xref: /illumos-gate/usr/src/uts/common/dtrace/dtrace.c (revision 6fb4854bed54ce82bd8610896b64ddebcd4af706)
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) 2011, Joyent, Inc. All rights reserved.
25  */
26 
27 /*
28  * DTrace - Dynamic Tracing for Solaris
29  *
30  * This is the implementation of the Solaris Dynamic Tracing framework
31  * (DTrace).  The user-visible interface to DTrace is described at length in
32  * the "Solaris Dynamic Tracing Guide".  The interfaces between the libdtrace
33  * library, the in-kernel DTrace framework, and the DTrace providers are
34  * described in the block comments in the <sys/dtrace.h> header file.  The
35  * internal architecture of DTrace is described in the block comments in the
36  * <sys/dtrace_impl.h> header file.  The comments contained within the DTrace
37  * implementation very much assume mastery of all of these sources; if one has
38  * an unanswered question about the implementation, one should consult them
39  * first.
40  *
41  * The functions here are ordered roughly as follows:
42  *
43  *   - Probe context functions
44  *   - Probe hashing functions
45  *   - Non-probe context utility functions
46  *   - Matching functions
47  *   - Provider-to-Framework API functions
48  *   - Probe management functions
49  *   - DIF object functions
50  *   - Format functions
51  *   - Predicate functions
52  *   - ECB functions
53  *   - Buffer functions
54  *   - Enabling functions
55  *   - DOF functions
56  *   - Anonymous enabling functions
57  *   - Consumer state functions
58  *   - Helper functions
59  *   - Hook functions
60  *   - Driver cookbook functions
61  *
62  * Each group of functions begins with a block comment labelled the "DTrace
63  * [Group] Functions", allowing one to find each block by searching forward
64  * on capital-f functions.
65  */
66 #include <sys/errno.h>
67 #include <sys/stat.h>
68 #include <sys/modctl.h>
69 #include <sys/conf.h>
70 #include <sys/systm.h>
71 #include <sys/ddi.h>
72 #include <sys/sunddi.h>
73 #include <sys/cpuvar.h>
74 #include <sys/kmem.h>
75 #include <sys/strsubr.h>
76 #include <sys/sysmacros.h>
77 #include <sys/dtrace_impl.h>
78 #include <sys/atomic.h>
79 #include <sys/cmn_err.h>
80 #include <sys/mutex_impl.h>
81 #include <sys/rwlock_impl.h>
82 #include <sys/ctf_api.h>
83 #include <sys/panic.h>
84 #include <sys/priv_impl.h>
85 #include <sys/policy.h>
86 #include <sys/cred_impl.h>
87 #include <sys/procfs_isa.h>
88 #include <sys/taskq.h>
89 #include <sys/mkdev.h>
90 #include <sys/kdi.h>
91 #include <sys/zone.h>
92 #include <sys/socket.h>
93 #include <netinet/in.h>
94 
95 /*
96  * DTrace Tunable Variables
97  *
98  * The following variables may be tuned by adding a line to /etc/system that
99  * includes both the name of the DTrace module ("dtrace") and the name of the
100  * variable.  For example:
101  *
102  *   set dtrace:dtrace_destructive_disallow = 1
103  *
104  * In general, the only variables that one should be tuning this way are those
105  * that affect system-wide DTrace behavior, and for which the default behavior
106  * is undesirable.  Most of these variables are tunable on a per-consumer
107  * basis using DTrace options, and need not be tuned on a system-wide basis.
108  * When tuning these variables, avoid pathological values; while some attempt
109  * is made to verify the integrity of these variables, they are not considered
110  * part of the supported interface to DTrace, and they are therefore not
111  * checked comprehensively.  Further, these variables should not be tuned
112  * dynamically via "mdb -kw" or other means; they should only be tuned via
113  * /etc/system.
114  */
115 int		dtrace_destructive_disallow = 0;
116 dtrace_optval_t	dtrace_nonroot_maxsize = (16 * 1024 * 1024);
117 size_t		dtrace_difo_maxsize = (256 * 1024);
118 dtrace_optval_t	dtrace_dof_maxsize = (256 * 1024);
119 size_t		dtrace_global_maxsize = (16 * 1024);
120 size_t		dtrace_actions_max = (16 * 1024);
121 size_t		dtrace_retain_max = 1024;
122 dtrace_optval_t	dtrace_helper_actions_max = 32;
123 dtrace_optval_t	dtrace_helper_providers_max = 32;
124 dtrace_optval_t	dtrace_dstate_defsize = (1 * 1024 * 1024);
125 size_t		dtrace_strsize_default = 256;
126 dtrace_optval_t	dtrace_cleanrate_default = 9900990;		/* 101 hz */
127 dtrace_optval_t	dtrace_cleanrate_min = 200000;			/* 5000 hz */
128 dtrace_optval_t	dtrace_cleanrate_max = (uint64_t)60 * NANOSEC;	/* 1/minute */
129 dtrace_optval_t	dtrace_aggrate_default = NANOSEC;		/* 1 hz */
130 dtrace_optval_t	dtrace_statusrate_default = NANOSEC;		/* 1 hz */
131 dtrace_optval_t dtrace_statusrate_max = (hrtime_t)10 * NANOSEC;	 /* 6/minute */
132 dtrace_optval_t	dtrace_switchrate_default = NANOSEC;		/* 1 hz */
133 dtrace_optval_t	dtrace_nspec_default = 1;
134 dtrace_optval_t	dtrace_specsize_default = 32 * 1024;
135 dtrace_optval_t dtrace_stackframes_default = 20;
136 dtrace_optval_t dtrace_ustackframes_default = 20;
137 dtrace_optval_t dtrace_jstackframes_default = 50;
138 dtrace_optval_t dtrace_jstackstrsize_default = 512;
139 int		dtrace_msgdsize_max = 128;
140 hrtime_t	dtrace_chill_max = 500 * (NANOSEC / MILLISEC);	/* 500 ms */
141 hrtime_t	dtrace_chill_interval = NANOSEC;		/* 1000 ms */
142 int		dtrace_devdepth_max = 32;
143 int		dtrace_err_verbose;
144 hrtime_t	dtrace_deadman_interval = NANOSEC;
145 hrtime_t	dtrace_deadman_timeout = (hrtime_t)10 * NANOSEC;
146 hrtime_t	dtrace_deadman_user = (hrtime_t)30 * NANOSEC;
147 hrtime_t	dtrace_unregister_defunct_reap = (hrtime_t)60 * NANOSEC;
148 
149 /*
150  * DTrace External Variables
151  *
152  * As dtrace(7D) is a kernel module, any DTrace variables are obviously
153  * available to DTrace consumers via the backtick (`) syntax.  One of these,
154  * dtrace_zero, is made deliberately so:  it is provided as a source of
155  * well-known, zero-filled memory.  While this variable is not documented,
156  * it is used by some translators as an implementation detail.
157  */
158 const char	dtrace_zero[256] = { 0 };	/* zero-filled memory */
159 
160 /*
161  * DTrace Internal Variables
162  */
163 static dev_info_t	*dtrace_devi;		/* device info */
164 static vmem_t		*dtrace_arena;		/* probe ID arena */
165 static vmem_t		*dtrace_minor;		/* minor number arena */
166 static taskq_t		*dtrace_taskq;		/* task queue */
167 static dtrace_probe_t	**dtrace_probes;	/* array of all probes */
168 static int		dtrace_nprobes;		/* number of probes */
169 static dtrace_provider_t *dtrace_provider;	/* provider list */
170 static dtrace_meta_t	*dtrace_meta_pid;	/* user-land meta provider */
171 static int		dtrace_opens;		/* number of opens */
172 static int		dtrace_helpers;		/* number of helpers */
173 static void		*dtrace_softstate;	/* softstate pointer */
174 static dtrace_hash_t	*dtrace_bymod;		/* probes hashed by module */
175 static dtrace_hash_t	*dtrace_byfunc;		/* probes hashed by function */
176 static dtrace_hash_t	*dtrace_byname;		/* probes hashed by name */
177 static dtrace_toxrange_t *dtrace_toxrange;	/* toxic range array */
178 static int		dtrace_toxranges;	/* number of toxic ranges */
179 static int		dtrace_toxranges_max;	/* size of toxic range array */
180 static dtrace_anon_t	dtrace_anon;		/* anonymous enabling */
181 static kmem_cache_t	*dtrace_state_cache;	/* cache for dynamic state */
182 static uint64_t		dtrace_vtime_references; /* number of vtimestamp refs */
183 static kthread_t	*dtrace_panicked;	/* panicking thread */
184 static dtrace_ecb_t	*dtrace_ecb_create_cache; /* cached created ECB */
185 static dtrace_genid_t	dtrace_probegen;	/* current probe generation */
186 static dtrace_helpers_t *dtrace_deferred_pid;	/* deferred helper list */
187 static dtrace_enabling_t *dtrace_retained;	/* list of retained enablings */
188 static dtrace_genid_t	dtrace_retained_gen;	/* current retained enab gen */
189 static dtrace_dynvar_t	dtrace_dynhash_sink;	/* end of dynamic hash chains */
190 static int		dtrace_dynvar_failclean; /* dynvars failed to clean */
191 
192 /*
193  * DTrace Locking
194  * DTrace is protected by three (relatively coarse-grained) locks:
195  *
196  * (1) dtrace_lock is required to manipulate essentially any DTrace state,
197  *     including enabling state, probes, ECBs, consumer state, helper state,
198  *     etc.  Importantly, dtrace_lock is _not_ required when in probe context;
199  *     probe context is lock-free -- synchronization is handled via the
200  *     dtrace_sync() cross call mechanism.
201  *
202  * (2) dtrace_provider_lock is required when manipulating provider state, or
203  *     when provider state must be held constant.
204  *
205  * (3) dtrace_meta_lock is required when manipulating meta provider state, or
206  *     when meta provider state must be held constant.
207  *
208  * The lock ordering between these three locks is dtrace_meta_lock before
209  * dtrace_provider_lock before dtrace_lock.  (In particular, there are
210  * several places where dtrace_provider_lock is held by the framework as it
211  * calls into the providers -- which then call back into the framework,
212  * grabbing dtrace_lock.)
213  *
214  * There are two other locks in the mix:  mod_lock and cpu_lock.  With respect
215  * to dtrace_provider_lock and dtrace_lock, cpu_lock continues its historical
216  * role as a coarse-grained lock; it is acquired before both of these locks.
217  * With respect to dtrace_meta_lock, its behavior is stranger:  cpu_lock must
218  * be acquired _between_ dtrace_meta_lock and any other DTrace locks.
219  * mod_lock is similar with respect to dtrace_provider_lock in that it must be
220  * acquired _between_ dtrace_provider_lock and dtrace_lock.
221  */
222 static kmutex_t		dtrace_lock;		/* probe state lock */
223 static kmutex_t		dtrace_provider_lock;	/* provider state lock */
224 static kmutex_t		dtrace_meta_lock;	/* meta-provider state lock */
225 
226 /*
227  * DTrace Provider Variables
228  *
229  * These are the variables relating to DTrace as a provider (that is, the
230  * provider of the BEGIN, END, and ERROR probes).
231  */
232 static dtrace_pattr_t	dtrace_provider_attr = {
233 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
234 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
235 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
236 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
237 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
238 };
239 
240 static void
241 dtrace_nullop(void)
242 {}
243 
244 static int
245 dtrace_enable_nullop(void)
246 {
247 	return (0);
248 }
249 
250 static dtrace_pops_t	dtrace_provider_ops = {
251 	(void (*)(void *, const dtrace_probedesc_t *))dtrace_nullop,
252 	(void (*)(void *, struct modctl *))dtrace_nullop,
253 	(int (*)(void *, dtrace_id_t, void *))dtrace_enable_nullop,
254 	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
255 	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
256 	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
257 	NULL,
258 	NULL,
259 	NULL,
260 	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop
261 };
262 
263 static dtrace_id_t	dtrace_probeid_begin;	/* special BEGIN probe */
264 static dtrace_id_t	dtrace_probeid_end;	/* special END probe */
265 dtrace_id_t		dtrace_probeid_error;	/* special ERROR probe */
266 
267 /*
268  * DTrace Helper Tracing Variables
269  */
270 uint32_t dtrace_helptrace_next = 0;
271 uint32_t dtrace_helptrace_nlocals;
272 char	*dtrace_helptrace_buffer;
273 int	dtrace_helptrace_bufsize = 512 * 1024;
274 
275 #ifdef DEBUG
276 int	dtrace_helptrace_enabled = 1;
277 #else
278 int	dtrace_helptrace_enabled = 0;
279 #endif
280 
281 /*
282  * DTrace Error Hashing
283  *
284  * On DEBUG kernels, DTrace will track the errors that has seen in a hash
285  * table.  This is very useful for checking coverage of tests that are
286  * expected to induce DIF or DOF processing errors, and may be useful for
287  * debugging problems in the DIF code generator or in DOF generation .  The
288  * error hash may be examined with the ::dtrace_errhash MDB dcmd.
289  */
290 #ifdef DEBUG
291 static dtrace_errhash_t	dtrace_errhash[DTRACE_ERRHASHSZ];
292 static const char *dtrace_errlast;
293 static kthread_t *dtrace_errthread;
294 static kmutex_t dtrace_errlock;
295 #endif
296 
297 /*
298  * DTrace Macros and Constants
299  *
300  * These are various macros that are useful in various spots in the
301  * implementation, along with a few random constants that have no meaning
302  * outside of the implementation.  There is no real structure to this cpp
303  * mishmash -- but is there ever?
304  */
305 #define	DTRACE_HASHSTR(hash, probe)	\
306 	dtrace_hash_str(*((char **)((uintptr_t)(probe) + (hash)->dth_stroffs)))
307 
308 #define	DTRACE_HASHNEXT(hash, probe)	\
309 	(dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_nextoffs)
310 
311 #define	DTRACE_HASHPREV(hash, probe)	\
312 	(dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_prevoffs)
313 
314 #define	DTRACE_HASHEQ(hash, lhs, rhs)	\
315 	(strcmp(*((char **)((uintptr_t)(lhs) + (hash)->dth_stroffs)), \
316 	    *((char **)((uintptr_t)(rhs) + (hash)->dth_stroffs))) == 0)
317 
318 #define	DTRACE_AGGHASHSIZE_SLEW		17
319 
320 #define	DTRACE_V4MAPPED_OFFSET		(sizeof (uint32_t) * 3)
321 
322 /*
323  * The key for a thread-local variable consists of the lower 61 bits of the
324  * t_did, plus the 3 bits of the highest active interrupt above LOCK_LEVEL.
325  * We add DIF_VARIABLE_MAX to t_did to assure that the thread key is never
326  * equal to a variable identifier.  This is necessary (but not sufficient) to
327  * assure that global associative arrays never collide with thread-local
328  * variables.  To guarantee that they cannot collide, we must also define the
329  * order for keying dynamic variables.  That order is:
330  *
331  *   [ key0 ] ... [ keyn ] [ variable-key ] [ tls-key ]
332  *
333  * Because the variable-key and the tls-key are in orthogonal spaces, there is
334  * no way for a global variable key signature to match a thread-local key
335  * signature.
336  */
337 #define	DTRACE_TLS_THRKEY(where) { \
338 	uint_t intr = 0; \
339 	uint_t actv = CPU->cpu_intr_actv >> (LOCK_LEVEL + 1); \
340 	for (; actv; actv >>= 1) \
341 		intr++; \
342 	ASSERT(intr < (1 << 3)); \
343 	(where) = ((curthread->t_did + DIF_VARIABLE_MAX) & \
344 	    (((uint64_t)1 << 61) - 1)) | ((uint64_t)intr << 61); \
345 }
346 
347 #define	DT_BSWAP_8(x)	((x) & 0xff)
348 #define	DT_BSWAP_16(x)	((DT_BSWAP_8(x) << 8) | DT_BSWAP_8((x) >> 8))
349 #define	DT_BSWAP_32(x)	((DT_BSWAP_16(x) << 16) | DT_BSWAP_16((x) >> 16))
350 #define	DT_BSWAP_64(x)	((DT_BSWAP_32(x) << 32) | DT_BSWAP_32((x) >> 32))
351 
352 #define	DT_MASK_LO 0x00000000FFFFFFFFULL
353 
354 #define	DTRACE_STORE(type, tomax, offset, what) \
355 	*((type *)((uintptr_t)(tomax) + (uintptr_t)offset)) = (type)(what);
356 
357 #ifndef __i386
358 #define	DTRACE_ALIGNCHECK(addr, size, flags)				\
359 	if (addr & (size - 1)) {					\
360 		*flags |= CPU_DTRACE_BADALIGN;				\
361 		cpu_core[CPU->cpu_id].cpuc_dtrace_illval = addr;	\
362 		return (0);						\
363 	}
364 #else
365 #define	DTRACE_ALIGNCHECK(addr, size, flags)
366 #endif
367 
368 /*
369  * Test whether a range of memory starting at testaddr of size testsz falls
370  * within the range of memory described by addr, sz.  We take care to avoid
371  * problems with overflow and underflow of the unsigned quantities, and
372  * disallow all negative sizes.  Ranges of size 0 are allowed.
373  */
374 #define	DTRACE_INRANGE(testaddr, testsz, baseaddr, basesz) \
375 	((testaddr) - (baseaddr) < (basesz) && \
376 	(testaddr) + (testsz) - (baseaddr) <= (basesz) && \
377 	(testaddr) + (testsz) >= (testaddr))
378 
379 /*
380  * Test whether alloc_sz bytes will fit in the scratch region.  We isolate
381  * alloc_sz on the righthand side of the comparison in order to avoid overflow
382  * or underflow in the comparison with it.  This is simpler than the INRANGE
383  * check above, because we know that the dtms_scratch_ptr is valid in the
384  * range.  Allocations of size zero are allowed.
385  */
386 #define	DTRACE_INSCRATCH(mstate, alloc_sz) \
387 	((mstate)->dtms_scratch_base + (mstate)->dtms_scratch_size - \
388 	(mstate)->dtms_scratch_ptr >= (alloc_sz))
389 
390 #define	DTRACE_LOADFUNC(bits)						\
391 /*CSTYLED*/								\
392 uint##bits##_t								\
393 dtrace_load##bits(uintptr_t addr)					\
394 {									\
395 	size_t size = bits / NBBY;					\
396 	/*CSTYLED*/							\
397 	uint##bits##_t rval;						\
398 	int i;								\
399 	volatile uint16_t *flags = (volatile uint16_t *)		\
400 	    &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;			\
401 									\
402 	DTRACE_ALIGNCHECK(addr, size, flags);				\
403 									\
404 	for (i = 0; i < dtrace_toxranges; i++) {			\
405 		if (addr >= dtrace_toxrange[i].dtt_limit)		\
406 			continue;					\
407 									\
408 		if (addr + size <= dtrace_toxrange[i].dtt_base)		\
409 			continue;					\
410 									\
411 		/*							\
412 		 * This address falls within a toxic region; return 0.	\
413 		 */							\
414 		*flags |= CPU_DTRACE_BADADDR;				\
415 		cpu_core[CPU->cpu_id].cpuc_dtrace_illval = addr;	\
416 		return (0);						\
417 	}								\
418 									\
419 	*flags |= CPU_DTRACE_NOFAULT;					\
420 	/*CSTYLED*/							\
421 	rval = *((volatile uint##bits##_t *)addr);			\
422 	*flags &= ~CPU_DTRACE_NOFAULT;					\
423 									\
424 	return (!(*flags & CPU_DTRACE_FAULT) ? rval : 0);		\
425 }
426 
427 #ifdef _LP64
428 #define	dtrace_loadptr	dtrace_load64
429 #else
430 #define	dtrace_loadptr	dtrace_load32
431 #endif
432 
433 #define	DTRACE_DYNHASH_FREE	0
434 #define	DTRACE_DYNHASH_SINK	1
435 #define	DTRACE_DYNHASH_VALID	2
436 
437 #define	DTRACE_MATCH_FAIL	-1
438 #define	DTRACE_MATCH_NEXT	0
439 #define	DTRACE_MATCH_DONE	1
440 #define	DTRACE_ANCHORED(probe)	((probe)->dtpr_func[0] != '\0')
441 #define	DTRACE_STATE_ALIGN	64
442 
443 #define	DTRACE_FLAGS2FLT(flags)						\
444 	(((flags) & CPU_DTRACE_BADADDR) ? DTRACEFLT_BADADDR :		\
445 	((flags) & CPU_DTRACE_ILLOP) ? DTRACEFLT_ILLOP :		\
446 	((flags) & CPU_DTRACE_DIVZERO) ? DTRACEFLT_DIVZERO :		\
447 	((flags) & CPU_DTRACE_KPRIV) ? DTRACEFLT_KPRIV :		\
448 	((flags) & CPU_DTRACE_UPRIV) ? DTRACEFLT_UPRIV :		\
449 	((flags) & CPU_DTRACE_TUPOFLOW) ?  DTRACEFLT_TUPOFLOW :		\
450 	((flags) & CPU_DTRACE_BADALIGN) ?  DTRACEFLT_BADALIGN :		\
451 	((flags) & CPU_DTRACE_NOSCRATCH) ?  DTRACEFLT_NOSCRATCH :	\
452 	((flags) & CPU_DTRACE_BADSTACK) ?  DTRACEFLT_BADSTACK :		\
453 	DTRACEFLT_UNKNOWN)
454 
455 #define	DTRACEACT_ISSTRING(act)						\
456 	((act)->dta_kind == DTRACEACT_DIFEXPR &&			\
457 	(act)->dta_difo->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING)
458 
459 static size_t dtrace_strlen(const char *, size_t);
460 static dtrace_probe_t *dtrace_probe_lookup_id(dtrace_id_t id);
461 static void dtrace_enabling_provide(dtrace_provider_t *);
462 static int dtrace_enabling_match(dtrace_enabling_t *, int *);
463 static void dtrace_enabling_matchall(void);
464 static void dtrace_enabling_reap(void);
465 static dtrace_state_t *dtrace_anon_grab(void);
466 static uint64_t dtrace_helper(int, dtrace_mstate_t *,
467     dtrace_state_t *, uint64_t, uint64_t);
468 static dtrace_helpers_t *dtrace_helpers_create(proc_t *);
469 static void dtrace_buffer_drop(dtrace_buffer_t *);
470 static int dtrace_buffer_consumed(dtrace_buffer_t *, hrtime_t when);
471 static intptr_t dtrace_buffer_reserve(dtrace_buffer_t *, size_t, size_t,
472     dtrace_state_t *, dtrace_mstate_t *);
473 static int dtrace_state_option(dtrace_state_t *, dtrace_optid_t,
474     dtrace_optval_t);
475 static int dtrace_ecb_create_enable(dtrace_probe_t *, void *);
476 static void dtrace_helper_provider_destroy(dtrace_helper_provider_t *);
477 
478 /*
479  * DTrace Probe Context Functions
480  *
481  * These functions are called from probe context.  Because probe context is
482  * any context in which C may be called, arbitrarily locks may be held,
483  * interrupts may be disabled, we may be in arbitrary dispatched state, etc.
484  * As a result, functions called from probe context may only call other DTrace
485  * support functions -- they may not interact at all with the system at large.
486  * (Note that the ASSERT macro is made probe-context safe by redefining it in
487  * terms of dtrace_assfail(), a probe-context safe function.) If arbitrary
488  * loads are to be performed from probe context, they _must_ be in terms of
489  * the safe dtrace_load*() variants.
490  *
491  * Some functions in this block are not actually called from probe context;
492  * for these functions, there will be a comment above the function reading
493  * "Note:  not called from probe context."
494  */
495 void
496 dtrace_panic(const char *format, ...)
497 {
498 	va_list alist;
499 
500 	va_start(alist, format);
501 	dtrace_vpanic(format, alist);
502 	va_end(alist);
503 }
504 
505 int
506 dtrace_assfail(const char *a, const char *f, int l)
507 {
508 	dtrace_panic("assertion failed: %s, file: %s, line: %d", a, f, l);
509 
510 	/*
511 	 * We just need something here that even the most clever compiler
512 	 * cannot optimize away.
513 	 */
514 	return (a[(uintptr_t)f]);
515 }
516 
517 /*
518  * Atomically increment a specified error counter from probe context.
519  */
520 static void
521 dtrace_error(uint32_t *counter)
522 {
523 	/*
524 	 * Most counters stored to in probe context are per-CPU counters.
525 	 * However, there are some error conditions that are sufficiently
526 	 * arcane that they don't merit per-CPU storage.  If these counters
527 	 * are incremented concurrently on different CPUs, scalability will be
528 	 * adversely affected -- but we don't expect them to be white-hot in a
529 	 * correctly constructed enabling...
530 	 */
531 	uint32_t oval, nval;
532 
533 	do {
534 		oval = *counter;
535 
536 		if ((nval = oval + 1) == 0) {
537 			/*
538 			 * If the counter would wrap, set it to 1 -- assuring
539 			 * that the counter is never zero when we have seen
540 			 * errors.  (The counter must be 32-bits because we
541 			 * aren't guaranteed a 64-bit compare&swap operation.)
542 			 * To save this code both the infamy of being fingered
543 			 * by a priggish news story and the indignity of being
544 			 * the target of a neo-puritan witch trial, we're
545 			 * carefully avoiding any colorful description of the
546 			 * likelihood of this condition -- but suffice it to
547 			 * say that it is only slightly more likely than the
548 			 * overflow of predicate cache IDs, as discussed in
549 			 * dtrace_predicate_create().
550 			 */
551 			nval = 1;
552 		}
553 	} while (dtrace_cas32(counter, oval, nval) != oval);
554 }
555 
556 /*
557  * Use the DTRACE_LOADFUNC macro to define functions for each of loading a
558  * uint8_t, a uint16_t, a uint32_t and a uint64_t.
559  */
560 DTRACE_LOADFUNC(8)
561 DTRACE_LOADFUNC(16)
562 DTRACE_LOADFUNC(32)
563 DTRACE_LOADFUNC(64)
564 
565 static int
566 dtrace_inscratch(uintptr_t dest, size_t size, dtrace_mstate_t *mstate)
567 {
568 	if (dest < mstate->dtms_scratch_base)
569 		return (0);
570 
571 	if (dest + size < dest)
572 		return (0);
573 
574 	if (dest + size > mstate->dtms_scratch_ptr)
575 		return (0);
576 
577 	return (1);
578 }
579 
580 static int
581 dtrace_canstore_statvar(uint64_t addr, size_t sz,
582     dtrace_statvar_t **svars, int nsvars)
583 {
584 	int i;
585 
586 	for (i = 0; i < nsvars; i++) {
587 		dtrace_statvar_t *svar = svars[i];
588 
589 		if (svar == NULL || svar->dtsv_size == 0)
590 			continue;
591 
592 		if (DTRACE_INRANGE(addr, sz, svar->dtsv_data, svar->dtsv_size))
593 			return (1);
594 	}
595 
596 	return (0);
597 }
598 
599 /*
600  * Check to see if the address is within a memory region to which a store may
601  * be issued.  This includes the DTrace scratch areas, and any DTrace variable
602  * region.  The caller of dtrace_canstore() is responsible for performing any
603  * alignment checks that are needed before stores are actually executed.
604  */
605 static int
606 dtrace_canstore(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
607     dtrace_vstate_t *vstate)
608 {
609 	/*
610 	 * First, check to see if the address is in scratch space...
611 	 */
612 	if (DTRACE_INRANGE(addr, sz, mstate->dtms_scratch_base,
613 	    mstate->dtms_scratch_size))
614 		return (1);
615 
616 	/*
617 	 * Now check to see if it's a dynamic variable.  This check will pick
618 	 * up both thread-local variables and any global dynamically-allocated
619 	 * variables.
620 	 */
621 	if (DTRACE_INRANGE(addr, sz, (uintptr_t)vstate->dtvs_dynvars.dtds_base,
622 	    vstate->dtvs_dynvars.dtds_size)) {
623 		dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
624 		uintptr_t base = (uintptr_t)dstate->dtds_base +
625 		    (dstate->dtds_hashsize * sizeof (dtrace_dynhash_t));
626 		uintptr_t chunkoffs;
627 
628 		/*
629 		 * Before we assume that we can store here, we need to make
630 		 * sure that it isn't in our metadata -- storing to our
631 		 * dynamic variable metadata would corrupt our state.  For
632 		 * the range to not include any dynamic variable metadata,
633 		 * it must:
634 		 *
635 		 *	(1) Start above the hash table that is at the base of
636 		 *	the dynamic variable space
637 		 *
638 		 *	(2) Have a starting chunk offset that is beyond the
639 		 *	dtrace_dynvar_t that is at the base of every chunk
640 		 *
641 		 *	(3) Not span a chunk boundary
642 		 *
643 		 */
644 		if (addr < base)
645 			return (0);
646 
647 		chunkoffs = (addr - base) % dstate->dtds_chunksize;
648 
649 		if (chunkoffs < sizeof (dtrace_dynvar_t))
650 			return (0);
651 
652 		if (chunkoffs + sz > dstate->dtds_chunksize)
653 			return (0);
654 
655 		return (1);
656 	}
657 
658 	/*
659 	 * Finally, check the static local and global variables.  These checks
660 	 * take the longest, so we perform them last.
661 	 */
662 	if (dtrace_canstore_statvar(addr, sz,
663 	    vstate->dtvs_locals, vstate->dtvs_nlocals))
664 		return (1);
665 
666 	if (dtrace_canstore_statvar(addr, sz,
667 	    vstate->dtvs_globals, vstate->dtvs_nglobals))
668 		return (1);
669 
670 	return (0);
671 }
672 
673 
674 /*
675  * Convenience routine to check to see if the address is within a memory
676  * region in which a load may be issued given the user's privilege level;
677  * if not, it sets the appropriate error flags and loads 'addr' into the
678  * illegal value slot.
679  *
680  * DTrace subroutines (DIF_SUBR_*) should use this helper to implement
681  * appropriate memory access protection.
682  */
683 static int
684 dtrace_canload(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
685     dtrace_vstate_t *vstate)
686 {
687 	volatile uintptr_t *illval = &cpu_core[CPU->cpu_id].cpuc_dtrace_illval;
688 
689 	/*
690 	 * If we hold the privilege to read from kernel memory, then
691 	 * everything is readable.
692 	 */
693 	if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
694 		return (1);
695 
696 	/*
697 	 * You can obviously read that which you can store.
698 	 */
699 	if (dtrace_canstore(addr, sz, mstate, vstate))
700 		return (1);
701 
702 	/*
703 	 * We're allowed to read from our own string table.
704 	 */
705 	if (DTRACE_INRANGE(addr, sz, (uintptr_t)mstate->dtms_difo->dtdo_strtab,
706 	    mstate->dtms_difo->dtdo_strlen))
707 		return (1);
708 
709 	DTRACE_CPUFLAG_SET(CPU_DTRACE_KPRIV);
710 	*illval = addr;
711 	return (0);
712 }
713 
714 /*
715  * Convenience routine to check to see if a given string is within a memory
716  * region in which a load may be issued given the user's privilege level;
717  * this exists so that we don't need to issue unnecessary dtrace_strlen()
718  * calls in the event that the user has all privileges.
719  */
720 static int
721 dtrace_strcanload(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
722     dtrace_vstate_t *vstate)
723 {
724 	size_t strsz;
725 
726 	/*
727 	 * If we hold the privilege to read from kernel memory, then
728 	 * everything is readable.
729 	 */
730 	if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
731 		return (1);
732 
733 	strsz = 1 + dtrace_strlen((char *)(uintptr_t)addr, sz);
734 	if (dtrace_canload(addr, strsz, mstate, vstate))
735 		return (1);
736 
737 	return (0);
738 }
739 
740 /*
741  * Convenience routine to check to see if a given variable is within a memory
742  * region in which a load may be issued given the user's privilege level.
743  */
744 static int
745 dtrace_vcanload(void *src, dtrace_diftype_t *type, dtrace_mstate_t *mstate,
746     dtrace_vstate_t *vstate)
747 {
748 	size_t sz;
749 	ASSERT(type->dtdt_flags & DIF_TF_BYREF);
750 
751 	/*
752 	 * If we hold the privilege to read from kernel memory, then
753 	 * everything is readable.
754 	 */
755 	if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
756 		return (1);
757 
758 	if (type->dtdt_kind == DIF_TYPE_STRING)
759 		sz = dtrace_strlen(src,
760 		    vstate->dtvs_state->dts_options[DTRACEOPT_STRSIZE]) + 1;
761 	else
762 		sz = type->dtdt_size;
763 
764 	return (dtrace_canload((uintptr_t)src, sz, mstate, vstate));
765 }
766 
767 /*
768  * Compare two strings using safe loads.
769  */
770 static int
771 dtrace_strncmp(char *s1, char *s2, size_t limit)
772 {
773 	uint8_t c1, c2;
774 	volatile uint16_t *flags;
775 
776 	if (s1 == s2 || limit == 0)
777 		return (0);
778 
779 	flags = (volatile uint16_t *)&cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
780 
781 	do {
782 		if (s1 == NULL) {
783 			c1 = '\0';
784 		} else {
785 			c1 = dtrace_load8((uintptr_t)s1++);
786 		}
787 
788 		if (s2 == NULL) {
789 			c2 = '\0';
790 		} else {
791 			c2 = dtrace_load8((uintptr_t)s2++);
792 		}
793 
794 		if (c1 != c2)
795 			return (c1 - c2);
796 	} while (--limit && c1 != '\0' && !(*flags & CPU_DTRACE_FAULT));
797 
798 	return (0);
799 }
800 
801 /*
802  * Compute strlen(s) for a string using safe memory accesses.  The additional
803  * len parameter is used to specify a maximum length to ensure completion.
804  */
805 static size_t
806 dtrace_strlen(const char *s, size_t lim)
807 {
808 	uint_t len;
809 
810 	for (len = 0; len != lim; len++) {
811 		if (dtrace_load8((uintptr_t)s++) == '\0')
812 			break;
813 	}
814 
815 	return (len);
816 }
817 
818 /*
819  * Check if an address falls within a toxic region.
820  */
821 static int
822 dtrace_istoxic(uintptr_t kaddr, size_t size)
823 {
824 	uintptr_t taddr, tsize;
825 	int i;
826 
827 	for (i = 0; i < dtrace_toxranges; i++) {
828 		taddr = dtrace_toxrange[i].dtt_base;
829 		tsize = dtrace_toxrange[i].dtt_limit - taddr;
830 
831 		if (kaddr - taddr < tsize) {
832 			DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
833 			cpu_core[CPU->cpu_id].cpuc_dtrace_illval = kaddr;
834 			return (1);
835 		}
836 
837 		if (taddr - kaddr < size) {
838 			DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
839 			cpu_core[CPU->cpu_id].cpuc_dtrace_illval = taddr;
840 			return (1);
841 		}
842 	}
843 
844 	return (0);
845 }
846 
847 /*
848  * Copy src to dst using safe memory accesses.  The src is assumed to be unsafe
849  * memory specified by the DIF program.  The dst is assumed to be safe memory
850  * that we can store to directly because it is managed by DTrace.  As with
851  * standard bcopy, overlapping copies are handled properly.
852  */
853 static void
854 dtrace_bcopy(const void *src, void *dst, size_t len)
855 {
856 	if (len != 0) {
857 		uint8_t *s1 = dst;
858 		const uint8_t *s2 = src;
859 
860 		if (s1 <= s2) {
861 			do {
862 				*s1++ = dtrace_load8((uintptr_t)s2++);
863 			} while (--len != 0);
864 		} else {
865 			s2 += len;
866 			s1 += len;
867 
868 			do {
869 				*--s1 = dtrace_load8((uintptr_t)--s2);
870 			} while (--len != 0);
871 		}
872 	}
873 }
874 
875 /*
876  * Copy src to dst using safe memory accesses, up to either the specified
877  * length, or the point that a nul byte is encountered.  The src is assumed to
878  * be unsafe memory specified by the DIF program.  The dst is assumed to be
879  * safe memory that we can store to directly because it is managed by DTrace.
880  * Unlike dtrace_bcopy(), overlapping regions are not handled.
881  */
882 static void
883 dtrace_strcpy(const void *src, void *dst, size_t len)
884 {
885 	if (len != 0) {
886 		uint8_t *s1 = dst, c;
887 		const uint8_t *s2 = src;
888 
889 		do {
890 			*s1++ = c = dtrace_load8((uintptr_t)s2++);
891 		} while (--len != 0 && c != '\0');
892 	}
893 }
894 
895 /*
896  * Copy src to dst, deriving the size and type from the specified (BYREF)
897  * variable type.  The src is assumed to be unsafe memory specified by the DIF
898  * program.  The dst is assumed to be DTrace variable memory that is of the
899  * specified type; we assume that we can store to directly.
900  */
901 static void
902 dtrace_vcopy(void *src, void *dst, dtrace_diftype_t *type)
903 {
904 	ASSERT(type->dtdt_flags & DIF_TF_BYREF);
905 
906 	if (type->dtdt_kind == DIF_TYPE_STRING) {
907 		dtrace_strcpy(src, dst, type->dtdt_size);
908 	} else {
909 		dtrace_bcopy(src, dst, type->dtdt_size);
910 	}
911 }
912 
913 /*
914  * Compare s1 to s2 using safe memory accesses.  The s1 data is assumed to be
915  * unsafe memory specified by the DIF program.  The s2 data is assumed to be
916  * safe memory that we can access directly because it is managed by DTrace.
917  */
918 static int
919 dtrace_bcmp(const void *s1, const void *s2, size_t len)
920 {
921 	volatile uint16_t *flags;
922 
923 	flags = (volatile uint16_t *)&cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
924 
925 	if (s1 == s2)
926 		return (0);
927 
928 	if (s1 == NULL || s2 == NULL)
929 		return (1);
930 
931 	if (s1 != s2 && len != 0) {
932 		const uint8_t *ps1 = s1;
933 		const uint8_t *ps2 = s2;
934 
935 		do {
936 			if (dtrace_load8((uintptr_t)ps1++) != *ps2++)
937 				return (1);
938 		} while (--len != 0 && !(*flags & CPU_DTRACE_FAULT));
939 	}
940 	return (0);
941 }
942 
943 /*
944  * Zero the specified region using a simple byte-by-byte loop.  Note that this
945  * is for safe DTrace-managed memory only.
946  */
947 static void
948 dtrace_bzero(void *dst, size_t len)
949 {
950 	uchar_t *cp;
951 
952 	for (cp = dst; len != 0; len--)
953 		*cp++ = 0;
954 }
955 
956 static void
957 dtrace_add_128(uint64_t *addend1, uint64_t *addend2, uint64_t *sum)
958 {
959 	uint64_t result[2];
960 
961 	result[0] = addend1[0] + addend2[0];
962 	result[1] = addend1[1] + addend2[1] +
963 	    (result[0] < addend1[0] || result[0] < addend2[0] ? 1 : 0);
964 
965 	sum[0] = result[0];
966 	sum[1] = result[1];
967 }
968 
969 /*
970  * Shift the 128-bit value in a by b. If b is positive, shift left.
971  * If b is negative, shift right.
972  */
973 static void
974 dtrace_shift_128(uint64_t *a, int b)
975 {
976 	uint64_t mask;
977 
978 	if (b == 0)
979 		return;
980 
981 	if (b < 0) {
982 		b = -b;
983 		if (b >= 64) {
984 			a[0] = a[1] >> (b - 64);
985 			a[1] = 0;
986 		} else {
987 			a[0] >>= b;
988 			mask = 1LL << (64 - b);
989 			mask -= 1;
990 			a[0] |= ((a[1] & mask) << (64 - b));
991 			a[1] >>= b;
992 		}
993 	} else {
994 		if (b >= 64) {
995 			a[1] = a[0] << (b - 64);
996 			a[0] = 0;
997 		} else {
998 			a[1] <<= b;
999 			mask = a[0] >> (64 - b);
1000 			a[1] |= mask;
1001 			a[0] <<= b;
1002 		}
1003 	}
1004 }
1005 
1006 /*
1007  * The basic idea is to break the 2 64-bit values into 4 32-bit values,
1008  * use native multiplication on those, and then re-combine into the
1009  * resulting 128-bit value.
1010  *
1011  * (hi1 << 32 + lo1) * (hi2 << 32 + lo2) =
1012  *     hi1 * hi2 << 64 +
1013  *     hi1 * lo2 << 32 +
1014  *     hi2 * lo1 << 32 +
1015  *     lo1 * lo2
1016  */
1017 static void
1018 dtrace_multiply_128(uint64_t factor1, uint64_t factor2, uint64_t *product)
1019 {
1020 	uint64_t hi1, hi2, lo1, lo2;
1021 	uint64_t tmp[2];
1022 
1023 	hi1 = factor1 >> 32;
1024 	hi2 = factor2 >> 32;
1025 
1026 	lo1 = factor1 & DT_MASK_LO;
1027 	lo2 = factor2 & DT_MASK_LO;
1028 
1029 	product[0] = lo1 * lo2;
1030 	product[1] = hi1 * hi2;
1031 
1032 	tmp[0] = hi1 * lo2;
1033 	tmp[1] = 0;
1034 	dtrace_shift_128(tmp, 32);
1035 	dtrace_add_128(product, tmp, product);
1036 
1037 	tmp[0] = hi2 * lo1;
1038 	tmp[1] = 0;
1039 	dtrace_shift_128(tmp, 32);
1040 	dtrace_add_128(product, tmp, product);
1041 }
1042 
1043 /*
1044  * This privilege check should be used by actions and subroutines to
1045  * verify that the user credentials of the process that enabled the
1046  * invoking ECB match the target credentials
1047  */
1048 static int
1049 dtrace_priv_proc_common_user(dtrace_state_t *state)
1050 {
1051 	cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1052 
1053 	/*
1054 	 * We should always have a non-NULL state cred here, since if cred
1055 	 * is null (anonymous tracing), we fast-path bypass this routine.
1056 	 */
1057 	ASSERT(s_cr != NULL);
1058 
1059 	if ((cr = CRED()) != NULL &&
1060 	    s_cr->cr_uid == cr->cr_uid &&
1061 	    s_cr->cr_uid == cr->cr_ruid &&
1062 	    s_cr->cr_uid == cr->cr_suid &&
1063 	    s_cr->cr_gid == cr->cr_gid &&
1064 	    s_cr->cr_gid == cr->cr_rgid &&
1065 	    s_cr->cr_gid == cr->cr_sgid)
1066 		return (1);
1067 
1068 	return (0);
1069 }
1070 
1071 /*
1072  * This privilege check should be used by actions and subroutines to
1073  * verify that the zone of the process that enabled the invoking ECB
1074  * matches the target credentials
1075  */
1076 static int
1077 dtrace_priv_proc_common_zone(dtrace_state_t *state)
1078 {
1079 	cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1080 
1081 	/*
1082 	 * We should always have a non-NULL state cred here, since if cred
1083 	 * is null (anonymous tracing), we fast-path bypass this routine.
1084 	 */
1085 	ASSERT(s_cr != NULL);
1086 
1087 	if ((cr = CRED()) != NULL &&
1088 	    s_cr->cr_zone == cr->cr_zone)
1089 		return (1);
1090 
1091 	return (0);
1092 }
1093 
1094 /*
1095  * This privilege check should be used by actions and subroutines to
1096  * verify that the process has not setuid or changed credentials.
1097  */
1098 static int
1099 dtrace_priv_proc_common_nocd()
1100 {
1101 	proc_t *proc;
1102 
1103 	if ((proc = ttoproc(curthread)) != NULL &&
1104 	    !(proc->p_flag & SNOCD))
1105 		return (1);
1106 
1107 	return (0);
1108 }
1109 
1110 static int
1111 dtrace_priv_proc_destructive(dtrace_state_t *state, dtrace_mstate_t *mstate)
1112 {
1113 	int action = state->dts_cred.dcr_action;
1114 
1115 	if (!(mstate->dtms_access & DTRACE_ACCESS_PROC))
1116 		goto bad;
1117 
1118 	if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE) == 0) &&
1119 	    dtrace_priv_proc_common_zone(state) == 0)
1120 		goto bad;
1121 
1122 	if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER) == 0) &&
1123 	    dtrace_priv_proc_common_user(state) == 0)
1124 		goto bad;
1125 
1126 	if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG) == 0) &&
1127 	    dtrace_priv_proc_common_nocd() == 0)
1128 		goto bad;
1129 
1130 	return (1);
1131 
1132 bad:
1133 	cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1134 
1135 	return (0);
1136 }
1137 
1138 static int
1139 dtrace_priv_proc_control(dtrace_state_t *state, dtrace_mstate_t *mstate)
1140 {
1141 	if (mstate->dtms_access & DTRACE_ACCESS_PROC) {
1142 		if (state->dts_cred.dcr_action & DTRACE_CRA_PROC_CONTROL)
1143 			return (1);
1144 
1145 		if (dtrace_priv_proc_common_zone(state) &&
1146 		    dtrace_priv_proc_common_user(state) &&
1147 		    dtrace_priv_proc_common_nocd())
1148 			return (1);
1149 	}
1150 
1151 	cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1152 
1153 	return (0);
1154 }
1155 
1156 static int
1157 dtrace_priv_proc(dtrace_state_t *state, dtrace_mstate_t *mstate)
1158 {
1159 	if ((mstate->dtms_access & DTRACE_ACCESS_PROC) &&
1160 	    (state->dts_cred.dcr_action & DTRACE_CRA_PROC))
1161 		return (1);
1162 
1163 	cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1164 
1165 	return (0);
1166 }
1167 
1168 static int
1169 dtrace_priv_kernel(dtrace_state_t *state)
1170 {
1171 	if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL)
1172 		return (1);
1173 
1174 	cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1175 
1176 	return (0);
1177 }
1178 
1179 static int
1180 dtrace_priv_kernel_destructive(dtrace_state_t *state)
1181 {
1182 	if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL_DESTRUCTIVE)
1183 		return (1);
1184 
1185 	cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1186 
1187 	return (0);
1188 }
1189 
1190 /*
1191  * Determine if the dte_cond of the specified ECB allows for processing of
1192  * the current probe to continue.  Note that this routine may allow continued
1193  * processing, but with access(es) stripped from the mstate's dtms_access
1194  * field.
1195  */
1196 static int
1197 dtrace_priv_probe(dtrace_state_t *state, dtrace_mstate_t *mstate,
1198     dtrace_ecb_t *ecb)
1199 {
1200 	dtrace_probe_t *probe = ecb->dte_probe;
1201 	dtrace_provider_t *prov = probe->dtpr_provider;
1202 	dtrace_pops_t *pops = &prov->dtpv_pops;
1203 	int mode = DTRACE_MODE_NOPRIV_DROP;
1204 
1205 	ASSERT(ecb->dte_cond);
1206 
1207 	if (pops->dtps_mode != NULL) {
1208 		mode = pops->dtps_mode(prov->dtpv_arg,
1209 		    probe->dtpr_id, probe->dtpr_arg);
1210 
1211 		ASSERT((mode & DTRACE_MODE_USER) ||
1212 		    (mode & DTRACE_MODE_KERNEL));
1213 		ASSERT((mode & DTRACE_MODE_NOPRIV_RESTRICT) ||
1214 		    (mode & DTRACE_MODE_NOPRIV_DROP));
1215 	}
1216 
1217 	/*
1218 	 * If the dte_cond bits indicate that this consumer is only allowed to
1219 	 * see user-mode firings of this probe, call the provider's dtps_mode()
1220 	 * entry point to check that the probe was fired while in a user
1221 	 * context.  If that's not the case, use the policy specified by the
1222 	 * provider to determine if we drop the probe or merely restrict
1223 	 * operation.
1224 	 */
1225 	if (ecb->dte_cond & DTRACE_COND_USERMODE) {
1226 		ASSERT(mode != DTRACE_MODE_NOPRIV_DROP);
1227 
1228 		if (!(mode & DTRACE_MODE_USER)) {
1229 			if (mode & DTRACE_MODE_NOPRIV_DROP)
1230 				return (0);
1231 
1232 			mstate->dtms_access &= ~DTRACE_ACCESS_ARGS;
1233 		}
1234 	}
1235 
1236 	/*
1237 	 * This is more subtle than it looks. We have to be absolutely certain
1238 	 * that CRED() isn't going to change out from under us so it's only
1239 	 * legit to examine that structure if we're in constrained situations.
1240 	 * Currently, the only times we'll this check is if a non-super-user
1241 	 * has enabled the profile or syscall providers -- providers that
1242 	 * allow visibility of all processes. For the profile case, the check
1243 	 * above will ensure that we're examining a user context.
1244 	 */
1245 	if (ecb->dte_cond & DTRACE_COND_OWNER) {
1246 		cred_t *cr;
1247 		cred_t *s_cr = state->dts_cred.dcr_cred;
1248 		proc_t *proc;
1249 
1250 		ASSERT(s_cr != NULL);
1251 
1252 		if ((cr = CRED()) == NULL ||
1253 		    s_cr->cr_uid != cr->cr_uid ||
1254 		    s_cr->cr_uid != cr->cr_ruid ||
1255 		    s_cr->cr_uid != cr->cr_suid ||
1256 		    s_cr->cr_gid != cr->cr_gid ||
1257 		    s_cr->cr_gid != cr->cr_rgid ||
1258 		    s_cr->cr_gid != cr->cr_sgid ||
1259 		    (proc = ttoproc(curthread)) == NULL ||
1260 		    (proc->p_flag & SNOCD)) {
1261 			if (mode & DTRACE_MODE_NOPRIV_DROP)
1262 				return (0);
1263 
1264 			mstate->dtms_access &= ~DTRACE_ACCESS_PROC;
1265 		}
1266 	}
1267 
1268 	/*
1269 	 * If our dte_cond is set to DTRACE_COND_ZONEOWNER and we are not
1270 	 * in our zone, check to see if our mode policy is to restrict rather
1271 	 * than to drop; if to restrict, strip away both DTRACE_ACCESS_PROC
1272 	 * and DTRACE_ACCESS_ARGS
1273 	 */
1274 	if (ecb->dte_cond & DTRACE_COND_ZONEOWNER) {
1275 		cred_t *cr;
1276 		cred_t *s_cr = state->dts_cred.dcr_cred;
1277 
1278 		ASSERT(s_cr != NULL);
1279 
1280 		if ((cr = CRED()) == NULL ||
1281 		    s_cr->cr_zone->zone_id != cr->cr_zone->zone_id) {
1282 			if (mode & DTRACE_MODE_NOPRIV_DROP)
1283 				return (0);
1284 
1285 			mstate->dtms_access &=
1286 			    ~(DTRACE_ACCESS_PROC | DTRACE_ACCESS_ARGS);
1287 		}
1288 	}
1289 
1290 	return (1);
1291 }
1292 
1293 /*
1294  * Note:  not called from probe context.  This function is called
1295  * asynchronously (and at a regular interval) from outside of probe context to
1296  * clean the dirty dynamic variable lists on all CPUs.  Dynamic variable
1297  * cleaning is explained in detail in <sys/dtrace_impl.h>.
1298  */
1299 void
1300 dtrace_dynvar_clean(dtrace_dstate_t *dstate)
1301 {
1302 	dtrace_dynvar_t *dirty;
1303 	dtrace_dstate_percpu_t *dcpu;
1304 	dtrace_dynvar_t **rinsep;
1305 	int i, j, work = 0;
1306 
1307 	for (i = 0; i < NCPU; i++) {
1308 		dcpu = &dstate->dtds_percpu[i];
1309 		rinsep = &dcpu->dtdsc_rinsing;
1310 
1311 		/*
1312 		 * If the dirty list is NULL, there is no dirty work to do.
1313 		 */
1314 		if (dcpu->dtdsc_dirty == NULL)
1315 			continue;
1316 
1317 		if (dcpu->dtdsc_rinsing != NULL) {
1318 			/*
1319 			 * If the rinsing list is non-NULL, then it is because
1320 			 * this CPU was selected to accept another CPU's
1321 			 * dirty list -- and since that time, dirty buffers
1322 			 * have accumulated.  This is a highly unlikely
1323 			 * condition, but we choose to ignore the dirty
1324 			 * buffers -- they'll be picked up a future cleanse.
1325 			 */
1326 			continue;
1327 		}
1328 
1329 		if (dcpu->dtdsc_clean != NULL) {
1330 			/*
1331 			 * If the clean list is non-NULL, then we're in a
1332 			 * situation where a CPU has done deallocations (we
1333 			 * have a non-NULL dirty list) but no allocations (we
1334 			 * also have a non-NULL clean list).  We can't simply
1335 			 * move the dirty list into the clean list on this
1336 			 * CPU, yet we also don't want to allow this condition
1337 			 * to persist, lest a short clean list prevent a
1338 			 * massive dirty list from being cleaned (which in
1339 			 * turn could lead to otherwise avoidable dynamic
1340 			 * drops).  To deal with this, we look for some CPU
1341 			 * with a NULL clean list, NULL dirty list, and NULL
1342 			 * rinsing list -- and then we borrow this CPU to
1343 			 * rinse our dirty list.
1344 			 */
1345 			for (j = 0; j < NCPU; j++) {
1346 				dtrace_dstate_percpu_t *rinser;
1347 
1348 				rinser = &dstate->dtds_percpu[j];
1349 
1350 				if (rinser->dtdsc_rinsing != NULL)
1351 					continue;
1352 
1353 				if (rinser->dtdsc_dirty != NULL)
1354 					continue;
1355 
1356 				if (rinser->dtdsc_clean != NULL)
1357 					continue;
1358 
1359 				rinsep = &rinser->dtdsc_rinsing;
1360 				break;
1361 			}
1362 
1363 			if (j == NCPU) {
1364 				/*
1365 				 * We were unable to find another CPU that
1366 				 * could accept this dirty list -- we are
1367 				 * therefore unable to clean it now.
1368 				 */
1369 				dtrace_dynvar_failclean++;
1370 				continue;
1371 			}
1372 		}
1373 
1374 		work = 1;
1375 
1376 		/*
1377 		 * Atomically move the dirty list aside.
1378 		 */
1379 		do {
1380 			dirty = dcpu->dtdsc_dirty;
1381 
1382 			/*
1383 			 * Before we zap the dirty list, set the rinsing list.
1384 			 * (This allows for a potential assertion in
1385 			 * dtrace_dynvar():  if a free dynamic variable appears
1386 			 * on a hash chain, either the dirty list or the
1387 			 * rinsing list for some CPU must be non-NULL.)
1388 			 */
1389 			*rinsep = dirty;
1390 			dtrace_membar_producer();
1391 		} while (dtrace_casptr(&dcpu->dtdsc_dirty,
1392 		    dirty, NULL) != dirty);
1393 	}
1394 
1395 	if (!work) {
1396 		/*
1397 		 * We have no work to do; we can simply return.
1398 		 */
1399 		return;
1400 	}
1401 
1402 	dtrace_sync();
1403 
1404 	for (i = 0; i < NCPU; i++) {
1405 		dcpu = &dstate->dtds_percpu[i];
1406 
1407 		if (dcpu->dtdsc_rinsing == NULL)
1408 			continue;
1409 
1410 		/*
1411 		 * We are now guaranteed that no hash chain contains a pointer
1412 		 * into this dirty list; we can make it clean.
1413 		 */
1414 		ASSERT(dcpu->dtdsc_clean == NULL);
1415 		dcpu->dtdsc_clean = dcpu->dtdsc_rinsing;
1416 		dcpu->dtdsc_rinsing = NULL;
1417 	}
1418 
1419 	/*
1420 	 * Before we actually set the state to be DTRACE_DSTATE_CLEAN, make
1421 	 * sure that all CPUs have seen all of the dtdsc_clean pointers.
1422 	 * This prevents a race whereby a CPU incorrectly decides that
1423 	 * the state should be something other than DTRACE_DSTATE_CLEAN
1424 	 * after dtrace_dynvar_clean() has completed.
1425 	 */
1426 	dtrace_sync();
1427 
1428 	dstate->dtds_state = DTRACE_DSTATE_CLEAN;
1429 }
1430 
1431 /*
1432  * Depending on the value of the op parameter, this function looks-up,
1433  * allocates or deallocates an arbitrarily-keyed dynamic variable.  If an
1434  * allocation is requested, this function will return a pointer to a
1435  * dtrace_dynvar_t corresponding to the allocated variable -- or NULL if no
1436  * variable can be allocated.  If NULL is returned, the appropriate counter
1437  * will be incremented.
1438  */
1439 dtrace_dynvar_t *
1440 dtrace_dynvar(dtrace_dstate_t *dstate, uint_t nkeys,
1441     dtrace_key_t *key, size_t dsize, dtrace_dynvar_op_t op,
1442     dtrace_mstate_t *mstate, dtrace_vstate_t *vstate)
1443 {
1444 	uint64_t hashval = DTRACE_DYNHASH_VALID;
1445 	dtrace_dynhash_t *hash = dstate->dtds_hash;
1446 	dtrace_dynvar_t *free, *new_free, *next, *dvar, *start, *prev = NULL;
1447 	processorid_t me = CPU->cpu_id, cpu = me;
1448 	dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[me];
1449 	size_t bucket, ksize;
1450 	size_t chunksize = dstate->dtds_chunksize;
1451 	uintptr_t kdata, lock, nstate;
1452 	uint_t i;
1453 
1454 	ASSERT(nkeys != 0);
1455 
1456 	/*
1457 	 * Hash the key.  As with aggregations, we use Jenkins' "One-at-a-time"
1458 	 * algorithm.  For the by-value portions, we perform the algorithm in
1459 	 * 16-bit chunks (as opposed to 8-bit chunks).  This speeds things up a
1460 	 * bit, and seems to have only a minute effect on distribution.  For
1461 	 * the by-reference data, we perform "One-at-a-time" iterating (safely)
1462 	 * over each referenced byte.  It's painful to do this, but it's much
1463 	 * better than pathological hash distribution.  The efficacy of the
1464 	 * hashing algorithm (and a comparison with other algorithms) may be
1465 	 * found by running the ::dtrace_dynstat MDB dcmd.
1466 	 */
1467 	for (i = 0; i < nkeys; i++) {
1468 		if (key[i].dttk_size == 0) {
1469 			uint64_t val = key[i].dttk_value;
1470 
1471 			hashval += (val >> 48) & 0xffff;
1472 			hashval += (hashval << 10);
1473 			hashval ^= (hashval >> 6);
1474 
1475 			hashval += (val >> 32) & 0xffff;
1476 			hashval += (hashval << 10);
1477 			hashval ^= (hashval >> 6);
1478 
1479 			hashval += (val >> 16) & 0xffff;
1480 			hashval += (hashval << 10);
1481 			hashval ^= (hashval >> 6);
1482 
1483 			hashval += val & 0xffff;
1484 			hashval += (hashval << 10);
1485 			hashval ^= (hashval >> 6);
1486 		} else {
1487 			/*
1488 			 * This is incredibly painful, but it beats the hell
1489 			 * out of the alternative.
1490 			 */
1491 			uint64_t j, size = key[i].dttk_size;
1492 			uintptr_t base = (uintptr_t)key[i].dttk_value;
1493 
1494 			if (!dtrace_canload(base, size, mstate, vstate))
1495 				break;
1496 
1497 			for (j = 0; j < size; j++) {
1498 				hashval += dtrace_load8(base + j);
1499 				hashval += (hashval << 10);
1500 				hashval ^= (hashval >> 6);
1501 			}
1502 		}
1503 	}
1504 
1505 	if (DTRACE_CPUFLAG_ISSET(CPU_DTRACE_FAULT))
1506 		return (NULL);
1507 
1508 	hashval += (hashval << 3);
1509 	hashval ^= (hashval >> 11);
1510 	hashval += (hashval << 15);
1511 
1512 	/*
1513 	 * There is a remote chance (ideally, 1 in 2^31) that our hashval
1514 	 * comes out to be one of our two sentinel hash values.  If this
1515 	 * actually happens, we set the hashval to be a value known to be a
1516 	 * non-sentinel value.
1517 	 */
1518 	if (hashval == DTRACE_DYNHASH_FREE || hashval == DTRACE_DYNHASH_SINK)
1519 		hashval = DTRACE_DYNHASH_VALID;
1520 
1521 	/*
1522 	 * Yes, it's painful to do a divide here.  If the cycle count becomes
1523 	 * important here, tricks can be pulled to reduce it.  (However, it's
1524 	 * critical that hash collisions be kept to an absolute minimum;
1525 	 * they're much more painful than a divide.)  It's better to have a
1526 	 * solution that generates few collisions and still keeps things
1527 	 * relatively simple.
1528 	 */
1529 	bucket = hashval % dstate->dtds_hashsize;
1530 
1531 	if (op == DTRACE_DYNVAR_DEALLOC) {
1532 		volatile uintptr_t *lockp = &hash[bucket].dtdh_lock;
1533 
1534 		for (;;) {
1535 			while ((lock = *lockp) & 1)
1536 				continue;
1537 
1538 			if (dtrace_casptr((void *)lockp,
1539 			    (void *)lock, (void *)(lock + 1)) == (void *)lock)
1540 				break;
1541 		}
1542 
1543 		dtrace_membar_producer();
1544 	}
1545 
1546 top:
1547 	prev = NULL;
1548 	lock = hash[bucket].dtdh_lock;
1549 
1550 	dtrace_membar_consumer();
1551 
1552 	start = hash[bucket].dtdh_chain;
1553 	ASSERT(start != NULL && (start->dtdv_hashval == DTRACE_DYNHASH_SINK ||
1554 	    start->dtdv_hashval != DTRACE_DYNHASH_FREE ||
1555 	    op != DTRACE_DYNVAR_DEALLOC));
1556 
1557 	for (dvar = start; dvar != NULL; dvar = dvar->dtdv_next) {
1558 		dtrace_tuple_t *dtuple = &dvar->dtdv_tuple;
1559 		dtrace_key_t *dkey = &dtuple->dtt_key[0];
1560 
1561 		if (dvar->dtdv_hashval != hashval) {
1562 			if (dvar->dtdv_hashval == DTRACE_DYNHASH_SINK) {
1563 				/*
1564 				 * We've reached the sink, and therefore the
1565 				 * end of the hash chain; we can kick out of
1566 				 * the loop knowing that we have seen a valid
1567 				 * snapshot of state.
1568 				 */
1569 				ASSERT(dvar->dtdv_next == NULL);
1570 				ASSERT(dvar == &dtrace_dynhash_sink);
1571 				break;
1572 			}
1573 
1574 			if (dvar->dtdv_hashval == DTRACE_DYNHASH_FREE) {
1575 				/*
1576 				 * We've gone off the rails:  somewhere along
1577 				 * the line, one of the members of this hash
1578 				 * chain was deleted.  Note that we could also
1579 				 * detect this by simply letting this loop run
1580 				 * to completion, as we would eventually hit
1581 				 * the end of the dirty list.  However, we
1582 				 * want to avoid running the length of the
1583 				 * dirty list unnecessarily (it might be quite
1584 				 * long), so we catch this as early as
1585 				 * possible by detecting the hash marker.  In
1586 				 * this case, we simply set dvar to NULL and
1587 				 * break; the conditional after the loop will
1588 				 * send us back to top.
1589 				 */
1590 				dvar = NULL;
1591 				break;
1592 			}
1593 
1594 			goto next;
1595 		}
1596 
1597 		if (dtuple->dtt_nkeys != nkeys)
1598 			goto next;
1599 
1600 		for (i = 0; i < nkeys; i++, dkey++) {
1601 			if (dkey->dttk_size != key[i].dttk_size)
1602 				goto next; /* size or type mismatch */
1603 
1604 			if (dkey->dttk_size != 0) {
1605 				if (dtrace_bcmp(
1606 				    (void *)(uintptr_t)key[i].dttk_value,
1607 				    (void *)(uintptr_t)dkey->dttk_value,
1608 				    dkey->dttk_size))
1609 					goto next;
1610 			} else {
1611 				if (dkey->dttk_value != key[i].dttk_value)
1612 					goto next;
1613 			}
1614 		}
1615 
1616 		if (op != DTRACE_DYNVAR_DEALLOC)
1617 			return (dvar);
1618 
1619 		ASSERT(dvar->dtdv_next == NULL ||
1620 		    dvar->dtdv_next->dtdv_hashval != DTRACE_DYNHASH_FREE);
1621 
1622 		if (prev != NULL) {
1623 			ASSERT(hash[bucket].dtdh_chain != dvar);
1624 			ASSERT(start != dvar);
1625 			ASSERT(prev->dtdv_next == dvar);
1626 			prev->dtdv_next = dvar->dtdv_next;
1627 		} else {
1628 			if (dtrace_casptr(&hash[bucket].dtdh_chain,
1629 			    start, dvar->dtdv_next) != start) {
1630 				/*
1631 				 * We have failed to atomically swing the
1632 				 * hash table head pointer, presumably because
1633 				 * of a conflicting allocation on another CPU.
1634 				 * We need to reread the hash chain and try
1635 				 * again.
1636 				 */
1637 				goto top;
1638 			}
1639 		}
1640 
1641 		dtrace_membar_producer();
1642 
1643 		/*
1644 		 * Now set the hash value to indicate that it's free.
1645 		 */
1646 		ASSERT(hash[bucket].dtdh_chain != dvar);
1647 		dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
1648 
1649 		dtrace_membar_producer();
1650 
1651 		/*
1652 		 * Set the next pointer to point at the dirty list, and
1653 		 * atomically swing the dirty pointer to the newly freed dvar.
1654 		 */
1655 		do {
1656 			next = dcpu->dtdsc_dirty;
1657 			dvar->dtdv_next = next;
1658 		} while (dtrace_casptr(&dcpu->dtdsc_dirty, next, dvar) != next);
1659 
1660 		/*
1661 		 * Finally, unlock this hash bucket.
1662 		 */
1663 		ASSERT(hash[bucket].dtdh_lock == lock);
1664 		ASSERT(lock & 1);
1665 		hash[bucket].dtdh_lock++;
1666 
1667 		return (NULL);
1668 next:
1669 		prev = dvar;
1670 		continue;
1671 	}
1672 
1673 	if (dvar == NULL) {
1674 		/*
1675 		 * If dvar is NULL, it is because we went off the rails:
1676 		 * one of the elements that we traversed in the hash chain
1677 		 * was deleted while we were traversing it.  In this case,
1678 		 * we assert that we aren't doing a dealloc (deallocs lock
1679 		 * the hash bucket to prevent themselves from racing with
1680 		 * one another), and retry the hash chain traversal.
1681 		 */
1682 		ASSERT(op != DTRACE_DYNVAR_DEALLOC);
1683 		goto top;
1684 	}
1685 
1686 	if (op != DTRACE_DYNVAR_ALLOC) {
1687 		/*
1688 		 * If we are not to allocate a new variable, we want to
1689 		 * return NULL now.  Before we return, check that the value
1690 		 * of the lock word hasn't changed.  If it has, we may have
1691 		 * seen an inconsistent snapshot.
1692 		 */
1693 		if (op == DTRACE_DYNVAR_NOALLOC) {
1694 			if (hash[bucket].dtdh_lock != lock)
1695 				goto top;
1696 		} else {
1697 			ASSERT(op == DTRACE_DYNVAR_DEALLOC);
1698 			ASSERT(hash[bucket].dtdh_lock == lock);
1699 			ASSERT(lock & 1);
1700 			hash[bucket].dtdh_lock++;
1701 		}
1702 
1703 		return (NULL);
1704 	}
1705 
1706 	/*
1707 	 * We need to allocate a new dynamic variable.  The size we need is the
1708 	 * size of dtrace_dynvar plus the size of nkeys dtrace_key_t's plus the
1709 	 * size of any auxiliary key data (rounded up to 8-byte alignment) plus
1710 	 * the size of any referred-to data (dsize).  We then round the final
1711 	 * size up to the chunksize for allocation.
1712 	 */
1713 	for (ksize = 0, i = 0; i < nkeys; i++)
1714 		ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
1715 
1716 	/*
1717 	 * This should be pretty much impossible, but could happen if, say,
1718 	 * strange DIF specified the tuple.  Ideally, this should be an
1719 	 * assertion and not an error condition -- but that requires that the
1720 	 * chunksize calculation in dtrace_difo_chunksize() be absolutely
1721 	 * bullet-proof.  (That is, it must not be able to be fooled by
1722 	 * malicious DIF.)  Given the lack of backwards branches in DIF,
1723 	 * solving this would presumably not amount to solving the Halting
1724 	 * Problem -- but it still seems awfully hard.
1725 	 */
1726 	if (sizeof (dtrace_dynvar_t) + sizeof (dtrace_key_t) * (nkeys - 1) +
1727 	    ksize + dsize > chunksize) {
1728 		dcpu->dtdsc_drops++;
1729 		return (NULL);
1730 	}
1731 
1732 	nstate = DTRACE_DSTATE_EMPTY;
1733 
1734 	do {
1735 retry:
1736 		free = dcpu->dtdsc_free;
1737 
1738 		if (free == NULL) {
1739 			dtrace_dynvar_t *clean = dcpu->dtdsc_clean;
1740 			void *rval;
1741 
1742 			if (clean == NULL) {
1743 				/*
1744 				 * We're out of dynamic variable space on
1745 				 * this CPU.  Unless we have tried all CPUs,
1746 				 * we'll try to allocate from a different
1747 				 * CPU.
1748 				 */
1749 				switch (dstate->dtds_state) {
1750 				case DTRACE_DSTATE_CLEAN: {
1751 					void *sp = &dstate->dtds_state;
1752 
1753 					if (++cpu >= NCPU)
1754 						cpu = 0;
1755 
1756 					if (dcpu->dtdsc_dirty != NULL &&
1757 					    nstate == DTRACE_DSTATE_EMPTY)
1758 						nstate = DTRACE_DSTATE_DIRTY;
1759 
1760 					if (dcpu->dtdsc_rinsing != NULL)
1761 						nstate = DTRACE_DSTATE_RINSING;
1762 
1763 					dcpu = &dstate->dtds_percpu[cpu];
1764 
1765 					if (cpu != me)
1766 						goto retry;
1767 
1768 					(void) dtrace_cas32(sp,
1769 					    DTRACE_DSTATE_CLEAN, nstate);
1770 
1771 					/*
1772 					 * To increment the correct bean
1773 					 * counter, take another lap.
