackdepth);
3318 
3319 	case DIF_VAR_CALLER:
3320 		if (!dtrace_priv_kernel(state))
3321 			return (0);
3322 		if (!(mstate->dtms_present & DTRACE_MSTATE_CALLER)) {
3323 			int aframes = mstate->dtms_probe->dtpr_aframes + 2;
3324 
3325 			if (!DTRACE_ANCHORED(mstate->dtms_probe)) {
3326 				/*
3327 				 * If this is an unanchored probe, we are
3328 				 * required to go through the slow path:
3329 				 * dtrace_caller() only guarantees correct
3330 				 * results for anchored probes.
3331 				 */
3332 				pc_t caller[2] = {0, 0};
3333 
3334 				dtrace_getpcstack(caller, 2, aframes,
3335 				    (uint32_t *)(uintptr_t)mstate->dtms_arg[0]);
3336 				mstate->dtms_caller = caller[1];
3337 			} else if ((mstate->dtms_caller =
3338 			    dtrace_caller(aframes)) == -1) {
3339 				/*
3340 				 * We have failed to do this the quick way;
3341 				 * we must resort to the slower approach of
3342 				 * calling dtrace_getpcstack().
3343 				 */
3344 				pc_t caller = 0;
3345 
3346 				dtrace_getpcstack(&caller, 1, aframes, NULL);
3347 				mstate->dtms_caller = caller;
3348 			}
3349 
3350 			mstate->dtms_present |= DTRACE_MSTATE_CALLER;
3351 		}
3352 		return (mstate->dtms_caller);
3353 
3354 	case DIF_VAR_UCALLER:
3355 		if (!dtrace_priv_proc(state))
3356 			return (0);
3357 
3358 		if (!(mstate->dtms_present & DTRACE_MSTATE_UCALLER)) {
3359 			uint64_t ustack[3];
3360 
3361 			/*
3362 			 * dtrace_getupcstack() fills in the first uint64_t
3363 			 * with the current PID.  The second uint64_t will
3364 			 * be the program counter at user-level.  The third
3365 			 * uint64_t will contain the caller, which is what
3366 			 * we're after.
3367 			 */
3368 			ustack[2] = 0;
3369 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3370 			dtrace_getupcstack(ustack, 3);
3371 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3372 			mstate->dtms_ucaller = ustack[2];
3373 			mstate->dtms_present |= DTRACE_MSTATE_UCALLER;
3374 		}
3375 
3376 		return (mstate->dtms_ucaller);
3377 
3378 	case DIF_VAR_PROBEPROV:
3379 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3380 		return (dtrace_dif_varstr(
3381 		    (uintptr_t)mstate->dtms_probe->dtpr_provider->dtpv_name,
3382 		    state, mstate));
3383 
3384 	case DIF_VAR_PROBEMOD:
3385 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3386 		return (dtrace_dif_varstr(
3387 		    (uintptr_t)mstate->dtms_probe->dtpr_mod,
3388 		    state, mstate));
3389 
3390 	case DIF_VAR_PROBEFUNC:
3391 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3392 		return (dtrace_dif_varstr(
3393 		    (uintptr_t)mstate->dtms_probe->dtpr_func,
3394 		    state, mstate));
3395 
3396 	case DIF_VAR_PROBENAME:
3397 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3398 		return (dtrace_dif_varstr(
3399 		    (uintptr_t)mstate->dtms_probe->dtpr_name,
3400 		    state, mstate));
3401 
3402 	case DIF_VAR_PID:
3403 		if (!dtrace_priv_proc(state))
3404 			return (0);
3405 
3406 #ifdef illumos
3407 		/*
3408 		 * Note that we are assuming that an unanchored probe is
3409 		 * always due to a high-level interrupt.  (And we're assuming
3410 		 * that there is only a single high level interrupt.)
3411 		 */
3412 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3413 			return (pid0.pid_id);
3414 
3415 		/*
3416 		 * It is always safe to dereference one's own t_procp pointer:
3417 		 * it always points to a valid, allocated proc structure.
3418 		 * Further, it is always safe to dereference the p_pidp member
3419 		 * of one's own proc structure.  (These are truisms becuase
3420 		 * threads and processes don't clean up their own state --
3421 		 * they leave that task to whomever reaps them.)
3422 		 */
3423 		return ((uint64_t)curthread->t_procp->p_pidp->pid_id);
3424 #else
3425 		return ((uint64_t)curproc->p_pid);
3426 #endif
3427 
3428 	case DIF_VAR_PPID:
3429 		if (!dtrace_priv_proc(state))
3430 			return (0);
3431 
3432 #ifdef illumos
3433 		/*
3434 		 * See comment in DIF_VAR_PID.
3435 		 */
3436 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3437 			return (pid0.pid_id);
3438 
3439 		/*
3440 		 * It is always safe to dereference one's own t_procp pointer:
3441 		 * it always points to a valid, allocated proc structure.
3442 		 * (This is true because threads don't clean up their own
3443 		 * state -- they leave that task to whomever reaps them.)
3444 		 */
3445 		return ((uint64_t)curthread->t_procp->p_ppid);
3446 #else
3447 		if (curproc->p_pid == proc0.p_pid)
3448 			return (curproc->p_pid);
3449 		else
3450 			return (curproc->p_pptr->p_pid);
3451 #endif
3452 
3453 	case DIF_VAR_TID:
3454 #ifdef illumos
3455 		/*
3456 		 * See comment in DIF_VAR_PID.
3457 		 */
3458 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3459 			return (0);
3460 #endif
3461 
3462 		return ((uint64_t)curthread->t_tid);
3463 
3464 	case DIF_VAR_EXECARGS: {
3465 		struct pargs *p_args = curthread->td_proc->p_args;
3466 
3467 		if (p_args == NULL)
3468 			return(0);
3469 
3470 		return (dtrace_dif_varstrz(
3471 		    (uintptr_t) p_args->ar_args, p_args->ar_length, state, mstate));
3472 	}
3473 
3474 	case DIF_VAR_EXECNAME:
3475 #ifdef illumos
3476 		if (!dtrace_priv_proc(state))
3477 			return (0);
3478 
3479 		/*
3480 		 * See comment in DIF_VAR_PID.
3481 		 */
3482 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3483 			return ((uint64_t)(uintptr_t)p0.p_user.u_comm);
3484 
3485 		/*
3486 		 * It is always safe to dereference one's own t_procp pointer:
3487 		 * it always points to a valid, allocated proc structure.
3488 		 * (This is true because threads don't clean up their own
3489 		 * state -- they leave that task to whomever reaps them.)
3490 		 */
3491 		return (dtrace_dif_varstr(
3492 		    (uintptr_t)curthread->t_procp->p_user.u_comm,
3493 		    state, mstate));
3494 #else
3495 		return (dtrace_dif_varstr(
3496 		    (uintptr_t) curthread->td_proc->p_comm, state, mstate));
3497 #endif
3498 
3499 	case DIF_VAR_ZONENAME:
3500 #ifdef illumos
3501 		if (!dtrace_priv_proc(state))
3502 			return (0);
3503 
3504 		/*
3505 		 * See comment in DIF_VAR_PID.
3506 		 */
3507 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3508 			return ((uint64_t)(uintptr_t)p0.p_zone->zone_name);
3509 
3510 		/*
3511 		 * It is always safe to dereference one's own t_procp pointer:
3512 		 * it always points to a valid, allocated proc structure.
3513 		 * (This is true because threads don't clean up their own
3514 		 * state -- they leave that task to whomever reaps them.)
3515 		 */
3516 		return (dtrace_dif_varstr(
3517 		    (uintptr_t)curthread->t_procp->p_zone->zone_name,
3518 		    state, mstate));
3519 #else
3520 		return (0);
3521 #endif
3522 
3523 	case DIF_VAR_UID:
3524 		if (!dtrace_priv_proc(state))
3525 			return (0);
3526 
3527 #ifdef illumos
3528 		/*
3529 		 * See comment in DIF_VAR_PID.
3530 		 */
3531 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3532 			return ((uint64_t)p0.p_cred->cr_uid);
3533 
3534 		/*
3535 		 * It is always safe to dereference one's own t_procp pointer:
3536 		 * it always points to a valid, allocated proc structure.
3537 		 * (This is true because threads don't clean up their own
3538 		 * state -- they leave that task to whomever reaps them.)
3539 		 *
3540 		 * Additionally, it is safe to dereference one's own process
3541 		 * credential, since this is never NULL after process birth.
3542 		 */
3543 		return ((uint64_t)curthread->t_procp->p_cred->cr_uid);
3544 #else
3545 		return ((uint64_t)curthread->td_ucred->cr_uid);
3546 #endif
3547 
3548 	case DIF_VAR_GID:
3549 		if (!dtrace_priv_proc(state))
3550 			return (0);
3551 
3552 #ifdef illumos
3553 		/*
3554 		 * See comment in DIF_VAR_PID.
3555 		 */
3556 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3557 			return ((uint64_t)p0.p_cred->cr_gid);
3558 
3559 		/*
3560 		 * It is always safe to dereference one's own t_procp pointer:
3561 		 * it always points to a valid, allocated proc structure.
3562 		 * (This is true because threads don't clean up their own
3563 		 * state -- they leave that task to whomever reaps them.)
3564 		 *
3565 		 * Additionally, it is safe to dereference one's own process
3566 		 * credential, since this is never NULL after process birth.
3567 		 */
3568 		return ((uint64_t)curthread->t_procp->p_cred->cr_gid);
3569 #else
3570 		return ((uint64_t)curthread->td_ucred->cr_gid);
3571 #endif
3572 
3573 	case DIF_VAR_ERRNO: {
3574 #ifdef illumos
3575 		klwp_t *lwp;
3576 		if (!dtrace_priv_proc(state))
3577 			return (0);
3578 
3579 		/*
3580 		 * See comment in DIF_VAR_PID.
3581 		 */
3582 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3583 			return (0);
3584 
3585 		/*
3586 		 * It is always safe to dereference one's own t_lwp pointer in
3587 		 * the event that this pointer is non-NULL.  (This is true
3588 		 * because threads and lwps don't clean up their own state --
3589 		 * they leave that task to whomever reaps them.)
3590 		 */
3591 		if ((lwp = curthread->t_lwp) == NULL)
3592 			return (0);
3593 
3594 		return ((uint64_t)lwp->lwp_errno);
3595 #else
3596 		return (curthread->td_errno);
3597 #endif
3598 	}
3599 #ifndef illumos
3600 	case DIF_VAR_CPU: {
3601 		return curcpu;
3602 	}
3603 #endif
3604 	default:
3605 		DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
3606 		return (0);
3607 	}
3608 }
3609 
3610 
3611 typedef enum dtrace_json_state {
3612 	DTRACE_JSON_REST = 1,
3613 	DTRACE_JSON_OBJECT,
3614 	DTRACE_JSON_STRING,
3615 	DTRACE_JSON_STRING_ESCAPE,
3616 	DTRACE_JSON_STRING_ESCAPE_UNICODE,
3617 	DTRACE_JSON_COLON,
3618 	DTRACE_JSON_COMMA,
3619 	DTRACE_JSON_VALUE,
3620 	DTRACE_JSON_IDENTIFIER,
3621 	DTRACE_JSON_NUMBER,
3622 	DTRACE_JSON_NUMBER_FRAC,
3623 	DTRACE_JSON_NUMBER_EXP,
3624 	DTRACE_JSON_COLLECT_OBJECT
3625 } dtrace_json_state_t;
3626 
3627 /*
3628  * This function possesses just enough knowledge about JSON to extract a single
3629  * value from a JSON string and store it in the scratch buffer.  It is able
3630  * to extract nested object values, and members of arrays by index.
3631  *
3632  * elemlist is a list of JSON keys, stored as packed NUL-terminated strings, to
3633  * be looked up as we descend into the object tree.  e.g.
3634  *
3635  *    foo[0].bar.baz[32] --> "foo" NUL "0" NUL "bar" NUL "baz" NUL "32" NUL
3636  *       with nelems = 5.
3637  *
3638  * The run time of this function must be bounded above by strsize to limit the
3639  * amount of work done in probe context.  As such, it is implemented as a
3640  * simple state machine, reading one character at a time using safe loads
3641  * until we find the requested element, hit a parsing error or run off the
3642  * end of the object or string.
3643  *
3644  * As there is no way for a subroutine to return an error without interrupting
3645  * clause execution, we simply return NULL in the event of a missing key or any
3646  * other error condition.  Each NULL return in this function is commented with
3647  * the error condition it represents -- parsing or otherwise.
3648  *
3649  * The set of states for the state machine closely matches the JSON
3650  * specification (http://json.org/).  Briefly:
3651  *
3652  *   DTRACE_JSON_REST:
3653  *     Skip whitespace until we find either a top-level Object, moving
3654  *     to DTRACE_JSON_OBJECT; or an Array, moving to DTRACE_JSON_VALUE.
3655  *
3656  *   DTRACE_JSON_OBJECT:
3657  *     Locate the next key String in an Object.  Sets a flag to denote
3658  *     the next String as a key string and moves to DTRACE_JSON_STRING.
3659  *
3660  *   DTRACE_JSON_COLON:
3661  *     Skip whitespace until we find the colon that separates key Strings
3662  *     from their values.  Once found, move to DTRACE_JSON_VALUE.
3663  *
3664  *   DTRACE_JSON_VALUE:
3665  *     Detects the type of the next value (String, Number, Identifier, Object
3666  *     or Array) and routes to the states that process that type.  Here we also
3667  *     deal with the element selector list if we are requested to traverse down
3668  *     into the object tree.
3669  *
3670  *   DTRACE_JSON_COMMA:
3671  *     Skip whitespace until we find the comma that separates key-value pairs
3672  *     in Objects (returning to DTRACE_JSON_OBJECT) or values in Arrays
3673  *     (similarly DTRACE_JSON_VALUE).  All following literal value processing
3674  *     states return to this state at the end of their value, unless otherwise
3675  *     noted.
3676  *
3677  *   DTRACE_JSON_NUMBER, DTRACE_JSON_NUMBER_FRAC, DTRACE_JSON_NUMBER_EXP:
3678  *     Processes a Number literal from the JSON, including any exponent
3679  *     component that may be present.  Numbers are returned as strings, which
3680  *     may be passed to strtoll() if an integer is required.
3681  *
3682  *   DTRACE_JSON_IDENTIFIER:
3683  *     Processes a "true", "false" or "null" literal in the JSON.
3684  *
3685  *   DTRACE_JSON_STRING, DTRACE_JSON_STRING_ESCAPE,
3686  *   DTRACE_JSON_STRING_ESCAPE_UNICODE:
3687  *     Processes a String literal from the JSON, whether the String denotes
3688  *     a key, a value or part of a larger Object.  Handles all escape sequences
3689  *     present in the specification, including four-digit unicode characters,
3690  *     but merely includes the escape sequence without converting it to the
3691  *     actual escaped character.  If the String is flagged as a key, we
3692  *     move to DTRACE_JSON_COLON rather than DTRACE_JSON_COMMA.
3693  *
3694  *   DTRACE_JSON_COLLECT_OBJECT:
3695  *     This state collects an entire Object (or Array), correctly handling
3696  *     embedded strings.  If the full element selector list matches this nested
3697  *     object, we return the Object in full as a string.  If not, we use this
3698  *     state to skip to the next value at this level and continue processing.
3699  *
3700  * NOTE: This function uses various macros from strtolctype.h to manipulate
3701  * digit values, etc -- these have all been checked to ensure they make
3702  * no additional function calls.
3703  */
3704 static char *
3705 dtrace_json(uint64_t size, uintptr_t json, char *elemlist, int nelems,
3706     char *dest)
3707 {
3708 	dtrace_json_state_t state = DTRACE_JSON_REST;
3709 	int64_t array_elem = INT64_MIN;
3710 	int64_t array_pos = 0;
3711 	uint8_t escape_unicount = 0;
3712 	boolean_t string_is_key = B_FALSE;
3713 	boolean_t collect_object = B_FALSE;
3714 	boolean_t found_key = B_FALSE;
3715 	boolean_t in_array = B_FALSE;
3716 	uint32_t braces = 0, brackets = 0;
3717 	char *elem = elemlist;
3718 	char *dd = dest;
3719 	uintptr_t cur;
3720 
3721 	for (cur = json; cur < json + size; cur++) {
3722 		char cc = dtrace_load8(cur);
3723 		if (cc == '\0')
3724 			return (NULL);
3725 
3726 		switch (state) {
3727 		case DTRACE_JSON_REST:
3728 			if (isspace(cc))
3729 				break;
3730 
3731 			if (cc == '{') {
3732 				state = DTRACE_JSON_OBJECT;
3733 				break;
3734 			}
3735 
3736 			if (cc == '[') {
3737 				in_array = B_TRUE;
3738 				array_pos = 0;
3739 				array_elem = dtrace_strtoll(elem, 10, size);
3740 				found_key = array_elem == 0 ? B_TRUE : B_FALSE;
3741 				state = DTRACE_JSON_VALUE;
3742 				break;
3743 			}
3744 
3745 			/*
3746 			 * ERROR: expected to find a top-level object or array.
3747 			 */
3748 			return (NULL);
3749 		case DTRACE_JSON_OBJECT:
3750 			if (isspace(cc))
3751 				break;
3752 
3753 			if (cc == '"') {
3754 				state = DTRACE_JSON_STRING;
3755 				string_is_key = B_TRUE;
3756 				break;
3757 			}
3758 
3759 			/*
3760 			 * ERROR: either the object did not start with a key
3761 			 * string, or we've run off the end of the object
3762 			 * without finding the requested key.
3763 			 */
3764 			return (NULL);
3765 		case DTRACE_JSON_STRING:
3766 			if (cc == '\\') {
3767 				*dd++ = '\\';
3768 				state = DTRACE_JSON_STRING_ESCAPE;
3769 				break;
3770 			}
3771 
3772 			if (cc == '"') {
3773 				if (collect_object) {
3774 					/*
3775 					 * We don't reset the dest here, as
3776 					 * the string is part of a larger
3777 					 * object being collected.
3778 					 */
3779 					*dd++ = cc;
3780 					collect_object = B_FALSE;
3781 					state = DTRACE_JSON_COLLECT_OBJECT;
3782 					break;
3783 				}
3784 				*dd = '\0';
3785 				dd = dest; /* reset string buffer */
3786 				if (string_is_key) {
3787 					if (dtrace_strncmp(dest, elem,
3788 					    size) == 0)
3789 						found_key = B_TRUE;
3790 				} else if (found_key) {
3791 					if (nelems > 1) {
3792 						/*
3793 						 * We expected an object, not
3794 						 * this string.
3795 						 */
3796 						return (NULL);
3797 					}
3798 					return (dest);
3799 				}
3800 				state = string_is_key ? DTRACE_JSON_COLON :
3801 				    DTRACE_JSON_COMMA;
3802 				string_is_key = B_FALSE;
3803 				break;
3804 			}
3805 
3806 			*dd++ = cc;
3807 			break;
3808 		case DTRACE_JSON_STRING_ESCAPE:
3809 			*dd++ = cc;
3810 			if (cc == 'u') {
3811 				escape_unicount = 0;
3812 				state = DTRACE_JSON_STRING_ESCAPE_UNICODE;
3813 			} else {
3814 				state = DTRACE_JSON_STRING;
3815 			}
3816 			break;
3817 		case DTRACE_JSON_STRING_ESCAPE_UNICODE:
3818 			if (!isxdigit(cc)) {
3819 				/*
3820 				 * ERROR: invalid unicode escape, expected
3821 				 * four valid hexidecimal digits.
3822 				 */
3823 				return (NULL);
3824 			}
3825 
3826 			*dd++ = cc;
3827 			if (++escape_unicount == 4)
3828 				state = DTRACE_JSON_STRING;
3829 			break;
3830 		case DTRACE_JSON_COLON:
3831 			if (isspace(cc))
3832 				break;
3833 
3834 			if (cc == ':') {
3835 				state = DTRACE_JSON_VALUE;
3836 				break;
3837 			}
3838 
3839 			/*
3840 			 * ERROR: expected a colon.
3841 			 */
3842 			return (NULL);
3843 		case DTRACE_JSON_COMMA:
3844 			if (isspace(cc))
3845 				break;
3846 
3847 			if (cc == ',') {
3848 				if (in_array) {
3849 					state = DTRACE_JSON_VALUE;
3850 					if (++array_pos == array_elem)
3851 						found_key = B_TRUE;
3852 				} else {
3853 					state = DTRACE_JSON_OBJECT;
3854 				}
3855 				break;
3856 			}
3857 
3858 			/*
3859 			 * ERROR: either we hit an unexpected character, or
3860 			 * we reached the end of the object or array without
3861 			 * finding the requested key.
3862 			 */
3863 			return (NULL);
3864 		case DTRACE_JSON_IDENTIFIER:
3865 			if (islower(cc)) {
3866 				*dd++ = cc;
3867 				break;
3868 			}
3869 
3870 			*dd = '\0';
3871 			dd = dest; /* reset string buffer */
3872 
3873 			if (dtrace_strncmp(dest, "true", 5) == 0 ||
3874 			    dtrace_strncmp(dest, "false", 6) == 0 ||
3875 			    dtrace_strncmp(dest, "null", 5) == 0) {
3876 				if (found_key) {
3877 					if (nelems > 1) {
3878 						/*
3879 						 * ERROR: We expected an object,
3880 						 * not this identifier.
3881 						 */
3882 						return (NULL);
3883 					}
3884 					return (dest);
3885 				} else {
3886 					cur--;
3887 					state = DTRACE_JSON_COMMA;
3888 					break;
3889 				}
3890 			}
3891 
3892 			/*
3893 			 * ERROR: we did not recognise the identifier as one
3894 			 * of those in the JSON specification.
3895 			 */
3896 			return (NULL);
3897 		case DTRACE_JSON_NUMBER:
3898 			if (cc == '.') {
3899 				*dd++ = cc;
3900 				state = DTRACE_JSON_NUMBER_FRAC;
3901 				break;
3902 			}
3903 
3904 			if (cc == 'x' || cc == 'X') {
3905 				/*
3906 				 * ERROR: specification explicitly excludes
3907 				 * hexidecimal or octal numbers.
3908 				 */
3909 				return (NULL);
3910 			}
3911 
3912 			/* FALLTHRU */
3913 		case DTRACE_JSON_NUMBER_FRAC:
3914 			if (cc == 'e' || cc == 'E') {
3915 				*dd++ = cc;
3916 				state = DTRACE_JSON_NUMBER_EXP;
3917 				break;
3918 			}
3919 
3920 			if (cc == '+' || cc == '-') {
3921 				/*
3922 				 * ERROR: expect sign as part of exponent only.
3923 				 */
3924 				return (NULL);
3925 			}
3926 			/* FALLTHRU */
3927 		case DTRACE_JSON_NUMBER_EXP:
3928 			if (isdigit(cc) || cc == '+' || cc == '-') {
3929 				*dd++ = cc;
3930 				break;
3931 			}
3932 
3933 			*dd = '\0';
3934 			dd = dest; /* reset string buffer */
3935 			if (found_key) {
3936 				if (nelems > 1) {
3937 					/*
3938 					 * ERROR: We expected an object, not
3939 					 * this number.
3940 					 */
3941 					return (NULL);
3942 				}
3943 				return (dest);
3944 			}
3945 
3946 			cur--;
3947 			state = DTRACE_JSON_COMMA;
3948 			break;
3949 		case DTRACE_JSON_VALUE:
3950 			if (isspace(cc))
3951 				break;
3952 
3953 			if (cc == '{' || cc == '[') {
3954 				if (nelems > 1 && found_key) {
3955 					in_array = cc == '[' ? B_TRUE : B_FALSE;
3956 					/*
3957 					 * If our element selector directs us
3958 					 * to descend into this nested object,
3959 					 * then move to the next selector
3960 					 * element in the list and restart the
3961 					 * state machine.
3962 					 */
3963 					while (*elem != '\0')
3964 						elem++;
3965 					elem++; /* skip the inter-element NUL */
3966 					nelems--;
3967 					dd = dest;
3968 					if (in_array) {
3969 						state = DTRACE_JSON_VALUE;
3970 						array_pos = 0;
3971 						array_elem = dtrace_strtoll(
3972 						    elem, 10, size);
3973 						found_key = array_elem == 0 ?
3974 						    B_TRUE : B_FALSE;
3975 					} else {
3976 						found_key = B_FALSE;
3977 						state = DTRACE_JSON_OBJECT;
3978 					}
3979 					break;
3980 				}
3981 
3982 				/*
3983 				 * Otherwise, we wish to either skip this
3984 				 * nested object or return it in full.
3985 				 */
3986 				if (cc == '[')
3987 					brackets = 1;
3988 				else
3989 					braces = 1;
3990 				*dd++ = cc;
3991 				state = DTRACE_JSON_COLLECT_OBJECT;
3992 				break;
3993 			}
3994 
3995 			if (cc == '"') {
3996 				state = DTRACE_JSON_STRING;
3997 				break;
3998 			}
3999 
4000 			if (islower(cc)) {
4001 				/*
4002 				 * Here we deal with true, false and null.
4003 				 */
4004 				*dd++ = cc;
4005 				state = DTRACE_JSON_IDENTIFIER;
4006 				break;
4007 			}
4008 
4009 			if (cc == '-' || isdigit(cc)) {
4010 				*dd++ = cc;
4011 				state = DTRACE_JSON_NUMBER;
4012 				break;
4013 			}
4014 
4015 			/*
4016 			 * ERROR: unexpected character at start of value.
4017 			 */
4018 			return (NULL);
4019 		case DTRACE_JSON_COLLECT_OBJECT:
4020 			if (cc == '\0')
4021 				/*
4022 				 * ERROR: unexpected end of input.
4023 				 */
4024 				return (NULL);
4025 
4026 			*dd++ = cc;
4027 			if (cc == '"') {
4028 				collect_object = B_TRUE;
4029 				state = DTRACE_JSON_STRING;
4030 				break;
4031 			}
4032 
4033 			if (cc == ']') {
4034 				if (brackets-- == 0) {
4035 					/*
4036 					 * ERROR: unbalanced brackets.
4037 					 */
4038 					return (NULL);
4039 				}
4040 			} else if (cc == '}') {
4041 				if (braces-- == 0) {
4042 					/*
4043 					 * ERROR: unbalanced braces.
4044 					 */
4045 					return (NULL);
4046 				}
4047 			} else if (cc == '{') {
4048 				braces++;
4049 			} else if (cc == '[') {
4050 				brackets++;
4051 			}
4052 
4053 			if (brackets == 0 && braces == 0) {
4054 				if (found_key) {
4055 					*dd = '\0';
4056 					return (dest);
4057 				}
4058 				dd = dest; /* reset string buffer */
4059 				state = DTRACE_JSON_COMMA;
4060 			}
4061 			break;
4062 		}
4063 	}
4064 	return (NULL);
4065 }
4066 
4067 /*
4068  * Emulate the execution of DTrace ID subroutines invoked by the call opcode.
4069  * Notice that we don't bother validating the proper number of arguments or
4070  * their types in the tuple stack.  This isn't needed because all argument
4071  * interpretation is safe because of our load safety -- the worst that can
4072  * happen is that a bogus program can obtain bogus results.
4073  */
4074 static void
4075 dtrace_dif_subr(uint_t subr, uint_t rd, uint64_t *regs,
4076     dtrace_key_t *tupregs, int nargs,
4077     dtrace_mstate_t *mstate, dtrace_state_t *state)
4078 {
4079 	volatile uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
4080 	volatile uintptr_t *illval = &cpu_core[curcpu].cpuc_dtrace_illval;
4081 	dtrace_vstate_t *vstate = &state->dts_vstate;
4082 
4083 #ifdef illumos
4084 	union {
4085 		mutex_impl_t mi;
4086 		uint64_t mx;
4087 	} m;
4088 
4089 	union {
4090 		krwlock_t ri;
4091 		uintptr_t rw;
4092 	} r;
4093 #else
4094 	struct thread *lowner;
4095 	union {
4096 		struct lock_object *li;
4097 		uintptr_t lx;
4098 	} l;
4099 #endif
4100 
4101 	switch (subr) {
4102 	case DIF_SUBR_RAND:
4103 		regs[rd] = (dtrace_gethrtime() * 2416 + 374441) % 1771875;
4104 		break;
4105 
4106 #ifdef illumos
4107 	case DIF_SUBR_MUTEX_OWNED:
4108 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
4109 		    mstate, vstate)) {
4110 			regs[rd] = 0;
4111 			break;
4112 		}
4113 
4114 		m.mx = dtrace_load64(tupregs[0].dttk_value);
4115 		if (MUTEX_TYPE_ADAPTIVE(&m.mi))
4116 			regs[rd] = MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER;
4117 		else
4118 			regs[rd] = LOCK_HELD(&m.mi.m_spin.m_spinlock);
4119 		break;
4120 
4121 	case DIF_SUBR_MUTEX_OWNER:
4122 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
4123 		    mstate, vstate)) {
4124 			regs[rd] = 0;
4125 			break;
4126 		}
4127 
4128 		m.mx = dtrace_load64(tupregs[0].dttk_value);
4129 		if (MUTEX_TYPE_ADAPTIVE(&m.mi) &&
4130 		    MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER)
4131 			regs[rd] = (uintptr_t)MUTEX_OWNER(&m.mi);
4132 		else
4133 			regs[rd] = 0;
4134 		break;
4135 
4136 	case DIF_SUBR_MUTEX_TYPE_ADAPTIVE:
4137 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
4138 		    mstate, vstate)) {
4139 			regs[rd] = 0;
4140 			break;
4141 		}
4142 
4143 		m.mx = dtrace_load64(tupregs[0].dttk_value);
4144 		regs[rd] = MUTEX_TYPE_ADAPTIVE(&m.mi);
4145 		break;
4146 
4147 	case DIF_SUBR_MUTEX_TYPE_SPIN:
4148 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
4149 		    mstate, vstate)) {
4150 			regs[rd] = 0;
4151 			break;
4152 		}
4153 
4154 		m.mx = dtrace_load64(tupregs[0].dttk_value);
4155 		regs[rd] = MUTEX_TYPE_SPIN(&m.mi);
4156 		break;
4157 
4158 	case DIF_SUBR_RW_READ_HELD: {
4159 		uintptr_t tmp;
4160 
4161 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
4162 		    mstate, vstate)) {
4163 			regs[rd] = 0;
4164 			break;
4165 		}
4166 
4167 		r.rw = dtrace_loadptr(tupregs[0].dttk_value);
4168 		regs[rd] = _RW_READ_HELD(&r.ri, tmp);
4169 		break;
4170 	}
4171 
4172 	case DIF_SUBR_RW_WRITE_HELD:
4173 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
4174 		    mstate, vstate)) {
4175 			regs[rd] = 0;
4176 			break;
4177 		}
4178 
4179 		r.rw = dtrace_loadptr(tupregs[0].dttk_value);
4180 		regs[rd] = _RW_WRITE_HELD(&r.ri);
4181 		break;
4182 
4183 	case DIF_SUBR_RW_ISWRITER:
4184 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
4185 		    mstate, vstate)) {
4186 			regs[rd] = 0;
4187 			break;
4188 		}
4189 
4190 		r.rw = dtrace_loadptr(tupregs[0].dttk_value);
4191 		regs[rd] = _RW_ISWRITER(&r.ri);
4192 		break;
4193 
4194 #else /* !illumos */
4195 	case DIF_SUBR_MUTEX_OWNED:
4196 		if (!dtrace_canload(tupregs[0].dttk_value,
4197 			sizeof (struct lock_object), mstate, vstate)) {
4198 			regs[rd] = 0;
4199 			break;
4200 		}
4201 		l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
4202 		regs[rd] = LOCK_CLASS(l.li)->lc_owner(l.li, &lowner);
4203 		break;
4204 
4205 	case DIF_SUBR_MUTEX_OWNER:
4206 		if (!dtrace_canload(tupregs[0].dttk_value,
4207 			sizeof (struct lock_object), mstate, vstate)) {
4208 			regs[rd] = 0;
4209 			break;
4210 		}
4211 		l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
4212 		LOCK_CLASS(l.li)->lc_owner(l.li, &lowner);
4213 		regs[rd] = (uintptr_t)lowner;
4214 		break;
4215 
4216 	case DIF_SUBR_MUTEX_TYPE_ADAPTIVE:
4217 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (struct mtx),
4218 		    mstate, vstate)) {
4219 			regs[rd] = 0;
4220 			break;
4221 		}
4222 		l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
4223 		/* XXX - should be only LC_SLEEPABLE? */
4224 		regs[rd] = (LOCK_CLASS(l.li)->lc_flags &
4225 		    (LC_SLEEPLOCK | LC_SLEEPABLE)) != 0;
4226 		break;
4227 
4228 	case DIF_SUBR_MUTEX_TYPE_SPIN:
4229 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (struct mtx),
4230 		    mstate, vstate)) {
4231 			regs[rd] = 0;
4232 			break;
4233 		}
4234 		l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
4235 		regs[rd] = (LOCK_CLASS(l.li)->lc_flags & LC_SPINLOCK) != 0;
4236 		break;
4237 
4238 	case DIF_SUBR_RW_READ_HELD:
4239 	case DIF_SUBR_SX_SHARED_HELD:
4240 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
4241 		    mstate, vstate)) {
4242 			regs[rd] = 0;
4243 			break;
4244 		}
4245 		l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
4246 		regs[rd] = LOCK_CLASS(l.li)->lc_owner(l.li, &lowner) &&
4247 		    lowner == NULL;
4248 		break;
4249 
4250 	case DIF_SUBR_RW_WRITE_HELD:
4251 	case DIF_SUBR_SX_EXCLUSIVE_HELD:
4252 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
4253 		    mstate, vstate)) {
4254 			regs[rd] = 0;
4255 			break;
4256 		}
4257 		l.lx = dtrace_loadptr(tupregs[0].dttk_value);
4258 		LOCK_CLASS(l.li)->lc_owner(l.li, &lowner);
4259 		regs[rd] = (lowner == curthread);
4260 		break;
4261 
4262 	case DIF_SUBR_RW_ISWRITER:
4263 	case DIF_SUBR_SX_ISEXCLUSIVE:
4264 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
4265 		    mstate, vstate)) {
4266 			regs[rd] = 0;
4267 			break;
4268 		}
4269 		l.lx = dtrace_loadptr(tupregs[0].dttk_value);
4270 		regs[rd] = LOCK_CLASS(l.li)->lc_owner(l.li, &lowner) &&
4271 		    lowner != NULL;
4272 		break;
4273 #endif /* illumos */
4274 
4275 	case DIF_SUBR_BCOPY: {
4276 		/*
4277 		 * We need to be sure that the destination is in the scratch
4278 		 * region -- no other region is allowed.
4279 		 */
4280 		uintptr_t src = tupregs[0].dttk_value;
4281 		uintptr_t dest = tupregs[1].dttk_value;
4282 		size_t size = tupregs[2].dttk_value;
4283 
4284 		if (!dtrace_inscratch(dest, size, mstate)) {
4285 			*flags |= CPU_DTRACE_BADADDR;
4286 			*illval = regs[rd];
4287 			break;
4288 		}
4289 
4290 		if (!dtrace_canload(src, size, mstate, vstate)) {
4291 			regs[rd] = 0;
4292 			break;
4293 		}
4294 
4295 		dtrace_bcopy((void *)src, (void *)dest, size);
4296 		break;
4297 	}
4298 
4299 	case DIF_SUBR_ALLOCA:
4300 	case DIF_SUBR_COPYIN: {
4301 		uintptr_t dest = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
4302 		uint64_t size =
4303 		    tupregs[subr == DIF_SUBR_ALLOCA ? 0 : 1].dttk_value;
4304 		size_t scratch_size = (dest - mstate->dtms_scratch_ptr) + size;
4305 
4306 		/*
4307 		 * This action doesn't require any credential checks since
4308 		 * probes will not activate in user contexts to which the
4309 		 * enabling user does not have permissions.
4310 		 */
4311 
4312 		/*
4313 		 * Rounding up the user allocation size could have overflowed
4314 		 * a large, bogus allocation (like -1ULL) to 0.
4315 		 */
4316 		if (scratch_size < size ||
4317 		    !DTRACE_INSCRATCH(mstate, scratch_size)) {
4318 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4319 			regs[rd] = 0;
4320 			break;
4321 		}
4322 
4323 		if (subr == DIF_SUBR_COPYIN) {
4324 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4325 			dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
4326 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4327 		}
4328 
4329 		mstate->dtms_scratch_ptr += scratch_size;
4330 		regs[rd] = dest;
4331 		break;
4332 	}
4333 
4334 	case DIF_SUBR_COPYINTO: {
4335 		uint64_t size = tupregs[1].dttk_value;
4336 		uintptr_t dest = tupregs[2].dttk_value;
4337 
4338 		/*
4339 		 * This action doesn't require any credential checks since
4340 		 * probes will not activate in user contexts to which the
4341 		 * enabling user does not have permissions.
4342 		 */
4343 		if (!dtrace_inscratch(dest, size, mstate)) {
4344 			*flags |= CPU_DTRACE_BADADDR;
4345 			*illval = regs[rd];
4346 			break;
4347 		}
4348 
4349 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4350 		dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
4351 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4352 		break;
4353 	}
4354 
4355 	case DIF_SUBR_COPYINSTR: {
4356 		uintptr_t dest = mstate->dtms_scratch_ptr;
4357 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4358 
4359 		if (nargs > 1 && tupregs[1].dttk_value < size)
4360 			size = tupregs[1].dttk_value + 1;
4361 
4362 		/*
4363 		 * This action doesn't require any credential checks since
4364 		 * probes will not activate in user contexts to which the
4365 		 * enabling user does not have permissions.
4366 		 */
4367 		if (!DTRACE_INSCRATCH(mstate, size)) {
4368 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4369 			regs[rd] = 0;
4370 			break;
4371 		}
4372 
4373 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4374 		dtrace_copyinstr(tupregs[0].dttk_value, dest, size, flags);
4375 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4376 
4377 		((char *)dest)[size - 1] = '\0';
4378 		mstate->dtms_scratch_ptr += size;
4379 		regs[rd] = dest;
4380 		break;
4381 	}
4382 
4383 #ifdef illumos
4384 	case DIF_SUBR_MSGSIZE:
4385 	case DIF_SUBR_MSGDSIZE: {
4386 		uintptr_t baddr = tupregs[0].dttk_value, daddr;
4387 		uintptr_t wptr, rptr;
4388 		size_t count = 0;
4389 		int cont = 0;
4390 
4391 		while (baddr != 0 && !(*flags & CPU_DTRACE_FAULT)) {
4392 
4393 			if (!dtrace_canload(baddr, sizeof (mblk_t), mstate,
4394 			    vstate)) {
4395 				regs[rd] = 0;
4396 				break;
4397 			}
4398 
4399 			wptr = dtrace_loadptr(baddr +
4400 			    offsetof(mblk_t, b_wptr));
4401 
4402 			rptr = dtrace_loadptr(baddr +
4403 			    offsetof(mblk_t, b_rptr));
4404 
4405 			if (wptr < rptr) {
4406 				*flags |= CPU_DTRACE_BADADDR;
4407 				*illval = tupregs[0].dttk_value;
4408 				break;
4409 			}
4410 
4411 			daddr = dtrace_loadptr(baddr +
4412 			    offsetof(mblk_t, b_datap));
4413 
4414 			baddr = dtrace_loadptr(baddr +
4415 			    offsetof(mblk_t, b_cont));
4416 
4417 			/*
4418 			 * We want to prevent against denial-of-service here,
4419 			 * so we're only going to search the list for
4420 			 * dtrace_msgdsize_max mblks.
4421 			 */
4422 			if (cont++ > dtrace_msgdsize_max) {
4423 				*flags |= CPU_DTRACE_ILLOP;
4424 				break;
4425 			}
4426 
4427 			if (subr == DIF_SUBR_MSGDSIZE) {
4428 				if (dtrace_load8(daddr +
4429 				    offsetof(dblk_t, db_type)) != M_DATA)
4430 					continue;
4431 			}
4432 
4433 			count += wptr - rptr;
4434 		}
4435 
4436 		if (!(*flags & CPU_DTRACE_FAULT))
4437 			regs[rd] = count;
4438 
4439 		break;
4440 	}
4441 #endif
4442 
4443 	case DIF_SUBR_PROGENYOF: {
4444 		pid_t pid = tupregs[0].dttk_value;
4445 		proc_t *p;
4446 		int rval = 0;
4447 
4448 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4449 
4450 		for (p = curthread->t_procp; p != NULL; p = p->p_parent) {
4451 #ifdef illumos
4452 			if (p->p_pidp->pid_id == pid) {
4453 #else
4454 			if (p->p_pid == pid) {
4455 #endif
4456 				rval = 1;
4457 				break;
4458 			}
4459 		}
4460 
4461 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4462 
4463 		regs[rd] = rval;
4464 		break;
4465 	}
4466 
4467 	case DIF_SUBR_SPECULATION:
4468 		regs[rd] = dtrace_speculation(state);
4469 		break;
4470 
4471 	case DIF_SUBR_COPYOUT: {
4472 		uintptr_t kaddr = tupregs[0].dttk_value;
4473 		uintptr_t uaddr = tupregs[1].dttk_value;
4474 		uint64_t size = tupregs[2].dttk_value;
4475 
4476 		if (!dtrace_destructive_disallow &&
4477 		    dtrace_priv_proc_control(state) &&
4478 		    !dtrace_istoxic(kaddr, size) &&
4479 		    dtrace_canload(kaddr, size, mstate, vstate)) {
4480 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4481 			dtrace_copyout(kaddr, uaddr, size, flags);
4482 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4483 		}
4484 		break;
4485 	}
4486 
4487 	case DIF_SUBR_COPYOUTSTR: {
4488 		uintptr_t kaddr = tupregs[0].dttk_value;
4489 		uintptr_t uaddr = tupregs[1].dttk_value;
4490 		uint64_t size = tupregs[2].dttk_value;
4491 
4492 		if (!dtrace_destructive_disallow &&
4493 		    dtrace_priv_proc_control(state) &&
4494 		    !dtrace_istoxic(kaddr, size) &&
4495 		    dtrace_strcanload(kaddr, size, mstate, vstate)) {
4496 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4497 			dtrace_copyoutstr(kaddr, uaddr, size, flags);
4498 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4499 		}
4500 		break;
4501 	}
4502 
4503 	case DIF_SUBR_STRLEN: {
4504 		size_t sz;
4505 		uintptr_t addr = (uintptr_t)tupregs[0].dttk_value;
4506 		sz = dtrace_strlen((char *)addr,
4507 		    state->dts_options[DTRACEOPT_STRSIZE]);
4508 
4509 		if (!dtrace_canload(addr, sz + 1, mstate, vstate)) {
4510 			regs[rd] = 0;
4511 			break;
4512 		}
4513 
4514 		regs[rd] = sz;
4515 
4516 		break;
4517 	}
4518 
4519 	case DIF_SUBR_STRCHR:
4520 	case DIF_SUBR_STRRCHR: {
4521 		/*
4522 		 * We're going to iterate over the string looking for the
4523 		 * specified character.  We will iterate until we have reached
4524 		 * the string length or we have found the character.  If this
4525 		 * is DIF_SUBR_STRRCHR, we will look for the last occurrence
4526 		 * of the specified character instead of the first.
4527 		 */
4528 		uintptr_t saddr = tupregs[0].dttk_value;
4529 		uintptr_t addr = tupregs[0].dttk_value;
4530 		uintptr_t limit = addr + state->dts_options[DTRACEOPT_STRSIZE];
4531 		char c, target = (char)tupregs[1].dttk_value;
4532 
4533 		for (regs[rd] = 0; addr < limit; addr++) {
4534 			if ((c = dtrace_load8(addr)) == target) {
4535 				regs[rd] = addr;
4536 
4537 				if (subr == DIF_SUBR_STRCHR)
4538 					break;
4539 			}
4540 
4541 			if (c == '\0')
4542 				break;
4543 		}
4544 
4545 		if (!dtrace_canload(saddr, addr - saddr, mstate, vstate)) {
4546 			regs[rd] = 0;
4547 			break;
4548 		}
4549 
4550 		break;
4551 	}
4552 
4553 	case DIF_SUBR_STRSTR:
4554 	case DIF_SUBR_INDEX:
4555 	case DIF_SUBR_RINDEX: {
4556 		/*
4557 		 * We're going to iterate over the string looking for the
4558 		 * specified string.  We will iterate until we have reached
4559 		 * the string length or we have found the string.  (Yes, this
4560 		 * is done in the most naive way possible -- but considering
4561 		 * that the string we're searching for is likely to be
4562 		 * relatively short, the complexity of Rabin-Karp or similar
4563 		 * hardly seems merited.)
4564 		 */
4565 		char *addr = (char *)(uintptr_t)tupregs[0].dttk_value;
4566 		char *substr = (char *)(uintptr_t)tupregs[1].dttk_value;
4567 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4568 		size_t len = dtrace_strlen(addr, size);
4569 		size_t sublen = dtrace_strlen(substr, size);
4570 		char *limit = addr + len, *orig = addr;
4571 		int notfound = subr == DIF_SUBR_STRSTR ? 0 : -1;
4572 		int inc = 1;
4573 
4574 		regs[rd] = notfound;
4575 
4576 		if (!dtrace_canload((uintptr_t)addr, len + 1, mstate, vstate)) {
4577 			regs[rd] = 0;
4578 			break;
4579 		}
4580 
4581 		if (!dtrace_canload((uintptr_t)substr, sublen + 1, mstate,
4582 		    vstate)) {
4583 			regs[rd] = 0;
4584 			break;
4585 		}
4586 
4587 		/*
4588 		 * strstr() and index()/rindex() have similar semantics if
4589 		 * both strings are the empty string: strstr() returns a
4590 		 * pointer to the (empty) string, and index() and rindex()
4591 		 * both return index 0 (regardless of any position argument).
4592 		 */
4593 		if (sublen == 0 && len == 0) {
4594 			if (subr == DIF_SUBR_STRSTR)
4595 				regs[rd] = (uintptr_t)addr;
4596 			else
4597 				regs[rd] = 0;
4598 			break;
4599 		}
4600 
4601 		if (subr != DIF_SUBR_STRSTR) {
4602 			if (subr == DIF_SUBR_RINDEX) {
4603 				limit = orig - 1;
4604 				addr += len;
4605 				inc = -1;
4606 			}
4607 
4608 			/*
4609 			 * Both index() and rindex() take an optional position
4610 			 * argument that denotes the starting position.
4611 			 */
4612 			if (nargs == 3) {
4613 				int64_t pos = (int64_t)tupregs[2].dttk_value;
4614 
4615 				/*
4616 				 * If the position argument to index() is
4617 				 * negative, Perl implicitly clamps it at
4618 				 * zero.  This semantic is a little surprising
4619 				 * given the special meaning of negative
4620 				 * positions to similar Perl functions like
4621 				 * substr(), but it appears to reflect a
4622 				 * notion that index() can start from a
4623 				 * negative index and increment its way up to
4624 				 * the string.  Given this notion, Perl's
4625 				 * rindex() is at least self-consistent in
4626 				 * that it implicitly clamps positions greater
4627 				 * than the string length to be the string
4628 				 * length.  Where Perl completely loses
4629 				 * coherence, however, is when the specified
4630 				 * substring is the empty string ("").  In
4631 				 * this case, even if the position is
4632 				 * negative, rindex() returns 0 -- and even if
4633 				 * the position is greater than the length,
4634 				 * index() returns the string length.  These
4635 				 * semantics violate the notion that index()
4636 				 * should never return a value less than the
4637 				 * specified position and that rindex() should
4638 				 * never return a value greater than the
4639 				 * specified position.  (One assumes that
4640 				 * these semantics are artifacts of Perl's
4641 				 * implementation and not the results of
4642 				 * deliberate design -- it beggars belief that
4643 				 * even Larry Wall could desire such oddness.)
4644 				 * While in the abstract one would wish for
4645 				 * consistent position semantics across
4646 				 * substr(), index() and rindex() -- or at the
4647 				 * very least self-consistent position
4648 				 * semantics for index() and rindex() -- we
4649 				 * instead opt to keep with the extant Perl
4650 				 * semantics, in all their broken glory.  (Do
4651 				 * we have more desire to maintain Perl's
4652 				 * semantics than Perl does?  Probably.)
4653 				 */
4654 				if (subr == DIF_SUBR_RINDEX) {
4655 					if (pos < 0) {
4656 						if (sublen == 0)
4657 							regs[rd] = 0;
4658 						break;
4659 					}
4660 
4661 					if (pos > len)
4662 						pos = len;
4663 				} else {
4664 					if (pos < 0)
4665 						pos = 0;
4666 
4667 					if (pos >= len) {
4668 						if (sublen == 0)
4669 							regs[rd] = len;
4670 						break;
4671 					}
4672 				}
4673 
4674 				addr = orig + pos;
4675 			}
4676 		}
4677 
4678 		for (regs[rd] = notfound; addr != limit; addr += inc) {
4679 			if (dtrace_strncmp(addr, substr, sublen) == 0) {
4680 				if (subr != DIF_SUBR_STRSTR) {
4681 					/*
4682 					 * As D index() and rindex() are
4683 					 * modeled on Perl (and not on awk),
4684 					 * we return a zero-based (and not a
4685 					 * one-based) index.  (For you Perl
4686 					 * weenies: no, we're not going to add
4687 					 * $[ -- and shouldn't you be at a con
4688 					 * or something?)
4689 					 */
4690 					regs[rd] = (uintptr_t)(addr - orig);
4691 					break;
4692 				}
4693 
4694 				ASSERT(subr == DIF_SUBR_STRSTR);
4695 				regs[rd] = (uintptr_t)addr;
4696 				break;
4697 			}
4698 		}
4699 
4700 		break;
4701 	}
4702 
4703 	case DIF_SUBR_STRTOK: {
4704 		uintptr_t addr = tupregs[0].dttk_value;
4705 		uintptr_t tokaddr = tupregs[1].dttk_value;
4706 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4707 		uintptr_t limit, toklimit = tokaddr + size;
4708 		uint8_t c = 0, tokmap[32];	 /* 256 / 8 */
4709 		char *dest = (char *)mstate->dtms_scratch_ptr;
4710 		int i;
4711 
4712 		/*
4713 		 * Check both the token buffer and (later) the input buffer,
4714 		 * since both could be non-scratch addresses.
4715 		 */
4716 		if (!dtrace_strcanload(tokaddr, size, mstate, vstate)) {
4717 			regs[rd] = 0;
4718 			break;
4719 		}
4720 
4721 		if (!DTRACE_INSCRATCH(mstate, size)) {
4722 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4723 			regs[rd] = 0;
4724 			break;
4725 		}
4726 
4727 		if (addr == 0) {
4728 			/*
4729 			 * If the address specified is NULL, we use our saved
4730 			 * strtok pointer from the mstate.  Note that this
4731 			 * means that the saved strtok pointer is _only_
4732 			 * valid within multiple enablings of the same probe --
4733 			 * it behaves like an implicit clause-local variable.
4734 			 */
4735 			addr = mstate->dtms_strtok;
4736 		} else {
4737 			/*
4738 			 * If the user-specified address is non-NULL we must
4739 			 * access check it.  This is the only time we have
4740 			 * a chance to do so, since this address may reside
4741 			 * in the string table of this clause-- future calls
4742 			 * (when we fetch addr from mstate->dtms_strtok)
4743 			 * would fail this access check.
4744 			 */
4745 			if (!dtrace_strcanload(addr, size, mstate, vstate)) {
4746 				regs[rd] = 0;
4747 				break;
4748 			}
4749 		}
4750 
4751 		/*
4752 		 * First, zero the token map, and then process the token
4753 		 * string -- setting a bit in the map for every character
4754 		 * found in the token string.
4755 		 */
4756 		for (i = 0; i < sizeof (tokmap); i++)
4757 			tokmap[i] = 0;
4758 
4759 		for (; tokaddr < toklimit; tokaddr++) {
4760 			if ((c = dtrace_load8(tokaddr)) == '\0')
4761 				break;
4762 
4763 			ASSERT((c >> 3) < sizeof (tokmap));
4764 			tokmap[c >> 3] |= (1 << (c & 0x7));
4765 		}
4766 
4767 		for (limit = addr + size; addr < limit; addr++) {
4768 			/*
4769 			 * We're looking for a character that is _not_ contained
4770 			 * in the token string.
4771 			 */
4772 			if ((c = dtrace_load8(addr)) == '\0')
4773 				break;
4774 
4775 			if (!(tokmap[c >> 3] & (1 << (c & 0x7))))
4776 				break;
4777 		}
4778 
4779 		if (c == '\0') {
4780 			/*
4781 			 * We reached the end of the string without finding
4782 			 * any character that was not in the token string.
4783 			 * We return NULL in this case, and we set the saved
4784 			 * address to NULL as well.
4785 			 */
4786 			regs[rd] = 0;
4787 			mstate->dtms_strtok = 0;
4788 			break;
4789 		}
4790 
4791 		/*
4792 		 * From here on, we're copying into the destination string.
4793 		 */
4794 		for (i = 0; addr < limit && i < size - 1; addr++) {
4795 			if ((c = dtrace_load8(addr)) == '\0')
4796 				break;
4797 
4798 			if (tokmap[c >> 3] & (1 << (c & 0x7)))
4799 				break;
4800 
4801 			ASSERT(i < size);
4802 			dest[i++] = c;
4803 		}
4804 
4805 		ASSERT(i < size);
4806 		dest[i] = '\0';
4807 		regs[rd] = (uintptr_t)dest;
4808 		mstate->dtms_scratch_ptr += size;
4809 		mstate->dtms_strtok = addr;
4810 		break;
4811 	}
4812 
4813 	case DIF_SUBR_SUBSTR: {
4814 		uintptr_t s = tupregs[0].dttk_value;
4815 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4816 		char *d = (char *)mstate->dtms_scratch_ptr;
4817 		int64_t index = (int64_t)tupregs[1].dttk_value;
4818 		int64_t remaining = (int64_t)tupregs[2].dttk_value;
4819 		size_t len = dtrace_strlen((char *)s, size);
4820 		int64_t i;
4821 
4822 		if (!dtrace_canload(s, len + 1, mstate, vstate)) {
4823 			regs[rd] = 0;
4824 			break;
4825 		}
4826 
4827 		if (!DTRACE_INSCRATCH(mstate, size)) {
4828 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4829 			regs[rd] = 0;
4830 			break;
4831 		}
4832 
4833 		if (nargs <= 2)
4834 			remaining = (int64_t)size;
4835 
4836 		if (index < 0) {
4837 			index += len;
4838 
4839 			if (index < 0 && index + remaining > 0) {
4840 				remaining += index;
4841 				index = 0;
4842 			}
4843 		}
4844 
4845 		if (index >= len || index < 0) {
4846 			remaining = 0;
4847 		} else if (remaining < 0) {
4848 			remaining += len - index;
4849 		} else if (index + remaining > size) {
4850 			remaining = size - index;
4851 		}
4852 
4853 		for (i = 0; i < remaining; i++) {
4854 			if ((d[i] = dtrace_load8(s + index + i)) == '\0')
4855 				break;
4856 		}
4857 
4858 		d[i] = '\0';
4859 
4860 		mstate->dtms_scratch_ptr += size;
4861 		regs[rd] = (uintptr_t)d;
4862 		break;
4863 	}
4864 
4865 	case DIF_SUBR_JSON: {
4866 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4867 		uintptr_t json = tupregs[0].dttk_value;
4868 		size_t jsonlen = dtrace_strlen((char *)json, size);
4869 		uintptr_t elem = tupregs[1].dttk_value;
4870 		size_t elemlen = dtrace_strlen((char *)elem, size);
4871 
4872 		char *dest = (char *)mstate->dtms_scratch_ptr;
4873 		char *elemlist = (char *)mstate->dtms_scratch_ptr + jsonlen + 1;
4874 		char *ee = elemlist;
4875 		int nelems = 1;
4876 		uintptr_t cur;
4877 
4878 		if (!dtrace_canload(json, jsonlen + 1, mstate, vstate) ||
4879 		    !dtrace_canload(elem, elemlen + 1, mstate, vstate)) {
4880 			regs[rd] = 0;
4881 			break;
4882 		}
4883 
4884 		if (!DTRACE_INSCRATCH(mstate, jsonlen + 1 + elemlen + 1)) {
4885 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4886 			regs[rd] = 0;
4887 			break;
4888 		}
4889 
4890 		/*
4891 		 * Read the element selector and split it up into a packed list
4892 		 * of strings.
4893 		 */
4894 		for (cur = elem; cur < elem + elemlen; cur++) {
4895 			char cc = dtrace_load8(cur);
4896 
4897 			if (cur == elem && cc == '[') {
4898 				/*
4899 				 * If the first element selector key is
4900 				 * actually an array index then ignore the
4901 				 * bracket.
4902 				 */
4903 				continue;
4904 			}
4905 
4906 			if (cc == ']')
4907 				continue;
4908 
4909 			if (cc == '.' || cc == '[') {
4910 				nelems++;
4911 				cc = '\0';
4912 			}
4913 
4914 			*ee++ = cc;
4915 		}
4916 		*ee++ = '\0';
4917 
4918 		if ((regs[rd] = (uintptr_t)dtrace_json(size, json, elemlist,
4919 		    nelems, dest)) != 0)
4920 			mstate->dtms_scratch_ptr += jsonlen + 1;
4921 		break;
4922 	}
4923 
4924 	case DIF_SUBR_TOUPPER:
4925 	case DIF_SUBR_TOLOWER: {
4926 		uintptr_t s = tupregs[0].dttk_value;
4927 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4928 		char *dest = (char *)mstate->dtms_scratch_ptr, c;
4929 		size_t len = dtrace_strlen((char *)s, size);
4930 		char lower, upper, convert;
4931 		int64_t i;
4932 
4933 		if (subr == DIF_SUBR_TOUPPER) {
4934 			lower = 'a';
4935 			upper = 'z';
4936 			convert = 'A';
4937 		} else {
4938 			lower = 'A';
4939 			upper = 'Z';
4940 			convert = 'a';
4941 		}
4942 
4943 		if (!dtrace_canload(s, len + 1, mstate, vstate)) {
4944 			regs[rd] = 0;
4945 			break;
4946 		}
4947 
4948 		if (!DTRACE_INSCRATCH(mstate, size)) {
4949 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4950 			regs[rd] = 0;
4951 			break;
4952 		}
4953 
4954 		for (i = 0; i < size - 1; i++) {
4955 			if ((c = dtrace_load8(s + i)) == '\0')
4956 				break;
4957 
4958 			if (c >= lower && c <= upper)
4959 				c = convert + (c - lower);
4960 
4961 			dest[i] = c;
4962 		}
4963 
4964 		ASSERT(i < size);
4965 		dest[i] = '\0';
4966 		regs[rd] = (uintptr_t)dest;
4967 		mstate->dtms_scratch_ptr += size;
4968 		break;
4969 	}
4970 
4971 #ifdef illumos
4972 	case DIF_SUBR_GETMAJOR:
4973 #ifdef _LP64
4974 		regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR64) & MAXMAJ64;
4975 #else
4976 		regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR) & MAXMAJ;
4977 #endif
4978 		break;
4979 
4980 	case DIF_SUBR_GETMINOR:
4981 #ifdef _LP64
4982 		regs[rd] = tupregs[0].dttk_value & MAXMIN64;
4983 #else
4984 		regs[rd] = tupregs[0].dttk_value & MAXMIN;
4985 #endif
4986 		break;
4987 
4988 	case DIF_SUBR_DDI_PATHNAME: {
4989 		/*
4990 		 * This one is a galactic mess.  We are going to roughly
4991 		 * emulate ddi_pathname(), but it's made more complicated
4992 		 * by the fact that we (a) want to include the minor name and
4993 		 * (b) must proceed iteratively instead of recursively.
4994 		 */
4995 		uintptr_t dest = mstate->dtms_scratch_ptr;
4996 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4997 		char *start = (char *)dest, *end = start + size - 1;
4998 		uintptr_t daddr = tupregs[0].dttk_value;
4999 		int64_t minor = (int64_t)tupregs[1].dttk_value;
5000 		char *s;
5001 		int i, len, depth = 0;
5002 
5003 		/*
5004 		 * Due to all the pointer jumping we do and context we must
5005 		 * rely upon, we just mandate that the user must have kernel
5006 		 * read privileges to use this routine.
5007 		 */
5008 		if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) == 0) {
5009 			*flags |= CPU_DTRACE_KPRIV;
5010 			*illval = daddr;
5011 			regs[rd] = 0;
5012 		}
5013 
5014 		if (!DTRACE_INSCRATCH(mstate, size)) {
5015 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5016 			regs[rd] = 0;
5017 			break;
5018 		}
5019 
5020 		*end = '\0';
5021 
5022 		/*
5023 		 * We want to have a name for the minor.  In order to do this,
5024 		 * we need to walk the minor list from the devinfo.  We want
5025 		 * to be sure that we don't infinitely walk a circular list,
5026 		 * so we check for circularity by sending a scout pointer
5027 		 * ahead two elements for every element that we iterate over;
5028 		 * if the list is circular, these will ultimately point to the
5029 		 * same element.  You may recognize this little trick as the
5030 		 * answer to a stupid interview question -- one that always
5031 		 * seems to be asked by those who had to have it laboriously
5032 		 * explained to them, and who can't even concisely describe
5033 		 * the conditions under which one would be forced to resort to
5034 		 * this technique.  Needless to say, those conditions are
5035 		 * found here -- and probably only here.  Is this the only use
5036 		 * of this infamous trick in shipping, production code?  If it
5037 		 * isn't, it probably should be...
5038 		 */
5039 		if (minor != -1) {
5040 			uintptr_t maddr = dtrace_loadptr(daddr +
5041 			    offsetof(struct dev_info, devi_minor));
5042 
5043 			uintptr_t next = offsetof(struct ddi_minor_data, next);
5044 			uintptr_t name = offsetof(struct ddi_minor_data,
5045 			    d_minor) + offsetof(struct ddi_minor, name);
5046 			uintptr_t dev = offsetof(struct ddi_minor_data,
5047 			    d_minor) + offsetof(struct ddi_minor, dev);
5048 			uintptr_t scout;
5049 
5050 			if (maddr != NULL)
5051 				scout = dtrace_loadptr(maddr + next);
5052 
5053 			while (maddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
5054 				uint64_t m;
5055 #ifdef _LP64
5056 				m = dtrace_load64(maddr + dev) & MAXMIN64;
5057 #else
5058 				m = dtrace_load32(maddr + dev) & MAXMIN;
5059 #endif
5060 				if (m != minor) {
5061 					maddr = dtrace_loadptr(maddr + next);
5062 
5063 					if (scout == NULL)
5064 						continue;
5065 
5066 					scout = dtrace_loadptr(scout + next);
5067 
5068 					if (scout == NULL)
5069 						continue;
5070 
5071 					scout = dtrace_loadptr(scout + next);
5072 
5073 					if (scout == NULL)
5074 						continue;
5075 
5076 					if (scout == maddr) {
5077 						*flags |= CPU_DTRACE_ILLOP;
5078 						break;
5079 					}
5080 
5081 					continue;
5082 				}
5083 
5084 				/*
5085 				 * We have the minor data.  Now we need to
5086 				 * copy the minor's name into the end of the
5087 				 * pathname.
5088 				 */
5089 				s = (char *)dtrace_loadptr(maddr + name);
5090 				len = dtrace_strlen(s, size);
5091 
5092 				if (*flags & CPU_DTRACE_FAULT)
5093 					break;
5094 
5095 				if (len != 0) {
5096 					if ((end -= (len + 1)) < start)
5097 						break;
5098 
5099 					*end = ':';
5100 				}
5101 
5102 				for (i = 1; i <= len; i++)
5103 					end[i] = dtrace_load8((uintptr_t)s++);
5104 				break;
5105 			}
5106 		}
5107 
5108 		while (daddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
5109 			ddi_node_state_t devi_state;
5110 
5111 			devi_state = dtrace_load32(daddr +
5112 			    offsetof(struct dev_info, devi_node_state));
5113 
5114 			if (*flags & CPU_DTRACE_FAULT)
5115 				break;
5116 
5117 			if (devi_state >= DS_INITIALIZED) {
5118 				s = (char *)dtrace_loadptr(daddr +
5119 				    offsetof(struct dev_info, devi_addr));
5120 				len = dtrace_strlen(s, size);
5121 
5122 				if (*flags & CPU_DTRACE_FAULT)
5123 					break;
5124 
5125 				if (len != 0) {
5126 					if ((end -= (len + 1)) < start)
5127 						break;
5128 
5129 					*end = '@';
5130 				}
5131 
5132 				for (i = 1; i <= len; i++)
5133 					end[i] = dtrace_load8((uintptr_t)s++);
5134 			}
5135 
5136 			/*
5137 			 * Now for the node name...
5138 			 */
5139 			s = (char *)dtrace_loadptr(daddr +
5140 			    offsetof(struct dev_info, devi_node_name));
5141 
5142 			daddr = dtrace_loadptr(daddr +
5143 			    offsetof(struct dev_info, devi_parent));
5144 
5145 			/*
5146 			 * If our parent is NULL (that is, if we're the root
5147 			 * node), we're going to use the special path
5148 			 * "devices".
5149 			 */
5150 			if (daddr == 0)
5151 				s = "devices";
5152 
5153 			len = dtrace_strlen(s, size);
5154 			if (*flags & CPU_DTRACE_FAULT)
5155 				break;
5156 
5157 			if ((end -= (len + 1)) < start)
5158 				break;
5159 
5160 			for (i = 1; i <= len; i++)
5161 				end[i] = dtrace_load8((uintptr_t)s++);
5162 			*end = '/';
5163 
5164 			if (depth++ > dtrace_devdepth_max) {
5165 				*flags |= CPU_DTRACE_ILLOP;
5166 				break;
5167 			}
5168 		}
5169 
5170 		if (end < start)
5171 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5172 
5173 		if (daddr == 0) {
5174 			regs[rd] = (uintptr_t)end;
5175 			mstate->dtms_scratch_ptr += size;
5176 		}
5177 
5178 		break;
5179 	}
5180 #endif
5181 
5182 	case DIF_SUBR_STRJOIN: {
5183 		char *d = (char *)mstate->dtms_scratch_ptr;
5184 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
5185 		uintptr_t s1 = tupregs[0].dttk_value;
5186 		uintptr_t s2 = tupregs[1].dttk_value;
5187 		int i = 0;
5188 
5189 		if (!dtrace_strcanload(s1, size, mstate, vstate) ||
5190 		    !dtrace_strcanload(s2, size, mstate, vstate)) {
5191 			regs[rd] = 0;
5192 			break;
5193 		}
5194 
5195 		if (!DTRACE_INSCRATCH(mstate, size)) {
5196 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5197 			regs[rd] = 0;
5198 			break;
5199 		}
5200 
5201 		for (;;) {
5202 			if (i >= size) {
5203 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5204 				regs[rd] = 0;
5205 				break;
5206 			}
5207 
5208 			if ((d[i++] = dtrace_load8(s1++)) == '\0') {
5209 				i--;
5210 				break;
5211 			}
5212 		}
5213 
5214 		for (;;) {
5215 			if (i >= size) {
5216 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5217 				regs[rd] = 0;
5218 				break;
5219 			}
5220 
5221 			if ((d[i++] = dtrace_load8(s2++)) == '\0')
5222 				break;
5223 		}
5224 
5225 		if (i < size) {
5226 			mstate->dtms_scratch_ptr += i;
5227 			regs[rd] = (uintptr_t)d;
5228 		}
5229 
5230 		break;
5231 	}
5232 
5233 	case DIF_SUBR_STRTOLL: {
5234 		uintptr_t s = tupregs[0].dttk_value;
5235 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
5236 		int base = 10;
5237 
5238 		if (nargs > 1) {
5239 			if ((base = tupregs[1].dttk_value) <= 1 ||
5240 			    base > ('z' - 'a' + 1) + ('9' - '0' + 1)) {
5241 				*flags |= CPU_DTRACE_ILLOP;
5242 				break;
5243 			}
5244 		}
5245 
5246 		if (!dtrace_strcanload(s, size, mstate, vstate)) {
5247 			regs[rd] = INT64_MIN;
5248 			break;
5249 		}
5250 
5251 		regs[rd] = dtrace_strtoll((char *)s, base, size);
5252 		break;
5253 	}
5254 
5255 	case DIF_SUBR_LLTOSTR: {
5256 		int64_t i = (int64_t)tupregs[0].dttk_value;
5257 		uint64_t val, digit;
5258 		uint64_t size = 65;	/* enough room for 2^64 in binary */
5259 		char *end = (char *)mstate->dtms_scratch_ptr + size - 1;
5260 		int base = 10;
5261 
5262 		if (nargs > 1) {
5263 			if ((base = tupregs[1].dttk_value) <= 1 ||
5264 			    base > ('z' - 'a' + 1) + ('9' - '0' + 1)) {
5265 				*flags |= CPU_DTRACE_ILLOP;
5266 				break;
5267 			}
5268 		}
5269 
5270 		val = (base == 10 && i < 0) ? i * -1 : i;
5271 
5272 		if (!DTRACE_INSCRATCH(mstate, size)) {
5273 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5274 			regs[rd] = 0;
5275 			break;
5276 		}
5277 
5278 		for (*end-- = '\0'; val; val /= base) {
5279 			if ((digit = val % base) <= '9' - '0') {
5280 				*end-- = '0' + digit;
5281 			} else {
5282 				*end-- = 'a' + (digit - ('9' - '0') - 1);
5283 			}
5284 		}
5285 
5286 		if (i == 0 && base == 16)
5287 			*end-- = '0';
5288 
5289 		if (base == 16)
5290 			*end-- = 'x';
5291 
5292 		if (i == 0 || base == 8 || base == 16)
5293 			*end-- = '0';
5294 
5295 		if (i < 0 && base == 10)
5296 			*end-- = '-';
5297 
5298 		regs[rd] = (uintptr_t)end + 1;
5299 		mstate->dtms_scratch_ptr += size;
5300 		break;
5301 	}
5302 
5303 	case DIF_SUBR_HTONS:
5304 	case DIF_SUBR_NTOHS:
5305 #if BYTE_ORDER == BIG_ENDIAN
5306 		regs[rd] = (uint16_t)tupregs[0].dttk_value;
5307 #else
5308 		regs[rd] = DT_BSWAP_16((uint16_t)tupregs[0].dttk_value);
5309 #endif
5310 		break;
5311 
5312 
5313 	case DIF_SUBR_HTONL:
5314 	case DIF_SUBR_NTOHL:
5315 #if BYTE_ORDER == BIG_ENDIAN
5316 		regs[rd] = (uint32_t)tupregs[0].dttk_value;
5317 #else
5318 		regs[rd] = DT_BSWAP_32((uint32_t)tupregs[0].dttk_value);
5319 #endif
5320 		break;
5321 
5322 
5323 	case DIF_SUBR_HTONLL:
5324 	case DIF_SUBR_NTOHLL:
5325 #if BYTE_ORDER == BIG_ENDIAN
5326 		regs[rd] = (uint64_t)tupregs[0].dttk_value;
5327 #else
5328 		regs[rd] = DT_BSWAP_64((uint64_t)tupregs[0].dttk_value);
5329 #endif
5330 		break;
5331 
5332 
5333 	case DIF_SUBR_DIRNAME:
5334 	case DIF_SUBR_BASENAME: {
5335 		char *dest = (char *)mstate->dtms_scratch_ptr;
5336 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
5337 		uintptr_t src = tupregs[0].dttk_value;
5338 		int i, j, len = dtrace_strlen((char *)src, size);
5339 		int lastbase = -1, firstbase = -1, lastdir = -1;
5340 		int start, end;
5341 
5342 		if (!dtrace_canload(src, len + 1, mstate, vstate)) {
5343 			regs[rd] = 0;
5344 			break;
5345 		}
5346 
5347 		if (!DTRACE_INSCRATCH(mstate, size)) {
5348 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5349 			regs[rd] = 0;
5350 			break;
5351 		}
5352 
5353 		/*
5354 		 * The basename and dirname for a zero-length string is
5355 		 * defined to be "."
