xref: /illumos-gate/usr/src/uts/common/dtrace/dtrace.c (revision 565657ca18725b8f8dbe5b93704cc1d173be9d65)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
24  * Copyright (c) 2016, Joyent, Inc. All rights reserved.
25  * Copyright (c) 2012, 2014 by Delphix. All rights reserved.
26  */
27 
28 /*
29  * DTrace - Dynamic Tracing for Solaris
30  *
31  * This is the implementation of the Solaris Dynamic Tracing framework
32  * (DTrace).  The user-visible interface to DTrace is described at length in
33  * the "Solaris Dynamic Tracing Guide".  The interfaces between the libdtrace
34  * library, the in-kernel DTrace framework, and the DTrace providers are
35  * described in the block comments in the <sys/dtrace.h> header file.  The
36  * internal architecture of DTrace is described in the block comments in the
37  * <sys/dtrace_impl.h> header file.  The comments contained within the DTrace
38  * implementation very much assume mastery of all of these sources; if one has
39  * an unanswered question about the implementation, one should consult them
40  * first.
41  *
42  * The functions here are ordered roughly as follows:
43  *
44  *   - Probe context functions
45  *   - Probe hashing functions
46  *   - Non-probe context utility functions
47  *   - Matching functions
48  *   - Provider-to-Framework API functions
49  *   - Probe management functions
50  *   - DIF object functions
51  *   - Format functions
52  *   - Predicate functions
53  *   - ECB functions
54  *   - Buffer functions
55  *   - Enabling functions
56  *   - DOF functions
57  *   - Anonymous enabling functions
58  *   - Consumer state functions
59  *   - Helper functions
60  *   - Hook functions
61  *   - Driver cookbook functions
62  *
63  * Each group of functions begins with a block comment labelled the "DTrace
64  * [Group] Functions", allowing one to find each block by searching forward
65  * on capital-f functions.
66  */
67 #include <sys/errno.h>
68 #include <sys/stat.h>
69 #include <sys/modctl.h>
70 #include <sys/conf.h>
71 #include <sys/systm.h>
72 #include <sys/ddi.h>
73 #include <sys/sunddi.h>
74 #include <sys/cpuvar.h>
75 #include <sys/kmem.h>
76 #include <sys/strsubr.h>
77 #include <sys/sysmacros.h>
78 #include <sys/dtrace_impl.h>
79 #include <sys/atomic.h>
80 #include <sys/cmn_err.h>
81 #include <sys/mutex_impl.h>
82 #include <sys/rwlock_impl.h>
83 #include <sys/ctf_api.h>
84 #include <sys/panic.h>
85 #include <sys/priv_impl.h>
86 #include <sys/policy.h>
87 #include <sys/cred_impl.h>
88 #include <sys/procfs_isa.h>
89 #include <sys/taskq.h>
90 #include <sys/mkdev.h>
91 #include <sys/kdi.h>
92 #include <sys/zone.h>
93 #include <sys/socket.h>
94 #include <netinet/in.h>
95 #include "strtolctype.h"
96 
97 /*
98  * DTrace Tunable Variables
99  *
100  * The following variables may be tuned by adding a line to /etc/system that
101  * includes both the name of the DTrace module ("dtrace") and the name of the
102  * variable.  For example:
103  *
104  *   set dtrace:dtrace_destructive_disallow = 1
105  *
106  * In general, the only variables that one should be tuning this way are those
107  * that affect system-wide DTrace behavior, and for which the default behavior
108  * is undesirable.  Most of these variables are tunable on a per-consumer
109  * basis using DTrace options, and need not be tuned on a system-wide basis.
110  * When tuning these variables, avoid pathological values; while some attempt
111  * is made to verify the integrity of these variables, they are not considered
112  * part of the supported interface to DTrace, and they are therefore not
113  * checked comprehensively.  Further, these variables should not be tuned
114  * dynamically via "mdb -kw" or other means; they should only be tuned via
115  * /etc/system.
116  */
117 int		dtrace_destructive_disallow = 0;
118 dtrace_optval_t	dtrace_nonroot_maxsize = (16 * 1024 * 1024);
119 size_t		dtrace_difo_maxsize = (256 * 1024);
120 dtrace_optval_t	dtrace_dof_maxsize = (8 * 1024 * 1024);
121 size_t		dtrace_statvar_maxsize = (16 * 1024);
122 size_t		dtrace_actions_max = (16 * 1024);
123 size_t		dtrace_retain_max = 1024;
124 dtrace_optval_t	dtrace_helper_actions_max = 1024;
125 dtrace_optval_t	dtrace_helper_providers_max = 32;
126 dtrace_optval_t	dtrace_dstate_defsize = (1 * 1024 * 1024);
127 size_t		dtrace_strsize_default = 256;
128 dtrace_optval_t	dtrace_cleanrate_default = 9900990;		/* 101 hz */
129 dtrace_optval_t	dtrace_cleanrate_min = 200000;			/* 5000 hz */
130 dtrace_optval_t	dtrace_cleanrate_max = (uint64_t)60 * NANOSEC;	/* 1/minute */
131 dtrace_optval_t	dtrace_aggrate_default = NANOSEC;		/* 1 hz */
132 dtrace_optval_t	dtrace_statusrate_default = NANOSEC;		/* 1 hz */
133 dtrace_optval_t dtrace_statusrate_max = (hrtime_t)10 * NANOSEC;	 /* 6/minute */
134 dtrace_optval_t	dtrace_switchrate_default = NANOSEC;		/* 1 hz */
135 dtrace_optval_t	dtrace_nspec_default = 1;
136 dtrace_optval_t	dtrace_specsize_default = 32 * 1024;
137 dtrace_optval_t dtrace_stackframes_default = 20;
138 dtrace_optval_t dtrace_ustackframes_default = 20;
139 dtrace_optval_t dtrace_jstackframes_default = 50;
140 dtrace_optval_t dtrace_jstackstrsize_default = 512;
141 int		dtrace_msgdsize_max = 128;
142 hrtime_t	dtrace_chill_max = MSEC2NSEC(500);		/* 500 ms */
143 hrtime_t	dtrace_chill_interval = NANOSEC;		/* 1000 ms */
144 int		dtrace_devdepth_max = 32;
145 int		dtrace_err_verbose;
146 hrtime_t	dtrace_deadman_interval = NANOSEC;
147 hrtime_t	dtrace_deadman_timeout = (hrtime_t)10 * NANOSEC;
148 hrtime_t	dtrace_deadman_user = (hrtime_t)30 * NANOSEC;
149 hrtime_t	dtrace_unregister_defunct_reap = (hrtime_t)60 * NANOSEC;
150 
151 /*
152  * DTrace External Variables
153  *
154  * As dtrace(7D) is a kernel module, any DTrace variables are obviously
155  * available to DTrace consumers via the backtick (`) syntax.  One of these,
156  * dtrace_zero, is made deliberately so:  it is provided as a source of
157  * well-known, zero-filled memory.  While this variable is not documented,
158  * it is used by some translators as an implementation detail.
159  */
160 const char	dtrace_zero[256] = { 0 };	/* zero-filled memory */
161 
162 /*
163  * DTrace Internal Variables
164  */
165 static dev_info_t	*dtrace_devi;		/* device info */
166 static vmem_t		*dtrace_arena;		/* probe ID arena */
167 static vmem_t		*dtrace_minor;		/* minor number arena */
168 static taskq_t		*dtrace_taskq;		/* task queue */
169 static dtrace_probe_t	**dtrace_probes;	/* array of all probes */
170 static int		dtrace_nprobes;		/* number of probes */
171 static dtrace_provider_t *dtrace_provider;	/* provider list */
172 static dtrace_meta_t	*dtrace_meta_pid;	/* user-land meta provider */
173 static int		dtrace_opens;		/* number of opens */
174 static int		dtrace_helpers;		/* number of helpers */
175 static int		dtrace_getf;		/* number of unpriv getf()s */
176 static void		*dtrace_softstate;	/* softstate pointer */
177 static dtrace_hash_t	*dtrace_bymod;		/* probes hashed by module */
178 static dtrace_hash_t	*dtrace_byfunc;		/* probes hashed by function */
179 static dtrace_hash_t	*dtrace_byname;		/* probes hashed by name */
180 static dtrace_toxrange_t *dtrace_toxrange;	/* toxic range array */
181 static int		dtrace_toxranges;	/* number of toxic ranges */
182 static int		dtrace_toxranges_max;	/* size of toxic range array */
183 static dtrace_anon_t	dtrace_anon;		/* anonymous enabling */
184 static kmem_cache_t	*dtrace_state_cache;	/* cache for dynamic state */
185 static uint64_t		dtrace_vtime_references; /* number of vtimestamp refs */
186 static kthread_t	*dtrace_panicked;	/* panicking thread */
187 static dtrace_ecb_t	*dtrace_ecb_create_cache; /* cached created ECB */
188 static dtrace_genid_t	dtrace_probegen;	/* current probe generation */
189 static dtrace_helpers_t *dtrace_deferred_pid;	/* deferred helper list */
190 static dtrace_enabling_t *dtrace_retained;	/* list of retained enablings */
191 static dtrace_genid_t	dtrace_retained_gen;	/* current retained enab gen */
192 static dtrace_dynvar_t	dtrace_dynhash_sink;	/* end of dynamic hash chains */
193 static int		dtrace_dynvar_failclean; /* dynvars failed to clean */
194 
195 /*
196  * DTrace Locking
197  * DTrace is protected by three (relatively coarse-grained) locks:
198  *
199  * (1) dtrace_lock is required to manipulate essentially any DTrace state,
200  *     including enabling state, probes, ECBs, consumer state, helper state,
201  *     etc.  Importantly, dtrace_lock is _not_ required when in probe context;
202  *     probe context is lock-free -- synchronization is handled via the
203  *     dtrace_sync() cross call mechanism.
204  *
205  * (2) dtrace_provider_lock is required when manipulating provider state, or
206  *     when provider state must be held constant.
207  *
208  * (3) dtrace_meta_lock is required when manipulating meta provider state, or
209  *     when meta provider state must be held constant.
210  *
211  * The lock ordering between these three locks is dtrace_meta_lock before
212  * dtrace_provider_lock before dtrace_lock.  (In particular, there are
213  * several places where dtrace_provider_lock is held by the framework as it
214  * calls into the providers -- which then call back into the framework,
215  * grabbing dtrace_lock.)
216  *
217  * There are two other locks in the mix:  mod_lock and cpu_lock.  With respect
218  * to dtrace_provider_lock and dtrace_lock, cpu_lock continues its historical
219  * role as a coarse-grained lock; it is acquired before both of these locks.
220  * With respect to dtrace_meta_lock, its behavior is stranger:  cpu_lock must
221  * be acquired _between_ dtrace_meta_lock and any other DTrace locks.
222  * mod_lock is similar with respect to dtrace_provider_lock in that it must be
223  * acquired _between_ dtrace_provider_lock and dtrace_lock.
224  */
225 static kmutex_t		dtrace_lock;		/* probe state lock */
226 static kmutex_t		dtrace_provider_lock;	/* provider state lock */
227 static kmutex_t		dtrace_meta_lock;	/* meta-provider state lock */
228 
229 /*
230  * DTrace Provider Variables
231  *
232  * These are the variables relating to DTrace as a provider (that is, the
233  * provider of the BEGIN, END, and ERROR probes).
234  */
235 static dtrace_pattr_t	dtrace_provider_attr = {
236 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
237 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
238 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
239 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
240 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
241 };
242 
243 static void
244 dtrace_nullop(void)
245 {}
246 
247 static int
248 dtrace_enable_nullop(void)
249 {
250 	return (0);
251 }
252 
253 static dtrace_pops_t	dtrace_provider_ops = {
254 	(void (*)(void *, const dtrace_probedesc_t *))dtrace_nullop,
255 	(void (*)(void *, struct modctl *))dtrace_nullop,
256 	(int (*)(void *, dtrace_id_t, void *))dtrace_enable_nullop,
257 	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
258 	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
259 	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
260 	NULL,
261 	NULL,
262 	NULL,
263 	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop
264 };
265 
266 static dtrace_id_t	dtrace_probeid_begin;	/* special BEGIN probe */
267 static dtrace_id_t	dtrace_probeid_end;	/* special END probe */
268 dtrace_id_t		dtrace_probeid_error;	/* special ERROR probe */
269 
270 /*
271  * DTrace Helper Tracing Variables
272  *
273  * These variables should be set dynamically to enable helper tracing.  The
274  * only variables that should be set are dtrace_helptrace_enable (which should
275  * be set to a non-zero value to allocate helper tracing buffers on the next
276  * open of /dev/dtrace) and dtrace_helptrace_disable (which should be set to a
277  * non-zero value to deallocate helper tracing buffers on the next close of
278  * /dev/dtrace).  When (and only when) helper tracing is disabled, the
279  * buffer size may also be set via dtrace_helptrace_bufsize.
280  */
281 int			dtrace_helptrace_enable = 0;
282 int			dtrace_helptrace_disable = 0;
283 int			dtrace_helptrace_bufsize = 16 * 1024 * 1024;
284 uint32_t		dtrace_helptrace_nlocals;
285 static dtrace_helptrace_t *dtrace_helptrace_buffer;
286 static uint32_t		dtrace_helptrace_next = 0;
287 static int		dtrace_helptrace_wrapped = 0;
288 
289 /*
290  * DTrace Error Hashing
291  *
292  * On DEBUG kernels, DTrace will track the errors that has seen in a hash
293  * table.  This is very useful for checking coverage of tests that are
294  * expected to induce DIF or DOF processing errors, and may be useful for
295  * debugging problems in the DIF code generator or in DOF generation .  The
296  * error hash may be examined with the ::dtrace_errhash MDB dcmd.
297  */
298 #ifdef DEBUG
299 static dtrace_errhash_t	dtrace_errhash[DTRACE_ERRHASHSZ];
300 static const char *dtrace_errlast;
301 static kthread_t *dtrace_errthread;
302 static kmutex_t dtrace_errlock;
303 #endif
304 
305 /*
306  * DTrace Macros and Constants
307  *
308  * These are various macros that are useful in various spots in the
309  * implementation, along with a few random constants that have no meaning
310  * outside of the implementation.  There is no real structure to this cpp
311  * mishmash -- but is there ever?
312  */
313 #define	DTRACE_HASHSTR(hash, probe)	\
314 	dtrace_hash_str(*((char **)((uintptr_t)(probe) + (hash)->dth_stroffs)))
315 
316 #define	DTRACE_HASHNEXT(hash, probe)	\
317 	(dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_nextoffs)
318 
319 #define	DTRACE_HASHPREV(hash, probe)	\
320 	(dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_prevoffs)
321 
322 #define	DTRACE_HASHEQ(hash, lhs, rhs)	\
323 	(strcmp(*((char **)((uintptr_t)(lhs) + (hash)->dth_stroffs)), \
324 	    *((char **)((uintptr_t)(rhs) + (hash)->dth_stroffs))) == 0)
325 
326 #define	DTRACE_AGGHASHSIZE_SLEW		17
327 
328 #define	DTRACE_V4MAPPED_OFFSET		(sizeof (uint32_t) * 3)
329 
330 /*
331  * The key for a thread-local variable consists of the lower 61 bits of the
332  * t_did, plus the 3 bits of the highest active interrupt above LOCK_LEVEL.
333  * We add DIF_VARIABLE_MAX to t_did to assure that the thread key is never
334  * equal to a variable identifier.  This is necessary (but not sufficient) to
335  * assure that global associative arrays never collide with thread-local
336  * variables.  To guarantee that they cannot collide, we must also define the
337  * order for keying dynamic variables.  That order is:
338  *
339  *   [ key0 ] ... [ keyn ] [ variable-key ] [ tls-key ]
340  *
341  * Because the variable-key and the tls-key are in orthogonal spaces, there is
342  * no way for a global variable key signature to match a thread-local key
343  * signature.
344  */
345 #define	DTRACE_TLS_THRKEY(where) { \
346 	uint_t intr = 0; \
347 	uint_t actv = CPU->cpu_intr_actv >> (LOCK_LEVEL + 1); \
348 	for (; actv; actv >>= 1) \
349 		intr++; \
350 	ASSERT(intr < (1 << 3)); \
351 	(where) = ((curthread->t_did + DIF_VARIABLE_MAX) & \
352 	    (((uint64_t)1 << 61) - 1)) | ((uint64_t)intr << 61); \
353 }
354 
355 #define	DT_BSWAP_8(x)	((x) & 0xff)
356 #define	DT_BSWAP_16(x)	((DT_BSWAP_8(x) << 8) | DT_BSWAP_8((x) >> 8))
357 #define	DT_BSWAP_32(x)	((DT_BSWAP_16(x) << 16) | DT_BSWAP_16((x) >> 16))
358 #define	DT_BSWAP_64(x)	((DT_BSWAP_32(x) << 32) | DT_BSWAP_32((x) >> 32))
359 
360 #define	DT_MASK_LO 0x00000000FFFFFFFFULL
361 
362 #define	DTRACE_STORE(type, tomax, offset, what) \
363 	*((type *)((uintptr_t)(tomax) + (uintptr_t)offset)) = (type)(what);
364 
365 #ifndef __x86
366 #define	DTRACE_ALIGNCHECK(addr, size, flags)				\
367 	if (addr & (size - 1)) {					\
368 		*flags |= CPU_DTRACE_BADALIGN;				\
369 		cpu_core[CPU->cpu_id].cpuc_dtrace_illval = addr;	\
370 		return (0);						\
371 	}
372 #else
373 #define	DTRACE_ALIGNCHECK(addr, size, flags)
374 #endif
375 
376 /*
377  * Test whether a range of memory starting at testaddr of size testsz falls
378  * within the range of memory described by addr, sz.  We take care to avoid
379  * problems with overflow and underflow of the unsigned quantities, and
380  * disallow all negative sizes.  Ranges of size 0 are allowed.
381  */
382 #define	DTRACE_INRANGE(testaddr, testsz, baseaddr, basesz) \
383 	((testaddr) - (uintptr_t)(baseaddr) < (basesz) && \
384 	(testaddr) + (testsz) - (uintptr_t)(baseaddr) <= (basesz) && \
385 	(testaddr) + (testsz) >= (testaddr))
386 
387 #define	DTRACE_RANGE_REMAIN(remp, addr, baseaddr, basesz)		\
388 do {									\
389 	if ((remp) != NULL) {						\
390 		*(remp) = (uintptr_t)(baseaddr) + (basesz) - (addr);	\
391 	}								\
392 _NOTE(CONSTCOND) } while (0)
393 
394 
395 /*
396  * Test whether alloc_sz bytes will fit in the scratch region.  We isolate
397  * alloc_sz on the righthand side of the comparison in order to avoid overflow
398  * or underflow in the comparison with it.  This is simpler than the INRANGE
399  * check above, because we know that the dtms_scratch_ptr is valid in the
400  * range.  Allocations of size zero are allowed.
401  */
402 #define	DTRACE_INSCRATCH(mstate, alloc_sz) \
403 	((mstate)->dtms_scratch_base + (mstate)->dtms_scratch_size - \
404 	(mstate)->dtms_scratch_ptr >= (alloc_sz))
405 
406 #define	DTRACE_LOADFUNC(bits)						\
407 /*CSTYLED*/								\
408 uint##bits##_t								\
409 dtrace_load##bits(uintptr_t addr)					\
410 {									\
411 	size_t size = bits / NBBY;					\
412 	/*CSTYLED*/							\
413 	uint##bits##_t rval;						\
414 	int i;								\
415 	volatile uint16_t *flags = (volatile uint16_t *)		\
416 	    &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;			\
417 									\
418 	DTRACE_ALIGNCHECK(addr, size, flags);				\
419 									\
420 	for (i = 0; i < dtrace_toxranges; i++) {			\
421 		if (addr >= dtrace_toxrange[i].dtt_limit)		\
422 			continue;					\
423 									\
424 		if (addr + size <= dtrace_toxrange[i].dtt_base)		\
425 			continue;					\
426 									\
427 		/*							\
428 		 * This address falls within a toxic region; return 0.	\
429 		 */							\
430 		*flags |= CPU_DTRACE_BADADDR;				\
431 		cpu_core[CPU->cpu_id].cpuc_dtrace_illval = addr;	\
432 		return (0);						\
433 	}								\
434 									\
435 	*flags |= CPU_DTRACE_NOFAULT;					\
436 	/*CSTYLED*/							\
437 	rval = *((volatile uint##bits##_t *)addr);			\
438 	*flags &= ~CPU_DTRACE_NOFAULT;					\
439 									\
440 	return (!(*flags & CPU_DTRACE_FAULT) ? rval : 0);		\
441 }
442 
443 #ifdef _LP64
444 #define	dtrace_loadptr	dtrace_load64
445 #else
446 #define	dtrace_loadptr	dtrace_load32
447 #endif
448 
449 #define	DTRACE_DYNHASH_FREE	0
450 #define	DTRACE_DYNHASH_SINK	1
451 #define	DTRACE_DYNHASH_VALID	2
452 
453 #define	DTRACE_MATCH_FAIL	-1
454 #define	DTRACE_MATCH_NEXT	0
455 #define	DTRACE_MATCH_DONE	1
456 #define	DTRACE_ANCHORED(probe)	((probe)->dtpr_func[0] != '\0')
457 #define	DTRACE_STATE_ALIGN	64
458 
459 #define	DTRACE_FLAGS2FLT(flags)						\
460 	(((flags) & CPU_DTRACE_BADADDR) ? DTRACEFLT_BADADDR :		\
461 	((flags) & CPU_DTRACE_ILLOP) ? DTRACEFLT_ILLOP :		\
462 	((flags) & CPU_DTRACE_DIVZERO) ? DTRACEFLT_DIVZERO :		\
463 	((flags) & CPU_DTRACE_KPRIV) ? DTRACEFLT_KPRIV :		\
464 	((flags) & CPU_DTRACE_UPRIV) ? DTRACEFLT_UPRIV :		\
465 	((flags) & CPU_DTRACE_TUPOFLOW) ?  DTRACEFLT_TUPOFLOW :		\
466 	((flags) & CPU_DTRACE_BADALIGN) ?  DTRACEFLT_BADALIGN :		\
467 	((flags) & CPU_DTRACE_NOSCRATCH) ?  DTRACEFLT_NOSCRATCH :	\
468 	((flags) & CPU_DTRACE_BADSTACK) ?  DTRACEFLT_BADSTACK :		\
469 	DTRACEFLT_UNKNOWN)
470 
471 #define	DTRACEACT_ISSTRING(act)						\
472 	((act)->dta_kind == DTRACEACT_DIFEXPR &&			\
473 	(act)->dta_difo->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING)
474 
475 static size_t dtrace_strlen(const char *, size_t);
476 static dtrace_probe_t *dtrace_probe_lookup_id(dtrace_id_t id);
477 static void dtrace_enabling_provide(dtrace_provider_t *);
478 static int dtrace_enabling_match(dtrace_enabling_t *, int *);
479 static void dtrace_enabling_matchall(void);
480 static void dtrace_enabling_reap(void);
481 static dtrace_state_t *dtrace_anon_grab(void);
482 static uint64_t dtrace_helper(int, dtrace_mstate_t *,
483     dtrace_state_t *, uint64_t, uint64_t);
484 static dtrace_helpers_t *dtrace_helpers_create(proc_t *);
485 static void dtrace_buffer_drop(dtrace_buffer_t *);
486 static int dtrace_buffer_consumed(dtrace_buffer_t *, hrtime_t when);
487 static intptr_t dtrace_buffer_reserve(dtrace_buffer_t *, size_t, size_t,
488     dtrace_state_t *, dtrace_mstate_t *);
489 static int dtrace_state_option(dtrace_state_t *, dtrace_optid_t,
490     dtrace_optval_t);
491 static int dtrace_ecb_create_enable(dtrace_probe_t *, void *);
492 static void dtrace_helper_provider_destroy(dtrace_helper_provider_t *);
493 static int dtrace_priv_proc(dtrace_state_t *, dtrace_mstate_t *);
494 static void dtrace_getf_barrier(void);
495 static int dtrace_canload_remains(uint64_t, size_t, size_t *,
496     dtrace_mstate_t *, dtrace_vstate_t *);
497 static int dtrace_canstore_remains(uint64_t, size_t, size_t *,
498     dtrace_mstate_t *, dtrace_vstate_t *);
499 
500 /*
501  * DTrace Probe Context Functions
502  *
503  * These functions are called from probe context.  Because probe context is
504  * any context in which C may be called, arbitrarily locks may be held,
505  * interrupts may be disabled, we may be in arbitrary dispatched state, etc.
506  * As a result, functions called from probe context may only call other DTrace
507  * support functions -- they may not interact at all with the system at large.
508  * (Note that the ASSERT macro is made probe-context safe by redefining it in
509  * terms of dtrace_assfail(), a probe-context safe function.) If arbitrary
510  * loads are to be performed from probe context, they _must_ be in terms of
511  * the safe dtrace_load*() variants.
512  *
513  * Some functions in this block are not actually called from probe context;
514  * for these functions, there will be a comment above the function reading
515  * "Note:  not called from probe context."
516  */
517 void
518 dtrace_panic(const char *format, ...)
519 {
520 	va_list alist;
521 
522 	va_start(alist, format);
523 	dtrace_vpanic(format, alist);
524 	va_end(alist);
525 }
526 
527 int
528 dtrace_assfail(const char *a, const char *f, int l)
529 {
530 	dtrace_panic("assertion failed: %s, file: %s, line: %d", a, f, l);
531 
532 	/*
533 	 * We just need something here that even the most clever compiler
534 	 * cannot optimize away.
535 	 */
536 	return (a[(uintptr_t)f]);
537 }
538 
539 /*
540  * Atomically increment a specified error counter from probe context.
541  */
542 static void
543 dtrace_error(uint32_t *counter)
544 {
545 	/*
546 	 * Most counters stored to in probe context are per-CPU counters.
547 	 * However, there are some error conditions that are sufficiently
548 	 * arcane that they don't merit per-CPU storage.  If these counters
549 	 * are incremented concurrently on different CPUs, scalability will be
550 	 * adversely affected -- but we don't expect them to be white-hot in a
551 	 * correctly constructed enabling...
552 	 */
553 	uint32_t oval, nval;
554 
555 	do {
556 		oval = *counter;
557 
558 		if ((nval = oval + 1) == 0) {
559 			/*
560 			 * If the counter would wrap, set it to 1 -- assuring
561 			 * that the counter is never zero when we have seen
562 			 * errors.  (The counter must be 32-bits because we
563 			 * aren't guaranteed a 64-bit compare&swap operation.)
564 			 * To save this code both the infamy of being fingered
565 			 * by a priggish news story and the indignity of being
566 			 * the target of a neo-puritan witch trial, we're
567 			 * carefully avoiding any colorful description of the
568 			 * likelihood of this condition -- but suffice it to
569 			 * say that it is only slightly more likely than the
570 			 * overflow of predicate cache IDs, as discussed in
571 			 * dtrace_predicate_create().
572 			 */
573 			nval = 1;
574 		}
575 	} while (dtrace_cas32(counter, oval, nval) != oval);
576 }
577 
578 /*
579  * Use the DTRACE_LOADFUNC macro to define functions for each of loading a
580  * uint8_t, a uint16_t, a uint32_t and a uint64_t.
581  */
582 /* BEGIN CSTYLED */
583 DTRACE_LOADFUNC(8)
584 DTRACE_LOADFUNC(16)
585 DTRACE_LOADFUNC(32)
586 DTRACE_LOADFUNC(64)
587 /* END CSTYLED */
588 
589 static int
590 dtrace_inscratch(uintptr_t dest, size_t size, dtrace_mstate_t *mstate)
591 {
592 	if (dest < mstate->dtms_scratch_base)
593 		return (0);
594 
595 	if (dest + size < dest)
596 		return (0);
597 
598 	if (dest + size > mstate->dtms_scratch_ptr)
599 		return (0);
600 
601 	return (1);
602 }
603 
604 static int
605 dtrace_canstore_statvar(uint64_t addr, size_t sz, size_t *remain,
606     dtrace_statvar_t **svars, int nsvars)
607 {
608 	int i;
609 	size_t maxglobalsize, maxlocalsize;
610 
611 	if (nsvars == 0)
612 		return (0);
613 
614 	maxglobalsize = dtrace_statvar_maxsize + sizeof (uint64_t);
615 	maxlocalsize = maxglobalsize * NCPU;
616 
617 	for (i = 0; i < nsvars; i++) {
618 		dtrace_statvar_t *svar = svars[i];
619 		uint8_t scope;
620 		size_t size;
621 
622 		if (svar == NULL || (size = svar->dtsv_size) == 0)
623 			continue;
624 
625 		scope = svar->dtsv_var.dtdv_scope;
626 
627 		/*
628 		 * We verify that our size is valid in the spirit of providing
629 		 * defense in depth:  we want to prevent attackers from using
630 		 * DTrace to escalate an orthogonal kernel heap corruption bug
631 		 * into the ability to store to arbitrary locations in memory.
632 		 */
633 		VERIFY((scope == DIFV_SCOPE_GLOBAL && size <= maxglobalsize) ||
634 		    (scope == DIFV_SCOPE_LOCAL && size <= maxlocalsize));
635 
636 		if (DTRACE_INRANGE(addr, sz, svar->dtsv_data,
637 		    svar->dtsv_size)) {
638 			DTRACE_RANGE_REMAIN(remain, addr, svar->dtsv_data,
639 			    svar->dtsv_size);
640 			return (1);
641 		}
642 	}
643 
644 	return (0);
645 }
646 
647 /*
648  * Check to see if the address is within a memory region to which a store may
649  * be issued.  This includes the DTrace scratch areas, and any DTrace variable
650  * region.  The caller of dtrace_canstore() is responsible for performing any
651  * alignment checks that are needed before stores are actually executed.
652  */
653 static int
654 dtrace_canstore(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
655     dtrace_vstate_t *vstate)
656 {
657 	return (dtrace_canstore_remains(addr, sz, NULL, mstate, vstate));
658 }
659 
660 /*
661  * Implementation of dtrace_canstore which communicates the upper bound of the
662  * allowed memory region.
663  */
664 static int
665 dtrace_canstore_remains(uint64_t addr, size_t sz, size_t *remain,
666     dtrace_mstate_t *mstate, dtrace_vstate_t *vstate)
667 {
668 	/*
669 	 * First, check to see if the address is in scratch space...
670 	 */
671 	if (DTRACE_INRANGE(addr, sz, mstate->dtms_scratch_base,
672 	    mstate->dtms_scratch_size)) {
673 		DTRACE_RANGE_REMAIN(remain, addr, mstate->dtms_scratch_base,
674 		    mstate->dtms_scratch_size);
675 		return (1);
676 	}
677 
678 	/*
679 	 * Now check to see if it's a dynamic variable.  This check will pick
680 	 * up both thread-local variables and any global dynamically-allocated
681 	 * variables.
682 	 */
683 	if (DTRACE_INRANGE(addr, sz, vstate->dtvs_dynvars.dtds_base,
684 	    vstate->dtvs_dynvars.dtds_size)) {
685 		dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
686 		uintptr_t base = (uintptr_t)dstate->dtds_base +
687 		    (dstate->dtds_hashsize * sizeof (dtrace_dynhash_t));
688 		uintptr_t chunkoffs;
689 		dtrace_dynvar_t *dvar;
690 
691 		/*
692 		 * Before we assume that we can store here, we need to make
693 		 * sure that it isn't in our metadata -- storing to our
694 		 * dynamic variable metadata would corrupt our state.  For
695 		 * the range to not include any dynamic variable metadata,
696 		 * it must:
697 		 *
698 		 *	(1) Start above the hash table that is at the base of
699 		 *	the dynamic variable space
700 		 *
701 		 *	(2) Have a starting chunk offset that is beyond the
702 		 *	dtrace_dynvar_t that is at the base of every chunk
703 		 *
704 		 *	(3) Not span a chunk boundary
705 		 *
706 		 *	(4) Not be in the tuple space of a dynamic variable
707 		 *
708 		 */
709 		if (addr < base)
710 			return (0);
711 
712 		chunkoffs = (addr - base) % dstate->dtds_chunksize;
713 
714 		if (chunkoffs < sizeof (dtrace_dynvar_t))
715 			return (0);
716 
717 		if (chunkoffs + sz > dstate->dtds_chunksize)
718 			return (0);
719 
720 		dvar = (dtrace_dynvar_t *)((uintptr_t)addr - chunkoffs);
721 
722 		if (dvar->dtdv_hashval == DTRACE_DYNHASH_FREE)
723 			return (0);
724 
725 		if (chunkoffs < sizeof (dtrace_dynvar_t) +
726 		    ((dvar->dtdv_tuple.dtt_nkeys - 1) * sizeof (dtrace_key_t)))
727 			return (0);
728 
729 		DTRACE_RANGE_REMAIN(remain, addr, dvar, dstate->dtds_chunksize);
730 		return (1);
731 	}
732 
733 	/*
734 	 * Finally, check the static local and global variables.  These checks
735 	 * take the longest, so we perform them last.
736 	 */
737 	if (dtrace_canstore_statvar(addr, sz, remain,
738 	    vstate->dtvs_locals, vstate->dtvs_nlocals))
739 		return (1);
740 
741 	if (dtrace_canstore_statvar(addr, sz, remain,
742 	    vstate->dtvs_globals, vstate->dtvs_nglobals))
743 		return (1);
744 
745 	return (0);
746 }
747 
748 
749 /*
750  * Convenience routine to check to see if the address is within a memory
751  * region in which a load may be issued given the user's privilege level;
752  * if not, it sets the appropriate error flags and loads 'addr' into the
753  * illegal value slot.
754  *
755  * DTrace subroutines (DIF_SUBR_*) should use this helper to implement
756  * appropriate memory access protection.
757  */
758 static int
759 dtrace_canload(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
760     dtrace_vstate_t *vstate)
761 {
762 	return (dtrace_canload_remains(addr, sz, NULL, mstate, vstate));
763 }
764 
765 /*
766  * Implementation of dtrace_canload which communicates the upper bound of the
767  * allowed memory region.
768  */
769 static int
770 dtrace_canload_remains(uint64_t addr, size_t sz, size_t *remain,
771     dtrace_mstate_t *mstate, dtrace_vstate_t *vstate)
772 {
773 	volatile uintptr_t *illval = &cpu_core[CPU->cpu_id].cpuc_dtrace_illval;
774 	file_t *fp;
775 
776 	/*
777 	 * If we hold the privilege to read from kernel memory, then
778 	 * everything is readable.
779 	 */
780 	if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0) {
781 		DTRACE_RANGE_REMAIN(remain, addr, addr, sz);
782 		return (1);
783 	}
784 
785 	/*
786 	 * You can obviously read that which you can store.
787 	 */
788 	if (dtrace_canstore_remains(addr, sz, remain, mstate, vstate))
789 		return (1);
790 
791 	/*
792 	 * We're allowed to read from our own string table.
793 	 */
794 	if (DTRACE_INRANGE(addr, sz, mstate->dtms_difo->dtdo_strtab,
795 	    mstate->dtms_difo->dtdo_strlen)) {
796 		DTRACE_RANGE_REMAIN(remain, addr,
797 		    mstate->dtms_difo->dtdo_strtab,
798 		    mstate->dtms_difo->dtdo_strlen);
799 		return (1);
800 	}
801 
802 	if (vstate->dtvs_state != NULL &&
803 	    dtrace_priv_proc(vstate->dtvs_state, mstate)) {
804 		proc_t *p;
805 
806 		/*
807 		 * When we have privileges to the current process, there are
808 		 * several context-related kernel structures that are safe to
809 		 * read, even absent the privilege to read from kernel memory.
810 		 * These reads are safe because these structures contain only
811 		 * state that (1) we're permitted to read, (2) is harmless or
812 		 * (3) contains pointers to additional kernel state that we're
813 		 * not permitted to read (and as such, do not present an
814 		 * opportunity for privilege escalation).  Finally (and
815 		 * critically), because of the nature of their relation with
816 		 * the current thread context, the memory associated with these
817 		 * structures cannot change over the duration of probe context,
818 		 * and it is therefore impossible for this memory to be
819 		 * deallocated and reallocated as something else while it's
820 		 * being operated upon.
821 		 */
822 		if (DTRACE_INRANGE(addr, sz, curthread, sizeof (kthread_t))) {
823 			DTRACE_RANGE_REMAIN(remain, addr, curthread,
824 			    sizeof (kthread_t));
825 			return (1);
826 		}
827 
828 		if ((p = curthread->t_procp) != NULL && DTRACE_INRANGE(addr,
829 		    sz, curthread->t_procp, sizeof (proc_t))) {
830 			DTRACE_RANGE_REMAIN(remain, addr, curthread->t_procp,
831 			    sizeof (proc_t));
832 			return (1);
833 		}
834 
835 		if (curthread->t_cred != NULL && DTRACE_INRANGE(addr, sz,
836 		    curthread->t_cred, sizeof (cred_t))) {
837 			DTRACE_RANGE_REMAIN(remain, addr, curthread->t_cred,
838 			    sizeof (cred_t));
839 			return (1);
840 		}
841 
842 		if (p != NULL && p->p_pidp != NULL && DTRACE_INRANGE(addr, sz,
843 		    &(p->p_pidp->pid_id), sizeof (pid_t))) {
844 			DTRACE_RANGE_REMAIN(remain, addr, &(p->p_pidp->pid_id),
845 			    sizeof (pid_t));
846 			return (1);
847 		}
848 
849 		if (curthread->t_cpu != NULL && DTRACE_INRANGE(addr, sz,
850 		    curthread->t_cpu, offsetof(cpu_t, cpu_pause_thread))) {
851 			DTRACE_RANGE_REMAIN(remain, addr, curthread->t_cpu,
852 			    offsetof(cpu_t, cpu_pause_thread));
853 			return (1);
854 		}
855 	}
856 
857 	if ((fp = mstate->dtms_getf) != NULL) {
858 		uintptr_t psz = sizeof (void *);
859 		vnode_t *vp;
860 		vnodeops_t *op;
861 
862 		/*
863 		 * When getf() returns a file_t, the enabling is implicitly
864 		 * granted the (transient) right to read the returned file_t
865 		 * as well as the v_path and v_op->vnop_name of the underlying
866 		 * vnode.  These accesses are allowed after a successful
867 		 * getf() because the members that they refer to cannot change
868 		 * once set -- and the barrier logic in the kernel's closef()
869 		 * path assures that the file_t and its referenced vode_t
870 		 * cannot themselves be stale (that is, it impossible for
871 		 * either dtms_getf itself or its f_vnode member to reference
872 		 * freed memory).
873 		 */
874 		if (DTRACE_INRANGE(addr, sz, fp, sizeof (file_t))) {
875 			DTRACE_RANGE_REMAIN(remain, addr, fp, sizeof (file_t));
876 			return (1);
877 		}
878 
879 		if ((vp = fp->f_vnode) != NULL) {
880 			size_t slen;
881 
882 			if (DTRACE_INRANGE(addr, sz, &vp->v_path, psz)) {
883 				DTRACE_RANGE_REMAIN(remain, addr, &vp->v_path,
884 				    psz);
885 				return (1);
886 			}
887 
888 			slen = strlen(vp->v_path) + 1;
889 			if (DTRACE_INRANGE(addr, sz, vp->v_path, slen)) {
890 				DTRACE_RANGE_REMAIN(remain, addr, vp->v_path,
891 				    slen);
892 				return (1);
893 			}
894 
895 			if (DTRACE_INRANGE(addr, sz, &vp->v_op, psz)) {
896 				DTRACE_RANGE_REMAIN(remain, addr, &vp->v_op,
897 				    psz);
898 				return (1);
899 			}
900 
901 			if ((op = vp->v_op) != NULL &&
902 			    DTRACE_INRANGE(addr, sz, &op->vnop_name, psz)) {
903 				DTRACE_RANGE_REMAIN(remain, addr,
904 				    &op->vnop_name, psz);
905 				return (1);
906 			}
907 
908 			if (op != NULL && op->vnop_name != NULL &&
909 			    DTRACE_INRANGE(addr, sz, op->vnop_name,
910 			    (slen = strlen(op->vnop_name) + 1))) {
911 				DTRACE_RANGE_REMAIN(remain, addr,
912 				    op->vnop_name, slen);
913 				return (1);
914 			}
915 		}
916 	}
917 
918 	DTRACE_CPUFLAG_SET(CPU_DTRACE_KPRIV);
919 	*illval = addr;
920 	return (0);
921 }
922 
923 /*
924  * Convenience routine to check to see if a given string is within a memory
925  * region in which a load may be issued given the user's privilege level;
926  * this exists so that we don't need to issue unnecessary dtrace_strlen()
927  * calls in the event that the user has all privileges.
928  */
929 static int
930 dtrace_strcanload(uint64_t addr, size_t sz, size_t *remain,
931     dtrace_mstate_t *mstate, dtrace_vstate_t *vstate)
932 {
933 	size_t rsize;
934 
935 	/*
936 	 * If we hold the privilege to read from kernel memory, then
937 	 * everything is readable.
938 	 */
939 	if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0) {
940 		DTRACE_RANGE_REMAIN(remain, addr, addr, sz);
941 		return (1);
942 	}
943 
944 	/*
945 	 * Even if the caller is uninterested in querying the remaining valid
946 	 * range, it is required to ensure that the access is allowed.
947 	 */
948 	if (remain == NULL) {
949 		remain = &rsize;
950 	}
951 	if (dtrace_canload_remains(addr, 0, remain, mstate, vstate)) {
952 		size_t strsz;
953 		/*
954 		 * Perform the strlen after determining the length of the
955 		 * memory region which is accessible.  This prevents timing
956 		 * information from being used to find NULs in memory which is
957 		 * not accessible to the caller.
958 		 */
959 		strsz = 1 + dtrace_strlen((char *)(uintptr_t)addr,
960 		    MIN(sz, *remain));
961 		if (strsz <= *remain) {
962 			return (1);
963 		}
964 	}
965 
966 	return (0);
967 }
968 
969 /*
970  * Convenience routine to check to see if a given variable is within a memory
971  * region in which a load may be issued given the user's privilege level.
972  */
973 static int
974 dtrace_vcanload(void *src, dtrace_diftype_t *type, size_t *remain,
975     dtrace_mstate_t *mstate, dtrace_vstate_t *vstate)
976 {
977 	size_t sz;
978 	ASSERT(type->dtdt_flags & DIF_TF_BYREF);
979 
980 	/*
981 	 * Calculate the max size before performing any checks since even
982 	 * DTRACE_ACCESS_KERNEL-credentialed callers expect that this function
983 	 * return the max length via 'remain'.
984 	 */
985 	if (type->dtdt_kind == DIF_TYPE_STRING) {
986 		dtrace_state_t *state = vstate->dtvs_state;
987 
988 		if (state != NULL) {
989 			sz = state->dts_options[DTRACEOPT_STRSIZE];
990 		} else {
991 			/*
992 			 * In helper context, we have a NULL state; fall back
993 			 * to using the system-wide default for the string size
994 			 * in this case.
995 			 */
996 			sz = dtrace_strsize_default;
997 		}
998 	} else {
999 		sz = type->dtdt_size;
1000 	}
1001 
1002 	/*
1003 	 * If we hold the privilege to read from kernel memory, then
1004 	 * everything is readable.
1005 	 */
1006 	if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0) {
1007 		DTRACE_RANGE_REMAIN(remain, (uintptr_t)src, src, sz);
1008 		return (1);
1009 	}
1010 
1011 	if (type->dtdt_kind == DIF_TYPE_STRING) {
1012 		return (dtrace_strcanload((uintptr_t)src, sz, remain, mstate,
1013 		    vstate));
1014 	}
1015 	return (dtrace_canload_remains((uintptr_t)src, sz, remain, mstate,
1016 	    vstate));
1017 }
1018 
1019 /*
1020  * Convert a string to a signed integer using safe loads.
1021  *
1022  * NOTE: This function uses various macros from strtolctype.h to manipulate
1023  * digit values, etc -- these have all been checked to ensure they make
1024  * no additional function calls.
1025  */
1026 static int64_t
1027 dtrace_strtoll(char *input, int base, size_t limit)
1028 {
1029 	uintptr_t pos = (uintptr_t)input;
1030 	int64_t val = 0;
1031 	int x;
1032 	boolean_t neg = B_FALSE;
1033 	char c, cc, ccc;
1034 	uintptr_t end = pos + limit;
1035 
1036 	/*
1037 	 * Consume any whitespace preceding digits.
1038 	 */
1039 	while ((c = dtrace_load8(pos)) == ' ' || c == '\t')
1040 		pos++;
1041 
1042 	/*
1043 	 * Handle an explicit sign if one is present.
1044 	 */
1045 	if (c == '-' || c == '+') {
1046 		if (c == '-')
1047 			neg = B_TRUE;
1048 		c = dtrace_load8(++pos);
1049 	}
1050 
1051 	/*
1052 	 * Check for an explicit hexadecimal prefix ("0x" or "0X") and skip it
1053 	 * if present.
1054 	 */
1055 	if (base == 16 && c == '0' && ((cc = dtrace_load8(pos + 1)) == 'x' ||
1056 	    cc == 'X') && isxdigit(ccc = dtrace_load8(pos + 2))) {
1057 		pos += 2;
1058 		c = ccc;
1059 	}
1060 
1061 	/*
1062 	 * Read in contiguous digits until the first non-digit character.
1063 	 */
1064 	for (; pos < end && c != '\0' && lisalnum(c) && (x = DIGIT(c)) < base;
1065 	    c = dtrace_load8(++pos))
1066 		val = val * base + x;
1067 
1068 	return (neg ? -val : val);
1069 }
1070 
1071 /*
1072  * Compare two strings using safe loads.
1073  */
1074 static int
1075 dtrace_strncmp(char *s1, char *s2, size_t limit)
1076 {
1077 	uint8_t c1, c2;
1078 	volatile uint16_t *flags;
1079 
1080 	if (s1 == s2 || limit == 0)
1081 		return (0);
1082 
1083 	flags = (volatile uint16_t *)&cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
1084 
1085 	do {
1086 		if (s1 == NULL) {
1087 			c1 = '\0';
1088 		} else {
1089 			c1 = dtrace_load8((uintptr_t)s1++);
1090 		}
1091 
1092 		if (s2 == NULL) {
1093 			c2 = '\0';
1094 		} else {
1095 			c2 = dtrace_load8((uintptr_t)s2++);
1096 		}
1097 
1098 		if (c1 != c2)
1099 			return (c1 - c2);
1100 	} while (--limit && c1 != '\0' && !(*flags & CPU_DTRACE_FAULT));
1101 
1102 	return (0);
1103 }
1104 
1105 /*
1106  * Compute strlen(s) for a string using safe memory accesses.  The additional
1107  * len parameter is used to specify a maximum length to ensure completion.
1108  */
1109 static size_t
1110 dtrace_strlen(const char *s, size_t lim)
1111 {
1112 	uint_t len;
1113 
1114 	for (len = 0; len != lim; len++) {
1115 		if (dtrace_load8((uintptr_t)s++) == '\0')
1116 			break;
1117 	}
1118 
1119 	return (len);
1120 }
1121 
1122 /*
1123  * Check if an address falls within a toxic region.
1124  */
1125 static int
1126 dtrace_istoxic(uintptr_t kaddr, size_t size)
1127 {
1128 	uintptr_t taddr, tsize;
1129 	int i;
1130 
1131 	for (i = 0; i < dtrace_toxranges; i++) {
1132 		taddr = dtrace_toxrange[i].dtt_base;
1133 		tsize = dtrace_toxrange[i].dtt_limit - taddr;
1134 
1135 		if (kaddr - taddr < tsize) {
1136 			DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
1137 			cpu_core[CPU->cpu_id].cpuc_dtrace_illval = kaddr;
1138 			return (1);
1139 		}
1140 
1141 		if (taddr - kaddr < size) {
1142 			DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
1143 			cpu_core[CPU->cpu_id].cpuc_dtrace_illval = taddr;
1144 			return (1);
1145 		}
1146 	}
1147 
1148 	return (0);
1149 }
1150 
1151 /*
1152  * Copy src to dst using safe memory accesses.  The src is assumed to be unsafe
1153  * memory specified by the DIF program.  The dst is assumed to be safe memory
1154  * that we can store to directly because it is managed by DTrace.  As with
1155  * standard bcopy, overlapping copies are handled properly.
1156  */
1157 static void
1158 dtrace_bcopy(const void *src, void *dst, size_t len)
1159 {
1160 	if (len != 0) {
1161 		uint8_t *s1 = dst;
1162 		const uint8_t *s2 = src;
1163 
1164 		if (s1 <= s2) {
1165 			do {
1166 				*s1++ = dtrace_load8((uintptr_t)s2++);
1167 			} while (--len != 0);
1168 		} else {
1169 			s2 += len;
1170 			s1 += len;
1171 
1172 			do {
1173 				*--s1 = dtrace_load8((uintptr_t)--s2);
1174 			} while (--len != 0);
1175 		}
1176 	}
1177 }
1178 
1179 /*
1180  * Copy src to dst using safe memory accesses, up to either the specified
1181  * length, or the point that a nul byte is encountered.  The src is assumed to
1182  * be unsafe memory specified by the DIF program.  The dst is assumed to be
1183  * safe memory that we can store to directly because it is managed by DTrace.
1184  * Unlike dtrace_bcopy(), overlapping regions are not handled.
1185  */
1186 static void
1187 dtrace_strcpy(const void *src, void *dst, size_t len)
1188 {
1189 	if (len != 0) {
1190 		uint8_t *s1 = dst, c;
1191 		const uint8_t *s2 = src;
1192 
1193 		do {
1194 			*s1++ = c = dtrace_load8((uintptr_t)s2++);
1195 		} while (--len != 0 && c != '\0');
1196 	}
1197 }
1198 
1199 /*
1200  * Copy src to dst, deriving the size and type from the specified (BYREF)
1201  * variable type.  The src is assumed to be unsafe memory specified by the DIF
1202  * program.  The dst is assumed to be DTrace variable memory that is of the
1203  * specified type; we assume that we can store to directly.
1204  */
1205 static void
1206 dtrace_vcopy(void *src, void *dst, dtrace_diftype_t *type, size_t limit)
1207 {
1208 	ASSERT(type->dtdt_flags & DIF_TF_BYREF);
1209 
1210 	if (type->dtdt_kind == DIF_TYPE_STRING) {
1211 		dtrace_strcpy(src, dst, MIN(type->dtdt_size, limit));
1212 	} else {
1213 		dtrace_bcopy(src, dst, MIN(type->dtdt_size, limit));
1214 	}
1215 }
1216 
1217 /*
1218  * Compare s1 to s2 using safe memory accesses.  The s1 data is assumed to be
1219  * unsafe memory specified by the DIF program.  The s2 data is assumed to be
1220  * safe memory that we can access directly because it is managed by DTrace.
1221  */
1222 static int
1223 dtrace_bcmp(const void *s1, const void *s2, size_t len)
1224 {
1225 	volatile uint16_t *flags;
1226 
1227 	flags = (volatile uint16_t *)&cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
1228 
1229 	if (s1 == s2)
1230 		return (0);
1231 
1232 	if (s1 == NULL || s2 == NULL)
1233 		return (1);
1234 
1235 	if (s1 != s2 && len != 0) {
1236 		const uint8_t *ps1 = s1;
1237 		const uint8_t *ps2 = s2;
1238 
1239 		do {
1240 			if (dtrace_load8((uintptr_t)ps1++) != *ps2++)
1241 				return (1);
1242 		} while (--len != 0 && !(*flags & CPU_DTRACE_FAULT));
1243 	}
1244 	return (0);
1245 }
1246 
1247 /*
1248  * Zero the specified region using a simple byte-by-byte loop.  Note that this
1249  * is for safe DTrace-managed memory only.
1250  */
1251 static void
1252 dtrace_bzero(void *dst, size_t len)
1253 {
1254 	uchar_t *cp;
1255 
1256 	for (cp = dst; len != 0; len--)
1257 		*cp++ = 0;
1258 }
1259 
1260 static void
1261 dtrace_add_128(uint64_t *addend1, uint64_t *addend2, uint64_t *sum)
1262 {
1263 	uint64_t result[2];
1264 
1265 	result[0] = addend1[0] + addend2[0];
1266 	result[1] = addend1[1] + addend2[1] +
1267 	    (result[0] < addend1[0] || result[0] < addend2[0] ? 1 : 0);
1268 
1269 	sum[0] = result[0];
1270 	sum[1] = result[1];
1271 }
1272 
1273 /*
1274  * Shift the 128-bit value in a by b. If b is positive, shift left.
1275  * If b is negative, shift right.
1276  */
1277 static void
1278 dtrace_shift_128(uint64_t *a, int b)
1279 {
1280 	uint64_t mask;
1281 
1282 	if (b == 0)
1283 		return;
1284 
1285 	if (b < 0) {
1286 		b = -b;
1287 		if (b >= 64) {
1288 			a[0] = a[1] >> (b - 64);
1289 			a[1] = 0;
1290 		} else {
1291 			a[0] >>= b;
1292 			mask = 1LL << (64 - b);
1293 			mask -= 1;
1294 			a[0] |= ((a[1] & mask) << (64 - b));
1295 			a[1] >>= b;
1296 		}
1297 	} else {
1298 		if (b >= 64) {
1299 			a[1] = a[0] << (b - 64);
1300 			a[0] = 0;
1301 		} else {
1302 			a[1] <<= b;
1303 			mask = a[0] >> (64 - b);
1304 			a[1] |= mask;
1305 			a[0] <<= b;
1306 		}
1307 	}
1308 }
1309 
1310 /*
1311  * The basic idea is to break the 2 64-bit values into 4 32-bit values,
1312  * use native multiplication on those, and then re-combine into the
1313  * resulting 128-bit value.
1314  *
1315  * (hi1 << 32 + lo1) * (hi2 << 32 + lo2) =
1316  *     hi1 * hi2 << 64 +
1317  *     hi1 * lo2 << 32 +
1318  *     hi2 * lo1 << 32 +
1319  *     lo1 * lo2
1320  */
1321 static void
1322 dtrace_multiply_128(uint64_t factor1, uint64_t factor2, uint64_t *product)
1323 {
1324 	uint64_t hi1, hi2, lo1, lo2;
1325 	uint64_t tmp[2];
1326 
1327 	hi1 = factor1 >> 32;
1328 	hi2 = factor2 >> 32;
1329 
1330 	lo1 = factor1 & DT_MASK_LO;
1331 	lo2 = factor2 & DT_MASK_LO;
1332 
1333 	product[0] = lo1 * lo2;
1334 	product[1] = hi1 * hi2;
1335 
1336 	tmp[0] = hi1 * lo2;
1337 	tmp[1] = 0;
1338 	dtrace_shift_128(tmp, 32);
1339 	dtrace_add_128(product, tmp, product);
1340 
1341 	tmp[0] = hi2 * lo1;
1342 	tmp[1] = 0;
1343 	dtrace_shift_128(tmp, 32);
1344 	dtrace_add_128(product, tmp, product);
1345 }
1346 
1347 /*
1348  * This privilege check should be used by actions and subroutines to
1349  * verify that the user credentials of the process that enabled the
1350  * invoking ECB match the target credentials
1351  */
1352 static int
1353 dtrace_priv_proc_common_user(dtrace_state_t *state)
1354 {
1355 	cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1356 
1357 	/*
1358 	 * We should always have a non-NULL state cred here, since if cred
1359 	 * is null (anonymous tracing), we fast-path bypass this routine.
1360 	 */
1361 	ASSERT(s_cr != NULL);
1362 
1363 	if ((cr = CRED()) != NULL &&
1364 	    s_cr->cr_uid == cr->cr_uid &&
1365 	    s_cr->cr_uid == cr->cr_ruid &&
1366 	    s_cr->cr_uid == cr->cr_suid &&
1367 	    s_cr->cr_gid == cr->cr_gid &&
1368 	    s_cr->cr_gid == cr->cr_rgid &&
1369 	    s_cr->cr_gid == cr->cr_sgid)
1370 		return (1);
1371 
1372 	return (0);
1373 }
1374 
1375 /*
1376  * This privilege check should be used by actions and subroutines to
1377  * verify that the zone of the process that enabled the invoking ECB
1378  * matches the target credentials
1379  */
1380 static int
1381 dtrace_priv_proc_common_zone(dtrace_state_t *state)
1382 {
1383 	cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1384 
1385 	/*
1386 	 * We should always have a non-NULL state cred here, since if cred
1387 	 * is null (anonymous tracing), we fast-path bypass this routine.
1388 	 */
1389 	ASSERT(s_cr != NULL);
1390 
1391 	if ((cr = CRED()) != NULL && s_cr->cr_zone == cr->cr_zone)
1392 		return (1);
1393 
1394 	return (0);
1395 }
1396 
1397 /*
1398  * This privilege check should be used by actions and subroutines to
1399  * verify that the process has not setuid or changed credentials.
1400  */
1401 static int
1402 dtrace_priv_proc_common_nocd()
1403 {
1404 	proc_t *proc;
1405 
1406 	if ((proc = ttoproc(curthread)) != NULL &&
1407 	    !(proc->p_flag & SNOCD))
1408 		return (1);
1409 
1410 	return (0);
1411 }
1412 
1413 static int
1414 dtrace_priv_proc_destructive(dtrace_state_t *state, dtrace_mstate_t *mstate)
1415 {
1416 	int action = state->dts_cred.dcr_action;
1417 
1418 	if (!(mstate->dtms_access & DTRACE_ACCESS_PROC))
1419 		goto bad;
1420 
1421 	if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE) == 0) &&
1422 	    dtrace_priv_proc_common_zone(state) == 0)
1423 		goto bad;
1424 
1425 	if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER) == 0) &&
1426 	    dtrace_priv_proc_common_user(state) == 0)
1427 		goto bad;
1428 
1429 	if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG) == 0) &&
1430 	    dtrace_priv_proc_common_nocd() == 0)
1431 		goto bad;
1432 
1433 	return (1);
1434 
1435 bad:
1436 	cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1437 
1438 	return (0);
1439 }
1440 
1441 static int
1442 dtrace_priv_proc_control(dtrace_state_t *state, dtrace_mstate_t *mstate)
1443 {
1444 	if (mstate->dtms_access & DTRACE_ACCESS_PROC) {
1445 		if (state->dts_cred.dcr_action & DTRACE_CRA_PROC_CONTROL)
1446 			return (1);
1447 
1448 		if (dtrace_priv_proc_common_zone(state) &&
1449 		    dtrace_priv_proc_common_user(state) &&
1450 		    dtrace_priv_proc_common_nocd())
1451 			return (1);
1452 	}
1453 
1454 	cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1455 
1456 	return (0);
1457 }
1458 
1459 static int
1460 dtrace_priv_proc(dtrace_state_t *state, dtrace_mstate_t *mstate)
1461 {
1462 	if ((mstate->dtms_access & DTRACE_ACCESS_PROC) &&
1463 	    (state->dts_cred.dcr_action & DTRACE_CRA_PROC))
1464 		return (1);
1465 
1466 	cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1467 
1468 	return (0);
1469 }
1470 
1471 static int
1472 dtrace_priv_kernel(dtrace_state_t *state)
1473 {
1474 	if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL)
1475 		return (1);
1476 
1477 	cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1478 
1479 	return (0);
1480 }
1481 
1482 static int
1483 dtrace_priv_kernel_destructive(dtrace_state_t *state)
1484 {
1485 	if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL_DESTRUCTIVE)
1486 		return (1);
1487 
1488 	cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1489 
1490 	return (0);
1491 }
1492 
1493 /*
1494  * Determine if the dte_cond of the specified ECB allows for processing of
1495  * the current probe to continue.  Note that this routine may allow continued
1496  * processing, but with access(es) stripped from the mstate's dtms_access
1497  * field.
1498  */
1499 static int
1500 dtrace_priv_probe(dtrace_state_t *state, dtrace_mstate_t *mstate,
1501     dtrace_ecb_t *ecb)
1502 {
1503 	dtrace_probe_t *probe = ecb->dte_probe;
1504 	dtrace_provider_t *prov = probe->dtpr_provider;
1505 	dtrace_pops_t *pops = &prov->dtpv_pops;
1506 	int mode = DTRACE_MODE_NOPRIV_DROP;
1507 
1508 	ASSERT(ecb->dte_cond);
1509 
1510 	if (pops->dtps_mode != NULL) {
1511 		mode = pops->dtps_mode(prov->dtpv_arg,
1512 		    probe->dtpr_id, probe->dtpr_arg);
1513 
1514 		ASSERT(mode & (DTRACE_MODE_USER | DTRACE_MODE_KERNEL));
1515 		ASSERT(mode & (DTRACE_MODE_NOPRIV_RESTRICT |
1516 		    DTRACE_MODE_NOPRIV_DROP));
1517 	}
1518 
1519 	/*
1520 	 * If the dte_cond bits indicate that this consumer is only allowed to
1521 	 * see user-mode firings of this probe, check that the probe was fired
1522 	 * while in a user context.  If that's not the case, use the policy
1523 	 * specified by the provider to determine if we drop the probe or
1524 	 * merely restrict operation.
1525 	 */
1526 	if (ecb->dte_cond & DTRACE_COND_USERMODE) {
1527 		ASSERT(mode != DTRACE_MODE_NOPRIV_DROP);
1528 
1529 		if (!(mode & DTRACE_MODE_USER)) {
1530 			if (mode & DTRACE_MODE_NOPRIV_DROP)
1531 				return (0);
1532 
1533 			mstate->dtms_access &= ~DTRACE_ACCESS_ARGS;
1534 		}
1535 	}
1536 
1537 	/*
1538 	 * This is more subtle than it looks. We have to be absolutely certain
1539 	 * that CRED() isn't going to change out from under us so it's only
1540 	 * legit to examine that structure if we're in constrained situations.
1541 	 * Currently, the only times we'll this check is if a non-super-user
1542 	 * has enabled the profile or syscall providers -- providers that
1543 	 * allow visibility of all processes. For the profile case, the check
1544 	 * above will ensure that we're examining a user context.
1545 	 */
1546 	if (ecb->dte_cond & DTRACE_COND_OWNER) {
1547 		cred_t *cr;
1548 		cred_t *s_cr = state->dts_cred.dcr_cred;
1549 		proc_t *proc;
1550 
1551 		ASSERT(s_cr != NULL);
1552 
1553 		if ((cr = CRED()) == NULL ||
1554 		    s_cr->cr_uid != cr->cr_uid ||
1555 		    s_cr->cr_uid != cr->cr_ruid ||
1556 		    s_cr->cr_uid != cr->cr_suid ||
1557 		    s_cr->cr_gid != cr->cr_gid ||
1558 		    s_cr->cr_gid != cr->cr_rgid ||
1559 		    s_cr->cr_gid != cr->cr_sgid ||
1560 		    (proc = ttoproc(curthread)) == NULL ||
1561 		    (proc->p_flag & SNOCD)) {
1562 			if (mode & DTRACE_MODE_NOPRIV_DROP)
1563 				return (0);
1564 
1565 			mstate->dtms_access &= ~DTRACE_ACCESS_PROC;
1566 		}
1567 	}
1568 
1569 	/*
1570 	 * If our dte_cond is set to DTRACE_COND_ZONEOWNER and we are not
1571 	 * in our zone, check to see if our mode policy is to restrict rather
1572 	 * than to drop; if to restrict, strip away both DTRACE_ACCESS_PROC
1573 	 * and DTRACE_ACCESS_ARGS
1574 	 */
1575 	if (ecb->dte_cond & DTRACE_COND_ZONEOWNER) {
1576 		cred_t *cr;
1577 		cred_t *s_cr = state->dts_cred.dcr_cred;
1578 
1579 		ASSERT(s_cr != NULL);
1580 
1581 		if ((cr = CRED()) == NULL ||
1582 		    s_cr->cr_zone->zone_id != cr->cr_zone->zone_id) {
1583 			if (mode & DTRACE_MODE_NOPRIV_DROP)
1584 				return (0);
1585 
1586 			mstate->dtms_access &=
1587 			    ~(DTRACE_ACCESS_PROC | DTRACE_ACCESS_ARGS);
1588 		}
1589 	}
1590 
1591 	/*
1592 	 * By merits of being in this code path at all, we have limited
1593 	 * privileges.  If the provider has indicated that limited privileges
1594 	 * are to denote restricted operation, strip off the ability to access
1595 	 * arguments.
1596 	 */
1597 	if (mode & DTRACE_MODE_LIMITEDPRIV_RESTRICT)
1598 		mstate->dtms_access &= ~DTRACE_ACCESS_ARGS;
1599 
1600 	return (1);
1601 }
1602 
1603 /*
1604  * Note:  not called from probe context.  This function is called
1605  * asynchronously (and at a regular interval) from outside of probe context to
1606  * clean the dirty dynamic variable lists on all CPUs.  Dynamic variable
1607  * cleaning is explained in detail in <sys/dtrace_impl.h>.
1608  */
1609 void
1610 dtrace_dynvar_clean(dtrace_dstate_t *dstate)
1611 {
1612 	dtrace_dynvar_t *dirty;
1613 	dtrace_dstate_percpu_t *dcpu;
1614 	dtrace_dynvar_t **rinsep;
1615 	int i, j, work = 0;
1616 
1617 	for (i = 0; i < NCPU; i++) {
1618 		dcpu = &dstate->dtds_percpu[i];
1619 		rinsep = &dcpu->dtdsc_rinsing;
1620 
1621 		/*
1622 		 * If the dirty list is NULL, there is no dirty work to do.
1623 		 */
1624 		if (dcpu->dtdsc_dirty == NULL)
1625 			continue;
1626 
1627 		if (dcpu->dtdsc_rinsing != NULL) {
1628 			/*
1629 			 * If the rinsing list is non-NULL, then it is because
1630 			 * this CPU was selected to accept another CPU's
1631 			 * dirty list -- and since that time, dirty buffers
1632 			 * have accumulated.  This is a highly unlikely
1633 			 * condition, but we choose to ignore the dirty
1634 			 * buffers -- they'll be picked up a future cleanse.
1635 			 */
1636 			continue;
1637 		}
1638 
1639 		if (dcpu->dtdsc_clean != NULL) {
1640 			/*
1641 			 * If the clean list is non-NULL, then we're in a
1642 			 * situation where a CPU has done deallocations (we
1643 			 * have a non-NULL dirty list) but no allocations (we
1644 			 * also have a non-NULL clean list).  We can't simply
1645 			 * move the dirty list into the clean list on this
1646 			 * CPU, yet we also don't want to allow this condition
1647 			 * to persist, lest a short clean list prevent a
1648 			 * massive dirty list from being cleaned (which in
1649 			 * turn could lead to otherwise avoidable dynamic
1650 			 * drops).  To deal with this, we look for some CPU
1651 			 * with a NULL clean list, NULL dirty list, and NULL
1652 			 * rinsing list -- and then we borrow this CPU to
1653 			 * rinse our dirty list.
1654 			 */
1655 			for (j = 0; j < NCPU; j++) {
1656 				dtrace_dstate_percpu_t *rinser;
1657 
1658 				rinser = &dstate->dtds_percpu[j];
1659 
1660 				if (rinser->dtdsc_rinsing != NULL)
1661 					continue;
1662 
1663 				if (rinser->dtdsc_dirty != NULL)
1664 					continue;
1665 
1666 				if (rinser->dtdsc_clean != NULL)
1667 					continue;
1668 
1669 				rinsep = &rinser->dtdsc_rinsing;
1670 				break;
1671 			}
1672 
1673 			if (j == NCPU) {
1674 				/*
1675 				 * We were unable to find another CPU that
1676 				 * could accept this dirty list -- we are
1677 				 * therefore unable to clean it now.
1678 				 */
1679 				dtrace_dynvar_failclean++;
1680 				continue;
1681 			}
1682 		}
1683 
1684 		work = 1;
1685 
1686 		/*
1687 		 * Atomically move the dirty list aside.
1688 		 */
1689 		do {
1690 			dirty = dcpu->dtdsc_dirty;
1691 
1692 			/*
1693 			 * Before we zap the dirty list, set the rinsing list.
1694 			 * (This allows for a potential assertion in
1695 			 * dtrace_dynvar():  if a free dynamic variable appears
1696 			 * on a hash chain, either the dirty list or the
1697 			 * rinsing list for some CPU must be non-NULL.)
1698 			 */
1699 			*rinsep = dirty;
1700 			dtrace_membar_producer();
1701 		} while (dtrace_casptr(&dcpu->dtdsc_dirty,
1702 		    dirty, NULL) != dirty);
1703 	}
1704 
1705 	if (!work) {
1706 		/*
1707 		 * We have no work to do; we can simply return.
1708 		 */
1709 		return;
1710 	}
1711 
1712 	dtrace_sync();
1713 
1714 	for (i = 0; i < NCPU; i++) {
1715 		dcpu = &dstate->dtds_percpu[i];
1716 
1717 		if (dcpu->dtdsc_rinsing == NULL)
1718 			continue;
1719 
1720 		/*
1721 		 * We are now guaranteed that no hash chain contains a pointer
1722 		 * into this dirty list; we can make it clean.
1723 		 */
1724 		ASSERT(dcpu->dtdsc_clean == NULL);
1725 		dcpu->dtdsc_clean = dcpu->dtdsc_rinsing;
1726 		dcpu->dtdsc_rinsing = NULL;
1727 	}
1728 
1729 	/*
1730 	 * Before we actually set the state to be DTRACE_DSTATE_CLEAN, make
1731 	 * sure that all CPUs have seen all of the dtdsc_clean pointers.
1732 	 * This prevents a race whereby a CPU incorrectly decides that
1733 	 * the state should be something other than DTRACE_DSTATE_CLEAN
1734 	 * after dtrace_dynvar_clean() has completed.
1735 	 */
1736 	dtrace_sync();
1737 
1738 	dstate->dtds_state = DTRACE_DSTATE_CLEAN;
1739 }
1740 
1741 /*
1742  * Depending on the value of the op parameter, this function looks-up,
1743  * allocates or deallocates an arbitrarily-keyed dynamic variable.  If an
1744  * allocation is requested, this function will return a pointer to a
1745  * dtrace_dynvar_t corresponding to the allocated variable -- or NULL if no
1746  * variable can be allocated.  If NULL is returned, the appropriate counter
1747  * will be incremented.
1748  */
1749 dtrace_dynvar_t *
1750 dtrace_dynvar(dtrace_dstate_t *dstate, uint_t nkeys,
1751     dtrace_key_t *key, size_t dsize, dtrace_dynvar_op_t op,
1752     dtrace_mstate_t *mstate, dtrace_vstate_t *vstate)
1753 {
1754 	uint64_t hashval = DTRACE_DYNHASH_VALID;
1755 	dtrace_dynhash_t *hash = dstate->dtds_hash;
1756 	dtrace_dynvar_t *free, *new_free, *next, *dvar, *start, *prev = NULL;
1757 	processorid_t me = CPU->cpu_id, cpu = me;
1758 	dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[me];
1759 	size_t bucket, ksize;
1760 	size_t chunksize = dstate->dtds_chunksize;
1761 	uintptr_t kdata, lock, nstate;
1762 	uint_t i;
1763 
1764 	ASSERT(nkeys != 0);
1765 
1766 	/*
1767 	 * Hash the key.  As with aggregations, we use Jenkins' "One-at-a-time"
1768 	 * algorithm.  For the by-value portions, we perform the algorithm in
1769 	 * 16-bit chunks (as opposed to 8-bit chunks).  This speeds things up a
1770 	 * bit, and seems to have only a minute effect on distribution.  For
1771 	 * the by-reference data, we perform "One-at-a-time" iterating (safely)
1772 	 * over each referenced byte.  It's painful to do this, but it's much
1773 	 * better than pathological hash distribution.  The efficacy of the
1774 	 * hashing algorithm (and a comparison with other algorithms) may be
1775 	 * found by running the ::dtrace_dynstat MDB dcmd.
1776 	 */
1777 	for (i = 0; i < nkeys; i++) {
1778 		if (key[i].dttk_size == 0) {
1779 			uint64_t val = key[i].dttk_value;
1780 
1781 			hashval += (val >> 48) & 0xffff;
1782 			hashval += (hashval << 10);
1783 			hashval ^= (hashval >> 6);
1784 
1785 			hashval += (val >> 32) & 0xffff;
1786 			hashval += (hashval << 10);
1787 			hashval ^= (hashval >> 6);
1788 
1789 			hashval += (val >> 16) & 0xffff;
1790 			hashval += (hashval << 10);
1791 			hashval ^= (hashval >> 6);
1792 
1793 			hashval += val & 0xffff;
1794 			hashval += (hashval << 10);
1795 			hashval ^= (hashval >> 6);
1796 		} else {
1797 			/*
1798 			 * This is incredibly painful, but it beats the hell
1799 			 * out of the alternative.
1800 			 */
1801 			uint64_t j, size = key[i].dttk_size;
1802 			uintptr_t base = (uintptr_t)key[i].dttk_value;
1803 
1804 			if (!dtrace_canload(base, size, mstate, vstate))
1805 				break;
1806 
1807 			for (j = 0; j < size; j++) {
1808 				hashval += dtrace_load8(base + j);
1809 				hashval += (hashval << 10);
1810 				hashval ^= (hashval >> 6);
1811 			}
1812 		}
1813 	}
1814 
1815 	if (DTRACE_CPUFLAG_ISSET(CPU_DTRACE_FAULT))
1816 		return (NULL);
1817 
1818 	hashval += (hashval << 3);
1819 	hashval ^= (hashval >> 11);
1820 	hashval += (hashval << 15);
1821 
1822 	/*
1823 	 * There is a remote chance (ideally, 1 in 2^31) that our hashval
1824 	 * comes out to be one of our two sentinel hash values.  If this
1825 	 * actually happens, we set the hashval to be a value known to be a
1826 	 * non-sentinel value.
1827 	 */
1828 	if (hashval == DTRACE_DYNHASH_FREE || hashval == DTRACE_DYNHASH_SINK)
1829 		hashval = DTRACE_DYNHASH_VALID;
1830 
1831 	/*
1832 	 * Yes, it's painful to do a divide here.  If the cycle count becomes
1833 	 * important here, tricks can be pulled to reduce it.  (However, it's
1834 	 * critical that hash collisions be kept to an absolute minimum;
1835 	 * they're much more painful than a divide.)  It's better to have a
1836 	 * solution that generates few collisions and still keeps things
1837 	 * relatively simple.
1838 	 */
1839 	bucket = hashval % dstate->dtds_hashsize;
1840 
1841 	if (op == DTRACE_DYNVAR_DEALLOC) {
1842 		volatile uintptr_t *lockp = &hash[bucket].dtdh_lock;
1843 
1844 		for (;;) {
1845 			while ((lock = *lockp) & 1)
1846 				continue;
1847 
1848 			if (dtrace_casptr((void *)lockp,
1849 			    (void *)lock, (void *)(lock + 1)) == (void *)lock)
1850 				break;
1851 		}
1852 
1853 		dtrace_membar_producer();
1854 	}
1855 
1856 top:
1857 	prev = NULL;
1858 	lock = hash[bucket].dtdh_lock;
1859 
1860 	dtrace_membar_consumer();
1861 
1862 	start = hash[bucket].dtdh_chain;
1863 	ASSERT(start != NULL && (start->dtdv_hashval == DTRACE_DYNHASH_SINK ||
1864 	    start->dtdv_hashval != DTRACE_DYNHASH_FREE ||
1865 	    op != DTRACE_DYNVAR_DEALLOC));
1866 
1867 	for (dvar = start; dvar != NULL; dvar = dvar->dtdv_next) {
1868 		dtrace_tuple_t *dtuple = &dvar->dtdv_tuple;
1869 		dtrace_key_t *dkey = &dtuple->dtt_key[0];
1870 
1871 		if (dvar->dtdv_hashval != hashval) {
1872 			if (dvar->dtdv_hashval == DTRACE_DYNHASH_SINK) {
1873 				/*
1874 				 * We've reached the sink, and therefore the
1875 				 * end of the hash chain; we can kick out of
1876 				 * the loop knowing that we have seen a valid
1877 				 * snapshot of state.
1878 				 */
1879 				ASSERT(dvar->dtdv_next == NULL);
1880 				ASSERT(dvar == &dtrace_dynhash_sink);
1881 				break;
1882 			}
1883 
1884 			if (dvar->dtdv_hashval == DTRACE_DYNHASH_FREE) {
1885 				/*
1886 				 * We've gone off the rails:  somewhere along
1887 				 * the line, one of the members of this hash
1888 				 * chain was deleted.  Note that we could also
1889 				 * detect this by simply letting this loop run
1890 				 * to completion, as we would eventually hit
1891 				 * the end of the dirty list.  However, we
1892 				 * want to avoid running the length of the
1893 				 * dirty list unnecessarily (it might be quite
1894 				 * long), so we catch this as early as
1895 				 * possible by detecting the hash marker.  In
1896 				 * this case, we simply set dvar to NULL and
1897 				 * break; the conditional after the loop will
1898 				 * send us back to top.
1899 				 */
1900 				dvar = NULL;
1901 				break;
1902 			}
1903 
1904 			goto next;
1905 		}
1906 
1907 		if (dtuple->dtt_nkeys != nkeys)
1908 			goto next;
1909 
1910 		for (i = 0; i < nkeys; i++, dkey++) {
1911 			if (dkey->dttk_size != key[i].dttk_size)
1912 				goto next; /* size or type mismatch */
1913 
1914 			if (dkey->dttk_size != 0) {
1915 				if (dtrace_bcmp(
1916 				    (void *)(uintptr_t)key[i].dttk_value,
1917 				    (void *)(uintptr_t)dkey->dttk_value,
1918 				    dkey->dttk_size))
1919 					goto next;
1920 			} else {
1921 				if (dkey->dttk_value != key[i].dttk_value)
1922 					goto next;
1923 			}
1924 		}
1925 
1926 		if (op != DTRACE_DYNVAR_DEALLOC)
1927 			return (dvar);
1928 
1929 		ASSERT(dvar->dtdv_next == NULL ||
1930 		    dvar->dtdv_next->dtdv_hashval != DTRACE_DYNHASH_FREE);
1931 
1932 		if (prev != NULL) {
1933 			ASSERT(hash[bucket].dtdh_chain != dvar);
1934 			ASSERT(start != dvar);
1935 			ASSERT(prev->dtdv_next == dvar);
1936 			prev->dtdv_next = dvar->dtdv_next;
1937 		} else {
1938 			if (dtrace_casptr(&hash[bucket].dtdh_chain,
1939 			    start, dvar->dtdv_next) != start) {
1940 				/*
1941 				 * We have failed to atomically swing the
1942 				 * hash table head pointer, presumably because
1943 				 * of a conflicting allocation on another CPU.
1944 				 * We need to reread the hash chain and try
1945 				 * again.
1946 				 */
1947 				goto top;
1948 			}
1949 		}
1950 
1951 		dtrace_membar_producer();
1952 
1953 		/*
1954 		 * Now set the hash value to indicate that it's free.
1955 		 */
1956 		ASSERT(hash[bucket].dtdh_chain != dvar);
1957 		dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
1958 
1959 		dtrace_membar_producer();
1960 
1961 		/*
1962 		 * Set the next pointer to point at the dirty list, and
1963 		 * atomically swing the dirty pointer to the newly freed dvar.
1964 		 */
1965 		do {
1966 			next = dcpu->dtdsc_dirty;
1967 			dvar->dtdv_next = next;
1968 		} while (dtrace_casptr(&dcpu->dtdsc_dirty, next, dvar) != next);
1969 
1970 		/*
1971 		 * Finally, unlock this hash bucket.
1972 		 */
1973 		ASSERT(hash[bucket].dtdh_lock == lock);
1974 		ASSERT(lock & 1);
1975 		hash[bucket].dtdh_lock++;
1976 
1977 		return (NULL);
1978 next:
1979 		prev = dvar;
1980 		continue;
1981 	}
1982 
1983 	if (dvar == NULL) {
1984 		/*
1985 		 * If dvar is NULL, it is because we went off the rails:
1986 		 * one of the elements that we traversed in the hash chain
1987 		 * was deleted while we were traversing it.  In this case,
1988 		 * we assert that we aren't doing a dealloc (deallocs lock
1989 		 * the hash bucket to prevent themselves from racing with
1990 		 * one another), and retry the hash chain traversal.
1991 		 */
1992 		ASSERT(op != DTRACE_DYNVAR_DEALLOC);
1993 		goto top;
1994 	}
1995 
1996 	if (op != DTRACE_DYNVAR_ALLOC) {
1997 		/*
1998 		 * If we are not to allocate a new variable, we want to
1999 		 * return NULL now.  Before we return, check that the value
2000 		 * of the lock word hasn't changed.  If it has, we may have
2001 		 * seen an inconsistent snapshot.
2002 		 */
2003 		if (op == DTRACE_DYNVAR_NOALLOC) {
2004 			if (hash[bucket].dtdh_lock != lock)
2005 				goto top;
2006 		} else {
2007 			ASSERT(op == DTRACE_DYNVAR_DEALLOC);
2008 			ASSERT(hash[bucket].dtdh_lock == lock);
2009 			ASSERT(lock & 1);
2010 			hash[bucket].dtdh_lock++;
2011 		}
2012 
2013 		return (NULL);
2014 	}
2015 
2016 	/*
2017 	 * We need to allocate a new dynamic variable.  The size we need is the
2018 	 * size of dtrace_dynvar plus the size of nkeys dtrace_key_t's plus the
2019 	 * size of any auxiliary key data (rounded up to 8-byte alignment) plus
2020 	 * the size of any referred-to data (dsize).  We then round the final
2021 	 * size up to the chunksize for allocation.
2022 	 */
2023 	for (ksize = 0, i = 0; i < nkeys; i++)
2024 		ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
2025 
2026 	/*
2027 	 * This should be pretty much impossible, but could happen if, say,
2028 	 * strange DIF specified the tuple.  Ideally, this should be an
2029 	 * assertion and not an error condition -- but that requires that the
2030 	 * chunksize calculation in dtrace_difo_chunksize() be absolutely
2031 	 * bullet-proof.  (That is, it must not be able to be fooled by
2032 	 * malicious DIF.)  Given the lack of backwards branches in DIF,
2033 	 * solving this would presumably not amount to solving the Halting
2034 	 * Problem -- but it still seems awfully hard.
2035 	 */
2036 	if (sizeof (dtrace_dynvar_t) + sizeof (dtrace_key_t) * (nkeys - 1) +
2037 	    ksize + dsize > chunksize) {
2038 		dcpu->dtdsc_drops++;
2039 		return (NULL);
2040 	}
2041 
2042 	nstate = DTRACE_DSTATE_EMPTY;
2043 
2044 	do {
2045 retry:
2046 		free = dcpu->dtdsc_free;
2047 
2048 		if (free == NULL) {
2049 			dtrace_dynvar_t *clean = dcpu->dtdsc_clean;
2050 			void *rval;
2051 
2052 			if (clean == NULL) {
2053 				/*
2054 				 * We're out of dynamic variable space on
2055 				 * this CPU.  Unless we have tried all CPUs,
2056 				 * we'll try to allocate from a different
2057 				 * CPU.