1774 					 */
1775 					goto retry;
1776 				}
1777 
1778 				case DTRACE_DSTATE_DIRTY:
1779 					dcpu->dtdsc_dirty_drops++;
1780 					break;
1781 
1782 				case DTRACE_DSTATE_RINSING:
1783 					dcpu->dtdsc_rinsing_drops++;
1784 					break;
1785 
1786 				case DTRACE_DSTATE_EMPTY:
1787 					dcpu->dtdsc_drops++;
1788 					break;
1789 				}
1790 
1791 				DTRACE_CPUFLAG_SET(CPU_DTRACE_DROP);
1792 				return (NULL);
1793 			}
1794 
1795 			/*
1796 			 * The clean list appears to be non-empty.  We want to
1797 			 * move the clean list to the free list; we start by
1798 			 * moving the clean pointer aside.
1799 			 */
1800 			if (dtrace_casptr(&dcpu->dtdsc_clean,
1801 			    clean, NULL) != clean) {
1802 				/*
1803 				 * We are in one of two situations:
1804 				 *
1805 				 *  (a)	The clean list was switched to the
1806 				 *	free list by another CPU.
1807 				 *
1808 				 *  (b)	The clean list was added to by the
1809 				 *	cleansing cyclic.
1810 				 *
1811 				 * In either of these situations, we can
1812 				 * just reattempt the free list allocation.
1813 				 */
1814 				goto retry;
1815 			}
1816 
1817 			ASSERT(clean->dtdv_hashval == DTRACE_DYNHASH_FREE);
1818 
1819 			/*
1820 			 * Now we'll move the clean list to our free list.
1821 			 * It's impossible for this to fail:  the only way
1822 			 * the free list can be updated is through this
1823 			 * code path, and only one CPU can own the clean list.
1824 			 * Thus, it would only be possible for this to fail if
1825 			 * this code were racing with dtrace_dynvar_clean().
1826 			 * (That is, if dtrace_dynvar_clean() updated the clean
1827 			 * list, and we ended up racing to update the free
1828 			 * list.)  This race is prevented by the dtrace_sync()
1829 			 * in dtrace_dynvar_clean() -- which flushes the
1830 			 * owners of the clean lists out before resetting
1831 			 * the clean lists.
1832 			 */
1833 			dcpu = &dstate->dtds_percpu[me];
1834 			rval = dtrace_casptr(&dcpu->dtdsc_free, NULL, clean);
1835 			ASSERT(rval == NULL);
1836 			goto retry;
1837 		}
1838 
1839 		dvar = free;
1840 		new_free = dvar->dtdv_next;
1841 	} while (dtrace_casptr(&dcpu->dtdsc_free, free, new_free) != free);
1842 
1843 	/*
1844 	 * We have now allocated a new chunk.  We copy the tuple keys into the
1845 	 * tuple array and copy any referenced key data into the data space
1846 	 * following the tuple array.  As we do this, we relocate dttk_value
1847 	 * in the final tuple to point to the key data address in the chunk.
1848 	 */
1849 	kdata = (uintptr_t)&dvar->dtdv_tuple.dtt_key[nkeys];
1850 	dvar->dtdv_data = (void *)(kdata + ksize);
1851 	dvar->dtdv_tuple.dtt_nkeys = nkeys;
1852 
1853 	for (i = 0; i < nkeys; i++) {
1854 		dtrace_key_t *dkey = &dvar->dtdv_tuple.dtt_key[i];
1855 		size_t kesize = key[i].dttk_size;
1856 
1857 		if (kesize != 0) {
1858 			dtrace_bcopy(
1859 			    (const void *)(uintptr_t)key[i].dttk_value,
1860 			    (void *)kdata, kesize);
1861 			dkey->dttk_value = kdata;
1862 			kdata += P2ROUNDUP(kesize, sizeof (uint64_t));
1863 		} else {
1864 			dkey->dttk_value = key[i].dttk_value;
1865 		}
1866 
1867 		dkey->dttk_size = kesize;
1868 	}
1869 
1870 	ASSERT(dvar->dtdv_hashval == DTRACE_DYNHASH_FREE);
1871 	dvar->dtdv_hashval = hashval;
1872 	dvar->dtdv_next = start;
1873 
1874 	if (dtrace_casptr(&hash[bucket].dtdh_chain, start, dvar) == start)
1875 		return (dvar);
1876 
1877 	/*
1878 	 * The cas has failed.  Either another CPU is adding an element to
1879 	 * this hash chain, or another CPU is deleting an element from this
1880 	 * hash chain.  The simplest way to deal with both of these cases
1881 	 * (though not necessarily the most efficient) is to free our
1882 	 * allocated block and tail-call ourselves.  Note that the free is
1883 	 * to the dirty list and _not_ to the free list.  This is to prevent
1884 	 * races with allocators, above.
1885 	 */
1886 	dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
1887 
1888 	dtrace_membar_producer();
1889 
1890 	do {
1891 		free = dcpu->dtdsc_dirty;
1892 		dvar->dtdv_next = free;
1893 	} while (dtrace_casptr(&dcpu->dtdsc_dirty, free, dvar) != free);
1894 
1895 	return (dtrace_dynvar(dstate, nkeys, key, dsize, op, mstate, vstate));
1896 }
1897 
1898 /*ARGSUSED*/
1899 static void
1900 dtrace_aggregate_min(uint64_t *oval, uint64_t nval, uint64_t arg)
1901 {
1902 	if ((int64_t)nval < (int64_t)*oval)
1903 		*oval = nval;
1904 }
1905 
1906 /*ARGSUSED*/
1907 static void
1908 dtrace_aggregate_max(uint64_t *oval, uint64_t nval, uint64_t arg)
1909 {
1910 	if ((int64_t)nval > (int64_t)*oval)
1911 		*oval = nval;
1912 }
1913 
1914 static void
1915 dtrace_aggregate_quantize(uint64_t *quanta, uint64_t nval, uint64_t incr)
1916 {
1917 	int i, zero = DTRACE_QUANTIZE_ZEROBUCKET;
1918 	int64_t val = (int64_t)nval;
1919 
1920 	if (val < 0) {
1921 		for (i = 0; i < zero; i++) {
1922 			if (val <= DTRACE_QUANTIZE_BUCKETVAL(i)) {
1923 				quanta[i] += incr;
1924 				return;
1925 			}
1926 		}
1927 	} else {
1928 		for (i = zero + 1; i < DTRACE_QUANTIZE_NBUCKETS; i++) {
1929 			if (val < DTRACE_QUANTIZE_BUCKETVAL(i)) {
1930 				quanta[i - 1] += incr;
1931 				return;
1932 			}
1933 		}
1934 
1935 		quanta[DTRACE_QUANTIZE_NBUCKETS - 1] += incr;
1936 		return;
1937 	}
1938 
1939 	ASSERT(0);
1940 }
1941 
1942 static void
1943 dtrace_aggregate_lquantize(uint64_t *lquanta, uint64_t nval, uint64_t incr)
1944 {
1945 	uint64_t arg = *lquanta++;
1946 	int32_t base = DTRACE_LQUANTIZE_BASE(arg);
1947 	uint16_t step = DTRACE_LQUANTIZE_STEP(arg);
1948 	uint16_t levels = DTRACE_LQUANTIZE_LEVELS(arg);
1949 	int32_t val = (int32_t)nval, level;
1950 
1951 	ASSERT(step != 0);
1952 	ASSERT(levels != 0);
1953 
1954 	if (val < base) {
1955 		/*
1956 		 * This is an underflow.
1957 		 */
1958 		lquanta[0] += incr;
1959 		return;
1960 	}
1961 
1962 	level = (val - base) / step;
1963 
1964 	if (level < levels) {
1965 		lquanta[level + 1] += incr;
1966 		return;
1967 	}
1968 
1969 	/*
1970 	 * This is an overflow.
1971 	 */
1972 	lquanta[levels + 1] += incr;
1973 }
1974 
1975 static int
1976 dtrace_aggregate_llquantize_bucket(uint16_t factor, uint16_t low,
1977     uint16_t high, uint16_t nsteps, int64_t value)
1978 {
1979 	int64_t this = 1, last, next;
1980 	int base = 1, order;
1981 
1982 	ASSERT(factor <= nsteps);
1983 	ASSERT(nsteps % factor == 0);
1984 
1985 	for (order = 0; order < low; order++)
1986 		this *= factor;
1987 
1988 	/*
1989 	 * If our value is less than our factor taken to the power of the
1990 	 * low order of magnitude, it goes into the zeroth bucket.
1991 	 */
1992 	if (value < (last = this))
1993 		return (0);
1994 
1995 	for (this *= factor; order <= high; order++) {
1996 		int nbuckets = this > nsteps ? nsteps : this;
1997 
1998 		if ((next = this * factor) < this) {
1999 			/*
2000 			 * We should not generally get log/linear quantizations
2001 			 * with a high magnitude that allows 64-bits to
2002 			 * overflow, but we nonetheless protect against this
2003 			 * by explicitly checking for overflow, and clamping
2004 			 * our value accordingly.
2005 			 */
2006 			value = this - 1;
2007 		}
2008 
2009 		if (value < this) {
2010 			/*
2011 			 * If our value lies within this order of magnitude,
2012 			 * determine its position by taking the offset within
2013 			 * the order of magnitude, dividing by the bucket
2014 			 * width, and adding to our (accumulated) base.
2015 			 */
2016 			return (base + (value - last) / (this / nbuckets));
2017 		}
2018 
2019 		base += nbuckets - (nbuckets / factor);
2020 		last = this;
2021 		this = next;
2022 	}
2023 
2024 	/*
2025 	 * Our value is greater than or equal to our factor taken to the
2026 	 * power of one plus the high magnitude -- return the top bucket.
2027 	 */
2028 	return (base);
2029 }
2030 
2031 static void
2032 dtrace_aggregate_llquantize(uint64_t *llquanta, uint64_t nval, uint64_t incr)
2033 {
2034 	uint64_t arg = *llquanta++;
2035 	uint16_t factor = DTRACE_LLQUANTIZE_FACTOR(arg);
2036 	uint16_t low = DTRACE_LLQUANTIZE_LOW(arg);
2037 	uint16_t high = DTRACE_LLQUANTIZE_HIGH(arg);
2038 	uint16_t nsteps = DTRACE_LLQUANTIZE_NSTEP(arg);
2039 
2040 	llquanta[dtrace_aggregate_llquantize_bucket(factor,
2041 	    low, high, nsteps, nval)] += incr;
2042 }
2043 
2044 /*ARGSUSED*/
2045 static void
2046 dtrace_aggregate_avg(uint64_t *data, uint64_t nval, uint64_t arg)
2047 {
2048 	data[0]++;
2049 	data[1] += nval;
2050 }
2051 
2052 /*ARGSUSED*/
2053 static void
2054 dtrace_aggregate_stddev(uint64_t *data, uint64_t nval, uint64_t arg)
2055 {
2056 	int64_t snval = (int64_t)nval;
2057 	uint64_t tmp[2];
2058 
2059 	data[0]++;
2060 	data[1] += nval;
2061 
2062 	/*
2063 	 * What we want to say here is:
2064 	 *
2065 	 * data[2] += nval * nval;
2066 	 *
2067 	 * But given that nval is 64-bit, we could easily overflow, so
2068 	 * we do this as 128-bit arithmetic.
2069 	 */
2070 	if (snval < 0)
2071 		snval = -snval;
2072 
2073 	dtrace_multiply_128((uint64_t)snval, (uint64_t)snval, tmp);
2074 	dtrace_add_128(data + 2, tmp, data + 2);
2075 }
2076 
2077 /*ARGSUSED*/
2078 static void
2079 dtrace_aggregate_count(uint64_t *oval, uint64_t nval, uint64_t arg)
2080 {
2081 	*oval = *oval + 1;
2082 }
2083 
2084 /*ARGSUSED*/
2085 static void
2086 dtrace_aggregate_sum(uint64_t *oval, uint64_t nval, uint64_t arg)
2087 {
2088 	*oval += nval;
2089 }
2090 
2091 /*
2092  * Aggregate given the tuple in the principal data buffer, and the aggregating
2093  * action denoted by the specified dtrace_aggregation_t.  The aggregation
2094  * buffer is specified as the buf parameter.  This routine does not return
2095  * failure; if there is no space in the aggregation buffer, the data will be
2096  * dropped, and a corresponding counter incremented.
2097  */
2098 static void
2099 dtrace_aggregate(dtrace_aggregation_t *agg, dtrace_buffer_t *dbuf,
2100     intptr_t offset, dtrace_buffer_t *buf, uint64_t expr, uint64_t arg)
2101 {
2102 	dtrace_recdesc_t *rec = &agg->dtag_action.dta_rec;
2103 	uint32_t i, ndx, size, fsize;
2104 	uint32_t align = sizeof (uint64_t) - 1;
2105 	dtrace_aggbuffer_t *agb;
2106 	dtrace_aggkey_t *key;
2107 	uint32_t hashval = 0, limit, isstr;
2108 	caddr_t tomax, data, kdata;
2109 	dtrace_actkind_t action;
2110 	dtrace_action_t *act;
2111 	uintptr_t offs;
2112 
2113 	if (buf == NULL)
2114 		return;
2115 
2116 	if (!agg->dtag_hasarg) {
2117 		/*
2118 		 * Currently, only quantize() and lquantize() take additional
2119 		 * arguments, and they have the same semantics:  an increment
2120 		 * value that defaults to 1 when not present.  If additional
2121 		 * aggregating actions take arguments, the setting of the
2122 		 * default argument value will presumably have to become more
2123 		 * sophisticated...
2124 		 */
2125 		arg = 1;
2126 	}
2127 
2128 	action = agg->dtag_action.dta_kind - DTRACEACT_AGGREGATION;
2129 	size = rec->dtrd_offset - agg->dtag_base;
2130 	fsize = size + rec->dtrd_size;
2131 
2132 	ASSERT(dbuf->dtb_tomax != NULL);
2133 	data = dbuf->dtb_tomax + offset + agg->dtag_base;
2134 
2135 	if ((tomax = buf->dtb_tomax) == NULL) {
2136 		dtrace_buffer_drop(buf);
2137 		return;
2138 	}
2139 
2140 	/*
2141 	 * The metastructure is always at the bottom of the buffer.
2142 	 */
2143 	agb = (dtrace_aggbuffer_t *)(tomax + buf->dtb_size -
2144 	    sizeof (dtrace_aggbuffer_t));
2145 
2146 	if (buf->dtb_offset == 0) {
2147 		/*
2148 		 * We just kludge up approximately 1/8th of the size to be
2149 		 * buckets.  If this guess ends up being routinely
2150 		 * off-the-mark, we may need to dynamically readjust this
2151 		 * based on past performance.
2152 		 */
2153 		uintptr_t hashsize = (buf->dtb_size >> 3) / sizeof (uintptr_t);
2154 
2155 		if ((uintptr_t)agb - hashsize * sizeof (dtrace_aggkey_t *) <
2156 		    (uintptr_t)tomax || hashsize == 0) {
2157 			/*
2158 			 * We've been given a ludicrously small buffer;
2159 			 * increment our drop count and leave.
2160 			 */
2161 			dtrace_buffer_drop(buf);
2162 			return;
2163 		}
2164 
2165 		/*
2166 		 * And now, a pathetic attempt to try to get a an odd (or
2167 		 * perchance, a prime) hash size for better hash distribution.
2168 		 */
2169 		if (hashsize > (DTRACE_AGGHASHSIZE_SLEW << 3))
2170 			hashsize -= DTRACE_AGGHASHSIZE_SLEW;
2171 
2172 		agb->dtagb_hashsize = hashsize;
2173 		agb->dtagb_hash = (dtrace_aggkey_t **)((uintptr_t)agb -
2174 		    agb->dtagb_hashsize * sizeof (dtrace_aggkey_t *));
2175 		agb->dtagb_free = (uintptr_t)agb->dtagb_hash;
2176 
2177 		for (i = 0; i < agb->dtagb_hashsize; i++)
2178 			agb->dtagb_hash[i] = NULL;
2179 	}
2180 
2181 	ASSERT(agg->dtag_first != NULL);
2182 	ASSERT(agg->dtag_first->dta_intuple);
2183 
2184 	/*
2185 	 * Calculate the hash value based on the key.  Note that we _don't_
2186 	 * include the aggid in the hashing (but we will store it as part of
2187 	 * the key).  The hashing algorithm is Bob Jenkins' "One-at-a-time"
2188 	 * algorithm: a simple, quick algorithm that has no known funnels, and
2189 	 * gets good distribution in practice.  The efficacy of the hashing
2190 	 * algorithm (and a comparison with other algorithms) may be found by
2191 	 * running the ::dtrace_aggstat MDB dcmd.
2192 	 */
2193 	for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
2194 		i = act->dta_rec.dtrd_offset - agg->dtag_base;
2195 		limit = i + act->dta_rec.dtrd_size;
2196 		ASSERT(limit <= size);
2197 		isstr = DTRACEACT_ISSTRING(act);
2198 
2199 		for (; i < limit; i++) {
2200 			hashval += data[i];
2201 			hashval += (hashval << 10);
2202 			hashval ^= (hashval >> 6);
2203 
2204 			if (isstr && data[i] == '\0')
2205 				break;
2206 		}
2207 	}
2208 
2209 	hashval += (hashval << 3);
2210 	hashval ^= (hashval >> 11);
2211 	hashval += (hashval << 15);
2212 
2213 	/*
2214 	 * Yes, the divide here is expensive -- but it's generally the least
2215 	 * of the performance issues given the amount of data that we iterate
2216 	 * over to compute hash values, compare data, etc.
2217 	 */
2218 	ndx = hashval % agb->dtagb_hashsize;
2219 
2220 	for (key = agb->dtagb_hash[ndx]; key != NULL; key = key->dtak_next) {
2221 		ASSERT((caddr_t)key >= tomax);
2222 		ASSERT((caddr_t)key < tomax + buf->dtb_size);
2223 
2224 		if (hashval != key->dtak_hashval || key->dtak_size != size)
2225 			continue;
2226 
2227 		kdata = key->dtak_data;
2228 		ASSERT(kdata >= tomax && kdata < tomax + buf->dtb_size);
2229 
2230 		for (act = agg->dtag_first; act->dta_intuple;
2231 		    act = act->dta_next) {
2232 			i = act->dta_rec.dtrd_offset - agg->dtag_base;
2233 			limit = i + act->dta_rec.dtrd_size;
2234 			ASSERT(limit <= size);
2235 			isstr = DTRACEACT_ISSTRING(act);
2236 
2237 			for (; i < limit; i++) {
2238 				if (kdata[i] != data[i])
2239 					goto next;
2240 
2241 				if (isstr && data[i] == '\0')
2242 					break;
2243 			}
2244 		}
2245 
2246 		if (action != key->dtak_action) {
2247 			/*
2248 			 * We are aggregating on the same value in the same
2249 			 * aggregation with two different aggregating actions.
2250 			 * (This should have been picked up in the compiler,
2251 			 * so we may be dealing with errant or devious DIF.)
2252 			 * This is an error condition; we indicate as much,
2253 			 * and return.
2254 			 */
2255 			DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
2256 			return;
2257 		}
2258 
2259 		/*
2260 		 * This is a hit:  we need to apply the aggregator to
2261 		 * the value at this key.
2262 		 */
2263 		agg->dtag_aggregate((uint64_t *)(kdata + size), expr, arg);
2264 		return;
2265 next:
2266 		continue;
2267 	}
2268 
2269 	/*
2270 	 * We didn't find it.  We need to allocate some zero-filled space,
2271 	 * link it into the hash table appropriately, and apply the aggregator
2272 	 * to the (zero-filled) value.
2273 	 */
2274 	offs = buf->dtb_offset;
2275 	while (offs & (align - 1))
2276 		offs += sizeof (uint32_t);
2277 
2278 	/*
2279 	 * If we don't have enough room to both allocate a new key _and_
2280 	 * its associated data, increment the drop count and return.
2281 	 */
2282 	if ((uintptr_t)tomax + offs + fsize >
2283 	    agb->dtagb_free - sizeof (dtrace_aggkey_t)) {
2284 		dtrace_buffer_drop(buf);
2285 		return;
2286 	}
2287 
2288 	/*CONSTCOND*/
2289 	ASSERT(!(sizeof (dtrace_aggkey_t) & (sizeof (uintptr_t) - 1)));
2290 	key = (dtrace_aggkey_t *)(agb->dtagb_free - sizeof (dtrace_aggkey_t));
2291 	agb->dtagb_free -= sizeof (dtrace_aggkey_t);
2292 
2293 	key->dtak_data = kdata = tomax + offs;
2294 	buf->dtb_offset = offs + fsize;
2295 
2296 	/*
2297 	 * Now copy the data across.
2298 	 */
2299 	*((dtrace_aggid_t *)kdata) = agg->dtag_id;
2300 
2301 	for (i = sizeof (dtrace_aggid_t); i < size; i++)
2302 		kdata[i] = data[i];
2303 
2304 	/*
2305 	 * Because strings are not zeroed out by default, we need to iterate
2306 	 * looking for actions that store strings, and we need to explicitly
2307 	 * pad these strings out with zeroes.
2308 	 */
2309 	for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
2310 		int nul;
2311 
2312 		if (!DTRACEACT_ISSTRING(act))
2313 			continue;
2314 
2315 		i = act->dta_rec.dtrd_offset - agg->dtag_base;
2316 		limit = i + act->dta_rec.dtrd_size;
2317 		ASSERT(limit <= size);
2318 
2319 		for (nul = 0; i < limit; i++) {
2320 			if (nul) {
2321 				kdata[i] = '\0';
2322 				continue;
2323 			}
2324 
2325 			if (data[i] != '\0')
2326 				continue;
2327 
2328 			nul = 1;
2329 		}
2330 	}
2331 
2332 	for (i = size; i < fsize; i++)
2333 		kdata[i] = 0;
2334 
2335 	key->dtak_hashval = hashval;
2336 	key->dtak_size = size;
2337 	key->dtak_action = action;
2338 	key->dtak_next = agb->dtagb_hash[ndx];
2339 	agb->dtagb_hash[ndx] = key;
2340 
2341 	/*
2342 	 * Finally, apply the aggregator.
2343 	 */
2344 	*((uint64_t *)(key->dtak_data + size)) = agg->dtag_initial;
2345 	agg->dtag_aggregate((uint64_t *)(key->dtak_data + size), expr, arg);
2346 }
2347 
2348 /*
2349  * Given consumer state, this routine finds a speculation in the INACTIVE
2350  * state and transitions it into the ACTIVE state.  If there is no speculation
2351  * in the INACTIVE state, 0 is returned.  In this case, no error counter is
2352  * incremented -- it is up to the caller to take appropriate action.
2353  */
2354 static int
2355 dtrace_speculation(dtrace_state_t *state)
2356 {
2357 	int i = 0;
2358 	dtrace_speculation_state_t current;
2359 	uint32_t *stat = &state->dts_speculations_unavail, count;
2360 
2361 	while (i < state->dts_nspeculations) {
2362 		dtrace_speculation_t *spec = &state->dts_speculations[i];
2363 
2364 		current = spec->dtsp_state;
2365 
2366 		if (current != DTRACESPEC_INACTIVE) {
2367 			if (current == DTRACESPEC_COMMITTINGMANY ||
2368 			    current == DTRACESPEC_COMMITTING ||
2369 			    current == DTRACESPEC_DISCARDING)
2370 				stat = &state->dts_speculations_busy;
2371 			i++;
2372 			continue;
2373 		}
2374 
2375 		if (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2376 		    current, DTRACESPEC_ACTIVE) == current)
2377 			return (i + 1);
2378 	}
2379 
2380 	/*
2381 	 * We couldn't find a speculation.  If we found as much as a single
2382 	 * busy speculation buffer, we'll attribute this failure as "busy"
2383 	 * instead of "unavail".
2384 	 */
2385 	do {
2386 		count = *stat;
2387 	} while (dtrace_cas32(stat, count, count + 1) != count);
2388 
2389 	return (0);
2390 }
2391 
2392 /*
2393  * This routine commits an active speculation.  If the specified speculation
2394  * is not in a valid state to perform a commit(), this routine will silently do
2395  * nothing.  The state of the specified speculation is transitioned according
2396  * to the state transition diagram outlined in <sys/dtrace_impl.h>
2397  */
2398 static void
2399 dtrace_speculation_commit(dtrace_state_t *state, processorid_t cpu,
2400     dtrace_specid_t which)
2401 {
2402 	dtrace_speculation_t *spec;
2403 	dtrace_buffer_t *src, *dest;
2404 	uintptr_t daddr, saddr, dlimit;
2405 	dtrace_speculation_state_t current, new;
2406 	intptr_t offs;
2407 
2408 	if (which == 0)
2409 		return;
2410 
2411 	if (which > state->dts_nspeculations) {
2412 		cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2413 		return;
2414 	}
2415 
2416 	spec = &state->dts_speculations[which - 1];
2417 	src = &spec->dtsp_buffer[cpu];
2418 	dest = &state->dts_buffer[cpu];
2419 
2420 	do {
2421 		current = spec->dtsp_state;
2422 
2423 		if (current == DTRACESPEC_COMMITTINGMANY)
2424 			break;
2425 
2426 		switch (current) {
2427 		case DTRACESPEC_INACTIVE:
2428 		case DTRACESPEC_DISCARDING:
2429 			return;
2430 
2431 		case DTRACESPEC_COMMITTING:
2432 			/*
2433 			 * This is only possible if we are (a) commit()'ing
2434 			 * without having done a prior speculate() on this CPU
2435 			 * and (b) racing with another commit() on a different
2436 			 * CPU.  There's nothing to do -- we just assert that
2437 			 * our offset is 0.
2438 			 */
2439 			ASSERT(src->dtb_offset == 0);
2440 			return;
2441 
2442 		case DTRACESPEC_ACTIVE:
2443 			new = DTRACESPEC_COMMITTING;
2444 			break;
2445 
2446 		case DTRACESPEC_ACTIVEONE:
2447 			/*
2448 			 * This speculation is active on one CPU.  If our
2449 			 * buffer offset is non-zero, we know that the one CPU
2450 			 * must be us.  Otherwise, we are committing on a
2451 			 * different CPU from the speculate(), and we must
2452 			 * rely on being asynchronously cleaned.
2453 			 */
2454 			if (src->dtb_offset != 0) {
2455 				new = DTRACESPEC_COMMITTING;
2456 				break;
2457 			}
2458 			/*FALLTHROUGH*/
2459 
2460 		case DTRACESPEC_ACTIVEMANY:
2461 			new = DTRACESPEC_COMMITTINGMANY;
2462 			break;
2463 
2464 		default:
2465 			ASSERT(0);
2466 		}
2467 	} while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2468 	    current, new) != current);
2469 
2470 	/*
2471 	 * We have set the state to indicate that we are committing this
2472 	 * speculation.  Now reserve the necessary space in the destination
2473 	 * buffer.
2474 	 */
2475 	if ((offs = dtrace_buffer_reserve(dest, src->dtb_offset,
2476 	    sizeof (uint64_t), state, NULL)) < 0) {
2477 		dtrace_buffer_drop(dest);
2478 		goto out;
2479 	}
2480 
2481 	/*
2482 	 * We have the space; copy the buffer across.  (Note that this is a
2483 	 * highly subobtimal bcopy(); in the unlikely event that this becomes
2484 	 * a serious performance issue, a high-performance DTrace-specific
2485 	 * bcopy() should obviously be invented.)
2486 	 */
2487 	daddr = (uintptr_t)dest->dtb_tomax + offs;
2488 	dlimit = daddr + src->dtb_offset;
2489 	saddr = (uintptr_t)src->dtb_tomax;
2490 
2491 	/*
2492 	 * First, the aligned portion.
2493 	 */
2494 	while (dlimit - daddr >= sizeof (uint64_t)) {
2495 		*((uint64_t *)daddr) = *((uint64_t *)saddr);
2496 
2497 		daddr += sizeof (uint64_t);
2498 		saddr += sizeof (uint64_t);
2499 	}
2500 
2501 	/*
2502 	 * Now any left-over bit...
2503 	 */
2504 	while (dlimit - daddr)
2505 		*((uint8_t *)daddr++) = *((uint8_t *)saddr++);
2506 
2507 	/*
2508 	 * Finally, commit the reserved space in the destination buffer.
2509 	 */
2510 	dest->dtb_offset = offs + src->dtb_offset;
2511 
2512 out:
2513 	/*
2514 	 * If we're lucky enough to be the only active CPU on this speculation
2515 	 * buffer, we can just set the state back to DTRACESPEC_INACTIVE.
2516 	 */
2517 	if (current == DTRACESPEC_ACTIVE ||
2518 	    (current == DTRACESPEC_ACTIVEONE && new == DTRACESPEC_COMMITTING)) {
2519 		uint32_t rval = dtrace_cas32((uint32_t *)&spec->dtsp_state,
2520 		    DTRACESPEC_COMMITTING, DTRACESPEC_INACTIVE);
2521 
2522 		ASSERT(rval == DTRACESPEC_COMMITTING);
2523 	}
2524 
2525 	src->dtb_offset = 0;
2526 	src->dtb_xamot_drops += src->dtb_drops;
2527 	src->dtb_drops = 0;
2528 }
2529 
2530 /*
2531  * This routine discards an active speculation.  If the specified speculation
2532  * is not in a valid state to perform a discard(), this routine will silently
2533  * do nothing.  The state of the specified speculation is transitioned
2534  * according to the state transition diagram outlined in <sys/dtrace_impl.h>
2535  */
2536 static void
2537 dtrace_speculation_discard(dtrace_state_t *state, processorid_t cpu,
2538     dtrace_specid_t which)
2539 {
2540 	dtrace_speculation_t *spec;
2541 	dtrace_speculation_state_t current, new;
2542 	dtrace_buffer_t *buf;
2543 
2544 	if (which == 0)
2545 		return;
2546 
2547 	if (which > state->dts_nspeculations) {
2548 		cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2549 		return;
2550 	}
2551 
2552 	spec = &state->dts_speculations[which - 1];
2553 	buf = &spec->dtsp_buffer[cpu];
2554 
2555 	do {
2556 		current = spec->dtsp_state;
2557 
2558 		switch (current) {
2559 		case DTRACESPEC_INACTIVE:
2560 		case DTRACESPEC_COMMITTINGMANY:
2561 		case DTRACESPEC_COMMITTING:
2562 		case DTRACESPEC_DISCARDING:
2563 			return;
2564 
2565 		case DTRACESPEC_ACTIVE:
2566 		case DTRACESPEC_ACTIVEMANY:
2567 			new = DTRACESPEC_DISCARDING;
2568 			break;
2569 
2570 		case DTRACESPEC_ACTIVEONE:
2571 			if (buf->dtb_offset != 0) {
2572 				new = DTRACESPEC_INACTIVE;
2573 			} else {
2574 				new = DTRACESPEC_DISCARDING;
2575 			}
2576 			break;
2577 
2578 		default:
2579 			ASSERT(0);
2580 		}
2581 	} while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2582 	    current, new) != current);
2583 
2584 	buf->dtb_offset = 0;
2585 	buf->dtb_drops = 0;
2586 }
2587 
2588 /*
2589  * Note:  not called from probe context.  This function is called
2590  * asynchronously from cross call context to clean any speculations that are
2591  * in the COMMITTINGMANY or DISCARDING states.  These speculations may not be
2592  * transitioned back to the INACTIVE state until all CPUs have cleaned the
2593  * speculation.
2594  */
2595 static void
2596 dtrace_speculation_clean_here(dtrace_state_t *state)
2597 {
2598 	dtrace_icookie_t cookie;
2599 	processorid_t cpu = CPU->cpu_id;
2600 	dtrace_buffer_t *dest = &state->dts_buffer[cpu];
2601 	dtrace_specid_t i;
2602 
2603 	cookie = dtrace_interrupt_disable();
2604 
2605 	if (dest->dtb_tomax == NULL) {
2606 		dtrace_interrupt_enable(cookie);
2607 		return;
2608 	}
2609 
2610 	for (i = 0; i < state->dts_nspeculations; i++) {
2611 		dtrace_speculation_t *spec = &state->dts_speculations[i];
2612 		dtrace_buffer_t *src = &spec->dtsp_buffer[cpu];
2613 
2614 		if (src->dtb_tomax == NULL)
2615 			continue;
2616 
2617 		if (spec->dtsp_state == DTRACESPEC_DISCARDING) {
2618 			src->dtb_offset = 0;
2619 			continue;
2620 		}
2621 
2622 		if (spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
2623 			continue;
2624 
2625 		if (src->dtb_offset == 0)
2626 			continue;
2627 
2628 		dtrace_speculation_commit(state, cpu, i + 1);
2629 	}
2630 
2631 	dtrace_interrupt_enable(cookie);
2632 }
2633 
2634 /*
2635  * Note:  not called from probe context.  This function is called
2636  * asynchronously (and at a regular interval) to clean any speculations that
2637  * are in the COMMITTINGMANY or DISCARDING states.  If it discovers that there
2638  * is work to be done, it cross calls all CPUs to perform that work;
2639  * COMMITMANY and DISCARDING speculations may not be transitioned back to the
2640  * INACTIVE state until they have been cleaned by all CPUs.
2641  */
2642 static void
2643 dtrace_speculation_clean(dtrace_state_t *state)
2644 {
2645 	int work = 0, rv;
2646 	dtrace_specid_t i;
2647 
2648 	for (i = 0; i < state->dts_nspeculations; i++) {
2649 		dtrace_speculation_t *spec = &state->dts_speculations[i];
2650 
2651 		ASSERT(!spec->dtsp_cleaning);
2652 
2653 		if (spec->dtsp_state != DTRACESPEC_DISCARDING &&
2654 		    spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
2655 			continue;
2656 
2657 		work++;
2658 		spec->dtsp_cleaning = 1;
2659 	}
2660 
2661 	if (!work)
2662 		return;
2663 
2664 	dtrace_xcall(DTRACE_CPUALL,
2665 	    (dtrace_xcall_t)dtrace_speculation_clean_here, state);
2666 
2667 	/*
2668 	 * We now know that all CPUs have committed or discarded their
2669 	 * speculation buffers, as appropriate.  We can now set the state
2670 	 * to inactive.
2671 	 */
2672 	for (i = 0; i < state->dts_nspeculations; i++) {
2673 		dtrace_speculation_t *spec = &state->dts_speculations[i];
2674 		dtrace_speculation_state_t current, new;
2675 
2676 		if (!spec->dtsp_cleaning)
2677 			continue;
2678 
2679 		current = spec->dtsp_state;
2680 		ASSERT(current == DTRACESPEC_DISCARDING ||
2681 		    current == DTRACESPEC_COMMITTINGMANY);
2682 
2683 		new = DTRACESPEC_INACTIVE;
2684 
2685 		rv = dtrace_cas32((uint32_t *)&spec->dtsp_state, current, new);
2686 		ASSERT(rv == current);
2687 		spec->dtsp_cleaning = 0;
2688 	}
2689 }
2690 
2691 /*
2692  * Called as part of a speculate() to get the speculative buffer associated
2693  * with a given speculation.  Returns NULL if the specified speculation is not
2694  * in an ACTIVE state.  If the speculation is in the ACTIVEONE state -- and
2695  * the active CPU is not the specified CPU -- the speculation will be
2696  * atomically transitioned into the ACTIVEMANY state.
2697  */
2698 static dtrace_buffer_t *
2699 dtrace_speculation_buffer(dtrace_state_t *state, processorid_t cpuid,
2700     dtrace_specid_t which)
2701 {
2702 	dtrace_speculation_t *spec;
2703 	dtrace_speculation_state_t current, new;
2704 	dtrace_buffer_t *buf;
2705 
2706 	if (which == 0)
2707 		return (NULL);
2708 
2709 	if (which > state->dts_nspeculations) {
2710 		cpu_core[cpuid].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2711 		return (NULL);
2712 	}
2713 
2714 	spec = &state->dts_speculations[which - 1];
2715 	buf = &spec->dtsp_buffer[cpuid];
2716 
2717 	do {
2718 		current = spec->dtsp_state;
2719 
2720 		switch (current) {
2721 		case DTRACESPEC_INACTIVE:
2722 		case DTRACESPEC_COMMITTINGMANY:
2723 		case DTRACESPEC_DISCARDING:
2724 			return (NULL);
2725 
2726 		case DTRACESPEC_COMMITTING:
2727 			ASSERT(buf->dtb_offset == 0);
2728 			return (NULL);
2729 
2730 		case DTRACESPEC_ACTIVEONE:
2731 			/*
2732 			 * This speculation is currently active on one CPU.
2733 			 * Check the offset in the buffer; if it's non-zero,
2734 			 * that CPU must be us (and we leave the state alone).
2735 			 * If it's zero, assume that we're starting on a new
2736 			 * CPU -- and change the state to indicate that the
2737 			 * speculation is active on more than one CPU.
2738 			 */
2739 			if (buf->dtb_offset != 0)
2740 				return (buf);
2741 
2742 			new = DTRACESPEC_ACTIVEMANY;
2743 			break;
2744 
2745 		case DTRACESPEC_ACTIVEMANY:
2746 			return (buf);
2747 
2748 		case DTRACESPEC_ACTIVE:
2749 			new = DTRACESPEC_ACTIVEONE;
2750 			break;
2751 
2752 		default:
2753 			ASSERT(0);
2754 		}
2755 	} while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2756 	    current, new) != current);
2757 
2758 	ASSERT(new == DTRACESPEC_ACTIVEONE || new == DTRACESPEC_ACTIVEMANY);
2759 	return (buf);
2760 }
2761 
2762 /*
2763  * Return a string.  In the event that the user lacks the privilege to access
2764  * arbitrary kernel memory, we copy the string out to scratch memory so that we
2765  * don't fail access checking.
2766  *
2767  * dtrace_dif_variable() uses this routine as a helper for various
2768  * builtin values such as 'execname' and 'probefunc.'
2769  */
2770 uintptr_t
2771 dtrace_dif_varstr(uintptr_t addr, dtrace_state_t *state,
2772     dtrace_mstate_t *mstate)
2773 {
2774 	uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
2775 	uintptr_t ret;
2776 	size_t strsz;
2777 
2778 	/*
2779 	 * The easy case: this probe is allowed to read all of memory, so
2780 	 * we can just return this as a vanilla pointer.
2781 	 */
2782 	if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
2783 		return (addr);
2784 
2785 	/*
2786 	 * This is the tougher case: we copy the string in question from
2787 	 * kernel memory into scratch memory and return it that way: this
2788 	 * ensures that we won't trip up when access checking tests the
2789 	 * BYREF return value.
2790 	 */
2791 	strsz = dtrace_strlen((char *)addr, size) + 1;
2792 
2793 	if (mstate->dtms_scratch_ptr + strsz >
2794 	    mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
2795 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
2796 		return (NULL);
2797 	}
2798 
2799 	dtrace_strcpy((const void *)addr, (void *)mstate->dtms_scratch_ptr,
2800 	    strsz);
2801 	ret = mstate->dtms_scratch_ptr;
2802 	mstate->dtms_scratch_ptr += strsz;
2803 	return (ret);
2804 }
2805 
2806 /*
2807  * This function implements the DIF emulator's variable lookups.  The emulator
2808  * passes a reserved variable identifier and optional built-in array index.
2809  */
2810 static uint64_t
2811 dtrace_dif_variable(dtrace_mstate_t *mstate, dtrace_state_t *state, uint64_t v,
2812     uint64_t ndx)
2813 {
2814 	/*
2815 	 * If we're accessing one of the uncached arguments, we'll turn this
2816 	 * into a reference in the args array.
2817 	 */
2818 	if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9) {
2819 		ndx = v - DIF_VAR_ARG0;
2820 		v = DIF_VAR_ARGS;
2821 	}
2822 
2823 	switch (v) {
2824 	case DIF_VAR_ARGS:
2825 		if (!(mstate->dtms_access & DTRACE_ACCESS_ARGS)) {
2826 			cpu_core[CPU->cpu_id].cpuc_dtrace_flags |=
2827 			    CPU_DTRACE_KPRIV;
2828 			return (0);
2829 		}
2830 
2831 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_ARGS);
2832 		if (ndx >= sizeof (mstate->dtms_arg) /
2833 		    sizeof (mstate->dtms_arg[0])) {
2834 			int aframes = mstate->dtms_probe->dtpr_aframes + 2;
2835 			dtrace_provider_t *pv;
2836 			uint64_t val;
2837 
2838 			pv = mstate->dtms_probe->dtpr_provider;
2839 			if (pv->dtpv_pops.dtps_getargval != NULL)
2840 				val = pv->dtpv_pops.dtps_getargval(pv->dtpv_arg,
2841 				    mstate->dtms_probe->dtpr_id,
2842 				    mstate->dtms_probe->dtpr_arg, ndx, aframes);
2843 			else
2844 				val = dtrace_getarg(ndx, aframes);
2845 
2846 			/*
2847 			 * This is regrettably required to keep the compiler
2848 			 * from tail-optimizing the call to dtrace_getarg().
2849 			 * The condition always evaluates to true, but the
2850 			 * compiler has no way of figuring that out a priori.
2851 			 * (None of this would be necessary if the compiler
2852 			 * could be relied upon to _always_ tail-optimize
2853 			 * the call to dtrace_getarg() -- but it can't.)
2854 			 */
2855 			if (mstate->dtms_probe != NULL)
2856 				return (val);
2857 
2858 			ASSERT(0);
2859 		}
2860 
2861 		return (mstate->dtms_arg[ndx]);
2862 
2863 	case DIF_VAR_UREGS: {
2864 		klwp_t *lwp;
2865 
2866 		if (!dtrace_priv_proc(state, mstate))
2867 			return (0);
2868 
2869 		if ((lwp = curthread->t_lwp) == NULL) {
2870 			DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
2871 			cpu_core[CPU->cpu_id].cpuc_dtrace_illval = NULL;
2872 			return (0);
2873 		}
2874 
2875 		return (dtrace_getreg(lwp->lwp_regs, ndx));
2876 	}
2877 
2878 	case DIF_VAR_VMREGS: {
2879 		uint64_t rval;
2880 
2881 		if (!dtrace_priv_kernel(state))
2882 			return (0);
2883 
2884 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
2885 
2886 		rval = dtrace_getvmreg(ndx,
2887 		    &cpu_core[CPU->cpu_id].cpuc_dtrace_flags);
2888 
2889 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
2890 
2891 		return (rval);
2892 	}
2893 
2894 	case DIF_VAR_CURTHREAD:
2895 		if (!dtrace_priv_kernel(state))
2896 			return (0);
2897 		return ((uint64_t)(uintptr_t)curthread);
2898 
2899 	case DIF_VAR_TIMESTAMP:
2900 		if (!(mstate->dtms_present & DTRACE_MSTATE_TIMESTAMP)) {
2901 			mstate->dtms_timestamp = dtrace_gethrtime();
2902 			mstate->dtms_present |= DTRACE_MSTATE_TIMESTAMP;
2903 		}
2904 		return (mstate->dtms_timestamp);
2905 
2906 	case DIF_VAR_VTIMESTAMP:
2907 		ASSERT(dtrace_vtime_references != 0);
2908 		return (curthread->t_dtrace_vtime);
2909 
2910 	case DIF_VAR_WALLTIMESTAMP:
2911 		if (!(mstate->dtms_present & DTRACE_MSTATE_WALLTIMESTAMP)) {
2912 			mstate->dtms_walltimestamp = dtrace_gethrestime();
2913 			mstate->dtms_present |= DTRACE_MSTATE_WALLTIMESTAMP;
2914 		}
2915 		return (mstate->dtms_walltimestamp);
2916 
2917 	case DIF_VAR_IPL:
2918 		if (!dtrace_priv_kernel(state))
2919 			return (0);
2920 		if (!(mstate->dtms_present & DTRACE_MSTATE_IPL)) {
2921 			mstate->dtms_ipl = dtrace_getipl();
2922 			mstate->dtms_present |= DTRACE_MSTATE_IPL;
2923 		}
2924 		return (mstate->dtms_ipl);
2925 
2926 	case DIF_VAR_EPID:
2927 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_EPID);
2928 		return (mstate->dtms_epid);
2929 
2930 	case DIF_VAR_ID:
2931 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
2932 		return (mstate->dtms_probe->dtpr_id);
2933 
2934 	case DIF_VAR_STACKDEPTH:
2935 		if (!dtrace_priv_kernel(state))
2936 			return (0);
2937 		if (!(mstate->dtms_present & DTRACE_MSTATE_STACKDEPTH)) {
2938 			int aframes = mstate->dtms_probe->dtpr_aframes + 2;
2939 
2940 			mstate->dtms_stackdepth = dtrace_getstackdepth(aframes);
2941 			mstate->dtms_present |= DTRACE_MSTATE_STACKDEPTH;
2942 		}
2943 		return (mstate->dtms_stackdepth);
2944 
2945 	case DIF_VAR_USTACKDEPTH:
2946 		if (!dtrace_priv_proc(state, mstate))
2947 			return (0);
2948 		if (!(mstate->dtms_present & DTRACE_MSTATE_USTACKDEPTH)) {
2949 			/*
2950 			 * See comment in DIF_VAR_PID.
2951 			 */
2952 			if (DTRACE_ANCHORED(mstate->dtms_probe) &&
2953 			    CPU_ON_INTR(CPU)) {
2954 				mstate->dtms_ustackdepth = 0;
2955 			} else {
2956 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
2957 				mstate->dtms_ustackdepth =
2958 				    dtrace_getustackdepth();
2959 				DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
2960 			}
2961 			mstate->dtms_present |= DTRACE_MSTATE_USTACKDEPTH;
2962 		}
2963 		return (mstate->dtms_ustackdepth);
2964 
2965 	case DIF_VAR_CALLER:
2966 		if (!dtrace_priv_kernel(state))
2967 			return (0);
2968 		if (!(mstate->dtms_present & DTRACE_MSTATE_CALLER)) {
2969 			int aframes = mstate->dtms_probe->dtpr_aframes + 2;
2970 
2971 			if (!DTRACE_ANCHORED(mstate->dtms_probe)) {
2972 				/*
2973 				 * If this is an unanchored probe, we are
2974 				 * required to go through the slow path:
2975 				 * dtrace_caller() only guarantees correct
2976 				 * results for anchored probes.
2977 				 */
2978 				pc_t caller[2];
2979 
2980 				dtrace_getpcstack(caller, 2, aframes,
2981 				    (uint32_t *)(uintptr_t)mstate->dtms_arg[0]);
2982 				mstate->dtms_caller = caller[1];
2983 			} else if ((mstate->dtms_caller =
2984 			    dtrace_caller(aframes)) == -1) {
2985 				/*
2986 				 * We have failed to do this the quick way;
2987 				 * we must resort to the slower approach of
2988 				 * calling dtrace_getpcstack().
2989 				 */
2990 				pc_t caller;
2991 
2992 				dtrace_getpcstack(&caller, 1, aframes, NULL);
2993 				mstate->dtms_caller = caller;
2994 			}
2995 
2996 			mstate->dtms_present |= DTRACE_MSTATE_CALLER;
2997 		}
2998 		return (mstate->dtms_caller);
2999 
3000 	case DIF_VAR_UCALLER:
3001 		if (!dtrace_priv_proc(state, mstate))
3002 			return (0);
3003 
3004 		if (!(mstate->dtms_present & DTRACE_MSTATE_UCALLER)) {
3005 			uint64_t ustack[3];
3006 
3007 			/*
3008 			 * dtrace_getupcstack() fills in the first uint64_t
3009 			 * with the current PID.  The second uint64_t will
3010 			 * be the program counter at user-level.  The third
3011 			 * uint64_t will contain the caller, which is what
3012 			 * we're after.
3013 			 */
3014 			ustack[2] = NULL;
3015 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3016 			dtrace_getupcstack(ustack, 3);
3017 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3018 			mstate->dtms_ucaller = ustack[2];
3019 			mstate->dtms_present |= DTRACE_MSTATE_UCALLER;
3020 		}
3021 
3022 		return (mstate->dtms_ucaller);
3023 
3024 	case DIF_VAR_PROBEPROV:
3025 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3026 		return (dtrace_dif_varstr(
3027 		    (uintptr_t)mstate->dtms_probe->dtpr_provider->dtpv_name,
3028 		    state, mstate));
3029 
3030 	case DIF_VAR_PROBEMOD:
3031 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3032 		return (dtrace_dif_varstr(
3033 		    (uintptr_t)mstate->dtms_probe->dtpr_mod,
3034 		    state, mstate));
3035 
3036 	case DIF_VAR_PROBEFUNC:
3037 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3038 		return (dtrace_dif_varstr(
3039 		    (uintptr_t)mstate->dtms_probe->dtpr_func,
3040 		    state, mstate));
3041 
3042 	case DIF_VAR_PROBENAME:
3043 		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3044 		return (dtrace_dif_varstr(
3045 		    (uintptr_t)mstate->dtms_probe->dtpr_name,
3046 		    state, mstate));
3047 
3048 	case DIF_VAR_PID:
3049 		if (!dtrace_priv_proc(state, mstate))
3050 			return (0);
3051 
3052 		/*
3053 		 * Note that we are assuming that an unanchored probe is
3054 		 * always due to a high-level interrupt.  (And we're assuming
3055 		 * that there is only a single high level interrupt.)
3056 		 */
3057 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3058 			return (pid0.pid_id);
3059 
3060 		/*
3061 		 * It is always safe to dereference one's own t_procp pointer:
3062 		 * it always points to a valid, allocated proc structure.
3063 		 * Further, it is always safe to dereference the p_pidp member
3064 		 * of one's own proc structure.  (These are truisms becuase
3065 		 * threads and processes don't clean up their own state --
3066 		 * they leave that task to whomever reaps them.)
3067 		 */
3068 		return ((uint64_t)curthread->t_procp->p_pidp->pid_id);
3069 
3070 	case DIF_VAR_PPID:
3071 		if (!dtrace_priv_proc(state, mstate))
3072 			return (0);
3073 
3074 		/*
3075 		 * See comment in DIF_VAR_PID.
3076 		 */
3077 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3078 			return (pid0.pid_id);
3079 
3080 		/*
3081 		 * It is always safe to dereference one's own t_procp pointer:
3082 		 * it always points to a valid, allocated proc structure.
3083 		 * (This is true because threads don't clean up their own
3084 		 * state -- they leave that task to whomever reaps them.)
3085 		 */
3086 		return ((uint64_t)curthread->t_procp->p_ppid);
3087 
3088 	case DIF_VAR_TID:
3089 		/*
3090 		 * See comment in DIF_VAR_PID.
3091 		 */
3092 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3093 			return (0);
3094 
3095 		return ((uint64_t)curthread->t_tid);
3096 
3097 	case DIF_VAR_EXECNAME:
3098 		if (!dtrace_priv_proc(state, mstate))
3099 			return (0);
3100 
3101 		/*
3102 		 * See comment in DIF_VAR_PID.
3103 		 */
3104 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3105 			return ((uint64_t)(uintptr_t)p0.p_user.u_comm);
3106 
3107 		/*
3108 		 * It is always safe to dereference one's own t_procp pointer:
3109 		 * it always points to a valid, allocated proc structure.
3110 		 * (This is true because threads don't clean up their own
3111 		 * state -- they leave that task to whomever reaps them.)
3112 		 */
3113 		return (dtrace_dif_varstr(
3114 		    (uintptr_t)curthread->t_procp->p_user.u_comm,
3115 		    state, mstate));
3116 
3117 	case DIF_VAR_ZONENAME:
3118 		if (!dtrace_priv_proc(state, mstate))
3119 			return (0);
3120 
3121 		/*
3122 		 * See comment in DIF_VAR_PID.
3123 		 */
3124 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3125 			return ((uint64_t)(uintptr_t)p0.p_zone->zone_name);
3126 
3127 		/*
3128 		 * It is always safe to dereference one's own t_procp pointer:
3129 		 * it always points to a valid, allocated proc structure.
3130 		 * (This is true because threads don't clean up their own
3131 		 * state -- they leave that task to whomever reaps them.)
3132 		 */
3133 		return (dtrace_dif_varstr(
3134 		    (uintptr_t)curthread->t_procp->p_zone->zone_name,
3135 		    state, mstate));
3136 
3137 	case DIF_VAR_UID:
3138 		if (!dtrace_priv_proc(state, mstate))
3139 			return (0);
3140 
3141 		/*
3142 		 * See comment in DIF_VAR_PID.
3143 		 */
3144 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3145 			return ((uint64_t)p0.p_cred->cr_uid);
3146 
3147 		/*
3148 		 * It is always safe to dereference one's own t_procp pointer:
3149 		 * it always points to a valid, allocated proc structure.
3150 		 * (This is true because threads don't clean up their own
3151 		 * state -- they leave that task to whomever reaps them.)
3152 		 *
3153 		 * Additionally, it is safe to dereference one's own process
3154 		 * credential, since this is never NULL after process birth.
3155 		 */
3156 		return ((uint64_t)curthread->t_procp->p_cred->cr_uid);
3157 
3158 	case DIF_VAR_GID:
3159 		if (!dtrace_priv_proc(state, mstate))
3160 			return (0);
3161 
3162 		/*
3163 		 * See comment in DIF_VAR_PID.
3164 		 */
3165 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3166 			return ((uint64_t)p0.p_cred->cr_gid);
3167 
3168 		/*
3169 		 * It is always safe to dereference one's own t_procp pointer:
3170 		 * it always points to a valid, allocated proc structure.
3171 		 * (This is true because threads don't clean up their own
3172 		 * state -- they leave that task to whomever reaps them.)
3173 		 *
3174 		 * Additionally, it is safe to dereference one's own process
3175 		 * credential, since this is never NULL after process birth.
3176 		 */
3177 		return ((uint64_t)curthread->t_procp->p_cred->cr_gid);
3178 
3179 	case DIF_VAR_ERRNO: {
3180 		klwp_t *lwp;
3181 		if (!dtrace_priv_proc(state, mstate))
3182 			return (0);
3183 
3184 		/*
3185 		 * See comment in DIF_VAR_PID.
3186 		 */
3187 		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3188 			return (0);
3189 
3190 		/*
3191 		 * It is always safe to dereference one's own t_lwp pointer in
3192 		 * the event that this pointer is non-NULL.  (This is true
3193 		 * because threads and lwps don't clean up their own state --
3194 		 * they leave that task to whomever reaps them.)
3195 		 */
3196 		if ((lwp = curthread->t_lwp) == NULL)
3197 			return (0);
3198 
3199 		return ((uint64_t)lwp->lwp_errno);
3200 	}
3201 	default:
3202 		DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
3203 		return (0);
3204 	}
3205 }
3206 
3207 /*
3208  * Emulate the execution of DTrace ID subroutines invoked by the call opcode.
3209  * Notice that we don't bother validating the proper number of arguments or
3210  * their types in the tuple stack.  This isn't needed because all argument
3211  * interpretation is safe because of our load safety -- the worst that can
3212  * happen is that a bogus program can obtain bogus results.
3213  */
3214 static void
3215 dtrace_dif_subr(uint_t subr, uint_t rd, uint64_t *regs,
3216     dtrace_key_t *tupregs, int nargs,
3217     dtrace_mstate_t *mstate, dtrace_state_t *state)
3218 {
3219 	volatile uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
3220 	volatile uintptr_t *illval = &cpu_core[CPU->cpu_id].cpuc_dtrace_illval;
3221 	dtrace_vstate_t *vstate = &state->dts_vstate;
3222 
3223 	union {
3224 		mutex_impl_t mi;
3225 		uint64_t mx;
3226 	} m;
3227 
3228 	union {
3229 		krwlock_t ri;
3230 		uintptr_t rw;
3231 	} r;
3232 
3233 	switch (subr) {
3234 	case DIF_SUBR_RAND:
3235 		regs[rd] = (dtrace_gethrtime() * 2416 + 374441) % 1771875;
3236 		break;
3237 
3238 	case DIF_SUBR_MUTEX_OWNED:
3239 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3240 		    mstate, vstate)) {
3241 			regs[rd] = NULL;
3242 			break;
3243 		}
3244 
3245 		m.mx = dtrace_load64(tupregs[0].dttk_value);
3246 		if (MUTEX_TYPE_ADAPTIVE(&m.mi))
3247 			regs[rd] = MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER;
3248 		else
3249 			regs[rd] = LOCK_HELD(&m.mi.m_spin.m_spinlock);
3250 		break;
3251 
3252 	case DIF_SUBR_MUTEX_OWNER:
3253 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3254 		    mstate, vstate)) {
3255 			regs[rd] = NULL;
3256 			break;
3257 		}
3258 
3259 		m.mx = dtrace_load64(tupregs[0].dttk_value);
3260 		if (MUTEX_TYPE_ADAPTIVE(&m.mi) &&
3261 		    MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER)
3262 			regs[rd] = (uintptr_t)MUTEX_OWNER(&m.mi);
3263 		else
3264 			regs[rd] = 0;
3265 		break;
3266 
3267 	case DIF_SUBR_MUTEX_TYPE_ADAPTIVE:
3268 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3269 		    mstate, vstate)) {
3270 			regs[rd] = NULL;
3271 			break;
3272 		}
3273 
3274 		m.mx = dtrace_load64(tupregs[0].dttk_value);
3275 		regs[rd] = MUTEX_TYPE_ADAPTIVE(&m.mi);
3276 		break;
3277 
3278 	case DIF_SUBR_MUTEX_TYPE_SPIN:
3279 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3280 		    mstate, vstate)) {
3281 			regs[rd] = NULL;
3282 			break;
3283 		}
3284 
3285 		m.mx = dtrace_load64(tupregs[0].dttk_value);
3286 		regs[rd] = MUTEX_TYPE_SPIN(&m.mi);
3287 		break;
3288 
3289 	case DIF_SUBR_RW_READ_HELD: {
3290 		uintptr_t tmp;
3291 
3292 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
3293 		    mstate, vstate)) {
3294 			regs[rd] = NULL;
3295 			break;
3296 		}
3297 
3298 		r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3299 		regs[rd] = _RW_READ_HELD(&r.ri, tmp);
3300 		break;
3301 	}
3302 
3303 	case DIF_SUBR_RW_WRITE_HELD:
3304 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
3305 		    mstate, vstate)) {
3306 			regs[rd] = NULL;
3307 			break;
3308 		}
3309 
3310 		r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3311 		regs[rd] = _RW_WRITE_HELD(&r.ri);
3312 		break;
3313 
3314 	case DIF_SUBR_RW_ISWRITER:
3315 		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
3316 		    mstate, vstate)) {
3317 			regs[rd] = NULL;
3318 			break;
3319 		}
3320 
3321 		r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3322 		regs[rd] = _RW_ISWRITER(&r.ri);
3323 		break;
3324 
3325 	case DIF_SUBR_BCOPY: {
3326 		/*
3327 		 * We need to be sure that the destination is in the scratch
3328 		 * region -- no other region is allowed.
3329 		 */
3330 		uintptr_t src = tupregs[0].dttk_value;
3331 		uintptr_t dest = tupregs[1].dttk_value;
3332 		size_t size = tupregs[2].dttk_value;
3333 
3334 		if (!dtrace_inscratch(dest, size, mstate)) {
3335 			*flags |= CPU_DTRACE_BADADDR;
3336 			*illval = regs[rd];
3337 			break;
3338 		}
3339 
3340 		if (!dtrace_canload(src, size, mstate, vstate)) {
3341 			regs[rd] = NULL;
3342 			break;
3343 		}
3344 
3345 		dtrace_bcopy((void *)src, (void *)dest, size);
3346 		break;
3347 	}
3348 
3349 	case DIF_SUBR_ALLOCA:
3350 	case DIF_SUBR_COPYIN: {
3351 		uintptr_t dest = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
3352 		uint64_t size =
3353 		    tupregs[subr == DIF_SUBR_ALLOCA ? 0 : 1].dttk_value;
3354 		size_t scratch_size = (dest - mstate->dtms_scratch_ptr) + size;
3355 
3356 		/*
3357 		 * This action doesn't require any credential checks since
3358 		 * probes will not activate in user contexts to which the
3359 		 * enabling user does not have permissions.
3360 		 */
3361 
3362 		/*
3363 		 * Rounding up the user allocation size could have overflowed
3364 		 * a large, bogus allocation (like -1ULL) to 0.
3365 		 */
3366 		if (scratch_size < size ||
3367 		    !DTRACE_INSCRATCH(mstate, scratch_size)) {
3368 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3369 			regs[rd] = NULL;
3370 			break;
3371 		}
3372 
3373 		if (subr == DIF_SUBR_COPYIN) {
3374 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3375 			dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
3376 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3377 		}
3378 
3379 		mstate->dtms_scratch_ptr += scratch_size;
3380 		regs[rd] = dest;
3381 		break;
3382 	}
3383 
3384 	case DIF_SUBR_COPYINTO: {
3385 		uint64_t size = tupregs[1].dttk_value;
3386 		uintptr_t dest = tupregs[2].dttk_value;
3387 
3388 		/*
3389 		 * This action doesn't require any credential checks since
3390 		 * probes will not activate in user contexts to which the
3391 		 * enabling user does not have permissions.
3392 		 */
3393 		if (!dtrace_inscratch(dest, size, mstate)) {
3394 			*flags |= CPU_DTRACE_BADADDR;
3395 			*illval = regs[rd];
3396 			break;
3397 		}
3398 
3399 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3400 		dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
3401 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3402 		break;
3403 	}
3404 
3405 	case DIF_SUBR_COPYINSTR: {
3406 		uintptr_t dest = mstate->dtms_scratch_ptr;
3407 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3408 
3409 		if (nargs > 1 && tupregs[1].dttk_value < size)
3410 			size = tupregs[1].dttk_value + 1;
3411 
3412 		/*
3413 		 * This action doesn't require any credential checks since
3414 		 * probes will not activate in user contexts to which the
3415 		 * enabling user does not have permissions.
3416 		 */
3417 		if (!DTRACE_INSCRATCH(mstate, size)) {
3418 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3419 			regs[rd] = NULL;
3420 			break;
3421 		}
3422 
3423 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3424 		dtrace_copyinstr(tupregs[0].dttk_value, dest, size, flags);
3425 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3426 
3427 		((char *)dest)[size - 1] = '\0';
3428 		mstate->dtms_scratch_ptr += size;
3429 		regs[rd] = dest;
3430 		break;
3431 	}
3432 
3433 	case DIF_SUBR_MSGSIZE:
3434 	case DIF_SUBR_MSGDSIZE: {
3435 		uintptr_t baddr = tupregs[0].dttk_value, daddr;
3436 		uintptr_t wptr, rptr;
3437 		size_t count = 0;
3438 		int cont = 0;
3439 
3440 		while (baddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
3441 
3442 			if (!dtrace_canload(baddr, sizeof (mblk_t), mstate,
3443 			    vstate)) {
3444 				regs[rd] = NULL;
3445 				break;
3446 			}
3447 
3448 			wptr = dtrace_loadptr(baddr +
3449 			    offsetof(mblk_t, b_wptr));
3450 
3451 			rptr = dtrace_loadptr(baddr +
3452 			    offsetof(mblk_t, b_rptr));
3453 
3454 			if (wptr < rptr) {
3455 				*flags |= CPU_DTRACE_BADADDR;
3456 				*illval = tupregs[0].dttk_value;
3457 				break;
3458 			}
3459 
3460 			daddr = dtrace_loadptr(baddr +
3461 			    offsetof(mblk_t, b_datap));
3462 
3463 			baddr = dtrace_loadptr(baddr +
3464 			    offsetof(mblk_t, b_cont));
3465 
3466 			/*
3467 			 * We want to prevent against denial-of-service here,
3468 			 * so we're only going to search the list for
3469 			 * dtrace_msgdsize_max mblks.
3470 			 */
3471 			if (cont++ > dtrace_msgdsize_max) {
3472 				*flags |= CPU_DTRACE_ILLOP;
3473 				break;
3474 			}
3475 
3476 			if (subr == DIF_SUBR_MSGDSIZE) {
3477 				if (dtrace_load8(daddr +
3478 				    offsetof(dblk_t, db_type)) != M_DATA)
3479 					continue;
3480 			}
3481 
3482 			count += wptr - rptr;
3483 		}
3484 
3485 		if (!(*flags & CPU_DTRACE_FAULT))
3486 			regs[rd] = count;
3487 
3488 		break;
3489 	}
3490 
3491 	case DIF_SUBR_PROGENYOF: {
3492 		pid_t pid = tupregs[0].dttk_value;
3493 		proc_t *p;
3494 		int rval = 0;
3495 
3496 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3497 
3498 		for (p = curthread->t_procp; p != NULL; p = p->p_parent) {
3499 			if (p->p_pidp->pid_id == pid) {
3500 				rval = 1;
3501 				break;
3502 			}
3503 		}
3504 
3505 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3506 
3507 		regs[rd] = rval;
3508 		break;
3509 	}
3510 
3511 	case DIF_SUBR_SPECULATION:
3512 		regs[rd] = dtrace_speculation(state);
3513 		break;
3514 
3515 	case DIF_SUBR_COPYOUT: {
3516 		uintptr_t kaddr = tupregs[0].dttk_value;
3517 		uintptr_t uaddr = tupregs[1].dttk_value;
3518 		uint64_t size = tupregs[2].dttk_value;
3519 
3520 		if (!dtrace_destructive_disallow &&
3521 		    dtrace_priv_proc_control(state, mstate) &&
3522 		    !dtrace_istoxic(kaddr, size)) {
3523 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3524 			dtrace_copyout(kaddr, uaddr, size, flags);
3525 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3526 		}
3527 		break;
3528 	}
3529 
3530 	case DIF_SUBR_COPYOUTSTR: {
3531 		uintptr_t kaddr = tupregs[0].dttk_value;
3532 		uintptr_t uaddr = tupregs[1].dttk_value;
3533 		uint64_t size = tupregs[2].dttk_value;
3534 
3535 		if (!dtrace_destructive_disallow &&
3536 		    dtrace_priv_proc_control(state, mstate) &&
3537 		    !dtrace_istoxic(kaddr, size)) {
3538 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3539 			dtrace_copyoutstr(kaddr, uaddr, size, flags);
3540 			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3541 		}
3542 		break;
3543 	}
3544 
3545 	case DIF_SUBR_STRLEN: {
3546 		size_t sz;
3547 		uintptr_t addr = (uintptr_t)tupregs[0].dttk_value;
3548 		sz = dtrace_strlen((char *)addr,
3549 		    state->dts_options[DTRACEOPT_STRSIZE]);
3550 
3551 		if (!dtrace_canload(addr, sz + 1, mstate, vstate)) {
3552 			regs[rd] = NULL;
3553 			break;
3554 		}
3555 
3556 		regs[rd] = sz;
3557 
3558 		break;
3559 	}
3560 
3561 	case DIF_SUBR_STRCHR:
3562 	case DIF_SUBR_STRRCHR: {
3563 		/*
3564 		 * We're going to iterate over the string looking for the
3565 		 * specified character.  We will iterate until we have reached
3566 		 * the string length or we have found the character.  If this
3567 		 * is DIF_SUBR_STRRCHR, we will look for the last occurrence
3568 		 * of the specified character instead of the first.
3569 		 */
3570 		uintptr_t saddr = tupregs[0].dttk_value;
3571 		uintptr_t addr = tupregs[0].dttk_value;
3572 		uintptr_t limit = addr + state->dts_options[DTRACEOPT_STRSIZE];
3573 		char c, target = (char)tupregs[1].dttk_value;
3574 
3575 		for (regs[rd] = NULL; addr < limit; addr++) {
3576 			if ((c = dtrace_load8(addr)) == target) {
3577 				regs[rd] = addr;
3578 
3579 				if (subr == DIF_SUBR_STRCHR)
3580 					break;
3581 			}
3582 
3583 			if (c == '\0')
3584 				break;
3585 		}
3586 
3587 		if (!dtrace_canload(saddr, addr - saddr, mstate, vstate)) {
3588 			regs[rd] = NULL;
3589 			break;
3590 		}
3591 
3592 		break;
3593 	}
3594 
3595 	case DIF_SUBR_STRSTR:
3596 	case DIF_SUBR_INDEX:
3597 	case DIF_SUBR_RINDEX: {
3598 		/*
3599 		 * We're going to iterate over the string looking for the
3600 		 * specified string.  We will iterate until we have reached
3601 		 * the string length or we have found the string.  (Yes, this
3602 		 * is done in the most naive way possible -- but considering
3603 		 * that the string we're searching for is likely to be
3604 		 * relatively short, the complexity of Rabin-Karp or similar
3605 		 * hardly seems merited.)
3606 		 */
3607 		char *addr = (char *)(uintptr_t)tupregs[0].dttk_value;
3608 		char *substr = (char *)(uintptr_t)tupregs[1].dttk_value;
3609 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3610 		size_t len = dtrace_strlen(addr, size);
3611 		size_t sublen = dtrace_strlen(substr, size);
3612 		char *limit = addr + len, *orig = addr;
3613 		int notfound = subr == DIF_SUBR_STRSTR ? 0 : -1;
3614 		int inc = 1;
3615 
3616 		regs[rd] = notfound;
3617 
3618 		if (!dtrace_canload((uintptr_t)addr, len + 1, mstate, vstate)) {
3619 			regs[rd] = NULL;
3620 			break;
3621 		}
3622 
3623 		if (!dtrace_canload((uintptr_t)substr, sublen + 1, mstate,
3624 		    vstate)) {
3625 			regs[rd] = NULL;
3626 			break;
3627 		}
3628 
3629 		/*
3630 		 * strstr() and index()/rindex() have similar semantics if
3631 		 * both strings are the empty string: strstr() returns a
3632 		 * pointer to the (empty) string, and index() and rindex()
3633 		 * both return index 0 (regardless of any position argument).