5356 		 */
5357 		if (len == 0) {
5358 			len = 1;
5359 			src = (uintptr_t)".";
5360 		}
5361 
5362 		/*
5363 		 * Start from the back of the string, moving back toward the
5364 		 * front until we see a character that isn't a slash.  That
5365 		 * character is the last character in the basename.
5366 		 */
5367 		for (i = len - 1; i >= 0; i--) {
5368 			if (dtrace_load8(src + i) != '/')
5369 				break;
5370 		}
5371 
5372 		if (i >= 0)
5373 			lastbase = i;
5374 
5375 		/*
5376 		 * Starting from the last character in the basename, move
5377 		 * towards the front until we find a slash.  The character
5378 		 * that we processed immediately before that is the first
5379 		 * character in the basename.
5380 		 */
5381 		for (; i >= 0; i--) {
5382 			if (dtrace_load8(src + i) == '/')
5383 				break;
5384 		}
5385 
5386 		if (i >= 0)
5387 			firstbase = i + 1;
5388 
5389 		/*
5390 		 * Now keep going until we find a non-slash character.  That
5391 		 * character is the last character in the dirname.
5392 		 */
5393 		for (; i >= 0; i--) {
5394 			if (dtrace_load8(src + i) != '/')
5395 				break;
5396 		}
5397 
5398 		if (i >= 0)
5399 			lastdir = i;
5400 
5401 		ASSERT(!(lastbase == -1 && firstbase != -1));
5402 		ASSERT(!(firstbase == -1 && lastdir != -1));
5403 
5404 		if (lastbase == -1) {
5405 			/*
5406 			 * We didn't find a non-slash character.  We know that
5407 			 * the length is non-zero, so the whole string must be
5408 			 * slashes.  In either the dirname or the basename
5409 			 * case, we return '/'.
5410 			 */
5411 			ASSERT(firstbase == -1);
5412 			firstbase = lastbase = lastdir = 0;
5413 		}
5414 
5415 		if (firstbase == -1) {
5416 			/*
5417 			 * The entire string consists only of a basename
5418 			 * component.  If we're looking for dirname, we need
5419 			 * to change our string to be just "."; if we're
5420 			 * looking for a basename, we'll just set the first
5421 			 * character of the basename to be 0.
5422 			 */
5423 			if (subr == DIF_SUBR_DIRNAME) {
5424 				ASSERT(lastdir == -1);
5425 				src = (uintptr_t)".";
5426 				lastdir = 0;
5427 			} else {
5428 				firstbase = 0;
5429 			}
5430 		}
5431 
5432 		if (subr == DIF_SUBR_DIRNAME) {
5433 			if (lastdir == -1) {
5434 				/*
5435 				 * We know that we have a slash in the name --
5436 				 * or lastdir would be set to 0, above.  And
5437 				 * because lastdir is -1, we know that this
5438 				 * slash must be the first character.  (That
5439 				 * is, the full string must be of the form
5440 				 * "/basename".)  In this case, the last
5441 				 * character of the directory name is 0.
5442 				 */
5443 				lastdir = 0;
5444 			}
5445 
5446 			start = 0;
5447 			end = lastdir;
5448 		} else {
5449 			ASSERT(subr == DIF_SUBR_BASENAME);
5450 			ASSERT(firstbase != -1 && lastbase != -1);
5451 			start = firstbase;
5452 			end = lastbase;
5453 		}
5454 
5455 		for (i = start, j = 0; i <= end && j < size - 1; i++, j++)
5456 			dest[j] = dtrace_load8(src + i);
5457 
5458 		dest[j] = '\0';
5459 		regs[rd] = (uintptr_t)dest;
5460 		mstate->dtms_scratch_ptr += size;
5461 		break;
5462 	}
5463 
5464 	case DIF_SUBR_GETF: {
5465 		uintptr_t fd = tupregs[0].dttk_value;
5466 		struct filedesc *fdp;
5467 		file_t *fp;
5468 
5469 		if (!dtrace_priv_proc(state)) {
5470 			regs[rd] = 0;
5471 			break;
5472 		}
5473 		fdp = curproc->p_fd;
5474 		FILEDESC_SLOCK(fdp);
5475 		fp = fget_locked(fdp, fd);
5476 		mstate->dtms_getf = fp;
5477 		regs[rd] = (uintptr_t)fp;
5478 		FILEDESC_SUNLOCK(fdp);
5479 		break;
5480 	}
5481 
5482 	case DIF_SUBR_CLEANPATH: {
5483 		char *dest = (char *)mstate->dtms_scratch_ptr, c;
5484 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
5485 		uintptr_t src = tupregs[0].dttk_value;
5486 		int i = 0, j = 0;
5487 #ifdef illumos
5488 		zone_t *z;
5489 #endif
5490 
5491 		if (!dtrace_strcanload(src, size, mstate, vstate)) {
5492 			regs[rd] = 0;
5493 			break;
5494 		}
5495 
5496 		if (!DTRACE_INSCRATCH(mstate, size)) {
5497 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5498 			regs[rd] = 0;
5499 			break;
5500 		}
5501 
5502 		/*
5503 		 * Move forward, loading each character.
5504 		 */
5505 		do {
5506 			c = dtrace_load8(src + i++);
5507 next:
5508 			if (j + 5 >= size)	/* 5 = strlen("/..c\0") */
5509 				break;
5510 
5511 			if (c != '/') {
5512 				dest[j++] = c;
5513 				continue;
5514 			}
5515 
5516 			c = dtrace_load8(src + i++);
5517 
5518 			if (c == '/') {
5519 				/*
5520 				 * We have two slashes -- we can just advance
5521 				 * to the next character.
5522 				 */
5523 				goto next;
5524 			}
5525 
5526 			if (c != '.') {
5527 				/*
5528 				 * This is not "." and it's not ".." -- we can
5529 				 * just store the "/" and this character and
5530 				 * drive on.
5531 				 */
5532 				dest[j++] = '/';
5533 				dest[j++] = c;
5534 				continue;
5535 			}
5536 
5537 			c = dtrace_load8(src + i++);
5538 
5539 			if (c == '/') {
5540 				/*
5541 				 * This is a "/./" component.  We're not going
5542 				 * to store anything in the destination buffer;
5543 				 * we're just going to go to the next component.
5544 				 */
5545 				goto next;
5546 			}
5547 
5548 			if (c != '.') {
5549 				/*
5550 				 * This is not ".." -- we can just store the
5551 				 * "/." and this character and continue
5552 				 * processing.
5553 				 */
5554 				dest[j++] = '/';
5555 				dest[j++] = '.';
5556 				dest[j++] = c;
5557 				continue;
5558 			}
5559 
5560 			c = dtrace_load8(src + i++);
5561 
5562 			if (c != '/' && c != '\0') {
5563 				/*
5564 				 * This is not ".." -- it's "..[mumble]".
5565 				 * We'll store the "/.." and this character
5566 				 * and continue processing.
5567 				 */
5568 				dest[j++] = '/';
5569 				dest[j++] = '.';
5570 				dest[j++] = '.';
5571 				dest[j++] = c;
5572 				continue;
5573 			}
5574 
5575 			/*
5576 			 * This is "/../" or "/..\0".  We need to back up
5577 			 * our destination pointer until we find a "/".
5578 			 */
5579 			i--;
5580 			while (j != 0 && dest[--j] != '/')
5581 				continue;
5582 
5583 			if (c == '\0')
5584 				dest[++j] = '/';
5585 		} while (c != '\0');
5586 
5587 		dest[j] = '\0';
5588 
5589 #ifdef illumos
5590 		if (mstate->dtms_getf != NULL &&
5591 		    !(mstate->dtms_access & DTRACE_ACCESS_KERNEL) &&
5592 		    (z = state->dts_cred.dcr_cred->cr_zone) != kcred->cr_zone) {
5593 			/*
5594 			 * If we've done a getf() as a part of this ECB and we
5595 			 * don't have kernel access (and we're not in the global
5596 			 * zone), check if the path we cleaned up begins with
5597 			 * the zone's root path, and trim it off if so.  Note
5598 			 * that this is an output cleanliness issue, not a
5599 			 * security issue: knowing one's zone root path does
5600 			 * not enable privilege escalation.
5601 			 */
5602 			if (strstr(dest, z->zone_rootpath) == dest)
5603 				dest += strlen(z->zone_rootpath) - 1;
5604 		}
5605 #endif
5606 
5607 		regs[rd] = (uintptr_t)dest;
5608 		mstate->dtms_scratch_ptr += size;
5609 		break;
5610 	}
5611 
5612 	case DIF_SUBR_INET_NTOA:
5613 	case DIF_SUBR_INET_NTOA6:
5614 	case DIF_SUBR_INET_NTOP: {
5615 		size_t size;
5616 		int af, argi, i;
5617 		char *base, *end;
5618 
5619 		if (subr == DIF_SUBR_INET_NTOP) {
5620 			af = (int)tupregs[0].dttk_value;
5621 			argi = 1;
5622 		} else {
5623 			af = subr == DIF_SUBR_INET_NTOA ? AF_INET: AF_INET6;
5624 			argi = 0;
5625 		}
5626 
5627 		if (af == AF_INET) {
5628 			ipaddr_t ip4;
5629 			uint8_t *ptr8, val;
5630 
5631 			/*
5632 			 * Safely load the IPv4 address.
5633 			 */
5634 			ip4 = dtrace_load32(tupregs[argi].dttk_value);
5635 
5636 			/*
5637 			 * Check an IPv4 string will fit in scratch.
5638 			 */
5639 			size = INET_ADDRSTRLEN;
5640 			if (!DTRACE_INSCRATCH(mstate, size)) {
5641 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5642 				regs[rd] = 0;
5643 				break;
5644 			}
5645 			base = (char *)mstate->dtms_scratch_ptr;
5646 			end = (char *)mstate->dtms_scratch_ptr + size - 1;
5647 
5648 			/*
5649 			 * Stringify as a dotted decimal quad.
5650 			 */
5651 			*end-- = '\0';
5652 			ptr8 = (uint8_t *)&ip4;
5653 			for (i = 3; i >= 0; i--) {
5654 				val = ptr8[i];
5655 
5656 				if (val == 0) {
5657 					*end-- = '0';
5658 				} else {
5659 					for (; val; val /= 10) {
5660 						*end-- = '0' + (val % 10);
5661 					}
5662 				}
5663 
5664 				if (i > 0)
5665 					*end-- = '.';
5666 			}
5667 			ASSERT(end + 1 >= base);
5668 
5669 		} else if (af == AF_INET6) {
5670 			struct in6_addr ip6;
5671 			int firstzero, tryzero, numzero, v6end;
5672 			uint16_t val;
5673 			const char digits[] = "0123456789abcdef";
5674 
5675 			/*
5676 			 * Stringify using RFC 1884 convention 2 - 16 bit
5677 			 * hexadecimal values with a zero-run compression.
5678 			 * Lower case hexadecimal digits are used.
5679 			 * 	eg, fe80::214:4fff:fe0b:76c8.
5680 			 * The IPv4 embedded form is returned for inet_ntop,
5681 			 * just the IPv4 string is returned for inet_ntoa6.
5682 			 */
5683 
5684 			/*
5685 			 * Safely load the IPv6 address.
5686 			 */
5687 			dtrace_bcopy(
5688 			    (void *)(uintptr_t)tupregs[argi].dttk_value,
5689 			    (void *)(uintptr_t)&ip6, sizeof (struct in6_addr));
5690 
5691 			/*
5692 			 * Check an IPv6 string will fit in scratch.
5693 			 */
5694 			size = INET6_ADDRSTRLEN;
5695 			if (!DTRACE_INSCRATCH(mstate, size)) {
5696 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5697 				regs[rd] = 0;
5698 				break;
5699 			}
5700 			base = (char *)mstate->dtms_scratch_ptr;
5701 			end = (char *)mstate->dtms_scratch_ptr + size - 1;
5702 			*end-- = '\0';
5703 
5704 			/*
5705 			 * Find the longest run of 16 bit zero values
5706 			 * for the single allowed zero compression - "::".
5707 			 */
5708 			firstzero = -1;
5709 			tryzero = -1;
5710 			numzero = 1;
5711 			for (i = 0; i < sizeof (struct in6_addr); i++) {
5712 #ifdef illumos
5713 				if (ip6._S6_un._S6_u8[i] == 0 &&
5714 #else
5715 				if (ip6.__u6_addr.__u6_addr8[i] == 0 &&
5716 #endif
5717 				    tryzero == -1 && i % 2 == 0) {
5718 					tryzero = i;
5719 					continue;
5720 				}
5721 
5722 				if (tryzero != -1 &&
5723 #ifdef illumos
5724 				    (ip6._S6_un._S6_u8[i] != 0 ||
5725 #else
5726 				    (ip6.__u6_addr.__u6_addr8[i] != 0 ||
5727 #endif
5728 				    i == sizeof (struct in6_addr) - 1)) {
5729 
5730 					if (i - tryzero <= numzero) {
5731 						tryzero = -1;
5732 						continue;
5733 					}
5734 
5735 					firstzero = tryzero;
5736 					numzero = i - i % 2 - tryzero;
5737 					tryzero = -1;
5738 
5739 #ifdef illumos
5740 					if (ip6._S6_un._S6_u8[i] == 0 &&
5741 #else
5742 					if (ip6.__u6_addr.__u6_addr8[i] == 0 &&
5743 #endif
5744 					    i == sizeof (struct in6_addr) - 1)
5745 						numzero += 2;
5746 				}
5747 			}
5748 			ASSERT(firstzero + numzero <= sizeof (struct in6_addr));
5749 
5750 			/*
5751 			 * Check for an IPv4 embedded address.
5752 			 */
5753 			v6end = sizeof (struct in6_addr) - 2;
5754 			if (IN6_IS_ADDR_V4MAPPED(&ip6) ||
5755 			    IN6_IS_ADDR_V4COMPAT(&ip6)) {
5756 				for (i = sizeof (struct in6_addr) - 1;
5757 				    i >= DTRACE_V4MAPPED_OFFSET; i--) {
5758 					ASSERT(end >= base);
5759 
5760 #ifdef illumos
5761 					val = ip6._S6_un._S6_u8[i];
5762 #else
5763 					val = ip6.__u6_addr.__u6_addr8[i];
5764 #endif
5765 
5766 					if (val == 0) {
5767 						*end-- = '0';
5768 					} else {
5769 						for (; val; val /= 10) {
5770 							*end-- = '0' + val % 10;
5771 						}
5772 					}
5773 
5774 					if (i > DTRACE_V4MAPPED_OFFSET)
5775 						*end-- = '.';
5776 				}
5777 
5778 				if (subr == DIF_SUBR_INET_NTOA6)
5779 					goto inetout;
5780 
5781 				/*
5782 				 * Set v6end to skip the IPv4 address that
5783 				 * we have already stringified.
5784 				 */
5785 				v6end = 10;
5786 			}
5787 
5788 			/*
5789 			 * Build the IPv6 string by working through the
5790 			 * address in reverse.
5791 			 */
5792 			for (i = v6end; i >= 0; i -= 2) {
5793 				ASSERT(end >= base);
5794 
5795 				if (i == firstzero + numzero - 2) {
5796 					*end-- = ':';
5797 					*end-- = ':';
5798 					i -= numzero - 2;
5799 					continue;
5800 				}
5801 
5802 				if (i < 14 && i != firstzero - 2)
5803 					*end-- = ':';
5804 
5805 #ifdef illumos
5806 				val = (ip6._S6_un._S6_u8[i] << 8) +
5807 				    ip6._S6_un._S6_u8[i + 1];
5808 #else
5809 				val = (ip6.__u6_addr.__u6_addr8[i] << 8) +
5810 				    ip6.__u6_addr.__u6_addr8[i + 1];
5811 #endif
5812 
5813 				if (val == 0) {
5814 					*end-- = '0';
5815 				} else {
5816 					for (; val; val /= 16) {
5817 						*end-- = digits[val % 16];
5818 					}
5819 				}
5820 			}
5821 			ASSERT(end + 1 >= base);
5822 
5823 		} else {
5824 			/*
5825 			 * The user didn't use AH_INET or AH_INET6.
5826 			 */
5827 			DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
5828 			regs[rd] = 0;
5829 			break;
5830 		}
5831 
5832 inetout:	regs[rd] = (uintptr_t)end + 1;
5833 		mstate->dtms_scratch_ptr += size;
5834 		break;
5835 	}
5836 
5837 	case DIF_SUBR_MEMREF: {
5838 		uintptr_t size = 2 * sizeof(uintptr_t);
5839 		uintptr_t *memref = (uintptr_t *) P2ROUNDUP(mstate->dtms_scratch_ptr, sizeof(uintptr_t));
5840 		size_t scratch_size = ((uintptr_t) memref - mstate->dtms_scratch_ptr) + size;
5841 
5842 		/* address and length */
5843 		memref[0] = tupregs[0].dttk_value;
5844 		memref[1] = tupregs[1].dttk_value;
5845 
5846 		regs[rd] = (uintptr_t) memref;
5847 		mstate->dtms_scratch_ptr += scratch_size;
5848 		break;
5849 	}
5850 
5851 #ifndef illumos
5852 	case DIF_SUBR_MEMSTR: {
5853 		char *str = (char *)mstate->dtms_scratch_ptr;
5854 		uintptr_t mem = tupregs[0].dttk_value;
5855 		char c = tupregs[1].dttk_value;
5856 		size_t size = tupregs[2].dttk_value;
5857 		uint8_t n;
5858 		int i;
5859 
5860 		regs[rd] = 0;
5861 
5862 		if (size == 0)
5863 			break;
5864 
5865 		if (!dtrace_canload(mem, size - 1, mstate, vstate))
5866 			break;
5867 
5868 		if (!DTRACE_INSCRATCH(mstate, size)) {
5869 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5870 			break;
5871 		}
5872 
5873 		if (dtrace_memstr_max != 0 && size > dtrace_memstr_max) {
5874 			*flags |= CPU_DTRACE_ILLOP;
5875 			break;
5876 		}
5877 
5878 		for (i = 0; i < size - 1; i++) {
5879 			n = dtrace_load8(mem++);
5880 			str[i] = (n == 0) ? c : n;
5881 		}
5882 		str[size - 1] = 0;
5883 
5884 		regs[rd] = (uintptr_t)str;
5885 		mstate->dtms_scratch_ptr += size;
5886 		break;
5887 	}
5888 #endif
5889 
5890 	case DIF_SUBR_TYPEREF: {
5891 		uintptr_t size = 4 * sizeof(uintptr_t);
5892 		uintptr_t *typeref = (uintptr_t *) P2ROUNDUP(mstate->dtms_scratch_ptr, sizeof(uintptr_t));
5893 		size_t scratch_size = ((uintptr_t) typeref - mstate->dtms_scratch_ptr) + size;
5894 
5895 		/* address, num_elements, type_str, type_len */
5896 		typeref[0] = tupregs[0].dttk_value;
5897 		typeref[1] = tupregs[1].dttk_value;
5898 		typeref[2] = tupregs[2].dttk_value;
5899 		typeref[3] = tupregs[3].dttk_value;
5900 
5901 		regs[rd] = (uintptr_t) typeref;
5902 		mstate->dtms_scratch_ptr += scratch_size;
5903 		break;
5904 	}
5905 	}
5906 }
5907 
5908 /*
5909  * Emulate the execution of DTrace IR instructions specified by the given
5910  * DIF object.  This function is deliberately void of assertions as all of
5911  * the necessary checks are handled by a call to dtrace_difo_validate().
5912  */
5913 static uint64_t
5914 dtrace_dif_emulate(dtrace_difo_t *difo, dtrace_mstate_t *mstate,
5915     dtrace_vstate_t *vstate, dtrace_state_t *state)
5916 {
5917 	const dif_instr_t *text = difo->dtdo_buf;
5918 	const uint_t textlen = difo->dtdo_len;
5919 	const char *strtab = difo->dtdo_strtab;
5920 	const uint64_t *inttab = difo->dtdo_inttab;
5921 
5922 	uint64_t rval = 0;
5923 	dtrace_statvar_t *svar;
5924 	dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
5925 	dtrace_difv_t *v;
5926 	volatile uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
5927 	volatile uintptr_t *illval = &cpu_core[curcpu].cpuc_dtrace_illval;
5928 
5929 	dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
5930 	uint64_t regs[DIF_DIR_NREGS];
5931 	uint64_t *tmp;
5932 
5933 	uint8_t cc_n = 0, cc_z = 0, cc_v = 0, cc_c = 0;
5934 	int64_t cc_r;
5935 	uint_t pc = 0, id, opc = 0;
5936 	uint8_t ttop = 0;
5937 	dif_instr_t instr;
5938 	uint_t r1, r2, rd;
5939 
5940 	/*
5941 	 * We stash the current DIF object into the machine state: we need it
5942 	 * for subsequent access checking.
5943 	 */
5944 	mstate->dtms_difo = difo;
5945 
5946 	regs[DIF_REG_R0] = 0; 		/* %r0 is fixed at zero */
5947 
5948 	while (pc < textlen && !(*flags & CPU_DTRACE_FAULT)) {
5949 		opc = pc;
5950 
5951 		instr = text[pc++];
5952 		r1 = DIF_INSTR_R1(instr);
5953 		r2 = DIF_INSTR_R2(instr);
5954 		rd = DIF_INSTR_RD(instr);
5955 
5956 		switch (DIF_INSTR_OP(instr)) {
5957 		case DIF_OP_OR:
5958 			regs[rd] = regs[r1] | regs[r2];
5959 			break;
5960 		case DIF_OP_XOR:
5961 			regs[rd] = regs[r1] ^ regs[r2];
5962 			break;
5963 		case DIF_OP_AND:
5964 			regs[rd] = regs[r1] & regs[r2];
5965 			break;
5966 		case DIF_OP_SLL:
5967 			regs[rd] = regs[r1] << regs[r2];
5968 			break;
5969 		case DIF_OP_SRL:
5970 			regs[rd] = regs[r1] >> regs[r2];
5971 			break;
5972 		case DIF_OP_SUB:
5973 			regs[rd] = regs[r1] - regs[r2];
5974 			break;
5975 		case DIF_OP_ADD:
5976 			regs[rd] = regs[r1] + regs[r2];
5977 			break;
5978 		case DIF_OP_MUL:
5979 			regs[rd] = regs[r1] * regs[r2];
5980 			break;
5981 		case DIF_OP_SDIV:
5982 			if (regs[r2] == 0) {
5983 				regs[rd] = 0;
5984 				*flags |= CPU_DTRACE_DIVZERO;
5985 			} else {
5986 				regs[rd] = (int64_t)regs[r1] /
5987 				    (int64_t)regs[r2];
5988 			}
5989 			break;
5990 
5991 		case DIF_OP_UDIV:
5992 			if (regs[r2] == 0) {
5993 				regs[rd] = 0;
5994 				*flags |= CPU_DTRACE_DIVZERO;
5995 			} else {
5996 				regs[rd] = regs[r1] / regs[r2];
5997 			}
5998 			break;
5999 
6000 		case DIF_OP_SREM:
6001 			if (regs[r2] == 0) {
6002 				regs[rd] = 0;
6003 				*flags |= CPU_DTRACE_DIVZERO;
6004 			} else {
6005 				regs[rd] = (int64_t)regs[r1] %
6006 				    (int64_t)regs[r2];
6007 			}
6008 			break;
6009 
6010 		case DIF_OP_UREM:
6011 			if (regs[r2] == 0) {
6012 				regs[rd] = 0;
6013 				*flags |= CPU_DTRACE_DIVZERO;
6014 			} else {
6015 				regs[rd] = regs[r1] % regs[r2];
6016 			}
6017 			break;
6018 
6019 		case DIF_OP_NOT:
6020 			regs[rd] = ~regs[r1];
6021 			break;
6022 		case DIF_OP_MOV:
6023 			regs[rd] = regs[r1];
6024 			break;
6025 		case DIF_OP_CMP:
6026 			cc_r = regs[r1] - regs[r2];
6027 			cc_n = cc_r < 0;
6028 			cc_z = cc_r == 0;
6029 			cc_v = 0;
6030 			cc_c = regs[r1] < regs[r2];
6031 			break;
6032 		case DIF_OP_TST:
6033 			cc_n = cc_v = cc_c = 0;
6034 			cc_z = regs[r1] == 0;
6035 			break;
6036 		case DIF_OP_BA:
6037 			pc = DIF_INSTR_LABEL(instr);
6038 			break;
6039 		case DIF_OP_BE:
6040 			if (cc_z)
6041 				pc = DIF_INSTR_LABEL(instr);
6042 			break;
6043 		case DIF_OP_BNE:
6044 			if (cc_z == 0)
6045 				pc = DIF_INSTR_LABEL(instr);
6046 			break;
6047 		case DIF_OP_BG:
6048 			if ((cc_z | (cc_n ^ cc_v)) == 0)
6049 				pc = DIF_INSTR_LABEL(instr);
6050 			break;
6051 		case DIF_OP_BGU:
6052 			if ((cc_c | cc_z) == 0)
6053 				pc = DIF_INSTR_LABEL(instr);
6054 			break;
6055 		case DIF_OP_BGE:
6056 			if ((cc_n ^ cc_v) == 0)
6057 				pc = DIF_INSTR_LABEL(instr);
6058 			break;
6059 		case DIF_OP_BGEU:
6060 			if (cc_c == 0)
6061 				pc = DIF_INSTR_LABEL(instr);
6062 			break;
6063 		case DIF_OP_BL:
6064 			if (cc_n ^ cc_v)
6065 				pc = DIF_INSTR_LABEL(instr);
6066 			break;
6067 		case DIF_OP_BLU:
6068 			if (cc_c)
6069 				pc = DIF_INSTR_LABEL(instr);
6070 			break;
6071 		case DIF_OP_BLE:
6072 			if (cc_z | (cc_n ^ cc_v))
6073 				pc = DIF_INSTR_LABEL(instr);
6074 			break;
6075 		case DIF_OP_BLEU:
6076 			if (cc_c | cc_z)
6077 				pc = DIF_INSTR_LABEL(instr);
6078 			break;
6079 		case DIF_OP_RLDSB:
6080 			if (!dtrace_canload(regs[r1], 1, mstate, vstate))
6081 				break;
6082 			/*FALLTHROUGH*/
6083 		case DIF_OP_LDSB:
6084 			regs[rd] = (int8_t)dtrace_load8(regs[r1]);
6085 			break;
6086 		case DIF_OP_RLDSH:
6087 			if (!dtrace_canload(regs[r1], 2, mstate, vstate))
6088 				break;
6089 			/*FALLTHROUGH*/
6090 		case DIF_OP_LDSH:
6091 			regs[rd] = (int16_t)dtrace_load16(regs[r1]);
6092 			break;
6093 		case DIF_OP_RLDSW:
6094 			if (!dtrace_canload(regs[r1], 4, mstate, vstate))
6095 				break;
6096 			/*FALLTHROUGH*/
6097 		case DIF_OP_LDSW:
6098 			regs[rd] = (int32_t)dtrace_load32(regs[r1]);
6099 			break;
6100 		case DIF_OP_RLDUB:
6101 			if (!dtrace_canload(regs[r1], 1, mstate, vstate))
6102 				break;
6103 			/*FALLTHROUGH*/
6104 		case DIF_OP_LDUB:
6105 			regs[rd] = dtrace_load8(regs[r1]);
6106 			break;
6107 		case DIF_OP_RLDUH:
6108 			if (!dtrace_canload(regs[r1], 2, mstate, vstate))
6109 				break;
6110 			/*FALLTHROUGH*/
6111 		case DIF_OP_LDUH:
6112 			regs[rd] = dtrace_load16(regs[r1]);
6113 			break;
6114 		case DIF_OP_RLDUW:
6115 			if (!dtrace_canload(regs[r1], 4, mstate, vstate))
6116 				break;
6117 			/*FALLTHROUGH*/
6118 		case DIF_OP_LDUW:
6119 			regs[rd] = dtrace_load32(regs[r1]);
6120 			break;
6121 		case DIF_OP_RLDX:
6122 			if (!dtrace_canload(regs[r1], 8, mstate, vstate))
6123 				break;
6124 			/*FALLTHROUGH*/
6125 		case DIF_OP_LDX:
6126 			regs[rd] = dtrace_load64(regs[r1]);
6127 			break;
6128 		case DIF_OP_ULDSB:
6129 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6130 			regs[rd] = (int8_t)
6131 			    dtrace_fuword8((void *)(uintptr_t)regs[r1]);
6132 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6133 			break;
6134 		case DIF_OP_ULDSH:
6135 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6136 			regs[rd] = (int16_t)
6137 			    dtrace_fuword16((void *)(uintptr_t)regs[r1]);
6138 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6139 			break;
6140 		case DIF_OP_ULDSW:
6141 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6142 			regs[rd] = (int32_t)
6143 			    dtrace_fuword32((void *)(uintptr_t)regs[r1]);
6144 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6145 			break;
6146 		case DIF_OP_ULDUB:
6147 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6148 			regs[rd] =
6149 			    dtrace_fuword8((void *)(uintptr_t)regs[r1]);
6150 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6151 			break;
6152 		case DIF_OP_ULDUH:
6153 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6154 			regs[rd] =
6155 			    dtrace_fuword16((void *)(uintptr_t)regs[r1]);
6156 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6157 			break;
6158 		case DIF_OP_ULDUW:
6159 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6160 			regs[rd] =
6161 			    dtrace_fuword32((void *)(uintptr_t)regs[r1]);
6162 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6163 			break;
6164 		case DIF_OP_ULDX:
6165 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6166 			regs[rd] =
6167 			    dtrace_fuword64((void *)(uintptr_t)regs[r1]);
6168 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6169 			break;
6170 		case DIF_OP_RET:
6171 			rval = regs[rd];
6172 			pc = textlen;
6173 			break;
6174 		case DIF_OP_NOP:
6175 			break;
6176 		case DIF_OP_SETX:
6177 			regs[rd] = inttab[DIF_INSTR_INTEGER(instr)];
6178 			break;
6179 		case DIF_OP_SETS:
6180 			regs[rd] = (uint64_t)(uintptr_t)
6181 			    (strtab + DIF_INSTR_STRING(instr));
6182 			break;
6183 		case DIF_OP_SCMP: {
6184 			size_t sz = state->dts_options[DTRACEOPT_STRSIZE];
6185 			uintptr_t s1 = regs[r1];
6186 			uintptr_t s2 = regs[r2];
6187 
6188 			if (s1 != 0 &&
6189 			    !dtrace_strcanload(s1, sz, mstate, vstate))
6190 				break;
6191 			if (s2 != 0 &&
6192 			    !dtrace_strcanload(s2, sz, mstate, vstate))
6193 				break;
6194 
6195 			cc_r = dtrace_strncmp((char *)s1, (char *)s2, sz);
6196 
6197 			cc_n = cc_r < 0;
6198 			cc_z = cc_r == 0;
6199 			cc_v = cc_c = 0;
6200 			break;
6201 		}
6202 		case DIF_OP_LDGA:
6203 			regs[rd] = dtrace_dif_variable(mstate, state,
6204 			    r1, regs[r2]);
6205 			break;
6206 		case DIF_OP_LDGS:
6207 			id = DIF_INSTR_VAR(instr);
6208 
6209 			if (id >= DIF_VAR_OTHER_UBASE) {
6210 				uintptr_t a;
6211 
6212 				id -= DIF_VAR_OTHER_UBASE;
6213 				svar = vstate->dtvs_globals[id];
6214 				ASSERT(svar != NULL);
6215 				v = &svar->dtsv_var;
6216 
6217 				if (!(v->dtdv_type.dtdt_flags & DIF_TF_BYREF)) {
6218 					regs[rd] = svar->dtsv_data;
6219 					break;
6220 				}
6221 
6222 				a = (uintptr_t)svar->dtsv_data;
6223 
6224 				if (*(uint8_t *)a == UINT8_MAX) {
6225 					/*
6226 					 * If the 0th byte is set to UINT8_MAX
6227 					 * then this is to be treated as a
6228 					 * reference to a NULL variable.
6229 					 */
6230 					regs[rd] = 0;
6231 				} else {
6232 					regs[rd] = a + sizeof (uint64_t);
6233 				}
6234 
6235 				break;
6236 			}
6237 
6238 			regs[rd] = dtrace_dif_variable(mstate, state, id, 0);
6239 			break;
6240 
6241 		case DIF_OP_STGS:
6242 			id = DIF_INSTR_VAR(instr);
6243 
6244 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
6245 			id -= DIF_VAR_OTHER_UBASE;
6246 
6247 			svar = vstate->dtvs_globals[id];
6248 			ASSERT(svar != NULL);
6249 			v = &svar->dtsv_var;
6250 
6251 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6252 				uintptr_t a = (uintptr_t)svar->dtsv_data;
6253 
6254 				ASSERT(a != 0);
6255 				ASSERT(svar->dtsv_size != 0);
6256 
6257 				if (regs[rd] == 0) {
6258 					*(uint8_t *)a = UINT8_MAX;
6259 					break;
6260 				} else {
6261 					*(uint8_t *)a = 0;
6262 					a += sizeof (uint64_t);
6263 				}
6264 				if (!dtrace_vcanload(
6265 				    (void *)(uintptr_t)regs[rd], &v->dtdv_type,
6266 				    mstate, vstate))
6267 					break;
6268 
6269 				dtrace_vcopy((void *)(uintptr_t)regs[rd],
6270 				    (void *)a, &v->dtdv_type);
6271 				break;
6272 			}
6273 
6274 			svar->dtsv_data = regs[rd];
6275 			break;
6276 
6277 		case DIF_OP_LDTA:
6278 			/*
6279 			 * There are no DTrace built-in thread-local arrays at
6280 			 * present.  This opcode is saved for future work.
6281 			 */
6282 			*flags |= CPU_DTRACE_ILLOP;
6283 			regs[rd] = 0;
6284 			break;
6285 
6286 		case DIF_OP_LDLS:
6287 			id = DIF_INSTR_VAR(instr);
6288 
6289 			if (id < DIF_VAR_OTHER_UBASE) {
6290 				/*
6291 				 * For now, this has no meaning.
6292 				 */
6293 				regs[rd] = 0;
6294 				break;
6295 			}
6296 
6297 			id -= DIF_VAR_OTHER_UBASE;
6298 
6299 			ASSERT(id < vstate->dtvs_nlocals);
6300 			ASSERT(vstate->dtvs_locals != NULL);
6301 
6302 			svar = vstate->dtvs_locals[id];
6303 			ASSERT(svar != NULL);
6304 			v = &svar->dtsv_var;
6305 
6306 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6307 				uintptr_t a = (uintptr_t)svar->dtsv_data;
6308 				size_t sz = v->dtdv_type.dtdt_size;
6309 
6310 				sz += sizeof (uint64_t);
6311 				ASSERT(svar->dtsv_size == NCPU * sz);
6312 				a += curcpu * sz;
6313 
6314 				if (*(uint8_t *)a == UINT8_MAX) {
6315 					/*
6316 					 * If the 0th byte is set to UINT8_MAX
6317 					 * then this is to be treated as a
6318 					 * reference to a NULL variable.
6319 					 */
6320 					regs[rd] = 0;
6321 				} else {
6322 					regs[rd] = a + sizeof (uint64_t);
6323 				}
6324 
6325 				break;
6326 			}
6327 
6328 			ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
6329 			tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
6330 			regs[rd] = tmp[curcpu];
6331 			break;
6332 
6333 		case DIF_OP_STLS:
6334 			id = DIF_INSTR_VAR(instr);
6335 
6336 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
6337 			id -= DIF_VAR_OTHER_UBASE;
6338 			ASSERT(id < vstate->dtvs_nlocals);
6339 
6340 			ASSERT(vstate->dtvs_locals != NULL);
6341 			svar = vstate->dtvs_locals[id];
6342 			ASSERT(svar != NULL);
6343 			v = &svar->dtsv_var;
6344 
6345 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6346 				uintptr_t a = (uintptr_t)svar->dtsv_data;
6347 				size_t sz = v->dtdv_type.dtdt_size;
6348 
6349 				sz += sizeof (uint64_t);
6350 				ASSERT(svar->dtsv_size == NCPU * sz);
6351 				a += curcpu * sz;
6352 
6353 				if (regs[rd] == 0) {
6354 					*(uint8_t *)a = UINT8_MAX;
6355 					break;
6356 				} else {
6357 					*(uint8_t *)a = 0;
6358 					a += sizeof (uint64_t);
6359 				}
6360 
6361 				if (!dtrace_vcanload(
6362 				    (void *)(uintptr_t)regs[rd], &v->dtdv_type,
6363 				    mstate, vstate))
6364 					break;
6365 
6366 				dtrace_vcopy((void *)(uintptr_t)regs[rd],
6367 				    (void *)a, &v->dtdv_type);
6368 				break;
6369 			}
6370 
6371 			ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
6372 			tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
6373 			tmp[curcpu] = regs[rd];
6374 			break;
6375 
6376 		case DIF_OP_LDTS: {
6377 			dtrace_dynvar_t *dvar;
6378 			dtrace_key_t *key;
6379 
6380 			id = DIF_INSTR_VAR(instr);
6381 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
6382 			id -= DIF_VAR_OTHER_UBASE;
6383 			v = &vstate->dtvs_tlocals[id];
6384 
6385 			key = &tupregs[DIF_DTR_NREGS];
6386 			key[0].dttk_value = (uint64_t)id;
6387 			key[0].dttk_size = 0;
6388 			DTRACE_TLS_THRKEY(key[1].dttk_value);
6389 			key[1].dttk_size = 0;
6390 
6391 			dvar = dtrace_dynvar(dstate, 2, key,
6392 			    sizeof (uint64_t), DTRACE_DYNVAR_NOALLOC,
6393 			    mstate, vstate);
6394 
6395 			if (dvar == NULL) {
6396 				regs[rd] = 0;
6397 				break;
6398 			}
6399 
6400 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6401 				regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
6402 			} else {
6403 				regs[rd] = *((uint64_t *)dvar->dtdv_data);
6404 			}
6405 
6406 			break;
6407 		}
6408 
6409 		case DIF_OP_STTS: {
6410 			dtrace_dynvar_t *dvar;
6411 			dtrace_key_t *key;
6412 
6413 			id = DIF_INSTR_VAR(instr);
6414 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
6415 			id -= DIF_VAR_OTHER_UBASE;
6416 
6417 			key = &tupregs[DIF_DTR_NREGS];
6418 			key[0].dttk_value = (uint64_t)id;
6419 			key[0].dttk_size = 0;
6420 			DTRACE_TLS_THRKEY(key[1].dttk_value);
6421 			key[1].dttk_size = 0;
6422 			v = &vstate->dtvs_tlocals[id];
6423 
6424 			dvar = dtrace_dynvar(dstate, 2, key,
6425 			    v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
6426 			    v->dtdv_type.dtdt_size : sizeof (uint64_t),
6427 			    regs[rd] ? DTRACE_DYNVAR_ALLOC :
6428 			    DTRACE_DYNVAR_DEALLOC, mstate, vstate);
6429 
6430 			/*
6431 			 * Given that we're storing to thread-local data,
6432 			 * we need to flush our predicate cache.
6433 			 */
6434 			curthread->t_predcache = 0;
6435 
6436 			if (dvar == NULL)
6437 				break;
6438 
6439 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6440 				if (!dtrace_vcanload(
6441 				    (void *)(uintptr_t)regs[rd],
6442 				    &v->dtdv_type, mstate, vstate))
6443 					break;
6444 
6445 				dtrace_vcopy((void *)(uintptr_t)regs[rd],
6446 				    dvar->dtdv_data, &v->dtdv_type);
6447 			} else {
6448 				*((uint64_t *)dvar->dtdv_data) = regs[rd];
6449 			}
6450 
6451 			break;
6452 		}
6453 
6454 		case DIF_OP_SRA:
6455 			regs[rd] = (int64_t)regs[r1] >> regs[r2];
6456 			break;
6457 
6458 		case DIF_OP_CALL:
6459 			dtrace_dif_subr(DIF_INSTR_SUBR(instr), rd,
6460 			    regs, tupregs, ttop, mstate, state);
6461 			break;
6462 
6463 		case DIF_OP_PUSHTR:
6464 			if (ttop == DIF_DTR_NREGS) {
6465 				*flags |= CPU_DTRACE_TUPOFLOW;
6466 				break;
6467 			}
6468 
6469 			if (r1 == DIF_TYPE_STRING) {
6470 				/*
6471 				 * If this is a string type and the size is 0,
6472 				 * we'll use the system-wide default string
6473 				 * size.  Note that we are _not_ looking at
6474 				 * the value of the DTRACEOPT_STRSIZE option;
6475 				 * had this been set, we would expect to have
6476 				 * a non-zero size value in the "pushtr".
6477 				 */
6478 				tupregs[ttop].dttk_size =
6479 				    dtrace_strlen((char *)(uintptr_t)regs[rd],
6480 				    regs[r2] ? regs[r2] :
6481 				    dtrace_strsize_default) + 1;
6482 			} else {
6483 				if (regs[r2] > LONG_MAX) {
6484 					*flags |= CPU_DTRACE_ILLOP;
6485 					break;
6486 				}
6487 
6488 				tupregs[ttop].dttk_size = regs[r2];
6489 			}
6490 
6491 			tupregs[ttop++].dttk_value = regs[rd];
6492 			break;
6493 
6494 		case DIF_OP_PUSHTV:
6495 			if (ttop == DIF_DTR_NREGS) {
6496 				*flags |= CPU_DTRACE_TUPOFLOW;
6497 				break;
6498 			}
6499 
6500 			tupregs[ttop].dttk_value = regs[rd];
6501 			tupregs[ttop++].dttk_size = 0;
6502 			break;
6503 
6504 		case DIF_OP_POPTS:
6505 			if (ttop != 0)
6506 				ttop--;
6507 			break;
6508 
6509 		case DIF_OP_FLUSHTS:
6510 			ttop = 0;
6511 			break;
6512 
6513 		case DIF_OP_LDGAA:
6514 		case DIF_OP_LDTAA: {
6515 			dtrace_dynvar_t *dvar;
6516 			dtrace_key_t *key = tupregs;
6517 			uint_t nkeys = ttop;
6518 
6519 			id = DIF_INSTR_VAR(instr);
6520 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
6521 			id -= DIF_VAR_OTHER_UBASE;
6522 
6523 			key[nkeys].dttk_value = (uint64_t)id;
6524 			key[nkeys++].dttk_size = 0;
6525 
6526 			if (DIF_INSTR_OP(instr) == DIF_OP_LDTAA) {
6527 				DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
6528 				key[nkeys++].dttk_size = 0;
6529 				v = &vstate->dtvs_tlocals[id];
6530 			} else {
6531 				v = &vstate->dtvs_globals[id]->dtsv_var;
6532 			}
6533 
6534 			dvar = dtrace_dynvar(dstate, nkeys, key,
6535 			    v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
6536 			    v->dtdv_type.dtdt_size : sizeof (uint64_t),
6537 			    DTRACE_DYNVAR_NOALLOC, mstate, vstate);
6538 
6539 			if (dvar == NULL) {
6540 				regs[rd] = 0;
6541 				break;
6542 			}
6543 
6544 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6545 				regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
6546 			} else {
6547 				regs[rd] = *((uint64_t *)dvar->dtdv_data);
6548 			}
6549 
6550 			break;
6551 		}
6552 
6553 		case DIF_OP_STGAA:
6554 		case DIF_OP_STTAA: {
6555 			dtrace_dynvar_t *dvar;
6556 			dtrace_key_t *key = tupregs;
6557 			uint_t nkeys = ttop;
6558 
6559 			id = DIF_INSTR_VAR(instr);
6560 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
6561 			id -= DIF_VAR_OTHER_UBASE;
6562 
6563 			key[nkeys].dttk_value = (uint64_t)id;
6564 			key[nkeys++].dttk_size = 0;
6565 
6566 			if (DIF_INSTR_OP(instr) == DIF_OP_STTAA) {
6567 				DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
6568 				key[nkeys++].dttk_size = 0;
6569 				v = &vstate->dtvs_tlocals[id];
6570 			} else {
6571 				v = &vstate->dtvs_globals[id]->dtsv_var;
6572 			}
6573 
6574 			dvar = dtrace_dynvar(dstate, nkeys, key,
6575 			    v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
6576 			    v->dtdv_type.dtdt_size : sizeof (uint64_t),
6577 			    regs[rd] ? DTRACE_DYNVAR_ALLOC :
6578 			    DTRACE_DYNVAR_DEALLOC, mstate, vstate);
6579 
6580 			if (dvar == NULL)
6581 				break;
6582 
6583 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6584 				if (!dtrace_vcanload(
6585 				    (void *)(uintptr_t)regs[rd], &v->dtdv_type,
6586 				    mstate, vstate))
6587 					break;
6588 
6589 				dtrace_vcopy((void *)(uintptr_t)regs[rd],
6590 				    dvar->dtdv_data, &v->dtdv_type);
6591 			} else {
6592 				*((uint64_t *)dvar->dtdv_data) = regs[rd];
6593 			}
6594 
6595 			break;
6596 		}
6597 
6598 		case DIF_OP_ALLOCS: {
6599 			uintptr_t ptr = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
6600 			size_t size = ptr - mstate->dtms_scratch_ptr + regs[r1];
6601 
6602 			/*
6603 			 * Rounding up the user allocation size could have
6604 			 * overflowed large, bogus allocations (like -1ULL) to
6605 			 * 0.
6606 			 */
6607 			if (size < regs[r1] ||
6608 			    !DTRACE_INSCRATCH(mstate, size)) {
6609 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
6610 				regs[rd] = 0;
6611 				break;
6612 			}
6613 
6614 			dtrace_bzero((void *) mstate->dtms_scratch_ptr, size);
6615 			mstate->dtms_scratch_ptr += size;
6616 			regs[rd] = ptr;
6617 			break;
6618 		}
6619 
6620 		case DIF_OP_COPYS:
6621 			if (!dtrace_canstore(regs[rd], regs[r2],
6622 			    mstate, vstate)) {
6623 				*flags |= CPU_DTRACE_BADADDR;
6624 				*illval = regs[rd];
6625 				break;
6626 			}
6627 
6628 			if (!dtrace_canload(regs[r1], regs[r2], mstate, vstate))
6629 				break;
6630 
6631 			dtrace_bcopy((void *)(uintptr_t)regs[r1],
6632 			    (void *)(uintptr_t)regs[rd], (size_t)regs[r2]);
6633 			break;
6634 
6635 		case DIF_OP_STB:
6636 			if (!dtrace_canstore(regs[rd], 1, mstate, vstate)) {
6637 				*flags |= CPU_DTRACE_BADADDR;
6638 				*illval = regs[rd];
6639 				break;
6640 			}
6641 			*((uint8_t *)(uintptr_t)regs[rd]) = (uint8_t)regs[r1];
6642 			break;
6643 
6644 		case DIF_OP_STH:
6645 			if (!dtrace_canstore(regs[rd], 2, mstate, vstate)) {
6646 				*flags |= CPU_DTRACE_BADADDR;
6647 				*illval = regs[rd];
6648 				break;
6649 			}
6650 			if (regs[rd] & 1) {
6651 				*flags |= CPU_DTRACE_BADALIGN;
6652 				*illval = regs[rd];
6653 				break;
6654 			}
6655 			*((uint16_t *)(uintptr_t)regs[rd]) = (uint16_t)regs[r1];
6656 			break;
6657 
6658 		case DIF_OP_STW:
6659 			if (!dtrace_canstore(regs[rd], 4, mstate, vstate)) {
6660 				*flags |= CPU_DTRACE_BADADDR;
6661 				*illval = regs[rd];
6662 				break;
6663 			}
6664 			if (regs[rd] & 3) {
6665 				*flags |= CPU_DTRACE_BADALIGN;
6666 				*illval = regs[rd];
6667 				break;
6668 			}
6669 			*((uint32_t *)(uintptr_t)regs[rd]) = (uint32_t)regs[r1];
6670 			break;
6671 
6672 		case DIF_OP_STX:
6673 			if (!dtrace_canstore(regs[rd], 8, mstate, vstate)) {
6674 				*flags |= CPU_DTRACE_BADADDR;
6675 				*illval = regs[rd];
6676 				break;
6677 			}
6678 			if (regs[rd] & 7) {
6679 				*flags |= CPU_DTRACE_BADALIGN;
6680 				*illval = regs[rd];
6681 				break;
6682 			}
6683 			*((uint64_t *)(uintptr_t)regs[rd]) = regs[r1];
6684 			break;
6685 		}
6686 	}
6687 
6688 	if (!(*flags & CPU_DTRACE_FAULT))
6689 		return (rval);
6690 
6691 	mstate->dtms_fltoffs = opc * sizeof (dif_instr_t);
6692 	mstate->dtms_present |= DTRACE_MSTATE_FLTOFFS;
6693 
6694 	return (0);
6695 }
6696 
6697 static void
6698 dtrace_action_breakpoint(dtrace_ecb_t *ecb)
6699 {
6700 	dtrace_probe_t *probe = ecb->dte_probe;
6701 	dtrace_provider_t *prov = probe->dtpr_provider;
6702 	char c[DTRACE_FULLNAMELEN + 80], *str;
6703 	char *msg = "dtrace: breakpoint action at probe ";
6704 	char *ecbmsg = " (ecb ";
6705 	uintptr_t mask = (0xf << (sizeof (uintptr_t) * NBBY / 4));
6706 	uintptr_t val = (uintptr_t)ecb;
6707 	int shift = (sizeof (uintptr_t) * NBBY) - 4, i = 0;
6708 
6709 	if (dtrace_destructive_disallow)
6710 		return;
6711 
6712 	/*
6713 	 * It's impossible to be taking action on the NULL probe.
6714 	 */
6715 	ASSERT(probe != NULL);
6716 
6717 	/*
6718 	 * This is a poor man's (destitute man's?) sprintf():  we want to
6719 	 * print the provider name, module name, function name and name of
6720 	 * the probe, along with the hex address of the ECB with the breakpoint
6721 	 * action -- all of which we must place in the character buffer by
6722 	 * hand.
6723 	 */
6724 	while (*msg != '\0')
6725 		c[i++] = *msg++;
6726 
6727 	for (str = prov->dtpv_name; *str != '\0'; str++)
6728 		c[i++] = *str;
6729 	c[i++] = ':';
6730 
6731 	for (str = probe->dtpr_mod; *str != '\0'; str++)
6732 		c[i++] = *str;
6733 	c[i++] = ':';
6734 
6735 	for (str = probe->dtpr_func; *str != '\0'; str++)
6736 		c[i++] = *str;
6737 	c[i++] = ':';
6738 
6739 	for (str = probe->dtpr_name; *str != '\0'; str++)
6740 		c[i++] = *str;
6741 
6742 	while (*ecbmsg != '\0')
6743 		c[i++] = *ecbmsg++;
6744 
6745 	while (shift >= 0) {
6746 		mask = (uintptr_t)0xf << shift;
6747 
6748 		if (val >= ((uintptr_t)1 << shift))
6749 			c[i++] = "0123456789abcdef"[(val & mask) >> shift];
6750 		shift -= 4;
6751 	}
6752 
6753 	c[i++] = ')';
6754 	c[i] = '\0';
6755 
6756 #ifdef illumos
6757 	debug_enter(c);
6758 #else
6759 	kdb_enter(KDB_WHY_DTRACE, "breakpoint action");
6760 #endif
6761 }
6762 
6763 static void
6764 dtrace_action_panic(dtrace_ecb_t *ecb)
6765 {
6766 	dtrace_probe_t *probe = ecb->dte_probe;
6767 
6768 	/*
6769 	 * It's impossible to be taking action on the NULL probe.
6770 	 */
6771 	ASSERT(probe != NULL);
6772 
6773 	if (dtrace_destructive_disallow)
6774 		return;
6775 
6776 	if (dtrace_panicked != NULL)
6777 		return;
6778 
6779 	if (dtrace_casptr(&dtrace_panicked, NULL, curthread) != NULL)
6780 		return;
6781 
6782 	/*
6783 	 * We won the right to panic.  (We want to be sure that only one
6784 	 * thread calls panic() from dtrace_probe(), and that panic() is
6785 	 * called exactly once.)
6786 	 */
6787 	dtrace_panic("dtrace: panic action at probe %s:%s:%s:%s (ecb %p)",
6788 	    probe->dtpr_provider->dtpv_name, probe->dtpr_mod,
6789 	    probe->dtpr_func, probe->dtpr_name, (void *)ecb);
6790 }
6791 
6792 static void
6793 dtrace_action_raise(uint64_t sig)
6794 {
6795 	if (dtrace_destructive_disallow)
6796 		return;
6797 
6798 	if (sig >= NSIG) {
6799 		DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
6800 		return;
6801 	}
6802 
6803 #ifdef illumos
6804 	/*
6805 	 * raise() has a queue depth of 1 -- we ignore all subsequent
6806 	 * invocations of the raise() action.
6807 	 */
6808 	if (curthread->t_dtrace_sig == 0)
6809 		curthread->t_dtrace_sig = (uint8_t)sig;
6810 
6811 	curthread->t_sig_check = 1;
6812 	aston(curthread);
6813 #else
6814 	struct proc *p = curproc;
6815 	PROC_LOCK(p);
6816 	kern_psignal(p, sig);
6817 	PROC_UNLOCK(p);
6818 #endif
6819 }
6820 
6821 static void
6822 dtrace_action_stop(void)
6823 {
6824 	if (dtrace_destructive_disallow)
6825 		return;
6826 
6827 #ifdef illumos
6828 	if (!curthread->t_dtrace_stop) {
6829 		curthread->t_dtrace_stop = 1;
6830 		curthread->t_sig_check = 1;
6831 		aston(curthread);
6832 	}
6833 #else
6834 	struct proc *p = curproc;
6835 	PROC_LOCK(p);
6836 	kern_psignal(p, SIGSTOP);
6837 	PROC_UNLOCK(p);
6838 #endif
6839 }
6840 
6841 static void
6842 dtrace_action_chill(dtrace_mstate_t *mstate, hrtime_t val)
6843 {
6844 	hrtime_t now;
6845 	volatile uint16_t *flags;
6846 #ifdef illumos
6847 	cpu_t *cpu = CPU;
6848 #else
6849 	cpu_t *cpu = &solaris_cpu[curcpu];
6850 #endif
6851 
6852 	if (dtrace_destructive_disallow)
6853 		return;
6854 
6855 	flags = (volatile uint16_t *)&cpu_core[curcpu].cpuc_dtrace_flags;
6856 
6857 	now = dtrace_gethrtime();
6858 
6859 	if (now - cpu->cpu_dtrace_chillmark > dtrace_chill_interval) {
6860 		/*
6861 		 * We need to advance the mark to the current time.
6862 		 */
6863 		cpu->cpu_dtrace_chillmark = now;
6864 		cpu->cpu_dtrace_chilled = 0;
6865 	}
6866 
6867 	/*
6868 	 * Now check to see if the requested chill time would take us over
6869 	 * the maximum amount of time allowed in the chill interval.  (Or
6870 	 * worse, if the calculation itself induces overflow.)
6871 	 */
6872 	if (cpu->cpu_dtrace_chilled + val > dtrace_chill_max ||
6873 	    cpu->cpu_dtrace_chilled + val < cpu->cpu_dtrace_chilled) {
6874 		*flags |= CPU_DTRACE_ILLOP;
6875 		return;
6876 	}
6877 
6878 	while (dtrace_gethrtime() - now < val)
6879 		continue;
6880 
6881 	/*
6882 	 * Normally, we assure that the value of the variable "timestamp" does
6883 	 * not change within an ECB.  The presence of chill() represents an
6884 	 * exception to this rule, however.
6885 	 */
6886 	mstate->dtms_present &= ~DTRACE_MSTATE_TIMESTAMP;
6887 	cpu->cpu_dtrace_chilled += val;
6888 }
6889 
6890 static void
6891 dtrace_action_ustack(dtrace_mstate_t *mstate, dtrace_state_t *state,
6892     uint64_t *buf, uint64_t arg)
6893 {
6894 	int nframes = DTRACE_USTACK_NFRAMES(arg);
6895 	int strsize = DTRACE_USTACK_STRSIZE(arg);
6896 	uint64_t *pcs = &buf[1], *fps;
6897 	char *str = (char *)&pcs[nframes];
6898 	int size, offs = 0, i, j;
6899 	uintptr_t old = mstate->dtms_scratch_ptr, saved;
6900 	uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
6901 	char *sym;
6902 
6903 	/*
6904 	 * Should be taking a faster path if string space has not been
6905 	 * allocated.
6906 	 */
6907 	ASSERT(strsize != 0);
6908 
6909 	/*
6910 	 * We will first allocate some temporary space for the frame pointers.
6911 	 */
6912 	fps = (uint64_t *)P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
6913 	size = (uintptr_t)fps - mstate->dtms_scratch_ptr +
6914 	    (nframes * sizeof (uint64_t));
6915 
6916 	if (!DTRACE_INSCRATCH(mstate, size)) {
6917 		/*
6918 		 * Not enough room for our frame pointers -- need to indicate
6919 		 * that we ran out of scratch space.
6920 		 */
6921 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
6922 		return;
6923 	}
6924 
6925 	mstate->dtms_scratch_ptr += size;
6926 	saved = mstate->dtms_scratch_ptr;
6927 
6928 	/*
6929 	 * Now get a stack with both program counters and frame pointers.
6930 	 */
6931 	DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6932 	dtrace_getufpstack(buf, fps, nframes + 1);
6933 	DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6934 
6935 	/*
6936 	 * If that faulted, we're cooked.
6937 	 */
6938 	if (*flags & CPU_DTRACE_FAULT)
6939 		goto out;
6940 
6941 	/*
6942 	 * Now we want to walk up the stack, calling the USTACK helper.  For
6943 	 * each iteration, we restore the scratch pointer.
6944 	 */
6945 	for (i = 0; i < nframes; i++) {
6946 		mstate->dtms_scratch_ptr = saved;
6947 
6948 		if (offs >= strsize)
6949 			break;
6950 
6951 		sym = (char *)(uintptr_t)dtrace_helper(
6952 		    DTRACE_HELPER_ACTION_USTACK,
6953 		    mstate, state, pcs[i], fps[i]);
6954 
6955 		/*
6956 		 * If we faulted while running the helper, we're going to
6957 		 * clear the fault and null out the corresponding string.
6958 		 */
6959 		if (*flags & CPU_DTRACE_FAULT) {
6960 			*flags &= ~CPU_DTRACE_FAULT;
6961 			str[offs++] = '\0';
6962 			continue;
6963 		}
6964 
6965 		if (sym == NULL) {
6966 			str[offs++] = '\0';
6967 			continue;
6968 		}
6969 
6970 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6971 
6972 		/*
6973 		 * Now copy in the string that the helper returned to us.
6974 		 */
6975 		for (j = 0; offs + j < strsize; j++) {
6976 			if ((str[offs + j] = sym[j]) == '\0')
6977 				break;
6978 		}
6979 
6980 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6981 
6982 		offs += j + 1;
6983 	}
6984 
6985 	if (offs >= strsize) {
6986 		/*
6987 		 * If we didn't have room for all of the strings, we don't
6988 		 * abort processing -- this needn't be a fatal error -- but we
6989 		 * still want to increment a counter (dts_stkstroverflows) to
6990 		 * allow this condition to be warned about.  (If this is from
6991 		 * a jstack() action, it is easily tuned via jstackstrsize.)
6992 		 */
6993 		dtrace_error(&state->dts_stkstroverflows);
6994 	}
6995 
6996 	while (offs < strsize)
6997 		str[offs++] = '\0';
6998 
6999 out:
7000 	mstate->dtms_scratch_ptr = old;
7001 }
7002 
7003 static void
7004 dtrace_store_by_ref(dtrace_difo_t *dp, caddr_t tomax, size_t size,
7005     size_t *valoffsp, uint64_t *valp, uint64_t end, int intuple, int dtkind)
7006 {
7007 	volatile uint16_t *flags;
7008 	uint64_t val = *valp;
7009 	size_t valoffs = *valoffsp;
7010 
7011 	flags = (volatile uint16_t *)&cpu_core[curcpu].cpuc_dtrace_flags;
7012 	ASSERT(dtkind == DIF_TF_BYREF || dtkind == DIF_TF_BYUREF);
7013 
7014 	/*
7015 	 * If this is a string, we're going to only load until we find the zero
7016 	 * byte -- after which we'll store zero bytes.
7017 	 */
7018 	if (dp->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING) {
7019 		char c = '\0' + 1;
7020 		size_t s;
7021 
7022 		for (s = 0; s < size; s++) {
7023 			if (c != '\0' && dtkind == DIF_TF_BYREF) {
7024 				c = dtrace_load8(val++);
7025 			} else if (c != '\0' && dtkind == DIF_TF_BYUREF) {
7026 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
7027 				c = dtrace_fuword8((void *)(uintptr_t)val++);
7028 				DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
7029 				if (*flags & CPU_DTRACE_FAULT)
7030 					break;
7031 			}
7032 
7033 			DTRACE_STORE(uint8_t, tomax, valoffs++, c);
7034 
7035 			if (c == '\0' && intuple)
7036 				break;
7037 		}
7038 	} else {
7039 		uint8_t c;
7040 		while (valoffs < end) {
7041 			if (dtkind == DIF_TF_BYREF) {
7042 				c = dtrace_load8(val++);
7043 			} else if (dtkind == DIF_TF_BYUREF) {
7044 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
7045 				c = dtrace_fuword8((void *)(uintptr_t)val++);
7046 				DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
7047 				if (*flags & CPU_DTRACE_FAULT)
7048 					break;
7049 			}
7050 
7051 			DTRACE_STORE(uint8_t, tomax,
7052 			    valoffs++, c);
7053 		}
7054 	}
7055 
7056 	*valp = val;
7057 	*valoffsp = valoffs;
7058 }
7059 
7060 /*
7061  * If you're looking for the epicenter of DTrace, you just found it.  This
7062  * is the function called by the provider to fire a probe -- from which all
7063  * subsequent probe-context DTrace activity emanates.
7064  */
7065 void
7066 dtrace_probe(dtrace_id_t id, uintptr_t arg0, uintptr_t arg1,
7067     uintptr_t arg2, uintptr_t arg3, uintptr_t arg4)
7068 {
7069 	processorid_t cpuid;
7070 	dtrace_icookie_t cookie;
7071 	dtrace_probe_t *probe;
7072 	dtrace_mstate_t mstate;
7073 	dtrace_ecb_t *ecb;
7074 	dtrace_action_t *act;
7075 	intptr_t offs;
7076 	size_t size;
7077 	int vtime, onintr;
7078 	volatile uint16_t *flags;
7079 	hrtime_t now;
7080 
7081 	if (panicstr != NULL)
7082 		return;
7083 
7084 #ifdef illumos
7085 	/*
7086 	 * Kick out immediately if this CPU is still being born (in which case
7087 	 * curthread will be set to -1) or the current thread can't allow
7088 	 * probes in its current context.
7089 	 */
7090 	if (((uintptr_t)curthread & 1) || (curthread->t_flag & T_DONTDTRACE))
7091 		return;
7092 #endif
7093 
7094 	cookie = dtrace_interrupt_disable();
7095 	probe = dtrace_probes[id - 1];
7096 	cpuid = curcpu;
7097 	onintr = CPU_ON_INTR(CPU);
7098 
7099 	if (!onintr && probe->dtpr_predcache != DTRACE_CACHEIDNONE &&
7100 	    probe->dtpr_predcache == curthread->t_predcache) {
7101 		/*
7102 		 * We have hit in the predicate cache; we know that
7103 		 * this predicate would evaluate to be false.
7104 		 */
7105 		dtrace_interrupt_enable(cookie);
7106 		return;
7107 	}
7108 
7109 #ifdef illumos
7110 	if (panic_quiesce) {
7111 #else
7112 	if (panicstr != NULL) {
7113 #endif
7114 		/*
7115 		 * We don't trace anything if we're panicking.
7116 		 */
7117 		dtrace_interrupt_enable(cookie);
7118 		return;
7119 	}
7120 
7121 	now = mstate.dtms_timestamp = dtrace_gethrtime();
7122 	mstate.dtms_present |= DTRACE_MSTATE_TIMESTAMP;
7123 	vtime = dtrace_vtime_references != 0;
7124 
7125 	if (vtime && curthread->t_dtrace_start)
7126 		curthread->t_dtrace_vtime += now - curthread->t_dtrace_start;
7127 
7128 	mstate.dtms_difo = NULL;
7129 	mstate.dtms_probe = probe;
7130 	mstate.dtms_strtok = 0;
7131 	mstate.dtms_arg[0] = arg0;
7132 	mstate.dtms_arg[1] = arg1;
7133 	mstate.dtms_arg[2] = arg2;
7134 	mstate.dtms_arg[3] = arg3;
7135 	mstate.dtms_arg[4] = arg4;
7136 
7137 	flags = (volatile uint16_t *)&cpu_core[cpuid].cpuc_dtrace_flags;
7138 
7139 	for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
7140 		dtrace_predicate_t *pred = ecb->dte_predicate;
7141 		dtrace_state_t *state = ecb->dte_state;
7142 		dtrace_buffer_t *buf = &state->dts_buffer[cpuid];
7143 		dtrace_buffer_t *aggbuf = &state->dts_aggbuffer[cpuid];
7144 		dtrace_vstate_t *vstate = &state->dts_vstate;
7145 		dtrace_provider_t *prov = probe->dtpr_provider;
7146 		uint64_t tracememsize = 0;
7147 		int committed = 0;
7148 		caddr_t tomax;
7149 
7150 		/*
7151 		 * A little subtlety with the following (seemingly innocuous)
7152 		 * declaration of the automatic 'val':  by looking at the
7153 		 * code, you might think that it could be declared in the
7154 		 * action processing loop, below.  (That is, it's only used in
7155 		 * the action processing loop.)  However, it must be declared
7156 		 * out of that scope because in the case of DIF expression
7157 		 * arguments to aggregating actions, one iteration of the
7158 		 * action loop will use the last iteration's value.
7159 		 */
7160 		uint64_t val = 0;
7161 
7162 		mstate.dtms_present = DTRACE_MSTATE_ARGS | DTRACE_MSTATE_PROBE;
7163 		mstate.dtms_getf = NULL;
7164 
7165 		*flags &= ~CPU_DTRACE_ERROR;
7166 
7167 		if (prov == dtrace_provider) {
7168 			/*
7169 			 * If dtrace itself is the provider of this probe,
7170 			 * we're only going to continue processing the ECB if
7171 			 * arg0 (the dtrace_state_t) is equal to the ECB's
7172 			 * creating state.  (This prevents disjoint consumers
7173 			 * from seeing one another's metaprobes.)
7174 			 */
7175 			if (arg0 != (uint64_t)(uintptr_t)state)
7176 				continue;
7177 		}
7178 
7179 		if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE) {
7180 			/*
7181 			 * We're not currently active.  If our provider isn't
7182 			 * the dtrace pseudo provider, we're not interested.
7183 			 */
7184 			if (prov != dtrace_provider)
7185 				continue;
7186 
7187 			/*
7188 			 * Now we must further check if we are in the BEGIN
7189 			 * probe.  If we are, we will only continue processing
7190 			 * if we're still in WARMUP -- if one BEGIN enabling
7191 			 * has invoked the exit() action, we don't want to
7192 			 * evaluate subsequent BEGIN enablings.