2058 				 */
2059 				switch (dstate->dtds_state) {
2060 				case DTRACE_DSTATE_CLEAN: {
2061 					void *sp = &dstate->dtds_state;
2062 
2063 					if (++cpu >= NCPU)
2064 						cpu = 0;
2065 
2066 					if (dcpu->dtdsc_dirty != NULL &&
2067 					    nstate == DTRACE_DSTATE_EMPTY)
2068 						nstate = DTRACE_DSTATE_DIRTY;
2069 
2070 					if (dcpu->dtdsc_rinsing != NULL)
2071 						nstate = DTRACE_DSTATE_RINSING;
2072 
2073 					dcpu = &dstate->dtds_percpu[cpu];
2074 
2075 					if (cpu != me)
2076 						goto retry;
2077 
2078 					(void) dtrace_cas32(sp,
2079 					    DTRACE_DSTATE_CLEAN, nstate);
2080 
2081 					/*
2082 					 * To increment the correct bean
2083 					 * counter, take another lap.
2084 					 */
2085 					goto retry;
2086 				}
2087 
2088 				case DTRACE_DSTATE_DIRTY:
2089 					dcpu->dtdsc_dirty_drops++;
2090 					break;
2091 
2092 				case DTRACE_DSTATE_RINSING:
2093 					dcpu->dtdsc_rinsing_drops++;
2094 					break;
2095 
2096 				case DTRACE_DSTATE_EMPTY:
2097 					dcpu->dtdsc_drops++;
2098 					break;
2099 				}
2100 
2101 				DTRACE_CPUFLAG_SET(CPU_DTRACE_DROP);
2102 				return (NULL);
2103 			}
2104 
2105 			/*
2106 			 * The clean list appears to be non-empty.  We want to
2107 			 * move the clean list to the free list; we start by
2108 			 * moving the clean pointer aside.
2109 			 */
2110 			if (dtrace_casptr(&dcpu->dtdsc_clean,
2111 			    clean, NULL) != clean) {
2112 				/*
2113 				 * We are in one of two situations:
2114 				 *
2115 				 *  (a)	The clean list was switched to the
2116 				 *	free list by another CPU.
2117 				 *
2118 				 *  (b)	The clean list was added to by the
2119 				 *	cleansing cyclic.
2120 				 *
2121 				 * In either of these situations, we can
2122 				 * just reattempt the free list allocation.
2123 				 */
2124 				goto retry;
2125 			}
2126 
2127 			ASSERT(clean->dtdv_hashval == DTRACE_DYNHASH_FREE);
2128 
2129 			/*
2130 			 * Now we'll move the clean list to our free list.
2131 			 * It's impossible for this to fail:  the only way
2132 			 * the free list can be updated is through this
2133 			 * code path, and only one CPU can own the clean list.
2134 			 * Thus, it would only be possible for this to fail if
2135 			 * this code were racing with dtrace_dynvar_clean().
2136 			 * (That is, if dtrace_dynvar_clean() updated the clean
2137 			 * list, and we ended up racing to update the free
2138 			 * list.)  This race is prevented by the dtrace_sync()
2139 			 * in dtrace_dynvar_clean() -- which flushes the
2140 			 * owners of the clean lists out before resetting
2141 			 * the clean lists.
2142 			 */
2143 			dcpu = &dstate->dtds_percpu[me];
2144 			rval = dtrace_casptr(&dcpu->dtdsc_free, NULL, clean);
2145 			ASSERT(rval == NULL);
2146 			goto retry;
2147 		}
2148 
2149 		dvar = free;
2150 		new_free = dvar->dtdv_next;
2151 	} while (dtrace_casptr(&dcpu->dtdsc_free, free, new_free) != free);
2152 
2153 	/*
2154 	 * We have now allocated a new chunk.  We copy the tuple keys into the
2155 	 * tuple array and copy any referenced key data into the data space
2156 	 * following the tuple array.  As we do this, we relocate dttk_value
2157 	 * in the final tuple to point to the key data address in the chunk.
2158 	 */
2159 	kdata = (uintptr_t)&dvar->dtdv_tuple.dtt_key[nkeys];
2160 	dvar->dtdv_data = (void *)(kdata + ksize);
2161 	dvar->dtdv_tuple.dtt_nkeys = nkeys;
2162 
2163 	for (i = 0; i < nkeys; i++) {
2164 		dtrace_key_t *dkey = &dvar->dtdv_tuple.dtt_key[i];
2165 		size_t kesize = key[i].dttk_size;
2166 
2167 		if (kesize != 0) {
2168 			dtrace_bcopy(
2169 			    (const void *)(uintptr_t)key[i].dttk_value,
2170 			    (void *)kdata, kesize);
2171 			dkey->dttk_value = kdata;
2172 			kdata += P2ROUNDUP(kesize, sizeof (uint64_t));
2173 		} else {
2174 			dkey->dttk_value = key[i].dttk_value;
2175 		}
2176 
2177 		dkey->dttk_size = kesize;
2178 	}
2179 
2180 	ASSERT(dvar->dtdv_hashval == DTRACE_DYNHASH_FREE);
2181 	dvar->dtdv_hashval = hashval;
2182 	dvar->dtdv_next = start;
2183 
2184 	if (dtrace_casptr(&hash[bucket].dtdh_chain, start, dvar) == start)
2185 		return (dvar);
2186 
2187 	/*
2188 	 * The cas has failed.  Either another CPU is adding an element to
2189 	 * this hash chain, or another CPU is deleting an element from this
2190 	 * hash chain.  The simplest way to deal with both of these cases
2191 	 * (though not necessarily the most efficient) is to free our
2192 	 * allocated block and re-attempt it all.  Note that the free is
2193 	 * to the dirty list and _not_ to the free list.  This is to prevent
2194 	 * races with allocators, above.
2195 	 */
2196 	dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
2197 
2198 	dtrace_membar_producer();
2199 
2200 	do {
2201 		free = dcpu->dtdsc_dirty;
2202 		dvar->dtdv_next = free;
2203 	} while (dtrace_casptr(&dcpu->dtdsc_dirty, free, dvar) != free);
2204 
2205 	goto top;
2206 }
2207 
2208 /*ARGSUSED*/
2209 static void
2210 dtrace_aggregate_min(uint64_t *oval, uint64_t nval, uint64_t arg)
2211 {
2212 	if ((int64_t)nval < (int64_t)*oval)
2213 		*oval = nval;
2214 }
2215 
2216 /*ARGSUSED*/
2217 static void
2218 dtrace_aggregate_max(uint64_t *oval, uint64_t nval, uint64_t arg)
2219 {
2220 	if ((int64_t)nval > (int64_t)*oval)
2221 		*oval = nval;
2222 }
2223 
2224 static void
2225 dtrace_aggregate_quantize(uint64_t *quanta, uint64_t nval, uint64_t incr)
2226 {
2227 	int i, zero = DTRACE_QUANTIZE_ZEROBUCKET;
2228 	int64_t val = (int64_t)nval;
2229 
2230 	if (val < 0) {
2231 		for (i = 0; i < zero; i++) {
2232 			if (val <= DTRACE_QUANTIZE_BUCKETVAL(i)) {
2233 				quanta[i] += incr;
2234 				return;
2235 			}
2236 		}
2237 	} else {
2238 		for (i = zero + 1; i < DTRACE_QUANTIZE_NBUCKETS; i++) {
2239 			if (val < DTRACE_QUANTIZE_BUCKETVAL(i)) {
2240 				quanta[i - 1] += incr;
2241 				return;
2242 			}
2243 		}
2244 
2245 		quanta[DTRACE_QUANTIZE_NBUCKETS - 1] += incr;
2246 		return;
2247 	}
2248 
2249 	ASSERT(0);
2250 }
2251 
2252 static void
2253 dtrace_aggregate_lquantize(uint64_t *lquanta, uint64_t nval, uint64_t incr)
2254 {
2255 	uint64_t arg = *lquanta++;
2256 	int32_t base = DTRACE_LQUANTIZE_BASE(arg);
2257 	uint16_t step = DTRACE_LQUANTIZE_STEP(arg);
2258 	uint16_t levels = DTRACE_LQUANTIZE_LEVELS(arg);
2259 	int32_t val = (int32_t)nval, level;
2260 
2261 	ASSERT(step != 0);
2262 	ASSERT(levels != 0);
2263 
2264 	if (val < base) {
2265 		/*
2266 		 * This is an underflow.
2267 		 */
2268 		lquanta[0] += incr;
2269 		return;
2270 	}
2271 
2272 	level = (val - base) / step;
2273 
2274 	if (level < levels) {
2275 		lquanta[level + 1] += incr;
2276 		return;
2277 	}
2278 
2279 	/*
2280 	 * This is an overflow.
2281 	 */
2282 	lquanta[levels + 1] += incr;
2283 }
2284 
2285 static int
2286 dtrace_aggregate_llquantize_bucket(uint16_t factor, uint16_t low,
2287     uint16_t high, uint16_t nsteps, int64_t value)
2288 {
2289 	int64_t this = 1, last, next;
2290 	int base = 1, order;
2291 
2292 	ASSERT(factor <= nsteps);
2293 	ASSERT(nsteps % factor == 0);
2294 
2295 	for (order = 0; order < low; order++)
2296 		this *= factor;
2297 
2298 	/*
2299 	 * If our value is less than our factor taken to the power of the
2300 	 * low order of magnitude, it goes into the zeroth bucket.
2301 	 */
2302 	if (value < (last = this))
2303 		return (0);
2304 
2305 	for (this *= factor; order <= high; order++) {
2306 		int nbuckets = this > nsteps ? nsteps : this;
2307 
2308 		if ((next = this * factor) < this) {
2309 			/*
2310 			 * We should not generally get log/linear quantizations
2311 			 * with a high magnitude that allows 64-bits to
2312 			 * overflow, but we nonetheless protect against this
2313 			 * by explicitly checking for overflow, and clamping
2314 			 * our value accordingly.
2315 			 */
2316 			value = this - 1;
2317 		}
2318 
2319 		if (value < this) {
2320 			/*
2321 			 * If our value lies within this order of magnitude,
2322 			 * determine its position by taking the offset within
2323 			 * the order of magnitude, dividing by the bucket
2324 			 * width, and adding to our (accumulated) base.
2325 			 */
2326 			return (base + (value - last) / (this / nbuckets));
2327 		}
2328 
2329 		base += nbuckets - (nbuckets / factor);
2330 		last = this;
2331 		this = next;
2332 	}
2333 
2334 	/*
2335 	 * Our value is greater than or equal to our factor taken to the
2336 	 * power of one plus the high magnitude -- return the top bucket.
2337 	 */
2338 	return (base);
2339 }
2340 
2341 static void
2342 dtrace_aggregate_llquantize(uint64_t *llquanta, uint64_t nval, uint64_t incr)
2343 {
2344 	uint64_t arg = *llquanta++;
2345 	uint16_t factor = DTRACE_LLQUANTIZE_FACTOR(arg);
2346 	uint16_t low = DTRACE_LLQUANTIZE_LOW(arg);
2347 	uint16_t high = DTRACE_LLQUANTIZE_HIGH(arg);
2348 	uint16_t nsteps = DTRACE_LLQUANTIZE_NSTEP(arg);
2349 
2350 	llquanta[dtrace_aggregate_llquantize_bucket(factor,
2351 	    low, high, nsteps, nval)] += incr;
2352 }
2353 
2354 /*ARGSUSED*/
2355 static void
2356 dtrace_aggregate_avg(uint64_t *data, uint64_t nval, uint64_t arg)
2357 {
2358 	data[0]++;
2359 	data[1] += nval;
2360 }
2361 
2362 /*ARGSUSED*/
2363 static void
2364 dtrace_aggregate_stddev(uint64_t *data, uint64_t nval, uint64_t arg)
2365 {
2366 	int64_t snval = (int64_t)nval;
2367 	uint64_t tmp[2];
2368 
2369 	data[0]++;
2370 	data[1] += nval;
2371 
2372 	/*
2373 	 * What we want to say here is:
2374 	 *
2375 	 * data[2] += nval * nval;
2376 	 *
2377 	 * But given that nval is 64-bit, we could easily overflow, so
2378 	 * we do this as 128-bit arithmetic.
2379 	 */
2380 	if (snval < 0)
2381 		snval = -snval;
2382 
2383 	dtrace_multiply_128((uint64_t)snval, (uint64_t)snval, tmp);
2384 	dtrace_add_128(data + 2, tmp, data + 2);
2385 }
2386 
2387 /*ARGSUSED*/
2388 static void
2389 dtrace_aggregate_count(uint64_t *oval, uint64_t nval, uint64_t arg)
2390 {
2391 	*oval = *oval + 1;
2392 }
2393 
2394 /*ARGSUSED*/
2395 static void
2396 dtrace_aggregate_sum(uint64_t *oval, uint64_t nval, uint64_t arg)
2397 {
2398 	*oval += nval;
2399 }
2400 
2401 /*
2402  * Aggregate given the tuple in the principal data buffer, and the aggregating
2403  * action denoted by the specified dtrace_aggregation_t.  The aggregation
2404  * buffer is specified as the buf parameter.  This routine does not return
2405  * failure; if there is no space in the aggregation buffer, the data will be
2406  * dropped, and a corresponding counter incremented.
2407  */
2408 static void
2409 dtrace_aggregate(dtrace_aggregation_t *agg, dtrace_buffer_t *dbuf,
2410     intptr_t offset, dtrace_buffer_t *buf, uint64_t expr, uint64_t arg)
2411 {
2412 	dtrace_recdesc_t *rec = &agg->dtag_action.dta_rec;
2413 	uint32_t i, ndx, size, fsize;
2414 	uint32_t align = sizeof (uint64_t) - 1;
2415 	dtrace_aggbuffer_t *agb;
2416 	dtrace_aggkey_t *key;
2417 	uint32_t hashval = 0, limit, isstr;
2418 	caddr_t tomax, data, kdata;
2419 	dtrace_actkind_t action;
2420 	dtrace_action_t *act;
2421 	uintptr_t offs;
2422 
2423 	if (buf == NULL)
2424 		return;
2425 
2426 	if (!agg->dtag_hasarg) {
2427 		/*
2428 		 * Currently, only quantize() and lquantize() take additional
2429 		 * arguments, and they have the same semantics:  an increment
2430 		 * value that defaults to 1 when not present.  If additional
2431 		 * aggregating actions take arguments, the setting of the
2432 		 * default argument value will presumably have to become more
2433 		 * sophisticated...
2434 		 */
2435 		arg = 1;
2436 	}
2437 
2438 	action = agg->dtag_action.dta_kind - DTRACEACT_AGGREGATION;
2439 	size = rec->dtrd_offset - agg->dtag_base;
2440 	fsize = size + rec->dtrd_size;
2441 
2442 	ASSERT(dbuf->dtb_tomax != NULL);
2443 	data = dbuf->dtb_tomax + offset + agg->dtag_base;
2444 
2445 	if ((tomax = buf->dtb_tomax) == NULL) {
2446 		dtrace_buffer_drop(buf);
2447 		return;
2448 	}
2449 
2450 	/*
2451 	 * The metastructure is always at the bottom of the buffer.
2452 	 */
2453 	agb = (dtrace_aggbuffer_t *)(tomax + buf->dtb_size -
2454 	    sizeof (dtrace_aggbuffer_t));
2455 
2456 	if (buf->dtb_offset == 0) {
2457 		/*
2458 		 * We just kludge up approximately 1/8th of the size to be
2459 		 * buckets.  If this guess ends up being routinely
2460 		 * off-the-mark, we may need to dynamically readjust this
2461 		 * based on past performance.
2462 		 */
2463 		uintptr_t hashsize = (buf->dtb_size >> 3) / sizeof (uintptr_t);
2464 
2465 		if ((uintptr_t)agb - hashsize * sizeof (dtrace_aggkey_t *) <
2466 		    (uintptr_t)tomax || hashsize == 0) {
2467 			/*
2468 			 * We've been given a ludicrously small buffer;
2469 			 * increment our drop count and leave.
2470 			 */
2471 			dtrace_buffer_drop(buf);
2472 			return;
2473 		}
2474 
2475 		/*
2476 		 * And now, a pathetic attempt to try to get a an odd (or
2477 		 * perchance, a prime) hash size for better hash distribution.
2478 		 */
2479 		if (hashsize > (DTRACE_AGGHASHSIZE_SLEW << 3))
2480 			hashsize -= DTRACE_AGGHASHSIZE_SLEW;
2481 
2482 		agb->dtagb_hashsize = hashsize;
2483 		agb->dtagb_hash = (dtrace_aggkey_t **)((uintptr_t)agb -
2484 		    agb->dtagb_hashsize * sizeof (dtrace_aggkey_t *));
2485 		agb->dtagb_free = (uintptr_t)agb->dtagb_hash;
2486 
2487 		for (i = 0; i < agb->dtagb_hashsize; i++)
2488 			agb->dtagb_hash[i] = NULL;
2489 	}
2490 
2491 	ASSERT(agg->dtag_first != NULL);
2492 	ASSERT(agg->dtag_first->dta_intuple);
2493 
2494 	/*
2495 	 * Calculate the hash value based on the key.  Note that we _don't_
2496 	 * include the aggid in the hashing (but we will store it as part of
2497 	 * the key).  The hashing algorithm is Bob Jenkins' "One-at-a-time"
2498 	 * algorithm: a simple, quick algorithm that has no known funnels, and
2499 	 * gets good distribution in practice.  The efficacy of the hashing
2500 	 * algorithm (and a comparison with other algorithms) may be found by
2501 	 * running the ::dtrace_aggstat MDB dcmd.
2502 	 */
2503 	for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
2504 		i = act->dta_rec.dtrd_offset - agg->dtag_base;
2505 		limit = i + act->dta_rec.dtrd_size;
2506 		ASSERT(limit <= size);
2507 		isstr = DTRACEACT_ISSTRING(act);
2508 
2509 		for (; i < limit; i++) {
2510 			hashval += data[i];
2511 			hashval += (hashval << 10);
2512 			hashval ^= (hashval >> 6);
2513 
2514 			if (isstr && data[i] == '\0')
2515 				break;
2516 		}
2517 	}
2518 
2519 	hashval += (hashval << 3);
2520 	hashval ^= (hashval >> 11);
2521 	hashval += (hashval << 15);
2522 
2523 	/*
2524 	 * Yes, the divide here is expensive -- but it's generally the least
2525 	 * of the performance issues given the amount of data that we iterate
2526 	 * over to compute hash values, compare data, etc.
2527 	 */
2528 	ndx = hashval % agb->dtagb_hashsize;
2529 
2530 	for (key = agb->dtagb_hash[ndx]; key != NULL; key = key->dtak_next) {
2531 		ASSERT((caddr_t)key >= tomax);
2532 		ASSERT((caddr_t)key < tomax + buf->dtb_size);
2533 
2534 		if (hashval != key->dtak_hashval || key->dtak_size != size)
2535 			continue;
2536 
2537 		kdata = key->dtak_data;
2538 		ASSERT(kdata >= tomax && kdata < tomax + buf->dtb_size);
2539 
2540 		for (act = agg->dtag_first; act->dta_intuple;
2541 		    act = act->dta_next) {
2542 			i = act->dta_rec.dtrd_offset - agg->dtag_base;
2543 			limit = i + act->dta_rec.dtrd_size;
2544 			ASSERT(limit <= size);
2545 			isstr = DTRACEACT_ISSTRING(act);
2546 
2547 			for (; i < limit; i++) {
2548 				if (kdata[i] != data[i])
2549 					goto next;
2550 
2551 				if (isstr && data[i] == '\0')
2552 					break;
2553 			}
2554 		}
2555 
2556 		if (action != key->dtak_action) {
2557 			/*
2558 			 * We are aggregating on the same value in the same
2559 			 * aggregation with two different aggregating actions.
2560 			 * (This should have been picked up in the compiler,
2561 			 * so we may be dealing with errant or devious DIF.)
2562 			 * This is an error condition; we indicate as much,
2563 			 * and return.
2564 			 */
2565 			DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
2566 			return;
2567 		}
2568 
2569 		/*
2570 		 * This is a hit:  we need to apply the aggregator to
2571 		 * the value at this key.
2572 		 */
2573 		agg->dtag_aggregate((uint64_t *)(kdata + size), expr, arg);
2574 		return;
2575 next:
2576 		continue;
2577 	}
2578 
2579 	/*
2580 	 * We didn't find it.  We need to allocate some zero-filled space,
2581 	 * link it into the hash table appropriately, and apply the aggregator
2582 	 * to the (zero-filled) value.
2583 	 */
2584 	offs = buf->dtb_offset;
2585 	while (offs & (align - 1))
2586 		offs += sizeof (uint32_t);
2587 
2588 	/*
2589 	 * If we don't have enough room to both allocate a new key _and_
2590 	 * its associated data, increment the drop count and return.
2591 	 */
2592 	if ((uintptr_t)tomax + offs + fsize >
2593 	    agb->dtagb_free - sizeof (dtrace_aggkey_t)) {
2594 		dtrace_buffer_drop(buf);
2595 		return;
2596 	}
2597 
2598 	/*CONSTCOND*/
2599 	ASSERT(!(sizeof (dtrace_aggkey_t) & (sizeof (uintptr_t) - 1)));
2600 	key = (dtrace_aggkey_t *)(agb->dtagb_free - sizeof (dtrace_aggkey_t));
2601 	agb->dtagb_free -= sizeof (dtrace_aggkey_t);
2602 
2603 	key->dtak_data = kdata = tomax + offs;
2604 	buf->dtb_offset = offs + fsize;
2605 
2606 	/*
2607 	 * Now copy the data across.
2608 	 */
2609 	*((dtrace_aggid_t *)kdata) = agg->dtag_id;
2610 
2611 	for (i = sizeof (dtrace_aggid_t); i < size; i++)
2612 		kdata[i] = data[i];
2613 
2614 	/*
2615 	 * Because strings are not zeroed out by default, we need to iterate
2616 	 * looking for actions that store strings, and we need to explicitly
2617 	 * pad these strings out with zeroes.
2618 	 */
2619 	for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
2620 		int nul;
2621 
2622 		if (!DTRACEACT_ISSTRING(act))
2623 			continue;
2624 
2625 		i = act->dta_rec.dtrd_offset - agg->dtag_base;
2626 		limit = i + act->dta_rec.dtrd_size;
2627 		ASSERT(limit <= size);
2628 
2629 		for (nul = 0; i < limit; i++) {
2630 			if (nul) {
2631 				kdata[i] = '\0';
2632 				continue;
2633 			}
2634 
2635 			if (data[i] != '\0')
2636 				continue;
2637 
2638 			nul = 1;
2639 		}
2640 	}
2641 
2642 	for (i = size; i < fsize; i++)
2643 		kdata[i] = 0;
2644 
2645 	key->dtak_hashval = hashval;
2646 	key->dtak_size = size;
2647 	key->dtak_action = action;
2648 	key->dtak_next = agb->dtagb_hash[ndx];
2649 	agb->dtagb_hash[ndx] = key;
2650 
2651 	/*
2652 	 * Finally, apply the aggregator.
2653 	 */
2654 	*((uint64_t *)(key->dtak_data + size)) = agg->dtag_initial;
2655 	agg->dtag_aggregate((uint64_t *)(key->dtak_data + size), expr, arg);
2656 }
2657 
2658 /*
2659  * Given consumer state, this routine finds a speculation in the INACTIVE
2660  * state and transitions it into the ACTIVE state.  If there is no speculation
2661  * in the INACTIVE state, 0 is returned.  In this case, no error counter is
2662  * incremented -- it is up to the caller to take appropriate action.
2663  */
2664 static int
2665 dtrace_speculation(dtrace_state_t *state)
2666 {
2667 	int i = 0;
2668 	dtrace_speculation_state_t current;
2669 	uint32_t *stat = &state->dts_speculations_unavail, count;
2670 
2671 	while (i < state->dts_nspeculations) {
2672 		dtrace_speculation_t *spec = &state->dts_speculations[i];
2673 
2674 		current = spec->dtsp_state;
2675 
2676 		if (current != DTRACESPEC_INACTIVE) {
2677 			if (current == DTRACESPEC_COMMITTINGMANY ||
2678 			    current == DTRACESPEC_COMMITTING ||
2679 			    current == DTRACESPEC_DISCARDING)
2680 				stat = &state->dts_speculations_busy;
2681 			i++;
2682 			continue;
2683 		}
2684 
2685 		if (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2686 		    current, DTRACESPEC_ACTIVE) == current)
2687 			return (i + 1);
2688 	}
2689 
2690 	/*
2691 	 * We couldn't find a speculation.  If we found as much as a single
2692 	 * busy speculation buffer, we'll attribute this failure as "busy"
2693 	 * instead of "unavail".
2694 	 */
2695 	do {
2696 		count = *stat;
2697 	} while (dtrace_cas32(stat, count, count + 1) != count);
2698 
2699 	return (0);
2700 }
2701 
2702 /*
2703  * This routine commits an active speculation.  If the specified speculation
2704  * is not in a valid state to perform a commit(), this routine will silently do
2705  * nothing.  The state of the specified speculation is transitioned according
2706  * to the state transition diagram outlined in <sys/dtrace_impl.h>
2707  */
2708 static void
2709 dtrace_speculation_commit(dtrace_state_t *state, processorid_t cpu,
2710     dtrace_specid_t which)
2711 {
2712 	dtrace_speculation_t *spec;
2713 	dtrace_buffer_t *src, *dest;
2714 	uintptr_t daddr, saddr, dlimit, slimit;
2715 	dtrace_speculation_state_t current, new;
2716 	intptr_t offs;
2717 	uint64_t timestamp;
2718 
2719 	if (which == 0)
2720 		return;
2721 
2722 	if (which > state->dts_nspeculations) {
2723 		cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2724 		return;
2725 	}
2726 
2727 	spec = &state->dts_speculations[which - 1];
2728 	src = &spec->dtsp_buffer[cpu];
2729 	dest = &state->dts_buffer[cpu];
2730 
2731 	do {
2732 		current = spec->dtsp_state;
2733 
2734 		if (current == DTRACESPEC_COMMITTINGMANY)
2735 			break;
2736 
2737 		switch (current) {
2738 		case DTRACESPEC_INACTIVE:
2739 		case DTRACESPEC_DISCARDING:
2740 			return;
2741 
2742 		case DTRACESPEC_COMMITTING:
2743 			/*
2744 			 * This is only possible if we are (a) commit()'ing
2745 			 * without having done a prior speculate() on this CPU
2746 			 * and (b) racing with another commit() on a different
2747 			 * CPU.  There's nothing to do -- we just assert that
2748 			 * our offset is 0.
2749 			 */
2750 			ASSERT(src->dtb_offset == 0);
2751 			return;
2752 
2753 		case DTRACESPEC_ACTIVE:
2754 			new = DTRACESPEC_COMMITTING;
2755 			break;
2756 
2757 		case DTRACESPEC_ACTIVEONE:
2758 			/*
2759 			 * This speculation is active on one CPU.  If our
2760 			 * buffer offset is non-zero, we know that the one CPU
2761 			 * must be us.  Otherwise, we are committing on a
2762 			 * different CPU from the speculate(), and we must
2763 			 * rely on being asynchronously cleaned.
2764 			 */
2765 			if (src->dtb_offset != 0) {
2766 				new = DTRACESPEC_COMMITTING;
2767 				break;
2768 			}
2769 			/*FALLTHROUGH*/
2770 
2771 		case DTRACESPEC_ACTIVEMANY:
2772 			new = DTRACESPEC_COMMITTINGMANY;
2773 			break;
2774 
2775 		default:
2776 			ASSERT(0);
2777 		}
2778 	} while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2779 	    current, new) != current);
2780 
2781 	/*
2782 	 * We have set the state to indicate that we are committing this
2783 	 * speculation.  Now reserve the necessary space in the destination
2784 	 * buffer.
2785 	 */
2786 	if ((offs = dtrace_buffer_reserve(dest, src->dtb_offset,
2787 	    sizeof (uint64_t), state, NULL)) < 0) {
2788 		dtrace_buffer_drop(dest);
2789 		goto out;
2790 	}
2791 
2792 	/*
2793 	 * We have sufficient space to copy the speculative buffer into the
2794 	 * primary buffer.  First, modify the speculative buffer, filling
2795 	 * in the timestamp of all entries with the current time.  The data
2796 	 * must have the commit() time rather than the time it was traced,
2797 	 * so that all entries in the primary buffer are in timestamp order.
2798 	 */
2799 	timestamp = dtrace_gethrtime();
2800 	saddr = (uintptr_t)src->dtb_tomax;
2801 	slimit = saddr + src->dtb_offset;
2802 	while (saddr < slimit) {
2803 		size_t size;
2804 		dtrace_rechdr_t *dtrh = (dtrace_rechdr_t *)saddr;
2805 
2806 		if (dtrh->dtrh_epid == DTRACE_EPIDNONE) {
2807 			saddr += sizeof (dtrace_epid_t);
2808 			continue;
2809 		}
2810 		ASSERT3U(dtrh->dtrh_epid, <=, state->dts_necbs);
2811 		size = state->dts_ecbs[dtrh->dtrh_epid - 1]->dte_size;
2812 
2813 		ASSERT3U(saddr + size, <=, slimit);
2814 		ASSERT3U(size, >=, sizeof (dtrace_rechdr_t));
2815 		ASSERT3U(DTRACE_RECORD_LOAD_TIMESTAMP(dtrh), ==, UINT64_MAX);
2816 
2817 		DTRACE_RECORD_STORE_TIMESTAMP(dtrh, timestamp);
2818 
2819 		saddr += size;
2820 	}
2821 
2822 	/*
2823 	 * Copy the buffer across.  (Note that this is a
2824 	 * highly subobtimal bcopy(); in the unlikely event that this becomes
2825 	 * a serious performance issue, a high-performance DTrace-specific
2826 	 * bcopy() should obviously be invented.)
2827 	 */
2828 	daddr = (uintptr_t)dest->dtb_tomax + offs;
2829 	dlimit = daddr + src->dtb_offset;
2830 	saddr = (uintptr_t)src->dtb_tomax;
2831 
2832 	/*
2833 	 * First, the aligned portion.
2834 	 */
2835 	while (dlimit - daddr >= sizeof (uint64_t)) {
2836 		*((uint64_t *)daddr) = *((uint64_t *)saddr);
2837 
2838 		daddr += sizeof (uint64_t);
2839 		saddr += sizeof (uint64_t);
2840 	}
2841 
2842 	/*
2843 	 * Now any left-over bit...
2844 	 */
2845 	while (dlimit - daddr)
2846 		*((uint8_t *)daddr++) = *((uint8_t *)saddr++);
2847 
2848 	/*
2849 	 * Finally, commit the reserved space in the destination buffer.
2850 	 */
2851 	dest->dtb_offset = offs + src->dtb_offset;
2852 
2853 out:
2854 	/*
2855 	 * If we're lucky enough to be the only active CPU on this speculation
2856 	 * buffer, we can just set the state back to DTRACESPEC_INACTIVE.
2857 	 */
2858 	if (current == DTRACESPEC_ACTIVE ||
2859 	    (current == DTRACESPEC_ACTIVEONE && new == DTRACESPEC_COMMITTING)) {
2860 		uint32_t rval = dtrace_cas32((uint32_t *)&spec->dtsp_state,
2861 		    DTRACESPEC_COMMITTING, DTRACESPEC_INACTIVE);
2862 
2863 		ASSERT(rval == DTRACESPEC_COMMITTING);
2864 	}
2865 
2866 	src->dtb_offset = 0;
2867 	src->dtb_xamot_drops += src->dtb_drops;
2868 	src->dtb_drops = 0;
2869 }
2870 
2871 /*
2872  * This routine discards an active speculation.  If the specified speculation
2873  * is not in a valid state to perform a discard(), this routine will silently
2874  * do nothing.  The state of the specified speculation is transitioned
2875  * according to the state transition diagram outlined in <sys/dtrace_impl.h>
2876  */
2877 static void
2878 dtrace_speculation_discard(dtrace_state_t *state, processorid_t cpu,
2879     dtrace_specid_t which)
2880 {
2881 	dtrace_speculation_t *spec;
2882 	dtrace_speculation_state_t current, new;
2883 	dtrace_buffer_t *buf;
2884 
2885 	if (which == 0)
2886 		return;
2887 
2888 	if (which > state->dts_nspeculations) {
2889 		cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2890 		return;
2891 	}
2892 
2893 	spec = &state->dts_speculations[which - 1];
2894 	buf = &spec->dtsp_buffer[cpu];
2895 
2896 	do {
2897 		current = spec->dtsp_state;
2898 
2899 		switch (current) {
2900 		case DTRACESPEC_INACTIVE:
2901 		case DTRACESPEC_COMMITTINGMANY:
2902 		case DTRACESPEC_COMMITTING:
2903 		case DTRACESPEC_DISCARDING:
2904 			return;
2905 
2906 		case DTRACESPEC_ACTIVE:
2907 		case DTRACESPEC_ACTIVEMANY:
2908 			new = DTRACESPEC_DISCARDING;
2909 			break;
2910 
2911 		case DTRACESPEC_ACTIVEONE:
2912 			if (buf->dtb_offset != 0) {
2913 				new = DTRACESPEC_INACTIVE;
2914 			} else {
2915 				new = DTRACESPEC_DISCARDING;
2916 			}
2917 			break;
2918 
2919 		default:
2920 			ASSERT(0);
2921 		}
2922 	} while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2923 	    current, new) != current);
2924 
2925 	buf->dtb_offset = 0;
2926 	buf->dtb_drops = 0;
2927 }
2928 
2929 /*
2930  * Note:  not called from probe context.  This function is called
2931  * asynchronously from cross call context to clean any speculations that are
2932  * in the COMMITTINGMANY or DISCARDING states.  These speculations may not be
2933  * transitioned back to the INACTIVE state until all CPUs have cleaned the
2934  * speculation.
2935  */
2936 static void
2937 dtrace_speculation_clean_here(dtrace_state_t *state)
2938 {
2939 	dtrace_icookie_t cookie;
2940 	processorid_t cpu = CPU->cpu_id;
2941 	dtrace_buffer_t *dest = &state->dts_buffer[cpu];
2942 	dtrace_specid_t i;
2943 
2944 	cookie = dtrace_interrupt_disable();
2945 
2946 	if (dest->dtb_tomax == NULL) {
2947 		dtrace_interrupt_enable(cookie);
2948 		return;
2949 	}
2950 
2951 	for (i = 0; i < state->dts_nspeculations; i++) {
2952 		dtrace_speculation_t *spec = &state->dts_speculations[i];
2953 		dtrace_buffer_t *src = &spec->dtsp_buffer[cpu];
2954 
2955 		if (src->dtb_tomax == NULL)
2956 			continue;
2957 
2958 		if (spec->dtsp_state == DTRACESPEC_DISCARDING) {
2959 			src->dtb_offset = 0;
2960 			continue;
2961 		}
2962 
2963 		if (spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
2964 			continue;
2965 
2966 		if (src->dtb_offset == 0)
2967 			continue;
2968 
2969 		dtrace_speculation_commit(state, cpu, i + 1);
2970 	}
2971 
2972 	dtrace_interrupt_enable(cookie);
2973 }
2974 
2975 /*
2976  * Note:  not called from probe context.  This function is called
2977  * asynchronously (and at a regular interval) to clean any speculations that
2978  * are in the COMMITTINGMANY or DISCARDING states.  If it discovers that there
2979  * is work to be done, it cross calls all CPUs to perform that work;
2980  * COMMITMANY and DISCARDING speculations may not be transitioned back to the
2981  * INACTIVE state until they have been cleaned by all CPUs.
2982  */
2983 static void
2984 dtrace_speculation_clean(dtrace_state_t *state)
2985 {
2986 	int work = 0, rv;
2987 	dtrace_specid_t i;
2988 
2989 	for (i = 0; i < state->dts_nspeculations; i++) {
2990 		dtrace_speculation_t *spec = &state->dts_speculations[i];
2991 
2992 		ASSERT(!spec->dtsp_cleaning);
2993 
2994 		if (spec->dtsp_state != DTRACESPEC_DISCARDING &&
2995 		    spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
2996 			continue;
2997 
2998 		work++;
2999 		spec->dtsp_cleaning = 1;
3000 	}
3001 
3002 	if (!work)
3003 		return;
3004 
3005 	dtrace_xcall(DTRACE_CPUALL,
3006 	    (dtrace_xcall_t)dtrace_speculation_clean_here, state);
3007 
3008 	/*
3009 	 * We now know that all CPUs have committed or discarded their
3010 	 * speculation buffers, as appropriate.  We can now set the state
3011 	 * to inactive.
3012 	 */
3013 	for (i = 0; i < state->dts_nspeculations; i++) {
3014 		dtrace_speculation_t *spec = &state->dts_speculations[i];
3015 		dtrace_speculation_state_t current, new;
3016 
3017 		if (!spec->dtsp_cleaning)
3018 			continue;
3019 
3020 		current = spec->dtsp_state;
3021 		ASSERT(current == DTRACESPEC_DISCARDING ||
3022 		    current == DTRACESPEC_COMMITTINGMANY);
3023 
3024 		new = DTRACESPEC_INACTIVE;
3025 
3026 		rv = dtrace_cas32((uint32_t *)&spec->dtsp_state, current, new);
3027 		ASSERT(rv == current);
3028 		spec->dtsp_cleaning = 0;
3029 	}
3030 }
3031 
3032 /*
3033  * Called as part of a speculate() to get the speculative buffer associated
3034  * with a given speculation.  Returns NULL if the specified speculation is not
3035  * in an ACTIVE state.  If the speculation is in the ACTIVEONE state -- and
3036  * the active CPU is not the specified CPU -- the speculation will be
3037  * atomically transitioned into the ACTIVEMANY state.
3038  */
3039 static dtrace_buffer_t *
3040 dtrace_speculation_buffer(dtrace_state_t *state, processorid_t cpuid,
3041     dtrace_specid_t which)
3042 {
3043 	dtrace_speculation_t *spec;
3044 	dtrace_speculation_state_t current, new;
3045 	dtrace_buffer_t *buf;
3046 
3047 	if (which == 0)
3048 		return (NULL);
3049 
3050 	if (which > state->dts_nspeculations) {
3051 		cpu_core[cpuid].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
3052 		return (NULL);
3053 	}
3054 
3055 	spec = &state->dts_speculations[which - 1];
3056 	buf = &spec->dtsp_buffer[cpuid];
3057 
3058 	do {
3059 		current = spec->dtsp_state;
3060 
3061 		switch (current) {
3062 		case DTRACESPEC_INACTIVE:
3063 		case DTRACESPEC_COMMITTINGMANY:
3064 		case DTRACESPEC_DISCARDING:
3065 			return (NULL);
3066 
3067 		case DTRACESPEC_COMMITTING:
3068 			ASSERT(buf->dtb_offset == 0);
3069 			return (NULL);
3070 
3071 		case DTRACESPEC_ACTIVEONE:
3072 			/*
3073 			 * This speculation is currently active on one CPU.
3074 			 * Check the offset in the buffer; if it's non-zero,
3075 			 * that CPU must be us (and we leave the state alone).
3076 			 * If it's zero, assume that we're starting on a new
3077 			 * CPU -- and change the state to indicate that the
3078 			 * speculation is active on more than one CPU.
3079 			 */
3080 			if (buf->dtb_offset != 0)
3081 				return (buf);
3082 
3083 			new = DTRACESPEC_ACTIVEMANY;
3084 			break;
3085 
3086 		case DTRACESPEC_ACTIVEMANY:
3087 			return (buf);
3088 
3089 		case DTRACESPEC_ACTIVE:
3090 			new = DTRACESPEC_ACTIVEONE;
3091 			break;
3092 
3093 		default:
3094 			ASSERT(0);
3095 		}
3096 	} while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
3097 	    current, new) != current);
3098 
3099 	ASSERT(new == DTRACESPEC_ACTIVEONE || new == DTRACESPEC_ACTIVEMANY);
3100 	return (buf);
3101 }
3102 
3103 /*
3104  * Return a string.  In the event that the user lacks the privilege to access
3105  * arbitrary kernel memory, we copy the string out to scratch memory so that we
3106  * don't fail access checking.
3107  *
3108  * dtrace_dif_variable() uses this routine as a helper for various
3109  * builtin values such as 'execname' and 'probefunc.'
3110  */
3111 uintptr_t
3112 dtrace_dif_varstr(uintptr_t addr, dtrace_state_t *state,
3113     dtrace_mstate_t *mstate)
3114 {
3115 	uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3116 	uintptr_t ret;
3117 	size_t strsz;
3118 
3119 	/*
3120 	 * The easy case: this probe is allowed to read all of memory, so
3121 	 * we can just return this as a vanilla pointer.
3122 	 */
3123 	if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
3124 		return (addr);
3125 
3126 	/*
3127 	 * This is the tougher case: we copy the string in question from
3128 	 * kernel memory into scratch memory and return it that way: this
3129 	 * ensures that we won't trip up when access checking tests the
3130 	 * BYREF return value.
3131 	 */
3132 	strsz = dtrace_strlen((char *)addr, size) + 1;
3133 
3134 	if (mstate->dtms_scratch_ptr + strsz >
3135 	    mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
3136 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3137 		return (NULL);
3138 	}
3139 
3140 	dtrace_strcpy((const void *)addr, (void *)mstate->dtms_scratch_ptr,
3141 	    strsz);
3142 	ret = mstate->dtms_scratch_ptr;
3143 	mstate->dtms_scratch_ptr += strsz;
3144 	return (ret);
3145 }
3146 
3147 /*
3148  * This function implements the DIF emulator's variable lookups.  The emulator
3149  * passes a reserved variable identifier and optional built-in array index.
3150  */
3151 static uint64_t
3152 dtrace_dif_variable(dtrace_mstate_t *mstate, dtrace_state_t *state, uint64_t v,
3153     uint64_t ndx)
3154 {
3155 	/*
3156 	 * If we're accessing one of the uncached arguments, we'll turn this
3157 	 * into a reference in the args array.
3158 	 */
3159 	if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9) {
3160 		ndx = v - DIF_VAR_ARG0;
3161 		v = DIF_VAR_ARGS;
3162 	}
3163 
3164 	switch (v) {
3165 	case DIF_VAR_ARGS:
3166 		if (!(mstate->dtms_access & DTRACE_ACCESS_ARGS)) {
3167 			cpu_core[CPU->cpu_id].cpuc_dtrace_flags |=
3168 			    CPU_DTRACE_KPRIV;
3169 			return (0);
3170 		}
3171 
3172 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_ARGS);
3173 		if (ndx >= sizeof (mstate->dtms_arg) /
3174 		    sizeof (mstate->dtms_arg[0])) {
3175 			int aframes = mstate->dtms_probe->dtpr_aframes + 2;
3176 			dtrace_provider_t *pv;
3177 			uint64_t val;
3178 
3179 			pv = mstate->dtms_probe->dtpr_provider;
3180 			if (pv->dtpv_pops.dtps_getargval != NULL)
3181 				val = pv->dtpv_pops.dtps_getargval(pv->dtpv_arg,
3182 				    mstate->dtms_probe->dtpr_id,
3183 				    mstate->dtms_probe->dtpr_arg, ndx, aframes);
3184 			else
3185 				val = dtrace_getarg(ndx, aframes);
3186 
3187 			/*
3188 			 * This is regrettably required to keep the compiler
3189 			 * from tail-optimizing the call to dtrace_getarg().
3190 			 * The condition always evaluates to true, but the
3191 			 * compiler has no way of figuring that out a priori.
3192 			 * (None of this would be necessary if the compiler
3193 			 * could be relied upon to _always_ tail-optimize
3194 			 * the call to dtrace_getarg() -- but it can't.)
3195 			 */
3196 			if (mstate->dtms_probe != NULL)
3197 				return (val);
3198 
3199 			ASSERT(0);
3200 		}
3201 
3202 		return (mstate->dtms_arg[ndx]);
3203 
3204 	case DIF_VAR_UREGS: {
3205 		klwp_t *lwp;
3206 
3207 		if (!dtrace_priv_proc(state, mstate))
3208 			return (0);
3209 
3210 		if ((lwp = curthread->t_lwp) == NULL) {
3211 			DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
3212 			cpu_core[CPU->cpu_id].cpuc_dtrace_illval = NULL;
3213 			return (0);
3214 		}
3215 
3216 		return (dtrace_getreg(lwp->lwp_regs, ndx));
3217 	}
3218 
3219 	case DIF_VAR_VMREGS: {
3220 		uint64_t rval;
3221 
3222 		if (!dtrace_priv_kernel(state))
3223 			return (0);
3224 
3225 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3226 
3227 		rval = dtrace_getvmreg(ndx,
3228 		    &cpu_core[CPU->cpu_id].cpuc_dtrace_flags);
3229 
3230 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3231 
3232 		return (rval);
3233 	}
3234 
3235 	case DIF_VAR_CURTHREAD:
3236 		if (!dtrace_priv_proc(state, mstate))
3237 			return (0);
3238 		return ((uint64_t)(uintptr_t)curthread);
3239 
3240 	case DIF_VAR_TIMESTAMP:
3241 		if (!(mstate->dtms_present & DTRACE_MSTATE_TIMESTAMP)) {
3242 			mstate->dtms_timestamp = dtrace_gethrtime();
3243 			mstate->dtms_present |= DTRACE_MSTATE_TIMESTAMP;
3244 		}
3245 		return (mstate->dtms_timestamp);
3246 
3247 	case DIF_VAR_VTIMESTAMP:
3248 		ASSERT(dtrace_vtime_references != 0);
3249 		return (curthread->t_dtrace_vtime);
3250 
3251 	case DIF_VAR_WALLTIMESTAMP:
3252 		if (!(mstate->dtms_present & DTRACE_MSTATE_WALLTIMESTAMP)) {
3253 			mstate->dtms_walltimestamp = dtrace_gethrestime();
3254 			mstate->dtms_present |= DTRACE_MSTATE_WALLTIMESTAMP;
3255 		}
3256 		return (mstate->dtms_walltimestamp);
3257 
3258 	case DIF_VAR_IPL:
3259 		if (!dtrace_priv_kernel(state))
3260 			return (0);
3261 		if (!(mstate->dtms_present & DTRACE_MSTATE_IPL)) {
3262 			mstate->dtms_ipl = dtrace_getipl();
3263 			mstate->dtms_present |= DTRACE_MSTATE_IPL;
3264 		}
3265 		return (mstate->dtms_ipl);
3266 
3267 	case DIF_VAR_EPID:
3268 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_EPID);
3269 		return (mstate->dtms_epid);
3270 
3271 	case DIF_VAR_ID:
3272 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3273 		return (mstate->dtms_probe->dtpr_id);
3274 
3275 	case DIF_VAR_STACKDEPTH:
3276 		if (!dtrace_priv_kernel(state))
3277 			return (0);
3278 		if (!(mstate->dtms_present & DTRACE_MSTATE_STACKDEPTH)) {
3279 			int aframes = mstate->dtms_probe->dtpr_aframes + 2;
3280 
3281 			mstate->dtms_stackdepth = dtrace_getstackdepth(aframes);
3282 			mstate->dtms_present |= DTRACE_MSTATE_STACKDEPTH;
3283 		}
3284 		return (mstate->dtms_stackdepth);
3285 
3286 	case DIF_VAR_USTACKDEPTH:
3287 		if (!dtrace_priv_proc(state, mstate))
3288 			return (0);
3289 		if (!(mstate->dtms_present & DTRACE_MSTATE_USTACKDEPTH)) {
3290 			/*
3291 			 * See comment in DIF_VAR_PID.
3292 			 */
3293 			if (DTRACE_ANCHORED(mstate->dtms_probe) &&
3294 			    CPU_ON_INTR(CPU)) {
3295 				mstate->dtms_ustackdepth = 0;
3296 			} else {
3297 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3298 				mstate->dtms_ustackdepth =
3299 				    dtrace_getustackdepth();
3300 				DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3301 			}
3302 			mstate->dtms_present |= DTRACE_MSTATE_USTACKDEPTH;
3303 		}
3304 		return (mstate->dtms_ustackdepth);
3305 
3306 	case DIF_VAR_CALLER:
3307 		if (!dtrace_priv_kernel(state))
3308 			return (0);
3309 		if (!(mstate->dtms_present & DTRACE_MSTATE_CALLER)) {
3310 			int aframes = mstate->dtms_probe->dtpr_aframes + 2;
3311 
3312 			if (!DTRACE_ANCHORED(mstate->dtms_probe)) {
3313 				/*
3314 				 * If this is an unanchored probe, we are
3315 				 * required to go through the slow path:
3316 				 * dtrace_caller() only guarantees correct
3317 				 * results for anchored probes.
3318 				 */
3319 				pc_t caller[2];
3320 
3321 				dtrace_getpcstack(caller, 2, aframes,
3322 				    (uint32_t *)(uintptr_t)mstate->dtms_arg[0]);
3323 				mstate->dtms_caller = caller[1];
3324 			} else if ((mstate->dtms_caller =
3325 			    dtrace_caller(aframes)) == -1) {
3326 				/*
3327 				 * We have failed to do this the quick way;
3328 				 * we must resort to the slower approach of
3329 				 * calling dtrace_getpcstack().
3330 				 */
3331 				pc_t caller;
3332 
3333 				dtrace_getpcstack(&caller, 1, aframes, NULL);
3334 				mstate->dtms_caller = caller;
3335 			}
3336 
3337 			mstate->dtms_present |= DTRACE_MSTATE_CALLER;
3338 		}
3339 		return (mstate->dtms_caller);
3340 
3341 	case DIF_VAR_UCALLER:
3342 		if (!dtrace_priv_proc(state, mstate))
3343 			return (0);
3344 
3345 		if (!(mstate->dtms_present & DTRACE_MSTATE_UCALLER)) {
3346 			uint64_t ustack[3];
3347 
3348 			/*
3349 			 * dtrace_getupcstack() fills in the first uint64_t
3350 			 * with the current PID.  The second uint64_t will
3351 			 * be the program counter at user-level.  The third
3352 			 * uint64_t will contain the caller, which is what
3353 			 * we're after.
3354 			 */
3355 			ustack[2] = NULL;
3356 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3357 			dtrace_getupcstack(ustack, 3);
3358 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3359 			mstate->dtms_ucaller = ustack[2];
3360 			mstate->dtms_present |= DTRACE_MSTATE_UCALLER;
3361 		}
3362 
3363 		return (mstate->dtms_ucaller);
3364 
3365 	case DIF_VAR_PROBEPROV:
3366 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3367 		return (dtrace_dif_varstr(
3368 		    (uintptr_t)mstate->dtms_probe->dtpr_provider->dtpv_name,
3369 		    state, mstate));
3370 
3371 	case DIF_VAR_PROBEMOD:
3372 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3373 		return (dtrace_dif_varstr(
3374 		    (uintptr_t)mstate->dtms_probe->dtpr_mod,
3375 		    state, mstate));
3376 
3377 	case DIF_VAR_PROBEFUNC:
3378 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3379 		return (dtrace_dif_varstr(
3380 		    (uintptr_t)mstate->dtms_probe->dtpr_func,
3381 		    state, mstate));
3382 
3383 	case DIF_VAR_PROBENAME:
3384 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3385 		return (dtrace_dif_varstr(
3386 		    (uintptr_t)mstate->dtms_probe->dtpr_name,
3387 		    state, mstate));
3388 
3389 	case DIF_VAR_PID:
3390 		if (!dtrace_priv_proc(state, mstate))
3391 			return (0);
3392 
3393 		/*
3394 		 * Note that we are assuming that an unanchored probe is
3395 		 * always due to a high-level interrupt.  (And we're assuming
3396 		 * that there is only a single high level interrupt.)
3397 		 */
3398 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3399 			return (pid0.pid_id);
3400 
3401 		/*
3402 		 * It is always safe to dereference one's own t_procp pointer:
3403 		 * it always points to a valid, allocated proc structure.
3404 		 * Further, it is always safe to dereference the p_pidp member
3405 		 * of one's own proc structure.  (These are truisms becuase
3406 		 * threads and processes don't clean up their own state --
3407 		 * they leave that task to whomever reaps them.)
3408 		 */
3409 		return ((uint64_t)curthread->t_procp->p_pidp->pid_id);
3410 
3411 	case DIF_VAR_PPID:
3412 		if (!dtrace_priv_proc(state, mstate))
3413 			return (0);
3414 
3415 		/*
3416 		 * See comment in DIF_VAR_PID.
3417 		 */
3418 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3419 			return (pid0.pid_id);
3420 
3421 		/*
3422 		 * It is always safe to dereference one's own t_procp pointer:
3423 		 * it always points to a valid, allocated proc structure.
3424 		 * (This is true because threads don't clean up their own
3425 		 * state -- they leave that task to whomever reaps them.)
3426 		 */
3427 		return ((uint64_t)curthread->t_procp->p_ppid);
3428 
3429 	case DIF_VAR_TID:
3430 		/*
3431 		 * See comment in DIF_VAR_PID.
3432 		 */
3433 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3434 			return (0);
3435 
3436 		return ((uint64_t)curthread->t_tid);
3437 
3438 	case DIF_VAR_EXECNAME:
3439 		if (!dtrace_priv_proc(state, mstate))
3440 			return (0);
3441 
3442 		/*
3443 		 * See comment in DIF_VAR_PID.
3444 		 */
3445 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3446 			return ((uint64_t)(uintptr_t)p0.p_user.u_comm);
3447 
3448 		/*
3449 		 * It is always safe to dereference one's own t_procp pointer:
3450 		 * it always points to a valid, allocated proc structure.
3451 		 * (This is true because threads don't clean up their own
3452 		 * state -- they leave that task to whomever reaps them.)
3453 		 */
3454 		return (dtrace_dif_varstr(
3455 		    (uintptr_t)curthread->t_procp->p_user.u_comm,
3456 		    state, mstate));
3457 
3458 	case DIF_VAR_ZONENAME:
3459 		if (!dtrace_priv_proc(state, mstate))
3460 			return (0);
3461 
3462 		/*
3463 		 * See comment in DIF_VAR_PID.
3464 		 */
3465 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3466 			return ((uint64_t)(uintptr_t)p0.p_zone->zone_name);
3467 
3468 		/*
3469 		 * It is always safe to dereference one's own t_procp pointer:
3470 		 * it always points to a valid, allocated proc structure.
3471 		 * (This is true because threads don't clean up their own
3472 		 * state -- they leave that task to whomever reaps them.)
3473 		 */
3474 		return (dtrace_dif_varstr(
3475 		    (uintptr_t)curthread->t_procp->p_zone->zone_name,
3476 		    state, mstate));
3477 
3478 	case DIF_VAR_UID:
3479 		if (!dtrace_priv_proc(state, mstate))
3480 			return (0);
3481 
3482 		/*
3483 		 * See comment in DIF_VAR_PID.
3484 		 */
3485 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3486 			return ((uint64_t)p0.p_cred->cr_uid);
3487 
3488 		/*
3489 		 * It is always safe to dereference one's own t_procp pointer:
3490 		 * it always points to a valid, allocated proc structure.
3491 		 * (This is true because threads don't clean up their own
3492 		 * state -- they leave that task to whomever reaps them.)
3493 		 *
3494 		 * Additionally, it is safe to dereference one's own process
3495 		 * credential, since this is never NULL after process birth.
3496 		 */
3497 		return ((uint64_t)curthread->t_procp->p_cred->cr_uid);
3498 
3499 	case DIF_VAR_GID:
3500 		if (!dtrace_priv_proc(state, mstate))
3501 			return (0);
3502 
3503 		/*
3504 		 * See comment in DIF_VAR_PID.
3505 		 */
3506 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3507 			return ((uint64_t)p0.p_cred->cr_gid);
3508 
3509 		/*
3510 		 * It is always safe to dereference one's own t_procp pointer:
3511 		 * it always points to a valid, allocated proc structure.
3512 		 * (This is true because threads don't clean up their own
3513 		 * state -- they leave that task to whomever reaps them.)
3514 		 *
3515 		 * Additionally, it is safe to dereference one's own process
3516 		 * credential, since this is never NULL after process birth.
3517 		 */
3518 		return ((uint64_t)curthread->t_procp->p_cred->cr_gid);
3519 
3520 	case DIF_VAR_ERRNO: {
3521 		klwp_t *lwp;
3522 		if (!dtrace_priv_proc(state, mstate))
3523 			return (0);
3524 
3525 		/*
3526 		 * See comment in DIF_VAR_PID.
3527 		 */
3528 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3529 			return (0);
3530 
3531 		/*
3532 		 * It is always safe to dereference one's own t_lwp pointer in
3533 		 * the event that this pointer is non-NULL.  (This is true
3534 		 * because threads and lwps don't clean up their own state --
3535 		 * they leave that task to whomever reaps them.)
3536 		 */
3537 		if ((lwp = curthread->t_lwp) == NULL)
3538 			return (0);
3539 
3540 		return ((uint64_t)lwp->lwp_errno);
3541 	}
3542 	default:
3543 		DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
3544 		return (0);
3545 	}
3546 }
3547 
3548 
3549 typedef enum dtrace_json_state {
3550 	DTRACE_JSON_REST = 1,
3551 	DTRACE_JSON_OBJECT,
3552 	DTRACE_JSON_STRING,
3553 	DTRACE_JSON_STRING_ESCAPE,
3554 	DTRACE_JSON_STRING_ESCAPE_UNICODE,
3555 	DTRACE_JSON_COLON,
3556 	DTRACE_JSON_COMMA,
3557 	DTRACE_JSON_VALUE,
3558 	DTRACE_JSON_IDENTIFIER,
3559 	DTRACE_JSON_NUMBER,
3560 	DTRACE_JSON_NUMBER_FRAC,
3561 	DTRACE_JSON_NUMBER_EXP,
3562 	DTRACE_JSON_COLLECT_OBJECT
3563 } dtrace_json_state_t;
3564 
3565 /*
3566  * This function possesses just enough knowledge about JSON to extract a single
3567  * value from a JSON string and store it in the scratch buffer.  It is able
3568  * to extract nested object values, and members of arrays by index.
3569  *
3570  * elemlist is a list of JSON keys, stored as packed NUL-terminated strings, to
3571  * be looked up as we descend into the object tree.  e.g.
3572  *
3573  *    foo[0].bar.baz[32] --> "foo" NUL "0" NUL "bar" NUL "baz" NUL "32" NUL
3574  *       with nelems = 5.
3575  *
3576  * The run time of this function must be bounded above by strsize to limit the
3577  * amount of work done in probe context.  As such, it is implemented as a
3578  * simple state machine, reading one character at a time using safe loads
3579  * until we find the requested element, hit a parsing error or run off the
3580  * end of the object or string.
3581  *
3582  * As there is no way for a subroutine to return an error without interrupting
3583  * clause execution, we simply return NULL in the event of a missing key or any
3584  * other error condition.  Each NULL return in this function is commented with
3585  * the error condition it represents -- parsing or otherwise.
3586  *
3587  * The set of states for the state machine closely matches the JSON
3588  * specification (http://json.org/).  Briefly:
3589  *
3590  *   DTRACE_JSON_REST:
3591  *     Skip whitespace until we find either a top-level Object, moving
3592  *     to DTRACE_JSON_OBJECT; or an Array, moving to DTRACE_JSON_VALUE.
3593  *
3594  *   DTRACE_JSON_OBJECT:
3595  *     Locate the next key String in an Object.  Sets a flag to denote
3596  *     the next String as a key string and moves to DTRACE_JSON_STRING.
3597  *
3598  *   DTRACE_JSON_COLON:
3599  *     Skip whitespace until we find the colon that separates key Strings
3600  *     from their values.  Once found, move to DTRACE_JSON_VALUE.
3601  *
3602  *   DTRACE_JSON_VALUE:
3603  *     Detects the type of the next value (String, Number, Identifier, Object
3604  *     or Array) and routes to the states that process that type.  Here we also
3605  *     deal with the element selector list if we are requested to traverse down
3606  *     into the object tree.
3607  *
3608  *   DTRACE_JSON_COMMA:
3609  *     Skip whitespace until we find the comma that separates key-value pairs
3610  *     in Objects (returning to DTRACE_JSON_OBJECT) or values in Arrays
3611  *     (similarly DTRACE_JSON_VALUE).  All following literal value processing
3612  *     states return to this state at the end of their value, unless otherwise
3613  *     noted.
3614  *
3615  *   DTRACE_JSON_NUMBER, DTRACE_JSON_NUMBER_FRAC, DTRACE_JSON_NUMBER_EXP:
3616  *     Processes a Number literal from the JSON, including any exponent
3617  *     component that may be present.  Numbers are returned as strings, which
3618  *     may be passed to strtoll() if an integer is required.
3619  *
3620  *   DTRACE_JSON_IDENTIFIER:
3621  *     Processes a "true", "false" or "null" literal in the JSON.
3622  *
3623  *   DTRACE_JSON_STRING, DTRACE_JSON_STRING_ESCAPE,
3624  *   DTRACE_JSON_STRING_ESCAPE_UNICODE:
3625  *     Processes a String literal from the JSON, whether the String denotes
3626  *     a key, a value or part of a larger Object.  Handles all escape sequences
3627  *     present in the specification, including four-digit unicode characters,
3628  *     but merely includes the escape sequence without converting it to the
3629  *     actual escaped character.  If the String is flagged as a key, we
3630  *     move to DTRACE_JSON_COLON rather than DTRACE_JSON_COMMA.
3631  *
3632  *   DTRACE_JSON_COLLECT_OBJECT:
3633  *     This state collects an entire Object (or Array), correctly handling
3634  *     embedded strings.  If the full element selector list matches this nested
3635  *     object, we return the Object in full as a string.  If not, we use this
3636  *     state to skip to the next value at this level and continue processing.
3637  *
3638  * NOTE: This function uses various macros from strtolctype.h to manipulate
3639  * digit values, etc -- these have all been checked to ensure they make
3640  * no additional function calls.
3641  */
3642 static char *
3643 dtrace_json(uint64_t size, uintptr_t json, char *elemlist, int nelems,
3644     char *dest)
3645 {
3646 	dtrace_json_state_t state = DTRACE_JSON_REST;
3647 	int64_t array_elem = INT64_MIN;
3648 	int64_t array_pos = 0;
3649 	uint8_t escape_unicount = 0;
3650 	boolean_t string_is_key = B_FALSE;
3651 	boolean_t collect_object = B_FALSE;
3652 	boolean_t found_key = B_FALSE;
3653 	boolean_t in_array = B_FALSE;
3654 	uint32_t braces = 0, brackets = 0;
3655 	char *elem = elemlist;
3656 	char *dd = dest;
3657 	uintptr_t cur;
3658 
3659 	for (cur = json; cur < json + size; cur++) {
3660 		char cc = dtrace_load8(cur);
3661 		if (cc == '\0')
3662 			return (NULL);
3663 
3664 		switch (state) {
3665 		case DTRACE_JSON_REST:
3666 			if (isspace(cc))
3667 				break;
3668 
3669 			if (cc == '{') {
3670 				state = DTRACE_JSON_OBJECT;
3671 				break;
3672 			}
3673 
3674 			if (cc == '[') {
3675 				in_array = B_TRUE;
3676 				array_pos = 0;
3677 				array_elem = dtrace_strtoll(elem, 10, size);
3678 				found_key = array_elem == 0 ? B_TRUE : B_FALSE;
3679 				state = DTRACE_JSON_VALUE;
3680 				break;
3681 			}
3682 
3683 			/*
3684 			 * ERROR: expected to find a top-level object or array.
3685 			 */
3686 			return (NULL);
3687 		case DTRACE_JSON_OBJECT:
3688 			if (isspace(cc))
3689 				break;
3690 
3691 			if (cc == '"') {
3692 				state = DTRACE_JSON_STRING;
3693 				string_is_key = B_TRUE;
3694 				break;
3695 			}
3696 
3697 			/*
3698 			 * ERROR: either the object did not start with a key
3699 			 * string, or we've run off the end of the object
3700 			 * without finding the requested key.
3701 			 */
3702 			return (NULL);
3703 		case DTRACE_JSON_STRING:
3704 			if (cc == '\\') {
3705 				*dd++ = '\\';
3706 				state = DTRACE_JSON_STRING_ESCAPE;
3707 				break;
3708 			}
3709 
3710 			if (cc == '"') {
3711 				if (collect_object) {
3712 					/*
3713 					 * We don't reset the dest here, as
3714 					 * the string is part of a larger
3715 					 * object being collected.
3716 					 */
3717 					*dd++ = cc;
3718 					collect_object = B_FALSE;
3719 					state = DTRACE_JSON_COLLECT_OBJECT;
3720 					break;
3721 				}
3722 				*dd = '\0';
3723 				dd = dest; /* reset string buffer */
3724 				if (string_is_key) {
3725 					if (dtrace_strncmp(dest, elem,
3726 					    size) == 0)
3727 						found_key = B_TRUE;
3728 				} else if (found_key) {
3729 					if (nelems > 1) {
3730 						/*
3731 						 * We expected an object, not
3732 						 * this string.
3733 						 */
3734 						return (NULL);
3735 					}
3736 					return (dest);
3737 				}
3738 				state = string_is_key ? DTRACE_JSON_COLON :
3739 				    DTRACE_JSON_COMMA;
3740 				string_is_key = B_FALSE;
3741 				break;
3742 			}
3743 
3744 			*dd++ = cc;
3745 			break;
3746 		case DTRACE_JSON_STRING_ESCAPE:
3747 			*dd++ = cc;
3748 			if (cc == 'u') {
3749 				escape_unicount = 0;
3750 				state = DTRACE_JSON_STRING_ESCAPE_UNICODE;
3751 			} else {
3752 				state = DTRACE_JSON_STRING;
3753 			}
3754 			break;
3755 		case DTRACE_JSON_STRING_ESCAPE_UNICODE:
3756 			if (!isxdigit(cc)) {
3757 				/*
3758 				 * ERROR: invalid unicode escape, expected
3759 				 * four valid hexidecimal digits.
3760 				 */
3761 				return (NULL);
3762 			}
3763 
3764 			*dd++ = cc;
3765 			if (++escape_unicount == 4)
3766 				state = DTRACE_JSON_STRING;
3767 			break;
3768 		case DTRACE_JSON_COLON:
3769 			if (isspace(cc))
3770 				break;
3771 
3772 			if (cc == ':') {
3773 				state = DTRACE_JSON_VALUE;
3774 				break;
3775 			}
3776 
3777 			/*
3778 			 * ERROR: expected a colon.
3779 			 */
3780 			return (NULL);
3781 		case DTRACE_JSON_COMMA:
3782 			if (isspace(cc))
3783 				break;
3784 
3785 			if (cc == ',') {
3786 				if (in_array) {
3787 					state = DTRACE_JSON_VALUE;
3788 					if (++array_pos == array_elem)
3789 						found_key = B_TRUE;
3790 				} else {
3791 					state = DTRACE_JSON_OBJECT;
3792 				}
3793 				break;
3794 			}
3795 
3796 			/*
3797 			 * ERROR: either we hit an unexpected character, or
3798 			 * we reached the end of the object or array without
3799 			 * finding the requested key.
3800 			 */
3801 			return (NULL);
3802 		case DTRACE_JSON_IDENTIFIER:
3803 			if (islower(cc)) {
3804 				*dd++ = cc;
3805 				break;
3806 			}
3807 
3808 			*dd = '\0';
3809 			dd = dest; /* reset string buffer */
3810 
3811 			if (dtrace_strncmp(dest, "true", 5) == 0 ||
3812 			    dtrace_strncmp(dest, "false", 6) == 0 ||
3813 			    dtrace_strncmp(dest, "null", 5) == 0) {
3814 				if (found_key) {
3815 					if (nelems > 1) {
3816 						/*
3817 						 * ERROR: We expected an object,
3818 						 * not this identifier.
3819 						 */
3820 						return (NULL);
3821 					}
3822 					return (dest);
3823 				} else {
3824 					cur--;
3825 					state = DTRACE_JSON_COMMA;
3826 					break;
3827 				}
3828 			}
3829 
3830 			/*
3831 			 * ERROR: we did not recognise the identifier as one
3832 			 * of those in the JSON specification.
3833 			 */
3834 			return (NULL);
3835 		case DTRACE_JSON_NUMBER:
3836 			if (cc == '.') {
3837 				*dd++ = cc;
3838 				state = DTRACE_JSON_NUMBER_FRAC;
3839 				break;
3840 			}
3841 
3842 			if (cc == 'x' || cc == 'X') {
3843 				/*
3844 				 * ERROR: specification explicitly excludes
3845 				 * hexidecimal or octal numbers.
3846 				 */
3847 				return (NULL);
3848 			}
3849 
3850 			/* FALLTHRU */
3851 		case DTRACE_JSON_NUMBER_FRAC:
3852 			if (cc == 'e' || cc == 'E') {
3853 				*dd++ = cc;
3854 				state = DTRACE_JSON_NUMBER_EXP;
3855 				break;
3856 			}
3857 
3858 			if (cc == '+' || cc == '-') {
3859 				/*
3860 				 * ERROR: expect sign as part of exponent only.
3861 				 */
3862 				return (NULL);
3863 			}
3864 			/* FALLTHRU */
3865 		case DTRACE_JSON_NUMBER_EXP:
3866 			if (isdigit(cc) || cc == '+' || cc == '-') {
3867 				*dd++ = cc;
3868 				break;
3869 			}
3870 
3871 			*dd = '\0';
3872 			dd = dest; /* reset string buffer */
3873 			if (found_key) {
3874 				if (nelems > 1) {
3875 					/*
3876 					 * ERROR: We expected an object, not
3877 					 * this number.
3878 					 */
3879 					return (NULL);
3880 				}
3881 				return (dest);
3882 			}
3883 
3884 			cur--;
3885 			state = DTRACE_JSON_COMMA;
3886 			break;
3887 		case DTRACE_JSON_VALUE:
3888 			if (isspace(cc))
3889 				break;
3890 
3891 			if (cc == '{' || cc == '[') {
3892 				if (nelems > 1 && found_key) {
3893 					in_array = cc == '[' ? B_TRUE : B_FALSE;
3894 					/*
3895 					 * If our element selector directs us
3896 					 * to descend into this nested object,
3897 					 * then move to the next selector
3898 					 * element in the list and restart the
3899 					 * state machine.
3900 					 */
3901 					while (*elem != '\0')
3902 						elem++;
3903 					elem++; /* skip the inter-element NUL */
3904 					nelems--;
3905 					dd = dest;
3906 					if (in_array) {
3907 						state = DTRACE_JSON_VALUE;
3908 						array_pos = 0;
3909 						array_elem = dtrace_strtoll(
3910 						    elem, 10, size);
3911 						found_key = array_elem == 0 ?
3912 						    B_TRUE : B_FALSE;
3913 					} else {
3914 						found_key = B_FALSE;
3915 						state = DTRACE_JSON_OBJECT;
3916 					}
3917 					break;
3918 				}
3919 
3920 				/*
3921 				 * Otherwise, we wish to either skip this
3922 				 * nested object or return it in full.
3923 				 */
3924 				if (cc == '[')
3925 					brackets = 1;
3926 				else
3927 					braces = 1;
3928 				*dd++ = cc;
3929 				state = DTRACE_JSON_COLLECT_OBJECT;
3930 				break;
3931 			}
3932 
3933 			if (cc == '"') {
3934 				state = DTRACE_JSON_STRING;
3935 				break;
3936 			}
3937 
3938 			if (islower(cc)) {
3939 				/*
3940 				 * Here we deal with true, false and null.
3941 				 */
3942 				*dd++ = cc;
3943 				state = DTRACE_JSON_IDENTIFIER;
3944 				break;
3945 			}
3946 
3947 			if (cc == '-' || isdigit(cc)) {
3948 				*dd++ = cc;
3949 				state = DTRACE_JSON_NUMBER;
3950 				break;
3951 			}
3952 
3953 			/*
3954 			 * ERROR: unexpected character at start of value.
3955 			 */
3956 			return (NULL);
3957 		case DTRACE_JSON_COLLECT_OBJECT:
3958 			if (cc == '\0')
3959 				/*
3960 				 * ERROR: unexpected end of input.
3961 				 */
3962 				return (NULL);
3963 
3964 			*dd++ = cc;
3965 			if (cc == '"') {
3966 				collect_object = B_TRUE;
3967 				state = DTRACE_JSON_STRING;
3968 				break;
3969 			}
3970 
3971 			if (cc == ']') {
3972 				if (brackets-- == 0) {
3973 					/*
3974 					 * ERROR: unbalanced brackets.
3975 					 */
3976 					return (NULL);
3977 				}
3978 			} else if (cc == '}') {
3979 				if (braces-- == 0) {
3980 					/*
3981 					 * ERROR: unbalanced braces.
3982 					 */
3983 					return (NULL);
3984 				}
3985 			} else if (cc == '{') {
3986 				braces++;
3987 			} else if (cc == '[') {
3988 				brackets++;
3989 			}
3990 
3991 			if (brackets == 0 && braces == 0) {
3992 				if (found_key) {
3993 					*dd = '\0';
3994 					return (dest);
3995 				}
3996 				dd = dest; /* reset string buffer */
3997 				state = DTRACE_JSON_COMMA;
3998 			}
3999 			break;
4000 		}
4001 	}
4002 	return (NULL);
4003 }
4004 
4005 /*
4006  * Emulate the execution of DTrace ID subroutines invoked by the call opcode.
4007  * Notice that we don't bother validating the proper number of arguments or
4008  * their types in the tuple stack.  This isn't needed because all argument
4009  * interpretation is safe because of our load safety -- the worst that can
4010  * happen is that a bogus program can obtain bogus results.
4011  */
4012 static void
4013 dtrace_dif_subr(uint_t subr, uint_t rd, uint64_t *regs,
4014     dtrace_key_t *tupregs, int nargs,
4015     dtrace_mstate_t *mstate, dtrace_state_t *state)
4016 {
4017 	volatile uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
4018 	volatile uintptr_t *illval = &cpu_core[CPU->cpu_id].cpuc_dtrace_illval;
4019 	dtrace_vstate_t *vstate = &state->dts_vstate;
4020 
4021 	union {
4022 		mutex_impl_t mi;
4023 		uint64_t mx;
4024 	} m;
4025 
4026 	union {
4027 		krwlock_t ri;
4028 		uintptr_t rw;
4029 	} r;
4030 
4031 	switch (subr) {
4032 	case DIF_SUBR_RAND:
4033 		regs[rd] = (dtrace_gethrtime() * 2416 + 374441) % 1771875;
4034 		break;
4035 
4036 	case DIF_SUBR_MUTEX_OWNED:
4037 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
4038 		    mstate, vstate)) {
4039 			regs[rd] = NULL;
4040 			break;
4041 		}
4042 
4043 		m.mx = dtrace_load64(tupregs[0].dttk_value);
4044 		if (MUTEX_TYPE_ADAPTIVE(&m.mi))
4045 			regs[rd] = MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER;
4046 		else
4047 			regs[rd] = LOCK_HELD(&m.mi.m_spin.m_spinlock);
4048 		break;
4049 
4050 	case DIF_SUBR_MUTEX_OWNER:
4051 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
4052 		    mstate, vstate)) {
4053 			regs[rd] = NULL;
4054 			break;
4055 		}
4056 
4057 		m.mx = dtrace_load64(tupregs[0].dttk_value);
4058 		if (MUTEX_TYPE_ADAPTIVE(&m.mi) &&
4059 		    MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER)
4060 			regs[rd] = (uintptr_t)MUTEX_OWNER(&m.mi);
4061 		else
4062 			regs[rd] = 0;
4063 		break;
4064 
4065 	case DIF_SUBR_MUTEX_TYPE_ADAPTIVE:
4066 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
4067 		    mstate, vstate)) {
4068 			regs[rd] = NULL;
4069 			break;
4070 		}
4071 
4072 		m.mx = dtrace_load64(tupregs[0].dttk_value);
4073 		regs[rd] = MUTEX_TYPE_ADAPTIVE(&m.mi);
4074 		break;
4075 
4076 	case DIF_SUBR_MUTEX_TYPE_SPIN:
4077 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
4078 		    mstate, vstate)) {
4079 			regs[rd] = NULL;
4080 			break;
4081 		}
4082 
4083 		m.mx = dtrace_load64(tupregs[0].dttk_value);
4084 		regs[rd] = MUTEX_TYPE_SPIN(&m.mi);
4085 		break;
4086 
4087 	case DIF_SUBR_RW_READ_HELD: {
4088 		uintptr_t tmp;
4089 
4090 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
4091 		    mstate, vstate)) {
4092 			regs[rd] = NULL;
4093 			break;
4094 		}
4095 
4096 		r.rw = dtrace_loadptr(tupregs[0].dttk_value);
4097 		regs[rd] = _RW_READ_HELD(&r.ri, tmp);
4098 		break;
4099 	}
4100 
4101 	case DIF_SUBR_RW_WRITE_HELD:
4102 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
4103 		    mstate, vstate)) {
4104 			regs[rd] = NULL;
4105 			break;
4106 		}
4107 
4108 		r.rw = dtrace_loadptr(tupregs[0].dttk_value);
4109 		regs[rd] = _RW_WRITE_HELD(&r.ri);
4110 		break;
4111 
4112 	case DIF_SUBR_RW_ISWRITER:
4113 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
4114 		    mstate, vstate)) {
4115 			regs[rd] = NULL;
4116 			break;
4117 		}
4118 
4119 		r.rw = dtrace_loadptr(tupregs[0].dttk_value);
4120 		regs[rd] = _RW_ISWRITER(&r.ri);
4121 		break;
4122 
4123 	case DIF_SUBR_BCOPY: {
4124 		/*
4125 		 * We need to be sure that the destination is in the scratch
4126 		 * region -- no other region is allowed.
4127 		 */
4128 		uintptr_t src = tupregs[0].dttk_value;
4129 		uintptr_t dest = tupregs[1].dttk_value;
4130 		size_t size = tupregs[2].dttk_value;
4131 
4132 		if (!dtrace_inscratch(dest, size, mstate)) {
4133 			*flags |= CPU_DTRACE_BADADDR;
4134 			*illval = regs[rd];
4135 			break;
4136 		}
4137 
4138 		if (!dtrace_canload(src, size, mstate, vstate)) {
4139 			regs[rd] = NULL;
4140 			break;
4141 		}
4142 
4143 		dtrace_bcopy((void *)src, (void *)dest, size);
4144 		break;
4145 	}
4146 
4147 	case DIF_SUBR_ALLOCA:
4148 	case DIF_SUBR_COPYIN: {
4149 		uintptr_t dest = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
4150 		uint64_t size =
4151 		    tupregs[subr == DIF_SUBR_ALLOCA ? 0 : 1].dttk_value;
4152 		size_t scratch_size = (dest - mstate->dtms_scratch_ptr) + size;
4153 
4154 		/*
4155 		 * This action doesn't require any credential checks since
4156 		 * probes will not activate in user contexts to which the
4157 		 * enabling user does not have permissions.
4158 		 */
4159 
4160 		/*
4161 		 * Rounding up the user allocation size could have overflowed
4162 		 * a large, bogus allocation (like -1ULL) to 0.
4163 		 */
4164 		if (scratch_size < size ||
4165 		    !DTRACE_INSCRATCH(mstate, scratch_size)) {
4166 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4167 			regs[rd] = NULL;
4168 			break;
4169 		}
4170 
4171 		if (subr == DIF_SUBR_COPYIN) {
4172 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4173 			dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
4174 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4175 		}
4176 
4177 		mstate->dtms_scratch_ptr += scratch_size;
4178 		regs[rd] = dest;
4179 		break;
4180 	}
4181 
4182 	case DIF_SUBR_COPYINTO: {
4183 		uint64_t size = tupregs[1].dttk_value;
4184 		uintptr_t dest = tupregs[2].dttk_value;
4185 
4186 		/*
4187 		 * This action doesn't require any credential checks since
4188 		 * probes will not activate in user contexts to which the
4189 		 * enabling user does not have permissions.
4190 		 */
4191 		if (!dtrace_inscratch(dest, size, mstate)) {
4192 			*flags |= CPU_DTRACE_BADADDR;
4193 			*illval = regs[rd];
4194 			break;
4195 		}
4196 
4197 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4198 		dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
4199 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4200 		break;
4201 	}
4202 
4203 	case DIF_SUBR_COPYINSTR: {
4204 		uintptr_t dest = mstate->dtms_scratch_ptr;
4205 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4206 
4207 		if (nargs > 1 && tupregs[1].dttk_value < size)
4208 			size = tupregs[1].dttk_value + 1;
4209 
4210 		/*
4211 		 * This action doesn't require any credential checks since
4212 		 * probes will not activate in user contexts to which the
4213 		 * enabling user does not have permissions.
4214 		 */
4215 		if (!DTRACE_INSCRATCH(mstate, size)) {
4216 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4217 			regs[rd] = NULL;
4218 			break;
4219 		}
4220 
4221 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4222 		dtrace_copyinstr(tupregs[0].dttk_value, dest, size, flags);
4223 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4224 
4225 		((char *)dest)[size - 1] = '\0';
4226 		mstate->dtms_scratch_ptr += size;
4227 		regs[rd] = dest;
4228 		break;
4229 	}
4230 
4231 	case DIF_SUBR_MSGSIZE:
4232 	case DIF_SUBR_MSGDSIZE: {
4233 		uintptr_t baddr = tupregs[0].dttk_value, daddr;
4234 		uintptr_t wptr, rptr;
4235 		size_t count = 0;
4236 		int cont = 0;
4237 
4238 		while (baddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
4239 
4240 			if (!dtrace_canload(baddr, sizeof (mblk_t), mstate,
4241 			    vstate)) {
4242 				regs[rd] = NULL;
4243 				break;
4244 			}
4245 
4246 			wptr = dtrace_loadptr(baddr +
4247 			    offsetof(mblk_t, b_wptr));
4248 
4249 			rptr = dtrace_loadptr(baddr +
4250 			    offsetof(mblk_t, b_rptr));
4251 
4252 			if (wptr < rptr) {
4253 				*flags |= CPU_DTRACE_BADADDR;
4254 				*illval = tupregs[0].dttk_value;
4255 				break;
4256 			}
4257 
4258 			daddr = dtrace_loadptr(baddr +
4259 			    offsetof(mblk_t, b_datap));
4260 
4261 			baddr = dtrace_loadptr(baddr +
4262 			    offsetof(mblk_t, b_cont));
4263 
4264 			/*
4265 			 * We want to prevent against denial-of-service here,
4266 			 * so we're only going to search the list for
4267 			 * dtrace_msgdsize_max mblks.