3634 		 */
3635 		if (sublen == 0 && len == 0) {
3636 			if (subr == DIF_SUBR_STRSTR)
3637 				regs[rd] = (uintptr_t)addr;
3638 			else
3639 				regs[rd] = 0;
3640 			break;
3641 		}
3642 
3643 		if (subr != DIF_SUBR_STRSTR) {
3644 			if (subr == DIF_SUBR_RINDEX) {
3645 				limit = orig - 1;
3646 				addr += len;
3647 				inc = -1;
3648 			}
3649 
3650 			/*
3651 			 * Both index() and rindex() take an optional position
3652 			 * argument that denotes the starting position.
3653 			 */
3654 			if (nargs == 3) {
3655 				int64_t pos = (int64_t)tupregs[2].dttk_value;
3656 
3657 				/*
3658 				 * If the position argument to index() is
3659 				 * negative, Perl implicitly clamps it at
3660 				 * zero.  This semantic is a little surprising
3661 				 * given the special meaning of negative
3662 				 * positions to similar Perl functions like
3663 				 * substr(), but it appears to reflect a
3664 				 * notion that index() can start from a
3665 				 * negative index and increment its way up to
3666 				 * the string.  Given this notion, Perl's
3667 				 * rindex() is at least self-consistent in
3668 				 * that it implicitly clamps positions greater
3669 				 * than the string length to be the string
3670 				 * length.  Where Perl completely loses
3671 				 * coherence, however, is when the specified
3672 				 * substring is the empty string ("").  In
3673 				 * this case, even if the position is
3674 				 * negative, rindex() returns 0 -- and even if
3675 				 * the position is greater than the length,
3676 				 * index() returns the string length.  These
3677 				 * semantics violate the notion that index()
3678 				 * should never return a value less than the
3679 				 * specified position and that rindex() should
3680 				 * never return a value greater than the
3681 				 * specified position.  (One assumes that
3682 				 * these semantics are artifacts of Perl's
3683 				 * implementation and not the results of
3684 				 * deliberate design -- it beggars belief that
3685 				 * even Larry Wall could desire such oddness.)
3686 				 * While in the abstract one would wish for
3687 				 * consistent position semantics across
3688 				 * substr(), index() and rindex() -- or at the
3689 				 * very least self-consistent position
3690 				 * semantics for index() and rindex() -- we
3691 				 * instead opt to keep with the extant Perl
3692 				 * semantics, in all their broken glory.  (Do
3693 				 * we have more desire to maintain Perl's
3694 				 * semantics than Perl does?  Probably.)
3695 				 */
3696 				if (subr == DIF_SUBR_RINDEX) {
3697 					if (pos < 0) {
3698 						if (sublen == 0)
3699 							regs[rd] = 0;
3700 						break;
3701 					}
3702 
3703 					if (pos > len)
3704 						pos = len;
3705 				} else {
3706 					if (pos < 0)
3707 						pos = 0;
3708 
3709 					if (pos >= len) {
3710 						if (sublen == 0)
3711 							regs[rd] = len;
3712 						break;
3713 					}
3714 				}
3715 
3716 				addr = orig + pos;
3717 			}
3718 		}
3719 
3720 		for (regs[rd] = notfound; addr != limit; addr += inc) {
3721 			if (dtrace_strncmp(addr, substr, sublen) == 0) {
3722 				if (subr != DIF_SUBR_STRSTR) {
3723 					/*
3724 					 * As D index() and rindex() are
3725 					 * modeled on Perl (and not on awk),
3726 					 * we return a zero-based (and not a
3727 					 * one-based) index.  (For you Perl
3728 					 * weenies: no, we're not going to add
3729 					 * $[ -- and shouldn't you be at a con
3730 					 * or something?)
3731 					 */
3732 					regs[rd] = (uintptr_t)(addr - orig);
3733 					break;
3734 				}
3735 
3736 				ASSERT(subr == DIF_SUBR_STRSTR);
3737 				regs[rd] = (uintptr_t)addr;
3738 				break;
3739 			}
3740 		}
3741 
3742 		break;
3743 	}
3744 
3745 	case DIF_SUBR_STRTOK: {
3746 		uintptr_t addr = tupregs[0].dttk_value;
3747 		uintptr_t tokaddr = tupregs[1].dttk_value;
3748 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3749 		uintptr_t limit, toklimit = tokaddr + size;
3750 		uint8_t c, tokmap[32];	 /* 256 / 8 */
3751 		char *dest = (char *)mstate->dtms_scratch_ptr;
3752 		int i;
3753 
3754 		/*
3755 		 * Check both the token buffer and (later) the input buffer,
3756 		 * since both could be non-scratch addresses.
3757 		 */
3758 		if (!dtrace_strcanload(tokaddr, size, mstate, vstate)) {
3759 			regs[rd] = NULL;
3760 			break;
3761 		}
3762 
3763 		if (!DTRACE_INSCRATCH(mstate, size)) {
3764 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3765 			regs[rd] = NULL;
3766 			break;
3767 		}
3768 
3769 		if (addr == NULL) {
3770 			/*
3771 			 * If the address specified is NULL, we use our saved
3772 			 * strtok pointer from the mstate.  Note that this
3773 			 * means that the saved strtok pointer is _only_
3774 			 * valid within multiple enablings of the same probe --
3775 			 * it behaves like an implicit clause-local variable.
3776 			 */
3777 			addr = mstate->dtms_strtok;
3778 		} else {
3779 			/*
3780 			 * If the user-specified address is non-NULL we must
3781 			 * access check it.  This is the only time we have
3782 			 * a chance to do so, since this address may reside
3783 			 * in the string table of this clause-- future calls
3784 			 * (when we fetch addr from mstate->dtms_strtok)
3785 			 * would fail this access check.
3786 			 */
3787 			if (!dtrace_strcanload(addr, size, mstate, vstate)) {
3788 				regs[rd] = NULL;
3789 				break;
3790 			}
3791 		}
3792 
3793 		/*
3794 		 * First, zero the token map, and then process the token
3795 		 * string -- setting a bit in the map for every character
3796 		 * found in the token string.
3797 		 */
3798 		for (i = 0; i < sizeof (tokmap); i++)
3799 			tokmap[i] = 0;
3800 
3801 		for (; tokaddr < toklimit; tokaddr++) {
3802 			if ((c = dtrace_load8(tokaddr)) == '\0')
3803 				break;
3804 
3805 			ASSERT((c >> 3) < sizeof (tokmap));
3806 			tokmap[c >> 3] |= (1 << (c & 0x7));
3807 		}
3808 
3809 		for (limit = addr + size; addr < limit; addr++) {
3810 			/*
3811 			 * We're looking for a character that is _not_ contained
3812 			 * in the token string.
3813 			 */
3814 			if ((c = dtrace_load8(addr)) == '\0')
3815 				break;
3816 
3817 			if (!(tokmap[c >> 3] & (1 << (c & 0x7))))
3818 				break;
3819 		}
3820 
3821 		if (c == '\0') {
3822 			/*
3823 			 * We reached the end of the string without finding
3824 			 * any character that was not in the token string.
3825 			 * We return NULL in this case, and we set the saved
3826 			 * address to NULL as well.
3827 			 */
3828 			regs[rd] = NULL;
3829 			mstate->dtms_strtok = NULL;
3830 			break;
3831 		}
3832 
3833 		/*
3834 		 * From here on, we're copying into the destination string.
3835 		 */
3836 		for (i = 0; addr < limit && i < size - 1; addr++) {
3837 			if ((c = dtrace_load8(addr)) == '\0')
3838 				break;
3839 
3840 			if (tokmap[c >> 3] & (1 << (c & 0x7)))
3841 				break;
3842 
3843 			ASSERT(i < size);
3844 			dest[i++] = c;
3845 		}
3846 
3847 		ASSERT(i < size);
3848 		dest[i] = '\0';
3849 		regs[rd] = (uintptr_t)dest;
3850 		mstate->dtms_scratch_ptr += size;
3851 		mstate->dtms_strtok = addr;
3852 		break;
3853 	}
3854 
3855 	case DIF_SUBR_SUBSTR: {
3856 		uintptr_t s = tupregs[0].dttk_value;
3857 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3858 		char *d = (char *)mstate->dtms_scratch_ptr;
3859 		int64_t index = (int64_t)tupregs[1].dttk_value;
3860 		int64_t remaining = (int64_t)tupregs[2].dttk_value;
3861 		size_t len = dtrace_strlen((char *)s, size);
3862 		int64_t i;
3863 
3864 		if (!dtrace_canload(s, len + 1, mstate, vstate)) {
3865 			regs[rd] = NULL;
3866 			break;
3867 		}
3868 
3869 		if (!DTRACE_INSCRATCH(mstate, size)) {
3870 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3871 			regs[rd] = NULL;
3872 			break;
3873 		}
3874 
3875 		if (nargs <= 2)
3876 			remaining = (int64_t)size;
3877 
3878 		if (index < 0) {
3879 			index += len;
3880 
3881 			if (index < 0 && index + remaining > 0) {
3882 				remaining += index;
3883 				index = 0;
3884 			}
3885 		}
3886 
3887 		if (index >= len || index < 0) {
3888 			remaining = 0;
3889 		} else if (remaining < 0) {
3890 			remaining += len - index;
3891 		} else if (index + remaining > size) {
3892 			remaining = size - index;
3893 		}
3894 
3895 		for (i = 0; i < remaining; i++) {
3896 			if ((d[i] = dtrace_load8(s + index + i)) == '\0')
3897 				break;
3898 		}
3899 
3900 		d[i] = '\0';
3901 
3902 		mstate->dtms_scratch_ptr += size;
3903 		regs[rd] = (uintptr_t)d;
3904 		break;
3905 	}
3906 
3907 	case DIF_SUBR_GETMAJOR:
3908 #ifdef _LP64
3909 		regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR64) & MAXMAJ64;
3910 #else
3911 		regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR) & MAXMAJ;
3912 #endif
3913 		break;
3914 
3915 	case DIF_SUBR_GETMINOR:
3916 #ifdef _LP64
3917 		regs[rd] = tupregs[0].dttk_value & MAXMIN64;
3918 #else
3919 		regs[rd] = tupregs[0].dttk_value & MAXMIN;
3920 #endif
3921 		break;
3922 
3923 	case DIF_SUBR_DDI_PATHNAME: {
3924 		/*
3925 		 * This one is a galactic mess.  We are going to roughly
3926 		 * emulate ddi_pathname(), but it's made more complicated
3927 		 * by the fact that we (a) want to include the minor name and
3928 		 * (b) must proceed iteratively instead of recursively.
3929 		 */
3930 		uintptr_t dest = mstate->dtms_scratch_ptr;
3931 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3932 		char *start = (char *)dest, *end = start + size - 1;
3933 		uintptr_t daddr = tupregs[0].dttk_value;
3934 		int64_t minor = (int64_t)tupregs[1].dttk_value;
3935 		char *s;
3936 		int i, len, depth = 0;
3937 
3938 		/*
3939 		 * Due to all the pointer jumping we do and context we must
3940 		 * rely upon, we just mandate that the user must have kernel
3941 		 * read privileges to use this routine.
3942 		 */
3943 		if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) == 0) {
3944 			*flags |= CPU_DTRACE_KPRIV;
3945 			*illval = daddr;
3946 			regs[rd] = NULL;
3947 		}
3948 
3949 		if (!DTRACE_INSCRATCH(mstate, size)) {
3950 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3951 			regs[rd] = NULL;
3952 			break;
3953 		}
3954 
3955 		*end = '\0';
3956 
3957 		/*
3958 		 * We want to have a name for the minor.  In order to do this,
3959 		 * we need to walk the minor list from the devinfo.  We want
3960 		 * to be sure that we don't infinitely walk a circular list,
3961 		 * so we check for circularity by sending a scout pointer
3962 		 * ahead two elements for every element that we iterate over;
3963 		 * if the list is circular, these will ultimately point to the
3964 		 * same element.  You may recognize this little trick as the
3965 		 * answer to a stupid interview question -- one that always
3966 		 * seems to be asked by those who had to have it laboriously
3967 		 * explained to them, and who can't even concisely describe
3968 		 * the conditions under which one would be forced to resort to
3969 		 * this technique.  Needless to say, those conditions are
3970 		 * found here -- and probably only here.  Is this the only use
3971 		 * of this infamous trick in shipping, production code?  If it
3972 		 * isn't, it probably should be...
3973 		 */
3974 		if (minor != -1) {
3975 			uintptr_t maddr = dtrace_loadptr(daddr +
3976 			    offsetof(struct dev_info, devi_minor));
3977 
3978 			uintptr_t next = offsetof(struct ddi_minor_data, next);
3979 			uintptr_t name = offsetof(struct ddi_minor_data,
3980 			    d_minor) + offsetof(struct ddi_minor, name);
3981 			uintptr_t dev = offsetof(struct ddi_minor_data,
3982 			    d_minor) + offsetof(struct ddi_minor, dev);
3983 			uintptr_t scout;
3984 
3985 			if (maddr != NULL)
3986 				scout = dtrace_loadptr(maddr + next);
3987 
3988 			while (maddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
3989 				uint64_t m;
3990 #ifdef _LP64
3991 				m = dtrace_load64(maddr + dev) & MAXMIN64;
3992 #else
3993 				m = dtrace_load32(maddr + dev) & MAXMIN;
3994 #endif
3995 				if (m != minor) {
3996 					maddr = dtrace_loadptr(maddr + next);
3997 
3998 					if (scout == NULL)
3999 						continue;
4000 
4001 					scout = dtrace_loadptr(scout + next);
4002 
4003 					if (scout == NULL)
4004 						continue;
4005 
4006 					scout = dtrace_loadptr(scout + next);
4007 
4008 					if (scout == NULL)
4009 						continue;
4010 
4011 					if (scout == maddr) {
4012 						*flags |= CPU_DTRACE_ILLOP;
4013 						break;
4014 					}
4015 
4016 					continue;
4017 				}
4018 
4019 				/*
4020 				 * We have the minor data.  Now we need to
4021 				 * copy the minor's name into the end of the
4022 				 * pathname.
4023 				 */
4024 				s = (char *)dtrace_loadptr(maddr + name);
4025 				len = dtrace_strlen(s, size);
4026 
4027 				if (*flags & CPU_DTRACE_FAULT)
4028 					break;
4029 
4030 				if (len != 0) {
4031 					if ((end -= (len + 1)) < start)
4032 						break;
4033 
4034 					*end = ':';
4035 				}
4036 
4037 				for (i = 1; i <= len; i++)
4038 					end[i] = dtrace_load8((uintptr_t)s++);
4039 				break;
4040 			}
4041 		}
4042 
4043 		while (daddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
4044 			ddi_node_state_t devi_state;
4045 
4046 			devi_state = dtrace_load32(daddr +
4047 			    offsetof(struct dev_info, devi_node_state));
4048 
4049 			if (*flags & CPU_DTRACE_FAULT)
4050 				break;
4051 
4052 			if (devi_state >= DS_INITIALIZED) {
4053 				s = (char *)dtrace_loadptr(daddr +
4054 				    offsetof(struct dev_info, devi_addr));
4055 				len = dtrace_strlen(s, size);
4056 
4057 				if (*flags & CPU_DTRACE_FAULT)
4058 					break;
4059 
4060 				if (len != 0) {
4061 					if ((end -= (len + 1)) < start)
4062 						break;
4063 
4064 					*end = '@';
4065 				}
4066 
4067 				for (i = 1; i <= len; i++)
4068 					end[i] = dtrace_load8((uintptr_t)s++);
4069 			}
4070 
4071 			/*
4072 			 * Now for the node name...
4073 			 */
4074 			s = (char *)dtrace_loadptr(daddr +
4075 			    offsetof(struct dev_info, devi_node_name));
4076 
4077 			daddr = dtrace_loadptr(daddr +
4078 			    offsetof(struct dev_info, devi_parent));
4079 
4080 			/*
4081 			 * If our parent is NULL (that is, if we're the root
4082 			 * node), we're going to use the special path
4083 			 * "devices".
4084 			 */
4085 			if (daddr == NULL)
4086 				s = "devices";
4087 
4088 			len = dtrace_strlen(s, size);
4089 			if (*flags & CPU_DTRACE_FAULT)
4090 				break;
4091 
4092 			if ((end -= (len + 1)) < start)
4093 				break;
4094 
4095 			for (i = 1; i <= len; i++)
4096 				end[i] = dtrace_load8((uintptr_t)s++);
4097 			*end = '/';
4098 
4099 			if (depth++ > dtrace_devdepth_max) {
4100 				*flags |= CPU_DTRACE_ILLOP;
4101 				break;
4102 			}
4103 		}
4104 
4105 		if (end < start)
4106 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4107 
4108 		if (daddr == NULL) {
4109 			regs[rd] = (uintptr_t)end;
4110 			mstate->dtms_scratch_ptr += size;
4111 		}
4112 
4113 		break;
4114 	}
4115 
4116 	case DIF_SUBR_STRJOIN: {
4117 		char *d = (char *)mstate->dtms_scratch_ptr;
4118 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4119 		uintptr_t s1 = tupregs[0].dttk_value;
4120 		uintptr_t s2 = tupregs[1].dttk_value;
4121 		int i = 0;
4122 
4123 		if (!dtrace_strcanload(s1, size, mstate, vstate) ||
4124 		    !dtrace_strcanload(s2, size, mstate, vstate)) {
4125 			regs[rd] = NULL;
4126 			break;
4127 		}
4128 
4129 		if (!DTRACE_INSCRATCH(mstate, size)) {
4130 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4131 			regs[rd] = NULL;
4132 			break;
4133 		}
4134 
4135 		for (;;) {
4136 			if (i >= size) {
4137 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4138 				regs[rd] = NULL;
4139 				break;
4140 			}
4141 
4142 			if ((d[i++] = dtrace_load8(s1++)) == '\0') {
4143 				i--;
4144 				break;
4145 			}
4146 		}
4147 
4148 		for (;;) {
4149 			if (i >= size) {
4150 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4151 				regs[rd] = NULL;
4152 				break;
4153 			}
4154 
4155 			if ((d[i++] = dtrace_load8(s2++)) == '\0')
4156 				break;
4157 		}
4158 
4159 		if (i < size) {
4160 			mstate->dtms_scratch_ptr += i;
4161 			regs[rd] = (uintptr_t)d;
4162 		}
4163 
4164 		break;
4165 	}
4166 
4167 	case DIF_SUBR_LLTOSTR: {
4168 		int64_t i = (int64_t)tupregs[0].dttk_value;
4169 		int64_t val = i < 0 ? i * -1 : i;
4170 		uint64_t size = 22;	/* enough room for 2^64 in decimal */
4171 		char *end = (char *)mstate->dtms_scratch_ptr + size - 1;
4172 
4173 		if (!DTRACE_INSCRATCH(mstate, size)) {
4174 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4175 			regs[rd] = NULL;
4176 			break;
4177 		}
4178 
4179 		for (*end-- = '\0'; val; val /= 10)
4180 			*end-- = '0' + (val % 10);
4181 
4182 		if (i == 0)
4183 			*end-- = '0';
4184 
4185 		if (i < 0)
4186 			*end-- = '-';
4187 
4188 		regs[rd] = (uintptr_t)end + 1;
4189 		mstate->dtms_scratch_ptr += size;
4190 		break;
4191 	}
4192 
4193 	case DIF_SUBR_HTONS:
4194 	case DIF_SUBR_NTOHS:
4195 #ifdef _BIG_ENDIAN
4196 		regs[rd] = (uint16_t)tupregs[0].dttk_value;
4197 #else
4198 		regs[rd] = DT_BSWAP_16((uint16_t)tupregs[0].dttk_value);
4199 #endif
4200 		break;
4201 
4202 
4203 	case DIF_SUBR_HTONL:
4204 	case DIF_SUBR_NTOHL:
4205 #ifdef _BIG_ENDIAN
4206 		regs[rd] = (uint32_t)tupregs[0].dttk_value;
4207 #else
4208 		regs[rd] = DT_BSWAP_32((uint32_t)tupregs[0].dttk_value);
4209 #endif
4210 		break;
4211 
4212 
4213 	case DIF_SUBR_HTONLL:
4214 	case DIF_SUBR_NTOHLL:
4215 #ifdef _BIG_ENDIAN
4216 		regs[rd] = (uint64_t)tupregs[0].dttk_value;
4217 #else
4218 		regs[rd] = DT_BSWAP_64((uint64_t)tupregs[0].dttk_value);
4219 #endif
4220 		break;
4221 
4222 
4223 	case DIF_SUBR_DIRNAME:
4224 	case DIF_SUBR_BASENAME: {
4225 		char *dest = (char *)mstate->dtms_scratch_ptr;
4226 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4227 		uintptr_t src = tupregs[0].dttk_value;
4228 		int i, j, len = dtrace_strlen((char *)src, size);
4229 		int lastbase = -1, firstbase = -1, lastdir = -1;
4230 		int start, end;
4231 
4232 		if (!dtrace_canload(src, len + 1, mstate, vstate)) {
4233 			regs[rd] = NULL;
4234 			break;
4235 		}
4236 
4237 		if (!DTRACE_INSCRATCH(mstate, size)) {
4238 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4239 			regs[rd] = NULL;
4240 			break;
4241 		}
4242 
4243 		/*
4244 		 * The basename and dirname for a zero-length string is
4245 		 * defined to be "."
4246 		 */
4247 		if (len == 0) {
4248 			len = 1;
4249 			src = (uintptr_t)".";
4250 		}
4251 
4252 		/*
4253 		 * Start from the back of the string, moving back toward the
4254 		 * front until we see a character that isn't a slash.  That
4255 		 * character is the last character in the basename.
4256 		 */
4257 		for (i = len - 1; i >= 0; i--) {
4258 			if (dtrace_load8(src + i) != '/')
4259 				break;
4260 		}
4261 
4262 		if (i >= 0)
4263 			lastbase = i;
4264 
4265 		/*
4266 		 * Starting from the last character in the basename, move
4267 		 * towards the front until we find a slash.  The character
4268 		 * that we processed immediately before that is the first
4269 		 * character in the basename.
4270 		 */
4271 		for (; i >= 0; i--) {
4272 			if (dtrace_load8(src + i) == '/')
4273 				break;
4274 		}
4275 
4276 		if (i >= 0)
4277 			firstbase = i + 1;
4278 
4279 		/*
4280 		 * Now keep going until we find a non-slash character.  That
4281 		 * character is the last character in the dirname.
4282 		 */
4283 		for (; i >= 0; i--) {
4284 			if (dtrace_load8(src + i) != '/')
4285 				break;
4286 		}
4287 
4288 		if (i >= 0)
4289 			lastdir = i;
4290 
4291 		ASSERT(!(lastbase == -1 && firstbase != -1));
4292 		ASSERT(!(firstbase == -1 && lastdir != -1));
4293 
4294 		if (lastbase == -1) {
4295 			/*
4296 			 * We didn't find a non-slash character.  We know that
4297 			 * the length is non-zero, so the whole string must be
4298 			 * slashes.  In either the dirname or the basename
4299 			 * case, we return '/'.
4300 			 */
4301 			ASSERT(firstbase == -1);
4302 			firstbase = lastbase = lastdir = 0;
4303 		}
4304 
4305 		if (firstbase == -1) {
4306 			/*
4307 			 * The entire string consists only of a basename
4308 			 * component.  If we're looking for dirname, we need
4309 			 * to change our string to be just "."; if we're
4310 			 * looking for a basename, we'll just set the first
4311 			 * character of the basename to be 0.
4312 			 */
4313 			if (subr == DIF_SUBR_DIRNAME) {
4314 				ASSERT(lastdir == -1);
4315 				src = (uintptr_t)".";
4316 				lastdir = 0;
4317 			} else {
4318 				firstbase = 0;
4319 			}
4320 		}
4321 
4322 		if (subr == DIF_SUBR_DIRNAME) {
4323 			if (lastdir == -1) {
4324 				/*
4325 				 * We know that we have a slash in the name --
4326 				 * or lastdir would be set to 0, above.  And
4327 				 * because lastdir is -1, we know that this
4328 				 * slash must be the first character.  (That
4329 				 * is, the full string must be of the form
4330 				 * "/basename".)  In this case, the last
4331 				 * character of the directory name is 0.
4332 				 */
4333 				lastdir = 0;
4334 			}
4335 
4336 			start = 0;
4337 			end = lastdir;
4338 		} else {
4339 			ASSERT(subr == DIF_SUBR_BASENAME);
4340 			ASSERT(firstbase != -1 && lastbase != -1);
4341 			start = firstbase;
4342 			end = lastbase;
4343 		}
4344 
4345 		for (i = start, j = 0; i <= end && j < size - 1; i++, j++)
4346 			dest[j] = dtrace_load8(src + i);
4347 
4348 		dest[j] = '\0';
4349 		regs[rd] = (uintptr_t)dest;
4350 		mstate->dtms_scratch_ptr += size;
4351 		break;
4352 	}
4353 
4354 	case DIF_SUBR_CLEANPATH: {
4355 		char *dest = (char *)mstate->dtms_scratch_ptr, c;
4356 		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4357 		uintptr_t src = tupregs[0].dttk_value;
4358 		int i = 0, j = 0;
4359 
4360 		if (!dtrace_strcanload(src, size, mstate, vstate)) {
4361 			regs[rd] = NULL;
4362 			break;
4363 		}
4364 
4365 		if (!DTRACE_INSCRATCH(mstate, size)) {
4366 			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4367 			regs[rd] = NULL;
4368 			break;
4369 		}
4370 
4371 		/*
4372 		 * Move forward, loading each character.
4373 		 */
4374 		do {
4375 			c = dtrace_load8(src + i++);
4376 next:
4377 			if (j + 5 >= size)	/* 5 = strlen("/..c\0") */
4378 				break;
4379 
4380 			if (c != '/') {
4381 				dest[j++] = c;
4382 				continue;
4383 			}
4384 
4385 			c = dtrace_load8(src + i++);
4386 
4387 			if (c == '/') {
4388 				/*
4389 				 * We have two slashes -- we can just advance
4390 				 * to the next character.
4391 				 */
4392 				goto next;
4393 			}
4394 
4395 			if (c != '.') {
4396 				/*
4397 				 * This is not "." and it's not ".." -- we can
4398 				 * just store the "/" and this character and
4399 				 * drive on.
4400 				 */
4401 				dest[j++] = '/';
4402 				dest[j++] = c;
4403 				continue;
4404 			}
4405 
4406 			c = dtrace_load8(src + i++);
4407 
4408 			if (c == '/') {
4409 				/*
4410 				 * This is a "/./" component.  We're not going
4411 				 * to store anything in the destination buffer;
4412 				 * we're just going to go to the next component.
4413 				 */
4414 				goto next;
4415 			}
4416 
4417 			if (c != '.') {
4418 				/*
4419 				 * This is not ".." -- we can just store the
4420 				 * "/." and this character and continue
4421 				 * processing.
4422 				 */
4423 				dest[j++] = '/';
4424 				dest[j++] = '.';
4425 				dest[j++] = c;
4426 				continue;
4427 			}
4428 
4429 			c = dtrace_load8(src + i++);
4430 
4431 			if (c != '/' && c != '\0') {
4432 				/*
4433 				 * This is not ".." -- it's "..[mumble]".
4434 				 * We'll store the "/.." and this character
4435 				 * and continue processing.
4436 				 */
4437 				dest[j++] = '/';
4438 				dest[j++] = '.';
4439 				dest[j++] = '.';
4440 				dest[j++] = c;
4441 				continue;
4442 			}
4443 
4444 			/*
4445 			 * This is "/../" or "/..\0".  We need to back up
4446 			 * our destination pointer until we find a "/".
4447 			 */
4448 			i--;
4449 			while (j != 0 && dest[--j] != '/')
4450 				continue;
4451 
4452 			if (c == '\0')
4453 				dest[++j] = '/';
4454 		} while (c != '\0');
4455 
4456 		dest[j] = '\0';
4457 		regs[rd] = (uintptr_t)dest;
4458 		mstate->dtms_scratch_ptr += size;
4459 		break;
4460 	}
4461 
4462 	case DIF_SUBR_INET_NTOA:
4463 	case DIF_SUBR_INET_NTOA6:
4464 	case DIF_SUBR_INET_NTOP: {
4465 		size_t size;
4466 		int af, argi, i;
4467 		char *base, *end;
4468 
4469 		if (subr == DIF_SUBR_INET_NTOP) {
4470 			af = (int)tupregs[0].dttk_value;
4471 			argi = 1;
4472 		} else {
4473 			af = subr == DIF_SUBR_INET_NTOA ? AF_INET: AF_INET6;
4474 			argi = 0;
4475 		}
4476 
4477 		if (af == AF_INET) {
4478 			ipaddr_t ip4;
4479 			uint8_t *ptr8, val;
4480 
4481 			/*
4482 			 * Safely load the IPv4 address.
4483 			 */
4484 			ip4 = dtrace_load32(tupregs[argi].dttk_value);
4485 
4486 			/*
4487 			 * Check an IPv4 string will fit in scratch.
4488 			 */
4489 			size = INET_ADDRSTRLEN;
4490 			if (!DTRACE_INSCRATCH(mstate, size)) {
4491 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4492 				regs[rd] = NULL;
4493 				break;
4494 			}
4495 			base = (char *)mstate->dtms_scratch_ptr;
4496 			end = (char *)mstate->dtms_scratch_ptr + size - 1;
4497 
4498 			/*
4499 			 * Stringify as a dotted decimal quad.
4500 			 */
4501 			*end-- = '\0';
4502 			ptr8 = (uint8_t *)&ip4;
4503 			for (i = 3; i >= 0; i--) {
4504 				val = ptr8[i];
4505 
4506 				if (val == 0) {
4507 					*end-- = '0';
4508 				} else {
4509 					for (; val; val /= 10) {
4510 						*end-- = '0' + (val % 10);
4511 					}
4512 				}
4513 
4514 				if (i > 0)
4515 					*end-- = '.';
4516 			}
4517 			ASSERT(end + 1 >= base);
4518 
4519 		} else if (af == AF_INET6) {
4520 			struct in6_addr ip6;
4521 			int firstzero, tryzero, numzero, v6end;
4522 			uint16_t val;
4523 			const char digits[] = "0123456789abcdef";
4524 
4525 			/*
4526 			 * Stringify using RFC 1884 convention 2 - 16 bit
4527 			 * hexadecimal values with a zero-run compression.
4528 			 * Lower case hexadecimal digits are used.
4529 			 * 	eg, fe80::214:4fff:fe0b:76c8.
4530 			 * The IPv4 embedded form is returned for inet_ntop,
4531 			 * just the IPv4 string is returned for inet_ntoa6.
4532 			 */
4533 
4534 			/*
4535 			 * Safely load the IPv6 address.
4536 			 */
4537 			dtrace_bcopy(
4538 			    (void *)(uintptr_t)tupregs[argi].dttk_value,
4539 			    (void *)(uintptr_t)&ip6, sizeof (struct in6_addr));
4540 
4541 			/*
4542 			 * Check an IPv6 string will fit in scratch.
4543 			 */
4544 			size = INET6_ADDRSTRLEN;
4545 			if (!DTRACE_INSCRATCH(mstate, size)) {
4546 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4547 				regs[rd] = NULL;
4548 				break;
4549 			}
4550 			base = (char *)mstate->dtms_scratch_ptr;
4551 			end = (char *)mstate->dtms_scratch_ptr + size - 1;
4552 			*end-- = '\0';
4553 
4554 			/*
4555 			 * Find the longest run of 16 bit zero values
4556 			 * for the single allowed zero compression - "::".
4557 			 */
4558 			firstzero = -1;
4559 			tryzero = -1;
4560 			numzero = 1;
4561 			for (i = 0; i < sizeof (struct in6_addr); i++) {
4562 				if (ip6._S6_un._S6_u8[i] == 0 &&
4563 				    tryzero == -1 && i % 2 == 0) {
4564 					tryzero = i;
4565 					continue;
4566 				}
4567 
4568 				if (tryzero != -1 &&
4569 				    (ip6._S6_un._S6_u8[i] != 0 ||
4570 				    i == sizeof (struct in6_addr) - 1)) {
4571 
4572 					if (i - tryzero <= numzero) {
4573 						tryzero = -1;
4574 						continue;
4575 					}
4576 
4577 					firstzero = tryzero;
4578 					numzero = i - i % 2 - tryzero;
4579 					tryzero = -1;
4580 
4581 					if (ip6._S6_un._S6_u8[i] == 0 &&
4582 					    i == sizeof (struct in6_addr) - 1)
4583 						numzero += 2;
4584 				}
4585 			}
4586 			ASSERT(firstzero + numzero <= sizeof (struct in6_addr));
4587 
4588 			/*
4589 			 * Check for an IPv4 embedded address.
4590 			 */
4591 			v6end = sizeof (struct in6_addr) - 2;
4592 			if (IN6_IS_ADDR_V4MAPPED(&ip6) ||
4593 			    IN6_IS_ADDR_V4COMPAT(&ip6)) {
4594 				for (i = sizeof (struct in6_addr) - 1;
4595 				    i >= DTRACE_V4MAPPED_OFFSET; i--) {
4596 					ASSERT(end >= base);
4597 
4598 					val = ip6._S6_un._S6_u8[i];
4599 
4600 					if (val == 0) {
4601 						*end-- = '0';
4602 					} else {
4603 						for (; val; val /= 10) {
4604 							*end-- = '0' + val % 10;
4605 						}
4606 					}
4607 
4608 					if (i > DTRACE_V4MAPPED_OFFSET)
4609 						*end-- = '.';
4610 				}
4611 
4612 				if (subr == DIF_SUBR_INET_NTOA6)
4613 					goto inetout;
4614 
4615 				/*
4616 				 * Set v6end to skip the IPv4 address that
4617 				 * we have already stringified.
4618 				 */
4619 				v6end = 10;
4620 			}
4621 
4622 			/*
4623 			 * Build the IPv6 string by working through the
4624 			 * address in reverse.
4625 			 */
4626 			for (i = v6end; i >= 0; i -= 2) {
4627 				ASSERT(end >= base);
4628 
4629 				if (i == firstzero + numzero - 2) {
4630 					*end-- = ':';
4631 					*end-- = ':';
4632 					i -= numzero - 2;
4633 					continue;
4634 				}
4635 
4636 				if (i < 14 && i != firstzero - 2)
4637 					*end-- = ':';
4638 
4639 				val = (ip6._S6_un._S6_u8[i] << 8) +
4640 				    ip6._S6_un._S6_u8[i + 1];
4641 
4642 				if (val == 0) {
4643 					*end-- = '0';
4644 				} else {
4645 					for (; val; val /= 16) {
4646 						*end-- = digits[val % 16];
4647 					}
4648 				}
4649 			}
4650 			ASSERT(end + 1 >= base);
4651 
4652 		} else {
4653 			/*
4654 			 * The user didn't use AH_INET or AH_INET6.
4655 			 */
4656 			DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
4657 			regs[rd] = NULL;
4658 			break;
4659 		}
4660 
4661 inetout:	regs[rd] = (uintptr_t)end + 1;
4662 		mstate->dtms_scratch_ptr += size;
4663 		break;
4664 	}
4665 
4666 	}
4667 }
4668 
4669 /*
4670  * Emulate the execution of DTrace IR instructions specified by the given
4671  * DIF object.  This function is deliberately void of assertions as all of
4672  * the necessary checks are handled by a call to dtrace_difo_validate().
4673  */
4674 static uint64_t
4675 dtrace_dif_emulate(dtrace_difo_t *difo, dtrace_mstate_t *mstate,
4676     dtrace_vstate_t *vstate, dtrace_state_t *state)
4677 {
4678 	const dif_instr_t *text = difo->dtdo_buf;
4679 	const uint_t textlen = difo->dtdo_len;
4680 	const char *strtab = difo->dtdo_strtab;
4681 	const uint64_t *inttab = difo->dtdo_inttab;
4682 
4683 	uint64_t rval = 0;
4684 	dtrace_statvar_t *svar;
4685 	dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
4686 	dtrace_difv_t *v;
4687 	volatile uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
4688 	volatile uintptr_t *illval = &cpu_core[CPU->cpu_id].cpuc_dtrace_illval;
4689 
4690 	dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
4691 	uint64_t regs[DIF_DIR_NREGS];
4692 	uint64_t *tmp;
4693 
4694 	uint8_t cc_n = 0, cc_z = 0, cc_v = 0, cc_c = 0;
4695 	int64_t cc_r;
4696 	uint_t pc = 0, id, opc;
4697 	uint8_t ttop = 0;
4698 	dif_instr_t instr;
4699 	uint_t r1, r2, rd;
4700 
4701 	/*
4702 	 * We stash the current DIF object into the machine state: we need it
4703 	 * for subsequent access checking.
4704 	 */
4705 	mstate->dtms_difo = difo;
4706 
4707 	regs[DIF_REG_R0] = 0; 		/* %r0 is fixed at zero */
4708 
4709 	while (pc < textlen && !(*flags & CPU_DTRACE_FAULT)) {
4710 		opc = pc;
4711 
4712 		instr = text[pc++];
4713 		r1 = DIF_INSTR_R1(instr);
4714 		r2 = DIF_INSTR_R2(instr);
4715 		rd = DIF_INSTR_RD(instr);
4716 
4717 		switch (DIF_INSTR_OP(instr)) {
4718 		case DIF_OP_OR:
4719 			regs[rd] = regs[r1] | regs[r2];
4720 			break;
4721 		case DIF_OP_XOR:
4722 			regs[rd] = regs[r1] ^ regs[r2];
4723 			break;
4724 		case DIF_OP_AND:
4725 			regs[rd] = regs[r1] & regs[r2];
4726 			break;
4727 		case DIF_OP_SLL:
4728 			regs[rd] = regs[r1] << regs[r2];
4729 			break;
4730 		case DIF_OP_SRL:
4731 			regs[rd] = regs[r1] >> regs[r2];
4732 			break;
4733 		case DIF_OP_SUB:
4734 			regs[rd] = regs[r1] - regs[r2];
4735 			break;
4736 		case DIF_OP_ADD:
4737 			regs[rd] = regs[r1] + regs[r2];
4738 			break;
4739 		case DIF_OP_MUL:
4740 			regs[rd] = regs[r1] * regs[r2];
4741 			break;
4742 		case DIF_OP_SDIV:
4743 			if (regs[r2] == 0) {
4744 				regs[rd] = 0;
4745 				*flags |= CPU_DTRACE_DIVZERO;
4746 			} else {
4747 				regs[rd] = (int64_t)regs[r1] /
4748 				    (int64_t)regs[r2];
4749 			}
4750 			break;
4751 
4752 		case DIF_OP_UDIV:
4753 			if (regs[r2] == 0) {
4754 				regs[rd] = 0;
4755 				*flags |= CPU_DTRACE_DIVZERO;
4756 			} else {
4757 				regs[rd] = regs[r1] / regs[r2];
4758 			}
4759 			break;
4760 
4761 		case DIF_OP_SREM:
4762 			if (regs[r2] == 0) {
4763 				regs[rd] = 0;
4764 				*flags |= CPU_DTRACE_DIVZERO;
4765 			} else {
4766 				regs[rd] = (int64_t)regs[r1] %
4767 				    (int64_t)regs[r2];
4768 			}
4769 			break;
4770 
4771 		case DIF_OP_UREM:
4772 			if (regs[r2] == 0) {
4773 				regs[rd] = 0;
4774 				*flags |= CPU_DTRACE_DIVZERO;
4775 			} else {
4776 				regs[rd] = regs[r1] % regs[r2];
4777 			}
4778 			break;
4779 
4780 		case DIF_OP_NOT:
4781 			regs[rd] = ~regs[r1];
4782 			break;
4783 		case DIF_OP_MOV:
4784 			regs[rd] = regs[r1];
4785 			break;
4786 		case DIF_OP_CMP:
4787 			cc_r = regs[r1] - regs[r2];
4788 			cc_n = cc_r < 0;
4789 			cc_z = cc_r == 0;
4790 			cc_v = 0;
4791 			cc_c = regs[r1] < regs[r2];
4792 			break;
4793 		case DIF_OP_TST:
4794 			cc_n = cc_v = cc_c = 0;
4795 			cc_z = regs[r1] == 0;
4796 			break;
4797 		case DIF_OP_BA:
4798 			pc = DIF_INSTR_LABEL(instr);
4799 			break;
4800 		case DIF_OP_BE:
4801 			if (cc_z)
4802 				pc = DIF_INSTR_LABEL(instr);
4803 			break;
4804 		case DIF_OP_BNE:
4805 			if (cc_z == 0)
4806 				pc = DIF_INSTR_LABEL(instr);
4807 			break;
4808 		case DIF_OP_BG:
4809 			if ((cc_z | (cc_n ^ cc_v)) == 0)
4810 				pc = DIF_INSTR_LABEL(instr);
4811 			break;
4812 		case DIF_OP_BGU:
4813 			if ((cc_c | cc_z) == 0)
4814 				pc = DIF_INSTR_LABEL(instr);
4815 			break;
4816 		case DIF_OP_BGE:
4817 			if ((cc_n ^ cc_v) == 0)
4818 				pc = DIF_INSTR_LABEL(instr);
4819 			break;
4820 		case DIF_OP_BGEU:
4821 			if (cc_c == 0)
4822 				pc = DIF_INSTR_LABEL(instr);
4823 			break;
4824 		case DIF_OP_BL:
4825 			if (cc_n ^ cc_v)
4826 				pc = DIF_INSTR_LABEL(instr);
4827 			break;
4828 		case DIF_OP_BLU:
4829 			if (cc_c)
4830 				pc = DIF_INSTR_LABEL(instr);
4831 			break;
4832 		case DIF_OP_BLE:
4833 			if (cc_z | (cc_n ^ cc_v))
4834 				pc = DIF_INSTR_LABEL(instr);
4835 			break;
4836 		case DIF_OP_BLEU:
4837 			if (cc_c | cc_z)
4838 				pc = DIF_INSTR_LABEL(instr);
4839 			break;
4840 		case DIF_OP_RLDSB:
4841 			if (!dtrace_canstore(regs[r1], 1, mstate, vstate)) {
4842 				*flags |= CPU_DTRACE_KPRIV;
4843 				*illval = regs[r1];
4844 				break;
4845 			}
4846 			/*FALLTHROUGH*/
4847 		case DIF_OP_LDSB:
4848 			regs[rd] = (int8_t)dtrace_load8(regs[r1]);
4849 			break;
4850 		case DIF_OP_RLDSH:
4851 			if (!dtrace_canstore(regs[r1], 2, mstate, vstate)) {
4852 				*flags |= CPU_DTRACE_KPRIV;
4853 				*illval = regs[r1];
4854 				break;
4855 			}
4856 			/*FALLTHROUGH*/
4857 		case DIF_OP_LDSH:
4858 			regs[rd] = (int16_t)dtrace_load16(regs[r1]);
4859 			break;
4860 		case DIF_OP_RLDSW:
4861 			if (!dtrace_canstore(regs[r1], 4, mstate, vstate)) {
4862 				*flags |= CPU_DTRACE_KPRIV;
4863 				*illval = regs[r1];
4864 				break;
4865 			}
4866 			/*FALLTHROUGH*/
4867 		case DIF_OP_LDSW:
4868 			regs[rd] = (int32_t)dtrace_load32(regs[r1]);
4869 			break;
4870 		case DIF_OP_RLDUB:
4871 			if (!dtrace_canstore(regs[r1], 1, mstate, vstate)) {
4872 				*flags |= CPU_DTRACE_KPRIV;
4873 				*illval = regs[r1];
4874 				break;
4875 			}
4876 			/*FALLTHROUGH*/
4877 		case DIF_OP_LDUB:
4878 			regs[rd] = dtrace_load8(regs[r1]);
4879 			break;
4880 		case DIF_OP_RLDUH:
4881 			if (!dtrace_canstore(regs[r1], 2, mstate, vstate)) {
4882 				*flags |= CPU_DTRACE_KPRIV;
4883 				*illval = regs[r1];
4884 				break;
4885 			}
4886 			/*FALLTHROUGH*/
4887 		case DIF_OP_LDUH:
4888 			regs[rd] = dtrace_load16(regs[r1]);
4889 			break;
4890 		case DIF_OP_RLDUW:
4891 			if (!dtrace_canstore(regs[r1], 4, mstate, vstate)) {
4892 				*flags |= CPU_DTRACE_KPRIV;
4893 				*illval = regs[r1];
4894 				break;
4895 			}
4896 			/*FALLTHROUGH*/
4897 		case DIF_OP_LDUW:
4898 			regs[rd] = dtrace_load32(regs[r1]);
4899 			break;
4900 		case DIF_OP_RLDX:
4901 			if (!dtrace_canstore(regs[r1], 8, mstate, vstate)) {
4902 				*flags |= CPU_DTRACE_KPRIV;
4903 				*illval = regs[r1];
4904 				break;
4905 			}
4906 			/*FALLTHROUGH*/
4907 		case DIF_OP_LDX:
4908 			regs[rd] = dtrace_load64(regs[r1]);
4909 			break;
4910 		case DIF_OP_ULDSB:
4911 			regs[rd] = (int8_t)
4912 			    dtrace_fuword8((void *)(uintptr_t)regs[r1]);
4913 			break;
4914 		case DIF_OP_ULDSH:
4915 			regs[rd] = (int16_t)
4916 			    dtrace_fuword16((void *)(uintptr_t)regs[r1]);
4917 			break;
4918 		case DIF_OP_ULDSW:
4919 			regs[rd] = (int32_t)
4920 			    dtrace_fuword32((void *)(uintptr_t)regs[r1]);
4921 			break;
4922 		case DIF_OP_ULDUB:
4923 			regs[rd] =
4924 			    dtrace_fuword8((void *)(uintptr_t)regs[r1]);
4925 			break;
4926 		case DIF_OP_ULDUH:
4927 			regs[rd] =
4928 			    dtrace_fuword16((void *)(uintptr_t)regs[r1]);
4929 			break;
4930 		case DIF_OP_ULDUW:
4931 			regs[rd] =
4932 			    dtrace_fuword32((void *)(uintptr_t)regs[r1]);
4933 			break;
4934 		case DIF_OP_ULDX:
4935 			regs[rd] =
4936 			    dtrace_fuword64((void *)(uintptr_t)regs[r1]);
4937 			break;
4938 		case DIF_OP_RET:
4939 			rval = regs[rd];
4940 			pc = textlen;
4941 			break;
4942 		case DIF_OP_NOP:
4943 			break;
4944 		case DIF_OP_SETX:
4945 			regs[rd] = inttab[DIF_INSTR_INTEGER(instr)];
4946 			break;
4947 		case DIF_OP_SETS:
4948 			regs[rd] = (uint64_t)(uintptr_t)
4949 			    (strtab + DIF_INSTR_STRING(instr));
4950 			break;
4951 		case DIF_OP_SCMP: {
4952 			size_t sz = state->dts_options[DTRACEOPT_STRSIZE];
4953 			uintptr_t s1 = regs[r1];
4954 			uintptr_t s2 = regs[r2];
4955 
4956 			if (s1 != NULL &&
4957 			    !dtrace_strcanload(s1, sz, mstate, vstate))
4958 				break;
4959 			if (s2 != NULL &&
4960 			    !dtrace_strcanload(s2, sz, mstate, vstate))
4961 				break;
4962 
4963 			cc_r = dtrace_strncmp((char *)s1, (char *)s2, sz);
4964 
4965 			cc_n = cc_r < 0;
4966 			cc_z = cc_r == 0;
4967 			cc_v = cc_c = 0;
4968 			break;
4969 		}
4970 		case DIF_OP_LDGA:
4971 			regs[rd] = dtrace_dif_variable(mstate, state,
4972 			    r1, regs[r2]);
4973 			break;
4974 		case DIF_OP_LDGS:
4975 			id = DIF_INSTR_VAR(instr);
4976 
4977 			if (id >= DIF_VAR_OTHER_UBASE) {
4978 				uintptr_t a;
4979 
4980 				id -= DIF_VAR_OTHER_UBASE;
4981 				svar = vstate->dtvs_globals[id];
4982 				ASSERT(svar != NULL);
4983 				v = &svar->dtsv_var;
4984 
4985 				if (!(v->dtdv_type.dtdt_flags & DIF_TF_BYREF)) {
4986 					regs[rd] = svar->dtsv_data;
4987 					break;
4988 				}
4989 
4990 				a = (uintptr_t)svar->dtsv_data;
4991 
4992 				if (*(uint8_t *)a == UINT8_MAX) {
4993 					/*
4994 					 * If the 0th byte is set to UINT8_MAX
4995 					 * then this is to be treated as a
4996 					 * reference to a NULL variable.
4997 					 */
4998 					regs[rd] = NULL;
4999 				} else {
5000 					regs[rd] = a + sizeof (uint64_t);
5001 				}
5002 
5003 				break;
5004 			}
5005 
5006 			regs[rd] = dtrace_dif_variable(mstate, state, id, 0);
5007 			break;
5008 
5009 		case DIF_OP_STGS:
5010 			id = DIF_INSTR_VAR(instr);
5011 
5012 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
5013 			id -= DIF_VAR_OTHER_UBASE;
5014 
5015 			svar = vstate->dtvs_globals[id];
5016 			ASSERT(svar != NULL);
5017 			v = &svar->dtsv_var;
5018 
5019 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5020 				uintptr_t a = (uintptr_t)svar->dtsv_data;
5021 
5022 				ASSERT(a != NULL);
5023 				ASSERT(svar->dtsv_size != 0);
5024 
5025 				if (regs[rd] == NULL) {
5026 					*(uint8_t *)a = UINT8_MAX;
5027 					break;
5028 				} else {
5029 					*(uint8_t *)a = 0;
5030 					a += sizeof (uint64_t);
5031 				}
5032 				if (!dtrace_vcanload(
5033 				    (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5034 				    mstate, vstate))
5035 					break;
5036 
5037 				dtrace_vcopy((void *)(uintptr_t)regs[rd],
5038 				    (void *)a, &v->dtdv_type);
5039 				break;
5040 			}
5041 
5042 			svar->dtsv_data = regs[rd];
5043 			break;
5044 
5045 		case DIF_OP_LDTA:
5046 			/*
5047 			 * There are no DTrace built-in thread-local arrays at
5048 			 * present.  This opcode is saved for future work.
5049 			 */
5050 			*flags |= CPU_DTRACE_ILLOP;
5051 			regs[rd] = 0;
5052 			break;
5053 
5054 		case DIF_OP_LDLS:
5055 			id = DIF_INSTR_VAR(instr);
5056 
5057 			if (id < DIF_VAR_OTHER_UBASE) {
5058 				/*
5059 				 * For now, this has no meaning.
5060 				 */
5061 				regs[rd] = 0;
5062 				break;
5063 			}
5064 
5065 			id -= DIF_VAR_OTHER_UBASE;
5066 
5067 			ASSERT(id < vstate->dtvs_nlocals);
5068 			ASSERT(vstate->dtvs_locals != NULL);
5069 
5070 			svar = vstate->dtvs_locals[id];
5071 			ASSERT(svar != NULL);
5072 			v = &svar->dtsv_var;
5073 
5074 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5075 				uintptr_t a = (uintptr_t)svar->dtsv_data;
5076 				size_t sz = v->dtdv_type.dtdt_size;
5077 
5078 				sz += sizeof (uint64_t);
5079 				ASSERT(svar->dtsv_size == NCPU * sz);
5080 				a += CPU->cpu_id * sz;
5081 
5082 				if (*(uint8_t *)a == UINT8_MAX) {
5083 					/*
5084 					 * If the 0th byte is set to UINT8_MAX
5085 					 * then this is to be treated as a
5086 					 * reference to a NULL variable.
5087 					 */
5088 					regs[rd] = NULL;
5089 				} else {
5090 					regs[rd] = a + sizeof (uint64_t);
5091 				}
5092 
5093 				break;
5094 			}
5095 
5096 			ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
5097 			tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
5098 			regs[rd] = tmp[CPU->cpu_id];
5099 			break;
5100 
5101 		case DIF_OP_STLS:
5102 			id = DIF_INSTR_VAR(instr);
5103 
5104 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
5105 			id -= DIF_VAR_OTHER_UBASE;
5106 			ASSERT(id < vstate->dtvs_nlocals);
5107 
5108 			ASSERT(vstate->dtvs_locals != NULL);
5109 			svar = vstate->dtvs_locals[id];
5110 			ASSERT(svar != NULL);
5111 			v = &svar->dtsv_var;
5112 
5113 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5114 				uintptr_t a = (uintptr_t)svar->dtsv_data;
5115 				size_t sz = v->dtdv_type.dtdt_size;
5116 
5117 				sz += sizeof (uint64_t);
5118 				ASSERT(svar->dtsv_size == NCPU * sz);
5119 				a += CPU->cpu_id * sz;
5120 
5121 				if (regs[rd] == NULL) {
5122 					*(uint8_t *)a = UINT8_MAX;
5123 					break;
5124 				} else {
5125 					*(uint8_t *)a = 0;
5126 					a += sizeof (uint64_t);
5127 				}
5128 
5129 				if (!dtrace_vcanload(
5130 				    (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5131 				    mstate, vstate))
5132 					break;
5133 
5134 				dtrace_vcopy((void *)(uintptr_t)regs[rd],
5135 				    (void *)a, &v->dtdv_type);
5136 				break;
5137 			}
5138 
5139 			ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
5140 			tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
5141 			tmp[CPU->cpu_id] = regs[rd];
5142 			break;
5143 
5144 		case DIF_OP_LDTS: {
5145 			dtrace_dynvar_t *dvar;
5146 			dtrace_key_t *key;
5147 
5148 			id = DIF_INSTR_VAR(instr);
5149 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
5150 			id -= DIF_VAR_OTHER_UBASE;
5151 			v = &vstate->dtvs_tlocals[id];
5152 
5153 			key = &tupregs[DIF_DTR_NREGS];
5154 			key[0].dttk_value = (uint64_t)id;
5155 			key[0].dttk_size = 0;
5156 			DTRACE_TLS_THRKEY(key[1].dttk_value);
5157 			key[1].dttk_size = 0;
5158 
5159 			dvar = dtrace_dynvar(dstate, 2, key,
5160 			    sizeof (uint64_t), DTRACE_DYNVAR_NOALLOC,
5161 			    mstate, vstate);
5162 
5163 			if (dvar == NULL) {
5164 				regs[rd] = 0;
5165 				break;
5166 			}
5167 
5168 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5169 				regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
5170 			} else {
5171 				regs[rd] = *((uint64_t *)dvar->dtdv_data);
5172 			}
5173 
5174 			break;
5175 		}
5176 
5177 		case DIF_OP_STTS: {
5178 			dtrace_dynvar_t *dvar;
5179 			dtrace_key_t *key;
5180 
5181 			id = DIF_INSTR_VAR(instr);
5182 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
5183 			id -= DIF_VAR_OTHER_UBASE;
5184 
5185 			key = &tupregs[DIF_DTR_NREGS];
5186 			key[0].dttk_value = (uint64_t)id;
5187 			key[0].dttk_size = 0;
5188 			DTRACE_TLS_THRKEY(key[1].dttk_value);
5189 			key[1].dttk_size = 0;
5190 			v = &vstate->dtvs_tlocals[id];
5191 
5192 			dvar = dtrace_dynvar(dstate, 2, key,
5193 			    v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
5194 			    v->dtdv_type.dtdt_size : sizeof (uint64_t),
5195 			    regs[rd] ? DTRACE_DYNVAR_ALLOC :
5196 			    DTRACE_DYNVAR_DEALLOC, mstate, vstate);
5197 
5198 			/*
5199 			 * Given that we're storing to thread-local data,
5200 			 * we need to flush our predicate cache.
5201 			 */
5202 			curthread->t_predcache = NULL;
5203 
5204 			if (dvar == NULL)
5205 				break;
5206 
5207 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5208 				if (!dtrace_vcanload(
5209 				    (void *)(uintptr_t)regs[rd],
5210 				    &v->dtdv_type, mstate, vstate))
5211 					break;
5212 
5213 				dtrace_vcopy((void *)(uintptr_t)regs[rd],
5214 				    dvar->dtdv_data, &v->dtdv_type);
5215 			} else {
5216 				*((uint64_t *)dvar->dtdv_data) = regs[rd];
5217 			}
5218 
5219 			break;
5220 		}
5221 
5222 		case DIF_OP_SRA:
5223 			regs[rd] = (int64_t)regs[r1] >> regs[r2];
5224 			break;
5225 
5226 		case DIF_OP_CALL:
5227 			dtrace_dif_subr(DIF_INSTR_SUBR(instr), rd,
5228 			    regs, tupregs, ttop, mstate, state);
5229 			break;
5230 
5231 		case DIF_OP_PUSHTR:
5232 			if (ttop == DIF_DTR_NREGS) {
5233 				*flags |= CPU_DTRACE_TUPOFLOW;
5234 				break;
5235 			}
5236 
5237 			if (r1 == DIF_TYPE_STRING) {
5238 				/*
5239 				 * If this is a string type and the size is 0,
5240 				 * we'll use the system-wide default string
5241 				 * size.  Note that we are _not_ looking at
5242 				 * the value of the DTRACEOPT_STRSIZE option;
5243 				 * had this been set, we would expect to have
5244 				 * a non-zero size value in the "pushtr".
5245 				 */
5246 				tupregs[ttop].dttk_size =
5247 				    dtrace_strlen((char *)(uintptr_t)regs[rd],
5248 				    regs[r2] ? regs[r2] :
5249 				    dtrace_strsize_default) + 1;
5250 			} else {
5251 				tupregs[ttop].dttk_size = regs[r2];
5252 			}
5253 
5254 			tupregs[ttop++].dttk_value = regs[rd];
5255 			break;
5256 
5257 		case DIF_OP_PUSHTV:
5258 			if (ttop == DIF_DTR_NREGS) {
5259 				*flags |= CPU_DTRACE_TUPOFLOW;
5260 				break;
5261 			}
5262 
5263 			tupregs[ttop].dttk_value = regs[rd];
5264 			tupregs[ttop++].dttk_size = 0;
5265 			break;
5266 
5267 		case DIF_OP_POPTS:
5268 			if (ttop != 0)
5269 				ttop--;
5270 			break;
5271 
5272 		case DIF_OP_FLUSHTS:
5273 			ttop = 0;
5274 			break;
5275 
5276 		case DIF_OP_LDGAA:
5277 		case DIF_OP_LDTAA: {
5278 			dtrace_dynvar_t *dvar;
5279 			dtrace_key_t *key = tupregs;
5280 			uint_t nkeys = ttop;
5281 
5282 			id = DIF_INSTR_VAR(instr);
5283 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
5284 			id -= DIF_VAR_OTHER_UBASE;
5285 
5286 			key[nkeys].dttk_value = (uint64_t)id;
5287 			key[nkeys++].dttk_size = 0;
5288 
5289 			if (DIF_INSTR_OP(instr) == DIF_OP_LDTAA) {
5290 				DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
5291 				key[nkeys++].dttk_size = 0;
5292 				v = &vstate->dtvs_tlocals[id];
5293 			} else {
5294 				v = &vstate->dtvs_globals[id]->dtsv_var;
5295 			}
5296 
5297 			dvar = dtrace_dynvar(dstate, nkeys, key,
5298 			    v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
5299 			    v->dtdv_type.dtdt_size : sizeof (uint64_t),
5300 			    DTRACE_DYNVAR_NOALLOC, mstate, vstate);
5301 
5302 			if (dvar == NULL) {
5303 				regs[rd] = 0;
5304 				break;
5305 			}
5306 
5307 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5308 				regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
5309 			} else {
5310 				regs[rd] = *((uint64_t *)dvar->dtdv_data);
5311 			}
5312 
5313 			break;
5314 		}
5315 
5316 		case DIF_OP_STGAA:
5317 		case DIF_OP_STTAA: {
5318 			dtrace_dynvar_t *dvar;
5319 			dtrace_key_t *key = tupregs;
5320 			uint_t nkeys = ttop;
5321 
5322 			id = DIF_INSTR_VAR(instr);
5323 			ASSERT(id >= DIF_VAR_OTHER_UBASE);
5324 			id -= DIF_VAR_OTHER_UBASE;
5325 
5326 			key[nkeys].dttk_value = (uint64_t)id;
5327 			key[nkeys++].dttk_size = 0;
5328 
5329 			if (DIF_INSTR_OP(instr) == DIF_OP_STTAA) {
5330 				DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
5331 				key[nkeys++].dttk_size = 0;
5332 				v = &vstate->dtvs_tlocals[id];
5333 			} else {
5334 				v = &vstate->dtvs_globals[id]->dtsv_var;
5335 			}
5336 
5337 			dvar = dtrace_dynvar(dstate, nkeys, key,
5338 			    v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
5339 			    v->dtdv_type.dtdt_size : sizeof (uint64_t),
5340 			    regs[rd] ? DTRACE_DYNVAR_ALLOC :
5341 			    DTRACE_DYNVAR_DEALLOC, mstate, vstate);
5342 
5343 			if (dvar == NULL)
5344 				break;
5345 
5346 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5347 				if (!dtrace_vcanload(
5348 				    (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5349 				    mstate, vstate))
5350 					break;
5351 
5352 				dtrace_vcopy((void *)(uintptr_t)regs[rd],
5353 				    dvar->dtdv_data, &v->dtdv_type);
5354 			} else {
5355 				*((uint64_t *)dvar->dtdv_data) = regs[rd];
5356 			}
5357 
5358 			break;
5359 		}
5360 
5361 		case DIF_OP_ALLOCS: {
5362 			uintptr_t ptr = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
5363 			size_t size = ptr - mstate->dtms_scratch_ptr + regs[r1];
5364 
5365 			/*
5366 			 * Rounding up the user allocation size could have
5367 			 * overflowed large, bogus allocations (like -1ULL) to
5368 			 * 0.
5369 			 */
5370 			if (size < regs[r1] ||
5371 			    !DTRACE_INSCRATCH(mstate, size)) {
5372 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5373 				regs[rd] = NULL;
5374 				break;
5375 			}
5376 
5377 			dtrace_bzero((void *) mstate->dtms_scratch_ptr, size);
5378 			mstate->dtms_scratch_ptr += size;
5379 			regs[rd] = ptr;
5380 			break;
5381 		}
5382 
5383 		case DIF_OP_COPYS:
5384 			if (!dtrace_canstore(regs[rd], regs[r2],
5385 			    mstate, vstate)) {
5386 				*flags |= CPU_DTRACE_BADADDR;
5387 				*illval = regs[rd];
5388 				break;
5389 			}
5390 
5391 			if (!dtrace_canload(regs[r1], regs[r2], mstate, vstate))
5392 				break;
5393 
5394 			dtrace_bcopy((void *)(uintptr_t)regs[r1],
5395 			    (void *)(uintptr_t)regs[rd], (size_t)regs[r2]);
5396 			break;
5397 
5398 		case DIF_OP_STB:
5399 			if (!dtrace_canstore(regs[rd], 1, mstate, vstate)) {
5400 				*flags |= CPU_DTRACE_BADADDR;
5401 				*illval = regs[rd];
5402 				break;
5403 			}
5404 			*((uint8_t *)(uintptr_t)regs[rd]) = (uint8_t)regs[r1];
5405 			break;
5406 
5407 		case DIF_OP_STH:
5408 			if (!dtrace_canstore(regs[rd], 2, mstate, vstate)) {
5409 				*flags |= CPU_DTRACE_BADADDR;
5410 				*illval = regs[rd];
5411 				break;
5412 			}
5413 			if (regs[rd] & 1) {
5414 				*flags |= CPU_DTRACE_BADALIGN;
5415 				*illval = regs[rd];
5416 				break;
5417 			}
5418 			*((uint16_t *)(uintptr_t)regs[rd]) = (uint16_t)regs[r1];
5419 			break;
5420 
5421 		case DIF_OP_STW:
5422 			if (!dtrace_canstore(regs[rd], 4, mstate, vstate)) {
5423 				*flags |= CPU_DTRACE_BADADDR;
5424 				*illval = regs[rd];
5425 				break;
5426 			}
5427 			if (regs[rd] & 3) {
5428 				*flags |= CPU_DTRACE_BADALIGN;
5429 				*illval = regs[rd];
5430 				break;
5431 			}
5432 			*((uint32_t *)(uintptr_t)regs[rd]) = (uint32_t)regs[r1];
5433 			break;
5434 
5435 		case DIF_OP_STX:
5436 			if (!dtrace_canstore(regs[rd], 8, mstate, vstate)) {
5437 				*flags |= CPU_DTRACE_BADADDR;
5438 				*illval = regs[rd];
5439 				break;
5440 			}
5441 			if (regs[rd] & 7) {
5442 				*flags |= CPU_DTRACE_BADALIGN;
5443 				*illval = regs[rd];
5444 				break;
5445 			}
5446 			*((uint64_t *)(uintptr_t)regs[rd]) = regs[r1];
5447 			break;
5448 		}
5449 	}
5450 
5451 	if (!(*flags & CPU_DTRACE_FAULT))
5452 		return (rval);
5453 
5454 	mstate->dtms_fltoffs = opc * sizeof (dif_instr_t);
5455 	mstate->dtms_present |= DTRACE_MSTATE_FLTOFFS;
5456 
5457 	return (0);
5458 }
5459 
5460 static void
5461 dtrace_action_breakpoint(dtrace_ecb_t *ecb)
5462 {
5463 	dtrace_probe_t *probe = ecb->dte_probe;
5464 	dtrace_provider_t *prov = probe->dtpr_provider;
5465 	char c[DTRACE_FULLNAMELEN + 80], *str;
5466 	char *msg = "dtrace: breakpoint action at probe ";
5467 	char *ecbmsg = " (ecb ";
5468 	uintptr_t mask = (0xf << (sizeof (uintptr_t) * NBBY / 4));
5469 	uintptr_t val = (uintptr_t)ecb;
5470 	int shift = (sizeof (uintptr_t) * NBBY) - 4, i = 0;
5471 
5472 	if (dtrace_destructive_disallow)
5473 		return;
5474 
5475 	/*
5476 	 * It's impossible to be taking action on the NULL probe.
5477 	 */
5478 	ASSERT(probe != NULL);
5479 
5480 	/*
5481 	 * This is a poor man's (destitute man's?) sprintf():  we want to
5482 	 * print the provider name, module name, function name and name of
5483 	 * the probe, along with the hex address of the ECB with the breakpoint
5484 	 * action -- all of which we must place in the character buffer by
5485 	 * hand.
5486 	 */
5487 	while (*msg != '\0')
5488 		c[i++] = *msg++;
5489 
5490 	for (str = prov->dtpv_name; *str != '\0'; str++)
5491 		c[i++] = *str;
5492 	c[i++] = ':';
5493 
5494 	for (str = probe->dtpr_mod; *str != '\0'; str++)
5495 		c[i++] = *str;
5496 	c[i++] = ':';
5497 
5498 	for (str = probe->dtpr_func; *str != '\0'; str++)
5499 		c[i++] = *str;
5500 	c[i++] = ':';
5501 
5502 	for (str = probe->dtpr_name; *str != '\0'; str++)
5503 		c[i++] = *str;
5504 
5505 	while (*ecbmsg != '\0')
5506 		c[i++] = *ecbmsg++;
5507 
5508 	while (shift >= 0) {
5509 		mask = (uintptr_t)0xf << shift;
5510 
5511 		if (val >= ((uintptr_t)1 << shift))
5512 			c[i++] = "0123456789abcdef"[(val & mask) >> shift];
5513 		shift -= 4;
5514 	}
5515 
5516 	c[i++] = ')';
5517 	c[i] = '\0';
5518 
5519 	debug_enter(c);
5520 }
5521 
5522 static void
5523 dtrace_action_panic(dtrace_ecb_t *ecb)
5524 {
5525 	dtrace_probe_t *probe = ecb->dte_probe;
5526 
5527 	/*
5528 	 * It's impossible to be taking action on the NULL probe.
5529 	 */
5530 	ASSERT(probe != NULL);
5531 
5532 	if (dtrace_destructive_disallow)
5533 		return;
5534 
5535 	if (dtrace_panicked != NULL)
5536 		return;
5537 
5538 	if (dtrace_casptr(&dtrace_panicked, NULL, curthread) != NULL)
5539 		return;
5540 
5541 	/*
5542 	 * We won the right to panic.  (We want to be sure that only one
5543 	 * thread calls panic() from dtrace_probe(), and that panic() is
5544 	 * called exactly once.)
5545 	 */
5546 	dtrace_panic("dtrace: panic action at probe %s:%s:%s:%s (ecb %p)",
5547 	    probe->dtpr_provider->dtpv_name, probe->dtpr_mod,
5548 	    probe->dtpr_func, probe->dtpr_name, (void *)ecb);
5549 }
5550 
5551 static void
5552 dtrace_action_raise(uint64_t sig)
5553 {
5554 	if (dtrace_destructive_disallow)
5555 		return;
5556 
5557 	if (sig >= NSIG) {
5558 		DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
5559 		return;
5560 	}
5561 
5562 	/*
5563 	 * raise() has a queue depth of 1 -- we ignore all subsequent
5564 	 * invocations of the raise() action.