7193 			 */
7194 			if (probe->dtpr_id == dtrace_probeid_begin &&
7195 			    state->dts_activity != DTRACE_ACTIVITY_WARMUP) {
7196 				ASSERT(state->dts_activity ==
7197 				    DTRACE_ACTIVITY_DRAINING);
7198 				continue;
7199 			}
7200 		}
7201 
7202 		if (ecb->dte_cond) {
7203 			/*
7204 			 * If the dte_cond bits indicate that this
7205 			 * consumer is only allowed to see user-mode firings
7206 			 * of this probe, call the provider's dtps_usermode()
7207 			 * entry point to check that the probe was fired
7208 			 * while in a user context. Skip this ECB if that's
7209 			 * not the case.
7210 			 */
7211 			if ((ecb->dte_cond & DTRACE_COND_USERMODE) &&
7212 			    prov->dtpv_pops.dtps_usermode(prov->dtpv_arg,
7213 			    probe->dtpr_id, probe->dtpr_arg) == 0)
7214 				continue;
7215 
7216 #ifdef illumos
7217 			/*
7218 			 * This is more subtle than it looks. We have to be
7219 			 * absolutely certain that CRED() isn't going to
7220 			 * change out from under us so it's only legit to
7221 			 * examine that structure if we're in constrained
7222 			 * situations. Currently, the only times we'll this
7223 			 * check is if a non-super-user has enabled the
7224 			 * profile or syscall providers -- providers that
7225 			 * allow visibility of all processes. For the
7226 			 * profile case, the check above will ensure that
7227 			 * we're examining a user context.
7228 			 */
7229 			if (ecb->dte_cond & DTRACE_COND_OWNER) {
7230 				cred_t *cr;
7231 				cred_t *s_cr =
7232 				    ecb->dte_state->dts_cred.dcr_cred;
7233 				proc_t *proc;
7234 
7235 				ASSERT(s_cr != NULL);
7236 
7237 				if ((cr = CRED()) == NULL ||
7238 				    s_cr->cr_uid != cr->cr_uid ||
7239 				    s_cr->cr_uid != cr->cr_ruid ||
7240 				    s_cr->cr_uid != cr->cr_suid ||
7241 				    s_cr->cr_gid != cr->cr_gid ||
7242 				    s_cr->cr_gid != cr->cr_rgid ||
7243 				    s_cr->cr_gid != cr->cr_sgid ||
7244 				    (proc = ttoproc(curthread)) == NULL ||
7245 				    (proc->p_flag & SNOCD))
7246 					continue;
7247 			}
7248 
7249 			if (ecb->dte_cond & DTRACE_COND_ZONEOWNER) {
7250 				cred_t *cr;
7251 				cred_t *s_cr =
7252 				    ecb->dte_state->dts_cred.dcr_cred;
7253 
7254 				ASSERT(s_cr != NULL);
7255 
7256 				if ((cr = CRED()) == NULL ||
7257 				    s_cr->cr_zone->zone_id !=
7258 				    cr->cr_zone->zone_id)
7259 					continue;
7260 			}
7261 #endif
7262 		}
7263 
7264 		if (now - state->dts_alive > dtrace_deadman_timeout) {
7265 			/*
7266 			 * We seem to be dead.  Unless we (a) have kernel
7267 			 * destructive permissions (b) have explicitly enabled
7268 			 * destructive actions and (c) destructive actions have
7269 			 * not been disabled, we're going to transition into
7270 			 * the KILLED state, from which no further processing
7271 			 * on this state will be performed.
7272 			 */
7273 			if (!dtrace_priv_kernel_destructive(state) ||
7274 			    !state->dts_cred.dcr_destructive ||
7275 			    dtrace_destructive_disallow) {
7276 				void *activity = &state->dts_activity;
7277 				dtrace_activity_t current;
7278 
7279 				do {
7280 					current = state->dts_activity;
7281 				} while (dtrace_cas32(activity, current,
7282 				    DTRACE_ACTIVITY_KILLED) != current);
7283 
7284 				continue;
7285 			}
7286 		}
7287 
7288 		if ((offs = dtrace_buffer_reserve(buf, ecb->dte_needed,
7289 		    ecb->dte_alignment, state, &mstate)) < 0)
7290 			continue;
7291 
7292 		tomax = buf->dtb_tomax;
7293 		ASSERT(tomax != NULL);
7294 
7295 		if (ecb->dte_size != 0) {
7296 			dtrace_rechdr_t dtrh;
7297 			if (!(mstate.dtms_present & DTRACE_MSTATE_TIMESTAMP)) {
7298 				mstate.dtms_timestamp = dtrace_gethrtime();
7299 				mstate.dtms_present |= DTRACE_MSTATE_TIMESTAMP;
7300 			}
7301 			ASSERT3U(ecb->dte_size, >=, sizeof (dtrace_rechdr_t));
7302 			dtrh.dtrh_epid = ecb->dte_epid;
7303 			DTRACE_RECORD_STORE_TIMESTAMP(&dtrh,
7304 			    mstate.dtms_timestamp);
7305 			*((dtrace_rechdr_t *)(tomax + offs)) = dtrh;
7306 		}
7307 
7308 		mstate.dtms_epid = ecb->dte_epid;
7309 		mstate.dtms_present |= DTRACE_MSTATE_EPID;
7310 
7311 		if (state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)
7312 			mstate.dtms_access = DTRACE_ACCESS_KERNEL;
7313 		else
7314 			mstate.dtms_access = 0;
7315 
7316 		if (pred != NULL) {
7317 			dtrace_difo_t *dp = pred->dtp_difo;
7318 			int rval;
7319 
7320 			rval = dtrace_dif_emulate(dp, &mstate, vstate, state);
7321 
7322 			if (!(*flags & CPU_DTRACE_ERROR) && !rval) {
7323 				dtrace_cacheid_t cid = probe->dtpr_predcache;
7324 
7325 				if (cid != DTRACE_CACHEIDNONE && !onintr) {
7326 					/*
7327 					 * Update the predicate cache...
7328 					 */
7329 					ASSERT(cid == pred->dtp_cacheid);
7330 					curthread->t_predcache = cid;
7331 				}
7332 
7333 				continue;
7334 			}
7335 		}
7336 
7337 		for (act = ecb->dte_action; !(*flags & CPU_DTRACE_ERROR) &&
7338 		    act != NULL; act = act->dta_next) {
7339 			size_t valoffs;
7340 			dtrace_difo_t *dp;
7341 			dtrace_recdesc_t *rec = &act->dta_rec;
7342 
7343 			size = rec->dtrd_size;
7344 			valoffs = offs + rec->dtrd_offset;
7345 
7346 			if (DTRACEACT_ISAGG(act->dta_kind)) {
7347 				uint64_t v = 0xbad;
7348 				dtrace_aggregation_t *agg;
7349 
7350 				agg = (dtrace_aggregation_t *)act;
7351 
7352 				if ((dp = act->dta_difo) != NULL)
7353 					v = dtrace_dif_emulate(dp,
7354 					    &mstate, vstate, state);
7355 
7356 				if (*flags & CPU_DTRACE_ERROR)
7357 					continue;
7358 
7359 				/*
7360 				 * Note that we always pass the expression
7361 				 * value from the previous iteration of the
7362 				 * action loop.  This value will only be used
7363 				 * if there is an expression argument to the
7364 				 * aggregating action, denoted by the
7365 				 * dtag_hasarg field.
7366 				 */
7367 				dtrace_aggregate(agg, buf,
7368 				    offs, aggbuf, v, val);
7369 				continue;
7370 			}
7371 
7372 			switch (act->dta_kind) {
7373 			case DTRACEACT_STOP:
7374 				if (dtrace_priv_proc_destructive(state))
7375 					dtrace_action_stop();
7376 				continue;
7377 
7378 			case DTRACEACT_BREAKPOINT:
7379 				if (dtrace_priv_kernel_destructive(state))
7380 					dtrace_action_breakpoint(ecb);
7381 				continue;
7382 
7383 			case DTRACEACT_PANIC:
7384 				if (dtrace_priv_kernel_destructive(state))
7385 					dtrace_action_panic(ecb);
7386 				continue;
7387 
7388 			case DTRACEACT_STACK:
7389 				if (!dtrace_priv_kernel(state))
7390 					continue;
7391 
7392 				dtrace_getpcstack((pc_t *)(tomax + valoffs),
7393 				    size / sizeof (pc_t), probe->dtpr_aframes,
7394 				    DTRACE_ANCHORED(probe) ? NULL :
7395 				    (uint32_t *)arg0);
7396 				continue;
7397 
7398 			case DTRACEACT_JSTACK:
7399 			case DTRACEACT_USTACK:
7400 				if (!dtrace_priv_proc(state))
7401 					continue;
7402 
7403 				/*
7404 				 * See comment in DIF_VAR_PID.
7405 				 */
7406 				if (DTRACE_ANCHORED(mstate.dtms_probe) &&
7407 				    CPU_ON_INTR(CPU)) {
7408 					int depth = DTRACE_USTACK_NFRAMES(
7409 					    rec->dtrd_arg) + 1;
7410 
7411 					dtrace_bzero((void *)(tomax + valoffs),
7412 					    DTRACE_USTACK_STRSIZE(rec->dtrd_arg)
7413 					    + depth * sizeof (uint64_t));
7414 
7415 					continue;
7416 				}
7417 
7418 				if (DTRACE_USTACK_STRSIZE(rec->dtrd_arg) != 0 &&
7419 				    curproc->p_dtrace_helpers != NULL) {
7420 					/*
7421 					 * This is the slow path -- we have
7422 					 * allocated string space, and we're
7423 					 * getting the stack of a process that
7424 					 * has helpers.  Call into a separate
7425 					 * routine to perform this processing.
7426 					 */
7427 					dtrace_action_ustack(&mstate, state,
7428 					    (uint64_t *)(tomax + valoffs),
7429 					    rec->dtrd_arg);
7430 					continue;
7431 				}
7432 
7433 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
7434 				dtrace_getupcstack((uint64_t *)
7435 				    (tomax + valoffs),
7436 				    DTRACE_USTACK_NFRAMES(rec->dtrd_arg) + 1);
7437 				DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
7438 				continue;
7439 
7440 			default:
7441 				break;
7442 			}
7443 
7444 			dp = act->dta_difo;
7445 			ASSERT(dp != NULL);
7446 
7447 			val = dtrace_dif_emulate(dp, &mstate, vstate, state);
7448 
7449 			if (*flags & CPU_DTRACE_ERROR)
7450 				continue;
7451 
7452 			switch (act->dta_kind) {
7453 			case DTRACEACT_SPECULATE: {
7454 				dtrace_rechdr_t *dtrh;
7455 
7456 				ASSERT(buf == &state->dts_buffer[cpuid]);
7457 				buf = dtrace_speculation_buffer(state,
7458 				    cpuid, val);
7459 
7460 				if (buf == NULL) {
7461 					*flags |= CPU_DTRACE_DROP;
7462 					continue;
7463 				}
7464 
7465 				offs = dtrace_buffer_reserve(buf,
7466 				    ecb->dte_needed, ecb->dte_alignment,
7467 				    state, NULL);
7468 
7469 				if (offs < 0) {
7470 					*flags |= CPU_DTRACE_DROP;
7471 					continue;
7472 				}
7473 
7474 				tomax = buf->dtb_tomax;
7475 				ASSERT(tomax != NULL);
7476 
7477 				if (ecb->dte_size == 0)
7478 					continue;
7479 
7480 				ASSERT3U(ecb->dte_size, >=,
7481 				    sizeof (dtrace_rechdr_t));
7482 				dtrh = ((void *)(tomax + offs));
7483 				dtrh->dtrh_epid = ecb->dte_epid;
7484 				/*
7485 				 * When the speculation is committed, all of
7486 				 * the records in the speculative buffer will
7487 				 * have their timestamps set to the commit
7488 				 * time.  Until then, it is set to a sentinel
7489 				 * value, for debugability.
7490 				 */
7491 				DTRACE_RECORD_STORE_TIMESTAMP(dtrh, UINT64_MAX);
7492 				continue;
7493 			}
7494 
7495 			case DTRACEACT_PRINTM: {
7496 				/* The DIF returns a 'memref'. */
7497 				uintptr_t *memref = (uintptr_t *)(uintptr_t) val;
7498 
7499 				/* Get the size from the memref. */
7500 				size = memref[1];
7501 
7502 				/*
7503 				 * Check if the size exceeds the allocated
7504 				 * buffer size.
7505 				 */
7506 				if (size + sizeof(uintptr_t) > dp->dtdo_rtype.dtdt_size) {
7507 					/* Flag a drop! */
7508 					*flags |= CPU_DTRACE_DROP;
7509 					continue;
7510 				}
7511 
7512 				/* Store the size in the buffer first. */
7513 				DTRACE_STORE(uintptr_t, tomax,
7514 				    valoffs, size);
7515 
7516 				/*
7517 				 * Offset the buffer address to the start
7518 				 * of the data.
7519 				 */
7520 				valoffs += sizeof(uintptr_t);
7521 
7522 				/*
7523 				 * Reset to the memory address rather than
7524 				 * the memref array, then let the BYREF
7525 				 * code below do the work to store the
7526 				 * memory data in the buffer.
7527 				 */
7528 				val = memref[0];
7529 				break;
7530 			}
7531 
7532 			case DTRACEACT_PRINTT: {
7533 				/* The DIF returns a 'typeref'. */
7534 				uintptr_t *typeref = (uintptr_t *)(uintptr_t) val;
7535 				char c = '\0' + 1;
7536 				size_t s;
7537 
7538 				/*
7539 				 * Get the type string length and round it
7540 				 * up so that the data that follows is
7541 				 * aligned for easy access.
7542 				 */
7543 				size_t typs = strlen((char *) typeref[2]) + 1;
7544 				typs = roundup(typs,  sizeof(uintptr_t));
7545 
7546 				/*
7547 				 *Get the size from the typeref using the
7548 				 * number of elements and the type size.
7549 				 */
7550 				size = typeref[1] * typeref[3];
7551 
7552 				/*
7553 				 * Check if the size exceeds the allocated
7554 				 * buffer size.
7555 				 */
7556 				if (size + typs + 2 * sizeof(uintptr_t) > dp->dtdo_rtype.dtdt_size) {
7557 					/* Flag a drop! */
7558 					*flags |= CPU_DTRACE_DROP;
7559 
7560 				}
7561 
7562 				/* Store the size in the buffer first. */
7563 				DTRACE_STORE(uintptr_t, tomax,
7564 				    valoffs, size);
7565 				valoffs += sizeof(uintptr_t);
7566 
7567 				/* Store the type size in the buffer. */
7568 				DTRACE_STORE(uintptr_t, tomax,
7569 				    valoffs, typeref[3]);
7570 				valoffs += sizeof(uintptr_t);
7571 
7572 				val = typeref[2];
7573 
7574 				for (s = 0; s < typs; s++) {
7575 					if (c != '\0')
7576 						c = dtrace_load8(val++);
7577 
7578 					DTRACE_STORE(uint8_t, tomax,
7579 					    valoffs++, c);
7580 				}
7581 
7582 				/*
7583 				 * Reset to the memory address rather than
7584 				 * the typeref array, then let the BYREF
7585 				 * code below do the work to store the
7586 				 * memory data in the buffer.
7587 				 */
7588 				val = typeref[0];
7589 				break;
7590 			}
7591 
7592 			case DTRACEACT_CHILL:
7593 				if (dtrace_priv_kernel_destructive(state))
7594 					dtrace_action_chill(&mstate, val);
7595 				continue;
7596 
7597 			case DTRACEACT_RAISE:
7598 				if (dtrace_priv_proc_destructive(state))
7599 					dtrace_action_raise(val);
7600 				continue;
7601 
7602 			case DTRACEACT_COMMIT:
7603 				ASSERT(!committed);
7604 
7605 				/*
7606 				 * We need to commit our buffer state.
7607 				 */
7608 				if (ecb->dte_size)
7609 					buf->dtb_offset = offs + ecb->dte_size;
7610 				buf = &state->dts_buffer[cpuid];
7611 				dtrace_speculation_commit(state, cpuid, val);
7612 				committed = 1;
7613 				continue;
7614 
7615 			case DTRACEACT_DISCARD:
7616 				dtrace_speculation_discard(state, cpuid, val);
7617 				continue;
7618 
7619 			case DTRACEACT_DIFEXPR:
7620 			case DTRACEACT_LIBACT:
7621 			case DTRACEACT_PRINTF:
7622 			case DTRACEACT_PRINTA:
7623 			case DTRACEACT_SYSTEM:
7624 			case DTRACEACT_FREOPEN:
7625 			case DTRACEACT_TRACEMEM:
7626 				break;
7627 
7628 			case DTRACEACT_TRACEMEM_DYNSIZE:
7629 				tracememsize = val;
7630 				break;
7631 
7632 			case DTRACEACT_SYM:
7633 			case DTRACEACT_MOD:
7634 				if (!dtrace_priv_kernel(state))
7635 					continue;
7636 				break;
7637 
7638 			case DTRACEACT_USYM:
7639 			case DTRACEACT_UMOD:
7640 			case DTRACEACT_UADDR: {
7641 #ifdef illumos
7642 				struct pid *pid = curthread->t_procp->p_pidp;
7643 #endif
7644 
7645 				if (!dtrace_priv_proc(state))
7646 					continue;
7647 
7648 				DTRACE_STORE(uint64_t, tomax,
7649 #ifdef illumos
7650 				    valoffs, (uint64_t)pid->pid_id);
7651 #else
7652 				    valoffs, (uint64_t) curproc->p_pid);
7653 #endif
7654 				DTRACE_STORE(uint64_t, tomax,
7655 				    valoffs + sizeof (uint64_t), val);
7656 
7657 				continue;
7658 			}
7659 
7660 			case DTRACEACT_EXIT: {
7661 				/*
7662 				 * For the exit action, we are going to attempt
7663 				 * to atomically set our activity to be
7664 				 * draining.  If this fails (either because
7665 				 * another CPU has beat us to the exit action,
7666 				 * or because our current activity is something
7667 				 * other than ACTIVE or WARMUP), we will
7668 				 * continue.  This assures that the exit action
7669 				 * can be successfully recorded at most once
7670 				 * when we're in the ACTIVE state.  If we're
7671 				 * encountering the exit() action while in
7672 				 * COOLDOWN, however, we want to honor the new
7673 				 * status code.  (We know that we're the only
7674 				 * thread in COOLDOWN, so there is no race.)
7675 				 */
7676 				void *activity = &state->dts_activity;
7677 				dtrace_activity_t current = state->dts_activity;
7678 
7679 				if (current == DTRACE_ACTIVITY_COOLDOWN)
7680 					break;
7681 
7682 				if (current != DTRACE_ACTIVITY_WARMUP)
7683 					current = DTRACE_ACTIVITY_ACTIVE;
7684 
7685 				if (dtrace_cas32(activity, current,
7686 				    DTRACE_ACTIVITY_DRAINING) != current) {
7687 					*flags |= CPU_DTRACE_DROP;
7688 					continue;
7689 				}
7690 
7691 				break;
7692 			}
7693 
7694 			default:
7695 				ASSERT(0);
7696 			}
7697 
7698 			if (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF ||
7699 			    dp->dtdo_rtype.dtdt_flags & DIF_TF_BYUREF) {
7700 				uintptr_t end = valoffs + size;
7701 
7702 				if (tracememsize != 0 &&
7703 				    valoffs + tracememsize < end) {
7704 					end = valoffs + tracememsize;
7705 					tracememsize = 0;
7706 				}
7707 
7708 				if (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF &&
7709 				    !dtrace_vcanload((void *)(uintptr_t)val,
7710 				    &dp->dtdo_rtype, &mstate, vstate))
7711 					continue;
7712 
7713 				dtrace_store_by_ref(dp, tomax, size, &valoffs,
7714 				    &val, end, act->dta_intuple,
7715 				    dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF ?
7716 				    DIF_TF_BYREF: DIF_TF_BYUREF);
7717 				continue;
7718 			}
7719 
7720 			switch (size) {
7721 			case 0:
7722 				break;
7723 
7724 			case sizeof (uint8_t):
7725 				DTRACE_STORE(uint8_t, tomax, valoffs, val);
7726 				break;
7727 			case sizeof (uint16_t):
7728 				DTRACE_STORE(uint16_t, tomax, valoffs, val);
7729 				break;
7730 			case sizeof (uint32_t):
7731 				DTRACE_STORE(uint32_t, tomax, valoffs, val);
7732 				break;
7733 			case sizeof (uint64_t):
7734 				DTRACE_STORE(uint64_t, tomax, valoffs, val);
7735 				break;
7736 			default:
7737 				/*
7738 				 * Any other size should have been returned by
7739 				 * reference, not by value.
7740 				 */
7741 				ASSERT(0);
7742 				break;
7743 			}
7744 		}
7745 
7746 		if (*flags & CPU_DTRACE_DROP)
7747 			continue;
7748 
7749 		if (*flags & CPU_DTRACE_FAULT) {
7750 			int ndx;
7751 			dtrace_action_t *err;
7752 
7753 			buf->dtb_errors++;
7754 
7755 			if (probe->dtpr_id == dtrace_probeid_error) {
7756 				/*
7757 				 * There's nothing we can do -- we had an
7758 				 * error on the error probe.  We bump an
7759 				 * error counter to at least indicate that
7760 				 * this condition happened.
7761 				 */
7762 				dtrace_error(&state->dts_dblerrors);
7763 				continue;
7764 			}
7765 
7766 			if (vtime) {
7767 				/*
7768 				 * Before recursing on dtrace_probe(), we
7769 				 * need to explicitly clear out our start
7770 				 * time to prevent it from being accumulated
7771 				 * into t_dtrace_vtime.
7772 				 */
7773 				curthread->t_dtrace_start = 0;
7774 			}
7775 
7776 			/*
7777 			 * Iterate over the actions to figure out which action
7778 			 * we were processing when we experienced the error.
7779 			 * Note that act points _past_ the faulting action; if
7780 			 * act is ecb->dte_action, the fault was in the
7781 			 * predicate, if it's ecb->dte_action->dta_next it's
7782 			 * in action #1, and so on.
7783 			 */
7784 			for (err = ecb->dte_action, ndx = 0;
7785 			    err != act; err = err->dta_next, ndx++)
7786 				continue;
7787 
7788 			dtrace_probe_error(state, ecb->dte_epid, ndx,
7789 			    (mstate.dtms_present & DTRACE_MSTATE_FLTOFFS) ?
7790 			    mstate.dtms_fltoffs : -1, DTRACE_FLAGS2FLT(*flags),
7791 			    cpu_core[cpuid].cpuc_dtrace_illval);
7792 
7793 			continue;
7794 		}
7795 
7796 		if (!committed)
7797 			buf->dtb_offset = offs + ecb->dte_size;
7798 	}
7799 
7800 	if (vtime)
7801 		curthread->t_dtrace_start = dtrace_gethrtime();
7802 
7803 	dtrace_interrupt_enable(cookie);
7804 }
7805 
7806 /*
7807  * DTrace Probe Hashing Functions
7808  *
7809  * The functions in this section (and indeed, the functions in remaining
7810  * sections) are not _called_ from probe context.  (Any exceptions to this are
7811  * marked with a "Note:".)  Rather, they are called from elsewhere in the
7812  * DTrace framework to look-up probes in, add probes to and remove probes from
7813  * the DTrace probe hashes.  (Each probe is hashed by each element of the
7814  * probe tuple -- allowing for fast lookups, regardless of what was
7815  * specified.)
7816  */
7817 static uint_t
7818 dtrace_hash_str(const char *p)
7819 {
7820 	unsigned int g;
7821 	uint_t hval = 0;
7822 
7823 	while (*p) {
7824 		hval = (hval << 4) + *p++;
7825 		if ((g = (hval & 0xf0000000)) != 0)
7826 			hval ^= g >> 24;
7827 		hval &= ~g;
7828 	}
7829 	return (hval);
7830 }
7831 
7832 static dtrace_hash_t *
7833 dtrace_hash_create(uintptr_t stroffs, uintptr_t nextoffs, uintptr_t prevoffs)
7834 {
7835 	dtrace_hash_t *hash = kmem_zalloc(sizeof (dtrace_hash_t), KM_SLEEP);
7836 
7837 	hash->dth_stroffs = stroffs;
7838 	hash->dth_nextoffs = nextoffs;
7839 	hash->dth_prevoffs = prevoffs;
7840 
7841 	hash->dth_size = 1;
7842 	hash->dth_mask = hash->dth_size - 1;
7843 
7844 	hash->dth_tab = kmem_zalloc(hash->dth_size *
7845 	    sizeof (dtrace_hashbucket_t *), KM_SLEEP);
7846 
7847 	return (hash);
7848 }
7849 
7850 static void
7851 dtrace_hash_destroy(dtrace_hash_t *hash)
7852 {
7853 #ifdef DEBUG
7854 	int i;
7855 
7856 	for (i = 0; i < hash->dth_size; i++)
7857 		ASSERT(hash->dth_tab[i] == NULL);
7858 #endif
7859 
7860 	kmem_free(hash->dth_tab,
7861 	    hash->dth_size * sizeof (dtrace_hashbucket_t *));
7862 	kmem_free(hash, sizeof (dtrace_hash_t));
7863 }
7864 
7865 static void
7866 dtrace_hash_resize(dtrace_hash_t *hash)
7867 {
7868 	int size = hash->dth_size, i, ndx;
7869 	int new_size = hash->dth_size << 1;
7870 	int new_mask = new_size - 1;
7871 	dtrace_hashbucket_t **new_tab, *bucket, *next;
7872 
7873 	ASSERT((new_size & new_mask) == 0);
7874 
7875 	new_tab = kmem_zalloc(new_size * sizeof (void *), KM_SLEEP);
7876 
7877 	for (i = 0; i < size; i++) {
7878 		for (bucket = hash->dth_tab[i]; bucket != NULL; bucket = next) {
7879 			dtrace_probe_t *probe = bucket->dthb_chain;
7880 
7881 			ASSERT(probe != NULL);
7882 			ndx = DTRACE_HASHSTR(hash, probe) & new_mask;
7883 
7884 			next = bucket->dthb_next;
7885 			bucket->dthb_next = new_tab[ndx];
7886 			new_tab[ndx] = bucket;
7887 		}
7888 	}
7889 
7890 	kmem_free(hash->dth_tab, hash->dth_size * sizeof (void *));
7891 	hash->dth_tab = new_tab;
7892 	hash->dth_size = new_size;
7893 	hash->dth_mask = new_mask;
7894 }
7895 
7896 static void
7897 dtrace_hash_add(dtrace_hash_t *hash, dtrace_probe_t *new)
7898 {
7899 	int hashval = DTRACE_HASHSTR(hash, new);
7900 	int ndx = hashval & hash->dth_mask;
7901 	dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
7902 	dtrace_probe_t **nextp, **prevp;
7903 
7904 	for (; bucket != NULL; bucket = bucket->dthb_next) {
7905 		if (DTRACE_HASHEQ(hash, bucket->dthb_chain, new))
7906 			goto add;
7907 	}
7908 
7909 	if ((hash->dth_nbuckets >> 1) > hash->dth_size) {
7910 		dtrace_hash_resize(hash);
7911 		dtrace_hash_add(hash, new);
7912 		return;
7913 	}
7914 
7915 	bucket = kmem_zalloc(sizeof (dtrace_hashbucket_t), KM_SLEEP);
7916 	bucket->dthb_next = hash->dth_tab[ndx];
7917 	hash->dth_tab[ndx] = bucket;
7918 	hash->dth_nbuckets++;
7919 
7920 add:
7921 	nextp = DTRACE_HASHNEXT(hash, new);
7922 	ASSERT(*nextp == NULL && *(DTRACE_HASHPREV(hash, new)) == NULL);
7923 	*nextp = bucket->dthb_chain;
7924 
7925 	if (bucket->dthb_chain != NULL) {
7926 		prevp = DTRACE_HASHPREV(hash, bucket->dthb_chain);
7927 		ASSERT(*prevp == NULL);
7928 		*prevp = new;
7929 	}
7930 
7931 	bucket->dthb_chain = new;
7932 	bucket->dthb_len++;
7933 }
7934 
7935 static dtrace_probe_t *
7936 dtrace_hash_lookup(dtrace_hash_t *hash, dtrace_probe_t *template)
7937 {
7938 	int hashval = DTRACE_HASHSTR(hash, template);
7939 	int ndx = hashval & hash->dth_mask;
7940 	dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
7941 
7942 	for (; bucket != NULL; bucket = bucket->dthb_next) {
7943 		if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
7944 			return (bucket->dthb_chain);
7945 	}
7946 
7947 	return (NULL);
7948 }
7949 
7950 static int
7951 dtrace_hash_collisions(dtrace_hash_t *hash, dtrace_probe_t *template)
7952 {
7953 	int hashval = DTRACE_HASHSTR(hash, template);
7954 	int ndx = hashval & hash->dth_mask;
7955 	dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
7956 
7957 	for (; bucket != NULL; bucket = bucket->dthb_next) {
7958 		if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
7959 			return (bucket->dthb_len);
7960 	}
7961 
7962 	return (0);
7963 }
7964 
7965 static void
7966 dtrace_hash_remove(dtrace_hash_t *hash, dtrace_probe_t *probe)
7967 {
7968 	int ndx = DTRACE_HASHSTR(hash, probe) & hash->dth_mask;
7969 	dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
7970 
7971 	dtrace_probe_t **prevp = DTRACE_HASHPREV(hash, probe);
7972 	dtrace_probe_t **nextp = DTRACE_HASHNEXT(hash, probe);
7973 
7974 	/*
7975 	 * Find the bucket that we're removing this probe from.
7976 	 */
7977 	for (; bucket != NULL; bucket = bucket->dthb_next) {
7978 		if (DTRACE_HASHEQ(hash, bucket->dthb_chain, probe))
7979 			break;
7980 	}
7981 
7982 	ASSERT(bucket != NULL);
7983 
7984 	if (*prevp == NULL) {
7985 		if (*nextp == NULL) {
7986 			/*
7987 			 * The removed probe was the only probe on this
7988 			 * bucket; we need to remove the bucket.
7989 			 */
7990 			dtrace_hashbucket_t *b = hash->dth_tab[ndx];
7991 
7992 			ASSERT(bucket->dthb_chain == probe);
7993 			ASSERT(b != NULL);
7994 
7995 			if (b == bucket) {
7996 				hash->dth_tab[ndx] = bucket->dthb_next;
7997 			} else {
7998 				while (b->dthb_next != bucket)
7999 					b = b->dthb_next;
8000 				b->dthb_next = bucket->dthb_next;
8001 			}
8002 
8003 			ASSERT(hash->dth_nbuckets > 0);
8004 			hash->dth_nbuckets--;
8005 			kmem_free(bucket, sizeof (dtrace_hashbucket_t));
8006 			return;
8007 		}
8008 
8009 		bucket->dthb_chain = *nextp;
8010 	} else {
8011 		*(DTRACE_HASHNEXT(hash, *prevp)) = *nextp;
8012 	}
8013 
8014 	if (*nextp != NULL)
8015 		*(DTRACE_HASHPREV(hash, *nextp)) = *prevp;
8016 }
8017 
8018 /*
8019  * DTrace Utility Functions
8020  *
8021  * These are random utility functions that are _not_ called from probe context.
8022  */
8023 static int
8024 dtrace_badattr(const dtrace_attribute_t *a)
8025 {
8026 	return (a->dtat_name > DTRACE_STABILITY_MAX ||
8027 	    a->dtat_data > DTRACE_STABILITY_MAX ||
8028 	    a->dtat_class > DTRACE_CLASS_MAX);
8029 }
8030 
8031 /*
8032  * Return a duplicate copy of a string.  If the specified string is NULL,
8033  * this function returns a zero-length string.
8034  */
8035 static char *
8036 dtrace_strdup(const char *str)
8037 {
8038 	char *new = kmem_zalloc((str != NULL ? strlen(str) : 0) + 1, KM_SLEEP);
8039 
8040 	if (str != NULL)
8041 		(void) strcpy(new, str);
8042 
8043 	return (new);
8044 }
8045 
8046 #define	DTRACE_ISALPHA(c)	\
8047 	(((c) >= 'a' && (c) <= 'z') || ((c) >= 'A' && (c) <= 'Z'))
8048 
8049 static int
8050 dtrace_badname(const char *s)
8051 {
8052 	char c;
8053 
8054 	if (s == NULL || (c = *s++) == '\0')
8055 		return (0);
8056 
8057 	if (!DTRACE_ISALPHA(c) && c != '-' && c != '_' && c != '.')
8058 		return (1);
8059 
8060 	while ((c = *s++) != '\0') {
8061 		if (!DTRACE_ISALPHA(c) && (c < '0' || c > '9') &&
8062 		    c != '-' && c != '_' && c != '.' && c != '`')
8063 			return (1);
8064 	}
8065 
8066 	return (0);
8067 }
8068 
8069 static void
8070 dtrace_cred2priv(cred_t *cr, uint32_t *privp, uid_t *uidp, zoneid_t *zoneidp)
8071 {
8072 	uint32_t priv;
8073 
8074 #ifdef illumos
8075 	if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
8076 		/*
8077 		 * For DTRACE_PRIV_ALL, the uid and zoneid don't matter.
8078 		 */
8079 		priv = DTRACE_PRIV_ALL;
8080 	} else {
8081 		*uidp = crgetuid(cr);
8082 		*zoneidp = crgetzoneid(cr);
8083 
8084 		priv = 0;
8085 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE))
8086 			priv |= DTRACE_PRIV_KERNEL | DTRACE_PRIV_USER;
8087 		else if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE))
8088 			priv |= DTRACE_PRIV_USER;
8089 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE))
8090 			priv |= DTRACE_PRIV_PROC;
8091 		if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
8092 			priv |= DTRACE_PRIV_OWNER;
8093 		if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
8094 			priv |= DTRACE_PRIV_ZONEOWNER;
8095 	}
8096 #else
8097 	priv = DTRACE_PRIV_ALL;
8098 #endif
8099 
8100 	*privp = priv;
8101 }
8102 
8103 #ifdef DTRACE_ERRDEBUG
8104 static void
8105 dtrace_errdebug(const char *str)
8106 {
8107 	int hval = dtrace_hash_str(str) % DTRACE_ERRHASHSZ;
8108 	int occupied = 0;
8109 
8110 	mutex_enter(&dtrace_errlock);
8111 	dtrace_errlast = str;
8112 	dtrace_errthread = curthread;
8113 
8114 	while (occupied++ < DTRACE_ERRHASHSZ) {
8115 		if (dtrace_errhash[hval].dter_msg == str) {
8116 			dtrace_errhash[hval].dter_count++;
8117 			goto out;
8118 		}
8119 
8120 		if (dtrace_errhash[hval].dter_msg != NULL) {
8121 			hval = (hval + 1) % DTRACE_ERRHASHSZ;
8122 			continue;
8123 		}
8124 
8125 		dtrace_errhash[hval].dter_msg = str;
8126 		dtrace_errhash[hval].dter_count = 1;
8127 		goto out;
8128 	}
8129 
8130 	panic("dtrace: undersized error hash");
8131 out:
8132 	mutex_exit(&dtrace_errlock);
8133 }
8134 #endif
8135 
8136 /*
8137  * DTrace Matching Functions
8138  *
8139  * These functions are used to match groups of probes, given some elements of
8140  * a probe tuple, or some globbed expressions for elements of a probe tuple.
8141  */
8142 static int
8143 dtrace_match_priv(const dtrace_probe_t *prp, uint32_t priv, uid_t uid,
8144     zoneid_t zoneid)
8145 {
8146 	if (priv != DTRACE_PRIV_ALL) {
8147 		uint32_t ppriv = prp->dtpr_provider->dtpv_priv.dtpp_flags;
8148 		uint32_t match = priv & ppriv;
8149 
8150 		/*
8151 		 * No PRIV_DTRACE_* privileges...
8152 		 */
8153 		if ((priv & (DTRACE_PRIV_PROC | DTRACE_PRIV_USER |
8154 		    DTRACE_PRIV_KERNEL)) == 0)
8155 			return (0);
8156 
8157 		/*
8158 		 * No matching bits, but there were bits to match...
8159 		 */
8160 		if (match == 0 && ppriv != 0)
8161 			return (0);
8162 
8163 		/*
8164 		 * Need to have permissions to the process, but don't...
8165 		 */
8166 		if (((ppriv & ~match) & DTRACE_PRIV_OWNER) != 0 &&
8167 		    uid != prp->dtpr_provider->dtpv_priv.dtpp_uid) {
8168 			return (0);
8169 		}
8170 
8171 		/*
8172 		 * Need to be in the same zone unless we possess the
8173 		 * privilege to examine all zones.
8174 		 */
8175 		if (((ppriv & ~match) & DTRACE_PRIV_ZONEOWNER) != 0 &&
8176 		    zoneid != prp->dtpr_provider->dtpv_priv.dtpp_zoneid) {
8177 			return (0);
8178 		}
8179 	}
8180 
8181 	return (1);
8182 }
8183 
8184 /*
8185  * dtrace_match_probe compares a dtrace_probe_t to a pre-compiled key, which
8186  * consists of input pattern strings and an ops-vector to evaluate them.
8187  * This function returns >0 for match, 0 for no match, and <0 for error.
8188  */
8189 static int
8190 dtrace_match_probe(const dtrace_probe_t *prp, const dtrace_probekey_t *pkp,
8191     uint32_t priv, uid_t uid, zoneid_t zoneid)
8192 {
8193 	dtrace_provider_t *pvp = prp->dtpr_provider;
8194 	int rv;
8195 
8196 	if (pvp->dtpv_defunct)
8197 		return (0);
8198 
8199 	if ((rv = pkp->dtpk_pmatch(pvp->dtpv_name, pkp->dtpk_prov, 0)) <= 0)
8200 		return (rv);
8201 
8202 	if ((rv = pkp->dtpk_mmatch(prp->dtpr_mod, pkp->dtpk_mod, 0)) <= 0)
8203 		return (rv);
8204 
8205 	if ((rv = pkp->dtpk_fmatch(prp->dtpr_func, pkp->dtpk_func, 0)) <= 0)
8206 		return (rv);
8207 
8208 	if ((rv = pkp->dtpk_nmatch(prp->dtpr_name, pkp->dtpk_name, 0)) <= 0)
8209 		return (rv);
8210 
8211 	if (dtrace_match_priv(prp, priv, uid, zoneid) == 0)
8212 		return (0);
8213 
8214 	return (rv);
8215 }
8216 
8217 /*
8218  * dtrace_match_glob() is a safe kernel implementation of the gmatch(3GEN)
8219  * interface for matching a glob pattern 'p' to an input string 's'.  Unlike
8220  * libc's version, the kernel version only applies to 8-bit ASCII strings.
8221  * In addition, all of the recursion cases except for '*' matching have been
8222  * unwound.  For '*', we still implement recursive evaluation, but a depth
8223  * counter is maintained and matching is aborted if we recurse too deep.
8224  * The function returns 0 if no match, >0 if match, and <0 if recursion error.
8225  */
8226 static int
8227 dtrace_match_glob(const char *s, const char *p, int depth)
8228 {
8229 	const char *olds;
8230 	char s1, c;
8231 	int gs;
8232 
8233 	if (depth > DTRACE_PROBEKEY_MAXDEPTH)
8234 		return (-1);
8235 
8236 	if (s == NULL)
8237 		s = ""; /* treat NULL as empty string */
8238 
8239 top:
8240 	olds = s;
8241 	s1 = *s++;
8242 
8243 	if (p == NULL)
8244 		return (0);
8245 
8246 	if ((c = *p++) == '\0')
8247 		return (s1 == '\0');
8248 
8249 	switch (c) {
8250 	case '[': {
8251 		int ok = 0, notflag = 0;
8252 		char lc = '\0';
8253 
8254 		if (s1 == '\0')
8255 			return (0);
8256 
8257 		if (*p == '!') {
8258 			notflag = 1;
8259 			p++;
8260 		}
8261 
8262 		if ((c = *p++) == '\0')
8263 			return (0);
8264 
8265 		do {
8266 			if (c == '-' && lc != '\0' && *p != ']') {
8267 				if ((c = *p++) == '\0')
8268 					return (0);
8269 				if (c == '\\' && (c = *p++) == '\0')
8270 					return (0);
8271 
8272 				if (notflag) {
8273 					if (s1 < lc || s1 > c)
8274 						ok++;
8275 					else
8276 						return (0);
8277 				} else if (lc <= s1 && s1 <= c)
8278 					ok++;
8279 
8280 			} else if (c == '\\' && (c = *p++) == '\0')
8281 				return (0);
8282 
8283 			lc = c; /* save left-hand 'c' for next iteration */
8284 
8285 			if (notflag) {
8286 				if (s1 != c)
8287 					ok++;
8288 				else
8289 					return (0);
8290 			} else if (s1 == c)
8291 				ok++;
8292 
8293 			if ((c = *p++) == '\0')
8294 				return (0);
8295 
8296 		} while (c != ']');
8297 
8298 		if (ok)
8299 			goto top;
8300 
8301 		return (0);
8302 	}
8303 
8304 	case '\\':
8305 		if ((c = *p++) == '\0')
8306 			return (0);
8307 		/*FALLTHRU*/
8308 
8309 	default:
8310 		if (c != s1)
8311 			return (0);
8312 		/*FALLTHRU*/
8313 
8314 	case '?':
8315 		if (s1 != '\0')
8316 			goto top;
8317 		return (0);
8318 
8319 	case '*':
8320 		while (*p == '*')
8321 			p++; /* consecutive *'s are identical to a single one */
8322 
8323 		if (*p == '\0')
8324 			return (1);
8325 
8326 		for (s = olds; *s != '\0'; s++) {
8327 			if ((gs = dtrace_match_glob(s, p, depth + 1)) != 0)
8328 				return (gs);
8329 		}
8330 
8331 		return (0);
8332 	}
8333 }
8334 
8335 /*ARGSUSED*/
8336 static int
8337 dtrace_match_string(const char *s, const char *p, int depth)
8338 {
8339 	return (s != NULL && strcmp(s, p) == 0);
8340 }
8341 
8342 /*ARGSUSED*/
8343 static int
8344 dtrace_match_nul(const char *s, const char *p, int depth)
8345 {
8346 	return (1); /* always match the empty pattern */
8347 }
8348 
8349 /*ARGSUSED*/
8350 static int
8351 dtrace_match_nonzero(const char *s, const char *p, int depth)
8352 {
8353 	return (s != NULL && s[0] != '\0');
8354 }
8355 
8356 static int
8357 dtrace_match(const dtrace_probekey_t *pkp, uint32_t priv, uid_t uid,
8358     zoneid_t zoneid, int (*matched)(dtrace_probe_t *, void *), void *arg)
8359 {
8360 	dtrace_probe_t template, *probe;
8361 	dtrace_hash_t *hash = NULL;
8362 	int len, best = INT_MAX, nmatched = 0;
8363 	dtrace_id_t i;
8364 
8365 	ASSERT(MUTEX_HELD(&dtrace_lock));
8366 
8367 	/*
8368 	 * If the probe ID is specified in the key, just lookup by ID and
8369 	 * invoke the match callback once if a matching probe is found.
8370 	 */
8371 	if (pkp->dtpk_id != DTRACE_IDNONE) {
8372 		if ((probe = dtrace_probe_lookup_id(pkp->dtpk_id)) != NULL &&
8373 		    dtrace_match_probe(probe, pkp, priv, uid, zoneid) > 0) {
8374 			(void) (*matched)(probe, arg);
8375 			nmatched++;
8376 		}
8377 		return (nmatched);
8378 	}
8379 
8380 	template.dtpr_mod = (char *)pkp->dtpk_mod;
8381 	template.dtpr_func = (char *)pkp->dtpk_func;
8382 	template.dtpr_name = (char *)pkp->dtpk_name;
8383 
8384 	/*
8385 	 * We want to find the most distinct of the module name, function
8386 	 * name, and name.  So for each one that is not a glob pattern or
8387 	 * empty string, we perform a lookup in the corresponding hash and
8388 	 * use the hash table with the fewest collisions to do our search.
8389 	 */
8390 	if (pkp->dtpk_mmatch == &dtrace_match_string &&
8391 	    (len = dtrace_hash_collisions(dtrace_bymod, &template)) < best) {
8392 		best = len;
8393 		hash = dtrace_bymod;
8394 	}
8395 
8396 	if (pkp->dtpk_fmatch == &dtrace_match_string &&
8397 	    (len = dtrace_hash_collisions(dtrace_byfunc, &template)) < best) {
8398 		best = len;
8399 		hash = dtrace_byfunc;
8400 	}
8401 
8402 	if (pkp->dtpk_nmatch == &dtrace_match_string &&
8403 	    (len = dtrace_hash_collisions(dtrace_byname, &template)) < best) {
8404 		best = len;
8405 		hash = dtrace_byname;
8406 	}
8407 
8408 	/*
8409 	 * If we did not select a hash table, iterate over every probe and
8410 	 * invoke our callback for each one that matches our input probe key.
8411 	 */
8412 	if (hash == NULL) {
8413 		for (i = 0; i < dtrace_nprobes; i++) {
8414 			if ((probe = dtrace_probes[i]) == NULL ||
8415 			    dtrace_match_probe(probe, pkp, priv, uid,
8416 			    zoneid) <= 0)
8417 				continue;
8418 
8419 			nmatched++;
8420 
8421 			if ((*matched)(probe, arg) != DTRACE_MATCH_NEXT)
8422 				break;
8423 		}
8424 
8425 		return (nmatched);
8426 	}
8427 
8428 	/*
8429 	 * If we selected a hash table, iterate over each probe of the same key
8430 	 * name and invoke the callback for every probe that matches the other
8431 	 * attributes of our input probe key.
8432 	 */
8433 	for (probe = dtrace_hash_lookup(hash, &template); probe != NULL;
8434 	    probe = *(DTRACE_HASHNEXT(hash, probe))) {
8435 
8436 		if (dtrace_match_probe(probe, pkp, priv, uid, zoneid) <= 0)
8437 			continue;
8438 
8439 		nmatched++;
8440 
8441 		if ((*matched)(probe, arg) != DTRACE_MATCH_NEXT)
8442 			break;
8443 	}
8444 
8445 	return (nmatched);
8446 }
8447 
8448 /*
8449  * Return the function pointer dtrace_probecmp() should use to compare the
8450  * specified pattern with a string.  For NULL or empty patterns, we select
8451  * dtrace_match_nul().  For glob pattern strings, we use dtrace_match_glob().
8452  * For non-empty non-glob strings, we use dtrace_match_string().
8453  */
8454 static dtrace_probekey_f *
8455 dtrace_probekey_func(const char *p)
8456 {
8457 	char c;
8458 
8459 	if (p == NULL || *p == '\0')
8460 		return (&dtrace_match_nul);
8461 
8462 	while ((c = *p++) != '\0') {
8463 		if (c == '[' || c == '?' || c == '*' || c == '\\')
8464 			return (&dtrace_match_glob);
8465 	}
8466 
8467 	return (&dtrace_match_string);
8468 }
8469 
8470 /*
8471  * Build a probe comparison key for use with dtrace_match_probe() from the
8472  * given probe description.  By convention, a null key only matches anchored
8473  * probes: if each field is the empty string, reset dtpk_fmatch to
8474  * dtrace_match_nonzero().
8475  */
8476 static void
8477 dtrace_probekey(dtrace_probedesc_t *pdp, dtrace_probekey_t *pkp)
8478 {
8479 	pkp->dtpk_prov = pdp->dtpd_provider;
8480 	pkp->dtpk_pmatch = dtrace_probekey_func(pdp->dtpd_provider);
8481 
8482 	pkp->dtpk_mod = pdp->dtpd_mod;
8483 	pkp->dtpk_mmatch = dtrace_probekey_func(pdp->dtpd_mod);
8484 
8485 	pkp->dtpk_func = pdp->dtpd_func;
8486 	pkp->dtpk_fmatch = dtrace_probekey_func(pdp->dtpd_func);
8487 
8488 	pkp->dtpk_name = pdp->dtpd_name;
8489 	pkp->dtpk_nmatch = dtrace_probekey_func(pdp->dtpd_name);
8490 
8491 	pkp->dtpk_id = pdp->dtpd_id;
8492 
8493 	if (pkp->dtpk_id == DTRACE_IDNONE &&
8494 	    pkp->dtpk_pmatch == &dtrace_match_nul &&
8495 	    pkp->dtpk_mmatch == &dtrace_match_nul &&
8496 	    pkp->dtpk_fmatch == &dtrace_match_nul &&
8497 	    pkp->dtpk_nmatch == &dtrace_match_nul)
8498 		pkp->dtpk_fmatch = &dtrace_match_nonzero;
8499 }
8500 
8501 /*
8502  * DTrace Provider-to-Framework API Functions
8503  *
8504  * These functions implement much of the Provider-to-Framework API, as
8505  * described in <sys/dtrace.h>.  The parts of the API not in this section are
8506  * the functions in the API for probe management (found below), and
8507  * dtrace_probe() itself (found above).
8508  */
8509 
8510 /*
8511  * Register the calling provider with the DTrace framework.  This should
8512  * generally be called by DTrace providers in their attach(9E) entry point.
8513  */
8514 int
8515 dtrace_register(const char *name, const dtrace_pattr_t *pap, uint32_t priv,
8516     cred_t *cr, const dtrace_pops_t *pops, void *arg, dtrace_provider_id_t *idp)
8517 {
8518 	dtrace_provider_t *provider;
8519 
8520 	if (name == NULL || pap == NULL || pops == NULL || idp == NULL) {
8521 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
8522 		    "arguments", name ? name : "<NULL>");
8523 		return (EINVAL);
8524 	}
8525 
8526 	if (name[0] == '\0' || dtrace_badname(name)) {
8527 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
8528 		    "provider name", name);
8529 		return (EINVAL);
8530 	}
8531 
8532 	if ((pops->dtps_provide == NULL && pops->dtps_provide_module == NULL) ||
8533 	    pops->dtps_enable == NULL || pops->dtps_disable == NULL ||
8534 	    pops->dtps_destroy == NULL ||
8535 	    ((pops->dtps_resume == NULL) != (pops->dtps_suspend == NULL))) {
8536 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
8537 		    "provider ops", name);
8538 		return (EINVAL);
8539 	}
8540 
8541 	if (dtrace_badattr(&pap->dtpa_provider) ||
8542 	    dtrace_badattr(&pap->dtpa_mod) ||
8543 	    dtrace_badattr(&pap->dtpa_func) ||
8544 	    dtrace_badattr(&pap->dtpa_name) ||
8545 	    dtrace_badattr(&pap->dtpa_args)) {
8546 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
8547 		    "provider attributes", name);
8548 		return (EINVAL);
8549 	}
8550 
8551 	if (priv & ~DTRACE_PRIV_ALL) {
8552 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
8553 		    "privilege attributes", name);
8554 		return (EINVAL);
8555 	}
8556 
8557 	if ((priv & DTRACE_PRIV_KERNEL) &&
8558 	    (priv & (DTRACE_PRIV_USER | DTRACE_PRIV_OWNER)) &&
8559 	    pops->dtps_usermode == NULL) {
8560 		cmn_err(CE_WARN, "failed to register provider '%s': need "
8561 		    "dtps_usermode() op for given privilege attributes", name);
8562 		return (EINVAL);
8563 	}
8564 
8565 	provider = kmem_zalloc(sizeof (dtrace_provider_t), KM_SLEEP);
8566 	provider->dtpv_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
8567 	(void) strcpy(provider->dtpv_name, name);
8568 
8569 	provider->dtpv_attr = *pap;
8570 	provider->dtpv_priv.dtpp_flags = priv;
8571 	if (cr != NULL) {
8572 		provider->dtpv_priv.dtpp_uid = crgetuid(cr);
8573 		provider->dtpv_priv.dtpp_zoneid = crgetzoneid(cr);
8574 	}
8575 	provider->dtpv_pops = *pops;
8576 
8577 	if (pops->dtps_provide == NULL) {
8578 		ASSERT(pops->dtps_provide_module != NULL);
8579 		provider->dtpv_pops.dtps_provide =
8580 		    (void (*)(void *, dtrace_probedesc_t *))dtrace_nullop;
8581 	}
8582 
8583 	if (pops->dtps_provide_module == NULL) {
8584 		ASSERT(pops->dtps_provide != NULL);
8585 		provider->dtpv_pops.dtps_provide_module =
8586 		    (void (*)(void *, modctl_t *))dtrace_nullop;
8587 	}
8588 
8589 	if (pops->dtps_suspend == NULL) {
8590 		ASSERT(pops->dtps_resume == NULL);
8591 		provider->dtpv_pops.dtps_suspend =
8592 		    (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
8593 		provider->dtpv_pops.dtps_resume =
8594 		    (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
8595 	}
8596 
8597 	provider->dtpv_arg = arg;
8598 	*idp = (dtrace_provider_id_t)provider;
8599 
8600 	if (pops == &dtrace_provider_ops) {
8601 		ASSERT(MUTEX_HELD(&dtrace_provider_lock));
8602 		ASSERT(MUTEX_HELD(&dtrace_lock));
8603 		ASSERT(dtrace_anon.dta_enabling == NULL);
8604 
8605 		/*
8606 		 * We make sure that the DTrace provider is at the head of
8607 		 * the provider chain.
8608 		 */
8609 		provider->dtpv_next = dtrace_provider;
8610 		dtrace_provider = provider;
8611 		return (0);
8612 	}
8613 
8614 	mutex_enter(&dtrace_provider_lock);
8615 	mutex_enter(&dtrace_lock);
8616 
8617 	/*
8618 	 * If there is at least one provider registered, we'll add this
8619 	 * provider after the first provider.
8620 	 */
8621 	if (dtrace_provider != NULL) {
8622 		provider->dtpv_next = dtrace_provider->dtpv_next;
8623 		dtrace_provider->dtpv_next = provider;
8624 	} else {
8625 		dtrace_provider = provider;
8626 	}
8627 
8628 	if (dtrace_retained != NULL) {
8629 		dtrace_enabling_provide(provider);
8630 
8631 		/*
8632 		 * Now we need to call dtrace_enabling_matchall() -- which
8633 		 * will acquire cpu_lock and dtrace_lock.  We therefore need
8634 		 * to drop all of our locks before calling into it...
8635 		 */
8636 		mutex_exit(&dtrace_lock);
8637 		mutex_exit(&dtrace_provider_lock);
8638 		dtrace_enabling_matchall();
8639 
8640 		return (0);
8641 	}
8642 
8643 	mutex_exit(&dtrace_lock);
8644 	mutex_exit(&dtrace_provider_lock);
8645 
8646 	return (0);
8647 }
8648 
8649 /*
8650  * Unregister the specified provider from the DTrace framework.  This should
8651  * generally be called by DTrace providers in their detach(9E) entry point.
8652  */
8653 int
8654 dtrace_unregister(dtrace_provider_id_t id)
8655 {
8656 	dtrace_provider_t *old = (dtrace_provider_t *)id;
8657 	dtrace_provider_t *prev = NULL;
8658 	int i, self = 0, noreap = 0;
8659 	dtrace_probe_t *probe, *first = NULL;
8660 
8661 	if (old->dtpv_pops.dtps_enable ==
8662 	    (void (*)(void *, dtrace_id_t, void *))dtrace_nullop) {
8663 		/*
8664 		 * If DTrace itself is the provider, we're called with locks
8665 		 * already held.
8666 		 */
8667 		ASSERT(old == dtrace_provider);
8668 #ifdef illumos
8669 		ASSERT(dtrace_devi != NULL);
8670 #endif
8671 		ASSERT(MUTEX_HELD(&dtrace_provider_lock));
8672 		ASSERT(MUTEX_HELD(&dtrace_lock));
8673 		self = 1;
8674 
8675 		if (dtrace_provider->dtpv_next != NULL) {
8676 			/*
8677 			 * There's another provider here; return failure.
8678 			 */
8679 			return (EBUSY);
8680 		}
8681 	} else {
8682 		mutex_enter(&dtrace_provider_lock);
8683 #ifdef illumos
8684 		mutex_enter(&mod_lock);
8685 #endif
8686 		mutex_enter(&dtrace_lock);
8687 	}
8688 
8689 	/*
8690 	 * If anyone has /dev/dtrace open, or if there are anonymous enabled
8691 	 * probes, we refuse to let providers slither away, unless this
8692 	 * provider has already been explicitly invalidated.
8693 	 */
8694 	if (!old->dtpv_defunct &&
8695 	    (dtrace_opens || (dtrace_anon.dta_state != NULL &&
8696 	    dtrace_anon.dta_state->dts_necbs > 0))) {
8697 		if (!self) {
8698 			mutex_exit(&dtrace_lock);
8699 #ifdef illumos
8700 			mutex_exit(&mod_lock);
8701 #endif
8702 			mutex_exit(&dtrace_provider_lock);
8703 		}
8704 		return (EBUSY);
8705 	}
8706 
8707 	/*
8708 	 * Attempt to destroy the probes associated with this provider.
8709 	 */
8710 	for (i = 0; i < dtrace_nprobes; i++) {
8711 		if ((probe = dtrace_probes[i]) == NULL)
8712 			continue;
8713 
8714 		if (probe->dtpr_provider != old)
8715 			continue;
8716 
8717 		if (probe->dtpr_ecb == NULL)
8718 			continue;
8719 
8720 		/*
8721 		 * If we are trying to unregister a defunct provider, and the
8722 		 * provider was made defunct within the interval dictated by
8723 		 * dtrace_unregister_defunct_reap, we'll (asynchronously)
8724 		 * attempt to reap our enablings.  To denote that the provider
8725 		 * should reattempt to unregister itself at some point in the
8726 		 * future, we will return a differentiable error code (EAGAIN
8727 		 * instead of EBUSY) in this case.
8728 		 */
8729 		if (dtrace_gethrtime() - old->dtpv_defunct >
8730 		    dtrace_unregister_defunct_reap)
8731 			noreap = 1;
8732 
8733 		if (!self) {
8734 			mutex_exit(&dtrace_lock);
8735 #ifdef illumos
8736 			mutex_exit(&mod_lock);
8737 #endif
8738 			mutex_exit(&dtrace_provider_lock);
8739 		}
8740 
8741 		if (noreap)
8742 			return (EBUSY);
8743 
8744 		(void) taskq_dispatch(dtrace_taskq,
8745 		    (task_func_t *)dtrace_enabling_reap, NULL, TQ_SLEEP);
8746 
8747 		return (EAGAIN);
8748 	}
8749 
8750 	/*
8751 	 * All of the probes for this provider are disabled; we can safely
8752 	 * remove all of them from their hash chains and from the probe array.
8753 	 */
8754 	for (i = 0; i < dtrace_nprobes; i++) {
8755 		if ((probe = dtrace_probes[i]) == NULL)
8756 			continue;
8757 
8758 		if (probe->dtpr_provider != old)
8759 			continue;
8760 
8761 		dtrace_probes[i] = NULL;
8762 
8763 		dtrace_hash_remove(dtrace_bymod, probe);
8764 		dtrace_hash_remove(dtrace_byfunc, probe);
8765 		dtrace_hash_remove(dtrace_byname, probe);
8766 
8767 		if (first == NULL) {
8768 			first = probe;
8769 			probe->dtpr_nextmod = NULL;
8770 		} else {
8771 			probe->dtpr_nextmod = first;
8772 			first = probe;
8773 		}
8774 	}
8775 
8776 	/*
8777 	 * The provider's probes have been removed from the hash chains and
8778 	 * from the probe array.  Now issue a dtrace_sync() to be sure that
8779 	 * everyone has cleared out from any probe array processing.
8780 	 */
8781 	dtrace_sync();
8782 
8783 	for (probe = first; probe != NULL; probe = first) {
8784 		first = probe->dtpr_nextmod;
8785 
8786 		old->dtpv_pops.dtps_destroy(old->dtpv_arg, probe->dtpr_id,
8787 		    probe->dtpr_arg);
8788 		kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
8789 		kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
8790 		kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
8791 #ifdef illumos
8792 		vmem_free(dtrace_arena, (void *)(uintptr_t)(probe->dtpr_id), 1);
8793 #else
8794 		free_unr(dtrace_arena, probe->dtpr_id);
8795 #endif
8796 		kmem_free(probe, sizeof (dtrace_probe_t));
8797 	}
8798 
8799 	if ((prev = dtrace_provider) == old) {
8800 #ifdef illumos
8801 		ASSERT(self || dtrace_devi == NULL);
8802 		ASSERT(old->dtpv_next == NULL || dtrace_devi == NULL);
8803 #endif
8804 		dtrace_provider = old->dtpv_next;
8805 	} else {
8806 		while (prev != NULL && prev->dtpv_next != old)
8807 			prev = prev->dtpv_next;
8808 
8809 		if (prev == NULL) {
8810 			panic("attempt to unregister non-existent "
8811 			    "dtrace provider %p\n", (void *)id);
8812 		}
8813 
8814 		prev->dtpv_next = old->dtpv_next;
8815 	}
8816 
8817 	if (!self) {
8818 		mutex_exit(&dtrace_lock);
8819 #ifdef illumos
8820 		mutex_exit(&mod_lock);
8821 #endif
8822 		mutex_exit(&dtrace_provider_lock);
8823 	}
8824 
8825 	kmem_free(old->dtpv_name, strlen(old->dtpv_name) + 1);
8826 	kmem_free(old, sizeof (dtrace_provider_t));
8827 
8828 	return (0);
8829 }
8830 
8831 /*
8832  * Invalidate the specified provider.  All subsequent probe lookups for the
8833  * specified provider will fail, but its probes will not be removed.
8834  */
8835 void
8836 dtrace_invalidate(dtrace_provider_id_t id)
8837 {
8838 	dtrace_provider_t *pvp = (dtrace_provider_t *)id;
8839 
8840 	ASSERT(pvp->dtpv_pops.dtps_enable !=
8841 	    (void (*)(void *, dtrace_id_t, void *))dtrace_nullop);
8842 
8843 	mutex_enter(&dtrace_provider_lock);
8844 	mutex_enter(&dtrace_lock);
8845 
8846 	pvp->dtpv_defunct = dtrace_gethrtime();
8847 
8848 	mutex_exit(&dtrace_lock);
8849 	mutex_exit(&dtrace_provider_lock);
8850 }
8851 
8852 /*
8853  * Indicate whether or not DTrace has attached.
8854  */
8855 int
8856 dtrace_attached(void)
8857 {
8858 	/*
8859 	 * dtrace_provider will be non-NULL iff the DTrace driver has
8860 	 * attached.  (It's non-NULL because DTrace is always itself a
8861 	 * provider.)
8862 	 */
8863 	return (dtrace_provider != NULL);
8864 }
8865 
8866 /*
8867  * Remove all the unenabled probes for the given provider.  This function is
8868  * not unlike dtrace_unregister(), except that it doesn't remove the provider
8869  * -- just as many of its associated probes as it can.
8870  */
8871 int
8872 dtrace_condense(dtrace_provider_id_t id)
8873 {
8874 	dtrace_provider_t *prov = (dtrace_provider_t *)id;
8875 	int i;
8876 	dtrace_probe_t *probe;
8877 
8878 	/*
8879 	 * Make sure this isn't the dtrace provider itself.
8880 	 */
8881 	ASSERT(prov->dtpv_pops.dtps_enable !=
8882 	    (void (*)(void *, dtrace_id_t, void *))dtrace_nullop);
8883 
8884 	mutex_enter(&dtrace_provider_lock);
8885 	mutex_enter(&dtrace_lock);
8886 
8887 	/*
8888 	 * Attempt to destroy the probes associated with this provider.
8889 	 */
8890 	for (i = 0; i < dtrace_nprobes; i++) {
8891 		if ((probe = dtrace_probes[i]) == NULL)
8892 			continue;
8893 
8894 		if (probe->dtpr_provider != prov)
8895 			continue;
8896 
8897 		if (probe->dtpr_ecb != NULL)
8898 			continue;
8899 
8900 		dtrace_probes[i] = NULL;
8901 
8902 		dtrace_hash_remove(dtrace_bymod, probe);
8903 		dtrace_hash_remove(dtrace_byfunc, probe);
8904 		dtrace_hash_remove(dtrace_byname, probe);
8905 
8906 		prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, i + 1,
8907 		    probe->dtpr_arg);
8908 		kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
8909 		kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
8910 		kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
8911 		kmem_free(probe, sizeof (dtrace_probe_t));
8912 #ifdef illumos
8913 		vmem_free(dtrace_arena, (void *)((uintptr_t)i + 1), 1);
8914 #else
8915 		free_unr(dtrace_arena, i + 1);
8916 #endif
8917 	}
8918 
8919 	mutex_exit(&dtrace_lock);
8920 	mutex_exit(&dtrace_provider_lock);
8921 
8922 	return (0);
8923 }
8924 
8925 /*
8926  * DTrace Probe Management Functions
8927  *
8928  * The functions in this section perform the DTrace probe management,
8929  * including functions to create probes, look-up probes, and call into the
8930  * providers to request that probes be provided.  Some of these functions are
8931  * in the Provider-to-Framework API; these functions can be identified by the
8932  * fact that they are not declared "static".
8933  */
8934 
8935 /*
8936  * Create a probe with the specified module name, function name, and name.
8937  */
8938 dtrace_id_t
8939 dtrace_probe_create(dtrace_provider_id_t prov, const char *mod,
8940     const char *func, const char *name, int aframes, void *arg)
8941 {
8942 	dtrace_probe_t *probe, **probes;
8943 	dtrace_provider_t *provider = (dtrace_provider_t *)prov;
8944 	dtrace_id_t id;
8945 
8946 	if (provider == dtrace_provider) {
8947 		ASSERT(MUTEX_HELD(&dtrace_lock));
8948 	} else {
8949 		mutex_enter(&dtrace_lock);
8950 	}
8951 
8952 #ifdef illumos
8953 	id = (dtrace_id_t)(uintptr_t)vmem_alloc(dtrace_arena, 1,
8954 	    VM_BESTFIT | VM_SLEEP);
8955 #else
8956 	id = alloc_unr(dtrace_arena);
8957 #endif
8958 	probe = kmem_zalloc(sizeof (dtrace_probe_t), KM_SLEEP);
8959 
8960 	probe->dtpr_id = id;
8961 	probe->dtpr_gen = dtrace_probegen++;
8962 	probe->dtpr_mod = dtrace_strdup(mod);
8963 	probe->dtpr_func = dtrace_strdup(func);
8964 	probe->dtpr_name = dtrace_strdup(name);
8965 	probe->dtpr_arg = arg;
8966 	probe->dtpr_aframes = aframes;
8967 	probe->dtpr_provider = provider;
8968 
8969 	dtrace_hash_add(dtrace_bymod, probe);
8970 	dtrace_hash_add(dtrace_byfunc, probe);
8971 	dtrace_hash_add(dtrace_byname, probe);
8972 
8973 	if (id - 1 >= dtrace_nprobes) {
8974 		size_t osize = dtrace_nprobes * sizeof (dtrace_probe_t *);
8975 		size_t nsize = osize << 1;
8976 
8977 		if (nsize == 0) {
8978 			ASSERT(osize == 0);
8979 			ASSERT(dtrace_probes == NULL);
8980 			nsize = sizeof (dtrace_probe_t *);
8981 		}
8982 
8983 		probes = kmem_zalloc(nsize, KM_SLEEP);
8984 
8985 		if (dtrace_probes == NULL) {
8986 			ASSERT(osize == 0);
8987 			dtrace_probes = probes;
8988 			dtrace_nprobes = 1;
8989 		} else {
8990 			dtrace_probe_t **oprobes = dtrace_probes;
8991 
8992 			bcopy(oprobes, probes, osize);
8993 			dtrace_membar_producer();
8994 			dtrace_probes = probes;
8995 
8996 			dtrace_sync();
8997 
8998 			/*
8999 			 * All CPUs are now seeing the new probes array; we can
9000 			 * safely free the old array.
9001 			 */
9002 			kmem_free(oprobes, osize);
9003 			dtrace_nprobes <<= 1;
9004 		}
9005 
9006 		ASSERT(id - 1 < dtrace_nprobes);
9007 	}
9008 
9009 	ASSERT(dtrace_probes[id - 1] == NULL);
9010 	dtrace_probes[id - 1] = probe;
9011 
9012 	if (provider != dtrace_provider)
9013 		mutex_exit(&dtrace_lock);
9014 
9015 	return (id);
9016 }
9017 
9018 static dtrace_probe_t *
9019 dtrace_probe_lookup_id(dtrace_id_t id)
9020 {
9021 	ASSERT(MUTEX_HELD(&dtrace_lock));
9022 
9023 	if (id == 0 || id > dtrace_nprobes)
9024 		return (NULL);
9025 
9026 	return (dtrace_probes[id - 1]);
9027 }
9028 
9029 static int
9030 dtrace_probe_lookup_match(dtrace_probe_t *probe, void *arg)
9031 {
9032 	*((dtrace_id_t *)arg) = probe->dtpr_id;
9033 
9034 	return (DTRACE_MATCH_DONE);
9035 }
9036 
9037 /*
9038  * Look up a probe based on provider and one or more of module name, function
9039  * name and probe name.
9040  */
9041 dtrace_id_t
9042 dtrace_probe_lookup(dtrace_provider_id_t prid, char *mod,
9043     char *func, char *name)
9044 {
9045 	dtrace_probekey_t pkey;
9046 	dtrace_id_t id;
9047 	int match;
9048 
9049 	pkey.dtpk_prov = ((dtrace_provider_t *)prid)->dtpv_name;
9050 	pkey.dtpk_pmatch = &dtrace_match_string;
9051 	pkey.dtpk_mod = mod;
9052 	pkey.dtpk_mmatch = mod ? &dtrace_match_string : &dtrace_match_nul;
9053 	pkey.dtpk_func = func;
9054 	pkey.dtpk_fmatch = func ? &dtrace_match_string : &dtrace_match_nul;
9055 	pkey.dtpk_name = name;
9056 	pkey.dtpk_nmatch = name ? &dtrace_match_string : &dtrace_match_nul;
9057 	pkey.dtpk_id = DTRACE_IDNONE;
9058 
9059 	mutex_enter(&dtrace_lock);
9060 	match = dtrace_match(&pkey, DTRACE_PRIV_ALL, 0, 0,
9061 	    dtrace_probe_lookup_match, &id);
9062 	mutex_exit(&dtrace_lock);
9063 
9064 	ASSERT(match == 1 || match == 0);
9065 	return (match ? id : 0);
9066 }
9067 
9068 /*
9069  * Returns the probe argument associated with the specified probe.