4268 			 */
4269 			if (cont++ > dtrace_msgdsize_max) {
4270 				*flags |= CPU_DTRACE_ILLOP;
4271 				break;
4272 			}
4273 
4274 			if (subr == DIF_SUBR_MSGDSIZE) {
4275 				if (dtrace_load8(daddr +
4276 				    offsetof(dblk_t, db_type)) != M_DATA)
4277 					continue;
4278 			}
4279 
4280 			count += wptr - rptr;
4281 		}
4282 
4283 		if (!(*flags & CPU_DTRACE_FAULT))
4284 			regs[rd] = count;
4285 
4286 		break;
4287 	}
4288 
4289 	case DIF_SUBR_PROGENYOF: {
4290 		pid_t pid = tupregs[0].dttk_value;
4291 		proc_t *p;
4292 		int rval = 0;
4293 
4294 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4295 
4296 		for (p = curthread->t_procp; p != NULL; p = p->p_parent) {
4297 			if (p->p_pidp->pid_id == pid) {
4298 				rval = 1;
4299 				break;
4300 			}
4301 		}
4302 
4303 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4304 
4305 		regs[rd] = rval;
4306 		break;
4307 	}
4308 
4309 	case DIF_SUBR_SPECULATION:
4310 		regs[rd] = dtrace_speculation(state);
4311 		break;
4312 
4313 	case DIF_SUBR_COPYOUT: {
4314 		uintptr_t kaddr = tupregs[0].dttk_value;
4315 		uintptr_t uaddr = tupregs[1].dttk_value;
4316 		uint64_t size = tupregs[2].dttk_value;
4317 
4318 		if (!dtrace_destructive_disallow &&
4319 		    dtrace_priv_proc_control(state, mstate) &&
4320 		    !dtrace_istoxic(kaddr, size) &&
4321 		    dtrace_canload(kaddr, size, mstate, vstate)) {
4322 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4323 			dtrace_copyout(kaddr, uaddr, size, flags);
4324 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4325 		}
4326 		break;
4327 	}
4328 
4329 	case DIF_SUBR_COPYOUTSTR: {
4330 		uintptr_t kaddr = tupregs[0].dttk_value;
4331 		uintptr_t uaddr = tupregs[1].dttk_value;
4332 		uint64_t size = tupregs[2].dttk_value;
4333 		size_t lim;
4334 
4335 		if (!dtrace_destructive_disallow &&
4336 		    dtrace_priv_proc_control(state, mstate) &&
4337 		    !dtrace_istoxic(kaddr, size) &&
4338 		    dtrace_strcanload(kaddr, size, &lim, mstate, vstate)) {
4339 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4340 			dtrace_copyoutstr(kaddr, uaddr, lim, flags);
4341 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4342 		}
4343 		break;
4344 	}
4345 
4346 	case DIF_SUBR_STRLEN: {
4347 		size_t size = state->dts_options[DTRACEOPT_STRSIZE];
4348 		uintptr_t addr = (uintptr_t)tupregs[0].dttk_value;
4349 		size_t lim;
4350 
4351 		if (!dtrace_strcanload(addr, size, &lim, mstate, vstate)) {
4352 			regs[rd] = NULL;
4353 			break;
4354 		}
4355 		regs[rd] = dtrace_strlen((char *)addr, lim);
4356 
4357 		break;
4358 	}
4359 
4360 	case DIF_SUBR_STRCHR:
4361 	case DIF_SUBR_STRRCHR: {
4362 		/*
4363 		 * We're going to iterate over the string looking for the
4364 		 * specified character.  We will iterate until we have reached
4365 		 * the string length or we have found the character.  If this
4366 		 * is DIF_SUBR_STRRCHR, we will look for the last occurrence
4367 		 * of the specified character instead of the first.
4368 		 */
4369 		uintptr_t addr = tupregs[0].dttk_value;
4370 		uintptr_t addr_limit;
4371 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4372 		size_t lim;
4373 		char c, target = (char)tupregs[1].dttk_value;
4374 
4375 		if (!dtrace_strcanload(addr, size, &lim, mstate, vstate)) {
4376 			regs[rd] = NULL;
4377 			break;
4378 		}
4379 		addr_limit = addr + lim;
4380 
4381 		for (regs[rd] = NULL; addr < addr_limit; addr++) {
4382 			if ((c = dtrace_load8(addr)) == target) {
4383 				regs[rd] = addr;
4384 
4385 				if (subr == DIF_SUBR_STRCHR)
4386 					break;
4387 			}
4388 			if (c == '\0')
4389 				break;
4390 		}
4391 
4392 		break;
4393 	}
4394 
4395 	case DIF_SUBR_STRSTR:
4396 	case DIF_SUBR_INDEX:
4397 	case DIF_SUBR_RINDEX: {
4398 		/*
4399 		 * We're going to iterate over the string looking for the
4400 		 * specified string.  We will iterate until we have reached
4401 		 * the string length or we have found the string.  (Yes, this
4402 		 * is done in the most naive way possible -- but considering
4403 		 * that the string we're searching for is likely to be
4404 		 * relatively short, the complexity of Rabin-Karp or similar
4405 		 * hardly seems merited.)
4406 		 */
4407 		char *addr = (char *)(uintptr_t)tupregs[0].dttk_value;
4408 		char *substr = (char *)(uintptr_t)tupregs[1].dttk_value;
4409 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4410 		size_t len = dtrace_strlen(addr, size);
4411 		size_t sublen = dtrace_strlen(substr, size);
4412 		char *limit = addr + len, *orig = addr;
4413 		int notfound = subr == DIF_SUBR_STRSTR ? 0 : -1;
4414 		int inc = 1;
4415 
4416 		regs[rd] = notfound;
4417 
4418 		if (!dtrace_canload((uintptr_t)addr, len + 1, mstate, vstate)) {
4419 			regs[rd] = NULL;
4420 			break;
4421 		}
4422 
4423 		if (!dtrace_canload((uintptr_t)substr, sublen + 1, mstate,
4424 		    vstate)) {
4425 			regs[rd] = NULL;
4426 			break;
4427 		}
4428 
4429 		/*
4430 		 * strstr() and index()/rindex() have similar semantics if
4431 		 * both strings are the empty string: strstr() returns a
4432 		 * pointer to the (empty) string, and index() and rindex()
4433 		 * both return index 0 (regardless of any position argument).
4434 		 */
4435 		if (sublen == 0 && len == 0) {
4436 			if (subr == DIF_SUBR_STRSTR)
4437 				regs[rd] = (uintptr_t)addr;
4438 			else
4439 				regs[rd] = 0;
4440 			break;
4441 		}
4442 
4443 		if (subr != DIF_SUBR_STRSTR) {
4444 			if (subr == DIF_SUBR_RINDEX) {
4445 				limit = orig - 1;
4446 				addr += len;
4447 				inc = -1;
4448 			}
4449 
4450 			/*
4451 			 * Both index() and rindex() take an optional position
4452 			 * argument that denotes the starting position.
4453 			 */
4454 			if (nargs == 3) {
4455 				int64_t pos = (int64_t)tupregs[2].dttk_value;
4456 
4457 				/*
4458 				 * If the position argument to index() is
4459 				 * negative, Perl implicitly clamps it at
4460 				 * zero.  This semantic is a little surprising
4461 				 * given the special meaning of negative
4462 				 * positions to similar Perl functions like
4463 				 * substr(), but it appears to reflect a
4464 				 * notion that index() can start from a
4465 				 * negative index and increment its way up to
4466 				 * the string.  Given this notion, Perl's
4467 				 * rindex() is at least self-consistent in
4468 				 * that it implicitly clamps positions greater
4469 				 * than the string length to be the string
4470 				 * length.  Where Perl completely loses
4471 				 * coherence, however, is when the specified
4472 				 * substring is the empty string ("").  In
4473 				 * this case, even if the position is
4474 				 * negative, rindex() returns 0 -- and even if
4475 				 * the position is greater than the length,
4476 				 * index() returns the string length.  These
4477 				 * semantics violate the notion that index()
4478 				 * should never return a value less than the
4479 				 * specified position and that rindex() should
4480 				 * never return a value greater than the
4481 				 * specified position.  (One assumes that
4482 				 * these semantics are artifacts of Perl's
4483 				 * implementation and not the results of
4484 				 * deliberate design -- it beggars belief that
4485 				 * even Larry Wall could desire such oddness.)
4486 				 * While in the abstract one would wish for
4487 				 * consistent position semantics across
4488 				 * substr(), index() and rindex() -- or at the
4489 				 * very least self-consistent position
4490 				 * semantics for index() and rindex() -- we
4491 				 * instead opt to keep with the extant Perl
4492 				 * semantics, in all their broken glory.  (Do
4493 				 * we have more desire to maintain Perl's
4494 				 * semantics than Perl does?  Probably.)
4495 				 */
4496 				if (subr == DIF_SUBR_RINDEX) {
4497 					if (pos < 0) {
4498 						if (sublen == 0)
4499 							regs[rd] = 0;
4500 						break;
4501 					}
4502 
4503 					if (pos > len)
4504 						pos = len;
4505 				} else {
4506 					if (pos < 0)
4507 						pos = 0;
4508 
4509 					if (pos >= len) {
4510 						if (sublen == 0)
4511 							regs[rd] = len;
4512 						break;
4513 					}
4514 				}
4515 
4516 				addr = orig + pos;
4517 			}
4518 		}
4519 
4520 		for (regs[rd] = notfound; addr != limit; addr += inc) {
4521 			if (dtrace_strncmp(addr, substr, sublen) == 0) {
4522 				if (subr != DIF_SUBR_STRSTR) {
4523 					/*
4524 					 * As D index() and rindex() are
4525 					 * modeled on Perl (and not on awk),
4526 					 * we return a zero-based (and not a
4527 					 * one-based) index.  (For you Perl
4528 					 * weenies: no, we're not going to add
4529 					 * $[ -- and shouldn't you be at a con
4530 					 * or something?)
4531 					 */
4532 					regs[rd] = (uintptr_t)(addr - orig);
4533 					break;
4534 				}
4535 
4536 				ASSERT(subr == DIF_SUBR_STRSTR);
4537 				regs[rd] = (uintptr_t)addr;
4538 				break;
4539 			}
4540 		}
4541 
4542 		break;
4543 	}
4544 
4545 	case DIF_SUBR_STRTOK: {
4546 		uintptr_t addr = tupregs[0].dttk_value;
4547 		uintptr_t tokaddr = tupregs[1].dttk_value;
4548 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4549 		uintptr_t limit, toklimit;
4550 		size_t clim;
4551 		uint8_t c, tokmap[32];	 /* 256 / 8 */
4552 		char *dest = (char *)mstate->dtms_scratch_ptr;
4553 		int i;
4554 
4555 		/*
4556 		 * Check both the token buffer and (later) the input buffer,
4557 		 * since both could be non-scratch addresses.
4558 		 */
4559 		if (!dtrace_strcanload(tokaddr, size, &clim, mstate, vstate)) {
4560 			regs[rd] = NULL;
4561 			break;
4562 		}
4563 		toklimit = tokaddr + clim;
4564 
4565 		if (!DTRACE_INSCRATCH(mstate, size)) {
4566 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4567 			regs[rd] = NULL;
4568 			break;
4569 		}
4570 
4571 		if (addr == NULL) {
4572 			/*
4573 			 * If the address specified is NULL, we use our saved
4574 			 * strtok pointer from the mstate.  Note that this
4575 			 * means that the saved strtok pointer is _only_
4576 			 * valid within multiple enablings of the same probe --
4577 			 * it behaves like an implicit clause-local variable.
4578 			 */
4579 			addr = mstate->dtms_strtok;
4580 			limit = mstate->dtms_strtok_limit;
4581 		} else {
4582 			/*
4583 			 * If the user-specified address is non-NULL we must
4584 			 * access check it.  This is the only time we have
4585 			 * a chance to do so, since this address may reside
4586 			 * in the string table of this clause-- future calls
4587 			 * (when we fetch addr from mstate->dtms_strtok)
4588 			 * would fail this access check.
4589 			 */
4590 			if (!dtrace_strcanload(addr, size, &clim, mstate,
4591 			    vstate)) {
4592 				regs[rd] = NULL;
4593 				break;
4594 			}
4595 			limit = addr + clim;
4596 		}
4597 
4598 		/*
4599 		 * First, zero the token map, and then process the token
4600 		 * string -- setting a bit in the map for every character
4601 		 * found in the token string.
4602 		 */
4603 		for (i = 0; i < sizeof (tokmap); i++)
4604 			tokmap[i] = 0;
4605 
4606 		for (; tokaddr < toklimit; tokaddr++) {
4607 			if ((c = dtrace_load8(tokaddr)) == '\0')
4608 				break;
4609 
4610 			ASSERT((c >> 3) < sizeof (tokmap));
4611 			tokmap[c >> 3] |= (1 << (c & 0x7));
4612 		}
4613 
4614 		for (; addr < limit; addr++) {
4615 			/*
4616 			 * We're looking for a character that is _not_
4617 			 * contained in the token string.
4618 			 */
4619 			if ((c = dtrace_load8(addr)) == '\0')
4620 				break;
4621 
4622 			if (!(tokmap[c >> 3] & (1 << (c & 0x7))))
4623 				break;
4624 		}
4625 
4626 		if (c == '\0') {
4627 			/*
4628 			 * We reached the end of the string without finding
4629 			 * any character that was not in the token string.
4630 			 * We return NULL in this case, and we set the saved
4631 			 * address to NULL as well.
4632 			 */
4633 			regs[rd] = NULL;
4634 			mstate->dtms_strtok = NULL;
4635 			mstate->dtms_strtok_limit = NULL;
4636 			break;
4637 		}
4638 
4639 		/*
4640 		 * From here on, we're copying into the destination string.
4641 		 */
4642 		for (i = 0; addr < limit && i < size - 1; addr++) {
4643 			if ((c = dtrace_load8(addr)) == '\0')
4644 				break;
4645 
4646 			if (tokmap[c >> 3] & (1 << (c & 0x7)))
4647 				break;
4648 
4649 			ASSERT(i < size);
4650 			dest[i++] = c;
4651 		}
4652 
4653 		ASSERT(i < size);
4654 		dest[i] = '\0';
4655 		regs[rd] = (uintptr_t)dest;
4656 		mstate->dtms_scratch_ptr += size;
4657 		mstate->dtms_strtok = addr;
4658 		mstate->dtms_strtok_limit = limit;
4659 		break;
4660 	}
4661 
4662 	case DIF_SUBR_SUBSTR: {
4663 		uintptr_t s = tupregs[0].dttk_value;
4664 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4665 		char *d = (char *)mstate->dtms_scratch_ptr;
4666 		int64_t index = (int64_t)tupregs[1].dttk_value;
4667 		int64_t remaining = (int64_t)tupregs[2].dttk_value;
4668 		size_t len = dtrace_strlen((char *)s, size);
4669 		int64_t i;
4670 
4671 		if (!dtrace_canload(s, len + 1, mstate, vstate)) {
4672 			regs[rd] = NULL;
4673 			break;
4674 		}
4675 
4676 		if (!DTRACE_INSCRATCH(mstate, size)) {
4677 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4678 			regs[rd] = NULL;
4679 			break;
4680 		}
4681 
4682 		if (nargs <= 2)
4683 			remaining = (int64_t)size;
4684 
4685 		if (index < 0) {
4686 			index += len;
4687 
4688 			if (index < 0 && index + remaining > 0) {
4689 				remaining += index;
4690 				index = 0;
4691 			}
4692 		}
4693 
4694 		if (index >= len || index < 0) {
4695 			remaining = 0;
4696 		} else if (remaining < 0) {
4697 			remaining += len - index;
4698 		} else if (index + remaining > size) {
4699 			remaining = size - index;
4700 		}
4701 
4702 		for (i = 0; i < remaining; i++) {
4703 			if ((d[i] = dtrace_load8(s + index + i)) == '\0')
4704 				break;
4705 		}
4706 
4707 		d[i] = '\0';
4708 
4709 		mstate->dtms_scratch_ptr += size;
4710 		regs[rd] = (uintptr_t)d;
4711 		break;
4712 	}
4713 
4714 	case DIF_SUBR_JSON: {
4715 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4716 		uintptr_t json = tupregs[0].dttk_value;
4717 		size_t jsonlen = dtrace_strlen((char *)json, size);
4718 		uintptr_t elem = tupregs[1].dttk_value;
4719 		size_t elemlen = dtrace_strlen((char *)elem, size);
4720 
4721 		char *dest = (char *)mstate->dtms_scratch_ptr;
4722 		char *elemlist = (char *)mstate->dtms_scratch_ptr + jsonlen + 1;
4723 		char *ee = elemlist;
4724 		int nelems = 1;
4725 		uintptr_t cur;
4726 
4727 		if (!dtrace_canload(json, jsonlen + 1, mstate, vstate) ||
4728 		    !dtrace_canload(elem, elemlen + 1, mstate, vstate)) {
4729 			regs[rd] = NULL;
4730 			break;
4731 		}
4732 
4733 		if (!DTRACE_INSCRATCH(mstate, jsonlen + 1 + elemlen + 1)) {
4734 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4735 			regs[rd] = NULL;
4736 			break;
4737 		}
4738 
4739 		/*
4740 		 * Read the element selector and split it up into a packed list
4741 		 * of strings.
4742 		 */
4743 		for (cur = elem; cur < elem + elemlen; cur++) {
4744 			char cc = dtrace_load8(cur);
4745 
4746 			if (cur == elem && cc == '[') {
4747 				/*
4748 				 * If the first element selector key is
4749 				 * actually an array index then ignore the
4750 				 * bracket.
4751 				 */
4752 				continue;
4753 			}
4754 
4755 			if (cc == ']')
4756 				continue;
4757 
4758 			if (cc == '.' || cc == '[') {
4759 				nelems++;
4760 				cc = '\0';
4761 			}
4762 
4763 			*ee++ = cc;
4764 		}
4765 		*ee++ = '\0';
4766 
4767 		if ((regs[rd] = (uintptr_t)dtrace_json(size, json, elemlist,
4768 		    nelems, dest)) != NULL)
4769 			mstate->dtms_scratch_ptr += jsonlen + 1;
4770 		break;
4771 	}
4772 
4773 	case DIF_SUBR_TOUPPER:
4774 	case DIF_SUBR_TOLOWER: {
4775 		uintptr_t s = tupregs[0].dttk_value;
4776 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4777 		char *dest = (char *)mstate->dtms_scratch_ptr, c;
4778 		size_t len = dtrace_strlen((char *)s, size);
4779 		char lower, upper, convert;
4780 		int64_t i;
4781 
4782 		if (subr == DIF_SUBR_TOUPPER) {
4783 			lower = 'a';
4784 			upper = 'z';
4785 			convert = 'A';
4786 		} else {
4787 			lower = 'A';
4788 			upper = 'Z';
4789 			convert = 'a';
4790 		}
4791 
4792 		if (!dtrace_canload(s, len + 1, mstate, vstate)) {
4793 			regs[rd] = NULL;
4794 			break;
4795 		}
4796 
4797 		if (!DTRACE_INSCRATCH(mstate, size)) {
4798 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4799 			regs[rd] = NULL;
4800 			break;
4801 		}
4802 
4803 		for (i = 0; i < size - 1; i++) {
4804 			if ((c = dtrace_load8(s + i)) == '\0')
4805 				break;
4806 
4807 			if (c >= lower && c <= upper)
4808 				c = convert + (c - lower);
4809 
4810 			dest[i] = c;
4811 		}
4812 
4813 		ASSERT(i < size);
4814 		dest[i] = '\0';
4815 		regs[rd] = (uintptr_t)dest;
4816 		mstate->dtms_scratch_ptr += size;
4817 		break;
4818 	}
4819 
4820 case DIF_SUBR_GETMAJOR:
4821 #ifdef _LP64
4822 		regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR64) & MAXMAJ64;
4823 #else
4824 		regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR) & MAXMAJ;
4825 #endif
4826 		break;
4827 
4828 	case DIF_SUBR_GETMINOR:
4829 #ifdef _LP64
4830 		regs[rd] = tupregs[0].dttk_value & MAXMIN64;
4831 #else
4832 		regs[rd] = tupregs[0].dttk_value & MAXMIN;
4833 #endif
4834 		break;
4835 
4836 	case DIF_SUBR_DDI_PATHNAME: {
4837 		/*
4838 		 * This one is a galactic mess.  We are going to roughly
4839 		 * emulate ddi_pathname(), but it's made more complicated
4840 		 * by the fact that we (a) want to include the minor name and
4841 		 * (b) must proceed iteratively instead of recursively.
4842 		 */
4843 		uintptr_t dest = mstate->dtms_scratch_ptr;
4844 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4845 		char *start = (char *)dest, *end = start + size - 1;
4846 		uintptr_t daddr = tupregs[0].dttk_value;
4847 		int64_t minor = (int64_t)tupregs[1].dttk_value;
4848 		char *s;
4849 		int i, len, depth = 0;
4850 
4851 		/*
4852 		 * Due to all the pointer jumping we do and context we must
4853 		 * rely upon, we just mandate that the user must have kernel
4854 		 * read privileges to use this routine.
4855 		 */
4856 		if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) == 0) {
4857 			*flags |= CPU_DTRACE_KPRIV;
4858 			*illval = daddr;
4859 			regs[rd] = NULL;
4860 		}
4861 
4862 		if (!DTRACE_INSCRATCH(mstate, size)) {
4863 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4864 			regs[rd] = NULL;
4865 			break;
4866 		}
4867 
4868 		*end = '\0';
4869 
4870 		/*
4871 		 * We want to have a name for the minor.  In order to do this,
4872 		 * we need to walk the minor list from the devinfo.  We want
4873 		 * to be sure that we don't infinitely walk a circular list,
4874 		 * so we check for circularity by sending a scout pointer
4875 		 * ahead two elements for every element that we iterate over;
4876 		 * if the list is circular, these will ultimately point to the
4877 		 * same element.  You may recognize this little trick as the
4878 		 * answer to a stupid interview question -- one that always
4879 		 * seems to be asked by those who had to have it laboriously
4880 		 * explained to them, and who can't even concisely describe
4881 		 * the conditions under which one would be forced to resort to
4882 		 * this technique.  Needless to say, those conditions are
4883 		 * found here -- and probably only here.  Is this the only use
4884 		 * of this infamous trick in shipping, production code?  If it
4885 		 * isn't, it probably should be...
4886 		 */
4887 		if (minor != -1) {
4888 			uintptr_t maddr = dtrace_loadptr(daddr +
4889 			    offsetof(struct dev_info, devi_minor));
4890 
4891 			uintptr_t next = offsetof(struct ddi_minor_data, next);
4892 			uintptr_t name = offsetof(struct ddi_minor_data,
4893 			    d_minor) + offsetof(struct ddi_minor, name);
4894 			uintptr_t dev = offsetof(struct ddi_minor_data,
4895 			    d_minor) + offsetof(struct ddi_minor, dev);
4896 			uintptr_t scout;
4897 
4898 			if (maddr != NULL)
4899 				scout = dtrace_loadptr(maddr + next);
4900 
4901 			while (maddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
4902 				uint64_t m;
4903 #ifdef _LP64
4904 				m = dtrace_load64(maddr + dev) & MAXMIN64;
4905 #else
4906 				m = dtrace_load32(maddr + dev) & MAXMIN;
4907 #endif
4908 				if (m != minor) {
4909 					maddr = dtrace_loadptr(maddr + next);
4910 
4911 					if (scout == NULL)
4912 						continue;
4913 
4914 					scout = dtrace_loadptr(scout + next);
4915 
4916 					if (scout == NULL)
4917 						continue;
4918 
4919 					scout = dtrace_loadptr(scout + next);
4920 
4921 					if (scout == NULL)
4922 						continue;
4923 
4924 					if (scout == maddr) {
4925 						*flags |= CPU_DTRACE_ILLOP;
4926 						break;
4927 					}
4928 
4929 					continue;
4930 				}
4931 
4932 				/*
4933 				 * We have the minor data.  Now we need to
4934 				 * copy the minor's name into the end of the
4935 				 * pathname.
4936 				 */
4937 				s = (char *)dtrace_loadptr(maddr + name);
4938 				len = dtrace_strlen(s, size);
4939 
4940 				if (*flags & CPU_DTRACE_FAULT)
4941 					break;
4942 
4943 				if (len != 0) {
4944 					if ((end -= (len + 1)) < start)
4945 						break;
4946 
4947 					*end = ':';
4948 				}
4949 
4950 				for (i = 1; i <= len; i++)
4951 					end[i] = dtrace_load8((uintptr_t)s++);
4952 				break;
4953 			}
4954 		}
4955 
4956 		while (daddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
4957 			ddi_node_state_t devi_state;
4958 
4959 			devi_state = dtrace_load32(daddr +
4960 			    offsetof(struct dev_info, devi_node_state));
4961 
4962 			if (*flags & CPU_DTRACE_FAULT)
4963 				break;
4964 
4965 			if (devi_state >= DS_INITIALIZED) {
4966 				s = (char *)dtrace_loadptr(daddr +
4967 				    offsetof(struct dev_info, devi_addr));
4968 				len = dtrace_strlen(s, size);
4969 
4970 				if (*flags & CPU_DTRACE_FAULT)
4971 					break;
4972 
4973 				if (len != 0) {
4974 					if ((end -= (len + 1)) < start)
4975 						break;
4976 
4977 					*end = '@';
4978 				}
4979 
4980 				for (i = 1; i <= len; i++)
4981 					end[i] = dtrace_load8((uintptr_t)s++);
4982 			}
4983 
4984 			/*
4985 			 * Now for the node name...
4986 			 */
4987 			s = (char *)dtrace_loadptr(daddr +
4988 			    offsetof(struct dev_info, devi_node_name));
4989 
4990 			daddr = dtrace_loadptr(daddr +
4991 			    offsetof(struct dev_info, devi_parent));
4992 
4993 			/*
4994 			 * If our parent is NULL (that is, if we're the root
4995 			 * node), we're going to use the special path
4996 			 * "devices".
4997 			 */
4998 			if (daddr == NULL)
4999 				s = "devices";
5000 
5001 			len = dtrace_strlen(s, size);
5002 			if (*flags & CPU_DTRACE_FAULT)
5003 				break;
5004 
5005 			if ((end -= (len + 1)) < start)
5006 				break;
5007 
5008 			for (i = 1; i <= len; i++)
5009 				end[i] = dtrace_load8((uintptr_t)s++);
5010 			*end = '/';
5011 
5012 			if (depth++ > dtrace_devdepth_max) {
5013 				*flags |= CPU_DTRACE_ILLOP;
5014 				break;
5015 			}
5016 		}
5017 
5018 		if (end < start)
5019 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5020 
5021 		if (daddr == NULL) {
5022 			regs[rd] = (uintptr_t)end;
5023 			mstate->dtms_scratch_ptr += size;
5024 		}
5025 
5026 		break;
5027 	}
5028 
5029 	case DIF_SUBR_STRJOIN: {
5030 		char *d = (char *)mstate->dtms_scratch_ptr;
5031 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
5032 		uintptr_t s1 = tupregs[0].dttk_value;
5033 		uintptr_t s2 = tupregs[1].dttk_value;
5034 		int i = 0, j = 0;
5035 		size_t lim1, lim2;
5036 		char c;
5037 
5038 		if (!dtrace_strcanload(s1, size, &lim1, mstate, vstate) ||
5039 		    !dtrace_strcanload(s2, size, &lim2, mstate, vstate)) {
5040 			regs[rd] = NULL;
5041 			break;
5042 		}
5043 
5044 		if (!DTRACE_INSCRATCH(mstate, size)) {
5045 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5046 			regs[rd] = NULL;
5047 			break;
5048 		}
5049 
5050 		for (;;) {
5051 			if (i >= size) {
5052 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5053 				regs[rd] = NULL;
5054 				break;
5055 			}
5056 			c = (i >= lim1) ? '\0' : dtrace_load8(s1++);
5057 			if ((d[i++] = c) == '\0') {
5058 				i--;
5059 				break;
5060 			}
5061 		}
5062 
5063 		for (;;) {
5064 			if (i >= size) {
5065 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5066 				regs[rd] = NULL;
5067 				break;
5068 			}
5069 
5070 			c = (j++ >= lim2) ? '\0' : dtrace_load8(s2++);
5071 			if ((d[i++] = c) == '\0')
5072 				break;
5073 		}
5074 
5075 		if (i < size) {
5076 			mstate->dtms_scratch_ptr += i;
5077 			regs[rd] = (uintptr_t)d;
5078 		}
5079 
5080 		break;
5081 	}
5082 
5083 	case DIF_SUBR_STRTOLL: {
5084 		uintptr_t s = tupregs[0].dttk_value;
5085 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
5086 		size_t lim;
5087 		int base = 10;
5088 
5089 		if (nargs > 1) {
5090 			if ((base = tupregs[1].dttk_value) <= 1 ||
5091 			    base > ('z' - 'a' + 1) + ('9' - '0' + 1)) {
5092 				*flags |= CPU_DTRACE_ILLOP;
5093 				break;
5094 			}
5095 		}
5096 
5097 		if (!dtrace_strcanload(s, size, &lim, mstate, vstate)) {
5098 			regs[rd] = INT64_MIN;
5099 			break;
5100 		}
5101 
5102 		regs[rd] = dtrace_strtoll((char *)s, base, lim);
5103 		break;
5104 	}
5105 
5106 	case DIF_SUBR_LLTOSTR: {
5107 		int64_t i = (int64_t)tupregs[0].dttk_value;
5108 		uint64_t val, digit;
5109 		uint64_t size = 65;	/* enough room for 2^64 in binary */
5110 		char *end = (char *)mstate->dtms_scratch_ptr + size - 1;
5111 		int base = 10;
5112 
5113 		if (nargs > 1) {
5114 			if ((base = tupregs[1].dttk_value) <= 1 ||
5115 			    base > ('z' - 'a' + 1) + ('9' - '0' + 1)) {
5116 				*flags |= CPU_DTRACE_ILLOP;
5117 				break;
5118 			}
5119 		}
5120 
5121 		val = (base == 10 && i < 0) ? i * -1 : i;
5122 
5123 		if (!DTRACE_INSCRATCH(mstate, size)) {
5124 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5125 			regs[rd] = NULL;
5126 			break;
5127 		}
5128 
5129 		for (*end-- = '\0'; val; val /= base) {
5130 			if ((digit = val % base) <= '9' - '0') {
5131 				*end-- = '0' + digit;
5132 			} else {
5133 				*end-- = 'a' + (digit - ('9' - '0') - 1);
5134 			}
5135 		}
5136 
5137 		if (i == 0 && base == 16)
5138 			*end-- = '0';
5139 
5140 		if (base == 16)
5141 			*end-- = 'x';
5142 
5143 		if (i == 0 || base == 8 || base == 16)
5144 			*end-- = '0';
5145 
5146 		if (i < 0 && base == 10)
5147 			*end-- = '-';
5148 
5149 		regs[rd] = (uintptr_t)end + 1;
5150 		mstate->dtms_scratch_ptr += size;
5151 		break;
5152 	}
5153 
5154 	case DIF_SUBR_HTONS:
5155 	case DIF_SUBR_NTOHS:
5156 #ifdef _BIG_ENDIAN
5157 		regs[rd] = (uint16_t)tupregs[0].dttk_value;
5158 #else
5159 		regs[rd] = DT_BSWAP_16((uint16_t)tupregs[0].dttk_value);
5160 #endif
5161 		break;
5162 
5163 
5164 	case DIF_SUBR_HTONL:
5165 	case DIF_SUBR_NTOHL:
5166 #ifdef _BIG_ENDIAN
5167 		regs[rd] = (uint32_t)tupregs[0].dttk_value;
5168 #else
5169 		regs[rd] = DT_BSWAP_32((uint32_t)tupregs[0].dttk_value);
5170 #endif
5171 		break;
5172 
5173 
5174 	case DIF_SUBR_HTONLL:
5175 	case DIF_SUBR_NTOHLL:
5176 #ifdef _BIG_ENDIAN
5177 		regs[rd] = (uint64_t)tupregs[0].dttk_value;
5178 #else
5179 		regs[rd] = DT_BSWAP_64((uint64_t)tupregs[0].dttk_value);
5180 #endif
5181 		break;
5182 
5183 
5184 	case DIF_SUBR_DIRNAME:
5185 	case DIF_SUBR_BASENAME: {
5186 		char *dest = (char *)mstate->dtms_scratch_ptr;
5187 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
5188 		uintptr_t src = tupregs[0].dttk_value;
5189 		int i, j, len = dtrace_strlen((char *)src, size);
5190 		int lastbase = -1, firstbase = -1, lastdir = -1;
5191 		int start, end;
5192 
5193 		if (!dtrace_canload(src, len + 1, mstate, vstate)) {
5194 			regs[rd] = NULL;
5195 			break;
5196 		}
5197 
5198 		if (!DTRACE_INSCRATCH(mstate, size)) {
5199 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5200 			regs[rd] = NULL;
5201 			break;
5202 		}
5203 
5204 		/*
5205 		 * The basename and dirname for a zero-length string is
5206 		 * defined to be "."
5207 		 */
5208 		if (len == 0) {
5209 			len = 1;
5210 			src = (uintptr_t)".";
5211 		}
5212 
5213 		/*
5214 		 * Start from the back of the string, moving back toward the
5215 		 * front until we see a character that isn't a slash.  That
5216 		 * character is the last character in the basename.
5217 		 */
5218 		for (i = len - 1; i >= 0; i--) {
5219 			if (dtrace_load8(src + i) != '/')
5220 				break;
5221 		}
5222 
5223 		if (i >= 0)
5224 			lastbase = i;
5225 
5226 		/*
5227 		 * Starting from the last character in the basename, move
5228 		 * towards the front until we find a slash.  The character
5229 		 * that we processed immediately before that is the first
5230 		 * character in the basename.
5231 		 */
5232 		for (; i >= 0; i--) {
5233 			if (dtrace_load8(src + i) == '/')
5234 				break;
5235 		}
5236 
5237 		if (i >= 0)
5238 			firstbase = i + 1;
5239 
5240 		/*
5241 		 * Now keep going until we find a non-slash character.  That
5242 		 * character is the last character in the dirname.
5243 		 */
5244 		for (; i >= 0; i--) {
5245 			if (dtrace_load8(src + i) != '/')
5246 				break;
5247 		}
5248 
5249 		if (i >= 0)
5250 			lastdir = i;
5251 
5252 		ASSERT(!(lastbase == -1 && firstbase != -1));
5253 		ASSERT(!(firstbase == -1 && lastdir != -1));
5254 
5255 		if (lastbase == -1) {
5256 			/*
5257 			 * We didn't find a non-slash character.  We know that
5258 			 * the length is non-zero, so the whole string must be
5259 			 * slashes.  In either the dirname or the basename
5260 			 * case, we return '/'.
5261 			 */
5262 			ASSERT(firstbase == -1);
5263 			firstbase = lastbase = lastdir = 0;
5264 		}
5265 
5266 		if (firstbase == -1) {
5267 			/*
5268 			 * The entire string consists only of a basename
5269 			 * component.  If we're looking for dirname, we need
5270 			 * to change our string to be just "."; if we're
5271 			 * looking for a basename, we'll just set the first
5272 			 * character of the basename to be 0.
5273 			 */
5274 			if (subr == DIF_SUBR_DIRNAME) {
5275 				ASSERT(lastdir == -1);
5276 				src = (uintptr_t)".";
5277 				lastdir = 0;
5278 			} else {
5279 				firstbase = 0;
5280 			}
5281 		}
5282 
5283 		if (subr == DIF_SUBR_DIRNAME) {
5284 			if (lastdir == -1) {
5285 				/*
5286 				 * We know that we have a slash in the name --
5287 				 * or lastdir would be set to 0, above.  And
5288 				 * because lastdir is -1, we know that this
5289 				 * slash must be the first character.  (That
5290 				 * is, the full string must be of the form
5291 				 * "/basename".)  In this case, the last
5292 				 * character of the directory name is 0.
5293 				 */
5294 				lastdir = 0;
5295 			}
5296 
5297 			start = 0;
5298 			end = lastdir;
5299 		} else {
5300 			ASSERT(subr == DIF_SUBR_BASENAME);
5301 			ASSERT(firstbase != -1 && lastbase != -1);
5302 			start = firstbase;
5303 			end = lastbase;
5304 		}
5305 
5306 		for (i = start, j = 0; i <= end && j < size - 1; i++, j++)
5307 			dest[j] = dtrace_load8(src + i);
5308 
5309 		dest[j] = '\0';
5310 		regs[rd] = (uintptr_t)dest;
5311 		mstate->dtms_scratch_ptr += size;
5312 		break;
5313 	}
5314 
5315 	case DIF_SUBR_GETF: {
5316 		uintptr_t fd = tupregs[0].dttk_value;
5317 		uf_info_t *finfo = &curthread->t_procp->p_user.u_finfo;
5318 		file_t *fp;
5319 
5320 		if (!dtrace_priv_proc(state, mstate)) {
5321 			regs[rd] = NULL;
5322 			break;
5323 		}
5324 
5325 		/*
5326 		 * This is safe because fi_nfiles only increases, and the
5327 		 * fi_list array is not freed when the array size doubles.
5328 		 * (See the comment in flist_grow() for details on the
5329 		 * management of the u_finfo structure.)
5330 		 */
5331 		fp = fd < finfo->fi_nfiles ? finfo->fi_list[fd].uf_file : NULL;
5332 
5333 		mstate->dtms_getf = fp;
5334 		regs[rd] = (uintptr_t)fp;
5335 		break;
5336 	}
5337 
5338 	case DIF_SUBR_CLEANPATH: {
5339 		char *dest = (char *)mstate->dtms_scratch_ptr, c;
5340 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
5341 		uintptr_t src = tupregs[0].dttk_value;
5342 		size_t lim;
5343 		int i = 0, j = 0;
5344 		zone_t *z;
5345 
5346 		if (!dtrace_strcanload(src, size, &lim, mstate, vstate)) {
5347 			regs[rd] = NULL;
5348 			break;
5349 		}
5350 
5351 		if (!DTRACE_INSCRATCH(mstate, size)) {
5352 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5353 			regs[rd] = NULL;
5354 			break;
5355 		}
5356 
5357 		/*
5358 		 * Move forward, loading each character.
5359 		 */
5360 		do {
5361 			c = (i >= lim) ? '\0' : dtrace_load8(src + i++);
5362 next:
5363 			if (j + 5 >= size)	/* 5 = strlen("/..c\0") */
5364 				break;
5365 
5366 			if (c != '/') {
5367 				dest[j++] = c;
5368 				continue;
5369 			}
5370 
5371 			c = (i >= lim) ? '\0' : dtrace_load8(src + i++);
5372 
5373 			if (c == '/') {
5374 				/*
5375 				 * We have two slashes -- we can just advance
5376 				 * to the next character.
5377 				 */
5378 				goto next;
5379 			}
5380 
5381 			if (c != '.') {
5382 				/*
5383 				 * This is not "." and it's not ".." -- we can
5384 				 * just store the "/" and this character and
5385 				 * drive on.
5386 				 */
5387 				dest[j++] = '/';
5388 				dest[j++] = c;
5389 				continue;
5390 			}
5391 
5392 			c = (i >= lim) ? '\0' : dtrace_load8(src + i++);
5393 
5394 			if (c == '/') {
5395 				/*
5396 				 * This is a "/./" component.  We're not going
5397 				 * to store anything in the destination buffer;
5398 				 * we're just going to go to the next component.
5399 				 */
5400 				goto next;
5401 			}
5402 
5403 			if (c != '.') {
5404 				/*
5405 				 * This is not ".." -- we can just store the
5406 				 * "/." and this character and continue
5407 				 * processing.
5408 				 */
5409 				dest[j++] = '/';
5410 				dest[j++] = '.';
5411 				dest[j++] = c;
5412 				continue;
5413 			}
5414 
5415 			c = (i >= lim) ? '\0' : dtrace_load8(src + i++);
5416 
5417 			if (c != '/' && c != '\0') {
5418 				/*
5419 				 * This is not ".." -- it's "..[mumble]".
5420 				 * We'll store the "/.." and this character
5421 				 * and continue processing.
5422 				 */
5423 				dest[j++] = '/';
5424 				dest[j++] = '.';
5425 				dest[j++] = '.';
5426 				dest[j++] = c;
5427 				continue;
5428 			}
5429 
5430 			/*
5431 			 * This is "/../" or "/..\0".  We need to back up
5432 			 * our destination pointer until we find a "/".
5433 			 */
5434 			i--;
5435 			while (j != 0 && dest[--j] != '/')
5436 				continue;
5437 
5438 			if (c == '\0')
5439 				dest[++j] = '/';
5440 		} while (c != '\0');
5441 
5442 		dest[j] = '\0';
5443 
5444 		if (mstate->dtms_getf != NULL &&
5445 		    !(mstate->dtms_access & DTRACE_ACCESS_KERNEL) &&
5446 		    (z = state->dts_cred.dcr_cred->cr_zone) != kcred->cr_zone) {
5447 			/*
5448 			 * If we've done a getf() as a part of this ECB and we
5449 			 * don't have kernel access (and we're not in the global
5450 			 * zone), check if the path we cleaned up begins with
5451 			 * the zone's root path, and trim it off if so.  Note
5452 			 * that this is an output cleanliness issue, not a
5453 			 * security issue: knowing one's zone root path does
5454 			 * not enable privilege escalation.
5455 			 */
5456 			if (strstr(dest, z->zone_rootpath) == dest)
5457 				dest += strlen(z->zone_rootpath) - 1;
5458 		}
5459 
5460 		regs[rd] = (uintptr_t)dest;
5461 		mstate->dtms_scratch_ptr += size;
5462 		break;
5463 	}
5464 
5465 	case DIF_SUBR_INET_NTOA:
5466 	case DIF_SUBR_INET_NTOA6:
5467 	case DIF_SUBR_INET_NTOP: {
5468 		size_t size;
5469 		int af, argi, i;
5470 		char *base, *end;
5471 
5472 		if (subr == DIF_SUBR_INET_NTOP) {
5473 			af = (int)tupregs[0].dttk_value;
5474 			argi = 1;
5475 		} else {
5476 			af = subr == DIF_SUBR_INET_NTOA ? AF_INET: AF_INET6;
5477 			argi = 0;
5478 		}
5479 
5480 		if (af == AF_INET) {
5481 			ipaddr_t ip4;
5482 			uint8_t *ptr8, val;
5483 
5484 			if (!dtrace_canload(tupregs[argi].dttk_value,
5485 			    sizeof (ipaddr_t), mstate, vstate)) {
5486 				regs[rd] = NULL;
5487 				break;
5488 			}
5489 
5490 			/*
5491 			 * Safely load the IPv4 address.
5492 			 */
5493 			ip4 = dtrace_load32(tupregs[argi].dttk_value);
5494 
5495 			/*
5496 			 * Check an IPv4 string will fit in scratch.
5497 			 */
5498 			size = INET_ADDRSTRLEN;
5499 			if (!DTRACE_INSCRATCH(mstate, size)) {
5500 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5501 				regs[rd] = NULL;
5502 				break;
5503 			}
5504 			base = (char *)mstate->dtms_scratch_ptr;
5505 			end = (char *)mstate->dtms_scratch_ptr + size - 1;
5506 
5507 			/*
5508 			 * Stringify as a dotted decimal quad.
5509 			 */
5510 			*end-- = '\0';
5511 			ptr8 = (uint8_t *)&ip4;
5512 			for (i = 3; i >= 0; i--) {
5513 				val = ptr8[i];
5514 
5515 				if (val == 0) {
5516 					*end-- = '0';
5517 				} else {
5518 					for (; val; val /= 10) {
5519 						*end-- = '0' + (val % 10);
5520 					}
5521 				}
5522 
5523 				if (i > 0)
5524 					*end-- = '.';
5525 			}
5526 			ASSERT(end + 1 >= base);
5527 
5528 		} else if (af == AF_INET6) {
5529 			struct in6_addr ip6;
5530 			int firstzero, tryzero, numzero, v6end;
5531 			uint16_t val;
5532 			const char digits[] = "0123456789abcdef";
5533 
5534 			/*
5535 			 * Stringify using RFC 1884 convention 2 - 16 bit
5536 			 * hexadecimal values with a zero-run compression.
5537 			 * Lower case hexadecimal digits are used.
5538 			 * 	eg, fe80::214:4fff:fe0b:76c8.
5539 			 * The IPv4 embedded form is returned for inet_ntop,
5540 			 * just the IPv4 string is returned for inet_ntoa6.
5541 			 */
5542 
5543 			if (!dtrace_canload(tupregs[argi].dttk_value,
5544 			    sizeof (struct in6_addr), mstate, vstate)) {
5545 				regs[rd] = NULL;
5546 				break;
5547 			}
5548 
5549 			/*
5550 			 * Safely load the IPv6 address.
5551 			 */
5552 			dtrace_bcopy(
5553 			    (void *)(uintptr_t)tupregs[argi].dttk_value,
5554 			    (void *)(uintptr_t)&ip6, sizeof (struct in6_addr));
5555 
5556 			/*
5557 			 * Check an IPv6 string will fit in scratch.
5558 			 */
5559 			size = INET6_ADDRSTRLEN;
5560 			if (!DTRACE_INSCRATCH(mstate, size)) {
5561 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5562 				regs[rd] = NULL;
5563 				break;
5564 			}
5565 			base = (char *)mstate->dtms_scratch_ptr;
5566 			end = (char *)mstate->dtms_scratch_ptr + size - 1;
5567 			*end-- = '\0';
5568 
5569 			/*
5570 			 * Find the longest run of 16 bit zero values
5571 			 * for the single allowed zero compression - "::".
5572 			 */
5573 			firstzero = -1;
5574 			tryzero = -1;
5575 			numzero = 1;
5576 			for (i = 0; i < sizeof (struct in6_addr); i++) {
5577 				if (ip6._S6_un._S6_u8[i] == 0 &&
5578 				    tryzero == -1 && i % 2 == 0) {
5579 					tryzero = i;
5580 					continue;
5581 				}
5582 
5583 				if (tryzero != -1 &&
5584 				    (ip6._S6_un._S6_u8[i] != 0 ||
5585 				    i == sizeof (struct in6_addr) - 1)) {
5586 
5587 					if (i - tryzero <= numzero) {
5588 						tryzero = -1;
5589 						continue;
5590 					}
5591 
5592 					firstzero = tryzero;
5593 					numzero = i - i % 2 - tryzero;
5594 					tryzero = -1;
5595 
5596 					if (ip6._S6_un._S6_u8[i] == 0 &&
5597 					    i == sizeof (struct in6_addr) - 1)
5598 						numzero += 2;
5599 				}
5600 			}
5601 			ASSERT(firstzero + numzero <= sizeof (struct in6_addr));
5602 
5603 			/*
5604 			 * Check for an IPv4 embedded address.
5605 			 */
5606 			v6end = sizeof (struct in6_addr) - 2;
5607 			if (IN6_IS_ADDR_V4MAPPED(&ip6) ||
5608 			    IN6_IS_ADDR_V4COMPAT(&ip6)) {
5609 				for (i = sizeof (struct in6_addr) - 1;
5610 				    i >= DTRACE_V4MAPPED_OFFSET; i--) {
5611 					ASSERT(end >= base);
5612 
5613 					val = ip6._S6_un._S6_u8[i];
5614 
5615 					if (val == 0) {
5616 						*end-- = '0';
5617 					} else {
5618 						for (; val; val /= 10) {
5619 							*end-- = '0' + val % 10;
5620 						}
5621 					}
5622 
5623 					if (i > DTRACE_V4MAPPED_OFFSET)
5624 						*end-- = '.';
5625 				}
5626 
5627 				if (subr == DIF_SUBR_INET_NTOA6)
5628 					goto inetout;
5629 
5630 				/*
5631 				 * Set v6end to skip the IPv4 address that
5632 				 * we have already stringified.
5633 				 */
5634 				v6end = 10;
5635 			}
5636 
5637 			/*
5638 			 * Build the IPv6 string by working through the
5639 			 * address in reverse.
5640 			 */
5641 			for (i = v6end; i >= 0; i -= 2) {
5642 				ASSERT(end >= base);
5643 
5644 				if (i == firstzero + numzero - 2) {
5645 					*end-- = ':';
5646 					*end-- = ':';
5647 					i -= numzero - 2;
5648 					continue;
5649 				}
5650 
5651 				if (i < 14 && i != firstzero - 2)
5652 					*end-- = ':';
5653 
5654 				val = (ip6._S6_un._S6_u8[i] << 8) +
5655 				    ip6._S6_un._S6_u8[i + 1];
5656 
5657 				if (val == 0) {
5658 					*end-- = '0';
5659 				} else {
5660 					for (; val; val /= 16) {
5661 						*end-- = digits[val % 16];
5662 					}
5663 				}
5664 			}
5665 			ASSERT(end + 1 >= base);
5666 
5667 		} else {
5668 			/*
5669 			 * The user didn't use AH_INET or AH_INET6.
5670 			 */
5671 			DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
5672 			regs[rd] = NULL;
5673 			break;
5674 		}
5675 
5676 inetout:	regs[rd] = (uintptr_t)end + 1;
5677 		mstate->dtms_scratch_ptr += size;
5678 		break;
5679 	}
5680 
5681 	}
5682 }
5683 
5684 /*
5685  * Emulate the execution of DTrace IR instructions specified by the given
5686  * DIF object.  This function is deliberately void of assertions as all of
5687  * the necessary checks are handled by a call to dtrace_difo_validate().
5688  */
5689 static uint64_t
5690 dtrace_dif_emulate(dtrace_difo_t *difo, dtrace_mstate_t *mstate,
5691     dtrace_vstate_t *vstate, dtrace_state_t *state)
5692 {
5693 	const dif_instr_t *text = difo->dtdo_buf;
5694 	const uint_t textlen = difo->dtdo_len;
5695 	const char *strtab = difo->dtdo_strtab;
5696 	const uint64_t *inttab = difo->dtdo_inttab;
5697 
5698 	uint64_t rval = 0;
5699 	dtrace_statvar_t *svar;
5700 	dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
5701 	dtrace_difv_t *v;
5702 	volatile uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
5703 	volatile uintptr_t *illval = &cpu_core[CPU->cpu_id].cpuc_dtrace_illval;
5704 
5705 	dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
5706 	uint64_t regs[DIF_DIR_NREGS];
5707 	uint64_t *tmp;
5708 
5709 	uint8_t cc_n = 0, cc_z = 0, cc_v = 0, cc_c = 0;
5710 	int64_t cc_r;
5711 	uint_t pc = 0, id, opc;
5712 	uint8_t ttop = 0;
5713 	dif_instr_t instr;
5714 	uint_t r1, r2, rd;
5715 
5716 	/*
5717 	 * We stash the current DIF object into the machine state: we need it
5718 	 * for subsequent access checking.
5719 	 */
5720 	mstate->dtms_difo = difo;
5721 
5722 	regs[DIF_REG_R0] = 0; 		/* %r0 is fixed at zero */
5723 
5724 	while (pc < textlen && !(*flags & CPU_DTRACE_FAULT)) {
5725 		opc = pc;
5726 
5727 		instr = text[pc++];
5728 		r1 = DIF_INSTR_R1(instr);
5729 		r2 = DIF_INSTR_R2(instr);
5730 		rd = DIF_INSTR_RD(instr);
5731 
5732 		switch (DIF_INSTR_OP(instr)) {
5733 		case DIF_OP_OR:
5734 			regs[rd] = regs[r1] | regs[r2];
5735 			break;
5736 		case DIF_OP_XOR:
5737 			regs[rd] = regs[r1] ^ regs[r2];
5738 			break;
5739 		case DIF_OP_AND:
5740 			regs[rd] = regs[r1] & regs[r2];
5741 			break;
5742 		case DIF_OP_SLL:
5743 			regs[rd] = regs[r1] << regs[r2];
5744 			break;
5745 		case DIF_OP_SRL:
5746 			regs[rd] = regs[r1] >> regs[r2];
5747 			break;
5748 		case DIF_OP_SUB:
5749 			regs[rd] = regs[r1] - regs[r2];
5750 			break;
5751 		case DIF_OP_ADD:
5752 			regs[rd] = regs[r1] + regs[r2];
5753 			break;
5754 		case DIF_OP_MUL:
5755 			regs[rd] = regs[r1] * regs[r2];
5756 			break;
5757 		case DIF_OP_SDIV:
5758 			if (regs[r2] == 0) {
5759 				regs[rd] = 0;
5760 				*flags |= CPU_DTRACE_DIVZERO;
5761 			} else {
5762 				regs[rd] = (int64_t)regs[r1] /
5763 				    (int64_t)regs[r2];
5764 			}
5765 			break;
5766 
5767 		case DIF_OP_UDIV:
5768 			if (regs[r2] == 0) {
5769 				regs[rd] = 0;
5770 				*flags |= CPU_DTRACE_DIVZERO;
5771 			} else {
5772 				regs[rd] = regs[r1] / regs[r2];
5773 			}
5774 			break;
5775 
5776 		case DIF_OP_SREM:
5777 			if (regs[r2] == 0) {
5778 				regs[rd] = 0;
5779 				*flags |= CPU_DTRACE_DIVZERO;
5780 			} else {
5781 				regs[rd] = (int64_t)regs[r1] %
5782 				    (int64_t)regs[r2];
5783 			}
5784 			break;
5785 
5786 		case DIF_OP_UREM:
5787 			if (regs[r2] == 0) {
5788 				regs[rd] = 0;
5789 				*flags |= CPU_DTRACE_DIVZERO;
5790 			} else {
5791 				regs[rd] = regs[r1] % regs[r2];
5792 			}
5793 			break;
5794 
5795 		case DIF_OP_NOT:
5796 			regs[rd] = ~regs[r1];
5797 			break;
5798 		case DIF_OP_MOV:
5799 			regs[rd] = regs[r1];
5800 			break;
5801 		case DIF_OP_CMP:
5802 			cc_r = regs[r1] - regs[r2];
5803 			cc_n = cc_r < 0;
5804 			cc_z = cc_r == 0;
5805 			cc_v = 0;
5806 			cc_c = regs[r1] < regs[r2];
5807 			break;
5808 		case DIF_OP_TST:
5809 			cc_n = cc_v = cc_c = 0;
5810 			cc_z = regs[r1] == 0;
5811 			break;
5812 		case DIF_OP_BA:
5813 			pc = DIF_INSTR_LABEL(instr);
5814 			break;
5815 		case DIF_OP_BE:
5816 			if (cc_z)
5817 				pc = DIF_INSTR_LABEL(instr);
5818 			break;
5819 		case DIF_OP_BNE:
5820 			if (cc_z == 0)
5821 				pc = DIF_INSTR_LABEL(instr);
5822 			break;
5823 		case DIF_OP_BG:
5824 			if ((cc_z | (cc_n ^ cc_v)) == 0)
5825 				pc = DIF_INSTR_LABEL(instr);
5826 			break;
5827 		case DIF_OP_BGU:
5828 			if ((cc_c | cc_z) == 0)
5829 				pc = DIF_INSTR_LABEL(instr);
5830 			break;
5831 		case DIF_OP_BGE:
5832 			if ((cc_n ^ cc_v) == 0)
5833 				pc = DIF_INSTR_LABEL(instr);
5834 			break;
5835 		case DIF_OP_BGEU:
5836 			if (cc_c == 0)
5837 				pc = DIF_INSTR_LABEL(instr);
5838 			break;
5839 		case DIF_OP_BL:
5840 			if (cc_n ^ cc_v)
5841 				pc = DIF_INSTR_LABEL(instr);
5842 			break;
5843 		case DIF_OP_BLU:
5844 			if (cc_c)
5845 				pc = DIF_INSTR_LABEL(instr);
5846 			break;
5847 		case DIF_OP_BLE:
5848 			if (cc_z | (cc_n ^ cc_v))
5849 				pc = DIF_INSTR_LABEL(instr);
5850 			break;
5851 		case DIF_OP_BLEU:
5852 			if (cc_c | cc_z)
5853 				pc = DIF_INSTR_LABEL(instr);
5854 			break;
5855 		case DIF_OP_RLDSB:
5856 			if (!dtrace_canload(regs[r1], 1, mstate, vstate))
5857 				break;
5858 			/*FALLTHROUGH*/
5859 		case DIF_OP_LDSB:
5860 			regs[rd] = (int8_t)dtrace_load8(regs[r1]);
5861 			break;
5862 		case DIF_OP_RLDSH:
5863 			if (!dtrace_canload(regs[r1], 2, mstate, vstate))
5864 				break;
5865 			/*FALLTHROUGH*/
5866 		case DIF_OP_LDSH:
5867 			regs[rd] = (int16_t)dtrace_load16(regs[r1]);
5868 			break;
5869 		case DIF_OP_RLDSW:
5870 			if (!dtrace_canload(regs[r1], 4, mstate, vstate))
5871 				break;
5872 			/*FALLTHROUGH*/
5873 		case DIF_OP_LDSW:
5874 			regs[rd] = (int32_t)dtrace_load32(regs[r1]);
5875 			break;
5876 		case DIF_OP_RLDUB:
5877 			if (!dtrace_canload(regs[r1], 1, mstate, vstate))
5878 				break;
5879 			/*FALLTHROUGH*/
5880 		case DIF_OP_LDUB:
5881 			regs[rd] = dtrace_load8(regs[r1]);
5882 			break;
5883 		case DIF_OP_RLDUH:
5884 			if (!dtrace_canload(regs[r1], 2, mstate, vstate))
5885 				break;
5886 			/*FALLTHROUGH*/
5887 		case DIF_OP_LDUH:
5888 			regs[rd] = dtrace_load16(regs[r1]);
5889 			break;
5890 		case DIF_OP_RLDUW:
5891 			if (!dtrace_canload(regs[r1], 4, mstate, vstate))
5892 				break;
5893 			/*FALLTHROUGH*/
5894 		case DIF_OP_LDUW:
5895 			regs[rd] = dtrace_load32(regs[r1]);
5896 			break;
5897 		case DIF_OP_RLDX:
5898 			if (!dtrace_canload(regs[r1], 8, mstate, vstate))
5899 				break;
5900 			/*FALLTHROUGH*/
5901 		case DIF_OP_LDX:
5902 			regs[rd] = dtrace_load64(regs[r1]);
5903 			break;
5904 		case DIF_OP_ULDSB:
5905 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5906 			regs[rd] = (int8_t)
5907 			    dtrace_fuword8((void *)(uintptr_t)regs[r1]);
5908 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5909 			break;
5910 		case DIF_OP_ULDSH:
5911 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5912 			regs[rd] = (int16_t)
5913 			    dtrace_fuword16((void *)(uintptr_t)regs[r1]);
5914 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5915 			break;
5916 		case DIF_OP_ULDSW:
5917 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5918 			regs[rd] = (int32_t)
5919 			    dtrace_fuword32((void *)(uintptr_t)regs[r1]);
5920 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5921 			break;
5922 		case DIF_OP_ULDUB:
5923 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5924 			regs[rd] =
5925 			    dtrace_fuword8((void *)(uintptr_t)regs[r1]);
5926 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5927 			break;
5928 		case DIF_OP_ULDUH:
5929 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5930 			regs[rd] =
5931 			    dtrace_fuword16((void *)(uintptr_t)regs[r1]);
5932 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5933 			break;
5934 		case DIF_OP_ULDUW:
5935 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5936 			regs[rd] =
5937 			    dtrace_fuword32((void *)(uintptr_t)regs[r1]);
5938 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5939 			break;
5940 		case DIF_OP_ULDX:
5941 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5942 			regs[rd] =
5943 			    dtrace_fuword64((void *)(uintptr_t)regs[r1]);
5944 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5945 			break;
5946 		case DIF_OP_RET:
5947 			rval = regs[rd];
5948 			pc = textlen;
5949 			break;
5950 		case DIF_OP_NOP:
5951 			break;
5952 		case DIF_OP_SETX:
5953 			regs[rd] = inttab[DIF_INSTR_INTEGER(instr)];
5954 			break;
5955 		case DIF_OP_SETS:
5956 			regs[rd] = (uint64_t)(uintptr_t)
5957 			    (strtab + DIF_INSTR_STRING(instr));
5958 			break;
5959 		case DIF_OP_SCMP: {
5960 			size_t sz = state->dts_options[DTRACEOPT_STRSIZE];
5961 			uintptr_t s1 = regs[r1];
5962 			uintptr_t s2 = regs[r2];
5963 			size_t lim1, lim2;
5964 
5965 			if (s1 != NULL &&
5966 			    !dtrace_strcanload(s1, sz, &lim1, mstate, vstate))
5967 				break;
5968 			if (s2 != NULL &&
5969 			    !dtrace_strcanload(s2, sz, &lim2, mstate, vstate))
5970 				break;
5971 
5972 			cc_r = dtrace_strncmp((char *)s1, (char *)s2,
5973 			    MIN(lim1, lim2));
5974 
5975 			cc_n = cc_r < 0;
5976 			cc_z = cc_r == 0;
5977 			cc_v = cc_c = 0;
5978 			break;
5979 		}
5980 		case DIF_OP_LDGA:
5981 			regs[rd] = dtrace_dif_variable(mstate, state,
5982 			    r1, regs[r2]);
5983 			break;
5984 		case DIF_OP_LDGS:
5985 			id = DIF_INSTR_VAR(instr);
5986 
5987 			if (id >= DIF_VAR_OTHER_UBASE) {
5988 				uintptr_t a;
5989 
5990 				id -= DIF_VAR_OTHER_UBASE;
5991 				svar = vstate->dtvs_globals[id];
5992 				ASSERT(svar != NULL);
5993 				v = &svar->dtsv_var;
5994 
5995 				if (!(v->dtdv_type.dtdt_flags & DIF_TF_BYREF)) {
5996 					regs[rd] = svar->dtsv_data;
5997 					break;
5998 				}
5999 
6000 				a = (uintptr_t)svar->dtsv_data;
6001 
6002 				if (*(uint8_t *)a == UINT8_MAX) {
6003 					/*
6004 					 * If the 0th byte is set to UINT8_MAX
6005 					 * then this is to be treated as a
6006 					 * reference to a NULL variable.
6007 					 */
6008 					regs[rd] = NULL;
6009 				} else {
6010 					regs[rd] = a + sizeof (uint64_t);
6011 				}
6012 
6013 				break;
6014 			}
6015 
6016 			regs[rd] = dtrace_dif_variable(mstate, state, id, 0);
6017 			break;
6018 
6019 		case DIF_OP_STGS:
6020 			id = DIF_INSTR_VAR(instr);
6021 
6022 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
6023 			id -= DIF_VAR_OTHER_UBASE;
6024 
6025 			VERIFY(id < vstate->dtvs_nglobals);
6026 			svar = vstate->dtvs_globals[id];
6027 			ASSERT(svar != NULL);
6028 			v = &svar->dtsv_var;
6029 
6030 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6031 				uintptr_t a = (uintptr_t)svar->dtsv_data;
6032 				size_t lim;
6033 
6034 				ASSERT(a != NULL);
6035 				ASSERT(svar->dtsv_size != 0);
6036 
6037 				if (regs[rd] == NULL) {
6038 					*(uint8_t *)a = UINT8_MAX;
6039 					break;
6040 				} else {
6041 					*(uint8_t *)a = 0;
6042 					a += sizeof (uint64_t);
6043 				}
6044 				if (!dtrace_vcanload(
6045 				    (void *)(uintptr_t)regs[rd], &v->dtdv_type,
6046 				    &lim, mstate, vstate))
6047 					break;
6048 
6049 				dtrace_vcopy((void *)(uintptr_t)regs[rd],
6050 				    (void *)a, &v->dtdv_type, lim);
6051 				break;
6052 			}
6053 
6054 			svar->dtsv_data = regs[rd];
6055 			break;
6056 
6057 		case DIF_OP_LDTA:
6058 			/*
6059 			 * There are no DTrace built-in thread-local arrays at
6060 			 * present.  This opcode is saved for future work.
6061 			 */
6062 			*flags |= CPU_DTRACE_ILLOP;
6063 			regs[rd] = 0;
6064 			break;
6065 
6066 		case DIF_OP_LDLS:
6067 			id = DIF_INSTR_VAR(instr);
6068 
6069 			if (id < DIF_VAR_OTHER_UBASE) {
6070 				/*
6071 				 * For now, this has no meaning.
6072 				 */
6073 				regs[rd] = 0;
6074 				break;
6075 			}
6076 
6077 			id -= DIF_VAR_OTHER_UBASE;
6078 
6079 			ASSERT(id < vstate->dtvs_nlocals);
6080 			ASSERT(vstate->dtvs_locals != NULL);
6081 
6082 			svar = vstate->dtvs_locals[id];
6083 			ASSERT(svar != NULL);
6084 			v = &svar->dtsv_var;
6085 
6086 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6087 				uintptr_t a = (uintptr_t)svar->dtsv_data;
6088 				size_t sz = v->dtdv_type.dtdt_size;
6089 
6090 				sz += sizeof (uint64_t);
6091 				ASSERT(svar->dtsv_size == NCPU * sz);
6092 				a += CPU->cpu_id * sz;
6093 
6094 				if (*(uint8_t *)a == UINT8_MAX) {
6095 					/*
6096 					 * If the 0th byte is set to UINT8_MAX
6097 					 * then this is to be treated as a
6098 					 * reference to a NULL variable.
6099 					 */
6100 					regs[rd] = NULL;
6101 				} else {
6102 					regs[rd] = a + sizeof (uint64_t);
6103 				}
6104 
6105 				break;
6106 			}
6107 
6108 			ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
6109 			tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
6110 			regs[rd] = tmp[CPU->cpu_id];
6111 			break;
6112 
6113 		case DIF_OP_STLS:
6114 			id = DIF_INSTR_VAR(instr);
6115 
6116 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
6117 			id -= DIF_VAR_OTHER_UBASE;
6118 			VERIFY(id < vstate->dtvs_nlocals);
6119 
6120 			ASSERT(vstate->dtvs_locals != NULL);
6121 			svar = vstate->dtvs_locals[id];
6122 			ASSERT(svar != NULL);
6123 			v = &svar->dtsv_var;
6124 
6125 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6126 				uintptr_t a = (uintptr_t)svar->dtsv_data;
6127 				size_t sz = v->dtdv_type.dtdt_size;
6128 				size_t lim;
6129 
6130 				sz += sizeof (uint64_t);
6131 				ASSERT(svar->dtsv_size == NCPU * sz);
6132 				a += CPU->cpu_id * sz;
6133 
6134 				if (regs[rd] == NULL) {
6135 					*(uint8_t *)a = UINT8_MAX;
6136 					break;
6137 				} else {
6138 					*(uint8_t *)a = 0;
6139 					a += sizeof (uint64_t);
6140 				}
6141 
6142 				if (!dtrace_vcanload(
6143 				    (void *)(uintptr_t)regs[rd], &v->dtdv_type,
6144 				    &lim, mstate, vstate))
6145 					break;
6146 
6147 				dtrace_vcopy((void *)(uintptr_t)regs[rd],
6148 				    (void *)a, &v->dtdv_type, lim);
6149 				break;
6150 			}
6151 
6152 			ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
6153 			tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
6154 			tmp[CPU->cpu_id] = regs[rd];
6155 			break;
6156 
6157 		case DIF_OP_LDTS: {
6158 			dtrace_dynvar_t *dvar;
6159 			dtrace_key_t *key;
6160 
6161 			id = DIF_INSTR_VAR(instr);
6162 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
6163 			id -= DIF_VAR_OTHER_UBASE;
6164 			v = &vstate->dtvs_tlocals[id];
6165 
6166 			key = &tupregs[DIF_DTR_NREGS];
6167 			key[0].dttk_value = (uint64_t)id;
6168 			key[0].dttk_size = 0;
6169 			DTRACE_TLS_THRKEY(key[1].dttk_value);
6170 			key[1].dttk_size = 0;
6171 
6172 			dvar = dtrace_dynvar(dstate, 2, key,
6173 			    sizeof (uint64_t), DTRACE_DYNVAR_NOALLOC,
6174 			    mstate, vstate);
6175 
6176 			if (dvar == NULL) {
6177 				regs[rd] = 0;
6178 				break;
6179 			}
6180 
6181 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6182 				regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
6183 			} else {
6184 				regs[rd] = *((uint64_t *)dvar->dtdv_data);
6185 			}
6186 
6187 			break;
6188 		}
6189 
6190 		case DIF_OP_STTS: {
6191 			dtrace_dynvar_t *dvar;
6192 			dtrace_key_t *key;
6193 
6194 			id = DIF_INSTR_VAR(instr);
6195 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
6196 			id -= DIF_VAR_OTHER_UBASE;
6197 			VERIFY(id < vstate->dtvs_ntlocals);
6198 
6199 			key = &tupregs[DIF_DTR_NREGS];
6200 			key[0].dttk_value = (uint64_t)id;
6201 			key[0].dttk_size = 0;
6202 			DTRACE_TLS_THRKEY(key[1].dttk_value);
6203 			key[1].dttk_size = 0;
6204 			v = &vstate->dtvs_tlocals[id];
6205 
6206 			dvar = dtrace_dynvar(dstate, 2, key,
6207 			    v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
6208 			    v->dtdv_type.dtdt_size : sizeof (uint64_t),
6209 			    regs[rd] ? DTRACE_DYNVAR_ALLOC :
6210 			    DTRACE_DYNVAR_DEALLOC, mstate, vstate);
6211 
6212 			/*
6213 			 * Given that we're storing to thread-local data,
6214 			 * we need to flush our predicate cache.
6215 			 */
6216 			curthread->t_predcache = NULL;
6217 
6218 			if (dvar == NULL)
6219 				break;
6220 
6221 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6222 				size_t lim;
6223 
6224 				if (!dtrace_vcanload(
6225 				    (void *)(uintptr_t)regs[rd],
6226 				    &v->dtdv_type, &lim, mstate, vstate))
6227 					break;
6228 
6229 				dtrace_vcopy((void *)(uintptr_t)regs[rd],
6230 				    dvar->dtdv_data, &v->dtdv_type, lim);
6231 			} else {
6232 				*((uint64_t *)dvar->dtdv_data) = regs[rd];
6233 			}
6234 
6235 			break;
6236 		}
6237 
6238 		case DIF_OP_SRA:
6239 			regs[rd] = (int64_t)regs[r1] >> regs[r2];
6240 			break;
6241 
6242 		case DIF_OP_CALL:
6243 			dtrace_dif_subr(DIF_INSTR_SUBR(instr), rd,
6244 			    regs, tupregs, ttop, mstate, state);
6245 			break;
6246 
6247 		case DIF_OP_PUSHTR:
6248 			if (ttop == DIF_DTR_NREGS) {
6249 				*flags |= CPU_DTRACE_TUPOFLOW;
6250 				break;
6251 			}
6252 
6253 			if (r1 == DIF_TYPE_STRING) {
6254 				/*
6255 				 * If this is a string type and the size is 0,
6256 				 * we'll use the system-wide default string
6257 				 * size.  Note that we are _not_ looking at
6258 				 * the value of the DTRACEOPT_STRSIZE option;
6259 				 * had this been set, we would expect to have
6260 				 * a non-zero size value in the "pushtr".
6261 				 */
6262 				tupregs[ttop].dttk_size =
6263 				    dtrace_strlen((char *)(uintptr_t)regs[rd],
6264 				    regs[r2] ? regs[r2] :
6265 				    dtrace_strsize_default) + 1;
6266 			} else {
6267 				if (regs[r2] > LONG_MAX) {
6268 					*flags |= CPU_DTRACE_ILLOP;
6269 					break;
6270 				}
6271 
6272 				tupregs[ttop].dttk_size = regs[r2];
6273 			}
6274 
6275 			tupregs[ttop++].dttk_value = regs[rd];
6276 			break;
6277 
6278 		case DIF_OP_PUSHTV:
6279 			if (ttop == DIF_DTR_NREGS) {
6280 				*flags |= CPU_DTRACE_TUPOFLOW;
6281 				break;
6282 			}
6283 
6284 			tupregs[ttop].dttk_value = regs[rd];
6285 			tupregs[ttop++].dttk_size = 0;
6286 			break;
6287 
6288 		case DIF_OP_POPTS:
6289 			if (ttop != 0)
6290 				ttop--;
6291 			break;
6292 
6293 		case DIF_OP_FLUSHTS:
6294 			ttop = 0;
6295 			break;
6296 
6297 		case DIF_OP_LDGAA:
6298 		case DIF_OP_LDTAA: {
6299 			dtrace_dynvar_t *dvar;
6300 			dtrace_key_t *key = tupregs;
6301 			uint_t nkeys = ttop;
6302 
6303 			id = DIF_INSTR_VAR(instr);
6304 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
6305 			id -= DIF_VAR_OTHER_UBASE;
6306 
6307 			key[nkeys].dttk_value = (uint64_t)id;
6308 			key[nkeys++].dttk_size = 0;
6309 
6310 			if (DIF_INSTR_OP(instr) == DIF_OP_LDTAA) {
6311 				DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
6312 				key[nkeys++].dttk_size = 0;
6313 				VERIFY(id < vstate->dtvs_ntlocals);
6314 				v = &vstate->dtvs_tlocals[id];
6315 			} else {
6316 				VERIFY(id < vstate->dtvs_nglobals);
6317 				v = &vstate->dtvs_globals[id]->dtsv_var;
6318 			}
6319 
6320 			dvar = dtrace_dynvar(dstate, nkeys, key,
6321 			    v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
6322 			    v->dtdv_type.dtdt_size : sizeof (uint64_t),
6323 			    DTRACE_DYNVAR_NOALLOC, mstate, vstate);
6324 
6325 			if (dvar == NULL) {
6326 				regs[rd] = 0;
6327 				break;
6328 			}
6329 
6330 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6331 				regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
6332 			} else {
6333 				regs[rd] = *((uint64_t *)dvar->dtdv_data);
6334 			}
6335 
6336 			break;
6337 		}
6338 
6339 		case DIF_OP_STGAA:
6340 		case DIF_OP_STTAA: {
6341 			dtrace_dynvar_t *dvar;
6342 			dtrace_key_t *key = tupregs;
6343 			uint_t nkeys = ttop;
6344 
6345 			id = DIF_INSTR_VAR(instr);
6346 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
6347 			id -= DIF_VAR_OTHER_UBASE;
6348 
6349 			key[nkeys].dttk_value = (uint64_t)id;
6350 			key[nkeys++].dttk_size = 0;
6351 
6352 			if (DIF_INSTR_OP(instr) == DIF_OP_STTAA) {
6353 				DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
6354 				key[nkeys++].dttk_size = 0;
6355 				VERIFY(id < vstate->dtvs_ntlocals);
6356 				v = &vstate->dtvs_tlocals[id];
6357 			} else {
6358 				VERIFY(id < vstate->dtvs_nglobals);
6359 				v = &vstate->dtvs_globals[id]->dtsv_var;
6360 			}
6361 
6362 			dvar = dtrace_dynvar(dstate, nkeys, key,
6363 			    v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
6364 			    v->dtdv_type.dtdt_size : sizeof (uint64_t),
6365 			    regs[rd] ? DTRACE_DYNVAR_ALLOC :
6366 			    DTRACE_DYNVAR_DEALLOC, mstate, vstate);
6367 
6368 			if (dvar == NULL)
6369 				break;
6370 
6371 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6372 				size_t lim;
6373 
6374 				if (!dtrace_vcanload(
6375 				    (void *)(uintptr_t)regs[rd], &v->dtdv_type,
6376 				    &lim, mstate, vstate))
6377 					break;
6378 
6379 				dtrace_vcopy((void *)(uintptr_t)regs[rd],
6380 				    dvar->dtdv_data, &v->dtdv_type, lim);
6381 			} else {
6382 				*((uint64_t *)dvar->dtdv_data) = regs[rd];
6383 			}
6384 
6385 			break;
6386 		}
6387 
6388 		case DIF_OP_ALLOCS: {
6389 			uintptr_t ptr = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
6390 			size_t size = ptr - mstate->dtms_scratch_ptr + regs[r1];
6391 
6392 			/*
6393 			 * Rounding up the user allocation size could have
6394 			 * overflowed large, bogus allocations (like -1ULL) to
6395 			 * 0.
6396 			 */
6397 			if (size < regs[r1] ||
6398 			    !DTRACE_INSCRATCH(mstate, size)) {
6399 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
6400 				regs[rd] = NULL;
6401 				break;
6402 			}
6403 
6404 			dtrace_bzero((void *) mstate->dtms_scratch_ptr, size);
6405 			mstate->dtms_scratch_ptr += size;
6406 			regs[rd] = ptr;
6407 			break;
6408 		}
6409 
6410 		case DIF_OP_COPYS:
6411 			if (!dtrace_canstore(regs[rd], regs[r2],
6412 			    mstate, vstate)) {
6413 				*flags |= CPU_DTRACE_BADADDR;
6414 				*illval = regs[rd];
6415 				break;
6416 			}
6417 
6418 			if (!dtrace_canload(regs[r1], regs[r2], mstate, vstate))
6419 				break;
6420 
6421 			dtrace_bcopy((void *)(uintptr_t)regs[r1],
6422 			    (void *)(uintptr_t)regs[rd], (size_t)regs[r2]);
6423 			break;
6424 
6425 		case DIF_OP_STB:
6426 			if (!dtrace_canstore(regs[rd], 1, mstate, vstate)) {
6427 				*flags |= CPU_DTRACE_BADADDR;
6428 				*illval = regs[rd];
6429 				break;
6430 			}
6431 			*((uint8_t *)(uintptr_t)regs[rd]) = (uint8_t)regs[r1];
6432 			break;
6433 
6434 		case DIF_OP_STH:
6435 			if (!dtrace_canstore(regs[rd], 2, mstate, vstate)) {
6436 				*flags |= CPU_DTRACE_BADADDR;
6437 				*illval = regs[rd];
6438 				break;
6439 			}
6440 			if (regs[rd] & 1) {
6441 				*flags |= CPU_DTRACE_BADALIGN;
6442 				*illval = regs[rd];
6443 				break;
6444 			}
6445 			*((uint16_t *)(uintptr_t)regs[rd]) = (uint16_t)regs[r1];
6446 			break;
6447 
6448 		case DIF_OP_STW:
6449 			if (!dtrace_canstore(regs[rd], 4, mstate, vstate)) {
6450 				*flags |= CPU_DTRACE_BADADDR;
6451 				*illval = regs[rd];
6452 				break;
6453 			}
6454 			if (regs[rd] & 3) {
6455 				*flags |= CPU_DTRACE_BADALIGN;
6456 				*illval = regs[rd];
6457 				break;
6458 			}
6459 			*((uint32_t *)(uintptr_t)regs[rd]) = (uint32_t)regs[r1];
6460 			break;
6461 
6462 		case DIF_OP_STX:
6463 			if (!dtrace_canstore(regs[rd], 8, mstate, vstate)) {
6464 				*flags |= CPU_DTRACE_BADADDR;
6465 				*illval = regs[rd];
6466 				break;
6467 			}
6468 			if (regs[rd] & 7) {
6469 				*flags |= CPU_DTRACE_BADALIGN;
6470 				*illval = regs[rd];
6471 				break;
6472 			}
6473 			*((uint64_t *)(uintptr_t)regs[rd]) = regs[r1];
6474 			break;
6475 		}
6476 	}
6477 
6478 	if (!(*flags & CPU_DTRACE_FAULT))
6479 		return (rval);
6480 
6481 	mstate->dtms_fltoffs = opc * sizeof (dif_instr_t);
6482 	mstate->dtms_present |= DTRACE_MSTATE_FLTOFFS;
6483 
6484 	return (0);
6485 }
6486 
6487 static void
6488 dtrace_action_breakpoint(dtrace_ecb_t *ecb)
6489 {
6490 	dtrace_probe_t *probe = ecb->dte_probe;
6491 	dtrace_provider_t *prov = probe->dtpr_provider;
6492 	char c[DTRACE_FULLNAMELEN + 80], *str;
6493 	char *msg = "dtrace: breakpoint action at probe ";
6494 	char *ecbmsg = " (ecb ";
6495 	uintptr_t mask = (0xf << (sizeof (uintptr_t) * NBBY / 4));
6496 	uintptr_t val = (uintptr_t)ecb;
6497 	int shift = (sizeof (uintptr_t) * NBBY) - 4, i = 0;
6498 
6499 	if (dtrace_destructive_disallow)
6500 		return;
6501 
6502 	/*
6503 	 * It's impossible to be taking action on the NULL probe.
6504 	 */
6505 	ASSERT(probe != NULL);
6506 
6507 	/*
6508 	 * This is a poor man's (destitute man's?) sprintf():  we want to
6509 	 * print the provider name, module name, function name and name of
6510 	 * the probe, along with the hex address of the ECB with the breakpoint
6511 	 * action -- all of which we must place in the character buffer by
6512 	 * hand.
6513 	 */
6514 	while (*msg != '\0')
6515 		c[i++] = *msg++;
6516 
6517 	for (str = prov->dtpv_name; *str != '\0'; str++)
6518 		c[i++] = *str;
6519 	c[i++] = ':';
6520 
6521 	for (str = probe->dtpr_mod; *str != '\0'; str++)
6522 		c[i++] = *str;
6523 	c[i++] = ':';
6524 
6525 	for (str = probe->dtpr_func; *str != '\0'; str++)
6526 		c[i++] = *str;
6527 	c[i++] = ':';
6528 
6529 	for (str = probe->dtpr_name; *str != '\0'; str++)
6530 		c[i++] = *str;
6531 
6532 	while (*ecbmsg != '\0')
6533 		c[i++] = *ecbmsg++;
6534 
6535 	while (shift >= 0) {
6536 		mask = (uintptr_t)0xf << shift;
6537 
6538 		if (val >= ((uintptr_t)1 << shift))
6539 			c[i++] = "0123456789abcdef"[(val & mask) >> shift];
6540 		shift -= 4;
6541 	}
6542 
6543 	c[i++] = ')';
6544 	c[i] = '\0';
6545 
6546 	debug_enter(c);
6547 }
6548 
6549 static void
6550 dtrace_action_panic(dtrace_ecb_t *ecb)
6551 {
6552 	dtrace_probe_t *probe = ecb->dte_probe;
6553 
6554 	/*
6555 	 * It's impossible to be taking action on the NULL probe.
6556 	 */
6557 	ASSERT(probe != NULL);
6558 
6559 	if (dtrace_destructive_disallow)
6560 		return;
6561 
6562 	if (dtrace_panicked != NULL)
6563 		return;
6564 
6565 	if (dtrace_casptr(&dtrace_panicked, NULL, curthread) != NULL)
6566 		return;
6567 
6568 	/*
6569 	 * We won the right to panic.  (We want to be sure that only one
6570 	 * thread calls panic() from dtrace_probe(), and that panic() is
6571 	 * called exactly once.)