5565 	 */
5566 	if (curthread->t_dtrace_sig == 0)
5567 		curthread->t_dtrace_sig = (uint8_t)sig;
5568 
5569 	curthread->t_sig_check = 1;
5570 	aston(curthread);
5571 }
5572 
5573 static void
5574 dtrace_action_stop(void)
5575 {
5576 	if (dtrace_destructive_disallow)
5577 		return;
5578 
5579 	if (!curthread->t_dtrace_stop) {
5580 		curthread->t_dtrace_stop = 1;
5581 		curthread->t_sig_check = 1;
5582 		aston(curthread);
5583 	}
5584 }
5585 
5586 static void
5587 dtrace_action_chill(dtrace_mstate_t *mstate, hrtime_t val)
5588 {
5589 	hrtime_t now;
5590 	volatile uint16_t *flags;
5591 	cpu_t *cpu = CPU;
5592 
5593 	if (dtrace_destructive_disallow)
5594 		return;
5595 
5596 	flags = (volatile uint16_t *)&cpu_core[cpu->cpu_id].cpuc_dtrace_flags;
5597 
5598 	now = dtrace_gethrtime();
5599 
5600 	if (now - cpu->cpu_dtrace_chillmark > dtrace_chill_interval) {
5601 		/*
5602 		 * We need to advance the mark to the current time.
5603 		 */
5604 		cpu->cpu_dtrace_chillmark = now;
5605 		cpu->cpu_dtrace_chilled = 0;
5606 	}
5607 
5608 	/*
5609 	 * Now check to see if the requested chill time would take us over
5610 	 * the maximum amount of time allowed in the chill interval.  (Or
5611 	 * worse, if the calculation itself induces overflow.)
5612 	 */
5613 	if (cpu->cpu_dtrace_chilled + val > dtrace_chill_max ||
5614 	    cpu->cpu_dtrace_chilled + val < cpu->cpu_dtrace_chilled) {
5615 		*flags |= CPU_DTRACE_ILLOP;
5616 		return;
5617 	}
5618 
5619 	while (dtrace_gethrtime() - now < val)
5620 		continue;
5621 
5622 	/*
5623 	 * Normally, we assure that the value of the variable "timestamp" does
5624 	 * not change within an ECB.  The presence of chill() represents an
5625 	 * exception to this rule, however.
5626 	 */
5627 	mstate->dtms_present &= ~DTRACE_MSTATE_TIMESTAMP;
5628 	cpu->cpu_dtrace_chilled += val;
5629 }
5630 
5631 static void
5632 dtrace_action_ustack(dtrace_mstate_t *mstate, dtrace_state_t *state,
5633     uint64_t *buf, uint64_t arg)
5634 {
5635 	int nframes = DTRACE_USTACK_NFRAMES(arg);
5636 	int strsize = DTRACE_USTACK_STRSIZE(arg);
5637 	uint64_t *pcs = &buf[1], *fps;
5638 	char *str = (char *)&pcs[nframes];
5639 	int size, offs = 0, i, j;
5640 	uintptr_t old = mstate->dtms_scratch_ptr, saved;
5641 	uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
5642 	char *sym;
5643 
5644 	/*
5645 	 * Should be taking a faster path if string space has not been
5646 	 * allocated.
5647 	 */
5648 	ASSERT(strsize != 0);
5649 
5650 	/*
5651 	 * We will first allocate some temporary space for the frame pointers.
5652 	 */
5653 	fps = (uint64_t *)P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
5654 	size = (uintptr_t)fps - mstate->dtms_scratch_ptr +
5655 	    (nframes * sizeof (uint64_t));
5656 
5657 	if (!DTRACE_INSCRATCH(mstate, size)) {
5658 		/*
5659 		 * Not enough room for our frame pointers -- need to indicate
5660 		 * that we ran out of scratch space.
5661 		 */
5662 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5663 		return;
5664 	}
5665 
5666 	mstate->dtms_scratch_ptr += size;
5667 	saved = mstate->dtms_scratch_ptr;
5668 
5669 	/*
5670 	 * Now get a stack with both program counters and frame pointers.
5671 	 */
5672 	DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5673 	dtrace_getufpstack(buf, fps, nframes + 1);
5674 	DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5675 
5676 	/*
5677 	 * If that faulted, we're cooked.
5678 	 */
5679 	if (*flags & CPU_DTRACE_FAULT)
5680 		goto out;
5681 
5682 	/*
5683 	 * Now we want to walk up the stack, calling the USTACK helper.  For
5684 	 * each iteration, we restore the scratch pointer.
5685 	 */
5686 	for (i = 0; i < nframes; i++) {
5687 		mstate->dtms_scratch_ptr = saved;
5688 
5689 		if (offs >= strsize)
5690 			break;
5691 
5692 		sym = (char *)(uintptr_t)dtrace_helper(
5693 		    DTRACE_HELPER_ACTION_USTACK,
5694 		    mstate, state, pcs[i], fps[i]);
5695 
5696 		/*
5697 		 * If we faulted while running the helper, we're going to
5698 		 * clear the fault and null out the corresponding string.
5699 		 */
5700 		if (*flags & CPU_DTRACE_FAULT) {
5701 			*flags &= ~CPU_DTRACE_FAULT;
5702 			str[offs++] = '\0';
5703 			continue;
5704 		}
5705 
5706 		if (sym == NULL) {
5707 			str[offs++] = '\0';
5708 			continue;
5709 		}
5710 
5711 		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5712 
5713 		/*
5714 		 * Now copy in the string that the helper returned to us.
5715 		 */
5716 		for (j = 0; offs + j < strsize; j++) {
5717 			if ((str[offs + j] = sym[j]) == '\0')
5718 				break;
5719 		}
5720 
5721 		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5722 
5723 		offs += j + 1;
5724 	}
5725 
5726 	if (offs >= strsize) {
5727 		/*
5728 		 * If we didn't have room for all of the strings, we don't
5729 		 * abort processing -- this needn't be a fatal error -- but we
5730 		 * still want to increment a counter (dts_stkstroverflows) to
5731 		 * allow this condition to be warned about.  (If this is from
5732 		 * a jstack() action, it is easily tuned via jstackstrsize.)
5733 		 */
5734 		dtrace_error(&state->dts_stkstroverflows);
5735 	}
5736 
5737 	while (offs < strsize)
5738 		str[offs++] = '\0';
5739 
5740 out:
5741 	mstate->dtms_scratch_ptr = old;
5742 }
5743 
5744 /*
5745  * If you're looking for the epicenter of DTrace, you just found it.  This
5746  * is the function called by the provider to fire a probe -- from which all
5747  * subsequent probe-context DTrace activity emanates.
5748  */
5749 void
5750 dtrace_probe(dtrace_id_t id, uintptr_t arg0, uintptr_t arg1,
5751     uintptr_t arg2, uintptr_t arg3, uintptr_t arg4)
5752 {
5753 	processorid_t cpuid;
5754 	dtrace_icookie_t cookie;
5755 	dtrace_probe_t *probe;
5756 	dtrace_mstate_t mstate;
5757 	dtrace_ecb_t *ecb;
5758 	dtrace_action_t *act;
5759 	intptr_t offs;
5760 	size_t size;
5761 	int vtime, onintr;
5762 	volatile uint16_t *flags;
5763 	hrtime_t now;
5764 
5765 	/*
5766 	 * Kick out immediately if this CPU is still being born (in which case
5767 	 * curthread will be set to -1) or the current thread can't allow
5768 	 * probes in its current context.
5769 	 */
5770 	if (((uintptr_t)curthread & 1) || (curthread->t_flag & T_DONTDTRACE))
5771 		return;
5772 
5773 	cookie = dtrace_interrupt_disable();
5774 	probe = dtrace_probes[id - 1];
5775 	cpuid = CPU->cpu_id;
5776 	onintr = CPU_ON_INTR(CPU);
5777 
5778 	if (!onintr && probe->dtpr_predcache != DTRACE_CACHEIDNONE &&
5779 	    probe->dtpr_predcache == curthread->t_predcache) {
5780 		/*
5781 		 * We have hit in the predicate cache; we know that
5782 		 * this predicate would evaluate to be false.
5783 		 */
5784 		dtrace_interrupt_enable(cookie);
5785 		return;
5786 	}
5787 
5788 	if (panic_quiesce) {
5789 		/*
5790 		 * We don't trace anything if we're panicking.
5791 		 */
5792 		dtrace_interrupt_enable(cookie);
5793 		return;
5794 	}
5795 
5796 	now = dtrace_gethrtime();
5797 	vtime = dtrace_vtime_references != 0;
5798 
5799 	if (vtime && curthread->t_dtrace_start)
5800 		curthread->t_dtrace_vtime += now - curthread->t_dtrace_start;
5801 
5802 	mstate.dtms_difo = NULL;
5803 	mstate.dtms_probe = probe;
5804 	mstate.dtms_strtok = NULL;
5805 	mstate.dtms_arg[0] = arg0;
5806 	mstate.dtms_arg[1] = arg1;
5807 	mstate.dtms_arg[2] = arg2;
5808 	mstate.dtms_arg[3] = arg3;
5809 	mstate.dtms_arg[4] = arg4;
5810 
5811 	flags = (volatile uint16_t *)&cpu_core[cpuid].cpuc_dtrace_flags;
5812 
5813 	for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
5814 		dtrace_predicate_t *pred = ecb->dte_predicate;
5815 		dtrace_state_t *state = ecb->dte_state;
5816 		dtrace_buffer_t *buf = &state->dts_buffer[cpuid];
5817 		dtrace_buffer_t *aggbuf = &state->dts_aggbuffer[cpuid];
5818 		dtrace_vstate_t *vstate = &state->dts_vstate;
5819 		dtrace_provider_t *prov = probe->dtpr_provider;
5820 		int committed = 0;
5821 		caddr_t tomax;
5822 
5823 		/*
5824 		 * A little subtlety with the following (seemingly innocuous)
5825 		 * declaration of the automatic 'val':  by looking at the
5826 		 * code, you might think that it could be declared in the
5827 		 * action processing loop, below.  (That is, it's only used in
5828 		 * the action processing loop.)  However, it must be declared
5829 		 * out of that scope because in the case of DIF expression
5830 		 * arguments to aggregating actions, one iteration of the
5831 		 * action loop will use the last iteration's value.
5832 		 */
5833 #ifdef lint
5834 		uint64_t val = 0;
5835 #else
5836 		uint64_t val;
5837 #endif
5838 
5839 		mstate.dtms_present = DTRACE_MSTATE_ARGS | DTRACE_MSTATE_PROBE;
5840 		mstate.dtms_access = DTRACE_ACCESS_ARGS | DTRACE_ACCESS_PROC;
5841 		*flags &= ~CPU_DTRACE_ERROR;
5842 
5843 		if (prov == dtrace_provider) {
5844 			/*
5845 			 * If dtrace itself is the provider of this probe,
5846 			 * we're only going to continue processing the ECB if
5847 			 * arg0 (the dtrace_state_t) is equal to the ECB's
5848 			 * creating state.  (This prevents disjoint consumers
5849 			 * from seeing one another's metaprobes.)
5850 			 */
5851 			if (arg0 != (uint64_t)(uintptr_t)state)
5852 				continue;
5853 		}
5854 
5855 		if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE) {
5856 			/*
5857 			 * We're not currently active.  If our provider isn't
5858 			 * the dtrace pseudo provider, we're not interested.
5859 			 */
5860 			if (prov != dtrace_provider)
5861 				continue;
5862 
5863 			/*
5864 			 * Now we must further check if we are in the BEGIN
5865 			 * probe.  If we are, we will only continue processing
5866 			 * if we're still in WARMUP -- if one BEGIN enabling
5867 			 * has invoked the exit() action, we don't want to
5868 			 * evaluate subsequent BEGIN enablings.
5869 			 */
5870 			if (probe->dtpr_id == dtrace_probeid_begin &&
5871 			    state->dts_activity != DTRACE_ACTIVITY_WARMUP) {
5872 				ASSERT(state->dts_activity ==
5873 				    DTRACE_ACTIVITY_DRAINING);
5874 				continue;
5875 			}
5876 		}
5877 
5878 		if (ecb->dte_cond && !dtrace_priv_probe(state, &mstate, ecb))
5879 			continue;
5880 
5881 		if (now - state->dts_alive > dtrace_deadman_timeout) {
5882 			/*
5883 			 * We seem to be dead.  Unless we (a) have kernel
5884 			 * destructive permissions (b) have expicitly enabled
5885 			 * destructive actions and (c) destructive actions have
5886 			 * not been disabled, we're going to transition into
5887 			 * the KILLED state, from which no further processing
5888 			 * on this state will be performed.
5889 			 */
5890 			if (!dtrace_priv_kernel_destructive(state) ||
5891 			    !state->dts_cred.dcr_destructive ||
5892 			    dtrace_destructive_disallow) {
5893 				void *activity = &state->dts_activity;
5894 				dtrace_activity_t current;
5895 
5896 				do {
5897 					current = state->dts_activity;
5898 				} while (dtrace_cas32(activity, current,
5899 				    DTRACE_ACTIVITY_KILLED) != current);
5900 
5901 				continue;
5902 			}
5903 		}
5904 
5905 		if ((offs = dtrace_buffer_reserve(buf, ecb->dte_needed,
5906 		    ecb->dte_alignment, state, &mstate)) < 0)
5907 			continue;
5908 
5909 		tomax = buf->dtb_tomax;
5910 		ASSERT(tomax != NULL);
5911 
5912 		if (ecb->dte_size != 0)
5913 			DTRACE_STORE(uint32_t, tomax, offs, ecb->dte_epid);
5914 
5915 		mstate.dtms_epid = ecb->dte_epid;
5916 		mstate.dtms_present |= DTRACE_MSTATE_EPID;
5917 
5918 		if (state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)
5919 			mstate.dtms_access |= DTRACE_ACCESS_KERNEL;
5920 
5921 		if (pred != NULL) {
5922 			dtrace_difo_t *dp = pred->dtp_difo;
5923 			int rval;
5924 
5925 			rval = dtrace_dif_emulate(dp, &mstate, vstate, state);
5926 
5927 			if (!(*flags & CPU_DTRACE_ERROR) && !rval) {
5928 				dtrace_cacheid_t cid = probe->dtpr_predcache;
5929 
5930 				if (cid != DTRACE_CACHEIDNONE && !onintr) {
5931 					/*
5932 					 * Update the predicate cache...
5933 					 */
5934 					ASSERT(cid == pred->dtp_cacheid);
5935 					curthread->t_predcache = cid;
5936 				}
5937 
5938 				continue;
5939 			}
5940 		}
5941 
5942 		for (act = ecb->dte_action; !(*flags & CPU_DTRACE_ERROR) &&
5943 		    act != NULL; act = act->dta_next) {
5944 			size_t valoffs;
5945 			dtrace_difo_t *dp;
5946 			dtrace_recdesc_t *rec = &act->dta_rec;
5947 
5948 			size = rec->dtrd_size;
5949 			valoffs = offs + rec->dtrd_offset;
5950 
5951 			if (DTRACEACT_ISAGG(act->dta_kind)) {
5952 				uint64_t v = 0xbad;
5953 				dtrace_aggregation_t *agg;
5954 
5955 				agg = (dtrace_aggregation_t *)act;
5956 
5957 				if ((dp = act->dta_difo) != NULL)
5958 					v = dtrace_dif_emulate(dp,
5959 					    &mstate, vstate, state);
5960 
5961 				if (*flags & CPU_DTRACE_ERROR)
5962 					continue;
5963 
5964 				/*
5965 				 * Note that we always pass the expression
5966 				 * value from the previous iteration of the
5967 				 * action loop.  This value will only be used
5968 				 * if there is an expression argument to the
5969 				 * aggregating action, denoted by the
5970 				 * dtag_hasarg field.
5971 				 */
5972 				dtrace_aggregate(agg, buf,
5973 				    offs, aggbuf, v, val);
5974 				continue;
5975 			}
5976 
5977 			switch (act->dta_kind) {
5978 			case DTRACEACT_STOP:
5979 				if (dtrace_priv_proc_destructive(state,
5980 				    &mstate))
5981 					dtrace_action_stop();
5982 				continue;
5983 
5984 			case DTRACEACT_BREAKPOINT:
5985 				if (dtrace_priv_kernel_destructive(state))
5986 					dtrace_action_breakpoint(ecb);
5987 				continue;
5988 
5989 			case DTRACEACT_PANIC:
5990 				if (dtrace_priv_kernel_destructive(state))
5991 					dtrace_action_panic(ecb);
5992 				continue;
5993 
5994 			case DTRACEACT_STACK:
5995 				if (!dtrace_priv_kernel(state))
5996 					continue;
5997 
5998 				dtrace_getpcstack((pc_t *)(tomax + valoffs),
5999 				    size / sizeof (pc_t), probe->dtpr_aframes,
6000 				    DTRACE_ANCHORED(probe) ? NULL :
6001 				    (uint32_t *)arg0);
6002 
6003 				continue;
6004 
6005 			case DTRACEACT_JSTACK:
6006 			case DTRACEACT_USTACK:
6007 				if (!dtrace_priv_proc(state, &mstate))
6008 					continue;
6009 
6010 				/*
6011 				 * See comment in DIF_VAR_PID.
6012 				 */
6013 				if (DTRACE_ANCHORED(mstate.dtms_probe) &&
6014 				    CPU_ON_INTR(CPU)) {
6015 					int depth = DTRACE_USTACK_NFRAMES(
6016 					    rec->dtrd_arg) + 1;
6017 
6018 					dtrace_bzero((void *)(tomax + valoffs),
6019 					    DTRACE_USTACK_STRSIZE(rec->dtrd_arg)
6020 					    + depth * sizeof (uint64_t));
6021 
6022 					continue;
6023 				}
6024 
6025 				if (DTRACE_USTACK_STRSIZE(rec->dtrd_arg) != 0 &&
6026 				    curproc->p_dtrace_helpers != NULL) {
6027 					/*
6028 					 * This is the slow path -- we have
6029 					 * allocated string space, and we're
6030 					 * getting the stack of a process that
6031 					 * has helpers.  Call into a separate
6032 					 * routine to perform this processing.
6033 					 */
6034 					dtrace_action_ustack(&mstate, state,
6035 					    (uint64_t *)(tomax + valoffs),
6036 					    rec->dtrd_arg);
6037 					continue;
6038 				}
6039 
6040 				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6041 				dtrace_getupcstack((uint64_t *)
6042 				    (tomax + valoffs),
6043 				    DTRACE_USTACK_NFRAMES(rec->dtrd_arg) + 1);
6044 				DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6045 				continue;
6046 
6047 			default:
6048 				break;
6049 			}
6050 
6051 			dp = act->dta_difo;
6052 			ASSERT(dp != NULL);
6053 
6054 			val = dtrace_dif_emulate(dp, &mstate, vstate, state);
6055 
6056 			if (*flags & CPU_DTRACE_ERROR)
6057 				continue;
6058 
6059 			switch (act->dta_kind) {
6060 			case DTRACEACT_SPECULATE:
6061 				ASSERT(buf == &state->dts_buffer[cpuid]);
6062 				buf = dtrace_speculation_buffer(state,
6063 				    cpuid, val);
6064 
6065 				if (buf == NULL) {
6066 					*flags |= CPU_DTRACE_DROP;
6067 					continue;
6068 				}
6069 
6070 				offs = dtrace_buffer_reserve(buf,
6071 				    ecb->dte_needed, ecb->dte_alignment,
6072 				    state, NULL);
6073 
6074 				if (offs < 0) {
6075 					*flags |= CPU_DTRACE_DROP;
6076 					continue;
6077 				}
6078 
6079 				tomax = buf->dtb_tomax;
6080 				ASSERT(tomax != NULL);
6081 
6082 				if (ecb->dte_size != 0)
6083 					DTRACE_STORE(uint32_t, tomax, offs,
6084 					    ecb->dte_epid);
6085 				continue;
6086 
6087 			case DTRACEACT_CHILL:
6088 				if (dtrace_priv_kernel_destructive(state))
6089 					dtrace_action_chill(&mstate, val);
6090 				continue;
6091 
6092 			case DTRACEACT_RAISE:
6093 				if (dtrace_priv_proc_destructive(state,
6094 				    &mstate))
6095 					dtrace_action_raise(val);
6096 				continue;
6097 
6098 			case DTRACEACT_COMMIT:
6099 				ASSERT(!committed);
6100 
6101 				/*
6102 				 * We need to commit our buffer state.
6103 				 */
6104 				if (ecb->dte_size)
6105 					buf->dtb_offset = offs + ecb->dte_size;
6106 				buf = &state->dts_buffer[cpuid];
6107 				dtrace_speculation_commit(state, cpuid, val);
6108 				committed = 1;
6109 				continue;
6110 
6111 			case DTRACEACT_DISCARD:
6112 				dtrace_speculation_discard(state, cpuid, val);
6113 				continue;
6114 
6115 			case DTRACEACT_DIFEXPR:
6116 			case DTRACEACT_LIBACT:
6117 			case DTRACEACT_PRINTF:
6118 			case DTRACEACT_PRINTA:
6119 			case DTRACEACT_SYSTEM:
6120 			case DTRACEACT_FREOPEN:
6121 				break;
6122 
6123 			case DTRACEACT_SYM:
6124 			case DTRACEACT_MOD:
6125 				if (!dtrace_priv_kernel(state))
6126 					continue;
6127 				break;
6128 
6129 			case DTRACEACT_USYM:
6130 			case DTRACEACT_UMOD:
6131 			case DTRACEACT_UADDR: {
6132 				struct pid *pid = curthread->t_procp->p_pidp;
6133 
6134 				if (!dtrace_priv_proc(state, &mstate))
6135 					continue;
6136 
6137 				DTRACE_STORE(uint64_t, tomax,
6138 				    valoffs, (uint64_t)pid->pid_id);
6139 				DTRACE_STORE(uint64_t, tomax,
6140 				    valoffs + sizeof (uint64_t), val);
6141 
6142 				continue;
6143 			}
6144 
6145 			case DTRACEACT_EXIT: {
6146 				/*
6147 				 * For the exit action, we are going to attempt
6148 				 * to atomically set our activity to be
6149 				 * draining.  If this fails (either because
6150 				 * another CPU has beat us to the exit action,
6151 				 * or because our current activity is something
6152 				 * other than ACTIVE or WARMUP), we will
6153 				 * continue.  This assures that the exit action
6154 				 * can be successfully recorded at most once
6155 				 * when we're in the ACTIVE state.  If we're
6156 				 * encountering the exit() action while in
6157 				 * COOLDOWN, however, we want to honor the new
6158 				 * status code.  (We know that we're the only
6159 				 * thread in COOLDOWN, so there is no race.)
6160 				 */
6161 				void *activity = &state->dts_activity;
6162 				dtrace_activity_t current = state->dts_activity;
6163 
6164 				if (current == DTRACE_ACTIVITY_COOLDOWN)
6165 					break;
6166 
6167 				if (current != DTRACE_ACTIVITY_WARMUP)
6168 					current = DTRACE_ACTIVITY_ACTIVE;
6169 
6170 				if (dtrace_cas32(activity, current,
6171 				    DTRACE_ACTIVITY_DRAINING) != current) {
6172 					*flags |= CPU_DTRACE_DROP;
6173 					continue;
6174 				}
6175 
6176 				break;
6177 			}
6178 
6179 			default:
6180 				ASSERT(0);
6181 			}
6182 
6183 			if (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF) {
6184 				uintptr_t end = valoffs + size;
6185 
6186 				if (!dtrace_vcanload((void *)(uintptr_t)val,
6187 				    &dp->dtdo_rtype, &mstate, vstate))
6188 					continue;
6189 
6190 				/*
6191 				 * If this is a string, we're going to only
6192 				 * load until we find the zero byte -- after
6193 				 * which we'll store zero bytes.
6194 				 */
6195 				if (dp->dtdo_rtype.dtdt_kind ==
6196 				    DIF_TYPE_STRING) {
6197 					char c = '\0' + 1;
6198 					int intuple = act->dta_intuple;
6199 					size_t s;
6200 
6201 					for (s = 0; s < size; s++) {
6202 						if (c != '\0')
6203 							c = dtrace_load8(val++);
6204 
6205 						DTRACE_STORE(uint8_t, tomax,
6206 						    valoffs++, c);
6207 
6208 						if (c == '\0' && intuple)
6209 							break;
6210 					}
6211 
6212 					continue;
6213 				}
6214 
6215 				while (valoffs < end) {
6216 					DTRACE_STORE(uint8_t, tomax, valoffs++,
6217 					    dtrace_load8(val++));
6218 				}
6219 
6220 				continue;
6221 			}
6222 
6223 			switch (size) {
6224 			case 0:
6225 				break;
6226 
6227 			case sizeof (uint8_t):
6228 				DTRACE_STORE(uint8_t, tomax, valoffs, val);
6229 				break;
6230 			case sizeof (uint16_t):
6231 				DTRACE_STORE(uint16_t, tomax, valoffs, val);
6232 				break;
6233 			case sizeof (uint32_t):
6234 				DTRACE_STORE(uint32_t, tomax, valoffs, val);
6235 				break;
6236 			case sizeof (uint64_t):
6237 				DTRACE_STORE(uint64_t, tomax, valoffs, val);
6238 				break;
6239 			default:
6240 				/*
6241 				 * Any other size should have been returned by
6242 				 * reference, not by value.
6243 				 */
6244 				ASSERT(0);
6245 				break;
6246 			}
6247 		}
6248 
6249 		if (*flags & CPU_DTRACE_DROP)
6250 			continue;
6251 
6252 		if (*flags & CPU_DTRACE_FAULT) {
6253 			int ndx;
6254 			dtrace_action_t *err;
6255 
6256 			buf->dtb_errors++;
6257 
6258 			if (probe->dtpr_id == dtrace_probeid_error) {
6259 				/*
6260 				 * There's nothing we can do -- we had an
6261 				 * error on the error probe.  We bump an
6262 				 * error counter to at least indicate that
6263 				 * this condition happened.
6264 				 */
6265 				dtrace_error(&state->dts_dblerrors);
6266 				continue;
6267 			}
6268 
6269 			if (vtime) {
6270 				/*
6271 				 * Before recursing on dtrace_probe(), we
6272 				 * need to explicitly clear out our start
6273 				 * time to prevent it from being accumulated
6274 				 * into t_dtrace_vtime.
6275 				 */
6276 				curthread->t_dtrace_start = 0;
6277 			}
6278 
6279 			/*
6280 			 * Iterate over the actions to figure out which action
6281 			 * we were processing when we experienced the error.
6282 			 * Note that act points _past_ the faulting action; if
6283 			 * act is ecb->dte_action, the fault was in the
6284 			 * predicate, if it's ecb->dte_action->dta_next it's
6285 			 * in action #1, and so on.
6286 			 */
6287 			for (err = ecb->dte_action, ndx = 0;
6288 			    err != act; err = err->dta_next, ndx++)
6289 				continue;
6290 
6291 			dtrace_probe_error(state, ecb->dte_epid, ndx,
6292 			    (mstate.dtms_present & DTRACE_MSTATE_FLTOFFS) ?
6293 			    mstate.dtms_fltoffs : -1, DTRACE_FLAGS2FLT(*flags),
6294 			    cpu_core[cpuid].cpuc_dtrace_illval);
6295 
6296 			continue;
6297 		}
6298 
6299 		if (!committed)
6300 			buf->dtb_offset = offs + ecb->dte_size;
6301 	}
6302 
6303 	if (vtime)
6304 		curthread->t_dtrace_start = dtrace_gethrtime();
6305 
6306 	dtrace_interrupt_enable(cookie);
6307 }
6308 
6309 /*
6310  * DTrace Probe Hashing Functions
6311  *
6312  * The functions in this section (and indeed, the functions in remaining
6313  * sections) are not _called_ from probe context.  (Any exceptions to this are
6314  * marked with a "Note:".)  Rather, they are called from elsewhere in the
6315  * DTrace framework to look-up probes in, add probes to and remove probes from
6316  * the DTrace probe hashes.  (Each probe is hashed by each element of the
6317  * probe tuple -- allowing for fast lookups, regardless of what was
6318  * specified.)
6319  */
6320 static uint_t
6321 dtrace_hash_str(char *p)
6322 {
6323 	unsigned int g;
6324 	uint_t hval = 0;
6325 
6326 	while (*p) {
6327 		hval = (hval << 4) + *p++;
6328 		if ((g = (hval & 0xf0000000)) != 0)
6329 			hval ^= g >> 24;
6330 		hval &= ~g;
6331 	}
6332 	return (hval);
6333 }
6334 
6335 static dtrace_hash_t *
6336 dtrace_hash_create(uintptr_t stroffs, uintptr_t nextoffs, uintptr_t prevoffs)
6337 {
6338 	dtrace_hash_t *hash = kmem_zalloc(sizeof (dtrace_hash_t), KM_SLEEP);
6339 
6340 	hash->dth_stroffs = stroffs;
6341 	hash->dth_nextoffs = nextoffs;
6342 	hash->dth_prevoffs = prevoffs;
6343 
6344 	hash->dth_size = 1;
6345 	hash->dth_mask = hash->dth_size - 1;
6346 
6347 	hash->dth_tab = kmem_zalloc(hash->dth_size *
6348 	    sizeof (dtrace_hashbucket_t *), KM_SLEEP);
6349 
6350 	return (hash);
6351 }
6352 
6353 static void
6354 dtrace_hash_destroy(dtrace_hash_t *hash)
6355 {
6356 #ifdef DEBUG
6357 	int i;
6358 
6359 	for (i = 0; i < hash->dth_size; i++)
6360 		ASSERT(hash->dth_tab[i] == NULL);
6361 #endif
6362 
6363 	kmem_free(hash->dth_tab,
6364 	    hash->dth_size * sizeof (dtrace_hashbucket_t *));
6365 	kmem_free(hash, sizeof (dtrace_hash_t));
6366 }
6367 
6368 static void
6369 dtrace_hash_resize(dtrace_hash_t *hash)
6370 {
6371 	int size = hash->dth_size, i, ndx;
6372 	int new_size = hash->dth_size << 1;
6373 	int new_mask = new_size - 1;
6374 	dtrace_hashbucket_t **new_tab, *bucket, *next;
6375 
6376 	ASSERT((new_size & new_mask) == 0);
6377 
6378 	new_tab = kmem_zalloc(new_size * sizeof (void *), KM_SLEEP);
6379 
6380 	for (i = 0; i < size; i++) {
6381 		for (bucket = hash->dth_tab[i]; bucket != NULL; bucket = next) {
6382 			dtrace_probe_t *probe = bucket->dthb_chain;
6383 
6384 			ASSERT(probe != NULL);
6385 			ndx = DTRACE_HASHSTR(hash, probe) & new_mask;
6386 
6387 			next = bucket->dthb_next;
6388 			bucket->dthb_next = new_tab[ndx];
6389 			new_tab[ndx] = bucket;
6390 		}
6391 	}
6392 
6393 	kmem_free(hash->dth_tab, hash->dth_size * sizeof (void *));
6394 	hash->dth_tab = new_tab;
6395 	hash->dth_size = new_size;
6396 	hash->dth_mask = new_mask;
6397 }
6398 
6399 static void
6400 dtrace_hash_add(dtrace_hash_t *hash, dtrace_probe_t *new)
6401 {
6402 	int hashval = DTRACE_HASHSTR(hash, new);
6403 	int ndx = hashval & hash->dth_mask;
6404 	dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6405 	dtrace_probe_t **nextp, **prevp;
6406 
6407 	for (; bucket != NULL; bucket = bucket->dthb_next) {
6408 		if (DTRACE_HASHEQ(hash, bucket->dthb_chain, new))
6409 			goto add;
6410 	}
6411 
6412 	if ((hash->dth_nbuckets >> 1) > hash->dth_size) {
6413 		dtrace_hash_resize(hash);
6414 		dtrace_hash_add(hash, new);
6415 		return;
6416 	}
6417 
6418 	bucket = kmem_zalloc(sizeof (dtrace_hashbucket_t), KM_SLEEP);
6419 	bucket->dthb_next = hash->dth_tab[ndx];
6420 	hash->dth_tab[ndx] = bucket;
6421 	hash->dth_nbuckets++;
6422 
6423 add:
6424 	nextp = DTRACE_HASHNEXT(hash, new);
6425 	ASSERT(*nextp == NULL && *(DTRACE_HASHPREV(hash, new)) == NULL);
6426 	*nextp = bucket->dthb_chain;
6427 
6428 	if (bucket->dthb_chain != NULL) {
6429 		prevp = DTRACE_HASHPREV(hash, bucket->dthb_chain);
6430 		ASSERT(*prevp == NULL);
6431 		*prevp = new;
6432 	}
6433 
6434 	bucket->dthb_chain = new;
6435 	bucket->dthb_len++;
6436 }
6437 
6438 static dtrace_probe_t *
6439 dtrace_hash_lookup(dtrace_hash_t *hash, dtrace_probe_t *template)
6440 {
6441 	int hashval = DTRACE_HASHSTR(hash, template);
6442 	int ndx = hashval & hash->dth_mask;
6443 	dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6444 
6445 	for (; bucket != NULL; bucket = bucket->dthb_next) {
6446 		if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
6447 			return (bucket->dthb_chain);
6448 	}
6449 
6450 	return (NULL);
6451 }
6452 
6453 static int
6454 dtrace_hash_collisions(dtrace_hash_t *hash, dtrace_probe_t *template)
6455 {
6456 	int hashval = DTRACE_HASHSTR(hash, template);
6457 	int ndx = hashval & hash->dth_mask;
6458 	dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6459 
6460 	for (; bucket != NULL; bucket = bucket->dthb_next) {
6461 		if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
6462 			return (bucket->dthb_len);
6463 	}
6464 
6465 	return (NULL);
6466 }
6467 
6468 static void
6469 dtrace_hash_remove(dtrace_hash_t *hash, dtrace_probe_t *probe)
6470 {
6471 	int ndx = DTRACE_HASHSTR(hash, probe) & hash->dth_mask;
6472 	dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6473 
6474 	dtrace_probe_t **prevp = DTRACE_HASHPREV(hash, probe);
6475 	dtrace_probe_t **nextp = DTRACE_HASHNEXT(hash, probe);
6476 
6477 	/*
6478 	 * Find the bucket that we're removing this probe from.
6479 	 */
6480 	for (; bucket != NULL; bucket = bucket->dthb_next) {
6481 		if (DTRACE_HASHEQ(hash, bucket->dthb_chain, probe))
6482 			break;
6483 	}
6484 
6485 	ASSERT(bucket != NULL);
6486 
6487 	if (*prevp == NULL) {
6488 		if (*nextp == NULL) {
6489 			/*
6490 			 * The removed probe was the only probe on this
6491 			 * bucket; we need to remove the bucket.
6492 			 */
6493 			dtrace_hashbucket_t *b = hash->dth_tab[ndx];
6494 
6495 			ASSERT(bucket->dthb_chain == probe);
6496 			ASSERT(b != NULL);
6497 
6498 			if (b == bucket) {
6499 				hash->dth_tab[ndx] = bucket->dthb_next;
6500 			} else {
6501 				while (b->dthb_next != bucket)
6502 					b = b->dthb_next;
6503 				b->dthb_next = bucket->dthb_next;
6504 			}
6505 
6506 			ASSERT(hash->dth_nbuckets > 0);
6507 			hash->dth_nbuckets--;
6508 			kmem_free(bucket, sizeof (dtrace_hashbucket_t));
6509 			return;
6510 		}
6511 
6512 		bucket->dthb_chain = *nextp;
6513 	} else {
6514 		*(DTRACE_HASHNEXT(hash, *prevp)) = *nextp;
6515 	}
6516 
6517 	if (*nextp != NULL)
6518 		*(DTRACE_HASHPREV(hash, *nextp)) = *prevp;
6519 }
6520 
6521 /*
6522  * DTrace Utility Functions
6523  *
6524  * These are random utility functions that are _not_ called from probe context.
6525  */
6526 static int
6527 dtrace_badattr(const dtrace_attribute_t *a)
6528 {
6529 	return (a->dtat_name > DTRACE_STABILITY_MAX ||
6530 	    a->dtat_data > DTRACE_STABILITY_MAX ||
6531 	    a->dtat_class > DTRACE_CLASS_MAX);
6532 }
6533 
6534 /*
6535  * Return a duplicate copy of a string.  If the specified string is NULL,
6536  * this function returns a zero-length string.
6537  */
6538 static char *
6539 dtrace_strdup(const char *str)
6540 {
6541 	char *new = kmem_zalloc((str != NULL ? strlen(str) : 0) + 1, KM_SLEEP);
6542 
6543 	if (str != NULL)
6544 		(void) strcpy(new, str);
6545 
6546 	return (new);
6547 }
6548 
6549 #define	DTRACE_ISALPHA(c)	\
6550 	(((c) >= 'a' && (c) <= 'z') || ((c) >= 'A' && (c) <= 'Z'))
6551 
6552 static int
6553 dtrace_badname(const char *s)
6554 {
6555 	char c;
6556 
6557 	if (s == NULL || (c = *s++) == '\0')
6558 		return (0);
6559 
6560 	if (!DTRACE_ISALPHA(c) && c != '-' && c != '_' && c != '.')
6561 		return (1);
6562 
6563 	while ((c = *s++) != '\0') {
6564 		if (!DTRACE_ISALPHA(c) && (c < '0' || c > '9') &&
6565 		    c != '-' && c != '_' && c != '.' && c != '`')
6566 			return (1);
6567 	}
6568 
6569 	return (0);
6570 }
6571 
6572 static void
6573 dtrace_cred2priv(cred_t *cr, uint32_t *privp, uid_t *uidp, zoneid_t *zoneidp)
6574 {
6575 	uint32_t priv;
6576 
6577 	if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
6578 		/*
6579 		 * For DTRACE_PRIV_ALL, the uid and zoneid don't matter.
6580 		 */
6581 		priv = DTRACE_PRIV_ALL;
6582 	} else {
6583 		*uidp = crgetuid(cr);
6584 		*zoneidp = crgetzoneid(cr);
6585 
6586 		priv = 0;
6587 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE))
6588 			priv |= DTRACE_PRIV_KERNEL | DTRACE_PRIV_USER;
6589 		else if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE))
6590 			priv |= DTRACE_PRIV_USER;
6591 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE))
6592 			priv |= DTRACE_PRIV_PROC;
6593 		if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
6594 			priv |= DTRACE_PRIV_OWNER;
6595 		if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
6596 			priv |= DTRACE_PRIV_ZONEOWNER;
6597 	}
6598 
6599 	*privp = priv;
6600 }
6601 
6602 #ifdef DTRACE_ERRDEBUG
6603 static void
6604 dtrace_errdebug(const char *str)
6605 {
6606 	int hval = dtrace_hash_str((char *)str) % DTRACE_ERRHASHSZ;
6607 	int occupied = 0;
6608 
6609 	mutex_enter(&dtrace_errlock);
6610 	dtrace_errlast = str;
6611 	dtrace_errthread = curthread;
6612 
6613 	while (occupied++ < DTRACE_ERRHASHSZ) {
6614 		if (dtrace_errhash[hval].dter_msg == str) {
6615 			dtrace_errhash[hval].dter_count++;
6616 			goto out;
6617 		}
6618 
6619 		if (dtrace_errhash[hval].dter_msg != NULL) {
6620 			hval = (hval + 1) % DTRACE_ERRHASHSZ;
6621 			continue;
6622 		}
6623 
6624 		dtrace_errhash[hval].dter_msg = str;
6625 		dtrace_errhash[hval].dter_count = 1;
6626 		goto out;
6627 	}
6628 
6629 	panic("dtrace: undersized error hash");
6630 out:
6631 	mutex_exit(&dtrace_errlock);
6632 }
6633 #endif
6634 
6635 /*
6636  * DTrace Matching Functions
6637  *
6638  * These functions are used to match groups of probes, given some elements of
6639  * a probe tuple, or some globbed expressions for elements of a probe tuple.
6640  */
6641 static int
6642 dtrace_match_priv(const dtrace_probe_t *prp, uint32_t priv, uid_t uid,
6643     zoneid_t zoneid)
6644 {
6645 	if (priv != DTRACE_PRIV_ALL) {
6646 		uint32_t ppriv = prp->dtpr_provider->dtpv_priv.dtpp_flags;
6647 		uint32_t match = priv & ppriv;
6648 
6649 		/*
6650 		 * No PRIV_DTRACE_* privileges...
6651 		 */
6652 		if ((priv & (DTRACE_PRIV_PROC | DTRACE_PRIV_USER |
6653 		    DTRACE_PRIV_KERNEL)) == 0)
6654 			return (0);
6655 
6656 		/*
6657 		 * No matching bits, but there were bits to match...
6658 		 */
6659 		if (match == 0 && ppriv != 0)
6660 			return (0);
6661 
6662 		/*
6663 		 * Need to have permissions to the process, but don't...
6664 		 */
6665 		if (((ppriv & ~match) & DTRACE_PRIV_OWNER) != 0 &&
6666 		    uid != prp->dtpr_provider->dtpv_priv.dtpp_uid) {
6667 			return (0);
6668 		}
6669 
6670 		/*
6671 		 * Need to be in the same zone unless we possess the
6672 		 * privilege to examine all zones.
6673 		 */
6674 		if (((ppriv & ~match) & DTRACE_PRIV_ZONEOWNER) != 0 &&
6675 		    zoneid != prp->dtpr_provider->dtpv_priv.dtpp_zoneid) {
6676 			return (0);
6677 		}
6678 	}
6679 
6680 	return (1);
6681 }
6682 
6683 /*
6684  * dtrace_match_probe compares a dtrace_probe_t to a pre-compiled key, which
6685  * consists of input pattern strings and an ops-vector to evaluate them.
6686  * This function returns >0 for match, 0 for no match, and <0 for error.
6687  */
6688 static int
6689 dtrace_match_probe(const dtrace_probe_t *prp, const dtrace_probekey_t *pkp,
6690     uint32_t priv, uid_t uid, zoneid_t zoneid)
6691 {
6692 	dtrace_provider_t *pvp = prp->dtpr_provider;
6693 	int rv;
6694 
6695 	if (pvp->dtpv_defunct)
6696 		return (0);
6697 
6698 	if ((rv = pkp->dtpk_pmatch(pvp->dtpv_name, pkp->dtpk_prov, 0)) <= 0)
6699 		return (rv);
6700 
6701 	if ((rv = pkp->dtpk_mmatch(prp->dtpr_mod, pkp->dtpk_mod, 0)) <= 0)
6702 		return (rv);
6703 
6704 	if ((rv = pkp->dtpk_fmatch(prp->dtpr_func, pkp->dtpk_func, 0)) <= 0)
6705 		return (rv);
6706 
6707 	if ((rv = pkp->dtpk_nmatch(prp->dtpr_name, pkp->dtpk_name, 0)) <= 0)
6708 		return (rv);
6709 
6710 	if (dtrace_match_priv(prp, priv, uid, zoneid) == 0)
6711 		return (0);
6712 
6713 	return (rv);
6714 }
6715 
6716 /*
6717  * dtrace_match_glob() is a safe kernel implementation of the gmatch(3GEN)
6718  * interface for matching a glob pattern 'p' to an input string 's'.  Unlike
6719  * libc's version, the kernel version only applies to 8-bit ASCII strings.
6720  * In addition, all of the recursion cases except for '*' matching have been
6721  * unwound.  For '*', we still implement recursive evaluation, but a depth
6722  * counter is maintained and matching is aborted if we recurse too deep.
6723  * The function returns 0 if no match, >0 if match, and <0 if recursion error.
6724  */
6725 static int
6726 dtrace_match_glob(const char *s, const char *p, int depth)
6727 {
6728 	const char *olds;
6729 	char s1, c;
6730 	int gs;
6731 
6732 	if (depth > DTRACE_PROBEKEY_MAXDEPTH)
6733 		return (-1);
6734 
6735 	if (s == NULL)
6736 		s = ""; /* treat NULL as empty string */
6737 
6738 top:
6739 	olds = s;
6740 	s1 = *s++;
6741 
6742 	if (p == NULL)
6743 		return (0);
6744 
6745 	if ((c = *p++) == '\0')
6746 		return (s1 == '\0');
6747 
6748 	switch (c) {
6749 	case '[': {
6750 		int ok = 0, notflag = 0;
6751 		char lc = '\0';
6752 
6753 		if (s1 == '\0')
6754 			return (0);
6755 
6756 		if (*p == '!') {
6757 			notflag = 1;
6758 			p++;
6759 		}
6760 
6761 		if ((c = *p++) == '\0')
6762 			return (0);
6763 
6764 		do {
6765 			if (c == '-' && lc != '\0' && *p != ']') {
6766 				if ((c = *p++) == '\0')
6767 					return (0);
6768 				if (c == '\\' && (c = *p++) == '\0')
6769 					return (0);
6770 
6771 				if (notflag) {
6772 					if (s1 < lc || s1 > c)
6773 						ok++;
6774 					else
6775 						return (0);
6776 				} else if (lc <= s1 && s1 <= c)
6777 					ok++;
6778 
6779 			} else if (c == '\\' && (c = *p++) == '\0')
6780 				return (0);
6781 
6782 			lc = c; /* save left-hand 'c' for next iteration */
6783 
6784 			if (notflag) {
6785 				if (s1 != c)
6786 					ok++;
6787 				else
6788 					return (0);
6789 			} else if (s1 == c)
6790 				ok++;
6791 
6792 			if ((c = *p++) == '\0')
6793 				return (0);
6794 
6795 		} while (c != ']');
6796 
6797 		if (ok)
6798 			goto top;
6799 
6800 		return (0);
6801 	}
6802 
6803 	case '\\':
6804 		if ((c = *p++) == '\0')
6805 			return (0);
6806 		/*FALLTHRU*/
6807 
6808 	default:
6809 		if (c != s1)
6810 			return (0);
6811 		/*FALLTHRU*/
6812 
6813 	case '?':
6814 		if (s1 != '\0')
6815 			goto top;
6816 		return (0);
6817 
6818 	case '*':
6819 		while (*p == '*')
6820 			p++; /* consecutive *'s are identical to a single one */
6821 
6822 		if (*p == '\0')
6823 			return (1);
6824 
6825 		for (s = olds; *s != '\0'; s++) {
6826 			if ((gs = dtrace_match_glob(s, p, depth + 1)) != 0)
6827 				return (gs);
6828 		}
6829 
6830 		return (0);
6831 	}
6832 }
6833 
6834 /*ARGSUSED*/
6835 static int
6836 dtrace_match_string(const char *s, const char *p, int depth)
6837 {
6838 	return (s != NULL && strcmp(s, p) == 0);
6839 }
6840 
6841 /*ARGSUSED*/
6842 static int
6843 dtrace_match_nul(const char *s, const char *p, int depth)
6844 {
6845 	return (1); /* always match the empty pattern */
6846 }
6847 
6848 /*ARGSUSED*/
6849 static int
6850 dtrace_match_nonzero(const char *s, const char *p, int depth)
6851 {
6852 	return (s != NULL && s[0] != '\0');
6853 }
6854 
6855 static int
6856 dtrace_match(const dtrace_probekey_t *pkp, uint32_t priv, uid_t uid,
6857     zoneid_t zoneid, int (*matched)(dtrace_probe_t *, void *), void *arg)
6858 {
6859 	dtrace_probe_t template, *probe;
6860 	dtrace_hash_t *hash = NULL;
6861 	int len, rc, best = INT_MAX, nmatched = 0;
6862 	dtrace_id_t i;
6863 
6864 	ASSERT(MUTEX_HELD(&dtrace_lock));
6865 
6866 	/*
6867 	 * If the probe ID is specified in the key, just lookup by ID and
6868 	 * invoke the match callback once if a matching probe is found.
6869 	 */
6870 	if (pkp->dtpk_id != DTRACE_IDNONE) {
6871 		if ((probe = dtrace_probe_lookup_id(pkp->dtpk_id)) != NULL &&
6872 		    dtrace_match_probe(probe, pkp, priv, uid, zoneid) > 0) {
6873 			if ((*matched)(probe, arg) == DTRACE_MATCH_FAIL)
6874 				return (DTRACE_MATCH_FAIL);
6875 			nmatched++;
6876 		}
6877 		return (nmatched);
6878 	}
6879 
6880 	template.dtpr_mod = (char *)pkp->dtpk_mod;
6881 	template.dtpr_func = (char *)pkp->dtpk_func;
6882 	template.dtpr_name = (char *)pkp->dtpk_name;
6883 
6884 	/*
6885 	 * We want to find the most distinct of the module name, function
6886 	 * name, and name.  So for each one that is not a glob pattern or
6887 	 * empty string, we perform a lookup in the corresponding hash and
6888 	 * use the hash table with the fewest collisions to do our search.
6889 	 */
6890 	if (pkp->dtpk_mmatch == &dtrace_match_string &&
6891 	    (len = dtrace_hash_collisions(dtrace_bymod, &template)) < best) {
6892 		best = len;
6893 		hash = dtrace_bymod;
6894 	}
6895 
6896 	if (pkp->dtpk_fmatch == &dtrace_match_string &&
6897 	    (len = dtrace_hash_collisions(dtrace_byfunc, &template)) < best) {
6898 		best = len;
6899 		hash = dtrace_byfunc;
6900 	}
6901 
6902 	if (pkp->dtpk_nmatch == &dtrace_match_string &&
6903 	    (len = dtrace_hash_collisions(dtrace_byname, &template)) < best) {
6904 		best = len;
6905 		hash = dtrace_byname;
6906 	}
6907 
6908 	/*
6909 	 * If we did not select a hash table, iterate over every probe and
6910 	 * invoke our callback for each one that matches our input probe key.
6911 	 */
6912 	if (hash == NULL) {
6913 		for (i = 0; i < dtrace_nprobes; i++) {
6914 			if ((probe = dtrace_probes[i]) == NULL ||
6915 			    dtrace_match_probe(probe, pkp, priv, uid,
6916 			    zoneid) <= 0)
6917 				continue;
6918 
6919 			nmatched++;
6920 
6921 			if ((rc = (*matched)(probe, arg)) !=
6922 			    DTRACE_MATCH_NEXT) {
6923 				if (rc == DTRACE_MATCH_FAIL)
6924 					return (DTRACE_MATCH_FAIL);
6925 				break;
6926 			}
6927 		}
6928 
6929 		return (nmatched);
6930 	}
6931 
6932 	/*
6933 	 * If we selected a hash table, iterate over each probe of the same key
6934 	 * name and invoke the callback for every probe that matches the other
6935 	 * attributes of our input probe key.
6936 	 */
6937 	for (probe = dtrace_hash_lookup(hash, &template); probe != NULL;
6938 	    probe = *(DTRACE_HASHNEXT(hash, probe))) {
6939 
6940 		if (dtrace_match_probe(probe, pkp, priv, uid, zoneid) <= 0)
6941 			continue;
6942 
6943 		nmatched++;
6944 
6945 		if ((rc = (*matched)(probe, arg)) != DTRACE_MATCH_NEXT) {
6946 			if (rc == DTRACE_MATCH_FAIL)
6947 				return (DTRACE_MATCH_FAIL);
6948 			break;
6949 		}
6950 	}
6951 
6952 	return (nmatched);
6953 }
6954 
6955 /*
6956  * Return the function pointer dtrace_probecmp() should use to compare the
6957  * specified pattern with a string.  For NULL or empty patterns, we select
6958  * dtrace_match_nul().  For glob pattern strings, we use dtrace_match_glob().
6959  * For non-empty non-glob strings, we use dtrace_match_string().
6960  */
6961 static dtrace_probekey_f *
6962 dtrace_probekey_func(const char *p)
6963 {
6964 	char c;
6965 
6966 	if (p == NULL || *p == '\0')
6967 		return (&dtrace_match_nul);
6968 
6969 	while ((c = *p++) != '\0') {
6970 		if (c == '[' || c == '?' || c == '*' || c == '\\')
6971 			return (&dtrace_match_glob);
6972 	}
6973 
6974 	return (&dtrace_match_string);
6975 }
6976 
6977 /*
6978  * Build a probe comparison key for use with dtrace_match_probe() from the
6979  * given probe description.  By convention, a null key only matches anchored
6980  * probes: if each field is the empty string, reset dtpk_fmatch to
6981  * dtrace_match_nonzero().
6982  */
6983 static void
6984 dtrace_probekey(const dtrace_probedesc_t *pdp, dtrace_probekey_t *pkp)
6985 {
6986 	pkp->dtpk_prov = pdp->dtpd_provider;
6987 	pkp->dtpk_pmatch = dtrace_probekey_func(pdp->dtpd_provider);
6988 
6989 	pkp->dtpk_mod = pdp->dtpd_mod;
6990 	pkp->dtpk_mmatch = dtrace_probekey_func(pdp->dtpd_mod);
6991 
6992 	pkp->dtpk_func = pdp->dtpd_func;
6993 	pkp->dtpk_fmatch = dtrace_probekey_func(pdp->dtpd_func);
6994 
6995 	pkp->dtpk_name = pdp->dtpd_name;
6996 	pkp->dtpk_nmatch = dtrace_probekey_func(pdp->dtpd_name);
6997 
6998 	pkp->dtpk_id = pdp->dtpd_id;
6999 
7000 	if (pkp->dtpk_id == DTRACE_IDNONE &&
7001 	    pkp->dtpk_pmatch == &dtrace_match_nul &&
7002 	    pkp->dtpk_mmatch == &dtrace_match_nul &&
7003 	    pkp->dtpk_fmatch == &dtrace_match_nul &&
7004 	    pkp->dtpk_nmatch == &dtrace_match_nul)
7005 		pkp->dtpk_fmatch = &dtrace_match_nonzero;
7006 }
7007 
7008 /*
7009  * DTrace Provider-to-Framework API Functions
7010  *
7011  * These functions implement much of the Provider-to-Framework API, as
7012  * described in <sys/dtrace.h>.  The parts of the API not in this section are
7013  * the functions in the API for probe management (found below), and
7014  * dtrace_probe() itself (found above).
7015  */
7016 
7017 /*
7018  * Register the calling provider with the DTrace framework.  This should
7019  * generally be called by DTrace providers in their attach(9E) entry point.
7020  */
7021 int
7022 dtrace_register(const char *name, const dtrace_pattr_t *pap, uint32_t priv,
7023     cred_t *cr, const dtrace_pops_t *pops, void *arg, dtrace_provider_id_t *idp)
7024 {
7025 	dtrace_provider_t *provider;
7026 
7027 	if (name == NULL || pap == NULL || pops == NULL || idp == NULL) {
7028 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7029 		    "arguments", name ? name : "<NULL>");
7030 		return (EINVAL);
7031 	}
7032 
7033 	if (name[0] == '\0' || dtrace_badname(name)) {
7034 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7035 		    "provider name", name);
7036 		return (EINVAL);
7037 	}
7038 
7039 	if ((pops->dtps_provide == NULL && pops->dtps_provide_module == NULL) ||
7040 	    pops->dtps_enable == NULL || pops->dtps_disable == NULL ||
7041 	    pops->dtps_destroy == NULL ||
7042 	    ((pops->dtps_resume == NULL) != (pops->dtps_suspend == NULL))) {
7043 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7044 		    "provider ops", name);
7045 		return (EINVAL);
7046 	}
7047 
7048 	if (dtrace_badattr(&pap->dtpa_provider) ||
7049 	    dtrace_badattr(&pap->dtpa_mod) ||
7050 	    dtrace_badattr(&pap->dtpa_func) ||
7051 	    dtrace_badattr(&pap->dtpa_name) ||
7052 	    dtrace_badattr(&pap->dtpa_args)) {
7053 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7054 		    "provider attributes", name);
7055 		return (EINVAL);
7056 	}
7057 
7058 	if (priv & ~DTRACE_PRIV_ALL) {
7059 		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7060 		    "privilege attributes", name);
7061 		return (EINVAL);
7062 	}
7063 
7064 	if ((priv & DTRACE_PRIV_KERNEL) &&
7065 	    (priv & (DTRACE_PRIV_USER | DTRACE_PRIV_OWNER)) &&
7066 	    pops->dtps_mode == NULL) {
7067 		cmn_err(CE_WARN, "failed to register provider '%s': need "
7068 		    "dtps_mode() op for given privilege attributes", name);
7069 		return (EINVAL);
7070 	}
7071 
7072 	provider = kmem_zalloc(sizeof (dtrace_provider_t), KM_SLEEP);
7073 	provider->dtpv_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
7074 	(void) strcpy(provider->dtpv_name, name);
7075 
7076 	provider->dtpv_attr = *pap;
7077 	provider->dtpv_priv.dtpp_flags = priv;
7078 	if (cr != NULL) {
7079 		provider->dtpv_priv.dtpp_uid = crgetuid(cr);
7080 		provider->dtpv_priv.dtpp_zoneid = crgetzoneid(cr);
7081 	}
7082 	provider->dtpv_pops = *pops;
7083 
7084 	if (pops->dtps_provide == NULL) {
7085 		ASSERT(pops->dtps_provide_module != NULL);
7086 		provider->dtpv_pops.dtps_provide =
7087 		    (void (*)(void *, const dtrace_probedesc_t *))dtrace_nullop;
7088 	}
7089 
7090 	if (pops->dtps_provide_module == NULL) {
7091 		ASSERT(pops->dtps_provide != NULL);
7092 		provider->dtpv_pops.dtps_provide_module =
7093 		    (void (*)(void *, struct modctl *))dtrace_nullop;
7094 	}
7095 
7096 	if (pops->dtps_suspend == NULL) {
7097 		ASSERT(pops->dtps_resume == NULL);
7098 		provider->dtpv_pops.dtps_suspend =
7099 		    (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
7100 		provider->dtpv_pops.dtps_resume =
7101 		    (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
7102 	}
7103 
7104 	provider->dtpv_arg = arg;
7105 	*idp = (dtrace_provider_id_t)provider;
7106 
7107 	if (pops == &dtrace_provider_ops) {
7108 		ASSERT(MUTEX_HELD(&dtrace_provider_lock));
7109 		ASSERT(MUTEX_HELD(&dtrace_lock));
7110 		ASSERT(dtrace_anon.dta_enabling == NULL);
7111 
7112 		/*
7113 		 * We make sure that the DTrace provider is at the head of
7114 		 * the provider chain.
7115 		 */
7116 		provider->dtpv_next = dtrace_provider;
7117 		dtrace_provider = provider;
7118 		return (0);
7119 	}
7120 
7121 	mutex_enter(&dtrace_provider_lock);
7122 	mutex_enter(&dtrace_lock);
7123 
7124 	/*
7125 	 * If there is at least one provider registered, we'll add this
7126 	 * provider after the first provider.
7127 	 */
7128 	if (dtrace_provider != NULL) {
7129 		provider->dtpv_next = dtrace_provider->dtpv_next;
7130 		dtrace_provider->dtpv_next = provider;
7131 	} else {
7132 		dtrace_provider = provider;
7133 	}
7134 
7135 	if (dtrace_retained != NULL) {
7136 		dtrace_enabling_provide(provider);
7137 
7138 		/*
7139 		 * Now we need to call dtrace_enabling_matchall() -- which
7140 		 * will acquire cpu_lock and dtrace_lock.  We therefore need
7141 		 * to drop all of our locks before calling into it...
7142 		 */
7143 		mutex_exit(&dtrace_lock);
7144 		mutex_exit(&dtrace_provider_lock);
7145 		dtrace_enabling_matchall();
7146 
7147 		return (0);
7148 	}
7149 
7150 	mutex_exit(&dtrace_lock);
7151 	mutex_exit(&dtrace_provider_lock);
7152 
7153 	return (0);
7154 }
7155 
7156 /*
7157  * Unregister the specified provider from the DTrace framework.  This should
7158  * generally be called by DTrace providers in their detach(9E) entry point.
7159  */
7160 int
7161 dtrace_unregister(dtrace_provider_id_t id)
7162 {
7163 	dtrace_provider_t *old = (dtrace_provider_t *)id;
7164 	dtrace_provider_t *prev = NULL;
7165 	int i, self = 0, noreap = 0;
7166 	dtrace_probe_t *probe, *first = NULL;
7167 
7168 	if (old->dtpv_pops.dtps_enable ==
7169 	    (int (*)(void *, dtrace_id_t, void *))dtrace_enable_nullop) {
7170 		/*
7171 		 * If DTrace itself is the provider, we're called with locks
7172 		 * already held.
7173 		 */
7174 		ASSERT(old == dtrace_provider);
7175 		ASSERT(dtrace_devi != NULL);
7176 		ASSERT(MUTEX_HELD(&dtrace_provider_lock));
7177 		ASSERT(MUTEX_HELD(&dtrace_lock));
7178 		self = 1;
7179 
7180 		if (dtrace_provider->dtpv_next != NULL) {
7181 			/*
7182 			 * There's another provider here; return failure.
7183 			 */
7184 			return (EBUSY);
7185 		}
7186 	} else {
7187 		mutex_enter(&dtrace_provider_lock);
7188 		mutex_enter(&mod_lock);
7189 		mutex_enter(&dtrace_lock);
7190 	}
7191 
7192 	/*
7193 	 * If anyone has /dev/dtrace open, or if there are anonymous enabled
7194 	 * probes, we refuse to let providers slither away, unless this
7195 	 * provider has already been explicitly invalidated.
7196 	 */
7197 	if (!old->dtpv_defunct &&
7198 	    (dtrace_opens || (dtrace_anon.dta_state != NULL &&
7199 	    dtrace_anon.dta_state->dts_necbs > 0))) {
7200 		if (!self) {
7201 			mutex_exit(&dtrace_lock);
7202 			mutex_exit(&mod_lock);
7203 			mutex_exit(&dtrace_provider_lock);
7204 		}
7205 		return (EBUSY);
7206 	}
7207 
7208 	/*
7209 	 * Attempt to destroy the probes associated with this provider.
7210 	 */
7211 	for (i = 0; i < dtrace_nprobes; i++) {
7212 		if ((probe = dtrace_probes[i]) == NULL)
7213 			continue;
7214 
7215 		if (probe->dtpr_provider != old)
7216 			continue;
7217 
7218 		if (probe->dtpr_ecb == NULL)
7219 			continue;
7220 
7221 		/*
7222 		 * If we are trying to unregister a defunct provider, and the
7223 		 * provider was made defunct within the interval dictated by
7224 		 * dtrace_unregister_defunct_reap, we'll (asynchronously)
7225 		 * attempt to reap our enablings.  To denote that the provider
7226 		 * should reattempt to unregister itself at some point in the
7227 		 * future, we will return a differentiable error code (EAGAIN
7228 		 * instead of EBUSY) in this case.
7229 		 */
7230 		if (dtrace_gethrtime() - old->dtpv_defunct >
7231 		    dtrace_unregister_defunct_reap)
7232 			noreap = 1;
7233 
7234 		if (!self) {
7235 			mutex_exit(&dtrace_lock);
7236 			mutex_exit(&mod_lock);
7237 			mutex_exit(&dtrace_provider_lock);
7238 		}
7239 
7240 		if (noreap)
7241 			return (EBUSY);
7242 
7243 		(void) taskq_dispatch(dtrace_taskq,
7244 		    (task_func_t *)dtrace_enabling_reap, NULL, TQ_SLEEP);
7245 
7246 		return (EAGAIN);
7247 	}
7248 
7249 	/*
7250 	 * All of the probes for this provider are disabled; we can safely
7251 	 * remove all of them from their hash chains and from the probe array.
7252 	 */
7253 	for (i = 0; i < dtrace_nprobes; i++) {
7254 		if ((probe = dtrace_probes[i]) == NULL)
7255 			continue;
7256 
7257 		if (probe->dtpr_provider != old)
7258 			continue;
7259 
7260 		dtrace_probes[i] = NULL;
7261 
7262 		dtrace_hash_remove(dtrace_bymod, probe);
7263 		dtrace_hash_remove(dtrace_byfunc, probe);
7264 		dtrace_hash_remove(dtrace_byname, probe);
7265 
7266 		if (first == NULL) {
7267 			first = probe;
7268 			probe->dtpr_nextmod = NULL;
7269 		} else {
7270 			probe->dtpr_nextmod = first;
7271 			first = probe;
7272 		}
7273 	}
7274 
7275 	/*
7276 	 * The provider's probes have been removed from the hash chains and
7277 	 * from the probe array.  Now issue a dtrace_sync() to be sure that
7278 	 * everyone has cleared out from any probe array processing.
7279 	 */
7280 	dtrace_sync();
7281 
7282 	for (probe = first; probe != NULL; probe = first) {
7283 		first = probe->dtpr_nextmod;
7284 
7285 		old->dtpv_pops.dtps_destroy(old->dtpv_arg, probe->dtpr_id,
7286 		    probe->dtpr_arg);
7287 		kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
7288 		kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
7289 		kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
7290 		vmem_free(dtrace_arena, (void *)(uintptr_t)(probe->dtpr_id), 1);
7291 		kmem_free(probe, sizeof (dtrace_probe_t));
7292 	}
7293 
7294 	if ((prev = dtrace_provider) == old) {
7295 		ASSERT(self || dtrace_devi == NULL);
7296 		ASSERT(old->dtpv_next == NULL || dtrace_devi == NULL);
7297 		dtrace_provider = old->dtpv_next;
7298 	} else {
7299 		while (prev != NULL && prev->dtpv_next != old)
7300 			prev = prev->dtpv_next;
7301 
7302 		if (prev == NULL) {
7303 			panic("attempt to unregister non-existent "
7304 			    "dtrace provider %p\n", (void *)id);
7305 		}
7306 
7307 		prev->dtpv_next = old->dtpv_next;
7308 	}
7309 
7310 	if (!self) {
7311 		mutex_exit(&dtrace_lock);
7312 		mutex_exit(&mod_lock);
7313 		mutex_exit(&dtrace_provider_lock);
7314 	}
7315 
7316 	kmem_free(old->dtpv_name, strlen(old->dtpv_name) + 1);
7317 	kmem_free(old, sizeof (dtrace_provider_t));
7318 
7319 	return (0);
7320 }
7321 
7322 /*
7323  * Invalidate the specified provider.  All subsequent probe lookups for the
7324  * specified provider will fail, but its probes will not be removed.
7325  */
7326 void
7327 dtrace_invalidate(dtrace_provider_id_t id)
7328 {
7329 	dtrace_provider_t *pvp = (dtrace_provider_t *)id;
7330 
7331 	ASSERT(pvp->dtpv_pops.dtps_enable !=
7332 	    (int (*)(void *, dtrace_id_t, void *))dtrace_enable_nullop);
7333 
7334 	mutex_enter(&dtrace_provider_lock);
7335 	mutex_enter(&dtrace_lock);
7336 
7337 	pvp->dtpv_defunct = dtrace_gethrtime();
7338 
7339 	mutex_exit(&dtrace_lock);
7340 	mutex_exit(&dtrace_provider_lock);
7341 }
7342 
7343 /*
7344  * Indicate whether or not DTrace has attached.
7345  */
7346 int
7347 dtrace_attached(void)
7348 {
7349 	/*
7350 	 * dtrace_provider will be non-NULL iff the DTrace driver has
7351 	 * attached.  (It's non-NULL because DTrace is always itself a
7352 	 * provider.)
7353 	 */
7354 	return (dtrace_provider != NULL);
7355 }
7356 
7357 /*
7358  * Remove all the unenabled probes for the given provider.  This function is
7359  * not unlike dtrace_unregister(), except that it doesn't remove the provider
7360  * -- just as many of its associated probes as it can.
7361  */
7362 int
7363 dtrace_condense(dtrace_provider_id_t id)
7364 {
7365 	dtrace_provider_t *prov = (dtrace_provider_t *)id;
7366 	int i;
7367 	dtrace_probe_t *probe;
7368 
7369 	/*
7370 	 * Make sure this isn't the dtrace provider itself.
7371 	 */
7372 	ASSERT(prov->dtpv_pops.dtps_enable !=
7373 	    (int (*)(void *, dtrace_id_t, void *))dtrace_enable_nullop);
7374 
7375 	mutex_enter(&dtrace_provider_lock);
7376 	mutex_enter(&dtrace_lock);
7377 
7378 	/*
7379 	 * Attempt to destroy the probes associated with this provider.
7380 	 */
7381 	for (i = 0; i < dtrace_nprobes; i++) {
7382 		if ((probe = dtrace_probes[i]) == NULL)
7383 			continue;
7384 
7385 		if (probe->dtpr_provider != prov)
7386 			continue;
7387 
7388 		if (probe->dtpr_ecb != NULL)
7389 			continue;
7390 
7391 		dtrace_probes[i] = NULL;
7392 
7393 		dtrace_hash_remove(dtrace_bymod, probe);
7394 		dtrace_hash_remove(dtrace_byfunc, probe);
7395 		dtrace_hash_remove(dtrace_byname, probe);
7396 
7397 		prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, i + 1,
7398 		    probe->dtpr_arg);
7399 		kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
7400 		kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
7401 		kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
7402 		kmem_free(probe, sizeof (dtrace_probe_t));
7403 		vmem_free(dtrace_arena, (void *)((uintptr_t)i + 1), 1);
7404 	}
7405 
7406 	mutex_exit(&dtrace_lock);
7407 	mutex_exit(&dtrace_provider_lock);
7408 
7409 	return (0);
7410 }
7411 
7412 /*
7413  * DTrace Probe Management Functions
7414  *
7415  * The functions in this section perform the DTrace probe management,
7416  * including functions to create probes, look-up probes, and call into the
7417  * providers to request that probes be provided.  Some of these functions are
7418  * in the Provider-to-Framework API; these functions can be identified by the
7419  * fact that they are not declared "static".
7420  */
7421 
7422 /*
7423  * Create a probe with the specified module name, function name, and name.