9070  */
9071 void *
9072 dtrace_probe_arg(dtrace_provider_id_t id, dtrace_id_t pid)
9073 {
9074 	dtrace_probe_t *probe;
9075 	void *rval = NULL;
9076 
9077 	mutex_enter(&dtrace_lock);
9078 
9079 	if ((probe = dtrace_probe_lookup_id(pid)) != NULL &&
9080 	    probe->dtpr_provider == (dtrace_provider_t *)id)
9081 		rval = probe->dtpr_arg;
9082 
9083 	mutex_exit(&dtrace_lock);
9084 
9085 	return (rval);
9086 }
9087 
9088 /*
9089  * Copy a probe into a probe description.
9090  */
9091 static void
9092 dtrace_probe_description(const dtrace_probe_t *prp, dtrace_probedesc_t *pdp)
9093 {
9094 	bzero(pdp, sizeof (dtrace_probedesc_t));
9095 	pdp->dtpd_id = prp->dtpr_id;
9096 
9097 	(void) strncpy(pdp->dtpd_provider,
9098 	    prp->dtpr_provider->dtpv_name, DTRACE_PROVNAMELEN - 1);
9099 
9100 	(void) strncpy(pdp->dtpd_mod, prp->dtpr_mod, DTRACE_MODNAMELEN - 1);
9101 	(void) strncpy(pdp->dtpd_func, prp->dtpr_func, DTRACE_FUNCNAMELEN - 1);
9102 	(void) strncpy(pdp->dtpd_name, prp->dtpr_name, DTRACE_NAMELEN - 1);
9103 }
9104 
9105 /*
9106  * Called to indicate that a probe -- or probes -- should be provided by a
9107  * specfied provider.  If the specified description is NULL, the provider will
9108  * be told to provide all of its probes.  (This is done whenever a new
9109  * consumer comes along, or whenever a retained enabling is to be matched.) If
9110  * the specified description is non-NULL, the provider is given the
9111  * opportunity to dynamically provide the specified probe, allowing providers
9112  * to support the creation of probes on-the-fly.  (So-called _autocreated_
9113  * probes.)  If the provider is NULL, the operations will be applied to all
9114  * providers; if the provider is non-NULL the operations will only be applied
9115  * to the specified provider.  The dtrace_provider_lock must be held, and the
9116  * dtrace_lock must _not_ be held -- the provider's dtps_provide() operation
9117  * will need to grab the dtrace_lock when it reenters the framework through
9118  * dtrace_probe_lookup(), dtrace_probe_create(), etc.
9119  */
9120 static void
9121 dtrace_probe_provide(dtrace_probedesc_t *desc, dtrace_provider_t *prv)
9122 {
9123 #ifdef illumos
9124 	modctl_t *ctl;
9125 #endif
9126 	int all = 0;
9127 
9128 	ASSERT(MUTEX_HELD(&dtrace_provider_lock));
9129 
9130 	if (prv == NULL) {
9131 		all = 1;
9132 		prv = dtrace_provider;
9133 	}
9134 
9135 	do {
9136 		/*
9137 		 * First, call the blanket provide operation.
9138 		 */
9139 		prv->dtpv_pops.dtps_provide(prv->dtpv_arg, desc);
9140 
9141 #ifdef illumos
9142 		/*
9143 		 * Now call the per-module provide operation.  We will grab
9144 		 * mod_lock to prevent the list from being modified.  Note
9145 		 * that this also prevents the mod_busy bits from changing.
9146 		 * (mod_busy can only be changed with mod_lock held.)
9147 		 */
9148 		mutex_enter(&mod_lock);
9149 
9150 		ctl = &modules;
9151 		do {
9152 			if (ctl->mod_busy || ctl->mod_mp == NULL)
9153 				continue;
9154 
9155 			prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
9156 
9157 		} while ((ctl = ctl->mod_next) != &modules);
9158 
9159 		mutex_exit(&mod_lock);
9160 #endif
9161 	} while (all && (prv = prv->dtpv_next) != NULL);
9162 }
9163 
9164 #ifdef illumos
9165 /*
9166  * Iterate over each probe, and call the Framework-to-Provider API function
9167  * denoted by offs.
9168  */
9169 static void
9170 dtrace_probe_foreach(uintptr_t offs)
9171 {
9172 	dtrace_provider_t *prov;
9173 	void (*func)(void *, dtrace_id_t, void *);
9174 	dtrace_probe_t *probe;
9175 	dtrace_icookie_t cookie;
9176 	int i;
9177 
9178 	/*
9179 	 * We disable interrupts to walk through the probe array.  This is
9180 	 * safe -- the dtrace_sync() in dtrace_unregister() assures that we
9181 	 * won't see stale data.
9182 	 */
9183 	cookie = dtrace_interrupt_disable();
9184 
9185 	for (i = 0; i < dtrace_nprobes; i++) {
9186 		if ((probe = dtrace_probes[i]) == NULL)
9187 			continue;
9188 
9189 		if (probe->dtpr_ecb == NULL) {
9190 			/*
9191 			 * This probe isn't enabled -- don't call the function.
9192 			 */
9193 			continue;
9194 		}
9195 
9196 		prov = probe->dtpr_provider;
9197 		func = *((void(**)(void *, dtrace_id_t, void *))
9198 		    ((uintptr_t)&prov->dtpv_pops + offs));
9199 
9200 		func(prov->dtpv_arg, i + 1, probe->dtpr_arg);
9201 	}
9202 
9203 	dtrace_interrupt_enable(cookie);
9204 }
9205 #endif
9206 
9207 static int
9208 dtrace_probe_enable(dtrace_probedesc_t *desc, dtrace_enabling_t *enab)
9209 {
9210 	dtrace_probekey_t pkey;
9211 	uint32_t priv;
9212 	uid_t uid;
9213 	zoneid_t zoneid;
9214 
9215 	ASSERT(MUTEX_HELD(&dtrace_lock));
9216 	dtrace_ecb_create_cache = NULL;
9217 
9218 	if (desc == NULL) {
9219 		/*
9220 		 * If we're passed a NULL description, we're being asked to
9221 		 * create an ECB with a NULL probe.
9222 		 */
9223 		(void) dtrace_ecb_create_enable(NULL, enab);
9224 		return (0);
9225 	}
9226 
9227 	dtrace_probekey(desc, &pkey);
9228 	dtrace_cred2priv(enab->dten_vstate->dtvs_state->dts_cred.dcr_cred,
9229 	    &priv, &uid, &zoneid);
9230 
9231 	return (dtrace_match(&pkey, priv, uid, zoneid, dtrace_ecb_create_enable,
9232 	    enab));
9233 }
9234 
9235 /*
9236  * DTrace Helper Provider Functions
9237  */
9238 static void
9239 dtrace_dofattr2attr(dtrace_attribute_t *attr, const dof_attr_t dofattr)
9240 {
9241 	attr->dtat_name = DOF_ATTR_NAME(dofattr);
9242 	attr->dtat_data = DOF_ATTR_DATA(dofattr);
9243 	attr->dtat_class = DOF_ATTR_CLASS(dofattr);
9244 }
9245 
9246 static void
9247 dtrace_dofprov2hprov(dtrace_helper_provdesc_t *hprov,
9248     const dof_provider_t *dofprov, char *strtab)
9249 {
9250 	hprov->dthpv_provname = strtab + dofprov->dofpv_name;
9251 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_provider,
9252 	    dofprov->dofpv_provattr);
9253 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_mod,
9254 	    dofprov->dofpv_modattr);
9255 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_func,
9256 	    dofprov->dofpv_funcattr);
9257 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_name,
9258 	    dofprov->dofpv_nameattr);
9259 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_args,
9260 	    dofprov->dofpv_argsattr);
9261 }
9262 
9263 static void
9264 dtrace_helper_provide_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
9265 {
9266 	uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
9267 	dof_hdr_t *dof = (dof_hdr_t *)daddr;
9268 	dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
9269 	dof_provider_t *provider;
9270 	dof_probe_t *probe;
9271 	uint32_t *off, *enoff;
9272 	uint8_t *arg;
9273 	char *strtab;
9274 	uint_t i, nprobes;
9275 	dtrace_helper_provdesc_t dhpv;
9276 	dtrace_helper_probedesc_t dhpb;
9277 	dtrace_meta_t *meta = dtrace_meta_pid;
9278 	dtrace_mops_t *mops = &meta->dtm_mops;
9279 	void *parg;
9280 
9281 	provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
9282 	str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
9283 	    provider->dofpv_strtab * dof->dofh_secsize);
9284 	prb_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
9285 	    provider->dofpv_probes * dof->dofh_secsize);
9286 	arg_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
9287 	    provider->dofpv_prargs * dof->dofh_secsize);
9288 	off_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
9289 	    provider->dofpv_proffs * dof->dofh_secsize);
9290 
9291 	strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
9292 	off = (uint32_t *)(uintptr_t)(daddr + off_sec->dofs_offset);
9293 	arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
9294 	enoff = NULL;
9295 
9296 	/*
9297 	 * See dtrace_helper_provider_validate().
9298 	 */
9299 	if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
9300 	    provider->dofpv_prenoffs != DOF_SECT_NONE) {
9301 		enoff_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
9302 		    provider->dofpv_prenoffs * dof->dofh_secsize);
9303 		enoff = (uint32_t *)(uintptr_t)(daddr + enoff_sec->dofs_offset);
9304 	}
9305 
9306 	nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
9307 
9308 	/*
9309 	 * Create the provider.
9310 	 */
9311 	dtrace_dofprov2hprov(&dhpv, provider, strtab);
9312 
9313 	if ((parg = mops->dtms_provide_pid(meta->dtm_arg, &dhpv, pid)) == NULL)
9314 		return;
9315 
9316 	meta->dtm_count++;
9317 
9318 	/*
9319 	 * Create the probes.
9320 	 */
9321 	for (i = 0; i < nprobes; i++) {
9322 		probe = (dof_probe_t *)(uintptr_t)(daddr +
9323 		    prb_sec->dofs_offset + i * prb_sec->dofs_entsize);
9324 
9325 		dhpb.dthpb_mod = dhp->dofhp_mod;
9326 		dhpb.dthpb_func = strtab + probe->dofpr_func;
9327 		dhpb.dthpb_name = strtab + probe->dofpr_name;
9328 		dhpb.dthpb_base = probe->dofpr_addr;
9329 		dhpb.dthpb_offs = off + probe->dofpr_offidx;
9330 		dhpb.dthpb_noffs = probe->dofpr_noffs;
9331 		if (enoff != NULL) {
9332 			dhpb.dthpb_enoffs = enoff + probe->dofpr_enoffidx;
9333 			dhpb.dthpb_nenoffs = probe->dofpr_nenoffs;
9334 		} else {
9335 			dhpb.dthpb_enoffs = NULL;
9336 			dhpb.dthpb_nenoffs = 0;
9337 		}
9338 		dhpb.dthpb_args = arg + probe->dofpr_argidx;
9339 		dhpb.dthpb_nargc = probe->dofpr_nargc;
9340 		dhpb.dthpb_xargc = probe->dofpr_xargc;
9341 		dhpb.dthpb_ntypes = strtab + probe->dofpr_nargv;
9342 		dhpb.dthpb_xtypes = strtab + probe->dofpr_xargv;
9343 
9344 		mops->dtms_create_probe(meta->dtm_arg, parg, &dhpb);
9345 	}
9346 }
9347 
9348 static void
9349 dtrace_helper_provide(dof_helper_t *dhp, pid_t pid)
9350 {
9351 	uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
9352 	dof_hdr_t *dof = (dof_hdr_t *)daddr;
9353 	int i;
9354 
9355 	ASSERT(MUTEX_HELD(&dtrace_meta_lock));
9356 
9357 	for (i = 0; i < dof->dofh_secnum; i++) {
9358 		dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
9359 		    dof->dofh_secoff + i * dof->dofh_secsize);
9360 
9361 		if (sec->dofs_type != DOF_SECT_PROVIDER)
9362 			continue;
9363 
9364 		dtrace_helper_provide_one(dhp, sec, pid);
9365 	}
9366 
9367 	/*
9368 	 * We may have just created probes, so we must now rematch against
9369 	 * any retained enablings.  Note that this call will acquire both
9370 	 * cpu_lock and dtrace_lock; the fact that we are holding
9371 	 * dtrace_meta_lock now is what defines the ordering with respect to
9372 	 * these three locks.
9373 	 */
9374 	dtrace_enabling_matchall();
9375 }
9376 
9377 static void
9378 dtrace_helper_provider_remove_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
9379 {
9380 	uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
9381 	dof_hdr_t *dof = (dof_hdr_t *)daddr;
9382 	dof_sec_t *str_sec;
9383 	dof_provider_t *provider;
9384 	char *strtab;
9385 	dtrace_helper_provdesc_t dhpv;
9386 	dtrace_meta_t *meta = dtrace_meta_pid;
9387 	dtrace_mops_t *mops = &meta->dtm_mops;
9388 
9389 	provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
9390 	str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
9391 	    provider->dofpv_strtab * dof->dofh_secsize);
9392 
9393 	strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
9394 
9395 	/*
9396 	 * Create the provider.
9397 	 */
9398 	dtrace_dofprov2hprov(&dhpv, provider, strtab);
9399 
9400 	mops->dtms_remove_pid(meta->dtm_arg, &dhpv, pid);
9401 
9402 	meta->dtm_count--;
9403 }
9404 
9405 static void
9406 dtrace_helper_provider_remove(dof_helper_t *dhp, pid_t pid)
9407 {
9408 	uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
9409 	dof_hdr_t *dof = (dof_hdr_t *)daddr;
9410 	int i;
9411 
9412 	ASSERT(MUTEX_HELD(&dtrace_meta_lock));
9413 
9414 	for (i = 0; i < dof->dofh_secnum; i++) {
9415 		dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
9416 		    dof->dofh_secoff + i * dof->dofh_secsize);
9417 
9418 		if (sec->dofs_type != DOF_SECT_PROVIDER)
9419 			continue;
9420 
9421 		dtrace_helper_provider_remove_one(dhp, sec, pid);
9422 	}
9423 }
9424 
9425 /*
9426  * DTrace Meta Provider-to-Framework API Functions
9427  *
9428  * These functions implement the Meta Provider-to-Framework API, as described
9429  * in <sys/dtrace.h>.
9430  */
9431 int
9432 dtrace_meta_register(const char *name, const dtrace_mops_t *mops, void *arg,
9433     dtrace_meta_provider_id_t *idp)
9434 {
9435 	dtrace_meta_t *meta;
9436 	dtrace_helpers_t *help, *next;
9437 	int i;
9438 
9439 	*idp = DTRACE_METAPROVNONE;
9440 
9441 	/*
9442 	 * We strictly don't need the name, but we hold onto it for
9443 	 * debuggability. All hail error queues!
9444 	 */
9445 	if (name == NULL) {
9446 		cmn_err(CE_WARN, "failed to register meta-provider: "
9447 		    "invalid name");
9448 		return (EINVAL);
9449 	}
9450 
9451 	if (mops == NULL ||
9452 	    mops->dtms_create_probe == NULL ||
9453 	    mops->dtms_provide_pid == NULL ||
9454 	    mops->dtms_remove_pid == NULL) {
9455 		cmn_err(CE_WARN, "failed to register meta-register %s: "
9456 		    "invalid ops", name);
9457 		return (EINVAL);
9458 	}
9459 
9460 	meta = kmem_zalloc(sizeof (dtrace_meta_t), KM_SLEEP);
9461 	meta->dtm_mops = *mops;
9462 	meta->dtm_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
9463 	(void) strcpy(meta->dtm_name, name);
9464 	meta->dtm_arg = arg;
9465 
9466 	mutex_enter(&dtrace_meta_lock);
9467 	mutex_enter(&dtrace_lock);
9468 
9469 	if (dtrace_meta_pid != NULL) {
9470 		mutex_exit(&dtrace_lock);
9471 		mutex_exit(&dtrace_meta_lock);
9472 		cmn_err(CE_WARN, "failed to register meta-register %s: "
9473 		    "user-land meta-provider exists", name);
9474 		kmem_free(meta->dtm_name, strlen(meta->dtm_name) + 1);
9475 		kmem_free(meta, sizeof (dtrace_meta_t));
9476 		return (EINVAL);
9477 	}
9478 
9479 	dtrace_meta_pid = meta;
9480 	*idp = (dtrace_meta_provider_id_t)meta;
9481 
9482 	/*
9483 	 * If there are providers and probes ready to go, pass them
9484 	 * off to the new meta provider now.
9485 	 */
9486 
9487 	help = dtrace_deferred_pid;
9488 	dtrace_deferred_pid = NULL;
9489 
9490 	mutex_exit(&dtrace_lock);
9491 
9492 	while (help != NULL) {
9493 		for (i = 0; i < help->dthps_nprovs; i++) {
9494 			dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
9495 			    help->dthps_pid);
9496 		}
9497 
9498 		next = help->dthps_next;
9499 		help->dthps_next = NULL;
9500 		help->dthps_prev = NULL;
9501 		help->dthps_deferred = 0;
9502 		help = next;
9503 	}
9504 
9505 	mutex_exit(&dtrace_meta_lock);
9506 
9507 	return (0);
9508 }
9509 
9510 int
9511 dtrace_meta_unregister(dtrace_meta_provider_id_t id)
9512 {
9513 	dtrace_meta_t **pp, *old = (dtrace_meta_t *)id;
9514 
9515 	mutex_enter(&dtrace_meta_lock);
9516 	mutex_enter(&dtrace_lock);
9517 
9518 	if (old == dtrace_meta_pid) {
9519 		pp = &dtrace_meta_pid;
9520 	} else {
9521 		panic("attempt to unregister non-existent "
9522 		    "dtrace meta-provider %p\n", (void *)old);
9523 	}
9524 
9525 	if (old->dtm_count != 0) {
9526 		mutex_exit(&dtrace_lock);
9527 		mutex_exit(&dtrace_meta_lock);
9528 		return (EBUSY);
9529 	}
9530 
9531 	*pp = NULL;
9532 
9533 	mutex_exit(&dtrace_lock);
9534 	mutex_exit(&dtrace_meta_lock);
9535 
9536 	kmem_free(old->dtm_name, strlen(old->dtm_name) + 1);
9537 	kmem_free(old, sizeof (dtrace_meta_t));
9538 
9539 	return (0);
9540 }
9541 
9542 
9543 /*
9544  * DTrace DIF Object Functions
9545  */
9546 static int
9547 dtrace_difo_err(uint_t pc, const char *format, ...)
9548 {
9549 	if (dtrace_err_verbose) {
9550 		va_list alist;
9551 
9552 		(void) uprintf("dtrace DIF object error: [%u]: ", pc);
9553 		va_start(alist, format);
9554 		(void) vuprintf(format, alist);
9555 		va_end(alist);
9556 	}
9557 
9558 #ifdef DTRACE_ERRDEBUG
9559 	dtrace_errdebug(format);
9560 #endif
9561 	return (1);
9562 }
9563 
9564 /*
9565  * Validate a DTrace DIF object by checking the IR instructions.  The following
9566  * rules are currently enforced by dtrace_difo_validate():
9567  *
9568  * 1. Each instruction must have a valid opcode
9569  * 2. Each register, string, variable, or subroutine reference must be valid
9570  * 3. No instruction can modify register %r0 (must be zero)
9571  * 4. All instruction reserved bits must be set to zero
9572  * 5. The last instruction must be a "ret" instruction
9573  * 6. All branch targets must reference a valid instruction _after_ the branch
9574  */
9575 static int
9576 dtrace_difo_validate(dtrace_difo_t *dp, dtrace_vstate_t *vstate, uint_t nregs,
9577     cred_t *cr)
9578 {
9579 	int err = 0, i;
9580 	int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
9581 	int kcheckload;
9582 	uint_t pc;
9583 
9584 	kcheckload = cr == NULL ||
9585 	    (vstate->dtvs_state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) == 0;
9586 
9587 	dp->dtdo_destructive = 0;
9588 
9589 	for (pc = 0; pc < dp->dtdo_len && err == 0; pc++) {
9590 		dif_instr_t instr = dp->dtdo_buf[pc];
9591 
9592 		uint_t r1 = DIF_INSTR_R1(instr);
9593 		uint_t r2 = DIF_INSTR_R2(instr);
9594 		uint_t rd = DIF_INSTR_RD(instr);
9595 		uint_t rs = DIF_INSTR_RS(instr);
9596 		uint_t label = DIF_INSTR_LABEL(instr);
9597 		uint_t v = DIF_INSTR_VAR(instr);
9598 		uint_t subr = DIF_INSTR_SUBR(instr);
9599 		uint_t type = DIF_INSTR_TYPE(instr);
9600 		uint_t op = DIF_INSTR_OP(instr);
9601 
9602 		switch (op) {
9603 		case DIF_OP_OR:
9604 		case DIF_OP_XOR:
9605 		case DIF_OP_AND:
9606 		case DIF_OP_SLL:
9607 		case DIF_OP_SRL:
9608 		case DIF_OP_SRA:
9609 		case DIF_OP_SUB:
9610 		case DIF_OP_ADD:
9611 		case DIF_OP_MUL:
9612 		case DIF_OP_SDIV:
9613 		case DIF_OP_UDIV:
9614 		case DIF_OP_SREM:
9615 		case DIF_OP_UREM:
9616 		case DIF_OP_COPYS:
9617 			if (r1 >= nregs)
9618 				err += efunc(pc, "invalid register %u\n", r1);
9619 			if (r2 >= nregs)
9620 				err += efunc(pc, "invalid register %u\n", r2);
9621 			if (rd >= nregs)
9622 				err += efunc(pc, "invalid register %u\n", rd);
9623 			if (rd == 0)
9624 				err += efunc(pc, "cannot write to %r0\n");
9625 			break;
9626 		case DIF_OP_NOT:
9627 		case DIF_OP_MOV:
9628 		case DIF_OP_ALLOCS:
9629 			if (r1 >= nregs)
9630 				err += efunc(pc, "invalid register %u\n", r1);
9631 			if (r2 != 0)
9632 				err += efunc(pc, "non-zero reserved bits\n");
9633 			if (rd >= nregs)
9634 				err += efunc(pc, "invalid register %u\n", rd);
9635 			if (rd == 0)
9636 				err += efunc(pc, "cannot write to %r0\n");
9637 			break;
9638 		case DIF_OP_LDSB:
9639 		case DIF_OP_LDSH:
9640 		case DIF_OP_LDSW:
9641 		case DIF_OP_LDUB:
9642 		case DIF_OP_LDUH:
9643 		case DIF_OP_LDUW:
9644 		case DIF_OP_LDX:
9645 			if (r1 >= nregs)
9646 				err += efunc(pc, "invalid register %u\n", r1);
9647 			if (r2 != 0)
9648 				err += efunc(pc, "non-zero reserved bits\n");
9649 			if (rd >= nregs)
9650 				err += efunc(pc, "invalid register %u\n", rd);
9651 			if (rd == 0)
9652 				err += efunc(pc, "cannot write to %r0\n");
9653 			if (kcheckload)
9654 				dp->dtdo_buf[pc] = DIF_INSTR_LOAD(op +
9655 				    DIF_OP_RLDSB - DIF_OP_LDSB, r1, rd);
9656 			break;
9657 		case DIF_OP_RLDSB:
9658 		case DIF_OP_RLDSH:
9659 		case DIF_OP_RLDSW:
9660 		case DIF_OP_RLDUB:
9661 		case DIF_OP_RLDUH:
9662 		case DIF_OP_RLDUW:
9663 		case DIF_OP_RLDX:
9664 			if (r1 >= nregs)
9665 				err += efunc(pc, "invalid register %u\n", r1);
9666 			if (r2 != 0)
9667 				err += efunc(pc, "non-zero reserved bits\n");
9668 			if (rd >= nregs)
9669 				err += efunc(pc, "invalid register %u\n", rd);
9670 			if (rd == 0)
9671 				err += efunc(pc, "cannot write to %r0\n");
9672 			break;
9673 		case DIF_OP_ULDSB:
9674 		case DIF_OP_ULDSH:
9675 		case DIF_OP_ULDSW:
9676 		case DIF_OP_ULDUB:
9677 		case DIF_OP_ULDUH:
9678 		case DIF_OP_ULDUW:
9679 		case DIF_OP_ULDX:
9680 			if (r1 >= nregs)
9681 				err += efunc(pc, "invalid register %u\n", r1);
9682 			if (r2 != 0)
9683 				err += efunc(pc, "non-zero reserved bits\n");
9684 			if (rd >= nregs)
9685 				err += efunc(pc, "invalid register %u\n", rd);
9686 			if (rd == 0)
9687 				err += efunc(pc, "cannot write to %r0\n");
9688 			break;
9689 		case DIF_OP_STB:
9690 		case DIF_OP_STH:
9691 		case DIF_OP_STW:
9692 		case DIF_OP_STX:
9693 			if (r1 >= nregs)
9694 				err += efunc(pc, "invalid register %u\n", r1);
9695 			if (r2 != 0)
9696 				err += efunc(pc, "non-zero reserved bits\n");
9697 			if (rd >= nregs)
9698 				err += efunc(pc, "invalid register %u\n", rd);
9699 			if (rd == 0)
9700 				err += efunc(pc, "cannot write to 0 address\n");
9701 			break;
9702 		case DIF_OP_CMP:
9703 		case DIF_OP_SCMP:
9704 			if (r1 >= nregs)
9705 				err += efunc(pc, "invalid register %u\n", r1);
9706 			if (r2 >= nregs)
9707 				err += efunc(pc, "invalid register %u\n", r2);
9708 			if (rd != 0)
9709 				err += efunc(pc, "non-zero reserved bits\n");
9710 			break;
9711 		case DIF_OP_TST:
9712 			if (r1 >= nregs)
9713 				err += efunc(pc, "invalid register %u\n", r1);
9714 			if (r2 != 0 || rd != 0)
9715 				err += efunc(pc, "non-zero reserved bits\n");
9716 			break;
9717 		case DIF_OP_BA:
9718 		case DIF_OP_BE:
9719 		case DIF_OP_BNE:
9720 		case DIF_OP_BG:
9721 		case DIF_OP_BGU:
9722 		case DIF_OP_BGE:
9723 		case DIF_OP_BGEU:
9724 		case DIF_OP_BL:
9725 		case DIF_OP_BLU:
9726 		case DIF_OP_BLE:
9727 		case DIF_OP_BLEU:
9728 			if (label >= dp->dtdo_len) {
9729 				err += efunc(pc, "invalid branch target %u\n",
9730 				    label);
9731 			}
9732 			if (label <= pc) {
9733 				err += efunc(pc, "backward branch to %u\n",
9734 				    label);
9735 			}
9736 			break;
9737 		case DIF_OP_RET:
9738 			if (r1 != 0 || r2 != 0)
9739 				err += efunc(pc, "non-zero reserved bits\n");
9740 			if (rd >= nregs)
9741 				err += efunc(pc, "invalid register %u\n", rd);
9742 			break;
9743 		case DIF_OP_NOP:
9744 		case DIF_OP_POPTS:
9745 		case DIF_OP_FLUSHTS:
9746 			if (r1 != 0 || r2 != 0 || rd != 0)
9747 				err += efunc(pc, "non-zero reserved bits\n");
9748 			break;
9749 		case DIF_OP_SETX:
9750 			if (DIF_INSTR_INTEGER(instr) >= dp->dtdo_intlen) {
9751 				err += efunc(pc, "invalid integer ref %u\n",
9752 				    DIF_INSTR_INTEGER(instr));
9753 			}
9754 			if (rd >= nregs)
9755 				err += efunc(pc, "invalid register %u\n", rd);
9756 			if (rd == 0)
9757 				err += efunc(pc, "cannot write to %r0\n");
9758 			break;
9759 		case DIF_OP_SETS:
9760 			if (DIF_INSTR_STRING(instr) >= dp->dtdo_strlen) {
9761 				err += efunc(pc, "invalid string ref %u\n",
9762 				    DIF_INSTR_STRING(instr));
9763 			}
9764 			if (rd >= nregs)
9765 				err += efunc(pc, "invalid register %u\n", rd);
9766 			if (rd == 0)
9767 				err += efunc(pc, "cannot write to %r0\n");
9768 			break;
9769 		case DIF_OP_LDGA:
9770 		case DIF_OP_LDTA:
9771 			if (r1 > DIF_VAR_ARRAY_MAX)
9772 				err += efunc(pc, "invalid array %u\n", r1);
9773 			if (r2 >= nregs)
9774 				err += efunc(pc, "invalid register %u\n", r2);
9775 			if (rd >= nregs)
9776 				err += efunc(pc, "invalid register %u\n", rd);
9777 			if (rd == 0)
9778 				err += efunc(pc, "cannot write to %r0\n");
9779 			break;
9780 		case DIF_OP_LDGS:
9781 		case DIF_OP_LDTS:
9782 		case DIF_OP_LDLS:
9783 		case DIF_OP_LDGAA:
9784 		case DIF_OP_LDTAA:
9785 			if (v < DIF_VAR_OTHER_MIN || v > DIF_VAR_OTHER_MAX)
9786 				err += efunc(pc, "invalid variable %u\n", v);
9787 			if (rd >= nregs)
9788 				err += efunc(pc, "invalid register %u\n", rd);
9789 			if (rd == 0)
9790 				err += efunc(pc, "cannot write to %r0\n");
9791 			break;
9792 		case DIF_OP_STGS:
9793 		case DIF_OP_STTS:
9794 		case DIF_OP_STLS:
9795 		case DIF_OP_STGAA:
9796 		case DIF_OP_STTAA:
9797 			if (v < DIF_VAR_OTHER_UBASE || v > DIF_VAR_OTHER_MAX)
9798 				err += efunc(pc, "invalid variable %u\n", v);
9799 			if (rs >= nregs)
9800 				err += efunc(pc, "invalid register %u\n", rd);
9801 			break;
9802 		case DIF_OP_CALL:
9803 			if (subr > DIF_SUBR_MAX)
9804 				err += efunc(pc, "invalid subr %u\n", subr);
9805 			if (rd >= nregs)
9806 				err += efunc(pc, "invalid register %u\n", rd);
9807 			if (rd == 0)
9808 				err += efunc(pc, "cannot write to %r0\n");
9809 
9810 			if (subr == DIF_SUBR_COPYOUT ||
9811 			    subr == DIF_SUBR_COPYOUTSTR) {
9812 				dp->dtdo_destructive = 1;
9813 			}
9814 
9815 			if (subr == DIF_SUBR_GETF) {
9816 				/*
9817 				 * If we have a getf() we need to record that
9818 				 * in our state.  Note that our state can be
9819 				 * NULL if this is a helper -- but in that
9820 				 * case, the call to getf() is itself illegal,
9821 				 * and will be caught (slightly later) when
9822 				 * the helper is validated.
9823 				 */
9824 				if (vstate->dtvs_state != NULL)
9825 					vstate->dtvs_state->dts_getf++;
9826 			}
9827 
9828 			break;
9829 		case DIF_OP_PUSHTR:
9830 			if (type != DIF_TYPE_STRING && type != DIF_TYPE_CTF)
9831 				err += efunc(pc, "invalid ref type %u\n", type);
9832 			if (r2 >= nregs)
9833 				err += efunc(pc, "invalid register %u\n", r2);
9834 			if (rs >= nregs)
9835 				err += efunc(pc, "invalid register %u\n", rs);
9836 			break;
9837 		case DIF_OP_PUSHTV:
9838 			if (type != DIF_TYPE_CTF)
9839 				err += efunc(pc, "invalid val type %u\n", type);
9840 			if (r2 >= nregs)
9841 				err += efunc(pc, "invalid register %u\n", r2);
9842 			if (rs >= nregs)
9843 				err += efunc(pc, "invalid register %u\n", rs);
9844 			break;
9845 		default:
9846 			err += efunc(pc, "invalid opcode %u\n",
9847 			    DIF_INSTR_OP(instr));
9848 		}
9849 	}
9850 
9851 	if (dp->dtdo_len != 0 &&
9852 	    DIF_INSTR_OP(dp->dtdo_buf[dp->dtdo_len - 1]) != DIF_OP_RET) {
9853 		err += efunc(dp->dtdo_len - 1,
9854 		    "expected 'ret' as last DIF instruction\n");
9855 	}
9856 
9857 	if (!(dp->dtdo_rtype.dtdt_flags & (DIF_TF_BYREF | DIF_TF_BYUREF))) {
9858 		/*
9859 		 * If we're not returning by reference, the size must be either
9860 		 * 0 or the size of one of the base types.
9861 		 */
9862 		switch (dp->dtdo_rtype.dtdt_size) {
9863 		case 0:
9864 		case sizeof (uint8_t):
9865 		case sizeof (uint16_t):
9866 		case sizeof (uint32_t):
9867 		case sizeof (uint64_t):
9868 			break;
9869 
9870 		default:
9871 			err += efunc(dp->dtdo_len - 1, "bad return size\n");
9872 		}
9873 	}
9874 
9875 	for (i = 0; i < dp->dtdo_varlen && err == 0; i++) {
9876 		dtrace_difv_t *v = &dp->dtdo_vartab[i], *existing = NULL;
9877 		dtrace_diftype_t *vt, *et;
9878 		uint_t id, ndx;
9879 
9880 		if (v->dtdv_scope != DIFV_SCOPE_GLOBAL &&
9881 		    v->dtdv_scope != DIFV_SCOPE_THREAD &&
9882 		    v->dtdv_scope != DIFV_SCOPE_LOCAL) {
9883 			err += efunc(i, "unrecognized variable scope %d\n",
9884 			    v->dtdv_scope);
9885 			break;
9886 		}
9887 
9888 		if (v->dtdv_kind != DIFV_KIND_ARRAY &&
9889 		    v->dtdv_kind != DIFV_KIND_SCALAR) {
9890 			err += efunc(i, "unrecognized variable type %d\n",
9891 			    v->dtdv_kind);
9892 			break;
9893 		}
9894 
9895 		if ((id = v->dtdv_id) > DIF_VARIABLE_MAX) {
9896 			err += efunc(i, "%d exceeds variable id limit\n", id);
9897 			break;
9898 		}
9899 
9900 		if (id < DIF_VAR_OTHER_UBASE)
9901 			continue;
9902 
9903 		/*
9904 		 * For user-defined variables, we need to check that this
9905 		 * definition is identical to any previous definition that we
9906 		 * encountered.
9907 		 */
9908 		ndx = id - DIF_VAR_OTHER_UBASE;
9909 
9910 		switch (v->dtdv_scope) {
9911 		case DIFV_SCOPE_GLOBAL:
9912 			if (ndx < vstate->dtvs_nglobals) {
9913 				dtrace_statvar_t *svar;
9914 
9915 				if ((svar = vstate->dtvs_globals[ndx]) != NULL)
9916 					existing = &svar->dtsv_var;
9917 			}
9918 
9919 			break;
9920 
9921 		case DIFV_SCOPE_THREAD:
9922 			if (ndx < vstate->dtvs_ntlocals)
9923 				existing = &vstate->dtvs_tlocals[ndx];
9924 			break;
9925 
9926 		case DIFV_SCOPE_LOCAL:
9927 			if (ndx < vstate->dtvs_nlocals) {
9928 				dtrace_statvar_t *svar;
9929 
9930 				if ((svar = vstate->dtvs_locals[ndx]) != NULL)
9931 					existing = &svar->dtsv_var;
9932 			}
9933 
9934 			break;
9935 		}
9936 
9937 		vt = &v->dtdv_type;
9938 
9939 		if (vt->dtdt_flags & DIF_TF_BYREF) {
9940 			if (vt->dtdt_size == 0) {
9941 				err += efunc(i, "zero-sized variable\n");
9942 				break;
9943 			}
9944 
9945 			if ((v->dtdv_scope == DIFV_SCOPE_GLOBAL ||
9946 			    v->dtdv_scope == DIFV_SCOPE_LOCAL) &&
9947 			    vt->dtdt_size > dtrace_statvar_maxsize) {
9948 				err += efunc(i, "oversized by-ref static\n");
9949 				break;
9950 			}
9951 		}
9952 
9953 		if (existing == NULL || existing->dtdv_id == 0)
9954 			continue;
9955 
9956 		ASSERT(existing->dtdv_id == v->dtdv_id);
9957 		ASSERT(existing->dtdv_scope == v->dtdv_scope);
9958 
9959 		if (existing->dtdv_kind != v->dtdv_kind)
9960 			err += efunc(i, "%d changed variable kind\n", id);
9961 
9962 		et = &existing->dtdv_type;
9963 
9964 		if (vt->dtdt_flags != et->dtdt_flags) {
9965 			err += efunc(i, "%d changed variable type flags\n", id);
9966 			break;
9967 		}
9968 
9969 		if (vt->dtdt_size != 0 && vt->dtdt_size != et->dtdt_size) {
9970 			err += efunc(i, "%d changed variable type size\n", id);
9971 			break;
9972 		}
9973 	}
9974 
9975 	return (err);
9976 }
9977 
9978 /*
9979  * Validate a DTrace DIF object that it is to be used as a helper.  Helpers
9980  * are much more constrained than normal DIFOs.  Specifically, they may
9981  * not:
9982  *
9983  * 1. Make calls to subroutines other than copyin(), copyinstr() or
9984  *    miscellaneous string routines
9985  * 2. Access DTrace variables other than the args[] array, and the
9986  *    curthread, pid, ppid, tid, execname, zonename, uid and gid variables.
9987  * 3. Have thread-local variables.
9988  * 4. Have dynamic variables.
9989  */
9990 static int
9991 dtrace_difo_validate_helper(dtrace_difo_t *dp)
9992 {
9993 	int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
9994 	int err = 0;
9995 	uint_t pc;
9996 
9997 	for (pc = 0; pc < dp->dtdo_len; pc++) {
9998 		dif_instr_t instr = dp->dtdo_buf[pc];
9999 
10000 		uint_t v = DIF_INSTR_VAR(instr);
10001 		uint_t subr = DIF_INSTR_SUBR(instr);
10002 		uint_t op = DIF_INSTR_OP(instr);
10003 
10004 		switch (op) {
10005 		case DIF_OP_OR:
10006 		case DIF_OP_XOR:
10007 		case DIF_OP_AND:
10008 		case DIF_OP_SLL:
10009 		case DIF_OP_SRL:
10010 		case DIF_OP_SRA:
10011 		case DIF_OP_SUB:
10012 		case DIF_OP_ADD:
10013 		case DIF_OP_MUL:
10014 		case DIF_OP_SDIV:
10015 		case DIF_OP_UDIV:
10016 		case DIF_OP_SREM:
10017 		case DIF_OP_UREM:
10018 		case DIF_OP_COPYS:
10019 		case DIF_OP_NOT:
10020 		case DIF_OP_MOV:
10021 		case DIF_OP_RLDSB:
10022 		case DIF_OP_RLDSH:
10023 		case DIF_OP_RLDSW:
10024 		case DIF_OP_RLDUB:
10025 		case DIF_OP_RLDUH:
10026 		case DIF_OP_RLDUW:
10027 		case DIF_OP_RLDX:
10028 		case DIF_OP_ULDSB:
10029 		case DIF_OP_ULDSH:
10030 		case DIF_OP_ULDSW:
10031 		case DIF_OP_ULDUB:
10032 		case DIF_OP_ULDUH:
10033 		case DIF_OP_ULDUW:
10034 		case DIF_OP_ULDX:
10035 		case DIF_OP_STB:
10036 		case DIF_OP_STH:
10037 		case DIF_OP_STW:
10038 		case DIF_OP_STX:
10039 		case DIF_OP_ALLOCS:
10040 		case DIF_OP_CMP:
10041 		case DIF_OP_SCMP:
10042 		case DIF_OP_TST:
10043 		case DIF_OP_BA:
10044 		case DIF_OP_BE:
10045 		case DIF_OP_BNE:
10046 		case DIF_OP_BG:
10047 		case DIF_OP_BGU:
10048 		case DIF_OP_BGE:
10049 		case DIF_OP_BGEU:
10050 		case DIF_OP_BL:
10051 		case DIF_OP_BLU:
10052 		case DIF_OP_BLE:
10053 		case DIF_OP_BLEU:
10054 		case DIF_OP_RET:
10055 		case DIF_OP_NOP:
10056 		case DIF_OP_POPTS:
10057 		case DIF_OP_FLUSHTS:
10058 		case DIF_OP_SETX:
10059 		case DIF_OP_SETS:
10060 		case DIF_OP_LDGA:
10061 		case DIF_OP_LDLS:
10062 		case DIF_OP_STGS:
10063 		case DIF_OP_STLS:
10064 		case DIF_OP_PUSHTR:
10065 		case DIF_OP_PUSHTV:
10066 			break;
10067 
10068 		case DIF_OP_LDGS:
10069 			if (v >= DIF_VAR_OTHER_UBASE)
10070 				break;
10071 
10072 			if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9)
10073 				break;
10074 
10075 			if (v == DIF_VAR_CURTHREAD || v == DIF_VAR_PID ||
10076 			    v == DIF_VAR_PPID || v == DIF_VAR_TID ||
10077 			    v == DIF_VAR_EXECARGS ||
10078 			    v == DIF_VAR_EXECNAME || v == DIF_VAR_ZONENAME ||
10079 			    v == DIF_VAR_UID || v == DIF_VAR_GID)
10080 				break;
10081 
10082 			err += efunc(pc, "illegal variable %u\n", v);
10083 			break;
10084 
10085 		case DIF_OP_LDTA:
10086 		case DIF_OP_LDTS:
10087 		case DIF_OP_LDGAA:
10088 		case DIF_OP_LDTAA:
10089 			err += efunc(pc, "illegal dynamic variable load\n");
10090 			break;
10091 
10092 		case DIF_OP_STTS:
10093 		case DIF_OP_STGAA:
10094 		case DIF_OP_STTAA:
10095 			err += efunc(pc, "illegal dynamic variable store\n");
10096 			break;
10097 
10098 		case DIF_OP_CALL:
10099 			if (subr == DIF_SUBR_ALLOCA ||
10100 			    subr == DIF_SUBR_BCOPY ||
10101 			    subr == DIF_SUBR_COPYIN ||
10102 			    subr == DIF_SUBR_COPYINTO ||
10103 			    subr == DIF_SUBR_COPYINSTR ||
10104 			    subr == DIF_SUBR_INDEX ||
10105 			    subr == DIF_SUBR_INET_NTOA ||
10106 			    subr == DIF_SUBR_INET_NTOA6 ||
10107 			    subr == DIF_SUBR_INET_NTOP ||
10108 			    subr == DIF_SUBR_JSON ||
10109 			    subr == DIF_SUBR_LLTOSTR ||
10110 			    subr == DIF_SUBR_STRTOLL ||
10111 			    subr == DIF_SUBR_RINDEX ||
10112 			    subr == DIF_SUBR_STRCHR ||
10113 			    subr == DIF_SUBR_STRJOIN ||
10114 			    subr == DIF_SUBR_STRRCHR ||
10115 			    subr == DIF_SUBR_STRSTR ||
10116 			    subr == DIF_SUBR_HTONS ||
10117 			    subr == DIF_SUBR_HTONL ||
10118 			    subr == DIF_SUBR_HTONLL ||
10119 			    subr == DIF_SUBR_NTOHS ||
10120 			    subr == DIF_SUBR_NTOHL ||
10121 			    subr == DIF_SUBR_NTOHLL ||
10122 			    subr == DIF_SUBR_MEMREF ||
10123 #ifndef illumos
10124 			    subr == DIF_SUBR_MEMSTR ||
10125 #endif
10126 			    subr == DIF_SUBR_TYPEREF)
10127 				break;
10128 
10129 			err += efunc(pc, "invalid subr %u\n", subr);
10130 			break;
10131 
10132 		default:
10133 			err += efunc(pc, "invalid opcode %u\n",
10134 			    DIF_INSTR_OP(instr));
10135 		}
10136 	}
10137 
10138 	return (err);
10139 }
10140 
10141 /*
10142  * Returns 1 if the expression in the DIF object can be cached on a per-thread
10143  * basis; 0 if not.
10144  */
10145 static int
10146 dtrace_difo_cacheable(dtrace_difo_t *dp)
10147 {
10148 	int i;
10149 
10150 	if (dp == NULL)
10151 		return (0);
10152 
10153 	for (i = 0; i < dp->dtdo_varlen; i++) {
10154 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
10155 
10156 		if (v->dtdv_scope != DIFV_SCOPE_GLOBAL)
10157 			continue;
10158 
10159 		switch (v->dtdv_id) {
10160 		case DIF_VAR_CURTHREAD:
10161 		case DIF_VAR_PID:
10162 		case DIF_VAR_TID:
10163 		case DIF_VAR_EXECARGS:
10164 		case DIF_VAR_EXECNAME:
10165 		case DIF_VAR_ZONENAME:
10166 			break;
10167 
10168 		default:
10169 			return (0);
10170 		}
10171 	}
10172 
10173 	/*
10174 	 * This DIF object may be cacheable.  Now we need to look for any
10175 	 * array loading instructions, any memory loading instructions, or
10176 	 * any stores to thread-local variables.
10177 	 */
10178 	for (i = 0; i < dp->dtdo_len; i++) {
10179 		uint_t op = DIF_INSTR_OP(dp->dtdo_buf[i]);
10180 
10181 		if ((op >= DIF_OP_LDSB && op <= DIF_OP_LDX) ||
10182 		    (op >= DIF_OP_ULDSB && op <= DIF_OP_ULDX) ||
10183 		    (op >= DIF_OP_RLDSB && op <= DIF_OP_RLDX) ||
10184 		    op == DIF_OP_LDGA || op == DIF_OP_STTS)
10185 			return (0);
10186 	}
10187 
10188 	return (1);
10189 }
10190 
10191 static void
10192 dtrace_difo_hold(dtrace_difo_t *dp)
10193 {
10194 	int i;
10195 
10196 	ASSERT(MUTEX_HELD(&dtrace_lock));
10197 
10198 	dp->dtdo_refcnt++;
10199 	ASSERT(dp->dtdo_refcnt != 0);
10200 
10201 	/*
10202 	 * We need to check this DIF object for references to the variable
10203 	 * DIF_VAR_VTIMESTAMP.
10204 	 */
10205 	for (i = 0; i < dp->dtdo_varlen; i++) {
10206 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
10207 
10208 		if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
10209 			continue;
10210 
10211 		if (dtrace_vtime_references++ == 0)
10212 			dtrace_vtime_enable();
10213 	}
10214 }
10215 
10216 /*
10217  * This routine calculates the dynamic variable chunksize for a given DIF
10218  * object.  The calculation is not fool-proof, and can probably be tricked by
10219  * malicious DIF -- but it works for all compiler-generated DIF.  Because this
10220  * calculation is likely imperfect, dtrace_dynvar() is able to gracefully fail
10221  * if a dynamic variable size exceeds the chunksize.
10222  */
10223 static void
10224 dtrace_difo_chunksize(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
10225 {
10226 	uint64_t sval = 0;
10227 	dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
10228 	const dif_instr_t *text = dp->dtdo_buf;
10229 	uint_t pc, srd = 0;
10230 	uint_t ttop = 0;
10231 	size_t size, ksize;
10232 	uint_t id, i;
10233 
10234 	for (pc = 0; pc < dp->dtdo_len; pc++) {
10235 		dif_instr_t instr = text[pc];
10236 		uint_t op = DIF_INSTR_OP(instr);
10237 		uint_t rd = DIF_INSTR_RD(instr);
10238 		uint_t r1 = DIF_INSTR_R1(instr);
10239 		uint_t nkeys = 0;
10240 		uchar_t scope = 0;
10241 
10242 		dtrace_key_t *key = tupregs;
10243 
10244 		switch (op) {
10245 		case DIF_OP_SETX:
10246 			sval = dp->dtdo_inttab[DIF_INSTR_INTEGER(instr)];
10247 			srd = rd;
10248 			continue;
10249 
10250 		case DIF_OP_STTS:
10251 			key = &tupregs[DIF_DTR_NREGS];
10252 			key[0].dttk_size = 0;
10253 			key[1].dttk_size = 0;
10254 			nkeys = 2;
10255 			scope = DIFV_SCOPE_THREAD;
10256 			break;
10257 
10258 		case DIF_OP_STGAA:
10259 		case DIF_OP_STTAA:
10260 			nkeys = ttop;
10261 
10262 			if (DIF_INSTR_OP(instr) == DIF_OP_STTAA)
10263 				key[nkeys++].dttk_size = 0;
10264 
10265 			key[nkeys++].dttk_size = 0;
10266 
10267 			if (op == DIF_OP_STTAA) {
10268 				scope = DIFV_SCOPE_THREAD;
10269 			} else {
10270 				scope = DIFV_SCOPE_GLOBAL;
10271 			}
10272 
10273 			break;
10274 
10275 		case DIF_OP_PUSHTR:
10276 			if (ttop == DIF_DTR_NREGS)
10277 				return;
10278 
10279 			if ((srd == 0 || sval == 0) && r1 == DIF_TYPE_STRING) {
10280 				/*
10281 				 * If the register for the size of the "pushtr"
10282 				 * is %r0 (or the value is 0) and the type is
10283 				 * a string, we'll use the system-wide default
10284 				 * string size.
10285 				 */
10286 				tupregs[ttop++].dttk_size =
10287 				    dtrace_strsize_default;
10288 			} else {
10289 				if (srd == 0)
10290 					return;
10291 
10292 				if (sval > LONG_MAX)
10293 					return;
10294 
10295 				tupregs[ttop++].dttk_size = sval;
10296 			}
10297 
10298 			break;
10299 
10300 		case DIF_OP_PUSHTV:
10301 			if (ttop == DIF_DTR_NREGS)
10302 				return;
10303 
10304 			tupregs[ttop++].dttk_size = 0;
10305 			break;
10306 
10307 		case DIF_OP_FLUSHTS:
10308 			ttop = 0;
10309 			break;
10310 
10311 		case DIF_OP_POPTS:
10312 			if (ttop != 0)
10313 				ttop--;
10314 			break;
10315 		}
10316 
10317 		sval = 0;
10318 		srd = 0;
10319 
10320 		if (nkeys == 0)
10321 			continue;
10322 
10323 		/*
10324 		 * We have a dynamic variable allocation; calculate its size.
10325 		 */
10326 		for (ksize = 0, i = 0; i < nkeys; i++)
10327 			ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
10328 
10329 		size = sizeof (dtrace_dynvar_t);
10330 		size += sizeof (dtrace_key_t) * (nkeys - 1);
10331 		size += ksize;
10332 
10333 		/*
10334 		 * Now we need to determine the size of the stored data.
10335 		 */
10336 		id = DIF_INSTR_VAR(instr);
10337 
10338 		for (i = 0; i < dp->dtdo_varlen; i++) {
10339 			dtrace_difv_t *v = &dp->dtdo_vartab[i];
10340 
10341 			if (v->dtdv_id == id && v->dtdv_scope == scope) {
10342 				size += v->dtdv_type.dtdt_size;
10343 				break;
10344 			}
10345 		}
10346 
10347 		if (i == dp->dtdo_varlen)
10348 			return;
10349 
10350 		/*
10351 		 * We have the size.  If this is larger than the chunk size
10352 		 * for our dynamic variable state, reset the chunk size.
10353 		 */
10354 		size = P2ROUNDUP(size, sizeof (uint64_t));
10355 
10356 		/*
10357 		 * Before setting the chunk size, check that we're not going
10358 		 * to set it to a negative value...
10359 		 */
10360 		if (size > LONG_MAX)
10361 			return;
10362 
10363 		/*
10364 		 * ...and make certain that we didn't badly overflow.
10365 		 */
10366 		if (size < ksize || size < sizeof (dtrace_dynvar_t))
10367 			return;
10368 
10369 		if (size > vstate->dtvs_dynvars.dtds_chunksize)
10370 			vstate->dtvs_dynvars.dtds_chunksize = size;
10371 	}
10372 }
10373 
10374 static void
10375 dtrace_difo_init(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
10376 {
10377 	int i, oldsvars, osz, nsz, otlocals, ntlocals;
10378 	uint_t id;
10379 
10380 	ASSERT(MUTEX_HELD(&dtrace_lock));
10381 	ASSERT(dp->dtdo_buf != NULL && dp->dtdo_len != 0);
10382 
10383 	for (i = 0; i < dp->dtdo_varlen; i++) {
10384 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
10385 		dtrace_statvar_t *svar, ***svarp = NULL;
10386 		size_t dsize = 0;
10387 		uint8_t scope = v->dtdv_scope;
10388 		int *np = NULL;
10389 
10390 		if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
10391 			continue;
10392 
10393 		id -= DIF_VAR_OTHER_UBASE;
10394 
10395 		switch (scope) {
10396 		case DIFV_SCOPE_THREAD:
10397 			while (id >= (otlocals = vstate->dtvs_ntlocals)) {
10398 				dtrace_difv_t *tlocals;
10399 
10400 				if ((ntlocals = (otlocals << 1)) == 0)
10401 					ntlocals = 1;
10402 
10403 				osz = otlocals * sizeof (dtrace_difv_t);
10404 				nsz = ntlocals * sizeof (dtrace_difv_t);
10405 
10406 				tlocals = kmem_zalloc(nsz, KM_SLEEP);
10407 
10408 				if (osz != 0) {
10409 					bcopy(vstate->dtvs_tlocals,
10410 					    tlocals, osz);
10411 					kmem_free(vstate->dtvs_tlocals, osz);
10412 				}
10413 
10414 				vstate->dtvs_tlocals = tlocals;
10415 				vstate->dtvs_ntlocals = ntlocals;
10416 			}
10417 
10418 			vstate->dtvs_tlocals[id] = *v;
10419 			continue;
10420 
10421 		case DIFV_SCOPE_LOCAL:
10422 			np = &vstate->dtvs_nlocals;
10423 			svarp = &vstate->dtvs_locals;
10424 
10425 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
10426 				dsize = NCPU * (v->dtdv_type.dtdt_size +
10427 				    sizeof (uint64_t));
10428 			else
10429 				dsize = NCPU * sizeof (uint64_t);
10430 
10431 			break;
10432 
10433 		case DIFV_SCOPE_GLOBAL:
10434 			np = &vstate->dtvs_nglobals;
10435 			svarp = &vstate->dtvs_globals;
10436 
10437 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
10438 				dsize = v->dtdv_type.dtdt_size +
10439 				    sizeof (uint64_t);
10440 
10441 			break;
10442 
10443 		default:
10444 			ASSERT(0);
10445 		}
10446 
10447 		while (id >= (oldsvars = *np)) {
10448 			dtrace_statvar_t **statics;
10449 			int newsvars, oldsize, newsize;
10450 
10451 			if ((newsvars = (oldsvars << 1)) == 0)
10452 				newsvars = 1;
10453 
10454 			oldsize = oldsvars * sizeof (dtrace_statvar_t *);
10455 			newsize = newsvars * sizeof (dtrace_statvar_t *);
10456 
10457 			statics = kmem_zalloc(newsize, KM_SLEEP);
10458 
10459 			if (oldsize != 0) {
10460 				bcopy(*svarp, statics, oldsize);
10461 				kmem_free(*svarp, oldsize);
10462 			}
10463 
10464 			*svarp = statics;
10465 			*np = newsvars;
10466 		}
10467 
10468 		if ((svar = (*svarp)[id]) == NULL) {
10469 			svar = kmem_zalloc(sizeof (dtrace_statvar_t), KM_SLEEP);
10470 			svar->dtsv_var = *v;
10471 
10472 			if ((svar->dtsv_size = dsize) != 0) {
10473 				svar->dtsv_data = (uint64_t)(uintptr_t)
10474 				    kmem_zalloc(dsize, KM_SLEEP);
10475 			}
10476 
10477 			(*svarp)[id] = svar;
10478 		}
10479 
10480 		svar->dtsv_refcnt++;
10481 	}
10482 
10483 	dtrace_difo_chunksize(dp, vstate);
10484 	dtrace_difo_hold(dp);
10485 }
10486 
10487 static dtrace_difo_t *
10488 dtrace_difo_duplicate(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
10489 {
10490 	dtrace_difo_t *new;
10491 	size_t sz;
10492 
10493 	ASSERT(dp->dtdo_buf != NULL);
10494 	ASSERT(dp->dtdo_refcnt != 0);
10495 
10496 	new = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
10497 
10498 	ASSERT(dp->dtdo_buf != NULL);
10499 	sz = dp->dtdo_len * sizeof (dif_instr_t);
10500 	new->dtdo_buf = kmem_alloc(sz, KM_SLEEP);
10501 	bcopy(dp->dtdo_buf, new->dtdo_buf, sz);
10502 	new->dtdo_len = dp->dtdo_len;
10503 
10504 	if (dp->dtdo_strtab != NULL) {
10505 		ASSERT(dp->dtdo_strlen != 0);
10506 		new->dtdo_strtab = kmem_alloc(dp->dtdo_strlen, KM_SLEEP);
10507 		bcopy(dp->dtdo_strtab, new->dtdo_strtab, dp->dtdo_strlen);
10508 		new->dtdo_strlen = dp->dtdo_strlen;
10509 	}
10510 
10511 	if (dp->dtdo_inttab != NULL) {
10512 		ASSERT(dp->dtdo_intlen != 0);
10513 		sz = dp->dtdo_intlen * sizeof (uint64_t);
10514 		new->dtdo_inttab = kmem_alloc(sz, KM_SLEEP);
10515 		bcopy(dp->dtdo_inttab, new->dtdo_inttab, sz);
10516 		new->dtdo_intlen = dp->dtdo_intlen;
10517 	}
10518 
10519 	if (dp->dtdo_vartab != NULL) {
10520 		ASSERT(dp->dtdo_varlen != 0);
10521 		sz = dp->dtdo_varlen * sizeof (dtrace_difv_t);
10522 		new->dtdo_vartab = kmem_alloc(sz, KM_SLEEP);
10523 		bcopy(dp->dtdo_vartab, new->dtdo_vartab, sz);
10524 		new->dtdo_varlen = dp->dtdo_varlen;
10525 	}
10526 
10527 	dtrace_difo_init(new, vstate);
10528 	return (new);
10529 }
10530 
10531 static void
10532 dtrace_difo_destroy(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
10533 {
10534 	int i;
10535 
10536 	ASSERT(dp->dtdo_refcnt == 0);
10537 
10538 	for (i = 0; i < dp->dtdo_varlen; i++) {
10539 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
10540 		dtrace_statvar_t *svar, **svarp = NULL;
10541 		uint_t id;
10542 		uint8_t scope = v->dtdv_scope;
10543 		int *np = NULL;
10544 
10545 		switch (scope) {
10546 		case DIFV_SCOPE_THREAD:
10547 			continue;
10548 
10549 		case DIFV_SCOPE_LOCAL:
10550 			np = &vstate->dtvs_nlocals;
10551 			svarp = vstate->dtvs_locals;
10552 			break;
10553 
10554 		case DIFV_SCOPE_GLOBAL:
10555 			np = &vstate->dtvs_nglobals;
10556 			svarp = vstate->dtvs_globals;
10557 			break;
10558 
10559 		default:
10560 			ASSERT(0);
10561 		}
10562 
10563 		if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
10564 			continue;
10565 
10566 		id -= DIF_VAR_OTHER_UBASE;
10567 		ASSERT(id < *np);
10568 
10569 		svar = svarp[id];
10570 		ASSERT(svar != NULL);
10571 		ASSERT(svar->dtsv_refcnt > 0);
10572 
10573 		if (--svar->dtsv_refcnt > 0)
10574 			continue;
10575 
10576 		if (svar->dtsv_size != 0) {
10577 			ASSERT(svar->dtsv_data != 0);
10578 			kmem_free((void *)(uintptr_t)svar->dtsv_data,
10579 			    svar->dtsv_size);
10580 		}
10581 
10582 		kmem_free(svar, sizeof (dtrace_statvar_t));
10583 		svarp[id] = NULL;
10584 	}
10585 
10586 	if (dp->dtdo_buf != NULL)
10587 		kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
10588 	if (dp->dtdo_inttab != NULL)
10589 		kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
10590 	if (dp->dtdo_strtab != NULL)
10591 		kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
10592 	if (dp->dtdo_vartab != NULL)
10593 		kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
10594 
10595 	kmem_free(dp, sizeof (dtrace_difo_t));
10596 }
10597 
10598 static void
10599 dtrace_difo_release(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
10600 {
10601 	int i;
10602 
10603 	ASSERT(MUTEX_HELD(&dtrace_lock));
10604 	ASSERT(dp->dtdo_refcnt != 0);
10605 
10606 	for (i = 0; i < dp->dtdo_varlen; i++) {
10607 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
10608 
10609 		if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
10610 			continue;
10611 
10612 		ASSERT(dtrace_vtime_references > 0);
10613 		if (--dtrace_vtime_references == 0)
10614 			dtrace_vtime_disable();
10615 	}
10616 
10617 	if (--dp->dtdo_refcnt == 0)
10618 		dtrace_difo_destroy(dp, vstate);
10619 }
10620 
10621 /*
10622  * DTrace Format Functions
10623  */
10624 static uint16_t
10625 dtrace_format_add(dtrace_state_t *state, char *str)
10626 {
10627 	char *fmt, **new;
10628 	uint16_t ndx, len = strlen(str) + 1;
10629 
10630 	fmt = kmem_zalloc(len, KM_SLEEP);
10631 	bcopy(str, fmt, len);
10632 
10633 	for (ndx = 0; ndx < state->dts_nformats; ndx++) {
10634 		if (state->dts_formats[ndx] == NULL) {
10635 			state->dts_formats[ndx] = fmt;
10636 			return (ndx + 1);
10637 		}
10638 	}
10639 
10640 	if (state->dts_nformats == USHRT_MAX) {
10641 		/*
10642 		 * This is only likely if a denial-of-service attack is being
10643 		 * attempted.  As such, it's okay to fail silently here.
10644 		 */
10645 		kmem_free(fmt, len);
10646 		return (0);
10647 	}
10648 
10649 	/*
10650 	 * For simplicity, we always resize the formats array to be exactly the
10651 	 * number of formats.
10652 	 */
10653 	ndx = state->dts_nformats++;
10654 	new = kmem_alloc((ndx + 1) * sizeof (char *), KM_SLEEP);
10655 
10656 	if (state->dts_formats != NULL) {
10657 		ASSERT(ndx != 0);
10658 		bcopy(state->dts_formats, new, ndx * sizeof (char *));
10659 		kmem_free(state->dts_formats, ndx * sizeof (char *));
10660 	}
10661 
10662 	state->dts_formats = new;
10663 	state->dts_formats[ndx] = fmt;
10664 
10665 	return (ndx + 1);
10666 }
10667 
10668 static void
10669 dtrace_format_remove(dtrace_state_t *state, uint16_t format)
10670 {
10671 	char *fmt;
10672 
10673 	ASSERT(state->dts_formats != NULL);
10674 	ASSERT(format <= state->dts_nformats);
10675 	ASSERT(state->dts_formats[format - 1] != NULL);
10676 
10677 	fmt = state->dts_formats[format - 1];
10678 	kmem_free(fmt, strlen(fmt) + 1);
10679 	state->dts_formats[format - 1] = NULL;
10680 }
10681 
10682 static void
10683 dtrace_format_destroy(dtrace_state_t *state)
10684 {
10685 	int i;
10686 
10687 	if (state->dts_nformats == 0) {
10688 		ASSERT(state->dts_formats == NULL);
10689 		return;
10690 	}
10691 
10692 	ASSERT(state->dts_formats != NULL);
10693 
10694 	for (i = 0; i < state->dts_nformats; i++) {
10695 		char *fmt = state->dts_formats[i];
10696 
10697 		if (fmt == NULL)
10698 			continue;
10699 
10700 		kmem_free(fmt, strlen(fmt) + 1);
10701 	}
10702 
10703 	kmem_free(state->dts_formats, state->dts_nformats * sizeof (char *));
10704 	state->dts_nformats = 0;
10705 	state->dts_formats = NULL;
10706 }
10707 
10708 /*
10709  * DTrace Predicate Functions
10710  */
10711 static dtrace_predicate_t *
10712 dtrace_predicate_create(dtrace_difo_t *dp)
10713 {
10714 	dtrace_predicate_t *pred;
10715 
10716 	ASSERT(MUTEX_HELD(&dtrace_lock));
10717 	ASSERT(dp->dtdo_refcnt != 0);
10718 
10719 	pred = kmem_zalloc(sizeof (dtrace_predicate_t), KM_SLEEP);
10720 	pred->dtp_difo = dp;
10721 	pred->dtp_refcnt = 1;
10722 
10723 	if (!dtrace_difo_cacheable(dp))
10724 		return (pred);
10725 
10726 	if (dtrace_predcache_id == DTRACE_CACHEIDNONE) {
10727 		/*
10728 		 * This is only theoretically possible -- we have had 2^32
10729 		 * cacheable predicates on this machine.  We cannot allow any
10730 		 * more predicates to become cacheable:  as unlikely as it is,
10731 		 * there may be a thread caching a (now stale) predicate cache
10732 		 * ID. (N.B.: the temptation is being successfully resisted to
10733 		 * have this cmn_err() "Holy shit -- we executed this code!")
10734 		 */
10735 		return (pred);
10736 	}
10737 
10738 	pred->dtp_cacheid = dtrace_predcache_id++;
10739 
10740 	return (pred);
10741 }
10742 
10743 static void
10744 dtrace_predicate_hold(dtrace_predicate_t *pred)
10745 {
10746 	ASSERT(MUTEX_HELD(&dtrace_lock));
10747 	ASSERT(pred->dtp_difo != NULL && pred->dtp_difo->dtdo_refcnt != 0);
10748 	ASSERT(pred->dtp_refcnt > 0);
10749 
10750 	pred->dtp_refcnt++;
10751 }
10752 
10753 static void
10754 dtrace_predicate_release(dtrace_predicate_t *pred, dtrace_vstate_t *vstate)
10755 {
10756 	dtrace_difo_t *dp = pred->dtp_difo;
10757 
10758 	ASSERT(MUTEX_HELD(&dtrace_lock));
10759 	ASSERT(dp != NULL && dp->dtdo_refcnt != 0);
10760 	ASSERT(pred->dtp_refcnt > 0);
10761 
10762 	if (--pred->dtp_refcnt == 0) {
10763 		dtrace_difo_release(pred->dtp_difo, vstate);
10764 		kmem_free(pred, sizeof (dtrace_predicate_t));
10765 	}
10766 }
10767 
10768 /*
10769  * DTrace Action Description Functions
10770  */
10771 static dtrace_actdesc_t *
10772 dtrace_actdesc_create(dtrace_actkind_t kind, uint32_t ntuple,
10773     uint64_t uarg, uint64_t arg)
10774 {
10775 	dtrace_actdesc_t *act;
10776 
10777 #ifdef illumos
10778 	ASSERT(!DTRACEACT_ISPRINTFLIKE(kind) || (arg != NULL &&
10779 	    arg >= KERNELBASE) || (arg == NULL && kind == DTRACEACT_PRINTA));
10780 #endif
10781 
10782 	act = kmem_zalloc(sizeof (dtrace_actdesc_t), KM_SLEEP);
10783 	act->dtad_kind = kind;
10784 	act->dtad_ntuple = ntuple;
10785 	act->dtad_uarg = uarg;
10786 	act->dtad_arg = arg;
10787 	act->dtad_refcnt = 1;
10788 
10789 	return (act);
10790 }
10791 
10792 static void
10793 dtrace_actdesc_hold(dtrace_actdesc_t *act)
10794 {
10795 	ASSERT(act->dtad_refcnt >= 1);
10796 	act->dtad_refcnt++;
10797 }
10798 
10799 static void
10800 dtrace_actdesc_release(dtrace_actdesc_t *act, dtrace_vstate_t *vstate)
10801 {
10802 	dtrace_actkind_t kind = act->dtad_kind;
10803 	dtrace_difo_t *dp;
10804 
10805 	ASSERT(act->dtad_refcnt >= 1);
10806 
10807 	if (--act->dtad_refcnt != 0)
10808 		return;
10809 
10810 	if ((dp = act->dtad_difo) != NULL)
10811 		dtrace_difo_release(dp, vstate);
10812 
10813 	if (DTRACEACT_ISPRINTFLIKE(kind)) {
10814 		char *str = (char *)(uintptr_t)act->dtad_arg;
10815 
10816 #ifdef illumos
10817 		ASSERT((str != NULL && (uintptr_t)str >= KERNELBASE) ||
10818 		    (str == NULL && act->dtad_kind == DTRACEACT_PRINTA));
10819 #endif
10820 
10821 		if (str != NULL)
10822 			kmem_free(str, strlen(str) + 1);
10823 	}
10824 
10825 	kmem_free(act, sizeof (dtrace_actdesc_t));
10826 }
10827 
10828 /*
10829  * DTrace ECB Functions
10830  */
10831 static dtrace_ecb_t *
10832 dtrace_ecb_add(dtrace_state_t *state, dtrace_probe_t *probe)
10833 {
10834 	dtrace_ecb_t *ecb;
10835 	dtrace_epid_t epid;
10836 
10837 	ASSERT(MUTEX_HELD(&dtrace_lock));
10838 
10839 	ecb = kmem_zalloc(sizeof (dtrace_ecb_t), KM_SLEEP);
10840 	ecb->dte_predicate = NULL;
10841 	ecb->dte_probe = probe;
10842 
10843 	/*
10844 	 * The default size is the size of the default action: recording
10845 	 * the header.