6572 	 */
6573 	dtrace_panic("dtrace: panic action at probe %s:%s:%s:%s (ecb %p)",
6574 	    probe->dtpr_provider->dtpv_name, probe->dtpr_mod,
6575 	    probe->dtpr_func, probe->dtpr_name, (void *)ecb);
6576 }
6577 
6578 static void
6579 dtrace_action_raise(uint64_t sig)
6580 {
6581 	if (dtrace_destructive_disallow)
6582 		return;
6583 
6584 	if (sig >= NSIG) {
6585 		DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
6586 		return;
6587 	}
6588 
6589 	/*
6590 	 * raise() has a queue depth of 1 -- we ignore all subsequent
6591 	 * invocations of the raise() action.
6592 	 */
6593 	if (curthread->t_dtrace_sig == 0)
6594 		curthread->t_dtrace_sig = (uint8_t)sig;
6595 
6596 	curthread->t_sig_check = 1;
6597 	aston(curthread);
6598 }
6599 
6600 static void
6601 dtrace_action_stop(void)
6602 {
6603 	if (dtrace_destructive_disallow)
6604 		return;
6605 
6606 	if (!curthread->t_dtrace_stop) {
6607 		curthread->t_dtrace_stop = 1;
6608 		curthread->t_sig_check = 1;
6609 		aston(curthread);
6610 	}
6611 }
6612 
6613 static void
6614 dtrace_action_chill(dtrace_mstate_t *mstate, hrtime_t val)
6615 {
6616 	hrtime_t now;
6617 	volatile uint16_t *flags;
6618 	cpu_t *cpu = CPU;
6619 
6620 	if (dtrace_destructive_disallow)
6621 		return;
6622 
6623 	flags = (volatile uint16_t *)&cpu_core[cpu->cpu_id].cpuc_dtrace_flags;
6624 
6625 	now = dtrace_gethrtime();
6626 
6627 	if (now - cpu->cpu_dtrace_chillmark > dtrace_chill_interval) {
6628 		/*
6629 		 * We need to advance the mark to the current time.
6630 		 */
6631 		cpu->cpu_dtrace_chillmark = now;
6632 		cpu->cpu_dtrace_chilled = 0;
6633 	}
6634 
6635 	/*
6636 	 * Now check to see if the requested chill time would take us over
6637 	 * the maximum amount of time allowed in the chill interval.  (Or
6638 	 * worse, if the calculation itself induces overflow.)
6639 	 */
6640 	if (cpu->cpu_dtrace_chilled + val > dtrace_chill_max ||
6641 	    cpu->cpu_dtrace_chilled + val < cpu->cpu_dtrace_chilled) {
6642 		*flags |= CPU_DTRACE_ILLOP;
6643 		return;
6644 	}
6645 
6646 	while (dtrace_gethrtime() - now < val)
6647 		continue;
6648 
6649 	/*
6650 	 * Normally, we assure that the value of the variable "timestamp" does
6651 	 * not change within an ECB.  The presence of chill() represents an
6652 	 * exception to this rule, however.
6653 	 */
6654 	mstate->dtms_present &= ~DTRACE_MSTATE_TIMESTAMP;
6655 	cpu->cpu_dtrace_chilled += val;
6656 }
6657 
6658 static void
6659 dtrace_action_ustack(dtrace_mstate_t *mstate, dtrace_state_t *state,
6660     uint64_t *buf, uint64_t arg)
6661 {
6662 	int nframes = DTRACE_USTACK_NFRAMES(arg);
6663 	int strsize = DTRACE_USTACK_STRSIZE(arg);
6664 	uint64_t *pcs = &buf[1], *fps;
6665 	char *str = (char *)&pcs[nframes];
6666 	int size, offs = 0, i, j;
6667 	size_t rem;
6668 	uintptr_t old = mstate->dtms_scratch_ptr, saved;
6669 	uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
6670 	char *sym;
6671 
6672 	/*
6673 	 * Should be taking a faster path if string space has not been
6674 	 * allocated.
6675 	 */
6676 	ASSERT(strsize != 0);
6677 
6678 	/*
6679 	 * We will first allocate some temporary space for the frame pointers.
6680 	 */
6681 	fps = (uint64_t *)P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
6682 	size = (uintptr_t)fps - mstate->dtms_scratch_ptr +
6683 	    (nframes * sizeof (uint64_t));
6684 
6685 	if (!DTRACE_INSCRATCH(mstate, size)) {
6686 		/*
6687 		 * Not enough room for our frame pointers -- need to indicate
6688 		 * that we ran out of scratch space.
6689 		 */
6690 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
6691 		return;
6692 	}
6693 
6694 	mstate->dtms_scratch_ptr += size;
6695 	saved = mstate->dtms_scratch_ptr;
6696 
6697 	/*
6698 	 * Now get a stack with both program counters and frame pointers.
6699 	 */
6700 	DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6701 	dtrace_getufpstack(buf, fps, nframes + 1);
6702 	DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6703 
6704 	/*
6705 	 * If that faulted, we're cooked.
6706 	 */
6707 	if (*flags & CPU_DTRACE_FAULT)
6708 		goto out;
6709 
6710 	/*
6711 	 * Now we want to walk up the stack, calling the USTACK helper.  For
6712 	 * each iteration, we restore the scratch pointer.
6713 	 */
6714 	for (i = 0; i < nframes; i++) {
6715 		mstate->dtms_scratch_ptr = saved;
6716 
6717 		if (offs >= strsize)
6718 			break;
6719 
6720 		sym = (char *)(uintptr_t)dtrace_helper(
6721 		    DTRACE_HELPER_ACTION_USTACK,
6722 		    mstate, state, pcs[i], fps[i]);
6723 
6724 		/*
6725 		 * If we faulted while running the helper, we're going to
6726 		 * clear the fault and null out the corresponding string.
6727 		 */
6728 		if (*flags & CPU_DTRACE_FAULT) {
6729 			*flags &= ~CPU_DTRACE_FAULT;
6730 			str[offs++] = '\0';
6731 			continue;
6732 		}
6733 
6734 		if (sym == NULL) {
6735 			str[offs++] = '\0';
6736 			continue;
6737 		}
6738 
6739 		if (!dtrace_strcanload((uintptr_t)sym, strsize, &rem, mstate,
6740 		    &(state->dts_vstate))) {
6741 			str[offs++] = '\0';
6742 			continue;
6743 		}
6744 
6745 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6746 
6747 		/*
6748 		 * Now copy in the string that the helper returned to us.
6749 		 */
6750 		for (j = 0; offs + j < strsize && j < rem; j++) {
6751 			if ((str[offs + j] = sym[j]) == '\0')
6752 				break;
6753 		}
6754 
6755 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6756 
6757 		offs += j + 1;
6758 	}
6759 
6760 	if (offs >= strsize) {
6761 		/*
6762 		 * If we didn't have room for all of the strings, we don't
6763 		 * abort processing -- this needn't be a fatal error -- but we
6764 		 * still want to increment a counter (dts_stkstroverflows) to
6765 		 * allow this condition to be warned about.  (If this is from
6766 		 * a jstack() action, it is easily tuned via jstackstrsize.)
6767 		 */
6768 		dtrace_error(&state->dts_stkstroverflows);
6769 	}
6770 
6771 	while (offs < strsize)
6772 		str[offs++] = '\0';
6773 
6774 out:
6775 	mstate->dtms_scratch_ptr = old;
6776 }
6777 
6778 static void
6779 dtrace_store_by_ref(dtrace_difo_t *dp, caddr_t tomax, size_t size,
6780     size_t *valoffsp, uint64_t *valp, uint64_t end, int intuple, int dtkind)
6781 {
6782 	volatile uint16_t *flags;
6783 	uint64_t val = *valp;
6784 	size_t valoffs = *valoffsp;
6785 
6786 	flags = (volatile uint16_t *)&cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
6787 	ASSERT(dtkind == DIF_TF_BYREF || dtkind == DIF_TF_BYUREF);
6788 
6789 	/*
6790 	 * If this is a string, we're going to only load until we find the zero
6791 	 * byte -- after which we'll store zero bytes.
6792 	 */
6793 	if (dp->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING) {
6794 		char c = '\0' + 1;
6795 		size_t s;
6796 
6797 		for (s = 0; s < size; s++) {
6798 			if (c != '\0' && dtkind == DIF_TF_BYREF) {
6799 				c = dtrace_load8(val++);
6800 			} else if (c != '\0' && dtkind == DIF_TF_BYUREF) {
6801 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6802 				c = dtrace_fuword8((void *)(uintptr_t)val++);
6803 				DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6804 				if (*flags & CPU_DTRACE_FAULT)
6805 					break;
6806 			}
6807 
6808 			DTRACE_STORE(uint8_t, tomax, valoffs++, c);
6809 
6810 			if (c == '\0' && intuple)
6811 				break;
6812 		}
6813 	} else {
6814 		uint8_t c;
6815 		while (valoffs < end) {
6816 			if (dtkind == DIF_TF_BYREF) {
6817 				c = dtrace_load8(val++);
6818 			} else if (dtkind == DIF_TF_BYUREF) {
6819 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6820 				c = dtrace_fuword8((void *)(uintptr_t)val++);
6821 				DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6822 				if (*flags & CPU_DTRACE_FAULT)
6823 					break;
6824 			}
6825 
6826 			DTRACE_STORE(uint8_t, tomax,
6827 			    valoffs++, c);
6828 		}
6829 	}
6830 
6831 	*valp = val;
6832 	*valoffsp = valoffs;
6833 }
6834 
6835 /*
6836  * If you're looking for the epicenter of DTrace, you just found it.  This
6837  * is the function called by the provider to fire a probe -- from which all
6838  * subsequent probe-context DTrace activity emanates.
6839  */
6840 void
6841 dtrace_probe(dtrace_id_t id, uintptr_t arg0, uintptr_t arg1,
6842     uintptr_t arg2, uintptr_t arg3, uintptr_t arg4)
6843 {
6844 	processorid_t cpuid;
6845 	dtrace_icookie_t cookie;
6846 	dtrace_probe_t *probe;
6847 	dtrace_mstate_t mstate;
6848 	dtrace_ecb_t *ecb;
6849 	dtrace_action_t *act;
6850 	intptr_t offs;
6851 	size_t size;
6852 	int vtime, onintr;
6853 	volatile uint16_t *flags;
6854 	hrtime_t now, end;
6855 
6856 	/*
6857 	 * Kick out immediately if this CPU is still being born (in which case
6858 	 * curthread will be set to -1) or the current thread can't allow
6859 	 * probes in its current context.
6860 	 */
6861 	if (((uintptr_t)curthread & 1) || (curthread->t_flag & T_DONTDTRACE))
6862 		return;
6863 
6864 	cookie = dtrace_interrupt_disable();
6865 	probe = dtrace_probes[id - 1];
6866 	cpuid = CPU->cpu_id;
6867 	onintr = CPU_ON_INTR(CPU);
6868 
6869 	CPU->cpu_dtrace_probes++;
6870 
6871 	if (!onintr && probe->dtpr_predcache != DTRACE_CACHEIDNONE &&
6872 	    probe->dtpr_predcache == curthread->t_predcache) {
6873 		/*
6874 		 * We have hit in the predicate cache; we know that
6875 		 * this predicate would evaluate to be false.
6876 		 */
6877 		dtrace_interrupt_enable(cookie);
6878 		return;
6879 	}
6880 
6881 	if (panic_quiesce) {
6882 		/*
6883 		 * We don't trace anything if we're panicking.
6884 		 */
6885 		dtrace_interrupt_enable(cookie);
6886 		return;
6887 	}
6888 
6889 	now = mstate.dtms_timestamp = dtrace_gethrtime();
6890 	mstate.dtms_present |= DTRACE_MSTATE_TIMESTAMP;
6891 	vtime = dtrace_vtime_references != 0;
6892 
6893 	if (vtime && curthread->t_dtrace_start)
6894 		curthread->t_dtrace_vtime += now - curthread->t_dtrace_start;
6895 
6896 	mstate.dtms_difo = NULL;
6897 	mstate.dtms_probe = probe;
6898 	mstate.dtms_strtok = NULL;
6899 	mstate.dtms_arg[0] = arg0;
6900 	mstate.dtms_arg[1] = arg1;
6901 	mstate.dtms_arg[2] = arg2;
6902 	mstate.dtms_arg[3] = arg3;
6903 	mstate.dtms_arg[4] = arg4;
6904 
6905 	flags = (volatile uint16_t *)&cpu_core[cpuid].cpuc_dtrace_flags;
6906 
6907 	for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
6908 		dtrace_predicate_t *pred = ecb->dte_predicate;
6909 		dtrace_state_t *state = ecb->dte_state;
6910 		dtrace_buffer_t *buf = &state->dts_buffer[cpuid];
6911 		dtrace_buffer_t *aggbuf = &state->dts_aggbuffer[cpuid];
6912 		dtrace_vstate_t *vstate = &state->dts_vstate;
6913 		dtrace_provider_t *prov = probe->dtpr_provider;
6914 		uint64_t tracememsize = 0;
6915 		int committed = 0;
6916 		caddr_t tomax;
6917 
6918 		/*
6919 		 * A little subtlety with the following (seemingly innocuous)
6920 		 * declaration of the automatic 'val':  by looking at the
6921 		 * code, you might think that it could be declared in the
6922 		 * action processing loop, below.  (That is, it's only used in
6923 		 * the action processing loop.)  However, it must be declared
6924 		 * out of that scope because in the case of DIF expression
6925 		 * arguments to aggregating actions, one iteration of the
6926 		 * action loop will use the last iteration's value.
6927 		 */
6928 #ifdef lint
6929 		uint64_t val = 0;
6930 #else
6931 		uint64_t val;
6932 #endif
6933 
6934 		mstate.dtms_present = DTRACE_MSTATE_ARGS | DTRACE_MSTATE_PROBE;
6935 		mstate.dtms_access = DTRACE_ACCESS_ARGS | DTRACE_ACCESS_PROC;
6936 		mstate.dtms_getf = NULL;
6937 
6938 		*flags &= ~CPU_DTRACE_ERROR;
6939 
6940 		if (prov == dtrace_provider) {
6941 			/*
6942 			 * If dtrace itself is the provider of this probe,
6943 			 * we're only going to continue processing the ECB if
6944 			 * arg0 (the dtrace_state_t) is equal to the ECB's
6945 			 * creating state.  (This prevents disjoint consumers
6946 			 * from seeing one another's metaprobes.)
6947 			 */
6948 			if (arg0 != (uint64_t)(uintptr_t)state)
6949 				continue;
6950 		}
6951 
6952 		if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE) {
6953 			/*
6954 			 * We're not currently active.  If our provider isn't
6955 			 * the dtrace pseudo provider, we're not interested.
6956 			 */
6957 			if (prov != dtrace_provider)
6958 				continue;
6959 
6960 			/*
6961 			 * Now we must further check if we are in the BEGIN
6962 			 * probe.  If we are, we will only continue processing
6963 			 * if we're still in WARMUP -- if one BEGIN enabling
6964 			 * has invoked the exit() action, we don't want to
6965 			 * evaluate subsequent BEGIN enablings.
6966 			 */
6967 			if (probe->dtpr_id == dtrace_probeid_begin &&
6968 			    state->dts_activity != DTRACE_ACTIVITY_WARMUP) {
6969 				ASSERT(state->dts_activity ==
6970 				    DTRACE_ACTIVITY_DRAINING);
6971 				continue;
6972 			}
6973 		}
6974 
6975 		if (ecb->dte_cond && !dtrace_priv_probe(state, &mstate, ecb))
6976 			continue;
6977 
6978 		if (now - state->dts_alive > dtrace_deadman_timeout) {
6979 			/*
6980 			 * We seem to be dead.  Unless we (a) have kernel
6981 			 * destructive permissions (b) have explicitly enabled
6982 			 * destructive actions and (c) destructive actions have
6983 			 * not been disabled, we're going to transition into
6984 			 * the KILLED state, from which no further processing
6985 			 * on this state will be performed.
6986 			 */
6987 			if (!dtrace_priv_kernel_destructive(state) ||
6988 			    !state->dts_cred.dcr_destructive ||
6989 			    dtrace_destructive_disallow) {
6990 				void *activity = &state->dts_activity;
6991 				dtrace_activity_t current;
6992 
6993 				do {
6994 					current = state->dts_activity;
6995 				} while (dtrace_cas32(activity, current,
6996 				    DTRACE_ACTIVITY_KILLED) != current);
6997 
6998 				continue;
6999 			}
7000 		}
7001 
7002 		if ((offs = dtrace_buffer_reserve(buf, ecb->dte_needed,
7003 		    ecb->dte_alignment, state, &mstate)) < 0)
7004 			continue;
7005 
7006 		tomax = buf->dtb_tomax;
7007 		ASSERT(tomax != NULL);
7008 
7009 		if (ecb->dte_size != 0) {
7010 			dtrace_rechdr_t dtrh;
7011 			if (!(mstate.dtms_present & DTRACE_MSTATE_TIMESTAMP)) {
7012 				mstate.dtms_timestamp = dtrace_gethrtime();
7013 				mstate.dtms_present |= DTRACE_MSTATE_TIMESTAMP;
7014 			}
7015 			ASSERT3U(ecb->dte_size, >=, sizeof (dtrace_rechdr_t));
7016 			dtrh.dtrh_epid = ecb->dte_epid;
7017 			DTRACE_RECORD_STORE_TIMESTAMP(&dtrh,
7018 			    mstate.dtms_timestamp);
7019 			*((dtrace_rechdr_t *)(tomax + offs)) = dtrh;
7020 		}
7021 
7022 		mstate.dtms_epid = ecb->dte_epid;
7023 		mstate.dtms_present |= DTRACE_MSTATE_EPID;
7024 
7025 		if (state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)
7026 			mstate.dtms_access |= DTRACE_ACCESS_KERNEL;
7027 
7028 		if (pred != NULL) {
7029 			dtrace_difo_t *dp = pred->dtp_difo;
7030 			int rval;
7031 
7032 			rval = dtrace_dif_emulate(dp, &mstate, vstate, state);
7033 
7034 			if (!(*flags & CPU_DTRACE_ERROR) && !rval) {
7035 				dtrace_cacheid_t cid = probe->dtpr_predcache;
7036 
7037 				if (cid != DTRACE_CACHEIDNONE && !onintr) {
7038 					/*
7039 					 * Update the predicate cache...
7040 					 */
7041 					ASSERT(cid == pred->dtp_cacheid);
7042 					curthread->t_predcache = cid;
7043 				}
7044 
7045 				continue;
7046 			}
7047 		}
7048 
7049 		for (act = ecb->dte_action; !(*flags & CPU_DTRACE_ERROR) &&
7050 		    act != NULL; act = act->dta_next) {
7051 			size_t valoffs;
7052 			dtrace_difo_t *dp;
7053 			dtrace_recdesc_t *rec = &act->dta_rec;
7054 
7055 			size = rec->dtrd_size;
7056 			valoffs = offs + rec->dtrd_offset;
7057 
7058 			if (DTRACEACT_ISAGG(act->dta_kind)) {
7059 				uint64_t v = 0xbad;
7060 				dtrace_aggregation_t *agg;
7061 
7062 				agg = (dtrace_aggregation_t *)act;
7063 
7064 				if ((dp = act->dta_difo) != NULL)
7065 					v = dtrace_dif_emulate(dp,
7066 					    &mstate, vstate, state);
7067 
7068 				if (*flags & CPU_DTRACE_ERROR)
7069 					continue;
7070 
7071 				/*
7072 				 * Note that we always pass the expression
7073 				 * value from the previous iteration of the
7074 				 * action loop.  This value will only be used
7075 				 * if there is an expression argument to the
7076 				 * aggregating action, denoted by the
7077 				 * dtag_hasarg field.
7078 				 */
7079 				dtrace_aggregate(agg, buf,
7080 				    offs, aggbuf, v, val);
7081 				continue;
7082 			}
7083 
7084 			switch (act->dta_kind) {
7085 			case DTRACEACT_STOP:
7086 				if (dtrace_priv_proc_destructive(state,
7087 				    &mstate))
7088 					dtrace_action_stop();
7089 				continue;
7090 
7091 			case DTRACEACT_BREAKPOINT:
7092 				if (dtrace_priv_kernel_destructive(state))
7093 					dtrace_action_breakpoint(ecb);
7094 				continue;
7095 
7096 			case DTRACEACT_PANIC:
7097 				if (dtrace_priv_kernel_destructive(state))
7098 					dtrace_action_panic(ecb);
7099 				continue;
7100 
7101 			case DTRACEACT_STACK:
7102 				if (!dtrace_priv_kernel(state))
7103 					continue;
7104 
7105 				dtrace_getpcstack((pc_t *)(tomax + valoffs),
7106 				    size / sizeof (pc_t), probe->dtpr_aframes,
7107 				    DTRACE_ANCHORED(probe) ? NULL :
7108 				    (uint32_t *)arg0);
7109 
7110 				continue;
7111 
7112 			case DTRACEACT_JSTACK:
7113 			case DTRACEACT_USTACK:
7114 				if (!dtrace_priv_proc(state, &mstate))
7115 					continue;
7116 
7117 				/*
7118 				 * See comment in DIF_VAR_PID.
7119 				 */
7120 				if (DTRACE_ANCHORED(mstate.dtms_probe) &&
7121 				    CPU_ON_INTR(CPU)) {
7122 					int depth = DTRACE_USTACK_NFRAMES(
7123 					    rec->dtrd_arg) + 1;
7124 
7125 					dtrace_bzero((void *)(tomax + valoffs),
7126 					    DTRACE_USTACK_STRSIZE(rec->dtrd_arg)
7127 					    + depth * sizeof (uint64_t));
7128 
7129 					continue;
7130 				}
7131 
7132 				if (DTRACE_USTACK_STRSIZE(rec->dtrd_arg) != 0 &&
7133 				    curproc->p_dtrace_helpers != NULL) {
7134 					/*
7135 					 * This is the slow path -- we have
7136 					 * allocated string space, and we're
7137 					 * getting the stack of a process that
7138 					 * has helpers.  Call into a separate
7139 					 * routine to perform this processing.
7140 					 */
7141 					dtrace_action_ustack(&mstate, state,
7142 					    (uint64_t *)(tomax + valoffs),
7143 					    rec->dtrd_arg);
7144 					continue;
7145 				}
7146 
7147 				/*
7148 				 * Clear the string space, since there's no
7149 				 * helper to do it for us.
7150 				 */
7151 				if (DTRACE_USTACK_STRSIZE(rec->dtrd_arg) != 0) {
7152 					int depth = DTRACE_USTACK_NFRAMES(
7153 					    rec->dtrd_arg);
7154 					size_t strsize = DTRACE_USTACK_STRSIZE(
7155 					    rec->dtrd_arg);
7156 					uint64_t *buf = (uint64_t *)(tomax +
7157 					    valoffs);
7158 					void *strspace = &buf[depth + 1];
7159 
7160 					dtrace_bzero(strspace,
7161 					    MIN(depth, strsize));
7162 				}
7163 
7164 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
7165 				dtrace_getupcstack((uint64_t *)
7166 				    (tomax + valoffs),
7167 				    DTRACE_USTACK_NFRAMES(rec->dtrd_arg) + 1);
7168 				DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
7169 				continue;
7170 
7171 			default:
7172 				break;
7173 			}
7174 
7175 			dp = act->dta_difo;
7176 			ASSERT(dp != NULL);
7177 
7178 			val = dtrace_dif_emulate(dp, &mstate, vstate, state);
7179 
7180 			if (*flags & CPU_DTRACE_ERROR)
7181 				continue;
7182 
7183 			switch (act->dta_kind) {
7184 			case DTRACEACT_SPECULATE: {
7185 				dtrace_rechdr_t *dtrh;
7186 
7187 				ASSERT(buf == &state->dts_buffer[cpuid]);
7188 				buf = dtrace_speculation_buffer(state,
7189 				    cpuid, val);
7190 
7191 				if (buf == NULL) {
7192 					*flags |= CPU_DTRACE_DROP;
7193 					continue;
7194 				}
7195 
7196 				offs = dtrace_buffer_reserve(buf,
7197 				    ecb->dte_needed, ecb->dte_alignment,
7198 				    state, NULL);
7199 
7200 				if (offs < 0) {
7201 					*flags |= CPU_DTRACE_DROP;
7202 					continue;
7203 				}
7204 
7205 				tomax = buf->dtb_tomax;
7206 				ASSERT(tomax != NULL);
7207 
7208 				if (ecb->dte_size == 0)
7209 					continue;
7210 
7211 				ASSERT3U(ecb->dte_size, >=,
7212 				    sizeof (dtrace_rechdr_t));
7213 				dtrh = ((void *)(tomax + offs));
7214 				dtrh->dtrh_epid = ecb->dte_epid;
7215 				/*
7216 				 * When the speculation is committed, all of
7217 				 * the records in the speculative buffer will
7218 				 * have their timestamps set to the commit
7219 				 * time.  Until then, it is set to a sentinel
7220 				 * value, for debugability.
7221 				 */
7222 				DTRACE_RECORD_STORE_TIMESTAMP(dtrh, UINT64_MAX);
7223 				continue;
7224 			}
7225 
7226 			case DTRACEACT_CHILL:
7227 				if (dtrace_priv_kernel_destructive(state))
7228 					dtrace_action_chill(&mstate, val);
7229 				continue;
7230 
7231 			case DTRACEACT_RAISE:
7232 				if (dtrace_priv_proc_destructive(state,
7233 				    &mstate))
7234 					dtrace_action_raise(val);
7235 				continue;
7236 
7237 			case DTRACEACT_COMMIT:
7238 				ASSERT(!committed);
7239 
7240 				/*
7241 				 * We need to commit our buffer state.
7242 				 */
7243 				if (ecb->dte_size)
7244 					buf->dtb_offset = offs + ecb->dte_size;
7245 				buf = &state->dts_buffer[cpuid];
7246 				dtrace_speculation_commit(state, cpuid, val);
7247 				committed = 1;
7248 				continue;
7249 
7250 			case DTRACEACT_DISCARD:
7251 				dtrace_speculation_discard(state, cpuid, val);
7252 				continue;
7253 
7254 			case DTRACEACT_DIFEXPR:
7255 			case DTRACEACT_LIBACT:
7256 			case DTRACEACT_PRINTF:
7257 			case DTRACEACT_PRINTA:
7258 			case DTRACEACT_SYSTEM:
7259 			case DTRACEACT_FREOPEN:
7260 			case DTRACEACT_TRACEMEM:
7261 				break;
7262 
7263 			case DTRACEACT_TRACEMEM_DYNSIZE:
7264 				tracememsize = val;
7265 				break;
7266 
7267 			case DTRACEACT_SYM:
7268 			case DTRACEACT_MOD:
7269 				if (!dtrace_priv_kernel(state))
7270 					continue;
7271 				break;
7272 
7273 			case DTRACEACT_USYM:
7274 			case DTRACEACT_UMOD:
7275 			case DTRACEACT_UADDR: {
7276 				struct pid *pid = curthread->t_procp->p_pidp;
7277 
7278 				if (!dtrace_priv_proc(state, &mstate))
7279 					continue;
7280 
7281 				DTRACE_STORE(uint64_t, tomax,
7282 				    valoffs, (uint64_t)pid->pid_id);
7283 				DTRACE_STORE(uint64_t, tomax,
7284 				    valoffs + sizeof (uint64_t), val);
7285 
7286 				continue;
7287 			}
7288 
7289 			case DTRACEACT_EXIT: {
7290 				/*
7291 				 * For the exit action, we are going to attempt
7292 				 * to atomically set our activity to be
7293 				 * draining.  If this fails (either because
7294 				 * another CPU has beat us to the exit action,
7295 				 * or because our current activity is something
7296 				 * other than ACTIVE or WARMUP), we will
7297 				 * continue.  This assures that the exit action
7298 				 * can be successfully recorded at most once
7299 				 * when we're in the ACTIVE state.  If we're
7300 				 * encountering the exit() action while in
7301 				 * COOLDOWN, however, we want to honor the new
7302 				 * status code.  (We know that we're the only
7303 				 * thread in COOLDOWN, so there is no race.)
7304 				 */
7305 				void *activity = &state->dts_activity;
7306 				dtrace_activity_t current = state->dts_activity;
7307 
7308 				if (current == DTRACE_ACTIVITY_COOLDOWN)
7309 					break;
7310 
7311 				if (current != DTRACE_ACTIVITY_WARMUP)
7312 					current = DTRACE_ACTIVITY_ACTIVE;
7313 
7314 				if (dtrace_cas32(activity, current,
7315 				    DTRACE_ACTIVITY_DRAINING) != current) {
7316 					*flags |= CPU_DTRACE_DROP;
7317 					continue;
7318 				}
7319 
7320 				break;
7321 			}
7322 
7323 			default:
7324 				ASSERT(0);
7325 			}
7326 
7327 			if (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF ||
7328 			    dp->dtdo_rtype.dtdt_flags & DIF_TF_BYUREF) {
7329 				uintptr_t end = valoffs + size;
7330 
7331 				if (tracememsize != 0 &&
7332 				    valoffs + tracememsize < end) {
7333 					end = valoffs + tracememsize;
7334 					tracememsize = 0;
7335 				}
7336 
7337 				if (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF &&
7338 				    !dtrace_vcanload((void *)(uintptr_t)val,
7339 				    &dp->dtdo_rtype, NULL, &mstate, vstate))
7340 					continue;
7341 
7342 				dtrace_store_by_ref(dp, tomax, size, &valoffs,
7343 				    &val, end, act->dta_intuple,
7344 				    dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF ?
7345 				    DIF_TF_BYREF: DIF_TF_BYUREF);
7346 				continue;
7347 			}
7348 
7349 			switch (size) {
7350 			case 0:
7351 				break;
7352 
7353 			case sizeof (uint8_t):
7354 				DTRACE_STORE(uint8_t, tomax, valoffs, val);
7355 				break;
7356 			case sizeof (uint16_t):
7357 				DTRACE_STORE(uint16_t, tomax, valoffs, val);
7358 				break;
7359 			case sizeof (uint32_t):
7360 				DTRACE_STORE(uint32_t, tomax, valoffs, val);
7361 				break;
7362 			case sizeof (uint64_t):
7363 				DTRACE_STORE(uint64_t, tomax, valoffs, val);
7364 				break;
7365 			default:
7366 				/*
7367 				 * Any other size should have been returned by
7368 				 * reference, not by value.
7369 				 */
7370 				ASSERT(0);
7371 				break;
7372 			}
7373 		}
7374 
7375 		if (*flags & CPU_DTRACE_DROP)
7376 			continue;
7377 
7378 		if (*flags & CPU_DTRACE_FAULT) {
7379 			int ndx;
7380 			dtrace_action_t *err;
7381 
7382 			buf->dtb_errors++;
7383 
7384 			if (probe->dtpr_id == dtrace_probeid_error) {
7385 				/*
7386 				 * There's nothing we can do -- we had an
7387 				 * error on the error probe.  We bump an
7388 				 * error counter to at least indicate that
7389 				 * this condition happened.
7390 				 */
7391 				dtrace_error(&state->dts_dblerrors);
7392 				continue;
7393 			}
7394 
7395 			if (vtime) {
7396 				/*
7397 				 * Before recursing on dtrace_probe(), we
7398 				 * need to explicitly clear out our start
7399 				 * time to prevent it from being accumulated
7400 				 * into t_dtrace_vtime.
7401 				 */
7402 				curthread->t_dtrace_start = 0;
7403 			}
7404 
7405 			/*
7406 			 * Iterate over the actions to figure out which action
7407 			 * we were processing when we experienced the error.
7408 			 * Note that act points _past_ the faulting action; if
7409 			 * act is ecb->dte_action, the fault was in the
7410 			 * predicate, if it's ecb->dte_action->dta_next it's
7411 			 * in action #1, and so on.
7412 			 */
7413 			for (err = ecb->dte_action, ndx = 0;
7414 			    err != act; err = err->dta_next, ndx++)
7415 				continue;
7416 
7417 			dtrace_probe_error(state, ecb->dte_epid, ndx,
7418 			    (mstate.dtms_present & DTRACE_MSTATE_FLTOFFS) ?
7419 			    mstate.dtms_fltoffs : -1, DTRACE_FLAGS2FLT(*flags),
7420 			    cpu_core[cpuid].cpuc_dtrace_illval);
7421 
7422 			continue;
7423 		}
7424 
7425 		if (!committed)
7426 			buf->dtb_offset = offs + ecb->dte_size;
7427 	}
7428 
7429 	end = dtrace_gethrtime();
7430 	if (vtime)
7431 		curthread->t_dtrace_start = end;
7432 
7433 	CPU->cpu_dtrace_nsec += end - now;
7434 
7435 	dtrace_interrupt_enable(cookie);
7436 }
7437 
7438 /*
7439  * DTrace Probe Hashing Functions
7440  *
7441  * The functions in this section (and indeed, the functions in remaining
7442  * sections) are not _called_ from probe context.  (Any exceptions to this are
7443  * marked with a "Note:".)  Rather, they are called from elsewhere in the
7444  * DTrace framework to look-up probes in, add probes to and remove probes from
7445  * the DTrace probe hashes.  (Each probe is hashed by each element of the
7446  * probe tuple -- allowing for fast lookups, regardless of what was
7447  * specified.)
7448  */
7449 static uint_t
7450 dtrace_hash_str(char *p)
7451 {
7452 	unsigned int g;
7453 	uint_t hval = 0;
7454 
7455 	while (*p) {
7456 		hval = (hval << 4) + *p++;
7457 		if ((g = (hval & 0xf0000000)) != 0)
7458 			hval ^= g >> 24;
7459 		hval &= ~g;
7460 	}
7461 	return (hval);
7462 }
7463 
7464 static dtrace_hash_t *
7465 dtrace_hash_create(uintptr_t stroffs, uintptr_t nextoffs, uintptr_t prevoffs)
7466 {
7467 	dtrace_hash_t *hash = kmem_zalloc(sizeof (dtrace_hash_t), KM_SLEEP);
7468 
7469 	hash->dth_stroffs = stroffs;
7470 	hash->dth_nextoffs = nextoffs;
7471 	hash->dth_prevoffs = prevoffs;
7472 
7473 	hash->dth_size = 1;
7474 	hash->dth_mask = hash->dth_size - 1;
7475 
7476 	hash->dth_tab = kmem_zalloc(hash->dth_size *
7477 	    sizeof (dtrace_hashbucket_t *), KM_SLEEP);
7478 
7479 	return (hash);
7480 }
7481 
7482 static void
7483 dtrace_hash_destroy(dtrace_hash_t *hash)
7484 {
7485 #ifdef DEBUG
7486 	int i;
7487 
7488 	for (i = 0; i < hash->dth_size; i++)
7489 		ASSERT(hash->dth_tab[i] == NULL);
7490 #endif
7491 
7492 	kmem_free(hash->dth_tab,
7493 	    hash->dth_size * sizeof (dtrace_hashbucket_t *));
7494 	kmem_free(hash, sizeof (dtrace_hash_t));
7495 }
7496 
7497 static void
7498 dtrace_hash_resize(dtrace_hash_t *hash)
7499 {
7500 	int size = hash->dth_size, i, ndx;
7501 	int new_size = hash->dth_size << 1;
7502 	int new_mask = new_size - 1;
7503 	dtrace_hashbucket_t **new_tab, *bucket, *next;
7504 
7505 	ASSERT((new_size & new_mask) == 0);
7506 
7507 	new_tab = kmem_zalloc(new_size * sizeof (void *), KM_SLEEP);
7508 
7509 	for (i = 0; i < size; i++) {
7510 		for (bucket = hash->dth_tab[i]; bucket != NULL; bucket = next) {
7511 			dtrace_probe_t *probe = bucket->dthb_chain;
7512 
7513 			ASSERT(probe != NULL);
7514 			ndx = DTRACE_HASHSTR(hash, probe) & new_mask;
7515 
7516 			next = bucket->dthb_next;
7517 			bucket->dthb_next = new_tab[ndx];
7518 			new_tab[ndx] = bucket;
7519 		}
7520 	}
7521 
7522 	kmem_free(hash->dth_tab, hash->dth_size * sizeof (void *));
7523 	hash->dth_tab = new_tab;
7524 	hash->dth_size = new_size;
7525 	hash->dth_mask = new_mask;
7526 }
7527 
7528 static void
7529 dtrace_hash_add(dtrace_hash_t *hash, dtrace_probe_t *new)
7530 {
7531 	int hashval = DTRACE_HASHSTR(hash, new);
7532 	int ndx = hashval & hash->dth_mask;
7533 	dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
7534 	dtrace_probe_t **nextp, **prevp;
7535 
7536 	for (; bucket != NULL; bucket = bucket->dthb_next) {
7537 		if (DTRACE_HASHEQ(hash, bucket->dthb_chain, new))
7538 			goto add;
7539 	}
7540 
7541 	if ((hash->dth_nbuckets >> 1) > hash->dth_size) {
7542 		dtrace_hash_resize(hash);
7543 		dtrace_hash_add(hash, new);
7544 		return;
7545 	}
7546 
7547 	bucket = kmem_zalloc(sizeof (dtrace_hashbucket_t), KM_SLEEP);
7548 	bucket->dthb_next = hash->dth_tab[ndx];
7549 	hash->dth_tab[ndx] = bucket;
7550 	hash->dth_nbuckets++;
7551 
7552 add:
7553 	nextp = DTRACE_HASHNEXT(hash, new);
7554 	ASSERT(*nextp == NULL && *(DTRACE_HASHPREV(hash, new)) == NULL);
7555 	*nextp = bucket->dthb_chain;
7556 
7557 	if (bucket->dthb_chain != NULL) {
7558 		prevp = DTRACE_HASHPREV(hash, bucket->dthb_chain);
7559 		ASSERT(*prevp == NULL);
7560 		*prevp = new;
7561 	}
7562 
7563 	bucket->dthb_chain = new;
7564 	bucket->dthb_len++;
7565 }
7566 
7567 static dtrace_probe_t *
7568 dtrace_hash_lookup(dtrace_hash_t *hash, dtrace_probe_t *template)
7569 {
7570 	int hashval = DTRACE_HASHSTR(hash, template);
7571 	int ndx = hashval & hash->dth_mask;
7572 	dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
7573 
7574 	for (; bucket != NULL; bucket = bucket->dthb_next) {
7575 		if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
7576 			return (bucket->dthb_chain);
7577 	}
7578 
7579 	return (NULL);
7580 }
7581 
7582 static int
7583 dtrace_hash_collisions(dtrace_hash_t *hash, dtrace_probe_t *template)
7584 {
7585 	int hashval = DTRACE_HASHSTR(hash, template);
7586 	int ndx = hashval & hash->dth_mask;
7587 	dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
7588 
7589 	for (; bucket != NULL; bucket = bucket->dthb_next) {
7590 		if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
7591 			return (bucket->dthb_len);
7592 	}
7593 
7594 	return (NULL);
7595 }
7596 
7597 static void
7598 dtrace_hash_remove(dtrace_hash_t *hash, dtrace_probe_t *probe)
7599 {
7600 	int ndx = DTRACE_HASHSTR(hash, probe) & hash->dth_mask;
7601 	dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
7602 
7603 	dtrace_probe_t **prevp = DTRACE_HASHPREV(hash, probe);
7604 	dtrace_probe_t **nextp = DTRACE_HASHNEXT(hash, probe);
7605 
7606 	/*
7607 	 * Find the bucket that we're removing this probe from.
7608 	 */
7609 	for (; bucket != NULL; bucket = bucket->dthb_next) {
7610 		if (DTRACE_HASHEQ(hash, bucket->dthb_chain, probe))
7611 			break;
7612 	}
7613 
7614 	ASSERT(bucket != NULL);
7615 
7616 	if (*prevp == NULL) {
7617 		if (*nextp == NULL) {
7618 			/*
7619 			 * The removed probe was the only probe on this
7620 			 * bucket; we need to remove the bucket.
7621 			 */
7622 			dtrace_hashbucket_t *b = hash->dth_tab[ndx];
7623 
7624 			ASSERT(bucket->dthb_chain == probe);
7625 			ASSERT(b != NULL);
7626 
7627 			if (b == bucket) {
7628 				hash->dth_tab[ndx] = bucket->dthb_next;
7629 			} else {
7630 				while (b->dthb_next != bucket)
7631 					b = b->dthb_next;
7632 				b->dthb_next = bucket->dthb_next;
7633 			}
7634 
7635 			ASSERT(hash->dth_nbuckets > 0);
7636 			hash->dth_nbuckets--;
7637 			kmem_free(bucket, sizeof (dtrace_hashbucket_t));
7638 			return;
7639 		}
7640 
7641 		bucket->dthb_chain = *nextp;
7642 	} else {
7643 		*(DTRACE_HASHNEXT(hash, *prevp)) = *nextp;
7644 	}
7645 
7646 	if (*nextp != NULL)
7647 		*(DTRACE_HASHPREV(hash, *nextp)) = *prevp;
7648 }
7649 
7650 /*
7651  * DTrace Utility Functions
7652  *
7653  * These are random utility functions that are _not_ called from probe context.
7654  */
7655 static int
7656 dtrace_badattr(const dtrace_attribute_t *a)
7657 {
7658 	return (a->dtat_name > DTRACE_STABILITY_MAX ||
7659 	    a->dtat_data > DTRACE_STABILITY_MAX ||
7660 	    a->dtat_class > DTRACE_CLASS_MAX);
7661 }
7662 
7663 /*
7664  * Return a duplicate copy of a string.  If the specified string is NULL,
7665  * this function returns a zero-length string.
7666  */
7667 static char *
7668 dtrace_strdup(const char *str)
7669 {
7670 	char *new = kmem_zalloc((str != NULL ? strlen(str) : 0) + 1, KM_SLEEP);
7671 
7672 	if (str != NULL)
7673 		(void) strcpy(new, str);
7674 
7675 	return (new);
7676 }
7677 
7678 #define	DTRACE_ISALPHA(c)	\
7679 	(((c) >= 'a' && (c) <= 'z') || ((c) >= 'A' && (c) <= 'Z'))
7680 
7681 static int
7682 dtrace_badname(const char *s)
7683 {
7684 	char c;
7685 
7686 	if (s == NULL || (c = *s++) == '\0')
7687 		return (0);
7688 
7689 	if (!DTRACE_ISALPHA(c) && c != '-' && c != '_' && c != '.')
7690 		return (1);
7691 
7692 	while ((c = *s++) != '\0') {
7693 		if (!DTRACE_ISALPHA(c) && (c < '0' || c > '9') &&
7694 		    c != '-' && c != '_' && c != '.' && c != '`')
7695 			return (1);
7696 	}
7697 
7698 	return (0);
7699 }
7700 
7701 static void
7702 dtrace_cred2priv(cred_t *cr, uint32_t *privp, uid_t *uidp, zoneid_t *zoneidp)
7703 {
7704 	uint32_t priv;
7705 
7706 	if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
7707 		/*
7708 		 * For DTRACE_PRIV_ALL, the uid and zoneid don't matter.
7709 		 */
7710 		priv = DTRACE_PRIV_ALL;
7711 	} else {
7712 		*uidp = crgetuid(cr);
7713 		*zoneidp = crgetzoneid(cr);
7714 
7715 		priv = 0;
7716 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE))
7717 			priv |= DTRACE_PRIV_KERNEL | DTRACE_PRIV_USER;
7718 		else if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE))
7719 			priv |= DTRACE_PRIV_USER;
7720 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE))
7721 			priv |= DTRACE_PRIV_PROC;
7722 		if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
7723 			priv |= DTRACE_PRIV_OWNER;
7724 		if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
7725 			priv |= DTRACE_PRIV_ZONEOWNER;
7726 	}
7727 
7728 	*privp = priv;
7729 }
7730 
7731 #ifdef DTRACE_ERRDEBUG
7732 static void
7733 dtrace_errdebug(const char *str)
7734 {
7735 	int hval = dtrace_hash_str((char *)str) % DTRACE_ERRHASHSZ;
7736 	int occupied = 0;
7737 
7738 	mutex_enter(&dtrace_errlock);
7739 	dtrace_errlast = str;
7740 	dtrace_errthread = curthread;
7741 
7742 	while (occupied++ < DTRACE_ERRHASHSZ) {
7743 		if (dtrace_errhash[hval].dter_msg == str) {
7744 			dtrace_errhash[hval].dter_count++;
7745 			goto out;
7746 		}
7747 
7748 		if (dtrace_errhash[hval].dter_msg != NULL) {
7749 			hval = (hval + 1) % DTRACE_ERRHASHSZ;
7750 			continue;
7751 		}
7752 
7753 		dtrace_errhash[hval].dter_msg = str;
7754 		dtrace_errhash[hval].dter_count = 1;
7755 		goto out;
7756 	}
7757 
7758 	panic("dtrace: undersized error hash");
7759 out:
7760 	mutex_exit(&dtrace_errlock);
7761 }
7762 #endif
7763 
7764 /*
7765  * DTrace Matching Functions
7766  *
7767  * These functions are used to match groups of probes, given some elements of
7768  * a probe tuple, or some globbed expressions for elements of a probe tuple.
7769  */
7770 static int
7771 dtrace_match_priv(const dtrace_probe_t *prp, uint32_t priv, uid_t uid,
7772     zoneid_t zoneid)
7773 {
7774 	if (priv != DTRACE_PRIV_ALL) {
7775 		uint32_t ppriv = prp->dtpr_provider->dtpv_priv.dtpp_flags;
7776 		uint32_t match = priv & ppriv;
7777 
7778 		/*
7779 		 * No PRIV_DTRACE_* privileges...
7780 		 */
7781 		if ((priv & (DTRACE_PRIV_PROC | DTRACE_PRIV_USER |
7782 		    DTRACE_PRIV_KERNEL)) == 0)
7783 			return (0);
7784 
7785 		/*
7786 		 * No matching bits, but there were bits to match...
7787 		 */
7788 		if (match == 0 && ppriv != 0)
7789 			return (0);
7790 
7791 		/*
7792 		 * Need to have permissions to the process, but don't...
7793 		 */
7794 		if (((ppriv & ~match) & DTRACE_PRIV_OWNER) != 0 &&
7795 		    uid != prp->dtpr_provider->dtpv_priv.dtpp_uid) {
7796 			return (0);
7797 		}
7798 
7799 		/*
7800 		 * Need to be in the same zone unless we possess the
7801 		 * privilege to examine all zones.
7802 		 */
7803 		if (((ppriv & ~match) & DTRACE_PRIV_ZONEOWNER) != 0 &&
7804 		    zoneid != prp->dtpr_provider->dtpv_priv.dtpp_zoneid) {
7805 			return (0);
7806 		}
7807 	}
7808 
7809 	return (1);
7810 }
7811 
7812 /*
7813  * dtrace_match_probe compares a dtrace_probe_t to a pre-compiled key, which
7814  * consists of input pattern strings and an ops-vector to evaluate them.
7815  * This function returns >0 for match, 0 for no match, and <0 for error.
7816  */
7817 static int
7818 dtrace_match_probe(const dtrace_probe_t *prp, const dtrace_probekey_t *pkp,
7819     uint32_t priv, uid_t uid, zoneid_t zoneid)
7820 {
7821 	dtrace_provider_t *pvp = prp->dtpr_provider;
7822 	int rv;
7823 
7824 	if (pvp->dtpv_defunct)
7825 		return (0);
7826 
7827 	if ((rv = pkp->dtpk_pmatch(pvp->dtpv_name, pkp->dtpk_prov, 0)) <= 0)
7828 		return (rv);
7829 
7830 	if ((rv = pkp->dtpk_mmatch(prp->dtpr_mod, pkp->dtpk_mod, 0)) <= 0)
7831 		return (rv);
7832 
7833 	if ((rv = pkp->dtpk_fmatch(prp->dtpr_func, pkp->dtpk_func, 0)) <= 0)
7834 		return (rv);
7835 
7836 	if ((rv = pkp->dtpk_nmatch(prp->dtpr_name, pkp->dtpk_name, 0)) <= 0)
7837 		return (rv);
7838 
7839 	if (dtrace_match_priv(prp, priv, uid, zoneid) == 0)
7840 		return (0);
7841 
7842 	return (rv);
7843 }
7844 
7845 /*
7846  * dtrace_match_glob() is a safe kernel implementation of the gmatch(3GEN)
7847  * interface for matching a glob pattern 'p' to an input string 's'.  Unlike
7848  * libc's version, the kernel version only applies to 8-bit ASCII strings.
7849  * In addition, all of the recursion cases except for '*' matching have been
7850  * unwound.  For '*', we still implement recursive evaluation, but a depth
7851  * counter is maintained and matching is aborted if we recurse too deep.
7852  * The function returns 0 if no match, >0 if match, and <0 if recursion error.
7853  */
7854 static int
7855 dtrace_match_glob(const char *s, const char *p, int depth)
7856 {
7857 	const char *olds;
7858 	char s1, c;
7859 	int gs;
7860 
7861 	if (depth > DTRACE_PROBEKEY_MAXDEPTH)
7862 		return (-1);
7863 
7864 	if (s == NULL)
7865 		s = ""; /* treat NULL as empty string */
7866 
7867 top:
7868 	olds = s;
7869 	s1 = *s++;
7870 
7871 	if (p == NULL)
7872 		return (0);
7873 
7874 	if ((c = *p++) == '\0')
7875 		return (s1 == '\0');
7876 
7877 	switch (c) {
7878 	case '[': {
7879 		int ok = 0, notflag = 0;
7880 		char lc = '\0';
7881 
7882 		if (s1 == '\0')
7883 			return (0);
7884 
7885 		if (*p == '!') {
7886 			notflag = 1;
7887 			p++;
7888 		}
7889 
7890 		if ((c = *p++) == '\0')
7891 			return (0);
7892 
7893 		do {
7894 			if (c == '-' && lc != '\0' && *p != ']') {
7895 				if ((c = *p++) == '\0')
7896 					return (0);
7897 				if (c == '\\' && (c = *p++) == '\0')
7898 					return (0);
7899 
7900 				if (notflag) {
7901 					if (s1 < lc || s1 > c)
7902 						ok++;
7903 					else
7904 						return (0);
7905 				} else if (lc <= s1 && s1 <= c)
7906 					ok++;
7907 
7908 			} else if (c == '\\' && (c = *p++) == '\0')
7909 				return (0);
7910 
7911 			lc = c; /* save left-hand 'c' for next iteration */
7912 
7913 			if (notflag) {
7914 				if (s1 != c)
7915 					ok++;
7916 				else
7917 					return (0);
7918 			} else if (s1 == c)
7919 				ok++;
7920 
7921 			if ((c = *p++) == '\0')
7922 				return (0);
7923 
7924 		} while (c != ']');
7925 
7926 		if (ok)
7927 			goto top;
7928 
7929 		return (0);
7930 	}
7931 
7932 	case '\\':
7933 		if ((c = *p++) == '\0')
7934 			return (0);
7935 		/*FALLTHRU*/
7936 
7937 	default:
7938 		if (c != s1)
7939 			return (0);
7940 		/*FALLTHRU*/
7941 
7942 	case '?':
7943 		if (s1 != '\0')
7944 			goto top;
7945 		return (0);
7946 
7947 	case '*':
7948 		while (*p == '*')
7949 			p++; /* consecutive *'s are identical to a single one */
7950 
7951 		if (*p == '\0')
7952 			return (1);
7953 
7954 		for (s = olds; *s != '\0'; s++) {
7955 			if ((gs = dtrace_match_glob(s, p, depth + 1)) != 0)
7956 				return (gs);
7957 		}
7958 
7959 		return (0);
7960 	}
7961 }
7962 
7963 /*ARGSUSED*/
7964 static int
7965 dtrace_match_string(const char *s, const char *p, int depth)
7966 {
7967 	return (s != NULL && strcmp(s, p) == 0);
7968 }
7969 
7970 /*ARGSUSED*/
7971 static int
7972 dtrace_match_nul(const char *s, const char *p, int depth)
7973 {
7974 	return (1); /* always match the empty pattern */
7975 }
7976 
7977 /*ARGSUSED*/
7978 static int
7979 dtrace_match_nonzero(const char *s, const char *p, int depth)
7980 {
7981 	return (s != NULL && s[0] != '\0');
7982 }
7983 
7984 static int
7985 dtrace_match(const dtrace_probekey_t *pkp, uint32_t priv, uid_t uid,
7986     zoneid_t zoneid, int (*matched)(dtrace_probe_t *, void *), void *arg)
7987 {
7988 	dtrace_probe_t template, *probe;
7989 	dtrace_hash_t *hash = NULL;
7990 	int len, rc, best = INT_MAX, nmatched = 0;
7991 	dtrace_id_t i;
7992 
7993 	ASSERT(MUTEX_HELD(&dtrace_lock));
7994 
7995 	/*
7996 	 * If the probe ID is specified in the key, just lookup by ID and
7997 	 * invoke the match callback once if a matching probe is found.
7998 	 */
7999 	if (pkp->dtpk_id != DTRACE_IDNONE) {
8000 		if ((probe = dtrace_probe_lookup_id(pkp->dtpk_id)) != NULL &&
8001 		    dtrace_match_probe(probe, pkp, priv, uid, zoneid) > 0) {
8002 			if ((*matched)(probe, arg) == DTRACE_MATCH_FAIL)
8003 				return (DTRACE_MATCH_FAIL);
8004 			nmatched++;
8005 		}
8006 		return (nmatched);
8007 	}
8008 
8009 	template.dtpr_mod = (char *)pkp->dtpk_mod;
8010 	template.dtpr_func = (char *)pkp->dtpk_func;
8011 	template.dtpr_name = (char *)pkp->dtpk_name;
8012 
8013 	/*
8014 	 * We want to find the most distinct of the module name, function
8015 	 * name, and name.  So for each one that is not a glob pattern or
8016 	 * empty string, we perform a lookup in the corresponding hash and
8017 	 * use the hash table with the fewest collisions to do our search.
8018 	 */
8019 	if (pkp->dtpk_mmatch == &dtrace_match_string &&
8020 	    (len = dtrace_hash_collisions(dtrace_bymod, &template)) < best) {
8021 		best = len;
8022 		hash = dtrace_bymod;
8023 	}
8024 
8025 	if (pkp->dtpk_fmatch == &dtrace_match_string &&
8026 	    (len = dtrace_hash_collisions(dtrace_byfunc, &template)) < best) {
8027 		best = len;
8028 		hash = dtrace_byfunc;
8029 	}
8030 
8031 	if (pkp->dtpk_nmatch == &dtrace_match_string &&
8032 	    (len = dtrace_hash_collisions(dtrace_byname, &template)) < best) {
8033 		best = len;
8034 		hash = dtrace_byname;
8035 	}
8036 
8037 	/*
8038 	 * If we did not select a hash table, iterate over every probe and
8039 	 * invoke our callback for each one that matches our input probe key.
8040 	 */
8041 	if (hash == NULL) {
8042 		for (i = 0; i < dtrace_nprobes; i++) {
8043 			if ((probe = dtrace_probes[i]) == NULL ||
8044 			    dtrace_match_probe(probe, pkp, priv, uid,
8045 			    zoneid) <= 0)
8046 				continue;
8047 
8048 			nmatched++;
8049 
8050 			if ((rc = (*matched)(probe, arg)) !=
8051 			    DTRACE_MATCH_NEXT) {
8052 				if (rc == DTRACE_MATCH_FAIL)
8053 					return (DTRACE_MATCH_FAIL);
8054 				break;
8055 			}
8056 		}
8057 
8058 		return (nmatched);
8059 	}
8060 
8061 	/*
8062 	 * If we selected a hash table, iterate over each probe of the same key
8063 	 * name and invoke the callback for every probe that matches the other
8064 	 * attributes of our input probe key.
8065 	 */
8066 	for (probe = dtrace_hash_lookup(hash, &template); probe != NULL;
8067 	    probe = *(DTRACE_HASHNEXT(hash, probe))) {
8068 
8069 		if (dtrace_match_probe(probe, pkp, priv, uid, zoneid) <= 0)
8070 			continue;
8071 
8072 		nmatched++;
8073 
8074 		if ((rc = (*matched)(probe, arg)) != DTRACE_MATCH_NEXT) {
8075 			if (rc == DTRACE_MATCH_FAIL)
8076 				return (DTRACE_MATCH_FAIL);
8077 			break;
8078 		}
8079 	}
8080 
8081 	return (nmatched);
8082 }
8083 
8084 /*
8085  * Return the function pointer dtrace_probecmp() should use to compare the
8086  * specified pattern with a string.  For NULL or empty patterns, we select
8087  * dtrace_match_nul().  For glob pattern strings, we use dtrace_match_glob().
8088  * For non-empty non-glob strings, we use dtrace_match_string().
8089  */
8090 static dtrace_probekey_f *
8091 dtrace_probekey_func(const char *p)
8092 {
8093 	char c;
8094 
8095 	if (p == NULL || *p == '\0')
8096 		return (&dtrace_match_nul);
8097 
8098 	while ((c = *p++) != '\0') {
8099 		if (c == '[' || c == '?' || c == '*' || c == '\\')
8100 			return (&dtrace_match_glob);
8101 	}
8102 
8103 	return (&dtrace_match_string);
8104 }
8105 
8106 /*
8107  * Build a probe comparison key for use with dtrace_match_probe() from the
8108  * given probe description.  By convention, a null key only matches anchored
8109  * probes: if each field is the empty string, reset dtpk_fmatch to
8110  * dtrace_match_nonzero().
8111  */
8112 static void
8113 dtrace_probekey(const dtrace_probedesc_t *pdp, dtrace_probekey_t *pkp)
8114 {
8115 	pkp->dtpk_prov = pdp->dtpd_provider;
8116 	pkp->dtpk_pmatch = dtrace_probekey_func(pdp->dtpd_provider);
8117 
8118 	pkp->dtpk_mod = pdp->dtpd_mod;
8119 	pkp->dtpk_mmatch = dtrace_probekey_func(pdp->dtpd_mod);
8120 
8121 	pkp->dtpk_func = pdp->dtpd_func;
8122 	pkp->dtpk_fmatch = dtrace_probekey_func(pdp->dtpd_func);
8123 
8124 	pkp->dtpk_name = pdp->dtpd_name;
8125 	pkp->dtpk_nmatch = dtrace_probekey_func(pdp->dtpd_name);
8126 
8127 	pkp->dtpk_id = pdp->dtpd_id;
8128 
8129 	if (pkp->dtpk_id == DTRACE_IDNONE &&
8130 	    pkp->dtpk_pmatch == &dtrace_match_nul &&
8131 	    pkp->dtpk_mmatch == &dtrace_match_nul &&
8132 	    pkp->dtpk_fmatch == &dtrace_match_nul &&
8133 	    pkp->dtpk_nmatch == &dtrace_match_nul)
8134 		pkp->dtpk_fmatch = &dtrace_match_nonzero;
8135 }
8136 
8137 /*
8138  * DTrace Provider-to-Framework API Functions
8139  *
8140  * These functions implement much of the Provider-to-Framework API, as
8141  * described in <sys/dtrace.h>.  The parts of the API not in this section are
8142  * the functions in the API for probe management (found below), and
8143  * dtrace_probe() itself (found above).
8144  */
8145 
8146 /*
8147  * Register the calling provider with the DTrace framework.  This should
8148  * generally be called by DTrace providers in their attach(9E) entry point.
8149  */
8150 int
8151 dtrace_register(const char *name, const dtrace_pattr_t *pap, uint32_t priv,
8152     cred_t *cr, const dtrace_pops_t *pops, void *arg, dtrace_provider_id_t *idp)
8153 {
8154 	dtrace_provider_t *provider;
8155 
8156 	if (name == NULL || pap == NULL || pops == NULL || idp == NULL) {
8157 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
8158 		    "arguments", name ? name : "<NULL>");
8159 		return (EINVAL);
8160 	}
8161 
8162 	if (name[0] == '\0' || dtrace_badname(name)) {
8163 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
8164 		    "provider name", name);
8165 		return (EINVAL);
8166 	}
8167 
8168 	if ((pops->dtps_provide == NULL && pops->dtps_provide_module == NULL) ||
8169 	    pops->dtps_enable == NULL || pops->dtps_disable == NULL ||
8170 	    pops->dtps_destroy == NULL ||
8171 	    ((pops->dtps_resume == NULL) != (pops->dtps_suspend == NULL))) {
8172 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
8173 		    "provider ops", name);
8174 		return (EINVAL);
8175 	}
8176 
8177 	if (dtrace_badattr(&pap->dtpa_provider) ||
8178 	    dtrace_badattr(&pap->dtpa_mod) ||
8179 	    dtrace_badattr(&pap->dtpa_func) ||
8180 	    dtrace_badattr(&pap->dtpa_name) ||
8181 	    dtrace_badattr(&pap->dtpa_args)) {
8182 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
8183 		    "provider attributes", name);
8184 		return (EINVAL);
8185 	}
8186 
8187 	if (priv & ~DTRACE_PRIV_ALL) {
8188 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
8189 		    "privilege attributes", name);
8190 		return (EINVAL);
8191 	}
8192 
8193 	if ((priv & DTRACE_PRIV_KERNEL) &&
8194 	    (priv & (DTRACE_PRIV_USER | DTRACE_PRIV_OWNER)) &&
8195 	    pops->dtps_mode == NULL) {
8196 		cmn_err(CE_WARN, "failed to register provider '%s': need "
8197 		    "dtps_mode() op for given privilege attributes", name);
8198 		return (EINVAL);
8199 	}
8200 
8201 	provider = kmem_zalloc(sizeof (dtrace_provider_t), KM_SLEEP);
8202 	provider->dtpv_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
8203 	(void) strcpy(provider->dtpv_name, name);
8204 
8205 	provider->dtpv_attr = *pap;
8206 	provider->dtpv_priv.dtpp_flags = priv;
8207 	if (cr != NULL) {
8208 		provider->dtpv_priv.dtpp_uid = crgetuid(cr);
8209 		provider->dtpv_priv.dtpp_zoneid = crgetzoneid(cr);
8210 	}
8211 	provider->dtpv_pops = *pops;
8212 
8213 	if (pops->dtps_provide == NULL) {
8214 		ASSERT(pops->dtps_provide_module != NULL);
8215 		provider->dtpv_pops.dtps_provide =
8216 		    (void (*)(void *, const dtrace_probedesc_t *))dtrace_nullop;
8217 	}
8218 
8219 	if (pops->dtps_provide_module == NULL) {
8220 		ASSERT(pops->dtps_provide != NULL);
8221 		provider->dtpv_pops.dtps_provide_module =
8222 		    (void (*)(void *, struct modctl *))dtrace_nullop;
8223 	}
8224 
8225 	if (pops->dtps_suspend == NULL) {
8226 		ASSERT(pops->dtps_resume == NULL);
8227 		provider->dtpv_pops.dtps_suspend =
8228 		    (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
8229 		provider->dtpv_pops.dtps_resume =
8230 		    (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
8231 	}
8232 
8233 	provider->dtpv_arg = arg;
8234 	*idp = (dtrace_provider_id_t)provider;
8235 
8236 	if (pops == &dtrace_provider_ops) {
8237 		ASSERT(MUTEX_HELD(&dtrace_provider_lock));
8238 		ASSERT(MUTEX_HELD(&dtrace_lock));
8239 		ASSERT(dtrace_anon.dta_enabling == NULL);
8240 
8241 		/*
8242 		 * We make sure that the DTrace provider is at the head of
8243 		 * the provider chain.
8244 		 */
8245 		provider->dtpv_next = dtrace_provider;
8246 		dtrace_provider = provider;
8247 		return (0);
8248 	}
8249 
8250 	mutex_enter(&dtrace_provider_lock);
8251 	mutex_enter(&dtrace_lock);
8252 
8253 	/*
8254 	 * If there is at least one provider registered, we'll add this
8255 	 * provider after the first provider.
8256 	 */
8257 	if (dtrace_provider != NULL) {
8258 		provider->dtpv_next = dtrace_provider->dtpv_next;
8259 		dtrace_provider->dtpv_next = provider;
8260 	} else {
8261 		dtrace_provider = provider;
8262 	}
8263 
8264 	if (dtrace_retained != NULL) {
8265 		dtrace_enabling_provide(provider);
8266 
8267 		/*
8268 		 * Now we need to call dtrace_enabling_matchall() -- which
8269 		 * will acquire cpu_lock and dtrace_lock.  We therefore need
8270 		 * to drop all of our locks before calling into it...
8271 		 */
8272 		mutex_exit(&dtrace_lock);
8273 		mutex_exit(&dtrace_provider_lock);
8274 		dtrace_enabling_matchall();
8275 
8276 		return (0);
8277 	}
8278 
8279 	mutex_exit(&dtrace_lock);
8280 	mutex_exit(&dtrace_provider_lock);
8281 
8282 	return (0);
8283 }
8284 
8285 /*
8286  * Unregister the specified provider from the DTrace framework.  This should
8287  * generally be called by DTrace providers in their detach(9E) entry point.
8288  */
8289 int
8290 dtrace_unregister(dtrace_provider_id_t id)
8291 {
8292 	dtrace_provider_t *old = (dtrace_provider_t *)id;
8293 	dtrace_provider_t *prev = NULL;
8294 	int i, self = 0, noreap = 0;
8295 	dtrace_probe_t *probe, *first = NULL;
8296 
8297 	if (old->dtpv_pops.dtps_enable ==
8298 	    (int (*)(void *, dtrace_id_t, void *))dtrace_enable_nullop) {
8299 		/*
8300 		 * If DTrace itself is the provider, we're called with locks
8301 		 * already held.
8302 		 */
8303 		ASSERT(old == dtrace_provider);
8304 		ASSERT(dtrace_devi != NULL);
8305 		ASSERT(MUTEX_HELD(&dtrace_provider_lock));
8306 		ASSERT(MUTEX_HELD(&dtrace_lock));
8307 		self = 1;
8308 
8309 		if (dtrace_provider->dtpv_next != NULL) {
8310 			/*
8311 			 * There's another provider here; return failure.
8312 			 */
8313 			return (EBUSY);
8314 		}
8315 	} else {
8316 		mutex_enter(&dtrace_provider_lock);
8317 		mutex_enter(&mod_lock);
8318 		mutex_enter(&dtrace_lock);
8319 	}
8320 
8321 	/*
8322 	 * If anyone has /dev/dtrace open, or if there are anonymous enabled
8323 	 * probes, we refuse to let providers slither away, unless this
8324 	 * provider has already been explicitly invalidated.
8325 	 */
8326 	if (!old->dtpv_defunct &&
8327 	    (dtrace_opens || (dtrace_anon.dta_state != NULL &&
8328 	    dtrace_anon.dta_state->dts_necbs > 0))) {
8329 		if (!self) {
8330 			mutex_exit(&dtrace_lock);
8331 			mutex_exit(&mod_lock);
8332 			mutex_exit(&dtrace_provider_lock);
8333 		}
8334 		return (EBUSY);
8335 	}
8336 
8337 	/*
8338 	 * Attempt to destroy the probes associated with this provider.
8339 	 */
8340 	for (i = 0; i < dtrace_nprobes; i++) {
8341 		if ((probe = dtrace_probes[i]) == NULL)
8342 			continue;
8343 
8344 		if (probe->dtpr_provider != old)
8345 			continue;
8346 
8347 		if (probe->dtpr_ecb == NULL)
8348 			continue;
8349 
8350 		/*
8351 		 * If we are trying to unregister a defunct provider, and the
8352 		 * provider was made defunct within the interval dictated by
8353 		 * dtrace_unregister_defunct_reap, we'll (asynchronously)
8354 		 * attempt to reap our enablings.  To denote that the provider
8355 		 * should reattempt to unregister itself at some point in the
8356 		 * future, we will return a differentiable error code (EAGAIN
8357 		 * instead of EBUSY) in this case.
8358 		 */
8359 		if (dtrace_gethrtime() - old->dtpv_defunct >
8360 		    dtrace_unregister_defunct_reap)
8361 			noreap = 1;
8362 
8363 		if (!self) {
8364 			mutex_exit(&dtrace_lock);
8365 			mutex_exit(&mod_lock);
8366 			mutex_exit(&dtrace_provider_lock);
8367 		}
8368 
8369 		if (noreap)
8370 			return (EBUSY);
8371 
8372 		(void) taskq_dispatch(dtrace_taskq,
8373 		    (task_func_t *)dtrace_enabling_reap, NULL, TQ_SLEEP);
8374 
8375 		return (EAGAIN);
8376 	}
8377 
8378 	/*
8379 	 * All of the probes for this provider are disabled; we can safely
8380 	 * remove all of them from their hash chains and from the probe array.
8381 	 */
8382 	for (i = 0; i < dtrace_nprobes; i++) {
8383 		if ((probe = dtrace_probes[i]) == NULL)
8384 			continue;
8385 
8386 		if (probe->dtpr_provider != old)
8387 			continue;
8388 
8389 		dtrace_probes[i] = NULL;
8390 
8391 		dtrace_hash_remove(dtrace_bymod, probe);
8392 		dtrace_hash_remove(dtrace_byfunc, probe);
8393 		dtrace_hash_remove(dtrace_byname, probe);
8394 
8395 		if (first == NULL) {
8396 			first = probe;
8397 			probe->dtpr_nextmod = NULL;
8398 		} else {
8399 			probe->dtpr_nextmod = first;
8400 			first = probe;
8401 		}
8402 	}
8403 
8404 	/*
8405 	 * The provider's probes have been removed from the hash chains and
8406 	 * from the probe array.  Now issue a dtrace_sync() to be sure that
8407 	 * everyone has cleared out from any probe array processing.
8408 	 */
8409 	dtrace_sync();
8410 
8411 	for (probe = first; probe != NULL; probe = first) {
8412 		first = probe->dtpr_nextmod;
8413 
8414 		old->dtpv_pops.dtps_destroy(old->dtpv_arg, probe->dtpr_id,
8415 		    probe->dtpr_arg);
8416 		kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
8417 		kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
8418 		kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
8419 		vmem_free(dtrace_arena, (void *)(uintptr_t)(probe->dtpr_id), 1);
8420 		kmem_free(probe, sizeof (dtrace_probe_t));
8421 	}
8422 
8423 	if ((prev = dtrace_provider) == old) {
8424 		ASSERT(self || dtrace_devi == NULL);
8425 		ASSERT(old->dtpv_next == NULL || dtrace_devi == NULL);
8426 		dtrace_provider = old->dtpv_next;
8427 	} else {
8428 		while (prev != NULL && prev->dtpv_next != old)
8429 			prev = prev->dtpv_next;
8430 
8431 		if (prev == NULL) {
8432 			panic("attempt to unregister non-existent "
8433 			    "dtrace provider %p\n", (void *)id);
8434 		}
8435 
8436 		prev->dtpv_next = old->dtpv_next;
8437 	}
8438 
8439 	if (!self) {
8440 		mutex_exit(&dtrace_lock);
8441 		mutex_exit(&mod_lock);
8442 		mutex_exit(&dtrace_provider_lock);
8443 	}
8444 
8445 	kmem_free(old->dtpv_name, strlen(old->dtpv_name) + 1);
8446 	kmem_free(old, sizeof (dtrace_provider_t));
8447 
8448 	return (0);
8449 }
8450 
8451 /*
8452  * Invalidate the specified provider.  All subsequent probe lookups for the
8453  * specified provider will fail, but its probes will not be removed.
8454  */
8455 void
8456 dtrace_invalidate(dtrace_provider_id_t id)
8457 {
8458 	dtrace_provider_t *pvp = (dtrace_provider_t *)id;
8459 
8460 	ASSERT(pvp->dtpv_pops.dtps_enable !=
8461 	    (int (*)(void *, dtrace_id_t, void *))dtrace_enable_nullop);
8462 
8463 	mutex_enter(&dtrace_provider_lock);
8464 	mutex_enter(&dtrace_lock);
8465 
8466 	pvp->dtpv_defunct = dtrace_gethrtime();
8467 
8468 	mutex_exit(&dtrace_lock);
8469 	mutex_exit(&dtrace_provider_lock);
8470 }
8471 
8472 /*
8473  * Indicate whether or not DTrace has attached.
8474  */
8475 int
8476 dtrace_attached(void)
8477 {
8478 	/*
8479 	 * dtrace_provider will be non-NULL iff the DTrace driver has
8480 	 * attached.  (It's non-NULL because DTrace is always itself a
8481 	 * provider.)
8482 	 */
8483 	return (dtrace_provider != NULL);
8484 }
8485 
8486 /*
8487  * Remove all the unenabled probes for the given provider.  This function is
8488  * not unlike dtrace_unregister(), except that it doesn't remove the provider
8489  * -- just as many of its associated probes as it can.
8490  */
8491 int
8492 dtrace_condense(dtrace_provider_id_t id)
8493 {
8494 	dtrace_provider_t *prov = (dtrace_provider_t *)id;
8495 	int i;
8496 	dtrace_probe_t *probe;
8497 
8498 	/*
8499 	 * Make sure this isn't the dtrace provider itself.
8500 	 */
8501 	ASSERT(prov->dtpv_pops.dtps_enable !=
8502 	    (int (*)(void *, dtrace_id_t, void *))dtrace_enable_nullop);
8503 
8504 	mutex_enter(&dtrace_provider_lock);
8505 	mutex_enter(&dtrace_lock);
8506 
8507 	/*
8508 	 * Attempt to destroy the probes associated with this provider.
8509 	 */
8510 	for (i = 0; i < dtrace_nprobes; i++) {
8511 		if ((probe = dtrace_probes[i]) == NULL)
8512 			continue;
8513 
8514 		if (probe->dtpr_provider != prov)
8515 			continue;
8516 
8517 		if (probe->dtpr_ecb != NULL)
8518 			continue;
8519 
8520 		dtrace_probes[i] = NULL;
8521 
8522 		dtrace_hash_remove(dtrace_bymod, probe);
8523 		dtrace_hash_remove(dtrace_byfunc, probe);
8524 		dtrace_hash_remove(dtrace_byname, probe);
8525 
8526 		prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, i + 1,
8527 		    probe->dtpr_arg);
8528 		kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
8529 		kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
8530 		kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
8531 		kmem_free(probe, sizeof (dtrace_probe_t));
8532 		vmem_free(dtrace_arena, (void *)((uintptr_t)i + 1), 1);
8533 	}
8534 
8535 	mutex_exit(&dtrace_lock);
8536 	mutex_exit(&dtrace_provider_lock);
8537 
8538 	return (0);
8539 }
8540 
8541 /*
8542  * DTrace Probe Management Functions
8543  *
8544  * The functions in this section perform the DTrace probe management,
8545  * including functions to create probes, look-up probes, and call into the
8546  * providers to request that probes be provided.  Some of these functions are
8547  * in the Provider-to-Framework API; these functions can be identified by the
8548  * fact that they are not declared "static".
8549  */
8550 
8551 /*
8552  * Create a probe with the specified module name, function name, and name.
8553  */
8554 dtrace_id_t
8555 dtrace_probe_create(dtrace_provider_id_t prov, const char *mod,
8556     const char *func, const char *name, int aframes, void *arg)
8557 {
8558 	dtrace_probe_t *probe, **probes;
8559 	dtrace_provider_t *provider = (dtrace_provider_t *)prov;
8560 	dtrace_id_t id;
8561 
8562 	if (provider == dtrace_provider) {
8563 		ASSERT(MUTEX_HELD(&dtrace_lock));
8564 	} else {
8565 		mutex_enter(&dtrace_lock);
8566 	}
8567 
8568 	id = (dtrace_id_t)(uintptr_t)vmem_alloc(dtrace_arena, 1,
8569 	    VM_BESTFIT | VM_SLEEP);
8570 	probe = kmem_zalloc(sizeof (dtrace_probe_t), KM_SLEEP);
8571 
8572 	probe->dtpr_id = id;
8573 	probe->dtpr_gen = dtrace_probegen++;
8574 	probe->dtpr_mod = dtrace_strdup(mod);
8575 	probe->dtpr_func = dtrace_strdup(func);
8576 	probe->dtpr_name = dtrace_strdup(name);
8577 	probe->dtpr_arg = arg;
8578 	probe->dtpr_aframes = aframes;
8579 	probe->dtpr_provider = provider;
8580 
8581 	dtrace_hash_add(dtrace_bymod, probe);
8582 	dtrace_hash_add(dtrace_byfunc, probe);
8583 	dtrace_hash_add(dtrace_byname, probe);
8584 
8585 	if (id - 1 >= dtrace_nprobes) {
8586 		size_t osize = dtrace_nprobes * sizeof (dtrace_probe_t *);
8587 		size_t nsize = osize << 1;
8588 
8589 		if (nsize == 0) {
8590 			ASSERT(osize == 0);
8591 			ASSERT(dtrace_probes == NULL);
8592 			nsize = sizeof (dtrace_probe_t *);
8593 		}
8594 
8595 		probes = kmem_zalloc(nsize, KM_SLEEP);
8596 
8597 		if (dtrace_probes == NULL) {
8598 			ASSERT(osize == 0);
8599 			dtrace_probes = probes;
8600 			dtrace_nprobes = 1;
8601 		} else {
8602 			dtrace_probe_t **oprobes = dtrace_probes;
8603 
8604 			bcopy(oprobes, probes, osize);
8605 			dtrace_membar_producer();
8606 			dtrace_probes = probes;
8607 
8608 			dtrace_sync();
8609 
8610 			/*
8611 			 * All CPUs are now seeing the new probes array; we can
8612 			 * safely free the old array.
8613 			 */
8614 			kmem_free(oprobes, osize);
8615 			dtrace_nprobes <<= 1;
8616 		}
8617 
8618 		ASSERT(id - 1 < dtrace_nprobes);
8619 	}
8620 
8621 	ASSERT(dtrace_probes[id - 1] == NULL);
8622 	dtrace_probes[id - 1] = probe;
8623 
8624 	if (provider != dtrace_provider)
8625 		mutex_exit(&dtrace_lock);
8626 
8627 	return (id);
8628 }
8629 
8630 static dtrace_probe_t *
8631 dtrace_probe_lookup_id(dtrace_id_t id)
8632 {
8633 	ASSERT(MUTEX_HELD(&dtrace_lock));
8634 
8635 	if (id == 0 || id > dtrace_nprobes)
8636 		return (NULL);
8637 
8638 	return (dtrace_probes[id - 1]);
8639 }
8640 
8641 static int
8642 dtrace_probe_lookup_match(dtrace_probe_t *probe, void *arg)
8643 {
8644 	*((dtrace_id_t *)arg) = probe->dtpr_id;
8645 
8646 	return (DTRACE_MATCH_DONE);
8647 }
8648 
8649 /*
8650  * Look up a probe based on provider and one or more of module name, function
8651  * name and probe name.
8652  */
8653 dtrace_id_t
8654 dtrace_probe_lookup(dtrace_provider_id_t prid, const char *mod,
8655     const char *func, const char *name)
8656 {
8657 	dtrace_probekey_t pkey;
8658 	dtrace_id_t id;
8659 	int match;
8660 
8661 	pkey.dtpk_prov = ((dtrace_provider_t *)prid)->dtpv_name;
8662 	pkey.dtpk_pmatch = &dtrace_match_string;
8663 	pkey.dtpk_mod = mod;
8664 	pkey.dtpk_mmatch = mod ? &dtrace_match_string : &dtrace_match_nul;
8665 	pkey.dtpk_func = func;
8666 	pkey.dtpk_fmatch = func ? &dtrace_match_string : &dtrace_match_nul;
8667 	pkey.dtpk_name = name;
8668 	pkey.dtpk_nmatch = name ? &dtrace_match_string : &dtrace_match_nul;
8669 	pkey.dtpk_id = DTRACE_IDNONE;
8670 
8671 	mutex_enter(&dtrace_lock);
8672 	match = dtrace_match(&pkey, DTRACE_PRIV_ALL, 0, 0,
8673 	    dtrace_probe_lookup_match, &id);
8674 	mutex_exit(&dtrace_lock);
8675 
8676 	ASSERT(match == 1 || match == 0);
8677 	return (match ? id : 0);
8678 }
8679 
8680 /*
8681  * Returns the probe argument associated with the specified probe.
8682  */
8683 void *
8684 dtrace_probe_arg(dtrace_provider_id_t id, dtrace_id_t pid)
8685 {
8686 	dtrace_probe_t *probe;
8687 	void *rval = NULL;
8688 
8689 	mutex_enter(&dtrace_lock);
8690 
8691 	if ((probe = dtrace_probe_lookup_id(pid)) != NULL &&
8692 	    probe->dtpr_provider == (dtrace_provider_t *)id)
8693 		rval = probe->dtpr_arg;
8694 
8695 	mutex_exit(&dtrace_lock);
8696 
8697 	return (rval);
8698 }
8699 
8700 /*
8701  * Copy a probe into a probe description.
8702  */
8703 static void
8704 dtrace_probe_description(const dtrace_probe_t *prp, dtrace_probedesc_t *pdp)
8705 {
8706 	bzero(pdp, sizeof (dtrace_probedesc_t));
8707 	pdp->dtpd_id = prp->dtpr_id;
8708 
8709 	(void) strncpy(pdp->dtpd_provider,
8710 	    prp->dtpr_provider->dtpv_name, DTRACE_PROVNAMELEN - 1);
8711 
8712 	(void) strncpy(pdp->dtpd_mod, prp->dtpr_mod, DTRACE_MODNAMELEN - 1);
8713 	(void) strncpy(pdp->dtpd_func, prp->dtpr_func, DTRACE_FUNCNAMELEN - 1);
8714 	(void) strncpy(pdp->dtpd_name, prp->dtpr_name, DTRACE_NAMELEN - 1);
8715 }
8716 
8717 /*
8718  * Called to indicate that a probe -- or probes -- should be provided by a
8719  * specfied provider.  If the specified description is NULL, the provider will
8720  * be told to provide all of its probes.  (This is done whenever a new
8721  * consumer comes along, or whenever a retained enabling is to be matched.) If
8722  * the specified description is non-NULL, the provider is given the
8723  * opportunity to dynamically provide the specified probe, allowing providers
8724  * to support the creation of probes on-the-fly.  (So-called _autocreated_
8725  * probes.)  If the provider is NULL, the operations will be applied to all
8726  * providers; if the provider is non-NULL the operations will only be applied
8727  * to the specified provider.  The dtrace_provider_lock must be held, and the
8728  * dtrace_lock must _not_ be held -- the provider's dtps_provide() operation
8729  * will need to grab the dtrace_lock when it reenters the framework through
8730  * dtrace_probe_lookup(), dtrace_probe_create(), etc.
8731  */
8732 static void
8733 dtrace_probe_provide(dtrace_probedesc_t *desc, dtrace_provider_t *prv)
8734 {
8735 	struct modctl *ctl;
8736 	int all = 0;
8737 
8738 	ASSERT(MUTEX_HELD(&dtrace_provider_lock));
8739 
8740 	if (prv == NULL) {
8741 		all = 1;
8742 		prv = dtrace_provider;
8743 	}
8744 
8745 	do {
8746 		/*
8747 		 * First, call the blanket provide operation.