7424  */
7425 dtrace_id_t
7426 dtrace_probe_create(dtrace_provider_id_t prov, const char *mod,
7427     const char *func, const char *name, int aframes, void *arg)
7428 {
7429 	dtrace_probe_t *probe, **probes;
7430 	dtrace_provider_t *provider = (dtrace_provider_t *)prov;
7431 	dtrace_id_t id;
7432 
7433 	if (provider == dtrace_provider) {
7434 		ASSERT(MUTEX_HELD(&dtrace_lock));
7435 	} else {
7436 		mutex_enter(&dtrace_lock);
7437 	}
7438 
7439 	id = (dtrace_id_t)(uintptr_t)vmem_alloc(dtrace_arena, 1,
7440 	    VM_BESTFIT | VM_SLEEP);
7441 	probe = kmem_zalloc(sizeof (dtrace_probe_t), KM_SLEEP);
7442 
7443 	probe->dtpr_id = id;
7444 	probe->dtpr_gen = dtrace_probegen++;
7445 	probe->dtpr_mod = dtrace_strdup(mod);
7446 	probe->dtpr_func = dtrace_strdup(func);
7447 	probe->dtpr_name = dtrace_strdup(name);
7448 	probe->dtpr_arg = arg;
7449 	probe->dtpr_aframes = aframes;
7450 	probe->dtpr_provider = provider;
7451 
7452 	dtrace_hash_add(dtrace_bymod, probe);
7453 	dtrace_hash_add(dtrace_byfunc, probe);
7454 	dtrace_hash_add(dtrace_byname, probe);
7455 
7456 	if (id - 1 >= dtrace_nprobes) {
7457 		size_t osize = dtrace_nprobes * sizeof (dtrace_probe_t *);
7458 		size_t nsize = osize << 1;
7459 
7460 		if (nsize == 0) {
7461 			ASSERT(osize == 0);
7462 			ASSERT(dtrace_probes == NULL);
7463 			nsize = sizeof (dtrace_probe_t *);
7464 		}
7465 
7466 		probes = kmem_zalloc(nsize, KM_SLEEP);
7467 
7468 		if (dtrace_probes == NULL) {
7469 			ASSERT(osize == 0);
7470 			dtrace_probes = probes;
7471 			dtrace_nprobes = 1;
7472 		} else {
7473 			dtrace_probe_t **oprobes = dtrace_probes;
7474 
7475 			bcopy(oprobes, probes, osize);
7476 			dtrace_membar_producer();
7477 			dtrace_probes = probes;
7478 
7479 			dtrace_sync();
7480 
7481 			/*
7482 			 * All CPUs are now seeing the new probes array; we can
7483 			 * safely free the old array.
7484 			 */
7485 			kmem_free(oprobes, osize);
7486 			dtrace_nprobes <<= 1;
7487 		}
7488 
7489 		ASSERT(id - 1 < dtrace_nprobes);
7490 	}
7491 
7492 	ASSERT(dtrace_probes[id - 1] == NULL);
7493 	dtrace_probes[id - 1] = probe;
7494 
7495 	if (provider != dtrace_provider)
7496 		mutex_exit(&dtrace_lock);
7497 
7498 	return (id);
7499 }
7500 
7501 static dtrace_probe_t *
7502 dtrace_probe_lookup_id(dtrace_id_t id)
7503 {
7504 	ASSERT(MUTEX_HELD(&dtrace_lock));
7505 
7506 	if (id == 0 || id > dtrace_nprobes)
7507 		return (NULL);
7508 
7509 	return (dtrace_probes[id - 1]);
7510 }
7511 
7512 static int
7513 dtrace_probe_lookup_match(dtrace_probe_t *probe, void *arg)
7514 {
7515 	*((dtrace_id_t *)arg) = probe->dtpr_id;
7516 
7517 	return (DTRACE_MATCH_DONE);
7518 }
7519 
7520 /*
7521  * Look up a probe based on provider and one or more of module name, function
7522  * name and probe name.
7523  */
7524 dtrace_id_t
7525 dtrace_probe_lookup(dtrace_provider_id_t prid, const char *mod,
7526     const char *func, const char *name)
7527 {
7528 	dtrace_probekey_t pkey;
7529 	dtrace_id_t id;
7530 	int match;
7531 
7532 	pkey.dtpk_prov = ((dtrace_provider_t *)prid)->dtpv_name;
7533 	pkey.dtpk_pmatch = &dtrace_match_string;
7534 	pkey.dtpk_mod = mod;
7535 	pkey.dtpk_mmatch = mod ? &dtrace_match_string : &dtrace_match_nul;
7536 	pkey.dtpk_func = func;
7537 	pkey.dtpk_fmatch = func ? &dtrace_match_string : &dtrace_match_nul;
7538 	pkey.dtpk_name = name;
7539 	pkey.dtpk_nmatch = name ? &dtrace_match_string : &dtrace_match_nul;
7540 	pkey.dtpk_id = DTRACE_IDNONE;
7541 
7542 	mutex_enter(&dtrace_lock);
7543 	match = dtrace_match(&pkey, DTRACE_PRIV_ALL, 0, 0,
7544 	    dtrace_probe_lookup_match, &id);
7545 	mutex_exit(&dtrace_lock);
7546 
7547 	ASSERT(match == 1 || match == 0);
7548 	return (match ? id : 0);
7549 }
7550 
7551 /*
7552  * Returns the probe argument associated with the specified probe.
7553  */
7554 void *
7555 dtrace_probe_arg(dtrace_provider_id_t id, dtrace_id_t pid)
7556 {
7557 	dtrace_probe_t *probe;
7558 	void *rval = NULL;
7559 
7560 	mutex_enter(&dtrace_lock);
7561 
7562 	if ((probe = dtrace_probe_lookup_id(pid)) != NULL &&
7563 	    probe->dtpr_provider == (dtrace_provider_t *)id)
7564 		rval = probe->dtpr_arg;
7565 
7566 	mutex_exit(&dtrace_lock);
7567 
7568 	return (rval);
7569 }
7570 
7571 /*
7572  * Copy a probe into a probe description.
7573  */
7574 static void
7575 dtrace_probe_description(const dtrace_probe_t *prp, dtrace_probedesc_t *pdp)
7576 {
7577 	bzero(pdp, sizeof (dtrace_probedesc_t));
7578 	pdp->dtpd_id = prp->dtpr_id;
7579 
7580 	(void) strncpy(pdp->dtpd_provider,
7581 	    prp->dtpr_provider->dtpv_name, DTRACE_PROVNAMELEN - 1);
7582 
7583 	(void) strncpy(pdp->dtpd_mod, prp->dtpr_mod, DTRACE_MODNAMELEN - 1);
7584 	(void) strncpy(pdp->dtpd_func, prp->dtpr_func, DTRACE_FUNCNAMELEN - 1);
7585 	(void) strncpy(pdp->dtpd_name, prp->dtpr_name, DTRACE_NAMELEN - 1);
7586 }
7587 
7588 /*
7589  * Called to indicate that a probe -- or probes -- should be provided by a
7590  * specfied provider.  If the specified description is NULL, the provider will
7591  * be told to provide all of its probes.  (This is done whenever a new
7592  * consumer comes along, or whenever a retained enabling is to be matched.) If
7593  * the specified description is non-NULL, the provider is given the
7594  * opportunity to dynamically provide the specified probe, allowing providers
7595  * to support the creation of probes on-the-fly.  (So-called _autocreated_
7596  * probes.)  If the provider is NULL, the operations will be applied to all
7597  * providers; if the provider is non-NULL the operations will only be applied
7598  * to the specified provider.  The dtrace_provider_lock must be held, and the
7599  * dtrace_lock must _not_ be held -- the provider's dtps_provide() operation
7600  * will need to grab the dtrace_lock when it reenters the framework through
7601  * dtrace_probe_lookup(), dtrace_probe_create(), etc.
7602  */
7603 static void
7604 dtrace_probe_provide(dtrace_probedesc_t *desc, dtrace_provider_t *prv)
7605 {
7606 	struct modctl *ctl;
7607 	int all = 0;
7608 
7609 	ASSERT(MUTEX_HELD(&dtrace_provider_lock));
7610 
7611 	if (prv == NULL) {
7612 		all = 1;
7613 		prv = dtrace_provider;
7614 	}
7615 
7616 	do {
7617 		/*
7618 		 * First, call the blanket provide operation.
7619 		 */
7620 		prv->dtpv_pops.dtps_provide(prv->dtpv_arg, desc);
7621 
7622 		/*
7623 		 * Now call the per-module provide operation.  We will grab
7624 		 * mod_lock to prevent the list from being modified.  Note
7625 		 * that this also prevents the mod_busy bits from changing.
7626 		 * (mod_busy can only be changed with mod_lock held.)
7627 		 */
7628 		mutex_enter(&mod_lock);
7629 
7630 		ctl = &modules;
7631 		do {
7632 			if (ctl->mod_busy || ctl->mod_mp == NULL)
7633 				continue;
7634 
7635 			prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
7636 
7637 		} while ((ctl = ctl->mod_next) != &modules);
7638 
7639 		mutex_exit(&mod_lock);
7640 	} while (all && (prv = prv->dtpv_next) != NULL);
7641 }
7642 
7643 /*
7644  * Iterate over each probe, and call the Framework-to-Provider API function
7645  * denoted by offs.
7646  */
7647 static void
7648 dtrace_probe_foreach(uintptr_t offs)
7649 {
7650 	dtrace_provider_t *prov;
7651 	void (*func)(void *, dtrace_id_t, void *);
7652 	dtrace_probe_t *probe;
7653 	dtrace_icookie_t cookie;
7654 	int i;
7655 
7656 	/*
7657 	 * We disable interrupts to walk through the probe array.  This is
7658 	 * safe -- the dtrace_sync() in dtrace_unregister() assures that we
7659 	 * won't see stale data.
7660 	 */
7661 	cookie = dtrace_interrupt_disable();
7662 
7663 	for (i = 0; i < dtrace_nprobes; i++) {
7664 		if ((probe = dtrace_probes[i]) == NULL)
7665 			continue;
7666 
7667 		if (probe->dtpr_ecb == NULL) {
7668 			/*
7669 			 * This probe isn't enabled -- don't call the function.
7670 			 */
7671 			continue;
7672 		}
7673 
7674 		prov = probe->dtpr_provider;
7675 		func = *((void(**)(void *, dtrace_id_t, void *))
7676 		    ((uintptr_t)&prov->dtpv_pops + offs));
7677 
7678 		func(prov->dtpv_arg, i + 1, probe->dtpr_arg);
7679 	}
7680 
7681 	dtrace_interrupt_enable(cookie);
7682 }
7683 
7684 static int
7685 dtrace_probe_enable(const dtrace_probedesc_t *desc, dtrace_enabling_t *enab)
7686 {
7687 	dtrace_probekey_t pkey;
7688 	uint32_t priv;
7689 	uid_t uid;
7690 	zoneid_t zoneid;
7691 
7692 	ASSERT(MUTEX_HELD(&dtrace_lock));
7693 	dtrace_ecb_create_cache = NULL;
7694 
7695 	if (desc == NULL) {
7696 		/*
7697 		 * If we're passed a NULL description, we're being asked to
7698 		 * create an ECB with a NULL probe.
7699 		 */
7700 		(void) dtrace_ecb_create_enable(NULL, enab);
7701 		return (0);
7702 	}
7703 
7704 	dtrace_probekey(desc, &pkey);
7705 	dtrace_cred2priv(enab->dten_vstate->dtvs_state->dts_cred.dcr_cred,
7706 	    &priv, &uid, &zoneid);
7707 
7708 	return (dtrace_match(&pkey, priv, uid, zoneid, dtrace_ecb_create_enable,
7709 	    enab));
7710 }
7711 
7712 /*
7713  * DTrace Helper Provider Functions
7714  */
7715 static void
7716 dtrace_dofattr2attr(dtrace_attribute_t *attr, const dof_attr_t dofattr)
7717 {
7718 	attr->dtat_name = DOF_ATTR_NAME(dofattr);
7719 	attr->dtat_data = DOF_ATTR_DATA(dofattr);
7720 	attr->dtat_class = DOF_ATTR_CLASS(dofattr);
7721 }
7722 
7723 static void
7724 dtrace_dofprov2hprov(dtrace_helper_provdesc_t *hprov,
7725     const dof_provider_t *dofprov, char *strtab)
7726 {
7727 	hprov->dthpv_provname = strtab + dofprov->dofpv_name;
7728 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_provider,
7729 	    dofprov->dofpv_provattr);
7730 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_mod,
7731 	    dofprov->dofpv_modattr);
7732 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_func,
7733 	    dofprov->dofpv_funcattr);
7734 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_name,
7735 	    dofprov->dofpv_nameattr);
7736 	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_args,
7737 	    dofprov->dofpv_argsattr);
7738 }
7739 
7740 static void
7741 dtrace_helper_provide_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
7742 {
7743 	uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
7744 	dof_hdr_t *dof = (dof_hdr_t *)daddr;
7745 	dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
7746 	dof_provider_t *provider;
7747 	dof_probe_t *probe;
7748 	uint32_t *off, *enoff;
7749 	uint8_t *arg;
7750 	char *strtab;
7751 	uint_t i, nprobes;
7752 	dtrace_helper_provdesc_t dhpv;
7753 	dtrace_helper_probedesc_t dhpb;
7754 	dtrace_meta_t *meta = dtrace_meta_pid;
7755 	dtrace_mops_t *mops = &meta->dtm_mops;
7756 	void *parg;
7757 
7758 	provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
7759 	str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
7760 	    provider->dofpv_strtab * dof->dofh_secsize);
7761 	prb_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
7762 	    provider->dofpv_probes * dof->dofh_secsize);
7763 	arg_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
7764 	    provider->dofpv_prargs * dof->dofh_secsize);
7765 	off_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
7766 	    provider->dofpv_proffs * dof->dofh_secsize);
7767 
7768 	strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
7769 	off = (uint32_t *)(uintptr_t)(daddr + off_sec->dofs_offset);
7770 	arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
7771 	enoff = NULL;
7772 
7773 	/*
7774 	 * See dtrace_helper_provider_validate().
7775 	 */
7776 	if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
7777 	    provider->dofpv_prenoffs != DOF_SECT_NONE) {
7778 		enoff_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
7779 		    provider->dofpv_prenoffs * dof->dofh_secsize);
7780 		enoff = (uint32_t *)(uintptr_t)(daddr + enoff_sec->dofs_offset);
7781 	}
7782 
7783 	nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
7784 
7785 	/*
7786 	 * Create the provider.
7787 	 */
7788 	dtrace_dofprov2hprov(&dhpv, provider, strtab);
7789 
7790 	if ((parg = mops->dtms_provide_pid(meta->dtm_arg, &dhpv, pid)) == NULL)
7791 		return;
7792 
7793 	meta->dtm_count++;
7794 
7795 	/*
7796 	 * Create the probes.
7797 	 */
7798 	for (i = 0; i < nprobes; i++) {
7799 		probe = (dof_probe_t *)(uintptr_t)(daddr +
7800 		    prb_sec->dofs_offset + i * prb_sec->dofs_entsize);
7801 
7802 		dhpb.dthpb_mod = dhp->dofhp_mod;
7803 		dhpb.dthpb_func = strtab + probe->dofpr_func;
7804 		dhpb.dthpb_name = strtab + probe->dofpr_name;
7805 		dhpb.dthpb_base = probe->dofpr_addr;
7806 		dhpb.dthpb_offs = off + probe->dofpr_offidx;
7807 		dhpb.dthpb_noffs = probe->dofpr_noffs;
7808 		if (enoff != NULL) {
7809 			dhpb.dthpb_enoffs = enoff + probe->dofpr_enoffidx;
7810 			dhpb.dthpb_nenoffs = probe->dofpr_nenoffs;
7811 		} else {
7812 			dhpb.dthpb_enoffs = NULL;
7813 			dhpb.dthpb_nenoffs = 0;
7814 		}
7815 		dhpb.dthpb_args = arg + probe->dofpr_argidx;
7816 		dhpb.dthpb_nargc = probe->dofpr_nargc;
7817 		dhpb.dthpb_xargc = probe->dofpr_xargc;
7818 		dhpb.dthpb_ntypes = strtab + probe->dofpr_nargv;
7819 		dhpb.dthpb_xtypes = strtab + probe->dofpr_xargv;
7820 
7821 		mops->dtms_create_probe(meta->dtm_arg, parg, &dhpb);
7822 	}
7823 }
7824 
7825 static void
7826 dtrace_helper_provide(dof_helper_t *dhp, pid_t pid)
7827 {
7828 	uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
7829 	dof_hdr_t *dof = (dof_hdr_t *)daddr;
7830 	int i;
7831 
7832 	ASSERT(MUTEX_HELD(&dtrace_meta_lock));
7833 
7834 	for (i = 0; i < dof->dofh_secnum; i++) {
7835 		dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
7836 		    dof->dofh_secoff + i * dof->dofh_secsize);
7837 
7838 		if (sec->dofs_type != DOF_SECT_PROVIDER)
7839 			continue;
7840 
7841 		dtrace_helper_provide_one(dhp, sec, pid);
7842 	}
7843 
7844 	/*
7845 	 * We may have just created probes, so we must now rematch against
7846 	 * any retained enablings.  Note that this call will acquire both
7847 	 * cpu_lock and dtrace_lock; the fact that we are holding
7848 	 * dtrace_meta_lock now is what defines the ordering with respect to
7849 	 * these three locks.
7850 	 */
7851 	dtrace_enabling_matchall();
7852 }
7853 
7854 static void
7855 dtrace_helper_provider_remove_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
7856 {
7857 	uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
7858 	dof_hdr_t *dof = (dof_hdr_t *)daddr;
7859 	dof_sec_t *str_sec;
7860 	dof_provider_t *provider;
7861 	char *strtab;
7862 	dtrace_helper_provdesc_t dhpv;
7863 	dtrace_meta_t *meta = dtrace_meta_pid;
7864 	dtrace_mops_t *mops = &meta->dtm_mops;
7865 
7866 	provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
7867 	str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
7868 	    provider->dofpv_strtab * dof->dofh_secsize);
7869 
7870 	strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
7871 
7872 	/*
7873 	 * Create the provider.
7874 	 */
7875 	dtrace_dofprov2hprov(&dhpv, provider, strtab);
7876 
7877 	mops->dtms_remove_pid(meta->dtm_arg, &dhpv, pid);
7878 
7879 	meta->dtm_count--;
7880 }
7881 
7882 static void
7883 dtrace_helper_provider_remove(dof_helper_t *dhp, pid_t pid)
7884 {
7885 	uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
7886 	dof_hdr_t *dof = (dof_hdr_t *)daddr;
7887 	int i;
7888 
7889 	ASSERT(MUTEX_HELD(&dtrace_meta_lock));
7890 
7891 	for (i = 0; i < dof->dofh_secnum; i++) {
7892 		dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
7893 		    dof->dofh_secoff + i * dof->dofh_secsize);
7894 
7895 		if (sec->dofs_type != DOF_SECT_PROVIDER)
7896 			continue;
7897 
7898 		dtrace_helper_provider_remove_one(dhp, sec, pid);
7899 	}
7900 }
7901 
7902 /*
7903  * DTrace Meta Provider-to-Framework API Functions
7904  *
7905  * These functions implement the Meta Provider-to-Framework API, as described
7906  * in <sys/dtrace.h>.
7907  */
7908 int
7909 dtrace_meta_register(const char *name, const dtrace_mops_t *mops, void *arg,
7910     dtrace_meta_provider_id_t *idp)
7911 {
7912 	dtrace_meta_t *meta;
7913 	dtrace_helpers_t *help, *next;
7914 	int i;
7915 
7916 	*idp = DTRACE_METAPROVNONE;
7917 
7918 	/*
7919 	 * We strictly don't need the name, but we hold onto it for
7920 	 * debuggability. All hail error queues!
7921 	 */
7922 	if (name == NULL) {
7923 		cmn_err(CE_WARN, "failed to register meta-provider: "
7924 		    "invalid name");
7925 		return (EINVAL);
7926 	}
7927 
7928 	if (mops == NULL ||
7929 	    mops->dtms_create_probe == NULL ||
7930 	    mops->dtms_provide_pid == NULL ||
7931 	    mops->dtms_remove_pid == NULL) {
7932 		cmn_err(CE_WARN, "failed to register meta-register %s: "
7933 		    "invalid ops", name);
7934 		return (EINVAL);
7935 	}
7936 
7937 	meta = kmem_zalloc(sizeof (dtrace_meta_t), KM_SLEEP);
7938 	meta->dtm_mops = *mops;
7939 	meta->dtm_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
7940 	(void) strcpy(meta->dtm_name, name);
7941 	meta->dtm_arg = arg;
7942 
7943 	mutex_enter(&dtrace_meta_lock);
7944 	mutex_enter(&dtrace_lock);
7945 
7946 	if (dtrace_meta_pid != NULL) {
7947 		mutex_exit(&dtrace_lock);
7948 		mutex_exit(&dtrace_meta_lock);
7949 		cmn_err(CE_WARN, "failed to register meta-register %s: "
7950 		    "user-land meta-provider exists", name);
7951 		kmem_free(meta->dtm_name, strlen(meta->dtm_name) + 1);
7952 		kmem_free(meta, sizeof (dtrace_meta_t));
7953 		return (EINVAL);
7954 	}
7955 
7956 	dtrace_meta_pid = meta;
7957 	*idp = (dtrace_meta_provider_id_t)meta;
7958 
7959 	/*
7960 	 * If there are providers and probes ready to go, pass them
7961 	 * off to the new meta provider now.
7962 	 */
7963 
7964 	help = dtrace_deferred_pid;
7965 	dtrace_deferred_pid = NULL;
7966 
7967 	mutex_exit(&dtrace_lock);
7968 
7969 	while (help != NULL) {
7970 		for (i = 0; i < help->dthps_nprovs; i++) {
7971 			dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
7972 			    help->dthps_pid);
7973 		}
7974 
7975 		next = help->dthps_next;
7976 		help->dthps_next = NULL;
7977 		help->dthps_prev = NULL;
7978 		help->dthps_deferred = 0;
7979 		help = next;
7980 	}
7981 
7982 	mutex_exit(&dtrace_meta_lock);
7983 
7984 	return (0);
7985 }
7986 
7987 int
7988 dtrace_meta_unregister(dtrace_meta_provider_id_t id)
7989 {
7990 	dtrace_meta_t **pp, *old = (dtrace_meta_t *)id;
7991 
7992 	mutex_enter(&dtrace_meta_lock);
7993 	mutex_enter(&dtrace_lock);
7994 
7995 	if (old == dtrace_meta_pid) {
7996 		pp = &dtrace_meta_pid;
7997 	} else {
7998 		panic("attempt to unregister non-existent "
7999 		    "dtrace meta-provider %p\n", (void *)old);
8000 	}
8001 
8002 	if (old->dtm_count != 0) {
8003 		mutex_exit(&dtrace_lock);
8004 		mutex_exit(&dtrace_meta_lock);
8005 		return (EBUSY);
8006 	}
8007 
8008 	*pp = NULL;
8009 
8010 	mutex_exit(&dtrace_lock);
8011 	mutex_exit(&dtrace_meta_lock);
8012 
8013 	kmem_free(old->dtm_name, strlen(old->dtm_name) + 1);
8014 	kmem_free(old, sizeof (dtrace_meta_t));
8015 
8016 	return (0);
8017 }
8018 
8019 
8020 /*
8021  * DTrace DIF Object Functions
8022  */
8023 static int
8024 dtrace_difo_err(uint_t pc, const char *format, ...)
8025 {
8026 	if (dtrace_err_verbose) {
8027 		va_list alist;
8028 
8029 		(void) uprintf("dtrace DIF object error: [%u]: ", pc);
8030 		va_start(alist, format);
8031 		(void) vuprintf(format, alist);
8032 		va_end(alist);
8033 	}
8034 
8035 #ifdef DTRACE_ERRDEBUG
8036 	dtrace_errdebug(format);
8037 #endif
8038 	return (1);
8039 }
8040 
8041 /*
8042  * Validate a DTrace DIF object by checking the IR instructions.  The following
8043  * rules are currently enforced by dtrace_difo_validate():
8044  *
8045  * 1. Each instruction must have a valid opcode
8046  * 2. Each register, string, variable, or subroutine reference must be valid
8047  * 3. No instruction can modify register %r0 (must be zero)
8048  * 4. All instruction reserved bits must be set to zero
8049  * 5. The last instruction must be a "ret" instruction
8050  * 6. All branch targets must reference a valid instruction _after_ the branch
8051  */
8052 static int
8053 dtrace_difo_validate(dtrace_difo_t *dp, dtrace_vstate_t *vstate, uint_t nregs,
8054     cred_t *cr)
8055 {
8056 	int err = 0, i;
8057 	int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
8058 	int kcheckload;
8059 	uint_t pc;
8060 
8061 	kcheckload = cr == NULL ||
8062 	    (vstate->dtvs_state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) == 0;
8063 
8064 	dp->dtdo_destructive = 0;
8065 
8066 	for (pc = 0; pc < dp->dtdo_len && err == 0; pc++) {
8067 		dif_instr_t instr = dp->dtdo_buf[pc];
8068 
8069 		uint_t r1 = DIF_INSTR_R1(instr);
8070 		uint_t r2 = DIF_INSTR_R2(instr);
8071 		uint_t rd = DIF_INSTR_RD(instr);
8072 		uint_t rs = DIF_INSTR_RS(instr);
8073 		uint_t label = DIF_INSTR_LABEL(instr);
8074 		uint_t v = DIF_INSTR_VAR(instr);
8075 		uint_t subr = DIF_INSTR_SUBR(instr);
8076 		uint_t type = DIF_INSTR_TYPE(instr);
8077 		uint_t op = DIF_INSTR_OP(instr);
8078 
8079 		switch (op) {
8080 		case DIF_OP_OR:
8081 		case DIF_OP_XOR:
8082 		case DIF_OP_AND:
8083 		case DIF_OP_SLL:
8084 		case DIF_OP_SRL:
8085 		case DIF_OP_SRA:
8086 		case DIF_OP_SUB:
8087 		case DIF_OP_ADD:
8088 		case DIF_OP_MUL:
8089 		case DIF_OP_SDIV:
8090 		case DIF_OP_UDIV:
8091 		case DIF_OP_SREM:
8092 		case DIF_OP_UREM:
8093 		case DIF_OP_COPYS:
8094 			if (r1 >= nregs)
8095 				err += efunc(pc, "invalid register %u\n", r1);
8096 			if (r2 >= nregs)
8097 				err += efunc(pc, "invalid register %u\n", r2);
8098 			if (rd >= nregs)
8099 				err += efunc(pc, "invalid register %u\n", rd);
8100 			if (rd == 0)
8101 				err += efunc(pc, "cannot write to %r0\n");
8102 			break;
8103 		case DIF_OP_NOT:
8104 		case DIF_OP_MOV:
8105 		case DIF_OP_ALLOCS:
8106 			if (r1 >= nregs)
8107 				err += efunc(pc, "invalid register %u\n", r1);
8108 			if (r2 != 0)
8109 				err += efunc(pc, "non-zero reserved bits\n");
8110 			if (rd >= nregs)
8111 				err += efunc(pc, "invalid register %u\n", rd);
8112 			if (rd == 0)
8113 				err += efunc(pc, "cannot write to %r0\n");
8114 			break;
8115 		case DIF_OP_LDSB:
8116 		case DIF_OP_LDSH:
8117 		case DIF_OP_LDSW:
8118 		case DIF_OP_LDUB:
8119 		case DIF_OP_LDUH:
8120 		case DIF_OP_LDUW:
8121 		case DIF_OP_LDX:
8122 			if (r1 >= nregs)
8123 				err += efunc(pc, "invalid register %u\n", r1);
8124 			if (r2 != 0)
8125 				err += efunc(pc, "non-zero reserved bits\n");
8126 			if (rd >= nregs)
8127 				err += efunc(pc, "invalid register %u\n", rd);
8128 			if (rd == 0)
8129 				err += efunc(pc, "cannot write to %r0\n");
8130 			if (kcheckload)
8131 				dp->dtdo_buf[pc] = DIF_INSTR_LOAD(op +
8132 				    DIF_OP_RLDSB - DIF_OP_LDSB, r1, rd);
8133 			break;
8134 		case DIF_OP_RLDSB:
8135 		case DIF_OP_RLDSH:
8136 		case DIF_OP_RLDSW:
8137 		case DIF_OP_RLDUB:
8138 		case DIF_OP_RLDUH:
8139 		case DIF_OP_RLDUW:
8140 		case DIF_OP_RLDX:
8141 			if (r1 >= nregs)
8142 				err += efunc(pc, "invalid register %u\n", r1);
8143 			if (r2 != 0)
8144 				err += efunc(pc, "non-zero reserved bits\n");
8145 			if (rd >= nregs)
8146 				err += efunc(pc, "invalid register %u\n", rd);
8147 			if (rd == 0)
8148 				err += efunc(pc, "cannot write to %r0\n");
8149 			break;
8150 		case DIF_OP_ULDSB:
8151 		case DIF_OP_ULDSH:
8152 		case DIF_OP_ULDSW:
8153 		case DIF_OP_ULDUB:
8154 		case DIF_OP_ULDUH:
8155 		case DIF_OP_ULDUW:
8156 		case DIF_OP_ULDX:
8157 			if (r1 >= nregs)
8158 				err += efunc(pc, "invalid register %u\n", r1);
8159 			if (r2 != 0)
8160 				err += efunc(pc, "non-zero reserved bits\n");
8161 			if (rd >= nregs)
8162 				err += efunc(pc, "invalid register %u\n", rd);
8163 			if (rd == 0)
8164 				err += efunc(pc, "cannot write to %r0\n");
8165 			break;
8166 		case DIF_OP_STB:
8167 		case DIF_OP_STH:
8168 		case DIF_OP_STW:
8169 		case DIF_OP_STX:
8170 			if (r1 >= nregs)
8171 				err += efunc(pc, "invalid register %u\n", r1);
8172 			if (r2 != 0)
8173 				err += efunc(pc, "non-zero reserved bits\n");
8174 			if (rd >= nregs)
8175 				err += efunc(pc, "invalid register %u\n", rd);
8176 			if (rd == 0)
8177 				err += efunc(pc, "cannot write to 0 address\n");
8178 			break;
8179 		case DIF_OP_CMP:
8180 		case DIF_OP_SCMP:
8181 			if (r1 >= nregs)
8182 				err += efunc(pc, "invalid register %u\n", r1);
8183 			if (r2 >= nregs)
8184 				err += efunc(pc, "invalid register %u\n", r2);
8185 			if (rd != 0)
8186 				err += efunc(pc, "non-zero reserved bits\n");
8187 			break;
8188 		case DIF_OP_TST:
8189 			if (r1 >= nregs)
8190 				err += efunc(pc, "invalid register %u\n", r1);
8191 			if (r2 != 0 || rd != 0)
8192 				err += efunc(pc, "non-zero reserved bits\n");
8193 			break;
8194 		case DIF_OP_BA:
8195 		case DIF_OP_BE:
8196 		case DIF_OP_BNE:
8197 		case DIF_OP_BG:
8198 		case DIF_OP_BGU:
8199 		case DIF_OP_BGE:
8200 		case DIF_OP_BGEU:
8201 		case DIF_OP_BL:
8202 		case DIF_OP_BLU:
8203 		case DIF_OP_BLE:
8204 		case DIF_OP_BLEU:
8205 			if (label >= dp->dtdo_len) {
8206 				err += efunc(pc, "invalid branch target %u\n",
8207 				    label);
8208 			}
8209 			if (label <= pc) {
8210 				err += efunc(pc, "backward branch to %u\n",
8211 				    label);
8212 			}
8213 			break;
8214 		case DIF_OP_RET:
8215 			if (r1 != 0 || r2 != 0)
8216 				err += efunc(pc, "non-zero reserved bits\n");
8217 			if (rd >= nregs)
8218 				err += efunc(pc, "invalid register %u\n", rd);
8219 			break;
8220 		case DIF_OP_NOP:
8221 		case DIF_OP_POPTS:
8222 		case DIF_OP_FLUSHTS:
8223 			if (r1 != 0 || r2 != 0 || rd != 0)
8224 				err += efunc(pc, "non-zero reserved bits\n");
8225 			break;
8226 		case DIF_OP_SETX:
8227 			if (DIF_INSTR_INTEGER(instr) >= dp->dtdo_intlen) {
8228 				err += efunc(pc, "invalid integer ref %u\n",
8229 				    DIF_INSTR_INTEGER(instr));
8230 			}
8231 			if (rd >= nregs)
8232 				err += efunc(pc, "invalid register %u\n", rd);
8233 			if (rd == 0)
8234 				err += efunc(pc, "cannot write to %r0\n");
8235 			break;
8236 		case DIF_OP_SETS:
8237 			if (DIF_INSTR_STRING(instr) >= dp->dtdo_strlen) {
8238 				err += efunc(pc, "invalid string ref %u\n",
8239 				    DIF_INSTR_STRING(instr));
8240 			}
8241 			if (rd >= nregs)
8242 				err += efunc(pc, "invalid register %u\n", rd);
8243 			if (rd == 0)
8244 				err += efunc(pc, "cannot write to %r0\n");
8245 			break;
8246 		case DIF_OP_LDGA:
8247 		case DIF_OP_LDTA:
8248 			if (r1 > DIF_VAR_ARRAY_MAX)
8249 				err += efunc(pc, "invalid array %u\n", r1);
8250 			if (r2 >= nregs)
8251 				err += efunc(pc, "invalid register %u\n", r2);
8252 			if (rd >= nregs)
8253 				err += efunc(pc, "invalid register %u\n", rd);
8254 			if (rd == 0)
8255 				err += efunc(pc, "cannot write to %r0\n");
8256 			break;
8257 		case DIF_OP_LDGS:
8258 		case DIF_OP_LDTS:
8259 		case DIF_OP_LDLS:
8260 		case DIF_OP_LDGAA:
8261 		case DIF_OP_LDTAA:
8262 			if (v < DIF_VAR_OTHER_MIN || v > DIF_VAR_OTHER_MAX)
8263 				err += efunc(pc, "invalid variable %u\n", v);
8264 			if (rd >= nregs)
8265 				err += efunc(pc, "invalid register %u\n", rd);
8266 			if (rd == 0)
8267 				err += efunc(pc, "cannot write to %r0\n");
8268 			break;
8269 		case DIF_OP_STGS:
8270 		case DIF_OP_STTS:
8271 		case DIF_OP_STLS:
8272 		case DIF_OP_STGAA:
8273 		case DIF_OP_STTAA:
8274 			if (v < DIF_VAR_OTHER_UBASE || v > DIF_VAR_OTHER_MAX)
8275 				err += efunc(pc, "invalid variable %u\n", v);
8276 			if (rs >= nregs)
8277 				err += efunc(pc, "invalid register %u\n", rd);
8278 			break;
8279 		case DIF_OP_CALL:
8280 			if (subr > DIF_SUBR_MAX)
8281 				err += efunc(pc, "invalid subr %u\n", subr);
8282 			if (rd >= nregs)
8283 				err += efunc(pc, "invalid register %u\n", rd);
8284 			if (rd == 0)
8285 				err += efunc(pc, "cannot write to %r0\n");
8286 
8287 			if (subr == DIF_SUBR_COPYOUT ||
8288 			    subr == DIF_SUBR_COPYOUTSTR) {
8289 				dp->dtdo_destructive = 1;
8290 			}
8291 			break;
8292 		case DIF_OP_PUSHTR:
8293 			if (type != DIF_TYPE_STRING && type != DIF_TYPE_CTF)
8294 				err += efunc(pc, "invalid ref type %u\n", type);
8295 			if (r2 >= nregs)
8296 				err += efunc(pc, "invalid register %u\n", r2);
8297 			if (rs >= nregs)
8298 				err += efunc(pc, "invalid register %u\n", rs);
8299 			break;
8300 		case DIF_OP_PUSHTV:
8301 			if (type != DIF_TYPE_CTF)
8302 				err += efunc(pc, "invalid val type %u\n", type);
8303 			if (r2 >= nregs)
8304 				err += efunc(pc, "invalid register %u\n", r2);
8305 			if (rs >= nregs)
8306 				err += efunc(pc, "invalid register %u\n", rs);
8307 			break;
8308 		default:
8309 			err += efunc(pc, "invalid opcode %u\n",
8310 			    DIF_INSTR_OP(instr));
8311 		}
8312 	}
8313 
8314 	if (dp->dtdo_len != 0 &&
8315 	    DIF_INSTR_OP(dp->dtdo_buf[dp->dtdo_len - 1]) != DIF_OP_RET) {
8316 		err += efunc(dp->dtdo_len - 1,
8317 		    "expected 'ret' as last DIF instruction\n");
8318 	}
8319 
8320 	if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF)) {
8321 		/*
8322 		 * If we're not returning by reference, the size must be either
8323 		 * 0 or the size of one of the base types.
8324 		 */
8325 		switch (dp->dtdo_rtype.dtdt_size) {
8326 		case 0:
8327 		case sizeof (uint8_t):
8328 		case sizeof (uint16_t):
8329 		case sizeof (uint32_t):
8330 		case sizeof (uint64_t):
8331 			break;
8332 
8333 		default:
8334 			err += efunc(dp->dtdo_len - 1, "bad return size\n");
8335 		}
8336 	}
8337 
8338 	for (i = 0; i < dp->dtdo_varlen && err == 0; i++) {
8339 		dtrace_difv_t *v = &dp->dtdo_vartab[i], *existing = NULL;
8340 		dtrace_diftype_t *vt, *et;
8341 		uint_t id, ndx;
8342 
8343 		if (v->dtdv_scope != DIFV_SCOPE_GLOBAL &&
8344 		    v->dtdv_scope != DIFV_SCOPE_THREAD &&
8345 		    v->dtdv_scope != DIFV_SCOPE_LOCAL) {
8346 			err += efunc(i, "unrecognized variable scope %d\n",
8347 			    v->dtdv_scope);
8348 			break;
8349 		}
8350 
8351 		if (v->dtdv_kind != DIFV_KIND_ARRAY &&
8352 		    v->dtdv_kind != DIFV_KIND_SCALAR) {
8353 			err += efunc(i, "unrecognized variable type %d\n",
8354 			    v->dtdv_kind);
8355 			break;
8356 		}
8357 
8358 		if ((id = v->dtdv_id) > DIF_VARIABLE_MAX) {
8359 			err += efunc(i, "%d exceeds variable id limit\n", id);
8360 			break;
8361 		}
8362 
8363 		if (id < DIF_VAR_OTHER_UBASE)
8364 			continue;
8365 
8366 		/*
8367 		 * For user-defined variables, we need to check that this
8368 		 * definition is identical to any previous definition that we
8369 		 * encountered.
8370 		 */
8371 		ndx = id - DIF_VAR_OTHER_UBASE;
8372 
8373 		switch (v->dtdv_scope) {
8374 		case DIFV_SCOPE_GLOBAL:
8375 			if (ndx < vstate->dtvs_nglobals) {
8376 				dtrace_statvar_t *svar;
8377 
8378 				if ((svar = vstate->dtvs_globals[ndx]) != NULL)
8379 					existing = &svar->dtsv_var;
8380 			}
8381 
8382 			break;
8383 
8384 		case DIFV_SCOPE_THREAD:
8385 			if (ndx < vstate->dtvs_ntlocals)
8386 				existing = &vstate->dtvs_tlocals[ndx];
8387 			break;
8388 
8389 		case DIFV_SCOPE_LOCAL:
8390 			if (ndx < vstate->dtvs_nlocals) {
8391 				dtrace_statvar_t *svar;
8392 
8393 				if ((svar = vstate->dtvs_locals[ndx]) != NULL)
8394 					existing = &svar->dtsv_var;
8395 			}
8396 
8397 			break;
8398 		}
8399 
8400 		vt = &v->dtdv_type;
8401 
8402 		if (vt->dtdt_flags & DIF_TF_BYREF) {
8403 			if (vt->dtdt_size == 0) {
8404 				err += efunc(i, "zero-sized variable\n");
8405 				break;
8406 			}
8407 
8408 			if (v->dtdv_scope == DIFV_SCOPE_GLOBAL &&
8409 			    vt->dtdt_size > dtrace_global_maxsize) {
8410 				err += efunc(i, "oversized by-ref global\n");
8411 				break;
8412 			}
8413 		}
8414 
8415 		if (existing == NULL || existing->dtdv_id == 0)
8416 			continue;
8417 
8418 		ASSERT(existing->dtdv_id == v->dtdv_id);
8419 		ASSERT(existing->dtdv_scope == v->dtdv_scope);
8420 
8421 		if (existing->dtdv_kind != v->dtdv_kind)
8422 			err += efunc(i, "%d changed variable kind\n", id);
8423 
8424 		et = &existing->dtdv_type;
8425 
8426 		if (vt->dtdt_flags != et->dtdt_flags) {
8427 			err += efunc(i, "%d changed variable type flags\n", id);
8428 			break;
8429 		}
8430 
8431 		if (vt->dtdt_size != 0 && vt->dtdt_size != et->dtdt_size) {
8432 			err += efunc(i, "%d changed variable type size\n", id);
8433 			break;
8434 		}
8435 	}
8436 
8437 	return (err);
8438 }
8439 
8440 /*
8441  * Validate a DTrace DIF object that it is to be used as a helper.  Helpers
8442  * are much more constrained than normal DIFOs.  Specifically, they may
8443  * not:
8444  *
8445  * 1. Make calls to subroutines other than copyin(), copyinstr() or
8446  *    miscellaneous string routines
8447  * 2. Access DTrace variables other than the args[] array, and the
8448  *    curthread, pid, ppid, tid, execname, zonename, uid and gid variables.
8449  * 3. Have thread-local variables.
8450  * 4. Have dynamic variables.
8451  */
8452 static int
8453 dtrace_difo_validate_helper(dtrace_difo_t *dp)
8454 {
8455 	int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
8456 	int err = 0;
8457 	uint_t pc;
8458 
8459 	for (pc = 0; pc < dp->dtdo_len; pc++) {
8460 		dif_instr_t instr = dp->dtdo_buf[pc];
8461 
8462 		uint_t v = DIF_INSTR_VAR(instr);
8463 		uint_t subr = DIF_INSTR_SUBR(instr);
8464 		uint_t op = DIF_INSTR_OP(instr);
8465 
8466 		switch (op) {
8467 		case DIF_OP_OR:
8468 		case DIF_OP_XOR:
8469 		case DIF_OP_AND:
8470 		case DIF_OP_SLL:
8471 		case DIF_OP_SRL:
8472 		case DIF_OP_SRA:
8473 		case DIF_OP_SUB:
8474 		case DIF_OP_ADD:
8475 		case DIF_OP_MUL:
8476 		case DIF_OP_SDIV:
8477 		case DIF_OP_UDIV:
8478 		case DIF_OP_SREM:
8479 		case DIF_OP_UREM:
8480 		case DIF_OP_COPYS:
8481 		case DIF_OP_NOT:
8482 		case DIF_OP_MOV:
8483 		case DIF_OP_RLDSB:
8484 		case DIF_OP_RLDSH:
8485 		case DIF_OP_RLDSW:
8486 		case DIF_OP_RLDUB:
8487 		case DIF_OP_RLDUH:
8488 		case DIF_OP_RLDUW:
8489 		case DIF_OP_RLDX:
8490 		case DIF_OP_ULDSB:
8491 		case DIF_OP_ULDSH:
8492 		case DIF_OP_ULDSW:
8493 		case DIF_OP_ULDUB:
8494 		case DIF_OP_ULDUH:
8495 		case DIF_OP_ULDUW:
8496 		case DIF_OP_ULDX:
8497 		case DIF_OP_STB:
8498 		case DIF_OP_STH:
8499 		case DIF_OP_STW:
8500 		case DIF_OP_STX:
8501 		case DIF_OP_ALLOCS:
8502 		case DIF_OP_CMP:
8503 		case DIF_OP_SCMP:
8504 		case DIF_OP_TST:
8505 		case DIF_OP_BA:
8506 		case DIF_OP_BE:
8507 		case DIF_OP_BNE:
8508 		case DIF_OP_BG:
8509 		case DIF_OP_BGU:
8510 		case DIF_OP_BGE:
8511 		case DIF_OP_BGEU:
8512 		case DIF_OP_BL:
8513 		case DIF_OP_BLU:
8514 		case DIF_OP_BLE:
8515 		case DIF_OP_BLEU:
8516 		case DIF_OP_RET:
8517 		case DIF_OP_NOP:
8518 		case DIF_OP_POPTS:
8519 		case DIF_OP_FLUSHTS:
8520 		case DIF_OP_SETX:
8521 		case DIF_OP_SETS:
8522 		case DIF_OP_LDGA:
8523 		case DIF_OP_LDLS:
8524 		case DIF_OP_STGS:
8525 		case DIF_OP_STLS:
8526 		case DIF_OP_PUSHTR:
8527 		case DIF_OP_PUSHTV:
8528 			break;
8529 
8530 		case DIF_OP_LDGS:
8531 			if (v >= DIF_VAR_OTHER_UBASE)
8532 				break;
8533 
8534 			if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9)
8535 				break;
8536 
8537 			if (v == DIF_VAR_CURTHREAD || v == DIF_VAR_PID ||
8538 			    v == DIF_VAR_PPID || v == DIF_VAR_TID ||
8539 			    v == DIF_VAR_EXECNAME || v == DIF_VAR_ZONENAME ||
8540 			    v == DIF_VAR_UID || v == DIF_VAR_GID)
8541 				break;
8542 
8543 			err += efunc(pc, "illegal variable %u\n", v);
8544 			break;
8545 
8546 		case DIF_OP_LDTA:
8547 		case DIF_OP_LDTS:
8548 		case DIF_OP_LDGAA:
8549 		case DIF_OP_LDTAA:
8550 			err += efunc(pc, "illegal dynamic variable load\n");
8551 			break;
8552 
8553 		case DIF_OP_STTS:
8554 		case DIF_OP_STGAA:
8555 		case DIF_OP_STTAA:
8556 			err += efunc(pc, "illegal dynamic variable store\n");
8557 			break;
8558 
8559 		case DIF_OP_CALL:
8560 			if (subr == DIF_SUBR_ALLOCA ||
8561 			    subr == DIF_SUBR_BCOPY ||
8562 			    subr == DIF_SUBR_COPYIN ||
8563 			    subr == DIF_SUBR_COPYINTO ||
8564 			    subr == DIF_SUBR_COPYINSTR ||
8565 			    subr == DIF_SUBR_INDEX ||
8566 			    subr == DIF_SUBR_INET_NTOA ||
8567 			    subr == DIF_SUBR_INET_NTOA6 ||
8568 			    subr == DIF_SUBR_INET_NTOP ||
8569 			    subr == DIF_SUBR_LLTOSTR ||
8570 			    subr == DIF_SUBR_RINDEX ||
8571 			    subr == DIF_SUBR_STRCHR ||
8572 			    subr == DIF_SUBR_STRJOIN ||
8573 			    subr == DIF_SUBR_STRRCHR ||
8574 			    subr == DIF_SUBR_STRSTR ||
8575 			    subr == DIF_SUBR_HTONS ||
8576 			    subr == DIF_SUBR_HTONL ||
8577 			    subr == DIF_SUBR_HTONLL ||
8578 			    subr == DIF_SUBR_NTOHS ||
8579 			    subr == DIF_SUBR_NTOHL ||
8580 			    subr == DIF_SUBR_NTOHLL)
8581 				break;
8582 
8583 			err += efunc(pc, "invalid subr %u\n", subr);
8584 			break;
8585 
8586 		default:
8587 			err += efunc(pc, "invalid opcode %u\n",
8588 			    DIF_INSTR_OP(instr));
8589 		}
8590 	}
8591 
8592 	return (err);
8593 }
8594 
8595 /*
8596  * Returns 1 if the expression in the DIF object can be cached on a per-thread
8597  * basis; 0 if not.
8598  */
8599 static int
8600 dtrace_difo_cacheable(dtrace_difo_t *dp)
8601 {
8602 	int i;
8603 
8604 	if (dp == NULL)
8605 		return (0);
8606 
8607 	for (i = 0; i < dp->dtdo_varlen; i++) {
8608 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
8609 
8610 		if (v->dtdv_scope != DIFV_SCOPE_GLOBAL)
8611 			continue;
8612 
8613 		switch (v->dtdv_id) {
8614 		case DIF_VAR_CURTHREAD:
8615 		case DIF_VAR_PID:
8616 		case DIF_VAR_TID:
8617 		case DIF_VAR_EXECNAME:
8618 		case DIF_VAR_ZONENAME:
8619 			break;
8620 
8621 		default:
8622 			return (0);
8623 		}
8624 	}
8625 
8626 	/*
8627 	 * This DIF object may be cacheable.  Now we need to look for any
8628 	 * array loading instructions, any memory loading instructions, or
8629 	 * any stores to thread-local variables.
8630 	 */
8631 	for (i = 0; i < dp->dtdo_len; i++) {
8632 		uint_t op = DIF_INSTR_OP(dp->dtdo_buf[i]);
8633 
8634 		if ((op >= DIF_OP_LDSB && op <= DIF_OP_LDX) ||
8635 		    (op >= DIF_OP_ULDSB && op <= DIF_OP_ULDX) ||
8636 		    (op >= DIF_OP_RLDSB && op <= DIF_OP_RLDX) ||
8637 		    op == DIF_OP_LDGA || op == DIF_OP_STTS)
8638 			return (0);
8639 	}
8640 
8641 	return (1);
8642 }
8643 
8644 static void
8645 dtrace_difo_hold(dtrace_difo_t *dp)
8646 {
8647 	int i;
8648 
8649 	ASSERT(MUTEX_HELD(&dtrace_lock));
8650 
8651 	dp->dtdo_refcnt++;
8652 	ASSERT(dp->dtdo_refcnt != 0);
8653 
8654 	/*
8655 	 * We need to check this DIF object for references to the variable
8656 	 * DIF_VAR_VTIMESTAMP.
8657 	 */
8658 	for (i = 0; i < dp->dtdo_varlen; i++) {
8659 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
8660 
8661 		if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
8662 			continue;
8663 
8664 		if (dtrace_vtime_references++ == 0)
8665 			dtrace_vtime_enable();
8666 	}
8667 }
8668 
8669 /*
8670  * This routine calculates the dynamic variable chunksize for a given DIF
8671  * object.  The calculation is not fool-proof, and can probably be tricked by
8672  * malicious DIF -- but it works for all compiler-generated DIF.  Because this
8673  * calculation is likely imperfect, dtrace_dynvar() is able to gracefully fail
8674  * if a dynamic variable size exceeds the chunksize.
8675  */
8676 static void
8677 dtrace_difo_chunksize(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
8678 {
8679 	uint64_t sval;
8680 	dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
8681 	const dif_instr_t *text = dp->dtdo_buf;
8682 	uint_t pc, srd = 0;
8683 	uint_t ttop = 0;
8684 	size_t size, ksize;
8685 	uint_t id, i;
8686 
8687 	for (pc = 0; pc < dp->dtdo_len; pc++) {
8688 		dif_instr_t instr = text[pc];
8689 		uint_t op = DIF_INSTR_OP(instr);
8690 		uint_t rd = DIF_INSTR_RD(instr);
8691 		uint_t r1 = DIF_INSTR_R1(instr);
8692 		uint_t nkeys = 0;
8693 		uchar_t scope;
8694 
8695 		dtrace_key_t *key = tupregs;
8696 
8697 		switch (op) {
8698 		case DIF_OP_SETX:
8699 			sval = dp->dtdo_inttab[DIF_INSTR_INTEGER(instr)];
8700 			srd = rd;
8701 			continue;
8702 
8703 		case DIF_OP_STTS:
8704 			key = &tupregs[DIF_DTR_NREGS];
8705 			key[0].dttk_size = 0;
8706 			key[1].dttk_size = 0;
8707 			nkeys = 2;
8708 			scope = DIFV_SCOPE_THREAD;
8709 			break;
8710 
8711 		case DIF_OP_STGAA:
8712 		case DIF_OP_STTAA:
8713 			nkeys = ttop;
8714 
8715 			if (DIF_INSTR_OP(instr) == DIF_OP_STTAA)
8716 				key[nkeys++].dttk_size = 0;
8717 
8718 			key[nkeys++].dttk_size = 0;
8719 
8720 			if (op == DIF_OP_STTAA) {
8721 				scope = DIFV_SCOPE_THREAD;
8722 			} else {
8723 				scope = DIFV_SCOPE_GLOBAL;
8724 			}
8725 
8726 			break;
8727 
8728 		case DIF_OP_PUSHTR:
8729 			if (ttop == DIF_DTR_NREGS)
8730 				return;
8731 
8732 			if ((srd == 0 || sval == 0) && r1 == DIF_TYPE_STRING) {
8733 				/*
8734 				 * If the register for the size of the "pushtr"
8735 				 * is %r0 (or the value is 0) and the type is
8736 				 * a string, we'll use the system-wide default
8737 				 * string size.
8738 				 */
8739 				tupregs[ttop++].dttk_size =
8740 				    dtrace_strsize_default;
8741 			} else {
8742 				if (srd == 0)
8743 					return;
8744 
8745 				tupregs[ttop++].dttk_size = sval;
8746 			}
8747 
8748 			break;
8749 
8750 		case DIF_OP_PUSHTV:
8751 			if (ttop == DIF_DTR_NREGS)
8752 				return;
8753 
8754 			tupregs[ttop++].dttk_size = 0;
8755 			break;
8756 
8757 		case DIF_OP_FLUSHTS:
8758 			ttop = 0;
8759 			break;
8760 
8761 		case DIF_OP_POPTS:
8762 			if (ttop != 0)
8763 				ttop--;
8764 			break;
8765 		}
8766 
8767 		sval = 0;
8768 		srd = 0;
8769 
8770 		if (nkeys == 0)
8771 			continue;
8772 
8773 		/*
8774 		 * We have a dynamic variable allocation; calculate its size.
8775 		 */
8776 		for (ksize = 0, i = 0; i < nkeys; i++)
8777 			ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
8778 
8779 		size = sizeof (dtrace_dynvar_t);
8780 		size += sizeof (dtrace_key_t) * (nkeys - 1);
8781 		size += ksize;
8782 
8783 		/*
8784 		 * Now we need to determine the size of the stored data.
8785 		 */
8786 		id = DIF_INSTR_VAR(instr);
8787 
8788 		for (i = 0; i < dp->dtdo_varlen; i++) {
8789 			dtrace_difv_t *v = &dp->dtdo_vartab[i];
8790 
8791 			if (v->dtdv_id == id && v->dtdv_scope == scope) {
8792 				size += v->dtdv_type.dtdt_size;
8793 				break;
8794 			}
8795 		}
8796 
8797 		if (i == dp->dtdo_varlen)
8798 			return;
8799 
8800 		/*
8801 		 * We have the size.  If this is larger than the chunk size
8802 		 * for our dynamic variable state, reset the chunk size.
8803 		 */
8804 		size = P2ROUNDUP(size, sizeof (uint64_t));
8805 
8806 		if (size > vstate->dtvs_dynvars.dtds_chunksize)
8807 			vstate->dtvs_dynvars.dtds_chunksize = size;
8808 	}
8809 }
8810 
8811 static void
8812 dtrace_difo_init(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
8813 {
8814 	int i, oldsvars, osz, nsz, otlocals, ntlocals;
8815 	uint_t id;
8816 
8817 	ASSERT(MUTEX_HELD(&dtrace_lock));
8818 	ASSERT(dp->dtdo_buf != NULL && dp->dtdo_len != 0);
8819 
8820 	for (i = 0; i < dp->dtdo_varlen; i++) {
8821 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
8822 		dtrace_statvar_t *svar, ***svarp;
8823 		size_t dsize = 0;
8824 		uint8_t scope = v->dtdv_scope;
8825 		int *np;
8826 
8827 		if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
8828 			continue;
8829 
8830 		id -= DIF_VAR_OTHER_UBASE;
8831 
8832 		switch (scope) {
8833 		case DIFV_SCOPE_THREAD:
8834 			while (id >= (otlocals = vstate->dtvs_ntlocals)) {
8835 				dtrace_difv_t *tlocals;
8836 
8837 				if ((ntlocals = (otlocals << 1)) == 0)
8838 					ntlocals = 1;
8839 
8840 				osz = otlocals * sizeof (dtrace_difv_t);
8841 				nsz = ntlocals * sizeof (dtrace_difv_t);
8842 
8843 				tlocals = kmem_zalloc(nsz, KM_SLEEP);
8844 
8845 				if (osz != 0) {
8846 					bcopy(vstate->dtvs_tlocals,
8847 					    tlocals, osz);
8848 					kmem_free(vstate->dtvs_tlocals, osz);
8849 				}
8850 
8851 				vstate->dtvs_tlocals = tlocals;
8852 				vstate->dtvs_ntlocals = ntlocals;
8853 			}
8854 
8855 			vstate->dtvs_tlocals[id] = *v;
8856 			continue;
8857 
8858 		case DIFV_SCOPE_LOCAL:
8859 			np = &vstate->dtvs_nlocals;
8860 			svarp = &vstate->dtvs_locals;
8861 
8862 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
8863 				dsize = NCPU * (v->dtdv_type.dtdt_size +
8864 				    sizeof (uint64_t));
8865 			else
8866 				dsize = NCPU * sizeof (uint64_t);
8867 
8868 			break;
8869 
8870 		case DIFV_SCOPE_GLOBAL:
8871 			np = &vstate->dtvs_nglobals;
8872 			svarp = &vstate->dtvs_globals;
8873 
8874 			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
8875 				dsize = v->dtdv_type.dtdt_size +
8876 				    sizeof (uint64_t);
8877 
8878 			break;
8879 
8880 		default:
8881 			ASSERT(0);
8882 		}
8883 
8884 		while (id >= (oldsvars = *np)) {
8885 			dtrace_statvar_t **statics;
8886 			int newsvars, oldsize, newsize;
8887 
8888 			if ((newsvars = (oldsvars << 1)) == 0)
8889 				newsvars = 1;
8890 
8891 			oldsize = oldsvars * sizeof (dtrace_statvar_t *);
8892 			newsize = newsvars * sizeof (dtrace_statvar_t *);
8893 
8894 			statics = kmem_zalloc(newsize, KM_SLEEP);
8895 
8896 			if (oldsize != 0) {
8897 				bcopy(*svarp, statics, oldsize);
8898 				kmem_free(*svarp, oldsize);
8899 			}
8900 
8901 			*svarp = statics;
8902 			*np = newsvars;
8903 		}
8904 
8905 		if ((svar = (*svarp)[id]) == NULL) {
8906 			svar = kmem_zalloc(sizeof (dtrace_statvar_t), KM_SLEEP);
8907 			svar->dtsv_var = *v;
8908 
8909 			if ((svar->dtsv_size = dsize) != 0) {
8910 				svar->dtsv_data = (uint64_t)(uintptr_t)
8911 				    kmem_zalloc(dsize, KM_SLEEP);
8912 			}
8913 
8914 			(*svarp)[id] = svar;
8915 		}
8916 
8917 		svar->dtsv_refcnt++;
8918 	}
8919 
8920 	dtrace_difo_chunksize(dp, vstate);
8921 	dtrace_difo_hold(dp);
8922 }
8923 
8924 static dtrace_difo_t *
8925 dtrace_difo_duplicate(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
8926 {
8927 	dtrace_difo_t *new;
8928 	size_t sz;
8929 
8930 	ASSERT(dp->dtdo_buf != NULL);
8931 	ASSERT(dp->dtdo_refcnt != 0);
8932 
8933 	new = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
8934 
8935 	ASSERT(dp->dtdo_buf != NULL);
8936 	sz = dp->dtdo_len * sizeof (dif_instr_t);
8937 	new->dtdo_buf = kmem_alloc(sz, KM_SLEEP);
8938 	bcopy(dp->dtdo_buf, new->dtdo_buf, sz);
8939 	new->dtdo_len = dp->dtdo_len;
8940 
8941 	if (dp->dtdo_strtab != NULL) {
8942 		ASSERT(dp->dtdo_strlen != 0);
8943 		new->dtdo_strtab = kmem_alloc(dp->dtdo_strlen, KM_SLEEP);
8944 		bcopy(dp->dtdo_strtab, new->dtdo_strtab, dp->dtdo_strlen);
8945 		new->dtdo_strlen = dp->dtdo_strlen;
8946 	}
8947 
8948 	if (dp->dtdo_inttab != NULL) {
8949 		ASSERT(dp->dtdo_intlen != 0);
8950 		sz = dp->dtdo_intlen * sizeof (uint64_t);
8951 		new->dtdo_inttab = kmem_alloc(sz, KM_SLEEP);
8952 		bcopy(dp->dtdo_inttab, new->dtdo_inttab, sz);
8953 		new->dtdo_intlen = dp->dtdo_intlen;
8954 	}
8955 
8956 	if (dp->dtdo_vartab != NULL) {
8957 		ASSERT(dp->dtdo_varlen != 0);
8958 		sz = dp->dtdo_varlen * sizeof (dtrace_difv_t);
8959 		new->dtdo_vartab = kmem_alloc(sz, KM_SLEEP);
8960 		bcopy(dp->dtdo_vartab, new->dtdo_vartab, sz);
8961 		new->dtdo_varlen = dp->dtdo_varlen;
8962 	}
8963 
8964 	dtrace_difo_init(new, vstate);
8965 	return (new);
8966 }
8967 
8968 static void
8969 dtrace_difo_destroy(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
8970 {
8971 	int i;
8972 
8973 	ASSERT(dp->dtdo_refcnt == 0);
8974 
8975 	for (i = 0; i < dp->dtdo_varlen; i++) {
8976 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
8977 		dtrace_statvar_t *svar, **svarp;
8978 		uint_t id;
8979 		uint8_t scope = v->dtdv_scope;
8980 		int *np;
8981 
8982 		switch (scope) {
8983 		case DIFV_SCOPE_THREAD:
8984 			continue;
8985 
8986 		case DIFV_SCOPE_LOCAL:
8987 			np = &vstate->dtvs_nlocals;
8988 			svarp = vstate->dtvs_locals;
8989 			break;
8990 
8991 		case DIFV_SCOPE_GLOBAL:
8992 			np = &vstate->dtvs_nglobals;
8993 			svarp = vstate->dtvs_globals;
8994 			break;
8995 
8996 		default:
8997 			ASSERT(0);
8998 		}
8999 
9000 		if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
9001 			continue;
9002 
9003 		id -= DIF_VAR_OTHER_UBASE;
9004 		ASSERT(id < *np);
9005 
9006 		svar = svarp[id];
9007 		ASSERT(svar != NULL);
9008 		ASSERT(svar->dtsv_refcnt > 0);
9009 
9010 		if (--svar->dtsv_refcnt > 0)
9011 			continue;
9012 
9013 		if (svar->dtsv_size != 0) {
9014 			ASSERT(svar->dtsv_data != NULL);
9015 			kmem_free((void *)(uintptr_t)svar->dtsv_data,
9016 			    svar->dtsv_size);
9017 		}
9018 
9019 		kmem_free(svar, sizeof (dtrace_statvar_t));
9020 		svarp[id] = NULL;
9021 	}
9022 
9023 	kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
9024 	kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
9025 	kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
9026 	kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
9027 
9028 	kmem_free(dp, sizeof (dtrace_difo_t));
9029 }
9030 
9031 static void
9032 dtrace_difo_release(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9033 {
9034 	int i;
9035 
9036 	ASSERT(MUTEX_HELD(&dtrace_lock));
9037 	ASSERT(dp->dtdo_refcnt != 0);
9038 
9039 	for (i = 0; i < dp->dtdo_varlen; i++) {
9040 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
9041 
9042 		if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
9043 			continue;
9044 
9045 		ASSERT(dtrace_vtime_references > 0);
9046 		if (--dtrace_vtime_references == 0)
9047 			dtrace_vtime_disable();
9048 	}
9049 
9050 	if (--dp->dtdo_refcnt == 0)
9051 		dtrace_difo_destroy(dp, vstate);
9052 }
9053 
9054 /*
9055  * DTrace Format Functions
9056  */
9057 static uint16_t
9058 dtrace_format_add(dtrace_state_t *state, char *str)
9059 {
9060 	char *fmt, **new;
9061 	uint16_t ndx, len = strlen(str) + 1;
9062 
9063 	fmt = kmem_zalloc(len, KM_SLEEP);
9064 	bcopy(str, fmt, len);
9065 
9066 	for (ndx = 0; ndx < state->dts_nformats; ndx++) {
9067 		if (state->dts_formats[ndx] == NULL) {
9068 			state->dts_formats[ndx] = fmt;
9069 			return (ndx + 1);
9070 		}
9071 	}
9072 
9073 	if (state->dts_nformats == USHRT_MAX) {
9074 		/*
9075 		 * This is only likely if a denial-of-service attack is being
9076 		 * attempted.  As such, it's okay to fail silently here.
9077 		 */
9078 		kmem_free(fmt, len);
9079 		return (0);
9080 	}
9081 
9082 	/*
9083 	 * For simplicity, we always resize the formats array to be exactly the
9084 	 * number of formats.
9085 	 */
9086 	ndx = state->dts_nformats++;
9087 	new = kmem_alloc((ndx + 1) * sizeof (char *), KM_SLEEP);
9088 
9089 	if (state->dts_formats != NULL) {
9090 		ASSERT(ndx != 0);
9091 		bcopy(state->dts_formats, new, ndx * sizeof (char *));
9092 		kmem_free(state->dts_formats, ndx * sizeof (char *));
9093 	}
9094 
9095 	state->dts_formats = new;
9096 	state->dts_formats[ndx] = fmt;
9097 
9098 	return (ndx + 1);
9099 }
9100 
9101 static void
9102 dtrace_format_remove(dtrace_state_t *state, uint16_t format)
9103 {
9104 	char *fmt;
9105 
9106 	ASSERT(state->dts_formats != NULL);
9107 	ASSERT(format <= state->dts_nformats);
9108 	ASSERT(state->dts_formats[format - 1] != NULL);
9109 
9110 	fmt = state->dts_formats[format - 1];
9111 	kmem_free(fmt, strlen(fmt) + 1);
9112 	state->dts_formats[format - 1] = NULL;
9113 }
9114 
9115 static void
9116 dtrace_format_destroy(dtrace_state_t *state)
9117 {
9118 	int i;
9119 
9120 	if (state->dts_nformats == 0) {
9121 		ASSERT(state->dts_formats == NULL);
9122 		return;
9123 	}
9124 
9125 	ASSERT(state->dts_formats != NULL);
9126 
9127 	for (i = 0; i < state->dts_nformats; i++) {
9128 		char *fmt = state->dts_formats[i];
9129 
9130 		if (fmt == NULL)
9131 			continue;
9132 
9133 		kmem_free(fmt, strlen(fmt) + 1);
9134 	}
9135 
9136 	kmem_free(state->dts_formats, state->dts_nformats * sizeof (char *));
9137 	state->dts_nformats = 0;
9138 	state->dts_formats = NULL;
9139 }
9140 
9141 /*
9142  * DTrace Predicate Functions
9143  */
9144 static dtrace_predicate_t *
9145 dtrace_predicate_create(dtrace_difo_t *dp)
9146 {
9147 	dtrace_predicate_t *pred;
9148 
9149 	ASSERT(MUTEX_HELD(&dtrace_lock));
9150 	ASSERT(dp->dtdo_refcnt != 0);
9151 
9152 	pred = kmem_zalloc(sizeof (dtrace_predicate_t), KM_SLEEP);
9153 	pred->dtp_difo = dp;
9154 	pred->dtp_refcnt = 1;
9155 
9156 	if (!dtrace_difo_cacheable(dp))
9157 		return (pred);
9158 
9159 	if (dtrace_predcache_id == DTRACE_CACHEIDNONE) {
9160 		/*
9161 		 * This is only theoretically possible -- we have had 2^32
9162 		 * cacheable predicates on this machine.  We cannot allow any
9163 		 * more predicates to become cacheable:  as unlikely as it is,
9164 		 * there may be a thread caching a (now stale) predicate cache
9165 		 * ID. (N.B.: the temptation is being successfully resisted to
9166 		 * have this cmn_err() "Holy shit -- we executed this code!")
9167 		 */
9168 		return (pred);
9169 	}
9170 
9171 	pred->dtp_cacheid = dtrace_predcache_id++;
9172 
9173 	return (pred);
9174 }
9175 
9176 static void
9177 dtrace_predicate_hold(dtrace_predicate_t *pred)
9178 {
9179 	ASSERT(MUTEX_HELD(&dtrace_lock));
9180 	ASSERT(pred->dtp_difo != NULL && pred->dtp_difo->dtdo_refcnt != 0);
9181 	ASSERT(pred->dtp_refcnt > 0);
9182 
9183 	pred->dtp_refcnt++;
9184 }
9185 
9186 static void
9187 dtrace_predicate_release(dtrace_predicate_t *pred, dtrace_vstate_t *vstate)
9188 {
9189 	dtrace_difo_t *dp = pred->dtp_difo;
9190 
9191 	ASSERT(MUTEX_HELD(&dtrace_lock));
9192 	ASSERT(dp != NULL && dp->dtdo_refcnt != 0);
9193 	ASSERT(pred->dtp_refcnt > 0);
9194 
9195 	if (--pred->dtp_refcnt == 0) {
9196 		dtrace_difo_release(pred->dtp_difo, vstate);
9197 		kmem_free(pred, sizeof (dtrace_predicate_t));
9198 	}
9199 }
9200 
9201 /*
9202  * DTrace Action Description Functions
9203  */
9204 static dtrace_actdesc_t *
9205 dtrace_actdesc_create(dtrace_actkind_t kind, uint32_t ntuple,
9206     uint64_t uarg, uint64_t arg)
9207 {
9208 	dtrace_actdesc_t *act;
9209 
9210 	ASSERT(!DTRACEACT_ISPRINTFLIKE(kind) || (arg != NULL &&
9211 	    arg >= KERNELBASE) || (arg == NULL && kind == DTRACEACT_PRINTA));
9212 
9213 	act = kmem_zalloc(sizeof (dtrace_actdesc_t), KM_SLEEP);
9214 	act->dtad_kind = kind;
9215 	act->dtad_ntuple = ntuple;
9216 	act->dtad_uarg = uarg;
9217 	act->dtad_arg = arg;
9218 	act->dtad_refcnt = 1;
9219 
9220 	return (act);
9221 }
9222 
9223 static void
9224 dtrace_actdesc_hold(dtrace_actdesc_t *act)
9225 {
9226 	ASSERT(act->dtad_refcnt >= 1);
9227 	act->dtad_refcnt++;
9228 }
9229 
9230 static void
9231 dtrace_actdesc_release(dtrace_actdesc_t *act, dtrace_vstate_t *vstate)
9232 {
9233 	dtrace_actkind_t kind = act->dtad_kind;
9234 	dtrace_difo_t *dp;
9235 
9236 	ASSERT(act->dtad_refcnt >= 1);
9237 
9238 	if (--act->dtad_refcnt != 0)
9239 		return;
9240 
9241 	if ((dp = act->dtad_difo) != NULL)
9242 		dtrace_difo_release(dp, vstate);
9243 
9244 	if (DTRACEACT_ISPRINTFLIKE(kind)) {
9245 		char *str = (char *)(uintptr_t)act->dtad_arg;
9246 
9247 		ASSERT((str != NULL && (uintptr_t)str >= KERNELBASE) ||
9248 		    (str == NULL && act->dtad_kind == DTRACEACT_PRINTA));
9249 
9250 		if (str != NULL)
9251 			kmem_free(str, strlen(str) + 1);
9252 	}
9253 
9254 	kmem_free(act, sizeof (dtrace_actdesc_t));
9255 }
9256 
9257 /*
9258  * DTrace ECB Functions
9259  */
9260 static dtrace_ecb_t *
9261 dtrace_ecb_add(dtrace_state_t *state, dtrace_probe_t *probe)
9262 {
9263 	dtrace_ecb_t *ecb;
9264 	dtrace_epid_t epid;
9265 
9266 	ASSERT(MUTEX_HELD(&dtrace_lock));
9267 
9268 	ecb = kmem_zalloc(sizeof (dtrace_ecb_t), KM_SLEEP);
9269 	ecb->dte_predicate = NULL;
9270 	ecb->dte_probe = probe;
9271 
9272 	/*
9273 	 * The default size is the size of the default action: recording
9274 	 * the epid.