10846 	 */
10847 	ecb->dte_size = ecb->dte_needed = sizeof (dtrace_rechdr_t);
10848 	ecb->dte_alignment = sizeof (dtrace_epid_t);
10849 
10850 	epid = state->dts_epid++;
10851 
10852 	if (epid - 1 >= state->dts_necbs) {
10853 		dtrace_ecb_t **oecbs = state->dts_ecbs, **ecbs;
10854 		int necbs = state->dts_necbs << 1;
10855 
10856 		ASSERT(epid == state->dts_necbs + 1);
10857 
10858 		if (necbs == 0) {
10859 			ASSERT(oecbs == NULL);
10860 			necbs = 1;
10861 		}
10862 
10863 		ecbs = kmem_zalloc(necbs * sizeof (*ecbs), KM_SLEEP);
10864 
10865 		if (oecbs != NULL)
10866 			bcopy(oecbs, ecbs, state->dts_necbs * sizeof (*ecbs));
10867 
10868 		dtrace_membar_producer();
10869 		state->dts_ecbs = ecbs;
10870 
10871 		if (oecbs != NULL) {
10872 			/*
10873 			 * If this state is active, we must dtrace_sync()
10874 			 * before we can free the old dts_ecbs array:  we're
10875 			 * coming in hot, and there may be active ring
10876 			 * buffer processing (which indexes into the dts_ecbs
10877 			 * array) on another CPU.
10878 			 */
10879 			if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
10880 				dtrace_sync();
10881 
10882 			kmem_free(oecbs, state->dts_necbs * sizeof (*ecbs));
10883 		}
10884 
10885 		dtrace_membar_producer();
10886 		state->dts_necbs = necbs;
10887 	}
10888 
10889 	ecb->dte_state = state;
10890 
10891 	ASSERT(state->dts_ecbs[epid - 1] == NULL);
10892 	dtrace_membar_producer();
10893 	state->dts_ecbs[(ecb->dte_epid = epid) - 1] = ecb;
10894 
10895 	return (ecb);
10896 }
10897 
10898 static void
10899 dtrace_ecb_enable(dtrace_ecb_t *ecb)
10900 {
10901 	dtrace_probe_t *probe = ecb->dte_probe;
10902 
10903 	ASSERT(MUTEX_HELD(&cpu_lock));
10904 	ASSERT(MUTEX_HELD(&dtrace_lock));
10905 	ASSERT(ecb->dte_next == NULL);
10906 
10907 	if (probe == NULL) {
10908 		/*
10909 		 * This is the NULL probe -- there's nothing to do.
10910 		 */
10911 		return;
10912 	}
10913 
10914 	if (probe->dtpr_ecb == NULL) {
10915 		dtrace_provider_t *prov = probe->dtpr_provider;
10916 
10917 		/*
10918 		 * We're the first ECB on this probe.
10919 		 */
10920 		probe->dtpr_ecb = probe->dtpr_ecb_last = ecb;
10921 
10922 		if (ecb->dte_predicate != NULL)
10923 			probe->dtpr_predcache = ecb->dte_predicate->dtp_cacheid;
10924 
10925 		prov->dtpv_pops.dtps_enable(prov->dtpv_arg,
10926 		    probe->dtpr_id, probe->dtpr_arg);
10927 	} else {
10928 		/*
10929 		 * This probe is already active.  Swing the last pointer to
10930 		 * point to the new ECB, and issue a dtrace_sync() to assure
10931 		 * that all CPUs have seen the change.
10932 		 */
10933 		ASSERT(probe->dtpr_ecb_last != NULL);
10934 		probe->dtpr_ecb_last->dte_next = ecb;
10935 		probe->dtpr_ecb_last = ecb;
10936 		probe->dtpr_predcache = 0;
10937 
10938 		dtrace_sync();
10939 	}
10940 }
10941 
10942 static void
10943 dtrace_ecb_resize(dtrace_ecb_t *ecb)
10944 {
10945 	dtrace_action_t *act;
10946 	uint32_t curneeded = UINT32_MAX;
10947 	uint32_t aggbase = UINT32_MAX;
10948 
10949 	/*
10950 	 * If we record anything, we always record the dtrace_rechdr_t.  (And
10951 	 * we always record it first.)
10952 	 */
10953 	ecb->dte_size = sizeof (dtrace_rechdr_t);
10954 	ecb->dte_alignment = sizeof (dtrace_epid_t);
10955 
10956 	for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
10957 		dtrace_recdesc_t *rec = &act->dta_rec;
10958 		ASSERT(rec->dtrd_size > 0 || rec->dtrd_alignment == 1);
10959 
10960 		ecb->dte_alignment = MAX(ecb->dte_alignment,
10961 		    rec->dtrd_alignment);
10962 
10963 		if (DTRACEACT_ISAGG(act->dta_kind)) {
10964 			dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
10965 
10966 			ASSERT(rec->dtrd_size != 0);
10967 			ASSERT(agg->dtag_first != NULL);
10968 			ASSERT(act->dta_prev->dta_intuple);
10969 			ASSERT(aggbase != UINT32_MAX);
10970 			ASSERT(curneeded != UINT32_MAX);
10971 
10972 			agg->dtag_base = aggbase;
10973 
10974 			curneeded = P2ROUNDUP(curneeded, rec->dtrd_alignment);
10975 			rec->dtrd_offset = curneeded;
10976 			curneeded += rec->dtrd_size;
10977 			ecb->dte_needed = MAX(ecb->dte_needed, curneeded);
10978 
10979 			aggbase = UINT32_MAX;
10980 			curneeded = UINT32_MAX;
10981 		} else if (act->dta_intuple) {
10982 			if (curneeded == UINT32_MAX) {
10983 				/*
10984 				 * This is the first record in a tuple.  Align
10985 				 * curneeded to be at offset 4 in an 8-byte
10986 				 * aligned block.
10987 				 */
10988 				ASSERT(act->dta_prev == NULL ||
10989 				    !act->dta_prev->dta_intuple);
10990 				ASSERT3U(aggbase, ==, UINT32_MAX);
10991 				curneeded = P2PHASEUP(ecb->dte_size,
10992 				    sizeof (uint64_t), sizeof (dtrace_aggid_t));
10993 
10994 				aggbase = curneeded - sizeof (dtrace_aggid_t);
10995 				ASSERT(IS_P2ALIGNED(aggbase,
10996 				    sizeof (uint64_t)));
10997 			}
10998 			curneeded = P2ROUNDUP(curneeded, rec->dtrd_alignment);
10999 			rec->dtrd_offset = curneeded;
11000 			curneeded += rec->dtrd_size;
11001 		} else {
11002 			/* tuples must be followed by an aggregation */
11003 			ASSERT(act->dta_prev == NULL ||
11004 			    !act->dta_prev->dta_intuple);
11005 
11006 			ecb->dte_size = P2ROUNDUP(ecb->dte_size,
11007 			    rec->dtrd_alignment);
11008 			rec->dtrd_offset = ecb->dte_size;
11009 			ecb->dte_size += rec->dtrd_size;
11010 			ecb->dte_needed = MAX(ecb->dte_needed, ecb->dte_size);
11011 		}
11012 	}
11013 
11014 	if ((act = ecb->dte_action) != NULL &&
11015 	    !(act->dta_kind == DTRACEACT_SPECULATE && act->dta_next == NULL) &&
11016 	    ecb->dte_size == sizeof (dtrace_rechdr_t)) {
11017 		/*
11018 		 * If the size is still sizeof (dtrace_rechdr_t), then all
11019 		 * actions store no data; set the size to 0.
11020 		 */
11021 		ecb->dte_size = 0;
11022 	}
11023 
11024 	ecb->dte_size = P2ROUNDUP(ecb->dte_size, sizeof (dtrace_epid_t));
11025 	ecb->dte_needed = P2ROUNDUP(ecb->dte_needed, (sizeof (dtrace_epid_t)));
11026 	ecb->dte_state->dts_needed = MAX(ecb->dte_state->dts_needed,
11027 	    ecb->dte_needed);
11028 }
11029 
11030 static dtrace_action_t *
11031 dtrace_ecb_aggregation_create(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
11032 {
11033 	dtrace_aggregation_t *agg;
11034 	size_t size = sizeof (uint64_t);
11035 	int ntuple = desc->dtad_ntuple;
11036 	dtrace_action_t *act;
11037 	dtrace_recdesc_t *frec;
11038 	dtrace_aggid_t aggid;
11039 	dtrace_state_t *state = ecb->dte_state;
11040 
11041 	agg = kmem_zalloc(sizeof (dtrace_aggregation_t), KM_SLEEP);
11042 	agg->dtag_ecb = ecb;
11043 
11044 	ASSERT(DTRACEACT_ISAGG(desc->dtad_kind));
11045 
11046 	switch (desc->dtad_kind) {
11047 	case DTRACEAGG_MIN:
11048 		agg->dtag_initial = INT64_MAX;
11049 		agg->dtag_aggregate = dtrace_aggregate_min;
11050 		break;
11051 
11052 	case DTRACEAGG_MAX:
11053 		agg->dtag_initial = INT64_MIN;
11054 		agg->dtag_aggregate = dtrace_aggregate_max;
11055 		break;
11056 
11057 	case DTRACEAGG_COUNT:
11058 		agg->dtag_aggregate = dtrace_aggregate_count;
11059 		break;
11060 
11061 	case DTRACEAGG_QUANTIZE:
11062 		agg->dtag_aggregate = dtrace_aggregate_quantize;
11063 		size = (((sizeof (uint64_t) * NBBY) - 1) * 2 + 1) *
11064 		    sizeof (uint64_t);
11065 		break;
11066 
11067 	case DTRACEAGG_LQUANTIZE: {
11068 		uint16_t step = DTRACE_LQUANTIZE_STEP(desc->dtad_arg);
11069 		uint16_t levels = DTRACE_LQUANTIZE_LEVELS(desc->dtad_arg);
11070 
11071 		agg->dtag_initial = desc->dtad_arg;
11072 		agg->dtag_aggregate = dtrace_aggregate_lquantize;
11073 
11074 		if (step == 0 || levels == 0)
11075 			goto err;
11076 
11077 		size = levels * sizeof (uint64_t) + 3 * sizeof (uint64_t);
11078 		break;
11079 	}
11080 
11081 	case DTRACEAGG_LLQUANTIZE: {
11082 		uint16_t factor = DTRACE_LLQUANTIZE_FACTOR(desc->dtad_arg);
11083 		uint16_t low = DTRACE_LLQUANTIZE_LOW(desc->dtad_arg);
11084 		uint16_t high = DTRACE_LLQUANTIZE_HIGH(desc->dtad_arg);
11085 		uint16_t nsteps = DTRACE_LLQUANTIZE_NSTEP(desc->dtad_arg);
11086 		int64_t v;
11087 
11088 		agg->dtag_initial = desc->dtad_arg;
11089 		agg->dtag_aggregate = dtrace_aggregate_llquantize;
11090 
11091 		if (factor < 2 || low >= high || nsteps < factor)
11092 			goto err;
11093 
11094 		/*
11095 		 * Now check that the number of steps evenly divides a power
11096 		 * of the factor.  (This assures both integer bucket size and
11097 		 * linearity within each magnitude.)
11098 		 */
11099 		for (v = factor; v < nsteps; v *= factor)
11100 			continue;
11101 
11102 		if ((v % nsteps) || (nsteps % factor))
11103 			goto err;
11104 
11105 		size = (dtrace_aggregate_llquantize_bucket(factor,
11106 		    low, high, nsteps, INT64_MAX) + 2) * sizeof (uint64_t);
11107 		break;
11108 	}
11109 
11110 	case DTRACEAGG_AVG:
11111 		agg->dtag_aggregate = dtrace_aggregate_avg;
11112 		size = sizeof (uint64_t) * 2;
11113 		break;
11114 
11115 	case DTRACEAGG_STDDEV:
11116 		agg->dtag_aggregate = dtrace_aggregate_stddev;
11117 		size = sizeof (uint64_t) * 4;
11118 		break;
11119 
11120 	case DTRACEAGG_SUM:
11121 		agg->dtag_aggregate = dtrace_aggregate_sum;
11122 		break;
11123 
11124 	default:
11125 		goto err;
11126 	}
11127 
11128 	agg->dtag_action.dta_rec.dtrd_size = size;
11129 
11130 	if (ntuple == 0)
11131 		goto err;
11132 
11133 	/*
11134 	 * We must make sure that we have enough actions for the n-tuple.
11135 	 */
11136 	for (act = ecb->dte_action_last; act != NULL; act = act->dta_prev) {
11137 		if (DTRACEACT_ISAGG(act->dta_kind))
11138 			break;
11139 
11140 		if (--ntuple == 0) {
11141 			/*
11142 			 * This is the action with which our n-tuple begins.
11143 			 */
11144 			agg->dtag_first = act;
11145 			goto success;
11146 		}
11147 	}
11148 
11149 	/*
11150 	 * This n-tuple is short by ntuple elements.  Return failure.
11151 	 */
11152 	ASSERT(ntuple != 0);
11153 err:
11154 	kmem_free(agg, sizeof (dtrace_aggregation_t));
11155 	return (NULL);
11156 
11157 success:
11158 	/*
11159 	 * If the last action in the tuple has a size of zero, it's actually
11160 	 * an expression argument for the aggregating action.
11161 	 */
11162 	ASSERT(ecb->dte_action_last != NULL);
11163 	act = ecb->dte_action_last;
11164 
11165 	if (act->dta_kind == DTRACEACT_DIFEXPR) {
11166 		ASSERT(act->dta_difo != NULL);
11167 
11168 		if (act->dta_difo->dtdo_rtype.dtdt_size == 0)
11169 			agg->dtag_hasarg = 1;
11170 	}
11171 
11172 	/*
11173 	 * We need to allocate an id for this aggregation.
11174 	 */
11175 #ifdef illumos
11176 	aggid = (dtrace_aggid_t)(uintptr_t)vmem_alloc(state->dts_aggid_arena, 1,
11177 	    VM_BESTFIT | VM_SLEEP);
11178 #else
11179 	aggid = alloc_unr(state->dts_aggid_arena);
11180 #endif
11181 
11182 	if (aggid - 1 >= state->dts_naggregations) {
11183 		dtrace_aggregation_t **oaggs = state->dts_aggregations;
11184 		dtrace_aggregation_t **aggs;
11185 		int naggs = state->dts_naggregations << 1;
11186 		int onaggs = state->dts_naggregations;
11187 
11188 		ASSERT(aggid == state->dts_naggregations + 1);
11189 
11190 		if (naggs == 0) {
11191 			ASSERT(oaggs == NULL);
11192 			naggs = 1;
11193 		}
11194 
11195 		aggs = kmem_zalloc(naggs * sizeof (*aggs), KM_SLEEP);
11196 
11197 		if (oaggs != NULL) {
11198 			bcopy(oaggs, aggs, onaggs * sizeof (*aggs));
11199 			kmem_free(oaggs, onaggs * sizeof (*aggs));
11200 		}
11201 
11202 		state->dts_aggregations = aggs;
11203 		state->dts_naggregations = naggs;
11204 	}
11205 
11206 	ASSERT(state->dts_aggregations[aggid - 1] == NULL);
11207 	state->dts_aggregations[(agg->dtag_id = aggid) - 1] = agg;
11208 
11209 	frec = &agg->dtag_first->dta_rec;
11210 	if (frec->dtrd_alignment < sizeof (dtrace_aggid_t))
11211 		frec->dtrd_alignment = sizeof (dtrace_aggid_t);
11212 
11213 	for (act = agg->dtag_first; act != NULL; act = act->dta_next) {
11214 		ASSERT(!act->dta_intuple);
11215 		act->dta_intuple = 1;
11216 	}
11217 
11218 	return (&agg->dtag_action);
11219 }
11220 
11221 static void
11222 dtrace_ecb_aggregation_destroy(dtrace_ecb_t *ecb, dtrace_action_t *act)
11223 {
11224 	dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
11225 	dtrace_state_t *state = ecb->dte_state;
11226 	dtrace_aggid_t aggid = agg->dtag_id;
11227 
11228 	ASSERT(DTRACEACT_ISAGG(act->dta_kind));
11229 #ifdef illumos
11230 	vmem_free(state->dts_aggid_arena, (void *)(uintptr_t)aggid, 1);
11231 #else
11232 	free_unr(state->dts_aggid_arena, aggid);
11233 #endif
11234 
11235 	ASSERT(state->dts_aggregations[aggid - 1] == agg);
11236 	state->dts_aggregations[aggid - 1] = NULL;
11237 
11238 	kmem_free(agg, sizeof (dtrace_aggregation_t));
11239 }
11240 
11241 static int
11242 dtrace_ecb_action_add(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
11243 {
11244 	dtrace_action_t *action, *last;
11245 	dtrace_difo_t *dp = desc->dtad_difo;
11246 	uint32_t size = 0, align = sizeof (uint8_t), mask;
11247 	uint16_t format = 0;
11248 	dtrace_recdesc_t *rec;
11249 	dtrace_state_t *state = ecb->dte_state;
11250 	dtrace_optval_t *opt = state->dts_options, nframes = 0, strsize;
11251 	uint64_t arg = desc->dtad_arg;
11252 
11253 	ASSERT(MUTEX_HELD(&dtrace_lock));
11254 	ASSERT(ecb->dte_action == NULL || ecb->dte_action->dta_refcnt == 1);
11255 
11256 	if (DTRACEACT_ISAGG(desc->dtad_kind)) {
11257 		/*
11258 		 * If this is an aggregating action, there must be neither
11259 		 * a speculate nor a commit on the action chain.
11260 		 */
11261 		dtrace_action_t *act;
11262 
11263 		for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
11264 			if (act->dta_kind == DTRACEACT_COMMIT)
11265 				return (EINVAL);
11266 
11267 			if (act->dta_kind == DTRACEACT_SPECULATE)
11268 				return (EINVAL);
11269 		}
11270 
11271 		action = dtrace_ecb_aggregation_create(ecb, desc);
11272 
11273 		if (action == NULL)
11274 			return (EINVAL);
11275 	} else {
11276 		if (DTRACEACT_ISDESTRUCTIVE(desc->dtad_kind) ||
11277 		    (desc->dtad_kind == DTRACEACT_DIFEXPR &&
11278 		    dp != NULL && dp->dtdo_destructive)) {
11279 			state->dts_destructive = 1;
11280 		}
11281 
11282 		switch (desc->dtad_kind) {
11283 		case DTRACEACT_PRINTF:
11284 		case DTRACEACT_PRINTA:
11285 		case DTRACEACT_SYSTEM:
11286 		case DTRACEACT_FREOPEN:
11287 		case DTRACEACT_DIFEXPR:
11288 			/*
11289 			 * We know that our arg is a string -- turn it into a
11290 			 * format.
11291 			 */
11292 			if (arg == 0) {
11293 				ASSERT(desc->dtad_kind == DTRACEACT_PRINTA ||
11294 				    desc->dtad_kind == DTRACEACT_DIFEXPR);
11295 				format = 0;
11296 			} else {
11297 				ASSERT(arg != 0);
11298 #ifdef illumos
11299 				ASSERT(arg > KERNELBASE);
11300 #endif
11301 				format = dtrace_format_add(state,
11302 				    (char *)(uintptr_t)arg);
11303 			}
11304 
11305 			/*FALLTHROUGH*/
11306 		case DTRACEACT_LIBACT:
11307 		case DTRACEACT_TRACEMEM:
11308 		case DTRACEACT_TRACEMEM_DYNSIZE:
11309 			if (dp == NULL)
11310 				return (EINVAL);
11311 
11312 			if ((size = dp->dtdo_rtype.dtdt_size) != 0)
11313 				break;
11314 
11315 			if (dp->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING) {
11316 				if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
11317 					return (EINVAL);
11318 
11319 				size = opt[DTRACEOPT_STRSIZE];
11320 			}
11321 
11322 			break;
11323 
11324 		case DTRACEACT_STACK:
11325 			if ((nframes = arg) == 0) {
11326 				nframes = opt[DTRACEOPT_STACKFRAMES];
11327 				ASSERT(nframes > 0);
11328 				arg = nframes;
11329 			}
11330 
11331 			size = nframes * sizeof (pc_t);
11332 			break;
11333 
11334 		case DTRACEACT_JSTACK:
11335 			if ((strsize = DTRACE_USTACK_STRSIZE(arg)) == 0)
11336 				strsize = opt[DTRACEOPT_JSTACKSTRSIZE];
11337 
11338 			if ((nframes = DTRACE_USTACK_NFRAMES(arg)) == 0)
11339 				nframes = opt[DTRACEOPT_JSTACKFRAMES];
11340 
11341 			arg = DTRACE_USTACK_ARG(nframes, strsize);
11342 
11343 			/*FALLTHROUGH*/
11344 		case DTRACEACT_USTACK:
11345 			if (desc->dtad_kind != DTRACEACT_JSTACK &&
11346 			    (nframes = DTRACE_USTACK_NFRAMES(arg)) == 0) {
11347 				strsize = DTRACE_USTACK_STRSIZE(arg);
11348 				nframes = opt[DTRACEOPT_USTACKFRAMES];
11349 				ASSERT(nframes > 0);
11350 				arg = DTRACE_USTACK_ARG(nframes, strsize);
11351 			}
11352 
11353 			/*
11354 			 * Save a slot for the pid.
11355 			 */
11356 			size = (nframes + 1) * sizeof (uint64_t);
11357 			size += DTRACE_USTACK_STRSIZE(arg);
11358 			size = P2ROUNDUP(size, (uint32_t)(sizeof (uintptr_t)));
11359 
11360 			break;
11361 
11362 		case DTRACEACT_SYM:
11363 		case DTRACEACT_MOD:
11364 			if (dp == NULL || ((size = dp->dtdo_rtype.dtdt_size) !=
11365 			    sizeof (uint64_t)) ||
11366 			    (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
11367 				return (EINVAL);
11368 			break;
11369 
11370 		case DTRACEACT_USYM:
11371 		case DTRACEACT_UMOD:
11372 		case DTRACEACT_UADDR:
11373 			if (dp == NULL ||
11374 			    (dp->dtdo_rtype.dtdt_size != sizeof (uint64_t)) ||
11375 			    (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
11376 				return (EINVAL);
11377 
11378 			/*
11379 			 * We have a slot for the pid, plus a slot for the
11380 			 * argument.  To keep things simple (aligned with
11381 			 * bitness-neutral sizing), we store each as a 64-bit
11382 			 * quantity.
11383 			 */
11384 			size = 2 * sizeof (uint64_t);
11385 			break;
11386 
11387 		case DTRACEACT_STOP:
11388 		case DTRACEACT_BREAKPOINT:
11389 		case DTRACEACT_PANIC:
11390 			break;
11391 
11392 		case DTRACEACT_CHILL:
11393 		case DTRACEACT_DISCARD:
11394 		case DTRACEACT_RAISE:
11395 			if (dp == NULL)
11396 				return (EINVAL);
11397 			break;
11398 
11399 		case DTRACEACT_EXIT:
11400 			if (dp == NULL ||
11401 			    (size = dp->dtdo_rtype.dtdt_size) != sizeof (int) ||
11402 			    (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
11403 				return (EINVAL);
11404 			break;
11405 
11406 		case DTRACEACT_SPECULATE:
11407 			if (ecb->dte_size > sizeof (dtrace_rechdr_t))
11408 				return (EINVAL);
11409 
11410 			if (dp == NULL)
11411 				return (EINVAL);
11412 
11413 			state->dts_speculates = 1;
11414 			break;
11415 
11416 		case DTRACEACT_PRINTM:
11417 		    	size = dp->dtdo_rtype.dtdt_size;
11418 			break;
11419 
11420 		case DTRACEACT_PRINTT:
11421 		    	size = dp->dtdo_rtype.dtdt_size;
11422 			break;
11423 
11424 		case DTRACEACT_COMMIT: {
11425 			dtrace_action_t *act = ecb->dte_action;
11426 
11427 			for (; act != NULL; act = act->dta_next) {
11428 				if (act->dta_kind == DTRACEACT_COMMIT)
11429 					return (EINVAL);
11430 			}
11431 
11432 			if (dp == NULL)
11433 				return (EINVAL);
11434 			break;
11435 		}
11436 
11437 		default:
11438 			return (EINVAL);
11439 		}
11440 
11441 		if (size != 0 || desc->dtad_kind == DTRACEACT_SPECULATE) {
11442 			/*
11443 			 * If this is a data-storing action or a speculate,
11444 			 * we must be sure that there isn't a commit on the
11445 			 * action chain.
11446 			 */
11447 			dtrace_action_t *act = ecb->dte_action;
11448 
11449 			for (; act != NULL; act = act->dta_next) {
11450 				if (act->dta_kind == DTRACEACT_COMMIT)
11451 					return (EINVAL);
11452 			}
11453 		}
11454 
11455 		action = kmem_zalloc(sizeof (dtrace_action_t), KM_SLEEP);
11456 		action->dta_rec.dtrd_size = size;
11457 	}
11458 
11459 	action->dta_refcnt = 1;
11460 	rec = &action->dta_rec;
11461 	size = rec->dtrd_size;
11462 
11463 	for (mask = sizeof (uint64_t) - 1; size != 0 && mask > 0; mask >>= 1) {
11464 		if (!(size & mask)) {
11465 			align = mask + 1;
11466 			break;
11467 		}
11468 	}
11469 
11470 	action->dta_kind = desc->dtad_kind;
11471 
11472 	if ((action->dta_difo = dp) != NULL)
11473 		dtrace_difo_hold(dp);
11474 
11475 	rec->dtrd_action = action->dta_kind;
11476 	rec->dtrd_arg = arg;
11477 	rec->dtrd_uarg = desc->dtad_uarg;
11478 	rec->dtrd_alignment = (uint16_t)align;
11479 	rec->dtrd_format = format;
11480 
11481 	if ((last = ecb->dte_action_last) != NULL) {
11482 		ASSERT(ecb->dte_action != NULL);
11483 		action->dta_prev = last;
11484 		last->dta_next = action;
11485 	} else {
11486 		ASSERT(ecb->dte_action == NULL);
11487 		ecb->dte_action = action;
11488 	}
11489 
11490 	ecb->dte_action_last = action;
11491 
11492 	return (0);
11493 }
11494 
11495 static void
11496 dtrace_ecb_action_remove(dtrace_ecb_t *ecb)
11497 {
11498 	dtrace_action_t *act = ecb->dte_action, *next;
11499 	dtrace_vstate_t *vstate = &ecb->dte_state->dts_vstate;
11500 	dtrace_difo_t *dp;
11501 	uint16_t format;
11502 
11503 	if (act != NULL && act->dta_refcnt > 1) {
11504 		ASSERT(act->dta_next == NULL || act->dta_next->dta_refcnt == 1);
11505 		act->dta_refcnt--;
11506 	} else {
11507 		for (; act != NULL; act = next) {
11508 			next = act->dta_next;
11509 			ASSERT(next != NULL || act == ecb->dte_action_last);
11510 			ASSERT(act->dta_refcnt == 1);
11511 
11512 			if ((format = act->dta_rec.dtrd_format) != 0)
11513 				dtrace_format_remove(ecb->dte_state, format);
11514 
11515 			if ((dp = act->dta_difo) != NULL)
11516 				dtrace_difo_release(dp, vstate);
11517 
11518 			if (DTRACEACT_ISAGG(act->dta_kind)) {
11519 				dtrace_ecb_aggregation_destroy(ecb, act);
11520 			} else {
11521 				kmem_free(act, sizeof (dtrace_action_t));
11522 			}
11523 		}
11524 	}
11525 
11526 	ecb->dte_action = NULL;
11527 	ecb->dte_action_last = NULL;
11528 	ecb->dte_size = 0;
11529 }
11530 
11531 static void
11532 dtrace_ecb_disable(dtrace_ecb_t *ecb)
11533 {
11534 	/*
11535 	 * We disable the ECB by removing it from its probe.
11536 	 */
11537 	dtrace_ecb_t *pecb, *prev = NULL;
11538 	dtrace_probe_t *probe = ecb->dte_probe;
11539 
11540 	ASSERT(MUTEX_HELD(&dtrace_lock));
11541 
11542 	if (probe == NULL) {
11543 		/*
11544 		 * This is the NULL probe; there is nothing to disable.
11545 		 */
11546 		return;
11547 	}
11548 
11549 	for (pecb = probe->dtpr_ecb; pecb != NULL; pecb = pecb->dte_next) {
11550 		if (pecb == ecb)
11551 			break;
11552 		prev = pecb;
11553 	}
11554 
11555 	ASSERT(pecb != NULL);
11556 
11557 	if (prev == NULL) {
11558 		probe->dtpr_ecb = ecb->dte_next;
11559 	} else {
11560 		prev->dte_next = ecb->dte_next;
11561 	}
11562 
11563 	if (ecb == probe->dtpr_ecb_last) {
11564 		ASSERT(ecb->dte_next == NULL);
11565 		probe->dtpr_ecb_last = prev;
11566 	}
11567 
11568 	/*
11569 	 * The ECB has been disconnected from the probe; now sync to assure
11570 	 * that all CPUs have seen the change before returning.
11571 	 */
11572 	dtrace_sync();
11573 
11574 	if (probe->dtpr_ecb == NULL) {
11575 		/*
11576 		 * That was the last ECB on the probe; clear the predicate
11577 		 * cache ID for the probe, disable it and sync one more time
11578 		 * to assure that we'll never hit it again.
11579 		 */
11580 		dtrace_provider_t *prov = probe->dtpr_provider;
11581 
11582 		ASSERT(ecb->dte_next == NULL);
11583 		ASSERT(probe->dtpr_ecb_last == NULL);
11584 		probe->dtpr_predcache = DTRACE_CACHEIDNONE;
11585 		prov->dtpv_pops.dtps_disable(prov->dtpv_arg,
11586 		    probe->dtpr_id, probe->dtpr_arg);
11587 		dtrace_sync();
11588 	} else {
11589 		/*
11590 		 * There is at least one ECB remaining on the probe.  If there
11591 		 * is _exactly_ one, set the probe's predicate cache ID to be
11592 		 * the predicate cache ID of the remaining ECB.
11593 		 */
11594 		ASSERT(probe->dtpr_ecb_last != NULL);
11595 		ASSERT(probe->dtpr_predcache == DTRACE_CACHEIDNONE);
11596 
11597 		if (probe->dtpr_ecb == probe->dtpr_ecb_last) {
11598 			dtrace_predicate_t *p = probe->dtpr_ecb->dte_predicate;
11599 
11600 			ASSERT(probe->dtpr_ecb->dte_next == NULL);
11601 
11602 			if (p != NULL)
11603 				probe->dtpr_predcache = p->dtp_cacheid;
11604 		}
11605 
11606 		ecb->dte_next = NULL;
11607 	}
11608 }
11609 
11610 static void
11611 dtrace_ecb_destroy(dtrace_ecb_t *ecb)
11612 {
11613 	dtrace_state_t *state = ecb->dte_state;
11614 	dtrace_vstate_t *vstate = &state->dts_vstate;
11615 	dtrace_predicate_t *pred;
11616 	dtrace_epid_t epid = ecb->dte_epid;
11617 
11618 	ASSERT(MUTEX_HELD(&dtrace_lock));
11619 	ASSERT(ecb->dte_next == NULL);
11620 	ASSERT(ecb->dte_probe == NULL || ecb->dte_probe->dtpr_ecb != ecb);
11621 
11622 	if ((pred = ecb->dte_predicate) != NULL)
11623 		dtrace_predicate_release(pred, vstate);
11624 
11625 	dtrace_ecb_action_remove(ecb);
11626 
11627 	ASSERT(state->dts_ecbs[epid - 1] == ecb);
11628 	state->dts_ecbs[epid - 1] = NULL;
11629 
11630 	kmem_free(ecb, sizeof (dtrace_ecb_t));
11631 }
11632 
11633 static dtrace_ecb_t *
11634 dtrace_ecb_create(dtrace_state_t *state, dtrace_probe_t *probe,
11635     dtrace_enabling_t *enab)
11636 {
11637 	dtrace_ecb_t *ecb;
11638 	dtrace_predicate_t *pred;
11639 	dtrace_actdesc_t *act;
11640 	dtrace_provider_t *prov;
11641 	dtrace_ecbdesc_t *desc = enab->dten_current;
11642 
11643 	ASSERT(MUTEX_HELD(&dtrace_lock));
11644 	ASSERT(state != NULL);
11645 
11646 	ecb = dtrace_ecb_add(state, probe);
11647 	ecb->dte_uarg = desc->dted_uarg;
11648 
11649 	if ((pred = desc->dted_pred.dtpdd_predicate) != NULL) {
11650 		dtrace_predicate_hold(pred);
11651 		ecb->dte_predicate = pred;
11652 	}
11653 
11654 	if (probe != NULL) {
11655 		/*
11656 		 * If the provider shows more leg than the consumer is old
11657 		 * enough to see, we need to enable the appropriate implicit
11658 		 * predicate bits to prevent the ecb from activating at
11659 		 * revealing times.
11660 		 *
11661 		 * Providers specifying DTRACE_PRIV_USER at register time
11662 		 * are stating that they need the /proc-style privilege
11663 		 * model to be enforced, and this is what DTRACE_COND_OWNER
11664 		 * and DTRACE_COND_ZONEOWNER will then do at probe time.
11665 		 */
11666 		prov = probe->dtpr_provider;
11667 		if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLPROC) &&
11668 		    (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
11669 			ecb->dte_cond |= DTRACE_COND_OWNER;
11670 
11671 		if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLZONE) &&
11672 		    (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
11673 			ecb->dte_cond |= DTRACE_COND_ZONEOWNER;
11674 
11675 		/*
11676 		 * If the provider shows us kernel innards and the user
11677 		 * is lacking sufficient privilege, enable the
11678 		 * DTRACE_COND_USERMODE implicit predicate.
11679 		 */
11680 		if (!(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) &&
11681 		    (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_KERNEL))
11682 			ecb->dte_cond |= DTRACE_COND_USERMODE;
11683 	}
11684 
11685 	if (dtrace_ecb_create_cache != NULL) {
11686 		/*
11687 		 * If we have a cached ecb, we'll use its action list instead
11688 		 * of creating our own (saving both time and space).
11689 		 */
11690 		dtrace_ecb_t *cached = dtrace_ecb_create_cache;
11691 		dtrace_action_t *act = cached->dte_action;
11692 
11693 		if (act != NULL) {
11694 			ASSERT(act->dta_refcnt > 0);
11695 			act->dta_refcnt++;
11696 			ecb->dte_action = act;
11697 			ecb->dte_action_last = cached->dte_action_last;
11698 			ecb->dte_needed = cached->dte_needed;
11699 			ecb->dte_size = cached->dte_size;
11700 			ecb->dte_alignment = cached->dte_alignment;
11701 		}
11702 
11703 		return (ecb);
11704 	}
11705 
11706 	for (act = desc->dted_action; act != NULL; act = act->dtad_next) {
11707 		if ((enab->dten_error = dtrace_ecb_action_add(ecb, act)) != 0) {
11708 			dtrace_ecb_destroy(ecb);
11709 			return (NULL);
11710 		}
11711 	}
11712 
11713 	dtrace_ecb_resize(ecb);
11714 
11715 	return (dtrace_ecb_create_cache = ecb);
11716 }
11717 
11718 static int
11719 dtrace_ecb_create_enable(dtrace_probe_t *probe, void *arg)
11720 {
11721 	dtrace_ecb_t *ecb;
11722 	dtrace_enabling_t *enab = arg;
11723 	dtrace_state_t *state = enab->dten_vstate->dtvs_state;
11724 
11725 	ASSERT(state != NULL);
11726 
11727 	if (probe != NULL && probe->dtpr_gen < enab->dten_probegen) {
11728 		/*
11729 		 * This probe was created in a generation for which this
11730 		 * enabling has previously created ECBs; we don't want to
11731 		 * enable it again, so just kick out.
11732 		 */
11733 		return (DTRACE_MATCH_NEXT);
11734 	}
11735 
11736 	if ((ecb = dtrace_ecb_create(state, probe, enab)) == NULL)
11737 		return (DTRACE_MATCH_DONE);
11738 
11739 	dtrace_ecb_enable(ecb);
11740 	return (DTRACE_MATCH_NEXT);
11741 }
11742 
11743 static dtrace_ecb_t *
11744 dtrace_epid2ecb(dtrace_state_t *state, dtrace_epid_t id)
11745 {
11746 	dtrace_ecb_t *ecb;
11747 
11748 	ASSERT(MUTEX_HELD(&dtrace_lock));
11749 
11750 	if (id == 0 || id > state->dts_necbs)
11751 		return (NULL);
11752 
11753 	ASSERT(state->dts_necbs > 0 && state->dts_ecbs != NULL);
11754 	ASSERT((ecb = state->dts_ecbs[id - 1]) == NULL || ecb->dte_epid == id);
11755 
11756 	return (state->dts_ecbs[id - 1]);
11757 }
11758 
11759 static dtrace_aggregation_t *
11760 dtrace_aggid2agg(dtrace_state_t *state, dtrace_aggid_t id)
11761 {
11762 	dtrace_aggregation_t *agg;
11763 
11764 	ASSERT(MUTEX_HELD(&dtrace_lock));
11765 
11766 	if (id == 0 || id > state->dts_naggregations)
11767 		return (NULL);
11768 
11769 	ASSERT(state->dts_naggregations > 0 && state->dts_aggregations != NULL);
11770 	ASSERT((agg = state->dts_aggregations[id - 1]) == NULL ||
11771 	    agg->dtag_id == id);
11772 
11773 	return (state->dts_aggregations[id - 1]);
11774 }
11775 
11776 /*
11777  * DTrace Buffer Functions
11778  *
11779  * The following functions manipulate DTrace buffers.  Most of these functions
11780  * are called in the context of establishing or processing consumer state;
11781  * exceptions are explicitly noted.
11782  */
11783 
11784 /*
11785  * Note:  called from cross call context.  This function switches the two
11786  * buffers on a given CPU.  The atomicity of this operation is assured by
11787  * disabling interrupts while the actual switch takes place; the disabling of
11788  * interrupts serializes the execution with any execution of dtrace_probe() on
11789  * the same CPU.
11790  */
11791 static void
11792 dtrace_buffer_switch(dtrace_buffer_t *buf)
11793 {
11794 	caddr_t tomax = buf->dtb_tomax;
11795 	caddr_t xamot = buf->dtb_xamot;
11796 	dtrace_icookie_t cookie;
11797 	hrtime_t now;
11798 
11799 	ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
11800 	ASSERT(!(buf->dtb_flags & DTRACEBUF_RING));
11801 
11802 	cookie = dtrace_interrupt_disable();
11803 	now = dtrace_gethrtime();
11804 	buf->dtb_tomax = xamot;
11805 	buf->dtb_xamot = tomax;
11806 	buf->dtb_xamot_drops = buf->dtb_drops;
11807 	buf->dtb_xamot_offset = buf->dtb_offset;
11808 	buf->dtb_xamot_errors = buf->dtb_errors;
11809 	buf->dtb_xamot_flags = buf->dtb_flags;
11810 	buf->dtb_offset = 0;
11811 	buf->dtb_drops = 0;
11812 	buf->dtb_errors = 0;
11813 	buf->dtb_flags &= ~(DTRACEBUF_ERROR | DTRACEBUF_DROPPED);
11814 	buf->dtb_interval = now - buf->dtb_switched;
11815 	buf->dtb_switched = now;
11816 	dtrace_interrupt_enable(cookie);
11817 }
11818 
11819 /*
11820  * Note:  called from cross call context.  This function activates a buffer
11821  * on a CPU.  As with dtrace_buffer_switch(), the atomicity of the operation
11822  * is guaranteed by the disabling of interrupts.
11823  */
11824 static void
11825 dtrace_buffer_activate(dtrace_state_t *state)
11826 {
11827 	dtrace_buffer_t *buf;
11828 	dtrace_icookie_t cookie = dtrace_interrupt_disable();
11829 
11830 	buf = &state->dts_buffer[curcpu];
11831 
11832 	if (buf->dtb_tomax != NULL) {
11833 		/*
11834 		 * We might like to assert that the buffer is marked inactive,
11835 		 * but this isn't necessarily true:  the buffer for the CPU
11836 		 * that processes the BEGIN probe has its buffer activated
11837 		 * manually.  In this case, we take the (harmless) action
11838 		 * re-clearing the bit INACTIVE bit.
11839 		 */
11840 		buf->dtb_flags &= ~DTRACEBUF_INACTIVE;
11841 	}
11842 
11843 	dtrace_interrupt_enable(cookie);
11844 }
11845 
11846 static int
11847 dtrace_buffer_alloc(dtrace_buffer_t *bufs, size_t size, int flags,
11848     processorid_t cpu, int *factor)
11849 {
11850 #ifdef illumos
11851 	cpu_t *cp;
11852 #endif
11853 	dtrace_buffer_t *buf;
11854 	int allocated = 0, desired = 0;
11855 
11856 #ifdef illumos
11857 	ASSERT(MUTEX_HELD(&cpu_lock));
11858 	ASSERT(MUTEX_HELD(&dtrace_lock));
11859 
11860 	*factor = 1;
11861 
11862 	if (size > dtrace_nonroot_maxsize &&
11863 	    !PRIV_POLICY_CHOICE(CRED(), PRIV_ALL, B_FALSE))
11864 		return (EFBIG);
11865 
11866 	cp = cpu_list;
11867 
11868 	do {
11869 		if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
11870 			continue;
11871 
11872 		buf = &bufs[cp->cpu_id];
11873 
11874 		/*
11875 		 * If there is already a buffer allocated for this CPU, it
11876 		 * is only possible that this is a DR event.  In this case,
11877 		 */
11878 		if (buf->dtb_tomax != NULL) {
11879 			ASSERT(buf->dtb_size == size);
11880 			continue;
11881 		}
11882 
11883 		ASSERT(buf->dtb_xamot == NULL);
11884 
11885 		if ((buf->dtb_tomax = kmem_zalloc(size,
11886 		    KM_NOSLEEP | KM_NORMALPRI)) == NULL)
11887 			goto err;
11888 
11889 		buf->dtb_size = size;
11890 		buf->dtb_flags = flags;
11891 		buf->dtb_offset = 0;
11892 		buf->dtb_drops = 0;
11893 
11894 		if (flags & DTRACEBUF_NOSWITCH)
11895 			continue;
11896 
11897 		if ((buf->dtb_xamot = kmem_zalloc(size,
11898 		    KM_NOSLEEP | KM_NORMALPRI)) == NULL)
11899 			goto err;
11900 	} while ((cp = cp->cpu_next) != cpu_list);
11901 
11902 	return (0);
11903 
11904 err:
11905 	cp = cpu_list;
11906 
11907 	do {
11908 		if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
11909 			continue;
11910 
11911 		buf = &bufs[cp->cpu_id];
11912 		desired += 2;
11913 
11914 		if (buf->dtb_xamot != NULL) {
11915 			ASSERT(buf->dtb_tomax != NULL);
11916 			ASSERT(buf->dtb_size == size);
11917 			kmem_free(buf->dtb_xamot, size);
11918 			allocated++;
11919 		}
11920 
11921 		if (buf->dtb_tomax != NULL) {
11922 			ASSERT(buf->dtb_size == size);
11923 			kmem_free(buf->dtb_tomax, size);
11924 			allocated++;
11925 		}
11926 
11927 		buf->dtb_tomax = NULL;
11928 		buf->dtb_xamot = NULL;
11929 		buf->dtb_size = 0;
11930 	} while ((cp = cp->cpu_next) != cpu_list);
11931 #else
11932 	int i;
11933 
11934 	*factor = 1;
11935 #if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \
11936     defined(__mips__) || defined(__powerpc__)
11937 	/*
11938 	 * FreeBSD isn't good at limiting the amount of memory we
11939 	 * ask to malloc, so let's place a limit here before trying
11940 	 * to do something that might well end in tears at bedtime.
11941 	 */
11942 	if (size > physmem * PAGE_SIZE / (128 * (mp_maxid + 1)))
11943 		return (ENOMEM);
11944 #endif
11945 
11946 	ASSERT(MUTEX_HELD(&dtrace_lock));
11947 	CPU_FOREACH(i) {
11948 		if (cpu != DTRACE_CPUALL && cpu != i)
11949 			continue;
11950 
11951 		buf = &bufs[i];
11952 
11953 		/*
11954 		 * If there is already a buffer allocated for this CPU, it
11955 		 * is only possible that this is a DR event.  In this case,
11956 		 * the buffer size must match our specified size.
11957 		 */
11958 		if (buf->dtb_tomax != NULL) {
11959 			ASSERT(buf->dtb_size == size);
11960 			continue;
11961 		}
11962 
11963 		ASSERT(buf->dtb_xamot == NULL);
11964 
11965 		if ((buf->dtb_tomax = kmem_zalloc(size,
11966 		    KM_NOSLEEP | KM_NORMALPRI)) == NULL)
11967 			goto err;
11968 
11969 		buf->dtb_size = size;
11970 		buf->dtb_flags = flags;
11971 		buf->dtb_offset = 0;
11972 		buf->dtb_drops = 0;
11973 
11974 		if (flags & DTRACEBUF_NOSWITCH)
11975 			continue;
11976 
11977 		if ((buf->dtb_xamot = kmem_zalloc(size,
11978 		    KM_NOSLEEP | KM_NORMALPRI)) == NULL)
11979 			goto err;
11980 	}
11981 
11982 	return (0);
11983 
11984 err:
11985 	/*
11986 	 * Error allocating memory, so free the buffers that were
11987 	 * allocated before the failed allocation.
11988 	 */
11989 	CPU_FOREACH(i) {
11990 		if (cpu != DTRACE_CPUALL && cpu != i)
11991 			continue;
11992 
11993 		buf = &bufs[i];
11994 		desired += 2;
11995 
11996 		if (buf->dtb_xamot != NULL) {
11997 			ASSERT(buf->dtb_tomax != NULL);
11998 			ASSERT(buf->dtb_size == size);
11999 			kmem_free(buf->dtb_xamot, size);
12000 			allocated++;
12001 		}
12002 
12003 		if (buf->dtb_tomax != NULL) {
12004 			ASSERT(buf->dtb_size == size);
12005 			kmem_free(buf->dtb_tomax, size);
12006 			allocated++;
12007 		}
12008 
12009 		buf->dtb_tomax = NULL;
12010 		buf->dtb_xamot = NULL;
12011 		buf->dtb_size = 0;
12012 
12013 	}
12014 #endif
12015 	*factor = desired / (allocated > 0 ? allocated : 1);
12016 
12017 	return (ENOMEM);
12018 }
12019 
12020 /*
12021  * Note:  called from probe context.  This function just increments the drop
12022  * count on a buffer.  It has been made a function to allow for the
12023  * possibility of understanding the source of mysterious drop counts.  (A
12024  * problem for which one may be particularly disappointed that DTrace cannot
12025  * be used to understand DTrace.)
12026  */
12027 static void
12028 dtrace_buffer_drop(dtrace_buffer_t *buf)
12029 {
12030 	buf->dtb_drops++;
12031 }
12032 
12033 /*
12034  * Note:  called from probe context.  This function is called to reserve space
12035  * in a buffer.  If mstate is non-NULL, sets the scratch base and size in the
12036  * mstate.  Returns the new offset in the buffer, or a negative value if an
12037  * error has occurred.
12038  */
12039 static intptr_t
12040 dtrace_buffer_reserve(dtrace_buffer_t *buf, size_t needed, size_t align,
12041     dtrace_state_t *state, dtrace_mstate_t *mstate)
12042 {
12043 	intptr_t offs = buf->dtb_offset, soffs;
12044 	intptr_t woffs;
12045 	caddr_t tomax;
12046 	size_t total;
12047 
12048 	if (buf->dtb_flags & DTRACEBUF_INACTIVE)
12049 		return (-1);
12050 
12051 	if ((tomax = buf->dtb_tomax) == NULL) {
12052 		dtrace_buffer_drop(buf);
12053 		return (-1);
12054 	}
12055 
12056 	if (!(buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL))) {
12057 		while (offs & (align - 1)) {
12058 			/*
12059 			 * Assert that our alignment is off by a number which
12060 			 * is itself sizeof (uint32_t) aligned.
12061 			 */
12062 			ASSERT(!((align - (offs & (align - 1))) &
12063 			    (sizeof (uint32_t) - 1)));
12064 			DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
12065 			offs += sizeof (uint32_t);
12066 		}
12067 
12068 		if ((soffs = offs + needed) > buf->dtb_size) {
12069 			dtrace_buffer_drop(buf);
12070 			return (-1);
12071 		}
12072 
12073 		if (mstate == NULL)
12074 			return (offs);
12075 
12076 		mstate->dtms_scratch_base = (uintptr_t)tomax + soffs;
12077 		mstate->dtms_scratch_size = buf->dtb_size - soffs;
12078 		mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
12079 
12080 		return (offs);
12081 	}
12082 
12083 	if (buf->dtb_flags & DTRACEBUF_FILL) {
12084 		if (state->dts_activity != DTRACE_ACTIVITY_COOLDOWN &&
12085 		    (buf->dtb_flags & DTRACEBUF_FULL))
12086 			return (-1);
12087 		goto out;
12088 	}
12089 
12090 	total = needed + (offs & (align - 1));
12091 
12092 	/*
12093 	 * For a ring buffer, life is quite a bit more complicated.  Before
12094 	 * we can store any padding, we need to adjust our wrapping offset.
12095 	 * (If we've never before wrapped or we're not about to, no adjustment
12096 	 * is required.)
12097 	 */
12098 	if ((buf->dtb_flags & DTRACEBUF_WRAPPED) ||
12099 	    offs + total > buf->dtb_size) {
12100 		woffs = buf->dtb_xamot_offset;
12101 
12102 		if (offs + total > buf->dtb_size) {
12103 			/*
12104 			 * We can't fit in the end of the buffer.  First, a
12105 			 * sanity check that we can fit in the buffer at all.
12106 			 */
12107 			if (total > buf->dtb_size) {
12108 				dtrace_buffer_drop(buf);
12109 				return (-1);
12110 			}
12111 
12112 			/*
12113 			 * We're going to be storing at the top of the buffer,
12114 			 * so now we need to deal with the wrapped offset.  We
12115 			 * only reset our wrapped offset to 0 if it is
12116 			 * currently greater than the current offset.  If it
12117 			 * is less than the current offset, it is because a
12118 			 * previous allocation induced a wrap -- but the
12119 			 * allocation didn't subsequently take the space due
12120 			 * to an error or false predicate evaluation.  In this
12121 			 * case, we'll just leave the wrapped offset alone: if
12122 			 * the wrapped offset hasn't been advanced far enough
12123 			 * for this allocation, it will be adjusted in the
12124 			 * lower loop.
12125 			 */
12126 			if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
12127 				if (woffs >= offs)
12128 					woffs = 0;
12129 			} else {
12130 				woffs = 0;
12131 			}
12132 
12133 			/*
12134 			 * Now we know that we're going to be storing to the
12135 			 * top of the buffer and that there is room for us
12136 			 * there.  We need to clear the buffer from the current
12137 			 * offset to the end (there may be old gunk there).
12138 			 */
12139 			while (offs < buf->dtb_size)
12140 				tomax[offs++] = 0;
12141 
12142 			/*
12143 			 * We need to set our offset to zero.  And because we
12144 			 * are wrapping, we need to set the bit indicating as
12145 			 * much.  We can also adjust our needed space back
12146 			 * down to the space required by the ECB -- we know
12147 			 * that the top of the buffer is aligned.
12148 			 */
12149 			offs = 0;
12150 			total = needed;
12151 			buf->dtb_flags |= DTRACEBUF_WRAPPED;
12152 		} else {
12153 			/*
12154 			 * There is room for us in the buffer, so we simply
12155 			 * need to check the wrapped offset.
12156 			 */
12157 			if (woffs < offs) {
12158 				/*
12159 				 * The wrapped offset is less than the offset.
12160 				 * This can happen if we allocated buffer space
12161 				 * that induced a wrap, but then we didn't
12162 				 * subsequently take the space due to an error
12163 				 * or false predicate evaluation.  This is
12164 				 * okay; we know that _this_ allocation isn't
12165 				 * going to induce a wrap.  We still can't
12166 				 * reset the wrapped offset to be zero,
12167 				 * however: the space may have been trashed in
12168 				 * the previous failed probe attempt.  But at
12169 				 * least the wrapped offset doesn't need to
12170 				 * be adjusted at all...
12171 				 */
12172 				goto out;
12173 			}
12174 		}
12175 
12176 		while (offs + total > woffs) {
12177 			dtrace_epid_t epid = *(uint32_t *)(tomax + woffs);
12178 			size_t size;
12179 
12180 			if (epid == DTRACE_EPIDNONE) {
12181 				size = sizeof (uint32_t);
12182 			} else {
12183 				ASSERT3U(epid, <=, state->dts_necbs);
12184 				ASSERT(state->dts_ecbs[epid - 1] != NULL);
12185 
12186 				size = state->dts_ecbs[epid - 1]->dte_size;
12187 			}
12188 
12189 			ASSERT(woffs + size <= buf->dtb_size);
12190 			ASSERT(size != 0);
12191 
12192 			if (woffs + size == buf->dtb_size) {
12193 				/*
12194 				 * We've reached the end of the buffer; we want
12195 				 * to set the wrapped offset to 0 and break
12196 				 * out.  However, if the offs is 0, then we're
12197 				 * in a strange edge-condition:  the amount of
12198 				 * space that we want to reserve plus the size
12199 				 * of the record that we're overwriting is
12200 				 * greater than the size of the buffer.  This
12201 				 * is problematic because if we reserve the
12202 				 * space but subsequently don't consume it (due
12203 				 * to a failed predicate or error) the wrapped
12204 				 * offset will be 0 -- yet the EPID at offset 0
12205 				 * will not be committed.  This situation is
12206 				 * relatively easy to deal with:  if we're in
12207 				 * this case, the buffer is indistinguishable
12208 				 * from one that hasn't wrapped; we need only
12209 				 * finish the job by clearing the wrapped bit,
12210 				 * explicitly setting the offset to be 0, and
12211 				 * zero'ing out the old data in the buffer.
12212 				 */
12213 				if (offs == 0) {
12214 					buf->dtb_flags &= ~DTRACEBUF_WRAPPED;
12215 					buf->dtb_offset = 0;
12216 					woffs = total;
12217 
12218 					while (woffs < buf->dtb_size)
12219 						tomax[woffs++] = 0;
12220 				}
12221 
12222 				woffs = 0;
12223 				break;
12224 			}
12225 
12226 			woffs += size;
12227 		}
12228 
12229 		/*
12230 		 * We have a wrapped offset.  It may be that the wrapped offset
12231 		 * has become zero -- that's okay.
12232 		 */
12233 		buf->dtb_xamot_offset = woffs;
12234 	}
12235 
12236 out:
12237 	/*
12238 	 * Now we can plow the buffer with any necessary padding.
12239 	 */
12240 	while (offs & (align - 1)) {
12241 		/*
12242 		 * Assert that our alignment is off by a number which
12243 		 * is itself sizeof (uint32_t) aligned.
12244 		 */
12245 		ASSERT(!((align - (offs & (align - 1))) &
12246 		    (sizeof (uint32_t) - 1)));
12247 		DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
12248 		offs += sizeof (uint32_t);
12249 	}
12250 
12251 	if (buf->dtb_flags & DTRACEBUF_FILL) {
12252 		if (offs + needed > buf->dtb_size - state->dts_reserve) {
12253 			buf->dtb_flags |= DTRACEBUF_FULL;
12254 			return (-1);
12255 		}
12256 	}
12257 
12258 	if (mstate == NULL)
12259 		return (offs);
12260 
12261 	/*
12262 	 * For ring buffers and fill buffers, the scratch space is always
12263 	 * the inactive buffer.
12264 	 */
12265 	mstate->dtms_scratch_base = (uintptr_t)buf->dtb_xamot;
12266 	mstate->dtms_scratch_size = buf->dtb_size;
12267 	mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
12268 
12269 	return (offs);
12270 }
12271 
12272 static void
12273 dtrace_buffer_polish(dtrace_buffer_t *buf)
12274 {
12275 	ASSERT(buf->dtb_flags & DTRACEBUF_RING);
12276 	ASSERT(MUTEX_HELD(&dtrace_lock));
12277 
12278 	if (!(buf->dtb_flags & DTRACEBUF_WRAPPED))
12279 		return;
12280 
12281 	/*
12282 	 * We need to polish the ring buffer.  There are three cases:
12283 	 *
12284 	 * - The first (and presumably most common) is that there is no gap
12285 	 *   between the buffer offset and the wrapped offset.  In this case,
12286 	 *   there is nothing in the buffer that isn't valid data; we can
12287 	 *   mark the buffer as polished and return.
12288 	 *
12289 	 * - The second (less common than the first but still more common
12290 	 *   than the third) is that there is a gap between the buffer offset
12291 	 *   and the wrapped offset, and the wrapped offset is larger than the
12292 	 *   buffer offset.  This can happen because of an alignment issue, or
12293 	 *   can happen because of a call to dtrace_buffer_reserve() that
12294 	 *   didn't subsequently consume the buffer space.  In this case,
12295 	 *   we need to zero the data from the buffer offset to the wrapped
12296 	 *   offset.
12297 	 *
12298 	 * - The third (and least common) is that there is a gap between the
12299 	 *   buffer offset and the wrapped offset, but the wrapped offset is
12300 	 *   _less_ than the buffer offset.  This can only happen because a
12301 	 *   call to dtrace_buffer_reserve() induced a wrap, but the space
12302 	 *   was not subsequently consumed.  In this case, we need to zero the
12303 	 *   space from the offset to the end of the buffer _and_ from the
12304 	 *   top of the buffer to the wrapped offset.
12305 	 */
12306 	if (buf->dtb_offset < buf->dtb_xamot_offset) {
12307 		bzero(buf->dtb_tomax + buf->dtb_offset,
12308 		    buf->dtb_xamot_offset - buf->dtb_offset);
12309 	}
12310 
12311 	if (buf->dtb_offset > buf->dtb_xamot_offset) {
12312 		bzero(buf->dtb_tomax + buf->dtb_offset,
12313 		    buf->dtb_size - buf->dtb_offset);
12314 		bzero(buf->dtb_tomax, buf->dtb_xamot_offset);
12315 	}
12316 }
12317 
12318 /*
12319  * This routine determines if data generated at the specified time has likely
12320  * been entirely consumed at user-level.  This routine is called to determine
12321  * if an ECB on a defunct probe (but for an active enabling) can be safely
12322  * disabled and destroyed.
12323  */
12324 static int
12325 dtrace_buffer_consumed(dtrace_buffer_t *bufs, hrtime_t when)
12326 {
12327 	int i;
12328 
12329 	for (i = 0; i < NCPU; i++) {
12330 		dtrace_buffer_t *buf = &bufs[i];
12331 
12332 		if (buf->dtb_size == 0)
12333 			continue;
12334 
12335 		if (buf->dtb_flags & DTRACEBUF_RING)
12336 			return (0);
12337 
12338 		if (!buf->dtb_switched && buf->dtb_offset != 0)
12339 			return (0);
12340 
12341 		if (buf->dtb_switched - buf->dtb_interval < when)
12342 			return (0);
12343 	}
12344 
12345 	return (1);
12346 }
12347 
12348 static void
12349 dtrace_buffer_free(dtrace_buffer_t *bufs)
12350 {
12351 	int i;
12352 
12353 	for (i = 0; i < NCPU; i++) {
12354 		dtrace_buffer_t *buf = &bufs[i];
12355 
12356 		if (buf->dtb_tomax == NULL) {
12357 			ASSERT(buf->dtb_xamot == NULL);
12358 			ASSERT(buf->dtb_size == 0);
12359 			continue;
12360 		}
12361 
12362 		if (buf->dtb_xamot != NULL) {
12363 			ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
12364 			kmem_free(buf->dtb_xamot, buf->dtb_size);
12365 		}
12366 
12367 		kmem_free(buf->dtb_tomax, buf->dtb_size);
12368 		buf->dtb_size = 0;
12369 		buf->dtb_tomax = NULL;
12370 		buf->dtb_xamot = NULL;
12371 	}
12372 }
12373 
12374 /*
12375  * DTrace Enabling Functions
12376  */
12377 static dtrace_enabling_t *
12378 dtrace_enabling_create(dtrace_vstate_t *vstate)
12379 {
12380 	dtrace_enabling_t *enab;
12381 
12382 	enab = kmem_zalloc(sizeof (dtrace_enabling_t), KM_SLEEP);
12383 	enab->dten_vstate = vstate;
12384 
12385 	return (enab);
12386 }
12387 
12388 static void
12389 dtrace_enabling_add(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb)
12390 {
12391 	dtrace_ecbdesc_t **ndesc;
12392 	size_t osize, nsize;
12393 
12394 	/*
12395 	 * We can't add to enablings after we've enabled them, or after we've
12396 	 * retained them.
12397 	 */
12398 	ASSERT(enab->dten_probegen == 0);
12399 	ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
12400 
12401 	if (enab->dten_ndesc < enab->dten_maxdesc) {
12402 		enab->dten_desc[enab->dten_ndesc++] = ecb;
12403 		return;
12404 	}
12405 
12406 	osize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
12407 
12408 	if (enab->dten_maxdesc == 0) {
12409 		enab->dten_maxdesc = 1;
12410 	} else {
12411 		enab->dten_maxdesc <<= 1;
12412 	}
12413 
12414 	ASSERT(enab->dten_ndesc < enab->dten_maxdesc);
12415 
12416 	nsize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
12417 	ndesc = kmem_zalloc(nsize, KM_SLEEP);
12418 	bcopy(enab->dten_desc, ndesc, osize);
12419 	if (enab->dten_desc != NULL)
12420 		kmem_free(enab->dten_desc, osize);
12421 
12422 	enab->dten_desc = ndesc;
12423 	enab->dten_desc[enab->dten_ndesc++] = ecb;
12424 }
12425 
12426 static void
12427 dtrace_enabling_addlike(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb,
12428     dtrace_probedesc_t *pd)
12429 {
12430 	dtrace_ecbdesc_t *new;
12431 	dtrace_predicate_t *pred;
12432 	dtrace_actdesc_t *act;
12433 
12434 	/*
12435 	 * We're going to create a new ECB description that matches the
12436 	 * specified ECB in every way, but has the specified probe description.
12437 	 */
12438 	new = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
12439 
12440 	if ((pred = ecb->dted_pred.dtpdd_predicate) != NULL)
12441 		dtrace_predicate_hold(pred);
12442 
12443 	for (act = ecb->dted_action; act != NULL; act = act->dtad_next)
12444 		dtrace_actdesc_hold(act);
12445 
12446 	new->dted_action = ecb->dted_action;
12447 	new->dted_pred = ecb->dted_pred;
12448 	new->dted_probe = *pd;
12449 	new->dted_uarg = ecb->dted_uarg;
12450 
12451 	dtrace_enabling_add(enab, new);
12452 }
12453 
12454 static void
12455 dtrace_enabling_dump(dtrace_enabling_t *enab)
12456 {
12457 	int i;
12458 
12459 	for (i = 0; i < enab->dten_ndesc; i++) {
12460 		dtrace_probedesc_t *desc = &enab->dten_desc[i]->dted_probe;
12461 
12462 		cmn_err(CE_NOTE, "enabling probe %d (%s:%s:%s:%s)", i,
12463 		    desc->dtpd_provider, desc->dtpd_mod,
12464 		    desc->dtpd_func, desc->dtpd_name);
12465 	}
12466 }
12467 
12468 static void
12469 dtrace_enabling_destroy(dtrace_enabling_t *enab)
12470 {
12471 	int i;
12472 	dtrace_ecbdesc_t *ep;
12473 	dtrace_vstate_t *vstate = enab->dten_vstate;
12474 
12475 	ASSERT(MUTEX_HELD(&dtrace_lock));
12476 
12477 	for (i = 0; i < enab->dten_ndesc; i++) {
12478 		dtrace_actdesc_t *act, *next;
12479 		dtrace_predicate_t *pred;
12480 
12481 		ep = enab->dten_desc[i];
12482 
12483 		if ((pred = ep->dted_pred.dtpdd_predicate) != NULL)
12484 			dtrace_predicate_release(pred, vstate);
12485 
12486 		for (act = ep->dted_action; act != NULL; act = next) {
12487 			next = act->dtad_next;
12488 			dtrace_actdesc_release(act, vstate);
12489 		}
12490 
12491 		kmem_free(ep, sizeof (dtrace_ecbdesc_t));
12492 	}
12493 
12494 	if (enab->dten_desc != NULL)
12495 		kmem_free(enab->dten_desc,
12496 		    enab->dten_maxdesc * sizeof (dtrace_enabling_t *));
12497 
12498 	/*
12499 	 * If this was a retained enabling, decrement the dts_nretained count
12500 	 * and take it off of the dtrace_retained list.
12501 	 */
12502 	if (enab->dten_prev != NULL || enab->dten_next != NULL ||
12503 	    dtrace_retained == enab) {
12504 		ASSERT(enab->dten_vstate->dtvs_state != NULL);
12505 		ASSERT(enab->dten_vstate->dtvs_state->dts_nretained > 0);
12506 		enab->dten_vstate->dtvs_state->dts_nretained--;
12507 		dtrace_retained_gen++;
12508 	}
12509 
12510 	if (enab->dten_prev == NULL) {
12511 		if (dtrace_retained == enab) {
12512 			dtrace_retained = enab->dten_next;
12513 
12514 			if (dtrace_retained != NULL)
12515 				dtrace_retained->dten_prev = NULL;
12516 		}
12517 	} else {
12518 		ASSERT(enab != dtrace_retained);
12519 		ASSERT(dtrace_retained != NULL);
12520 		enab->dten_prev->dten_next = enab->dten_next;
12521 	}
12522 
12523 	if (enab->dten_next != NULL) {
12524 		ASSERT(dtrace_retained != NULL);
12525 		enab->dten_next->dten_prev = enab->dten_prev;
12526 	}
12527 
12528 	kmem_free(enab, sizeof (dtrace_enabling_t));
12529 }
12530 
12531 static int
12532 dtrace_enabling_retain(dtrace_enabling_t *enab)
12533 {
12534 	dtrace_state_t *state;
12535 
12536 	ASSERT(MUTEX_HELD(&dtrace_lock));
12537 	ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
12538 	ASSERT(enab->dten_vstate != NULL);
12539 
12540 	state = enab->dten_vstate->dtvs_state;
12541 	ASSERT(state != NULL);
12542 
12543 	/*
12544 	 * We only allow each state to retain dtrace_retain_max enablings.
12545 	 */
12546 	if (state->dts_nretained >= dtrace_retain_max)
12547 		return (ENOSPC);
12548 
12549 	state->dts_nretained++;
12550 	dtrace_retained_gen++;
12551 
12552 	if (dtrace_retained == NULL) {
12553 		dtrace_retained = enab;
12554 		return (0);
12555 	}
12556 
12557 	enab->dten_next = dtrace_retained;
12558 	dtrace_retained->dten_prev = enab;
12559 	dtrace_retained = enab;
12560 
12561 	return (0);
12562 }
12563 
12564 static int
12565 dtrace_enabling_replicate(dtrace_state_t *state, dtrace_probedesc_t *match,
12566     dtrace_probedesc_t *create)
12567 {
12568 	dtrace_enabling_t *new, *enab;
12569 	int found = 0, err = ENOENT;
12570 
12571 	ASSERT(MUTEX_HELD(&dtrace_lock));
12572 	ASSERT(strlen(match->dtpd_provider) < DTRACE_PROVNAMELEN);
12573 	ASSERT(strlen(match->dtpd_mod) < DTRACE_MODNAMELEN);
12574 	ASSERT(strlen(match->dtpd_func) < DTRACE_FUNCNAMELEN);
12575 	ASSERT(strlen(match->dtpd_name) < DTRACE_NAMELEN);
12576 
12577 	new = dtrace_enabling_create(&state->dts_vstate);
12578 
12579 	/*
12580 	 * Iterate over all retained enablings, looking for enablings that
12581 	 * match the specified state.
12582 	 */
12583 	for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
12584 		int i;
12585 
12586 		/*
12587 		 * dtvs_state can only be NULL for helper enablings -- and
12588 		 * helper enablings can't be retained.
12589 		 */
12590 		ASSERT(enab->dten_vstate->dtvs_state != NULL);
12591 
12592 		if (enab->dten_vstate->dtvs_state != state)
12593 			continue;
12594 
12595 		/*
12596 		 * Now iterate over each probe description; we're looking for
12597 		 * an exact match to the specified probe description.
12598 		 */
12599 		for (i = 0; i < enab->dten_ndesc; i++) {
12600 			dtrace_ecbdesc_t *ep = enab->dten_desc[i];
12601 			dtrace_probedesc_t *pd = &ep->dted_probe;
12602 
12603 			if (strcmp(pd->dtpd_provider, match->dtpd_provider))
12604 				continue;
12605 
12606 			if (strcmp(pd->dtpd_mod, match->dtpd_mod))
12607 				continue;
12608 
12609 			if (strcmp(pd->dtpd_func, match->dtpd_func))
12610 				continue;
12611 
12612 			if (strcmp(pd->dtpd_name, match->dtpd_name))
12613 				continue;
12614 
12615 			/*
12616 			 * We have a winning probe!  Add it to our growing
12617 			 * enabling.
12618 			 */
12619 			found = 1;
12620 			dtrace_enabling_addlike(new, ep, create);
12621 		}
12622 	}
12623 
12624 	if (!found || (err = dtrace_enabling_retain(new)) != 0) {
12625 		dtrace_enabling_destroy(new);
12626 		return (err);
12627 	}
12628 
12629 	return (0);
12630 }
12631 
12632 static void
12633 dtrace_enabling_retract(dtrace_state_t *state)
12634 {
12635 	dtrace_enabling_t *enab, *next;
12636 
12637 	ASSERT(MUTEX_HELD(&dtrace_lock));
12638 
12639 	/*
12640 	 * Iterate over all retained enablings, destroy the enablings retained
12641 	 * for the specified state.