8748 		 */
8749 		prv->dtpv_pops.dtps_provide(prv->dtpv_arg, desc);
8750 
8751 		/*
8752 		 * Now call the per-module provide operation.  We will grab
8753 		 * mod_lock to prevent the list from being modified.  Note
8754 		 * that this also prevents the mod_busy bits from changing.
8755 		 * (mod_busy can only be changed with mod_lock held.)
8756 		 */
8757 		mutex_enter(&mod_lock);
8758 
8759 		ctl = &modules;
8760 		do {
8761 			if (ctl->mod_busy || ctl->mod_mp == NULL)
8762 				continue;
8763 
8764 			prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
8765 
8766 		} while ((ctl = ctl->mod_next) != &modules);
8767 
8768 		mutex_exit(&mod_lock);
8769 	} while (all && (prv = prv->dtpv_next) != NULL);
8770 }
8771 
8772 /*
8773  * Iterate over each probe, and call the Framework-to-Provider API function
8774  * denoted by offs.
8775  */
8776 static void
8777 dtrace_probe_foreach(uintptr_t offs)
8778 {
8779 	dtrace_provider_t *prov;
8780 	void (*func)(void *, dtrace_id_t, void *);
8781 	dtrace_probe_t *probe;
8782 	dtrace_icookie_t cookie;
8783 	int i;
8784 
8785 	/*
8786 	 * We disable interrupts to walk through the probe array.  This is
8787 	 * safe -- the dtrace_sync() in dtrace_unregister() assures that we
8788 	 * won't see stale data.
8789 	 */
8790 	cookie = dtrace_interrupt_disable();
8791 
8792 	for (i = 0; i < dtrace_nprobes; i++) {
8793 		if ((probe = dtrace_probes[i]) == NULL)
8794 			continue;
8795 
8796 		if (probe->dtpr_ecb == NULL) {
8797 			/*
8798 			 * This probe isn't enabled -- don't call the function.
8799 			 */
8800 			continue;
8801 		}
8802 
8803 		prov = probe->dtpr_provider;
8804 		func = *((void(**)(void *, dtrace_id_t, void *))
8805 		    ((uintptr_t)&prov->dtpv_pops + offs));
8806 
8807 		func(prov->dtpv_arg, i + 1, probe->dtpr_arg);
8808 	}
8809 
8810 	dtrace_interrupt_enable(cookie);
8811 }
8812 
8813 static int
8814 dtrace_probe_enable(const dtrace_probedesc_t *desc, dtrace_enabling_t *enab)
8815 {
8816 	dtrace_probekey_t pkey;
8817 	uint32_t priv;
8818 	uid_t uid;
8819 	zoneid_t zoneid;
8820 
8821 	ASSERT(MUTEX_HELD(&dtrace_lock));
8822 	dtrace_ecb_create_cache = NULL;
8823 
8824 	if (desc == NULL) {
8825 		/*
8826 		 * If we're passed a NULL description, we're being asked to
8827 		 * create an ECB with a NULL probe.
8828 		 */
8829 		(void) dtrace_ecb_create_enable(NULL, enab);
8830 		return (0);
8831 	}
8832 
8833 	dtrace_probekey(desc, &pkey);
8834 	dtrace_cred2priv(enab->dten_vstate->dtvs_state->dts_cred.dcr_cred,
8835 	    &priv, &uid, &zoneid);
8836 
8837 	return (dtrace_match(&pkey, priv, uid, zoneid, dtrace_ecb_create_enable,
8838 	    enab));
8839 }
8840 
8841 /*
8842  * DTrace Helper Provider Functions
8843  */
8844 static void
8845 dtrace_dofattr2attr(dtrace_attribute_t *attr, const dof_attr_t dofattr)
8846 {
8847 	attr->dtat_name = DOF_ATTR_NAME(dofattr);
8848 	attr->dtat_data = DOF_ATTR_DATA(dofattr);
8849 	attr->dtat_class = DOF_ATTR_CLASS(dofattr);
8850 }
8851 
8852 static void
8853 dtrace_dofprov2hprov(dtrace_helper_provdesc_t *hprov,
8854     const dof_provider_t *dofprov, char *strtab)
8855 {
8856 	hprov->dthpv_provname = strtab + dofprov->dofpv_name;
8857 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_provider,
8858 	    dofprov->dofpv_provattr);
8859 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_mod,
8860 	    dofprov->dofpv_modattr);
8861 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_func,
8862 	    dofprov->dofpv_funcattr);
8863 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_name,
8864 	    dofprov->dofpv_nameattr);
8865 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_args,
8866 	    dofprov->dofpv_argsattr);
8867 }
8868 
8869 static void
8870 dtrace_helper_provide_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
8871 {
8872 	uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8873 	dof_hdr_t *dof = (dof_hdr_t *)daddr;
8874 	dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
8875 	dof_provider_t *provider;
8876 	dof_probe_t *probe;
8877 	uint32_t *off, *enoff;
8878 	uint8_t *arg;
8879 	char *strtab;
8880 	uint_t i, nprobes;
8881 	dtrace_helper_provdesc_t dhpv;
8882 	dtrace_helper_probedesc_t dhpb;
8883 	dtrace_meta_t *meta = dtrace_meta_pid;
8884 	dtrace_mops_t *mops = &meta->dtm_mops;
8885 	void *parg;
8886 
8887 	provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
8888 	str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8889 	    provider->dofpv_strtab * dof->dofh_secsize);
8890 	prb_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8891 	    provider->dofpv_probes * dof->dofh_secsize);
8892 	arg_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8893 	    provider->dofpv_prargs * dof->dofh_secsize);
8894 	off_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8895 	    provider->dofpv_proffs * dof->dofh_secsize);
8896 
8897 	strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
8898 	off = (uint32_t *)(uintptr_t)(daddr + off_sec->dofs_offset);
8899 	arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
8900 	enoff = NULL;
8901 
8902 	/*
8903 	 * See dtrace_helper_provider_validate().
8904 	 */
8905 	if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
8906 	    provider->dofpv_prenoffs != DOF_SECT_NONE) {
8907 		enoff_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8908 		    provider->dofpv_prenoffs * dof->dofh_secsize);
8909 		enoff = (uint32_t *)(uintptr_t)(daddr + enoff_sec->dofs_offset);
8910 	}
8911 
8912 	nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
8913 
8914 	/*
8915 	 * Create the provider.
8916 	 */
8917 	dtrace_dofprov2hprov(&dhpv, provider, strtab);
8918 
8919 	if ((parg = mops->dtms_provide_pid(meta->dtm_arg, &dhpv, pid)) == NULL)
8920 		return;
8921 
8922 	meta->dtm_count++;
8923 
8924 	/*
8925 	 * Create the probes.
8926 	 */
8927 	for (i = 0; i < nprobes; i++) {
8928 		probe = (dof_probe_t *)(uintptr_t)(daddr +
8929 		    prb_sec->dofs_offset + i * prb_sec->dofs_entsize);
8930 
8931 		dhpb.dthpb_mod = dhp->dofhp_mod;
8932 		dhpb.dthpb_func = strtab + probe->dofpr_func;
8933 		dhpb.dthpb_name = strtab + probe->dofpr_name;
8934 		dhpb.dthpb_base = probe->dofpr_addr;
8935 		dhpb.dthpb_offs = off + probe->dofpr_offidx;
8936 		dhpb.dthpb_noffs = probe->dofpr_noffs;
8937 		if (enoff != NULL) {
8938 			dhpb.dthpb_enoffs = enoff + probe->dofpr_enoffidx;
8939 			dhpb.dthpb_nenoffs = probe->dofpr_nenoffs;
8940 		} else {
8941 			dhpb.dthpb_enoffs = NULL;
8942 			dhpb.dthpb_nenoffs = 0;
8943 		}
8944 		dhpb.dthpb_args = arg + probe->dofpr_argidx;
8945 		dhpb.dthpb_nargc = probe->dofpr_nargc;
8946 		dhpb.dthpb_xargc = probe->dofpr_xargc;
8947 		dhpb.dthpb_ntypes = strtab + probe->dofpr_nargv;
8948 		dhpb.dthpb_xtypes = strtab + probe->dofpr_xargv;
8949 
8950 		mops->dtms_create_probe(meta->dtm_arg, parg, &dhpb);
8951 	}
8952 }
8953 
8954 static void
8955 dtrace_helper_provide(dof_helper_t *dhp, pid_t pid)
8956 {
8957 	uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8958 	dof_hdr_t *dof = (dof_hdr_t *)daddr;
8959 	int i;
8960 
8961 	ASSERT(MUTEX_HELD(&dtrace_meta_lock));
8962 
8963 	for (i = 0; i < dof->dofh_secnum; i++) {
8964 		dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
8965 		    dof->dofh_secoff + i * dof->dofh_secsize);
8966 
8967 		if (sec->dofs_type != DOF_SECT_PROVIDER)
8968 			continue;
8969 
8970 		dtrace_helper_provide_one(dhp, sec, pid);
8971 	}
8972 
8973 	/*
8974 	 * We may have just created probes, so we must now rematch against
8975 	 * any retained enablings.  Note that this call will acquire both
8976 	 * cpu_lock and dtrace_lock; the fact that we are holding
8977 	 * dtrace_meta_lock now is what defines the ordering with respect to
8978 	 * these three locks.
8979 	 */
8980 	dtrace_enabling_matchall();
8981 }
8982 
8983 static void
8984 dtrace_helper_provider_remove_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
8985 {
8986 	uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8987 	dof_hdr_t *dof = (dof_hdr_t *)daddr;
8988 	dof_sec_t *str_sec;
8989 	dof_provider_t *provider;
8990 	char *strtab;
8991 	dtrace_helper_provdesc_t dhpv;
8992 	dtrace_meta_t *meta = dtrace_meta_pid;
8993 	dtrace_mops_t *mops = &meta->dtm_mops;
8994 
8995 	provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
8996 	str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8997 	    provider->dofpv_strtab * dof->dofh_secsize);
8998 
8999 	strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
9000 
9001 	/*
9002 	 * Create the provider.
9003 	 */
9004 	dtrace_dofprov2hprov(&dhpv, provider, strtab);
9005 
9006 	mops->dtms_remove_pid(meta->dtm_arg, &dhpv, pid);
9007 
9008 	meta->dtm_count--;
9009 }
9010 
9011 static void
9012 dtrace_helper_provider_remove(dof_helper_t *dhp, pid_t pid)
9013 {
9014 	uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
9015 	dof_hdr_t *dof = (dof_hdr_t *)daddr;
9016 	int i;
9017 
9018 	ASSERT(MUTEX_HELD(&dtrace_meta_lock));
9019 
9020 	for (i = 0; i < dof->dofh_secnum; i++) {
9021 		dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
9022 		    dof->dofh_secoff + i * dof->dofh_secsize);
9023 
9024 		if (sec->dofs_type != DOF_SECT_PROVIDER)
9025 			continue;
9026 
9027 		dtrace_helper_provider_remove_one(dhp, sec, pid);
9028 	}
9029 }
9030 
9031 /*
9032  * DTrace Meta Provider-to-Framework API Functions
9033  *
9034  * These functions implement the Meta Provider-to-Framework API, as described
9035  * in <sys/dtrace.h>.
9036  */
9037 int
9038 dtrace_meta_register(const char *name, const dtrace_mops_t *mops, void *arg,
9039     dtrace_meta_provider_id_t *idp)
9040 {
9041 	dtrace_meta_t *meta;
9042 	dtrace_helpers_t *help, *next;
9043 	int i;
9044 
9045 	*idp = DTRACE_METAPROVNONE;
9046 
9047 	/*
9048 	 * We strictly don't need the name, but we hold onto it for
9049 	 * debuggability. All hail error queues!
9050 	 */
9051 	if (name == NULL) {
9052 		cmn_err(CE_WARN, "failed to register meta-provider: "
9053 		    "invalid name");
9054 		return (EINVAL);
9055 	}
9056 
9057 	if (mops == NULL ||
9058 	    mops->dtms_create_probe == NULL ||
9059 	    mops->dtms_provide_pid == NULL ||
9060 	    mops->dtms_remove_pid == NULL) {
9061 		cmn_err(CE_WARN, "failed to register meta-register %s: "
9062 		    "invalid ops", name);
9063 		return (EINVAL);
9064 	}
9065 
9066 	meta = kmem_zalloc(sizeof (dtrace_meta_t), KM_SLEEP);
9067 	meta->dtm_mops = *mops;
9068 	meta->dtm_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
9069 	(void) strcpy(meta->dtm_name, name);
9070 	meta->dtm_arg = arg;
9071 
9072 	mutex_enter(&dtrace_meta_lock);
9073 	mutex_enter(&dtrace_lock);
9074 
9075 	if (dtrace_meta_pid != NULL) {
9076 		mutex_exit(&dtrace_lock);
9077 		mutex_exit(&dtrace_meta_lock);
9078 		cmn_err(CE_WARN, "failed to register meta-register %s: "
9079 		    "user-land meta-provider exists", name);
9080 		kmem_free(meta->dtm_name, strlen(meta->dtm_name) + 1);
9081 		kmem_free(meta, sizeof (dtrace_meta_t));
9082 		return (EINVAL);
9083 	}
9084 
9085 	dtrace_meta_pid = meta;
9086 	*idp = (dtrace_meta_provider_id_t)meta;
9087 
9088 	/*
9089 	 * If there are providers and probes ready to go, pass them
9090 	 * off to the new meta provider now.
9091 	 */
9092 
9093 	help = dtrace_deferred_pid;
9094 	dtrace_deferred_pid = NULL;
9095 
9096 	mutex_exit(&dtrace_lock);
9097 
9098 	while (help != NULL) {
9099 		for (i = 0; i < help->dthps_nprovs; i++) {
9100 			dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
9101 			    help->dthps_pid);
9102 		}
9103 
9104 		next = help->dthps_next;
9105 		help->dthps_next = NULL;
9106 		help->dthps_prev = NULL;
9107 		help->dthps_deferred = 0;
9108 		help = next;
9109 	}
9110 
9111 	mutex_exit(&dtrace_meta_lock);
9112 
9113 	return (0);
9114 }
9115 
9116 int
9117 dtrace_meta_unregister(dtrace_meta_provider_id_t id)
9118 {
9119 	dtrace_meta_t **pp, *old = (dtrace_meta_t *)id;
9120 
9121 	mutex_enter(&dtrace_meta_lock);
9122 	mutex_enter(&dtrace_lock);
9123 
9124 	if (old == dtrace_meta_pid) {
9125 		pp = &dtrace_meta_pid;
9126 	} else {
9127 		panic("attempt to unregister non-existent "
9128 		    "dtrace meta-provider %p\n", (void *)old);
9129 	}
9130 
9131 	if (old->dtm_count != 0) {
9132 		mutex_exit(&dtrace_lock);
9133 		mutex_exit(&dtrace_meta_lock);
9134 		return (EBUSY);
9135 	}
9136 
9137 	*pp = NULL;
9138 
9139 	mutex_exit(&dtrace_lock);
9140 	mutex_exit(&dtrace_meta_lock);
9141 
9142 	kmem_free(old->dtm_name, strlen(old->dtm_name) + 1);
9143 	kmem_free(old, sizeof (dtrace_meta_t));
9144 
9145 	return (0);
9146 }
9147 
9148 
9149 /*
9150  * DTrace DIF Object Functions
9151  */
9152 static int
9153 dtrace_difo_err(uint_t pc, const char *format, ...)
9154 {
9155 	if (dtrace_err_verbose) {
9156 		va_list alist;
9157 
9158 		(void) uprintf("dtrace DIF object error: [%u]: ", pc);
9159 		va_start(alist, format);
9160 		(void) vuprintf(format, alist);
9161 		va_end(alist);
9162 	}
9163 
9164 #ifdef DTRACE_ERRDEBUG
9165 	dtrace_errdebug(format);
9166 #endif
9167 	return (1);
9168 }
9169 
9170 /*
9171  * Validate a DTrace DIF object by checking the IR instructions.  The following
9172  * rules are currently enforced by dtrace_difo_validate():
9173  *
9174  * 1. Each instruction must have a valid opcode
9175  * 2. Each register, string, variable, or subroutine reference must be valid
9176  * 3. No instruction can modify register %r0 (must be zero)
9177  * 4. All instruction reserved bits must be set to zero
9178  * 5. The last instruction must be a "ret" instruction
9179  * 6. All branch targets must reference a valid instruction _after_ the branch
9180  */
9181 static int
9182 dtrace_difo_validate(dtrace_difo_t *dp, dtrace_vstate_t *vstate, uint_t nregs,
9183     cred_t *cr)
9184 {
9185 	int err = 0, i;
9186 	int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
9187 	int kcheckload;
9188 	uint_t pc;
9189 	int maxglobal = -1, maxlocal = -1, maxtlocal = -1;
9190 
9191 	kcheckload = cr == NULL ||
9192 	    (vstate->dtvs_state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) == 0;
9193 
9194 	dp->dtdo_destructive = 0;
9195 
9196 	for (pc = 0; pc < dp->dtdo_len && err == 0; pc++) {
9197 		dif_instr_t instr = dp->dtdo_buf[pc];
9198 
9199 		uint_t r1 = DIF_INSTR_R1(instr);
9200 		uint_t r2 = DIF_INSTR_R2(instr);
9201 		uint_t rd = DIF_INSTR_RD(instr);
9202 		uint_t rs = DIF_INSTR_RS(instr);
9203 		uint_t label = DIF_INSTR_LABEL(instr);
9204 		uint_t v = DIF_INSTR_VAR(instr);
9205 		uint_t subr = DIF_INSTR_SUBR(instr);
9206 		uint_t type = DIF_INSTR_TYPE(instr);
9207 		uint_t op = DIF_INSTR_OP(instr);
9208 
9209 		switch (op) {
9210 		case DIF_OP_OR:
9211 		case DIF_OP_XOR:
9212 		case DIF_OP_AND:
9213 		case DIF_OP_SLL:
9214 		case DIF_OP_SRL:
9215 		case DIF_OP_SRA:
9216 		case DIF_OP_SUB:
9217 		case DIF_OP_ADD:
9218 		case DIF_OP_MUL:
9219 		case DIF_OP_SDIV:
9220 		case DIF_OP_UDIV:
9221 		case DIF_OP_SREM:
9222 		case DIF_OP_UREM:
9223 		case DIF_OP_COPYS:
9224 			if (r1 >= nregs)
9225 				err += efunc(pc, "invalid register %u\n", r1);
9226 			if (r2 >= nregs)
9227 				err += efunc(pc, "invalid register %u\n", r2);
9228 			if (rd >= nregs)
9229 				err += efunc(pc, "invalid register %u\n", rd);
9230 			if (rd == 0)
9231 				err += efunc(pc, "cannot write to %r0\n");
9232 			break;
9233 		case DIF_OP_NOT:
9234 		case DIF_OP_MOV:
9235 		case DIF_OP_ALLOCS:
9236 			if (r1 >= nregs)
9237 				err += efunc(pc, "invalid register %u\n", r1);
9238 			if (r2 != 0)
9239 				err += efunc(pc, "non-zero reserved bits\n");
9240 			if (rd >= nregs)
9241 				err += efunc(pc, "invalid register %u\n", rd);
9242 			if (rd == 0)
9243 				err += efunc(pc, "cannot write to %r0\n");
9244 			break;
9245 		case DIF_OP_LDSB:
9246 		case DIF_OP_LDSH:
9247 		case DIF_OP_LDSW:
9248 		case DIF_OP_LDUB:
9249 		case DIF_OP_LDUH:
9250 		case DIF_OP_LDUW:
9251 		case DIF_OP_LDX:
9252 			if (r1 >= nregs)
9253 				err += efunc(pc, "invalid register %u\n", r1);
9254 			if (r2 != 0)
9255 				err += efunc(pc, "non-zero reserved bits\n");
9256 			if (rd >= nregs)
9257 				err += efunc(pc, "invalid register %u\n", rd);
9258 			if (rd == 0)
9259 				err += efunc(pc, "cannot write to %r0\n");
9260 			if (kcheckload)
9261 				dp->dtdo_buf[pc] = DIF_INSTR_LOAD(op +
9262 				    DIF_OP_RLDSB - DIF_OP_LDSB, r1, rd);
9263 			break;
9264 		case DIF_OP_RLDSB:
9265 		case DIF_OP_RLDSH:
9266 		case DIF_OP_RLDSW:
9267 		case DIF_OP_RLDUB:
9268 		case DIF_OP_RLDUH:
9269 		case DIF_OP_RLDUW:
9270 		case DIF_OP_RLDX:
9271 			if (r1 >= nregs)
9272 				err += efunc(pc, "invalid register %u\n", r1);
9273 			if (r2 != 0)
9274 				err += efunc(pc, "non-zero reserved bits\n");
9275 			if (rd >= nregs)
9276 				err += efunc(pc, "invalid register %u\n", rd);
9277 			if (rd == 0)
9278 				err += efunc(pc, "cannot write to %r0\n");
9279 			break;
9280 		case DIF_OP_ULDSB:
9281 		case DIF_OP_ULDSH:
9282 		case DIF_OP_ULDSW:
9283 		case DIF_OP_ULDUB:
9284 		case DIF_OP_ULDUH:
9285 		case DIF_OP_ULDUW:
9286 		case DIF_OP_ULDX:
9287 			if (r1 >= nregs)
9288 				err += efunc(pc, "invalid register %u\n", r1);
9289 			if (r2 != 0)
9290 				err += efunc(pc, "non-zero reserved bits\n");
9291 			if (rd >= nregs)
9292 				err += efunc(pc, "invalid register %u\n", rd);
9293 			if (rd == 0)
9294 				err += efunc(pc, "cannot write to %r0\n");
9295 			break;
9296 		case DIF_OP_STB:
9297 		case DIF_OP_STH:
9298 		case DIF_OP_STW:
9299 		case DIF_OP_STX:
9300 			if (r1 >= nregs)
9301 				err += efunc(pc, "invalid register %u\n", r1);
9302 			if (r2 != 0)
9303 				err += efunc(pc, "non-zero reserved bits\n");
9304 			if (rd >= nregs)
9305 				err += efunc(pc, "invalid register %u\n", rd);
9306 			if (rd == 0)
9307 				err += efunc(pc, "cannot write to 0 address\n");
9308 			break;
9309 		case DIF_OP_CMP:
9310 		case DIF_OP_SCMP:
9311 			if (r1 >= nregs)
9312 				err += efunc(pc, "invalid register %u\n", r1);
9313 			if (r2 >= nregs)
9314 				err += efunc(pc, "invalid register %u\n", r2);
9315 			if (rd != 0)
9316 				err += efunc(pc, "non-zero reserved bits\n");
9317 			break;
9318 		case DIF_OP_TST:
9319 			if (r1 >= nregs)
9320 				err += efunc(pc, "invalid register %u\n", r1);
9321 			if (r2 != 0 || rd != 0)
9322 				err += efunc(pc, "non-zero reserved bits\n");
9323 			break;
9324 		case DIF_OP_BA:
9325 		case DIF_OP_BE:
9326 		case DIF_OP_BNE:
9327 		case DIF_OP_BG:
9328 		case DIF_OP_BGU:
9329 		case DIF_OP_BGE:
9330 		case DIF_OP_BGEU:
9331 		case DIF_OP_BL:
9332 		case DIF_OP_BLU:
9333 		case DIF_OP_BLE:
9334 		case DIF_OP_BLEU:
9335 			if (label >= dp->dtdo_len) {
9336 				err += efunc(pc, "invalid branch target %u\n",
9337 				    label);
9338 			}
9339 			if (label <= pc) {
9340 				err += efunc(pc, "backward branch to %u\n",
9341 				    label);
9342 			}
9343 			break;
9344 		case DIF_OP_RET:
9345 			if (r1 != 0 || r2 != 0)
9346 				err += efunc(pc, "non-zero reserved bits\n");
9347 			if (rd >= nregs)
9348 				err += efunc(pc, "invalid register %u\n", rd);
9349 			break;
9350 		case DIF_OP_NOP:
9351 		case DIF_OP_POPTS:
9352 		case DIF_OP_FLUSHTS:
9353 			if (r1 != 0 || r2 != 0 || rd != 0)
9354 				err += efunc(pc, "non-zero reserved bits\n");
9355 			break;
9356 		case DIF_OP_SETX:
9357 			if (DIF_INSTR_INTEGER(instr) >= dp->dtdo_intlen) {
9358 				err += efunc(pc, "invalid integer ref %u\n",
9359 				    DIF_INSTR_INTEGER(instr));
9360 			}
9361 			if (rd >= nregs)
9362 				err += efunc(pc, "invalid register %u\n", rd);
9363 			if (rd == 0)
9364 				err += efunc(pc, "cannot write to %r0\n");
9365 			break;
9366 		case DIF_OP_SETS:
9367 			if (DIF_INSTR_STRING(instr) >= dp->dtdo_strlen) {
9368 				err += efunc(pc, "invalid string ref %u\n",
9369 				    DIF_INSTR_STRING(instr));
9370 			}
9371 			if (rd >= nregs)
9372 				err += efunc(pc, "invalid register %u\n", rd);
9373 			if (rd == 0)
9374 				err += efunc(pc, "cannot write to %r0\n");
9375 			break;
9376 		case DIF_OP_LDGA:
9377 		case DIF_OP_LDTA:
9378 			if (r1 > DIF_VAR_ARRAY_MAX)
9379 				err += efunc(pc, "invalid array %u\n", r1);
9380 			if (r2 >= nregs)
9381 				err += efunc(pc, "invalid register %u\n", r2);
9382 			if (rd >= nregs)
9383 				err += efunc(pc, "invalid register %u\n", rd);
9384 			if (rd == 0)
9385 				err += efunc(pc, "cannot write to %r0\n");
9386 			break;
9387 		case DIF_OP_LDGS:
9388 		case DIF_OP_LDTS:
9389 		case DIF_OP_LDLS:
9390 		case DIF_OP_LDGAA:
9391 		case DIF_OP_LDTAA:
9392 			if (v < DIF_VAR_OTHER_MIN || v > DIF_VAR_OTHER_MAX)
9393 				err += efunc(pc, "invalid variable %u\n", v);
9394 			if (rd >= nregs)
9395 				err += efunc(pc, "invalid register %u\n", rd);
9396 			if (rd == 0)
9397 				err += efunc(pc, "cannot write to %r0\n");
9398 			break;
9399 		case DIF_OP_STGS:
9400 		case DIF_OP_STTS:
9401 		case DIF_OP_STLS:
9402 		case DIF_OP_STGAA:
9403 		case DIF_OP_STTAA:
9404 			if (v < DIF_VAR_OTHER_UBASE || v > DIF_VAR_OTHER_MAX)
9405 				err += efunc(pc, "invalid variable %u\n", v);
9406 			if (rs >= nregs)
9407 				err += efunc(pc, "invalid register %u\n", rd);
9408 			break;
9409 		case DIF_OP_CALL:
9410 			if (subr > DIF_SUBR_MAX)
9411 				err += efunc(pc, "invalid subr %u\n", subr);
9412 			if (rd >= nregs)
9413 				err += efunc(pc, "invalid register %u\n", rd);
9414 			if (rd == 0)
9415 				err += efunc(pc, "cannot write to %r0\n");
9416 
9417 			if (subr == DIF_SUBR_COPYOUT ||
9418 			    subr == DIF_SUBR_COPYOUTSTR) {
9419 				dp->dtdo_destructive = 1;
9420 			}
9421 
9422 			if (subr == DIF_SUBR_GETF) {
9423 				/*
9424 				 * If we have a getf() we need to record that
9425 				 * in our state.  Note that our state can be
9426 				 * NULL if this is a helper -- but in that
9427 				 * case, the call to getf() is itself illegal,
9428 				 * and will be caught (slightly later) when
9429 				 * the helper is validated.
9430 				 */
9431 				if (vstate->dtvs_state != NULL)
9432 					vstate->dtvs_state->dts_getf++;
9433 			}
9434 
9435 			break;
9436 		case DIF_OP_PUSHTR:
9437 			if (type != DIF_TYPE_STRING && type != DIF_TYPE_CTF)
9438 				err += efunc(pc, "invalid ref type %u\n", type);
9439 			if (r2 >= nregs)
9440 				err += efunc(pc, "invalid register %u\n", r2);
9441 			if (rs >= nregs)
9442 				err += efunc(pc, "invalid register %u\n", rs);
9443 			break;
9444 		case DIF_OP_PUSHTV:
9445 			if (type != DIF_TYPE_CTF)
9446 				err += efunc(pc, "invalid val type %u\n", type);
9447 			if (r2 >= nregs)
9448 				err += efunc(pc, "invalid register %u\n", r2);
9449 			if (rs >= nregs)
9450 				err += efunc(pc, "invalid register %u\n", rs);
9451 			break;
9452 		default:
9453 			err += efunc(pc, "invalid opcode %u\n",
9454 			    DIF_INSTR_OP(instr));
9455 		}
9456 	}
9457 
9458 	if (dp->dtdo_len != 0 &&
9459 	    DIF_INSTR_OP(dp->dtdo_buf[dp->dtdo_len - 1]) != DIF_OP_RET) {
9460 		err += efunc(dp->dtdo_len - 1,
9461 		    "expected 'ret' as last DIF instruction\n");
9462 	}
9463 
9464 	if (!(dp->dtdo_rtype.dtdt_flags & (DIF_TF_BYREF | DIF_TF_BYUREF))) {
9465 		/*
9466 		 * If we're not returning by reference, the size must be either
9467 		 * 0 or the size of one of the base types.
9468 		 */
9469 		switch (dp->dtdo_rtype.dtdt_size) {
9470 		case 0:
9471 		case sizeof (uint8_t):
9472 		case sizeof (uint16_t):
9473 		case sizeof (uint32_t):
9474 		case sizeof (uint64_t):
9475 			break;
9476 
9477 		default:
9478 			err += efunc(dp->dtdo_len - 1, "bad return size\n");
9479 		}
9480 	}
9481 
9482 	for (i = 0; i < dp->dtdo_varlen && err == 0; i++) {
9483 		dtrace_difv_t *v = &dp->dtdo_vartab[i], *existing = NULL;
9484 		dtrace_diftype_t *vt, *et;
9485 		uint_t id, ndx;
9486 
9487 		if (v->dtdv_scope != DIFV_SCOPE_GLOBAL &&
9488 		    v->dtdv_scope != DIFV_SCOPE_THREAD &&
9489 		    v->dtdv_scope != DIFV_SCOPE_LOCAL) {
9490 			err += efunc(i, "unrecognized variable scope %d\n",
9491 			    v->dtdv_scope);
9492 			break;
9493 		}
9494 
9495 		if (v->dtdv_kind != DIFV_KIND_ARRAY &&
9496 		    v->dtdv_kind != DIFV_KIND_SCALAR) {
9497 			err += efunc(i, "unrecognized variable type %d\n",
9498 			    v->dtdv_kind);
9499 			break;
9500 		}
9501 
9502 		if ((id = v->dtdv_id) > DIF_VARIABLE_MAX) {
9503 			err += efunc(i, "%d exceeds variable id limit\n", id);
9504 			break;
9505 		}
9506 
9507 		if (id < DIF_VAR_OTHER_UBASE)
9508 			continue;
9509 
9510 		/*
9511 		 * For user-defined variables, we need to check that this
9512 		 * definition is identical to any previous definition that we
9513 		 * encountered.
9514 		 */
9515 		ndx = id - DIF_VAR_OTHER_UBASE;
9516 
9517 		switch (v->dtdv_scope) {
9518 		case DIFV_SCOPE_GLOBAL:
9519 			if (maxglobal == -1 || ndx > maxglobal)
9520 				maxglobal = ndx;
9521 
9522 			if (ndx < vstate->dtvs_nglobals) {
9523 				dtrace_statvar_t *svar;
9524 
9525 				if ((svar = vstate->dtvs_globals[ndx]) != NULL)
9526 					existing = &svar->dtsv_var;
9527 			}
9528 
9529 			break;
9530 
9531 		case DIFV_SCOPE_THREAD:
9532 			if (maxtlocal == -1 || ndx > maxtlocal)
9533 				maxtlocal = ndx;
9534 
9535 			if (ndx < vstate->dtvs_ntlocals)
9536 				existing = &vstate->dtvs_tlocals[ndx];
9537 			break;
9538 
9539 		case DIFV_SCOPE_LOCAL:
9540 			if (maxlocal == -1 || ndx > maxlocal)
9541 				maxlocal = ndx;
9542 
9543 			if (ndx < vstate->dtvs_nlocals) {
9544 				dtrace_statvar_t *svar;
9545 
9546 				if ((svar = vstate->dtvs_locals[ndx]) != NULL)
9547 					existing = &svar->dtsv_var;
9548 			}
9549 
9550 			break;
9551 		}
9552 
9553 		vt = &v->dtdv_type;
9554 
9555 		if (vt->dtdt_flags & DIF_TF_BYREF) {
9556 			if (vt->dtdt_size == 0) {
9557 				err += efunc(i, "zero-sized variable\n");
9558 				break;
9559 			}
9560 
9561 			if ((v->dtdv_scope == DIFV_SCOPE_GLOBAL ||
9562 			    v->dtdv_scope == DIFV_SCOPE_LOCAL) &&
9563 			    vt->dtdt_size > dtrace_statvar_maxsize) {
9564 				err += efunc(i, "oversized by-ref static\n");
9565 				break;
9566 			}
9567 		}
9568 
9569 		if (existing == NULL || existing->dtdv_id == 0)
9570 			continue;
9571 
9572 		ASSERT(existing->dtdv_id == v->dtdv_id);
9573 		ASSERT(existing->dtdv_scope == v->dtdv_scope);
9574 
9575 		if (existing->dtdv_kind != v->dtdv_kind)
9576 			err += efunc(i, "%d changed variable kind\n", id);
9577 
9578 		et = &existing->dtdv_type;
9579 
9580 		if (vt->dtdt_flags != et->dtdt_flags) {
9581 			err += efunc(i, "%d changed variable type flags\n", id);
9582 			break;
9583 		}
9584 
9585 		if (vt->dtdt_size != 0 && vt->dtdt_size != et->dtdt_size) {
9586 			err += efunc(i, "%d changed variable type size\n", id);
9587 			break;
9588 		}
9589 	}
9590 
9591 	for (pc = 0; pc < dp->dtdo_len && err == 0; pc++) {
9592 		dif_instr_t instr = dp->dtdo_buf[pc];
9593 
9594 		uint_t v = DIF_INSTR_VAR(instr);
9595 		uint_t op = DIF_INSTR_OP(instr);
9596 
9597 		switch (op) {
9598 		case DIF_OP_LDGS:
9599 		case DIF_OP_LDGAA:
9600 		case DIF_OP_STGS:
9601 		case DIF_OP_STGAA:
9602 			if (v > DIF_VAR_OTHER_UBASE + maxglobal)
9603 				err += efunc(pc, "invalid variable %u\n", v);
9604 			break;
9605 		case DIF_OP_LDTS:
9606 		case DIF_OP_LDTAA:
9607 		case DIF_OP_STTS:
9608 		case DIF_OP_STTAA:
9609 			if (v > DIF_VAR_OTHER_UBASE + maxtlocal)
9610 				err += efunc(pc, "invalid variable %u\n", v);
9611 			break;
9612 		case DIF_OP_LDLS:
9613 		case DIF_OP_STLS:
9614 			if (v > DIF_VAR_OTHER_UBASE + maxlocal)
9615 				err += efunc(pc, "invalid variable %u\n", v);
9616 			break;
9617 		default:
9618 			break;
9619 		}
9620 	}
9621 
9622 	return (err);
9623 }
9624 
9625 /*
9626  * Validate a DTrace DIF object that it is to be used as a helper.  Helpers
9627  * are much more constrained than normal DIFOs.  Specifically, they may
9628  * not:
9629  *
9630  * 1. Make calls to subroutines other than copyin(), copyinstr() or
9631  *    miscellaneous string routines
9632  * 2. Access DTrace variables other than the args[] array, and the
9633  *    curthread, pid, ppid, tid, execname, zonename, uid and gid variables.
9634  * 3. Have thread-local variables.
9635  * 4. Have dynamic variables.
9636  */
9637 static int
9638 dtrace_difo_validate_helper(dtrace_difo_t *dp)
9639 {
9640 	int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
9641 	int err = 0;
9642 	uint_t pc;
9643 
9644 	for (pc = 0; pc < dp->dtdo_len; pc++) {
9645 		dif_instr_t instr = dp->dtdo_buf[pc];
9646 
9647 		uint_t v = DIF_INSTR_VAR(instr);
9648 		uint_t subr = DIF_INSTR_SUBR(instr);
9649 		uint_t op = DIF_INSTR_OP(instr);
9650 
9651 		switch (op) {
9652 		case DIF_OP_OR:
9653 		case DIF_OP_XOR:
9654 		case DIF_OP_AND:
9655 		case DIF_OP_SLL:
9656 		case DIF_OP_SRL:
9657 		case DIF_OP_SRA:
9658 		case DIF_OP_SUB:
9659 		case DIF_OP_ADD:
9660 		case DIF_OP_MUL:
9661 		case DIF_OP_SDIV:
9662 		case DIF_OP_UDIV:
9663 		case DIF_OP_SREM:
9664 		case DIF_OP_UREM:
9665 		case DIF_OP_COPYS:
9666 		case DIF_OP_NOT:
9667 		case DIF_OP_MOV:
9668 		case DIF_OP_RLDSB:
9669 		case DIF_OP_RLDSH:
9670 		case DIF_OP_RLDSW:
9671 		case DIF_OP_RLDUB:
9672 		case DIF_OP_RLDUH:
9673 		case DIF_OP_RLDUW:
9674 		case DIF_OP_RLDX:
9675 		case DIF_OP_ULDSB:
9676 		case DIF_OP_ULDSH:
9677 		case DIF_OP_ULDSW:
9678 		case DIF_OP_ULDUB:
9679 		case DIF_OP_ULDUH:
9680 		case DIF_OP_ULDUW:
9681 		case DIF_OP_ULDX:
9682 		case DIF_OP_STB:
9683 		case DIF_OP_STH:
9684 		case DIF_OP_STW:
9685 		case DIF_OP_STX:
9686 		case DIF_OP_ALLOCS:
9687 		case DIF_OP_CMP:
9688 		case DIF_OP_SCMP:
9689 		case DIF_OP_TST:
9690 		case DIF_OP_BA:
9691 		case DIF_OP_BE:
9692 		case DIF_OP_BNE:
9693 		case DIF_OP_BG:
9694 		case DIF_OP_BGU:
9695 		case DIF_OP_BGE:
9696 		case DIF_OP_BGEU:
9697 		case DIF_OP_BL:
9698 		case DIF_OP_BLU:
9699 		case DIF_OP_BLE:
9700 		case DIF_OP_BLEU:
9701 		case DIF_OP_RET:
9702 		case DIF_OP_NOP:
9703 		case DIF_OP_POPTS:
9704 		case DIF_OP_FLUSHTS:
9705 		case DIF_OP_SETX:
9706 		case DIF_OP_SETS:
9707 		case DIF_OP_LDGA:
9708 		case DIF_OP_LDLS:
9709 		case DIF_OP_STGS:
9710 		case DIF_OP_STLS:
9711 		case DIF_OP_PUSHTR:
9712 		case DIF_OP_PUSHTV:
9713 			break;
9714 
9715 		case DIF_OP_LDGS:
9716 			if (v >= DIF_VAR_OTHER_UBASE)
9717 				break;
9718 
9719 			if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9)
9720 				break;
9721 
9722 			if (v == DIF_VAR_CURTHREAD || v == DIF_VAR_PID ||
9723 			    v == DIF_VAR_PPID || v == DIF_VAR_TID ||
9724 			    v == DIF_VAR_EXECNAME || v == DIF_VAR_ZONENAME ||
9725 			    v == DIF_VAR_UID || v == DIF_VAR_GID)
9726 				break;
9727 
9728 			err += efunc(pc, "illegal variable %u\n", v);
9729 			break;
9730 
9731 		case DIF_OP_LDTA:
9732 		case DIF_OP_LDTS:
9733 		case DIF_OP_LDGAA:
9734 		case DIF_OP_LDTAA:
9735 			err += efunc(pc, "illegal dynamic variable load\n");
9736 			break;
9737 
9738 		case DIF_OP_STTS:
9739 		case DIF_OP_STGAA:
9740 		case DIF_OP_STTAA:
9741 			err += efunc(pc, "illegal dynamic variable store\n");
9742 			break;
9743 
9744 		case DIF_OP_CALL:
9745 			if (subr == DIF_SUBR_ALLOCA ||
9746 			    subr == DIF_SUBR_BCOPY ||
9747 			    subr == DIF_SUBR_COPYIN ||
9748 			    subr == DIF_SUBR_COPYINTO ||
9749 			    subr == DIF_SUBR_COPYINSTR ||
9750 			    subr == DIF_SUBR_INDEX ||
9751 			    subr == DIF_SUBR_INET_NTOA ||
9752 			    subr == DIF_SUBR_INET_NTOA6 ||
9753 			    subr == DIF_SUBR_INET_NTOP ||
9754 			    subr == DIF_SUBR_JSON ||
9755 			    subr == DIF_SUBR_LLTOSTR ||
9756 			    subr == DIF_SUBR_STRTOLL ||
9757 			    subr == DIF_SUBR_RINDEX ||
9758 			    subr == DIF_SUBR_STRCHR ||
9759 			    subr == DIF_SUBR_STRJOIN ||
9760 			    subr == DIF_SUBR_STRRCHR ||
9761 			    subr == DIF_SUBR_STRSTR ||
9762 			    subr == DIF_SUBR_HTONS ||
9763 			    subr == DIF_SUBR_HTONL ||
9764 			    subr == DIF_SUBR_HTONLL ||
9765 			    subr == DIF_SUBR_NTOHS ||
9766 			    subr == DIF_SUBR_NTOHL ||
9767 			    subr == DIF_SUBR_NTOHLL)
9768 				break;
9769 
9770 			err += efunc(pc, "invalid subr %u\n", subr);
9771 			break;
9772 
9773 		default:
9774 			err += efunc(pc, "invalid opcode %u\n",
9775 			    DIF_INSTR_OP(instr));
9776 		}
9777 	}
9778 
9779 	return (err);
9780 }
9781 
9782 /*
9783  * Returns 1 if the expression in the DIF object can be cached on a per-thread
9784  * basis; 0 if not.
9785  */
9786 static int
9787 dtrace_difo_cacheable(dtrace_difo_t *dp)
9788 {
9789 	int i;
9790 
9791 	if (dp == NULL)
9792 		return (0);
9793 
9794 	for (i = 0; i < dp->dtdo_varlen; i++) {
9795 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
9796 
9797 		if (v->dtdv_scope != DIFV_SCOPE_GLOBAL)
9798 			continue;
9799 
9800 		switch (v->dtdv_id) {
9801 		case DIF_VAR_CURTHREAD:
9802 		case DIF_VAR_PID:
9803 		case DIF_VAR_TID:
9804 		case DIF_VAR_EXECNAME:
9805 		case DIF_VAR_ZONENAME:
9806 			break;
9807 
9808 		default:
9809 			return (0);
9810 		}
9811 	}
9812 
9813 	/*
9814 	 * This DIF object may be cacheable.  Now we need to look for any
9815 	 * array loading instructions, any memory loading instructions, or
9816 	 * any stores to thread-local variables.
9817 	 */
9818 	for (i = 0; i < dp->dtdo_len; i++) {
9819 		uint_t op = DIF_INSTR_OP(dp->dtdo_buf[i]);
9820 
9821 		if ((op >= DIF_OP_LDSB && op <= DIF_OP_LDX) ||
9822 		    (op >= DIF_OP_ULDSB && op <= DIF_OP_ULDX) ||
9823 		    (op >= DIF_OP_RLDSB && op <= DIF_OP_RLDX) ||
9824 		    op == DIF_OP_LDGA || op == DIF_OP_STTS)
9825 			return (0);
9826 	}
9827 
9828 	return (1);
9829 }
9830 
9831 static void
9832 dtrace_difo_hold(dtrace_difo_t *dp)
9833 {
9834 	int i;
9835 
9836 	ASSERT(MUTEX_HELD(&dtrace_lock));
9837 
9838 	dp->dtdo_refcnt++;
9839 	ASSERT(dp->dtdo_refcnt != 0);
9840 
9841 	/*
9842 	 * We need to check this DIF object for references to the variable
9843 	 * DIF_VAR_VTIMESTAMP.
9844 	 */
9845 	for (i = 0; i < dp->dtdo_varlen; i++) {
9846 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
9847 
9848 		if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
9849 			continue;
9850 
9851 		if (dtrace_vtime_references++ == 0)
9852 			dtrace_vtime_enable();
9853 	}
9854 }
9855 
9856 /*
9857  * This routine calculates the dynamic variable chunksize for a given DIF
9858  * object.  The calculation is not fool-proof, and can probably be tricked by
9859  * malicious DIF -- but it works for all compiler-generated DIF.  Because this
9860  * calculation is likely imperfect, dtrace_dynvar() is able to gracefully fail
9861  * if a dynamic variable size exceeds the chunksize.
9862  */
9863 static void
9864 dtrace_difo_chunksize(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9865 {
9866 	uint64_t sval;
9867 	dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
9868 	const dif_instr_t *text = dp->dtdo_buf;
9869 	uint_t pc, srd = 0;
9870 	uint_t ttop = 0;
9871 	size_t size, ksize;
9872 	uint_t id, i;
9873 
9874 	for (pc = 0; pc < dp->dtdo_len; pc++) {
9875 		dif_instr_t instr = text[pc];
9876 		uint_t op = DIF_INSTR_OP(instr);
9877 		uint_t rd = DIF_INSTR_RD(instr);
9878 		uint_t r1 = DIF_INSTR_R1(instr);
9879 		uint_t nkeys = 0;
9880 		uchar_t scope;
9881 
9882 		dtrace_key_t *key = tupregs;
9883 
9884 		switch (op) {
9885 		case DIF_OP_SETX:
9886 			sval = dp->dtdo_inttab[DIF_INSTR_INTEGER(instr)];
9887 			srd = rd;
9888 			continue;
9889 
9890 		case DIF_OP_STTS:
9891 			key = &tupregs[DIF_DTR_NREGS];
9892 			key[0].dttk_size = 0;
9893 			key[1].dttk_size = 0;
9894 			nkeys = 2;
9895 			scope = DIFV_SCOPE_THREAD;
9896 			break;
9897 
9898 		case DIF_OP_STGAA:
9899 		case DIF_OP_STTAA:
9900 			nkeys = ttop;
9901 
9902 			if (DIF_INSTR_OP(instr) == DIF_OP_STTAA)
9903 				key[nkeys++].dttk_size = 0;
9904 
9905 			key[nkeys++].dttk_size = 0;
9906 
9907 			if (op == DIF_OP_STTAA) {
9908 				scope = DIFV_SCOPE_THREAD;
9909 			} else {
9910 				scope = DIFV_SCOPE_GLOBAL;
9911 			}
9912 
9913 			break;
9914 
9915 		case DIF_OP_PUSHTR:
9916 			if (ttop == DIF_DTR_NREGS)
9917 				return;
9918 
9919 			if ((srd == 0 || sval == 0) && r1 == DIF_TYPE_STRING) {
9920 				/*
9921 				 * If the register for the size of the "pushtr"
9922 				 * is %r0 (or the value is 0) and the type is
9923 				 * a string, we'll use the system-wide default
9924 				 * string size.
9925 				 */
9926 				tupregs[ttop++].dttk_size =
9927 				    dtrace_strsize_default;
9928 			} else {
9929 				if (srd == 0)
9930 					return;
9931 
9932 				if (sval > LONG_MAX)
9933 					return;
9934 
9935 				tupregs[ttop++].dttk_size = sval;
9936 			}
9937 
9938 			break;
9939 
9940 		case DIF_OP_PUSHTV:
9941 			if (ttop == DIF_DTR_NREGS)
9942 				return;
9943 
9944 			tupregs[ttop++].dttk_size = 0;
9945 			break;
9946 
9947 		case DIF_OP_FLUSHTS:
9948 			ttop = 0;
9949 			break;
9950 
9951 		case DIF_OP_POPTS:
9952 			if (ttop != 0)
9953 				ttop--;
9954 			break;
9955 		}
9956 
9957 		sval = 0;
9958 		srd = 0;
9959 
9960 		if (nkeys == 0)
9961 			continue;
9962 
9963 		/*
9964 		 * We have a dynamic variable allocation; calculate its size.
9965 		 */
9966 		for (ksize = 0, i = 0; i < nkeys; i++)
9967 			ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
9968 
9969 		size = sizeof (dtrace_dynvar_t);
9970 		size += sizeof (dtrace_key_t) * (nkeys - 1);
9971 		size += ksize;
9972 
9973 		/*
9974 		 * Now we need to determine the size of the stored data.
9975 		 */
9976 		id = DIF_INSTR_VAR(instr);
9977 
9978 		for (i = 0; i < dp->dtdo_varlen; i++) {
9979 			dtrace_difv_t *v = &dp->dtdo_vartab[i];
9980 
9981 			if (v->dtdv_id == id && v->dtdv_scope == scope) {
9982 				size += v->dtdv_type.dtdt_size;
9983 				break;
9984 			}
9985 		}
9986 
9987 		if (i == dp->dtdo_varlen)
9988 			return;
9989 
9990 		/*
9991 		 * We have the size.  If this is larger than the chunk size
9992 		 * for our dynamic variable state, reset the chunk size.
9993 		 */
9994 		size = P2ROUNDUP(size, sizeof (uint64_t));
9995 
9996 		/*
9997 		 * Before setting the chunk size, check that we're not going
9998 		 * to set it to a negative value...
9999 		 */
10000 		if (size > LONG_MAX)
10001 			return;
10002 
10003 		/*
10004 		 * ...and make certain that we didn't badly overflow.
10005 		 */
10006 		if (size < ksize || size < sizeof (dtrace_dynvar_t))
10007 			return;
10008 
10009 		if (size > vstate->dtvs_dynvars.dtds_chunksize)
10010 			vstate->dtvs_dynvars.dtds_chunksize = size;
10011 	}
10012 }
10013 
10014 static void
10015 dtrace_difo_init(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
10016 {
10017 	int i, oldsvars, osz, nsz, otlocals, ntlocals;
10018 	uint_t id;
10019 
10020 	ASSERT(MUTEX_HELD(&dtrace_lock));
10021 	ASSERT(dp->dtdo_buf != NULL && dp->dtdo_len != 0);
10022 
10023 	for (i = 0; i < dp->dtdo_varlen; i++) {
10024 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
10025 		dtrace_statvar_t *svar, ***svarp;
10026 		size_t dsize = 0;
10027 		uint8_t scope = v->dtdv_scope;
10028 		int *np;
10029 
10030 		if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
10031 			continue;
10032 
10033 		id -= DIF_VAR_OTHER_UBASE;
10034 
10035 		switch (scope) {
10036 		case DIFV_SCOPE_THREAD:
10037 			while (id >= (otlocals = vstate->dtvs_ntlocals)) {
10038 				dtrace_difv_t *tlocals;
10039 
10040 				if ((ntlocals = (otlocals << 1)) == 0)
10041 					ntlocals = 1;
10042 
10043 				osz = otlocals * sizeof (dtrace_difv_t);
10044 				nsz = ntlocals * sizeof (dtrace_difv_t);
10045 
10046 				tlocals = kmem_zalloc(nsz, KM_SLEEP);
10047 
10048 				if (osz != 0) {
10049 					bcopy(vstate->dtvs_tlocals,
10050 					    tlocals, osz);
10051 					kmem_free(vstate->dtvs_tlocals, osz);
10052 				}
10053 
10054 				vstate->dtvs_tlocals = tlocals;
10055 				vstate->dtvs_ntlocals = ntlocals;
10056 			}
10057 
10058 			vstate->dtvs_tlocals[id] = *v;
10059 			continue;
10060 
10061 		case DIFV_SCOPE_LOCAL:
10062 			np = &vstate->dtvs_nlocals;
10063 			svarp = &vstate->dtvs_locals;
10064 
10065 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
10066 				dsize = NCPU * (v->dtdv_type.dtdt_size +
10067 				    sizeof (uint64_t));
10068 			else
10069 				dsize = NCPU * sizeof (uint64_t);
10070 
10071 			break;
10072 
10073 		case DIFV_SCOPE_GLOBAL:
10074 			np = &vstate->dtvs_nglobals;
10075 			svarp = &vstate->dtvs_globals;
10076 
10077 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
10078 				dsize = v->dtdv_type.dtdt_size +
10079 				    sizeof (uint64_t);
10080 
10081 			break;
10082 
10083 		default:
10084 			ASSERT(0);
10085 		}
10086 
10087 		while (id >= (oldsvars = *np)) {
10088 			dtrace_statvar_t **statics;
10089 			int newsvars, oldsize, newsize;
10090 
10091 			if ((newsvars = (oldsvars << 1)) == 0)
10092 				newsvars = 1;
10093 
10094 			oldsize = oldsvars * sizeof (dtrace_statvar_t *);
10095 			newsize = newsvars * sizeof (dtrace_statvar_t *);
10096 
10097 			statics = kmem_zalloc(newsize, KM_SLEEP);
10098 
10099 			if (oldsize != 0) {
10100 				bcopy(*svarp, statics, oldsize);
10101 				kmem_free(*svarp, oldsize);
10102 			}
10103 
10104 			*svarp = statics;
10105 			*np = newsvars;
10106 		}
10107 
10108 		if ((svar = (*svarp)[id]) == NULL) {
10109 			svar = kmem_zalloc(sizeof (dtrace_statvar_t), KM_SLEEP);
10110 			svar->dtsv_var = *v;
10111 
10112 			if ((svar->dtsv_size = dsize) != 0) {
10113 				svar->dtsv_data = (uint64_t)(uintptr_t)
10114 				    kmem_zalloc(dsize, KM_SLEEP);
10115 			}
10116 
10117 			(*svarp)[id] = svar;
10118 		}
10119 
10120 		svar->dtsv_refcnt++;
10121 	}
10122 
10123 	dtrace_difo_chunksize(dp, vstate);
10124 	dtrace_difo_hold(dp);
10125 }
10126 
10127 static dtrace_difo_t *
10128 dtrace_difo_duplicate(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
10129 {
10130 	dtrace_difo_t *new;
10131 	size_t sz;
10132 
10133 	ASSERT(dp->dtdo_buf != NULL);
10134 	ASSERT(dp->dtdo_refcnt != 0);
10135 
10136 	new = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
10137 
10138 	ASSERT(dp->dtdo_buf != NULL);
10139 	sz = dp->dtdo_len * sizeof (dif_instr_t);
10140 	new->dtdo_buf = kmem_alloc(sz, KM_SLEEP);
10141 	bcopy(dp->dtdo_buf, new->dtdo_buf, sz);
10142 	new->dtdo_len = dp->dtdo_len;
10143 
10144 	if (dp->dtdo_strtab != NULL) {
10145 		ASSERT(dp->dtdo_strlen != 0);
10146 		new->dtdo_strtab = kmem_alloc(dp->dtdo_strlen, KM_SLEEP);
10147 		bcopy(dp->dtdo_strtab, new->dtdo_strtab, dp->dtdo_strlen);
10148 		new->dtdo_strlen = dp->dtdo_strlen;
10149 	}
10150 
10151 	if (dp->dtdo_inttab != NULL) {
10152 		ASSERT(dp->dtdo_intlen != 0);
10153 		sz = dp->dtdo_intlen * sizeof (uint64_t);
10154 		new->dtdo_inttab = kmem_alloc(sz, KM_SLEEP);
10155 		bcopy(dp->dtdo_inttab, new->dtdo_inttab, sz);
10156 		new->dtdo_intlen = dp->dtdo_intlen;
10157 	}
10158 
10159 	if (dp->dtdo_vartab != NULL) {
10160 		ASSERT(dp->dtdo_varlen != 0);
10161 		sz = dp->dtdo_varlen * sizeof (dtrace_difv_t);
10162 		new->dtdo_vartab = kmem_alloc(sz, KM_SLEEP);
10163 		bcopy(dp->dtdo_vartab, new->dtdo_vartab, sz);
10164 		new->dtdo_varlen = dp->dtdo_varlen;
10165 	}
10166 
10167 	dtrace_difo_init(new, vstate);
10168 	return (new);
10169 }
10170 
10171 static void
10172 dtrace_difo_destroy(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
10173 {
10174 	int i;
10175 
10176 	ASSERT(dp->dtdo_refcnt == 0);
10177 
10178 	for (i = 0; i < dp->dtdo_varlen; i++) {
10179 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
10180 		dtrace_statvar_t *svar, **svarp;
10181 		uint_t id;
10182 		uint8_t scope = v->dtdv_scope;
10183 		int *np;
10184 
10185 		switch (scope) {
10186 		case DIFV_SCOPE_THREAD:
10187 			continue;
10188 
10189 		case DIFV_SCOPE_LOCAL:
10190 			np = &vstate->dtvs_nlocals;
10191 			svarp = vstate->dtvs_locals;
10192 			break;
10193 
10194 		case DIFV_SCOPE_GLOBAL:
10195 			np = &vstate->dtvs_nglobals;
10196 			svarp = vstate->dtvs_globals;
10197 			break;
10198 
10199 		default:
10200 			ASSERT(0);
10201 		}
10202 
10203 		if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
10204 			continue;
10205 
10206 		id -= DIF_VAR_OTHER_UBASE;
10207 		ASSERT(id < *np);
10208 
10209 		svar = svarp[id];
10210 		ASSERT(svar != NULL);
10211 		ASSERT(svar->dtsv_refcnt > 0);
10212 
10213 		if (--svar->dtsv_refcnt > 0)
10214 			continue;
10215 
10216 		if (svar->dtsv_size != 0) {
10217 			ASSERT(svar->dtsv_data != NULL);
10218 			kmem_free((void *)(uintptr_t)svar->dtsv_data,
10219 			    svar->dtsv_size);
10220 		}
10221 
10222 		kmem_free(svar, sizeof (dtrace_statvar_t));
10223 		svarp[id] = NULL;
10224 	}
10225 
10226 	kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
10227 	kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
10228 	kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
10229 	kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
10230 
10231 	kmem_free(dp, sizeof (dtrace_difo_t));
10232 }
10233 
10234 static void
10235 dtrace_difo_release(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
10236 {
10237 	int i;
10238 
10239 	ASSERT(MUTEX_HELD(&dtrace_lock));
10240 	ASSERT(dp->dtdo_refcnt != 0);
10241 
10242 	for (i = 0; i < dp->dtdo_varlen; i++) {
10243 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
10244 
10245 		if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
10246 			continue;
10247 
10248 		ASSERT(dtrace_vtime_references > 0);
10249 		if (--dtrace_vtime_references == 0)
10250 			dtrace_vtime_disable();
10251 	}
10252 
10253 	if (--dp->dtdo_refcnt == 0)
10254 		dtrace_difo_destroy(dp, vstate);
10255 }
10256 
10257 /*
10258  * DTrace Format Functions
10259  */
10260 static uint16_t
10261 dtrace_format_add(dtrace_state_t *state, char *str)
10262 {
10263 	char *fmt, **new;
10264 	uint16_t ndx, len = strlen(str) + 1;
10265 
10266 	fmt = kmem_zalloc(len, KM_SLEEP);
10267 	bcopy(str, fmt, len);
10268 
10269 	for (ndx = 0; ndx < state->dts_nformats; ndx++) {
10270 		if (state->dts_formats[ndx] == NULL) {
10271 			state->dts_formats[ndx] = fmt;
10272 			return (ndx + 1);
10273 		}
10274 	}
10275 
10276 	if (state->dts_nformats == USHRT_MAX) {
10277 		/*
10278 		 * This is only likely if a denial-of-service attack is being
10279 		 * attempted.  As such, it's okay to fail silently here.
10280 		 */
10281 		kmem_free(fmt, len);
10282 		return (0);
10283 	}
10284 
10285 	/*
10286 	 * For simplicity, we always resize the formats array to be exactly the
10287 	 * number of formats.
10288 	 */
10289 	ndx = state->dts_nformats++;
10290 	new = kmem_alloc((ndx + 1) * sizeof (char *), KM_SLEEP);
10291 
10292 	if (state->dts_formats != NULL) {
10293 		ASSERT(ndx != 0);
10294 		bcopy(state->dts_formats, new, ndx * sizeof (char *));
10295 		kmem_free(state->dts_formats, ndx * sizeof (char *));
10296 	}
10297 
10298 	state->dts_formats = new;
10299 	state->dts_formats[ndx] = fmt;
10300 
10301 	return (ndx + 1);
10302 }
10303 
10304 static void
10305 dtrace_format_remove(dtrace_state_t *state, uint16_t format)
10306 {
10307 	char *fmt;
10308 
10309 	ASSERT(state->dts_formats != NULL);
10310 	ASSERT(format <= state->dts_nformats);
10311 	ASSERT(state->dts_formats[format - 1] != NULL);
10312 
10313 	fmt = state->dts_formats[format - 1];
10314 	kmem_free(fmt, strlen(fmt) + 1);
10315 	state->dts_formats[format - 1] = NULL;
10316 }
10317 
10318 static void
10319 dtrace_format_destroy(dtrace_state_t *state)
10320 {
10321 	int i;
10322 
10323 	if (state->dts_nformats == 0) {
10324 		ASSERT(state->dts_formats == NULL);
10325 		return;
10326 	}
10327 
10328 	ASSERT(state->dts_formats != NULL);
10329 
10330 	for (i = 0; i < state->dts_nformats; i++) {
10331 		char *fmt = state->dts_formats[i];
10332 
10333 		if (fmt == NULL)
10334 			continue;
10335 
10336 		kmem_free(fmt, strlen(fmt) + 1);
10337 	}
10338 
10339 	kmem_free(state->dts_formats, state->dts_nformats * sizeof (char *));
10340 	state->dts_nformats = 0;
10341 	state->dts_formats = NULL;
10342 }
10343 
10344 /*
10345  * DTrace Predicate Functions
10346  */
10347 static dtrace_predicate_t *
10348 dtrace_predicate_create(dtrace_difo_t *dp)
10349 {
10350 	dtrace_predicate_t *pred;
10351 
10352 	ASSERT(MUTEX_HELD(&dtrace_lock));
10353 	ASSERT(dp->dtdo_refcnt != 0);
10354 
10355 	pred = kmem_zalloc(sizeof (dtrace_predicate_t), KM_SLEEP);
10356 	pred->dtp_difo = dp;
10357 	pred->dtp_refcnt = 1;
10358 
10359 	if (!dtrace_difo_cacheable(dp))
10360 		return (pred);
10361 
10362 	if (dtrace_predcache_id == DTRACE_CACHEIDNONE) {
10363 		/*
10364 		 * This is only theoretically possible -- we have had 2^32
10365 		 * cacheable predicates on this machine.  We cannot allow any
10366 		 * more predicates to become cacheable:  as unlikely as it is,
10367 		 * there may be a thread caching a (now stale) predicate cache
10368 		 * ID. (N.B.: the temptation is being successfully resisted to
10369 		 * have this cmn_err() "Holy shit -- we executed this code!")
10370 		 */
10371 		return (pred);
10372 	}
10373 
10374 	pred->dtp_cacheid = dtrace_predcache_id++;
10375 
10376 	return (pred);
10377 }
10378 
10379 static void
10380 dtrace_predicate_hold(dtrace_predicate_t *pred)
10381 {
10382 	ASSERT(MUTEX_HELD(&dtrace_lock));
10383 	ASSERT(pred->dtp_difo != NULL && pred->dtp_difo->dtdo_refcnt != 0);
10384 	ASSERT(pred->dtp_refcnt > 0);
10385 
10386 	pred->dtp_refcnt++;
10387 }
10388 
10389 static void
10390 dtrace_predicate_release(dtrace_predicate_t *pred, dtrace_vstate_t *vstate)
10391 {
10392 	dtrace_difo_t *dp = pred->dtp_difo;
10393 
10394 	ASSERT(MUTEX_HELD(&dtrace_lock));
10395 	ASSERT(dp != NULL && dp->dtdo_refcnt != 0);
10396 	ASSERT(pred->dtp_refcnt > 0);
10397 
10398 	if (--pred->dtp_refcnt == 0) {
10399 		dtrace_difo_release(pred->dtp_difo, vstate);
10400 		kmem_free(pred, sizeof (dtrace_predicate_t));
10401 	}
10402 }
10403 
10404 /*
10405  * DTrace Action Description Functions
10406  */
10407 static dtrace_actdesc_t *
10408 dtrace_actdesc_create(dtrace_actkind_t kind, uint32_t ntuple,
10409     uint64_t uarg, uint64_t arg)
10410 {
10411 	dtrace_actdesc_t *act;
10412 
10413 	ASSERT(!DTRACEACT_ISPRINTFLIKE(kind) || (arg != NULL &&
10414 	    arg >= KERNELBASE) || (arg == NULL && kind == DTRACEACT_PRINTA));
10415 
10416 	act = kmem_zalloc(sizeof (dtrace_actdesc_t), KM_SLEEP);
10417 	act->dtad_kind = kind;
10418 	act->dtad_ntuple = ntuple;
10419 	act->dtad_uarg = uarg;
10420 	act->dtad_arg = arg;
10421 	act->dtad_refcnt = 1;
10422 
10423 	return (act);
10424 }
10425 
10426 static void
10427 dtrace_actdesc_hold(dtrace_actdesc_t *act)
10428 {
10429 	ASSERT(act->dtad_refcnt >= 1);
10430 	act->dtad_refcnt++;
10431 }
10432 
10433 static void
10434 dtrace_actdesc_release(dtrace_actdesc_t *act, dtrace_vstate_t *vstate)
10435 {
10436 	dtrace_actkind_t kind = act->dtad_kind;
10437 	dtrace_difo_t *dp;
10438 
10439 	ASSERT(act->dtad_refcnt >= 1);
10440 
10441 	if (--act->dtad_refcnt != 0)
10442 		return;
10443 
10444 	if ((dp = act->dtad_difo) != NULL)
10445 		dtrace_difo_release(dp, vstate);
10446 
10447 	if (DTRACEACT_ISPRINTFLIKE(kind)) {
10448 		char *str = (char *)(uintptr_t)act->dtad_arg;
10449 
10450 		ASSERT((str != NULL && (uintptr_t)str >= KERNELBASE) ||
10451 		    (str == NULL && act->dtad_kind == DTRACEACT_PRINTA));
10452 
10453 		if (str != NULL)
10454 			kmem_free(str, strlen(str) + 1);
10455 	}
10456 
10457 	kmem_free(act, sizeof (dtrace_actdesc_t));
10458 }
10459 
10460 /*
10461  * DTrace ECB Functions
10462  */
10463 static dtrace_ecb_t *
10464 dtrace_ecb_add(dtrace_state_t *state, dtrace_probe_t *probe)
10465 {
10466 	dtrace_ecb_t *ecb;
10467 	dtrace_epid_t epid;
10468 
10469 	ASSERT(MUTEX_HELD(&dtrace_lock));
10470 
10471 	ecb = kmem_zalloc(sizeof (dtrace_ecb_t), KM_SLEEP);
10472 	ecb->dte_predicate = NULL;
10473 	ecb->dte_probe = probe;
10474 
10475 	/*
10476 	 * The default size is the size of the default action: recording
10477 	 * the header.
10478 	 */
10479 	ecb->dte_size = ecb->dte_needed = sizeof (dtrace_rechdr_t);
10480 	ecb->dte_alignment = sizeof (dtrace_epid_t);
10481 
10482 	epid = state->dts_epid++;
10483 
10484 	if (epid - 1 >= state->dts_necbs) {
10485 		dtrace_ecb_t **oecbs = state->dts_ecbs, **ecbs;
10486 		int necbs = state->dts_necbs << 1;
10487 
10488 		ASSERT(epid == state->dts_necbs + 1);
10489 
10490 		if (necbs == 0) {
10491 			ASSERT(oecbs == NULL);
10492 			necbs = 1;
10493 		}
10494 
10495 		ecbs = kmem_zalloc(necbs * sizeof (*ecbs), KM_SLEEP);
10496 
10497 		if (oecbs != NULL)
10498 			bcopy(oecbs, ecbs, state->dts_necbs * sizeof (*ecbs));
10499 
10500 		dtrace_membar_producer();
10501 		state->dts_ecbs = ecbs;
10502 
10503 		if (oecbs != NULL) {
10504 			/*
10505 			 * If this state is active, we must dtrace_sync()
10506 			 * before we can free the old dts_ecbs array:  we're
10507 			 * coming in hot, and there may be active ring
10508 			 * buffer processing (which indexes into the dts_ecbs
10509 			 * array) on another CPU.
10510 			 */
10511 			if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
10512 				dtrace_sync();
10513 
10514 			kmem_free(oecbs, state->dts_necbs * sizeof (*ecbs));
10515 		}
10516 
10517 		dtrace_membar_producer();
10518 		state->dts_necbs = necbs;
10519 	}
10520 
10521 	ecb->dte_state = state;
10522 
10523 	ASSERT(state->dts_ecbs[epid - 1] == NULL);
10524 	dtrace_membar_producer();
10525 	state->dts_ecbs[(ecb->dte_epid = epid) - 1] = ecb;
10526 
10527 	return (ecb);
10528 }
10529 
10530 static int
10531 dtrace_ecb_enable(dtrace_ecb_t *ecb)
10532 {
10533 	dtrace_probe_t *probe = ecb->dte_probe;
10534 
10535 	ASSERT(MUTEX_HELD(&cpu_lock));
10536 	ASSERT(MUTEX_HELD(&dtrace_lock));
10537 	ASSERT(ecb->dte_next == NULL);
10538 
10539 	if (probe == NULL) {
10540 		/*
10541 		 * This is the NULL probe -- there's nothing to do.
10542 		 */
10543 		return (0);
10544 	}
10545 
10546 	if (probe->dtpr_ecb == NULL) {
10547 		dtrace_provider_t *prov = probe->dtpr_provider;
10548 
10549 		/*
10550 		 * We're the first ECB on this probe.
10551 		 */
10552 		probe->dtpr_ecb = probe->dtpr_ecb_last = ecb;
10553 
10554 		if (ecb->dte_predicate != NULL)
10555 			probe->dtpr_predcache = ecb->dte_predicate->dtp_cacheid;
10556 
10557 		return (prov->dtpv_pops.dtps_enable(prov->dtpv_arg,
10558 		    probe->dtpr_id, probe->dtpr_arg));
10559 	} else {
10560 		/*
10561 		 * This probe is already active.  Swing the last pointer to
10562 		 * point to the new ECB, and issue a dtrace_sync() to assure
10563 		 * that all CPUs have seen the change.
10564 		 */
10565 		ASSERT(probe->dtpr_ecb_last != NULL);
10566 		probe->dtpr_ecb_last->dte_next = ecb;
10567 		probe->dtpr_ecb_last = ecb;
10568 		probe->dtpr_predcache = 0;
10569 
10570 		dtrace_sync();
10571 		return (0);
10572 	}
10573 }
10574 
10575 static int
10576 dtrace_ecb_resize(dtrace_ecb_t *ecb)
10577 {
10578 	dtrace_action_t *act;
10579 	uint32_t curneeded = UINT32_MAX;
10580 	uint32_t aggbase = UINT32_MAX;
10581 
10582 	/*
10583 	 * If we record anything, we always record the dtrace_rechdr_t.  (And
10584 	 * we always record it first.)
10585 	 */
10586 	ecb->dte_size = sizeof (dtrace_rechdr_t);
10587 	ecb->dte_alignment = sizeof (dtrace_epid_t);
10588 
10589 	for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
10590 		dtrace_recdesc_t *rec = &act->dta_rec;
10591 		ASSERT(rec->dtrd_size > 0 || rec->dtrd_alignment == 1);
10592 
10593 		ecb->dte_alignment = MAX(ecb->dte_alignment,
10594 		    rec->dtrd_alignment);
10595 
10596 		if (DTRACEACT_ISAGG(act->dta_kind)) {
10597 			dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
10598 
10599 			ASSERT(rec->dtrd_size != 0);
10600 			ASSERT(agg->dtag_first != NULL);
10601 			ASSERT(act->dta_prev->dta_intuple);
10602 			ASSERT(aggbase != UINT32_MAX);
10603 			ASSERT(curneeded != UINT32_MAX);
10604 
10605 			agg->dtag_base = aggbase;
10606 
10607 			curneeded = P2ROUNDUP(curneeded, rec->dtrd_alignment);
10608 			rec->dtrd_offset = curneeded;
10609 			if (curneeded + rec->dtrd_size < curneeded)
10610 				return (EINVAL);
10611 			curneeded += rec->dtrd_size;
10612 			ecb->dte_needed = MAX(ecb->dte_needed, curneeded);
10613 
10614 			aggbase = UINT32_MAX;
10615 			curneeded = UINT32_MAX;
10616 		} else if (act->dta_intuple) {
10617 			if (curneeded == UINT32_MAX) {
10618 				/*
10619 				 * This is the first record in a tuple.  Align
10620 				 * curneeded to be at offset 4 in an 8-byte
10621 				 * aligned block.
10622 				 */
10623 				ASSERT(act->dta_prev == NULL ||
10624 				    !act->dta_prev->dta_intuple);
10625 				ASSERT3U(aggbase, ==, UINT32_MAX);
10626 				curneeded = P2PHASEUP(ecb->dte_size,
10627 				    sizeof (uint64_t), sizeof (dtrace_aggid_t));
10628 
10629 				aggbase = curneeded - sizeof (dtrace_aggid_t);
10630 				ASSERT(IS_P2ALIGNED(aggbase,
10631 				    sizeof (uint64_t)));
10632 			}
10633 			curneeded = P2ROUNDUP(curneeded, rec->dtrd_alignment);
10634 			rec->dtrd_offset = curneeded;
10635 			if (curneeded + rec->dtrd_size < curneeded)
10636 				return (EINVAL);
10637 			curneeded += rec->dtrd_size;
10638 		} else {
10639 			/* tuples must be followed by an aggregation */
10640 			ASSERT(act->dta_prev == NULL ||
10641 			    !act->dta_prev->dta_intuple);
10642 
10643 			ecb->dte_size = P2ROUNDUP(ecb->dte_size,
10644 			    rec->dtrd_alignment);
10645 			rec->dtrd_offset = ecb->dte_size;
10646 			if (ecb->dte_size + rec->dtrd_size < ecb->dte_size)
10647 				return (EINVAL);
10648 			ecb->dte_size += rec->dtrd_size;
10649 			ecb->dte_needed = MAX(ecb->dte_needed, ecb->dte_size);
10650 		}
10651 	}
10652 
10653 	if ((act = ecb->dte_action) != NULL &&
10654 	    !(act->dta_kind == DTRACEACT_SPECULATE && act->dta_next == NULL) &&
10655 	    ecb->dte_size == sizeof (dtrace_rechdr_t)) {
10656 		/*
10657 		 * If the size is still sizeof (dtrace_rechdr_t), then all
10658 		 * actions store no data; set the size to 0.
10659 		 */
10660 		ecb->dte_size = 0;
10661 	}
10662 
10663 	ecb->dte_size = P2ROUNDUP(ecb->dte_size, sizeof (dtrace_epid_t));
10664 	ecb->dte_needed = P2ROUNDUP(ecb->dte_needed, (sizeof (dtrace_epid_t)));
10665 	ecb->dte_state->dts_needed = MAX(ecb->dte_state->dts_needed,
10666 	    ecb->dte_needed);
10667 	return (0);
10668 }
10669 
10670 static dtrace_action_t *
10671 dtrace_ecb_aggregation_create(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
10672 {
10673 	dtrace_aggregation_t *agg;
10674 	size_t size = sizeof (uint64_t);
10675 	int ntuple = desc->dtad_ntuple;
10676 	dtrace_action_t *act;
10677 	dtrace_recdesc_t *frec;
10678 	dtrace_aggid_t aggid;
10679 	dtrace_state_t *state = ecb->dte_state;
10680 
10681 	agg = kmem_zalloc(sizeof (dtrace_aggregation_t), KM_SLEEP);
10682 	agg->dtag_ecb = ecb;
10683 
10684 	ASSERT(DTRACEACT_ISAGG(desc->dtad_kind));
10685 
10686 	switch (desc->dtad_kind) {
10687 	case DTRACEAGG_MIN:
10688 		agg->dtag_initial = INT64_MAX;
10689 		agg->dtag_aggregate = dtrace_aggregate_min;
10690 		break;
10691 
10692 	case DTRACEAGG_MAX:
10693 		agg->dtag_initial = INT64_MIN;
10694 		agg->dtag_aggregate = dtrace_aggregate_max;
10695 		break;
10696 
10697 	case DTRACEAGG_COUNT:
10698 		agg->dtag_aggregate = dtrace_aggregate_count;
10699 		break;
10700 
10701 	case DTRACEAGG_QUANTIZE:
10702 		agg->dtag_aggregate = dtrace_aggregate_quantize;
10703 		size = (((sizeof (uint64_t) * NBBY) - 1) * 2 + 1) *
10704 		    sizeof (uint64_t);
10705 		break;
10706 
10707 	case DTRACEAGG_LQUANTIZE: {
10708 		uint16_t step = DTRACE_LQUANTIZE_STEP(desc->dtad_arg);
10709 		uint16_t levels = DTRACE_LQUANTIZE_LEVELS(desc->dtad_arg);
10710 
10711 		agg->dtag_initial = desc->dtad_arg;
10712 		agg->dtag_aggregate = dtrace_aggregate_lquantize;
10713 
10714 		if (step == 0 || levels == 0)
10715 			goto err;
10716 
10717 		size = levels * sizeof (uint64_t) + 3 * sizeof (uint64_t);
10718 		break;
10719 	}
10720 
10721 	case DTRACEAGG_LLQUANTIZE: {
10722 		uint16_t factor = DTRACE_LLQUANTIZE_FACTOR(desc->dtad_arg);
10723 		uint16_t low = DTRACE_LLQUANTIZE_LOW(desc->dtad_arg);
10724 		uint16_t high = DTRACE_LLQUANTIZE_HIGH(desc->dtad_arg);
10725 		uint16_t nsteps = DTRACE_LLQUANTIZE_NSTEP(desc->dtad_arg);
10726 		int64_t v;
10727 
10728 		agg->dtag_initial = desc->dtad_arg;
10729 		agg->dtag_aggregate = dtrace_aggregate_llquantize;
10730 
10731 		if (factor < 2 || low >= high || nsteps < factor)
10732 			goto err;
10733 
10734 		/*
10735 		 * Now check that the number of steps evenly divides a power
10736 		 * of the factor.  (This assures both integer bucket size and
10737 		 * linearity within each magnitude.)
10738 		 */
10739 		for (v = factor; v < nsteps; v *= factor)
10740 			continue;
10741 
10742 		if ((v % nsteps) || (nsteps % factor))
10743 			goto err;
10744 
10745 		size = (dtrace_aggregate_llquantize_bucket(factor,
10746 		    low, high, nsteps, INT64_MAX) + 2) * sizeof (uint64_t);
10747 		break;
10748 	}
10749 
10750 	case DTRACEAGG_AVG:
10751 		agg->dtag_aggregate = dtrace_aggregate_avg;
10752 		size = sizeof (uint64_t) * 2;
10753 		break;
10754 
10755 	case DTRACEAGG_STDDEV:
10756 		agg->dtag_aggregate = dtrace_aggregate_stddev;
10757 		size = sizeof (uint64_t) * 4;
10758 		break;
10759 
10760 	case DTRACEAGG_SUM:
10761 		agg->dtag_aggregate = dtrace_aggregate_sum;
10762 		break;
10763 
10764 	default:
10765 		goto err;
10766 	}
10767 
10768 	agg->dtag_action.dta_rec.dtrd_size = size;
10769 
10770 	if (ntuple == 0)
10771 		goto err;
10772 
10773 	/*
10774 	 * We must make sure that we have enough actions for the n-tuple.
10775 	 */
10776 	for (act = ecb->dte_action_last; act != NULL; act = act->dta_prev) {
10777 		if (DTRACEACT_ISAGG(act->dta_kind))
10778 			break;
10779 
10780 		if (--ntuple == 0) {
10781 			/*
10782 			 * This is the action with which our n-tuple begins.
10783 			 */
10784 			agg->dtag_first = act;
10785 			goto success;
10786 		}
10787 	}
10788 
10789 	/*
10790 	 * This n-tuple is short by ntuple elements.  Return failure.
10791 	 */
10792 	ASSERT(ntuple != 0);
10793 err:
10794 	kmem_free(agg, sizeof (dtrace_aggregation_t));
10795 	return (NULL);
10796 
10797 success:
10798 	/*
10799 	 * If the last action in the tuple has a size of zero, it's actually
10800 	 * an expression argument for the aggregating action.
10801 	 */
10802 	ASSERT(ecb->dte_action_last != NULL);
10803 	act = ecb->dte_action_last;
10804 
10805 	if (act->dta_kind == DTRACEACT_DIFEXPR) {
10806 		ASSERT(act->dta_difo != NULL);
10807 
10808 		if (act->dta_difo->dtdo_rtype.dtdt_size == 0)
10809 			agg->dtag_hasarg = 1;
10810 	}
10811 
10812 	/*
10813 	 * We need to allocate an id for this aggregation.