9275 	 */
9276 	ecb->dte_size = ecb->dte_needed = sizeof (dtrace_epid_t);
9277 	ecb->dte_alignment = sizeof (dtrace_epid_t);
9278 
9279 	epid = state->dts_epid++;
9280 
9281 	if (epid - 1 >= state->dts_necbs) {
9282 		dtrace_ecb_t **oecbs = state->dts_ecbs, **ecbs;
9283 		int necbs = state->dts_necbs << 1;
9284 
9285 		ASSERT(epid == state->dts_necbs + 1);
9286 
9287 		if (necbs == 0) {
9288 			ASSERT(oecbs == NULL);
9289 			necbs = 1;
9290 		}
9291 
9292 		ecbs = kmem_zalloc(necbs * sizeof (*ecbs), KM_SLEEP);
9293 
9294 		if (oecbs != NULL)
9295 			bcopy(oecbs, ecbs, state->dts_necbs * sizeof (*ecbs));
9296 
9297 		dtrace_membar_producer();
9298 		state->dts_ecbs = ecbs;
9299 
9300 		if (oecbs != NULL) {
9301 			/*
9302 			 * If this state is active, we must dtrace_sync()
9303 			 * before we can free the old dts_ecbs array:  we're
9304 			 * coming in hot, and there may be active ring
9305 			 * buffer processing (which indexes into the dts_ecbs
9306 			 * array) on another CPU.
9307 			 */
9308 			if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
9309 				dtrace_sync();
9310 
9311 			kmem_free(oecbs, state->dts_necbs * sizeof (*ecbs));
9312 		}
9313 
9314 		dtrace_membar_producer();
9315 		state->dts_necbs = necbs;
9316 	}
9317 
9318 	ecb->dte_state = state;
9319 
9320 	ASSERT(state->dts_ecbs[epid - 1] == NULL);
9321 	dtrace_membar_producer();
9322 	state->dts_ecbs[(ecb->dte_epid = epid) - 1] = ecb;
9323 
9324 	return (ecb);
9325 }
9326 
9327 static int
9328 dtrace_ecb_enable(dtrace_ecb_t *ecb)
9329 {
9330 	dtrace_probe_t *probe = ecb->dte_probe;
9331 
9332 	ASSERT(MUTEX_HELD(&cpu_lock));
9333 	ASSERT(MUTEX_HELD(&dtrace_lock));
9334 	ASSERT(ecb->dte_next == NULL);
9335 
9336 	if (probe == NULL) {
9337 		/*
9338 		 * This is the NULL probe -- there's nothing to do.
9339 		 */
9340 		return (0);
9341 	}
9342 
9343 	if (probe->dtpr_ecb == NULL) {
9344 		dtrace_provider_t *prov = probe->dtpr_provider;
9345 
9346 		/*
9347 		 * We're the first ECB on this probe.
9348 		 */
9349 		probe->dtpr_ecb = probe->dtpr_ecb_last = ecb;
9350 
9351 		if (ecb->dte_predicate != NULL)
9352 			probe->dtpr_predcache = ecb->dte_predicate->dtp_cacheid;
9353 
9354 		return (prov->dtpv_pops.dtps_enable(prov->dtpv_arg,
9355 		    probe->dtpr_id, probe->dtpr_arg));
9356 	} else {
9357 		/*
9358 		 * This probe is already active.  Swing the last pointer to
9359 		 * point to the new ECB, and issue a dtrace_sync() to assure
9360 		 * that all CPUs have seen the change.
9361 		 */
9362 		ASSERT(probe->dtpr_ecb_last != NULL);
9363 		probe->dtpr_ecb_last->dte_next = ecb;
9364 		probe->dtpr_ecb_last = ecb;
9365 		probe->dtpr_predcache = 0;
9366 
9367 		dtrace_sync();
9368 		return (0);
9369 	}
9370 }
9371 
9372 static void
9373 dtrace_ecb_resize(dtrace_ecb_t *ecb)
9374 {
9375 	uint32_t maxalign = sizeof (dtrace_epid_t);
9376 	uint32_t align = sizeof (uint8_t), offs, diff;
9377 	dtrace_action_t *act;
9378 	int wastuple = 0;
9379 	uint32_t aggbase = UINT32_MAX;
9380 	dtrace_state_t *state = ecb->dte_state;
9381 
9382 	/*
9383 	 * If we record anything, we always record the epid.  (And we always
9384 	 * record it first.)
9385 	 */
9386 	offs = sizeof (dtrace_epid_t);
9387 	ecb->dte_size = ecb->dte_needed = sizeof (dtrace_epid_t);
9388 
9389 	for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
9390 		dtrace_recdesc_t *rec = &act->dta_rec;
9391 
9392 		if ((align = rec->dtrd_alignment) > maxalign)
9393 			maxalign = align;
9394 
9395 		if (!wastuple && act->dta_intuple) {
9396 			/*
9397 			 * This is the first record in a tuple.  Align the
9398 			 * offset to be at offset 4 in an 8-byte aligned
9399 			 * block.
9400 			 */
9401 			diff = offs + sizeof (dtrace_aggid_t);
9402 
9403 			if (diff = (diff & (sizeof (uint64_t) - 1)))
9404 				offs += sizeof (uint64_t) - diff;
9405 
9406 			aggbase = offs - sizeof (dtrace_aggid_t);
9407 			ASSERT(!(aggbase & (sizeof (uint64_t) - 1)));
9408 		}
9409 
9410 		/*LINTED*/
9411 		if (rec->dtrd_size != 0 && (diff = (offs & (align - 1)))) {
9412 			/*
9413 			 * The current offset is not properly aligned; align it.
9414 			 */
9415 			offs += align - diff;
9416 		}
9417 
9418 		rec->dtrd_offset = offs;
9419 
9420 		if (offs + rec->dtrd_size > ecb->dte_needed) {
9421 			ecb->dte_needed = offs + rec->dtrd_size;
9422 
9423 			if (ecb->dte_needed > state->dts_needed)
9424 				state->dts_needed = ecb->dte_needed;
9425 		}
9426 
9427 		if (DTRACEACT_ISAGG(act->dta_kind)) {
9428 			dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
9429 			dtrace_action_t *first = agg->dtag_first, *prev;
9430 
9431 			ASSERT(rec->dtrd_size != 0 && first != NULL);
9432 			ASSERT(wastuple);
9433 			ASSERT(aggbase != UINT32_MAX);
9434 
9435 			agg->dtag_base = aggbase;
9436 
9437 			while ((prev = first->dta_prev) != NULL &&
9438 			    DTRACEACT_ISAGG(prev->dta_kind)) {
9439 				agg = (dtrace_aggregation_t *)prev;
9440 				first = agg->dtag_first;
9441 			}
9442 
9443 			if (prev != NULL) {
9444 				offs = prev->dta_rec.dtrd_offset +
9445 				    prev->dta_rec.dtrd_size;
9446 			} else {
9447 				offs = sizeof (dtrace_epid_t);
9448 			}
9449 			wastuple = 0;
9450 		} else {
9451 			if (!act->dta_intuple)
9452 				ecb->dte_size = offs + rec->dtrd_size;
9453 
9454 			offs += rec->dtrd_size;
9455 		}
9456 
9457 		wastuple = act->dta_intuple;
9458 	}
9459 
9460 	if ((act = ecb->dte_action) != NULL &&
9461 	    !(act->dta_kind == DTRACEACT_SPECULATE && act->dta_next == NULL) &&
9462 	    ecb->dte_size == sizeof (dtrace_epid_t)) {
9463 		/*
9464 		 * If the size is still sizeof (dtrace_epid_t), then all
9465 		 * actions store no data; set the size to 0.
9466 		 */
9467 		ecb->dte_alignment = maxalign;
9468 		ecb->dte_size = 0;
9469 
9470 		/*
9471 		 * If the needed space is still sizeof (dtrace_epid_t), then
9472 		 * all actions need no additional space; set the needed
9473 		 * size to 0.
9474 		 */
9475 		if (ecb->dte_needed == sizeof (dtrace_epid_t))
9476 			ecb->dte_needed = 0;
9477 
9478 		return;
9479 	}
9480 
9481 	/*
9482 	 * Set our alignment, and make sure that the dte_size and dte_needed
9483 	 * are aligned to the size of an EPID.
9484 	 */
9485 	ecb->dte_alignment = maxalign;
9486 	ecb->dte_size = (ecb->dte_size + (sizeof (dtrace_epid_t) - 1)) &
9487 	    ~(sizeof (dtrace_epid_t) - 1);
9488 	ecb->dte_needed = (ecb->dte_needed + (sizeof (dtrace_epid_t) - 1)) &
9489 	    ~(sizeof (dtrace_epid_t) - 1);
9490 	ASSERT(ecb->dte_size <= ecb->dte_needed);
9491 }
9492 
9493 static dtrace_action_t *
9494 dtrace_ecb_aggregation_create(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
9495 {
9496 	dtrace_aggregation_t *agg;
9497 	size_t size = sizeof (uint64_t);
9498 	int ntuple = desc->dtad_ntuple;
9499 	dtrace_action_t *act;
9500 	dtrace_recdesc_t *frec;
9501 	dtrace_aggid_t aggid;
9502 	dtrace_state_t *state = ecb->dte_state;
9503 
9504 	agg = kmem_zalloc(sizeof (dtrace_aggregation_t), KM_SLEEP);
9505 	agg->dtag_ecb = ecb;
9506 
9507 	ASSERT(DTRACEACT_ISAGG(desc->dtad_kind));
9508 
9509 	switch (desc->dtad_kind) {
9510 	case DTRACEAGG_MIN:
9511 		agg->dtag_initial = INT64_MAX;
9512 		agg->dtag_aggregate = dtrace_aggregate_min;
9513 		break;
9514 
9515 	case DTRACEAGG_MAX:
9516 		agg->dtag_initial = INT64_MIN;
9517 		agg->dtag_aggregate = dtrace_aggregate_max;
9518 		break;
9519 
9520 	case DTRACEAGG_COUNT:
9521 		agg->dtag_aggregate = dtrace_aggregate_count;
9522 		break;
9523 
9524 	case DTRACEAGG_QUANTIZE:
9525 		agg->dtag_aggregate = dtrace_aggregate_quantize;
9526 		size = (((sizeof (uint64_t) * NBBY) - 1) * 2 + 1) *
9527 		    sizeof (uint64_t);
9528 		break;
9529 
9530 	case DTRACEAGG_LQUANTIZE: {
9531 		uint16_t step = DTRACE_LQUANTIZE_STEP(desc->dtad_arg);
9532 		uint16_t levels = DTRACE_LQUANTIZE_LEVELS(desc->dtad_arg);
9533 
9534 		agg->dtag_initial = desc->dtad_arg;
9535 		agg->dtag_aggregate = dtrace_aggregate_lquantize;
9536 
9537 		if (step == 0 || levels == 0)
9538 			goto err;
9539 
9540 		size = levels * sizeof (uint64_t) + 3 * sizeof (uint64_t);
9541 		break;
9542 	}
9543 
9544 	case DTRACEAGG_LLQUANTIZE: {
9545 		uint16_t factor = DTRACE_LLQUANTIZE_FACTOR(desc->dtad_arg);
9546 		uint16_t low = DTRACE_LLQUANTIZE_LOW(desc->dtad_arg);
9547 		uint16_t high = DTRACE_LLQUANTIZE_HIGH(desc->dtad_arg);
9548 		uint16_t nsteps = DTRACE_LLQUANTIZE_NSTEP(desc->dtad_arg);
9549 		int64_t v;
9550 
9551 		agg->dtag_initial = desc->dtad_arg;
9552 		agg->dtag_aggregate = dtrace_aggregate_llquantize;
9553 
9554 		if (factor < 2 || low >= high || nsteps < factor)
9555 			goto err;
9556 
9557 		/*
9558 		 * Now check that the number of steps evenly divides a power
9559 		 * of the factor.  (This assures both integer bucket size and
9560 		 * linearity within each magnitude.)
9561 		 */
9562 		for (v = factor; v < nsteps; v *= factor)
9563 			continue;
9564 
9565 		if ((v % nsteps) || (nsteps % factor))
9566 			goto err;
9567 
9568 		size = (dtrace_aggregate_llquantize_bucket(factor,
9569 		    low, high, nsteps, INT64_MAX) + 2) * sizeof (uint64_t);
9570 		break;
9571 	}
9572 
9573 	case DTRACEAGG_AVG:
9574 		agg->dtag_aggregate = dtrace_aggregate_avg;
9575 		size = sizeof (uint64_t) * 2;
9576 		break;
9577 
9578 	case DTRACEAGG_STDDEV:
9579 		agg->dtag_aggregate = dtrace_aggregate_stddev;
9580 		size = sizeof (uint64_t) * 4;
9581 		break;
9582 
9583 	case DTRACEAGG_SUM:
9584 		agg->dtag_aggregate = dtrace_aggregate_sum;
9585 		break;
9586 
9587 	default:
9588 		goto err;
9589 	}
9590 
9591 	agg->dtag_action.dta_rec.dtrd_size = size;
9592 
9593 	if (ntuple == 0)
9594 		goto err;
9595 
9596 	/*
9597 	 * We must make sure that we have enough actions for the n-tuple.
9598 	 */
9599 	for (act = ecb->dte_action_last; act != NULL; act = act->dta_prev) {
9600 		if (DTRACEACT_ISAGG(act->dta_kind))
9601 			break;
9602 
9603 		if (--ntuple == 0) {
9604 			/*
9605 			 * This is the action with which our n-tuple begins.
9606 			 */
9607 			agg->dtag_first = act;
9608 			goto success;
9609 		}
9610 	}
9611 
9612 	/*
9613 	 * This n-tuple is short by ntuple elements.  Return failure.
9614 	 */
9615 	ASSERT(ntuple != 0);
9616 err:
9617 	kmem_free(agg, sizeof (dtrace_aggregation_t));
9618 	return (NULL);
9619 
9620 success:
9621 	/*
9622 	 * If the last action in the tuple has a size of zero, it's actually
9623 	 * an expression argument for the aggregating action.
9624 	 */
9625 	ASSERT(ecb->dte_action_last != NULL);
9626 	act = ecb->dte_action_last;
9627 
9628 	if (act->dta_kind == DTRACEACT_DIFEXPR) {
9629 		ASSERT(act->dta_difo != NULL);
9630 
9631 		if (act->dta_difo->dtdo_rtype.dtdt_size == 0)
9632 			agg->dtag_hasarg = 1;
9633 	}
9634 
9635 	/*
9636 	 * We need to allocate an id for this aggregation.
9637 	 */
9638 	aggid = (dtrace_aggid_t)(uintptr_t)vmem_alloc(state->dts_aggid_arena, 1,
9639 	    VM_BESTFIT | VM_SLEEP);
9640 
9641 	if (aggid - 1 >= state->dts_naggregations) {
9642 		dtrace_aggregation_t **oaggs = state->dts_aggregations;
9643 		dtrace_aggregation_t **aggs;
9644 		int naggs = state->dts_naggregations << 1;
9645 		int onaggs = state->dts_naggregations;
9646 
9647 		ASSERT(aggid == state->dts_naggregations + 1);
9648 
9649 		if (naggs == 0) {
9650 			ASSERT(oaggs == NULL);
9651 			naggs = 1;
9652 		}
9653 
9654 		aggs = kmem_zalloc(naggs * sizeof (*aggs), KM_SLEEP);
9655 
9656 		if (oaggs != NULL) {
9657 			bcopy(oaggs, aggs, onaggs * sizeof (*aggs));
9658 			kmem_free(oaggs, onaggs * sizeof (*aggs));
9659 		}
9660 
9661 		state->dts_aggregations = aggs;
9662 		state->dts_naggregations = naggs;
9663 	}
9664 
9665 	ASSERT(state->dts_aggregations[aggid - 1] == NULL);
9666 	state->dts_aggregations[(agg->dtag_id = aggid) - 1] = agg;
9667 
9668 	frec = &agg->dtag_first->dta_rec;
9669 	if (frec->dtrd_alignment < sizeof (dtrace_aggid_t))
9670 		frec->dtrd_alignment = sizeof (dtrace_aggid_t);
9671 
9672 	for (act = agg->dtag_first; act != NULL; act = act->dta_next) {
9673 		ASSERT(!act->dta_intuple);
9674 		act->dta_intuple = 1;
9675 	}
9676 
9677 	return (&agg->dtag_action);
9678 }
9679 
9680 static void
9681 dtrace_ecb_aggregation_destroy(dtrace_ecb_t *ecb, dtrace_action_t *act)
9682 {
9683 	dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
9684 	dtrace_state_t *state = ecb->dte_state;
9685 	dtrace_aggid_t aggid = agg->dtag_id;
9686 
9687 	ASSERT(DTRACEACT_ISAGG(act->dta_kind));
9688 	vmem_free(state->dts_aggid_arena, (void *)(uintptr_t)aggid, 1);
9689 
9690 	ASSERT(state->dts_aggregations[aggid - 1] == agg);
9691 	state->dts_aggregations[aggid - 1] = NULL;
9692 
9693 	kmem_free(agg, sizeof (dtrace_aggregation_t));
9694 }
9695 
9696 static int
9697 dtrace_ecb_action_add(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
9698 {
9699 	dtrace_action_t *action, *last;
9700 	dtrace_difo_t *dp = desc->dtad_difo;
9701 	uint32_t size = 0, align = sizeof (uint8_t), mask;
9702 	uint16_t format = 0;
9703 	dtrace_recdesc_t *rec;
9704 	dtrace_state_t *state = ecb->dte_state;
9705 	dtrace_optval_t *opt = state->dts_options, nframes, strsize;
9706 	uint64_t arg = desc->dtad_arg;
9707 
9708 	ASSERT(MUTEX_HELD(&dtrace_lock));
9709 	ASSERT(ecb->dte_action == NULL || ecb->dte_action->dta_refcnt == 1);
9710 
9711 	if (DTRACEACT_ISAGG(desc->dtad_kind)) {
9712 		/*
9713 		 * If this is an aggregating action, there must be neither
9714 		 * a speculate nor a commit on the action chain.
9715 		 */
9716 		dtrace_action_t *act;
9717 
9718 		for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
9719 			if (act->dta_kind == DTRACEACT_COMMIT)
9720 				return (EINVAL);
9721 
9722 			if (act->dta_kind == DTRACEACT_SPECULATE)
9723 				return (EINVAL);
9724 		}
9725 
9726 		action = dtrace_ecb_aggregation_create(ecb, desc);
9727 
9728 		if (action == NULL)
9729 			return (EINVAL);
9730 	} else {
9731 		if (DTRACEACT_ISDESTRUCTIVE(desc->dtad_kind) ||
9732 		    (desc->dtad_kind == DTRACEACT_DIFEXPR &&
9733 		    dp != NULL && dp->dtdo_destructive)) {
9734 			state->dts_destructive = 1;
9735 		}
9736 
9737 		switch (desc->dtad_kind) {
9738 		case DTRACEACT_PRINTF:
9739 		case DTRACEACT_PRINTA:
9740 		case DTRACEACT_SYSTEM:
9741 		case DTRACEACT_FREOPEN:
9742 			/*
9743 			 * We know that our arg is a string -- turn it into a
9744 			 * format.
9745 			 */
9746 			if (arg == NULL) {
9747 				ASSERT(desc->dtad_kind == DTRACEACT_PRINTA);
9748 				format = 0;
9749 			} else {
9750 				ASSERT(arg != NULL);
9751 				ASSERT(arg > KERNELBASE);
9752 				format = dtrace_format_add(state,
9753 				    (char *)(uintptr_t)arg);
9754 			}
9755 
9756 			/*FALLTHROUGH*/
9757 		case DTRACEACT_LIBACT:
9758 		case DTRACEACT_DIFEXPR:
9759 			if (dp == NULL)
9760 				return (EINVAL);
9761 
9762 			if ((size = dp->dtdo_rtype.dtdt_size) != 0)
9763 				break;
9764 
9765 			if (dp->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING) {
9766 				if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
9767 					return (EINVAL);
9768 
9769 				size = opt[DTRACEOPT_STRSIZE];
9770 			}
9771 
9772 			break;
9773 
9774 		case DTRACEACT_STACK:
9775 			if ((nframes = arg) == 0) {
9776 				nframes = opt[DTRACEOPT_STACKFRAMES];
9777 				ASSERT(nframes > 0);
9778 				arg = nframes;
9779 			}
9780 
9781 			size = nframes * sizeof (pc_t);
9782 			break;
9783 
9784 		case DTRACEACT_JSTACK:
9785 			if ((strsize = DTRACE_USTACK_STRSIZE(arg)) == 0)
9786 				strsize = opt[DTRACEOPT_JSTACKSTRSIZE];
9787 
9788 			if ((nframes = DTRACE_USTACK_NFRAMES(arg)) == 0)
9789 				nframes = opt[DTRACEOPT_JSTACKFRAMES];
9790 
9791 			arg = DTRACE_USTACK_ARG(nframes, strsize);
9792 
9793 			/*FALLTHROUGH*/
9794 		case DTRACEACT_USTACK:
9795 			if (desc->dtad_kind != DTRACEACT_JSTACK &&
9796 			    (nframes = DTRACE_USTACK_NFRAMES(arg)) == 0) {
9797 				strsize = DTRACE_USTACK_STRSIZE(arg);
9798 				nframes = opt[DTRACEOPT_USTACKFRAMES];
9799 				ASSERT(nframes > 0);
9800 				arg = DTRACE_USTACK_ARG(nframes, strsize);
9801 			}
9802 
9803 			/*
9804 			 * Save a slot for the pid.
9805 			 */
9806 			size = (nframes + 1) * sizeof (uint64_t);
9807 			size += DTRACE_USTACK_STRSIZE(arg);
9808 			size = P2ROUNDUP(size, (uint32_t)(sizeof (uintptr_t)));
9809 
9810 			break;
9811 
9812 		case DTRACEACT_SYM:
9813 		case DTRACEACT_MOD:
9814 			if (dp == NULL || ((size = dp->dtdo_rtype.dtdt_size) !=
9815 			    sizeof (uint64_t)) ||
9816 			    (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
9817 				return (EINVAL);
9818 			break;
9819 
9820 		case DTRACEACT_USYM:
9821 		case DTRACEACT_UMOD:
9822 		case DTRACEACT_UADDR:
9823 			if (dp == NULL ||
9824 			    (dp->dtdo_rtype.dtdt_size != sizeof (uint64_t)) ||
9825 			    (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
9826 				return (EINVAL);
9827 
9828 			/*
9829 			 * We have a slot for the pid, plus a slot for the
9830 			 * argument.  To keep things simple (aligned with
9831 			 * bitness-neutral sizing), we store each as a 64-bit
9832 			 * quantity.
9833 			 */
9834 			size = 2 * sizeof (uint64_t);
9835 			break;
9836 
9837 		case DTRACEACT_STOP:
9838 		case DTRACEACT_BREAKPOINT:
9839 		case DTRACEACT_PANIC:
9840 			break;
9841 
9842 		case DTRACEACT_CHILL:
9843 		case DTRACEACT_DISCARD:
9844 		case DTRACEACT_RAISE:
9845 			if (dp == NULL)
9846 				return (EINVAL);
9847 			break;
9848 
9849 		case DTRACEACT_EXIT:
9850 			if (dp == NULL ||
9851 			    (size = dp->dtdo_rtype.dtdt_size) != sizeof (int) ||
9852 			    (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
9853 				return (EINVAL);
9854 			break;
9855 
9856 		case DTRACEACT_SPECULATE:
9857 			if (ecb->dte_size > sizeof (dtrace_epid_t))
9858 				return (EINVAL);
9859 
9860 			if (dp == NULL)
9861 				return (EINVAL);
9862 
9863 			state->dts_speculates = 1;
9864 			break;
9865 
9866 		case DTRACEACT_COMMIT: {
9867 			dtrace_action_t *act = ecb->dte_action;
9868 
9869 			for (; act != NULL; act = act->dta_next) {
9870 				if (act->dta_kind == DTRACEACT_COMMIT)
9871 					return (EINVAL);
9872 			}
9873 
9874 			if (dp == NULL)
9875 				return (EINVAL);
9876 			break;
9877 		}
9878 
9879 		default:
9880 			return (EINVAL);
9881 		}
9882 
9883 		if (size != 0 || desc->dtad_kind == DTRACEACT_SPECULATE) {
9884 			/*
9885 			 * If this is a data-storing action or a speculate,
9886 			 * we must be sure that there isn't a commit on the
9887 			 * action chain.
9888 			 */
9889 			dtrace_action_t *act = ecb->dte_action;
9890 
9891 			for (; act != NULL; act = act->dta_next) {
9892 				if (act->dta_kind == DTRACEACT_COMMIT)
9893 					return (EINVAL);
9894 			}
9895 		}
9896 
9897 		action = kmem_zalloc(sizeof (dtrace_action_t), KM_SLEEP);
9898 		action->dta_rec.dtrd_size = size;
9899 	}
9900 
9901 	action->dta_refcnt = 1;
9902 	rec = &action->dta_rec;
9903 	size = rec->dtrd_size;
9904 
9905 	for (mask = sizeof (uint64_t) - 1; size != 0 && mask > 0; mask >>= 1) {
9906 		if (!(size & mask)) {
9907 			align = mask + 1;
9908 			break;
9909 		}
9910 	}
9911 
9912 	action->dta_kind = desc->dtad_kind;
9913 
9914 	if ((action->dta_difo = dp) != NULL)
9915 		dtrace_difo_hold(dp);
9916 
9917 	rec->dtrd_action = action->dta_kind;
9918 	rec->dtrd_arg = arg;
9919 	rec->dtrd_uarg = desc->dtad_uarg;
9920 	rec->dtrd_alignment = (uint16_t)align;
9921 	rec->dtrd_format = format;
9922 
9923 	if ((last = ecb->dte_action_last) != NULL) {
9924 		ASSERT(ecb->dte_action != NULL);
9925 		action->dta_prev = last;
9926 		last->dta_next = action;
9927 	} else {
9928 		ASSERT(ecb->dte_action == NULL);
9929 		ecb->dte_action = action;
9930 	}
9931 
9932 	ecb->dte_action_last = action;
9933 
9934 	return (0);
9935 }
9936 
9937 static void
9938 dtrace_ecb_action_remove(dtrace_ecb_t *ecb)
9939 {
9940 	dtrace_action_t *act = ecb->dte_action, *next;
9941 	dtrace_vstate_t *vstate = &ecb->dte_state->dts_vstate;
9942 	dtrace_difo_t *dp;
9943 	uint16_t format;
9944 
9945 	if (act != NULL && act->dta_refcnt > 1) {
9946 		ASSERT(act->dta_next == NULL || act->dta_next->dta_refcnt == 1);
9947 		act->dta_refcnt--;
9948 	} else {
9949 		for (; act != NULL; act = next) {
9950 			next = act->dta_next;
9951 			ASSERT(next != NULL || act == ecb->dte_action_last);
9952 			ASSERT(act->dta_refcnt == 1);
9953 
9954 			if ((format = act->dta_rec.dtrd_format) != 0)
9955 				dtrace_format_remove(ecb->dte_state, format);
9956 
9957 			if ((dp = act->dta_difo) != NULL)
9958 				dtrace_difo_release(dp, vstate);
9959 
9960 			if (DTRACEACT_ISAGG(act->dta_kind)) {
9961 				dtrace_ecb_aggregation_destroy(ecb, act);
9962 			} else {
9963 				kmem_free(act, sizeof (dtrace_action_t));
9964 			}
9965 		}
9966 	}
9967 
9968 	ecb->dte_action = NULL;
9969 	ecb->dte_action_last = NULL;
9970 	ecb->dte_size = sizeof (dtrace_epid_t);
9971 }
9972 
9973 static void
9974 dtrace_ecb_disable(dtrace_ecb_t *ecb)
9975 {
9976 	/*
9977 	 * We disable the ECB by removing it from its probe.
9978 	 */
9979 	dtrace_ecb_t *pecb, *prev = NULL;
9980 	dtrace_probe_t *probe = ecb->dte_probe;
9981 
9982 	ASSERT(MUTEX_HELD(&dtrace_lock));
9983 
9984 	if (probe == NULL) {
9985 		/*
9986 		 * This is the NULL probe; there is nothing to disable.
9987 		 */
9988 		return;
9989 	}
9990 
9991 	for (pecb = probe->dtpr_ecb; pecb != NULL; pecb = pecb->dte_next) {
9992 		if (pecb == ecb)
9993 			break;
9994 		prev = pecb;
9995 	}
9996 
9997 	ASSERT(pecb != NULL);
9998 
9999 	if (prev == NULL) {
10000 		probe->dtpr_ecb = ecb->dte_next;
10001 	} else {
10002 		prev->dte_next = ecb->dte_next;
10003 	}
10004 
10005 	if (ecb == probe->dtpr_ecb_last) {
10006 		ASSERT(ecb->dte_next == NULL);
10007 		probe->dtpr_ecb_last = prev;
10008 	}
10009 
10010 	/*
10011 	 * The ECB has been disconnected from the probe; now sync to assure
10012 	 * that all CPUs have seen the change before returning.
10013 	 */
10014 	dtrace_sync();
10015 
10016 	if (probe->dtpr_ecb == NULL) {
10017 		/*
10018 		 * That was the last ECB on the probe; clear the predicate
10019 		 * cache ID for the probe, disable it and sync one more time
10020 		 * to assure that we'll never hit it again.
10021 		 */
10022 		dtrace_provider_t *prov = probe->dtpr_provider;
10023 
10024 		ASSERT(ecb->dte_next == NULL);
10025 		ASSERT(probe->dtpr_ecb_last == NULL);
10026 		probe->dtpr_predcache = DTRACE_CACHEIDNONE;
10027 		prov->dtpv_pops.dtps_disable(prov->dtpv_arg,
10028 		    probe->dtpr_id, probe->dtpr_arg);
10029 		dtrace_sync();
10030 	} else {
10031 		/*
10032 		 * There is at least one ECB remaining on the probe.  If there
10033 		 * is _exactly_ one, set the probe's predicate cache ID to be
10034 		 * the predicate cache ID of the remaining ECB.
10035 		 */
10036 		ASSERT(probe->dtpr_ecb_last != NULL);
10037 		ASSERT(probe->dtpr_predcache == DTRACE_CACHEIDNONE);
10038 
10039 		if (probe->dtpr_ecb == probe->dtpr_ecb_last) {
10040 			dtrace_predicate_t *p = probe->dtpr_ecb->dte_predicate;
10041 
10042 			ASSERT(probe->dtpr_ecb->dte_next == NULL);
10043 
10044 			if (p != NULL)
10045 				probe->dtpr_predcache = p->dtp_cacheid;
10046 		}
10047 
10048 		ecb->dte_next = NULL;
10049 	}
10050 }
10051 
10052 static void
10053 dtrace_ecb_destroy(dtrace_ecb_t *ecb)
10054 {
10055 	dtrace_state_t *state = ecb->dte_state;
10056 	dtrace_vstate_t *vstate = &state->dts_vstate;
10057 	dtrace_predicate_t *pred;
10058 	dtrace_epid_t epid = ecb->dte_epid;
10059 
10060 	ASSERT(MUTEX_HELD(&dtrace_lock));
10061 	ASSERT(ecb->dte_next == NULL);
10062 	ASSERT(ecb->dte_probe == NULL || ecb->dte_probe->dtpr_ecb != ecb);
10063 
10064 	if ((pred = ecb->dte_predicate) != NULL)
10065 		dtrace_predicate_release(pred, vstate);
10066 
10067 	dtrace_ecb_action_remove(ecb);
10068 
10069 	ASSERT(state->dts_ecbs[epid - 1] == ecb);
10070 	state->dts_ecbs[epid - 1] = NULL;
10071 
10072 	kmem_free(ecb, sizeof (dtrace_ecb_t));
10073 }
10074 
10075 static dtrace_ecb_t *
10076 dtrace_ecb_create(dtrace_state_t *state, dtrace_probe_t *probe,
10077     dtrace_enabling_t *enab)
10078 {
10079 	dtrace_ecb_t *ecb;
10080 	dtrace_predicate_t *pred;
10081 	dtrace_actdesc_t *act;
10082 	dtrace_provider_t *prov;
10083 	dtrace_ecbdesc_t *desc = enab->dten_current;
10084 
10085 	ASSERT(MUTEX_HELD(&dtrace_lock));
10086 	ASSERT(state != NULL);
10087 
10088 	ecb = dtrace_ecb_add(state, probe);
10089 	ecb->dte_uarg = desc->dted_uarg;
10090 
10091 	if ((pred = desc->dted_pred.dtpdd_predicate) != NULL) {
10092 		dtrace_predicate_hold(pred);
10093 		ecb->dte_predicate = pred;
10094 	}
10095 
10096 	if (probe != NULL) {
10097 		/*
10098 		 * If the provider shows more leg than the consumer is old
10099 		 * enough to see, we need to enable the appropriate implicit
10100 		 * predicate bits to prevent the ecb from activating at
10101 		 * revealing times.
10102 		 *
10103 		 * Providers specifying DTRACE_PRIV_USER at register time
10104 		 * are stating that they need the /proc-style privilege
10105 		 * model to be enforced, and this is what DTRACE_COND_OWNER
10106 		 * and DTRACE_COND_ZONEOWNER will then do at probe time.
10107 		 */
10108 		prov = probe->dtpr_provider;
10109 		if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLPROC) &&
10110 		    (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
10111 			ecb->dte_cond |= DTRACE_COND_OWNER;
10112 
10113 		if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLZONE) &&
10114 		    (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
10115 			ecb->dte_cond |= DTRACE_COND_ZONEOWNER;
10116 
10117 		/*
10118 		 * If the provider shows us kernel innards and the user
10119 		 * is lacking sufficient privilege, enable the
10120 		 * DTRACE_COND_USERMODE implicit predicate.
10121 		 */
10122 		if (!(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) &&
10123 		    (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_KERNEL))
10124 			ecb->dte_cond |= DTRACE_COND_USERMODE;
10125 	}
10126 
10127 	if (dtrace_ecb_create_cache != NULL) {
10128 		/*
10129 		 * If we have a cached ecb, we'll use its action list instead
10130 		 * of creating our own (saving both time and space).
10131 		 */
10132 		dtrace_ecb_t *cached = dtrace_ecb_create_cache;
10133 		dtrace_action_t *act = cached->dte_action;
10134 
10135 		if (act != NULL) {
10136 			ASSERT(act->dta_refcnt > 0);
10137 			act->dta_refcnt++;
10138 			ecb->dte_action = act;
10139 			ecb->dte_action_last = cached->dte_action_last;
10140 			ecb->dte_needed = cached->dte_needed;
10141 			ecb->dte_size = cached->dte_size;
10142 			ecb->dte_alignment = cached->dte_alignment;
10143 		}
10144 
10145 		return (ecb);
10146 	}
10147 
10148 	for (act = desc->dted_action; act != NULL; act = act->dtad_next) {
10149 		if ((enab->dten_error = dtrace_ecb_action_add(ecb, act)) != 0) {
10150 			dtrace_ecb_destroy(ecb);
10151 			return (NULL);
10152 		}
10153 	}
10154 
10155 	dtrace_ecb_resize(ecb);
10156 
10157 	return (dtrace_ecb_create_cache = ecb);
10158 }
10159 
10160 static int
10161 dtrace_ecb_create_enable(dtrace_probe_t *probe, void *arg)
10162 {
10163 	dtrace_ecb_t *ecb;
10164 	dtrace_enabling_t *enab = arg;
10165 	dtrace_state_t *state = enab->dten_vstate->dtvs_state;
10166 
10167 	ASSERT(state != NULL);
10168 
10169 	if (probe != NULL && probe->dtpr_gen < enab->dten_probegen) {
10170 		/*
10171 		 * This probe was created in a generation for which this
10172 		 * enabling has previously created ECBs; we don't want to
10173 		 * enable it again, so just kick out.
10174 		 */
10175 		return (DTRACE_MATCH_NEXT);
10176 	}
10177 
10178 	if ((ecb = dtrace_ecb_create(state, probe, enab)) == NULL)
10179 		return (DTRACE_MATCH_DONE);
10180 
10181 	if (dtrace_ecb_enable(ecb) < 0)
10182 		return (DTRACE_MATCH_FAIL);
10183 
10184 	return (DTRACE_MATCH_NEXT);
10185 }
10186 
10187 static dtrace_ecb_t *
10188 dtrace_epid2ecb(dtrace_state_t *state, dtrace_epid_t id)
10189 {
10190 	dtrace_ecb_t *ecb;
10191 
10192 	ASSERT(MUTEX_HELD(&dtrace_lock));
10193 
10194 	if (id == 0 || id > state->dts_necbs)
10195 		return (NULL);
10196 
10197 	ASSERT(state->dts_necbs > 0 && state->dts_ecbs != NULL);
10198 	ASSERT((ecb = state->dts_ecbs[id - 1]) == NULL || ecb->dte_epid == id);
10199 
10200 	return (state->dts_ecbs[id - 1]);
10201 }
10202 
10203 static dtrace_aggregation_t *
10204 dtrace_aggid2agg(dtrace_state_t *state, dtrace_aggid_t id)
10205 {
10206 	dtrace_aggregation_t *agg;
10207 
10208 	ASSERT(MUTEX_HELD(&dtrace_lock));
10209 
10210 	if (id == 0 || id > state->dts_naggregations)
10211 		return (NULL);
10212 
10213 	ASSERT(state->dts_naggregations > 0 && state->dts_aggregations != NULL);
10214 	ASSERT((agg = state->dts_aggregations[id - 1]) == NULL ||
10215 	    agg->dtag_id == id);
10216 
10217 	return (state->dts_aggregations[id - 1]);
10218 }
10219 
10220 /*
10221  * DTrace Buffer Functions
10222  *
10223  * The following functions manipulate DTrace buffers.  Most of these functions
10224  * are called in the context of establishing or processing consumer state;
10225  * exceptions are explicitly noted.
10226  */
10227 
10228 /*
10229  * Note:  called from cross call context.  This function switches the two
10230  * buffers on a given CPU.  The atomicity of this operation is assured by
10231  * disabling interrupts while the actual switch takes place; the disabling of
10232  * interrupts serializes the execution with any execution of dtrace_probe() on
10233  * the same CPU.
10234  */
10235 static void
10236 dtrace_buffer_switch(dtrace_buffer_t *buf)
10237 {
10238 	caddr_t tomax = buf->dtb_tomax;
10239 	caddr_t xamot = buf->dtb_xamot;
10240 	dtrace_icookie_t cookie;
10241 	hrtime_t now = dtrace_gethrtime();
10242 
10243 	ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
10244 	ASSERT(!(buf->dtb_flags & DTRACEBUF_RING));
10245 
10246 	cookie = dtrace_interrupt_disable();
10247 	buf->dtb_tomax = xamot;
10248 	buf->dtb_xamot = tomax;
10249 	buf->dtb_xamot_drops = buf->dtb_drops;
10250 	buf->dtb_xamot_offset = buf->dtb_offset;
10251 	buf->dtb_xamot_errors = buf->dtb_errors;
10252 	buf->dtb_xamot_flags = buf->dtb_flags;
10253 	buf->dtb_offset = 0;
10254 	buf->dtb_drops = 0;
10255 	buf->dtb_errors = 0;
10256 	buf->dtb_flags &= ~(DTRACEBUF_ERROR | DTRACEBUF_DROPPED);
10257 	buf->dtb_interval = now - buf->dtb_switched;
10258 	buf->dtb_switched = now;
10259 	dtrace_interrupt_enable(cookie);
10260 }
10261 
10262 /*
10263  * Note:  called from cross call context.  This function activates a buffer
10264  * on a CPU.  As with dtrace_buffer_switch(), the atomicity of the operation
10265  * is guaranteed by the disabling of interrupts.
10266  */
10267 static void
10268 dtrace_buffer_activate(dtrace_state_t *state)
10269 {
10270 	dtrace_buffer_t *buf;
10271 	dtrace_icookie_t cookie = dtrace_interrupt_disable();
10272 
10273 	buf = &state->dts_buffer[CPU->cpu_id];
10274 
10275 	if (buf->dtb_tomax != NULL) {
10276 		/*
10277 		 * We might like to assert that the buffer is marked inactive,
10278 		 * but this isn't necessarily true:  the buffer for the CPU
10279 		 * that processes the BEGIN probe has its buffer activated
10280 		 * manually.  In this case, we take the (harmless) action
10281 		 * re-clearing the bit INACTIVE bit.
10282 		 */
10283 		buf->dtb_flags &= ~DTRACEBUF_INACTIVE;
10284 	}
10285 
10286 	dtrace_interrupt_enable(cookie);
10287 }
10288 
10289 static int
10290 dtrace_buffer_alloc(dtrace_buffer_t *bufs, size_t size, int flags,
10291     processorid_t cpu, int *factor)
10292 {
10293 	cpu_t *cp;
10294 	dtrace_buffer_t *buf;
10295 	int allocated = 0, desired = 0;
10296 
10297 	ASSERT(MUTEX_HELD(&cpu_lock));
10298 	ASSERT(MUTEX_HELD(&dtrace_lock));
10299 
10300 	*factor = 1;
10301 
10302 	if (size > dtrace_nonroot_maxsize &&
10303 	    !PRIV_POLICY_CHOICE(CRED(), PRIV_ALL, B_FALSE))
10304 		return (EFBIG);
10305 
10306 	cp = cpu_list;
10307 
10308 	do {
10309 		if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
10310 			continue;
10311 
10312 		buf = &bufs[cp->cpu_id];
10313 
10314 		/*
10315 		 * If there is already a buffer allocated for this CPU, it
10316 		 * is only possible that this is a DR event.  In this case,
10317 		 * the buffer size must match our specified size.
10318 		 */
10319 		if (buf->dtb_tomax != NULL) {
10320 			ASSERT(buf->dtb_size == size);
10321 			continue;
10322 		}
10323 
10324 		ASSERT(buf->dtb_xamot == NULL);
10325 
10326 		if ((buf->dtb_tomax = kmem_zalloc(size,
10327 		    KM_NOSLEEP | KM_NORMALPRI)) == NULL)
10328 			goto err;
10329 
10330 		buf->dtb_size = size;
10331 		buf->dtb_flags = flags;
10332 		buf->dtb_offset = 0;
10333 		buf->dtb_drops = 0;
10334 
10335 		if (flags & DTRACEBUF_NOSWITCH)
10336 			continue;
10337 
10338 		if ((buf->dtb_xamot = kmem_zalloc(size,
10339 		    KM_NOSLEEP | KM_NORMALPRI)) == NULL)
10340 			goto err;
10341 	} while ((cp = cp->cpu_next) != cpu_list);
10342 
10343 	return (0);
10344 
10345 err:
10346 	cp = cpu_list;
10347 
10348 	do {
10349 		if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
10350 			continue;
10351 
10352 		buf = &bufs[cp->cpu_id];
10353 		desired += 2;
10354 
10355 		if (buf->dtb_xamot != NULL) {
10356 			ASSERT(buf->dtb_tomax != NULL);
10357 			ASSERT(buf->dtb_size == size);
10358 			kmem_free(buf->dtb_xamot, size);
10359 			allocated++;
10360 		}
10361 
10362 		if (buf->dtb_tomax != NULL) {
10363 			ASSERT(buf->dtb_size == size);
10364 			kmem_free(buf->dtb_tomax, size);
10365 			allocated++;
10366 		}
10367 
10368 		buf->dtb_tomax = NULL;
10369 		buf->dtb_xamot = NULL;
10370 		buf->dtb_size = 0;
10371 	} while ((cp = cp->cpu_next) != cpu_list);
10372 
10373 	*factor = desired / (allocated > 0 ? allocated : 1);
10374 
10375 	return (ENOMEM);
10376 }
10377 
10378 /*
10379  * Note:  called from probe context.  This function just increments the drop
10380  * count on a buffer.  It has been made a function to allow for the
10381  * possibility of understanding the source of mysterious drop counts.  (A
10382  * problem for which one may be particularly disappointed that DTrace cannot
10383  * be used to understand DTrace.)
10384  */
10385 static void
10386 dtrace_buffer_drop(dtrace_buffer_t *buf)
10387 {
10388 	buf->dtb_drops++;
10389 }
10390 
10391 /*
10392  * Note:  called from probe context.  This function is called to reserve space
10393  * in a buffer.  If mstate is non-NULL, sets the scratch base and size in the
10394  * mstate.  Returns the new offset in the buffer, or a negative value if an
10395  * error has occurred.
10396  */
10397 static intptr_t
10398 dtrace_buffer_reserve(dtrace_buffer_t *buf, size_t needed, size_t align,
10399     dtrace_state_t *state, dtrace_mstate_t *mstate)
10400 {
10401 	intptr_t offs = buf->dtb_offset, soffs;
10402 	intptr_t woffs;
10403 	caddr_t tomax;
10404 	size_t total;
10405 
10406 	if (buf->dtb_flags & DTRACEBUF_INACTIVE)
10407 		return (-1);
10408 
10409 	if ((tomax = buf->dtb_tomax) == NULL) {
10410 		dtrace_buffer_drop(buf);
10411 		return (-1);
10412 	}
10413 
10414 	if (!(buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL))) {
10415 		while (offs & (align - 1)) {
10416 			/*
10417 			 * Assert that our alignment is off by a number which
10418 			 * is itself sizeof (uint32_t) aligned.
10419 			 */
10420 			ASSERT(!((align - (offs & (align - 1))) &
10421 			    (sizeof (uint32_t) - 1)));
10422 			DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
10423 			offs += sizeof (uint32_t);
10424 		}
10425 
10426 		if ((soffs = offs + needed) > buf->dtb_size) {
10427 			dtrace_buffer_drop(buf);
10428 			return (-1);
10429 		}
10430 
10431 		if (mstate == NULL)
10432 			return (offs);
10433 
10434 		mstate->dtms_scratch_base = (uintptr_t)tomax + soffs;
10435 		mstate->dtms_scratch_size = buf->dtb_size - soffs;
10436 		mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
10437 
10438 		return (offs);
10439 	}
10440 
10441 	if (buf->dtb_flags & DTRACEBUF_FILL) {
10442 		if (state->dts_activity != DTRACE_ACTIVITY_COOLDOWN &&
10443 		    (buf->dtb_flags & DTRACEBUF_FULL))
10444 			return (-1);
10445 		goto out;
10446 	}
10447 
10448 	total = needed + (offs & (align - 1));
10449 
10450 	/*
10451 	 * For a ring buffer, life is quite a bit more complicated.  Before
10452 	 * we can store any padding, we need to adjust our wrapping offset.
10453 	 * (If we've never before wrapped or we're not about to, no adjustment
10454 	 * is required.)
10455 	 */
10456 	if ((buf->dtb_flags & DTRACEBUF_WRAPPED) ||
10457 	    offs + total > buf->dtb_size) {
10458 		woffs = buf->dtb_xamot_offset;
10459 
10460 		if (offs + total > buf->dtb_size) {
10461 			/*
10462 			 * We can't fit in the end of the buffer.  First, a
10463 			 * sanity check that we can fit in the buffer at all.
10464 			 */
10465 			if (total > buf->dtb_size) {
10466 				dtrace_buffer_drop(buf);
10467 				return (-1);
10468 			}
10469 
10470 			/*
10471 			 * We're going to be storing at the top of the buffer,
10472 			 * so now we need to deal with the wrapped offset.  We
10473 			 * only reset our wrapped offset to 0 if it is
10474 			 * currently greater than the current offset.  If it
10475 			 * is less than the current offset, it is because a
10476 			 * previous allocation induced a wrap -- but the
10477 			 * allocation didn't subsequently take the space due
10478 			 * to an error or false predicate evaluation.  In this
10479 			 * case, we'll just leave the wrapped offset alone: if
10480 			 * the wrapped offset hasn't been advanced far enough
10481 			 * for this allocation, it will be adjusted in the
10482 			 * lower loop.
10483 			 */
10484 			if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
10485 				if (woffs >= offs)
10486 					woffs = 0;
10487 			} else {
10488 				woffs = 0;
10489 			}
10490 
10491 			/*
10492 			 * Now we know that we're going to be storing to the
10493 			 * top of the buffer and that there is room for us
10494 			 * there.  We need to clear the buffer from the current
10495 			 * offset to the end (there may be old gunk there).
10496 			 */
10497 			while (offs < buf->dtb_size)
10498 				tomax[offs++] = 0;
10499 
10500 			/*
10501 			 * We need to set our offset to zero.  And because we
10502 			 * are wrapping, we need to set the bit indicating as
10503 			 * much.  We can also adjust our needed space back
10504 			 * down to the space required by the ECB -- we know
10505 			 * that the top of the buffer is aligned.
10506 			 */
10507 			offs = 0;
10508 			total = needed;
10509 			buf->dtb_flags |= DTRACEBUF_WRAPPED;
10510 		} else {
10511 			/*
10512 			 * There is room for us in the buffer, so we simply
10513 			 * need to check the wrapped offset.
10514 			 */
10515 			if (woffs < offs) {
10516 				/*
10517 				 * The wrapped offset is less than the offset.
10518 				 * This can happen if we allocated buffer space
10519 				 * that induced a wrap, but then we didn't
10520 				 * subsequently take the space due to an error
10521 				 * or false predicate evaluation.  This is
10522 				 * okay; we know that _this_ allocation isn't
10523 				 * going to induce a wrap.  We still can't
10524 				 * reset the wrapped offset to be zero,
10525 				 * however: the space may have been trashed in
10526 				 * the previous failed probe attempt.  But at
10527 				 * least the wrapped offset doesn't need to
10528 				 * be adjusted at all...
10529 				 */
10530 				goto out;
10531 			}
10532 		}
10533 
10534 		while (offs + total > woffs) {
10535 			dtrace_epid_t epid = *(uint32_t *)(tomax + woffs);
10536 			size_t size;
10537 
10538 			if (epid == DTRACE_EPIDNONE) {
10539 				size = sizeof (uint32_t);
10540 			} else {
10541 				ASSERT(epid <= state->dts_necbs);
10542 				ASSERT(state->dts_ecbs[epid - 1] != NULL);
10543 
10544 				size = state->dts_ecbs[epid - 1]->dte_size;
10545 			}
10546 
10547 			ASSERT(woffs + size <= buf->dtb_size);
10548 			ASSERT(size != 0);
10549 
10550 			if (woffs + size == buf->dtb_size) {
10551 				/*
10552 				 * We've reached the end of the buffer; we want
10553 				 * to set the wrapped offset to 0 and break
10554 				 * out.  However, if the offs is 0, then we're
10555 				 * in a strange edge-condition:  the amount of
10556 				 * space that we want to reserve plus the size
10557 				 * of the record that we're overwriting is
10558 				 * greater than the size of the buffer.  This
10559 				 * is problematic because if we reserve the
10560 				 * space but subsequently don't consume it (due
10561 				 * to a failed predicate or error) the wrapped
10562 				 * offset will be 0 -- yet the EPID at offset 0
10563 				 * will not be committed.  This situation is
10564 				 * relatively easy to deal with:  if we're in
10565 				 * this case, the buffer is indistinguishable
10566 				 * from one that hasn't wrapped; we need only
10567 				 * finish the job by clearing the wrapped bit,
10568 				 * explicitly setting the offset to be 0, and
10569 				 * zero'ing out the old data in the buffer.
10570 				 */
10571 				if (offs == 0) {
10572 					buf->dtb_flags &= ~DTRACEBUF_WRAPPED;
10573 					buf->dtb_offset = 0;
10574 					woffs = total;
10575 
10576 					while (woffs < buf->dtb_size)
10577 						tomax[woffs++] = 0;
10578 				}
10579 
10580 				woffs = 0;
10581 				break;
10582 			}
10583 
10584 			woffs += size;
10585 		}
10586 
10587 		/*
10588 		 * We have a wrapped offset.  It may be that the wrapped offset
10589 		 * has become zero -- that's okay.
10590 		 */
10591 		buf->dtb_xamot_offset = woffs;
10592 	}
10593 
10594 out:
10595 	/*
10596 	 * Now we can plow the buffer with any necessary padding.
10597 	 */
10598 	while (offs & (align - 1)) {
10599 		/*
10600 		 * Assert that our alignment is off by a number which
10601 		 * is itself sizeof (uint32_t) aligned.
10602 		 */
10603 		ASSERT(!((align - (offs & (align - 1))) &
10604 		    (sizeof (uint32_t) - 1)));
10605 		DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
10606 		offs += sizeof (uint32_t);
10607 	}
10608 
10609 	if (buf->dtb_flags & DTRACEBUF_FILL) {
10610 		if (offs + needed > buf->dtb_size - state->dts_reserve) {
10611 			buf->dtb_flags |= DTRACEBUF_FULL;
10612 			return (-1);
10613 		}
10614 	}
10615 
10616 	if (mstate == NULL)
10617 		return (offs);
10618 
10619 	/*
10620 	 * For ring buffers and fill buffers, the scratch space is always
10621 	 * the inactive buffer.
10622 	 */
10623 	mstate->dtms_scratch_base = (uintptr_t)buf->dtb_xamot;
10624 	mstate->dtms_scratch_size = buf->dtb_size;
10625 	mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
10626 
10627 	return (offs);
10628 }
10629 
10630 static void
10631 dtrace_buffer_polish(dtrace_buffer_t *buf)
10632 {
10633 	ASSERT(buf->dtb_flags & DTRACEBUF_RING);
10634 	ASSERT(MUTEX_HELD(&dtrace_lock));
10635 
10636 	if (!(buf->dtb_flags & DTRACEBUF_WRAPPED))
10637 		return;
10638 
10639 	/*
10640 	 * We need to polish the ring buffer.  There are three cases:
10641 	 *
10642 	 * - The first (and presumably most common) is that there is no gap
10643 	 *   between the buffer offset and the wrapped offset.  In this case,
10644 	 *   there is nothing in the buffer that isn't valid data; we can
10645 	 *   mark the buffer as polished and return.
10646 	 *
10647 	 * - The second (less common than the first but still more common
10648 	 *   than the third) is that there is a gap between the buffer offset
10649 	 *   and the wrapped offset, and the wrapped offset is larger than the
10650 	 *   buffer offset.  This can happen because of an alignment issue, or
10651 	 *   can happen because of a call to dtrace_buffer_reserve() that
10652 	 *   didn't subsequently consume the buffer space.  In this case,
10653 	 *   we need to zero the data from the buffer offset to the wrapped
10654 	 *   offset.
10655 	 *
10656 	 * - The third (and least common) is that there is a gap between the
10657 	 *   buffer offset and the wrapped offset, but the wrapped offset is
10658 	 *   _less_ than the buffer offset.  This can only happen because a
10659 	 *   call to dtrace_buffer_reserve() induced a wrap, but the space
10660 	 *   was not subsequently consumed.  In this case, we need to zero the
10661 	 *   space from the offset to the end of the buffer _and_ from the
10662 	 *   top of the buffer to the wrapped offset.
10663 	 */
10664 	if (buf->dtb_offset < buf->dtb_xamot_offset) {
10665 		bzero(buf->dtb_tomax + buf->dtb_offset,
10666 		    buf->dtb_xamot_offset - buf->dtb_offset);
10667 	}
10668 
10669 	if (buf->dtb_offset > buf->dtb_xamot_offset) {
10670 		bzero(buf->dtb_tomax + buf->dtb_offset,
10671 		    buf->dtb_size - buf->dtb_offset);
10672 		bzero(buf->dtb_tomax, buf->dtb_xamot_offset);
10673 	}
10674 }
10675 
10676 /*
10677  * This routine determines if data generated at the specified time has likely
10678  * been entirely consumed at user-level.  This routine is called to determine
10679  * if an ECB on a defunct probe (but for an active enabling) can be safely
10680  * disabled and destroyed.
10681  */
10682 static int
10683 dtrace_buffer_consumed(dtrace_buffer_t *bufs, hrtime_t when)
10684 {
10685 	int i;
10686 
10687 	for (i = 0; i < NCPU; i++) {
10688 		dtrace_buffer_t *buf = &bufs[i];
10689 
10690 		if (buf->dtb_size == 0)
10691 			continue;
10692 
10693 		if (buf->dtb_flags & DTRACEBUF_RING)
10694 			return (0);
10695 
10696 		if (!buf->dtb_switched && buf->dtb_offset != 0)
10697 			return (0);
10698 
10699 		if (buf->dtb_switched - buf->dtb_interval < when)
10700 			return (0);
10701 	}
10702 
10703 	return (1);
10704 }
10705 
10706 static void
10707 dtrace_buffer_free(dtrace_buffer_t *bufs)
10708 {
10709 	int i;
10710 
10711 	for (i = 0; i < NCPU; i++) {
10712 		dtrace_buffer_t *buf = &bufs[i];
10713 
10714 		if (buf->dtb_tomax == NULL) {
10715 			ASSERT(buf->dtb_xamot == NULL);
10716 			ASSERT(buf->dtb_size == 0);
10717 			continue;
10718 		}
10719 
10720 		if (buf->dtb_xamot != NULL) {
10721 			ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
10722 			kmem_free(buf->dtb_xamot, buf->dtb_size);
10723 		}
10724 
10725 		kmem_free(buf->dtb_tomax, buf->dtb_size);
10726 		buf->dtb_size = 0;
10727 		buf->dtb_tomax = NULL;
10728 		buf->dtb_xamot = NULL;
10729 	}
10730 }
10731 
10732 /*
10733  * DTrace Enabling Functions
10734  */
10735 static dtrace_enabling_t *
10736 dtrace_enabling_create(dtrace_vstate_t *vstate)
10737 {
10738 	dtrace_enabling_t *enab;
10739 
10740 	enab = kmem_zalloc(sizeof (dtrace_enabling_t), KM_SLEEP);
10741 	enab->dten_vstate = vstate;
10742 
10743 	return (enab);
10744 }
10745 
10746 static void
10747 dtrace_enabling_add(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb)
10748 {
10749 	dtrace_ecbdesc_t **ndesc;
10750 	size_t osize, nsize;
10751 
10752 	/*
10753 	 * We can't add to enablings after we've enabled them, or after we've
10754 	 * retained them.
10755 	 */
10756 	ASSERT(enab->dten_probegen == 0);
10757 	ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
10758 
10759 	if (enab->dten_ndesc < enab->dten_maxdesc) {
10760 		enab->dten_desc[enab->dten_ndesc++] = ecb;
10761 		return;
10762 	}
10763 
10764 	osize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
10765 
10766 	if (enab->dten_maxdesc == 0) {
10767 		enab->dten_maxdesc = 1;
10768 	} else {
10769 		enab->dten_maxdesc <<= 1;
10770 	}
10771 
10772 	ASSERT(enab->dten_ndesc < enab->dten_maxdesc);
10773 
10774 	nsize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
10775 	ndesc = kmem_zalloc(nsize, KM_SLEEP);
10776 	bcopy(enab->dten_desc, ndesc, osize);
10777 	kmem_free(enab->dten_desc, osize);
10778 
10779 	enab->dten_desc = ndesc;
10780 	enab->dten_desc[enab->dten_ndesc++] = ecb;
10781 }
10782 
10783 static void
10784 dtrace_enabling_addlike(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb,
10785     dtrace_probedesc_t *pd)
10786 {
10787 	dtrace_ecbdesc_t *new;
10788 	dtrace_predicate_t *pred;
10789 	dtrace_actdesc_t *act;
10790 
10791 	/*
10792 	 * We're going to create a new ECB description that matches the
10793 	 * specified ECB in every way, but has the specified probe description.
10794 	 */
10795 	new = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
10796 
10797 	if ((pred = ecb->dted_pred.dtpdd_predicate) != NULL)
10798 		dtrace_predicate_hold(pred);
10799 
10800 	for (act = ecb->dted_action; act != NULL; act = act->dtad_next)
10801 		dtrace_actdesc_hold(act);
10802 
10803 	new->dted_action = ecb->dted_action;
10804 	new->dted_pred = ecb->dted_pred;
10805 	new->dted_probe = *pd;
10806 	new->dted_uarg = ecb->dted_uarg;
10807 
10808 	dtrace_enabling_add(enab, new);
10809 }
10810 
10811 static void
10812 dtrace_enabling_dump(dtrace_enabling_t *enab)
10813 {
10814 	int i;
10815 
10816 	for (i = 0; i < enab->dten_ndesc; i++) {
10817 		dtrace_probedesc_t *desc = &enab->dten_desc[i]->dted_probe;
10818 
10819 		cmn_err(CE_NOTE, "enabling probe %d (%s:%s:%s:%s)", i,
10820 		    desc->dtpd_provider, desc->dtpd_mod,
10821 		    desc->dtpd_func, desc->dtpd_name);
10822 	}
10823 }
10824 
10825 static void
10826 dtrace_enabling_destroy(dtrace_enabling_t *enab)
10827 {
10828 	int i;
10829 	dtrace_ecbdesc_t *ep;
10830 	dtrace_vstate_t *vstate = enab->dten_vstate;
10831 
10832 	ASSERT(MUTEX_HELD(&dtrace_lock));
10833 
10834 	for (i = 0; i < enab->dten_ndesc; i++) {
10835 		dtrace_actdesc_t *act, *next;
10836 		dtrace_predicate_t *pred;
10837 
10838 		ep = enab->dten_desc[i];
10839 
10840 		if ((pred = ep->dted_pred.dtpdd_predicate) != NULL)
10841 			dtrace_predicate_release(pred, vstate);
10842 
10843 		for (act = ep->dted_action; act != NULL; act = next) {
10844 			next = act->dtad_next;
10845 			dtrace_actdesc_release(act, vstate);
10846 		}
10847 
10848 		kmem_free(ep, sizeof (dtrace_ecbdesc_t));
10849 	}
10850 
10851 	kmem_free(enab->dten_desc,
10852 	    enab->dten_maxdesc * sizeof (dtrace_enabling_t *));
10853 
10854 	/*
10855 	 * If this was a retained enabling, decrement the dts_nretained count
10856 	 * and take it off of the dtrace_retained list.
10857 	 */
10858 	if (enab->dten_prev != NULL || enab->dten_next != NULL ||
10859 	    dtrace_retained == enab) {
10860 		ASSERT(enab->dten_vstate->dtvs_state != NULL);
10861 		ASSERT(enab->dten_vstate->dtvs_state->dts_nretained > 0);
10862 		enab->dten_vstate->dtvs_state->dts_nretained--;
10863 		dtrace_retained_gen++;
10864 	}
10865 
10866 	if (enab->dten_prev == NULL) {
10867 		if (dtrace_retained == enab) {
10868 			dtrace_retained = enab->dten_next;
10869 
10870 			if (dtrace_retained != NULL)
10871 				dtrace_retained->dten_prev = NULL;
10872 		}
10873 	} else {
10874 		ASSERT(enab != dtrace_retained);
10875 		ASSERT(dtrace_retained != NULL);
10876 		enab->dten_prev->dten_next = enab->dten_next;
10877 	}
10878 
10879 	if (enab->dten_next != NULL) {
10880 		ASSERT(dtrace_retained != NULL);
10881 		enab->dten_next->dten_prev = enab->dten_prev;
10882 	}
10883 
10884 	kmem_free(enab, sizeof (dtrace_enabling_t));
10885 }
10886 
10887 static int
10888 dtrace_enabling_retain(dtrace_enabling_t *enab)
10889 {
10890 	dtrace_state_t *state;
10891 
10892 	ASSERT(MUTEX_HELD(&dtrace_lock));
10893 	ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
10894 	ASSERT(enab->dten_vstate != NULL);
10895 
10896 	state = enab->dten_vstate->dtvs_state;
10897 	ASSERT(state != NULL);
10898 
10899 	/*
10900 	 * We only allow each state to retain dtrace_retain_max enablings.
10901 	 */
10902 	if (state->dts_nretained >= dtrace_retain_max)
10903 		return (ENOSPC);
10904 
10905 	state->dts_nretained++;
10906 	dtrace_retained_gen++;
10907 
10908 	if (dtrace_retained == NULL) {
10909 		dtrace_retained = enab;
10910 		return (0);
10911 	}
10912 
10913 	enab->dten_next = dtrace_retained;
10914 	dtrace_retained->dten_prev = enab;
10915 	dtrace_retained = enab;
10916 
10917 	return (0);
10918 }
10919 
10920 static int
10921 dtrace_enabling_replicate(dtrace_state_t *state, dtrace_probedesc_t *match,
10922     dtrace_probedesc_t *create)
10923 {
10924 	dtrace_enabling_t *new, *enab;
10925 	int found = 0, err = ENOENT;
10926 
10927 	ASSERT(MUTEX_HELD(&dtrace_lock));
10928 	ASSERT(strlen(match->dtpd_provider) < DTRACE_PROVNAMELEN);
10929 	ASSERT(strlen(match->dtpd_mod) < DTRACE_MODNAMELEN);
10930 	ASSERT(strlen(match->dtpd_func) < DTRACE_FUNCNAMELEN);
10931 	ASSERT(strlen(match->dtpd_name) < DTRACE_NAMELEN);
10932 
10933 	new = dtrace_enabling_create(&state->dts_vstate);
10934 
10935 	/*
10936 	 * Iterate over all retained enablings, looking for enablings that
10937 	 * match the specified state.
10938 	 */
10939 	for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
10940 		int i;
10941 
10942 		/*
10943 		 * dtvs_state can only be NULL for helper enablings -- and
10944 		 * helper enablings can't be retained.
10945 		 */
10946 		ASSERT(enab->dten_vstate->dtvs_state != NULL);
10947 
10948 		if (enab->dten_vstate->dtvs_state != state)
10949 			continue;
10950 
10951 		/*
10952 		 * Now iterate over each probe description; we're looking for
10953 		 * an exact match to the specified probe description.
10954 		 */
10955 		for (i = 0; i < enab->dten_ndesc; i++) {
10956 			dtrace_ecbdesc_t *ep = enab->dten_desc[i];
10957 			dtrace_probedesc_t *pd = &ep->dted_probe;
10958 
10959 			if (strcmp(pd->dtpd_provider, match->dtpd_provider))
10960 				continue;
10961 
10962 			if (strcmp(pd->dtpd_mod, match->dtpd_mod))
10963 				continue;
10964 
10965 			if (strcmp(pd->dtpd_func, match->dtpd_func))
10966 				continue;
10967 
10968 			if (strcmp(pd->dtpd_name, match->dtpd_name))
10969 				continue;
10970 
10971 			/*
10972 			 * We have a winning probe!  Add it to our growing
10973 			 * enabling.
10974 			 */
10975 			found = 1;
10976 			dtrace_enabling_addlike(new, ep, create);
10977 		}
10978 	}
10979 
10980 	if (!found || (err = dtrace_enabling_retain(new)) != 0) {
10981 		dtrace_enabling_destroy(new);
10982 		return (err);
10983 	}
10984 
10985 	return (0);
10986 }
10987 
10988 static void
10989 dtrace_enabling_retract(dtrace_state_t *state)
10990 {
10991 	dtrace_enabling_t *enab, *next;
10992 
10993 	ASSERT(MUTEX_HELD(&dtrace_lock));
10994 
10995 	/*
10996 	 * Iterate over all retained enablings, destroy the enablings retained
10997 	 * for the specified state.
10998 	 */
10999 	for (enab = dtrace_retained; enab != NULL; enab = next) {
11000 		next = enab->dten_next;
11001 
11002 		/*
11003 		 * dtvs_state can only be NULL for helper enablings -- and
11004 		 * helper enablings can't be retained.