12642 	 */
12643 	for (enab = dtrace_retained; enab != NULL; enab = next) {
12644 		next = enab->dten_next;
12645 
12646 		/*
12647 		 * dtvs_state can only be NULL for helper enablings -- and
12648 		 * helper enablings can't be retained.
12649 		 */
12650 		ASSERT(enab->dten_vstate->dtvs_state != NULL);
12651 
12652 		if (enab->dten_vstate->dtvs_state == state) {
12653 			ASSERT(state->dts_nretained > 0);
12654 			dtrace_enabling_destroy(enab);
12655 		}
12656 	}
12657 
12658 	ASSERT(state->dts_nretained == 0);
12659 }
12660 
12661 static int
12662 dtrace_enabling_match(dtrace_enabling_t *enab, int *nmatched)
12663 {
12664 	int i = 0;
12665 	int matched = 0;
12666 
12667 	ASSERT(MUTEX_HELD(&cpu_lock));
12668 	ASSERT(MUTEX_HELD(&dtrace_lock));
12669 
12670 	for (i = 0; i < enab->dten_ndesc; i++) {
12671 		dtrace_ecbdesc_t *ep = enab->dten_desc[i];
12672 
12673 		enab->dten_current = ep;
12674 		enab->dten_error = 0;
12675 
12676 		matched += dtrace_probe_enable(&ep->dted_probe, enab);
12677 
12678 		if (enab->dten_error != 0) {
12679 			/*
12680 			 * If we get an error half-way through enabling the
12681 			 * probes, we kick out -- perhaps with some number of
12682 			 * them enabled.  Leaving enabled probes enabled may
12683 			 * be slightly confusing for user-level, but we expect
12684 			 * that no one will attempt to actually drive on in
12685 			 * the face of such errors.  If this is an anonymous
12686 			 * enabling (indicated with a NULL nmatched pointer),
12687 			 * we cmn_err() a message.  We aren't expecting to
12688 			 * get such an error -- such as it can exist at all,
12689 			 * it would be a result of corrupted DOF in the driver
12690 			 * properties.
12691 			 */
12692 			if (nmatched == NULL) {
12693 				cmn_err(CE_WARN, "dtrace_enabling_match() "
12694 				    "error on %p: %d", (void *)ep,
12695 				    enab->dten_error);
12696 			}
12697 
12698 			return (enab->dten_error);
12699 		}
12700 	}
12701 
12702 	enab->dten_probegen = dtrace_probegen;
12703 	if (nmatched != NULL)
12704 		*nmatched = matched;
12705 
12706 	return (0);
12707 }
12708 
12709 static void
12710 dtrace_enabling_matchall(void)
12711 {
12712 	dtrace_enabling_t *enab;
12713 
12714 	mutex_enter(&cpu_lock);
12715 	mutex_enter(&dtrace_lock);
12716 
12717 	/*
12718 	 * Iterate over all retained enablings to see if any probes match
12719 	 * against them.  We only perform this operation on enablings for which
12720 	 * we have sufficient permissions by virtue of being in the global zone
12721 	 * or in the same zone as the DTrace client.  Because we can be called
12722 	 * after dtrace_detach() has been called, we cannot assert that there
12723 	 * are retained enablings.  We can safely load from dtrace_retained,
12724 	 * however:  the taskq_destroy() at the end of dtrace_detach() will
12725 	 * block pending our completion.
12726 	 */
12727 	for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
12728 #ifdef illumos
12729 		cred_t *cr = enab->dten_vstate->dtvs_state->dts_cred.dcr_cred;
12730 
12731 		if (INGLOBALZONE(curproc) ||
12732 		    cr != NULL && getzoneid() == crgetzoneid(cr))
12733 #endif
12734 			(void) dtrace_enabling_match(enab, NULL);
12735 	}
12736 
12737 	mutex_exit(&dtrace_lock);
12738 	mutex_exit(&cpu_lock);
12739 }
12740 
12741 /*
12742  * If an enabling is to be enabled without having matched probes (that is, if
12743  * dtrace_state_go() is to be called on the underlying dtrace_state_t), the
12744  * enabling must be _primed_ by creating an ECB for every ECB description.
12745  * This must be done to assure that we know the number of speculations, the
12746  * number of aggregations, the minimum buffer size needed, etc. before we
12747  * transition out of DTRACE_ACTIVITY_INACTIVE.  To do this without actually
12748  * enabling any probes, we create ECBs for every ECB decription, but with a
12749  * NULL probe -- which is exactly what this function does.
12750  */
12751 static void
12752 dtrace_enabling_prime(dtrace_state_t *state)
12753 {
12754 	dtrace_enabling_t *enab;
12755 	int i;
12756 
12757 	for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
12758 		ASSERT(enab->dten_vstate->dtvs_state != NULL);
12759 
12760 		if (enab->dten_vstate->dtvs_state != state)
12761 			continue;
12762 
12763 		/*
12764 		 * We don't want to prime an enabling more than once, lest
12765 		 * we allow a malicious user to induce resource exhaustion.
12766 		 * (The ECBs that result from priming an enabling aren't
12767 		 * leaked -- but they also aren't deallocated until the
12768 		 * consumer state is destroyed.)
12769 		 */
12770 		if (enab->dten_primed)
12771 			continue;
12772 
12773 		for (i = 0; i < enab->dten_ndesc; i++) {
12774 			enab->dten_current = enab->dten_desc[i];
12775 			(void) dtrace_probe_enable(NULL, enab);
12776 		}
12777 
12778 		enab->dten_primed = 1;
12779 	}
12780 }
12781 
12782 /*
12783  * Called to indicate that probes should be provided due to retained
12784  * enablings.  This is implemented in terms of dtrace_probe_provide(), but it
12785  * must take an initial lap through the enabling calling the dtps_provide()
12786  * entry point explicitly to allow for autocreated probes.
12787  */
12788 static void
12789 dtrace_enabling_provide(dtrace_provider_t *prv)
12790 {
12791 	int i, all = 0;
12792 	dtrace_probedesc_t desc;
12793 	dtrace_genid_t gen;
12794 
12795 	ASSERT(MUTEX_HELD(&dtrace_lock));
12796 	ASSERT(MUTEX_HELD(&dtrace_provider_lock));
12797 
12798 	if (prv == NULL) {
12799 		all = 1;
12800 		prv = dtrace_provider;
12801 	}
12802 
12803 	do {
12804 		dtrace_enabling_t *enab;
12805 		void *parg = prv->dtpv_arg;
12806 
12807 retry:
12808 		gen = dtrace_retained_gen;
12809 		for (enab = dtrace_retained; enab != NULL;
12810 		    enab = enab->dten_next) {
12811 			for (i = 0; i < enab->dten_ndesc; i++) {
12812 				desc = enab->dten_desc[i]->dted_probe;
12813 				mutex_exit(&dtrace_lock);
12814 				prv->dtpv_pops.dtps_provide(parg, &desc);
12815 				mutex_enter(&dtrace_lock);
12816 				/*
12817 				 * Process the retained enablings again if
12818 				 * they have changed while we weren't holding
12819 				 * dtrace_lock.
12820 				 */
12821 				if (gen != dtrace_retained_gen)
12822 					goto retry;
12823 			}
12824 		}
12825 	} while (all && (prv = prv->dtpv_next) != NULL);
12826 
12827 	mutex_exit(&dtrace_lock);
12828 	dtrace_probe_provide(NULL, all ? NULL : prv);
12829 	mutex_enter(&dtrace_lock);
12830 }
12831 
12832 /*
12833  * Called to reap ECBs that are attached to probes from defunct providers.
12834  */
12835 static void
12836 dtrace_enabling_reap(void)
12837 {
12838 	dtrace_provider_t *prov;
12839 	dtrace_probe_t *probe;
12840 	dtrace_ecb_t *ecb;
12841 	hrtime_t when;
12842 	int i;
12843 
12844 	mutex_enter(&cpu_lock);
12845 	mutex_enter(&dtrace_lock);
12846 
12847 	for (i = 0; i < dtrace_nprobes; i++) {
12848 		if ((probe = dtrace_probes[i]) == NULL)
12849 			continue;
12850 
12851 		if (probe->dtpr_ecb == NULL)
12852 			continue;
12853 
12854 		prov = probe->dtpr_provider;
12855 
12856 		if ((when = prov->dtpv_defunct) == 0)
12857 			continue;
12858 
12859 		/*
12860 		 * We have ECBs on a defunct provider:  we want to reap these
12861 		 * ECBs to allow the provider to unregister.  The destruction
12862 		 * of these ECBs must be done carefully:  if we destroy the ECB
12863 		 * and the consumer later wishes to consume an EPID that
12864 		 * corresponds to the destroyed ECB (and if the EPID metadata
12865 		 * has not been previously consumed), the consumer will abort
12866 		 * processing on the unknown EPID.  To reduce (but not, sadly,
12867 		 * eliminate) the possibility of this, we will only destroy an
12868 		 * ECB for a defunct provider if, for the state that
12869 		 * corresponds to the ECB:
12870 		 *
12871 		 *  (a)	There is no speculative tracing (which can effectively
12872 		 *	cache an EPID for an arbitrary amount of time).
12873 		 *
12874 		 *  (b)	The principal buffers have been switched twice since the
12875 		 *	provider became defunct.
12876 		 *
12877 		 *  (c)	The aggregation buffers are of zero size or have been
12878 		 *	switched twice since the provider became defunct.
12879 		 *
12880 		 * We use dts_speculates to determine (a) and call a function
12881 		 * (dtrace_buffer_consumed()) to determine (b) and (c).  Note
12882 		 * that as soon as we've been unable to destroy one of the ECBs
12883 		 * associated with the probe, we quit trying -- reaping is only
12884 		 * fruitful in as much as we can destroy all ECBs associated
12885 		 * with the defunct provider's probes.
12886 		 */
12887 		while ((ecb = probe->dtpr_ecb) != NULL) {
12888 			dtrace_state_t *state = ecb->dte_state;
12889 			dtrace_buffer_t *buf = state->dts_buffer;
12890 			dtrace_buffer_t *aggbuf = state->dts_aggbuffer;
12891 
12892 			if (state->dts_speculates)
12893 				break;
12894 
12895 			if (!dtrace_buffer_consumed(buf, when))
12896 				break;
12897 
12898 			if (!dtrace_buffer_consumed(aggbuf, when))
12899 				break;
12900 
12901 			dtrace_ecb_disable(ecb);
12902 			ASSERT(probe->dtpr_ecb != ecb);
12903 			dtrace_ecb_destroy(ecb);
12904 		}
12905 	}
12906 
12907 	mutex_exit(&dtrace_lock);
12908 	mutex_exit(&cpu_lock);
12909 }
12910 
12911 /*
12912  * DTrace DOF Functions
12913  */
12914 /*ARGSUSED*/
12915 static void
12916 dtrace_dof_error(dof_hdr_t *dof, const char *str)
12917 {
12918 	if (dtrace_err_verbose)
12919 		cmn_err(CE_WARN, "failed to process DOF: %s", str);
12920 
12921 #ifdef DTRACE_ERRDEBUG
12922 	dtrace_errdebug(str);
12923 #endif
12924 }
12925 
12926 /*
12927  * Create DOF out of a currently enabled state.  Right now, we only create
12928  * DOF containing the run-time options -- but this could be expanded to create
12929  * complete DOF representing the enabled state.
12930  */
12931 static dof_hdr_t *
12932 dtrace_dof_create(dtrace_state_t *state)
12933 {
12934 	dof_hdr_t *dof;
12935 	dof_sec_t *sec;
12936 	dof_optdesc_t *opt;
12937 	int i, len = sizeof (dof_hdr_t) +
12938 	    roundup(sizeof (dof_sec_t), sizeof (uint64_t)) +
12939 	    sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
12940 
12941 	ASSERT(MUTEX_HELD(&dtrace_lock));
12942 
12943 	dof = kmem_zalloc(len, KM_SLEEP);
12944 	dof->dofh_ident[DOF_ID_MAG0] = DOF_MAG_MAG0;
12945 	dof->dofh_ident[DOF_ID_MAG1] = DOF_MAG_MAG1;
12946 	dof->dofh_ident[DOF_ID_MAG2] = DOF_MAG_MAG2;
12947 	dof->dofh_ident[DOF_ID_MAG3] = DOF_MAG_MAG3;
12948 
12949 	dof->dofh_ident[DOF_ID_MODEL] = DOF_MODEL_NATIVE;
12950 	dof->dofh_ident[DOF_ID_ENCODING] = DOF_ENCODE_NATIVE;
12951 	dof->dofh_ident[DOF_ID_VERSION] = DOF_VERSION;
12952 	dof->dofh_ident[DOF_ID_DIFVERS] = DIF_VERSION;
12953 	dof->dofh_ident[DOF_ID_DIFIREG] = DIF_DIR_NREGS;
12954 	dof->dofh_ident[DOF_ID_DIFTREG] = DIF_DTR_NREGS;
12955 
12956 	dof->dofh_flags = 0;
12957 	dof->dofh_hdrsize = sizeof (dof_hdr_t);
12958 	dof->dofh_secsize = sizeof (dof_sec_t);
12959 	dof->dofh_secnum = 1;	/* only DOF_SECT_OPTDESC */
12960 	dof->dofh_secoff = sizeof (dof_hdr_t);
12961 	dof->dofh_loadsz = len;
12962 	dof->dofh_filesz = len;
12963 	dof->dofh_pad = 0;
12964 
12965 	/*
12966 	 * Fill in the option section header...
12967 	 */
12968 	sec = (dof_sec_t *)((uintptr_t)dof + sizeof (dof_hdr_t));
12969 	sec->dofs_type = DOF_SECT_OPTDESC;
12970 	sec->dofs_align = sizeof (uint64_t);
12971 	sec->dofs_flags = DOF_SECF_LOAD;
12972 	sec->dofs_entsize = sizeof (dof_optdesc_t);
12973 
12974 	opt = (dof_optdesc_t *)((uintptr_t)sec +
12975 	    roundup(sizeof (dof_sec_t), sizeof (uint64_t)));
12976 
12977 	sec->dofs_offset = (uintptr_t)opt - (uintptr_t)dof;
12978 	sec->dofs_size = sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
12979 
12980 	for (i = 0; i < DTRACEOPT_MAX; i++) {
12981 		opt[i].dofo_option = i;
12982 		opt[i].dofo_strtab = DOF_SECIDX_NONE;
12983 		opt[i].dofo_value = state->dts_options[i];
12984 	}
12985 
12986 	return (dof);
12987 }
12988 
12989 static dof_hdr_t *
12990 dtrace_dof_copyin(uintptr_t uarg, int *errp)
12991 {
12992 	dof_hdr_t hdr, *dof;
12993 
12994 	ASSERT(!MUTEX_HELD(&dtrace_lock));
12995 
12996 	/*
12997 	 * First, we're going to copyin() the sizeof (dof_hdr_t).
12998 	 */
12999 	if (copyin((void *)uarg, &hdr, sizeof (hdr)) != 0) {
13000 		dtrace_dof_error(NULL, "failed to copyin DOF header");
13001 		*errp = EFAULT;
13002 		return (NULL);
13003 	}
13004 
13005 	/*
13006 	 * Now we'll allocate the entire DOF and copy it in -- provided
13007 	 * that the length isn't outrageous.
13008 	 */
13009 	if (hdr.dofh_loadsz >= dtrace_dof_maxsize) {
13010 		dtrace_dof_error(&hdr, "load size exceeds maximum");
13011 		*errp = E2BIG;
13012 		return (NULL);
13013 	}
13014 
13015 	if (hdr.dofh_loadsz < sizeof (hdr)) {
13016 		dtrace_dof_error(&hdr, "invalid load size");
13017 		*errp = EINVAL;
13018 		return (NULL);
13019 	}
13020 
13021 	dof = kmem_alloc(hdr.dofh_loadsz, KM_SLEEP);
13022 
13023 	if (copyin((void *)uarg, dof, hdr.dofh_loadsz) != 0 ||
13024 	    dof->dofh_loadsz != hdr.dofh_loadsz) {
13025 		kmem_free(dof, hdr.dofh_loadsz);
13026 		*errp = EFAULT;
13027 		return (NULL);
13028 	}
13029 
13030 	return (dof);
13031 }
13032 
13033 #ifndef illumos
13034 static __inline uchar_t
13035 dtrace_dof_char(char c) {
13036 	switch (c) {
13037 	case '0':
13038 	case '1':
13039 	case '2':
13040 	case '3':
13041 	case '4':
13042 	case '5':
13043 	case '6':
13044 	case '7':
13045 	case '8':
13046 	case '9':
13047 		return (c - '0');
13048 	case 'A':
13049 	case 'B':
13050 	case 'C':
13051 	case 'D':
13052 	case 'E':
13053 	case 'F':
13054 		return (c - 'A' + 10);
13055 	case 'a':
13056 	case 'b':
13057 	case 'c':
13058 	case 'd':
13059 	case 'e':
13060 	case 'f':
13061 		return (c - 'a' + 10);
13062 	}
13063 	/* Should not reach here. */
13064 	return (0);
13065 }
13066 #endif
13067 
13068 static dof_hdr_t *
13069 dtrace_dof_property(const char *name)
13070 {
13071 	uchar_t *buf;
13072 	uint64_t loadsz;
13073 	unsigned int len, i;
13074 	dof_hdr_t *dof;
13075 
13076 #ifdef illumos
13077 	/*
13078 	 * Unfortunately, array of values in .conf files are always (and
13079 	 * only) interpreted to be integer arrays.  We must read our DOF
13080 	 * as an integer array, and then squeeze it into a byte array.
13081 	 */
13082 	if (ddi_prop_lookup_int_array(DDI_DEV_T_ANY, dtrace_devi, 0,
13083 	    (char *)name, (int **)&buf, &len) != DDI_PROP_SUCCESS)
13084 		return (NULL);
13085 
13086 	for (i = 0; i < len; i++)
13087 		buf[i] = (uchar_t)(((int *)buf)[i]);
13088 
13089 	if (len < sizeof (dof_hdr_t)) {
13090 		ddi_prop_free(buf);
13091 		dtrace_dof_error(NULL, "truncated header");
13092 		return (NULL);
13093 	}
13094 
13095 	if (len < (loadsz = ((dof_hdr_t *)buf)->dofh_loadsz)) {
13096 		ddi_prop_free(buf);
13097 		dtrace_dof_error(NULL, "truncated DOF");
13098 		return (NULL);
13099 	}
13100 
13101 	if (loadsz >= dtrace_dof_maxsize) {
13102 		ddi_prop_free(buf);
13103 		dtrace_dof_error(NULL, "oversized DOF");
13104 		return (NULL);
13105 	}
13106 
13107 	dof = kmem_alloc(loadsz, KM_SLEEP);
13108 	bcopy(buf, dof, loadsz);
13109 	ddi_prop_free(buf);
13110 #else
13111 	char *p;
13112 	char *p_env;
13113 
13114 	if ((p_env = kern_getenv(name)) == NULL)
13115 		return (NULL);
13116 
13117 	len = strlen(p_env) / 2;
13118 
13119 	buf = kmem_alloc(len, KM_SLEEP);
13120 
13121 	dof = (dof_hdr_t *) buf;
13122 
13123 	p = p_env;
13124 
13125 	for (i = 0; i < len; i++) {
13126 		buf[i] = (dtrace_dof_char(p[0]) << 4) |
13127 		     dtrace_dof_char(p[1]);
13128 		p += 2;
13129 	}
13130 
13131 	freeenv(p_env);
13132 
13133 	if (len < sizeof (dof_hdr_t)) {
13134 		kmem_free(buf, 0);
13135 		dtrace_dof_error(NULL, "truncated header");
13136 		return (NULL);
13137 	}
13138 
13139 	if (len < (loadsz = dof->dofh_loadsz)) {
13140 		kmem_free(buf, 0);
13141 		dtrace_dof_error(NULL, "truncated DOF");
13142 		return (NULL);
13143 	}
13144 
13145 	if (loadsz >= dtrace_dof_maxsize) {
13146 		kmem_free(buf, 0);
13147 		dtrace_dof_error(NULL, "oversized DOF");
13148 		return (NULL);
13149 	}
13150 #endif
13151 
13152 	return (dof);
13153 }
13154 
13155 static void
13156 dtrace_dof_destroy(dof_hdr_t *dof)
13157 {
13158 	kmem_free(dof, dof->dofh_loadsz);
13159 }
13160 
13161 /*
13162  * Return the dof_sec_t pointer corresponding to a given section index.  If the
13163  * index is not valid, dtrace_dof_error() is called and NULL is returned.  If
13164  * a type other than DOF_SECT_NONE is specified, the header is checked against
13165  * this type and NULL is returned if the types do not match.
13166  */
13167 static dof_sec_t *
13168 dtrace_dof_sect(dof_hdr_t *dof, uint32_t type, dof_secidx_t i)
13169 {
13170 	dof_sec_t *sec = (dof_sec_t *)(uintptr_t)
13171 	    ((uintptr_t)dof + dof->dofh_secoff + i * dof->dofh_secsize);
13172 
13173 	if (i >= dof->dofh_secnum) {
13174 		dtrace_dof_error(dof, "referenced section index is invalid");
13175 		return (NULL);
13176 	}
13177 
13178 	if (!(sec->dofs_flags & DOF_SECF_LOAD)) {
13179 		dtrace_dof_error(dof, "referenced section is not loadable");
13180 		return (NULL);
13181 	}
13182 
13183 	if (type != DOF_SECT_NONE && type != sec->dofs_type) {
13184 		dtrace_dof_error(dof, "referenced section is the wrong type");
13185 		return (NULL);
13186 	}
13187 
13188 	return (sec);
13189 }
13190 
13191 static dtrace_probedesc_t *
13192 dtrace_dof_probedesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_probedesc_t *desc)
13193 {
13194 	dof_probedesc_t *probe;
13195 	dof_sec_t *strtab;
13196 	uintptr_t daddr = (uintptr_t)dof;
13197 	uintptr_t str;
13198 	size_t size;
13199 
13200 	if (sec->dofs_type != DOF_SECT_PROBEDESC) {
13201 		dtrace_dof_error(dof, "invalid probe section");
13202 		return (NULL);
13203 	}
13204 
13205 	if (sec->dofs_align != sizeof (dof_secidx_t)) {
13206 		dtrace_dof_error(dof, "bad alignment in probe description");
13207 		return (NULL);
13208 	}
13209 
13210 	if (sec->dofs_offset + sizeof (dof_probedesc_t) > dof->dofh_loadsz) {
13211 		dtrace_dof_error(dof, "truncated probe description");
13212 		return (NULL);
13213 	}
13214 
13215 	probe = (dof_probedesc_t *)(uintptr_t)(daddr + sec->dofs_offset);
13216 	strtab = dtrace_dof_sect(dof, DOF_SECT_STRTAB, probe->dofp_strtab);
13217 
13218 	if (strtab == NULL)
13219 		return (NULL);
13220 
13221 	str = daddr + strtab->dofs_offset;
13222 	size = strtab->dofs_size;
13223 
13224 	if (probe->dofp_provider >= strtab->dofs_size) {
13225 		dtrace_dof_error(dof, "corrupt probe provider");
13226 		return (NULL);
13227 	}
13228 
13229 	(void) strncpy(desc->dtpd_provider,
13230 	    (char *)(str + probe->dofp_provider),
13231 	    MIN(DTRACE_PROVNAMELEN - 1, size - probe->dofp_provider));
13232 
13233 	if (probe->dofp_mod >= strtab->dofs_size) {
13234 		dtrace_dof_error(dof, "corrupt probe module");
13235 		return (NULL);
13236 	}
13237 
13238 	(void) strncpy(desc->dtpd_mod, (char *)(str + probe->dofp_mod),
13239 	    MIN(DTRACE_MODNAMELEN - 1, size - probe->dofp_mod));
13240 
13241 	if (probe->dofp_func >= strtab->dofs_size) {
13242 		dtrace_dof_error(dof, "corrupt probe function");
13243 		return (NULL);
13244 	}
13245 
13246 	(void) strncpy(desc->dtpd_func, (char *)(str + probe->dofp_func),
13247 	    MIN(DTRACE_FUNCNAMELEN - 1, size - probe->dofp_func));
13248 
13249 	if (probe->dofp_name >= strtab->dofs_size) {
13250 		dtrace_dof_error(dof, "corrupt probe name");
13251 		return (NULL);
13252 	}
13253 
13254 	(void) strncpy(desc->dtpd_name, (char *)(str + probe->dofp_name),
13255 	    MIN(DTRACE_NAMELEN - 1, size - probe->dofp_name));
13256 
13257 	return (desc);
13258 }
13259 
13260 static dtrace_difo_t *
13261 dtrace_dof_difo(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
13262     cred_t *cr)
13263 {
13264 	dtrace_difo_t *dp;
13265 	size_t ttl = 0;
13266 	dof_difohdr_t *dofd;
13267 	uintptr_t daddr = (uintptr_t)dof;
13268 	size_t max = dtrace_difo_maxsize;
13269 	int i, l, n;
13270 
13271 	static const struct {
13272 		int section;
13273 		int bufoffs;
13274 		int lenoffs;
13275 		int entsize;
13276 		int align;
13277 		const char *msg;
13278 	} difo[] = {
13279 		{ DOF_SECT_DIF, offsetof(dtrace_difo_t, dtdo_buf),
13280 		offsetof(dtrace_difo_t, dtdo_len), sizeof (dif_instr_t),
13281 		sizeof (dif_instr_t), "multiple DIF sections" },
13282 
13283 		{ DOF_SECT_INTTAB, offsetof(dtrace_difo_t, dtdo_inttab),
13284 		offsetof(dtrace_difo_t, dtdo_intlen), sizeof (uint64_t),
13285 		sizeof (uint64_t), "multiple integer tables" },
13286 
13287 		{ DOF_SECT_STRTAB, offsetof(dtrace_difo_t, dtdo_strtab),
13288 		offsetof(dtrace_difo_t, dtdo_strlen), 0,
13289 		sizeof (char), "multiple string tables" },
13290 
13291 		{ DOF_SECT_VARTAB, offsetof(dtrace_difo_t, dtdo_vartab),
13292 		offsetof(dtrace_difo_t, dtdo_varlen), sizeof (dtrace_difv_t),
13293 		sizeof (uint_t), "multiple variable tables" },
13294 
13295 		{ DOF_SECT_NONE, 0, 0, 0, 0, NULL }
13296 	};
13297 
13298 	if (sec->dofs_type != DOF_SECT_DIFOHDR) {
13299 		dtrace_dof_error(dof, "invalid DIFO header section");
13300 		return (NULL);
13301 	}
13302 
13303 	if (sec->dofs_align != sizeof (dof_secidx_t)) {
13304 		dtrace_dof_error(dof, "bad alignment in DIFO header");
13305 		return (NULL);
13306 	}
13307 
13308 	if (sec->dofs_size < sizeof (dof_difohdr_t) ||
13309 	    sec->dofs_size % sizeof (dof_secidx_t)) {
13310 		dtrace_dof_error(dof, "bad size in DIFO header");
13311 		return (NULL);
13312 	}
13313 
13314 	dofd = (dof_difohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
13315 	n = (sec->dofs_size - sizeof (*dofd)) / sizeof (dof_secidx_t) + 1;
13316 
13317 	dp = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
13318 	dp->dtdo_rtype = dofd->dofd_rtype;
13319 
13320 	for (l = 0; l < n; l++) {
13321 		dof_sec_t *subsec;
13322 		void **bufp;
13323 		uint32_t *lenp;
13324 
13325 		if ((subsec = dtrace_dof_sect(dof, DOF_SECT_NONE,
13326 		    dofd->dofd_links[l])) == NULL)
13327 			goto err; /* invalid section link */
13328 
13329 		if (ttl + subsec->dofs_size > max) {
13330 			dtrace_dof_error(dof, "exceeds maximum size");
13331 			goto err;
13332 		}
13333 
13334 		ttl += subsec->dofs_size;
13335 
13336 		for (i = 0; difo[i].section != DOF_SECT_NONE; i++) {
13337 			if (subsec->dofs_type != difo[i].section)
13338 				continue;
13339 
13340 			if (!(subsec->dofs_flags & DOF_SECF_LOAD)) {
13341 				dtrace_dof_error(dof, "section not loaded");
13342 				goto err;
13343 			}
13344 
13345 			if (subsec->dofs_align != difo[i].align) {
13346 				dtrace_dof_error(dof, "bad alignment");
13347 				goto err;
13348 			}
13349 
13350 			bufp = (void **)((uintptr_t)dp + difo[i].bufoffs);
13351 			lenp = (uint32_t *)((uintptr_t)dp + difo[i].lenoffs);
13352 
13353 			if (*bufp != NULL) {
13354 				dtrace_dof_error(dof, difo[i].msg);
13355 				goto err;
13356 			}
13357 
13358 			if (difo[i].entsize != subsec->dofs_entsize) {
13359 				dtrace_dof_error(dof, "entry size mismatch");
13360 				goto err;
13361 			}
13362 
13363 			if (subsec->dofs_entsize != 0 &&
13364 			    (subsec->dofs_size % subsec->dofs_entsize) != 0) {
13365 				dtrace_dof_error(dof, "corrupt entry size");
13366 				goto err;
13367 			}
13368 
13369 			*lenp = subsec->dofs_size;
13370 			*bufp = kmem_alloc(subsec->dofs_size, KM_SLEEP);
13371 			bcopy((char *)(uintptr_t)(daddr + subsec->dofs_offset),
13372 			    *bufp, subsec->dofs_size);
13373 
13374 			if (subsec->dofs_entsize != 0)
13375 				*lenp /= subsec->dofs_entsize;
13376 
13377 			break;
13378 		}
13379 
13380 		/*
13381 		 * If we encounter a loadable DIFO sub-section that is not
13382 		 * known to us, assume this is a broken program and fail.
13383 		 */
13384 		if (difo[i].section == DOF_SECT_NONE &&
13385 		    (subsec->dofs_flags & DOF_SECF_LOAD)) {
13386 			dtrace_dof_error(dof, "unrecognized DIFO subsection");
13387 			goto err;
13388 		}
13389 	}
13390 
13391 	if (dp->dtdo_buf == NULL) {
13392 		/*
13393 		 * We can't have a DIF object without DIF text.
13394 		 */
13395 		dtrace_dof_error(dof, "missing DIF text");
13396 		goto err;
13397 	}
13398 
13399 	/*
13400 	 * Before we validate the DIF object, run through the variable table
13401 	 * looking for the strings -- if any of their size are under, we'll set
13402 	 * their size to be the system-wide default string size.  Note that
13403 	 * this should _not_ happen if the "strsize" option has been set --
13404 	 * in this case, the compiler should have set the size to reflect the
13405 	 * setting of the option.
13406 	 */
13407 	for (i = 0; i < dp->dtdo_varlen; i++) {
13408 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
13409 		dtrace_diftype_t *t = &v->dtdv_type;
13410 
13411 		if (v->dtdv_id < DIF_VAR_OTHER_UBASE)
13412 			continue;
13413 
13414 		if (t->dtdt_kind == DIF_TYPE_STRING && t->dtdt_size == 0)
13415 			t->dtdt_size = dtrace_strsize_default;
13416 	}
13417 
13418 	if (dtrace_difo_validate(dp, vstate, DIF_DIR_NREGS, cr) != 0)
13419 		goto err;
13420 
13421 	dtrace_difo_init(dp, vstate);
13422 	return (dp);
13423 
13424 err:
13425 	kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
13426 	kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
13427 	kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
13428 	kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
13429 
13430 	kmem_free(dp, sizeof (dtrace_difo_t));
13431 	return (NULL);
13432 }
13433 
13434 static dtrace_predicate_t *
13435 dtrace_dof_predicate(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
13436     cred_t *cr)
13437 {
13438 	dtrace_difo_t *dp;
13439 
13440 	if ((dp = dtrace_dof_difo(dof, sec, vstate, cr)) == NULL)
13441 		return (NULL);
13442 
13443 	return (dtrace_predicate_create(dp));
13444 }
13445 
13446 static dtrace_actdesc_t *
13447 dtrace_dof_actdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
13448     cred_t *cr)
13449 {
13450 	dtrace_actdesc_t *act, *first = NULL, *last = NULL, *next;
13451 	dof_actdesc_t *desc;
13452 	dof_sec_t *difosec;
13453 	size_t offs;
13454 	uintptr_t daddr = (uintptr_t)dof;
13455 	uint64_t arg;
13456 	dtrace_actkind_t kind;
13457 
13458 	if (sec->dofs_type != DOF_SECT_ACTDESC) {
13459 		dtrace_dof_error(dof, "invalid action section");
13460 		return (NULL);
13461 	}
13462 
13463 	if (sec->dofs_offset + sizeof (dof_actdesc_t) > dof->dofh_loadsz) {
13464 		dtrace_dof_error(dof, "truncated action description");
13465 		return (NULL);
13466 	}
13467 
13468 	if (sec->dofs_align != sizeof (uint64_t)) {
13469 		dtrace_dof_error(dof, "bad alignment in action description");
13470 		return (NULL);
13471 	}
13472 
13473 	if (sec->dofs_size < sec->dofs_entsize) {
13474 		dtrace_dof_error(dof, "section entry size exceeds total size");
13475 		return (NULL);
13476 	}
13477 
13478 	if (sec->dofs_entsize != sizeof (dof_actdesc_t)) {
13479 		dtrace_dof_error(dof, "bad entry size in action description");
13480 		return (NULL);
13481 	}
13482 
13483 	if (sec->dofs_size / sec->dofs_entsize > dtrace_actions_max) {
13484 		dtrace_dof_error(dof, "actions exceed dtrace_actions_max");
13485 		return (NULL);
13486 	}
13487 
13488 	for (offs = 0; offs < sec->dofs_size; offs += sec->dofs_entsize) {
13489 		desc = (dof_actdesc_t *)(daddr +
13490 		    (uintptr_t)sec->dofs_offset + offs);
13491 		kind = (dtrace_actkind_t)desc->dofa_kind;
13492 
13493 		if ((DTRACEACT_ISPRINTFLIKE(kind) &&
13494 		    (kind != DTRACEACT_PRINTA ||
13495 		    desc->dofa_strtab != DOF_SECIDX_NONE)) ||
13496 		    (kind == DTRACEACT_DIFEXPR &&
13497 		    desc->dofa_strtab != DOF_SECIDX_NONE)) {
13498 			dof_sec_t *strtab;
13499 			char *str, *fmt;
13500 			uint64_t i;
13501 
13502 			/*
13503 			 * The argument to these actions is an index into the
13504 			 * DOF string table.  For printf()-like actions, this
13505 			 * is the format string.  For print(), this is the
13506 			 * CTF type of the expression result.
13507 			 */
13508 			if ((strtab = dtrace_dof_sect(dof,
13509 			    DOF_SECT_STRTAB, desc->dofa_strtab)) == NULL)
13510 				goto err;
13511 
13512 			str = (char *)((uintptr_t)dof +
13513 			    (uintptr_t)strtab->dofs_offset);
13514 
13515 			for (i = desc->dofa_arg; i < strtab->dofs_size; i++) {
13516 				if (str[i] == '\0')
13517 					break;
13518 			}
13519 
13520 			if (i >= strtab->dofs_size) {
13521 				dtrace_dof_error(dof, "bogus format string");
13522 				goto err;
13523 			}
13524 
13525 			if (i == desc->dofa_arg) {
13526 				dtrace_dof_error(dof, "empty format string");
13527 				goto err;
13528 			}
13529 
13530 			i -= desc->dofa_arg;
13531 			fmt = kmem_alloc(i + 1, KM_SLEEP);
13532 			bcopy(&str[desc->dofa_arg], fmt, i + 1);
13533 			arg = (uint64_t)(uintptr_t)fmt;
13534 		} else {
13535 			if (kind == DTRACEACT_PRINTA) {
13536 				ASSERT(desc->dofa_strtab == DOF_SECIDX_NONE);
13537 				arg = 0;
13538 			} else {
13539 				arg = desc->dofa_arg;
13540 			}
13541 		}
13542 
13543 		act = dtrace_actdesc_create(kind, desc->dofa_ntuple,
13544 		    desc->dofa_uarg, arg);
13545 
13546 		if (last != NULL) {
13547 			last->dtad_next = act;
13548 		} else {
13549 			first = act;
13550 		}
13551 
13552 		last = act;
13553 
13554 		if (desc->dofa_difo == DOF_SECIDX_NONE)
13555 			continue;
13556 
13557 		if ((difosec = dtrace_dof_sect(dof,
13558 		    DOF_SECT_DIFOHDR, desc->dofa_difo)) == NULL)
13559 			goto err;
13560 
13561 		act->dtad_difo = dtrace_dof_difo(dof, difosec, vstate, cr);
13562 
13563 		if (act->dtad_difo == NULL)
13564 			goto err;
13565 	}
13566 
13567 	ASSERT(first != NULL);
13568 	return (first);
13569 
13570 err:
13571 	for (act = first; act != NULL; act = next) {
13572 		next = act->dtad_next;
13573 		dtrace_actdesc_release(act, vstate);
13574 	}
13575 
13576 	return (NULL);
13577 }
13578 
13579 static dtrace_ecbdesc_t *
13580 dtrace_dof_ecbdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
13581     cred_t *cr)
13582 {
13583 	dtrace_ecbdesc_t *ep;
13584 	dof_ecbdesc_t *ecb;
13585 	dtrace_probedesc_t *desc;
13586 	dtrace_predicate_t *pred = NULL;
13587 
13588 	if (sec->dofs_size < sizeof (dof_ecbdesc_t)) {
13589 		dtrace_dof_error(dof, "truncated ECB description");
13590 		return (NULL);
13591 	}
13592 
13593 	if (sec->dofs_align != sizeof (uint64_t)) {
13594 		dtrace_dof_error(dof, "bad alignment in ECB description");
13595 		return (NULL);
13596 	}
13597 
13598 	ecb = (dof_ecbdesc_t *)((uintptr_t)dof + (uintptr_t)sec->dofs_offset);
13599 	sec = dtrace_dof_sect(dof, DOF_SECT_PROBEDESC, ecb->dofe_probes);
13600 
13601 	if (sec == NULL)
13602 		return (NULL);
13603 
13604 	ep = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
13605 	ep->dted_uarg = ecb->dofe_uarg;
13606 	desc = &ep->dted_probe;
13607 
13608 	if (dtrace_dof_probedesc(dof, sec, desc) == NULL)
13609 		goto err;
13610 
13611 	if (ecb->dofe_pred != DOF_SECIDX_NONE) {
13612 		if ((sec = dtrace_dof_sect(dof,
13613 		    DOF_SECT_DIFOHDR, ecb->dofe_pred)) == NULL)
13614 			goto err;
13615 
13616 		if ((pred = dtrace_dof_predicate(dof, sec, vstate, cr)) == NULL)
13617 			goto err;
13618 
13619 		ep->dted_pred.dtpdd_predicate = pred;
13620 	}
13621 
13622 	if (ecb->dofe_actions != DOF_SECIDX_NONE) {
13623 		if ((sec = dtrace_dof_sect(dof,
13624 		    DOF_SECT_ACTDESC, ecb->dofe_actions)) == NULL)
13625 			goto err;
13626 
13627 		ep->dted_action = dtrace_dof_actdesc(dof, sec, vstate, cr);
13628 
13629 		if (ep->dted_action == NULL)
13630 			goto err;
13631 	}
13632 
13633 	return (ep);
13634 
13635 err:
13636 	if (pred != NULL)
13637 		dtrace_predicate_release(pred, vstate);
13638 	kmem_free(ep, sizeof (dtrace_ecbdesc_t));
13639 	return (NULL);
13640 }
13641 
13642 /*
13643  * Apply the relocations from the specified 'sec' (a DOF_SECT_URELHDR) to the
13644  * specified DOF.  At present, this amounts to simply adding 'ubase' to the
13645  * site of any user SETX relocations to account for load object base address.
13646  * In the future, if we need other relocations, this function can be extended.
13647  */
13648 static int
13649 dtrace_dof_relocate(dof_hdr_t *dof, dof_sec_t *sec, uint64_t ubase)
13650 {
13651 	uintptr_t daddr = (uintptr_t)dof;
13652 	dof_relohdr_t *dofr =
13653 	    (dof_relohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
13654 	dof_sec_t *ss, *rs, *ts;
13655 	dof_relodesc_t *r;
13656 	uint_t i, n;
13657 
13658 	if (sec->dofs_size < sizeof (dof_relohdr_t) ||
13659 	    sec->dofs_align != sizeof (dof_secidx_t)) {
13660 		dtrace_dof_error(dof, "invalid relocation header");
13661 		return (-1);
13662 	}
13663 
13664 	ss = dtrace_dof_sect(dof, DOF_SECT_STRTAB, dofr->dofr_strtab);
13665 	rs = dtrace_dof_sect(dof, DOF_SECT_RELTAB, dofr->dofr_relsec);
13666 	ts = dtrace_dof_sect(dof, DOF_SECT_NONE, dofr->dofr_tgtsec);
13667 
13668 	if (ss == NULL || rs == NULL || ts == NULL)
13669 		return (-1); /* dtrace_dof_error() has been called already */
13670 
13671 	if (rs->dofs_entsize < sizeof (dof_relodesc_t) ||
13672 	    rs->dofs_align != sizeof (uint64_t)) {
13673 		dtrace_dof_error(dof, "invalid relocation section");
13674 		return (-1);
13675 	}
13676 
13677 	r = (dof_relodesc_t *)(uintptr_t)(daddr + rs->dofs_offset);
13678 	n = rs->dofs_size / rs->dofs_entsize;
13679 
13680 	for (i = 0; i < n; i++) {
13681 		uintptr_t taddr = daddr + ts->dofs_offset + r->dofr_offset;
13682 
13683 		switch (r->dofr_type) {
13684 		case DOF_RELO_NONE:
13685 			break;
13686 		case DOF_RELO_SETX:
13687 			if (r->dofr_offset >= ts->dofs_size || r->dofr_offset +
13688 			    sizeof (uint64_t) > ts->dofs_size) {
13689 				dtrace_dof_error(dof, "bad relocation offset");
13690 				return (-1);
13691 			}
13692 
13693 			if (!IS_P2ALIGNED(taddr, sizeof (uint64_t))) {
13694 				dtrace_dof_error(dof, "misaligned setx relo");
13695 				return (-1);
13696 			}
13697 
13698 			*(uint64_t *)taddr += ubase;
13699 			break;
13700 		default:
13701 			dtrace_dof_error(dof, "invalid relocation type");
13702 			return (-1);
13703 		}
13704 
13705 		r = (dof_relodesc_t *)((uintptr_t)r + rs->dofs_entsize);
13706 	}
13707 
13708 	return (0);
13709 }
13710 
13711 /*
13712  * The dof_hdr_t passed to dtrace_dof_slurp() should be a partially validated
13713  * header:  it should be at the front of a memory region that is at least
13714  * sizeof (dof_hdr_t) in size -- and then at least dof_hdr.dofh_loadsz in
13715  * size.  It need not be validated in any other way.
13716  */
13717 static int
13718 dtrace_dof_slurp(dof_hdr_t *dof, dtrace_vstate_t *vstate, cred_t *cr,
13719     dtrace_enabling_t **enabp, uint64_t ubase, int noprobes)
13720 {
13721 	uint64_t len = dof->dofh_loadsz, seclen;
13722 	uintptr_t daddr = (uintptr_t)dof;
13723 	dtrace_ecbdesc_t *ep;
13724 	dtrace_enabling_t *enab;
13725 	uint_t i;
13726 
13727 	ASSERT(MUTEX_HELD(&dtrace_lock));
13728 	ASSERT(dof->dofh_loadsz >= sizeof (dof_hdr_t));
13729 
13730 	/*
13731 	 * Check the DOF header identification bytes.  In addition to checking
13732 	 * valid settings, we also verify that unused bits/bytes are zeroed so
13733 	 * we can use them later without fear of regressing existing binaries.
13734 	 */
13735 	if (bcmp(&dof->dofh_ident[DOF_ID_MAG0],
13736 	    DOF_MAG_STRING, DOF_MAG_STRLEN) != 0) {
13737 		dtrace_dof_error(dof, "DOF magic string mismatch");
13738 		return (-1);
13739 	}
13740 
13741 	if (dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_ILP32 &&
13742 	    dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_LP64) {
13743 		dtrace_dof_error(dof, "DOF has invalid data model");
13744 		return (-1);
13745 	}
13746 
13747 	if (dof->dofh_ident[DOF_ID_ENCODING] != DOF_ENCODE_NATIVE) {
13748 		dtrace_dof_error(dof, "DOF encoding mismatch");
13749 		return (-1);
13750 	}
13751 
13752 	if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
13753 	    dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_2) {
13754 		dtrace_dof_error(dof, "DOF version mismatch");
13755 		return (-1);
13756 	}
13757 
13758 	if (dof->dofh_ident[DOF_ID_DIFVERS] != DIF_VERSION_2) {
13759 		dtrace_dof_error(dof, "DOF uses unsupported instruction set");
13760 		return (-1);
13761 	}
13762 
13763 	if (dof->dofh_ident[DOF_ID_DIFIREG] > DIF_DIR_NREGS) {
13764 		dtrace_dof_error(dof, "DOF uses too many integer registers");
13765 		return (-1);
13766 	}
13767 
13768 	if (dof->dofh_ident[DOF_ID_DIFTREG] > DIF_DTR_NREGS) {
13769 		dtrace_dof_error(dof, "DOF uses too many tuple registers");
13770 		return (-1);
13771 	}
13772 
13773 	for (i = DOF_ID_PAD; i < DOF_ID_SIZE; i++) {
13774 		if (dof->dofh_ident[i] != 0) {
13775 			dtrace_dof_error(dof, "DOF has invalid ident byte set");
13776 			return (-1);
13777 		}
13778 	}
13779 
13780 	if (dof->dofh_flags & ~DOF_FL_VALID) {
13781 		dtrace_dof_error(dof, "DOF has invalid flag bits set");
13782 		return (-1);
13783 	}
13784 
13785 	if (dof->dofh_secsize == 0) {
13786 		dtrace_dof_error(dof, "zero section header size");
13787 		return (-1);
13788 	}
13789 
13790 	/*
13791 	 * Check that the section headers don't exceed the amount of DOF
13792 	 * data.  Note that we cast the section size and number of sections
13793 	 * to uint64_t's to prevent possible overflow in the multiplication.
13794 	 */
13795 	seclen = (uint64_t)dof->dofh_secnum * (uint64_t)dof->dofh_secsize;
13796 
13797 	if (dof->dofh_secoff > len || seclen > len ||
13798 	    dof->dofh_secoff + seclen > len) {
13799 		dtrace_dof_error(dof, "truncated section headers");
13800 		return (-1);
13801 	}
13802 
13803 	if (!IS_P2ALIGNED(dof->dofh_secoff, sizeof (uint64_t))) {
13804 		dtrace_dof_error(dof, "misaligned section headers");
13805 		return (-1);
13806 	}
13807 
13808 	if (!IS_P2ALIGNED(dof->dofh_secsize, sizeof (uint64_t))) {
13809 		dtrace_dof_error(dof, "misaligned section size");
13810 		return (-1);
13811 	}
13812 
13813 	/*
13814 	 * Take an initial pass through the section headers to be sure that
13815 	 * the headers don't have stray offsets.  If the 'noprobes' flag is
13816 	 * set, do not permit sections relating to providers, probes, or args.
13817 	 */
13818 	for (i = 0; i < dof->dofh_secnum; i++) {
13819 		dof_sec_t *sec = (dof_sec_t *)(daddr +
13820 		    (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
13821 
13822 		if (noprobes) {
13823 			switch (sec->dofs_type) {
13824 			case DOF_SECT_PROVIDER:
13825 			case DOF_SECT_PROBES:
13826 			case DOF_SECT_PRARGS:
13827 			case DOF_SECT_PROFFS:
13828 				dtrace_dof_error(dof, "illegal sections "
13829 				    "for enabling");
13830 				return (-1);
13831 			}
13832 		}
13833 
13834 		if (DOF_SEC_ISLOADABLE(sec->dofs_type) &&
13835 		    !(sec->dofs_flags & DOF_SECF_LOAD)) {
13836 			dtrace_dof_error(dof, "loadable section with load "
13837 			    "flag unset");
13838 			return (-1);
13839 		}
13840 
13841 		if (!(sec->dofs_flags & DOF_SECF_LOAD))
13842 			continue; /* just ignore non-loadable sections */
13843 
13844 		if (!ISP2(sec->dofs_align)) {
13845 			dtrace_dof_error(dof, "bad section alignment");
13846 			return (-1);
13847 		}
13848 
13849 		if (sec->dofs_offset & (sec->dofs_align - 1)) {
13850 			dtrace_dof_error(dof, "misaligned section");
13851 			return (-1);
13852 		}
13853 
13854 		if (sec->dofs_offset > len || sec->dofs_size > len ||
13855 		    sec->dofs_offset + sec->dofs_size > len) {
13856 			dtrace_dof_error(dof, "corrupt section header");
13857 			return (-1);
13858 		}
13859 
13860 		if (sec->dofs_type == DOF_SECT_STRTAB && *((char *)daddr +
13861 		    sec->dofs_offset + sec->dofs_size - 1) != '\0') {
13862 			dtrace_dof_error(dof, "non-terminating string table");
13863 			return (-1);
13864 		}
13865 	}
13866 
13867 	/*
13868 	 * Take a second pass through the sections and locate and perform any
13869 	 * relocations that are present.  We do this after the first pass to
13870 	 * be sure that all sections have had their headers validated.
13871 	 */
13872 	for (i = 0; i < dof->dofh_secnum; i++) {
13873 		dof_sec_t *sec = (dof_sec_t *)(daddr +
13874 		    (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
13875 
13876 		if (!(sec->dofs_flags & DOF_SECF_LOAD))
13877 			continue; /* skip sections that are not loadable */
13878 
13879 		switch (sec->dofs_type) {
13880 		case DOF_SECT_URELHDR:
13881 			if (dtrace_dof_relocate(dof, sec, ubase) != 0)
13882 				return (-1);
13883 			break;
13884 		}
13885 	}
13886 
13887 	if ((enab = *enabp) == NULL)
13888 		enab = *enabp = dtrace_enabling_create(vstate);
13889 
13890 	for (i = 0; i < dof->dofh_secnum; i++) {
13891 		dof_sec_t *sec = (dof_sec_t *)(daddr +
13892 		    (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
13893 
13894 		if (sec->dofs_type != DOF_SECT_ECBDESC)
13895 			continue;
13896 
13897 		if ((ep = dtrace_dof_ecbdesc(dof, sec, vstate, cr)) == NULL) {
13898 			dtrace_enabling_destroy(enab);
13899 			*enabp = NULL;
13900 			return (-1);
13901 		}
13902 
13903 		dtrace_enabling_add(enab, ep);
13904 	}
13905 
13906 	return (0);
13907 }
13908 
13909 /*
13910  * Process DOF for any options.  This routine assumes that the DOF has been
13911  * at least processed by dtrace_dof_slurp().
13912  */
13913 static int
13914 dtrace_dof_options(dof_hdr_t *dof, dtrace_state_t *state)
13915 {
13916 	int i, rval;
13917 	uint32_t entsize;
13918 	size_t offs;
13919 	dof_optdesc_t *desc;
13920 
13921 	for (i = 0; i < dof->dofh_secnum; i++) {
13922 		dof_sec_t *sec = (dof_sec_t *)((uintptr_t)dof +
13923 		    (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
13924 
13925 		if (sec->dofs_type != DOF_SECT_OPTDESC)
13926 			continue;
13927 
13928 		if (sec->dofs_align != sizeof (uint64_t)) {
13929 			dtrace_dof_error(dof, "bad alignment in "
13930 			    "option description");
13931 			return (EINVAL);
13932 		}
13933 
13934 		if ((entsize = sec->dofs_entsize) == 0) {
13935 			dtrace_dof_error(dof, "zeroed option entry size");
13936 			return (EINVAL);
13937 		}
13938 
13939 		if (entsize < sizeof (dof_optdesc_t)) {
13940 			dtrace_dof_error(dof, "bad option entry size");
13941 			return (EINVAL);
13942 		}
13943 
13944 		for (offs = 0; offs < sec->dofs_size; offs += entsize) {
13945 			desc = (dof_optdesc_t *)((uintptr_t)dof +
13946 			    (uintptr_t)sec->dofs_offset + offs);
13947 
13948 			if (desc->dofo_strtab != DOF_SECIDX_NONE) {
13949 				dtrace_dof_error(dof, "non-zero option string");
13950 				return (EINVAL);
13951 			}
13952 
13953 			if (desc->dofo_value == DTRACEOPT_UNSET) {
13954 				dtrace_dof_error(dof, "unset option");
13955 				return (EINVAL);
13956 			}
13957 
13958 			if ((rval = dtrace_state_option(state,
13959 			    desc->dofo_option, desc->dofo_value)) != 0) {
13960 				dtrace_dof_error(dof, "rejected option");
13961 				return (rval);
13962 			}
13963 		}
13964 	}
13965 
13966 	return (0);
13967 }
13968 
13969 /*
13970  * DTrace Consumer State Functions
13971  */
13972 static int
13973 dtrace_dstate_init(dtrace_dstate_t *dstate, size_t size)
13974 {
13975 	size_t hashsize, maxper, min, chunksize = dstate->dtds_chunksize;
13976 	void *base;
13977 	uintptr_t limit;
13978 	dtrace_dynvar_t *dvar, *next, *start;
13979 	int i;
13980 
13981 	ASSERT(MUTEX_HELD(&dtrace_lock));
13982 	ASSERT(dstate->dtds_base == NULL && dstate->dtds_percpu == NULL);
13983 
13984 	bzero(dstate, sizeof (dtrace_dstate_t));
13985 
13986 	if ((dstate->dtds_chunksize = chunksize) == 0)
13987 		dstate->dtds_chunksize = DTRACE_DYNVAR_CHUNKSIZE;
13988 
13989 	VERIFY(dstate->dtds_chunksize < LONG_MAX);
13990 
13991 	if (size < (min = dstate->dtds_chunksize + sizeof (dtrace_dynhash_t)))
13992 		size = min;
13993 
13994 	if ((base = kmem_zalloc(size, KM_NOSLEEP | KM_NORMALPRI)) == NULL)
13995 		return (ENOMEM);
13996 
13997 	dstate->dtds_size = size;
13998 	dstate->dtds_base = base;
13999 	dstate->dtds_percpu = kmem_cache_alloc(dtrace_state_cache, KM_SLEEP);
14000 	bzero(dstate->dtds_percpu, NCPU * sizeof (dtrace_dstate_percpu_t));
14001 
14002 	hashsize = size / (dstate->dtds_chunksize + sizeof (dtrace_dynhash_t));
14003 
14004 	if (hashsize != 1 && (hashsize & 1))
14005 		hashsize--;
14006 
14007 	dstate->dtds_hashsize = hashsize;
14008 	dstate->dtds_hash = dstate->dtds_base;
14009 
14010 	/*
14011 	 * Set all of our hash buckets to point to the single sink, and (if
14012 	 * it hasn't already been set), set the sink's hash value to be the
14013 	 * sink sentinel value.  The sink is needed for dynamic variable
14014 	 * lookups to know that they have iterated over an entire, valid hash
14015 	 * chain.
14016 	 */
14017 	for (i = 0; i < hashsize; i++)
14018 		dstate->dtds_hash[i].dtdh_chain = &dtrace_dynhash_sink;
14019 
14020 	if (dtrace_dynhash_sink.dtdv_hashval != DTRACE_DYNHASH_SINK)
14021 		dtrace_dynhash_sink.dtdv_hashval = DTRACE_DYNHASH_SINK;
14022 
14023 	/*
14024 	 * Determine number of active CPUs.  Divide free list evenly among
14025 	 * active CPUs.
14026 	 */
14027 	start = (dtrace_dynvar_t *)
14028 	    ((uintptr_t)base + hashsize * sizeof (dtrace_dynhash_t));
14029 	limit = (uintptr_t)base + size;
14030 
14031 	VERIFY((uintptr_t)start < limit);
14032 	VERIFY((uintptr_t)start >= (uintptr_t)base);
14033 
14034 	maxper = (limit - (uintptr_t)start) / NCPU;
14035 	maxper = (maxper / dstate->dtds_chunksize) * dstate->dtds_chunksize;
14036 
14037 #ifndef illumos
14038 	CPU_FOREACH(i) {
14039 #else
14040 	for (i = 0; i < NCPU; i++) {
14041 #endif
14042 		dstate->dtds_percpu[i].dtdsc_free = dvar = start;
14043 
14044 		/*
14045 		 * If we don't even have enough chunks to make it once through
14046 		 * NCPUs, we're just going to allocate everything to the first
14047 		 * CPU.  And if we're on the last CPU, we're going to allocate
14048 		 * whatever is left over.  In either case, we set the limit to
14049 		 * be the limit of the dynamic variable space.
14050 		 */
14051 		if (maxper == 0 || i == NCPU - 1) {
14052 			limit = (uintptr_t)base + size;
14053 			start = NULL;
14054 		} else {
14055 			limit = (uintptr_t)start + maxper;
14056 			start = (dtrace_dynvar_t *)limit;
14057 		}
14058 
14059 		VERIFY(limit <= (uintptr_t)base + size);
14060 
14061 		for (;;) {
14062 			next = (dtrace_dynvar_t *)((uintptr_t)dvar +
14063 			    dstate->dtds_chunksize);
14064 
14065 			if ((uintptr_t)next + dstate->dtds_chunksize >= limit)
14066 				break;
14067 
14068 			VERIFY((uintptr_t)dvar >= (uintptr_t)base &&
14069 			    (uintptr_t)dvar <= (uintptr_t)base + size);
14070 			dvar->dtdv_next = next;
14071 			dvar = next;
14072 		}
14073 
14074 		if (maxper == 0)
14075 			break;
14076 	}
14077 
14078 	return (0);
14079 }
14080 
14081 static void
14082 dtrace_dstate_fini(dtrace_dstate_t *dstate)
14083 {
14084 	ASSERT(MUTEX_HELD(&cpu_lock));
14085 
14086 	if (dstate->dtds_base == NULL)
14087 		return;
14088 
14089 	kmem_free(dstate->dtds_base, dstate->dtds_size);
14090 	kmem_cache_free(dtrace_state_cache, dstate->dtds_percpu);
14091 }
14092 
14093 static void
14094 dtrace_vstate_fini(dtrace_vstate_t *vstate)
14095 {
14096 	/*
14097 	 * Logical XOR, where are you?
14098 	 */
14099 	ASSERT((vstate->dtvs_nglobals == 0) ^ (vstate->dtvs_globals != NULL));
14100 
14101 	if (vstate->dtvs_nglobals > 0) {
14102 		kmem_free(vstate->dtvs_globals, vstate->dtvs_nglobals *
14103 		    sizeof (dtrace_statvar_t *));
14104 	}
14105 
14106 	if (vstate->dtvs_ntlocals > 0) {
14107 		kmem_free(vstate->dtvs_tlocals, vstate->dtvs_ntlocals *
14108 		    sizeof (dtrace_difv_t));
14109 	}
14110 
14111 	ASSERT((vstate->dtvs_nlocals == 0) ^ (vstate->dtvs_locals != NULL));
14112 
14113 	if (vstate->dtvs_nlocals > 0) {
14114 		kmem_free(vstate->dtvs_locals, vstate->dtvs_nlocals *
14115 		    sizeof (dtrace_statvar_t *));
14116 	}
14117 }
14118 
14119 #ifdef illumos
14120 static void
14121 dtrace_state_clean(dtrace_state_t *state)
14122 {
14123 	if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE)
14124 		return;
14125 
14126 	dtrace_dynvar_clean(&state->dts_vstate.dtvs_dynvars);
14127 	dtrace_speculation_clean(state);
14128 }
14129 
14130 static void
14131 dtrace_state_deadman(dtrace_state_t *state)
14132 {
14133 	hrtime_t now;
14134 
14135 	dtrace_sync();
14136 
14137 	now = dtrace_gethrtime();
14138 
14139 	if (state != dtrace_anon.dta_state &&
14140 	    now - state->dts_laststatus >= dtrace_deadman_user)
14141 		return;
14142 
14143 	/*
14144 	 * We must be sure that dts_alive never appears to be less than the
14145 	 * value upon entry to dtrace_state_deadman(), and because we lack a
14146 	 * dtrace_cas64(), we cannot store to it atomically.  We thus instead
14147 	 * store INT64_MAX to it, followed by a memory barrier, followed by
14148 	 * the new value.  This assures that dts_alive never appears to be
14149 	 * less than its true value, regardless of the order in which the
14150 	 * stores to the underlying storage are issued.
14151 	 */
14152 	state->dts_alive = INT64_MAX;
14153 	dtrace_membar_producer();
14154 	state->dts_alive = now;
14155 }
14156 #else	/* !illumos */
14157 static void
14158 dtrace_state_clean(void *arg)
14159 {
14160 	dtrace_state_t *state = arg;
14161 	dtrace_optval_t *opt = state->dts_options;
14162 
14163 	if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE)
14164 		return;
14165 
14166 	dtrace_dynvar_clean(&state->dts_vstate.dtvs_dynvars);
14167 	dtrace_speculation_clean(state);
14168 
14169 	callout_reset(&state->dts_cleaner, hz * opt[DTRACEOPT_CLEANRATE] / NANOSEC,
14170 	    dtrace_state_clean, state);
14171 }
14172 
14173 static void
14174 dtrace_state_deadman(void *arg)
14175 {
14176 	dtrace_state_t *state = arg;
14177 	hrtime_t now;
14178 
14179 	dtrace_sync();
14180 
14181 	dtrace_debug_output();
14182 
14183 	now = dtrace_gethrtime();
14184 
14185 	if (state != dtrace_anon.dta_state &&
14186 	    now - state->dts_laststatus >= dtrace_deadman_user)
14187 		return;
14188 
14189 	/*
14190 	 * We must be sure that dts_alive never appears to be less than the
14191 	 * value upon entry to dtrace_state_deadman(), and because we lack a
14192 	 * dtrace_cas64(), we cannot store to it atomically.  We thus instead
14193 	 * store INT64_MAX to it, followed by a memory barrier, followed by
14194 	 * the new value.  This assures that dts_alive never appears to be
14195 	 * less than its true value, regardless of the order in which the
14196 	 * stores to the underlying storage are issued.
14197 	 */
14198 	state->dts_alive = INT64_MAX;
14199 	dtrace_membar_producer();
14200 	state->dts_alive = now;
14201 
14202 	callout_reset(&state->dts_deadman, hz * dtrace_deadman_interval / NANOSEC,
14203 	    dtrace_state_deadman, state);
14204 }
14205 #endif	/* illumos */
14206 
14207 static dtrace_state_t *
14208 #ifdef illumos
14209 dtrace_state_create(dev_t *devp, cred_t *cr)
14210 #else
14211 dtrace_state_create(struct cdev *dev)
14212 #endif
14213 {
14214 #ifdef illumos
14215 	minor_t minor;
14216 	major_t major;
14217 #else
14218 	cred_t *cr = NULL;
14219 	int m = 0;
14220 #endif
14221 	char c[30];
14222 	dtrace_state_t *state;
14223 	dtrace_optval_t *opt;
14224 	int bufsize = NCPU * sizeof (dtrace_buffer_t), i;
14225 
14226 	ASSERT(MUTEX_HELD(&dtrace_lock));
14227 	ASSERT(MUTEX_HELD(&cpu_lock));
14228 
14229 #ifdef illumos
14230 	minor = (minor_t)(uintptr_t)vmem_alloc(dtrace_minor, 1,
14231 	    VM_BESTFIT | VM_SLEEP);
14232 
14233 	if (ddi_soft_state_zalloc(dtrace_softstate, minor) != DDI_SUCCESS) {
14234 		vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
14235 		return (NULL);
14236 	}
14237 
14238 	state = ddi_get_soft_state(dtrace_softstate, minor);
14239 #else
14240 	if (dev != NULL) {
14241 		cr = dev->si_cred;
14242 		m = dev2unit(dev);
14243 	}
14244 
14245 	/* Allocate memory for the state. */
14246 	state = kmem_zalloc(sizeof(dtrace_state_t), KM_SLEEP);
14247 #endif
14248 
14249 	state->dts_epid = DTRACE_EPIDNONE + 1;
14250 
14251 	(void) snprintf(c, sizeof (c), "dtrace_aggid_%d", m);
14252 #ifdef illumos
14253 	state->dts_aggid_arena = vmem_create(c, (void *)1, UINT32_MAX, 1,
14254 	    NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
14255 
14256 	if (devp != NULL) {
14257 		major = getemajor(*devp);
14258 	} else {
14259 		major = ddi_driver_major(dtrace_devi);
14260 	}
14261 
14262 	state->dts_dev = makedevice(major, minor);
14263 
14264 	if (devp != NULL)
14265 		*devp = state->dts_dev;
14266 #else
14267 	state->dts_aggid_arena = new_unrhdr(1, INT_MAX, &dtrace_unr_mtx);
14268 	state->dts_dev = dev;
14269 #endif
14270 
14271 	/*
14272 	 * We allocate NCPU buffers.  On the one hand, this can be quite
14273 	 * a bit of memory per instance (nearly 36K on a Starcat).  On the
14274 	 * other hand, it saves an additional memory reference in the probe
14275 	 * path.
14276 	 */
14277 	state->dts_buffer = kmem_zalloc(bufsize, KM_SLEEP);
14278 	state->dts_aggbuffer = kmem_zalloc(bufsize, KM_SLEEP);
14279 
14280 #ifdef illumos
14281 	state->dts_cleaner = CYCLIC_NONE;
14282 	state->dts_deadman = CYCLIC_NONE;
14283 #else
14284 	callout_init(&state->dts_cleaner, 1);
14285 	callout_init(&state->dts_deadman, 1);
14286 #endif
14287 	state->dts_vstate.dtvs_state = state;
14288 
14289 	for (i = 0; i < DTRACEOPT_MAX; i++)
14290 		state->dts_options[i] = DTRACEOPT_UNSET;
14291 
14292 	/*
14293 	 * Set the default options.
14294 	 */
14295 	opt = state->dts_options;
14296 	opt[DTRACEOPT_BUFPOLICY] = DTRACEOPT_BUFPOLICY_SWITCH;
14297 	opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_AUTO;
14298 	opt[DTRACEOPT_NSPEC] = dtrace_nspec_default;
14299 	opt[DTRACEOPT_SPECSIZE] = dtrace_specsize_default;
14300 	opt[DTRACEOPT_CPU] = (dtrace_optval_t)DTRACE_CPUALL;
14301 	opt[DTRACEOPT_STRSIZE] = dtrace_strsize_default;
14302 	opt[DTRACEOPT_STACKFRAMES] = dtrace_stackframes_default;
14303 	opt[DTRACEOPT_USTACKFRAMES] = dtrace_ustackframes_default;
14304 	opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_default;
14305 	opt[DTRACEOPT_AGGRATE] = dtrace_aggrate_default;
14306 	opt[DTRACEOPT_SWITCHRATE] = dtrace_switchrate_default;
14307 	opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_default;
14308 	opt[DTRACEOPT_JSTACKFRAMES] = dtrace_jstackframes_default;
14309 	opt[DTRACEOPT_JSTACKSTRSIZE] = dtrace_jstackstrsize_default;
14310 
14311 	state->dts_activity = DTRACE_ACTIVITY_INACTIVE;
14312 
14313 	/*
14314 	 * Depending on the user credentials, we set flag bits which alter probe
14315 	 * visibility or the amount of destructiveness allowed.  In the case of
14316 	 * actual anonymous tracing, or the possession of all privileges, all of
14317 	 * the normal checks are bypassed.
14318 	 */
14319 	if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
14320 		state->dts_cred.dcr_visible = DTRACE_CRV_ALL;
14321 		state->dts_cred.dcr_action = DTRACE_CRA_ALL;
14322 	} else {
14323 		/*
14324 		 * Set up the credentials for this instantiation.  We take a
14325 		 * hold on the credential to prevent it from disappearing on
14326 		 * us; this in turn prevents the zone_t referenced by this
14327 		 * credential from disappearing.  This means that we can
14328 		 * examine the credential and the zone from probe context.
14329 		 */
14330 		crhold(cr);
14331 		state->dts_cred.dcr_cred = cr;
14332 
14333 		/*
14334 		 * CRA_PROC means "we have *some* privilege for dtrace" and
14335 		 * unlocks the use of variables like pid, zonename, etc.
14336 		 */
14337 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE) ||
14338 		    PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
14339 			state->dts_cred.dcr_action |= DTRACE_CRA_PROC;
14340 		}
14341 
14342 		/*
14343 		 * dtrace_user allows use of syscall and profile providers.
14344 		 * If the user also has proc_owner and/or proc_zone, we
14345 		 * extend the scope to include additional visibility and
14346 		 * destructive power.
14347 		 */
14348 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE)) {
14349 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE)) {
14350 				state->dts_cred.dcr_visible |=
14351 				    DTRACE_CRV_ALLPROC;
14352 
14353 				state->dts_cred.dcr_action |=
14354 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
14355 			}
14356 
14357 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE)) {
14358 				state->dts_cred.dcr_visible |=
14359 				    DTRACE_CRV_ALLZONE;
14360 
14361 				state->dts_cred.dcr_action |=
14362 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
14363 			}
14364 
14365 			/*
14366 			 * If we have all privs in whatever zone this is,
14367 			 * we can do destructive things to processes which
14368 			 * have altered credentials.
14369 			 */
14370 #ifdef illumos
14371 			if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
14372 			    cr->cr_zone->zone_privset)) {
14373 				state->dts_cred.dcr_action |=
14374 				    DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
14375 			}
14376 #endif
14377 		}
14378 
14379 		/*
14380 		 * Holding the dtrace_kernel privilege also implies that
14381 		 * the user has the dtrace_user privilege from a visibility
14382 		 * perspective.  But without further privileges, some
14383 		 * destructive actions are not available.