10814 	 */
10815 	aggid = (dtrace_aggid_t)(uintptr_t)vmem_alloc(state->dts_aggid_arena, 1,
10816 	    VM_BESTFIT | VM_SLEEP);
10817 
10818 	if (aggid - 1 >= state->dts_naggregations) {
10819 		dtrace_aggregation_t **oaggs = state->dts_aggregations;
10820 		dtrace_aggregation_t **aggs;
10821 		int naggs = state->dts_naggregations << 1;
10822 		int onaggs = state->dts_naggregations;
10823 
10824 		ASSERT(aggid == state->dts_naggregations + 1);
10825 
10826 		if (naggs == 0) {
10827 			ASSERT(oaggs == NULL);
10828 			naggs = 1;
10829 		}
10830 
10831 		aggs = kmem_zalloc(naggs * sizeof (*aggs), KM_SLEEP);
10832 
10833 		if (oaggs != NULL) {
10834 			bcopy(oaggs, aggs, onaggs * sizeof (*aggs));
10835 			kmem_free(oaggs, onaggs * sizeof (*aggs));
10836 		}
10837 
10838 		state->dts_aggregations = aggs;
10839 		state->dts_naggregations = naggs;
10840 	}
10841 
10842 	ASSERT(state->dts_aggregations[aggid - 1] == NULL);
10843 	state->dts_aggregations[(agg->dtag_id = aggid) - 1] = agg;
10844 
10845 	frec = &agg->dtag_first->dta_rec;
10846 	if (frec->dtrd_alignment < sizeof (dtrace_aggid_t))
10847 		frec->dtrd_alignment = sizeof (dtrace_aggid_t);
10848 
10849 	for (act = agg->dtag_first; act != NULL; act = act->dta_next) {
10850 		ASSERT(!act->dta_intuple);
10851 		act->dta_intuple = 1;
10852 	}
10853 
10854 	return (&agg->dtag_action);
10855 }
10856 
10857 static void
10858 dtrace_ecb_aggregation_destroy(dtrace_ecb_t *ecb, dtrace_action_t *act)
10859 {
10860 	dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
10861 	dtrace_state_t *state = ecb->dte_state;
10862 	dtrace_aggid_t aggid = agg->dtag_id;
10863 
10864 	ASSERT(DTRACEACT_ISAGG(act->dta_kind));
10865 	vmem_free(state->dts_aggid_arena, (void *)(uintptr_t)aggid, 1);
10866 
10867 	ASSERT(state->dts_aggregations[aggid - 1] == agg);
10868 	state->dts_aggregations[aggid - 1] = NULL;
10869 
10870 	kmem_free(agg, sizeof (dtrace_aggregation_t));
10871 }
10872 
10873 static int
10874 dtrace_ecb_action_add(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
10875 {
10876 	dtrace_action_t *action, *last;
10877 	dtrace_difo_t *dp = desc->dtad_difo;
10878 	uint32_t size = 0, align = sizeof (uint8_t), mask;
10879 	uint16_t format = 0;
10880 	dtrace_recdesc_t *rec;
10881 	dtrace_state_t *state = ecb->dte_state;
10882 	dtrace_optval_t *opt = state->dts_options, nframes, strsize;
10883 	uint64_t arg = desc->dtad_arg;
10884 
10885 	ASSERT(MUTEX_HELD(&dtrace_lock));
10886 	ASSERT(ecb->dte_action == NULL || ecb->dte_action->dta_refcnt == 1);
10887 
10888 	if (DTRACEACT_ISAGG(desc->dtad_kind)) {
10889 		/*
10890 		 * If this is an aggregating action, there must be neither
10891 		 * a speculate nor a commit on the action chain.
10892 		 */
10893 		dtrace_action_t *act;
10894 
10895 		for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
10896 			if (act->dta_kind == DTRACEACT_COMMIT)
10897 				return (EINVAL);
10898 
10899 			if (act->dta_kind == DTRACEACT_SPECULATE)
10900 				return (EINVAL);
10901 		}
10902 
10903 		action = dtrace_ecb_aggregation_create(ecb, desc);
10904 
10905 		if (action == NULL)
10906 			return (EINVAL);
10907 	} else {
10908 		if (DTRACEACT_ISDESTRUCTIVE(desc->dtad_kind) ||
10909 		    (desc->dtad_kind == DTRACEACT_DIFEXPR &&
10910 		    dp != NULL && dp->dtdo_destructive)) {
10911 			state->dts_destructive = 1;
10912 		}
10913 
10914 		switch (desc->dtad_kind) {
10915 		case DTRACEACT_PRINTF:
10916 		case DTRACEACT_PRINTA:
10917 		case DTRACEACT_SYSTEM:
10918 		case DTRACEACT_FREOPEN:
10919 		case DTRACEACT_DIFEXPR:
10920 			/*
10921 			 * We know that our arg is a string -- turn it into a
10922 			 * format.
10923 			 */
10924 			if (arg == NULL) {
10925 				ASSERT(desc->dtad_kind == DTRACEACT_PRINTA ||
10926 				    desc->dtad_kind == DTRACEACT_DIFEXPR);
10927 				format = 0;
10928 			} else {
10929 				ASSERT(arg != NULL);
10930 				ASSERT(arg > KERNELBASE);
10931 				format = dtrace_format_add(state,
10932 				    (char *)(uintptr_t)arg);
10933 			}
10934 
10935 			/*FALLTHROUGH*/
10936 		case DTRACEACT_LIBACT:
10937 		case DTRACEACT_TRACEMEM:
10938 		case DTRACEACT_TRACEMEM_DYNSIZE:
10939 			if (dp == NULL)
10940 				return (EINVAL);
10941 
10942 			if ((size = dp->dtdo_rtype.dtdt_size) != 0)
10943 				break;
10944 
10945 			if (dp->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING) {
10946 				if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10947 					return (EINVAL);
10948 
10949 				size = opt[DTRACEOPT_STRSIZE];
10950 			}
10951 
10952 			break;
10953 
10954 		case DTRACEACT_STACK:
10955 			if ((nframes = arg) == 0) {
10956 				nframes = opt[DTRACEOPT_STACKFRAMES];
10957 				ASSERT(nframes > 0);
10958 				arg = nframes;
10959 			}
10960 
10961 			size = nframes * sizeof (pc_t);
10962 			break;
10963 
10964 		case DTRACEACT_JSTACK:
10965 			if ((strsize = DTRACE_USTACK_STRSIZE(arg)) == 0)
10966 				strsize = opt[DTRACEOPT_JSTACKSTRSIZE];
10967 
10968 			if ((nframes = DTRACE_USTACK_NFRAMES(arg)) == 0)
10969 				nframes = opt[DTRACEOPT_JSTACKFRAMES];
10970 
10971 			arg = DTRACE_USTACK_ARG(nframes, strsize);
10972 
10973 			/*FALLTHROUGH*/
10974 		case DTRACEACT_USTACK:
10975 			if (desc->dtad_kind != DTRACEACT_JSTACK &&
10976 			    (nframes = DTRACE_USTACK_NFRAMES(arg)) == 0) {
10977 				strsize = DTRACE_USTACK_STRSIZE(arg);
10978 				nframes = opt[DTRACEOPT_USTACKFRAMES];
10979 				ASSERT(nframes > 0);
10980 				arg = DTRACE_USTACK_ARG(nframes, strsize);
10981 			}
10982 
10983 			/*
10984 			 * Save a slot for the pid.
10985 			 */
10986 			size = (nframes + 1) * sizeof (uint64_t);
10987 			size += DTRACE_USTACK_STRSIZE(arg);
10988 			size = P2ROUNDUP(size, (uint32_t)(sizeof (uintptr_t)));
10989 
10990 			break;
10991 
10992 		case DTRACEACT_SYM:
10993 		case DTRACEACT_MOD:
10994 			if (dp == NULL || ((size = dp->dtdo_rtype.dtdt_size) !=
10995 			    sizeof (uint64_t)) ||
10996 			    (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10997 				return (EINVAL);
10998 			break;
10999 
11000 		case DTRACEACT_USYM:
11001 		case DTRACEACT_UMOD:
11002 		case DTRACEACT_UADDR:
11003 			if (dp == NULL ||
11004 			    (dp->dtdo_rtype.dtdt_size != sizeof (uint64_t)) ||
11005 			    (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
11006 				return (EINVAL);
11007 
11008 			/*
11009 			 * We have a slot for the pid, plus a slot for the
11010 			 * argument.  To keep things simple (aligned with
11011 			 * bitness-neutral sizing), we store each as a 64-bit
11012 			 * quantity.
11013 			 */
11014 			size = 2 * sizeof (uint64_t);
11015 			break;
11016 
11017 		case DTRACEACT_STOP:
11018 		case DTRACEACT_BREAKPOINT:
11019 		case DTRACEACT_PANIC:
11020 			break;
11021 
11022 		case DTRACEACT_CHILL:
11023 		case DTRACEACT_DISCARD:
11024 		case DTRACEACT_RAISE:
11025 			if (dp == NULL)
11026 				return (EINVAL);
11027 			break;
11028 
11029 		case DTRACEACT_EXIT:
11030 			if (dp == NULL ||
11031 			    (size = dp->dtdo_rtype.dtdt_size) != sizeof (int) ||
11032 			    (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
11033 				return (EINVAL);
11034 			break;
11035 
11036 		case DTRACEACT_SPECULATE:
11037 			if (ecb->dte_size > sizeof (dtrace_rechdr_t))
11038 				return (EINVAL);
11039 
11040 			if (dp == NULL)
11041 				return (EINVAL);
11042 
11043 			state->dts_speculates = 1;
11044 			break;
11045 
11046 		case DTRACEACT_COMMIT: {
11047 			dtrace_action_t *act = ecb->dte_action;
11048 
11049 			for (; act != NULL; act = act->dta_next) {
11050 				if (act->dta_kind == DTRACEACT_COMMIT)
11051 					return (EINVAL);
11052 			}
11053 
11054 			if (dp == NULL)
11055 				return (EINVAL);
11056 			break;
11057 		}
11058 
11059 		default:
11060 			return (EINVAL);
11061 		}
11062 
11063 		if (size != 0 || desc->dtad_kind == DTRACEACT_SPECULATE) {
11064 			/*
11065 			 * If this is a data-storing action or a speculate,
11066 			 * we must be sure that there isn't a commit on the
11067 			 * action chain.
11068 			 */
11069 			dtrace_action_t *act = ecb->dte_action;
11070 
11071 			for (; act != NULL; act = act->dta_next) {
11072 				if (act->dta_kind == DTRACEACT_COMMIT)
11073 					return (EINVAL);
11074 			}
11075 		}
11076 
11077 		action = kmem_zalloc(sizeof (dtrace_action_t), KM_SLEEP);
11078 		action->dta_rec.dtrd_size = size;
11079 	}
11080 
11081 	action->dta_refcnt = 1;
11082 	rec = &action->dta_rec;
11083 	size = rec->dtrd_size;
11084 
11085 	for (mask = sizeof (uint64_t) - 1; size != 0 && mask > 0; mask >>= 1) {
11086 		if (!(size & mask)) {
11087 			align = mask + 1;
11088 			break;
11089 		}
11090 	}
11091 
11092 	action->dta_kind = desc->dtad_kind;
11093 
11094 	if ((action->dta_difo = dp) != NULL)
11095 		dtrace_difo_hold(dp);
11096 
11097 	rec->dtrd_action = action->dta_kind;
11098 	rec->dtrd_arg = arg;
11099 	rec->dtrd_uarg = desc->dtad_uarg;
11100 	rec->dtrd_alignment = (uint16_t)align;
11101 	rec->dtrd_format = format;
11102 
11103 	if ((last = ecb->dte_action_last) != NULL) {
11104 		ASSERT(ecb->dte_action != NULL);
11105 		action->dta_prev = last;
11106 		last->dta_next = action;
11107 	} else {
11108 		ASSERT(ecb->dte_action == NULL);
11109 		ecb->dte_action = action;
11110 	}
11111 
11112 	ecb->dte_action_last = action;
11113 
11114 	return (0);
11115 }
11116 
11117 static void
11118 dtrace_ecb_action_remove(dtrace_ecb_t *ecb)
11119 {
11120 	dtrace_action_t *act = ecb->dte_action, *next;
11121 	dtrace_vstate_t *vstate = &ecb->dte_state->dts_vstate;
11122 	dtrace_difo_t *dp;
11123 	uint16_t format;
11124 
11125 	if (act != NULL && act->dta_refcnt > 1) {
11126 		ASSERT(act->dta_next == NULL || act->dta_next->dta_refcnt == 1);
11127 		act->dta_refcnt--;
11128 	} else {
11129 		for (; act != NULL; act = next) {
11130 			next = act->dta_next;
11131 			ASSERT(next != NULL || act == ecb->dte_action_last);
11132 			ASSERT(act->dta_refcnt == 1);
11133 
11134 			if ((format = act->dta_rec.dtrd_format) != 0)
11135 				dtrace_format_remove(ecb->dte_state, format);
11136 
11137 			if ((dp = act->dta_difo) != NULL)
11138 				dtrace_difo_release(dp, vstate);
11139 
11140 			if (DTRACEACT_ISAGG(act->dta_kind)) {
11141 				dtrace_ecb_aggregation_destroy(ecb, act);
11142 			} else {
11143 				kmem_free(act, sizeof (dtrace_action_t));
11144 			}
11145 		}
11146 	}
11147 
11148 	ecb->dte_action = NULL;
11149 	ecb->dte_action_last = NULL;
11150 	ecb->dte_size = 0;
11151 }
11152 
11153 static void
11154 dtrace_ecb_disable(dtrace_ecb_t *ecb)
11155 {
11156 	/*
11157 	 * We disable the ECB by removing it from its probe.
11158 	 */
11159 	dtrace_ecb_t *pecb, *prev = NULL;
11160 	dtrace_probe_t *probe = ecb->dte_probe;
11161 
11162 	ASSERT(MUTEX_HELD(&dtrace_lock));
11163 
11164 	if (probe == NULL) {
11165 		/*
11166 		 * This is the NULL probe; there is nothing to disable.
11167 		 */
11168 		return;
11169 	}
11170 
11171 	for (pecb = probe->dtpr_ecb; pecb != NULL; pecb = pecb->dte_next) {
11172 		if (pecb == ecb)
11173 			break;
11174 		prev = pecb;
11175 	}
11176 
11177 	ASSERT(pecb != NULL);
11178 
11179 	if (prev == NULL) {
11180 		probe->dtpr_ecb = ecb->dte_next;
11181 	} else {
11182 		prev->dte_next = ecb->dte_next;
11183 	}
11184 
11185 	if (ecb == probe->dtpr_ecb_last) {
11186 		ASSERT(ecb->dte_next == NULL);
11187 		probe->dtpr_ecb_last = prev;
11188 	}
11189 
11190 	/*
11191 	 * The ECB has been disconnected from the probe; now sync to assure
11192 	 * that all CPUs have seen the change before returning.
11193 	 */
11194 	dtrace_sync();
11195 
11196 	if (probe->dtpr_ecb == NULL) {
11197 		/*
11198 		 * That was the last ECB on the probe; clear the predicate
11199 		 * cache ID for the probe, disable it and sync one more time
11200 		 * to assure that we'll never hit it again.
11201 		 */
11202 		dtrace_provider_t *prov = probe->dtpr_provider;
11203 
11204 		ASSERT(ecb->dte_next == NULL);
11205 		ASSERT(probe->dtpr_ecb_last == NULL);
11206 		probe->dtpr_predcache = DTRACE_CACHEIDNONE;
11207 		prov->dtpv_pops.dtps_disable(prov->dtpv_arg,
11208 		    probe->dtpr_id, probe->dtpr_arg);
11209 		dtrace_sync();
11210 	} else {
11211 		/*
11212 		 * There is at least one ECB remaining on the probe.  If there
11213 		 * is _exactly_ one, set the probe's predicate cache ID to be
11214 		 * the predicate cache ID of the remaining ECB.
11215 		 */
11216 		ASSERT(probe->dtpr_ecb_last != NULL);
11217 		ASSERT(probe->dtpr_predcache == DTRACE_CACHEIDNONE);
11218 
11219 		if (probe->dtpr_ecb == probe->dtpr_ecb_last) {
11220 			dtrace_predicate_t *p = probe->dtpr_ecb->dte_predicate;
11221 
11222 			ASSERT(probe->dtpr_ecb->dte_next == NULL);
11223 
11224 			if (p != NULL)
11225 				probe->dtpr_predcache = p->dtp_cacheid;
11226 		}
11227 
11228 		ecb->dte_next = NULL;
11229 	}
11230 }
11231 
11232 static void
11233 dtrace_ecb_destroy(dtrace_ecb_t *ecb)
11234 {
11235 	dtrace_state_t *state = ecb->dte_state;
11236 	dtrace_vstate_t *vstate = &state->dts_vstate;
11237 	dtrace_predicate_t *pred;
11238 	dtrace_epid_t epid = ecb->dte_epid;
11239 
11240 	ASSERT(MUTEX_HELD(&dtrace_lock));
11241 	ASSERT(ecb->dte_next == NULL);
11242 	ASSERT(ecb->dte_probe == NULL || ecb->dte_probe->dtpr_ecb != ecb);
11243 
11244 	if ((pred = ecb->dte_predicate) != NULL)
11245 		dtrace_predicate_release(pred, vstate);
11246 
11247 	dtrace_ecb_action_remove(ecb);
11248 
11249 	ASSERT(state->dts_ecbs[epid - 1] == ecb);
11250 	state->dts_ecbs[epid - 1] = NULL;
11251 
11252 	kmem_free(ecb, sizeof (dtrace_ecb_t));
11253 }
11254 
11255 static dtrace_ecb_t *
11256 dtrace_ecb_create(dtrace_state_t *state, dtrace_probe_t *probe,
11257     dtrace_enabling_t *enab)
11258 {
11259 	dtrace_ecb_t *ecb;
11260 	dtrace_predicate_t *pred;
11261 	dtrace_actdesc_t *act;
11262 	dtrace_provider_t *prov;
11263 	dtrace_ecbdesc_t *desc = enab->dten_current;
11264 
11265 	ASSERT(MUTEX_HELD(&dtrace_lock));
11266 	ASSERT(state != NULL);
11267 
11268 	ecb = dtrace_ecb_add(state, probe);
11269 	ecb->dte_uarg = desc->dted_uarg;
11270 
11271 	if ((pred = desc->dted_pred.dtpdd_predicate) != NULL) {
11272 		dtrace_predicate_hold(pred);
11273 		ecb->dte_predicate = pred;
11274 	}
11275 
11276 	if (probe != NULL) {
11277 		/*
11278 		 * If the provider shows more leg than the consumer is old
11279 		 * enough to see, we need to enable the appropriate implicit
11280 		 * predicate bits to prevent the ecb from activating at
11281 		 * revealing times.
11282 		 *
11283 		 * Providers specifying DTRACE_PRIV_USER at register time
11284 		 * are stating that they need the /proc-style privilege
11285 		 * model to be enforced, and this is what DTRACE_COND_OWNER
11286 		 * and DTRACE_COND_ZONEOWNER will then do at probe time.
11287 		 */
11288 		prov = probe->dtpr_provider;
11289 		if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLPROC) &&
11290 		    (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
11291 			ecb->dte_cond |= DTRACE_COND_OWNER;
11292 
11293 		if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLZONE) &&
11294 		    (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
11295 			ecb->dte_cond |= DTRACE_COND_ZONEOWNER;
11296 
11297 		/*
11298 		 * If the provider shows us kernel innards and the user
11299 		 * is lacking sufficient privilege, enable the
11300 		 * DTRACE_COND_USERMODE implicit predicate.
11301 		 */
11302 		if (!(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) &&
11303 		    (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_KERNEL))
11304 			ecb->dte_cond |= DTRACE_COND_USERMODE;
11305 	}
11306 
11307 	if (dtrace_ecb_create_cache != NULL) {
11308 		/*
11309 		 * If we have a cached ecb, we'll use its action list instead
11310 		 * of creating our own (saving both time and space).
11311 		 */
11312 		dtrace_ecb_t *cached = dtrace_ecb_create_cache;
11313 		dtrace_action_t *act = cached->dte_action;
11314 
11315 		if (act != NULL) {
11316 			ASSERT(act->dta_refcnt > 0);
11317 			act->dta_refcnt++;
11318 			ecb->dte_action = act;
11319 			ecb->dte_action_last = cached->dte_action_last;
11320 			ecb->dte_needed = cached->dte_needed;
11321 			ecb->dte_size = cached->dte_size;
11322 			ecb->dte_alignment = cached->dte_alignment;
11323 		}
11324 
11325 		return (ecb);
11326 	}
11327 
11328 	for (act = desc->dted_action; act != NULL; act = act->dtad_next) {
11329 		if ((enab->dten_error = dtrace_ecb_action_add(ecb, act)) != 0) {
11330 			dtrace_ecb_destroy(ecb);
11331 			return (NULL);
11332 		}
11333 	}
11334 
11335 	if ((enab->dten_error = dtrace_ecb_resize(ecb)) != 0) {
11336 		dtrace_ecb_destroy(ecb);
11337 		return (NULL);
11338 	}
11339 
11340 	return (dtrace_ecb_create_cache = ecb);
11341 }
11342 
11343 static int
11344 dtrace_ecb_create_enable(dtrace_probe_t *probe, void *arg)
11345 {
11346 	dtrace_ecb_t *ecb;
11347 	dtrace_enabling_t *enab = arg;
11348 	dtrace_state_t *state = enab->dten_vstate->dtvs_state;
11349 
11350 	ASSERT(state != NULL);
11351 
11352 	if (probe != NULL && probe->dtpr_gen < enab->dten_probegen) {
11353 		/*
11354 		 * This probe was created in a generation for which this
11355 		 * enabling has previously created ECBs; we don't want to
11356 		 * enable it again, so just kick out.
11357 		 */
11358 		return (DTRACE_MATCH_NEXT);
11359 	}
11360 
11361 	if ((ecb = dtrace_ecb_create(state, probe, enab)) == NULL)
11362 		return (DTRACE_MATCH_DONE);
11363 
11364 	if (dtrace_ecb_enable(ecb) < 0)
11365 		return (DTRACE_MATCH_FAIL);
11366 
11367 	return (DTRACE_MATCH_NEXT);
11368 }
11369 
11370 static dtrace_ecb_t *
11371 dtrace_epid2ecb(dtrace_state_t *state, dtrace_epid_t id)
11372 {
11373 	dtrace_ecb_t *ecb;
11374 
11375 	ASSERT(MUTEX_HELD(&dtrace_lock));
11376 
11377 	if (id == 0 || id > state->dts_necbs)
11378 		return (NULL);
11379 
11380 	ASSERT(state->dts_necbs > 0 && state->dts_ecbs != NULL);
11381 	ASSERT((ecb = state->dts_ecbs[id - 1]) == NULL || ecb->dte_epid == id);
11382 
11383 	return (state->dts_ecbs[id - 1]);
11384 }
11385 
11386 static dtrace_aggregation_t *
11387 dtrace_aggid2agg(dtrace_state_t *state, dtrace_aggid_t id)
11388 {
11389 	dtrace_aggregation_t *agg;
11390 
11391 	ASSERT(MUTEX_HELD(&dtrace_lock));
11392 
11393 	if (id == 0 || id > state->dts_naggregations)
11394 		return (NULL);
11395 
11396 	ASSERT(state->dts_naggregations > 0 && state->dts_aggregations != NULL);
11397 	ASSERT((agg = state->dts_aggregations[id - 1]) == NULL ||
11398 	    agg->dtag_id == id);
11399 
11400 	return (state->dts_aggregations[id - 1]);
11401 }
11402 
11403 /*
11404  * DTrace Buffer Functions
11405  *
11406  * The following functions manipulate DTrace buffers.  Most of these functions
11407  * are called in the context of establishing or processing consumer state;
11408  * exceptions are explicitly noted.
11409  */
11410 
11411 /*
11412  * Note:  called from cross call context.  This function switches the two
11413  * buffers on a given CPU.  The atomicity of this operation is assured by
11414  * disabling interrupts while the actual switch takes place; the disabling of
11415  * interrupts serializes the execution with any execution of dtrace_probe() on
11416  * the same CPU.
11417  */
11418 static void
11419 dtrace_buffer_switch(dtrace_buffer_t *buf)
11420 {
11421 	caddr_t tomax = buf->dtb_tomax;
11422 	caddr_t xamot = buf->dtb_xamot;
11423 	dtrace_icookie_t cookie;
11424 	hrtime_t now;
11425 
11426 	ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
11427 	ASSERT(!(buf->dtb_flags & DTRACEBUF_RING));
11428 
11429 	cookie = dtrace_interrupt_disable();
11430 	now = dtrace_gethrtime();
11431 	buf->dtb_tomax = xamot;
11432 	buf->dtb_xamot = tomax;
11433 	buf->dtb_xamot_drops = buf->dtb_drops;
11434 	buf->dtb_xamot_offset = buf->dtb_offset;
11435 	buf->dtb_xamot_errors = buf->dtb_errors;
11436 	buf->dtb_xamot_flags = buf->dtb_flags;
11437 	buf->dtb_offset = 0;
11438 	buf->dtb_drops = 0;
11439 	buf->dtb_errors = 0;
11440 	buf->dtb_flags &= ~(DTRACEBUF_ERROR | DTRACEBUF_DROPPED);
11441 	buf->dtb_interval = now - buf->dtb_switched;
11442 	buf->dtb_switched = now;
11443 	dtrace_interrupt_enable(cookie);
11444 }
11445 
11446 /*
11447  * Note:  called from cross call context.  This function activates a buffer
11448  * on a CPU.  As with dtrace_buffer_switch(), the atomicity of the operation
11449  * is guaranteed by the disabling of interrupts.
11450  */
11451 static void
11452 dtrace_buffer_activate(dtrace_state_t *state)
11453 {
11454 	dtrace_buffer_t *buf;
11455 	dtrace_icookie_t cookie = dtrace_interrupt_disable();
11456 
11457 	buf = &state->dts_buffer[CPU->cpu_id];
11458 
11459 	if (buf->dtb_tomax != NULL) {
11460 		/*
11461 		 * We might like to assert that the buffer is marked inactive,
11462 		 * but this isn't necessarily true:  the buffer for the CPU
11463 		 * that processes the BEGIN probe has its buffer activated
11464 		 * manually.  In this case, we take the (harmless) action
11465 		 * re-clearing the bit INACTIVE bit.
11466 		 */
11467 		buf->dtb_flags &= ~DTRACEBUF_INACTIVE;
11468 	}
11469 
11470 	dtrace_interrupt_enable(cookie);
11471 }
11472 
11473 static int
11474 dtrace_buffer_alloc(dtrace_buffer_t *bufs, size_t size, int flags,
11475     processorid_t cpu, int *factor)
11476 {
11477 	cpu_t *cp;
11478 	dtrace_buffer_t *buf;
11479 	int allocated = 0, desired = 0;
11480 
11481 	ASSERT(MUTEX_HELD(&cpu_lock));
11482 	ASSERT(MUTEX_HELD(&dtrace_lock));
11483 
11484 	*factor = 1;
11485 
11486 	if (size > dtrace_nonroot_maxsize &&
11487 	    !PRIV_POLICY_CHOICE(CRED(), PRIV_ALL, B_FALSE))
11488 		return (EFBIG);
11489 
11490 	cp = cpu_list;
11491 
11492 	do {
11493 		if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
11494 			continue;
11495 
11496 		buf = &bufs[cp->cpu_id];
11497 
11498 		/*
11499 		 * If there is already a buffer allocated for this CPU, it
11500 		 * is only possible that this is a DR event.  In this case,
11501 		 * the buffer size must match our specified size.
11502 		 */
11503 		if (buf->dtb_tomax != NULL) {
11504 			ASSERT(buf->dtb_size == size);
11505 			continue;
11506 		}
11507 
11508 		ASSERT(buf->dtb_xamot == NULL);
11509 
11510 		if ((buf->dtb_tomax = kmem_zalloc(size,
11511 		    KM_NOSLEEP | KM_NORMALPRI)) == NULL)
11512 			goto err;
11513 
11514 		buf->dtb_size = size;
11515 		buf->dtb_flags = flags;
11516 		buf->dtb_offset = 0;
11517 		buf->dtb_drops = 0;
11518 
11519 		if (flags & DTRACEBUF_NOSWITCH)
11520 			continue;
11521 
11522 		if ((buf->dtb_xamot = kmem_zalloc(size,
11523 		    KM_NOSLEEP | KM_NORMALPRI)) == NULL)
11524 			goto err;
11525 	} while ((cp = cp->cpu_next) != cpu_list);
11526 
11527 	return (0);
11528 
11529 err:
11530 	cp = cpu_list;
11531 
11532 	do {
11533 		if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
11534 			continue;
11535 
11536 		buf = &bufs[cp->cpu_id];
11537 		desired += 2;
11538 
11539 		if (buf->dtb_xamot != NULL) {
11540 			ASSERT(buf->dtb_tomax != NULL);
11541 			ASSERT(buf->dtb_size == size);
11542 			kmem_free(buf->dtb_xamot, size);
11543 			allocated++;
11544 		}
11545 
11546 		if (buf->dtb_tomax != NULL) {
11547 			ASSERT(buf->dtb_size == size);
11548 			kmem_free(buf->dtb_tomax, size);
11549 			allocated++;
11550 		}
11551 
11552 		buf->dtb_tomax = NULL;
11553 		buf->dtb_xamot = NULL;
11554 		buf->dtb_size = 0;
11555 	} while ((cp = cp->cpu_next) != cpu_list);
11556 
11557 	*factor = desired / (allocated > 0 ? allocated : 1);
11558 
11559 	return (ENOMEM);
11560 }
11561 
11562 /*
11563  * Note:  called from probe context.  This function just increments the drop
11564  * count on a buffer.  It has been made a function to allow for the
11565  * possibility of understanding the source of mysterious drop counts.  (A
11566  * problem for which one may be particularly disappointed that DTrace cannot
11567  * be used to understand DTrace.)
11568  */
11569 static void
11570 dtrace_buffer_drop(dtrace_buffer_t *buf)
11571 {
11572 	buf->dtb_drops++;
11573 }
11574 
11575 /*
11576  * Note:  called from probe context.  This function is called to reserve space
11577  * in a buffer.  If mstate is non-NULL, sets the scratch base and size in the
11578  * mstate.  Returns the new offset in the buffer, or a negative value if an
11579  * error has occurred.
11580  */
11581 static intptr_t
11582 dtrace_buffer_reserve(dtrace_buffer_t *buf, size_t needed, size_t align,
11583     dtrace_state_t *state, dtrace_mstate_t *mstate)
11584 {
11585 	intptr_t offs = buf->dtb_offset, soffs;
11586 	intptr_t woffs;
11587 	caddr_t tomax;
11588 	size_t total;
11589 
11590 	if (buf->dtb_flags & DTRACEBUF_INACTIVE)
11591 		return (-1);
11592 
11593 	if ((tomax = buf->dtb_tomax) == NULL) {
11594 		dtrace_buffer_drop(buf);
11595 		return (-1);
11596 	}
11597 
11598 	if (!(buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL))) {
11599 		while (offs & (align - 1)) {
11600 			/*
11601 			 * Assert that our alignment is off by a number which
11602 			 * is itself sizeof (uint32_t) aligned.
11603 			 */
11604 			ASSERT(!((align - (offs & (align - 1))) &
11605 			    (sizeof (uint32_t) - 1)));
11606 			DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
11607 			offs += sizeof (uint32_t);
11608 		}
11609 
11610 		if ((soffs = offs + needed) > buf->dtb_size) {
11611 			dtrace_buffer_drop(buf);
11612 			return (-1);
11613 		}
11614 
11615 		if (mstate == NULL)
11616 			return (offs);
11617 
11618 		mstate->dtms_scratch_base = (uintptr_t)tomax + soffs;
11619 		mstate->dtms_scratch_size = buf->dtb_size - soffs;
11620 		mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
11621 
11622 		return (offs);
11623 	}
11624 
11625 	if (buf->dtb_flags & DTRACEBUF_FILL) {
11626 		if (state->dts_activity != DTRACE_ACTIVITY_COOLDOWN &&
11627 		    (buf->dtb_flags & DTRACEBUF_FULL))
11628 			return (-1);
11629 		goto out;
11630 	}
11631 
11632 	total = needed + (offs & (align - 1));
11633 
11634 	/*
11635 	 * For a ring buffer, life is quite a bit more complicated.  Before
11636 	 * we can store any padding, we need to adjust our wrapping offset.
11637 	 * (If we've never before wrapped or we're not about to, no adjustment
11638 	 * is required.)
11639 	 */
11640 	if ((buf->dtb_flags & DTRACEBUF_WRAPPED) ||
11641 	    offs + total > buf->dtb_size) {
11642 		woffs = buf->dtb_xamot_offset;
11643 
11644 		if (offs + total > buf->dtb_size) {
11645 			/*
11646 			 * We can't fit in the end of the buffer.  First, a
11647 			 * sanity check that we can fit in the buffer at all.
11648 			 */
11649 			if (total > buf->dtb_size) {
11650 				dtrace_buffer_drop(buf);
11651 				return (-1);
11652 			}
11653 
11654 			/*
11655 			 * We're going to be storing at the top of the buffer,
11656 			 * so now we need to deal with the wrapped offset.  We
11657 			 * only reset our wrapped offset to 0 if it is
11658 			 * currently greater than the current offset.  If it
11659 			 * is less than the current offset, it is because a
11660 			 * previous allocation induced a wrap -- but the
11661 			 * allocation didn't subsequently take the space due
11662 			 * to an error or false predicate evaluation.  In this
11663 			 * case, we'll just leave the wrapped offset alone: if
11664 			 * the wrapped offset hasn't been advanced far enough
11665 			 * for this allocation, it will be adjusted in the
11666 			 * lower loop.
11667 			 */
11668 			if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
11669 				if (woffs >= offs)
11670 					woffs = 0;
11671 			} else {
11672 				woffs = 0;
11673 			}
11674 
11675 			/*
11676 			 * Now we know that we're going to be storing to the
11677 			 * top of the buffer and that there is room for us
11678 			 * there.  We need to clear the buffer from the current
11679 			 * offset to the end (there may be old gunk there).
11680 			 */
11681 			while (offs < buf->dtb_size)
11682 				tomax[offs++] = 0;
11683 
11684 			/*
11685 			 * We need to set our offset to zero.  And because we
11686 			 * are wrapping, we need to set the bit indicating as
11687 			 * much.  We can also adjust our needed space back
11688 			 * down to the space required by the ECB -- we know
11689 			 * that the top of the buffer is aligned.
11690 			 */
11691 			offs = 0;
11692 			total = needed;
11693 			buf->dtb_flags |= DTRACEBUF_WRAPPED;
11694 		} else {
11695 			/*
11696 			 * There is room for us in the buffer, so we simply
11697 			 * need to check the wrapped offset.
11698 			 */
11699 			if (woffs < offs) {
11700 				/*
11701 				 * The wrapped offset is less than the offset.
11702 				 * This can happen if we allocated buffer space
11703 				 * that induced a wrap, but then we didn't
11704 				 * subsequently take the space due to an error
11705 				 * or false predicate evaluation.  This is
11706 				 * okay; we know that _this_ allocation isn't
11707 				 * going to induce a wrap.  We still can't
11708 				 * reset the wrapped offset to be zero,
11709 				 * however: the space may have been trashed in
11710 				 * the previous failed probe attempt.  But at
11711 				 * least the wrapped offset doesn't need to
11712 				 * be adjusted at all...
11713 				 */
11714 				goto out;
11715 			}
11716 		}
11717 
11718 		while (offs + total > woffs) {
11719 			dtrace_epid_t epid = *(uint32_t *)(tomax + woffs);
11720 			size_t size;
11721 
11722 			if (epid == DTRACE_EPIDNONE) {
11723 				size = sizeof (uint32_t);
11724 			} else {
11725 				ASSERT3U(epid, <=, state->dts_necbs);
11726 				ASSERT(state->dts_ecbs[epid - 1] != NULL);
11727 
11728 				size = state->dts_ecbs[epid - 1]->dte_size;
11729 			}
11730 
11731 			ASSERT(woffs + size <= buf->dtb_size);
11732 			ASSERT(size != 0);
11733 
11734 			if (woffs + size == buf->dtb_size) {
11735 				/*
11736 				 * We've reached the end of the buffer; we want
11737 				 * to set the wrapped offset to 0 and break
11738 				 * out.  However, if the offs is 0, then we're
11739 				 * in a strange edge-condition:  the amount of
11740 				 * space that we want to reserve plus the size
11741 				 * of the record that we're overwriting is
11742 				 * greater than the size of the buffer.  This
11743 				 * is problematic because if we reserve the
11744 				 * space but subsequently don't consume it (due
11745 				 * to a failed predicate or error) the wrapped
11746 				 * offset will be 0 -- yet the EPID at offset 0
11747 				 * will not be committed.  This situation is
11748 				 * relatively easy to deal with:  if we're in
11749 				 * this case, the buffer is indistinguishable
11750 				 * from one that hasn't wrapped; we need only
11751 				 * finish the job by clearing the wrapped bit,
11752 				 * explicitly setting the offset to be 0, and
11753 				 * zero'ing out the old data in the buffer.
11754 				 */
11755 				if (offs == 0) {
11756 					buf->dtb_flags &= ~DTRACEBUF_WRAPPED;
11757 					buf->dtb_offset = 0;
11758 					woffs = total;
11759 
11760 					while (woffs < buf->dtb_size)
11761 						tomax[woffs++] = 0;
11762 				}
11763 
11764 				woffs = 0;
11765 				break;
11766 			}
11767 
11768 			woffs += size;
11769 		}
11770 
11771 		/*
11772 		 * We have a wrapped offset.  It may be that the wrapped offset
11773 		 * has become zero -- that's okay.
11774 		 */
11775 		buf->dtb_xamot_offset = woffs;
11776 	}
11777 
11778 out:
11779 	/*
11780 	 * Now we can plow the buffer with any necessary padding.
11781 	 */
11782 	while (offs & (align - 1)) {
11783 		/*
11784 		 * Assert that our alignment is off by a number which
11785 		 * is itself sizeof (uint32_t) aligned.
11786 		 */
11787 		ASSERT(!((align - (offs & (align - 1))) &
11788 		    (sizeof (uint32_t) - 1)));
11789 		DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
11790 		offs += sizeof (uint32_t);
11791 	}
11792 
11793 	if (buf->dtb_flags & DTRACEBUF_FILL) {
11794 		if (offs + needed > buf->dtb_size - state->dts_reserve) {
11795 			buf->dtb_flags |= DTRACEBUF_FULL;
11796 			return (-1);
11797 		}
11798 	}
11799 
11800 	if (mstate == NULL)
11801 		return (offs);
11802 
11803 	/*
11804 	 * For ring buffers and fill buffers, the scratch space is always
11805 	 * the inactive buffer.
11806 	 */
11807 	mstate->dtms_scratch_base = (uintptr_t)buf->dtb_xamot;
11808 	mstate->dtms_scratch_size = buf->dtb_size;
11809 	mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
11810 
11811 	return (offs);
11812 }
11813 
11814 static void
11815 dtrace_buffer_polish(dtrace_buffer_t *buf)
11816 {
11817 	ASSERT(buf->dtb_flags & DTRACEBUF_RING);
11818 	ASSERT(MUTEX_HELD(&dtrace_lock));
11819 
11820 	if (!(buf->dtb_flags & DTRACEBUF_WRAPPED))
11821 		return;
11822 
11823 	/*
11824 	 * We need to polish the ring buffer.  There are three cases:
11825 	 *
11826 	 * - The first (and presumably most common) is that there is no gap
11827 	 *   between the buffer offset and the wrapped offset.  In this case,
11828 	 *   there is nothing in the buffer that isn't valid data; we can
11829 	 *   mark the buffer as polished and return.
11830 	 *
11831 	 * - The second (less common than the first but still more common
11832 	 *   than the third) is that there is a gap between the buffer offset
11833 	 *   and the wrapped offset, and the wrapped offset is larger than the
11834 	 *   buffer offset.  This can happen because of an alignment issue, or
11835 	 *   can happen because of a call to dtrace_buffer_reserve() that
11836 	 *   didn't subsequently consume the buffer space.  In this case,
11837 	 *   we need to zero the data from the buffer offset to the wrapped
11838 	 *   offset.
11839 	 *
11840 	 * - The third (and least common) is that there is a gap between the
11841 	 *   buffer offset and the wrapped offset, but the wrapped offset is
11842 	 *   _less_ than the buffer offset.  This can only happen because a
11843 	 *   call to dtrace_buffer_reserve() induced a wrap, but the space
11844 	 *   was not subsequently consumed.  In this case, we need to zero the
11845 	 *   space from the offset to the end of the buffer _and_ from the
11846 	 *   top of the buffer to the wrapped offset.
11847 	 */
11848 	if (buf->dtb_offset < buf->dtb_xamot_offset) {
11849 		bzero(buf->dtb_tomax + buf->dtb_offset,
11850 		    buf->dtb_xamot_offset - buf->dtb_offset);
11851 	}
11852 
11853 	if (buf->dtb_offset > buf->dtb_xamot_offset) {
11854 		bzero(buf->dtb_tomax + buf->dtb_offset,
11855 		    buf->dtb_size - buf->dtb_offset);
11856 		bzero(buf->dtb_tomax, buf->dtb_xamot_offset);
11857 	}
11858 }
11859 
11860 /*
11861  * This routine determines if data generated at the specified time has likely
11862  * been entirely consumed at user-level.  This routine is called to determine
11863  * if an ECB on a defunct probe (but for an active enabling) can be safely
11864  * disabled and destroyed.
11865  */
11866 static int
11867 dtrace_buffer_consumed(dtrace_buffer_t *bufs, hrtime_t when)
11868 {
11869 	int i;
11870 
11871 	for (i = 0; i < NCPU; i++) {
11872 		dtrace_buffer_t *buf = &bufs[i];
11873 
11874 		if (buf->dtb_size == 0)
11875 			continue;
11876 
11877 		if (buf->dtb_flags & DTRACEBUF_RING)
11878 			return (0);
11879 
11880 		if (!buf->dtb_switched && buf->dtb_offset != 0)
11881 			return (0);
11882 
11883 		if (buf->dtb_switched - buf->dtb_interval < when)
11884 			return (0);
11885 	}
11886 
11887 	return (1);
11888 }
11889 
11890 static void
11891 dtrace_buffer_free(dtrace_buffer_t *bufs)
11892 {
11893 	int i;
11894 
11895 	for (i = 0; i < NCPU; i++) {
11896 		dtrace_buffer_t *buf = &bufs[i];
11897 
11898 		if (buf->dtb_tomax == NULL) {
11899 			ASSERT(buf->dtb_xamot == NULL);
11900 			ASSERT(buf->dtb_size == 0);
11901 			continue;
11902 		}
11903 
11904 		if (buf->dtb_xamot != NULL) {
11905 			ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
11906 			kmem_free(buf->dtb_xamot, buf->dtb_size);
11907 		}
11908 
11909 		kmem_free(buf->dtb_tomax, buf->dtb_size);
11910 		buf->dtb_size = 0;
11911 		buf->dtb_tomax = NULL;
11912 		buf->dtb_xamot = NULL;
11913 	}
11914 }
11915 
11916 /*
11917  * DTrace Enabling Functions
11918  */
11919 static dtrace_enabling_t *
11920 dtrace_enabling_create(dtrace_vstate_t *vstate)
11921 {
11922 	dtrace_enabling_t *enab;
11923 
11924 	enab = kmem_zalloc(sizeof (dtrace_enabling_t), KM_SLEEP);
11925 	enab->dten_vstate = vstate;
11926 
11927 	return (enab);
11928 }
11929 
11930 static void
11931 dtrace_enabling_add(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb)
11932 {
11933 	dtrace_ecbdesc_t **ndesc;
11934 	size_t osize, nsize;
11935 
11936 	/*
11937 	 * We can't add to enablings after we've enabled them, or after we've
11938 	 * retained them.
11939 	 */
11940 	ASSERT(enab->dten_probegen == 0);
11941 	ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
11942 
11943 	if (enab->dten_ndesc < enab->dten_maxdesc) {
11944 		enab->dten_desc[enab->dten_ndesc++] = ecb;
11945 		return;
11946 	}
11947 
11948 	osize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
11949 
11950 	if (enab->dten_maxdesc == 0) {
11951 		enab->dten_maxdesc = 1;
11952 	} else {
11953 		enab->dten_maxdesc <<= 1;
11954 	}
11955 
11956 	ASSERT(enab->dten_ndesc < enab->dten_maxdesc);
11957 
11958 	nsize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
11959 	ndesc = kmem_zalloc(nsize, KM_SLEEP);
11960 	bcopy(enab->dten_desc, ndesc, osize);
11961 	kmem_free(enab->dten_desc, osize);
11962 
11963 	enab->dten_desc = ndesc;
11964 	enab->dten_desc[enab->dten_ndesc++] = ecb;
11965 }
11966 
11967 static void
11968 dtrace_enabling_addlike(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb,
11969     dtrace_probedesc_t *pd)
11970 {
11971 	dtrace_ecbdesc_t *new;
11972 	dtrace_predicate_t *pred;
11973 	dtrace_actdesc_t *act;
11974 
11975 	/*
11976 	 * We're going to create a new ECB description that matches the
11977 	 * specified ECB in every way, but has the specified probe description.
11978 	 */
11979 	new = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
11980 
11981 	if ((pred = ecb->dted_pred.dtpdd_predicate) != NULL)
11982 		dtrace_predicate_hold(pred);
11983 
11984 	for (act = ecb->dted_action; act != NULL; act = act->dtad_next)
11985 		dtrace_actdesc_hold(act);
11986 
11987 	new->dted_action = ecb->dted_action;
11988 	new->dted_pred = ecb->dted_pred;
11989 	new->dted_probe = *pd;
11990 	new->dted_uarg = ecb->dted_uarg;
11991 
11992 	dtrace_enabling_add(enab, new);
11993 }
11994 
11995 static void
11996 dtrace_enabling_dump(dtrace_enabling_t *enab)
11997 {
11998 	int i;
11999 
12000 	for (i = 0; i < enab->dten_ndesc; i++) {
12001 		dtrace_probedesc_t *desc = &enab->dten_desc[i]->dted_probe;
12002 
12003 		cmn_err(CE_NOTE, "enabling probe %d (%s:%s:%s:%s)", i,
12004 		    desc->dtpd_provider, desc->dtpd_mod,
12005 		    desc->dtpd_func, desc->dtpd_name);
12006 	}
12007 }
12008 
12009 static void
12010 dtrace_enabling_destroy(dtrace_enabling_t *enab)
12011 {
12012 	int i;
12013 	dtrace_ecbdesc_t *ep;
12014 	dtrace_vstate_t *vstate = enab->dten_vstate;
12015 
12016 	ASSERT(MUTEX_HELD(&dtrace_lock));
12017 
12018 	for (i = 0; i < enab->dten_ndesc; i++) {
12019 		dtrace_actdesc_t *act, *next;
12020 		dtrace_predicate_t *pred;
12021 
12022 		ep = enab->dten_desc[i];
12023 
12024 		if ((pred = ep->dted_pred.dtpdd_predicate) != NULL)
12025 			dtrace_predicate_release(pred, vstate);
12026 
12027 		for (act = ep->dted_action; act != NULL; act = next) {
12028 			next = act->dtad_next;
12029 			dtrace_actdesc_release(act, vstate);
12030 		}
12031 
12032 		kmem_free(ep, sizeof (dtrace_ecbdesc_t));
12033 	}
12034 
12035 	kmem_free(enab->dten_desc,
12036 	    enab->dten_maxdesc * sizeof (dtrace_enabling_t *));
12037 
12038 	/*
12039 	 * If this was a retained enabling, decrement the dts_nretained count
12040 	 * and take it off of the dtrace_retained list.
12041 	 */
12042 	if (enab->dten_prev != NULL || enab->dten_next != NULL ||
12043 	    dtrace_retained == enab) {
12044 		ASSERT(enab->dten_vstate->dtvs_state != NULL);
12045 		ASSERT(enab->dten_vstate->dtvs_state->dts_nretained > 0);
12046 		enab->dten_vstate->dtvs_state->dts_nretained--;
12047 		dtrace_retained_gen++;
12048 	}
12049 
12050 	if (enab->dten_prev == NULL) {
12051 		if (dtrace_retained == enab) {
12052 			dtrace_retained = enab->dten_next;
12053 
12054 			if (dtrace_retained != NULL)
12055 				dtrace_retained->dten_prev = NULL;
12056 		}
12057 	} else {
12058 		ASSERT(enab != dtrace_retained);
12059 		ASSERT(dtrace_retained != NULL);
12060 		enab->dten_prev->dten_next = enab->dten_next;
12061 	}
12062 
12063 	if (enab->dten_next != NULL) {
12064 		ASSERT(dtrace_retained != NULL);
12065 		enab->dten_next->dten_prev = enab->dten_prev;
12066 	}
12067 
12068 	kmem_free(enab, sizeof (dtrace_enabling_t));
12069 }
12070 
12071 static int
12072 dtrace_enabling_retain(dtrace_enabling_t *enab)
12073 {
12074 	dtrace_state_t *state;
12075 
12076 	ASSERT(MUTEX_HELD(&dtrace_lock));
12077 	ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
12078 	ASSERT(enab->dten_vstate != NULL);
12079 
12080 	state = enab->dten_vstate->dtvs_state;
12081 	ASSERT(state != NULL);
12082 
12083 	/*
12084 	 * We only allow each state to retain dtrace_retain_max enablings.
12085 	 */
12086 	if (state->dts_nretained >= dtrace_retain_max)
12087 		return (ENOSPC);
12088 
12089 	state->dts_nretained++;
12090 	dtrace_retained_gen++;
12091 
12092 	if (dtrace_retained == NULL) {
12093 		dtrace_retained = enab;
12094 		return (0);
12095 	}
12096 
12097 	enab->dten_next = dtrace_retained;
12098 	dtrace_retained->dten_prev = enab;
12099 	dtrace_retained = enab;
12100 
12101 	return (0);
12102 }
12103 
12104 static int
12105 dtrace_enabling_replicate(dtrace_state_t *state, dtrace_probedesc_t *match,
12106     dtrace_probedesc_t *create)
12107 {
12108 	dtrace_enabling_t *new, *enab;
12109 	int found = 0, err = ENOENT;
12110 
12111 	ASSERT(MUTEX_HELD(&dtrace_lock));
12112 	ASSERT(strlen(match->dtpd_provider) < DTRACE_PROVNAMELEN);
12113 	ASSERT(strlen(match->dtpd_mod) < DTRACE_MODNAMELEN);
12114 	ASSERT(strlen(match->dtpd_func) < DTRACE_FUNCNAMELEN);
12115 	ASSERT(strlen(match->dtpd_name) < DTRACE_NAMELEN);
12116 
12117 	new = dtrace_enabling_create(&state->dts_vstate);
12118 
12119 	/*
12120 	 * Iterate over all retained enablings, looking for enablings that
12121 	 * match the specified state.
12122 	 */
12123 	for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
12124 		int i;
12125 
12126 		/*
12127 		 * dtvs_state can only be NULL for helper enablings -- and
12128 		 * helper enablings can't be retained.
12129 		 */
12130 		ASSERT(enab->dten_vstate->dtvs_state != NULL);
12131 
12132 		if (enab->dten_vstate->dtvs_state != state)
12133 			continue;
12134 
12135 		/*
12136 		 * Now iterate over each probe description; we're looking for
12137 		 * an exact match to the specified probe description.
12138 		 */
12139 		for (i = 0; i < enab->dten_ndesc; i++) {
12140 			dtrace_ecbdesc_t *ep = enab->dten_desc[i];
12141 			dtrace_probedesc_t *pd = &ep->dted_probe;
12142 
12143 			if (strcmp(pd->dtpd_provider, match->dtpd_provider))
12144 				continue;
12145 
12146 			if (strcmp(pd->dtpd_mod, match->dtpd_mod))
12147 				continue;
12148 
12149 			if (strcmp(pd->dtpd_func, match->dtpd_func))
12150 				continue;
12151 
12152 			if (strcmp(pd->dtpd_name, match->dtpd_name))
12153 				continue;
12154 
12155 			/*
12156 			 * We have a winning probe!  Add it to our growing
12157 			 * enabling.
12158 			 */
12159 			found = 1;
12160 			dtrace_enabling_addlike(new, ep, create);
12161 		}
12162 	}
12163 
12164 	if (!found || (err = dtrace_enabling_retain(new)) != 0) {
12165 		dtrace_enabling_destroy(new);
12166 		return (err);
12167 	}
12168 
12169 	return (0);
12170 }
12171 
12172 static void
12173 dtrace_enabling_retract(dtrace_state_t *state)
12174 {
12175 	dtrace_enabling_t *enab, *next;
12176 
12177 	ASSERT(MUTEX_HELD(&dtrace_lock));
12178 
12179 	/*
12180 	 * Iterate over all retained enablings, destroy the enablings retained
12181 	 * for the specified state.
12182 	 */
12183 	for (enab = dtrace_retained; enab != NULL; enab = next) {
12184 		next = enab->dten_next;
12185 
12186 		/*
12187 		 * dtvs_state can only be NULL for helper enablings -- and
12188 		 * helper enablings can't be retained.
12189 		 */
12190 		ASSERT(enab->dten_vstate->dtvs_state != NULL);
12191 
12192 		if (enab->dten_vstate->dtvs_state == state) {
12193 			ASSERT(state->dts_nretained > 0);
12194 			dtrace_enabling_destroy(enab);
12195 		}
12196 	}
12197 
12198 	ASSERT(state->dts_nretained == 0);
12199 }
12200 
12201 static int
12202 dtrace_enabling_match(dtrace_enabling_t *enab, int *nmatched)
12203 {
12204 	int i = 0;
12205 	int total_matched = 0, matched = 0;
12206 
12207 	ASSERT(MUTEX_HELD(&cpu_lock));
12208 	ASSERT(MUTEX_HELD(&dtrace_lock));
12209 
12210 	for (i = 0; i < enab->dten_ndesc; i++) {
12211 		dtrace_ecbdesc_t *ep = enab->dten_desc[i];
12212 
12213 		enab->dten_current = ep;
12214 		enab->dten_error = 0;
12215 
12216 		/*
12217 		 * If a provider failed to enable a probe then get out and
12218 		 * let the consumer know we failed.
12219 		 */
12220 		if ((matched = dtrace_probe_enable(&ep->dted_probe, enab)) < 0)
12221 			return (EBUSY);
12222 
12223 		total_matched += matched;
12224 
12225 		if (enab->dten_error != 0) {
12226 			/*
12227 			 * If we get an error half-way through enabling the
12228 			 * probes, we kick out -- perhaps with some number of
12229 			 * them enabled.  Leaving enabled probes enabled may
12230 			 * be slightly confusing for user-level, but we expect
12231 			 * that no one will attempt to actually drive on in
12232 			 * the face of such errors.  If this is an anonymous
12233 			 * enabling (indicated with a NULL nmatched pointer),
12234 			 * we cmn_err() a message.  We aren't expecting to
12235 			 * get such an error -- such as it can exist at all,
12236 			 * it would be a result of corrupted DOF in the driver
12237 			 * properties.
12238 			 */
12239 			if (nmatched == NULL) {
12240 				cmn_err(CE_WARN, "dtrace_enabling_match() "
12241 				    "error on %p: %d", (void *)ep,
12242 				    enab->dten_error);
12243 			}
12244 
12245 			return (enab->dten_error);
12246 		}
12247 	}
12248 
12249 	enab->dten_probegen = dtrace_probegen;
12250 	if (nmatched != NULL)
12251 		*nmatched = total_matched;
12252 
12253 	return (0);
12254 }
12255 
12256 static void
12257 dtrace_enabling_matchall(void)
12258 {
12259 	dtrace_enabling_t *enab;
12260 
12261 	mutex_enter(&cpu_lock);
12262 	mutex_enter(&dtrace_lock);
12263 
12264 	/*
12265 	 * Iterate over all retained enablings to see if any probes match
12266 	 * against them.  We only perform this operation on enablings for which
12267 	 * we have sufficient permissions by virtue of being in the global zone
12268 	 * or in the same zone as the DTrace client.  Because we can be called
12269 	 * after dtrace_detach() has been called, we cannot assert that there
12270 	 * are retained enablings.  We can safely load from dtrace_retained,
12271 	 * however:  the taskq_destroy() at the end of dtrace_detach() will
12272 	 * block pending our completion.
12273 	 */
12274 	for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
12275 		dtrace_cred_t *dcr = &enab->dten_vstate->dtvs_state->dts_cred;
12276 		cred_t *cr = dcr->dcr_cred;
12277 		zoneid_t zone = cr != NULL ? crgetzoneid(cr) : 0;
12278 
12279 		if ((dcr->dcr_visible & DTRACE_CRV_ALLZONE) || (cr != NULL &&
12280 		    (zone == GLOBAL_ZONEID || getzoneid() == zone)))
12281 			(void) dtrace_enabling_match(enab, NULL);
12282 	}
12283 
12284 	mutex_exit(&dtrace_lock);
12285 	mutex_exit(&cpu_lock);
12286 }
12287 
12288 /*
12289  * If an enabling is to be enabled without having matched probes (that is, if
12290  * dtrace_state_go() is to be called on the underlying dtrace_state_t), the
12291  * enabling must be _primed_ by creating an ECB for every ECB description.
12292  * This must be done to assure that we know the number of speculations, the
12293  * number of aggregations, the minimum buffer size needed, etc. before we
12294  * transition out of DTRACE_ACTIVITY_INACTIVE.  To do this without actually
12295  * enabling any probes, we create ECBs for every ECB decription, but with a
12296  * NULL probe -- which is exactly what this function does.
12297  */
12298 static void
12299 dtrace_enabling_prime(dtrace_state_t *state)
12300 {
12301 	dtrace_enabling_t *enab;
12302 	int i;
12303 
12304 	for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
12305 		ASSERT(enab->dten_vstate->dtvs_state != NULL);
12306 
12307 		if (enab->dten_vstate->dtvs_state != state)
12308 			continue;
12309 
12310 		/*
12311 		 * We don't want to prime an enabling more than once, lest
12312 		 * we allow a malicious user to induce resource exhaustion.
12313 		 * (The ECBs that result from priming an enabling aren't
12314 		 * leaked -- but they also aren't deallocated until the
12315 		 * consumer state is destroyed.)
12316 		 */
12317 		if (enab->dten_primed)
12318 			continue;
12319 
12320 		for (i = 0; i < enab->dten_ndesc; i++) {
12321 			enab->dten_current = enab->dten_desc[i];
12322 			(void) dtrace_probe_enable(NULL, enab);
12323 		}
12324 
12325 		enab->dten_primed = 1;
12326 	}
12327 }
12328 
12329 /*
12330  * Called to indicate that probes should be provided due to retained
12331  * enablings.  This is implemented in terms of dtrace_probe_provide(), but it
12332  * must take an initial lap through the enabling calling the dtps_provide()
12333  * entry point explicitly to allow for autocreated probes.
12334  */
12335 static void
12336 dtrace_enabling_provide(dtrace_provider_t *prv)
12337 {
12338 	int i, all = 0;
12339 	dtrace_probedesc_t desc;
12340 	dtrace_genid_t gen;
12341 
12342 	ASSERT(MUTEX_HELD(&dtrace_lock));
12343 	ASSERT(MUTEX_HELD(&dtrace_provider_lock));
12344 
12345 	if (prv == NULL) {
12346 		all = 1;
12347 		prv = dtrace_provider;
12348 	}
12349 
12350 	do {
12351 		dtrace_enabling_t *enab;
12352 		void *parg = prv->dtpv_arg;
12353 
12354 retry:
12355 		gen = dtrace_retained_gen;
12356 		for (enab = dtrace_retained; enab != NULL;
12357 		    enab = enab->dten_next) {
12358 			for (i = 0; i < enab->dten_ndesc; i++) {
12359 				desc = enab->dten_desc[i]->dted_probe;
12360 				mutex_exit(&dtrace_lock);
12361 				prv->dtpv_pops.dtps_provide(parg, &desc);
12362 				mutex_enter(&dtrace_lock);
12363 				/*
12364 				 * Process the retained enablings again if
12365 				 * they have changed while we weren't holding
12366 				 * dtrace_lock.
12367 				 */
12368 				if (gen != dtrace_retained_gen)
12369 					goto retry;
12370 			}
12371 		}
12372 	} while (all && (prv = prv->dtpv_next) != NULL);
12373 
12374 	mutex_exit(&dtrace_lock);
12375 	dtrace_probe_provide(NULL, all ? NULL : prv);
12376 	mutex_enter(&dtrace_lock);
12377 }
12378 
12379 /*
12380  * Called to reap ECBs that are attached to probes from defunct providers.
12381  */
12382 static void
12383 dtrace_enabling_reap(void)
12384 {
12385 	dtrace_provider_t *prov;
12386 	dtrace_probe_t *probe;
12387 	dtrace_ecb_t *ecb;
12388 	hrtime_t when;
12389 	int i;
12390 
12391 	mutex_enter(&cpu_lock);
12392 	mutex_enter(&dtrace_lock);
12393 
12394 	for (i = 0; i < dtrace_nprobes; i++) {
12395 		if ((probe = dtrace_probes[i]) == NULL)
12396 			continue;
12397 
12398 		if (probe->dtpr_ecb == NULL)
12399 			continue;
12400 
12401 		prov = probe->dtpr_provider;
12402 
12403 		if ((when = prov->dtpv_defunct) == 0)
12404 			continue;
12405 
12406 		/*
12407 		 * We have ECBs on a defunct provider:  we want to reap these
12408 		 * ECBs to allow the provider to unregister.  The destruction
12409 		 * of these ECBs must be done carefully:  if we destroy the ECB
12410 		 * and the consumer later wishes to consume an EPID that
12411 		 * corresponds to the destroyed ECB (and if the EPID metadata
12412 		 * has not been previously consumed), the consumer will abort
12413 		 * processing on the unknown EPID.  To reduce (but not, sadly,
12414 		 * eliminate) the possibility of this, we will only destroy an
12415 		 * ECB for a defunct provider if, for the state that
12416 		 * corresponds to the ECB:
12417 		 *
12418 		 *  (a)	There is no speculative tracing (which can effectively
12419 		 *	cache an EPID for an arbitrary amount of time).
12420 		 *
12421 		 *  (b)	The principal buffers have been switched twice since the
12422 		 *	provider became defunct.
12423 		 *
12424 		 *  (c)	The aggregation buffers are of zero size or have been
12425 		 *	switched twice since the provider became defunct.
12426 		 *
12427 		 * We use dts_speculates to determine (a) and call a function
12428 		 * (dtrace_buffer_consumed()) to determine (b) and (c).  Note
12429 		 * that as soon as we've been unable to destroy one of the ECBs
12430 		 * associated with the probe, we quit trying -- reaping is only
12431 		 * fruitful in as much as we can destroy all ECBs associated
12432 		 * with the defunct provider's probes.
12433 		 */
12434 		while ((ecb = probe->dtpr_ecb) != NULL) {
12435 			dtrace_state_t *state = ecb->dte_state;
12436 			dtrace_buffer_t *buf = state->dts_buffer;
12437 			dtrace_buffer_t *aggbuf = state->dts_aggbuffer;
12438 
12439 			if (state->dts_speculates)
12440 				break;
12441 
12442 			if (!dtrace_buffer_consumed(buf, when))
12443 				break;
12444 
12445 			if (!dtrace_buffer_consumed(aggbuf, when))
12446 				break;
12447 
12448 			dtrace_ecb_disable(ecb);
12449 			ASSERT(probe->dtpr_ecb != ecb);
12450 			dtrace_ecb_destroy(ecb);
12451 		}
12452 	}
12453 
12454 	mutex_exit(&dtrace_lock);
12455 	mutex_exit(&cpu_lock);
12456 }
12457 
12458 /*
12459  * DTrace DOF Functions
12460  */
12461 /*ARGSUSED*/
12462 static void
12463 dtrace_dof_error(dof_hdr_t *dof, const char *str)
12464 {
12465 	if (dtrace_err_verbose)
12466 		cmn_err(CE_WARN, "failed to process DOF: %s", str);
12467 
12468 #ifdef DTRACE_ERRDEBUG
12469 	dtrace_errdebug(str);
12470 #endif
12471 }
12472 
12473 /*
12474  * Create DOF out of a currently enabled state.  Right now, we only create
12475  * DOF containing the run-time options -- but this could be expanded to create
12476  * complete DOF representing the enabled state.
12477  */
12478 static dof_hdr_t *
12479 dtrace_dof_create(dtrace_state_t *state)
12480 {
12481 	dof_hdr_t *dof;
12482 	dof_sec_t *sec;
12483 	dof_optdesc_t *opt;
12484 	int i, len = sizeof (dof_hdr_t) +
12485 	    roundup(sizeof (dof_sec_t), sizeof (uint64_t)) +
12486 	    sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
12487 
12488 	ASSERT(MUTEX_HELD(&dtrace_lock));
12489 
12490 	dof = kmem_zalloc(len, KM_SLEEP);
12491 	dof->dofh_ident[DOF_ID_MAG0] = DOF_MAG_MAG0;
12492 	dof->dofh_ident[DOF_ID_MAG1] = DOF_MAG_MAG1;
12493 	dof->dofh_ident[DOF_ID_MAG2] = DOF_MAG_MAG2;
12494 	dof->dofh_ident[DOF_ID_MAG3] = DOF_MAG_MAG3;
12495 
12496 	dof->dofh_ident[DOF_ID_MODEL] = DOF_MODEL_NATIVE;
12497 	dof->dofh_ident[DOF_ID_ENCODING] = DOF_ENCODE_NATIVE;
12498 	dof->dofh_ident[DOF_ID_VERSION] = DOF_VERSION;
12499 	dof->dofh_ident[DOF_ID_DIFVERS] = DIF_VERSION;
12500 	dof->dofh_ident[DOF_ID_DIFIREG] = DIF_DIR_NREGS;
12501 	dof->dofh_ident[DOF_ID_DIFTREG] = DIF_DTR_NREGS;
12502 
12503 	dof->dofh_flags = 0;
12504 	dof->dofh_hdrsize = sizeof (dof_hdr_t);
12505 	dof->dofh_secsize = sizeof (dof_sec_t);
12506 	dof->dofh_secnum = 1;	/* only DOF_SECT_OPTDESC */
12507 	dof->dofh_secoff = sizeof (dof_hdr_t);
12508 	dof->dofh_loadsz = len;
12509 	dof->dofh_filesz = len;
12510 	dof->dofh_pad = 0;
12511 
12512 	/*
12513 	 * Fill in the option section header...
12514 	 */
12515 	sec = (dof_sec_t *)((uintptr_t)dof + sizeof (dof_hdr_t));
12516 	sec->dofs_type = DOF_SECT_OPTDESC;
12517 	sec->dofs_align = sizeof (uint64_t);
12518 	sec->dofs_flags = DOF_SECF_LOAD;
12519 	sec->dofs_entsize = sizeof (dof_optdesc_t);
12520 
12521 	opt = (dof_optdesc_t *)((uintptr_t)sec +
12522 	    roundup(sizeof (dof_sec_t), sizeof (uint64_t)));
12523 
12524 	sec->dofs_offset = (uintptr_t)opt - (uintptr_t)dof;
12525 	sec->dofs_size = sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
12526 
12527 	for (i = 0; i < DTRACEOPT_MAX; i++) {
12528 		opt[i].dofo_option = i;
12529 		opt[i].dofo_strtab = DOF_SECIDX_NONE;
12530 		opt[i].dofo_value = state->dts_options[i];
12531 	}
12532 
12533 	return (dof);
12534 }
12535 
12536 static dof_hdr_t *
12537 dtrace_dof_copyin(uintptr_t uarg, int *errp)
12538 {
12539 	dof_hdr_t hdr, *dof;
12540 
12541 	ASSERT(!MUTEX_HELD(&dtrace_lock));
12542 
12543 	/*
12544 	 * First, we're going to copyin() the sizeof (dof_hdr_t).
12545 	 */
12546 	if (copyin((void *)uarg, &hdr, sizeof (hdr)) != 0) {
12547 		dtrace_dof_error(NULL, "failed to copyin DOF header");
12548 		*errp = EFAULT;
12549 		return (NULL);
12550 	}
12551 
12552 	/*
12553 	 * Now we'll allocate the entire DOF and copy it in -- provided
12554 	 * that the length isn't outrageous.
12555 	 */
12556 	if (hdr.dofh_loadsz >= dtrace_dof_maxsize) {
12557 		dtrace_dof_error(&hdr, "load size exceeds maximum");
12558 		*errp = E2BIG;
12559 		return (NULL);
12560 	}
12561 
12562 	if (hdr.dofh_loadsz < sizeof (hdr)) {
12563 		dtrace_dof_error(&hdr, "invalid load size");
12564 		*errp = EINVAL;
12565 		return (NULL);
12566 	}
12567 
12568 	dof = kmem_alloc(hdr.dofh_loadsz, KM_SLEEP);
12569 
12570 	if (copyin((void *)uarg, dof, hdr.dofh_loadsz) != 0 ||
12571 	    dof->dofh_loadsz != hdr.dofh_loadsz) {
12572 		kmem_free(dof, hdr.dofh_loadsz);
12573 		*errp = EFAULT;
12574 		return (NULL);
12575 	}
12576 
12577 	return (dof);
12578 }
12579 
12580 static dof_hdr_t *
12581 dtrace_dof_property(const char *name)
12582 {
12583 	uchar_t *buf;
12584 	uint64_t loadsz;
12585 	unsigned int len, i;
12586 	dof_hdr_t *dof;
12587 
12588 	/*
12589 	 * Unfortunately, array of values in .conf files are always (and
12590 	 * only) interpreted to be integer arrays.  We must read our DOF
12591 	 * as an integer array, and then squeeze it into a byte array.
12592 	 */
12593 	if (ddi_prop_lookup_int_array(DDI_DEV_T_ANY, dtrace_devi, 0,
12594 	    (char *)name, (int **)&buf, &len) != DDI_PROP_SUCCESS)
12595 		return (NULL);
12596 
12597 	for (i = 0; i < len; i++)
12598 		buf[i] = (uchar_t)(((int *)buf)[i]);
12599 
12600 	if (len < sizeof (dof_hdr_t)) {
12601 		ddi_prop_free(buf);
12602 		dtrace_dof_error(NULL, "truncated header");
12603 		return (NULL);
12604 	}
12605 
12606 	if (len < (loadsz = ((dof_hdr_t *)buf)->dofh_loadsz)) {
12607 		ddi_prop_free(buf);
12608 		dtrace_dof_error(NULL, "truncated DOF");
12609 		return (NULL);
12610 	}
12611 
12612 	if (loadsz >= dtrace_dof_maxsize) {
12613 		ddi_prop_free(buf);
12614 		dtrace_dof_error(NULL, "oversized DOF");
12615 		return (NULL);
12616 	}
12617 
12618 	dof = kmem_alloc(loadsz, KM_SLEEP);
12619 	bcopy(buf, dof, loadsz);
12620 	ddi_prop_free(buf);
12621 
12622 	return (dof);
12623 }
12624 
12625 static void
12626 dtrace_dof_destroy(dof_hdr_t *dof)
12627 {
12628 	kmem_free(dof, dof->dofh_loadsz);
12629 }
12630 
12631 /*
12632  * Return the dof_sec_t pointer corresponding to a given section index.  If the
12633  * index is not valid, dtrace_dof_error() is called and NULL is returned.  If
12634  * a type other than DOF_SECT_NONE is specified, the header is checked against
12635  * this type and NULL is returned if the types do not match.
12636  */
12637 static dof_sec_t *
12638 dtrace_dof_sect(dof_hdr_t *dof, uint32_t type, dof_secidx_t i)
12639 {
12640 	dof_sec_t *sec = (dof_sec_t *)(uintptr_t)
12641 	    ((uintptr_t)dof + dof->dofh_secoff + i * dof->dofh_secsize);
12642 
12643 	if (i >= dof->dofh_secnum) {
12644 		dtrace_dof_error(dof, "referenced section index is invalid");
12645 		return (NULL);
12646 	}
12647 
12648 	if (!(sec->dofs_flags & DOF_SECF_LOAD)) {
12649 		dtrace_dof_error(dof, "referenced section is not loadable");
12650 		return (NULL);
12651 	}
12652 
12653 	if (type != DOF_SECT_NONE && type != sec->dofs_type) {
12654 		dtrace_dof_error(dof, "referenced section is the wrong type");
12655 		return (NULL);
12656 	}
12657 
12658 	return (sec);
12659 }
12660 
12661 static dtrace_probedesc_t *
12662 dtrace_dof_probedesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_probedesc_t *desc)
12663 {
12664 	dof_probedesc_t *probe;
12665 	dof_sec_t *strtab;
12666 	uintptr_t daddr = (uintptr_t)dof;
12667 	uintptr_t str;
12668 	size_t size;
12669 
12670 	if (sec->dofs_type != DOF_SECT_PROBEDESC) {
12671 		dtrace_dof_error(dof, "invalid probe section");
12672 		return (NULL);
12673 	}
12674 
12675 	if (sec->dofs_align != sizeof (dof_secidx_t)) {
12676 		dtrace_dof_error(dof, "bad alignment in probe description");
12677 		return (NULL);
12678 	}
12679 
12680 	if (sec->dofs_offset + sizeof (dof_probedesc_t) > dof->dofh_loadsz) {
12681 		dtrace_dof_error(dof, "truncated probe description");
12682 		return (NULL);
12683 	}
12684 
12685 	probe = (dof_probedesc_t *)(uintptr_t)(daddr + sec->dofs_offset);
12686 	strtab = dtrace_dof_sect(dof, DOF_SECT_STRTAB, probe->dofp_strtab);
12687 
12688 	if (strtab == NULL)
12689 		return (NULL);
12690 
12691 	str = daddr + strtab->dofs_offset;
12692 	size = strtab->dofs_size;
12693 
12694 	if (probe->dofp_provider >= strtab->dofs_size) {
12695 		dtrace_dof_error(dof, "corrupt probe provider");
12696 		return (NULL);
12697 	}
12698 
12699 	(void) strncpy(desc->dtpd_provider,
12700 	    (char *)(str + probe->dofp_provider),
12701 	    MIN(DTRACE_PROVNAMELEN - 1, size - probe->dofp_provider));
12702 
12703 	if (probe->dofp_mod >= strtab->dofs_size) {
12704 		dtrace_dof_error(dof, "corrupt probe module");
12705 		return (NULL);
12706 	}
12707 
12708 	(void) strncpy(desc->dtpd_mod, (char *)(str + probe->dofp_mod),
12709 	    MIN(DTRACE_MODNAMELEN - 1, size - probe->dofp_mod));
12710 
12711 	if (probe->dofp_func >= strtab->dofs_size) {
12712 		dtrace_dof_error(dof, "corrupt probe function");
12713 		return (NULL);
12714 	}
12715 
12716 	(void) strncpy(desc->dtpd_func, (char *)(str + probe->dofp_func),
12717 	    MIN(DTRACE_FUNCNAMELEN - 1, size - probe->dofp_func));
12718 
12719 	if (probe->dofp_name >= strtab->dofs_size) {
12720 		dtrace_dof_error(dof, "corrupt probe name");
12721 		return (NULL);
12722 	}
12723 
12724 	(void) strncpy(desc->dtpd_name, (char *)(str + probe->dofp_name),
12725 	    MIN(DTRACE_NAMELEN - 1, size - probe->dofp_name));
12726 
12727 	return (desc);
12728 }
12729 
12730 static dtrace_difo_t *
12731 dtrace_dof_difo(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12732     cred_t *cr)
12733 {
12734 	dtrace_difo_t *dp;
12735 	size_t ttl = 0;
12736 	dof_difohdr_t *dofd;
12737 	uintptr_t daddr = (uintptr_t)dof;
12738 	size_t max = dtrace_difo_maxsize;
12739 	int i, l, n;
12740 
12741 	static const struct {
12742 		int section;
12743 		int bufoffs;
12744 		int lenoffs;
12745 		int entsize;
12746 		int align;
12747 		const char *msg;
12748 	} difo[] = {
12749 		{ DOF_SECT_DIF, offsetof(dtrace_difo_t, dtdo_buf),
12750 		offsetof(dtrace_difo_t, dtdo_len), sizeof (dif_instr_t),
12751 		sizeof (dif_instr_t), "multiple DIF sections" },
12752 
12753 		{ DOF_SECT_INTTAB, offsetof(dtrace_difo_t, dtdo_inttab),
12754 		offsetof(dtrace_difo_t, dtdo_intlen), sizeof (uint64_t),
12755 		sizeof (uint64_t), "multiple integer tables" },
12756 
12757 		{ DOF_SECT_STRTAB, offsetof(dtrace_difo_t, dtdo_strtab),
12758 		offsetof(dtrace_difo_t, dtdo_strlen), 0,
12759 		sizeof (char), "multiple string tables" },
12760 
12761 		{ DOF_SECT_VARTAB, offsetof(dtrace_difo_t, dtdo_vartab),
12762 		offsetof(dtrace_difo_t, dtdo_varlen), sizeof (dtrace_difv_t),
12763 		sizeof (uint_t), "multiple variable tables" },
12764 
12765 		{ DOF_SECT_NONE, 0, 0, 0, NULL }
12766 	};
12767 
12768 	if (sec->dofs_type != DOF_SECT_DIFOHDR) {
12769 		dtrace_dof_error(dof, "invalid DIFO header section");
12770 		return (NULL);
12771 	}
12772 
12773 	if (sec->dofs_align != sizeof (dof_secidx_t)) {
12774 		dtrace_dof_error(dof, "bad alignment in DIFO header");
12775 		return (NULL);
12776 	}
12777 
12778 	if (sec->dofs_size < sizeof (dof_difohdr_t) ||
12779 	    sec->dofs_size % sizeof (dof_secidx_t)) {
12780 		dtrace_dof_error(dof, "bad size in DIFO header");
12781 		return (NULL);
12782 	}
12783 
12784 	dofd = (dof_difohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
12785 	n = (sec->dofs_size - sizeof (*dofd)) / sizeof (dof_secidx_t) + 1;
12786 
12787 	dp = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
12788 	dp->dtdo_rtype = dofd->dofd_rtype;
12789 
12790 	for (l = 0; l < n; l++) {
12791 		dof_sec_t *subsec;
12792 		void **bufp;
12793 		uint32_t *lenp;
12794 
12795 		if ((subsec = dtrace_dof_sect(dof, DOF_SECT_NONE,
12796 		    dofd->dofd_links[l])) == NULL)
12797 			goto err; /* invalid section link */
12798 
12799 		if (ttl + subsec->dofs_size > max) {
12800 			dtrace_dof_error(dof, "exceeds maximum size");
12801 			goto err;
12802 		}
12803 
12804 		ttl += subsec->dofs_size;
12805 
12806 		for (i = 0; difo[i].section != DOF_SECT_NONE; i++) {
12807 			if (subsec->dofs_type != difo[i].section)
12808 				continue;
12809 
12810 			if (!(subsec->dofs_flags & DOF_SECF_LOAD)) {
12811 				dtrace_dof_error(dof, "section not loaded");
12812 				goto err;
12813 			}
12814 
12815 			if (subsec->dofs_align != difo[i].align) {
12816 				dtrace_dof_error(dof, "bad alignment");
12817 				goto err;
12818 			}
12819 
12820 			bufp = (void **)((uintptr_t)dp + difo[i].bufoffs);
12821 			lenp = (uint32_t *)((uintptr_t)dp + difo[i].lenoffs);
12822 
12823 			if (*bufp != NULL) {
12824 				dtrace_dof_error(dof, difo[i].msg);
12825 				goto err;
12826 			}
12827 
12828 			if (difo[i].entsize != subsec->dofs_entsize) {
12829 				dtrace_dof_error(dof, "entry size mismatch");
12830 				goto err;
12831 			}
12832 
12833 			if (subsec->dofs_entsize != 0 &&
12834 			    (subsec->dofs_size % subsec->dofs_entsize) != 0) {
12835 				dtrace_dof_error(dof, "corrupt entry size");
12836 				goto err;
12837 			}
12838 
12839 			*lenp = subsec->dofs_size;
12840 			*bufp = kmem_alloc(subsec->dofs_size, KM_SLEEP);
12841 			bcopy((char *)(uintptr_t)(daddr + subsec->dofs_offset),
12842 			    *bufp, subsec->dofs_size);
12843 
12844 			if (subsec->dofs_entsize != 0)
12845 				*lenp /= subsec->dofs_entsize;
12846 
12847 			break;
12848 		}
12849 
12850 		/*
12851 		 * If we encounter a loadable DIFO sub-section that is not
12852 		 * known to us, assume this is a broken program and fail.
12853 		 */
12854 		if (difo[i].section == DOF_SECT_NONE &&
12855 		    (subsec->dofs_flags & DOF_SECF_LOAD)) {
12856 			dtrace_dof_error(dof, "unrecognized DIFO subsection");
12857 			goto err;
12858 		}
12859 	}
12860 
12861 	if (dp->dtdo_buf == NULL) {
12862 		/*
12863 		 * We can't have a DIF object without DIF text.
12864 		 */
12865 		dtrace_dof_error(dof, "missing DIF text");
12866 		goto err;
12867 	}
12868 
12869 	/*
12870 	 * Before we validate the DIF object, run through the variable table
12871 	 * looking for the strings -- if any of their size are under, we'll set
12872 	 * their size to be the system-wide default string size.  Note that
12873 	 * this should _not_ happen if the "strsize" option has been set --
12874 	 * in this case, the compiler should have set the size to reflect the
12875 	 * setting of the option.
12876 	 */
12877 	for (i = 0; i < dp->dtdo_varlen; i++) {
12878 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
12879 		dtrace_diftype_t *t = &v->dtdv_type;
12880 
12881 		if (v->dtdv_id < DIF_VAR_OTHER_UBASE)
12882 			continue;
12883 
12884 		if (t->dtdt_kind == DIF_TYPE_STRING && t->dtdt_size == 0)
12885 			t->dtdt_size = dtrace_strsize_default;
12886 	}
12887 
12888 	if (dtrace_difo_validate(dp, vstate, DIF_DIR_NREGS, cr) != 0)
12889 		goto err;
12890 
12891 	dtrace_difo_init(dp, vstate);
12892 	return (dp);
12893 
12894 err:
12895 	kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
12896 	kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
12897 	kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
12898 	kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
12899 
12900 	kmem_free(dp, sizeof (dtrace_difo_t));
12901 	return (NULL);
12902 }
12903 
12904 static dtrace_predicate_t *
12905 dtrace_dof_predicate(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12906     cred_t *cr)
12907 {
12908 	dtrace_difo_t *dp;
12909 
12910 	if ((dp = dtrace_dof_difo(dof, sec, vstate, cr)) == NULL)
12911 		return (NULL);
12912 
12913 	return (dtrace_predicate_create(dp));
12914 }
12915 
12916 static dtrace_actdesc_t *
12917 dtrace_dof_actdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12918     cred_t *cr)
12919 {
12920 	dtrace_actdesc_t *act, *first = NULL, *last = NULL, *next;
12921 	dof_actdesc_t *desc;
12922 	dof_sec_t *difosec;
12923 	size_t offs;
12924 	uintptr_t daddr = (uintptr_t)dof;
12925 	uint64_t arg;
12926 	dtrace_actkind_t kind;
12927 
12928 	if (sec->dofs_type != DOF_SECT_ACTDESC) {
12929 		dtrace_dof_error(dof, "invalid action section");
12930 		return (NULL);
12931 	}
12932 
12933 	if (sec->dofs_offset + sizeof (dof_actdesc_t) > dof->dofh_loadsz) {
12934 		dtrace_dof_error(dof, "truncated action description");
12935 		return (NULL);
12936 	}
12937 
12938 	if (sec->dofs_align != sizeof (uint64_t)) {
12939 		dtrace_dof_error(dof, "bad alignment in action description");
12940 		return (NULL);
12941 	}
12942 
12943 	if (sec->dofs_size < sec->dofs_entsize) {
12944 		dtrace_dof_error(dof, "section entry size exceeds total size");
12945 		return (NULL);
12946 	}
12947 
12948 	if (sec->dofs_entsize != sizeof (dof_actdesc_t)) {
12949 		dtrace_dof_error(dof, "bad entry size in action description");
12950 		return (NULL);
12951 	}
12952 
12953 	if (sec->dofs_size / sec->dofs_entsize > dtrace_actions_max) {
12954 		dtrace_dof_error(dof, "actions exceed dtrace_actions_max");
12955 		return (NULL);
12956 	}
12957 
12958 	for (offs = 0; offs < sec->dofs_size; offs += sec->dofs_entsize) {
12959 		desc = (dof_actdesc_t *)(daddr +
12960 		    (uintptr_t)sec->dofs_offset + offs);
12961 		kind = (dtrace_actkind_t)desc->dofa_kind;
12962 
12963 		if ((DTRACEACT_ISPRINTFLIKE(kind) &&
12964 		    (kind != DTRACEACT_PRINTA ||
12965 		    desc->dofa_strtab != DOF_SECIDX_NONE)) ||
12966 		    (kind == DTRACEACT_DIFEXPR &&
12967 		    desc->dofa_strtab != DOF_SECIDX_NONE)) {
12968 			dof_sec_t *strtab;
12969 			char *str, *fmt;
12970 			uint64_t i;
12971 
12972 			/*
12973 			 * The argument to these actions is an index into the
12974 			 * DOF string table.  For printf()-like actions, this
12975 			 * is the format string.  For print(), this is the
12976 			 * CTF type of the expression result.