11005 		 */
11006 		ASSERT(enab->dten_vstate->dtvs_state != NULL);
11007 
11008 		if (enab->dten_vstate->dtvs_state == state) {
11009 			ASSERT(state->dts_nretained > 0);
11010 			dtrace_enabling_destroy(enab);
11011 		}
11012 	}
11013 
11014 	ASSERT(state->dts_nretained == 0);
11015 }
11016 
11017 static int
11018 dtrace_enabling_match(dtrace_enabling_t *enab, int *nmatched)
11019 {
11020 	int i = 0;
11021 	int total_matched = 0, matched = 0;
11022 
11023 	ASSERT(MUTEX_HELD(&cpu_lock));
11024 	ASSERT(MUTEX_HELD(&dtrace_lock));
11025 
11026 	for (i = 0; i < enab->dten_ndesc; i++) {
11027 		dtrace_ecbdesc_t *ep = enab->dten_desc[i];
11028 
11029 		enab->dten_current = ep;
11030 		enab->dten_error = 0;
11031 
11032 		/*
11033 		 * If a provider failed to enable a probe then get out and
11034 		 * let the consumer know we failed.
11035 		 */
11036 		if ((matched = dtrace_probe_enable(&ep->dted_probe, enab)) < 0)
11037 			return (EBUSY);
11038 
11039 		total_matched += matched;
11040 
11041 		if (enab->dten_error != 0) {
11042 			/*
11043 			 * If we get an error half-way through enabling the
11044 			 * probes, we kick out -- perhaps with some number of
11045 			 * them enabled.  Leaving enabled probes enabled may
11046 			 * be slightly confusing for user-level, but we expect
11047 			 * that no one will attempt to actually drive on in
11048 			 * the face of such errors.  If this is an anonymous
11049 			 * enabling (indicated with a NULL nmatched pointer),
11050 			 * we cmn_err() a message.  We aren't expecting to
11051 			 * get such an error -- such as it can exist at all,
11052 			 * it would be a result of corrupted DOF in the driver
11053 			 * properties.
11054 			 */
11055 			if (nmatched == NULL) {
11056 				cmn_err(CE_WARN, "dtrace_enabling_match() "
11057 				    "error on %p: %d", (void *)ep,
11058 				    enab->dten_error);
11059 			}
11060 
11061 			return (enab->dten_error);
11062 		}
11063 	}
11064 
11065 	enab->dten_probegen = dtrace_probegen;
11066 	if (nmatched != NULL)
11067 		*nmatched = total_matched;
11068 
11069 	return (0);
11070 }
11071 
11072 static void
11073 dtrace_enabling_matchall(void)
11074 {
11075 	dtrace_enabling_t *enab;
11076 
11077 	mutex_enter(&cpu_lock);
11078 	mutex_enter(&dtrace_lock);
11079 
11080 	/*
11081 	 * Iterate over all retained enablings to see if any probes match
11082 	 * against them.  We only perform this operation on enablings for which
11083 	 * we have sufficient permissions by virtue of being in the global zone
11084 	 * or in the same zone as the DTrace client.  Because we can be called
11085 	 * after dtrace_detach() has been called, we cannot assert that there
11086 	 * are retained enablings.  We can safely load from dtrace_retained,
11087 	 * however:  the taskq_destroy() at the end of dtrace_detach() will
11088 	 * block pending our completion.
11089 	 */
11090 	for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11091 		dtrace_cred_t *dcr = &enab->dten_vstate->dtvs_state->dts_cred;
11092 		cred_t *cr = dcr->dcr_cred;
11093 		zoneid_t zone = cr != NULL ? crgetzoneid(cr) : 0;
11094 
11095 		if ((dcr->dcr_visible & DTRACE_CRV_ALLZONE) || (cr != NULL &&
11096 		    (zone == GLOBAL_ZONEID || getzoneid() == zone)))
11097 			(void) dtrace_enabling_match(enab, NULL);
11098 	}
11099 
11100 	mutex_exit(&dtrace_lock);
11101 	mutex_exit(&cpu_lock);
11102 }
11103 
11104 /*
11105  * If an enabling is to be enabled without having matched probes (that is, if
11106  * dtrace_state_go() is to be called on the underlying dtrace_state_t), the
11107  * enabling must be _primed_ by creating an ECB for every ECB description.
11108  * This must be done to assure that we know the number of speculations, the
11109  * number of aggregations, the minimum buffer size needed, etc. before we
11110  * transition out of DTRACE_ACTIVITY_INACTIVE.  To do this without actually
11111  * enabling any probes, we create ECBs for every ECB decription, but with a
11112  * NULL probe -- which is exactly what this function does.
11113  */
11114 static void
11115 dtrace_enabling_prime(dtrace_state_t *state)
11116 {
11117 	dtrace_enabling_t *enab;
11118 	int i;
11119 
11120 	for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11121 		ASSERT(enab->dten_vstate->dtvs_state != NULL);
11122 
11123 		if (enab->dten_vstate->dtvs_state != state)
11124 			continue;
11125 
11126 		/*
11127 		 * We don't want to prime an enabling more than once, lest
11128 		 * we allow a malicious user to induce resource exhaustion.
11129 		 * (The ECBs that result from priming an enabling aren't
11130 		 * leaked -- but they also aren't deallocated until the
11131 		 * consumer state is destroyed.)
11132 		 */
11133 		if (enab->dten_primed)
11134 			continue;
11135 
11136 		for (i = 0; i < enab->dten_ndesc; i++) {
11137 			enab->dten_current = enab->dten_desc[i];
11138 			(void) dtrace_probe_enable(NULL, enab);
11139 		}
11140 
11141 		enab->dten_primed = 1;
11142 	}
11143 }
11144 
11145 /*
11146  * Called to indicate that probes should be provided due to retained
11147  * enablings.  This is implemented in terms of dtrace_probe_provide(), but it
11148  * must take an initial lap through the enabling calling the dtps_provide()
11149  * entry point explicitly to allow for autocreated probes.
11150  */
11151 static void
11152 dtrace_enabling_provide(dtrace_provider_t *prv)
11153 {
11154 	int i, all = 0;
11155 	dtrace_probedesc_t desc;
11156 	dtrace_genid_t gen;
11157 
11158 	ASSERT(MUTEX_HELD(&dtrace_lock));
11159 	ASSERT(MUTEX_HELD(&dtrace_provider_lock));
11160 
11161 	if (prv == NULL) {
11162 		all = 1;
11163 		prv = dtrace_provider;
11164 	}
11165 
11166 	do {
11167 		dtrace_enabling_t *enab;
11168 		void *parg = prv->dtpv_arg;
11169 
11170 retry:
11171 		gen = dtrace_retained_gen;
11172 		for (enab = dtrace_retained; enab != NULL;
11173 		    enab = enab->dten_next) {
11174 			for (i = 0; i < enab->dten_ndesc; i++) {
11175 				desc = enab->dten_desc[i]->dted_probe;
11176 				mutex_exit(&dtrace_lock);
11177 				prv->dtpv_pops.dtps_provide(parg, &desc);
11178 				mutex_enter(&dtrace_lock);
11179 				/*
11180 				 * Process the retained enablings again if
11181 				 * they have changed while we weren't holding
11182 				 * dtrace_lock.
11183 				 */
11184 				if (gen != dtrace_retained_gen)
11185 					goto retry;
11186 			}
11187 		}
11188 	} while (all && (prv = prv->dtpv_next) != NULL);
11189 
11190 	mutex_exit(&dtrace_lock);
11191 	dtrace_probe_provide(NULL, all ? NULL : prv);
11192 	mutex_enter(&dtrace_lock);
11193 }
11194 
11195 /*
11196  * Called to reap ECBs that are attached to probes from defunct providers.
11197  */
11198 static void
11199 dtrace_enabling_reap(void)
11200 {
11201 	dtrace_provider_t *prov;
11202 	dtrace_probe_t *probe;
11203 	dtrace_ecb_t *ecb;
11204 	hrtime_t when;
11205 	int i;
11206 
11207 	mutex_enter(&cpu_lock);
11208 	mutex_enter(&dtrace_lock);
11209 
11210 	for (i = 0; i < dtrace_nprobes; i++) {
11211 		if ((probe = dtrace_probes[i]) == NULL)
11212 			continue;
11213 
11214 		if (probe->dtpr_ecb == NULL)
11215 			continue;
11216 
11217 		prov = probe->dtpr_provider;
11218 
11219 		if ((when = prov->dtpv_defunct) == 0)
11220 			continue;
11221 
11222 		/*
11223 		 * We have ECBs on a defunct provider:  we want to reap these
11224 		 * ECBs to allow the provider to unregister.  The destruction
11225 		 * of these ECBs must be done carefully:  if we destroy the ECB
11226 		 * and the consumer later wishes to consume an EPID that
11227 		 * corresponds to the destroyed ECB (and if the EPID metadata
11228 		 * has not been previously consumed), the consumer will abort
11229 		 * processing on the unknown EPID.  To reduce (but not, sadly,
11230 		 * eliminate) the possibility of this, we will only destroy an
11231 		 * ECB for a defunct provider if, for the state that
11232 		 * corresponds to the ECB:
11233 		 *
11234 		 *  (a)	There is no speculative tracing (which can effectively
11235 		 *	cache an EPID for an arbitrary amount of time).
11236 		 *
11237 		 *  (b)	The principal buffers have been switched twice since the
11238 		 *	provider became defunct.
11239 		 *
11240 		 *  (c)	The aggregation buffers are of zero size or have been
11241 		 *	switched twice since the provider became defunct.
11242 		 *
11243 		 * We use dts_speculates to determine (a) and call a function
11244 		 * (dtrace_buffer_consumed()) to determine (b) and (c).  Note
11245 		 * that as soon as we've been unable to destroy one of the ECBs
11246 		 * associated with the probe, we quit trying -- reaping is only
11247 		 * fruitful in as much as we can destroy all ECBs associated
11248 		 * with the defunct provider's probes.
11249 		 */
11250 		while ((ecb = probe->dtpr_ecb) != NULL) {
11251 			dtrace_state_t *state = ecb->dte_state;
11252 			dtrace_buffer_t *buf = state->dts_buffer;
11253 			dtrace_buffer_t *aggbuf = state->dts_aggbuffer;
11254 
11255 			if (state->dts_speculates)
11256 				break;
11257 
11258 			if (!dtrace_buffer_consumed(buf, when))
11259 				break;
11260 
11261 			if (!dtrace_buffer_consumed(aggbuf, when))
11262 				break;
11263 
11264 			dtrace_ecb_disable(ecb);
11265 			ASSERT(probe->dtpr_ecb != ecb);
11266 			dtrace_ecb_destroy(ecb);
11267 		}
11268 	}
11269 
11270 	mutex_exit(&dtrace_lock);
11271 	mutex_exit(&cpu_lock);
11272 }
11273 
11274 /*
11275  * DTrace DOF Functions
11276  */
11277 /*ARGSUSED*/
11278 static void
11279 dtrace_dof_error(dof_hdr_t *dof, const char *str)
11280 {
11281 	if (dtrace_err_verbose)
11282 		cmn_err(CE_WARN, "failed to process DOF: %s", str);
11283 
11284 #ifdef DTRACE_ERRDEBUG
11285 	dtrace_errdebug(str);
11286 #endif
11287 }
11288 
11289 /*
11290  * Create DOF out of a currently enabled state.  Right now, we only create
11291  * DOF containing the run-time options -- but this could be expanded to create
11292  * complete DOF representing the enabled state.
11293  */
11294 static dof_hdr_t *
11295 dtrace_dof_create(dtrace_state_t *state)
11296 {
11297 	dof_hdr_t *dof;
11298 	dof_sec_t *sec;
11299 	dof_optdesc_t *opt;
11300 	int i, len = sizeof (dof_hdr_t) +
11301 	    roundup(sizeof (dof_sec_t), sizeof (uint64_t)) +
11302 	    sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
11303 
11304 	ASSERT(MUTEX_HELD(&dtrace_lock));
11305 
11306 	dof = kmem_zalloc(len, KM_SLEEP);
11307 	dof->dofh_ident[DOF_ID_MAG0] = DOF_MAG_MAG0;
11308 	dof->dofh_ident[DOF_ID_MAG1] = DOF_MAG_MAG1;
11309 	dof->dofh_ident[DOF_ID_MAG2] = DOF_MAG_MAG2;
11310 	dof->dofh_ident[DOF_ID_MAG3] = DOF_MAG_MAG3;
11311 
11312 	dof->dofh_ident[DOF_ID_MODEL] = DOF_MODEL_NATIVE;
11313 	dof->dofh_ident[DOF_ID_ENCODING] = DOF_ENCODE_NATIVE;
11314 	dof->dofh_ident[DOF_ID_VERSION] = DOF_VERSION;
11315 	dof->dofh_ident[DOF_ID_DIFVERS] = DIF_VERSION;
11316 	dof->dofh_ident[DOF_ID_DIFIREG] = DIF_DIR_NREGS;
11317 	dof->dofh_ident[DOF_ID_DIFTREG] = DIF_DTR_NREGS;
11318 
11319 	dof->dofh_flags = 0;
11320 	dof->dofh_hdrsize = sizeof (dof_hdr_t);
11321 	dof->dofh_secsize = sizeof (dof_sec_t);
11322 	dof->dofh_secnum = 1;	/* only DOF_SECT_OPTDESC */
11323 	dof->dofh_secoff = sizeof (dof_hdr_t);
11324 	dof->dofh_loadsz = len;
11325 	dof->dofh_filesz = len;
11326 	dof->dofh_pad = 0;
11327 
11328 	/*
11329 	 * Fill in the option section header...
11330 	 */
11331 	sec = (dof_sec_t *)((uintptr_t)dof + sizeof (dof_hdr_t));
11332 	sec->dofs_type = DOF_SECT_OPTDESC;
11333 	sec->dofs_align = sizeof (uint64_t);
11334 	sec->dofs_flags = DOF_SECF_LOAD;
11335 	sec->dofs_entsize = sizeof (dof_optdesc_t);
11336 
11337 	opt = (dof_optdesc_t *)((uintptr_t)sec +
11338 	    roundup(sizeof (dof_sec_t), sizeof (uint64_t)));
11339 
11340 	sec->dofs_offset = (uintptr_t)opt - (uintptr_t)dof;
11341 	sec->dofs_size = sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
11342 
11343 	for (i = 0; i < DTRACEOPT_MAX; i++) {
11344 		opt[i].dofo_option = i;
11345 		opt[i].dofo_strtab = DOF_SECIDX_NONE;
11346 		opt[i].dofo_value = state->dts_options[i];
11347 	}
11348 
11349 	return (dof);
11350 }
11351 
11352 static dof_hdr_t *
11353 dtrace_dof_copyin(uintptr_t uarg, int *errp)
11354 {
11355 	dof_hdr_t hdr, *dof;
11356 
11357 	ASSERT(!MUTEX_HELD(&dtrace_lock));
11358 
11359 	/*
11360 	 * First, we're going to copyin() the sizeof (dof_hdr_t).
11361 	 */
11362 	if (copyin((void *)uarg, &hdr, sizeof (hdr)) != 0) {
11363 		dtrace_dof_error(NULL, "failed to copyin DOF header");
11364 		*errp = EFAULT;
11365 		return (NULL);
11366 	}
11367 
11368 	/*
11369 	 * Now we'll allocate the entire DOF and copy it in -- provided
11370 	 * that the length isn't outrageous.
11371 	 */
11372 	if (hdr.dofh_loadsz >= dtrace_dof_maxsize) {
11373 		dtrace_dof_error(&hdr, "load size exceeds maximum");
11374 		*errp = E2BIG;
11375 		return (NULL);
11376 	}
11377 
11378 	if (hdr.dofh_loadsz < sizeof (hdr)) {
11379 		dtrace_dof_error(&hdr, "invalid load size");
11380 		*errp = EINVAL;
11381 		return (NULL);
11382 	}
11383 
11384 	dof = kmem_alloc(hdr.dofh_loadsz, KM_SLEEP);
11385 
11386 	if (copyin((void *)uarg, dof, hdr.dofh_loadsz) != 0 ||
11387 	    dof->dofh_loadsz != hdr.dofh_loadsz) {
11388 		kmem_free(dof, hdr.dofh_loadsz);
11389 		*errp = EFAULT;
11390 		return (NULL);
11391 	}
11392 
11393 	return (dof);
11394 }
11395 
11396 static dof_hdr_t *
11397 dtrace_dof_property(const char *name)
11398 {
11399 	uchar_t *buf;
11400 	uint64_t loadsz;
11401 	unsigned int len, i;
11402 	dof_hdr_t *dof;
11403 
11404 	/*
11405 	 * Unfortunately, array of values in .conf files are always (and
11406 	 * only) interpreted to be integer arrays.  We must read our DOF
11407 	 * as an integer array, and then squeeze it into a byte array.
11408 	 */
11409 	if (ddi_prop_lookup_int_array(DDI_DEV_T_ANY, dtrace_devi, 0,
11410 	    (char *)name, (int **)&buf, &len) != DDI_PROP_SUCCESS)
11411 		return (NULL);
11412 
11413 	for (i = 0; i < len; i++)
11414 		buf[i] = (uchar_t)(((int *)buf)[i]);
11415 
11416 	if (len < sizeof (dof_hdr_t)) {
11417 		ddi_prop_free(buf);
11418 		dtrace_dof_error(NULL, "truncated header");
11419 		return (NULL);
11420 	}
11421 
11422 	if (len < (loadsz = ((dof_hdr_t *)buf)->dofh_loadsz)) {
11423 		ddi_prop_free(buf);
11424 		dtrace_dof_error(NULL, "truncated DOF");
11425 		return (NULL);
11426 	}
11427 
11428 	if (loadsz >= dtrace_dof_maxsize) {
11429 		ddi_prop_free(buf);
11430 		dtrace_dof_error(NULL, "oversized DOF");
11431 		return (NULL);
11432 	}
11433 
11434 	dof = kmem_alloc(loadsz, KM_SLEEP);
11435 	bcopy(buf, dof, loadsz);
11436 	ddi_prop_free(buf);
11437 
11438 	return (dof);
11439 }
11440 
11441 static void
11442 dtrace_dof_destroy(dof_hdr_t *dof)
11443 {
11444 	kmem_free(dof, dof->dofh_loadsz);
11445 }
11446 
11447 /*
11448  * Return the dof_sec_t pointer corresponding to a given section index.  If the
11449  * index is not valid, dtrace_dof_error() is called and NULL is returned.  If
11450  * a type other than DOF_SECT_NONE is specified, the header is checked against
11451  * this type and NULL is returned if the types do not match.
11452  */
11453 static dof_sec_t *
11454 dtrace_dof_sect(dof_hdr_t *dof, uint32_t type, dof_secidx_t i)
11455 {
11456 	dof_sec_t *sec = (dof_sec_t *)(uintptr_t)
11457 	    ((uintptr_t)dof + dof->dofh_secoff + i * dof->dofh_secsize);
11458 
11459 	if (i >= dof->dofh_secnum) {
11460 		dtrace_dof_error(dof, "referenced section index is invalid");
11461 		return (NULL);
11462 	}
11463 
11464 	if (!(sec->dofs_flags & DOF_SECF_LOAD)) {
11465 		dtrace_dof_error(dof, "referenced section is not loadable");
11466 		return (NULL);
11467 	}
11468 
11469 	if (type != DOF_SECT_NONE && type != sec->dofs_type) {
11470 		dtrace_dof_error(dof, "referenced section is the wrong type");
11471 		return (NULL);
11472 	}
11473 
11474 	return (sec);
11475 }
11476 
11477 static dtrace_probedesc_t *
11478 dtrace_dof_probedesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_probedesc_t *desc)
11479 {
11480 	dof_probedesc_t *probe;
11481 	dof_sec_t *strtab;
11482 	uintptr_t daddr = (uintptr_t)dof;
11483 	uintptr_t str;
11484 	size_t size;
11485 
11486 	if (sec->dofs_type != DOF_SECT_PROBEDESC) {
11487 		dtrace_dof_error(dof, "invalid probe section");
11488 		return (NULL);
11489 	}
11490 
11491 	if (sec->dofs_align != sizeof (dof_secidx_t)) {
11492 		dtrace_dof_error(dof, "bad alignment in probe description");
11493 		return (NULL);
11494 	}
11495 
11496 	if (sec->dofs_offset + sizeof (dof_probedesc_t) > dof->dofh_loadsz) {
11497 		dtrace_dof_error(dof, "truncated probe description");
11498 		return (NULL);
11499 	}
11500 
11501 	probe = (dof_probedesc_t *)(uintptr_t)(daddr + sec->dofs_offset);
11502 	strtab = dtrace_dof_sect(dof, DOF_SECT_STRTAB, probe->dofp_strtab);
11503 
11504 	if (strtab == NULL)
11505 		return (NULL);
11506 
11507 	str = daddr + strtab->dofs_offset;
11508 	size = strtab->dofs_size;
11509 
11510 	if (probe->dofp_provider >= strtab->dofs_size) {
11511 		dtrace_dof_error(dof, "corrupt probe provider");
11512 		return (NULL);
11513 	}
11514 
11515 	(void) strncpy(desc->dtpd_provider,
11516 	    (char *)(str + probe->dofp_provider),
11517 	    MIN(DTRACE_PROVNAMELEN - 1, size - probe->dofp_provider));
11518 
11519 	if (probe->dofp_mod >= strtab->dofs_size) {
11520 		dtrace_dof_error(dof, "corrupt probe module");
11521 		return (NULL);
11522 	}
11523 
11524 	(void) strncpy(desc->dtpd_mod, (char *)(str + probe->dofp_mod),
11525 	    MIN(DTRACE_MODNAMELEN - 1, size - probe->dofp_mod));
11526 
11527 	if (probe->dofp_func >= strtab->dofs_size) {
11528 		dtrace_dof_error(dof, "corrupt probe function");
11529 		return (NULL);
11530 	}
11531 
11532 	(void) strncpy(desc->dtpd_func, (char *)(str + probe->dofp_func),
11533 	    MIN(DTRACE_FUNCNAMELEN - 1, size - probe->dofp_func));
11534 
11535 	if (probe->dofp_name >= strtab->dofs_size) {
11536 		dtrace_dof_error(dof, "corrupt probe name");
11537 		return (NULL);
11538 	}
11539 
11540 	(void) strncpy(desc->dtpd_name, (char *)(str + probe->dofp_name),
11541 	    MIN(DTRACE_NAMELEN - 1, size - probe->dofp_name));
11542 
11543 	return (desc);
11544 }
11545 
11546 static dtrace_difo_t *
11547 dtrace_dof_difo(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
11548     cred_t *cr)
11549 {
11550 	dtrace_difo_t *dp;
11551 	size_t ttl = 0;
11552 	dof_difohdr_t *dofd;
11553 	uintptr_t daddr = (uintptr_t)dof;
11554 	size_t max = dtrace_difo_maxsize;
11555 	int i, l, n;
11556 
11557 	static const struct {
11558 		int section;
11559 		int bufoffs;
11560 		int lenoffs;
11561 		int entsize;
11562 		int align;
11563 		const char *msg;
11564 	} difo[] = {
11565 		{ DOF_SECT_DIF, offsetof(dtrace_difo_t, dtdo_buf),
11566 		offsetof(dtrace_difo_t, dtdo_len), sizeof (dif_instr_t),
11567 		sizeof (dif_instr_t), "multiple DIF sections" },
11568 
11569 		{ DOF_SECT_INTTAB, offsetof(dtrace_difo_t, dtdo_inttab),
11570 		offsetof(dtrace_difo_t, dtdo_intlen), sizeof (uint64_t),
11571 		sizeof (uint64_t), "multiple integer tables" },
11572 
11573 		{ DOF_SECT_STRTAB, offsetof(dtrace_difo_t, dtdo_strtab),
11574 		offsetof(dtrace_difo_t, dtdo_strlen), 0,
11575 		sizeof (char), "multiple string tables" },
11576 
11577 		{ DOF_SECT_VARTAB, offsetof(dtrace_difo_t, dtdo_vartab),
11578 		offsetof(dtrace_difo_t, dtdo_varlen), sizeof (dtrace_difv_t),
11579 		sizeof (uint_t), "multiple variable tables" },
11580 
11581 		{ DOF_SECT_NONE, 0, 0, 0, NULL }
11582 	};
11583 
11584 	if (sec->dofs_type != DOF_SECT_DIFOHDR) {
11585 		dtrace_dof_error(dof, "invalid DIFO header section");
11586 		return (NULL);
11587 	}
11588 
11589 	if (sec->dofs_align != sizeof (dof_secidx_t)) {
11590 		dtrace_dof_error(dof, "bad alignment in DIFO header");
11591 		return (NULL);
11592 	}
11593 
11594 	if (sec->dofs_size < sizeof (dof_difohdr_t) ||
11595 	    sec->dofs_size % sizeof (dof_secidx_t)) {
11596 		dtrace_dof_error(dof, "bad size in DIFO header");
11597 		return (NULL);
11598 	}
11599 
11600 	dofd = (dof_difohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
11601 	n = (sec->dofs_size - sizeof (*dofd)) / sizeof (dof_secidx_t) + 1;
11602 
11603 	dp = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
11604 	dp->dtdo_rtype = dofd->dofd_rtype;
11605 
11606 	for (l = 0; l < n; l++) {
11607 		dof_sec_t *subsec;
11608 		void **bufp;
11609 		uint32_t *lenp;
11610 
11611 		if ((subsec = dtrace_dof_sect(dof, DOF_SECT_NONE,
11612 		    dofd->dofd_links[l])) == NULL)
11613 			goto err; /* invalid section link */
11614 
11615 		if (ttl + subsec->dofs_size > max) {
11616 			dtrace_dof_error(dof, "exceeds maximum size");
11617 			goto err;
11618 		}
11619 
11620 		ttl += subsec->dofs_size;
11621 
11622 		for (i = 0; difo[i].section != DOF_SECT_NONE; i++) {
11623 			if (subsec->dofs_type != difo[i].section)
11624 				continue;
11625 
11626 			if (!(subsec->dofs_flags & DOF_SECF_LOAD)) {
11627 				dtrace_dof_error(dof, "section not loaded");
11628 				goto err;
11629 			}
11630 
11631 			if (subsec->dofs_align != difo[i].align) {
11632 				dtrace_dof_error(dof, "bad alignment");
11633 				goto err;
11634 			}
11635 
11636 			bufp = (void **)((uintptr_t)dp + difo[i].bufoffs);
11637 			lenp = (uint32_t *)((uintptr_t)dp + difo[i].lenoffs);
11638 
11639 			if (*bufp != NULL) {
11640 				dtrace_dof_error(dof, difo[i].msg);
11641 				goto err;
11642 			}
11643 
11644 			if (difo[i].entsize != subsec->dofs_entsize) {
11645 				dtrace_dof_error(dof, "entry size mismatch");
11646 				goto err;
11647 			}
11648 
11649 			if (subsec->dofs_entsize != 0 &&
11650 			    (subsec->dofs_size % subsec->dofs_entsize) != 0) {
11651 				dtrace_dof_error(dof, "corrupt entry size");
11652 				goto err;
11653 			}
11654 
11655 			*lenp = subsec->dofs_size;
11656 			*bufp = kmem_alloc(subsec->dofs_size, KM_SLEEP);
11657 			bcopy((char *)(uintptr_t)(daddr + subsec->dofs_offset),
11658 			    *bufp, subsec->dofs_size);
11659 
11660 			if (subsec->dofs_entsize != 0)
11661 				*lenp /= subsec->dofs_entsize;
11662 
11663 			break;
11664 		}
11665 
11666 		/*
11667 		 * If we encounter a loadable DIFO sub-section that is not
11668 		 * known to us, assume this is a broken program and fail.
11669 		 */
11670 		if (difo[i].section == DOF_SECT_NONE &&
11671 		    (subsec->dofs_flags & DOF_SECF_LOAD)) {
11672 			dtrace_dof_error(dof, "unrecognized DIFO subsection");
11673 			goto err;
11674 		}
11675 	}
11676 
11677 	if (dp->dtdo_buf == NULL) {
11678 		/*
11679 		 * We can't have a DIF object without DIF text.
11680 		 */
11681 		dtrace_dof_error(dof, "missing DIF text");
11682 		goto err;
11683 	}
11684 
11685 	/*
11686 	 * Before we validate the DIF object, run through the variable table
11687 	 * looking for the strings -- if any of their size are under, we'll set
11688 	 * their size to be the system-wide default string size.  Note that
11689 	 * this should _not_ happen if the "strsize" option has been set --
11690 	 * in this case, the compiler should have set the size to reflect the
11691 	 * setting of the option.
11692 	 */
11693 	for (i = 0; i < dp->dtdo_varlen; i++) {
11694 		dtrace_difv_t *v = &dp->dtdo_vartab[i];
11695 		dtrace_diftype_t *t = &v->dtdv_type;
11696 
11697 		if (v->dtdv_id < DIF_VAR_OTHER_UBASE)
11698 			continue;
11699 
11700 		if (t->dtdt_kind == DIF_TYPE_STRING && t->dtdt_size == 0)
11701 			t->dtdt_size = dtrace_strsize_default;
11702 	}
11703 
11704 	if (dtrace_difo_validate(dp, vstate, DIF_DIR_NREGS, cr) != 0)
11705 		goto err;
11706 
11707 	dtrace_difo_init(dp, vstate);
11708 	return (dp);
11709 
11710 err:
11711 	kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
11712 	kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
11713 	kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
11714 	kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
11715 
11716 	kmem_free(dp, sizeof (dtrace_difo_t));
11717 	return (NULL);
11718 }
11719 
11720 static dtrace_predicate_t *
11721 dtrace_dof_predicate(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
11722     cred_t *cr)
11723 {
11724 	dtrace_difo_t *dp;
11725 
11726 	if ((dp = dtrace_dof_difo(dof, sec, vstate, cr)) == NULL)
11727 		return (NULL);
11728 
11729 	return (dtrace_predicate_create(dp));
11730 }
11731 
11732 static dtrace_actdesc_t *
11733 dtrace_dof_actdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
11734     cred_t *cr)
11735 {
11736 	dtrace_actdesc_t *act, *first = NULL, *last = NULL, *next;
11737 	dof_actdesc_t *desc;
11738 	dof_sec_t *difosec;
11739 	size_t offs;
11740 	uintptr_t daddr = (uintptr_t)dof;
11741 	uint64_t arg;
11742 	dtrace_actkind_t kind;
11743 
11744 	if (sec->dofs_type != DOF_SECT_ACTDESC) {
11745 		dtrace_dof_error(dof, "invalid action section");
11746 		return (NULL);
11747 	}
11748 
11749 	if (sec->dofs_offset + sizeof (dof_actdesc_t) > dof->dofh_loadsz) {
11750 		dtrace_dof_error(dof, "truncated action description");
11751 		return (NULL);
11752 	}
11753 
11754 	if (sec->dofs_align != sizeof (uint64_t)) {
11755 		dtrace_dof_error(dof, "bad alignment in action description");
11756 		return (NULL);
11757 	}
11758 
11759 	if (sec->dofs_size < sec->dofs_entsize) {
11760 		dtrace_dof_error(dof, "section entry size exceeds total size");
11761 		return (NULL);
11762 	}
11763 
11764 	if (sec->dofs_entsize != sizeof (dof_actdesc_t)) {
11765 		dtrace_dof_error(dof, "bad entry size in action description");
11766 		return (NULL);
11767 	}
11768 
11769 	if (sec->dofs_size / sec->dofs_entsize > dtrace_actions_max) {
11770 		dtrace_dof_error(dof, "actions exceed dtrace_actions_max");
11771 		return (NULL);
11772 	}
11773 
11774 	for (offs = 0; offs < sec->dofs_size; offs += sec->dofs_entsize) {
11775 		desc = (dof_actdesc_t *)(daddr +
11776 		    (uintptr_t)sec->dofs_offset + offs);
11777 		kind = (dtrace_actkind_t)desc->dofa_kind;
11778 
11779 		if (DTRACEACT_ISPRINTFLIKE(kind) &&
11780 		    (kind != DTRACEACT_PRINTA ||
11781 		    desc->dofa_strtab != DOF_SECIDX_NONE)) {
11782 			dof_sec_t *strtab;
11783 			char *str, *fmt;
11784 			uint64_t i;
11785 
11786 			/*
11787 			 * printf()-like actions must have a format string.
11788 			 */
11789 			if ((strtab = dtrace_dof_sect(dof,
11790 			    DOF_SECT_STRTAB, desc->dofa_strtab)) == NULL)
11791 				goto err;
11792 
11793 			str = (char *)((uintptr_t)dof +
11794 			    (uintptr_t)strtab->dofs_offset);
11795 
11796 			for (i = desc->dofa_arg; i < strtab->dofs_size; i++) {
11797 				if (str[i] == '\0')
11798 					break;
11799 			}
11800 
11801 			if (i >= strtab->dofs_size) {
11802 				dtrace_dof_error(dof, "bogus format string");
11803 				goto err;
11804 			}
11805 
11806 			if (i == desc->dofa_arg) {
11807 				dtrace_dof_error(dof, "empty format string");
11808 				goto err;
11809 			}
11810 
11811 			i -= desc->dofa_arg;
11812 			fmt = kmem_alloc(i + 1, KM_SLEEP);
11813 			bcopy(&str[desc->dofa_arg], fmt, i + 1);
11814 			arg = (uint64_t)(uintptr_t)fmt;
11815 		} else {
11816 			if (kind == DTRACEACT_PRINTA) {
11817 				ASSERT(desc->dofa_strtab == DOF_SECIDX_NONE);
11818 				arg = 0;
11819 			} else {
11820 				arg = desc->dofa_arg;
11821 			}
11822 		}
11823 
11824 		act = dtrace_actdesc_create(kind, desc->dofa_ntuple,
11825 		    desc->dofa_uarg, arg);
11826 
11827 		if (last != NULL) {
11828 			last->dtad_next = act;
11829 		} else {
11830 			first = act;
11831 		}
11832 
11833 		last = act;
11834 
11835 		if (desc->dofa_difo == DOF_SECIDX_NONE)
11836 			continue;
11837 
11838 		if ((difosec = dtrace_dof_sect(dof,
11839 		    DOF_SECT_DIFOHDR, desc->dofa_difo)) == NULL)
11840 			goto err;
11841 
11842 		act->dtad_difo = dtrace_dof_difo(dof, difosec, vstate, cr);
11843 
11844 		if (act->dtad_difo == NULL)
11845 			goto err;
11846 	}
11847 
11848 	ASSERT(first != NULL);
11849 	return (first);
11850 
11851 err:
11852 	for (act = first; act != NULL; act = next) {
11853 		next = act->dtad_next;
11854 		dtrace_actdesc_release(act, vstate);
11855 	}
11856 
11857 	return (NULL);
11858 }
11859 
11860 static dtrace_ecbdesc_t *
11861 dtrace_dof_ecbdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
11862     cred_t *cr)
11863 {
11864 	dtrace_ecbdesc_t *ep;
11865 	dof_ecbdesc_t *ecb;
11866 	dtrace_probedesc_t *desc;
11867 	dtrace_predicate_t *pred = NULL;
11868 
11869 	if (sec->dofs_size < sizeof (dof_ecbdesc_t)) {
11870 		dtrace_dof_error(dof, "truncated ECB description");
11871 		return (NULL);
11872 	}
11873 
11874 	if (sec->dofs_align != sizeof (uint64_t)) {
11875 		dtrace_dof_error(dof, "bad alignment in ECB description");
11876 		return (NULL);
11877 	}
11878 
11879 	ecb = (dof_ecbdesc_t *)((uintptr_t)dof + (uintptr_t)sec->dofs_offset);
11880 	sec = dtrace_dof_sect(dof, DOF_SECT_PROBEDESC, ecb->dofe_probes);
11881 
11882 	if (sec == NULL)
11883 		return (NULL);
11884 
11885 	ep = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
11886 	ep->dted_uarg = ecb->dofe_uarg;
11887 	desc = &ep->dted_probe;
11888 
11889 	if (dtrace_dof_probedesc(dof, sec, desc) == NULL)
11890 		goto err;
11891 
11892 	if (ecb->dofe_pred != DOF_SECIDX_NONE) {
11893 		if ((sec = dtrace_dof_sect(dof,
11894 		    DOF_SECT_DIFOHDR, ecb->dofe_pred)) == NULL)
11895 			goto err;
11896 
11897 		if ((pred = dtrace_dof_predicate(dof, sec, vstate, cr)) == NULL)
11898 			goto err;
11899 
11900 		ep->dted_pred.dtpdd_predicate = pred;
11901 	}
11902 
11903 	if (ecb->dofe_actions != DOF_SECIDX_NONE) {
11904 		if ((sec = dtrace_dof_sect(dof,
11905 		    DOF_SECT_ACTDESC, ecb->dofe_actions)) == NULL)
11906 			goto err;
11907 
11908 		ep->dted_action = dtrace_dof_actdesc(dof, sec, vstate, cr);
11909 
11910 		if (ep->dted_action == NULL)
11911 			goto err;
11912 	}
11913 
11914 	return (ep);
11915 
11916 err:
11917 	if (pred != NULL)
11918 		dtrace_predicate_release(pred, vstate);
11919 	kmem_free(ep, sizeof (dtrace_ecbdesc_t));
11920 	return (NULL);
11921 }
11922 
11923 /*
11924  * Apply the relocations from the specified 'sec' (a DOF_SECT_URELHDR) to the
11925  * specified DOF.  At present, this amounts to simply adding 'ubase' to the
11926  * site of any user SETX relocations to account for load object base address.
11927  * In the future, if we need other relocations, this function can be extended.
11928  */
11929 static int
11930 dtrace_dof_relocate(dof_hdr_t *dof, dof_sec_t *sec, uint64_t ubase)
11931 {
11932 	uintptr_t daddr = (uintptr_t)dof;
11933 	dof_relohdr_t *dofr =
11934 	    (dof_relohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
11935 	dof_sec_t *ss, *rs, *ts;
11936 	dof_relodesc_t *r;
11937 	uint_t i, n;
11938 
11939 	if (sec->dofs_size < sizeof (dof_relohdr_t) ||
11940 	    sec->dofs_align != sizeof (dof_secidx_t)) {
11941 		dtrace_dof_error(dof, "invalid relocation header");
11942 		return (-1);
11943 	}
11944 
11945 	ss = dtrace_dof_sect(dof, DOF_SECT_STRTAB, dofr->dofr_strtab);
11946 	rs = dtrace_dof_sect(dof, DOF_SECT_RELTAB, dofr->dofr_relsec);
11947 	ts = dtrace_dof_sect(dof, DOF_SECT_NONE, dofr->dofr_tgtsec);
11948 
11949 	if (ss == NULL || rs == NULL || ts == NULL)
11950 		return (-1); /* dtrace_dof_error() has been called already */
11951 
11952 	if (rs->dofs_entsize < sizeof (dof_relodesc_t) ||
11953 	    rs->dofs_align != sizeof (uint64_t)) {
11954 		dtrace_dof_error(dof, "invalid relocation section");
11955 		return (-1);
11956 	}
11957 
11958 	r = (dof_relodesc_t *)(uintptr_t)(daddr + rs->dofs_offset);
11959 	n = rs->dofs_size / rs->dofs_entsize;
11960 
11961 	for (i = 0; i < n; i++) {
11962 		uintptr_t taddr = daddr + ts->dofs_offset + r->dofr_offset;
11963 
11964 		switch (r->dofr_type) {
11965 		case DOF_RELO_NONE:
11966 			break;
11967 		case DOF_RELO_SETX:
11968 			if (r->dofr_offset >= ts->dofs_size || r->dofr_offset +
11969 			    sizeof (uint64_t) > ts->dofs_size) {
11970 				dtrace_dof_error(dof, "bad relocation offset");
11971 				return (-1);
11972 			}
11973 
11974 			if (!IS_P2ALIGNED(taddr, sizeof (uint64_t))) {
11975 				dtrace_dof_error(dof, "misaligned setx relo");
11976 				return (-1);
11977 			}
11978 
11979 			*(uint64_t *)taddr += ubase;
11980 			break;
11981 		default:
11982 			dtrace_dof_error(dof, "invalid relocation type");
11983 			return (-1);
11984 		}
11985 
11986 		r = (dof_relodesc_t *)((uintptr_t)r + rs->dofs_entsize);
11987 	}
11988 
11989 	return (0);
11990 }
11991 
11992 /*
11993  * The dof_hdr_t passed to dtrace_dof_slurp() should be a partially validated
11994  * header:  it should be at the front of a memory region that is at least
11995  * sizeof (dof_hdr_t) in size -- and then at least dof_hdr.dofh_loadsz in
11996  * size.  It need not be validated in any other way.
11997  */
11998 static int
11999 dtrace_dof_slurp(dof_hdr_t *dof, dtrace_vstate_t *vstate, cred_t *cr,
12000     dtrace_enabling_t **enabp, uint64_t ubase, int noprobes)
12001 {
12002 	uint64_t len = dof->dofh_loadsz, seclen;
12003 	uintptr_t daddr = (uintptr_t)dof;
12004 	dtrace_ecbdesc_t *ep;
12005 	dtrace_enabling_t *enab;
12006 	uint_t i;
12007 
12008 	ASSERT(MUTEX_HELD(&dtrace_lock));
12009 	ASSERT(dof->dofh_loadsz >= sizeof (dof_hdr_t));
12010 
12011 	/*
12012 	 * Check the DOF header identification bytes.  In addition to checking
12013 	 * valid settings, we also verify that unused bits/bytes are zeroed so
12014 	 * we can use them later without fear of regressing existing binaries.
12015 	 */
12016 	if (bcmp(&dof->dofh_ident[DOF_ID_MAG0],
12017 	    DOF_MAG_STRING, DOF_MAG_STRLEN) != 0) {
12018 		dtrace_dof_error(dof, "DOF magic string mismatch");
12019 		return (-1);
12020 	}
12021 
12022 	if (dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_ILP32 &&
12023 	    dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_LP64) {
12024 		dtrace_dof_error(dof, "DOF has invalid data model");
12025 		return (-1);
12026 	}
12027 
12028 	if (dof->dofh_ident[DOF_ID_ENCODING] != DOF_ENCODE_NATIVE) {
12029 		dtrace_dof_error(dof, "DOF encoding mismatch");
12030 		return (-1);
12031 	}
12032 
12033 	if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
12034 	    dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_2) {
12035 		dtrace_dof_error(dof, "DOF version mismatch");
12036 		return (-1);
12037 	}
12038 
12039 	if (dof->dofh_ident[DOF_ID_DIFVERS] != DIF_VERSION_2) {
12040 		dtrace_dof_error(dof, "DOF uses unsupported instruction set");
12041 		return (-1);
12042 	}
12043 
12044 	if (dof->dofh_ident[DOF_ID_DIFIREG] > DIF_DIR_NREGS) {
12045 		dtrace_dof_error(dof, "DOF uses too many integer registers");
12046 		return (-1);
12047 	}
12048 
12049 	if (dof->dofh_ident[DOF_ID_DIFTREG] > DIF_DTR_NREGS) {
12050 		dtrace_dof_error(dof, "DOF uses too many tuple registers");
12051 		return (-1);
12052 	}
12053 
12054 	for (i = DOF_ID_PAD; i < DOF_ID_SIZE; i++) {
12055 		if (dof->dofh_ident[i] != 0) {
12056 			dtrace_dof_error(dof, "DOF has invalid ident byte set");
12057 			return (-1);
12058 		}
12059 	}
12060 
12061 	if (dof->dofh_flags & ~DOF_FL_VALID) {
12062 		dtrace_dof_error(dof, "DOF has invalid flag bits set");
12063 		return (-1);
12064 	}
12065 
12066 	if (dof->dofh_secsize == 0) {
12067 		dtrace_dof_error(dof, "zero section header size");
12068 		return (-1);
12069 	}
12070 
12071 	/*
12072 	 * Check that the section headers don't exceed the amount of DOF
12073 	 * data.  Note that we cast the section size and number of sections
12074 	 * to uint64_t's to prevent possible overflow in the multiplication.
12075 	 */
12076 	seclen = (uint64_t)dof->dofh_secnum * (uint64_t)dof->dofh_secsize;
12077 
12078 	if (dof->dofh_secoff > len || seclen > len ||
12079 	    dof->dofh_secoff + seclen > len) {
12080 		dtrace_dof_error(dof, "truncated section headers");
12081 		return (-1);
12082 	}
12083 
12084 	if (!IS_P2ALIGNED(dof->dofh_secoff, sizeof (uint64_t))) {
12085 		dtrace_dof_error(dof, "misaligned section headers");
12086 		return (-1);
12087 	}
12088 
12089 	if (!IS_P2ALIGNED(dof->dofh_secsize, sizeof (uint64_t))) {
12090 		dtrace_dof_error(dof, "misaligned section size");
12091 		return (-1);
12092 	}
12093 
12094 	/*
12095 	 * Take an initial pass through the section headers to be sure that
12096 	 * the headers don't have stray offsets.  If the 'noprobes' flag is
12097 	 * set, do not permit sections relating to providers, probes, or args.
12098 	 */
12099 	for (i = 0; i < dof->dofh_secnum; i++) {
12100 		dof_sec_t *sec = (dof_sec_t *)(daddr +
12101 		    (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12102 
12103 		if (noprobes) {
12104 			switch (sec->dofs_type) {
12105 			case DOF_SECT_PROVIDER:
12106 			case DOF_SECT_PROBES:
12107 			case DOF_SECT_PRARGS:
12108 			case DOF_SECT_PROFFS:
12109 				dtrace_dof_error(dof, "illegal sections "
12110 				    "for enabling");
12111 				return (-1);
12112 			}
12113 		}
12114 
12115 		if (DOF_SEC_ISLOADABLE(sec->dofs_type) &&
12116 		    !(sec->dofs_flags & DOF_SECF_LOAD)) {
12117 			dtrace_dof_error(dof, "loadable section with load "
12118 			    "flag unset");
12119 			return (-1);
12120 		}
12121 
12122 		if (!(sec->dofs_flags & DOF_SECF_LOAD))
12123 			continue; /* just ignore non-loadable sections */
12124 
12125 		if (sec->dofs_align & (sec->dofs_align - 1)) {
12126 			dtrace_dof_error(dof, "bad section alignment");
12127 			return (-1);
12128 		}
12129 
12130 		if (sec->dofs_offset & (sec->dofs_align - 1)) {
12131 			dtrace_dof_error(dof, "misaligned section");
12132 			return (-1);
12133 		}
12134 
12135 		if (sec->dofs_offset > len || sec->dofs_size > len ||
12136 		    sec->dofs_offset + sec->dofs_size > len) {
12137 			dtrace_dof_error(dof, "corrupt section header");
12138 			return (-1);
12139 		}
12140 
12141 		if (sec->dofs_type == DOF_SECT_STRTAB && *((char *)daddr +
12142 		    sec->dofs_offset + sec->dofs_size - 1) != '\0') {
12143 			dtrace_dof_error(dof, "non-terminating string table");
12144 			return (-1);
12145 		}
12146 	}
12147 
12148 	/*
12149 	 * Take a second pass through the sections and locate and perform any
12150 	 * relocations that are present.  We do this after the first pass to
12151 	 * be sure that all sections have had their headers validated.
12152 	 */
12153 	for (i = 0; i < dof->dofh_secnum; i++) {
12154 		dof_sec_t *sec = (dof_sec_t *)(daddr +
12155 		    (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12156 
12157 		if (!(sec->dofs_flags & DOF_SECF_LOAD))
12158 			continue; /* skip sections that are not loadable */
12159 
12160 		switch (sec->dofs_type) {
12161 		case DOF_SECT_URELHDR:
12162 			if (dtrace_dof_relocate(dof, sec, ubase) != 0)
12163 				return (-1);
12164 			break;
12165 		}
12166 	}
12167 
12168 	if ((enab = *enabp) == NULL)
12169 		enab = *enabp = dtrace_enabling_create(vstate);
12170 
12171 	for (i = 0; i < dof->dofh_secnum; i++) {
12172 		dof_sec_t *sec = (dof_sec_t *)(daddr +
12173 		    (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12174 
12175 		if (sec->dofs_type != DOF_SECT_ECBDESC)
12176 			continue;
12177 
12178 		if ((ep = dtrace_dof_ecbdesc(dof, sec, vstate, cr)) == NULL) {
12179 			dtrace_enabling_destroy(enab);
12180 			*enabp = NULL;
12181 			return (-1);
12182 		}
12183 
12184 		dtrace_enabling_add(enab, ep);
12185 	}
12186 
12187 	return (0);
12188 }
12189 
12190 /*
12191  * Process DOF for any options.  This routine assumes that the DOF has been
12192  * at least processed by dtrace_dof_slurp().
12193  */
12194 static int
12195 dtrace_dof_options(dof_hdr_t *dof, dtrace_state_t *state)
12196 {
12197 	int i, rval;
12198 	uint32_t entsize;
12199 	size_t offs;
12200 	dof_optdesc_t *desc;
12201 
12202 	for (i = 0; i < dof->dofh_secnum; i++) {
12203 		dof_sec_t *sec = (dof_sec_t *)((uintptr_t)dof +
12204 		    (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12205 
12206 		if (sec->dofs_type != DOF_SECT_OPTDESC)
12207 			continue;
12208 
12209 		if (sec->dofs_align != sizeof (uint64_t)) {
12210 			dtrace_dof_error(dof, "bad alignment in "
12211 			    "option description");
12212 			return (EINVAL);
12213 		}
12214 
12215 		if ((entsize = sec->dofs_entsize) == 0) {
12216 			dtrace_dof_error(dof, "zeroed option entry size");
12217 			return (EINVAL);
12218 		}
12219 
12220 		if (entsize < sizeof (dof_optdesc_t)) {
12221 			dtrace_dof_error(dof, "bad option entry size");
12222 			return (EINVAL);
12223 		}
12224 
12225 		for (offs = 0; offs < sec->dofs_size; offs += entsize) {
12226 			desc = (dof_optdesc_t *)((uintptr_t)dof +
12227 			    (uintptr_t)sec->dofs_offset + offs);
12228 
12229 			if (desc->dofo_strtab != DOF_SECIDX_NONE) {
12230 				dtrace_dof_error(dof, "non-zero option string");
12231 				return (EINVAL);
12232 			}
12233 
12234 			if (desc->dofo_value == DTRACEOPT_UNSET) {
12235 				dtrace_dof_error(dof, "unset option");
12236 				return (EINVAL);
12237 			}
12238 
12239 			if ((rval = dtrace_state_option(state,
12240 			    desc->dofo_option, desc->dofo_value)) != 0) {
12241 				dtrace_dof_error(dof, "rejected option");
12242 				return (rval);
12243 			}
12244 		}
12245 	}
12246 
12247 	return (0);
12248 }
12249 
12250 /*
12251  * DTrace Consumer State Functions
12252  */
12253 int
12254 dtrace_dstate_init(dtrace_dstate_t *dstate, size_t size)
12255 {
12256 	size_t hashsize, maxper, min, chunksize = dstate->dtds_chunksize;
12257 	void *base;
12258 	uintptr_t limit;
12259 	dtrace_dynvar_t *dvar, *next, *start;
12260 	int i;
12261 
12262 	ASSERT(MUTEX_HELD(&dtrace_lock));
12263 	ASSERT(dstate->dtds_base == NULL && dstate->dtds_percpu == NULL);
12264 
12265 	bzero(dstate, sizeof (dtrace_dstate_t));
12266 
12267 	if ((dstate->dtds_chunksize = chunksize) == 0)
12268 		dstate->dtds_chunksize = DTRACE_DYNVAR_CHUNKSIZE;
12269 
12270 	if (size < (min = dstate->dtds_chunksize + sizeof (dtrace_dynhash_t)))
12271 		size = min;
12272 
12273 	if ((base = kmem_zalloc(size, KM_NOSLEEP | KM_NORMALPRI)) == NULL)
12274 		return (ENOMEM);
12275 
12276 	dstate->dtds_size = size;
12277 	dstate->dtds_base = base;
12278 	dstate->dtds_percpu = kmem_cache_alloc(dtrace_state_cache, KM_SLEEP);
12279 	bzero(dstate->dtds_percpu, NCPU * sizeof (dtrace_dstate_percpu_t));
12280 
12281 	hashsize = size / (dstate->dtds_chunksize + sizeof (dtrace_dynhash_t));
12282 
12283 	if (hashsize != 1 && (hashsize & 1))
12284 		hashsize--;
12285 
12286 	dstate->dtds_hashsize = hashsize;
12287 	dstate->dtds_hash = dstate->dtds_base;
12288 
12289 	/*
12290 	 * Set all of our hash buckets to point to the single sink, and (if
12291 	 * it hasn't already been set), set the sink's hash value to be the
12292 	 * sink sentinel value.  The sink is needed for dynamic variable
12293 	 * lookups to know that they have iterated over an entire, valid hash
12294 	 * chain.
12295 	 */
12296 	for (i = 0; i < hashsize; i++)
12297 		dstate->dtds_hash[i].dtdh_chain = &dtrace_dynhash_sink;
12298 
12299 	if (dtrace_dynhash_sink.dtdv_hashval != DTRACE_DYNHASH_SINK)
12300 		dtrace_dynhash_sink.dtdv_hashval = DTRACE_DYNHASH_SINK;
12301 
12302 	/*
12303 	 * Determine number of active CPUs.  Divide free list evenly among
12304 	 * active CPUs.
12305 	 */
12306 	start = (dtrace_dynvar_t *)
12307 	    ((uintptr_t)base + hashsize * sizeof (dtrace_dynhash_t));
12308 	limit = (uintptr_t)base + size;
12309 
12310 	maxper = (limit - (uintptr_t)start) / NCPU;
12311 	maxper = (maxper / dstate->dtds_chunksize) * dstate->dtds_chunksize;
12312 
12313 	for (i = 0; i < NCPU; i++) {
12314 		dstate->dtds_percpu[i].dtdsc_free = dvar = start;
12315 
12316 		/*
12317 		 * If we don't even have enough chunks to make it once through
12318 		 * NCPUs, we're just going to allocate everything to the first
12319 		 * CPU.  And if we're on the last CPU, we're going to allocate
12320 		 * whatever is left over.  In either case, we set the limit to
12321 		 * be the limit of the dynamic variable space.
12322 		 */
12323 		if (maxper == 0 || i == NCPU - 1) {
12324 			limit = (uintptr_t)base + size;
12325 			start = NULL;
12326 		} else {
12327 			limit = (uintptr_t)start + maxper;
12328 			start = (dtrace_dynvar_t *)limit;
12329 		}
12330 
12331 		ASSERT(limit <= (uintptr_t)base + size);
12332 
12333 		for (;;) {
12334 			next = (dtrace_dynvar_t *)((uintptr_t)dvar +
12335 			    dstate->dtds_chunksize);
12336 
12337 			if ((uintptr_t)next + dstate->dtds_chunksize >= limit)
12338 				break;
12339 
12340 			dvar->dtdv_next = next;
12341 			dvar = next;
12342 		}
12343 
12344 		if (maxper == 0)
12345 			break;
12346 	}
12347 
12348 	return (0);
12349 }
12350 
12351 void
12352 dtrace_dstate_fini(dtrace_dstate_t *dstate)
12353 {
12354 	ASSERT(MUTEX_HELD(&cpu_lock));
12355 
12356 	if (dstate->dtds_base == NULL)
12357 		return;
12358 
12359 	kmem_free(dstate->dtds_base, dstate->dtds_size);
12360 	kmem_cache_free(dtrace_state_cache, dstate->dtds_percpu);
12361 }
12362 
12363 static void
12364 dtrace_vstate_fini(dtrace_vstate_t *vstate)
12365 {
12366 	/*
12367 	 * Logical XOR, where are you?
12368 	 */
12369 	ASSERT((vstate->dtvs_nglobals == 0) ^ (vstate->dtvs_globals != NULL));
12370 
12371 	if (vstate->dtvs_nglobals > 0) {
12372 		kmem_free(vstate->dtvs_globals, vstate->dtvs_nglobals *
12373 		    sizeof (dtrace_statvar_t *));
12374 	}
12375 
12376 	if (vstate->dtvs_ntlocals > 0) {
12377 		kmem_free(vstate->dtvs_tlocals, vstate->dtvs_ntlocals *
12378 		    sizeof (dtrace_difv_t));
12379 	}
12380 
12381 	ASSERT((vstate->dtvs_nlocals == 0) ^ (vstate->dtvs_locals != NULL));
12382 
12383 	if (vstate->dtvs_nlocals > 0) {
12384 		kmem_free(vstate->dtvs_locals, vstate->dtvs_nlocals *
12385 		    sizeof (dtrace_statvar_t *));
12386 	}
12387 }
12388 
12389 static void
12390 dtrace_state_clean(dtrace_state_t *state)
12391 {
12392 	if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE)
12393 		return;
12394 
12395 	dtrace_dynvar_clean(&state->dts_vstate.dtvs_dynvars);
12396 	dtrace_speculation_clean(state);
12397 }
12398 
12399 static void
12400 dtrace_state_deadman(dtrace_state_t *state)
12401 {
12402 	hrtime_t now;
12403 
12404 	dtrace_sync();
12405 
12406 	now = dtrace_gethrtime();
12407 
12408 	if (state != dtrace_anon.dta_state &&
12409 	    now - state->dts_laststatus >= dtrace_deadman_user)
12410 		return;
12411 
12412 	/*
12413 	 * We must be sure that dts_alive never appears to be less than the
12414 	 * value upon entry to dtrace_state_deadman(), and because we lack a
12415 	 * dtrace_cas64(), we cannot store to it atomically.  We thus instead
12416 	 * store INT64_MAX to it, followed by a memory barrier, followed by
12417 	 * the new value.  This assures that dts_alive never appears to be
12418 	 * less than its true value, regardless of the order in which the
12419 	 * stores to the underlying storage are issued.
12420 	 */
12421 	state->dts_alive = INT64_MAX;
12422 	dtrace_membar_producer();
12423 	state->dts_alive = now;
12424 }
12425 
12426 dtrace_state_t *
12427 dtrace_state_create(dev_t *devp, cred_t *cr)
12428 {
12429 	minor_t minor;
12430 	major_t major;
12431 	char c[30];
12432 	dtrace_state_t *state;
12433 	dtrace_optval_t *opt;
12434 	int bufsize = NCPU * sizeof (dtrace_buffer_t), i;
12435 
12436 	ASSERT(MUTEX_HELD(&dtrace_lock));
12437 	ASSERT(MUTEX_HELD(&cpu_lock));
12438 
12439 	minor = (minor_t)(uintptr_t)vmem_alloc(dtrace_minor, 1,
12440 	    VM_BESTFIT | VM_SLEEP);
12441 
12442 	if (ddi_soft_state_zalloc(dtrace_softstate, minor) != DDI_SUCCESS) {
12443 		vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
12444 		return (NULL);
12445 	}
12446 
12447 	state = ddi_get_soft_state(dtrace_softstate, minor);
12448 	state->dts_epid = DTRACE_EPIDNONE + 1;
12449 
12450 	(void) snprintf(c, sizeof (c), "dtrace_aggid_%d", minor);
12451 	state->dts_aggid_arena = vmem_create(c, (void *)1, UINT32_MAX, 1,
12452 	    NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
12453 
12454 	if (devp != NULL) {
12455 		major = getemajor(*devp);
12456 	} else {
12457 		major = ddi_driver_major(dtrace_devi);
12458 	}
12459 
12460 	state->dts_dev = makedevice(major, minor);
12461 
12462 	if (devp != NULL)
12463 		*devp = state->dts_dev;
12464 
12465 	/*
12466 	 * We allocate NCPU buffers.  On the one hand, this can be quite
12467 	 * a bit of memory per instance (nearly 36K on a Starcat).  On the
12468 	 * other hand, it saves an additional memory reference in the probe
12469 	 * path.
12470 	 */
12471 	state->dts_buffer = kmem_zalloc(bufsize, KM_SLEEP);
12472 	state->dts_aggbuffer = kmem_zalloc(bufsize, KM_SLEEP);
12473 	state->dts_cleaner = CYCLIC_NONE;
12474 	state->dts_deadman = CYCLIC_NONE;
12475 	state->dts_vstate.dtvs_state = state;
12476 
12477 	for (i = 0; i < DTRACEOPT_MAX; i++)
12478 		state->dts_options[i] = DTRACEOPT_UNSET;
12479 
12480 	/*
12481 	 * Set the default options.
12482 	 */
12483 	opt = state->dts_options;
12484 	opt[DTRACEOPT_BUFPOLICY] = DTRACEOPT_BUFPOLICY_SWITCH;
12485 	opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_AUTO;
12486 	opt[DTRACEOPT_NSPEC] = dtrace_nspec_default;
12487 	opt[DTRACEOPT_SPECSIZE] = dtrace_specsize_default;
12488 	opt[DTRACEOPT_CPU] = (dtrace_optval_t)DTRACE_CPUALL;
12489 	opt[DTRACEOPT_STRSIZE] = dtrace_strsize_default;
12490 	opt[DTRACEOPT_STACKFRAMES] = dtrace_stackframes_default;
12491 	opt[DTRACEOPT_USTACKFRAMES] = dtrace_ustackframes_default;
12492 	opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_default;
12493 	opt[DTRACEOPT_AGGRATE] = dtrace_aggrate_default;
12494 	opt[DTRACEOPT_SWITCHRATE] = dtrace_switchrate_default;
12495 	opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_default;
12496 	opt[DTRACEOPT_JSTACKFRAMES] = dtrace_jstackframes_default;
12497 	opt[DTRACEOPT_JSTACKSTRSIZE] = dtrace_jstackstrsize_default;
12498 
12499 	state->dts_activity = DTRACE_ACTIVITY_INACTIVE;
12500 
12501 	/*
12502 	 * Depending on the user credentials, we set flag bits which alter probe
12503 	 * visibility or the amount of destructiveness allowed.  In the case of
12504 	 * actual anonymous tracing, or the possession of all privileges, all of
12505 	 * the normal checks are bypassed.
12506 	 */
12507 	if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
12508 		state->dts_cred.dcr_visible = DTRACE_CRV_ALL;
12509 		state->dts_cred.dcr_action = DTRACE_CRA_ALL;
12510 	} else {
12511 		/*
12512 		 * Set up the credentials for this instantiation.  We take a
12513 		 * hold on the credential to prevent it from disappearing on
12514 		 * us; this in turn prevents the zone_t referenced by this
12515 		 * credential from disappearing.  This means that we can
12516 		 * examine the credential and the zone from probe context.
12517 		 */
12518 		crhold(cr);
12519 		state->dts_cred.dcr_cred = cr;
12520 
12521 		/*
12522 		 * CRA_PROC means "we have *some* privilege for dtrace" and
12523 		 * unlocks the use of variables like pid, zonename, etc.
12524 		 */
12525 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE) ||
12526 		    PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
12527 			state->dts_cred.dcr_action |= DTRACE_CRA_PROC;
12528 		}
12529 
12530 		/*
12531 		 * dtrace_user allows use of syscall and profile providers.
12532 		 * If the user also has proc_owner and/or proc_zone, we
12533 		 * extend the scope to include additional visibility and
12534 		 * destructive power.
12535 		 */
12536 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE)) {
12537 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE)) {
12538 				state->dts_cred.dcr_visible |=
12539 				    DTRACE_CRV_ALLPROC;
12540 
12541 				state->dts_cred.dcr_action |=
12542 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
12543 			}
12544 
12545 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE)) {
12546 				state->dts_cred.dcr_visible |=
12547 				    DTRACE_CRV_ALLZONE;
12548 
12549 				state->dts_cred.dcr_action |=
12550 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
12551 			}
12552 
12553 			/*
12554 			 * If we have all privs in whatever zone this is,
12555 			 * we can do destructive things to processes which
12556 			 * have altered credentials.
12557 			 */
12558 			if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
12559 			    cr->cr_zone->zone_privset)) {
12560 				state->dts_cred.dcr_action |=
12561 				    DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
12562 			}
12563 		}
12564 
12565 		/*
12566 		 * Holding the dtrace_kernel privilege also implies that
12567 		 * the user has the dtrace_user privilege from a visibility
12568 		 * perspective.  But without further privileges, some
12569 		 * destructive actions are not available.
12570 		 */
12571 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE)) {
12572 			/*
12573 			 * Make all probes in all zones visible.  However,
12574 			 * this doesn't mean that all actions become available
12575 			 * to all zones.
12576 			 */
12577 			state->dts_cred.dcr_visible |= DTRACE_CRV_KERNEL |
12578 			    DTRACE_CRV_ALLPROC | DTRACE_CRV_ALLZONE;
12579 
12580 			state->dts_cred.dcr_action |= DTRACE_CRA_KERNEL |
12581 			    DTRACE_CRA_PROC;
12582 			/*
12583 			 * Holding proc_owner means that destructive actions
12584 			 * for *this* zone are allowed.
12585 			 */
12586 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
12587 				state->dts_cred.dcr_action |=
12588 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
12589 
12590 			/*
12591 			 * Holding proc_zone means that destructive actions
12592 			 * for this user/group ID in all zones is allowed.
12593 			 */
12594 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
12595 				state->dts_cred.dcr_action |=
12596 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
12597 
12598 			/*
12599 			 * If we have all privs in whatever zone this is,
12600 			 * we can do destructive things to processes which
12601 			 * have altered credentials.
12602 			 */
12603 			if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
12604 			    cr->cr_zone->zone_privset)) {
12605 				state->dts_cred.dcr_action |=
12606 				    DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
12607 			}
12608 		}
12609 
12610 		/*
12611 		 * Holding the dtrace_proc privilege gives control over fasttrap
12612 		 * and pid providers.  We need to grant wider destructive
12613 		 * privileges in the event that the user has proc_owner and/or
12614 		 * proc_zone.
12615 		 */
12616 		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
12617 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
12618 				state->dts_cred.dcr_action |=
12619 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
12620 
12621 			if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
12622 				state->dts_cred.dcr_action |=
12623 				    DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
12624 		}
12625 	}
12626 
12627 	return (state);
12628 }
12629 
12630 static int
12631 dtrace_state_buffer(dtrace_state_t *state, dtrace_buffer_t *buf, int which)
12632 {
12633 	dtrace_optval_t *opt = state->dts_options, size;
12634 	processorid_t cpu;
12635 	int flags = 0, rval, factor, divisor = 1;
12636 
12637 	ASSERT(MUTEX_HELD(&dtrace_lock));
12638 	ASSERT(MUTEX_HELD(&cpu_lock));
12639 	ASSERT(which < DTRACEOPT_MAX);
12640 	ASSERT(state->dts_activity == DTRACE_ACTIVITY_INACTIVE ||
12641 	    (state == dtrace_anon.dta_state &&
12642 	    state->dts_activity == DTRACE_ACTIVITY_ACTIVE));
12643 
12644 	if (opt[which] == DTRACEOPT_UNSET || opt[which] == 0)
12645 		return (0);
12646 
12647 	if (opt[DTRACEOPT_CPU] != DTRACEOPT_UNSET)
12648 		cpu = opt[DTRACEOPT_CPU];
12649 
12650 	if (which == DTRACEOPT_SPECSIZE)
12651 		flags |= DTRACEBUF_NOSWITCH;
12652 
12653 	if (which == DTRACEOPT_BUFSIZE) {
12654 		if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_RING)
12655 			flags |= DTRACEBUF_RING;
12656 
12657 		if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_FILL)
12658 			flags |= DTRACEBUF_FILL;
12659 
12660 		if (state != dtrace_anon.dta_state ||
12661 		    state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
12662 			flags |= DTRACEBUF_INACTIVE;
12663 	}
12664 
12665 	for (size = opt[which]; size >= sizeof (uint64_t); size /= divisor) {
12666 		/*
12667 		 * The size must be 8-byte aligned.  If the size is not 8-byte
12668 		 * aligned, drop it down by the difference.
12669 		 */
12670 		if (size & (sizeof (uint64_t) - 1))
12671 			size -= size & (sizeof (uint64_t) - 1);
12672 
12673 		if (size < state->dts_reserve) {
12674 			/*
12675 			 * Buffers always must be large enough to accommodate
12676 			 * their prereserved space.  We return E2BIG instead
12677 			 * of ENOMEM in this case to allow for user-level
12678 			 * software to differentiate the cases.
12679 			 */
12680 			return (E2BIG);
12681 		}
12682 
12683 		rval = dtrace_buffer_alloc(buf, size, flags, cpu, &factor);
12684 
12685 		if (rval != ENOMEM) {
12686 			opt[which] = size;
12687 			return (rval);
12688 		}
12689 
12690 		if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
12691 			return (rval);
12692 
12693 		for (divisor = 2; divisor < factor; divisor <<= 1)
12694 			continue;
12695 	}
12696 
12697 	return (ENOMEM);
12698 }
12699 
12700 static int
12701 dtrace_state_buffers(dtrace_state_t *state)
12702 {
12703 	dtrace_speculation_t *spec = state->dts_speculations;
12704 	int rval, i;
12705 
12706 	if ((rval = dtrace_state_buffer(state, state->dts_buffer,
12707 	    DTRACEOPT_BUFSIZE)) != 0)
12708 		return (rval);
12709 
12710 	if ((rval = dtrace_state_buffer(state, state->dts_aggbuffer,
12711 	    DTRACEOPT_AGGSIZE)) != 0)
12712 		return (rval);
12713 
12714 	for (i = 0; i < state->dts_nspeculations; i++) {
12715 		if ((rval = dtrace_state_buffer(state,
12716 		    spec[i].dtsp_buffer, DTRACEOPT_SPECSIZE)) != 0)
12717 			return (rval);
12718 	}
12719 
12720 	return (0);
12721 }
12722 
12723 static void
12724 dtrace_state_prereserve(dtrace_state_t *state)
12725 {
12726 	dtrace_ecb_t *ecb;
12727 	dtrace_probe_t *probe;
12728 
12729 	state->dts_reserve = 0;
12730 
12731 	if (state->dts_options[DTRACEOPT_BUFPOLICY] != DTRACEOPT_BUFPOLICY_FILL)
12732 		return;
12733 
12734 	/*
12735 	 * If our buffer policy is a "fill" buffer policy, we need to set the
12736 	 * prereserved space to be the space required by the END probes.