14384 		 */
14385 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE)) {
14386 			/*
14387 			 * Make all probes in all zones visible.  However,
14388 			 * this doesn't mean that all actions become available
14389 			 * to all zones.
14390 			 */
14391 			state->dts_cred.dcr_visible |= DTRACE_CRV_KERNEL |
14392 			    DTRACE_CRV_ALLPROC | DTRACE_CRV_ALLZONE;
14393 
14394 			state->dts_cred.dcr_action |= DTRACE_CRA_KERNEL |
14395 			    DTRACE_CRA_PROC;
14396 			/*
14397 			 * Holding proc_owner means that destructive actions
14398 			 * for *this* zone are allowed.
14399 			 */
14400 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
14401 				state->dts_cred.dcr_action |=
14402 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
14403 
14404 			/*
14405 			 * Holding proc_zone means that destructive actions
14406 			 * for this user/group ID in all zones is allowed.
14407 			 */
14408 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
14409 				state->dts_cred.dcr_action |=
14410 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
14411 
14412 #ifdef illumos
14413 			/*
14414 			 * If we have all privs in whatever zone this is,
14415 			 * we can do destructive things to processes which
14416 			 * have altered credentials.
14417 			 */
14418 			if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
14419 			    cr->cr_zone->zone_privset)) {
14420 				state->dts_cred.dcr_action |=
14421 				    DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
14422 			}
14423 #endif
14424 		}
14425 
14426 		/*
14427 		 * Holding the dtrace_proc privilege gives control over fasttrap
14428 		 * and pid providers.  We need to grant wider destructive
14429 		 * privileges in the event that the user has proc_owner and/or
14430 		 * proc_zone.
14431 		 */
14432 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
14433 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
14434 				state->dts_cred.dcr_action |=
14435 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
14436 
14437 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
14438 				state->dts_cred.dcr_action |=
14439 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
14440 		}
14441 	}
14442 
14443 	return (state);
14444 }
14445 
14446 static int
14447 dtrace_state_buffer(dtrace_state_t *state, dtrace_buffer_t *buf, int which)
14448 {
14449 	dtrace_optval_t *opt = state->dts_options, size;
14450 	processorid_t cpu = 0;;
14451 	int flags = 0, rval, factor, divisor = 1;
14452 
14453 	ASSERT(MUTEX_HELD(&dtrace_lock));
14454 	ASSERT(MUTEX_HELD(&cpu_lock));
14455 	ASSERT(which < DTRACEOPT_MAX);
14456 	ASSERT(state->dts_activity == DTRACE_ACTIVITY_INACTIVE ||
14457 	    (state == dtrace_anon.dta_state &&
14458 	    state->dts_activity == DTRACE_ACTIVITY_ACTIVE));
14459 
14460 	if (opt[which] == DTRACEOPT_UNSET || opt[which] == 0)
14461 		return (0);
14462 
14463 	if (opt[DTRACEOPT_CPU] != DTRACEOPT_UNSET)
14464 		cpu = opt[DTRACEOPT_CPU];
14465 
14466 	if (which == DTRACEOPT_SPECSIZE)
14467 		flags |= DTRACEBUF_NOSWITCH;
14468 
14469 	if (which == DTRACEOPT_BUFSIZE) {
14470 		if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_RING)
14471 			flags |= DTRACEBUF_RING;
14472 
14473 		if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_FILL)
14474 			flags |= DTRACEBUF_FILL;
14475 
14476 		if (state != dtrace_anon.dta_state ||
14477 		    state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
14478 			flags |= DTRACEBUF_INACTIVE;
14479 	}
14480 
14481 	for (size = opt[which]; size >= sizeof (uint64_t); size /= divisor) {
14482 		/*
14483 		 * The size must be 8-byte aligned.  If the size is not 8-byte
14484 		 * aligned, drop it down by the difference.
14485 		 */
14486 		if (size & (sizeof (uint64_t) - 1))
14487 			size -= size & (sizeof (uint64_t) - 1);
14488 
14489 		if (size < state->dts_reserve) {
14490 			/*
14491 			 * Buffers always must be large enough to accommodate
14492 			 * their prereserved space.  We return E2BIG instead
14493 			 * of ENOMEM in this case to allow for user-level
14494 			 * software to differentiate the cases.
14495 			 */
14496 			return (E2BIG);
14497 		}
14498 
14499 		rval = dtrace_buffer_alloc(buf, size, flags, cpu, &factor);
14500 
14501 		if (rval != ENOMEM) {
14502 			opt[which] = size;
14503 			return (rval);
14504 		}
14505 
14506 		if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
14507 			return (rval);
14508 
14509 		for (divisor = 2; divisor < factor; divisor <<= 1)
14510 			continue;
14511 	}
14512 
14513 	return (ENOMEM);
14514 }
14515 
14516 static int
14517 dtrace_state_buffers(dtrace_state_t *state)
14518 {
14519 	dtrace_speculation_t *spec = state->dts_speculations;
14520 	int rval, i;
14521 
14522 	if ((rval = dtrace_state_buffer(state, state->dts_buffer,
14523 	    DTRACEOPT_BUFSIZE)) != 0)
14524 		return (rval);
14525 
14526 	if ((rval = dtrace_state_buffer(state, state->dts_aggbuffer,
14527 	    DTRACEOPT_AGGSIZE)) != 0)
14528 		return (rval);
14529 
14530 	for (i = 0; i < state->dts_nspeculations; i++) {
14531 		if ((rval = dtrace_state_buffer(state,
14532 		    spec[i].dtsp_buffer, DTRACEOPT_SPECSIZE)) != 0)
14533 			return (rval);
14534 	}
14535 
14536 	return (0);
14537 }
14538 
14539 static void
14540 dtrace_state_prereserve(dtrace_state_t *state)
14541 {
14542 	dtrace_ecb_t *ecb;
14543 	dtrace_probe_t *probe;
14544 
14545 	state->dts_reserve = 0;
14546 
14547 	if (state->dts_options[DTRACEOPT_BUFPOLICY] != DTRACEOPT_BUFPOLICY_FILL)
14548 		return;
14549 
14550 	/*
14551 	 * If our buffer policy is a "fill" buffer policy, we need to set the
14552 	 * prereserved space to be the space required by the END probes.
14553 	 */
14554 	probe = dtrace_probes[dtrace_probeid_end - 1];
14555 	ASSERT(probe != NULL);
14556 
14557 	for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
14558 		if (ecb->dte_state != state)
14559 			continue;
14560 
14561 		state->dts_reserve += ecb->dte_needed + ecb->dte_alignment;
14562 	}
14563 }
14564 
14565 static int
14566 dtrace_state_go(dtrace_state_t *state, processorid_t *cpu)
14567 {
14568 	dtrace_optval_t *opt = state->dts_options, sz, nspec;
14569 	dtrace_speculation_t *spec;
14570 	dtrace_buffer_t *buf;
14571 #ifdef illumos
14572 	cyc_handler_t hdlr;
14573 	cyc_time_t when;
14574 #endif
14575 	int rval = 0, i, bufsize = NCPU * sizeof (dtrace_buffer_t);
14576 	dtrace_icookie_t cookie;
14577 
14578 	mutex_enter(&cpu_lock);
14579 	mutex_enter(&dtrace_lock);
14580 
14581 	if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
14582 		rval = EBUSY;
14583 		goto out;
14584 	}
14585 
14586 	/*
14587 	 * Before we can perform any checks, we must prime all of the
14588 	 * retained enablings that correspond to this state.
14589 	 */
14590 	dtrace_enabling_prime(state);
14591 
14592 	if (state->dts_destructive && !state->dts_cred.dcr_destructive) {
14593 		rval = EACCES;
14594 		goto out;
14595 	}
14596 
14597 	dtrace_state_prereserve(state);
14598 
14599 	/*
14600 	 * Now we want to do is try to allocate our speculations.
14601 	 * We do not automatically resize the number of speculations; if
14602 	 * this fails, we will fail the operation.
14603 	 */
14604 	nspec = opt[DTRACEOPT_NSPEC];
14605 	ASSERT(nspec != DTRACEOPT_UNSET);
14606 
14607 	if (nspec > INT_MAX) {
14608 		rval = ENOMEM;
14609 		goto out;
14610 	}
14611 
14612 	spec = kmem_zalloc(nspec * sizeof (dtrace_speculation_t),
14613 	    KM_NOSLEEP | KM_NORMALPRI);
14614 
14615 	if (spec == NULL) {
14616 		rval = ENOMEM;
14617 		goto out;
14618 	}
14619 
14620 	state->dts_speculations = spec;
14621 	state->dts_nspeculations = (int)nspec;
14622 
14623 	for (i = 0; i < nspec; i++) {
14624 		if ((buf = kmem_zalloc(bufsize,
14625 		    KM_NOSLEEP | KM_NORMALPRI)) == NULL) {
14626 			rval = ENOMEM;
14627 			goto err;
14628 		}
14629 
14630 		spec[i].dtsp_buffer = buf;
14631 	}
14632 
14633 	if (opt[DTRACEOPT_GRABANON] != DTRACEOPT_UNSET) {
14634 		if (dtrace_anon.dta_state == NULL) {
14635 			rval = ENOENT;
14636 			goto out;
14637 		}
14638 
14639 		if (state->dts_necbs != 0) {
14640 			rval = EALREADY;
14641 			goto out;
14642 		}
14643 
14644 		state->dts_anon = dtrace_anon_grab();
14645 		ASSERT(state->dts_anon != NULL);
14646 		state = state->dts_anon;
14647 
14648 		/*
14649 		 * We want "grabanon" to be set in the grabbed state, so we'll
14650 		 * copy that option value from the grabbing state into the
14651 		 * grabbed state.
14652 		 */
14653 		state->dts_options[DTRACEOPT_GRABANON] =
14654 		    opt[DTRACEOPT_GRABANON];
14655 
14656 		*cpu = dtrace_anon.dta_beganon;
14657 
14658 		/*
14659 		 * If the anonymous state is active (as it almost certainly
14660 		 * is if the anonymous enabling ultimately matched anything),
14661 		 * we don't allow any further option processing -- but we
14662 		 * don't return failure.
14663 		 */
14664 		if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
14665 			goto out;
14666 	}
14667 
14668 	if (opt[DTRACEOPT_AGGSIZE] != DTRACEOPT_UNSET &&
14669 	    opt[DTRACEOPT_AGGSIZE] != 0) {
14670 		if (state->dts_aggregations == NULL) {
14671 			/*
14672 			 * We're not going to create an aggregation buffer
14673 			 * because we don't have any ECBs that contain
14674 			 * aggregations -- set this option to 0.
14675 			 */
14676 			opt[DTRACEOPT_AGGSIZE] = 0;
14677 		} else {
14678 			/*
14679 			 * If we have an aggregation buffer, we must also have
14680 			 * a buffer to use as scratch.
14681 			 */
14682 			if (opt[DTRACEOPT_BUFSIZE] == DTRACEOPT_UNSET ||
14683 			    opt[DTRACEOPT_BUFSIZE] < state->dts_needed) {
14684 				opt[DTRACEOPT_BUFSIZE] = state->dts_needed;
14685 			}
14686 		}
14687 	}
14688 
14689 	if (opt[DTRACEOPT_SPECSIZE] != DTRACEOPT_UNSET &&
14690 	    opt[DTRACEOPT_SPECSIZE] != 0) {
14691 		if (!state->dts_speculates) {
14692 			/*
14693 			 * We're not going to create speculation buffers
14694 			 * because we don't have any ECBs that actually
14695 			 * speculate -- set the speculation size to 0.
14696 			 */
14697 			opt[DTRACEOPT_SPECSIZE] = 0;
14698 		}
14699 	}
14700 
14701 	/*
14702 	 * The bare minimum size for any buffer that we're actually going to
14703 	 * do anything to is sizeof (uint64_t).
14704 	 */
14705 	sz = sizeof (uint64_t);
14706 
14707 	if ((state->dts_needed != 0 && opt[DTRACEOPT_BUFSIZE] < sz) ||
14708 	    (state->dts_speculates && opt[DTRACEOPT_SPECSIZE] < sz) ||
14709 	    (state->dts_aggregations != NULL && opt[DTRACEOPT_AGGSIZE] < sz)) {
14710 		/*
14711 		 * A buffer size has been explicitly set to 0 (or to a size
14712 		 * that will be adjusted to 0) and we need the space -- we
14713 		 * need to return failure.  We return ENOSPC to differentiate
14714 		 * it from failing to allocate a buffer due to failure to meet
14715 		 * the reserve (for which we return E2BIG).
14716 		 */
14717 		rval = ENOSPC;
14718 		goto out;
14719 	}
14720 
14721 	if ((rval = dtrace_state_buffers(state)) != 0)
14722 		goto err;
14723 
14724 	if ((sz = opt[DTRACEOPT_DYNVARSIZE]) == DTRACEOPT_UNSET)
14725 		sz = dtrace_dstate_defsize;
14726 
14727 	do {
14728 		rval = dtrace_dstate_init(&state->dts_vstate.dtvs_dynvars, sz);
14729 
14730 		if (rval == 0)
14731 			break;
14732 
14733 		if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
14734 			goto err;
14735 	} while (sz >>= 1);
14736 
14737 	opt[DTRACEOPT_DYNVARSIZE] = sz;
14738 
14739 	if (rval != 0)
14740 		goto err;
14741 
14742 	if (opt[DTRACEOPT_STATUSRATE] > dtrace_statusrate_max)
14743 		opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_max;
14744 
14745 	if (opt[DTRACEOPT_CLEANRATE] == 0)
14746 		opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
14747 
14748 	if (opt[DTRACEOPT_CLEANRATE] < dtrace_cleanrate_min)
14749 		opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_min;
14750 
14751 	if (opt[DTRACEOPT_CLEANRATE] > dtrace_cleanrate_max)
14752 		opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
14753 
14754 	state->dts_alive = state->dts_laststatus = dtrace_gethrtime();
14755 #ifdef illumos
14756 	hdlr.cyh_func = (cyc_func_t)dtrace_state_clean;
14757 	hdlr.cyh_arg = state;
14758 	hdlr.cyh_level = CY_LOW_LEVEL;
14759 
14760 	when.cyt_when = 0;
14761 	when.cyt_interval = opt[DTRACEOPT_CLEANRATE];
14762 
14763 	state->dts_cleaner = cyclic_add(&hdlr, &when);
14764 
14765 	hdlr.cyh_func = (cyc_func_t)dtrace_state_deadman;
14766 	hdlr.cyh_arg = state;
14767 	hdlr.cyh_level = CY_LOW_LEVEL;
14768 
14769 	when.cyt_when = 0;
14770 	when.cyt_interval = dtrace_deadman_interval;
14771 
14772 	state->dts_deadman = cyclic_add(&hdlr, &when);
14773 #else
14774 	callout_reset(&state->dts_cleaner, hz * opt[DTRACEOPT_CLEANRATE] / NANOSEC,
14775 	    dtrace_state_clean, state);
14776 	callout_reset(&state->dts_deadman, hz * dtrace_deadman_interval / NANOSEC,
14777 	    dtrace_state_deadman, state);
14778 #endif
14779 
14780 	state->dts_activity = DTRACE_ACTIVITY_WARMUP;
14781 
14782 #ifdef illumos
14783 	if (state->dts_getf != 0 &&
14784 	    !(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)) {
14785 		/*
14786 		 * We don't have kernel privs but we have at least one call
14787 		 * to getf(); we need to bump our zone's count, and (if
14788 		 * this is the first enabling to have an unprivileged call
14789 		 * to getf()) we need to hook into closef().
14790 		 */
14791 		state->dts_cred.dcr_cred->cr_zone->zone_dtrace_getf++;
14792 
14793 		if (dtrace_getf++ == 0) {
14794 			ASSERT(dtrace_closef == NULL);
14795 			dtrace_closef = dtrace_getf_barrier;
14796 		}
14797 	}
14798 #endif
14799 
14800 	/*
14801 	 * Now it's time to actually fire the BEGIN probe.  We need to disable
14802 	 * interrupts here both to record the CPU on which we fired the BEGIN
14803 	 * probe (the data from this CPU will be processed first at user
14804 	 * level) and to manually activate the buffer for this CPU.
14805 	 */
14806 	cookie = dtrace_interrupt_disable();
14807 	*cpu = curcpu;
14808 	ASSERT(state->dts_buffer[*cpu].dtb_flags & DTRACEBUF_INACTIVE);
14809 	state->dts_buffer[*cpu].dtb_flags &= ~DTRACEBUF_INACTIVE;
14810 
14811 	dtrace_probe(dtrace_probeid_begin,
14812 	    (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
14813 	dtrace_interrupt_enable(cookie);
14814 	/*
14815 	 * We may have had an exit action from a BEGIN probe; only change our
14816 	 * state to ACTIVE if we're still in WARMUP.
14817 	 */
14818 	ASSERT(state->dts_activity == DTRACE_ACTIVITY_WARMUP ||
14819 	    state->dts_activity == DTRACE_ACTIVITY_DRAINING);
14820 
14821 	if (state->dts_activity == DTRACE_ACTIVITY_WARMUP)
14822 		state->dts_activity = DTRACE_ACTIVITY_ACTIVE;
14823 
14824 	/*
14825 	 * Regardless of whether or not now we're in ACTIVE or DRAINING, we
14826 	 * want each CPU to transition its principal buffer out of the
14827 	 * INACTIVE state.  Doing this assures that no CPU will suddenly begin
14828 	 * processing an ECB halfway down a probe's ECB chain; all CPUs will
14829 	 * atomically transition from processing none of a state's ECBs to
14830 	 * processing all of them.
14831 	 */
14832 	dtrace_xcall(DTRACE_CPUALL,
14833 	    (dtrace_xcall_t)dtrace_buffer_activate, state);
14834 	goto out;
14835 
14836 err:
14837 	dtrace_buffer_free(state->dts_buffer);
14838 	dtrace_buffer_free(state->dts_aggbuffer);
14839 
14840 	if ((nspec = state->dts_nspeculations) == 0) {
14841 		ASSERT(state->dts_speculations == NULL);
14842 		goto out;
14843 	}
14844 
14845 	spec = state->dts_speculations;
14846 	ASSERT(spec != NULL);
14847 
14848 	for (i = 0; i < state->dts_nspeculations; i++) {
14849 		if ((buf = spec[i].dtsp_buffer) == NULL)
14850 			break;
14851 
14852 		dtrace_buffer_free(buf);
14853 		kmem_free(buf, bufsize);
14854 	}
14855 
14856 	kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
14857 	state->dts_nspeculations = 0;
14858 	state->dts_speculations = NULL;
14859 
14860 out:
14861 	mutex_exit(&dtrace_lock);
14862 	mutex_exit(&cpu_lock);
14863 
14864 	return (rval);
14865 }
14866 
14867 static int
14868 dtrace_state_stop(dtrace_state_t *state, processorid_t *cpu)
14869 {
14870 	dtrace_icookie_t cookie;
14871 
14872 	ASSERT(MUTEX_HELD(&dtrace_lock));
14873 
14874 	if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE &&
14875 	    state->dts_activity != DTRACE_ACTIVITY_DRAINING)
14876 		return (EINVAL);
14877 
14878 	/*
14879 	 * We'll set the activity to DTRACE_ACTIVITY_DRAINING, and issue a sync
14880 	 * to be sure that every CPU has seen it.  See below for the details
14881 	 * on why this is done.
14882 	 */
14883 	state->dts_activity = DTRACE_ACTIVITY_DRAINING;
14884 	dtrace_sync();
14885 
14886 	/*
14887 	 * By this point, it is impossible for any CPU to be still processing
14888 	 * with DTRACE_ACTIVITY_ACTIVE.  We can thus set our activity to
14889 	 * DTRACE_ACTIVITY_COOLDOWN and know that we're not racing with any
14890 	 * other CPU in dtrace_buffer_reserve().  This allows dtrace_probe()
14891 	 * and callees to know that the activity is DTRACE_ACTIVITY_COOLDOWN
14892 	 * iff we're in the END probe.
14893 	 */
14894 	state->dts_activity = DTRACE_ACTIVITY_COOLDOWN;
14895 	dtrace_sync();
14896 	ASSERT(state->dts_activity == DTRACE_ACTIVITY_COOLDOWN);
14897 
14898 	/*
14899 	 * Finally, we can release the reserve and call the END probe.  We
14900 	 * disable interrupts across calling the END probe to allow us to
14901 	 * return the CPU on which we actually called the END probe.  This
14902 	 * allows user-land to be sure that this CPU's principal buffer is
14903 	 * processed last.
14904 	 */
14905 	state->dts_reserve = 0;
14906 
14907 	cookie = dtrace_interrupt_disable();
14908 	*cpu = curcpu;
14909 	dtrace_probe(dtrace_probeid_end,
14910 	    (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
14911 	dtrace_interrupt_enable(cookie);
14912 
14913 	state->dts_activity = DTRACE_ACTIVITY_STOPPED;
14914 	dtrace_sync();
14915 
14916 #ifdef illumos
14917 	if (state->dts_getf != 0 &&
14918 	    !(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)) {
14919 		/*
14920 		 * We don't have kernel privs but we have at least one call
14921 		 * to getf(); we need to lower our zone's count, and (if
14922 		 * this is the last enabling to have an unprivileged call
14923 		 * to getf()) we need to clear the closef() hook.
14924 		 */
14925 		ASSERT(state->dts_cred.dcr_cred->cr_zone->zone_dtrace_getf > 0);
14926 		ASSERT(dtrace_closef == dtrace_getf_barrier);
14927 		ASSERT(dtrace_getf > 0);
14928 
14929 		state->dts_cred.dcr_cred->cr_zone->zone_dtrace_getf--;
14930 
14931 		if (--dtrace_getf == 0)
14932 			dtrace_closef = NULL;
14933 	}
14934 #endif
14935 
14936 	return (0);
14937 }
14938 
14939 static int
14940 dtrace_state_option(dtrace_state_t *state, dtrace_optid_t option,
14941     dtrace_optval_t val)
14942 {
14943 	ASSERT(MUTEX_HELD(&dtrace_lock));
14944 
14945 	if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
14946 		return (EBUSY);
14947 
14948 	if (option >= DTRACEOPT_MAX)
14949 		return (EINVAL);
14950 
14951 	if (option != DTRACEOPT_CPU && val < 0)
14952 		return (EINVAL);
14953 
14954 	switch (option) {
14955 	case DTRACEOPT_DESTRUCTIVE:
14956 		if (dtrace_destructive_disallow)
14957 			return (EACCES);
14958 
14959 		state->dts_cred.dcr_destructive = 1;
14960 		break;
14961 
14962 	case DTRACEOPT_BUFSIZE:
14963 	case DTRACEOPT_DYNVARSIZE:
14964 	case DTRACEOPT_AGGSIZE:
14965 	case DTRACEOPT_SPECSIZE:
14966 	case DTRACEOPT_STRSIZE:
14967 		if (val < 0)
14968 			return (EINVAL);
14969 
14970 		if (val >= LONG_MAX) {
14971 			/*
14972 			 * If this is an otherwise negative value, set it to
14973 			 * the highest multiple of 128m less than LONG_MAX.
14974 			 * Technically, we're adjusting the size without
14975 			 * regard to the buffer resizing policy, but in fact,
14976 			 * this has no effect -- if we set the buffer size to
14977 			 * ~LONG_MAX and the buffer policy is ultimately set to
14978 			 * be "manual", the buffer allocation is guaranteed to
14979 			 * fail, if only because the allocation requires two
14980 			 * buffers.  (We set the the size to the highest
14981 			 * multiple of 128m because it ensures that the size
14982 			 * will remain a multiple of a megabyte when
14983 			 * repeatedly halved -- all the way down to 15m.)
14984 			 */
14985 			val = LONG_MAX - (1 << 27) + 1;
14986 		}
14987 	}
14988 
14989 	state->dts_options[option] = val;
14990 
14991 	return (0);
14992 }
14993 
14994 static void
14995 dtrace_state_destroy(dtrace_state_t *state)
14996 {
14997 	dtrace_ecb_t *ecb;
14998 	dtrace_vstate_t *vstate = &state->dts_vstate;
14999 #ifdef illumos
15000 	minor_t minor = getminor(state->dts_dev);
15001 #endif
15002 	int i, bufsize = NCPU * sizeof (dtrace_buffer_t);
15003 	dtrace_speculation_t *spec = state->dts_speculations;
15004 	int nspec = state->dts_nspeculations;
15005 	uint32_t match;
15006 
15007 	ASSERT(MUTEX_HELD(&dtrace_lock));
15008 	ASSERT(MUTEX_HELD(&cpu_lock));
15009 
15010 	/*
15011 	 * First, retract any retained enablings for this state.
15012 	 */
15013 	dtrace_enabling_retract(state);
15014 	ASSERT(state->dts_nretained == 0);
15015 
15016 	if (state->dts_activity == DTRACE_ACTIVITY_ACTIVE ||
15017 	    state->dts_activity == DTRACE_ACTIVITY_DRAINING) {
15018 		/*
15019 		 * We have managed to come into dtrace_state_destroy() on a
15020 		 * hot enabling -- almost certainly because of a disorderly
15021 		 * shutdown of a consumer.  (That is, a consumer that is
15022 		 * exiting without having called dtrace_stop().) In this case,
15023 		 * we're going to set our activity to be KILLED, and then
15024 		 * issue a sync to be sure that everyone is out of probe
15025 		 * context before we start blowing away ECBs.
15026 		 */
15027 		state->dts_activity = DTRACE_ACTIVITY_KILLED;
15028 		dtrace_sync();
15029 	}
15030 
15031 	/*
15032 	 * Release the credential hold we took in dtrace_state_create().
15033 	 */
15034 	if (state->dts_cred.dcr_cred != NULL)
15035 		crfree(state->dts_cred.dcr_cred);
15036 
15037 	/*
15038 	 * Now we can safely disable and destroy any enabled probes.  Because
15039 	 * any DTRACE_PRIV_KERNEL probes may actually be slowing our progress
15040 	 * (especially if they're all enabled), we take two passes through the
15041 	 * ECBs:  in the first, we disable just DTRACE_PRIV_KERNEL probes, and
15042 	 * in the second we disable whatever is left over.
15043 	 */
15044 	for (match = DTRACE_PRIV_KERNEL; ; match = 0) {
15045 		for (i = 0; i < state->dts_necbs; i++) {
15046 			if ((ecb = state->dts_ecbs[i]) == NULL)
15047 				continue;
15048 
15049 			if (match && ecb->dte_probe != NULL) {
15050 				dtrace_probe_t *probe = ecb->dte_probe;
15051 				dtrace_provider_t *prov = probe->dtpr_provider;
15052 
15053 				if (!(prov->dtpv_priv.dtpp_flags & match))
15054 					continue;
15055 			}
15056 
15057 			dtrace_ecb_disable(ecb);
15058 			dtrace_ecb_destroy(ecb);
15059 		}
15060 
15061 		if (!match)
15062 			break;
15063 	}
15064 
15065 	/*
15066 	 * Before we free the buffers, perform one more sync to assure that
15067 	 * every CPU is out of probe context.
15068 	 */
15069 	dtrace_sync();
15070 
15071 	dtrace_buffer_free(state->dts_buffer);
15072 	dtrace_buffer_free(state->dts_aggbuffer);
15073 
15074 	for (i = 0; i < nspec; i++)
15075 		dtrace_buffer_free(spec[i].dtsp_buffer);
15076 
15077 #ifdef illumos
15078 	if (state->dts_cleaner != CYCLIC_NONE)
15079 		cyclic_remove(state->dts_cleaner);
15080 
15081 	if (state->dts_deadman != CYCLIC_NONE)
15082 		cyclic_remove(state->dts_deadman);
15083 #else
15084 	callout_stop(&state->dts_cleaner);
15085 	callout_drain(&state->dts_cleaner);
15086 	callout_stop(&state->dts_deadman);
15087 	callout_drain(&state->dts_deadman);
15088 #endif
15089 
15090 	dtrace_dstate_fini(&vstate->dtvs_dynvars);
15091 	dtrace_vstate_fini(vstate);
15092 	if (state->dts_ecbs != NULL)
15093 		kmem_free(state->dts_ecbs, state->dts_necbs * sizeof (dtrace_ecb_t *));
15094 
15095 	if (state->dts_aggregations != NULL) {
15096 #ifdef DEBUG
15097 		for (i = 0; i < state->dts_naggregations; i++)
15098 			ASSERT(state->dts_aggregations[i] == NULL);
15099 #endif
15100 		ASSERT(state->dts_naggregations > 0);
15101 		kmem_free(state->dts_aggregations,
15102 		    state->dts_naggregations * sizeof (dtrace_aggregation_t *));
15103 	}
15104 
15105 	kmem_free(state->dts_buffer, bufsize);
15106 	kmem_free(state->dts_aggbuffer, bufsize);
15107 
15108 	for (i = 0; i < nspec; i++)
15109 		kmem_free(spec[i].dtsp_buffer, bufsize);
15110 
15111 	if (spec != NULL)
15112 		kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
15113 
15114 	dtrace_format_destroy(state);
15115 
15116 	if (state->dts_aggid_arena != NULL) {
15117 #ifdef illumos
15118 		vmem_destroy(state->dts_aggid_arena);
15119 #else
15120 		delete_unrhdr(state->dts_aggid_arena);
15121 #endif
15122 		state->dts_aggid_arena = NULL;
15123 	}
15124 #ifdef illumos
15125 	ddi_soft_state_free(dtrace_softstate, minor);
15126 	vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
15127 #endif
15128 }
15129 
15130 /*
15131  * DTrace Anonymous Enabling Functions
15132  */
15133 static dtrace_state_t *
15134 dtrace_anon_grab(void)
15135 {
15136 	dtrace_state_t *state;
15137 
15138 	ASSERT(MUTEX_HELD(&dtrace_lock));
15139 
15140 	if ((state = dtrace_anon.dta_state) == NULL) {
15141 		ASSERT(dtrace_anon.dta_enabling == NULL);
15142 		return (NULL);
15143 	}
15144 
15145 	ASSERT(dtrace_anon.dta_enabling != NULL);
15146 	ASSERT(dtrace_retained != NULL);
15147 
15148 	dtrace_enabling_destroy(dtrace_anon.dta_enabling);
15149 	dtrace_anon.dta_enabling = NULL;
15150 	dtrace_anon.dta_state = NULL;
15151 
15152 	return (state);
15153 }
15154 
15155 static void
15156 dtrace_anon_property(void)
15157 {
15158 	int i, rv;
15159 	dtrace_state_t *state;
15160 	dof_hdr_t *dof;
15161 	char c[32];		/* enough for "dof-data-" + digits */
15162 
15163 	ASSERT(MUTEX_HELD(&dtrace_lock));
15164 	ASSERT(MUTEX_HELD(&cpu_lock));
15165 
15166 	for (i = 0; ; i++) {
15167 		(void) snprintf(c, sizeof (c), "dof-data-%d", i);
15168 
15169 		dtrace_err_verbose = 1;
15170 
15171 		if ((dof = dtrace_dof_property(c)) == NULL) {
15172 			dtrace_err_verbose = 0;
15173 			break;
15174 		}
15175 
15176 #ifdef illumos
15177 		/*
15178 		 * We want to create anonymous state, so we need to transition
15179 		 * the kernel debugger to indicate that DTrace is active.  If
15180 		 * this fails (e.g. because the debugger has modified text in
15181 		 * some way), we won't continue with the processing.
15182 		 */
15183 		if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
15184 			cmn_err(CE_NOTE, "kernel debugger active; anonymous "
15185 			    "enabling ignored.");
15186 			dtrace_dof_destroy(dof);
15187 			break;
15188 		}
15189 #endif
15190 
15191 		/*
15192 		 * If we haven't allocated an anonymous state, we'll do so now.
15193 		 */
15194 		if ((state = dtrace_anon.dta_state) == NULL) {
15195 #ifdef illumos
15196 			state = dtrace_state_create(NULL, NULL);
15197 #else
15198 			state = dtrace_state_create(NULL);
15199 #endif
15200 			dtrace_anon.dta_state = state;
15201 
15202 			if (state == NULL) {
15203 				/*
15204 				 * This basically shouldn't happen:  the only
15205 				 * failure mode from dtrace_state_create() is a
15206 				 * failure of ddi_soft_state_zalloc() that
15207 				 * itself should never happen.  Still, the
15208 				 * interface allows for a failure mode, and
15209 				 * we want to fail as gracefully as possible:
15210 				 * we'll emit an error message and cease
15211 				 * processing anonymous state in this case.
15212 				 */
15213 				cmn_err(CE_WARN, "failed to create "
15214 				    "anonymous state");
15215 				dtrace_dof_destroy(dof);
15216 				break;
15217 			}
15218 		}
15219 
15220 		rv = dtrace_dof_slurp(dof, &state->dts_vstate, CRED(),
15221 		    &dtrace_anon.dta_enabling, 0, B_TRUE);
15222 
15223 		if (rv == 0)
15224 			rv = dtrace_dof_options(dof, state);
15225 
15226 		dtrace_err_verbose = 0;
15227 		dtrace_dof_destroy(dof);
15228 
15229 		if (rv != 0) {
15230 			/*
15231 			 * This is malformed DOF; chuck any anonymous state
15232 			 * that we created.
15233 			 */
15234 			ASSERT(dtrace_anon.dta_enabling == NULL);
15235 			dtrace_state_destroy(state);
15236 			dtrace_anon.dta_state = NULL;
15237 			break;
15238 		}
15239 
15240 		ASSERT(dtrace_anon.dta_enabling != NULL);
15241 	}
15242 
15243 	if (dtrace_anon.dta_enabling != NULL) {
15244 		int rval;
15245 
15246 		/*
15247 		 * dtrace_enabling_retain() can only fail because we are
15248 		 * trying to retain more enablings than are allowed -- but
15249 		 * we only have one anonymous enabling, and we are guaranteed
15250 		 * to be allowed at least one retained enabling; we assert
15251 		 * that dtrace_enabling_retain() returns success.
15252 		 */
15253 		rval = dtrace_enabling_retain(dtrace_anon.dta_enabling);
15254 		ASSERT(rval == 0);
15255 
15256 		dtrace_enabling_dump(dtrace_anon.dta_enabling);
15257 	}
15258 }
15259 
15260 /*
15261  * DTrace Helper Functions
15262  */
15263 static void
15264 dtrace_helper_trace(dtrace_helper_action_t *helper,
15265     dtrace_mstate_t *mstate, dtrace_vstate_t *vstate, int where)
15266 {
15267 	uint32_t size, next, nnext, i;
15268 	dtrace_helptrace_t *ent, *buffer;
15269 	uint16_t flags = cpu_core[curcpu].cpuc_dtrace_flags;
15270 
15271 	if ((buffer = dtrace_helptrace_buffer) == NULL)
15272 		return;
15273 
15274 	ASSERT(vstate->dtvs_nlocals <= dtrace_helptrace_nlocals);
15275 
15276 	/*
15277 	 * What would a tracing framework be without its own tracing
15278 	 * framework?  (Well, a hell of a lot simpler, for starters...)
15279 	 */
15280 	size = sizeof (dtrace_helptrace_t) + dtrace_helptrace_nlocals *
15281 	    sizeof (uint64_t) - sizeof (uint64_t);
15282 
15283 	/*
15284 	 * Iterate until we can allocate a slot in the trace buffer.
15285 	 */
15286 	do {
15287 		next = dtrace_helptrace_next;
15288 
15289 		if (next + size < dtrace_helptrace_bufsize) {
15290 			nnext = next + size;
15291 		} else {
15292 			nnext = size;
15293 		}
15294 	} while (dtrace_cas32(&dtrace_helptrace_next, next, nnext) != next);
15295 
15296 	/*
15297 	 * We have our slot; fill it in.
15298 	 */
15299 	if (nnext == size) {
15300 		dtrace_helptrace_wrapped++;
15301 		next = 0;
15302 	}
15303 
15304 	ent = (dtrace_helptrace_t *)((uintptr_t)buffer + next);
15305 	ent->dtht_helper = helper;
15306 	ent->dtht_where = where;
15307 	ent->dtht_nlocals = vstate->dtvs_nlocals;
15308 
15309 	ent->dtht_fltoffs = (mstate->dtms_present & DTRACE_MSTATE_FLTOFFS) ?
15310 	    mstate->dtms_fltoffs : -1;
15311 	ent->dtht_fault = DTRACE_FLAGS2FLT(flags);
15312 	ent->dtht_illval = cpu_core[curcpu].cpuc_dtrace_illval;
15313 
15314 	for (i = 0; i < vstate->dtvs_nlocals; i++) {
15315 		dtrace_statvar_t *svar;
15316 
15317 		if ((svar = vstate->dtvs_locals[i]) == NULL)
15318 			continue;
15319 
15320 		ASSERT(svar->dtsv_size >= NCPU * sizeof (uint64_t));
15321 		ent->dtht_locals[i] =
15322 		    ((uint64_t *)(uintptr_t)svar->dtsv_data)[curcpu];
15323 	}
15324 }
15325 
15326 static uint64_t
15327 dtrace_helper(int which, dtrace_mstate_t *mstate,
15328     dtrace_state_t *state, uint64_t arg0, uint64_t arg1)
15329 {
15330 	uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
15331 	uint64_t sarg0 = mstate->dtms_arg[0];
15332 	uint64_t sarg1 = mstate->dtms_arg[1];
15333 	uint64_t rval = 0;
15334 	dtrace_helpers_t *helpers = curproc->p_dtrace_helpers;
15335 	dtrace_helper_action_t *helper;
15336 	dtrace_vstate_t *vstate;
15337 	dtrace_difo_t *pred;
15338 	int i, trace = dtrace_helptrace_buffer != NULL;
15339 
15340 	ASSERT(which >= 0 && which < DTRACE_NHELPER_ACTIONS);
15341 
15342 	if (helpers == NULL)
15343 		return (0);
15344 
15345 	if ((helper = helpers->dthps_actions[which]) == NULL)
15346 		return (0);
15347 
15348 	vstate = &helpers->dthps_vstate;
15349 	mstate->dtms_arg[0] = arg0;
15350 	mstate->dtms_arg[1] = arg1;
15351 
15352 	/*
15353 	 * Now iterate over each helper.  If its predicate evaluates to 'true',
15354 	 * we'll call the corresponding actions.  Note that the below calls
15355 	 * to dtrace_dif_emulate() may set faults in machine state.  This is
15356 	 * okay:  our caller (the outer dtrace_dif_emulate()) will simply plow
15357 	 * the stored DIF offset with its own (which is the desired behavior).
15358 	 * Also, note the calls to dtrace_dif_emulate() may allocate scratch
15359 	 * from machine state; this is okay, too.
15360 	 */
15361 	for (; helper != NULL; helper = helper->dtha_next) {
15362 		if ((pred = helper->dtha_predicate) != NULL) {
15363 			if (trace)
15364 				dtrace_helper_trace(helper, mstate, vstate, 0);
15365 
15366 			if (!dtrace_dif_emulate(pred, mstate, vstate, state))
15367 				goto next;
15368 
15369 			if (*flags & CPU_DTRACE_FAULT)
15370 				goto err;
15371 		}
15372 
15373 		for (i = 0; i < helper->dtha_nactions; i++) {
15374 			if (trace)
15375 				dtrace_helper_trace(helper,
15376 				    mstate, vstate, i + 1);
15377 
15378 			rval = dtrace_dif_emulate(helper->dtha_actions[i],
15379 			    mstate, vstate, state);
15380 
15381 			if (*flags & CPU_DTRACE_FAULT)
15382 				goto err;
15383 		}
15384 
15385 next:
15386 		if (trace)
15387 			dtrace_helper_trace(helper, mstate, vstate,
15388 			    DTRACE_HELPTRACE_NEXT);
15389 	}
15390 
15391 	if (trace)
15392 		dtrace_helper_trace(helper, mstate, vstate,
15393 		    DTRACE_HELPTRACE_DONE);
15394 
15395 	/*
15396 	 * Restore the arg0 that we saved upon entry.
15397 	 */
15398 	mstate->dtms_arg[0] = sarg0;
15399 	mstate->dtms_arg[1] = sarg1;
15400 
15401 	return (rval);
15402 
15403 err:
15404 	if (trace)
15405 		dtrace_helper_trace(helper, mstate, vstate,
15406 		    DTRACE_HELPTRACE_ERR);
15407 
15408 	/*
15409 	 * Restore the arg0 that we saved upon entry.
15410 	 */
15411 	mstate->dtms_arg[0] = sarg0;
15412 	mstate->dtms_arg[1] = sarg1;
15413 
15414 	return (0);
15415 }
15416 
15417 static void
15418 dtrace_helper_action_destroy(dtrace_helper_action_t *helper,
15419     dtrace_vstate_t *vstate)
15420 {
15421 	int i;
15422 
15423 	if (helper->dtha_predicate != NULL)
15424 		dtrace_difo_release(helper->dtha_predicate, vstate);
15425 
15426 	for (i = 0; i < helper->dtha_nactions; i++) {
15427 		ASSERT(helper->dtha_actions[i] != NULL);
15428 		dtrace_difo_release(helper->dtha_actions[i], vstate);
15429 	}
15430 
15431 	kmem_free(helper->dtha_actions,
15432 	    helper->dtha_nactions * sizeof (dtrace_difo_t *));
15433 	kmem_free(helper, sizeof (dtrace_helper_action_t));
15434 }
15435 
15436 static int
15437 dtrace_helper_destroygen(dtrace_helpers_t *help, int gen)
15438 {
15439 	proc_t *p = curproc;
15440 	dtrace_vstate_t *vstate;
15441 	int i;
15442 
15443 	if (help == NULL)
15444 		help = p->p_dtrace_helpers;
15445 
15446 	ASSERT(MUTEX_HELD(&dtrace_lock));
15447 
15448 	if (help == NULL || gen > help->dthps_generation)
15449 		return (EINVAL);
15450 
15451 	vstate = &help->dthps_vstate;
15452 
15453 	for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
15454 		dtrace_helper_action_t *last = NULL, *h, *next;
15455 
15456 		for (h = help->dthps_actions[i]; h != NULL; h = next) {
15457 			next = h->dtha_next;
15458 
15459 			if (h->dtha_generation == gen) {
15460 				if (last != NULL) {
15461 					last->dtha_next = next;
15462 				} else {
15463 					help->dthps_actions[i] = next;
15464 				}
15465 
15466 				dtrace_helper_action_destroy(h, vstate);
15467 			} else {
15468 				last = h;
15469 			}
15470 		}
15471 	}
15472 
15473 	/*
15474 	 * Interate until we've cleared out all helper providers with the
15475 	 * given generation number.
15476 	 */
15477 	for (;;) {
15478 		dtrace_helper_provider_t *prov;
15479 
15480 		/*
15481 		 * Look for a helper provider with the right generation. We
15482 		 * have to start back at the beginning of the list each time
15483 		 * because we drop dtrace_lock. It's unlikely that we'll make
15484 		 * more than two passes.
15485 		 */
15486 		for (i = 0; i < help->dthps_nprovs; i++) {
15487 			prov = help->dthps_provs[i];
15488 
15489 			if (prov->dthp_generation == gen)
15490 				break;
15491 		}
15492 
15493 		/*
15494 		 * If there were no matches, we're done.
15495 		 */
15496 		if (i == help->dthps_nprovs)
15497 			break;
15498 
15499 		/*
15500 		 * Move the last helper provider into this slot.
15501 		 */
15502 		help->dthps_nprovs--;
15503 		help->dthps_provs[i] = help->dthps_provs[help->dthps_nprovs];
15504 		help->dthps_provs[help->dthps_nprovs] = NULL;
15505 
15506 		mutex_exit(&dtrace_lock);
15507 
15508 		/*
15509 		 * If we have a meta provider, remove this helper provider.
15510 		 */
15511 		mutex_enter(&dtrace_meta_lock);
15512 		if (dtrace_meta_pid != NULL) {
15513 			ASSERT(dtrace_deferred_pid == NULL);
15514 			dtrace_helper_provider_remove(&prov->dthp_prov,
15515 			    p->p_pid);
15516 		}
15517 		mutex_exit(&dtrace_meta_lock);
15518 
15519 		dtrace_helper_provider_destroy(prov);
15520 
15521 		mutex_enter(&dtrace_lock);
15522 	}
15523 
15524 	return (0);
15525 }
15526 
15527 static int
15528 dtrace_helper_validate(dtrace_helper_action_t *helper)
15529 {
15530 	int err = 0, i;
15531 	dtrace_difo_t *dp;
15532 
15533 	if ((dp = helper->dtha_predicate) != NULL)
15534 		err += dtrace_difo_validate_helper(dp);
15535 
15536 	for (i = 0; i < helper->dtha_nactions; i++)
15537 		err += dtrace_difo_validate_helper(helper->dtha_actions[i]);
15538 
15539 	return (err == 0);
15540 }
15541 
15542 static int
15543 dtrace_helper_action_add(int which, dtrace_ecbdesc_t *ep,
15544     dtrace_helpers_t *help)
15545 {
15546 	dtrace_helper_action_t *helper, *last;
15547 	dtrace_actdesc_t *act;
15548 	dtrace_vstate_t *vstate;
15549 	dtrace_predicate_t *pred;
15550 	int count = 0, nactions = 0, i;
15551 
15552 	if (which < 0 || which >= DTRACE_NHELPER_ACTIONS)
15553 		return (EINVAL);
15554 
15555 	last = help->dthps_actions[which];
15556 	vstate = &help->dthps_vstate;
15557 
15558 	for (count = 0; last != NULL; last = last->dtha_next) {
15559 		count++;
15560 		if (last->dtha_next == NULL)
15561 			break;
15562 	}
15563 
15564 	/*
15565 	 * If we already have dtrace_helper_actions_max helper actions for this
15566 	 * helper action type, we'll refuse to add a new one.
15567 	 */
15568 	if (count >= dtrace_helper_actions_max)
15569 		return (ENOSPC);
15570 
15571 	helper = kmem_zalloc(sizeof (dtrace_helper_action_t), KM_SLEEP);
15572 	helper->dtha_generation = help->dthps_generation;
15573 
15574 	if ((pred = ep->dted_pred.dtpdd_predicate) != NULL) {
15575 		ASSERT(pred->dtp_difo != NULL);
15576 		dtrace_difo_hold(pred->dtp_difo);
15577 		helper->dtha_predicate = pred->dtp_difo;
15578 	}
15579 
15580 	for (act = ep->dted_action; act != NULL; act = act->dtad_next) {
15581 		if (act->dtad_kind != DTRACEACT_DIFEXPR)
15582 			goto err;
15583 
15584 		if (act->dtad_difo == NULL)
15585 			goto err;
15586 
15587 		nactions++;
15588 	}
15589 
15590 	helper->dtha_actions = kmem_zalloc(sizeof (dtrace_difo_t *) *
15591 	    (helper->dtha_nactions = nactions), KM_SLEEP);
15592 
15593 	for (act = ep->dted_action, i = 0; act != NULL; act = act->dtad_next) {
15594 		dtrace_difo_hold(act->dtad_difo);
15595 		helper->dtha_actions[i++] = act->dtad_difo;
15596 	}
15597 
15598 	if (!dtrace_helper_validate(helper))
15599 		goto err;
15600 
15601 	if (last == NULL) {
15602 		help->dthps_actions[which] = helper;
15603 	} else {
15604 		last->dtha_next = helper;
15605 	}
15606 
15607 	if (vstate->dtvs_nlocals > dtrace_helptrace_nlocals) {
15608 		dtrace_helptrace_nlocals = vstate->dtvs_nlocals;
15609 		dtrace_helptrace_next = 0;
15610 	}
15611 
15612 	return (0);
15613 err:
15614 	dtrace_helper_action_destroy(helper, vstate);
15615 	return (EINVAL);
15616 }
15617 
15618 static void
15619 dtrace_helper_provider_register(proc_t *p, dtrace_helpers_t *help,
15620     dof_helper_t *dofhp)
15621 {
15622 	ASSERT(MUTEX_NOT_HELD(&dtrace_lock));
15623 
15624 	mutex_enter(&dtrace_meta_lock);
15625 	mutex_enter(&dtrace_lock);
15626 
15627 	if (!dtrace_attached() || dtrace_meta_pid == NULL) {
15628 		/*
15629 		 * If the dtrace module is loaded but not attached, or if
15630 		 * there aren't isn't a meta provider registered to deal with
15631 		 * these provider descriptions, we need to postpone creating
15632 		 * the actual providers until later.
15633 		 */
15634 
15635 		if (help->dthps_next == NULL && help->dthps_prev == NULL &&
15636 		    dtrace_deferred_pid != help) {
15637 			help->dthps_deferred = 1;
15638 			help->dthps_pid = p->p_pid;
15639 			help->dthps_next = dtrace_deferred_pid;
15640 			help->dthps_prev = NULL;
15641 			if (dtrace_deferred_pid != NULL)
15642 				dtrace_deferred_pid->dthps_prev = help;
15643 			dtrace_deferred_pid = help;
15644 		}
15645 
15646 		mutex_exit(&dtrace_lock);
15647 
15648 	} else if (dofhp != NULL) {
15649 		/*
15650 		 * If the dtrace module is loaded and we have a particular
15651 		 * helper provider description, pass that off to the
15652 		 * meta provider.
15653 		 */
15654 
15655 		mutex_exit(&dtrace_lock);
15656 
15657 		dtrace_helper_provide(dofhp, p->p_pid);
15658 
15659 	} else {
15660 		/*
15661 		 * Otherwise, just pass all the helper provider descriptions
15662 		 * off to the meta provider.
15663 		 */
15664 
15665 		int i;
15666 		mutex_exit(&dtrace_lock);
15667 
15668 		for (i = 0; i < help->dthps_nprovs; i++) {
15669 			dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
15670 			    p->p_pid);
15671 		}
15672 	}
15673 
15674 	mutex_exit(&dtrace_meta_lock);
15675 }
15676 
15677 static int
15678 dtrace_helper_provider_add(dof_helper_t *dofhp, dtrace_helpers_t *help, int gen)
15679 {
15680 	dtrace_helper_provider_t *hprov, **tmp_provs;
15681 	uint_t tmp_maxprovs, i;
15682 
15683 	ASSERT(MUTEX_HELD(&dtrace_lock));
15684 	ASSERT(help != NULL);
15685 
15686 	/*
15687 	 * If we already have dtrace_helper_providers_max helper providers,
15688 	 * we're refuse to add a new one.
15689 	 */
15690 	if (help->dthps_nprovs >= dtrace_helper_providers_max)
15691 		return (ENOSPC);
15692 
15693 	/*
15694 	 * Check to make sure this isn't a duplicate.
15695 	 */
15696 	for (i = 0; i < help->dthps_nprovs; i++) {
15697 		if (dofhp->dofhp_dof ==
15698 		    help->dthps_provs[i]->dthp_prov.dofhp_dof)
15699 			return (EALREADY);
15700 	}
15701 
15702 	hprov = kmem_zalloc(sizeof (dtrace_helper_provider_t), KM_SLEEP);
15703 	hprov->dthp_prov = *dofhp;
15704 	hprov->dthp_ref = 1;
15705 	hprov->dthp_generation = gen;
15706 
15707 	/*
15708 	 * Allocate a bigger table for helper providers if it's already full.
15709 	 */
15710 	if (help->dthps_maxprovs == help->dthps_nprovs) {
15711 		tmp_maxprovs = help->dthps_maxprovs;
15712 		tmp_provs = help->dthps_provs;
15713 
15714 		if (help->dthps_maxprovs == 0)
15715 			help->dthps_maxprovs = 2;
15716 		else
15717 			help->dthps_maxprovs *= 2;
15718 		if (help->dthps_maxprovs > dtrace_helper_providers_max)
15719 			help->dthps_maxprovs = dtrace_helper_providers_max;
15720 
15721 		ASSERT(tmp_maxprovs < help->dthps_maxprovs);
15722 
15723 		help->dthps_provs = kmem_zalloc(help->dthps_maxprovs *
15724 		    sizeof (dtrace_helper_provider_t *), KM_SLEEP);
15725 
15726 		if (tmp_provs != NULL) {
15727 			bcopy(tmp_provs, help->dthps_provs, tmp_maxprovs *
15728 			    sizeof (dtrace_helper_provider_t *));
15729 			kmem_free(tmp_provs, tmp_maxprovs *
15730 			    sizeof (dtrace_helper_provider_t *));
15731 		}
15732 	}
15733 
15734 	help->dthps_provs[help->dthps_nprovs] = hprov;
15735 	help->dthps_nprovs++;
15736 
15737 	return (0);
15738 }
15739 
15740 static void
15741 dtrace_helper_provider_destroy(dtrace_helper_provider_t *hprov)
15742 {
15743 	mutex_enter(&dtrace_lock);
15744 
15745 	if (--hprov->dthp_ref == 0) {
15746 		dof_hdr_t *dof;
15747 		mutex_exit(&dtrace_lock);
15748 		dof = (dof_hdr_t *)(uintptr_t)hprov->dthp_prov.dofhp_dof;
15749 		dtrace_dof_destroy(dof);
15750 		kmem_free(hprov, sizeof (dtrace_helper_provider_t));
15751 	} else {
15752 		mutex_exit(&dtrace_lock);
15753 	}
15754 }
15755 
15756 static int
15757 dtrace_helper_provider_validate(dof_hdr_t *dof, dof_sec_t *sec)
15758 {
15759 	uintptr_t daddr = (uintptr_t)dof;
15760 	dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
15761 	dof_provider_t *provider;
15762 	dof_probe_t *probe;
15763 	uint8_t *arg;
15764 	char *strtab, *typestr;
15765 	dof_stridx_t typeidx;
15766 	size_t typesz;
15767 	uint_t nprobes, j, k;
15768 
15769 	ASSERT(sec->dofs_type == DOF_SECT_PROVIDER);
15770 
15771 	if (sec->dofs_offset & (sizeof (uint_t) - 1)) {
15772 		dtrace_dof_error(dof, "misaligned section offset");
15773 		return (-1);
15774 	}
15775 
15776 	/*
15777 	 * The section needs to be large enough to contain the DOF provider
15778 	 * structure appropriate for the given version.
15779 	 */
15780 	if (sec->dofs_size <
15781 	    ((dof->dofh_ident[DOF_ID_VERSION] == DOF_VERSION_1) ?
15782 	    offsetof(dof_provider_t, dofpv_prenoffs) :
15783 	    sizeof (dof_provider_t))) {
15784 		dtrace_dof_error(dof, "provider section too small");
15785 		return (-1);
15786 	}
15787 
15788 	provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
15789 	str_sec = dtrace_dof_sect(dof, DOF_SECT_STRTAB, provider->dofpv_strtab);
15790 	prb_sec = dtrace_dof_sect(dof, DOF_SECT_PROBES, provider->dofpv_probes);
15791 	arg_sec = dtrace_dof_sect(dof, DOF_SECT_PRARGS, provider->dofpv_prargs);
15792 	off_sec = dtrace_dof_sect(dof, DOF_SECT_PROFFS, provider->dofpv_proffs);
15793 
15794 	if (str_sec == NULL || prb_sec == NULL ||
15795 	    arg_sec == NULL || off_sec == NULL)
15796 		return (-1);
15797 
15798 	enoff_sec = NULL;
15799 
15800 	if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
15801 	    provider->dofpv_prenoffs != DOF_SECT_NONE &&
15802 	    (enoff_sec = dtrace_dof_sect(dof, DOF_SECT_PRENOFFS,
15803 	    provider->dofpv_prenoffs)) == NULL)
15804 		return (-1);
15805 
15806 	strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
15807 
15808 	if (provider->dofpv_name >= str_sec->dofs_size ||
15809 	    strlen(strtab + provider->dofpv_name) >= DTRACE_PROVNAMELEN) {
15810 		dtrace_dof_error(dof, "invalid provider name");
15811 		return (-1);
15812 	}
15813 
15814 	if (prb_sec->dofs_entsize == 0 ||
15815 	    prb_sec->dofs_entsize > prb_sec->dofs_size) {
15816 		dtrace_dof_error(dof, "invalid entry size");
15817 		return (-1);
15818 	}
15819 
15820 	if (prb_sec->dofs_entsize & (sizeof (uintptr_t) - 1)) {
15821 		dtrace_dof_error(dof, "misaligned entry size");
15822 		return (-1);
15823 	}
15824 
15825 	if (off_sec->dofs_entsize != sizeof (uint32_t)) {
15826 		dtrace_dof_error(dof, "invalid entry size");
15827 		return (-1);
15828 	}
15829 
15830 	if (off_sec->dofs_offset & (sizeof (uint32_t) - 1)) {
15831 		dtrace_dof_error(dof, "misaligned section offset");
15832 		return (-1);
15833 	}
15834 
15835 	if (arg_sec->dofs_entsize != sizeof (uint8_t)) {
15836 		dtrace_dof_error(dof, "invalid entry size");
15837 		return (-1);
15838 	}
15839 
15840 	arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
15841 
15842 	nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
15843 
15844 	/*
15845 	 * Take a pass through the probes to check for errors.
15846 	 */
15847 	for (j = 0; j < nprobes; j++) {
15848 		probe = (dof_probe_t *)(uintptr_t)(daddr +
15849 		    prb_sec->dofs_offset + j * prb_sec->dofs_entsize);
15850 
15851 		if (probe->dofpr_func >= str_sec->dofs_size) {
15852 			dtrace_dof_error(dof, "invalid function name");
15853 			return (-1);
15854 		}
15855 
15856 		if (strlen(strtab + probe->dofpr_func) >= DTRACE_FUNCNAMELEN) {
15857 			dtrace_dof_error(dof, "function name too long");
15858 			return (-1);
15859 		}
15860 
15861 		if (probe->dofpr_name >= str_sec->dofs_size ||
15862 		    strlen(strtab + probe->dofpr_name) >= DTRACE_NAMELEN) {
15863 			dtrace_dof_error(dof, "invalid probe name");
15864 			return (-1);
15865 		}
15866 
15867 		/*
15868 		 * The offset count must not wrap the index, and the offsets
15869 		 * must also not overflow the section's data.
15870 		 */
15871 		if (probe->dofpr_offidx + probe->dofpr_noffs <
15872 		    probe->dofpr_offidx ||
15873 		    (probe->dofpr_offidx + probe->dofpr_noffs) *
15874 		    off_sec->dofs_entsize > off_sec->dofs_size) {
15875 			dtrace_dof_error(dof, "invalid probe offset");
15876 			return (-1);
15877 		}
15878 
15879 		if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1) {
15880 			/*
15881 			 * If there's no is-enabled offset section, make sure
15882 			 * there aren't any is-enabled offsets. Otherwise
15883 			 * perform the same checks as for probe offsets
15884 			 * (immediately above).
15885 			 */
15886 			if (enoff_sec == NULL) {
15887 				if (probe->dofpr_enoffidx != 0 ||
15888 				    probe->dofpr_nenoffs != 0) {
15889 					dtrace_dof_error(dof, "is-enabled "
15890 					    "offsets with null section");
15891 					return (-1);
15892 				}
15893 			} else if (probe->dofpr_enoffidx +
15894 			    probe->dofpr_nenoffs < probe->dofpr_enoffidx ||
15895 			    (probe->dofpr_enoffidx + probe->dofpr_nenoffs) *
15896 			    enoff_sec->dofs_entsize > enoff_sec->dofs_size) {
15897 				dtrace_dof_error(dof, "invalid is-enabled "
15898 				    "offset");
15899 				return (-1);
15900 			}
15901 
15902 			if (probe->dofpr_noffs + probe->dofpr_nenoffs == 0) {
15903 				dtrace_dof_error(dof, "zero probe and "
15904 				    "is-enabled offsets");
15905 				return (-1);
15906 			}
15907 		} else if (probe->dofpr_noffs == 0) {
15908 			dtrace_dof_error(dof, "zero probe offsets");
15909 			return (-1);
15910 		}
15911 
15912 		if (probe->dofpr_argidx + probe->dofpr_xargc <
15913 		    probe->dofpr_argidx ||
15914 		    (probe->dofpr_argidx + probe->dofpr_xargc) *
15915 		    arg_sec->dofs_entsize > arg_sec->dofs_size) {
15916 			dtrace_dof_error(dof, "invalid args");
15917 			return (-1);
15918 		}
15919 
15920 		typeidx = probe->dofpr_nargv;
15921 		typestr = strtab + probe->dofpr_nargv;
15922 		for (k = 0; k < probe->dofpr_nargc; k++) {
15923 			if (typeidx >= str_sec->dofs_size) {
15924 				dtrace_dof_error(dof, "bad "
15925 				    "native argument type");
15926 				return (-1);
15927 			}
15928 
15929 			typesz = strlen(typestr) + 1;
15930 			if (typesz > DTRACE_ARGTYPELEN) {
15931 				dtrace_dof_error(dof, "native "
15932 				    "argument type too long");
15933 				return (-1);
15934 			}
15935 			typeidx += typesz;
15936 			typestr += typesz;
15937 		}
15938 
15939 		typeidx = probe->dofpr_xargv;
15940 		typestr = strtab + probe->dofpr_xargv;
15941 		for (k = 0; k < probe->dofpr_xargc; k++) {
15942 			if (arg[probe->dofpr_argidx + k] > probe->dofpr_nargc) {
15943 				dtrace_dof_error(dof, "bad "
15944 				    "native argument index");
15945 				return (-1);
15946 			}
15947 
15948 			if (typeidx >= str_sec->dofs_size) {
15949 				dtrace_dof_error(dof, "bad "
15950 				    "translated argument type");
15951 				return (-1);
15952 			}
15953 
15954 			typesz = strlen(typestr) + 1;
15955 			if (typesz > DTRACE_ARGTYPELEN) {
15956 				dtrace_dof_error(dof, "translated argument "
15957 				    "type too long");
15958 				return (-1);
15959 			}
15960 
15961 			typeidx += typesz;
15962 			typestr += typesz;
15963 		}
15964 	}
15965 
15966 	return (0);
15967 }
15968 
15969 static int
15970 dtrace_helper_slurp(dof_hdr_t *dof, dof_helper_t *dhp)
15971 {
15972 	dtrace_helpers_t *help;
15973 	dtrace_vstate_t *vstate;
15974 	dtrace_enabling_t *enab = NULL;
15975 	proc_t *p = curproc;
15976 	int i, gen, rv, nhelpers = 0, nprovs = 0, destroy = 1;
15977 	uintptr_t daddr = (uintptr_t)dof;
15978 
15979 	ASSERT(MUTEX_HELD(&dtrace_lock));
15980 
15981 #ifdef __FreeBSD__
15982 	if (dhp->dofhp_pid != p->p_pid) {
15983 		if ((p = pfind(dhp->dofhp_pid)) == NULL)
15984 			return (-1);
15985 		if (!P_SHOULDSTOP(p) ||
15986 		    (p->p_flag & P_TRACED) == 0 ||
15987 		    p->p_pptr->p_pid != curproc->p_pid) {
15988 			PROC_UNLOCK(p);
15989 			return (-1);
15990 		}
15991 		PROC_UNLOCK(p);
15992 	}
15993 #endif
15994 
15995 	if ((help = p->p_dtrace_helpers) == NULL)
15996 		help = dtrace_helpers_create(p);
15997 
15998 	vstate = &help->dthps_vstate;
15999 
16000 	if ((rv = dtrace_dof_slurp(dof, vstate, NULL, &enab,
16001 	    dhp != NULL ? dhp->dofhp_addr : 0, B_FALSE)) != 0) {
16002 		dtrace_dof_destroy(dof);
16003 		return (rv);
16004 	}
16005 
16006 	/*
16007 	 * Look for helper providers and validate their descriptions.
16008 	 */
16009 	if (dhp != NULL) {
16010 		for (i = 0; i < dof->dofh_secnum; i++) {
16011 			dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
16012 			    dof->dofh_secoff + i * dof->dofh_secsize);
16013 
16014 			if (sec->dofs_type != DOF_SECT_PROVIDER)
16015 				continue;
16016 
16017 			if (dtrace_helper_provider_validate(dof, sec) != 0) {
16018 				dtrace_enabling_destroy(enab);
16019 				dtrace_dof_destroy(dof);
16020 				return (-1);
16021 			}
16022 
16023 			nprovs++;
16024 		}
16025 	}
16026 
16027 	/*
16028 	 * Now we need to walk through the ECB descriptions in the enabling.
16029 	 */
16030 	for (i = 0; i < enab->dten_ndesc; i++) {
16031 		dtrace_ecbdesc_t *ep = enab->dten_desc[i];
16032 		dtrace_probedesc_t *desc = &ep->dted_probe;
16033 
16034 		if (strcmp(desc->dtpd_provider, "dtrace") != 0)
16035 			continue;
16036 
16037 		if (strcmp(desc->dtpd_mod, "helper") != 0)
16038 			continue;
16039 
16040 		if (strcmp(desc->dtpd_func, "ustack") != 0)
16041 			continue;
16042 
16043 		if ((rv = dtrace_helper_action_add(DTRACE_HELPER_ACTION_USTACK,
16044 		    ep, help)) != 0) {
16045 			/*
16046 			 * Adding this helper action failed -- we are now going
16047 			 * to rip out the entire generation and return failure.
16048 			 */
16049 			(void) dtrace_helper_destroygen(help,
16050 			    help->dthps_generation);
16051 			dtrace_enabling_destroy(enab);
16052 			dtrace_dof_destroy(dof);
16053 			return (-1);
16054 		}
16055 
16056 		nhelpers++;
16057 	}
16058 
16059 	if (nhelpers < enab->dten_ndesc)
16060 		dtrace_dof_error(dof, "unmatched helpers");
16061 
16062 	gen = help->dthps_generation++;
16063 	dtrace_enabling_destroy(enab);
16064 
16065 	if (dhp != NULL && nprovs > 0) {
16066 		dhp->dofhp_dof = (uint64_t)(uintptr_t)dof;
16067 		if (dtrace_helper_provider_add(dhp, help, gen) == 0) {
16068 			mutex_exit(&dtrace_lock);
16069 			dtrace_helper_provider_register(p, help, dhp);
16070 			mutex_enter(&dtrace_lock);
16071 
16072 			destroy = 0;
16073 		}
16074 	}
16075 
16076 	if (destroy)
16077 		dtrace_dof_destroy(dof);
16078 
16079 	return (gen);
16080 }
16081 
16082 static dtrace_helpers_t *
16083 dtrace_helpers_create(proc_t *p)
16084 {
16085 	dtrace_helpers_t *help;
16086 
16087 	ASSERT(MUTEX_HELD(&dtrace_lock));
16088 	ASSERT(p->p_dtrace_helpers == NULL);
16089 
16090 	help = kmem_zalloc(sizeof (dtrace_helpers_t), KM_SLEEP);
16091 	help->dthps_actions = kmem_zalloc(sizeof (dtrace_helper_action_t *) *
16092 	    DTRACE_NHELPER_ACTIONS, KM_SLEEP);
16093 
16094 	p->p_dtrace_helpers = help;
16095 	dtrace_helpers++;
16096 
16097 	return (help);
16098 }
16099 
16100 #ifdef illumos
16101 static
16102 #endif
16103 void
16104 dtrace_helpers_destroy(proc_t *p)
16105 {
16106 	dtrace_helpers_t *help;
16107 	dtrace_vstate_t *vstate;
16108 #ifdef illumos
16109 	proc_t *p = curproc;
16110 #endif
16111 	int i;
16112 
16113 	mutex_enter(&dtrace_lock);
16114 
16115 	ASSERT(p->p_dtrace_helpers != NULL);
16116 	ASSERT(dtrace_helpers > 0);
16117 
16118 	help = p->p_dtrace_helpers;
16119 	vstate = &help->dthps_vstate;
16120 
16121 	/*
16122 	 * We're now going to lose the help from this process.
16123 	 */
16124 	p->p_dtrace_helpers = NULL;
16125 	dtrace_sync();
16126 
16127 	/*
16128 	 * Destory the helper actions.
16129 	 */
16130 	for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
16131 		dtrace_helper_action_t *h, *next;
16132 
16133 		for (h = help->dthps_actions[i]; h != NULL; h = next) {
16134 			next = h->dtha_next;
16135 			dtrace_helper_action_destroy(h, vstate);
16136 			h = next;
16137 		}
16138 	}
16139 
16140 	mutex_exit(&dtrace_lock);
16141 
16142 	/*
16143 	 * Destroy the helper providers.
16144 	 */
16145 	if (help->dthps_maxprovs > 0) {
16146 		mutex_enter(&dtrace_meta_lock);
16147 		if (dtrace_meta_pid != NULL) {
16148 			ASSERT(dtrace_deferred_pid == NULL);
16149 
16150 			for (i = 0; i < help->dthps_nprovs; i++) {
16151 				dtrace_helper_provider_remove(
16152 				    &help->dthps_provs[i]->dthp_prov, p->p_pid);
16153 			}
16154 		} else {
16155 			mutex_enter(&dtrace_lock);
16156 			ASSERT(help->dthps_deferred == 0 ||
16157 			    help->dthps_next != NULL ||
16158 			    help->dthps_prev != NULL ||
16159 			    help == dtrace_deferred_pid);
16160 
16161 			/*
16162 			 * Remove the helper from the deferred list.