12977 			 */
12978 			if ((strtab = dtrace_dof_sect(dof,
12979 			    DOF_SECT_STRTAB, desc->dofa_strtab)) == NULL)
12980 				goto err;
12981 
12982 			str = (char *)((uintptr_t)dof +
12983 			    (uintptr_t)strtab->dofs_offset);
12984 
12985 			for (i = desc->dofa_arg; i < strtab->dofs_size; i++) {
12986 				if (str[i] == '\0')
12987 					break;
12988 			}
12989 
12990 			if (i >= strtab->dofs_size) {
12991 				dtrace_dof_error(dof, "bogus format string");
12992 				goto err;
12993 			}
12994 
12995 			if (i == desc->dofa_arg) {
12996 				dtrace_dof_error(dof, "empty format string");
12997 				goto err;
12998 			}
12999 
13000 			i -= desc->dofa_arg;
13001 			fmt = kmem_alloc(i + 1, KM_SLEEP);
13002 			bcopy(&str[desc->dofa_arg], fmt, i + 1);
13003 			arg = (uint64_t)(uintptr_t)fmt;
13004 		} else {
13005 			if (kind == DTRACEACT_PRINTA) {
13006 				ASSERT(desc->dofa_strtab == DOF_SECIDX_NONE);
13007 				arg = 0;
13008 			} else {
13009 				arg = desc->dofa_arg;
13010 			}
13011 		}
13012 
13013 		act = dtrace_actdesc_create(kind, desc->dofa_ntuple,
13014 		    desc->dofa_uarg, arg);
13015 
13016 		if (last != NULL) {
13017 			last->dtad_next = act;
13018 		} else {
13019 			first = act;
13020 		}
13021 
13022 		last = act;
13023 
13024 		if (desc->dofa_difo == DOF_SECIDX_NONE)
13025 			continue;
13026 
13027 		if ((difosec = dtrace_dof_sect(dof,
13028 		    DOF_SECT_DIFOHDR, desc->dofa_difo)) == NULL)
13029 			goto err;
13030 
13031 		act->dtad_difo = dtrace_dof_difo(dof, difosec, vstate, cr);
13032 
13033 		if (act->dtad_difo == NULL)
13034 			goto err;
13035 	}
13036 
13037 	ASSERT(first != NULL);
13038 	return (first);
13039 
13040 err:
13041 	for (act = first; act != NULL; act = next) {
13042 		next = act->dtad_next;
13043 		dtrace_actdesc_release(act, vstate);
13044 	}
13045 
13046 	return (NULL);
13047 }
13048 
13049 static dtrace_ecbdesc_t *
13050 dtrace_dof_ecbdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
13051     cred_t *cr)
13052 {
13053 	dtrace_ecbdesc_t *ep;
13054 	dof_ecbdesc_t *ecb;
13055 	dtrace_probedesc_t *desc;
13056 	dtrace_predicate_t *pred = NULL;
13057 
13058 	if (sec->dofs_size < sizeof (dof_ecbdesc_t)) {
13059 		dtrace_dof_error(dof, "truncated ECB description");
13060 		return (NULL);
13061 	}
13062 
13063 	if (sec->dofs_align != sizeof (uint64_t)) {
13064 		dtrace_dof_error(dof, "bad alignment in ECB description");
13065 		return (NULL);
13066 	}
13067 
13068 	ecb = (dof_ecbdesc_t *)((uintptr_t)dof + (uintptr_t)sec->dofs_offset);
13069 	sec = dtrace_dof_sect(dof, DOF_SECT_PROBEDESC, ecb->dofe_probes);
13070 
13071 	if (sec == NULL)
13072 		return (NULL);
13073 
13074 	ep = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
13075 	ep->dted_uarg = ecb->dofe_uarg;
13076 	desc = &ep->dted_probe;
13077 
13078 	if (dtrace_dof_probedesc(dof, sec, desc) == NULL)
13079 		goto err;
13080 
13081 	if (ecb->dofe_pred != DOF_SECIDX_NONE) {
13082 		if ((sec = dtrace_dof_sect(dof,
13083 		    DOF_SECT_DIFOHDR, ecb->dofe_pred)) == NULL)
13084 			goto err;
13085 
13086 		if ((pred = dtrace_dof_predicate(dof, sec, vstate, cr)) == NULL)
13087 			goto err;
13088 
13089 		ep->dted_pred.dtpdd_predicate = pred;
13090 	}
13091 
13092 	if (ecb->dofe_actions != DOF_SECIDX_NONE) {
13093 		if ((sec = dtrace_dof_sect(dof,
13094 		    DOF_SECT_ACTDESC, ecb->dofe_actions)) == NULL)
13095 			goto err;
13096 
13097 		ep->dted_action = dtrace_dof_actdesc(dof, sec, vstate, cr);
13098 
13099 		if (ep->dted_action == NULL)
13100 			goto err;
13101 	}
13102 
13103 	return (ep);
13104 
13105 err:
13106 	if (pred != NULL)
13107 		dtrace_predicate_release(pred, vstate);
13108 	kmem_free(ep, sizeof (dtrace_ecbdesc_t));
13109 	return (NULL);
13110 }
13111 
13112 /*
13113  * Apply the relocations from the specified 'sec' (a DOF_SECT_URELHDR) to the
13114  * specified DOF.  At present, this amounts to simply adding 'ubase' to the
13115  * site of any user SETX relocations to account for load object base address.
13116  * In the future, if we need other relocations, this function can be extended.
13117  */
13118 static int
13119 dtrace_dof_relocate(dof_hdr_t *dof, dof_sec_t *sec, uint64_t ubase)
13120 {
13121 	uintptr_t daddr = (uintptr_t)dof;
13122 	dof_relohdr_t *dofr =
13123 	    (dof_relohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
13124 	dof_sec_t *ss, *rs, *ts;
13125 	dof_relodesc_t *r;
13126 	uint_t i, n;
13127 
13128 	if (sec->dofs_size < sizeof (dof_relohdr_t) ||
13129 	    sec->dofs_align != sizeof (dof_secidx_t)) {
13130 		dtrace_dof_error(dof, "invalid relocation header");
13131 		return (-1);
13132 	}
13133 
13134 	ss = dtrace_dof_sect(dof, DOF_SECT_STRTAB, dofr->dofr_strtab);
13135 	rs = dtrace_dof_sect(dof, DOF_SECT_RELTAB, dofr->dofr_relsec);
13136 	ts = dtrace_dof_sect(dof, DOF_SECT_NONE, dofr->dofr_tgtsec);
13137 
13138 	if (ss == NULL || rs == NULL || ts == NULL)
13139 		return (-1); /* dtrace_dof_error() has been called already */
13140 
13141 	if (rs->dofs_entsize < sizeof (dof_relodesc_t) ||
13142 	    rs->dofs_align != sizeof (uint64_t)) {
13143 		dtrace_dof_error(dof, "invalid relocation section");
13144 		return (-1);
13145 	}
13146 
13147 	r = (dof_relodesc_t *)(uintptr_t)(daddr + rs->dofs_offset);
13148 	n = rs->dofs_size / rs->dofs_entsize;
13149 
13150 	for (i = 0; i < n; i++) {
13151 		uintptr_t taddr = daddr + ts->dofs_offset + r->dofr_offset;
13152 
13153 		switch (r->dofr_type) {
13154 		case DOF_RELO_NONE:
13155 			break;
13156 		case DOF_RELO_SETX:
13157 			if (r->dofr_offset >= ts->dofs_size || r->dofr_offset +
13158 			    sizeof (uint64_t) > ts->dofs_size) {
13159 				dtrace_dof_error(dof, "bad relocation offset");
13160 				return (-1);
13161 			}
13162 
13163 			if (!IS_P2ALIGNED(taddr, sizeof (uint64_t))) {
13164 				dtrace_dof_error(dof, "misaligned setx relo");
13165 				return (-1);
13166 			}
13167 
13168 			*(uint64_t *)taddr += ubase;
13169 			break;
13170 		default:
13171 			dtrace_dof_error(dof, "invalid relocation type");
13172 			return (-1);
13173 		}
13174 
13175 		r = (dof_relodesc_t *)((uintptr_t)r + rs->dofs_entsize);
13176 	}
13177 
13178 	return (0);
13179 }
13180 
13181 /*
13182  * The dof_hdr_t passed to dtrace_dof_slurp() should be a partially validated
13183  * header:  it should be at the front of a memory region that is at least
13184  * sizeof (dof_hdr_t) in size -- and then at least dof_hdr.dofh_loadsz in
13185  * size.  It need not be validated in any other way.
13186  */
13187 static int
13188 dtrace_dof_slurp(dof_hdr_t *dof, dtrace_vstate_t *vstate, cred_t *cr,
13189     dtrace_enabling_t **enabp, uint64_t ubase, int noprobes)
13190 {
13191 	uint64_t len = dof->dofh_loadsz, seclen;
13192 	uintptr_t daddr = (uintptr_t)dof;
13193 	dtrace_ecbdesc_t *ep;
13194 	dtrace_enabling_t *enab;
13195 	uint_t i;
13196 
13197 	ASSERT(MUTEX_HELD(&dtrace_lock));
13198 	ASSERT(dof->dofh_loadsz >= sizeof (dof_hdr_t));
13199 
13200 	/*
13201 	 * Check the DOF header identification bytes.  In addition to checking
13202 	 * valid settings, we also verify that unused bits/bytes are zeroed so
13203 	 * we can use them later without fear of regressing existing binaries.
13204 	 */
13205 	if (bcmp(&dof->dofh_ident[DOF_ID_MAG0],
13206 	    DOF_MAG_STRING, DOF_MAG_STRLEN) != 0) {
13207 		dtrace_dof_error(dof, "DOF magic string mismatch");
13208 		return (-1);
13209 	}
13210 
13211 	if (dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_ILP32 &&
13212 	    dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_LP64) {
13213 		dtrace_dof_error(dof, "DOF has invalid data model");
13214 		return (-1);
13215 	}
13216 
13217 	if (dof->dofh_ident[DOF_ID_ENCODING] != DOF_ENCODE_NATIVE) {
13218 		dtrace_dof_error(dof, "DOF encoding mismatch");
13219 		return (-1);
13220 	}
13221 
13222 	if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
13223 	    dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_2) {
13224 		dtrace_dof_error(dof, "DOF version mismatch");
13225 		return (-1);
13226 	}
13227 
13228 	if (dof->dofh_ident[DOF_ID_DIFVERS] != DIF_VERSION_2) {
13229 		dtrace_dof_error(dof, "DOF uses unsupported instruction set");
13230 		return (-1);
13231 	}
13232 
13233 	if (dof->dofh_ident[DOF_ID_DIFIREG] > DIF_DIR_NREGS) {
13234 		dtrace_dof_error(dof, "DOF uses too many integer registers");
13235 		return (-1);
13236 	}
13237 
13238 	if (dof->dofh_ident[DOF_ID_DIFTREG] > DIF_DTR_NREGS) {
13239 		dtrace_dof_error(dof, "DOF uses too many tuple registers");
13240 		return (-1);
13241 	}
13242 
13243 	for (i = DOF_ID_PAD; i < DOF_ID_SIZE; i++) {
13244 		if (dof->dofh_ident[i] != 0) {
13245 			dtrace_dof_error(dof, "DOF has invalid ident byte set");
13246 			return (-1);
13247 		}
13248 	}
13249 
13250 	if (dof->dofh_flags & ~DOF_FL_VALID) {
13251 		dtrace_dof_error(dof, "DOF has invalid flag bits set");
13252 		return (-1);
13253 	}
13254 
13255 	if (dof->dofh_secsize == 0) {
13256 		dtrace_dof_error(dof, "zero section header size");
13257 		return (-1);
13258 	}
13259 
13260 	/*
13261 	 * Check that the section headers don't exceed the amount of DOF
13262 	 * data.  Note that we cast the section size and number of sections
13263 	 * to uint64_t's to prevent possible overflow in the multiplication.
13264 	 */
13265 	seclen = (uint64_t)dof->dofh_secnum * (uint64_t)dof->dofh_secsize;
13266 
13267 	if (dof->dofh_secoff > len || seclen > len ||
13268 	    dof->dofh_secoff + seclen > len) {
13269 		dtrace_dof_error(dof, "truncated section headers");
13270 		return (-1);
13271 	}
13272 
13273 	if (!IS_P2ALIGNED(dof->dofh_secoff, sizeof (uint64_t))) {
13274 		dtrace_dof_error(dof, "misaligned section headers");
13275 		return (-1);
13276 	}
13277 
13278 	if (!IS_P2ALIGNED(dof->dofh_secsize, sizeof (uint64_t))) {
13279 		dtrace_dof_error(dof, "misaligned section size");
13280 		return (-1);
13281 	}
13282 
13283 	/*
13284 	 * Take an initial pass through the section headers to be sure that
13285 	 * the headers don't have stray offsets.  If the 'noprobes' flag is
13286 	 * set, do not permit sections relating to providers, probes, or args.
13287 	 */
13288 	for (i = 0; i < dof->dofh_secnum; i++) {
13289 		dof_sec_t *sec = (dof_sec_t *)(daddr +
13290 		    (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
13291 
13292 		if (noprobes) {
13293 			switch (sec->dofs_type) {
13294 			case DOF_SECT_PROVIDER:
13295 			case DOF_SECT_PROBES:
13296 			case DOF_SECT_PRARGS:
13297 			case DOF_SECT_PROFFS:
13298 				dtrace_dof_error(dof, "illegal sections "
13299 				    "for enabling");
13300 				return (-1);
13301 			}
13302 		}
13303 
13304 		if (DOF_SEC_ISLOADABLE(sec->dofs_type) &&
13305 		    !(sec->dofs_flags & DOF_SECF_LOAD)) {
13306 			dtrace_dof_error(dof, "loadable section with load "
13307 			    "flag unset");
13308 			return (-1);
13309 		}
13310 
13311 		if (!(sec->dofs_flags & DOF_SECF_LOAD))
13312 			continue; /* just ignore non-loadable sections */
13313 
13314 		if (!ISP2(sec->dofs_align)) {
13315 			dtrace_dof_error(dof, "bad section alignment");
13316 			return (-1);
13317 		}
13318 
13319 		if (sec->dofs_offset & (sec->dofs_align - 1)) {
13320 			dtrace_dof_error(dof, "misaligned section");
13321 			return (-1);
13322 		}
13323 
13324 		if (sec->dofs_offset > len || sec->dofs_size > len ||
13325 		    sec->dofs_offset + sec->dofs_size > len) {
13326 			dtrace_dof_error(dof, "corrupt section header");
13327 			return (-1);
13328 		}
13329 
13330 		if (sec->dofs_type == DOF_SECT_STRTAB && *((char *)daddr +
13331 		    sec->dofs_offset + sec->dofs_size - 1) != '\0') {
13332 			dtrace_dof_error(dof, "non-terminating string table");
13333 			return (-1);
13334 		}
13335 	}
13336 
13337 	/*
13338 	 * Take a second pass through the sections and locate and perform any
13339 	 * relocations that are present.  We do this after the first pass to
13340 	 * be sure that all sections have had their headers validated.
13341 	 */
13342 	for (i = 0; i < dof->dofh_secnum; i++) {
13343 		dof_sec_t *sec = (dof_sec_t *)(daddr +
13344 		    (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
13345 
13346 		if (!(sec->dofs_flags & DOF_SECF_LOAD))
13347 			continue; /* skip sections that are not loadable */
13348 
13349 		switch (sec->dofs_type) {
13350 		case DOF_SECT_URELHDR:
13351 			if (dtrace_dof_relocate(dof, sec, ubase) != 0)
13352 				return (-1);
13353 			break;
13354 		}
13355 	}
13356 
13357 	if ((enab = *enabp) == NULL)
13358 		enab = *enabp = dtrace_enabling_create(vstate);
13359 
13360 	for (i = 0; i < dof->dofh_secnum; i++) {
13361 		dof_sec_t *sec = (dof_sec_t *)(daddr +
13362 		    (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
13363 
13364 		if (sec->dofs_type != DOF_SECT_ECBDESC)
13365 			continue;
13366 
13367 		if ((ep = dtrace_dof_ecbdesc(dof, sec, vstate, cr)) == NULL) {
13368 			dtrace_enabling_destroy(enab);
13369 			*enabp = NULL;
13370 			return (-1);
13371 		}
13372 
13373 		dtrace_enabling_add(enab, ep);
13374 	}
13375 
13376 	return (0);
13377 }
13378 
13379 /*
13380  * Process DOF for any options.  This routine assumes that the DOF has been
13381  * at least processed by dtrace_dof_slurp().
13382  */
13383 static int
13384 dtrace_dof_options(dof_hdr_t *dof, dtrace_state_t *state)
13385 {
13386 	int i, rval;
13387 	uint32_t entsize;
13388 	size_t offs;
13389 	dof_optdesc_t *desc;
13390 
13391 	for (i = 0; i < dof->dofh_secnum; i++) {
13392 		dof_sec_t *sec = (dof_sec_t *)((uintptr_t)dof +
13393 		    (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
13394 
13395 		if (sec->dofs_type != DOF_SECT_OPTDESC)
13396 			continue;
13397 
13398 		if (sec->dofs_align != sizeof (uint64_t)) {
13399 			dtrace_dof_error(dof, "bad alignment in "
13400 			    "option description");
13401 			return (EINVAL);
13402 		}
13403 
13404 		if ((entsize = sec->dofs_entsize) == 0) {
13405 			dtrace_dof_error(dof, "zeroed option entry size");
13406 			return (EINVAL);
13407 		}
13408 
13409 		if (entsize < sizeof (dof_optdesc_t)) {
13410 			dtrace_dof_error(dof, "bad option entry size");
13411 			return (EINVAL);
13412 		}
13413 
13414 		for (offs = 0; offs < sec->dofs_size; offs += entsize) {
13415 			desc = (dof_optdesc_t *)((uintptr_t)dof +
13416 			    (uintptr_t)sec->dofs_offset + offs);
13417 
13418 			if (desc->dofo_strtab != DOF_SECIDX_NONE) {
13419 				dtrace_dof_error(dof, "non-zero option string");
13420 				return (EINVAL);
13421 			}
13422 
13423 			if (desc->dofo_value == DTRACEOPT_UNSET) {
13424 				dtrace_dof_error(dof, "unset option");
13425 				return (EINVAL);
13426 			}
13427 
13428 			if ((rval = dtrace_state_option(state,
13429 			    desc->dofo_option, desc->dofo_value)) != 0) {
13430 				dtrace_dof_error(dof, "rejected option");
13431 				return (rval);
13432 			}
13433 		}
13434 	}
13435 
13436 	return (0);
13437 }
13438 
13439 /*
13440  * DTrace Consumer State Functions
13441  */
13442 int
13443 dtrace_dstate_init(dtrace_dstate_t *dstate, size_t size)
13444 {
13445 	size_t hashsize, maxper, min, chunksize = dstate->dtds_chunksize;
13446 	void *base;
13447 	uintptr_t limit;
13448 	dtrace_dynvar_t *dvar, *next, *start;
13449 	int i;
13450 
13451 	ASSERT(MUTEX_HELD(&dtrace_lock));
13452 	ASSERT(dstate->dtds_base == NULL && dstate->dtds_percpu == NULL);
13453 
13454 	bzero(dstate, sizeof (dtrace_dstate_t));
13455 
13456 	if ((dstate->dtds_chunksize = chunksize) == 0)
13457 		dstate->dtds_chunksize = DTRACE_DYNVAR_CHUNKSIZE;
13458 
13459 	VERIFY(dstate->dtds_chunksize < LONG_MAX);
13460 
13461 	if (size < (min = dstate->dtds_chunksize + sizeof (dtrace_dynhash_t)))
13462 		size = min;
13463 
13464 	if ((base = kmem_zalloc(size, KM_NOSLEEP | KM_NORMALPRI)) == NULL)
13465 		return (ENOMEM);
13466 
13467 	dstate->dtds_size = size;
13468 	dstate->dtds_base = base;
13469 	dstate->dtds_percpu = kmem_cache_alloc(dtrace_state_cache, KM_SLEEP);
13470 	bzero(dstate->dtds_percpu, NCPU * sizeof (dtrace_dstate_percpu_t));
13471 
13472 	hashsize = size / (dstate->dtds_chunksize + sizeof (dtrace_dynhash_t));
13473 
13474 	if (hashsize != 1 && (hashsize & 1))
13475 		hashsize--;
13476 
13477 	dstate->dtds_hashsize = hashsize;
13478 	dstate->dtds_hash = dstate->dtds_base;
13479 
13480 	/*
13481 	 * Set all of our hash buckets to point to the single sink, and (if
13482 	 * it hasn't already been set), set the sink's hash value to be the
13483 	 * sink sentinel value.  The sink is needed for dynamic variable
13484 	 * lookups to know that they have iterated over an entire, valid hash
13485 	 * chain.
13486 	 */
13487 	for (i = 0; i < hashsize; i++)
13488 		dstate->dtds_hash[i].dtdh_chain = &dtrace_dynhash_sink;
13489 
13490 	if (dtrace_dynhash_sink.dtdv_hashval != DTRACE_DYNHASH_SINK)
13491 		dtrace_dynhash_sink.dtdv_hashval = DTRACE_DYNHASH_SINK;
13492 
13493 	/*
13494 	 * Determine number of active CPUs.  Divide free list evenly among
13495 	 * active CPUs.
13496 	 */
13497 	start = (dtrace_dynvar_t *)
13498 	    ((uintptr_t)base + hashsize * sizeof (dtrace_dynhash_t));
13499 	limit = (uintptr_t)base + size;
13500 
13501 	VERIFY((uintptr_t)start < limit);
13502 	VERIFY((uintptr_t)start >= (uintptr_t)base);
13503 
13504 	maxper = (limit - (uintptr_t)start) / NCPU;
13505 	maxper = (maxper / dstate->dtds_chunksize) * dstate->dtds_chunksize;
13506 
13507 	for (i = 0; i < NCPU; i++) {
13508 		dstate->dtds_percpu[i].dtdsc_free = dvar = start;
13509 
13510 		/*
13511 		 * If we don't even have enough chunks to make it once through
13512 		 * NCPUs, we're just going to allocate everything to the first
13513 		 * CPU.  And if we're on the last CPU, we're going to allocate
13514 		 * whatever is left over.  In either case, we set the limit to
13515 		 * be the limit of the dynamic variable space.
13516 		 */
13517 		if (maxper == 0 || i == NCPU - 1) {
13518 			limit = (uintptr_t)base + size;
13519 			start = NULL;
13520 		} else {
13521 			limit = (uintptr_t)start + maxper;
13522 			start = (dtrace_dynvar_t *)limit;
13523 		}
13524 
13525 		VERIFY(limit <= (uintptr_t)base + size);
13526 
13527 		for (;;) {
13528 			next = (dtrace_dynvar_t *)((uintptr_t)dvar +
13529 			    dstate->dtds_chunksize);
13530 
13531 			if ((uintptr_t)next + dstate->dtds_chunksize >= limit)
13532 				break;
13533 
13534 			VERIFY((uintptr_t)dvar >= (uintptr_t)base &&
13535 			    (uintptr_t)dvar <= (uintptr_t)base + size);
13536 			dvar->dtdv_next = next;
13537 			dvar = next;
13538 		}
13539 
13540 		if (maxper == 0)
13541 			break;
13542 	}
13543 
13544 	return (0);
13545 }
13546 
13547 void
13548 dtrace_dstate_fini(dtrace_dstate_t *dstate)
13549 {
13550 	ASSERT(MUTEX_HELD(&cpu_lock));
13551 
13552 	if (dstate->dtds_base == NULL)
13553 		return;
13554 
13555 	kmem_free(dstate->dtds_base, dstate->dtds_size);
13556 	kmem_cache_free(dtrace_state_cache, dstate->dtds_percpu);
13557 }
13558 
13559 static void
13560 dtrace_vstate_fini(dtrace_vstate_t *vstate)
13561 {
13562 	/*
13563 	 * Logical XOR, where are you?
13564 	 */
13565 	ASSERT((vstate->dtvs_nglobals == 0) ^ (vstate->dtvs_globals != NULL));
13566 
13567 	if (vstate->dtvs_nglobals > 0) {
13568 		kmem_free(vstate->dtvs_globals, vstate->dtvs_nglobals *
13569 		    sizeof (dtrace_statvar_t *));
13570 	}
13571 
13572 	if (vstate->dtvs_ntlocals > 0) {
13573 		kmem_free(vstate->dtvs_tlocals, vstate->dtvs_ntlocals *
13574 		    sizeof (dtrace_difv_t));
13575 	}
13576 
13577 	ASSERT((vstate->dtvs_nlocals == 0) ^ (vstate->dtvs_locals != NULL));
13578 
13579 	if (vstate->dtvs_nlocals > 0) {
13580 		kmem_free(vstate->dtvs_locals, vstate->dtvs_nlocals *
13581 		    sizeof (dtrace_statvar_t *));
13582 	}
13583 }
13584 
13585 static void
13586 dtrace_state_clean(dtrace_state_t *state)
13587 {
13588 	if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE)
13589 		return;
13590 
13591 	dtrace_dynvar_clean(&state->dts_vstate.dtvs_dynvars);
13592 	dtrace_speculation_clean(state);
13593 }
13594 
13595 static void
13596 dtrace_state_deadman(dtrace_state_t *state)
13597 {
13598 	hrtime_t now;
13599 
13600 	dtrace_sync();
13601 
13602 	now = dtrace_gethrtime();
13603 
13604 	if (state != dtrace_anon.dta_state &&
13605 	    now - state->dts_laststatus >= dtrace_deadman_user)
13606 		return;
13607 
13608 	/*
13609 	 * We must be sure that dts_alive never appears to be less than the
13610 	 * value upon entry to dtrace_state_deadman(), and because we lack a
13611 	 * dtrace_cas64(), we cannot store to it atomically.  We thus instead
13612 	 * store INT64_MAX to it, followed by a memory barrier, followed by
13613 	 * the new value.  This assures that dts_alive never appears to be
13614 	 * less than its true value, regardless of the order in which the
13615 	 * stores to the underlying storage are issued.
13616 	 */
13617 	state->dts_alive = INT64_MAX;
13618 	dtrace_membar_producer();
13619 	state->dts_alive = now;
13620 }
13621 
13622 dtrace_state_t *
13623 dtrace_state_create(dev_t *devp, cred_t *cr)
13624 {
13625 	minor_t minor;
13626 	major_t major;
13627 	char c[30];
13628 	dtrace_state_t *state;
13629 	dtrace_optval_t *opt;
13630 	int bufsize = NCPU * sizeof (dtrace_buffer_t), i;
13631 
13632 	ASSERT(MUTEX_HELD(&dtrace_lock));
13633 	ASSERT(MUTEX_HELD(&cpu_lock));
13634 
13635 	minor = (minor_t)(uintptr_t)vmem_alloc(dtrace_minor, 1,
13636 	    VM_BESTFIT | VM_SLEEP);
13637 
13638 	if (ddi_soft_state_zalloc(dtrace_softstate, minor) != DDI_SUCCESS) {
13639 		vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
13640 		return (NULL);
13641 	}
13642 
13643 	state = ddi_get_soft_state(dtrace_softstate, minor);
13644 	state->dts_epid = DTRACE_EPIDNONE + 1;
13645 
13646 	(void) snprintf(c, sizeof (c), "dtrace_aggid_%d", minor);
13647 	state->dts_aggid_arena = vmem_create(c, (void *)1, UINT32_MAX, 1,
13648 	    NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
13649 
13650 	if (devp != NULL) {
13651 		major = getemajor(*devp);
13652 	} else {
13653 		major = ddi_driver_major(dtrace_devi);
13654 	}
13655 
13656 	state->dts_dev = makedevice(major, minor);
13657 
13658 	if (devp != NULL)
13659 		*devp = state->dts_dev;
13660 
13661 	/*
13662 	 * We allocate NCPU buffers.  On the one hand, this can be quite
13663 	 * a bit of memory per instance (nearly 36K on a Starcat).  On the
13664 	 * other hand, it saves an additional memory reference in the probe
13665 	 * path.
13666 	 */
13667 	state->dts_buffer = kmem_zalloc(bufsize, KM_SLEEP);
13668 	state->dts_aggbuffer = kmem_zalloc(bufsize, KM_SLEEP);
13669 	state->dts_cleaner = CYCLIC_NONE;
13670 	state->dts_deadman = CYCLIC_NONE;
13671 	state->dts_vstate.dtvs_state = state;
13672 
13673 	for (i = 0; i < DTRACEOPT_MAX; i++)
13674 		state->dts_options[i] = DTRACEOPT_UNSET;
13675 
13676 	/*
13677 	 * Set the default options.
13678 	 */
13679 	opt = state->dts_options;
13680 	opt[DTRACEOPT_BUFPOLICY] = DTRACEOPT_BUFPOLICY_SWITCH;
13681 	opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_AUTO;
13682 	opt[DTRACEOPT_NSPEC] = dtrace_nspec_default;
13683 	opt[DTRACEOPT_SPECSIZE] = dtrace_specsize_default;
13684 	opt[DTRACEOPT_CPU] = (dtrace_optval_t)DTRACE_CPUALL;
13685 	opt[DTRACEOPT_STRSIZE] = dtrace_strsize_default;
13686 	opt[DTRACEOPT_STACKFRAMES] = dtrace_stackframes_default;
13687 	opt[DTRACEOPT_USTACKFRAMES] = dtrace_ustackframes_default;
13688 	opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_default;
13689 	opt[DTRACEOPT_AGGRATE] = dtrace_aggrate_default;
13690 	opt[DTRACEOPT_SWITCHRATE] = dtrace_switchrate_default;
13691 	opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_default;
13692 	opt[DTRACEOPT_JSTACKFRAMES] = dtrace_jstackframes_default;
13693 	opt[DTRACEOPT_JSTACKSTRSIZE] = dtrace_jstackstrsize_default;
13694 
13695 	state->dts_activity = DTRACE_ACTIVITY_INACTIVE;
13696 
13697 	/*
13698 	 * Depending on the user credentials, we set flag bits which alter probe
13699 	 * visibility or the amount of destructiveness allowed.  In the case of
13700 	 * actual anonymous tracing, or the possession of all privileges, all of
13701 	 * the normal checks are bypassed.
13702 	 */
13703 	if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
13704 		state->dts_cred.dcr_visible = DTRACE_CRV_ALL;
13705 		state->dts_cred.dcr_action = DTRACE_CRA_ALL;
13706 	} else {
13707 		/*
13708 		 * Set up the credentials for this instantiation.  We take a
13709 		 * hold on the credential to prevent it from disappearing on
13710 		 * us; this in turn prevents the zone_t referenced by this
13711 		 * credential from disappearing.  This means that we can
13712 		 * examine the credential and the zone from probe context.
13713 		 */
13714 		crhold(cr);
13715 		state->dts_cred.dcr_cred = cr;
13716 
13717 		/*
13718 		 * CRA_PROC means "we have *some* privilege for dtrace" and
13719 		 * unlocks the use of variables like pid, zonename, etc.
13720 		 */
13721 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE) ||
13722 		    PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
13723 			state->dts_cred.dcr_action |= DTRACE_CRA_PROC;
13724 		}
13725 
13726 		/*
13727 		 * dtrace_user allows use of syscall and profile providers.
13728 		 * If the user also has proc_owner and/or proc_zone, we
13729 		 * extend the scope to include additional visibility and
13730 		 * destructive power.
13731 		 */
13732 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE)) {
13733 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE)) {
13734 				state->dts_cred.dcr_visible |=
13735 				    DTRACE_CRV_ALLPROC;
13736 
13737 				state->dts_cred.dcr_action |=
13738 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
13739 			}
13740 
13741 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE)) {
13742 				state->dts_cred.dcr_visible |=
13743 				    DTRACE_CRV_ALLZONE;
13744 
13745 				state->dts_cred.dcr_action |=
13746 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
13747 			}
13748 
13749 			/*
13750 			 * If we have all privs in whatever zone this is,
13751 			 * we can do destructive things to processes which
13752 			 * have altered credentials.
13753 			 */
13754 			if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
13755 			    cr->cr_zone->zone_privset)) {
13756 				state->dts_cred.dcr_action |=
13757 				    DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
13758 			}
13759 		}
13760 
13761 		/*
13762 		 * Holding the dtrace_kernel privilege also implies that
13763 		 * the user has the dtrace_user privilege from a visibility
13764 		 * perspective.  But without further privileges, some
13765 		 * destructive actions are not available.
13766 		 */
13767 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE)) {
13768 			/*
13769 			 * Make all probes in all zones visible.  However,
13770 			 * this doesn't mean that all actions become available
13771 			 * to all zones.
13772 			 */
13773 			state->dts_cred.dcr_visible |= DTRACE_CRV_KERNEL |
13774 			    DTRACE_CRV_ALLPROC | DTRACE_CRV_ALLZONE;
13775 
13776 			state->dts_cred.dcr_action |= DTRACE_CRA_KERNEL |
13777 			    DTRACE_CRA_PROC;
13778 			/*
13779 			 * Holding proc_owner means that destructive actions
13780 			 * for *this* zone are allowed.
13781 			 */
13782 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
13783 				state->dts_cred.dcr_action |=
13784 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
13785 
13786 			/*
13787 			 * Holding proc_zone means that destructive actions
13788 			 * for this user/group ID in all zones is allowed.
13789 			 */
13790 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
13791 				state->dts_cred.dcr_action |=
13792 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
13793 
13794 			/*
13795 			 * If we have all privs in whatever zone this is,
13796 			 * we can do destructive things to processes which
13797 			 * have altered credentials.
13798 			 */
13799 			if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
13800 			    cr->cr_zone->zone_privset)) {
13801 				state->dts_cred.dcr_action |=
13802 				    DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
13803 			}
13804 		}
13805 
13806 		/*
13807 		 * Holding the dtrace_proc privilege gives control over fasttrap
13808 		 * and pid providers.  We need to grant wider destructive
13809 		 * privileges in the event that the user has proc_owner and/or
13810 		 * proc_zone.
13811 		 */
13812 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
13813 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
13814 				state->dts_cred.dcr_action |=
13815 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
13816 
13817 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
13818 				state->dts_cred.dcr_action |=
13819 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
13820 		}
13821 	}
13822 
13823 	return (state);
13824 }
13825 
13826 static int
13827 dtrace_state_buffer(dtrace_state_t *state, dtrace_buffer_t *buf, int which)
13828 {
13829 	dtrace_optval_t *opt = state->dts_options, size;
13830 	processorid_t cpu;
13831 	int flags = 0, rval, factor, divisor = 1;
13832 
13833 	ASSERT(MUTEX_HELD(&dtrace_lock));
13834 	ASSERT(MUTEX_HELD(&cpu_lock));
13835 	ASSERT(which < DTRACEOPT_MAX);
13836 	ASSERT(state->dts_activity == DTRACE_ACTIVITY_INACTIVE ||
13837 	    (state == dtrace_anon.dta_state &&
13838 	    state->dts_activity == DTRACE_ACTIVITY_ACTIVE));
13839 
13840 	if (opt[which] == DTRACEOPT_UNSET || opt[which] == 0)
13841 		return (0);
13842 
13843 	if (opt[DTRACEOPT_CPU] != DTRACEOPT_UNSET)
13844 		cpu = opt[DTRACEOPT_CPU];
13845 
13846 	if (which == DTRACEOPT_SPECSIZE)
13847 		flags |= DTRACEBUF_NOSWITCH;
13848 
13849 	if (which == DTRACEOPT_BUFSIZE) {
13850 		if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_RING)
13851 			flags |= DTRACEBUF_RING;
13852 
13853 		if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_FILL)
13854 			flags |= DTRACEBUF_FILL;
13855 
13856 		if (state != dtrace_anon.dta_state ||
13857 		    state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
13858 			flags |= DTRACEBUF_INACTIVE;
13859 	}
13860 
13861 	for (size = opt[which]; size >= sizeof (uint64_t); size /= divisor) {
13862 		/*
13863 		 * The size must be 8-byte aligned.  If the size is not 8-byte
13864 		 * aligned, drop it down by the difference.
13865 		 */
13866 		if (size & (sizeof (uint64_t) - 1))
13867 			size -= size & (sizeof (uint64_t) - 1);
13868 
13869 		if (size < state->dts_reserve) {
13870 			/*
13871 			 * Buffers always must be large enough to accommodate
13872 			 * their prereserved space.  We return E2BIG instead
13873 			 * of ENOMEM in this case to allow for user-level
13874 			 * software to differentiate the cases.
13875 			 */
13876 			return (E2BIG);
13877 		}
13878 
13879 		rval = dtrace_buffer_alloc(buf, size, flags, cpu, &factor);
13880 
13881 		if (rval != ENOMEM) {
13882 			opt[which] = size;
13883 			return (rval);
13884 		}
13885 
13886 		if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
13887 			return (rval);
13888 
13889 		for (divisor = 2; divisor < factor; divisor <<= 1)
13890 			continue;
13891 	}
13892 
13893 	return (ENOMEM);
13894 }
13895 
13896 static int
13897 dtrace_state_buffers(dtrace_state_t *state)
13898 {
13899 	dtrace_speculation_t *spec = state->dts_speculations;
13900 	int rval, i;
13901 
13902 	if ((rval = dtrace_state_buffer(state, state->dts_buffer,
13903 	    DTRACEOPT_BUFSIZE)) != 0)
13904 		return (rval);
13905 
13906 	if ((rval = dtrace_state_buffer(state, state->dts_aggbuffer,
13907 	    DTRACEOPT_AGGSIZE)) != 0)
13908 		return (rval);
13909 
13910 	for (i = 0; i < state->dts_nspeculations; i++) {
13911 		if ((rval = dtrace_state_buffer(state,
13912 		    spec[i].dtsp_buffer, DTRACEOPT_SPECSIZE)) != 0)
13913 			return (rval);
13914 	}
13915 
13916 	return (0);
13917 }
13918 
13919 static void
13920 dtrace_state_prereserve(dtrace_state_t *state)
13921 {
13922 	dtrace_ecb_t *ecb;
13923 	dtrace_probe_t *probe;
13924 
13925 	state->dts_reserve = 0;
13926 
13927 	if (state->dts_options[DTRACEOPT_BUFPOLICY] != DTRACEOPT_BUFPOLICY_FILL)
13928 		return;
13929 
13930 	/*
13931 	 * If our buffer policy is a "fill" buffer policy, we need to set the
13932 	 * prereserved space to be the space required by the END probes.
13933 	 */
13934 	probe = dtrace_probes[dtrace_probeid_end - 1];
13935 	ASSERT(probe != NULL);
13936 
13937 	for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
13938 		if (ecb->dte_state != state)
13939 			continue;
13940 
13941 		state->dts_reserve += ecb->dte_needed + ecb->dte_alignment;
13942 	}
13943 }
13944 
13945 static int
13946 dtrace_state_go(dtrace_state_t *state, processorid_t *cpu)
13947 {
13948 	dtrace_optval_t *opt = state->dts_options, sz, nspec;
13949 	dtrace_speculation_t *spec;
13950 	dtrace_buffer_t *buf;
13951 	cyc_handler_t hdlr;
13952 	cyc_time_t when;
13953 	int rval = 0, i, bufsize = NCPU * sizeof (dtrace_buffer_t);
13954 	dtrace_icookie_t cookie;
13955 
13956 	mutex_enter(&cpu_lock);
13957 	mutex_enter(&dtrace_lock);
13958 
13959 	if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
13960 		rval = EBUSY;
13961 		goto out;
13962 	}
13963 
13964 	/*
13965 	 * Before we can perform any checks, we must prime all of the
13966 	 * retained enablings that correspond to this state.
13967 	 */
13968 	dtrace_enabling_prime(state);
13969 
13970 	if (state->dts_destructive && !state->dts_cred.dcr_destructive) {
13971 		rval = EACCES;
13972 		goto out;
13973 	}
13974 
13975 	dtrace_state_prereserve(state);
13976 
13977 	/*
13978 	 * Now we want to do is try to allocate our speculations.
13979 	 * We do not automatically resize the number of speculations; if
13980 	 * this fails, we will fail the operation.
13981 	 */
13982 	nspec = opt[DTRACEOPT_NSPEC];
13983 	ASSERT(nspec != DTRACEOPT_UNSET);
13984 
13985 	if (nspec > INT_MAX) {
13986 		rval = ENOMEM;
13987 		goto out;
13988 	}
13989 
13990 	spec = kmem_zalloc(nspec * sizeof (dtrace_speculation_t),
13991 	    KM_NOSLEEP | KM_NORMALPRI);
13992 
13993 	if (spec == NULL) {
13994 		rval = ENOMEM;
13995 		goto out;
13996 	}
13997 
13998 	state->dts_speculations = spec;
13999 	state->dts_nspeculations = (int)nspec;
14000 
14001 	for (i = 0; i < nspec; i++) {
14002 		if ((buf = kmem_zalloc(bufsize,
14003 		    KM_NOSLEEP | KM_NORMALPRI)) == NULL) {
14004 			rval = ENOMEM;
14005 			goto err;
14006 		}
14007 
14008 		spec[i].dtsp_buffer = buf;
14009 	}
14010 
14011 	if (opt[DTRACEOPT_GRABANON] != DTRACEOPT_UNSET) {
14012 		if (dtrace_anon.dta_state == NULL) {
14013 			rval = ENOENT;
14014 			goto out;
14015 		}
14016 
14017 		if (state->dts_necbs != 0) {
14018 			rval = EALREADY;
14019 			goto out;
14020 		}
14021 
14022 		state->dts_anon = dtrace_anon_grab();
14023 		ASSERT(state->dts_anon != NULL);
14024 		state = state->dts_anon;
14025 
14026 		/*
14027 		 * We want "grabanon" to be set in the grabbed state, so we'll
14028 		 * copy that option value from the grabbing state into the
14029 		 * grabbed state.
14030 		 */
14031 		state->dts_options[DTRACEOPT_GRABANON] =
14032 		    opt[DTRACEOPT_GRABANON];
14033 
14034 		*cpu = dtrace_anon.dta_beganon;
14035 
14036 		/*
14037 		 * If the anonymous state is active (as it almost certainly
14038 		 * is if the anonymous enabling ultimately matched anything),
14039 		 * we don't allow any further option processing -- but we
14040 		 * don't return failure.
14041 		 */
14042 		if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
14043 			goto out;
14044 	}
14045 
14046 	if (opt[DTRACEOPT_AGGSIZE] != DTRACEOPT_UNSET &&
14047 	    opt[DTRACEOPT_AGGSIZE] != 0) {
14048 		if (state->dts_aggregations == NULL) {
14049 			/*
14050 			 * We're not going to create an aggregation buffer
14051 			 * because we don't have any ECBs that contain
14052 			 * aggregations -- set this option to 0.
14053 			 */
14054 			opt[DTRACEOPT_AGGSIZE] = 0;
14055 		} else {
14056 			/*
14057 			 * If we have an aggregation buffer, we must also have
14058 			 * a buffer to use as scratch.
14059 			 */
14060 			if (opt[DTRACEOPT_BUFSIZE] == DTRACEOPT_UNSET ||
14061 			    opt[DTRACEOPT_BUFSIZE] < state->dts_needed) {
14062 				opt[DTRACEOPT_BUFSIZE] = state->dts_needed;
14063 			}
14064 		}
14065 	}
14066 
14067 	if (opt[DTRACEOPT_SPECSIZE] != DTRACEOPT_UNSET &&
14068 	    opt[DTRACEOPT_SPECSIZE] != 0) {
14069 		if (!state->dts_speculates) {
14070 			/*
14071 			 * We're not going to create speculation buffers
14072 			 * because we don't have any ECBs that actually
14073 			 * speculate -- set the speculation size to 0.
14074 			 */
14075 			opt[DTRACEOPT_SPECSIZE] = 0;
14076 		}
14077 	}
14078 
14079 	/*
14080 	 * The bare minimum size for any buffer that we're actually going to
14081 	 * do anything to is sizeof (uint64_t).
14082 	 */
14083 	sz = sizeof (uint64_t);
14084 
14085 	if ((state->dts_needed != 0 && opt[DTRACEOPT_BUFSIZE] < sz) ||
14086 	    (state->dts_speculates && opt[DTRACEOPT_SPECSIZE] < sz) ||
14087 	    (state->dts_aggregations != NULL && opt[DTRACEOPT_AGGSIZE] < sz)) {
14088 		/*
14089 		 * A buffer size has been explicitly set to 0 (or to a size
14090 		 * that will be adjusted to 0) and we need the space -- we
14091 		 * need to return failure.  We return ENOSPC to differentiate
14092 		 * it from failing to allocate a buffer due to failure to meet
14093 		 * the reserve (for which we return E2BIG).
14094 		 */
14095 		rval = ENOSPC;
14096 		goto out;
14097 	}
14098 
14099 	if ((rval = dtrace_state_buffers(state)) != 0)
14100 		goto err;
14101 
14102 	if ((sz = opt[DTRACEOPT_DYNVARSIZE]) == DTRACEOPT_UNSET)
14103 		sz = dtrace_dstate_defsize;
14104 
14105 	do {
14106 		rval = dtrace_dstate_init(&state->dts_vstate.dtvs_dynvars, sz);
14107 
14108 		if (rval == 0)
14109 			break;
14110 
14111 		if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
14112 			goto err;
14113 	} while (sz >>= 1);
14114 
14115 	opt[DTRACEOPT_DYNVARSIZE] = sz;
14116 
14117 	if (rval != 0)
14118 		goto err;
14119 
14120 	if (opt[DTRACEOPT_STATUSRATE] > dtrace_statusrate_max)
14121 		opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_max;
14122 
14123 	if (opt[DTRACEOPT_CLEANRATE] == 0)
14124 		opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
14125 
14126 	if (opt[DTRACEOPT_CLEANRATE] < dtrace_cleanrate_min)
14127 		opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_min;
14128 
14129 	if (opt[DTRACEOPT_CLEANRATE] > dtrace_cleanrate_max)
14130 		opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
14131 
14132 	hdlr.cyh_func = (cyc_func_t)dtrace_state_clean;
14133 	hdlr.cyh_arg = state;
14134 	hdlr.cyh_level = CY_LOW_LEVEL;
14135 
14136 	when.cyt_when = 0;
14137 	when.cyt_interval = opt[DTRACEOPT_CLEANRATE];
14138 
14139 	state->dts_cleaner = cyclic_add(&hdlr, &when);
14140 
14141 	hdlr.cyh_func = (cyc_func_t)dtrace_state_deadman;
14142 	hdlr.cyh_arg = state;
14143 	hdlr.cyh_level = CY_LOW_LEVEL;
14144 
14145 	when.cyt_when = 0;
14146 	when.cyt_interval = dtrace_deadman_interval;
14147 
14148 	state->dts_alive = state->dts_laststatus = dtrace_gethrtime();
14149 	state->dts_deadman = cyclic_add(&hdlr, &when);
14150 
14151 	state->dts_activity = DTRACE_ACTIVITY_WARMUP;
14152 
14153 	if (state->dts_getf != 0 &&
14154 	    !(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)) {
14155 		/*
14156 		 * We don't have kernel privs but we have at least one call
14157 		 * to getf(); we need to bump our zone's count, and (if
14158 		 * this is the first enabling to have an unprivileged call
14159 		 * to getf()) we need to hook into closef().
14160 		 */
14161 		state->dts_cred.dcr_cred->cr_zone->zone_dtrace_getf++;
14162 
14163 		if (dtrace_getf++ == 0) {
14164 			ASSERT(dtrace_closef == NULL);
14165 			dtrace_closef = dtrace_getf_barrier;
14166 		}
14167 	}
14168 
14169 	/*
14170 	 * Now it's time to actually fire the BEGIN probe.  We need to disable
14171 	 * interrupts here both to record the CPU on which we fired the BEGIN
14172 	 * probe (the data from this CPU will be processed first at user
14173 	 * level) and to manually activate the buffer for this CPU.
14174 	 */
14175 	cookie = dtrace_interrupt_disable();
14176 	*cpu = CPU->cpu_id;
14177 	ASSERT(state->dts_buffer[*cpu].dtb_flags & DTRACEBUF_INACTIVE);
14178 	state->dts_buffer[*cpu].dtb_flags &= ~DTRACEBUF_INACTIVE;
14179 
14180 	dtrace_probe(dtrace_probeid_begin,
14181 	    (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
14182 	dtrace_interrupt_enable(cookie);
14183 	/*
14184 	 * We may have had an exit action from a BEGIN probe; only change our
14185 	 * state to ACTIVE if we're still in WARMUP.
14186 	 */
14187 	ASSERT(state->dts_activity == DTRACE_ACTIVITY_WARMUP ||
14188 	    state->dts_activity == DTRACE_ACTIVITY_DRAINING);
14189 
14190 	if (state->dts_activity == DTRACE_ACTIVITY_WARMUP)
14191 		state->dts_activity = DTRACE_ACTIVITY_ACTIVE;
14192 
14193 	/*
14194 	 * Regardless of whether or not now we're in ACTIVE or DRAINING, we
14195 	 * want each CPU to transition its principal buffer out of the
14196 	 * INACTIVE state.  Doing this assures that no CPU will suddenly begin
14197 	 * processing an ECB halfway down a probe's ECB chain; all CPUs will
14198 	 * atomically transition from processing none of a state's ECBs to
14199 	 * processing all of them.
14200 	 */
14201 	dtrace_xcall(DTRACE_CPUALL,
14202 	    (dtrace_xcall_t)dtrace_buffer_activate, state);
14203 	goto out;
14204 
14205 err:
14206 	dtrace_buffer_free(state->dts_buffer);
14207 	dtrace_buffer_free(state->dts_aggbuffer);
14208 
14209 	if ((nspec = state->dts_nspeculations) == 0) {
14210 		ASSERT(state->dts_speculations == NULL);
14211 		goto out;
14212 	}
14213 
14214 	spec = state->dts_speculations;
14215 	ASSERT(spec != NULL);
14216 
14217 	for (i = 0; i < state->dts_nspeculations; i++) {
14218 		if ((buf = spec[i].dtsp_buffer) == NULL)
14219 			break;
14220 
14221 		dtrace_buffer_free(buf);
14222 		kmem_free(buf, bufsize);
14223 	}
14224 
14225 	kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
14226 	state->dts_nspeculations = 0;
14227 	state->dts_speculations = NULL;
14228 
14229 out:
14230 	mutex_exit(&dtrace_lock);
14231 	mutex_exit(&cpu_lock);
14232 
14233 	return (rval);
14234 }
14235 
14236 static int
14237 dtrace_state_stop(dtrace_state_t *state, processorid_t *cpu)
14238 {
14239 	dtrace_icookie_t cookie;
14240 
14241 	ASSERT(MUTEX_HELD(&dtrace_lock));
14242 
14243 	if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE &&
14244 	    state->dts_activity != DTRACE_ACTIVITY_DRAINING)
14245 		return (EINVAL);
14246 
14247 	/*
14248 	 * We'll set the activity to DTRACE_ACTIVITY_DRAINING, and issue a sync
14249 	 * to be sure that every CPU has seen it.  See below for the details
14250 	 * on why this is done.
14251 	 */
14252 	state->dts_activity = DTRACE_ACTIVITY_DRAINING;
14253 	dtrace_sync();
14254 
14255 	/*
14256 	 * By this point, it is impossible for any CPU to be still processing
14257 	 * with DTRACE_ACTIVITY_ACTIVE.  We can thus set our activity to
14258 	 * DTRACE_ACTIVITY_COOLDOWN and know that we're not racing with any
14259 	 * other CPU in dtrace_buffer_reserve().  This allows dtrace_probe()
14260 	 * and callees to know that the activity is DTRACE_ACTIVITY_COOLDOWN
14261 	 * iff we're in the END probe.
14262 	 */
14263 	state->dts_activity = DTRACE_ACTIVITY_COOLDOWN;
14264 	dtrace_sync();
14265 	ASSERT(state->dts_activity == DTRACE_ACTIVITY_COOLDOWN);
14266 
14267 	/*
14268 	 * Finally, we can release the reserve and call the END probe.  We
14269 	 * disable interrupts across calling the END probe to allow us to
14270 	 * return the CPU on which we actually called the END probe.  This
14271 	 * allows user-land to be sure that this CPU's principal buffer is
14272 	 * processed last.
14273 	 */
14274 	state->dts_reserve = 0;
14275 
14276 	cookie = dtrace_interrupt_disable();
14277 	*cpu = CPU->cpu_id;
14278 	dtrace_probe(dtrace_probeid_end,
14279 	    (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
14280 	dtrace_interrupt_enable(cookie);
14281 
14282 	state->dts_activity = DTRACE_ACTIVITY_STOPPED;
14283 	dtrace_sync();
14284 
14285 	if (state->dts_getf != 0 &&
14286 	    !(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)) {
14287 		/*
14288 		 * We don't have kernel privs but we have at least one call
14289 		 * to getf(); we need to lower our zone's count, and (if
14290 		 * this is the last enabling to have an unprivileged call
14291 		 * to getf()) we need to clear the closef() hook.
14292 		 */
14293 		ASSERT(state->dts_cred.dcr_cred->cr_zone->zone_dtrace_getf > 0);
14294 		ASSERT(dtrace_closef == dtrace_getf_barrier);
14295 		ASSERT(dtrace_getf > 0);
14296 
14297 		state->dts_cred.dcr_cred->cr_zone->zone_dtrace_getf--;
14298 
14299 		if (--dtrace_getf == 0)
14300 			dtrace_closef = NULL;
14301 	}
14302 
14303 	return (0);
14304 }
14305 
14306 static int
14307 dtrace_state_option(dtrace_state_t *state, dtrace_optid_t option,
14308     dtrace_optval_t val)
14309 {
14310 	ASSERT(MUTEX_HELD(&dtrace_lock));
14311 
14312 	if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
14313 		return (EBUSY);
14314 
14315 	if (option >= DTRACEOPT_MAX)
14316 		return (EINVAL);
14317 
14318 	if (option != DTRACEOPT_CPU && val < 0)
14319 		return (EINVAL);
14320 
14321 	switch (option) {
14322 	case DTRACEOPT_DESTRUCTIVE:
14323 		if (dtrace_destructive_disallow)
14324 			return (EACCES);
14325 
14326 		state->dts_cred.dcr_destructive = 1;
14327 		break;
14328 
14329 	case DTRACEOPT_BUFSIZE:
14330 	case DTRACEOPT_DYNVARSIZE:
14331 	case DTRACEOPT_AGGSIZE:
14332 	case DTRACEOPT_SPECSIZE:
14333 	case DTRACEOPT_STRSIZE:
14334 		if (val < 0)
14335 			return (EINVAL);
14336 
14337 		if (val >= LONG_MAX) {
14338 			/*
14339 			 * If this is an otherwise negative value, set it to
14340 			 * the highest multiple of 128m less than LONG_MAX.
14341 			 * Technically, we're adjusting the size without
14342 			 * regard to the buffer resizing policy, but in fact,
14343 			 * this has no effect -- if we set the buffer size to
14344 			 * ~LONG_MAX and the buffer policy is ultimately set to
14345 			 * be "manual", the buffer allocation is guaranteed to
14346 			 * fail, if only because the allocation requires two
14347 			 * buffers.  (We set the the size to the highest
14348 			 * multiple of 128m because it ensures that the size
14349 			 * will remain a multiple of a megabyte when
14350 			 * repeatedly halved -- all the way down to 15m.)
14351 			 */
14352 			val = LONG_MAX - (1 << 27) + 1;
14353 		}
14354 	}
14355 
14356 	state->dts_options[option] = val;
14357 
14358 	return (0);
14359 }
14360 
14361 static void
14362 dtrace_state_destroy(dtrace_state_t *state)
14363 {
14364 	dtrace_ecb_t *ecb;
14365 	dtrace_vstate_t *vstate = &state->dts_vstate;
14366 	minor_t minor = getminor(state->dts_dev);
14367 	int i, bufsize = NCPU * sizeof (dtrace_buffer_t);
14368 	dtrace_speculation_t *spec = state->dts_speculations;
14369 	int nspec = state->dts_nspeculations;
14370 	uint32_t match;
14371 
14372 	ASSERT(MUTEX_HELD(&dtrace_lock));
14373 	ASSERT(MUTEX_HELD(&cpu_lock));
14374 
14375 	/*
14376 	 * First, retract any retained enablings for this state.
14377 	 */
14378 	dtrace_enabling_retract(state);
14379 	ASSERT(state->dts_nretained == 0);
14380 
14381 	if (state->dts_activity == DTRACE_ACTIVITY_ACTIVE ||
14382 	    state->dts_activity == DTRACE_ACTIVITY_DRAINING) {
14383 		/*
14384 		 * We have managed to come into dtrace_state_destroy() on a
14385 		 * hot enabling -- almost certainly because of a disorderly
14386 		 * shutdown of a consumer.  (That is, a consumer that is
14387 		 * exiting without having called dtrace_stop().) In this case,
14388 		 * we're going to set our activity to be KILLED, and then
14389 		 * issue a sync to be sure that everyone is out of probe
14390 		 * context before we start blowing away ECBs.
14391 		 */
14392 		state->dts_activity = DTRACE_ACTIVITY_KILLED;
14393 		dtrace_sync();
14394 	}
14395 
14396 	/*
14397 	 * Release the credential hold we took in dtrace_state_create().
14398 	 */
14399 	if (state->dts_cred.dcr_cred != NULL)
14400 		crfree(state->dts_cred.dcr_cred);
14401 
14402 	/*
14403 	 * Now we can safely disable and destroy any enabled probes.  Because
14404 	 * any DTRACE_PRIV_KERNEL probes may actually be slowing our progress
14405 	 * (especially if they're all enabled), we take two passes through the
14406 	 * ECBs:  in the first, we disable just DTRACE_PRIV_KERNEL probes, and
14407 	 * in the second we disable whatever is left over.
14408 	 */
14409 	for (match = DTRACE_PRIV_KERNEL; ; match = 0) {
14410 		for (i = 0; i < state->dts_necbs; i++) {
14411 			if ((ecb = state->dts_ecbs[i]) == NULL)
14412 				continue;
14413 
14414 			if (match && ecb->dte_probe != NULL) {
14415 				dtrace_probe_t *probe = ecb->dte_probe;
14416 				dtrace_provider_t *prov = probe->dtpr_provider;
14417 
14418 				if (!(prov->dtpv_priv.dtpp_flags & match))
14419 					continue;
14420 			}
14421 
14422 			dtrace_ecb_disable(ecb);
14423 			dtrace_ecb_destroy(ecb);
14424 		}
14425 
14426 		if (!match)
14427 			break;
14428 	}
14429 
14430 	/*
14431 	 * Before we free the buffers, perform one more sync to assure that
14432 	 * every CPU is out of probe context.
14433 	 */
14434 	dtrace_sync();
14435 
14436 	dtrace_buffer_free(state->dts_buffer);
14437 	dtrace_buffer_free(state->dts_aggbuffer);
14438 
14439 	for (i = 0; i < nspec; i++)
14440 		dtrace_buffer_free(spec[i].dtsp_buffer);
14441 
14442 	if (state->dts_cleaner != CYCLIC_NONE)
14443 		cyclic_remove(state->dts_cleaner);
14444 
14445 	if (state->dts_deadman != CYCLIC_NONE)
14446 		cyclic_remove(state->dts_deadman);
14447 
14448 	dtrace_dstate_fini(&vstate->dtvs_dynvars);
14449 	dtrace_vstate_fini(vstate);
14450 	kmem_free(state->dts_ecbs, state->dts_necbs * sizeof (dtrace_ecb_t *));
14451 
14452 	if (state->dts_aggregations != NULL) {
14453 #ifdef DEBUG
14454 		for (i = 0; i < state->dts_naggregations; i++)
14455 			ASSERT(state->dts_aggregations[i] == NULL);
14456 #endif
14457 		ASSERT(state->dts_naggregations > 0);
14458 		kmem_free(state->dts_aggregations,
14459 		    state->dts_naggregations * sizeof (dtrace_aggregation_t *));
14460 	}
14461 
14462 	kmem_free(state->dts_buffer, bufsize);
14463 	kmem_free(state->dts_aggbuffer, bufsize);
14464 
14465 	for (i = 0; i < nspec; i++)
14466 		kmem_free(spec[i].dtsp_buffer, bufsize);
14467 
14468 	kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
14469 
14470 	dtrace_format_destroy(state);
14471 
14472 	vmem_destroy(state->dts_aggid_arena);
14473 	ddi_soft_state_free(dtrace_softstate, minor);
14474 	vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
14475 }
14476 
14477 /*
14478  * DTrace Anonymous Enabling Functions
14479  */
14480 static dtrace_state_t *
14481 dtrace_anon_grab(void)
14482 {
14483 	dtrace_state_t *state;
14484 
14485 	ASSERT(MUTEX_HELD(&dtrace_lock));
14486 
14487 	if ((state = dtrace_anon.dta_state) == NULL) {
14488 		ASSERT(dtrace_anon.dta_enabling == NULL);
14489 		return (NULL);
14490 	}
14491 
14492 	ASSERT(dtrace_anon.dta_enabling != NULL);
14493 	ASSERT(dtrace_retained != NULL);
14494 
14495 	dtrace_enabling_destroy(dtrace_anon.dta_enabling);
14496 	dtrace_anon.dta_enabling = NULL;
14497 	dtrace_anon.dta_state = NULL;
14498 
14499 	return (state);
14500 }
14501 
14502 static void
14503 dtrace_anon_property(void)
14504 {
14505 	int i, rv;
14506 	dtrace_state_t *state;
14507 	dof_hdr_t *dof;
14508 	char c[32];		/* enough for "dof-data-" + digits */
14509 
14510 	ASSERT(MUTEX_HELD(&dtrace_lock));
14511 	ASSERT(MUTEX_HELD(&cpu_lock));
14512 
14513 	for (i = 0; ; i++) {
14514 		(void) snprintf(c, sizeof (c), "dof-data-%d", i);
14515 
14516 		dtrace_err_verbose = 1;
14517 
14518 		if ((dof = dtrace_dof_property(c)) == NULL) {
14519 			dtrace_err_verbose = 0;
14520 			break;
14521 		}
14522 
14523 		/*
14524 		 * We want to create anonymous state, so we need to transition
14525 		 * the kernel debugger to indicate that DTrace is active.  If
14526 		 * this fails (e.g. because the debugger has modified text in
14527 		 * some way), we won't continue with the processing.
14528 		 */
14529 		if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
14530 			cmn_err(CE_NOTE, "kernel debugger active; anonymous "
14531 			    "enabling ignored.");
14532 			dtrace_dof_destroy(dof);
14533 			break;
14534 		}
14535 
14536 		/*
14537 		 * If we haven't allocated an anonymous state, we'll do so now.
14538 		 */
14539 		if ((state = dtrace_anon.dta_state) == NULL) {
14540 			state = dtrace_state_create(NULL, NULL);
14541 			dtrace_anon.dta_state = state;
14542 
14543 			if (state == NULL) {
14544 				/*
14545 				 * This basically shouldn't happen:  the only
14546 				 * failure mode from dtrace_state_create() is a
14547 				 * failure of ddi_soft_state_zalloc() that
14548 				 * itself should never happen.  Still, the
14549 				 * interface allows for a failure mode, and
14550 				 * we want to fail as gracefully as possible:
14551 				 * we'll emit an error message and cease
14552 				 * processing anonymous state in this case.
14553 				 */
14554 				cmn_err(CE_WARN, "failed to create "
14555 				    "anonymous state");
14556 				dtrace_dof_destroy(dof);
14557 				break;
14558 			}
14559 		}
14560 
14561 		rv = dtrace_dof_slurp(dof, &state->dts_vstate, CRED(),
14562 		    &dtrace_anon.dta_enabling, 0, B_TRUE);
14563 
14564 		if (rv == 0)
14565 			rv = dtrace_dof_options(dof, state);
14566 
14567 		dtrace_err_verbose = 0;
14568 		dtrace_dof_destroy(dof);
14569 
14570 		if (rv != 0) {
14571 			/*
14572 			 * This is malformed DOF; chuck any anonymous state
14573 			 * that we created.
14574 			 */
14575 			ASSERT(dtrace_anon.dta_enabling == NULL);
14576 			dtrace_state_destroy(state);
14577 			dtrace_anon.dta_state = NULL;
14578 			break;
14579 		}
14580 
14581 		ASSERT(dtrace_anon.dta_enabling != NULL);
14582 	}
14583 
14584 	if (dtrace_anon.dta_enabling != NULL) {
14585 		int rval;
14586 
14587 		/*
14588 		 * dtrace_enabling_retain() can only fail because we are
14589 		 * trying to retain more enablings than are allowed -- but
14590 		 * we only have one anonymous enabling, and we are guaranteed
14591 		 * to be allowed at least one retained enabling; we assert
14592 		 * that dtrace_enabling_retain() returns success.
14593 		 */
14594 		rval = dtrace_enabling_retain(dtrace_anon.dta_enabling);
14595 		ASSERT(rval == 0);
14596 
14597 		dtrace_enabling_dump(dtrace_anon.dta_enabling);
14598 	}
14599 }
14600 
14601 /*
14602  * DTrace Helper Functions
14603  */
14604 static void
14605 dtrace_helper_trace(dtrace_helper_action_t *helper,
14606     dtrace_mstate_t *mstate, dtrace_vstate_t *vstate, int where)
14607 {
14608 	uint32_t size, next, nnext, i;
14609 	dtrace_helptrace_t *ent, *buffer;
14610 	uint16_t flags = cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
14611 
14612 	if ((buffer = dtrace_helptrace_buffer) == NULL)
14613 		return;
14614 
14615 	ASSERT(vstate->dtvs_nlocals <= dtrace_helptrace_nlocals);
14616 
14617 	/*
14618 	 * What would a tracing framework be without its own tracing
14619 	 * framework?  (Well, a hell of a lot simpler, for starters...)
14620 	 */
14621 	size = sizeof (dtrace_helptrace_t) + dtrace_helptrace_nlocals *
14622 	    sizeof (uint64_t) - sizeof (uint64_t);
14623 
14624 	/*
14625 	 * Iterate until we can allocate a slot in the trace buffer.
14626 	 */
14627 	do {
14628 		next = dtrace_helptrace_next;
14629 
14630 		if (next + size < dtrace_helptrace_bufsize) {
14631 			nnext = next + size;
14632 		} else {
14633 			nnext = size;
14634 		}
14635 	} while (dtrace_cas32(&dtrace_helptrace_next, next, nnext) != next);
14636 
14637 	/*
14638 	 * We have our slot; fill it in.
14639 	 */
14640 	if (nnext == size) {
14641 		dtrace_helptrace_wrapped++;
14642 		next = 0;
14643 	}
14644 
14645 	ent = (dtrace_helptrace_t *)((uintptr_t)buffer + next);
14646 	ent->dtht_helper = helper;
14647 	ent->dtht_where = where;
14648 	ent->dtht_nlocals = vstate->dtvs_nlocals;
14649 
14650 	ent->dtht_fltoffs = (mstate->dtms_present & DTRACE_MSTATE_FLTOFFS) ?
14651 	    mstate->dtms_fltoffs : -1;
14652 	ent->dtht_fault = DTRACE_FLAGS2FLT(flags);
14653 	ent->dtht_illval = cpu_core[CPU->cpu_id].cpuc_dtrace_illval;
14654 
14655 	for (i = 0; i < vstate->dtvs_nlocals; i++) {
14656 		dtrace_statvar_t *svar;
14657 
14658 		if ((svar = vstate->dtvs_locals[i]) == NULL)
14659 			continue;
14660 
14661 		ASSERT(svar->dtsv_size >= NCPU * sizeof (uint64_t));
14662 		ent->dtht_locals[i] =
14663 		    ((uint64_t *)(uintptr_t)svar->dtsv_data)[CPU->cpu_id];
14664 	}
14665 }
14666 
14667 static uint64_t
14668 dtrace_helper(int which, dtrace_mstate_t *mstate,
14669     dtrace_state_t *state, uint64_t arg0, uint64_t arg1)
14670 {
14671 	uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
14672 	uint64_t sarg0 = mstate->dtms_arg[0];
14673 	uint64_t sarg1 = mstate->dtms_arg[1];
14674 	uint64_t rval;
14675 	dtrace_helpers_t *helpers = curproc->p_dtrace_helpers;
14676 	dtrace_helper_action_t *helper;
14677 	dtrace_vstate_t *vstate;
14678 	dtrace_difo_t *pred;
14679 	int i, trace = dtrace_helptrace_buffer != NULL;
14680 
14681 	ASSERT(which >= 0 && which < DTRACE_NHELPER_ACTIONS);
14682 
14683 	if (helpers == NULL)
14684 		return (0);
14685 
14686 	if ((helper = helpers->dthps_actions[which]) == NULL)
14687 		return (0);
14688 
14689 	vstate = &helpers->dthps_vstate;
14690 	mstate->dtms_arg[0] = arg0;
14691 	mstate->dtms_arg[1] = arg1;
14692 
14693 	/*
14694 	 * Now iterate over each helper.  If its predicate evaluates to 'true',
14695 	 * we'll call the corresponding actions.  Note that the below calls
14696 	 * to dtrace_dif_emulate() may set faults in machine state.  This is
14697 	 * okay:  our caller (the outer dtrace_dif_emulate()) will simply plow
14698 	 * the stored DIF offset with its own (which is the desired behavior).
14699 	 * Also, note the calls to dtrace_dif_emulate() may allocate scratch
14700 	 * from machine state; this is okay, too.
14701 	 */
14702 	for (; helper != NULL; helper = helper->dtha_next) {
14703 		if ((pred = helper->dtha_predicate) != NULL) {
14704 			if (trace)
14705 				dtrace_helper_trace(helper, mstate, vstate, 0);
14706 
14707 			if (!dtrace_dif_emulate(pred, mstate, vstate, state))
14708 				goto next;
14709 
14710 			if (*flags & CPU_DTRACE_FAULT)
14711 				goto err;
14712 		}
14713 
14714 		for (i = 0; i < helper->dtha_nactions; i++) {
14715 			if (trace)
14716 				dtrace_helper_trace(helper,
14717 				    mstate, vstate, i + 1);
14718 
14719 			rval = dtrace_dif_emulate(helper->dtha_actions[i],
14720 			    mstate, vstate, state);
14721 
14722 			if (*flags & CPU_DTRACE_FAULT)
14723 				goto err;
14724 		}
14725 
14726 next:
14727 		if (trace)
14728 			dtrace_helper_trace(helper, mstate, vstate,
14729 			    DTRACE_HELPTRACE_NEXT);
14730 	}
14731 
14732 	if (trace)
14733 		dtrace_helper_trace(helper, mstate, vstate,
14734 		    DTRACE_HELPTRACE_DONE);
14735 
14736 	/*
14737 	 * Restore the arg0 that we saved upon entry.
14738 	 */
14739 	mstate->dtms_arg[0] = sarg0;
14740 	mstate->dtms_arg[1] = sarg1;
14741 
14742 	return (rval);
14743 
14744 err:
14745 	if (trace)
14746 		dtrace_helper_trace(helper, mstate, vstate,
14747 		    DTRACE_HELPTRACE_ERR);
14748 
14749 	/*
14750 	 * Restore the arg0 that we saved upon entry.
14751 	 */
14752 	mstate->dtms_arg[0] = sarg0;
14753 	mstate->dtms_arg[1] = sarg1;
14754 
14755 	return (NULL);
14756 }
14757 
14758 static void
14759 dtrace_helper_action_destroy(dtrace_helper_action_t *helper,
14760     dtrace_vstate_t *vstate)
14761 {
14762 	int i;
14763 
14764 	if (helper->dtha_predicate != NULL)
14765 		dtrace_difo_release(helper->dtha_predicate, vstate);
14766 
14767 	for (i = 0; i < helper->dtha_nactions; i++) {
14768 		ASSERT(helper->dtha_actions[i] != NULL);
14769 		dtrace_difo_release(helper->dtha_actions[i], vstate);
14770 	}
14771 
14772 	kmem_free(helper->dtha_actions,
14773 	    helper->dtha_nactions * sizeof (dtrace_difo_t *));
14774 	kmem_free(helper, sizeof (dtrace_helper_action_t));
14775 }
14776 
14777 static int
14778 dtrace_helper_destroygen(int gen)
14779 {
14780 	proc_t *p = curproc;
14781 	dtrace_helpers_t *help = p->p_dtrace_helpers;
14782 	dtrace_vstate_t *vstate;
14783 	int i;
14784 
14785 	ASSERT(MUTEX_HELD(&dtrace_lock));
14786 
14787 	if (help == NULL || gen > help->dthps_generation)
14788 		return (EINVAL);
14789 
14790 	vstate = &help->dthps_vstate;
14791 
14792 	for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
14793 		dtrace_helper_action_t *last = NULL, *h, *next;
14794 
14795 		for (h = help->dthps_actions[i]; h != NULL; h = next) {
14796 			next = h->dtha_next;
14797 
14798 			if (h->dtha_generation == gen) {
14799 				if (last != NULL) {
14800 					last->dtha_next = next;
14801 				} else {
14802 					help->dthps_actions[i] = next;
14803 				}
14804 
14805 				dtrace_helper_action_destroy(h, vstate);
14806 			} else {
14807 				last = h;
14808 			}
14809 		}
14810 	}
14811 
14812 	/*
14813 	 * Interate until we've cleared out all helper providers with the
14814 	 * given generation number.
14815 	 */
14816 	for (;;) {
14817 		dtrace_helper_provider_t *prov;
14818 
14819 		/*
14820 		 * Look for a helper provider with the right generation. We
14821 		 * have to start back at the beginning of the list each time
14822 		 * because we drop dtrace_lock. It's unlikely that we'll make
14823 		 * more than two passes.
14824 		 */
14825 		for (i = 0; i < help->dthps_nprovs; i++) {
14826 			prov = help->dthps_provs[i];
14827 
14828 			if (prov->dthp_generation == gen)
14829 				break;
14830 		}
14831 
14832 		/*
14833 		 * If there were no matches, we're done.
14834 		 */
14835 		if (i == help->dthps_nprovs)
14836 			break;
14837 
14838 		/*
14839 		 * Move the last helper provider into this slot.
14840 		 */
14841 		help->dthps_nprovs--;
14842 		help->dthps_provs[i] = help->dthps_provs[help->dthps_nprovs];
14843 		help->dthps_provs[help->dthps_nprovs] = NULL;
14844 
14845 		mutex_exit(&dtrace_lock);
14846 
14847 		/*
14848 		 * If we have a meta provider, remove this helper provider.
14849 		 */
14850 		mutex_enter(&dtrace_meta_lock);
14851 		if (dtrace_meta_pid != NULL) {
14852 			ASSERT(dtrace_deferred_pid == NULL);
14853 			dtrace_helper_provider_remove(&prov->dthp_prov,
14854 			    p->p_pid);
14855 		}
14856 		mutex_exit(&dtrace_meta_lock);
14857 
14858 		dtrace_helper_provider_destroy(prov);
14859 
14860 		mutex_enter(&dtrace_lock);
14861 	}
14862 
14863 	return (0);
14864 }
14865 
14866 static int
14867 dtrace_helper_validate(dtrace_helper_action_t *helper)
14868 {
14869 	int err = 0, i;
14870 	dtrace_difo_t *dp;
14871 
14872 	if ((dp = helper->dtha_predicate) != NULL)
14873 		err += dtrace_difo_validate_helper(dp);
14874 
14875 	for (i = 0; i < helper->dtha_nactions; i++)
14876 		err += dtrace_difo_validate_helper(helper->dtha_actions[i]);
14877 
14878 	return (err == 0);
14879 }
14880 
14881 static int
14882 dtrace_helper_action_add(int which, dtrace_ecbdesc_t *ep)
14883 {
14884 	dtrace_helpers_t *help;
14885 	dtrace_helper_action_t *helper, *last;
14886 	dtrace_actdesc_t *act;
14887 	dtrace_vstate_t *vstate;
14888 	dtrace_predicate_t *pred;
14889 	int count = 0, nactions = 0, i;
14890 
14891 	if (which < 0 || which >= DTRACE_NHELPER_ACTIONS)
14892 		return (EINVAL);
14893 
14894 	help = curproc->p_dtrace_helpers;
14895 	last = help->dthps_actions[which];
14896 	vstate = &help->dthps_vstate;
14897 
14898 	for (count = 0; last != NULL; last = last->dtha_next) {
14899 		count++;
14900 		if (last->dtha_next == NULL)
14901 			break;
14902 	}
14903 
14904 	/*
14905 	 * If we already have dtrace_helper_actions_max helper actions for this
14906 	 * helper action type, we'll refuse to add a new one.
14907 	 */
14908 	if (count >= dtrace_helper_actions_max)
14909 		return (ENOSPC);
14910 
14911 	helper = kmem_zalloc(sizeof (dtrace_helper_action_t), KM_SLEEP);
14912 	helper->dtha_generation = help->dthps_generation;
14913 
14914 	if ((pred = ep->dted_pred.dtpdd_predicate) != NULL) {
14915 		ASSERT(pred->dtp_difo != NULL);
14916 		dtrace_difo_hold(pred->dtp_difo);
14917 		helper->dtha_predicate = pred->dtp_difo;
14918 	}
14919 
14920 	for (act = ep->dted_action; act != NULL; act = act->dtad_next) {
14921 		if (act->dtad_kind != DTRACEACT_DIFEXPR)
14922 			goto err;
14923 
14924 		if (act->dtad_difo == NULL)
14925 			goto err;
14926 
14927 		nactions++;
14928 	}
14929 
14930 	helper->dtha_actions = kmem_zalloc(sizeof (dtrace_difo_t *) *
14931 	    (helper->dtha_nactions = nactions), KM_SLEEP);
14932 
14933 	for (act = ep->dted_action, i = 0; act != NULL; act = act->dtad_next) {
14934 		dtrace_difo_hold(act->dtad_difo);
14935 		helper->dtha_actions[i++] = act->dtad_difo;
14936 	}
14937 
14938 	if (!dtrace_helper_validate(helper))
14939 		goto err;
14940 
14941 	if (last == NULL) {
14942 		help->dthps_actions[which] = helper;
14943 	} else {
14944 		last->dtha_next = helper;
14945 	}
14946 
14947 	if (vstate->dtvs_nlocals > dtrace_helptrace_nlocals) {
14948 		dtrace_helptrace_nlocals = vstate->dtvs_nlocals;
14949 		dtrace_helptrace_next = 0;
14950 	}
14951 
14952 	return (0);
14953 err:
14954 	dtrace_helper_action_destroy(helper, vstate);
14955 	return (EINVAL);
14956 }
14957 
14958 static void
14959 dtrace_helper_provider_register(proc_t *p, dtrace_helpers_t *help,
14960     dof_helper_t *dofhp)
14961 {
14962 	ASSERT(MUTEX_NOT_HELD(&dtrace_lock));
14963 
14964 	mutex_enter(&dtrace_meta_lock);
14965 	mutex_enter(&dtrace_lock);
14966 
14967 	if (!dtrace_attached() || dtrace_meta_pid == NULL) {
14968 		/*
14969 		 * If the dtrace module is loaded but not attached, or if
14970 		 * there aren't isn't a meta provider registered to deal with
14971 		 * these provider descriptions, we need to postpone creating
14972 		 * the actual providers until later.
14973 		 */
14974 
14975 		if (help->dthps_next == NULL && help->dthps_prev == NULL &&
14976 		    dtrace_deferred_pid != help) {
14977 			help->dthps_deferred = 1;
14978 			help->dthps_pid = p->p_pid;
14979 			help->dthps_next = dtrace_deferred_pid;
14980 			help->dthps_prev = NULL;
14981 			if (dtrace_deferred_pid != NULL)
14982 				dtrace_deferred_pid->dthps_prev = help;
14983 			dtrace_deferred_pid = help;
14984 		}
14985 
14986 		mutex_exit(&dtrace_lock);
14987 
14988 	} else if (dofhp != NULL) {
14989 		/*
14990 		 * If the dtrace module is loaded and we have a particular
14991 		 * helper provider description, pass that off to the
14992 		 * meta provider.
14993 		 */
14994 
14995 		mutex_exit(&dtrace_lock);
14996 
14997 		dtrace_helper_provide(dofhp, p->p_pid);
14998 
14999 	} else {
15000 		/*
15001 		 * Otherwise, just pass all the helper provider descriptions
15002 		 * off to the meta provider.
15003 		 */
15004 
15005 		int i;
15006 		mutex_exit(&dtrace_lock);
15007 
15008 		for (i = 0; i < help->dthps_nprovs; i++) {
15009 			dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
15010 			    p->p_pid);
15011 		}
15012 	}
15013 
15014 	mutex_exit(&dtrace_meta_lock);
15015 }
15016 
15017 static int
15018 dtrace_helper_provider_add(dof_helper_t *dofhp, int gen)
15019 {
15020 	dtrace_helpers_t *help;
15021 	dtrace_helper_provider_t *hprov, **tmp_provs;
15022 	uint_t tmp_maxprovs, i;
15023 
15024 	ASSERT(MUTEX_HELD(&dtrace_lock));
15025 
15026 	help = curproc->p_dtrace_helpers;
15027 	ASSERT(help != NULL);
15028 
15029 	/*
15030 	 * If we already have dtrace_helper_providers_max helper providers,
15031 	 * we're refuse to add a new one.
15032 	 */
15033 	if (help->dthps_nprovs >= dtrace_helper_providers_max)
15034 		return (ENOSPC);
15035 
15036 	/*
15037 	 * Check to make sure this isn't a duplicate.