16163 			 */
16164 			if (help->dthps_next != NULL)
16165 				help->dthps_next->dthps_prev = help->dthps_prev;
16166 			if (help->dthps_prev != NULL)
16167 				help->dthps_prev->dthps_next = help->dthps_next;
16168 			if (dtrace_deferred_pid == help) {
16169 				dtrace_deferred_pid = help->dthps_next;
16170 				ASSERT(help->dthps_prev == NULL);
16171 			}
16172 
16173 			mutex_exit(&dtrace_lock);
16174 		}
16175 
16176 		mutex_exit(&dtrace_meta_lock);
16177 
16178 		for (i = 0; i < help->dthps_nprovs; i++) {
16179 			dtrace_helper_provider_destroy(help->dthps_provs[i]);
16180 		}
16181 
16182 		kmem_free(help->dthps_provs, help->dthps_maxprovs *
16183 		    sizeof (dtrace_helper_provider_t *));
16184 	}
16185 
16186 	mutex_enter(&dtrace_lock);
16187 
16188 	dtrace_vstate_fini(&help->dthps_vstate);
16189 	kmem_free(help->dthps_actions,
16190 	    sizeof (dtrace_helper_action_t *) * DTRACE_NHELPER_ACTIONS);
16191 	kmem_free(help, sizeof (dtrace_helpers_t));
16192 
16193 	--dtrace_helpers;
16194 	mutex_exit(&dtrace_lock);
16195 }
16196 
16197 #ifdef illumos
16198 static
16199 #endif
16200 void
16201 dtrace_helpers_duplicate(proc_t *from, proc_t *to)
16202 {
16203 	dtrace_helpers_t *help, *newhelp;
16204 	dtrace_helper_action_t *helper, *new, *last;
16205 	dtrace_difo_t *dp;
16206 	dtrace_vstate_t *vstate;
16207 	int i, j, sz, hasprovs = 0;
16208 
16209 	mutex_enter(&dtrace_lock);
16210 	ASSERT(from->p_dtrace_helpers != NULL);
16211 	ASSERT(dtrace_helpers > 0);
16212 
16213 	help = from->p_dtrace_helpers;
16214 	newhelp = dtrace_helpers_create(to);
16215 	ASSERT(to->p_dtrace_helpers != NULL);
16216 
16217 	newhelp->dthps_generation = help->dthps_generation;
16218 	vstate = &newhelp->dthps_vstate;
16219 
16220 	/*
16221 	 * Duplicate the helper actions.
16222 	 */
16223 	for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
16224 		if ((helper = help->dthps_actions[i]) == NULL)
16225 			continue;
16226 
16227 		for (last = NULL; helper != NULL; helper = helper->dtha_next) {
16228 			new = kmem_zalloc(sizeof (dtrace_helper_action_t),
16229 			    KM_SLEEP);
16230 			new->dtha_generation = helper->dtha_generation;
16231 
16232 			if ((dp = helper->dtha_predicate) != NULL) {
16233 				dp = dtrace_difo_duplicate(dp, vstate);
16234 				new->dtha_predicate = dp;
16235 			}
16236 
16237 			new->dtha_nactions = helper->dtha_nactions;
16238 			sz = sizeof (dtrace_difo_t *) * new->dtha_nactions;
16239 			new->dtha_actions = kmem_alloc(sz, KM_SLEEP);
16240 
16241 			for (j = 0; j < new->dtha_nactions; j++) {
16242 				dtrace_difo_t *dp = helper->dtha_actions[j];
16243 
16244 				ASSERT(dp != NULL);
16245 				dp = dtrace_difo_duplicate(dp, vstate);
16246 				new->dtha_actions[j] = dp;
16247 			}
16248 
16249 			if (last != NULL) {
16250 				last->dtha_next = new;
16251 			} else {
16252 				newhelp->dthps_actions[i] = new;
16253 			}
16254 
16255 			last = new;
16256 		}
16257 	}
16258 
16259 	/*
16260 	 * Duplicate the helper providers and register them with the
16261 	 * DTrace framework.
16262 	 */
16263 	if (help->dthps_nprovs > 0) {
16264 		newhelp->dthps_nprovs = help->dthps_nprovs;
16265 		newhelp->dthps_maxprovs = help->dthps_nprovs;
16266 		newhelp->dthps_provs = kmem_alloc(newhelp->dthps_nprovs *
16267 		    sizeof (dtrace_helper_provider_t *), KM_SLEEP);
16268 		for (i = 0; i < newhelp->dthps_nprovs; i++) {
16269 			newhelp->dthps_provs[i] = help->dthps_provs[i];
16270 			newhelp->dthps_provs[i]->dthp_ref++;
16271 		}
16272 
16273 		hasprovs = 1;
16274 	}
16275 
16276 	mutex_exit(&dtrace_lock);
16277 
16278 	if (hasprovs)
16279 		dtrace_helper_provider_register(to, newhelp, NULL);
16280 }
16281 
16282 /*
16283  * DTrace Hook Functions
16284  */
16285 static void
16286 dtrace_module_loaded(modctl_t *ctl)
16287 {
16288 	dtrace_provider_t *prv;
16289 
16290 	mutex_enter(&dtrace_provider_lock);
16291 #ifdef illumos
16292 	mutex_enter(&mod_lock);
16293 #endif
16294 
16295 #ifdef illumos
16296 	ASSERT(ctl->mod_busy);
16297 #endif
16298 
16299 	/*
16300 	 * We're going to call each providers per-module provide operation
16301 	 * specifying only this module.
16302 	 */
16303 	for (prv = dtrace_provider; prv != NULL; prv = prv->dtpv_next)
16304 		prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
16305 
16306 #ifdef illumos
16307 	mutex_exit(&mod_lock);
16308 #endif
16309 	mutex_exit(&dtrace_provider_lock);
16310 
16311 	/*
16312 	 * If we have any retained enablings, we need to match against them.
16313 	 * Enabling probes requires that cpu_lock be held, and we cannot hold
16314 	 * cpu_lock here -- it is legal for cpu_lock to be held when loading a
16315 	 * module.  (In particular, this happens when loading scheduling
16316 	 * classes.)  So if we have any retained enablings, we need to dispatch
16317 	 * our task queue to do the match for us.
16318 	 */
16319 	mutex_enter(&dtrace_lock);
16320 
16321 	if (dtrace_retained == NULL) {
16322 		mutex_exit(&dtrace_lock);
16323 		return;
16324 	}
16325 
16326 	(void) taskq_dispatch(dtrace_taskq,
16327 	    (task_func_t *)dtrace_enabling_matchall, NULL, TQ_SLEEP);
16328 
16329 	mutex_exit(&dtrace_lock);
16330 
16331 	/*
16332 	 * And now, for a little heuristic sleaze:  in general, we want to
16333 	 * match modules as soon as they load.  However, we cannot guarantee
16334 	 * this, because it would lead us to the lock ordering violation
16335 	 * outlined above.  The common case, of course, is that cpu_lock is
16336 	 * _not_ held -- so we delay here for a clock tick, hoping that that's
16337 	 * long enough for the task queue to do its work.  If it's not, it's
16338 	 * not a serious problem -- it just means that the module that we
16339 	 * just loaded may not be immediately instrumentable.
16340 	 */
16341 	delay(1);
16342 }
16343 
16344 static void
16345 #ifdef illumos
16346 dtrace_module_unloaded(modctl_t *ctl)
16347 #else
16348 dtrace_module_unloaded(modctl_t *ctl, int *error)
16349 #endif
16350 {
16351 	dtrace_probe_t template, *probe, *first, *next;
16352 	dtrace_provider_t *prov;
16353 #ifndef illumos
16354 	char modname[DTRACE_MODNAMELEN];
16355 	size_t len;
16356 #endif
16357 
16358 #ifdef illumos
16359 	template.dtpr_mod = ctl->mod_modname;
16360 #else
16361 	/* Handle the fact that ctl->filename may end in ".ko". */
16362 	strlcpy(modname, ctl->filename, sizeof(modname));
16363 	len = strlen(ctl->filename);
16364 	if (len > 3 && strcmp(modname + len - 3, ".ko") == 0)
16365 		modname[len - 3] = '\0';
16366 	template.dtpr_mod = modname;
16367 #endif
16368 
16369 	mutex_enter(&dtrace_provider_lock);
16370 #ifdef illumos
16371 	mutex_enter(&mod_lock);
16372 #endif
16373 	mutex_enter(&dtrace_lock);
16374 
16375 #ifndef illumos
16376 	if (ctl->nenabled > 0) {
16377 		/* Don't allow unloads if a probe is enabled. */
16378 		mutex_exit(&dtrace_provider_lock);
16379 		mutex_exit(&dtrace_lock);
16380 		*error = -1;
16381 		printf(
16382 	"kldunload: attempt to unload module that has DTrace probes enabled\n");
16383 		return;
16384 	}
16385 #endif
16386 
16387 	if (dtrace_bymod == NULL) {
16388 		/*
16389 		 * The DTrace module is loaded (obviously) but not attached;
16390 		 * we don't have any work to do.
16391 		 */
16392 		mutex_exit(&dtrace_provider_lock);
16393 #ifdef illumos
16394 		mutex_exit(&mod_lock);
16395 #endif
16396 		mutex_exit(&dtrace_lock);
16397 		return;
16398 	}
16399 
16400 	for (probe = first = dtrace_hash_lookup(dtrace_bymod, &template);
16401 	    probe != NULL; probe = probe->dtpr_nextmod) {
16402 		if (probe->dtpr_ecb != NULL) {
16403 			mutex_exit(&dtrace_provider_lock);
16404 #ifdef illumos
16405 			mutex_exit(&mod_lock);
16406 #endif
16407 			mutex_exit(&dtrace_lock);
16408 
16409 			/*
16410 			 * This shouldn't _actually_ be possible -- we're
16411 			 * unloading a module that has an enabled probe in it.
16412 			 * (It's normally up to the provider to make sure that
16413 			 * this can't happen.)  However, because dtps_enable()
16414 			 * doesn't have a failure mode, there can be an
16415 			 * enable/unload race.  Upshot:  we don't want to
16416 			 * assert, but we're not going to disable the
16417 			 * probe, either.
16418 			 */
16419 			if (dtrace_err_verbose) {
16420 #ifdef illumos
16421 				cmn_err(CE_WARN, "unloaded module '%s' had "
16422 				    "enabled probes", ctl->mod_modname);
16423 #else
16424 				cmn_err(CE_WARN, "unloaded module '%s' had "
16425 				    "enabled probes", modname);
16426 #endif
16427 			}
16428 
16429 			return;
16430 		}
16431 	}
16432 
16433 	probe = first;
16434 
16435 	for (first = NULL; probe != NULL; probe = next) {
16436 		ASSERT(dtrace_probes[probe->dtpr_id - 1] == probe);
16437 
16438 		dtrace_probes[probe->dtpr_id - 1] = NULL;
16439 
16440 		next = probe->dtpr_nextmod;
16441 		dtrace_hash_remove(dtrace_bymod, probe);
16442 		dtrace_hash_remove(dtrace_byfunc, probe);
16443 		dtrace_hash_remove(dtrace_byname, probe);
16444 
16445 		if (first == NULL) {
16446 			first = probe;
16447 			probe->dtpr_nextmod = NULL;
16448 		} else {
16449 			probe->dtpr_nextmod = first;
16450 			first = probe;
16451 		}
16452 	}
16453 
16454 	/*
16455 	 * We've removed all of the module's probes from the hash chains and
16456 	 * from the probe array.  Now issue a dtrace_sync() to be sure that
16457 	 * everyone has cleared out from any probe array processing.
16458 	 */
16459 	dtrace_sync();
16460 
16461 	for (probe = first; probe != NULL; probe = first) {
16462 		first = probe->dtpr_nextmod;
16463 		prov = probe->dtpr_provider;
16464 		prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, probe->dtpr_id,
16465 		    probe->dtpr_arg);
16466 		kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
16467 		kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
16468 		kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
16469 #ifdef illumos
16470 		vmem_free(dtrace_arena, (void *)(uintptr_t)probe->dtpr_id, 1);
16471 #else
16472 		free_unr(dtrace_arena, probe->dtpr_id);
16473 #endif
16474 		kmem_free(probe, sizeof (dtrace_probe_t));
16475 	}
16476 
16477 	mutex_exit(&dtrace_lock);
16478 #ifdef illumos
16479 	mutex_exit(&mod_lock);
16480 #endif
16481 	mutex_exit(&dtrace_provider_lock);
16482 }
16483 
16484 #ifndef illumos
16485 static void
16486 dtrace_kld_load(void *arg __unused, linker_file_t lf)
16487 {
16488 
16489 	dtrace_module_loaded(lf);
16490 }
16491 
16492 static void
16493 dtrace_kld_unload_try(void *arg __unused, linker_file_t lf, int *error)
16494 {
16495 
16496 	if (*error != 0)
16497 		/* We already have an error, so don't do anything. */
16498 		return;
16499 	dtrace_module_unloaded(lf, error);
16500 }
16501 #endif
16502 
16503 #ifdef illumos
16504 static void
16505 dtrace_suspend(void)
16506 {
16507 	dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_suspend));
16508 }
16509 
16510 static void
16511 dtrace_resume(void)
16512 {
16513 	dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_resume));
16514 }
16515 #endif
16516 
16517 static int
16518 dtrace_cpu_setup(cpu_setup_t what, processorid_t cpu)
16519 {
16520 	ASSERT(MUTEX_HELD(&cpu_lock));
16521 	mutex_enter(&dtrace_lock);
16522 
16523 	switch (what) {
16524 	case CPU_CONFIG: {
16525 		dtrace_state_t *state;
16526 		dtrace_optval_t *opt, rs, c;
16527 
16528 		/*
16529 		 * For now, we only allocate a new buffer for anonymous state.
16530 		 */
16531 		if ((state = dtrace_anon.dta_state) == NULL)
16532 			break;
16533 
16534 		if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
16535 			break;
16536 
16537 		opt = state->dts_options;
16538 		c = opt[DTRACEOPT_CPU];
16539 
16540 		if (c != DTRACE_CPUALL && c != DTRACEOPT_UNSET && c != cpu)
16541 			break;
16542 
16543 		/*
16544 		 * Regardless of what the actual policy is, we're going to
16545 		 * temporarily set our resize policy to be manual.  We're
16546 		 * also going to temporarily set our CPU option to denote
16547 		 * the newly configured CPU.
16548 		 */
16549 		rs = opt[DTRACEOPT_BUFRESIZE];
16550 		opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_MANUAL;
16551 		opt[DTRACEOPT_CPU] = (dtrace_optval_t)cpu;
16552 
16553 		(void) dtrace_state_buffers(state);
16554 
16555 		opt[DTRACEOPT_BUFRESIZE] = rs;
16556 		opt[DTRACEOPT_CPU] = c;
16557 
16558 		break;
16559 	}
16560 
16561 	case CPU_UNCONFIG:
16562 		/*
16563 		 * We don't free the buffer in the CPU_UNCONFIG case.  (The
16564 		 * buffer will be freed when the consumer exits.)
16565 		 */
16566 		break;
16567 
16568 	default:
16569 		break;
16570 	}
16571 
16572 	mutex_exit(&dtrace_lock);
16573 	return (0);
16574 }
16575 
16576 #ifdef illumos
16577 static void
16578 dtrace_cpu_setup_initial(processorid_t cpu)
16579 {
16580 	(void) dtrace_cpu_setup(CPU_CONFIG, cpu);
16581 }
16582 #endif
16583 
16584 static void
16585 dtrace_toxrange_add(uintptr_t base, uintptr_t limit)
16586 {
16587 	if (dtrace_toxranges >= dtrace_toxranges_max) {
16588 		int osize, nsize;
16589 		dtrace_toxrange_t *range;
16590 
16591 		osize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
16592 
16593 		if (osize == 0) {
16594 			ASSERT(dtrace_toxrange == NULL);
16595 			ASSERT(dtrace_toxranges_max == 0);
16596 			dtrace_toxranges_max = 1;
16597 		} else {
16598 			dtrace_toxranges_max <<= 1;
16599 		}
16600 
16601 		nsize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
16602 		range = kmem_zalloc(nsize, KM_SLEEP);
16603 
16604 		if (dtrace_toxrange != NULL) {
16605 			ASSERT(osize != 0);
16606 			bcopy(dtrace_toxrange, range, osize);
16607 			kmem_free(dtrace_toxrange, osize);
16608 		}
16609 
16610 		dtrace_toxrange = range;
16611 	}
16612 
16613 	ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_base == 0);
16614 	ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_limit == 0);
16615 
16616 	dtrace_toxrange[dtrace_toxranges].dtt_base = base;
16617 	dtrace_toxrange[dtrace_toxranges].dtt_limit = limit;
16618 	dtrace_toxranges++;
16619 }
16620 
16621 static void
16622 dtrace_getf_barrier()
16623 {
16624 #ifdef illumos
16625 	/*
16626 	 * When we have unprivileged (that is, non-DTRACE_CRV_KERNEL) enablings
16627 	 * that contain calls to getf(), this routine will be called on every
16628 	 * closef() before either the underlying vnode is released or the
16629 	 * file_t itself is freed.  By the time we are here, it is essential
16630 	 * that the file_t can no longer be accessed from a call to getf()
16631 	 * in probe context -- that assures that a dtrace_sync() can be used
16632 	 * to clear out any enablings referring to the old structures.
16633 	 */
16634 	if (curthread->t_procp->p_zone->zone_dtrace_getf != 0 ||
16635 	    kcred->cr_zone->zone_dtrace_getf != 0)
16636 		dtrace_sync();
16637 #endif
16638 }
16639 
16640 /*
16641  * DTrace Driver Cookbook Functions
16642  */
16643 #ifdef illumos
16644 /*ARGSUSED*/
16645 static int
16646 dtrace_attach(dev_info_t *devi, ddi_attach_cmd_t cmd)
16647 {
16648 	dtrace_provider_id_t id;
16649 	dtrace_state_t *state = NULL;
16650 	dtrace_enabling_t *enab;
16651 
16652 	mutex_enter(&cpu_lock);
16653 	mutex_enter(&dtrace_provider_lock);
16654 	mutex_enter(&dtrace_lock);
16655 
16656 	if (ddi_soft_state_init(&dtrace_softstate,
16657 	    sizeof (dtrace_state_t), 0) != 0) {
16658 		cmn_err(CE_NOTE, "/dev/dtrace failed to initialize soft state");
16659 		mutex_exit(&cpu_lock);
16660 		mutex_exit(&dtrace_provider_lock);
16661 		mutex_exit(&dtrace_lock);
16662 		return (DDI_FAILURE);
16663 	}
16664 
16665 	if (ddi_create_minor_node(devi, DTRACEMNR_DTRACE, S_IFCHR,
16666 	    DTRACEMNRN_DTRACE, DDI_PSEUDO, NULL) == DDI_FAILURE ||
16667 	    ddi_create_minor_node(devi, DTRACEMNR_HELPER, S_IFCHR,
16668 	    DTRACEMNRN_HELPER, DDI_PSEUDO, NULL) == DDI_FAILURE) {
16669 		cmn_err(CE_NOTE, "/dev/dtrace couldn't create minor nodes");
16670 		ddi_remove_minor_node(devi, NULL);
16671 		ddi_soft_state_fini(&dtrace_softstate);
16672 		mutex_exit(&cpu_lock);
16673 		mutex_exit(&dtrace_provider_lock);
16674 		mutex_exit(&dtrace_lock);
16675 		return (DDI_FAILURE);
16676 	}
16677 
16678 	ddi_report_dev(devi);
16679 	dtrace_devi = devi;
16680 
16681 	dtrace_modload = dtrace_module_loaded;
16682 	dtrace_modunload = dtrace_module_unloaded;
16683 	dtrace_cpu_init = dtrace_cpu_setup_initial;
16684 	dtrace_helpers_cleanup = dtrace_helpers_destroy;
16685 	dtrace_helpers_fork = dtrace_helpers_duplicate;
16686 	dtrace_cpustart_init = dtrace_suspend;
16687 	dtrace_cpustart_fini = dtrace_resume;
16688 	dtrace_debugger_init = dtrace_suspend;
16689 	dtrace_debugger_fini = dtrace_resume;
16690 
16691 	register_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
16692 
16693 	ASSERT(MUTEX_HELD(&cpu_lock));
16694 
16695 	dtrace_arena = vmem_create("dtrace", (void *)1, UINT32_MAX, 1,
16696 	    NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
16697 	dtrace_minor = vmem_create("dtrace_minor", (void *)DTRACEMNRN_CLONE,
16698 	    UINT32_MAX - DTRACEMNRN_CLONE, 1, NULL, NULL, NULL, 0,
16699 	    VM_SLEEP | VMC_IDENTIFIER);
16700 	dtrace_taskq = taskq_create("dtrace_taskq", 1, maxclsyspri,
16701 	    1, INT_MAX, 0);
16702 
16703 	dtrace_state_cache = kmem_cache_create("dtrace_state_cache",
16704 	    sizeof (dtrace_dstate_percpu_t) * NCPU, DTRACE_STATE_ALIGN,
16705 	    NULL, NULL, NULL, NULL, NULL, 0);
16706 
16707 	ASSERT(MUTEX_HELD(&cpu_lock));
16708 	dtrace_bymod = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_mod),
16709 	    offsetof(dtrace_probe_t, dtpr_nextmod),
16710 	    offsetof(dtrace_probe_t, dtpr_prevmod));
16711 
16712 	dtrace_byfunc = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_func),
16713 	    offsetof(dtrace_probe_t, dtpr_nextfunc),
16714 	    offsetof(dtrace_probe_t, dtpr_prevfunc));
16715 
16716 	dtrace_byname = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_name),
16717 	    offsetof(dtrace_probe_t, dtpr_nextname),
16718 	    offsetof(dtrace_probe_t, dtpr_prevname));
16719 
16720 	if (dtrace_retain_max < 1) {
16721 		cmn_err(CE_WARN, "illegal value (%lu) for dtrace_retain_max; "
16722 		    "setting to 1", dtrace_retain_max);
16723 		dtrace_retain_max = 1;
16724 	}
16725 
16726 	/*
16727 	 * Now discover our toxic ranges.
16728 	 */
16729 	dtrace_toxic_ranges(dtrace_toxrange_add);
16730 
16731 	/*
16732 	 * Before we register ourselves as a provider to our own framework,
16733 	 * we would like to assert that dtrace_provider is NULL -- but that's
16734 	 * not true if we were loaded as a dependency of a DTrace provider.
16735 	 * Once we've registered, we can assert that dtrace_provider is our
16736 	 * pseudo provider.
16737 	 */
16738 	(void) dtrace_register("dtrace", &dtrace_provider_attr,
16739 	    DTRACE_PRIV_NONE, 0, &dtrace_provider_ops, NULL, &id);
16740 
16741 	ASSERT(dtrace_provider != NULL);
16742 	ASSERT((dtrace_provider_id_t)dtrace_provider == id);
16743 
16744 	dtrace_probeid_begin = dtrace_probe_create((dtrace_provider_id_t)
16745 	    dtrace_provider, NULL, NULL, "BEGIN", 0, NULL);
16746 	dtrace_probeid_end = dtrace_probe_create((dtrace_provider_id_t)
16747 	    dtrace_provider, NULL, NULL, "END", 0, NULL);
16748 	dtrace_probeid_error = dtrace_probe_create((dtrace_provider_id_t)
16749 	    dtrace_provider, NULL, NULL, "ERROR", 1, NULL);
16750 
16751 	dtrace_anon_property();
16752 	mutex_exit(&cpu_lock);
16753 
16754 	/*
16755 	 * If there are already providers, we must ask them to provide their
16756 	 * probes, and then match any anonymous enabling against them.  Note
16757 	 * that there should be no other retained enablings at this time:
16758 	 * the only retained enablings at this time should be the anonymous
16759 	 * enabling.
16760 	 */
16761 	if (dtrace_anon.dta_enabling != NULL) {
16762 		ASSERT(dtrace_retained == dtrace_anon.dta_enabling);
16763 
16764 		dtrace_enabling_provide(NULL);
16765 		state = dtrace_anon.dta_state;
16766 
16767 		/*
16768 		 * We couldn't hold cpu_lock across the above call to
16769 		 * dtrace_enabling_provide(), but we must hold it to actually
16770 		 * enable the probes.  We have to drop all of our locks, pick
16771 		 * up cpu_lock, and regain our locks before matching the
16772 		 * retained anonymous enabling.
16773 		 */
16774 		mutex_exit(&dtrace_lock);
16775 		mutex_exit(&dtrace_provider_lock);
16776 
16777 		mutex_enter(&cpu_lock);
16778 		mutex_enter(&dtrace_provider_lock);
16779 		mutex_enter(&dtrace_lock);
16780 
16781 		if ((enab = dtrace_anon.dta_enabling) != NULL)
16782 			(void) dtrace_enabling_match(enab, NULL);
16783 
16784 		mutex_exit(&cpu_lock);
16785 	}
16786 
16787 	mutex_exit(&dtrace_lock);
16788 	mutex_exit(&dtrace_provider_lock);
16789 
16790 	if (state != NULL) {
16791 		/*
16792 		 * If we created any anonymous state, set it going now.
16793 		 */
16794 		(void) dtrace_state_go(state, &dtrace_anon.dta_beganon);
16795 	}
16796 
16797 	return (DDI_SUCCESS);
16798 }
16799 #endif	/* illumos */
16800 
16801 #ifndef illumos
16802 static void dtrace_dtr(void *);
16803 #endif
16804 
16805 /*ARGSUSED*/
16806 static int
16807 #ifdef illumos
16808 dtrace_open(dev_t *devp, int flag, int otyp, cred_t *cred_p)
16809 #else
16810 dtrace_open(struct cdev *dev, int oflags, int devtype, struct thread *td)
16811 #endif
16812 {
16813 	dtrace_state_t *state;
16814 	uint32_t priv;
16815 	uid_t uid;
16816 	zoneid_t zoneid;
16817 
16818 #ifdef illumos
16819 	if (getminor(*devp) == DTRACEMNRN_HELPER)
16820 		return (0);
16821 
16822 	/*
16823 	 * If this wasn't an open with the "helper" minor, then it must be
16824 	 * the "dtrace" minor.
16825 	 */
16826 	if (getminor(*devp) == DTRACEMNRN_DTRACE)
16827 		return (ENXIO);
16828 #else
16829 	cred_t *cred_p = NULL;
16830 	cred_p = dev->si_cred;
16831 
16832 	/*
16833 	 * If no DTRACE_PRIV_* bits are set in the credential, then the
16834 	 * caller lacks sufficient permission to do anything with DTrace.
16835 	 */
16836 	dtrace_cred2priv(cred_p, &priv, &uid, &zoneid);
16837 	if (priv == DTRACE_PRIV_NONE) {
16838 #endif
16839 
16840 		return (EACCES);
16841 	}
16842 
16843 	/*
16844 	 * Ask all providers to provide all their probes.
16845 	 */
16846 	mutex_enter(&dtrace_provider_lock);
16847 	dtrace_probe_provide(NULL, NULL);
16848 	mutex_exit(&dtrace_provider_lock);
16849 
16850 	mutex_enter(&cpu_lock);
16851 	mutex_enter(&dtrace_lock);
16852 	dtrace_opens++;
16853 	dtrace_membar_producer();
16854 
16855 #ifdef illumos
16856 	/*
16857 	 * If the kernel debugger is active (that is, if the kernel debugger
16858 	 * modified text in some way), we won't allow the open.
16859 	 */
16860 	if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
16861 		dtrace_opens--;
16862 		mutex_exit(&cpu_lock);
16863 		mutex_exit(&dtrace_lock);
16864 		return (EBUSY);
16865 	}
16866 
16867 	if (dtrace_helptrace_enable && dtrace_helptrace_buffer == NULL) {
16868 		/*
16869 		 * If DTrace helper tracing is enabled, we need to allocate the
16870 		 * trace buffer and initialize the values.
16871 		 */
16872 		dtrace_helptrace_buffer =
16873 		    kmem_zalloc(dtrace_helptrace_bufsize, KM_SLEEP);
16874 		dtrace_helptrace_next = 0;
16875 		dtrace_helptrace_wrapped = 0;
16876 		dtrace_helptrace_enable = 0;
16877 	}
16878 
16879 	state = dtrace_state_create(devp, cred_p);
16880 #else
16881 	state = dtrace_state_create(dev);
16882 	devfs_set_cdevpriv(state, dtrace_dtr);
16883 #endif
16884 
16885 	mutex_exit(&cpu_lock);
16886 
16887 	if (state == NULL) {
16888 #ifdef illumos
16889 		if (--dtrace_opens == 0 && dtrace_anon.dta_enabling == NULL)
16890 			(void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
16891 #else
16892 		--dtrace_opens;
16893 #endif
16894 		mutex_exit(&dtrace_lock);
16895 		return (EAGAIN);
16896 	}
16897 
16898 	mutex_exit(&dtrace_lock);
16899 
16900 	return (0);
16901 }
16902 
16903 /*ARGSUSED*/
16904 #ifdef illumos
16905 static int
16906 dtrace_close(dev_t dev, int flag, int otyp, cred_t *cred_p)
16907 #else
16908 static void
16909 dtrace_dtr(void *data)
16910 #endif
16911 {
16912 #ifdef illumos
16913 	minor_t minor = getminor(dev);
16914 	dtrace_state_t *state;
16915 #endif
16916 	dtrace_helptrace_t *buf = NULL;
16917 
16918 #ifdef illumos
16919 	if (minor == DTRACEMNRN_HELPER)
16920 		return (0);
16921 
16922 	state = ddi_get_soft_state(dtrace_softstate, minor);
16923 #else
16924 	dtrace_state_t *state = data;
16925 #endif
16926 
16927 	mutex_enter(&cpu_lock);
16928 	mutex_enter(&dtrace_lock);
16929 
16930 #ifdef illumos
16931 	if (state->dts_anon)
16932 #else
16933 	if (state != NULL && state->dts_anon)
16934 #endif
16935 	{
16936 		/*
16937 		 * There is anonymous state. Destroy that first.
16938 		 */
16939 		ASSERT(dtrace_anon.dta_state == NULL);
16940 		dtrace_state_destroy(state->dts_anon);
16941 	}
16942 
16943 	if (dtrace_helptrace_disable) {
16944 		/*
16945 		 * If we have been told to disable helper tracing, set the
16946 		 * buffer to NULL before calling into dtrace_state_destroy();
16947 		 * we take advantage of its dtrace_sync() to know that no
16948 		 * CPU is in probe context with enabled helper tracing
16949 		 * after it returns.
16950 		 */
16951 		buf = dtrace_helptrace_buffer;
16952 		dtrace_helptrace_buffer = NULL;
16953 	}
16954 
16955 #ifdef illumos
16956 	dtrace_state_destroy(state);
16957 #else
16958 	if (state != NULL) {
16959 		dtrace_state_destroy(state);
16960 		kmem_free(state, 0);
16961 	}
16962 #endif
16963 	ASSERT(dtrace_opens > 0);
16964 
16965 #ifdef illumos
16966 	/*
16967 	 * Only relinquish control of the kernel debugger interface when there
16968 	 * are no consumers and no anonymous enablings.
16969 	 */
16970 	if (--dtrace_opens == 0 && dtrace_anon.dta_enabling == NULL)
16971 		(void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
16972 #else
16973 	--dtrace_opens;
16974 #endif
16975 
16976 	if (buf != NULL) {
16977 		kmem_free(buf, dtrace_helptrace_bufsize);
16978 		dtrace_helptrace_disable = 0;
16979 	}
16980 
16981 	mutex_exit(&dtrace_lock);
16982 	mutex_exit(&cpu_lock);
16983 
16984 #ifdef illumos
16985 	return (0);
16986 #endif
16987 }
16988 
16989 #ifdef illumos
16990 /*ARGSUSED*/
16991 static int
16992 dtrace_ioctl_helper(int cmd, intptr_t arg, int *rv)
16993 {
16994 	int rval;
16995 	dof_helper_t help, *dhp = NULL;
16996 
16997 	switch (cmd) {
16998 	case DTRACEHIOC_ADDDOF:
16999 		if (copyin((void *)arg, &help, sizeof (help)) != 0) {
17000 			dtrace_dof_error(NULL, "failed to copyin DOF helper");
17001 			return (EFAULT);
17002 		}
17003 
17004 		dhp = &help;
17005 		arg = (intptr_t)help.dofhp_dof;
17006 		/*FALLTHROUGH*/
17007 
17008 	case DTRACEHIOC_ADD: {
17009 		dof_hdr_t *dof = dtrace_dof_copyin(arg, &rval);
17010 
17011 		if (dof == NULL)
17012 			return (rval);
17013 
17014 		mutex_enter(&dtrace_lock);
17015 
17016 		/*
17017 		 * dtrace_helper_slurp() takes responsibility for the dof --
17018 		 * it may free it now or it may save it and free it later.
17019 		 */
17020 		if ((rval = dtrace_helper_slurp(dof, dhp)) != -1) {
17021 			*rv = rval;
17022 			rval = 0;
17023 		} else {
17024 			rval = EINVAL;
17025 		}
17026 
17027 		mutex_exit(&dtrace_lock);
17028 		return (rval);
17029 	}
17030 
17031 	case DTRACEHIOC_REMOVE: {
17032 		mutex_enter(&dtrace_lock);
17033 		rval = dtrace_helper_destroygen(NULL, arg);
17034 		mutex_exit(&dtrace_lock);
17035 
17036 		return (rval);
17037 	}
17038 
17039 	default:
17040 		break;
17041 	}
17042 
17043 	return (ENOTTY);
17044 }
17045 
17046 /*ARGSUSED*/
17047 static int
17048 dtrace_ioctl(dev_t dev, int cmd, intptr_t arg, int md, cred_t *cr, int *rv)
17049 {
17050 	minor_t minor = getminor(dev);
17051 	dtrace_state_t *state;
17052 	int rval;
17053 
17054 	if (minor == DTRACEMNRN_HELPER)
17055 		return (dtrace_ioctl_helper(cmd, arg, rv));
17056 
17057 	state = ddi_get_soft_state(dtrace_softstate, minor);
17058 
17059 	if (state->dts_anon) {
17060 		ASSERT(dtrace_anon.dta_state == NULL);
17061 		state = state->dts_anon;
17062 	}
17063 
17064 	switch (cmd) {
17065 	case DTRACEIOC_PROVIDER: {
17066 		dtrace_providerdesc_t pvd;
17067 		dtrace_provider_t *pvp;
17068 
17069 		if (copyin((void *)arg, &pvd, sizeof (pvd)) != 0)
17070 			return (EFAULT);
17071 
17072 		pvd.dtvd_name[DTRACE_PROVNAMELEN - 1] = '\0';
17073 		mutex_enter(&dtrace_provider_lock);
17074 
17075 		for (pvp = dtrace_provider; pvp != NULL; pvp = pvp->dtpv_next) {
17076 			if (strcmp(pvp->dtpv_name, pvd.dtvd_name) == 0)
17077 				break;
17078 		}
17079 
17080 		mutex_exit(&dtrace_provider_lock);
17081 
17082 		if (pvp == NULL)
17083 			return (ESRCH);
17084 
17085 		bcopy(&pvp->dtpv_priv, &pvd.dtvd_priv, sizeof (dtrace_ppriv_t));
17086 		bcopy(&pvp->dtpv_attr, &pvd.dtvd_attr, sizeof (dtrace_pattr_t));
17087 
17088 		if (copyout(&pvd, (void *)arg, sizeof (pvd)) != 0)
17089 			return (EFAULT);
17090 
17091 		return (0);
17092 	}
17093 
17094 	case DTRACEIOC_EPROBE: {
17095 		dtrace_eprobedesc_t epdesc;
17096 		dtrace_ecb_t *ecb;
17097 		dtrace_action_t *act;
17098 		void *buf;
17099 		size_t size;
17100 		uintptr_t dest;
17101 		int nrecs;
17102 
17103 		if (copyin((void *)arg, &epdesc, sizeof (epdesc)) != 0)
17104 			return (EFAULT);
17105 
17106 		mutex_enter(&dtrace_lock);
17107 
17108 		if ((ecb = dtrace_epid2ecb(state, epdesc.dtepd_epid)) == NULL) {
17109 			mutex_exit(&dtrace_lock);
17110 			return (EINVAL);
17111 		}
17112 
17113 		if (ecb->dte_probe == NULL) {
17114 			mutex_exit(&dtrace_lock);
17115 			return (EINVAL);
17116 		}
17117 
17118 		epdesc.dtepd_probeid = ecb->dte_probe->dtpr_id;
17119 		epdesc.dtepd_uarg = ecb->dte_uarg;
17120 		epdesc.dtepd_size = ecb->dte_size;
17121 
17122 		nrecs = epdesc.dtepd_nrecs;
17123 		epdesc.dtepd_nrecs = 0;
17124 		for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
17125 			if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
17126 				continue;
17127 
17128 			epdesc.dtepd_nrecs++;
17129 		}
17130 
17131 		/*
17132 		 * Now that we have the size, we need to allocate a temporary
17133 		 * buffer in which to store the complete description.  We need
17134 		 * the temporary buffer to be able to drop dtrace_lock()
17135 		 * across the copyout(), below.
17136 		 */
17137 		size = sizeof (dtrace_eprobedesc_t) +
17138 		    (epdesc.dtepd_nrecs * sizeof (dtrace_recdesc_t));
17139 
17140 		buf = kmem_alloc(size, KM_SLEEP);
17141 		dest = (uintptr_t)buf;
17142 
17143 		bcopy(&epdesc, (void *)dest, sizeof (epdesc));
17144 		dest += offsetof(dtrace_eprobedesc_t, dtepd_rec[0]);
17145 
17146 		for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
17147 			if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
17148 				continue;
17149 
17150 			if (nrecs-- == 0)
17151 				break;
17152 
17153 			bcopy(&act->dta_rec, (void *)dest,
17154 			    sizeof (dtrace_recdesc_t));
17155 			dest += sizeof (dtrace_recdesc_t);
17156 		}
17157 
17158 		mutex_exit(&dtrace_lock);
17159 
17160 		if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
17161 			kmem_free(buf, size);
17162 			return (EFAULT);
17163 		}
17164 
17165 		kmem_free(buf, size);
17166 		return (0);
17167 	}
17168 
17169 	case DTRACEIOC_AGGDESC: {
17170 		dtrace_aggdesc_t aggdesc;
17171 		dtrace_action_t *act;
17172 		dtrace_aggregation_t *agg;
17173 		int nrecs;
17174 		uint32_t offs;
17175 		dtrace_recdesc_t *lrec;
17176 		void *buf;
17177 		size_t size;
17178 		uintptr_t dest;
17179 
17180 		if (copyin((void *)arg, &aggdesc, sizeof (aggdesc)) != 0)
17181 			return (EFAULT);
17182 
17183 		mutex_enter(&dtrace_lock);
17184 
17185 		if ((agg = dtrace_aggid2agg(state, aggdesc.dtagd_id)) == NULL) {
17186 			mutex_exit(&dtrace_lock);
17187 			return (EINVAL);
17188 		}
17189 
17190 		aggdesc.dtagd_epid = agg->dtag_ecb->dte_epid;
17191 
17192 		nrecs = aggdesc.dtagd_nrecs;
17193 		aggdesc.dtagd_nrecs = 0;
17194 
17195 		offs = agg->dtag_base;
17196 		lrec = &agg->dtag_action.dta_rec;
17197 		aggdesc.dtagd_size = lrec->dtrd_offset + lrec->dtrd_size - offs;
17198 
17199 		for (act = agg->dtag_first; ; act = act->dta_next) {
17200 			ASSERT(act->dta_intuple ||
17201 			    DTRACEACT_ISAGG(act->dta_kind));
17202 
17203 			/*
17204 			 * If this action has a record size of zero, it
17205 			 * denotes an argument to the aggregating action.
17206 			 * Because the presence of this record doesn't (or
17207 			 * shouldn't) affect the way the data is interpreted,
17208 			 * we don't copy it out to save user-level the
17209 			 * confusion of dealing with a zero-length record.
17210 			 */
17211 			if (act->dta_rec.dtrd_size == 0) {
17212 				ASSERT(agg->dtag_hasarg);
17213 				continue;
17214 			}
17215 
17216 			aggdesc.dtagd_nrecs++;
17217 
17218 			if (act == &agg->dtag_action)
17219 				break;
17220 		}
17221 
17222 		/*
17223 		 * Now that we have the size, we need to allocate a temporary
17224 		 * buffer in which to store the complete description.  We need
17225 		 * the temporary buffer to be able to drop dtrace_lock()
17226 		 * across the copyout(), below.
17227 		 */
17228 		size = sizeof (dtrace_aggdesc_t) +
17229 		    (aggdesc.dtagd_nrecs * sizeof (dtrace_recdesc_t));
17230 
17231 		buf = kmem_alloc(size, KM_SLEEP);
17232 		dest = (uintptr_t)buf;
17233 
17234 		bcopy(&aggdesc, (void *)dest, sizeof (aggdesc));
17235 		dest += offsetof(dtrace_aggdesc_t, dtagd_rec[0]);
17236 
17237 		for (act = agg->dtag_first; ; act = act->dta_next) {
17238 			dtrace_recdesc_t rec = act->dta_rec;
17239 
17240 			/*
17241 			 * See the comment in the above loop for why we pass
17242 			 * over zero-length records.
17243 			 */
17244 			if (rec.dtrd_size == 0) {
17245 				ASSERT(agg->dtag_hasarg);
17246 				continue;
17247 			}
17248 
17249 			if (nrecs-- == 0)
17250 				break;
17251 
17252 			rec.dtrd_offset -= offs;
17253 			bcopy(&rec, (void *)dest, sizeof (rec));
17254 			dest += sizeof (dtrace_recdesc_t);
17255 
17256 			if (act == &agg->dtag_action)
17257 				break;
17258 		}
17259 
17260 		mutex_exit(&dtrace_lock);
17261 
17262 		if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
17263 			kmem_free(buf, size);
17264 			return (EFAULT);
17265 		}
17266 
17267 		kmem_free(buf, size);
17268 		return (0);
17269 	}
17270 
17271 	case DTRACEIOC_ENABLE: {
17272 		dof_hdr_t *dof;
17273 		dtrace_enabling_t *enab = NULL;
17274 		dtrace_vstate_t *vstate;
17275 		int err = 0;
17276 
17277 		*rv = 0;
17278 
17279 		/*
17280 		 * If a NULL argument has been passed, we take this as our
17281 		 * cue to reevaluate our enablings.
17282 		 */
17283 		if (arg == NULL) {
17284 			dtrace_enabling_matchall();
17285 
17286 			return (0);
17287 		}
17288 
17289 		if ((dof = dtrace_dof_copyin(arg, &rval)) == NULL)
17290 			return (rval);
17291 
17292 		mutex_enter(&cpu_lock);
17293 		mutex_enter(&dtrace_lock);
17294 		vstate = &state->dts_vstate;
17295 
17296 		if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
17297 			mutex_exit(&dtrace_lock);
17298 			mutex_exit(&cpu_lock);
17299 			dtrace_dof_destroy(dof);
17300 			return (EBUSY);
17301 		}
17302 
17303 		if (dtrace_dof_slurp(dof, vstate, cr, &enab, 0, B_TRUE) != 0) {
17304 			mutex_exit(&dtrace_lock);
17305 			mutex_exit(&cpu_lock);
17306 			dtrace_dof_destroy(dof);
17307 			return (EINVAL);
17308 		}
17309 
17310 		if ((rval = dtrace_dof_options(dof, state)) != 0) {
17311 			dtrace_enabling_destroy(enab);
17312 			mutex_exit(&dtrace_lock);
17313 			mutex_exit(&cpu_lock);
17314 			dtrace_dof_destroy(dof);
17315 			return (rval);
17316 		}
17317 
17318 		if ((err = dtrace_enabling_match(enab, rv)) == 0) {
17319 			err = dtrace_enabling_retain(enab);
17320 		} else {
17321 			dtrace_enabling_destroy(enab);
17322 		}
17323 
17324 		mutex_exit(&cpu_lock);
17325 		mutex_exit(&dtrace_lock);
17326 		dtrace_dof_destroy(dof);
17327 
17328 		return (err);
17329 	}
17330 
17331 	case DTRACEIOC_REPLICATE: {
17332 		dtrace_repldesc_t desc;
17333 		dtrace_probedesc_t *match = &desc.dtrpd_match;
17334 		dtrace_probedesc_t *create = &desc.dtrpd_create;
17335 		int err;
17336 
17337 		if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
17338 			return (EFAULT);
17339 
17340 		match->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
17341 		match->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
17342 		match->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
17343 		match->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
17344 
17345 		create->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
17346 		create->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
17347 		create->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
17348 		create->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
17349 
17350 		mutex_enter(&dtrace_lock);
17351 		err = dtrace_enabling_replicate(state, match, create);
17352 		mutex_exit(&dtrace_lock);
17353 
17354 		return (err);
17355 	}
17356 
17357 	case DTRACEIOC_PROBEMATCH:
17358 	case DTRACEIOC_PROBES: {
17359 		dtrace_probe_t *probe = NULL;
17360 		dtrace_probedesc_t desc;
17361 		dtrace_probekey_t pkey;
17362 		dtrace_id_t i;
17363 		int m = 0;
17364 		uint32_t priv;
17365 		uid_t uid;
17366 		zoneid_t zoneid;
17367 
17368 		if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
17369 			return (EFAULT);
17370 
17371 		desc.dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
17372 		desc.dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
17373 		desc.dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
17374 		desc.dtpd_name[DTRACE_NAMELEN - 1] = '\0';
17375 
17376 		/*
17377 		 * Before we attempt to match this probe, we want to give
17378 		 * all providers the opportunity to provide it.
17379 		 */
17380 		if (desc.dtpd_id == DTRACE_IDNONE) {
17381 			mutex_enter(&dtrace_provider_lock);
17382 			dtrace_probe_provide(&desc, NULL);
17383 			mutex_exit(&dtrace_provider_lock);
17384 			desc.dtpd_id++;
17385 		}
17386 
17387 		if (cmd == DTRACEIOC_PROBEMATCH)  {
17388 			dtrace_probekey(&desc, &pkey);
17389 			pkey.dtpk_id = DTRACE_IDNONE;
17390 		}
17391 
17392 		dtrace_cred2priv(cr, &priv, &uid, &zoneid);
17393 
17394 		mutex_enter(&dtrace_lock);
17395 
17396 		if (cmd == DTRACEIOC_PROBEMATCH) {
17397 			for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
17398 				if ((probe = dtrace_probes[i - 1]) != NULL &&
17399 				    (m = dtrace_match_probe(probe, &pkey,
17400 				    priv, uid, zoneid)) != 0)
17401 					break;
17402 			}
17403 
17404 			if (m < 0) {
17405 				mutex_exit(&dtrace_lock);
17406 				return (EINVAL);
17407 			}
17408 
17409 		} else {
17410 			for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
17411 				if ((probe = dtrace_probes[i - 1]) != NULL &&
17412 				    dtrace_match_priv(probe, priv, uid, zoneid))
17413 					break;
17414 			}
17415 		}
17416 
17417 		if (probe == NULL) {
17418 			mutex_exit(&dtrace_lock);
17419 			return (ESRCH);
17420 		}
17421 
17422 		dtrace_probe_description(probe, &desc);
17423 		mutex_exit(&dtrace_lock);
17424 
17425 		if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
17426 			return (EFAULT);
17427 
17428 		return (0);
17429 	}
17430 
17431 	case DTRACEIOC_PROBEARG: {
17432 		dtrace_argdesc_t desc;
17433 		dtrace_probe_t *probe;
17434 		dtrace_provider_t *prov;
17435 
17436 		if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
17437 			return (EFAULT);
17438 
17439 		if (desc.dtargd_id == DTRACE_IDNONE)
17440 			return (EINVAL);
17441 
17442 		if (desc.dtargd_ndx == DTRACE_ARGNONE)
17443 			return (EINVAL);
17444 
17445 		mutex_enter(&dtrace_provider_lock);
17446 		mutex_enter(&mod_lock);
17447 		mutex_enter(&dtrace_lock);
17448 
17449 		if (desc.dtargd_id > dtrace_nprobes) {
17450 			mutex_exit(&dtrace_lock);
17451 			mutex_exit(&mod_lock);
17452 			mutex_exit(&dtrace_provider_lock);
17453 			return (EINVAL);
17454 		}
17455 
17456 		if ((probe = dtrace_probes[desc.dtargd_id - 1]) == NULL) {
17457 			mutex_exit(&dtrace_lock);
17458 			mutex_exit(&mod_lock);
17459 			mutex_exit(&dtrace_provider_lock);
17460 			return (EINVAL);
17461 		}
17462 
17463 		mutex_exit(&dtrace_lock);
17464 
17465 		prov = probe->dtpr_provider;
17466 
17467 		if (prov->dtpv_pops.dtps_getargdesc == NULL) {
17468 			/*
17469 			 * There isn't any typed information for this probe.
17470 			 * Set the argument number to DTRACE_ARGNONE.
17471 			 */
17472 			desc.dtargd_ndx = DTRACE_ARGNONE;
17473 		} else {
17474 			desc.dtargd_native[0] = '\0';
17475 			desc.dtargd_xlate[0] = '\0';
17476 			desc.dtargd_mapping = desc.dtargd_ndx;
17477 
17478 			prov->dtpv_pops.dtps_getargdesc(prov->dtpv_arg,
17479 			    probe->dtpr_id, probe->dtpr_arg, &desc);
17480 		}
17481 
17482 		mutex_exit(&mod_lock);
17483 		mutex_exit(&dtrace_provider_lock);
17484 
17485 		if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
17486 			return (EFAULT);
17487 
17488 		return (0);
17489 	}
17490 
17491 	case DTRACEIOC_GO: {
17492 		processorid_t cpuid;
17493 		rval = dtrace_state_go(state, &cpuid);
17494 
17495 		if (rval != 0)
17496 			return (rval);
17497 
17498 		if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
17499 			return (EFAULT);
17500 
17501 		return (0);
17502 	}
17503 
17504 	case DTRACEIOC_STOP: {
17505 		processorid_t cpuid;
17506 
17507 		mutex_enter(&dtrace_lock);
17508 		rval = dtrace_state_stop(state, &cpuid);
17509 		mutex_exit(&dtrace_lock);
17510 
17511 		if (rval != 0)
17512 			return (rval);
17513 
17514 		if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
17515 			return (EFAULT);
17516 
17517 		return (0);
17518 	}
17519 
17520 	case DTRACEIOC_DOFGET: {
17521 		dof_hdr_t hdr, *dof;
17522 		uint64_t len;
17523 
17524 		if (copyin((void *)arg, &hdr, sizeof (hdr)) != 0)
17525 			return (EFAULT);
17526 
17527 		mutex_enter(&dtrace_lock);
17528 		dof = dtrace_dof_create(state);
17529 		mutex_exit(&dtrace_lock);
17530 
17531 		len = MIN(hdr.dofh_loadsz, dof->dofh_loadsz);
17532 		rval = copyout(dof, (void *)arg, len);
17533 		dtrace_dof_destroy(dof);
17534 
17535 		return (rval == 0 ? 0 : EFAULT);
17536 	}
17537 
17538 	case DTRACEIOC_AGGSNAP:
17539 	case DTRACEIOC_BUFSNAP: {
17540 		dtrace_bufdesc_t desc;
17541 		caddr_t cached;
17542 		dtrace_buffer_t *buf;
17543 
17544 		if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
17545 			return (EFAULT);
17546 
17547 		if (desc.dtbd_cpu < 0 || desc.dtbd_cpu >= NCPU)
17548 			return (EINVAL);
17549 
17550 		mutex_enter(&dtrace_lock);
17551 
17552 		if (cmd == DTRACEIOC_BUFSNAP) {
17553 			buf = &state->dts_buffer[desc.dtbd_cpu];
17554 		} else {
17555 			buf = &state->dts_aggbuffer[desc.dtbd_cpu];
17556 		}
17557 
17558 		if (buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL)) {
17559 			size_t sz = buf->dtb_offset;
17560 
17561 			if (state->dts_activity != DTRACE_ACTIVITY_STOPPED) {
17562 				mutex_exit(&dtrace_lock);
17563 				return (EBUSY);
17564 			}
17565 
17566 			/*
17567 			 * If this buffer has already been consumed, we're
17568 			 * going to indicate that there's nothing left here
17569 			 * to consume.
17570 			 */
17571 			if (buf->dtb_flags & DTRACEBUF_CONSUMED) {
17572 				mutex_exit(&dtrace_lock);
17573 
17574 				desc.dtbd_size = 0;
17575 				desc.dtbd_drops = 0;
17576 				desc.dtbd_errors = 0;
17577 				desc.dtbd_oldest = 0;
17578 				sz = sizeof (desc);
17579 
17580 				if (copyout(&desc, (void *)arg, sz) != 0)
17581 					return (EFAULT);
17582 
17583 				return (0);
17584 			}
17585 
17586 			/*
17587 			 * If this is a ring buffer that has wrapped, we want
17588 			 * to copy the whole thing out.
17589 			 */
17590 			if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
17591 				dtrace_buffer_polish(buf);
17592 				sz = buf->dtb_size;
17593 			}
17594 
17595 			if (copyout(buf->dtb_tomax, desc.dtbd_data, sz) != 0) {
17596 				mutex_exit(&dtrace_lock);
17597 				return (EFAULT);
17598 			}
17599 
17600 			desc.dtbd_size = sz;
17601 			desc.dtbd_drops = buf->dtb_drops;
17602 			desc.dtbd_errors = buf->dtb_errors;
17603 			desc.dtbd_oldest = buf->dtb_xamot_offset;
17604 			desc.dtbd_timestamp = dtrace_gethrtime();
17605 
17606 			mutex_exit(&dtrace_lock);
17607 
17608 			if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
17609 				return (EFAULT);
17610 
17611 			buf->dtb_flags |= DTRACEBUF_CONSUMED;
17612 
17613 			return (0);
17614 		}
17615 
17616 		if (buf->dtb_tomax == NULL) {
17617 			ASSERT(buf->dtb_xamot == NULL);
17618 			mutex_exit(&dtrace_lock);
17619 			return (ENOENT);
17620 		}
17621 
17622 		cached = buf->dtb_tomax;
17623 		ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
17624 
17625 		dtrace_xcall(desc.dtbd_cpu,
17626 		    (dtrace_xcall_t)dtrace_buffer_switch, buf);
17627 
17628 		state->dts_errors += buf->dtb_xamot_errors;
17629 
17630 		/*
17631 		 * If the buffers did not actually switch, then the cross call
17632 		 * did not take place -- presumably because the given CPU is
17633 		 * not in the ready set.  If this is the case, we'll return
17634 		 * ENOENT.
17635 		 */
17636 		if (buf->dtb_tomax == cached) {
17637 			ASSERT(buf->dtb_xamot != cached);
17638 			mutex_exit(&dtrace_lock);
17639 			return (ENOENT);
17640 		}
17641 
17642 		ASSERT(cached == buf->dtb_xamot);
17643 
17644 		/*
17645 		 * We have our snapshot; now copy it out.
17646 		 */
17647 		if (copyout(buf->dtb_xamot, desc.dtbd_data,
17648 		    buf->dtb_xamot_offset) != 0) {
17649 			mutex_exit(&dtrace_lock);
17650 			return (EFAULT);
17651 		}
17652 
17653 		desc.dtbd_size = buf->dtb_xamot_offset;
17654 		desc.dtbd_drops = buf->dtb_xamot_drops;
17655 		desc.dtbd_errors = buf->dtb_xamot_errors;
17656 		desc.dtbd_oldest = 0;
17657 		desc.dtbd_timestamp = buf->dtb_switched;
17658 
17659 		mutex_exit(&dtrace_lock);
17660 
17661 		/*
17662 		 * Finally, copy out the buffer description.
17663 		 */
17664 		if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
17665 			return (EFAULT);
17666 
17667 		return (0);
17668 	}
17669 
17670 	case DTRACEIOC_CONF: {
17671 		dtrace_conf_t conf;
17672 
17673 		bzero(&conf, sizeof (conf));
17674 		conf.dtc_difversion = DIF_VERSION;
17675 		conf.dtc_difintregs = DIF_DIR_NREGS;
17676 		conf.dtc_diftupregs = DIF_DTR_NREGS;
17677 		conf.dtc_ctfmodel = CTF_MODEL_NATIVE;
17678 
17679 		if (copyout(&conf, (void *)arg, sizeof (conf)) != 0)
17680 			return (EFAULT);
17681 
17682 		return (0);
17683 	}
17684 
17685 	case DTRACEIOC_STATUS: {
17686 		dtrace_status_t stat;
17687 		dtrace_dstate_t *dstate;
17688 		int i, j;
17689 		uint64_t nerrs;
17690 
17691 		/*
17692 		 * See the comment in dtrace_state_deadman() for the reason
17693 		 * for setting dts_laststatus to INT64_MAX before setting
17694 		 * it to the correct value.
17695 		 */
17696 		state->dts_laststatus = INT64_MAX;
17697 		dtrace_membar_producer();
17698 		state->dts_laststatus = dtrace_gethrtime();
17699 
17700 		bzero(&stat, sizeof (stat));
17701 
17702 		mutex_enter(&dtrace_lock);
17703 
17704 		if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE) {
17705 			mutex_exit(&dtrace_lock);
17706 			return (ENOENT);
17707 		}
17708 
17709 		if (state->dts_activity == DTRACE_ACTIVITY_DRAINING)
17710 			stat.dtst_exiting = 1;
17711 
17712 		nerrs = state->dts_errors;
17713 		dstate = &state->dts_vstate.dtvs_dynvars;
17714 
17715 		for (i = 0; i < NCPU; i++) {
17716 			dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[i];
17717 
17718 			stat.dtst_dyndrops += dcpu->dtdsc_drops;
17719 			stat.dtst_dyndrops_dirty += dcpu->dtdsc_dirty_drops;
17720 			stat.dtst_dyndrops_rinsing += dcpu->dtdsc_rinsing_drops;
17721 
17722 			if (state->dts_buffer[i].dtb_flags & DTRACEBUF_FULL)
17723 				stat.dtst_filled++;
17724 
17725 			nerrs += state->dts_buffer[i].dtb_errors;
17726 
17727 			for (j = 0; j < state->dts_nspeculations; j++) {
17728 				dtrace_speculation_t *spec;
17729 				dtrace_buffer_t *buf;
17730 
17731 				spec = &state->dts_speculations[j];
17732 				buf = &spec->dtsp_buffer[i];
17733 				stat.dtst_specdrops += buf->dtb_xamot_drops;
17734 			}
17735 		}
17736 
17737 		stat.dtst_specdrops_busy = state->dts_speculations_busy;
17738 		stat.dtst_specdrops_unavail = state->dts_speculations_unavail;
17739 		stat.dtst_stkstroverflows = state->dts_stkstroverflows;
17740 		stat.dtst_dblerrors = state->dts_dblerrors;
17741 		stat.dtst_killed =
17742 		    (state->dts_activity == DTRACE_ACTIVITY_KILLED);
17743 		stat.dtst_errors = nerrs;
17744 
17745 		mutex_exit(&dtrace_lock);
17746 
17747 		if (copyout(&stat, (void *)arg, sizeof (stat)) != 0)
17748 			return (EFAULT);
17749 
17750 		return (0);
17751 	}
17752 
17753 	case DTRACEIOC_FORMAT: {
17754 		dtrace_fmtdesc_t fmt;
17755 		char *str;
17756 		int len;
17757 
17758 		if (copyin((void *)arg, &fmt, sizeof (fmt)) != 0)
17759 			return (EFAULT);
17760 
17761 		mutex_enter(&dtrace_lock);
17762 
17763 		if (fmt.dtfd_format == 0 ||
17764 		    fmt.dtfd_format > state->dts_nformats) {
17765 			mutex_exit(&dtrace_lock);
17766 			return (EINVAL);
17767 		}
17768 
17769 		/*
17770 		 * Format strings are allocated contiguously and they are
17771 		 * never freed; if a format index is less than the number
17772 		 * of formats, we can assert that the format map is non-NULL
17773 		 * and that the format for the specified index is non-NULL.
17774 		 */
17775 		ASSERT(state->dts_formats != NULL);
17776 		str = state->dts_formats[fmt.dtfd_format - 1];
17777 		ASSERT(str != NULL);
17778 
17779 		len = strlen(str) + 1;
17780 
17781 		if (len > fmt.dtfd_length) {
17782 			fmt.dtfd_length = len;
17783 
17784 			if (copyout(&fmt, (void *)arg, sizeof (fmt)) != 0) {
17785 				mutex_exit(&dtrace_lock);
17786 				return (EINVAL);
17787 			}
17788 		} else {
17789 			if (copyout(str, fmt.dtfd_string, len) != 0) {
17790 				mutex_exit(&dtrace_lock);
17791 				return (EINVAL);
17792 			}
17793 		}
17794 
17795 		mutex_exit(&dtrace_lock);
17796 		return (0);
17797 	}
17798 
17799 	default:
17800 		break;
17801 	}
17802 
17803 	return (ENOTTY);
17804 }
17805 
17806 /*ARGSUSED*/
17807 static int
17808 dtrace_detach(dev_info_t *dip, ddi_detach_cmd_t cmd)
17809 {
17810 	dtrace_state_t *state;
17811 
17812 	switch (cmd) {
17813 	case DDI_DETACH:
17814 		break;
17815 
17816 	case DDI_SUSPEND:
17817 		return (DDI_SUCCESS);
17818 
17819 	default:
17820 		return (DDI_FAILURE);
17821 	}
17822 
17823 	mutex_enter(&cpu_lock);
17824 	mutex_enter(&dtrace_provider_lock);
17825 	mutex_enter(&dtrace_lock);
17826 
17827 	ASSERT(dtrace_opens == 0);
17828 
17829 	if (dtrace_helpers > 0) {
17830 		mutex_exit(&dtrace_provider_lock);
17831 		mutex_exit(&dtrace_lock);
17832 		mutex_exit(&cpu_lock);
17833 		return (DDI_FAILURE);
17834 	}
17835 
17836 	if (dtrace_unregister((dtrace_provider_id_t)dtrace_provider) != 0) {
17837 		mutex_exit(&dtrace_provider_lock);
17838 		mutex_exit(&dtrace_lock);
17839 		mutex_exit(&cpu_lock);
17840 		return (DDI_FAILURE);
17841 	}
17842 
17843 	dtrace_provider = NULL;
17844 
17845 	if ((state = dtrace_anon_grab()) != NULL) {
17846 		/*
17847 		 * If there were ECBs on this state, the provider should
17848 		 * have not been allowed to detach; assert that there is
17849 		 * none.
17850 		 */
17851 		ASSERT(state->dts_necbs == 0);
17852 		dtrace_state_destroy(state);
17853 
17854 		/*
17855 		 * If we're being detached with anonymous state, we need to
17856 		 * indicate to the kernel debugger that DTrace is now inactive.
17857 		 */
17858 		(void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
17859 	}
17860 
17861 	bzero(&dtrace_anon, sizeof (dtrace_anon_t));
17862 	unregister_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
17863 	dtrace_cpu_init = NULL;
17864 	dtrace_helpers_cleanup = NULL;
17865 	dtrace_helpers_fork = NULL;
17866 	dtrace_cpustart_init = NULL;
17867 	dtrace_cpustart_fini = NULL;
17868 	dtrace_debugger_init = NULL;
17869 	dtrace_debugger_fini = NULL;
17870 	dtrace_modload = NULL;
17871 	dtrace_modunload = NULL;
17872 
17873 	ASSERT(dtrace_getf == 0);
17874 	ASSERT(dtrace_closef == NULL);
17875 
17876 	mutex_exit(&cpu_lock);
17877 
17878 	kmem_free(dtrace_probes, dtrace_nprobes * sizeof (dtrace_probe_t *));
17879 	dtrace_probes = NULL;
17880 	dtrace_nprobes = 0;
17881 
17882 	dtrace_hash_destroy(dtrace_bymod);
17883 	dtrace_hash_destroy(dtrace_byfunc);
17884 	dtrace_hash_destroy(dtrace_byname);
17885 	dtrace_bymod = NULL;
17886 	dtrace_byfunc = NULL;
17887 	dtrace_byname = NULL;
17888 
17889 	kmem_cache_destroy(dtrace_state_cache);
17890 	vmem_destroy(dtrace_minor);
17891 	vmem_destroy(dtrace_arena);
17892 
17893 	if (dtrace_toxrange != NULL) {
17894 		kmem_free(dtrace_toxrange,
17895 		    dtrace_toxranges_max * sizeof (dtrace_toxrange_t));
17896 		dtrace_toxrange = NULL;
17897 		dtrace_toxranges = 0;
17898 		dtrace_toxranges_max = 0;
17899 	}
17900 
17901 	ddi_remove_minor_node(dtrace_devi, NULL);
17902 	dtrace_devi = NULL;
17903 
17904 	ddi_soft_state_fini(&dtrace_softstate);
17905 
17906 	ASSERT(dtrace_vtime_references == 0);
17907 	ASSERT(dtrace_opens == 0);
17908 	ASSERT(dtrace_retained == NULL);
17909 
17910 	mutex_exit(&dtrace_lock);
17911 	mutex_exit(&dtrace_provider_lock);
17912 
17913 	/*
17914 	 * We don't destroy the task queue until after we have dropped our
17915 	 * locks (taskq_destroy() may block on running tasks).  To prevent
17916 	 * attempting to do work after we have effectively detached but before
17917 	 * the task queue has been destroyed, all tasks dispatched via the
17918 	 * task queue must check that DTrace is still attached before
17919 	 * performing any operation.
17920 	 */
17921 	taskq_destroy(dtrace_taskq);
17922 	dtrace_taskq = NULL;
17923 
17924 	return (DDI_SUCCESS);
17925 }
17926 #endif
17927 
17928 #ifdef illumos
17929 /*ARGSUSED*/
17930 static int
17931 dtrace_info(dev_info_t *dip, ddi_info_cmd_t infocmd, void *arg, void **result)
17932 {
17933 	int error;
17934 
17935 	switch (infocmd) {
17936 	case DDI_INFO_DEVT2DEVINFO:
17937 		*result = (void *)dtrace_devi;
17938 		error = DDI_SUCCESS;
17939 		break;
17940 	case DDI_INFO_DEVT2INSTANCE:
17941 		*result = (void *)0;
17942 		error = DDI_SUCCESS;
17943 		break;
17944 	default:
17945 		error = DDI_FAILURE;
17946 	}
17947 	return (error);
17948 }
17949 #endif
17950 
17951 #ifdef illumos
17952 static struct cb_ops dtrace_cb_ops = {
17953 	dtrace_open,		/* open */
17954 	dtrace_close,		/* close */
17955 	nulldev,		/* strategy */
17956 	nulldev,		/* print */
17957 	nodev,			/* dump */
17958 	nodev,			/* read */
17959 	nodev,			/* write */
17960 	dtrace_ioctl,		/* ioctl */
17961 	nodev,			/* devmap */
17962 	nodev,			/* mmap */
17963 	nodev,			/* segmap */
17964 	nochpoll,		/* poll */
17965 	ddi_prop_op,		/* cb_prop_op */
17966 	0,			/* streamtab  */
17967 	D_NEW | D_MP		/* Driver compatibility flag */
17968 };
17969 
17970 static struct dev_ops dtrace_ops = {
17971 	DEVO_REV,		/* devo_rev */
17972 	0,			/* refcnt */
17973 	dtrace_info,		/* get_dev_info */
17974 	nulldev,		/* identify */
17975 	nulldev,		/* probe */
17976 	dtrace_attach,		/* attach */
17977 	dtrace_detach,		/* detach */
17978 	nodev,			/* reset */
17979 	&dtrace_cb_ops,		/* driver operations */
17980 	NULL,			/* bus operations */
17981 	nodev			/* dev power */
17982 };
17983 
17984 static struct modldrv modldrv = {
17985 	&mod_driverops,		/* module type (this is a pseudo driver) */
17986 	"Dynamic Tracing",	/* name of module */
17987 	&dtrace_ops,		/* driver ops */
17988 };
17989 
17990 static struct modlinkage modlinkage = {
17991 	MODREV_1,
17992 	(void *)&modldrv,
17993 	NULL
17994 };
17995 
17996 int
17997 _init(void)
17998 {
17999 	return (mod_install(&modlinkage));
18000 }
18001 
18002 int
18003 _info(struct modinfo *modinfop)
18004 {
18005 	return (mod_info(&modlinkage, modinfop));
18006 }
18007 
18008 int
18009 _fini(void)
18010 {
18011 	return (mod_remove(&modlinkage));
18012 }
18013 #else
18014 
18015 static d_ioctl_t	dtrace_ioctl;
18016 static d_ioctl_t	dtrace_ioctl_helper;
18017 static void		dtrace_load(void *);
18018 static int		dtrace_unload(void);
18019 static struct cdev	*dtrace_dev;
18020 static struct cdev	*helper_dev;
18021 
18022 void dtrace_invop_init(void);
18023 void dtrace_invop_uninit(void);
18024 
18025 static struct cdevsw dtrace_cdevsw = {
18026 	.d_version	= D_VERSION,
18027 	.d_ioctl	= dtrace_ioctl,
18028 	.d_open		= dtrace_open,
18029 	.d_name		= "dtrace",
18030 };
18031 
18032 static struct cdevsw helper_cdevsw = {
18033 	.d_version	= D_VERSION,
18034 	.d_ioctl	= dtrace_ioctl_helper,
18035 	.d_name		= "helper",
18036 };
18037 
18038 #include <dtrace_anon.c>
18039 #include <dtrace_ioctl.c>
18040 #include <dtrace_load.c>
18041 #include <dtrace_modevent.c>
18042 #include <dtrace_sysctl.c>
18043 #include <dtrace_unload.c>
18044 #include <dtrace_vtime.c>
18045 #include <dtrace_hacks.c>
18046 #include <dtrace_isa.c>
18047 
18048 SYSINIT(dtrace_load, SI_SUB_DTRACE, SI_ORDER_FIRST, dtrace_load, NULL);
18049 SYSUNINIT(dtrace_unload, SI_SUB_DTRACE, SI_ORDER_FIRST, dtrace_unload, NULL);
18050 SYSINIT(dtrace_anon_init, SI_SUB_DTRACE_ANON, SI_ORDER_FIRST, dtrace_anon_init, NULL);
18051 
18052 DEV_MODULE(dtrace, dtrace_modevent, NULL);
18053 MODULE_VERSION(dtrace, 1);
18054 MODULE_DEPEND(dtrace, opensolaris, 1, 1, 1);
18055 #endif
18056