15038 	 */
15039 	for (i = 0; i < help->dthps_nprovs; i++) {
15040 		if (dofhp->dofhp_addr ==
15041 		    help->dthps_provs[i]->dthp_prov.dofhp_addr)
15042 			return (EALREADY);
15043 	}
15044 
15045 	hprov = kmem_zalloc(sizeof (dtrace_helper_provider_t), KM_SLEEP);
15046 	hprov->dthp_prov = *dofhp;
15047 	hprov->dthp_ref = 1;
15048 	hprov->dthp_generation = gen;
15049 
15050 	/*
15051 	 * Allocate a bigger table for helper providers if it's already full.
15052 	 */
15053 	if (help->dthps_maxprovs == help->dthps_nprovs) {
15054 		tmp_maxprovs = help->dthps_maxprovs;
15055 		tmp_provs = help->dthps_provs;
15056 
15057 		if (help->dthps_maxprovs == 0)
15058 			help->dthps_maxprovs = 2;
15059 		else
15060 			help->dthps_maxprovs *= 2;
15061 		if (help->dthps_maxprovs > dtrace_helper_providers_max)
15062 			help->dthps_maxprovs = dtrace_helper_providers_max;
15063 
15064 		ASSERT(tmp_maxprovs < help->dthps_maxprovs);
15065 
15066 		help->dthps_provs = kmem_zalloc(help->dthps_maxprovs *
15067 		    sizeof (dtrace_helper_provider_t *), KM_SLEEP);
15068 
15069 		if (tmp_provs != NULL) {
15070 			bcopy(tmp_provs, help->dthps_provs, tmp_maxprovs *
15071 			    sizeof (dtrace_helper_provider_t *));
15072 			kmem_free(tmp_provs, tmp_maxprovs *
15073 			    sizeof (dtrace_helper_provider_t *));
15074 		}
15075 	}
15076 
15077 	help->dthps_provs[help->dthps_nprovs] = hprov;
15078 	help->dthps_nprovs++;
15079 
15080 	return (0);
15081 }
15082 
15083 static void
15084 dtrace_helper_provider_destroy(dtrace_helper_provider_t *hprov)
15085 {
15086 	mutex_enter(&dtrace_lock);
15087 
15088 	if (--hprov->dthp_ref == 0) {
15089 		dof_hdr_t *dof;
15090 		mutex_exit(&dtrace_lock);
15091 		dof = (dof_hdr_t *)(uintptr_t)hprov->dthp_prov.dofhp_dof;
15092 		dtrace_dof_destroy(dof);
15093 		kmem_free(hprov, sizeof (dtrace_helper_provider_t));
15094 	} else {
15095 		mutex_exit(&dtrace_lock);
15096 	}
15097 }
15098 
15099 static int
15100 dtrace_helper_provider_validate(dof_hdr_t *dof, dof_sec_t *sec)
15101 {
15102 	uintptr_t daddr = (uintptr_t)dof;
15103 	dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
15104 	dof_provider_t *provider;
15105 	dof_probe_t *probe;
15106 	uint8_t *arg;
15107 	char *strtab, *typestr;
15108 	dof_stridx_t typeidx;
15109 	size_t typesz;
15110 	uint_t nprobes, j, k;
15111 
15112 	ASSERT(sec->dofs_type == DOF_SECT_PROVIDER);
15113 
15114 	if (sec->dofs_offset & (sizeof (uint_t) - 1)) {
15115 		dtrace_dof_error(dof, "misaligned section offset");
15116 		return (-1);
15117 	}
15118 
15119 	/*
15120 	 * The section needs to be large enough to contain the DOF provider
15121 	 * structure appropriate for the given version.
15122 	 */
15123 	if (sec->dofs_size <
15124 	    ((dof->dofh_ident[DOF_ID_VERSION] == DOF_VERSION_1) ?
15125 	    offsetof(dof_provider_t, dofpv_prenoffs) :
15126 	    sizeof (dof_provider_t))) {
15127 		dtrace_dof_error(dof, "provider section too small");
15128 		return (-1);
15129 	}
15130 
15131 	provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
15132 	str_sec = dtrace_dof_sect(dof, DOF_SECT_STRTAB, provider->dofpv_strtab);
15133 	prb_sec = dtrace_dof_sect(dof, DOF_SECT_PROBES, provider->dofpv_probes);
15134 	arg_sec = dtrace_dof_sect(dof, DOF_SECT_PRARGS, provider->dofpv_prargs);
15135 	off_sec = dtrace_dof_sect(dof, DOF_SECT_PROFFS, provider->dofpv_proffs);
15136 
15137 	if (str_sec == NULL || prb_sec == NULL ||
15138 	    arg_sec == NULL || off_sec == NULL)
15139 		return (-1);
15140 
15141 	enoff_sec = NULL;
15142 
15143 	if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
15144 	    provider->dofpv_prenoffs != DOF_SECT_NONE &&
15145 	    (enoff_sec = dtrace_dof_sect(dof, DOF_SECT_PRENOFFS,
15146 	    provider->dofpv_prenoffs)) == NULL)
15147 		return (-1);
15148 
15149 	strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
15150 
15151 	if (provider->dofpv_name >= str_sec->dofs_size ||
15152 	    strlen(strtab + provider->dofpv_name) >= DTRACE_PROVNAMELEN) {
15153 		dtrace_dof_error(dof, "invalid provider name");
15154 		return (-1);
15155 	}
15156 
15157 	if (prb_sec->dofs_entsize == 0 ||
15158 	    prb_sec->dofs_entsize > prb_sec->dofs_size) {
15159 		dtrace_dof_error(dof, "invalid entry size");
15160 		return (-1);
15161 	}
15162 
15163 	if (prb_sec->dofs_entsize & (sizeof (uintptr_t) - 1)) {
15164 		dtrace_dof_error(dof, "misaligned entry size");
15165 		return (-1);
15166 	}
15167 
15168 	if (off_sec->dofs_entsize != sizeof (uint32_t)) {
15169 		dtrace_dof_error(dof, "invalid entry size");
15170 		return (-1);
15171 	}
15172 
15173 	if (off_sec->dofs_offset & (sizeof (uint32_t) - 1)) {
15174 		dtrace_dof_error(dof, "misaligned section offset");
15175 		return (-1);
15176 	}
15177 
15178 	if (arg_sec->dofs_entsize != sizeof (uint8_t)) {
15179 		dtrace_dof_error(dof, "invalid entry size");
15180 		return (-1);
15181 	}
15182 
15183 	arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
15184 
15185 	nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
15186 
15187 	/*
15188 	 * Take a pass through the probes to check for errors.
15189 	 */
15190 	for (j = 0; j < nprobes; j++) {
15191 		probe = (dof_probe_t *)(uintptr_t)(daddr +
15192 		    prb_sec->dofs_offset + j * prb_sec->dofs_entsize);
15193 
15194 		if (probe->dofpr_func >= str_sec->dofs_size) {
15195 			dtrace_dof_error(dof, "invalid function name");
15196 			return (-1);
15197 		}
15198 
15199 		if (strlen(strtab + probe->dofpr_func) >= DTRACE_FUNCNAMELEN) {
15200 			dtrace_dof_error(dof, "function name too long");
15201 			return (-1);
15202 		}
15203 
15204 		if (probe->dofpr_name >= str_sec->dofs_size ||
15205 		    strlen(strtab + probe->dofpr_name) >= DTRACE_NAMELEN) {
15206 			dtrace_dof_error(dof, "invalid probe name");
15207 			return (-1);
15208 		}
15209 
15210 		/*
15211 		 * The offset count must not wrap the index, and the offsets
15212 		 * must also not overflow the section's data.
15213 		 */
15214 		if (probe->dofpr_offidx + probe->dofpr_noffs <
15215 		    probe->dofpr_offidx ||
15216 		    (probe->dofpr_offidx + probe->dofpr_noffs) *
15217 		    off_sec->dofs_entsize > off_sec->dofs_size) {
15218 			dtrace_dof_error(dof, "invalid probe offset");
15219 			return (-1);
15220 		}
15221 
15222 		if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1) {
15223 			/*
15224 			 * If there's no is-enabled offset section, make sure
15225 			 * there aren't any is-enabled offsets. Otherwise
15226 			 * perform the same checks as for probe offsets
15227 			 * (immediately above).
15228 			 */
15229 			if (enoff_sec == NULL) {
15230 				if (probe->dofpr_enoffidx != 0 ||
15231 				    probe->dofpr_nenoffs != 0) {
15232 					dtrace_dof_error(dof, "is-enabled "
15233 					    "offsets with null section");
15234 					return (-1);
15235 				}
15236 			} else if (probe->dofpr_enoffidx +
15237 			    probe->dofpr_nenoffs < probe->dofpr_enoffidx ||
15238 			    (probe->dofpr_enoffidx + probe->dofpr_nenoffs) *
15239 			    enoff_sec->dofs_entsize > enoff_sec->dofs_size) {
15240 				dtrace_dof_error(dof, "invalid is-enabled "
15241 				    "offset");
15242 				return (-1);
15243 			}
15244 
15245 			if (probe->dofpr_noffs + probe->dofpr_nenoffs == 0) {
15246 				dtrace_dof_error(dof, "zero probe and "
15247 				    "is-enabled offsets");
15248 				return (-1);
15249 			}
15250 		} else if (probe->dofpr_noffs == 0) {
15251 			dtrace_dof_error(dof, "zero probe offsets");
15252 			return (-1);
15253 		}
15254 
15255 		if (probe->dofpr_argidx + probe->dofpr_xargc <
15256 		    probe->dofpr_argidx ||
15257 		    (probe->dofpr_argidx + probe->dofpr_xargc) *
15258 		    arg_sec->dofs_entsize > arg_sec->dofs_size) {
15259 			dtrace_dof_error(dof, "invalid args");
15260 			return (-1);
15261 		}
15262 
15263 		typeidx = probe->dofpr_nargv;
15264 		typestr = strtab + probe->dofpr_nargv;
15265 		for (k = 0; k < probe->dofpr_nargc; k++) {
15266 			if (typeidx >= str_sec->dofs_size) {
15267 				dtrace_dof_error(dof, "bad "
15268 				    "native argument type");
15269 				return (-1);
15270 			}
15271 
15272 			typesz = strlen(typestr) + 1;
15273 			if (typesz > DTRACE_ARGTYPELEN) {
15274 				dtrace_dof_error(dof, "native "
15275 				    "argument type too long");
15276 				return (-1);
15277 			}
15278 			typeidx += typesz;
15279 			typestr += typesz;
15280 		}
15281 
15282 		typeidx = probe->dofpr_xargv;
15283 		typestr = strtab + probe->dofpr_xargv;
15284 		for (k = 0; k < probe->dofpr_xargc; k++) {
15285 			if (arg[probe->dofpr_argidx + k] > probe->dofpr_nargc) {
15286 				dtrace_dof_error(dof, "bad "
15287 				    "native argument index");
15288 				return (-1);
15289 			}
15290 
15291 			if (typeidx >= str_sec->dofs_size) {
15292 				dtrace_dof_error(dof, "bad "
15293 				    "translated argument type");
15294 				return (-1);
15295 			}
15296 
15297 			typesz = strlen(typestr) + 1;
15298 			if (typesz > DTRACE_ARGTYPELEN) {
15299 				dtrace_dof_error(dof, "translated argument "
15300 				    "type too long");
15301 				return (-1);
15302 			}
15303 
15304 			typeidx += typesz;
15305 			typestr += typesz;
15306 		}
15307 	}
15308 
15309 	return (0);
15310 }
15311 
15312 static int
15313 dtrace_helper_slurp(dof_hdr_t *dof, dof_helper_t *dhp)
15314 {
15315 	dtrace_helpers_t *help;
15316 	dtrace_vstate_t *vstate;
15317 	dtrace_enabling_t *enab = NULL;
15318 	int i, gen, rv, nhelpers = 0, nprovs = 0, destroy = 1;
15319 	uintptr_t daddr = (uintptr_t)dof;
15320 
15321 	ASSERT(MUTEX_HELD(&dtrace_lock));
15322 
15323 	if ((help = curproc->p_dtrace_helpers) == NULL)
15324 		help = dtrace_helpers_create(curproc);
15325 
15326 	vstate = &help->dthps_vstate;
15327 
15328 	if ((rv = dtrace_dof_slurp(dof, vstate, NULL, &enab,
15329 	    dhp != NULL ? dhp->dofhp_addr : 0, B_FALSE)) != 0) {
15330 		dtrace_dof_destroy(dof);
15331 		return (rv);
15332 	}
15333 
15334 	/*
15335 	 * Look for helper providers and validate their descriptions.
15336 	 */
15337 	if (dhp != NULL) {
15338 		for (i = 0; i < dof->dofh_secnum; i++) {
15339 			dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
15340 			    dof->dofh_secoff + i * dof->dofh_secsize);
15341 
15342 			if (sec->dofs_type != DOF_SECT_PROVIDER)
15343 				continue;
15344 
15345 			if (dtrace_helper_provider_validate(dof, sec) != 0) {
15346 				dtrace_enabling_destroy(enab);
15347 				dtrace_dof_destroy(dof);
15348 				return (-1);
15349 			}
15350 
15351 			nprovs++;
15352 		}
15353 	}
15354 
15355 	/*
15356 	 * Now we need to walk through the ECB descriptions in the enabling.
15357 	 */
15358 	for (i = 0; i < enab->dten_ndesc; i++) {
15359 		dtrace_ecbdesc_t *ep = enab->dten_desc[i];
15360 		dtrace_probedesc_t *desc = &ep->dted_probe;
15361 
15362 		if (strcmp(desc->dtpd_provider, "dtrace") != 0)
15363 			continue;
15364 
15365 		if (strcmp(desc->dtpd_mod, "helper") != 0)
15366 			continue;
15367 
15368 		if (strcmp(desc->dtpd_func, "ustack") != 0)
15369 			continue;
15370 
15371 		if ((rv = dtrace_helper_action_add(DTRACE_HELPER_ACTION_USTACK,
15372 		    ep)) != 0) {
15373 			/*
15374 			 * Adding this helper action failed -- we are now going
15375 			 * to rip out the entire generation and return failure.
15376 			 */
15377 			(void) dtrace_helper_destroygen(help->dthps_generation);
15378 			dtrace_enabling_destroy(enab);
15379 			dtrace_dof_destroy(dof);
15380 			return (-1);
15381 		}
15382 
15383 		nhelpers++;
15384 	}
15385 
15386 	if (nhelpers < enab->dten_ndesc)
15387 		dtrace_dof_error(dof, "unmatched helpers");
15388 
15389 	gen = help->dthps_generation++;
15390 	dtrace_enabling_destroy(enab);
15391 
15392 	if (dhp != NULL && nprovs > 0) {
15393 		/*
15394 		 * Now that this is in-kernel, we change the sense of the
15395 		 * members:  dofhp_dof denotes the in-kernel copy of the DOF
15396 		 * and dofhp_addr denotes the address at user-level.
15397 		 */
15398 		dhp->dofhp_addr = dhp->dofhp_dof;
15399 		dhp->dofhp_dof = (uint64_t)(uintptr_t)dof;
15400 
15401 		if (dtrace_helper_provider_add(dhp, gen) == 0) {
15402 			mutex_exit(&dtrace_lock);
15403 			dtrace_helper_provider_register(curproc, help, dhp);
15404 			mutex_enter(&dtrace_lock);
15405 
15406 			destroy = 0;
15407 		}
15408 	}
15409 
15410 	if (destroy)
15411 		dtrace_dof_destroy(dof);
15412 
15413 	return (gen);
15414 }
15415 
15416 static dtrace_helpers_t *
15417 dtrace_helpers_create(proc_t *p)
15418 {
15419 	dtrace_helpers_t *help;
15420 
15421 	ASSERT(MUTEX_HELD(&dtrace_lock));
15422 	ASSERT(p->p_dtrace_helpers == NULL);
15423 
15424 	help = kmem_zalloc(sizeof (dtrace_helpers_t), KM_SLEEP);
15425 	help->dthps_actions = kmem_zalloc(sizeof (dtrace_helper_action_t *) *
15426 	    DTRACE_NHELPER_ACTIONS, KM_SLEEP);
15427 
15428 	p->p_dtrace_helpers = help;
15429 	dtrace_helpers++;
15430 
15431 	return (help);
15432 }
15433 
15434 static void
15435 dtrace_helpers_destroy(void)
15436 {
15437 	dtrace_helpers_t *help;
15438 	dtrace_vstate_t *vstate;
15439 	proc_t *p = curproc;
15440 	int i;
15441 
15442 	mutex_enter(&dtrace_lock);
15443 
15444 	ASSERT(p->p_dtrace_helpers != NULL);
15445 	ASSERT(dtrace_helpers > 0);
15446 
15447 	help = p->p_dtrace_helpers;
15448 	vstate = &help->dthps_vstate;
15449 
15450 	/*
15451 	 * We're now going to lose the help from this process.
15452 	 */
15453 	p->p_dtrace_helpers = NULL;
15454 	dtrace_sync();
15455 
15456 	/*
15457 	 * Destory the helper actions.
15458 	 */
15459 	for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
15460 		dtrace_helper_action_t *h, *next;
15461 
15462 		for (h = help->dthps_actions[i]; h != NULL; h = next) {
15463 			next = h->dtha_next;
15464 			dtrace_helper_action_destroy(h, vstate);
15465 			h = next;
15466 		}
15467 	}
15468 
15469 	mutex_exit(&dtrace_lock);
15470 
15471 	/*
15472 	 * Destroy the helper providers.
15473 	 */
15474 	if (help->dthps_maxprovs > 0) {
15475 		mutex_enter(&dtrace_meta_lock);
15476 		if (dtrace_meta_pid != NULL) {
15477 			ASSERT(dtrace_deferred_pid == NULL);
15478 
15479 			for (i = 0; i < help->dthps_nprovs; i++) {
15480 				dtrace_helper_provider_remove(
15481 				    &help->dthps_provs[i]->dthp_prov, p->p_pid);
15482 			}
15483 		} else {
15484 			mutex_enter(&dtrace_lock);
15485 			ASSERT(help->dthps_deferred == 0 ||
15486 			    help->dthps_next != NULL ||
15487 			    help->dthps_prev != NULL ||
15488 			    help == dtrace_deferred_pid);
15489 
15490 			/*
15491 			 * Remove the helper from the deferred list.
15492 			 */
15493 			if (help->dthps_next != NULL)
15494 				help->dthps_next->dthps_prev = help->dthps_prev;
15495 			if (help->dthps_prev != NULL)
15496 				help->dthps_prev->dthps_next = help->dthps_next;
15497 			if (dtrace_deferred_pid == help) {
15498 				dtrace_deferred_pid = help->dthps_next;
15499 				ASSERT(help->dthps_prev == NULL);
15500 			}
15501 
15502 			mutex_exit(&dtrace_lock);
15503 		}
15504 
15505 		mutex_exit(&dtrace_meta_lock);
15506 
15507 		for (i = 0; i < help->dthps_nprovs; i++) {
15508 			dtrace_helper_provider_destroy(help->dthps_provs[i]);
15509 		}
15510 
15511 		kmem_free(help->dthps_provs, help->dthps_maxprovs *
15512 		    sizeof (dtrace_helper_provider_t *));
15513 	}
15514 
15515 	mutex_enter(&dtrace_lock);
15516 
15517 	dtrace_vstate_fini(&help->dthps_vstate);
15518 	kmem_free(help->dthps_actions,
15519 	    sizeof (dtrace_helper_action_t *) * DTRACE_NHELPER_ACTIONS);
15520 	kmem_free(help, sizeof (dtrace_helpers_t));
15521 
15522 	--dtrace_helpers;
15523 	mutex_exit(&dtrace_lock);
15524 }
15525 
15526 static void
15527 dtrace_helpers_duplicate(proc_t *from, proc_t *to)
15528 {
15529 	dtrace_helpers_t *help, *newhelp;
15530 	dtrace_helper_action_t *helper, *new, *last;
15531 	dtrace_difo_t *dp;
15532 	dtrace_vstate_t *vstate;
15533 	int i, j, sz, hasprovs = 0;
15534 
15535 	mutex_enter(&dtrace_lock);
15536 	ASSERT(from->p_dtrace_helpers != NULL);
15537 	ASSERT(dtrace_helpers > 0);
15538 
15539 	help = from->p_dtrace_helpers;
15540 	newhelp = dtrace_helpers_create(to);
15541 	ASSERT(to->p_dtrace_helpers != NULL);
15542 
15543 	newhelp->dthps_generation = help->dthps_generation;
15544 	vstate = &newhelp->dthps_vstate;
15545 
15546 	/*
15547 	 * Duplicate the helper actions.
15548 	 */
15549 	for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
15550 		if ((helper = help->dthps_actions[i]) == NULL)
15551 			continue;
15552 
15553 		for (last = NULL; helper != NULL; helper = helper->dtha_next) {
15554 			new = kmem_zalloc(sizeof (dtrace_helper_action_t),
15555 			    KM_SLEEP);
15556 			new->dtha_generation = helper->dtha_generation;
15557 
15558 			if ((dp = helper->dtha_predicate) != NULL) {
15559 				dp = dtrace_difo_duplicate(dp, vstate);
15560 				new->dtha_predicate = dp;
15561 			}
15562 
15563 			new->dtha_nactions = helper->dtha_nactions;
15564 			sz = sizeof (dtrace_difo_t *) * new->dtha_nactions;
15565 			new->dtha_actions = kmem_alloc(sz, KM_SLEEP);
15566 
15567 			for (j = 0; j < new->dtha_nactions; j++) {
15568 				dtrace_difo_t *dp = helper->dtha_actions[j];
15569 
15570 				ASSERT(dp != NULL);
15571 				dp = dtrace_difo_duplicate(dp, vstate);
15572 				new->dtha_actions[j] = dp;
15573 			}
15574 
15575 			if (last != NULL) {
15576 				last->dtha_next = new;
15577 			} else {
15578 				newhelp->dthps_actions[i] = new;
15579 			}
15580 
15581 			last = new;
15582 		}
15583 	}
15584 
15585 	/*
15586 	 * Duplicate the helper providers and register them with the
15587 	 * DTrace framework.
15588 	 */
15589 	if (help->dthps_nprovs > 0) {
15590 		newhelp->dthps_nprovs = help->dthps_nprovs;
15591 		newhelp->dthps_maxprovs = help->dthps_nprovs;
15592 		newhelp->dthps_provs = kmem_alloc(newhelp->dthps_nprovs *
15593 		    sizeof (dtrace_helper_provider_t *), KM_SLEEP);
15594 		for (i = 0; i < newhelp->dthps_nprovs; i++) {
15595 			newhelp->dthps_provs[i] = help->dthps_provs[i];
15596 			newhelp->dthps_provs[i]->dthp_ref++;
15597 		}
15598 
15599 		hasprovs = 1;
15600 	}
15601 
15602 	mutex_exit(&dtrace_lock);
15603 
15604 	if (hasprovs)
15605 		dtrace_helper_provider_register(to, newhelp, NULL);
15606 }
15607 
15608 /*
15609  * DTrace Hook Functions
15610  */
15611 static void
15612 dtrace_module_loaded(struct modctl *ctl)
15613 {
15614 	dtrace_provider_t *prv;
15615 
15616 	mutex_enter(&dtrace_provider_lock);
15617 	mutex_enter(&mod_lock);
15618 
15619 	ASSERT(ctl->mod_busy);
15620 
15621 	/*
15622 	 * We're going to call each providers per-module provide operation
15623 	 * specifying only this module.
15624 	 */
15625 	for (prv = dtrace_provider; prv != NULL; prv = prv->dtpv_next)
15626 		prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
15627 
15628 	mutex_exit(&mod_lock);
15629 	mutex_exit(&dtrace_provider_lock);
15630 
15631 	/*
15632 	 * If we have any retained enablings, we need to match against them.
15633 	 * Enabling probes requires that cpu_lock be held, and we cannot hold
15634 	 * cpu_lock here -- it is legal for cpu_lock to be held when loading a
15635 	 * module.  (In particular, this happens when loading scheduling
15636 	 * classes.)  So if we have any retained enablings, we need to dispatch
15637 	 * our task queue to do the match for us.
15638 	 */
15639 	mutex_enter(&dtrace_lock);
15640 
15641 	if (dtrace_retained == NULL) {
15642 		mutex_exit(&dtrace_lock);
15643 		return;
15644 	}
15645 
15646 	(void) taskq_dispatch(dtrace_taskq,
15647 	    (task_func_t *)dtrace_enabling_matchall, NULL, TQ_SLEEP);
15648 
15649 	mutex_exit(&dtrace_lock);
15650 
15651 	/*
15652 	 * And now, for a little heuristic sleaze:  in general, we want to
15653 	 * match modules as soon as they load.  However, we cannot guarantee
15654 	 * this, because it would lead us to the lock ordering violation
15655 	 * outlined above.  The common case, of course, is that cpu_lock is
15656 	 * _not_ held -- so we delay here for a clock tick, hoping that that's
15657 	 * long enough for the task queue to do its work.  If it's not, it's
15658 	 * not a serious problem -- it just means that the module that we
15659 	 * just loaded may not be immediately instrumentable.
15660 	 */
15661 	delay(1);
15662 }
15663 
15664 static void
15665 dtrace_module_unloaded(struct modctl *ctl)
15666 {
15667 	dtrace_probe_t template, *probe, *first, *next;
15668 	dtrace_provider_t *prov;
15669 
15670 	template.dtpr_mod = ctl->mod_modname;
15671 
15672 	mutex_enter(&dtrace_provider_lock);
15673 	mutex_enter(&mod_lock);
15674 	mutex_enter(&dtrace_lock);
15675 
15676 	if (dtrace_bymod == NULL) {
15677 		/*
15678 		 * The DTrace module is loaded (obviously) but not attached;
15679 		 * we don't have any work to do.
15680 		 */
15681 		mutex_exit(&dtrace_provider_lock);
15682 		mutex_exit(&mod_lock);
15683 		mutex_exit(&dtrace_lock);
15684 		return;
15685 	}
15686 
15687 	for (probe = first = dtrace_hash_lookup(dtrace_bymod, &template);
15688 	    probe != NULL; probe = probe->dtpr_nextmod) {
15689 		if (probe->dtpr_ecb != NULL) {
15690 			mutex_exit(&dtrace_provider_lock);
15691 			mutex_exit(&mod_lock);
15692 			mutex_exit(&dtrace_lock);
15693 
15694 			/*
15695 			 * This shouldn't _actually_ be possible -- we're
15696 			 * unloading a module that has an enabled probe in it.
15697 			 * (It's normally up to the provider to make sure that
15698 			 * this can't happen.)  However, because dtps_enable()
15699 			 * doesn't have a failure mode, there can be an
15700 			 * enable/unload race.  Upshot:  we don't want to
15701 			 * assert, but we're not going to disable the
15702 			 * probe, either.
15703 			 */
15704 			if (dtrace_err_verbose) {
15705 				cmn_err(CE_WARN, "unloaded module '%s' had "
15706 				    "enabled probes", ctl->mod_modname);
15707 			}
15708 
15709 			return;
15710 		}
15711 	}
15712 
15713 	probe = first;
15714 
15715 	for (first = NULL; probe != NULL; probe = next) {
15716 		ASSERT(dtrace_probes[probe->dtpr_id - 1] == probe);
15717 
15718 		dtrace_probes[probe->dtpr_id - 1] = NULL;
15719 
15720 		next = probe->dtpr_nextmod;
15721 		dtrace_hash_remove(dtrace_bymod, probe);
15722 		dtrace_hash_remove(dtrace_byfunc, probe);
15723 		dtrace_hash_remove(dtrace_byname, probe);
15724 
15725 		if (first == NULL) {
15726 			first = probe;
15727 			probe->dtpr_nextmod = NULL;
15728 		} else {
15729 			probe->dtpr_nextmod = first;
15730 			first = probe;
15731 		}
15732 	}
15733 
15734 	/*
15735 	 * We've removed all of the module's probes from the hash chains and
15736 	 * from the probe array.  Now issue a dtrace_sync() to be sure that
15737 	 * everyone has cleared out from any probe array processing.
15738 	 */
15739 	dtrace_sync();
15740 
15741 	for (probe = first; probe != NULL; probe = first) {
15742 		first = probe->dtpr_nextmod;
15743 		prov = probe->dtpr_provider;
15744 		prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, probe->dtpr_id,
15745 		    probe->dtpr_arg);
15746 		kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
15747 		kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
15748 		kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
15749 		vmem_free(dtrace_arena, (void *)(uintptr_t)probe->dtpr_id, 1);
15750 		kmem_free(probe, sizeof (dtrace_probe_t));
15751 	}
15752 
15753 	mutex_exit(&dtrace_lock);
15754 	mutex_exit(&mod_lock);
15755 	mutex_exit(&dtrace_provider_lock);
15756 }
15757 
15758 void
15759 dtrace_suspend(void)
15760 {
15761 	dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_suspend));
15762 }
15763 
15764 void
15765 dtrace_resume(void)
15766 {
15767 	dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_resume));
15768 }
15769 
15770 static int
15771 dtrace_cpu_setup(cpu_setup_t what, processorid_t cpu)
15772 {
15773 	ASSERT(MUTEX_HELD(&cpu_lock));
15774 	mutex_enter(&dtrace_lock);
15775 
15776 	switch (what) {
15777 	case CPU_CONFIG: {
15778 		dtrace_state_t *state;
15779 		dtrace_optval_t *opt, rs, c;
15780 
15781 		/*
15782 		 * For now, we only allocate a new buffer for anonymous state.
15783 		 */
15784 		if ((state = dtrace_anon.dta_state) == NULL)
15785 			break;
15786 
15787 		if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
15788 			break;
15789 
15790 		opt = state->dts_options;
15791 		c = opt[DTRACEOPT_CPU];
15792 
15793 		if (c != DTRACE_CPUALL && c != DTRACEOPT_UNSET && c != cpu)
15794 			break;
15795 
15796 		/*
15797 		 * Regardless of what the actual policy is, we're going to
15798 		 * temporarily set our resize policy to be manual.  We're
15799 		 * also going to temporarily set our CPU option to denote
15800 		 * the newly configured CPU.
15801 		 */
15802 		rs = opt[DTRACEOPT_BUFRESIZE];
15803 		opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_MANUAL;
15804 		opt[DTRACEOPT_CPU] = (dtrace_optval_t)cpu;
15805 
15806 		(void) dtrace_state_buffers(state);
15807 
15808 		opt[DTRACEOPT_BUFRESIZE] = rs;
15809 		opt[DTRACEOPT_CPU] = c;
15810 
15811 		break;
15812 	}
15813 
15814 	case CPU_UNCONFIG:
15815 		/*
15816 		 * We don't free the buffer in the CPU_UNCONFIG case.  (The
15817 		 * buffer will be freed when the consumer exits.)
15818 		 */
15819 		break;
15820 
15821 	default:
15822 		break;
15823 	}
15824 
15825 	mutex_exit(&dtrace_lock);
15826 	return (0);
15827 }
15828 
15829 static void
15830 dtrace_cpu_setup_initial(processorid_t cpu)
15831 {
15832 	(void) dtrace_cpu_setup(CPU_CONFIG, cpu);
15833 }
15834 
15835 static void
15836 dtrace_toxrange_add(uintptr_t base, uintptr_t limit)
15837 {
15838 	if (dtrace_toxranges >= dtrace_toxranges_max) {
15839 		int osize, nsize;
15840 		dtrace_toxrange_t *range;
15841 
15842 		osize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
15843 
15844 		if (osize == 0) {
15845 			ASSERT(dtrace_toxrange == NULL);
15846 			ASSERT(dtrace_toxranges_max == 0);
15847 			dtrace_toxranges_max = 1;
15848 		} else {
15849 			dtrace_toxranges_max <<= 1;
15850 		}
15851 
15852 		nsize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
15853 		range = kmem_zalloc(nsize, KM_SLEEP);
15854 
15855 		if (dtrace_toxrange != NULL) {
15856 			ASSERT(osize != 0);
15857 			bcopy(dtrace_toxrange, range, osize);
15858 			kmem_free(dtrace_toxrange, osize);
15859 		}
15860 
15861 		dtrace_toxrange = range;
15862 	}
15863 
15864 	ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_base == NULL);
15865 	ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_limit == NULL);
15866 
15867 	dtrace_toxrange[dtrace_toxranges].dtt_base = base;
15868 	dtrace_toxrange[dtrace_toxranges].dtt_limit = limit;
15869 	dtrace_toxranges++;
15870 }
15871 
15872 static void
15873 dtrace_getf_barrier()
15874 {
15875 	/*
15876 	 * When we have unprivileged (that is, non-DTRACE_CRV_KERNEL) enablings
15877 	 * that contain calls to getf(), this routine will be called on every
15878 	 * closef() before either the underlying vnode is released or the
15879 	 * file_t itself is freed.  By the time we are here, it is essential
15880 	 * that the file_t can no longer be accessed from a call to getf()
15881 	 * in probe context -- that assures that a dtrace_sync() can be used
15882 	 * to clear out any enablings referring to the old structures.
15883 	 */
15884 	if (curthread->t_procp->p_zone->zone_dtrace_getf != 0 ||
15885 	    kcred->cr_zone->zone_dtrace_getf != 0)
15886 		dtrace_sync();
15887 }
15888 
15889 /*
15890  * DTrace Driver Cookbook Functions
15891  */
15892 /*ARGSUSED*/
15893 static int
15894 dtrace_attach(dev_info_t *devi, ddi_attach_cmd_t cmd)
15895 {
15896 	dtrace_provider_id_t id;
15897 	dtrace_state_t *state = NULL;
15898 	dtrace_enabling_t *enab;
15899 
15900 	mutex_enter(&cpu_lock);
15901 	mutex_enter(&dtrace_provider_lock);
15902 	mutex_enter(&dtrace_lock);
15903 
15904 	if (ddi_soft_state_init(&dtrace_softstate,
15905 	    sizeof (dtrace_state_t), 0) != 0) {
15906 		cmn_err(CE_NOTE, "/dev/dtrace failed to initialize soft state");
15907 		mutex_exit(&cpu_lock);
15908 		mutex_exit(&dtrace_provider_lock);
15909 		mutex_exit(&dtrace_lock);
15910 		return (DDI_FAILURE);
15911 	}
15912 
15913 	if (ddi_create_minor_node(devi, DTRACEMNR_DTRACE, S_IFCHR,
15914 	    DTRACEMNRN_DTRACE, DDI_PSEUDO, NULL) == DDI_FAILURE ||
15915 	    ddi_create_minor_node(devi, DTRACEMNR_HELPER, S_IFCHR,
15916 	    DTRACEMNRN_HELPER, DDI_PSEUDO, NULL) == DDI_FAILURE) {
15917 		cmn_err(CE_NOTE, "/dev/dtrace couldn't create minor nodes");
15918 		ddi_remove_minor_node(devi, NULL);
15919 		ddi_soft_state_fini(&dtrace_softstate);
15920 		mutex_exit(&cpu_lock);
15921 		mutex_exit(&dtrace_provider_lock);
15922 		mutex_exit(&dtrace_lock);
15923 		return (DDI_FAILURE);
15924 	}
15925 
15926 	ddi_report_dev(devi);
15927 	dtrace_devi = devi;
15928 
15929 	dtrace_modload = dtrace_module_loaded;
15930 	dtrace_modunload = dtrace_module_unloaded;
15931 	dtrace_cpu_init = dtrace_cpu_setup_initial;
15932 	dtrace_helpers_cleanup = dtrace_helpers_destroy;
15933 	dtrace_helpers_fork = dtrace_helpers_duplicate;
15934 	dtrace_cpustart_init = dtrace_suspend;
15935 	dtrace_cpustart_fini = dtrace_resume;
15936 	dtrace_debugger_init = dtrace_suspend;
15937 	dtrace_debugger_fini = dtrace_resume;
15938 
15939 	register_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
15940 
15941 	ASSERT(MUTEX_HELD(&cpu_lock));
15942 
15943 	dtrace_arena = vmem_create("dtrace", (void *)1, UINT32_MAX, 1,
15944 	    NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
15945 	dtrace_minor = vmem_create("dtrace_minor", (void *)DTRACEMNRN_CLONE,
15946 	    UINT32_MAX - DTRACEMNRN_CLONE, 1, NULL, NULL, NULL, 0,
15947 	    VM_SLEEP | VMC_IDENTIFIER);
15948 	dtrace_taskq = taskq_create("dtrace_taskq", 1, maxclsyspri,
15949 	    1, INT_MAX, 0);
15950 
15951 	dtrace_state_cache = kmem_cache_create("dtrace_state_cache",
15952 	    sizeof (dtrace_dstate_percpu_t) * NCPU, DTRACE_STATE_ALIGN,
15953 	    NULL, NULL, NULL, NULL, NULL, 0);
15954 
15955 	ASSERT(MUTEX_HELD(&cpu_lock));
15956 	dtrace_bymod = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_mod),
15957 	    offsetof(dtrace_probe_t, dtpr_nextmod),
15958 	    offsetof(dtrace_probe_t, dtpr_prevmod));
15959 
15960 	dtrace_byfunc = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_func),
15961 	    offsetof(dtrace_probe_t, dtpr_nextfunc),
15962 	    offsetof(dtrace_probe_t, dtpr_prevfunc));
15963 
15964 	dtrace_byname = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_name),
15965 	    offsetof(dtrace_probe_t, dtpr_nextname),
15966 	    offsetof(dtrace_probe_t, dtpr_prevname));
15967 
15968 	if (dtrace_retain_max < 1) {
15969 		cmn_err(CE_WARN, "illegal value (%lu) for dtrace_retain_max; "
15970 		    "setting to 1", dtrace_retain_max);
15971 		dtrace_retain_max = 1;
15972 	}
15973 
15974 	/*
15975 	 * Now discover our toxic ranges.
15976 	 */
15977 	dtrace_toxic_ranges(dtrace_toxrange_add);
15978 
15979 	/*
15980 	 * Before we register ourselves as a provider to our own framework,
15981 	 * we would like to assert that dtrace_provider is NULL -- but that's
15982 	 * not true if we were loaded as a dependency of a DTrace provider.
15983 	 * Once we've registered, we can assert that dtrace_provider is our
15984 	 * pseudo provider.
15985 	 */
15986 	(void) dtrace_register("dtrace", &dtrace_provider_attr,
15987 	    DTRACE_PRIV_NONE, 0, &dtrace_provider_ops, NULL, &id);
15988 
15989 	ASSERT(dtrace_provider != NULL);
15990 	ASSERT((dtrace_provider_id_t)dtrace_provider == id);
15991 
15992 	dtrace_probeid_begin = dtrace_probe_create((dtrace_provider_id_t)
15993 	    dtrace_provider, NULL, NULL, "BEGIN", 0, NULL);
15994 	dtrace_probeid_end = dtrace_probe_create((dtrace_provider_id_t)
15995 	    dtrace_provider, NULL, NULL, "END", 0, NULL);
15996 	dtrace_probeid_error = dtrace_probe_create((dtrace_provider_id_t)
15997 	    dtrace_provider, NULL, NULL, "ERROR", 1, NULL);
15998 
15999 	dtrace_anon_property();
16000 	mutex_exit(&cpu_lock);
16001 
16002 	/*
16003 	 * If there are already providers, we must ask them to provide their
16004 	 * probes, and then match any anonymous enabling against them.  Note
16005 	 * that there should be no other retained enablings at this time:
16006 	 * the only retained enablings at this time should be the anonymous
16007 	 * enabling.
16008 	 */
16009 	if (dtrace_anon.dta_enabling != NULL) {
16010 		ASSERT(dtrace_retained == dtrace_anon.dta_enabling);
16011 
16012 		dtrace_enabling_provide(NULL);
16013 		state = dtrace_anon.dta_state;
16014 
16015 		/*
16016 		 * We couldn't hold cpu_lock across the above call to
16017 		 * dtrace_enabling_provide(), but we must hold it to actually
16018 		 * enable the probes.  We have to drop all of our locks, pick
16019 		 * up cpu_lock, and regain our locks before matching the
16020 		 * retained anonymous enabling.
16021 		 */
16022 		mutex_exit(&dtrace_lock);
16023 		mutex_exit(&dtrace_provider_lock);
16024 
16025 		mutex_enter(&cpu_lock);
16026 		mutex_enter(&dtrace_provider_lock);
16027 		mutex_enter(&dtrace_lock);
16028 
16029 		if ((enab = dtrace_anon.dta_enabling) != NULL)
16030 			(void) dtrace_enabling_match(enab, NULL);
16031 
16032 		mutex_exit(&cpu_lock);
16033 	}
16034 
16035 	mutex_exit(&dtrace_lock);
16036 	mutex_exit(&dtrace_provider_lock);
16037 
16038 	if (state != NULL) {
16039 		/*
16040 		 * If we created any anonymous state, set it going now.
16041 		 */
16042 		(void) dtrace_state_go(state, &dtrace_anon.dta_beganon);
16043 	}
16044 
16045 	return (DDI_SUCCESS);
16046 }
16047 
16048 /*ARGSUSED*/
16049 static int
16050 dtrace_open(dev_t *devp, int flag, int otyp, cred_t *cred_p)
16051 {
16052 	dtrace_state_t *state;
16053 	uint32_t priv;
16054 	uid_t uid;
16055 	zoneid_t zoneid;
16056 
16057 	if (getminor(*devp) == DTRACEMNRN_HELPER)
16058 		return (0);
16059 
16060 	/*
16061 	 * If this wasn't an open with the "helper" minor, then it must be
16062 	 * the "dtrace" minor.
16063 	 */
16064 	if (getminor(*devp) != DTRACEMNRN_DTRACE)
16065 		return (ENXIO);
16066 
16067 	/*
16068 	 * If no DTRACE_PRIV_* bits are set in the credential, then the
16069 	 * caller lacks sufficient permission to do anything with DTrace.
16070 	 */
16071 	dtrace_cred2priv(cred_p, &priv, &uid, &zoneid);
16072 	if (priv == DTRACE_PRIV_NONE)
16073 		return (EACCES);
16074 
16075 	/*
16076 	 * Ask all providers to provide all their probes.
16077 	 */
16078 	mutex_enter(&dtrace_provider_lock);
16079 	dtrace_probe_provide(NULL, NULL);
16080 	mutex_exit(&dtrace_provider_lock);
16081 
16082 	mutex_enter(&cpu_lock);
16083 	mutex_enter(&dtrace_lock);
16084 	dtrace_opens++;
16085 	dtrace_membar_producer();
16086 
16087 	/*
16088 	 * If the kernel debugger is active (that is, if the kernel debugger
16089 	 * modified text in some way), we won't allow the open.
16090 	 */
16091 	if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
16092 		dtrace_opens--;
16093 		mutex_exit(&cpu_lock);
16094 		mutex_exit(&dtrace_lock);
16095 		return (EBUSY);
16096 	}
16097 
16098 	if (dtrace_helptrace_enable && dtrace_helptrace_buffer == NULL) {
16099 		/*
16100 		 * If DTrace helper tracing is enabled, we need to allocate the
16101 		 * trace buffer and initialize the values.
16102 		 */
16103 		dtrace_helptrace_buffer =
16104 		    kmem_zalloc(dtrace_helptrace_bufsize, KM_SLEEP);
16105 		dtrace_helptrace_next = 0;
16106 		dtrace_helptrace_wrapped = 0;
16107 		dtrace_helptrace_enable = 0;
16108 	}
16109 
16110 	state = dtrace_state_create(devp, cred_p);
16111 	mutex_exit(&cpu_lock);
16112 
16113 	if (state == NULL) {
16114 		if (--dtrace_opens == 0 && dtrace_anon.dta_enabling == NULL)
16115 			(void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
16116 		mutex_exit(&dtrace_lock);
16117 		return (EAGAIN);
16118 	}
16119 
16120 	mutex_exit(&dtrace_lock);
16121 
16122 	return (0);
16123 }
16124 
16125 /*ARGSUSED*/
16126 static int
16127 dtrace_close(dev_t dev, int flag, int otyp, cred_t *cred_p)
16128 {
16129 	minor_t minor = getminor(dev);
16130 	dtrace_state_t *state;
16131 	dtrace_helptrace_t *buf = NULL;
16132 
16133 	if (minor == DTRACEMNRN_HELPER)
16134 		return (0);
16135 
16136 	state = ddi_get_soft_state(dtrace_softstate, minor);
16137 
16138 	mutex_enter(&cpu_lock);
16139 	mutex_enter(&dtrace_lock);
16140 
16141 	if (state->dts_anon) {
16142 		/*
16143 		 * There is anonymous state. Destroy that first.
16144 		 */
16145 		ASSERT(dtrace_anon.dta_state == NULL);
16146 		dtrace_state_destroy(state->dts_anon);
16147 	}
16148 
16149 	if (dtrace_helptrace_disable) {
16150 		/*
16151 		 * If we have been told to disable helper tracing, set the
16152 		 * buffer to NULL before calling into dtrace_state_destroy();
16153 		 * we take advantage of its dtrace_sync() to know that no
16154 		 * CPU is in probe context with enabled helper tracing
16155 		 * after it returns.
16156 		 */
16157 		buf = dtrace_helptrace_buffer;
16158 		dtrace_helptrace_buffer = NULL;
16159 	}
16160 
16161 	dtrace_state_destroy(state);
16162 	ASSERT(dtrace_opens > 0);
16163 
16164 	/*
16165 	 * Only relinquish control of the kernel debugger interface when there
16166 	 * are no consumers and no anonymous enablings.
16167 	 */
16168 	if (--dtrace_opens == 0 && dtrace_anon.dta_enabling == NULL)
16169 		(void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
16170 
16171 	if (buf != NULL) {
16172 		kmem_free(buf, dtrace_helptrace_bufsize);
16173 		dtrace_helptrace_disable = 0;
16174 	}
16175 
16176 	mutex_exit(&dtrace_lock);
16177 	mutex_exit(&cpu_lock);
16178 
16179 	return (0);
16180 }
16181 
16182 /*ARGSUSED*/
16183 static int
16184 dtrace_ioctl_helper(int cmd, intptr_t arg, int *rv)
16185 {
16186 	int rval;
16187 	dof_helper_t help, *dhp = NULL;
16188 
16189 	switch (cmd) {
16190 	case DTRACEHIOC_ADDDOF:
16191 		if (copyin((void *)arg, &help, sizeof (help)) != 0) {
16192 			dtrace_dof_error(NULL, "failed to copyin DOF helper");
16193 			return (EFAULT);
16194 		}
16195 
16196 		dhp = &help;
16197 		arg = (intptr_t)help.dofhp_dof;
16198 		/*FALLTHROUGH*/
16199 
16200 	case DTRACEHIOC_ADD: {
16201 		dof_hdr_t *dof = dtrace_dof_copyin(arg, &rval);
16202 
16203 		if (dof == NULL)
16204 			return (rval);
16205 
16206 		mutex_enter(&dtrace_lock);
16207 
16208 		/*
16209 		 * dtrace_helper_slurp() takes responsibility for the dof --
16210 		 * it may free it now or it may save it and free it later.
16211 		 */
16212 		if ((rval = dtrace_helper_slurp(dof, dhp)) != -1) {
16213 			*rv = rval;
16214 			rval = 0;
16215 		} else {
16216 			rval = EINVAL;
16217 		}
16218 
16219 		mutex_exit(&dtrace_lock);
16220 		return (rval);
16221 	}
16222 
16223 	case DTRACEHIOC_REMOVE: {
16224 		mutex_enter(&dtrace_lock);
16225 		rval = dtrace_helper_destroygen(arg);
16226 		mutex_exit(&dtrace_lock);
16227 
16228 		return (rval);
16229 	}
16230 
16231 	default:
16232 		break;
16233 	}
16234 
16235 	return (ENOTTY);
16236 }
16237 
16238 /*ARGSUSED*/
16239 static int
16240 dtrace_ioctl(dev_t dev, int cmd, intptr_t arg, int md, cred_t *cr, int *rv)
16241 {
16242 	minor_t minor = getminor(dev);
16243 	dtrace_state_t *state;
16244 	int rval;
16245 
16246 	if (minor == DTRACEMNRN_HELPER)
16247 		return (dtrace_ioctl_helper(cmd, arg, rv));
16248 
16249 	state = ddi_get_soft_state(dtrace_softstate, minor);
16250 
16251 	if (state->dts_anon) {
16252 		ASSERT(dtrace_anon.dta_state == NULL);
16253 		state = state->dts_anon;
16254 	}
16255 
16256 	switch (cmd) {
16257 	case DTRACEIOC_PROVIDER: {
16258 		dtrace_providerdesc_t pvd;
16259 		dtrace_provider_t *pvp;
16260 
16261 		if (copyin((void *)arg, &pvd, sizeof (pvd)) != 0)
16262 			return (EFAULT);
16263 
16264 		pvd.dtvd_name[DTRACE_PROVNAMELEN - 1] = '\0';
16265 		mutex_enter(&dtrace_provider_lock);
16266 
16267 		for (pvp = dtrace_provider; pvp != NULL; pvp = pvp->dtpv_next) {
16268 			if (strcmp(pvp->dtpv_name, pvd.dtvd_name) == 0)
16269 				break;
16270 		}
16271 
16272 		mutex_exit(&dtrace_provider_lock);
16273 
16274 		if (pvp == NULL)
16275 			return (ESRCH);
16276 
16277 		bcopy(&pvp->dtpv_priv, &pvd.dtvd_priv, sizeof (dtrace_ppriv_t));
16278 		bcopy(&pvp->dtpv_attr, &pvd.dtvd_attr, sizeof (dtrace_pattr_t));
16279 		if (copyout(&pvd, (void *)arg, sizeof (pvd)) != 0)
16280 			return (EFAULT);
16281 
16282 		return (0);
16283 	}
16284 
16285 	case DTRACEIOC_EPROBE: {
16286 		dtrace_eprobedesc_t epdesc;
16287 		dtrace_ecb_t *ecb;
16288 		dtrace_action_t *act;
16289 		void *buf;
16290 		size_t size;
16291 		uintptr_t dest;
16292 		int nrecs;
16293 
16294 		if (copyin((void *)arg, &epdesc, sizeof (epdesc)) != 0)
16295 			return (EFAULT);
16296 
16297 		mutex_enter(&dtrace_lock);
16298 
16299 		if ((ecb = dtrace_epid2ecb(state, epdesc.dtepd_epid)) == NULL) {
16300 			mutex_exit(&dtrace_lock);
16301 			return (EINVAL);
16302 		}
16303 
16304 		if (ecb->dte_probe == NULL) {
16305 			mutex_exit(&dtrace_lock);
16306 			return (EINVAL);
16307 		}
16308 
16309 		epdesc.dtepd_probeid = ecb->dte_probe->dtpr_id;
16310 		epdesc.dtepd_uarg = ecb->dte_uarg;
16311 		epdesc.dtepd_size = ecb->dte_size;
16312 
16313 		nrecs = epdesc.dtepd_nrecs;
16314 		epdesc.dtepd_nrecs = 0;
16315 		for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
16316 			if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
16317 				continue;
16318 
16319 			epdesc.dtepd_nrecs++;
16320 		}
16321 
16322 		/*
16323 		 * Now that we have the size, we need to allocate a temporary
16324 		 * buffer in which to store the complete description.  We need
16325 		 * the temporary buffer to be able to drop dtrace_lock()
16326 		 * across the copyout(), below.
16327 		 */
16328 		size = sizeof (dtrace_eprobedesc_t) +
16329 		    (epdesc.dtepd_nrecs * sizeof (dtrace_recdesc_t));
16330 
16331 		buf = kmem_alloc(size, KM_SLEEP);
16332 		dest = (uintptr_t)buf;
16333 
16334 		bcopy(&epdesc, (void *)dest, sizeof (epdesc));
16335 		dest += offsetof(dtrace_eprobedesc_t, dtepd_rec[0]);
16336 
16337 		for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
16338 			if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
16339 				continue;
16340 
16341 			if (nrecs-- == 0)
16342 				break;
16343 
16344 			bcopy(&act->dta_rec, (void *)dest,
16345 			    sizeof (dtrace_recdesc_t));
16346 			dest += sizeof (dtrace_recdesc_t);
16347 		}
16348 
16349 		mutex_exit(&dtrace_lock);
16350 
16351 		if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
16352 			kmem_free(buf, size);
16353 			return (EFAULT);
16354 		}
16355 
16356 		kmem_free(buf, size);
16357 		return (0);
16358 	}
16359 
16360 	case DTRACEIOC_AGGDESC: {
16361 		dtrace_aggdesc_t aggdesc;
16362 		dtrace_action_t *act;
16363 		dtrace_aggregation_t *agg;
16364 		int nrecs;
16365 		uint32_t offs;
16366 		dtrace_recdesc_t *lrec;
16367 		void *buf;
16368 		size_t size;
16369 		uintptr_t dest;
16370 
16371 		if (copyin((void *)arg, &aggdesc, sizeof (aggdesc)) != 0)
16372 			return (EFAULT);
16373 
16374 		mutex_enter(&dtrace_lock);
16375 
16376 		if ((agg = dtrace_aggid2agg(state, aggdesc.dtagd_id)) == NULL) {
16377 			mutex_exit(&dtrace_lock);
16378 			return (EINVAL);
16379 		}
16380 
16381 		aggdesc.dtagd_epid = agg->dtag_ecb->dte_epid;
16382 
16383 		nrecs = aggdesc.dtagd_nrecs;
16384 		aggdesc.dtagd_nrecs = 0;
16385 
16386 		offs = agg->dtag_base;
16387 		lrec = &agg->dtag_action.dta_rec;
16388 		aggdesc.dtagd_size = lrec->dtrd_offset + lrec->dtrd_size - offs;
16389 
16390 		for (act = agg->dtag_first; ; act = act->dta_next) {
16391 			ASSERT(act->dta_intuple ||
16392 			    DTRACEACT_ISAGG(act->dta_kind));
16393 
16394 			/*
16395 			 * If this action has a record size of zero, it
16396 			 * denotes an argument to the aggregating action.
16397 			 * Because the presence of this record doesn't (or
16398 			 * shouldn't) affect the way the data is interpreted,
16399 			 * we don't copy it out to save user-level the
16400 			 * confusion of dealing with a zero-length record.
16401 			 */
16402 			if (act->dta_rec.dtrd_size == 0) {
16403 				ASSERT(agg->dtag_hasarg);
16404 				continue;
16405 			}
16406 
16407 			aggdesc.dtagd_nrecs++;
16408 
16409 			if (act == &agg->dtag_action)
16410 				break;
16411 		}
16412 
16413 		/*
16414 		 * Now that we have the size, we need to allocate a temporary
16415 		 * buffer in which to store the complete description.  We need
16416 		 * the temporary buffer to be able to drop dtrace_lock()
16417 		 * across the copyout(), below.
16418 		 */
16419 		size = sizeof (dtrace_aggdesc_t) +
16420 		    (aggdesc.dtagd_nrecs * sizeof (dtrace_recdesc_t));
16421 
16422 		buf = kmem_alloc(size, KM_SLEEP);
16423 		dest = (uintptr_t)buf;
16424 
16425 		bcopy(&aggdesc, (void *)dest, sizeof (aggdesc));
16426 		dest += offsetof(dtrace_aggdesc_t, dtagd_rec[0]);
16427 
16428 		for (act = agg->dtag_first; ; act = act->dta_next) {
16429 			dtrace_recdesc_t rec = act->dta_rec;
16430 
16431 			/*
16432 			 * See the comment in the above loop for why we pass
16433 			 * over zero-length records.
16434 			 */
16435 			if (rec.dtrd_size == 0) {
16436 				ASSERT(agg->dtag_hasarg);
16437 				continue;
16438 			}
16439 
16440 			if (nrecs-- == 0)
16441 				break;
16442 
16443 			rec.dtrd_offset -= offs;
16444 			bcopy(&rec, (void *)dest, sizeof (rec));
16445 			dest += sizeof (dtrace_recdesc_t);
16446 
16447 			if (act == &agg->dtag_action)
16448 				break;
16449 		}
16450 
16451 		mutex_exit(&dtrace_lock);
16452 
16453 		if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
16454 			kmem_free(buf, size);
16455 			return (EFAULT);
16456 		}
16457 
16458 		kmem_free(buf, size);
16459 		return (0);
16460 	}
16461 
16462 	case DTRACEIOC_ENABLE: {
16463 		dof_hdr_t *dof;
16464 		dtrace_enabling_t *enab = NULL;
16465 		dtrace_vstate_t *vstate;
16466 		int err = 0;
16467 
16468 		*rv = 0;
16469 
16470 		/*
16471 		 * If a NULL argument has been passed, we take this as our
16472 		 * cue to reevaluate our enablings.
16473 		 */
16474 		if (arg == NULL) {
16475 			dtrace_enabling_matchall();
16476 
16477 			return (0);
16478 		}
16479 
16480 		if ((dof = dtrace_dof_copyin(arg, &rval)) == NULL)
16481 			return (rval);
16482 
16483 		mutex_enter(&cpu_lock);
16484 		mutex_enter(&dtrace_lock);
16485 		vstate = &state->dts_vstate;
16486 
16487 		if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
16488 			mutex_exit(&dtrace_lock);
16489 			mutex_exit(&cpu_lock);
16490 			dtrace_dof_destroy(dof);
16491 			return (EBUSY);
16492 		}
16493 
16494 		if (dtrace_dof_slurp(dof, vstate, cr, &enab, 0, B_TRUE) != 0) {
16495 			mutex_exit(&dtrace_lock);
16496 			mutex_exit(&cpu_lock);
16497 			dtrace_dof_destroy(dof);
16498 			return (EINVAL);
16499 		}
16500 
16501 		if ((rval = dtrace_dof_options(dof, state)) != 0) {
16502 			dtrace_enabling_destroy(enab);
16503 			mutex_exit(&dtrace_lock);
16504 			mutex_exit(&cpu_lock);
16505 			dtrace_dof_destroy(dof);
16506 			return (rval);
16507 		}
16508 
16509 		if ((err = dtrace_enabling_match(enab, rv)) == 0) {
16510 			err = dtrace_enabling_retain(enab);
16511 		} else {
16512 			dtrace_enabling_destroy(enab);
16513 		}
16514 
16515 		mutex_exit(&cpu_lock);
16516 		mutex_exit(&dtrace_lock);
16517 		dtrace_dof_destroy(dof);
16518 
16519 		return (err);
16520 	}
16521 
16522 	case DTRACEIOC_REPLICATE: {
16523 		dtrace_repldesc_t desc;
16524 		dtrace_probedesc_t *match = &desc.dtrpd_match;
16525 		dtrace_probedesc_t *create = &desc.dtrpd_create;
16526 		int err;
16527 
16528 		if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16529 			return (EFAULT);
16530 
16531 		match->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
16532 		match->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
16533 		match->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
16534 		match->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
16535 
16536 		create->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
16537 		create->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
16538 		create->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
16539 		create->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
16540 
16541 		mutex_enter(&dtrace_lock);
16542 		err = dtrace_enabling_replicate(state, match, create);
16543 		mutex_exit(&dtrace_lock);
16544 
16545 		return (err);
16546 	}
16547 
16548 	case DTRACEIOC_PROBEMATCH:
16549 	case DTRACEIOC_PROBES: {
16550 		dtrace_probe_t *probe = NULL;
16551 		dtrace_probedesc_t desc;
16552 		dtrace_probekey_t pkey;
16553 		dtrace_id_t i;
16554 		int m = 0;
16555 		uint32_t priv;
16556 		uid_t uid;
16557 		zoneid_t zoneid;
16558 
16559 		if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16560 			return (EFAULT);
16561 
16562 		desc.dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
16563 		desc.dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
16564 		desc.dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
16565 		desc.dtpd_name[DTRACE_NAMELEN - 1] = '\0';
16566 
16567 		/*
16568 		 * Before we attempt to match this probe, we want to give
16569 		 * all providers the opportunity to provide it.
16570 		 */
16571 		if (desc.dtpd_id == DTRACE_IDNONE) {
16572 			mutex_enter(&dtrace_provider_lock);
16573 			dtrace_probe_provide(&desc, NULL);
16574 			mutex_exit(&dtrace_provider_lock);
16575 			desc.dtpd_id++;
16576 		}
16577 
16578 		if (cmd == DTRACEIOC_PROBEMATCH)  {
16579 			dtrace_probekey(&desc, &pkey);
16580 			pkey.dtpk_id = DTRACE_IDNONE;
16581 		}
16582 
16583 		dtrace_cred2priv(cr, &priv, &uid, &zoneid);
16584 
16585 		mutex_enter(&dtrace_lock);
16586 
16587 		if (cmd == DTRACEIOC_PROBEMATCH) {
16588 			for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
16589 				if ((probe = dtrace_probes[i - 1]) != NULL &&
16590 				    (m = dtrace_match_probe(probe, &pkey,
16591 				    priv, uid, zoneid)) != 0)
16592 					break;
16593 			}
16594 
16595 			if (m < 0) {
16596 				mutex_exit(&dtrace_lock);
16597 				return (EINVAL);
16598 			}
16599 
16600 		} else {
16601 			for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
16602 				if ((probe = dtrace_probes[i - 1]) != NULL &&
16603 				    dtrace_match_priv(probe, priv, uid, zoneid))
16604 					break;
16605 			}
16606 		}
16607 
16608 		if (probe == NULL) {
16609 			mutex_exit(&dtrace_lock);
16610 			return (ESRCH);
16611 		}
16612 
16613 		dtrace_probe_description(probe, &desc);
16614 		mutex_exit(&dtrace_lock);
16615 
16616 		if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16617 			return (EFAULT);
16618 
16619 		return (0);
16620 	}
16621 
16622 	case DTRACEIOC_PROBEARG: {
16623 		dtrace_argdesc_t desc;
16624 		dtrace_probe_t *probe;
16625 		dtrace_provider_t *prov;
16626 
16627 		if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16628 			return (EFAULT);
16629 
16630 		if (desc.dtargd_id == DTRACE_IDNONE)
16631 			return (EINVAL);
16632 
16633 		if (desc.dtargd_ndx == DTRACE_ARGNONE)
16634 			return (EINVAL);
16635 
16636 		mutex_enter(&dtrace_provider_lock);
16637 		mutex_enter(&mod_lock);
16638 		mutex_enter(&dtrace_lock);
16639 
16640 		if (desc.dtargd_id > dtrace_nprobes) {
16641 			mutex_exit(&dtrace_lock);
16642 			mutex_exit(&mod_lock);
16643 			mutex_exit(&dtrace_provider_lock);
16644 			return (EINVAL);
16645 		}
16646 
16647 		if ((probe = dtrace_probes[desc.dtargd_id - 1]) == NULL) {
16648 			mutex_exit(&dtrace_lock);
16649 			mutex_exit(&mod_lock);
16650 			mutex_exit(&dtrace_provider_lock);
16651 			return (EINVAL);
16652 		}
16653 
16654 		mutex_exit(&dtrace_lock);
16655 
16656 		prov = probe->dtpr_provider;
16657 
16658 		if (prov->dtpv_pops.dtps_getargdesc == NULL) {
16659 			/*
16660 			 * There isn't any typed information for this probe.
16661 			 * Set the argument number to DTRACE_ARGNONE.
16662 			 */
16663 			desc.dtargd_ndx = DTRACE_ARGNONE;
16664 		} else {
16665 			desc.dtargd_native[0] = '\0';
16666 			desc.dtargd_xlate[0] = '\0';
16667 			desc.dtargd_mapping = desc.dtargd_ndx;
16668 
16669 			prov->dtpv_pops.dtps_getargdesc(prov->dtpv_arg,
16670 			    probe->dtpr_id, probe->dtpr_arg, &desc);
16671 		}
16672 
16673 		mutex_exit(&mod_lock);
16674 		mutex_exit(&dtrace_provider_lock);
16675 
16676 		if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16677 			return (EFAULT);
16678 
16679 		return (0);
16680 	}
16681 
16682 	case DTRACEIOC_GO: {
16683 		processorid_t cpuid;
16684 		rval = dtrace_state_go(state, &cpuid);
16685 
16686 		if (rval != 0)
16687 			return (rval);
16688 
16689 		if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
16690 			return (EFAULT);
16691 
16692 		return (0);
16693 	}
16694 
16695 	case DTRACEIOC_STOP: {
16696 		processorid_t cpuid;
16697 
16698 		mutex_enter(&dtrace_lock);
16699 		rval = dtrace_state_stop(state, &cpuid);
16700 		mutex_exit(&dtrace_lock);
16701 
16702 		if (rval != 0)
16703 			return (rval);
16704 
16705 		if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
16706 			return (EFAULT);
16707 
16708 		return (0);
16709 	}
16710 
16711 	case DTRACEIOC_DOFGET: {
16712 		dof_hdr_t hdr, *dof;
16713 		uint64_t len;
16714 
16715 		if (copyin((void *)arg, &hdr, sizeof (hdr)) != 0)
16716 			return (EFAULT);
16717 
16718 		mutex_enter(&dtrace_lock);
16719 		dof = dtrace_dof_create(state);
16720 		mutex_exit(&dtrace_lock);
16721 
16722 		len = MIN(hdr.dofh_loadsz, dof->dofh_loadsz);
16723 		rval = copyout(dof, (void *)arg, len);
16724 		dtrace_dof_destroy(dof);
16725 
16726 		return (rval == 0 ? 0 : EFAULT);
16727 	}
16728 
16729 	case DTRACEIOC_AGGSNAP:
16730 	case DTRACEIOC_BUFSNAP: {
16731 		dtrace_bufdesc_t desc;
16732 		caddr_t cached;
16733 		dtrace_buffer_t *buf;
16734 
16735 		if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16736 			return (EFAULT);
16737 
16738 		if (desc.dtbd_cpu < 0 || desc.dtbd_cpu >= NCPU)
16739 			return (EINVAL);
16740 
16741 		mutex_enter(&dtrace_lock);
16742 
16743 		if (cmd == DTRACEIOC_BUFSNAP) {
16744 			buf = &state->dts_buffer[desc.dtbd_cpu];
16745 		} else {
16746 			buf = &state->dts_aggbuffer[desc.dtbd_cpu];
16747 		}
16748 
16749 		if (buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL)) {
16750 			size_t sz = buf->dtb_offset;
16751 
16752 			if (state->dts_activity != DTRACE_ACTIVITY_STOPPED) {
16753 				mutex_exit(&dtrace_lock);
16754 				return (EBUSY);
16755 			}
16756 
16757 			/*
16758 			 * If this buffer has already been consumed, we're
16759 			 * going to indicate that there's nothing left here
16760 			 * to consume.
16761 			 */
16762 			if (buf->dtb_flags & DTRACEBUF_CONSUMED) {
16763 				mutex_exit(&dtrace_lock);
16764 
16765 				desc.dtbd_size = 0;
16766 				desc.dtbd_drops = 0;
16767 				desc.dtbd_errors = 0;
16768 				desc.dtbd_oldest = 0;
16769 				sz = sizeof (desc);
16770 
16771 				if (copyout(&desc, (void *)arg, sz) != 0)
16772 					return (EFAULT);
16773 
16774 				return (0);
16775 			}
16776 
16777 			/*
16778 			 * If this is a ring buffer that has wrapped, we want
16779 			 * to copy the whole thing out.
16780 			 */
16781 			if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
16782 				dtrace_buffer_polish(buf);
16783 				sz = buf->dtb_size;
16784 			}
16785 
16786 			if (copyout(buf->dtb_tomax, desc.dtbd_data, sz) != 0) {
16787 				mutex_exit(&dtrace_lock);
16788 				return (EFAULT);
16789 			}
16790 
16791 			desc.dtbd_size = sz;
16792 			desc.dtbd_drops = buf->dtb_drops;
16793 			desc.dtbd_errors = buf->dtb_errors;
16794 			desc.dtbd_oldest = buf->dtb_xamot_offset;
16795 			desc.dtbd_timestamp = dtrace_gethrtime();
16796 
16797 			mutex_exit(&dtrace_lock);
16798 
16799 			if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16800 				return (EFAULT);
16801 
16802 			buf->dtb_flags |= DTRACEBUF_CONSUMED;
16803 
16804 			return (0);
16805 		}
16806 
16807 		if (buf->dtb_tomax == NULL) {
16808 			ASSERT(buf->dtb_xamot == NULL);
16809 			mutex_exit(&dtrace_lock);
16810 			return (ENOENT);
16811 		}
16812 
16813 		cached = buf->dtb_tomax;
16814 		ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
16815 
16816 		dtrace_xcall(desc.dtbd_cpu,
16817 		    (dtrace_xcall_t)dtrace_buffer_switch, buf);
16818 
16819 		state->dts_errors += buf->dtb_xamot_errors;
16820 
16821 		/*
16822 		 * If the buffers did not actually switch, then the cross call
16823 		 * did not take place -- presumably because the given CPU is
16824 		 * not in the ready set.  If this is the case, we'll return
16825 		 * ENOENT.
16826 		 */
16827 		if (buf->dtb_tomax == cached) {
16828 			ASSERT(buf->dtb_xamot != cached);
16829 			mutex_exit(&dtrace_lock);
16830 			return (ENOENT);
16831 		}
16832 
16833 		ASSERT(cached == buf->dtb_xamot);
16834 
16835 		/*
16836 		 * We have our snapshot; now copy it out.
16837 		 */
16838 		if (copyout(buf->dtb_xamot, desc.dtbd_data,
16839 		    buf->dtb_xamot_offset) != 0) {
16840 			mutex_exit(&dtrace_lock);
16841 			return (EFAULT);
16842 		}
16843 
16844 		desc.dtbd_size = buf->dtb_xamot_offset;
16845 		desc.dtbd_drops = buf->dtb_xamot_drops;
16846 		desc.dtbd_errors = buf->dtb_xamot_errors;
16847 		desc.dtbd_oldest = 0;
16848 		desc.dtbd_timestamp = buf->dtb_switched;
16849 
16850 		mutex_exit(&dtrace_lock);
16851 
16852 		/*
16853 		 * Finally, copy out the buffer description.
16854 		 */
16855 		if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16856 			return (EFAULT);
16857 
16858 		return (0);
16859 	}
16860 
16861 	case DTRACEIOC_CONF: {
16862 		dtrace_conf_t conf;
16863 
16864 		bzero(&conf, sizeof (conf));
16865 		conf.dtc_difversion = DIF_VERSION;
16866 		conf.dtc_difintregs = DIF_DIR_NREGS;
16867 		conf.dtc_diftupregs = DIF_DTR_NREGS;
16868 		conf.dtc_ctfmodel = CTF_MODEL_NATIVE;
16869 
16870 		if (copyout(&conf, (void *)arg, sizeof (conf)) != 0)
16871 			return (EFAULT);
16872 
16873 		return (0);
16874 	}
16875 
16876 	case DTRACEIOC_STATUS: {
16877 		dtrace_status_t stat;
16878 		dtrace_dstate_t *dstate;
16879 		int i, j;
16880 		uint64_t nerrs;
16881 
16882 		/*
16883 		 * See the comment in dtrace_state_deadman() for the reason
16884 		 * for setting dts_laststatus to INT64_MAX before setting
16885 		 * it to the correct value.
16886 		 */
16887 		state->dts_laststatus = INT64_MAX;
16888 		dtrace_membar_producer();
16889 		state->dts_laststatus = dtrace_gethrtime();
16890 
16891 		bzero(&stat, sizeof (stat));
16892 
16893 		mutex_enter(&dtrace_lock);
16894 
16895 		if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE) {
16896 			mutex_exit(&dtrace_lock);
16897 			return (ENOENT);
16898 		}
16899 
16900 		if (state->dts_activity == DTRACE_ACTIVITY_DRAINING)
16901 			stat.dtst_exiting = 1;
16902 
16903 		nerrs = state->dts_errors;
16904 		dstate = &state->dts_vstate.dtvs_dynvars;
16905 
16906 		for (i = 0; i < NCPU; i++) {
16907 			dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[i];
16908 
16909 			stat.dtst_dyndrops += dcpu->dtdsc_drops;
16910 			stat.dtst_dyndrops_dirty += dcpu->dtdsc_dirty_drops;
16911 			stat.dtst_dyndrops_rinsing += dcpu->dtdsc_rinsing_drops;
16912 
16913 			if (state->dts_buffer[i].dtb_flags & DTRACEBUF_FULL)
16914 				stat.dtst_filled++;
16915 
16916 			nerrs += state->dts_buffer[i].dtb_errors;
16917 
16918 			for (j = 0; j < state->dts_nspeculations; j++) {
16919 				dtrace_speculation_t *spec;
16920 				dtrace_buffer_t *buf;
16921 
16922 				spec = &state->dts_speculations[j];
16923 				buf = &spec->dtsp_buffer[i];
16924 				stat.dtst_specdrops += buf->dtb_xamot_drops;
16925 			}
16926 		}
16927 
16928 		stat.dtst_specdrops_busy = state->dts_speculations_busy;
16929 		stat.dtst_specdrops_unavail = state->dts_speculations_unavail;
16930 		stat.dtst_stkstroverflows = state->dts_stkstroverflows;
16931 		stat.dtst_dblerrors = state->dts_dblerrors;
16932 		stat.dtst_killed =
16933 		    (state->dts_activity == DTRACE_ACTIVITY_KILLED);
16934 		stat.dtst_errors = nerrs;
16935 
16936 		mutex_exit(&dtrace_lock);
16937 
16938 		if (copyout(&stat, (void *)arg, sizeof (stat)) != 0)
16939 			return (EFAULT);
16940 
16941 		return (0);
16942 	}
16943 
16944 	case DTRACEIOC_FORMAT: {
16945 		dtrace_fmtdesc_t fmt;
16946 		char *str;
16947 		int len;
16948 
16949 		if (copyin((void *)arg, &fmt, sizeof (fmt)) != 0)
16950 			return (EFAULT);
16951 
16952 		mutex_enter(&dtrace_lock);
16953 
16954 		if (fmt.dtfd_format == 0 ||
16955 		    fmt.dtfd_format > state->dts_nformats) {
16956 			mutex_exit(&dtrace_lock);
16957 			return (EINVAL);
16958 		}
16959 
16960 		/*
16961 		 * Format strings are allocated contiguously and they are
16962 		 * never freed; if a format index is less than the number
16963 		 * of formats, we can assert that the format map is non-NULL
16964 		 * and that the format for the specified index is non-NULL.
16965 		 */
16966 		ASSERT(state->dts_formats != NULL);
16967 		str = state->dts_formats[fmt.dtfd_format - 1];
16968 		ASSERT(str != NULL);
16969 
16970 		len = strlen(str) + 1;
16971 
16972 		if (len > fmt.dtfd_length) {
16973 			fmt.dtfd_length = len;
16974 
16975 			if (copyout(&fmt, (void *)arg, sizeof (fmt)) != 0) {
16976 				mutex_exit(&dtrace_lock);
16977 				return (EINVAL);
16978 			}
16979 		} else {
16980 			if (copyout(str, fmt.dtfd_string, len) != 0) {
16981 				mutex_exit(&dtrace_lock);
16982 				return (EINVAL);
16983 			}
16984 		}
16985 
16986 		mutex_exit(&dtrace_lock);
16987 		return (0);
16988 	}
16989 
16990 	default:
16991 		break;
16992 	}
16993 
16994 	return (ENOTTY);
16995 }
16996 
16997 /*ARGSUSED*/
16998 static int
16999 dtrace_detach(dev_info_t *dip, ddi_detach_cmd_t cmd)
17000 {
17001 	dtrace_state_t *state;
17002 
17003 	switch (cmd) {
17004 	case DDI_DETACH:
17005 		break;
17006 
17007 	case DDI_SUSPEND:
17008 		return (DDI_SUCCESS);
17009 
17010 	default:
17011 		return (DDI_FAILURE);
17012 	}
17013 
17014 	mutex_enter(&cpu_lock);
17015 	mutex_enter(&dtrace_provider_lock);
17016 	mutex_enter(&dtrace_lock);
17017 
17018 	ASSERT(dtrace_opens == 0);
17019 
17020 	if (dtrace_helpers > 0) {
17021 		mutex_exit(&dtrace_provider_lock);
17022 		mutex_exit(&dtrace_lock);
17023 		mutex_exit(&cpu_lock);
17024 		return (DDI_FAILURE);
17025 	}
17026 
17027 	if (dtrace_unregister((dtrace_provider_id_t)dtrace_provider) != 0) {
17028 		mutex_exit(&dtrace_provider_lock);
17029 		mutex_exit(&dtrace_lock);
17030 		mutex_exit(&cpu_lock);
17031 		return (DDI_FAILURE);
17032 	}
17033 
17034 	dtrace_provider = NULL;
17035 
17036 	if ((state = dtrace_anon_grab()) != NULL) {
17037 		/*
17038 		 * If there were ECBs on this state, the provider should
17039 		 * have not been allowed to detach; assert that there is
17040 		 * none.
17041 		 */
17042 		ASSERT(state->dts_necbs == 0);
17043 		dtrace_state_destroy(state);
17044 
17045 		/*
17046 		 * If we're being detached with anonymous state, we need to
17047 		 * indicate to the kernel debugger that DTrace is now inactive.
17048 		 */
17049 		(void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
17050 	}
17051 
17052 	bzero(&dtrace_anon, sizeof (dtrace_anon_t));
17053 	unregister_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
17054 	dtrace_cpu_init = NULL;
17055 	dtrace_helpers_cleanup = NULL;
17056 	dtrace_helpers_fork = NULL;
17057 	dtrace_cpustart_init = NULL;
17058 	dtrace_cpustart_fini = NULL;
17059 	dtrace_debugger_init = NULL;
17060 	dtrace_debugger_fini = NULL;
17061 	dtrace_modload = NULL;
17062 	dtrace_modunload = NULL;
17063 
17064 	ASSERT(dtrace_getf == 0);
17065 	ASSERT(dtrace_closef == NULL);
17066 
17067 	mutex_exit(&cpu_lock);
17068 
17069 	kmem_free(dtrace_probes, dtrace_nprobes * sizeof (dtrace_probe_t *));
17070 	dtrace_probes = NULL;
17071 	dtrace_nprobes = 0;
17072 
17073 	dtrace_hash_destroy(dtrace_bymod);
17074 	dtrace_hash_destroy(dtrace_byfunc);
17075 	dtrace_hash_destroy(dtrace_byname);
17076 	dtrace_bymod = NULL;
17077 	dtrace_byfunc = NULL;
17078 	dtrace_byname = NULL;
17079 
17080 	kmem_cache_destroy(dtrace_state_cache);
17081 	vmem_destroy(dtrace_minor);
17082 	vmem_destroy(dtrace_arena);
17083 
17084 	if (dtrace_toxrange != NULL) {
17085 		kmem_free(dtrace_toxrange,
17086 		    dtrace_toxranges_max * sizeof (dtrace_toxrange_t));
17087 		dtrace_toxrange = NULL;
17088 		dtrace_toxranges = 0;
17089 		dtrace_toxranges_max = 0;
17090 	}
17091 
17092 	ddi_remove_minor_node(dtrace_devi, NULL);
17093 	dtrace_devi = NULL;
17094 
17095 	ddi_soft_state_fini(&dtrace_softstate);
17096 
17097 	ASSERT(dtrace_vtime_references == 0);
17098 	ASSERT(dtrace_opens == 0);
17099 	ASSERT(dtrace_retained == NULL);
17100 
17101 	mutex_exit(&dtrace_lock);
17102 	mutex_exit(&dtrace_provider_lock);
17103 
17104 	/*
17105 	 * We don't destroy the task queue until after we have dropped our
17106 	 * locks (taskq_destroy() may block on running tasks).  To prevent
17107 	 * attempting to do work after we have effectively detached but before
17108 	 * the task queue has been destroyed, all tasks dispatched via the
17109 	 * task queue must check that DTrace is still attached before
17110 	 * performing any operation.
17111 	 */
17112 	taskq_destroy(dtrace_taskq);
17113 	dtrace_taskq = NULL;
17114 
17115 	return (DDI_SUCCESS);
17116 }
17117 
17118 /*ARGSUSED*/
17119 static int
17120 dtrace_info(dev_info_t *dip, ddi_info_cmd_t infocmd, void *arg, void **result)
17121 {
17122 	int error;
17123 
17124 	switch (infocmd) {
17125 	case DDI_INFO_DEVT2DEVINFO:
17126 		*result = (void *)dtrace_devi;
17127 		error = DDI_SUCCESS;
17128 		break;
17129 	case DDI_INFO_DEVT2INSTANCE:
17130 		*result = (void *)0;
17131 		error = DDI_SUCCESS;
17132 		break;
17133 	default:
17134 		error = DDI_FAILURE;
17135 	}
17136 	return (error);
17137 }
17138 
17139 static struct cb_ops dtrace_cb_ops = {
17140 	dtrace_open,		/* open */
17141 	dtrace_close,		/* close */
17142 	nulldev,		/* strategy */
17143 	nulldev,		/* print */
17144 	nodev,			/* dump */
17145 	nodev,			/* read */
17146 	nodev,			/* write */
17147 	dtrace_ioctl,		/* ioctl */
17148 	nodev,			/* devmap */
17149 	nodev,			/* mmap */
17150 	nodev,			/* segmap */
17151 	nochpoll,		/* poll */
17152 	ddi_prop_op,		/* cb_prop_op */
17153 	0,			/* streamtab  */
17154 	D_NEW | D_MP		/* Driver compatibility flag */
17155 };
17156 
17157 static struct dev_ops dtrace_ops = {
17158 	DEVO_REV,		/* devo_rev */
17159 	0,			/* refcnt */
17160 	dtrace_info,		/* get_dev_info */
17161 	nulldev,		/* identify */
17162 	nulldev,		/* probe */
17163 	dtrace_attach,		/* attach */
17164 	dtrace_detach,		/* detach */
17165 	nodev,			/* reset */
17166 	&dtrace_cb_ops,		/* driver operations */
17167 	NULL,			/* bus operations */
17168 	nodev,			/* dev power */
17169 	ddi_quiesce_not_needed,		/* quiesce */
17170 };
17171 
17172 static struct modldrv modldrv = {
17173 	&mod_driverops,		/* module type (this is a pseudo driver) */
17174 	"Dynamic Tracing",	/* name of module */
17175 	&dtrace_ops,		/* driver ops */
17176 };
17177 
17178 static struct modlinkage modlinkage = {
17179 	MODREV_1,
17180 	(void *)&modldrv,
17181 	NULL
17182 };
17183 
17184 int
17185 _init(void)
17186 {
17187 	return (mod_install(&modlinkage));
17188 }
17189 
17190 int
17191 _info(struct modinfo *modinfop)
17192 {
17193 	return (mod_info(&modlinkage, modinfop));
17194 }
17195 
17196 int
17197 _fini(void)
17198 {
17199 	return (mod_remove(&modlinkage));
17200 }
17201