12737 	 */
12738 	probe = dtrace_probes[dtrace_probeid_end - 1];
12739 	ASSERT(probe != NULL);
12740 
12741 	for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
12742 		if (ecb->dte_state != state)
12743 			continue;
12744 
12745 		state->dts_reserve += ecb->dte_needed + ecb->dte_alignment;
12746 	}
12747 }
12748 
12749 static int
12750 dtrace_state_go(dtrace_state_t *state, processorid_t *cpu)
12751 {
12752 	dtrace_optval_t *opt = state->dts_options, sz, nspec;
12753 	dtrace_speculation_t *spec;
12754 	dtrace_buffer_t *buf;
12755 	cyc_handler_t hdlr;
12756 	cyc_time_t when;
12757 	int rval = 0, i, bufsize = NCPU * sizeof (dtrace_buffer_t);
12758 	dtrace_icookie_t cookie;
12759 
12760 	mutex_enter(&cpu_lock);
12761 	mutex_enter(&dtrace_lock);
12762 
12763 	if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
12764 		rval = EBUSY;
12765 		goto out;
12766 	}
12767 
12768 	/*
12769 	 * Before we can perform any checks, we must prime all of the
12770 	 * retained enablings that correspond to this state.
12771 	 */
12772 	dtrace_enabling_prime(state);
12773 
12774 	if (state->dts_destructive && !state->dts_cred.dcr_destructive) {
12775 		rval = EACCES;
12776 		goto out;
12777 	}
12778 
12779 	dtrace_state_prereserve(state);
12780 
12781 	/*
12782 	 * Now we want to do is try to allocate our speculations.
12783 	 * We do not automatically resize the number of speculations; if
12784 	 * this fails, we will fail the operation.
12785 	 */
12786 	nspec = opt[DTRACEOPT_NSPEC];
12787 	ASSERT(nspec != DTRACEOPT_UNSET);
12788 
12789 	if (nspec > INT_MAX) {
12790 		rval = ENOMEM;
12791 		goto out;
12792 	}
12793 
12794 	spec = kmem_zalloc(nspec * sizeof (dtrace_speculation_t),
12795 	    KM_NOSLEEP | KM_NORMALPRI);
12796 
12797 	if (spec == NULL) {
12798 		rval = ENOMEM;
12799 		goto out;
12800 	}
12801 
12802 	state->dts_speculations = spec;
12803 	state->dts_nspeculations = (int)nspec;
12804 
12805 	for (i = 0; i < nspec; i++) {
12806 		if ((buf = kmem_zalloc(bufsize,
12807 		    KM_NOSLEEP | KM_NORMALPRI)) == NULL) {
12808 			rval = ENOMEM;
12809 			goto err;
12810 		}
12811 
12812 		spec[i].dtsp_buffer = buf;
12813 	}
12814 
12815 	if (opt[DTRACEOPT_GRABANON] != DTRACEOPT_UNSET) {
12816 		if (dtrace_anon.dta_state == NULL) {
12817 			rval = ENOENT;
12818 			goto out;
12819 		}
12820 
12821 		if (state->dts_necbs != 0) {
12822 			rval = EALREADY;
12823 			goto out;
12824 		}
12825 
12826 		state->dts_anon = dtrace_anon_grab();
12827 		ASSERT(state->dts_anon != NULL);
12828 		state = state->dts_anon;
12829 
12830 		/*
12831 		 * We want "grabanon" to be set in the grabbed state, so we'll
12832 		 * copy that option value from the grabbing state into the
12833 		 * grabbed state.
12834 		 */
12835 		state->dts_options[DTRACEOPT_GRABANON] =
12836 		    opt[DTRACEOPT_GRABANON];
12837 
12838 		*cpu = dtrace_anon.dta_beganon;
12839 
12840 		/*
12841 		 * If the anonymous state is active (as it almost certainly
12842 		 * is if the anonymous enabling ultimately matched anything),
12843 		 * we don't allow any further option processing -- but we
12844 		 * don't return failure.
12845 		 */
12846 		if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
12847 			goto out;
12848 	}
12849 
12850 	if (opt[DTRACEOPT_AGGSIZE] != DTRACEOPT_UNSET &&
12851 	    opt[DTRACEOPT_AGGSIZE] != 0) {
12852 		if (state->dts_aggregations == NULL) {
12853 			/*
12854 			 * We're not going to create an aggregation buffer
12855 			 * because we don't have any ECBs that contain
12856 			 * aggregations -- set this option to 0.
12857 			 */
12858 			opt[DTRACEOPT_AGGSIZE] = 0;
12859 		} else {
12860 			/*
12861 			 * If we have an aggregation buffer, we must also have
12862 			 * a buffer to use as scratch.
12863 			 */
12864 			if (opt[DTRACEOPT_BUFSIZE] == DTRACEOPT_UNSET ||
12865 			    opt[DTRACEOPT_BUFSIZE] < state->dts_needed) {
12866 				opt[DTRACEOPT_BUFSIZE] = state->dts_needed;
12867 			}
12868 		}
12869 	}
12870 
12871 	if (opt[DTRACEOPT_SPECSIZE] != DTRACEOPT_UNSET &&
12872 	    opt[DTRACEOPT_SPECSIZE] != 0) {
12873 		if (!state->dts_speculates) {
12874 			/*
12875 			 * We're not going to create speculation buffers
12876 			 * because we don't have any ECBs that actually
12877 			 * speculate -- set the speculation size to 0.
12878 			 */
12879 			opt[DTRACEOPT_SPECSIZE] = 0;
12880 		}
12881 	}
12882 
12883 	/*
12884 	 * The bare minimum size for any buffer that we're actually going to
12885 	 * do anything to is sizeof (uint64_t).
12886 	 */
12887 	sz = sizeof (uint64_t);
12888 
12889 	if ((state->dts_needed != 0 && opt[DTRACEOPT_BUFSIZE] < sz) ||
12890 	    (state->dts_speculates && opt[DTRACEOPT_SPECSIZE] < sz) ||
12891 	    (state->dts_aggregations != NULL && opt[DTRACEOPT_AGGSIZE] < sz)) {
12892 		/*
12893 		 * A buffer size has been explicitly set to 0 (or to a size
12894 		 * that will be adjusted to 0) and we need the space -- we
12895 		 * need to return failure.  We return ENOSPC to differentiate
12896 		 * it from failing to allocate a buffer due to failure to meet
12897 		 * the reserve (for which we return E2BIG).
12898 		 */
12899 		rval = ENOSPC;
12900 		goto out;
12901 	}
12902 
12903 	if ((rval = dtrace_state_buffers(state)) != 0)
12904 		goto err;
12905 
12906 	if ((sz = opt[DTRACEOPT_DYNVARSIZE]) == DTRACEOPT_UNSET)
12907 		sz = dtrace_dstate_defsize;
12908 
12909 	do {
12910 		rval = dtrace_dstate_init(&state->dts_vstate.dtvs_dynvars, sz);
12911 
12912 		if (rval == 0)
12913 			break;
12914 
12915 		if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
12916 			goto err;
12917 	} while (sz >>= 1);
12918 
12919 	opt[DTRACEOPT_DYNVARSIZE] = sz;
12920 
12921 	if (rval != 0)
12922 		goto err;
12923 
12924 	if (opt[DTRACEOPT_STATUSRATE] > dtrace_statusrate_max)
12925 		opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_max;
12926 
12927 	if (opt[DTRACEOPT_CLEANRATE] == 0)
12928 		opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
12929 
12930 	if (opt[DTRACEOPT_CLEANRATE] < dtrace_cleanrate_min)
12931 		opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_min;
12932 
12933 	if (opt[DTRACEOPT_CLEANRATE] > dtrace_cleanrate_max)
12934 		opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
12935 
12936 	hdlr.cyh_func = (cyc_func_t)dtrace_state_clean;
12937 	hdlr.cyh_arg = state;
12938 	hdlr.cyh_level = CY_LOW_LEVEL;
12939 
12940 	when.cyt_when = 0;
12941 	when.cyt_interval = opt[DTRACEOPT_CLEANRATE];
12942 
12943 	state->dts_cleaner = cyclic_add(&hdlr, &when);
12944 
12945 	hdlr.cyh_func = (cyc_func_t)dtrace_state_deadman;
12946 	hdlr.cyh_arg = state;
12947 	hdlr.cyh_level = CY_LOW_LEVEL;
12948 
12949 	when.cyt_when = 0;
12950 	when.cyt_interval = dtrace_deadman_interval;
12951 
12952 	state->dts_alive = state->dts_laststatus = dtrace_gethrtime();
12953 	state->dts_deadman = cyclic_add(&hdlr, &when);
12954 
12955 	state->dts_activity = DTRACE_ACTIVITY_WARMUP;
12956 
12957 	/*
12958 	 * Now it's time to actually fire the BEGIN probe.  We need to disable
12959 	 * interrupts here both to record the CPU on which we fired the BEGIN
12960 	 * probe (the data from this CPU will be processed first at user
12961 	 * level) and to manually activate the buffer for this CPU.
12962 	 */
12963 	cookie = dtrace_interrupt_disable();
12964 	*cpu = CPU->cpu_id;
12965 	ASSERT(state->dts_buffer[*cpu].dtb_flags & DTRACEBUF_INACTIVE);
12966 	state->dts_buffer[*cpu].dtb_flags &= ~DTRACEBUF_INACTIVE;
12967 
12968 	dtrace_probe(dtrace_probeid_begin,
12969 	    (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
12970 	dtrace_interrupt_enable(cookie);
12971 	/*
12972 	 * We may have had an exit action from a BEGIN probe; only change our
12973 	 * state to ACTIVE if we're still in WARMUP.
12974 	 */
12975 	ASSERT(state->dts_activity == DTRACE_ACTIVITY_WARMUP ||
12976 	    state->dts_activity == DTRACE_ACTIVITY_DRAINING);
12977 
12978 	if (state->dts_activity == DTRACE_ACTIVITY_WARMUP)
12979 		state->dts_activity = DTRACE_ACTIVITY_ACTIVE;
12980 
12981 	/*
12982 	 * Regardless of whether or not now we're in ACTIVE or DRAINING, we
12983 	 * want each CPU to transition its principal buffer out of the
12984 	 * INACTIVE state.  Doing this assures that no CPU will suddenly begin
12985 	 * processing an ECB halfway down a probe's ECB chain; all CPUs will
12986 	 * atomically transition from processing none of a state's ECBs to
12987 	 * processing all of them.
12988 	 */
12989 	dtrace_xcall(DTRACE_CPUALL,
12990 	    (dtrace_xcall_t)dtrace_buffer_activate, state);
12991 	goto out;
12992 
12993 err:
12994 	dtrace_buffer_free(state->dts_buffer);
12995 	dtrace_buffer_free(state->dts_aggbuffer);
12996 
12997 	if ((nspec = state->dts_nspeculations) == 0) {
12998 		ASSERT(state->dts_speculations == NULL);
12999 		goto out;
13000 	}
13001 
13002 	spec = state->dts_speculations;
13003 	ASSERT(spec != NULL);
13004 
13005 	for (i = 0; i < state->dts_nspeculations; i++) {
13006 		if ((buf = spec[i].dtsp_buffer) == NULL)
13007 			break;
13008 
13009 		dtrace_buffer_free(buf);
13010 		kmem_free(buf, bufsize);
13011 	}
13012 
13013 	kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
13014 	state->dts_nspeculations = 0;
13015 	state->dts_speculations = NULL;
13016 
13017 out:
13018 	mutex_exit(&dtrace_lock);
13019 	mutex_exit(&cpu_lock);
13020 
13021 	return (rval);
13022 }
13023 
13024 static int
13025 dtrace_state_stop(dtrace_state_t *state, processorid_t *cpu)
13026 {
13027 	dtrace_icookie_t cookie;
13028 
13029 	ASSERT(MUTEX_HELD(&dtrace_lock));
13030 
13031 	if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE &&
13032 	    state->dts_activity != DTRACE_ACTIVITY_DRAINING)
13033 		return (EINVAL);
13034 
13035 	/*
13036 	 * We'll set the activity to DTRACE_ACTIVITY_DRAINING, and issue a sync
13037 	 * to be sure that every CPU has seen it.  See below for the details
13038 	 * on why this is done.
13039 	 */
13040 	state->dts_activity = DTRACE_ACTIVITY_DRAINING;
13041 	dtrace_sync();
13042 
13043 	/*
13044 	 * By this point, it is impossible for any CPU to be still processing
13045 	 * with DTRACE_ACTIVITY_ACTIVE.  We can thus set our activity to
13046 	 * DTRACE_ACTIVITY_COOLDOWN and know that we're not racing with any
13047 	 * other CPU in dtrace_buffer_reserve().  This allows dtrace_probe()
13048 	 * and callees to know that the activity is DTRACE_ACTIVITY_COOLDOWN
13049 	 * iff we're in the END probe.
13050 	 */
13051 	state->dts_activity = DTRACE_ACTIVITY_COOLDOWN;
13052 	dtrace_sync();
13053 	ASSERT(state->dts_activity == DTRACE_ACTIVITY_COOLDOWN);
13054 
13055 	/*
13056 	 * Finally, we can release the reserve and call the END probe.  We
13057 	 * disable interrupts across calling the END probe to allow us to
13058 	 * return the CPU on which we actually called the END probe.  This
13059 	 * allows user-land to be sure that this CPU's principal buffer is
13060 	 * processed last.
13061 	 */
13062 	state->dts_reserve = 0;
13063 
13064 	cookie = dtrace_interrupt_disable();
13065 	*cpu = CPU->cpu_id;
13066 	dtrace_probe(dtrace_probeid_end,
13067 	    (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
13068 	dtrace_interrupt_enable(cookie);
13069 
13070 	state->dts_activity = DTRACE_ACTIVITY_STOPPED;
13071 	dtrace_sync();
13072 
13073 	return (0);
13074 }
13075 
13076 static int
13077 dtrace_state_option(dtrace_state_t *state, dtrace_optid_t option,
13078     dtrace_optval_t val)
13079 {
13080 	ASSERT(MUTEX_HELD(&dtrace_lock));
13081 
13082 	if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
13083 		return (EBUSY);
13084 
13085 	if (option >= DTRACEOPT_MAX)
13086 		return (EINVAL);
13087 
13088 	if (option != DTRACEOPT_CPU && val < 0)
13089 		return (EINVAL);
13090 
13091 	switch (option) {
13092 	case DTRACEOPT_DESTRUCTIVE:
13093 		if (dtrace_destructive_disallow)
13094 			return (EACCES);
13095 
13096 		state->dts_cred.dcr_destructive = 1;
13097 		break;
13098 
13099 	case DTRACEOPT_BUFSIZE:
13100 	case DTRACEOPT_DYNVARSIZE:
13101 	case DTRACEOPT_AGGSIZE:
13102 	case DTRACEOPT_SPECSIZE:
13103 	case DTRACEOPT_STRSIZE:
13104 		if (val < 0)
13105 			return (EINVAL);
13106 
13107 		if (val >= LONG_MAX) {
13108 			/*
13109 			 * If this is an otherwise negative value, set it to
13110 			 * the highest multiple of 128m less than LONG_MAX.
13111 			 * Technically, we're adjusting the size without
13112 			 * regard to the buffer resizing policy, but in fact,
13113 			 * this has no effect -- if we set the buffer size to
13114 			 * ~LONG_MAX and the buffer policy is ultimately set to
13115 			 * be "manual", the buffer allocation is guaranteed to
13116 			 * fail, if only because the allocation requires two
13117 			 * buffers.  (We set the the size to the highest
13118 			 * multiple of 128m because it ensures that the size
13119 			 * will remain a multiple of a megabyte when
13120 			 * repeatedly halved -- all the way down to 15m.)
13121 			 */
13122 			val = LONG_MAX - (1 << 27) + 1;
13123 		}
13124 	}
13125 
13126 	state->dts_options[option] = val;
13127 
13128 	return (0);
13129 }
13130 
13131 static void
13132 dtrace_state_destroy(dtrace_state_t *state)
13133 {
13134 	dtrace_ecb_t *ecb;
13135 	dtrace_vstate_t *vstate = &state->dts_vstate;
13136 	minor_t minor = getminor(state->dts_dev);
13137 	int i, bufsize = NCPU * sizeof (dtrace_buffer_t);
13138 	dtrace_speculation_t *spec = state->dts_speculations;
13139 	int nspec = state->dts_nspeculations;
13140 	uint32_t match;
13141 
13142 	ASSERT(MUTEX_HELD(&dtrace_lock));
13143 	ASSERT(MUTEX_HELD(&cpu_lock));
13144 
13145 	/*
13146 	 * First, retract any retained enablings for this state.
13147 	 */
13148 	dtrace_enabling_retract(state);
13149 	ASSERT(state->dts_nretained == 0);
13150 
13151 	if (state->dts_activity == DTRACE_ACTIVITY_ACTIVE ||
13152 	    state->dts_activity == DTRACE_ACTIVITY_DRAINING) {
13153 		/*
13154 		 * We have managed to come into dtrace_state_destroy() on a
13155 		 * hot enabling -- almost certainly because of a disorderly
13156 		 * shutdown of a consumer.  (That is, a consumer that is
13157 		 * exiting without having called dtrace_stop().) In this case,
13158 		 * we're going to set our activity to be KILLED, and then
13159 		 * issue a sync to be sure that everyone is out of probe
13160 		 * context before we start blowing away ECBs.
13161 		 */
13162 		state->dts_activity = DTRACE_ACTIVITY_KILLED;
13163 		dtrace_sync();
13164 	}
13165 
13166 	/*
13167 	 * Release the credential hold we took in dtrace_state_create().
13168 	 */
13169 	if (state->dts_cred.dcr_cred != NULL)
13170 		crfree(state->dts_cred.dcr_cred);
13171 
13172 	/*
13173 	 * Now we can safely disable and destroy any enabled probes.  Because
13174 	 * any DTRACE_PRIV_KERNEL probes may actually be slowing our progress
13175 	 * (especially if they're all enabled), we take two passes through the
13176 	 * ECBs:  in the first, we disable just DTRACE_PRIV_KERNEL probes, and
13177 	 * in the second we disable whatever is left over.
13178 	 */
13179 	for (match = DTRACE_PRIV_KERNEL; ; match = 0) {
13180 		for (i = 0; i < state->dts_necbs; i++) {
13181 			if ((ecb = state->dts_ecbs[i]) == NULL)
13182 				continue;
13183 
13184 			if (match && ecb->dte_probe != NULL) {
13185 				dtrace_probe_t *probe = ecb->dte_probe;
13186 				dtrace_provider_t *prov = probe->dtpr_provider;
13187 
13188 				if (!(prov->dtpv_priv.dtpp_flags & match))
13189 					continue;
13190 			}
13191 
13192 			dtrace_ecb_disable(ecb);
13193 			dtrace_ecb_destroy(ecb);
13194 		}
13195 
13196 		if (!match)
13197 			break;
13198 	}
13199 
13200 	/*
13201 	 * Before we free the buffers, perform one more sync to assure that
13202 	 * every CPU is out of probe context.
13203 	 */
13204 	dtrace_sync();
13205 
13206 	dtrace_buffer_free(state->dts_buffer);
13207 	dtrace_buffer_free(state->dts_aggbuffer);
13208 
13209 	for (i = 0; i < nspec; i++)
13210 		dtrace_buffer_free(spec[i].dtsp_buffer);
13211 
13212 	if (state->dts_cleaner != CYCLIC_NONE)
13213 		cyclic_remove(state->dts_cleaner);
13214 
13215 	if (state->dts_deadman != CYCLIC_NONE)
13216 		cyclic_remove(state->dts_deadman);
13217 
13218 	dtrace_dstate_fini(&vstate->dtvs_dynvars);
13219 	dtrace_vstate_fini(vstate);
13220 	kmem_free(state->dts_ecbs, state->dts_necbs * sizeof (dtrace_ecb_t *));
13221 
13222 	if (state->dts_aggregations != NULL) {
13223 #ifdef DEBUG
13224 		for (i = 0; i < state->dts_naggregations; i++)
13225 			ASSERT(state->dts_aggregations[i] == NULL);
13226 #endif
13227 		ASSERT(state->dts_naggregations > 0);
13228 		kmem_free(state->dts_aggregations,
13229 		    state->dts_naggregations * sizeof (dtrace_aggregation_t *));
13230 	}
13231 
13232 	kmem_free(state->dts_buffer, bufsize);
13233 	kmem_free(state->dts_aggbuffer, bufsize);
13234 
13235 	for (i = 0; i < nspec; i++)
13236 		kmem_free(spec[i].dtsp_buffer, bufsize);
13237 
13238 	kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
13239 
13240 	dtrace_format_destroy(state);
13241 
13242 	vmem_destroy(state->dts_aggid_arena);
13243 	ddi_soft_state_free(dtrace_softstate, minor);
13244 	vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
13245 }
13246 
13247 /*
13248  * DTrace Anonymous Enabling Functions
13249  */
13250 static dtrace_state_t *
13251 dtrace_anon_grab(void)
13252 {
13253 	dtrace_state_t *state;
13254 
13255 	ASSERT(MUTEX_HELD(&dtrace_lock));
13256 
13257 	if ((state = dtrace_anon.dta_state) == NULL) {
13258 		ASSERT(dtrace_anon.dta_enabling == NULL);
13259 		return (NULL);
13260 	}
13261 
13262 	ASSERT(dtrace_anon.dta_enabling != NULL);
13263 	ASSERT(dtrace_retained != NULL);
13264 
13265 	dtrace_enabling_destroy(dtrace_anon.dta_enabling);
13266 	dtrace_anon.dta_enabling = NULL;
13267 	dtrace_anon.dta_state = NULL;
13268 
13269 	return (state);
13270 }
13271 
13272 static void
13273 dtrace_anon_property(void)
13274 {
13275 	int i, rv;
13276 	dtrace_state_t *state;
13277 	dof_hdr_t *dof;
13278 	char c[32];		/* enough for "dof-data-" + digits */
13279 
13280 	ASSERT(MUTEX_HELD(&dtrace_lock));
13281 	ASSERT(MUTEX_HELD(&cpu_lock));
13282 
13283 	for (i = 0; ; i++) {
13284 		(void) snprintf(c, sizeof (c), "dof-data-%d", i);
13285 
13286 		dtrace_err_verbose = 1;
13287 
13288 		if ((dof = dtrace_dof_property(c)) == NULL) {
13289 			dtrace_err_verbose = 0;
13290 			break;
13291 		}
13292 
13293 		/*
13294 		 * We want to create anonymous state, so we need to transition
13295 		 * the kernel debugger to indicate that DTrace is active.  If
13296 		 * this fails (e.g. because the debugger has modified text in
13297 		 * some way), we won't continue with the processing.
13298 		 */
13299 		if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
13300 			cmn_err(CE_NOTE, "kernel debugger active; anonymous "
13301 			    "enabling ignored.");
13302 			dtrace_dof_destroy(dof);
13303 			break;
13304 		}
13305 
13306 		/*
13307 		 * If we haven't allocated an anonymous state, we'll do so now.
13308 		 */
13309 		if ((state = dtrace_anon.dta_state) == NULL) {
13310 			state = dtrace_state_create(NULL, NULL);
13311 			dtrace_anon.dta_state = state;
13312 
13313 			if (state == NULL) {
13314 				/*
13315 				 * This basically shouldn't happen:  the only
13316 				 * failure mode from dtrace_state_create() is a
13317 				 * failure of ddi_soft_state_zalloc() that
13318 				 * itself should never happen.  Still, the
13319 				 * interface allows for a failure mode, and
13320 				 * we want to fail as gracefully as possible:
13321 				 * we'll emit an error message and cease
13322 				 * processing anonymous state in this case.
13323 				 */
13324 				cmn_err(CE_WARN, "failed to create "
13325 				    "anonymous state");
13326 				dtrace_dof_destroy(dof);
13327 				break;
13328 			}
13329 		}
13330 
13331 		rv = dtrace_dof_slurp(dof, &state->dts_vstate, CRED(),
13332 		    &dtrace_anon.dta_enabling, 0, B_TRUE);
13333 
13334 		if (rv == 0)
13335 			rv = dtrace_dof_options(dof, state);
13336 
13337 		dtrace_err_verbose = 0;
13338 		dtrace_dof_destroy(dof);
13339 
13340 		if (rv != 0) {
13341 			/*
13342 			 * This is malformed DOF; chuck any anonymous state
13343 			 * that we created.
13344 			 */
13345 			ASSERT(dtrace_anon.dta_enabling == NULL);
13346 			dtrace_state_destroy(state);
13347 			dtrace_anon.dta_state = NULL;
13348 			break;
13349 		}
13350 
13351 		ASSERT(dtrace_anon.dta_enabling != NULL);
13352 	}
13353 
13354 	if (dtrace_anon.dta_enabling != NULL) {
13355 		int rval;
13356 
13357 		/*
13358 		 * dtrace_enabling_retain() can only fail because we are
13359 		 * trying to retain more enablings than are allowed -- but
13360 		 * we only have one anonymous enabling, and we are guaranteed
13361 		 * to be allowed at least one retained enabling; we assert
13362 		 * that dtrace_enabling_retain() returns success.
13363 		 */
13364 		rval = dtrace_enabling_retain(dtrace_anon.dta_enabling);
13365 		ASSERT(rval == 0);
13366 
13367 		dtrace_enabling_dump(dtrace_anon.dta_enabling);
13368 	}
13369 }
13370 
13371 /*
13372  * DTrace Helper Functions
13373  */
13374 static void
13375 dtrace_helper_trace(dtrace_helper_action_t *helper,
13376     dtrace_mstate_t *mstate, dtrace_vstate_t *vstate, int where)
13377 {
13378 	uint32_t size, next, nnext, i;
13379 	dtrace_helptrace_t *ent;
13380 	uint16_t flags = cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
13381 
13382 	if (!dtrace_helptrace_enabled)
13383 		return;
13384 
13385 	ASSERT(vstate->dtvs_nlocals <= dtrace_helptrace_nlocals);
13386 
13387 	/*
13388 	 * What would a tracing framework be without its own tracing
13389 	 * framework?  (Well, a hell of a lot simpler, for starters...)
13390 	 */
13391 	size = sizeof (dtrace_helptrace_t) + dtrace_helptrace_nlocals *
13392 	    sizeof (uint64_t) - sizeof (uint64_t);
13393 
13394 	/*
13395 	 * Iterate until we can allocate a slot in the trace buffer.
13396 	 */
13397 	do {
13398 		next = dtrace_helptrace_next;
13399 
13400 		if (next + size < dtrace_helptrace_bufsize) {
13401 			nnext = next + size;
13402 		} else {
13403 			nnext = size;
13404 		}
13405 	} while (dtrace_cas32(&dtrace_helptrace_next, next, nnext) != next);
13406 
13407 	/*
13408 	 * We have our slot; fill it in.
13409 	 */
13410 	if (nnext == size)
13411 		next = 0;
13412 
13413 	ent = (dtrace_helptrace_t *)&dtrace_helptrace_buffer[next];
13414 	ent->dtht_helper = helper;
13415 	ent->dtht_where = where;
13416 	ent->dtht_nlocals = vstate->dtvs_nlocals;
13417 
13418 	ent->dtht_fltoffs = (mstate->dtms_present & DTRACE_MSTATE_FLTOFFS) ?
13419 	    mstate->dtms_fltoffs : -1;
13420 	ent->dtht_fault = DTRACE_FLAGS2FLT(flags);
13421 	ent->dtht_illval = cpu_core[CPU->cpu_id].cpuc_dtrace_illval;
13422 
13423 	for (i = 0; i < vstate->dtvs_nlocals; i++) {
13424 		dtrace_statvar_t *svar;
13425 
13426 		if ((svar = vstate->dtvs_locals[i]) == NULL)
13427 			continue;
13428 
13429 		ASSERT(svar->dtsv_size >= NCPU * sizeof (uint64_t));
13430 		ent->dtht_locals[i] =
13431 		    ((uint64_t *)(uintptr_t)svar->dtsv_data)[CPU->cpu_id];
13432 	}
13433 }
13434 
13435 static uint64_t
13436 dtrace_helper(int which, dtrace_mstate_t *mstate,
13437     dtrace_state_t *state, uint64_t arg0, uint64_t arg1)
13438 {
13439 	uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
13440 	uint64_t sarg0 = mstate->dtms_arg[0];
13441 	uint64_t sarg1 = mstate->dtms_arg[1];
13442 	uint64_t rval;
13443 	dtrace_helpers_t *helpers = curproc->p_dtrace_helpers;
13444 	dtrace_helper_action_t *helper;
13445 	dtrace_vstate_t *vstate;
13446 	dtrace_difo_t *pred;
13447 	int i, trace = dtrace_helptrace_enabled;
13448 
13449 	ASSERT(which >= 0 && which < DTRACE_NHELPER_ACTIONS);
13450 
13451 	if (helpers == NULL)
13452 		return (0);
13453 
13454 	if ((helper = helpers->dthps_actions[which]) == NULL)
13455 		return (0);
13456 
13457 	vstate = &helpers->dthps_vstate;
13458 	mstate->dtms_arg[0] = arg0;
13459 	mstate->dtms_arg[1] = arg1;
13460 
13461 	/*
13462 	 * Now iterate over each helper.  If its predicate evaluates to 'true',
13463 	 * we'll call the corresponding actions.  Note that the below calls
13464 	 * to dtrace_dif_emulate() may set faults in machine state.  This is
13465 	 * okay:  our caller (the outer dtrace_dif_emulate()) will simply plow
13466 	 * the stored DIF offset with its own (which is the desired behavior).
13467 	 * Also, note the calls to dtrace_dif_emulate() may allocate scratch
13468 	 * from machine state; this is okay, too.
13469 	 */
13470 	for (; helper != NULL; helper = helper->dtha_next) {
13471 		if ((pred = helper->dtha_predicate) != NULL) {
13472 			if (trace)
13473 				dtrace_helper_trace(helper, mstate, vstate, 0);
13474 
13475 			if (!dtrace_dif_emulate(pred, mstate, vstate, state))
13476 				goto next;
13477 
13478 			if (*flags & CPU_DTRACE_FAULT)
13479 				goto err;
13480 		}
13481 
13482 		for (i = 0; i < helper->dtha_nactions; i++) {
13483 			if (trace)
13484 				dtrace_helper_trace(helper,
13485 				    mstate, vstate, i + 1);
13486 
13487 			rval = dtrace_dif_emulate(helper->dtha_actions[i],
13488 			    mstate, vstate, state);
13489 
13490 			if (*flags & CPU_DTRACE_FAULT)
13491 				goto err;
13492 		}
13493 
13494 next:
13495 		if (trace)
13496 			dtrace_helper_trace(helper, mstate, vstate,
13497 			    DTRACE_HELPTRACE_NEXT);
13498 	}
13499 
13500 	if (trace)
13501 		dtrace_helper_trace(helper, mstate, vstate,
13502 		    DTRACE_HELPTRACE_DONE);
13503 
13504 	/*
13505 	 * Restore the arg0 that we saved upon entry.
13506 	 */
13507 	mstate->dtms_arg[0] = sarg0;
13508 	mstate->dtms_arg[1] = sarg1;
13509 
13510 	return (rval);
13511 
13512 err:
13513 	if (trace)
13514 		dtrace_helper_trace(helper, mstate, vstate,
13515 		    DTRACE_HELPTRACE_ERR);
13516 
13517 	/*
13518 	 * Restore the arg0 that we saved upon entry.
13519 	 */
13520 	mstate->dtms_arg[0] = sarg0;
13521 	mstate->dtms_arg[1] = sarg1;
13522 
13523 	return (NULL);
13524 }
13525 
13526 static void
13527 dtrace_helper_action_destroy(dtrace_helper_action_t *helper,
13528     dtrace_vstate_t *vstate)
13529 {
13530 	int i;
13531 
13532 	if (helper->dtha_predicate != NULL)
13533 		dtrace_difo_release(helper->dtha_predicate, vstate);
13534 
13535 	for (i = 0; i < helper->dtha_nactions; i++) {
13536 		ASSERT(helper->dtha_actions[i] != NULL);
13537 		dtrace_difo_release(helper->dtha_actions[i], vstate);
13538 	}
13539 
13540 	kmem_free(helper->dtha_actions,
13541 	    helper->dtha_nactions * sizeof (dtrace_difo_t *));
13542 	kmem_free(helper, sizeof (dtrace_helper_action_t));
13543 }
13544 
13545 static int
13546 dtrace_helper_destroygen(int gen)
13547 {
13548 	proc_t *p = curproc;
13549 	dtrace_helpers_t *help = p->p_dtrace_helpers;
13550 	dtrace_vstate_t *vstate;
13551 	int i;
13552 
13553 	ASSERT(MUTEX_HELD(&dtrace_lock));
13554 
13555 	if (help == NULL || gen > help->dthps_generation)
13556 		return (EINVAL);
13557 
13558 	vstate = &help->dthps_vstate;
13559 
13560 	for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
13561 		dtrace_helper_action_t *last = NULL, *h, *next;
13562 
13563 		for (h = help->dthps_actions[i]; h != NULL; h = next) {
13564 			next = h->dtha_next;
13565 
13566 			if (h->dtha_generation == gen) {
13567 				if (last != NULL) {
13568 					last->dtha_next = next;
13569 				} else {
13570 					help->dthps_actions[i] = next;
13571 				}
13572 
13573 				dtrace_helper_action_destroy(h, vstate);
13574 			} else {
13575 				last = h;
13576 			}
13577 		}
13578 	}
13579 
13580 	/*
13581 	 * Interate until we've cleared out all helper providers with the
13582 	 * given generation number.
13583 	 */
13584 	for (;;) {
13585 		dtrace_helper_provider_t *prov;
13586 
13587 		/*
13588 		 * Look for a helper provider with the right generation. We
13589 		 * have to start back at the beginning of the list each time
13590 		 * because we drop dtrace_lock. It's unlikely that we'll make
13591 		 * more than two passes.
13592 		 */
13593 		for (i = 0; i < help->dthps_nprovs; i++) {
13594 			prov = help->dthps_provs[i];
13595 
13596 			if (prov->dthp_generation == gen)
13597 				break;
13598 		}
13599 
13600 		/*
13601 		 * If there were no matches, we're done.
13602 		 */
13603 		if (i == help->dthps_nprovs)
13604 			break;
13605 
13606 		/*
13607 		 * Move the last helper provider into this slot.
13608 		 */
13609 		help->dthps_nprovs--;
13610 		help->dthps_provs[i] = help->dthps_provs[help->dthps_nprovs];
13611 		help->dthps_provs[help->dthps_nprovs] = NULL;
13612 
13613 		mutex_exit(&dtrace_lock);
13614 
13615 		/*
13616 		 * If we have a meta provider, remove this helper provider.
13617 		 */
13618 		mutex_enter(&dtrace_meta_lock);
13619 		if (dtrace_meta_pid != NULL) {
13620 			ASSERT(dtrace_deferred_pid == NULL);
13621 			dtrace_helper_provider_remove(&prov->dthp_prov,
13622 			    p->p_pid);
13623 		}
13624 		mutex_exit(&dtrace_meta_lock);
13625 
13626 		dtrace_helper_provider_destroy(prov);
13627 
13628 		mutex_enter(&dtrace_lock);
13629 	}
13630 
13631 	return (0);
13632 }
13633 
13634 static int
13635 dtrace_helper_validate(dtrace_helper_action_t *helper)
13636 {
13637 	int err = 0, i;
13638 	dtrace_difo_t *dp;
13639 
13640 	if ((dp = helper->dtha_predicate) != NULL)
13641 		err += dtrace_difo_validate_helper(dp);
13642 
13643 	for (i = 0; i < helper->dtha_nactions; i++)
13644 		err += dtrace_difo_validate_helper(helper->dtha_actions[i]);
13645 
13646 	return (err == 0);
13647 }
13648 
13649 static int
13650 dtrace_helper_action_add(int which, dtrace_ecbdesc_t *ep)
13651 {
13652 	dtrace_helpers_t *help;
13653 	dtrace_helper_action_t *helper, *last;
13654 	dtrace_actdesc_t *act;
13655 	dtrace_vstate_t *vstate;
13656 	dtrace_predicate_t *pred;
13657 	int count = 0, nactions = 0, i;
13658 
13659 	if (which < 0 || which >= DTRACE_NHELPER_ACTIONS)
13660 		return (EINVAL);
13661 
13662 	help = curproc->p_dtrace_helpers;
13663 	last = help->dthps_actions[which];
13664 	vstate = &help->dthps_vstate;
13665 
13666 	for (count = 0; last != NULL; last = last->dtha_next) {
13667 		count++;
13668 		if (last->dtha_next == NULL)
13669 			break;
13670 	}
13671 
13672 	/*
13673 	 * If we already have dtrace_helper_actions_max helper actions for this
13674 	 * helper action type, we'll refuse to add a new one.
13675 	 */
13676 	if (count >= dtrace_helper_actions_max)
13677 		return (ENOSPC);
13678 
13679 	helper = kmem_zalloc(sizeof (dtrace_helper_action_t), KM_SLEEP);
13680 	helper->dtha_generation = help->dthps_generation;
13681 
13682 	if ((pred = ep->dted_pred.dtpdd_predicate) != NULL) {
13683 		ASSERT(pred->dtp_difo != NULL);
13684 		dtrace_difo_hold(pred->dtp_difo);
13685 		helper->dtha_predicate = pred->dtp_difo;
13686 	}
13687 
13688 	for (act = ep->dted_action; act != NULL; act = act->dtad_next) {
13689 		if (act->dtad_kind != DTRACEACT_DIFEXPR)
13690 			goto err;
13691 
13692 		if (act->dtad_difo == NULL)
13693 			goto err;
13694 
13695 		nactions++;
13696 	}
13697 
13698 	helper->dtha_actions = kmem_zalloc(sizeof (dtrace_difo_t *) *
13699 	    (helper->dtha_nactions = nactions), KM_SLEEP);
13700 
13701 	for (act = ep->dted_action, i = 0; act != NULL; act = act->dtad_next) {
13702 		dtrace_difo_hold(act->dtad_difo);
13703 		helper->dtha_actions[i++] = act->dtad_difo;
13704 	}
13705 
13706 	if (!dtrace_helper_validate(helper))
13707 		goto err;
13708 
13709 	if (last == NULL) {
13710 		help->dthps_actions[which] = helper;
13711 	} else {
13712 		last->dtha_next = helper;
13713 	}
13714 
13715 	if (vstate->dtvs_nlocals > dtrace_helptrace_nlocals) {
13716 		dtrace_helptrace_nlocals = vstate->dtvs_nlocals;
13717 		dtrace_helptrace_next = 0;
13718 	}
13719 
13720 	return (0);
13721 err:
13722 	dtrace_helper_action_destroy(helper, vstate);
13723 	return (EINVAL);
13724 }
13725 
13726 static void
13727 dtrace_helper_provider_register(proc_t *p, dtrace_helpers_t *help,
13728     dof_helper_t *dofhp)
13729 {
13730 	ASSERT(MUTEX_NOT_HELD(&dtrace_lock));
13731 
13732 	mutex_enter(&dtrace_meta_lock);
13733 	mutex_enter(&dtrace_lock);
13734 
13735 	if (!dtrace_attached() || dtrace_meta_pid == NULL) {
13736 		/*
13737 		 * If the dtrace module is loaded but not attached, or if
13738 		 * there aren't isn't a meta provider registered to deal with
13739 		 * these provider descriptions, we need to postpone creating
13740 		 * the actual providers until later.
13741 		 */
13742 
13743 		if (help->dthps_next == NULL && help->dthps_prev == NULL &&
13744 		    dtrace_deferred_pid != help) {
13745 			help->dthps_deferred = 1;
13746 			help->dthps_pid = p->p_pid;
13747 			help->dthps_next = dtrace_deferred_pid;
13748 			help->dthps_prev = NULL;
13749 			if (dtrace_deferred_pid != NULL)
13750 				dtrace_deferred_pid->dthps_prev = help;
13751 			dtrace_deferred_pid = help;
13752 		}
13753 
13754 		mutex_exit(&dtrace_lock);
13755 
13756 	} else if (dofhp != NULL) {
13757 		/*
13758 		 * If the dtrace module is loaded and we have a particular
13759 		 * helper provider description, pass that off to the
13760 		 * meta provider.
13761 		 */
13762 
13763 		mutex_exit(&dtrace_lock);
13764 
13765 		dtrace_helper_provide(dofhp, p->p_pid);
13766 
13767 	} else {
13768 		/*
13769 		 * Otherwise, just pass all the helper provider descriptions
13770 		 * off to the meta provider.
13771 		 */
13772 
13773 		int i;
13774 		mutex_exit(&dtrace_lock);
13775 
13776 		for (i = 0; i < help->dthps_nprovs; i++) {
13777 			dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
13778 			    p->p_pid);
13779 		}
13780 	}
13781 
13782 	mutex_exit(&dtrace_meta_lock);
13783 }
13784 
13785 static int
13786 dtrace_helper_provider_add(dof_helper_t *dofhp, int gen)
13787 {
13788 	dtrace_helpers_t *help;
13789 	dtrace_helper_provider_t *hprov, **tmp_provs;
13790 	uint_t tmp_maxprovs, i;
13791 
13792 	ASSERT(MUTEX_HELD(&dtrace_lock));
13793 
13794 	help = curproc->p_dtrace_helpers;
13795 	ASSERT(help != NULL);
13796 
13797 	/*
13798 	 * If we already have dtrace_helper_providers_max helper providers,
13799 	 * we're refuse to add a new one.
13800 	 */
13801 	if (help->dthps_nprovs >= dtrace_helper_providers_max)
13802 		return (ENOSPC);
13803 
13804 	/*
13805 	 * Check to make sure this isn't a duplicate.
13806 	 */
13807 	for (i = 0; i < help->dthps_nprovs; i++) {
13808 		if (dofhp->dofhp_addr ==
13809 		    help->dthps_provs[i]->dthp_prov.dofhp_addr)
13810 			return (EALREADY);
13811 	}
13812 
13813 	hprov = kmem_zalloc(sizeof (dtrace_helper_provider_t), KM_SLEEP);
13814 	hprov->dthp_prov = *dofhp;
13815 	hprov->dthp_ref = 1;
13816 	hprov->dthp_generation = gen;
13817 
13818 	/*
13819 	 * Allocate a bigger table for helper providers if it's already full.
13820 	 */
13821 	if (help->dthps_maxprovs == help->dthps_nprovs) {
13822 		tmp_maxprovs = help->dthps_maxprovs;
13823 		tmp_provs = help->dthps_provs;
13824 
13825 		if (help->dthps_maxprovs == 0)
13826 			help->dthps_maxprovs = 2;
13827 		else
13828 			help->dthps_maxprovs *= 2;
13829 		if (help->dthps_maxprovs > dtrace_helper_providers_max)
13830 			help->dthps_maxprovs = dtrace_helper_providers_max;
13831 
13832 		ASSERT(tmp_maxprovs < help->dthps_maxprovs);
13833 
13834 		help->dthps_provs = kmem_zalloc(help->dthps_maxprovs *
13835 		    sizeof (dtrace_helper_provider_t *), KM_SLEEP);
13836 
13837 		if (tmp_provs != NULL) {
13838 			bcopy(tmp_provs, help->dthps_provs, tmp_maxprovs *
13839 			    sizeof (dtrace_helper_provider_t *));
13840 			kmem_free(tmp_provs, tmp_maxprovs *
13841 			    sizeof (dtrace_helper_provider_t *));
13842 		}
13843 	}
13844 
13845 	help->dthps_provs[help->dthps_nprovs] = hprov;
13846 	help->dthps_nprovs++;
13847 
13848 	return (0);
13849 }
13850 
13851 static void
13852 dtrace_helper_provider_destroy(dtrace_helper_provider_t *hprov)
13853 {
13854 	mutex_enter(&dtrace_lock);
13855 
13856 	if (--hprov->dthp_ref == 0) {
13857 		dof_hdr_t *dof;
13858 		mutex_exit(&dtrace_lock);
13859 		dof = (dof_hdr_t *)(uintptr_t)hprov->dthp_prov.dofhp_dof;
13860 		dtrace_dof_destroy(dof);
13861 		kmem_free(hprov, sizeof (dtrace_helper_provider_t));
13862 	} else {
13863 		mutex_exit(&dtrace_lock);
13864 	}
13865 }
13866 
13867 static int
13868 dtrace_helper_provider_validate(dof_hdr_t *dof, dof_sec_t *sec)
13869 {
13870 	uintptr_t daddr = (uintptr_t)dof;
13871 	dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
13872 	dof_provider_t *provider;
13873 	dof_probe_t *probe;
13874 	uint8_t *arg;
13875 	char *strtab, *typestr;
13876 	dof_stridx_t typeidx;
13877 	size_t typesz;
13878 	uint_t nprobes, j, k;
13879 
13880 	ASSERT(sec->dofs_type == DOF_SECT_PROVIDER);
13881 
13882 	if (sec->dofs_offset & (sizeof (uint_t) - 1)) {
13883 		dtrace_dof_error(dof, "misaligned section offset");
13884 		return (-1);
13885 	}
13886 
13887 	/*
13888 	 * The section needs to be large enough to contain the DOF provider
13889 	 * structure appropriate for the given version.
13890 	 */
13891 	if (sec->dofs_size <
13892 	    ((dof->dofh_ident[DOF_ID_VERSION] == DOF_VERSION_1) ?
13893 	    offsetof(dof_provider_t, dofpv_prenoffs) :
13894 	    sizeof (dof_provider_t))) {
13895 		dtrace_dof_error(dof, "provider section too small");
13896 		return (-1);
13897 	}
13898 
13899 	provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
13900 	str_sec = dtrace_dof_sect(dof, DOF_SECT_STRTAB, provider->dofpv_strtab);
13901 	prb_sec = dtrace_dof_sect(dof, DOF_SECT_PROBES, provider->dofpv_probes);
13902 	arg_sec = dtrace_dof_sect(dof, DOF_SECT_PRARGS, provider->dofpv_prargs);
13903 	off_sec = dtrace_dof_sect(dof, DOF_SECT_PROFFS, provider->dofpv_proffs);
13904 
13905 	if (str_sec == NULL || prb_sec == NULL ||
13906 	    arg_sec == NULL || off_sec == NULL)
13907 		return (-1);
13908 
13909 	enoff_sec = NULL;
13910 
13911 	if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
13912 	    provider->dofpv_prenoffs != DOF_SECT_NONE &&
13913 	    (enoff_sec = dtrace_dof_sect(dof, DOF_SECT_PRENOFFS,
13914 	    provider->dofpv_prenoffs)) == NULL)
13915 		return (-1);
13916 
13917 	strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
13918 
13919 	if (provider->dofpv_name >= str_sec->dofs_size ||
13920 	    strlen(strtab + provider->dofpv_name) >= DTRACE_PROVNAMELEN) {
13921 		dtrace_dof_error(dof, "invalid provider name");
13922 		return (-1);
13923 	}
13924 
13925 	if (prb_sec->dofs_entsize == 0 ||
13926 	    prb_sec->dofs_entsize > prb_sec->dofs_size) {
13927 		dtrace_dof_error(dof, "invalid entry size");
13928 		return (-1);
13929 	}
13930 
13931 	if (prb_sec->dofs_entsize & (sizeof (uintptr_t) - 1)) {
13932 		dtrace_dof_error(dof, "misaligned entry size");
13933 		return (-1);
13934 	}
13935 
13936 	if (off_sec->dofs_entsize != sizeof (uint32_t)) {
13937 		dtrace_dof_error(dof, "invalid entry size");
13938 		return (-1);
13939 	}
13940 
13941 	if (off_sec->dofs_offset & (sizeof (uint32_t) - 1)) {
13942 		dtrace_dof_error(dof, "misaligned section offset");
13943 		return (-1);
13944 	}
13945 
13946 	if (arg_sec->dofs_entsize != sizeof (uint8_t)) {
13947 		dtrace_dof_error(dof, "invalid entry size");
13948 		return (-1);
13949 	}
13950 
13951 	arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
13952 
13953 	nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
13954 
13955 	/*
13956 	 * Take a pass through the probes to check for errors.
13957 	 */
13958 	for (j = 0; j < nprobes; j++) {
13959 		probe = (dof_probe_t *)(uintptr_t)(daddr +
13960 		    prb_sec->dofs_offset + j * prb_sec->dofs_entsize);
13961 
13962 		if (probe->dofpr_func >= str_sec->dofs_size) {
13963 			dtrace_dof_error(dof, "invalid function name");
13964 			return (-1);
13965 		}
13966 
13967 		if (strlen(strtab + probe->dofpr_func) >= DTRACE_FUNCNAMELEN) {
13968 			dtrace_dof_error(dof, "function name too long");
13969 			return (-1);
13970 		}
13971 
13972 		if (probe->dofpr_name >= str_sec->dofs_size ||
13973 		    strlen(strtab + probe->dofpr_name) >= DTRACE_NAMELEN) {
13974 			dtrace_dof_error(dof, "invalid probe name");
13975 			return (-1);
13976 		}
13977 
13978 		/*
13979 		 * The offset count must not wrap the index, and the offsets
13980 		 * must also not overflow the section's data.
13981 		 */
13982 		if (probe->dofpr_offidx + probe->dofpr_noffs <
13983 		    probe->dofpr_offidx ||
13984 		    (probe->dofpr_offidx + probe->dofpr_noffs) *
13985 		    off_sec->dofs_entsize > off_sec->dofs_size) {
13986 			dtrace_dof_error(dof, "invalid probe offset");
13987 			return (-1);
13988 		}
13989 
13990 		if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1) {
13991 			/*
13992 			 * If there's no is-enabled offset section, make sure
13993 			 * there aren't any is-enabled offsets. Otherwise
13994 			 * perform the same checks as for probe offsets
13995 			 * (immediately above).
13996 			 */
13997 			if (enoff_sec == NULL) {
13998 				if (probe->dofpr_enoffidx != 0 ||
13999 				    probe->dofpr_nenoffs != 0) {
14000 					dtrace_dof_error(dof, "is-enabled "
14001 					    "offsets with null section");
14002 					return (-1);
14003 				}
14004 			} else if (probe->dofpr_enoffidx +
14005 			    probe->dofpr_nenoffs < probe->dofpr_enoffidx ||
14006 			    (probe->dofpr_enoffidx + probe->dofpr_nenoffs) *
14007 			    enoff_sec->dofs_entsize > enoff_sec->dofs_size) {
14008 				dtrace_dof_error(dof, "invalid is-enabled "
14009 				    "offset");
14010 				return (-1);
14011 			}
14012 
14013 			if (probe->dofpr_noffs + probe->dofpr_nenoffs == 0) {
14014 				dtrace_dof_error(dof, "zero probe and "
14015 				    "is-enabled offsets");
14016 				return (-1);
14017 			}
14018 		} else if (probe->dofpr_noffs == 0) {
14019 			dtrace_dof_error(dof, "zero probe offsets");
14020 			return (-1);
14021 		}
14022 
14023 		if (probe->dofpr_argidx + probe->dofpr_xargc <
14024 		    probe->dofpr_argidx ||
14025 		    (probe->dofpr_argidx + probe->dofpr_xargc) *
14026 		    arg_sec->dofs_entsize > arg_sec->dofs_size) {
14027 			dtrace_dof_error(dof, "invalid args");
14028 			return (-1);
14029 		}
14030 
14031 		typeidx = probe->dofpr_nargv;
14032 		typestr = strtab + probe->dofpr_nargv;
14033 		for (k = 0; k < probe->dofpr_nargc; k++) {
14034 			if (typeidx >= str_sec->dofs_size) {
14035 				dtrace_dof_error(dof, "bad "
14036 				    "native argument type");
14037 				return (-1);
14038 			}
14039 
14040 			typesz = strlen(typestr) + 1;
14041 			if (typesz > DTRACE_ARGTYPELEN) {
14042 				dtrace_dof_error(dof, "native "
14043 				    "argument type too long");
14044 				return (-1);
14045 			}
14046 			typeidx += typesz;
14047 			typestr += typesz;
14048 		}
14049 
14050 		typeidx = probe->dofpr_xargv;
14051 		typestr = strtab + probe->dofpr_xargv;
14052 		for (k = 0; k < probe->dofpr_xargc; k++) {
14053 			if (arg[probe->dofpr_argidx + k] > probe->dofpr_nargc) {
14054 				dtrace_dof_error(dof, "bad "
14055 				    "native argument index");
14056 				return (-1);
14057 			}
14058 
14059 			if (typeidx >= str_sec->dofs_size) {
14060 				dtrace_dof_error(dof, "bad "
14061 				    "translated argument type");
14062 				return (-1);
14063 			}
14064 
14065 			typesz = strlen(typestr) + 1;
14066 			if (typesz > DTRACE_ARGTYPELEN) {
14067 				dtrace_dof_error(dof, "translated argument "
14068 				    "type too long");
14069 				return (-1);
14070 			}
14071 
14072 			typeidx += typesz;
14073 			typestr += typesz;
14074 		}
14075 	}
14076 
14077 	return (0);
14078 }
14079 
14080 static int
14081 dtrace_helper_slurp(dof_hdr_t *dof, dof_helper_t *dhp)
14082 {
14083 	dtrace_helpers_t *help;
14084 	dtrace_vstate_t *vstate;
14085 	dtrace_enabling_t *enab = NULL;
14086 	int i, gen, rv, nhelpers = 0, nprovs = 0, destroy = 1;
14087 	uintptr_t daddr = (uintptr_t)dof;
14088 
14089 	ASSERT(MUTEX_HELD(&dtrace_lock));
14090 
14091 	if ((help = curproc->p_dtrace_helpers) == NULL)
14092 		help = dtrace_helpers_create(curproc);
14093 
14094 	vstate = &help->dthps_vstate;
14095 
14096 	if ((rv = dtrace_dof_slurp(dof, vstate, NULL, &enab,
14097 	    dhp != NULL ? dhp->dofhp_addr : 0, B_FALSE)) != 0) {
14098 		dtrace_dof_destroy(dof);
14099 		return (rv);
14100 	}
14101 
14102 	/*
14103 	 * Look for helper providers and validate their descriptions.
14104 	 */
14105 	if (dhp != NULL) {
14106 		for (i = 0; i < dof->dofh_secnum; i++) {
14107 			dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
14108 			    dof->dofh_secoff + i * dof->dofh_secsize);
14109 
14110 			if (sec->dofs_type != DOF_SECT_PROVIDER)
14111 				continue;
14112 
14113 			if (dtrace_helper_provider_validate(dof, sec) != 0) {
14114 				dtrace_enabling_destroy(enab);
14115 				dtrace_dof_destroy(dof);
14116 				return (-1);
14117 			}
14118 
14119 			nprovs++;
14120 		}
14121 	}
14122 
14123 	/*
14124 	 * Now we need to walk through the ECB descriptions in the enabling.
14125 	 */
14126 	for (i = 0; i < enab->dten_ndesc; i++) {
14127 		dtrace_ecbdesc_t *ep = enab->dten_desc[i];
14128 		dtrace_probedesc_t *desc = &ep->dted_probe;
14129 
14130 		if (strcmp(desc->dtpd_provider, "dtrace") != 0)
14131 			continue;
14132 
14133 		if (strcmp(desc->dtpd_mod, "helper") != 0)
14134 			continue;
14135 
14136 		if (strcmp(desc->dtpd_func, "ustack") != 0)
14137 			continue;
14138 
14139 		if ((rv = dtrace_helper_action_add(DTRACE_HELPER_ACTION_USTACK,
14140 		    ep)) != 0) {
14141 			/*
14142 			 * Adding this helper action failed -- we are now going
14143 			 * to rip out the entire generation and return failure.
14144 			 */
14145 			(void) dtrace_helper_destroygen(help->dthps_generation);
14146 			dtrace_enabling_destroy(enab);
14147 			dtrace_dof_destroy(dof);
14148 			return (-1);
14149 		}
14150 
14151 		nhelpers++;
14152 	}
14153 
14154 	if (nhelpers < enab->dten_ndesc)
14155 		dtrace_dof_error(dof, "unmatched helpers");
14156 
14157 	gen = help->dthps_generation++;
14158 	dtrace_enabling_destroy(enab);
14159 
14160 	if (dhp != NULL && nprovs > 0) {
14161 		dhp->dofhp_dof = (uint64_t)(uintptr_t)dof;
14162 		if (dtrace_helper_provider_add(dhp, gen) == 0) {
14163 			mutex_exit(&dtrace_lock);
14164 			dtrace_helper_provider_register(curproc, help, dhp);
14165 			mutex_enter(&dtrace_lock);
14166 
14167 			destroy = 0;
14168 		}
14169 	}
14170 
14171 	if (destroy)
14172 		dtrace_dof_destroy(dof);
14173 
14174 	return (gen);
14175 }
14176 
14177 static dtrace_helpers_t *
14178 dtrace_helpers_create(proc_t *p)
14179 {
14180 	dtrace_helpers_t *help;
14181 
14182 	ASSERT(MUTEX_HELD(&dtrace_lock));
14183 	ASSERT(p->p_dtrace_helpers == NULL);
14184 
14185 	help = kmem_zalloc(sizeof (dtrace_helpers_t), KM_SLEEP);
14186 	help->dthps_actions = kmem_zalloc(sizeof (dtrace_helper_action_t *) *
14187 	    DTRACE_NHELPER_ACTIONS, KM_SLEEP);
14188 
14189 	p->p_dtrace_helpers = help;
14190 	dtrace_helpers++;
14191 
14192 	return (help);
14193 }
14194 
14195 static void
14196 dtrace_helpers_destroy(void)
14197 {
14198 	dtrace_helpers_t *help;
14199 	dtrace_vstate_t *vstate;
14200 	proc_t *p = curproc;
14201 	int i;
14202 
14203 	mutex_enter(&dtrace_lock);
14204 
14205 	ASSERT(p->p_dtrace_helpers != NULL);
14206 	ASSERT(dtrace_helpers > 0);
14207 
14208 	help = p->p_dtrace_helpers;
14209 	vstate = &help->dthps_vstate;
14210 
14211 	/*
14212 	 * We're now going to lose the help from this process.
14213 	 */
14214 	p->p_dtrace_helpers = NULL;
14215 	dtrace_sync();
14216 
14217 	/*
14218 	 * Destory the helper actions.
14219 	 */
14220 	for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
14221 		dtrace_helper_action_t *h, *next;
14222 
14223 		for (h = help->dthps_actions[i]; h != NULL; h = next) {
14224 			next = h->dtha_next;
14225 			dtrace_helper_action_destroy(h, vstate);
14226 			h = next;
14227 		}
14228 	}
14229 
14230 	mutex_exit(&dtrace_lock);
14231 
14232 	/*
14233 	 * Destroy the helper providers.
14234 	 */
14235 	if (help->dthps_maxprovs > 0) {
14236 		mutex_enter(&dtrace_meta_lock);
14237 		if (dtrace_meta_pid != NULL) {
14238 			ASSERT(dtrace_deferred_pid == NULL);
14239 
14240 			for (i = 0; i < help->dthps_nprovs; i++) {
14241 				dtrace_helper_provider_remove(
14242 				    &help->dthps_provs[i]->dthp_prov, p->p_pid);
14243 			}
14244 		} else {
14245 			mutex_enter(&dtrace_lock);
14246 			ASSERT(help->dthps_deferred == 0 ||
14247 			    help->dthps_next != NULL ||
14248 			    help->dthps_prev != NULL ||
14249 			    help == dtrace_deferred_pid);
14250 
14251 			/*
14252 			 * Remove the helper from the deferred list.
14253 			 */
14254 			if (help->dthps_next != NULL)
14255 				help->dthps_next->dthps_prev = help->dthps_prev;
14256 			if (help->dthps_prev != NULL)
14257 				help->dthps_prev->dthps_next = help->dthps_next;
14258 			if (dtrace_deferred_pid == help) {
14259 				dtrace_deferred_pid = help->dthps_next;
14260 				ASSERT(help->dthps_prev == NULL);
14261 			}
14262 
14263 			mutex_exit(&dtrace_lock);
14264 		}
14265 
14266 		mutex_exit(&dtrace_meta_lock);
14267 
14268 		for (i = 0; i < help->dthps_nprovs; i++) {
14269 			dtrace_helper_provider_destroy(help->dthps_provs[i]);
14270 		}
14271 
14272 		kmem_free(help->dthps_provs, help->dthps_maxprovs *
14273 		    sizeof (dtrace_helper_provider_t *));
14274 	}
14275 
14276 	mutex_enter(&dtrace_lock);
14277 
14278 	dtrace_vstate_fini(&help->dthps_vstate);
14279 	kmem_free(help->dthps_actions,
14280 	    sizeof (dtrace_helper_action_t *) * DTRACE_NHELPER_ACTIONS);
14281 	kmem_free(help, sizeof (dtrace_helpers_t));
14282 
14283 	--dtrace_helpers;
14284 	mutex_exit(&dtrace_lock);
14285 }
14286 
14287 static void
14288 dtrace_helpers_duplicate(proc_t *from, proc_t *to)
14289 {
14290 	dtrace_helpers_t *help, *newhelp;
14291 	dtrace_helper_action_t *helper, *new, *last;
14292 	dtrace_difo_t *dp;
14293 	dtrace_vstate_t *vstate;
14294 	int i, j, sz, hasprovs = 0;
14295 
14296 	mutex_enter(&dtrace_lock);
14297 	ASSERT(from->p_dtrace_helpers != NULL);
14298 	ASSERT(dtrace_helpers > 0);
14299 
14300 	help = from->p_dtrace_helpers;
14301 	newhelp = dtrace_helpers_create(to);
14302 	ASSERT(to->p_dtrace_helpers != NULL);
14303 
14304 	newhelp->dthps_generation = help->dthps_generation;
14305 	vstate = &newhelp->dthps_vstate;
14306 
14307 	/*
14308 	 * Duplicate the helper actions.
14309 	 */
14310 	for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
14311 		if ((helper = help->dthps_actions[i]) == NULL)
14312 			continue;
14313 
14314 		for (last = NULL; helper != NULL; helper = helper->dtha_next) {
14315 			new = kmem_zalloc(sizeof (dtrace_helper_action_t),
14316 			    KM_SLEEP);
14317 			new->dtha_generation = helper->dtha_generation;
14318 
14319 			if ((dp = helper->dtha_predicate) != NULL) {
14320 				dp = dtrace_difo_duplicate(dp, vstate);
14321 				new->dtha_predicate = dp;
14322 			}
14323 
14324 			new->dtha_nactions = helper->dtha_nactions;
14325 			sz = sizeof (dtrace_difo_t *) * new->dtha_nactions;
14326 			new->dtha_actions = kmem_alloc(sz, KM_SLEEP);
14327 
14328 			for (j = 0; j < new->dtha_nactions; j++) {
14329 				dtrace_difo_t *dp = helper->dtha_actions[j];
14330 
14331 				ASSERT(dp != NULL);
14332 				dp = dtrace_difo_duplicate(dp, vstate);
14333 				new->dtha_actions[j] = dp;
14334 			}
14335 
14336 			if (last != NULL) {
14337 				last->dtha_next = new;
14338 			} else {
14339 				newhelp->dthps_actions[i] = new;
14340 			}
14341 
14342 			last = new;
14343 		}
14344 	}
14345 
14346 	/*
14347 	 * Duplicate the helper providers and register them with the
14348 	 * DTrace framework.
14349 	 */
14350 	if (help->dthps_nprovs > 0) {
14351 		newhelp->dthps_nprovs = help->dthps_nprovs;
14352 		newhelp->dthps_maxprovs = help->dthps_nprovs;
14353 		newhelp->dthps_provs = kmem_alloc(newhelp->dthps_nprovs *
14354 		    sizeof (dtrace_helper_provider_t *), KM_SLEEP);
14355 		for (i = 0; i < newhelp->dthps_nprovs; i++) {
14356 			newhelp->dthps_provs[i] = help->dthps_provs[i];
14357 			newhelp->dthps_provs[i]->dthp_ref++;
14358 		}
14359 
14360 		hasprovs = 1;
14361 	}
14362 
14363 	mutex_exit(&dtrace_lock);
14364 
14365 	if (hasprovs)
14366 		dtrace_helper_provider_register(to, newhelp, NULL);
14367 }
14368 
14369 /*
14370  * DTrace Hook Functions
14371  */
14372 static void
14373 dtrace_module_loaded(struct modctl *ctl)
14374 {
14375 	dtrace_provider_t *prv;
14376 
14377 	mutex_enter(&dtrace_provider_lock);
14378 	mutex_enter(&mod_lock);
14379 
14380 	ASSERT(ctl->mod_busy);
14381 
14382 	/*
14383 	 * We're going to call each providers per-module provide operation
14384 	 * specifying only this module.
14385 	 */
14386 	for (prv = dtrace_provider; prv != NULL; prv = prv->dtpv_next)
14387 		prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
14388 
14389 	mutex_exit(&mod_lock);
14390 	mutex_exit(&dtrace_provider_lock);
14391 
14392 	/*
14393 	 * If we have any retained enablings, we need to match against them.
14394 	 * Enabling probes requires that cpu_lock be held, and we cannot hold
14395 	 * cpu_lock here -- it is legal for cpu_lock to be held when loading a
14396 	 * module.  (In particular, this happens when loading scheduling
14397 	 * classes.)  So if we have any retained enablings, we need to dispatch
14398 	 * our task queue to do the match for us.
14399 	 */
14400 	mutex_enter(&dtrace_lock);
14401 
14402 	if (dtrace_retained == NULL) {
14403 		mutex_exit(&dtrace_lock);
14404 		return;
14405 	}
14406 
14407 	(void) taskq_dispatch(dtrace_taskq,
14408 	    (task_func_t *)dtrace_enabling_matchall, NULL, TQ_SLEEP);
14409 
14410 	mutex_exit(&dtrace_lock);
14411 
14412 	/*
14413 	 * And now, for a little heuristic sleaze:  in general, we want to
14414 	 * match modules as soon as they load.  However, we cannot guarantee
14415 	 * this, because it would lead us to the lock ordering violation
14416 	 * outlined above.  The common case, of course, is that cpu_lock is
14417 	 * _not_ held -- so we delay here for a clock tick, hoping that that's
14418 	 * long enough for the task queue to do its work.  If it's not, it's
14419 	 * not a serious problem -- it just means that the module that we
14420 	 * just loaded may not be immediately instrumentable.
14421 	 */
14422 	delay(1);
14423 }
14424 
14425 static void
14426 dtrace_module_unloaded(struct modctl *ctl)
14427 {
14428 	dtrace_probe_t template, *probe, *first, *next;
14429 	dtrace_provider_t *prov;
14430 
14431 	template.dtpr_mod = ctl->mod_modname;
14432 
14433 	mutex_enter(&dtrace_provider_lock);
14434 	mutex_enter(&mod_lock);
14435 	mutex_enter(&dtrace_lock);
14436 
14437 	if (dtrace_bymod == NULL) {
14438 		/*
14439 		 * The DTrace module is loaded (obviously) but not attached;
14440 		 * we don't have any work to do.
14441 		 */
14442 		mutex_exit(&dtrace_provider_lock);
14443 		mutex_exit(&mod_lock);
14444 		mutex_exit(&dtrace_lock);
14445 		return;
14446 	}
14447 
14448 	for (probe = first = dtrace_hash_lookup(dtrace_bymod, &template);
14449 	    probe != NULL; probe = probe->dtpr_nextmod) {
14450 		if (probe->dtpr_ecb != NULL) {
14451 			mutex_exit(&dtrace_provider_lock);
14452 			mutex_exit(&mod_lock);
14453 			mutex_exit(&dtrace_lock);
14454 
14455 			/*
14456 			 * This shouldn't _actually_ be possible -- we're
14457 			 * unloading a module that has an enabled probe in it.
14458 			 * (It's normally up to the provider to make sure that
14459 			 * this can't happen.)  However, because dtps_enable()
14460 			 * doesn't have a failure mode, there can be an
14461 			 * enable/unload race.  Upshot:  we don't want to
14462 			 * assert, but we're not going to disable the
14463 			 * probe, either.
14464 			 */
14465 			if (dtrace_err_verbose) {
14466 				cmn_err(CE_WARN, "unloaded module '%s' had "
14467 				    "enabled probes", ctl->mod_modname);
14468 			}
14469 
14470 			return;
14471 		}
14472 	}
14473 
14474 	probe = first;
14475 
14476 	for (first = NULL; probe != NULL; probe = next) {
14477 		ASSERT(dtrace_probes[probe->dtpr_id - 1] == probe);
14478 
14479 		dtrace_probes[probe->dtpr_id - 1] = NULL;
14480 
14481 		next = probe->dtpr_nextmod;
14482 		dtrace_hash_remove(dtrace_bymod, probe);
14483 		dtrace_hash_remove(dtrace_byfunc, probe);
14484 		dtrace_hash_remove(dtrace_byname, probe);
14485 
14486 		if (first == NULL) {
14487 			first = probe;
14488 			probe->dtpr_nextmod = NULL;
14489 		} else {
14490 			probe->dtpr_nextmod = first;
14491 			first = probe;
14492 		}
14493 	}
14494 
14495 	/*
14496 	 * We've removed all of the module's probes from the hash chains and
14497 	 * from the probe array.  Now issue a dtrace_sync() to be sure that
14498 	 * everyone has cleared out from any probe array processing.
14499 	 */
14500 	dtrace_sync();
14501 
14502 	for (probe = first; probe != NULL; probe = first) {
14503 		first = probe->dtpr_nextmod;
14504 		prov = probe->dtpr_provider;
14505 		prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, probe->dtpr_id,
14506 		    probe->dtpr_arg);
14507 		kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
14508 		kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
14509 		kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
14510 		vmem_free(dtrace_arena, (void *)(uintptr_t)probe->dtpr_id, 1);
14511 		kmem_free(probe, sizeof (dtrace_probe_t));
14512 	}
14513 
14514 	mutex_exit(&dtrace_lock);
14515 	mutex_exit(&mod_lock);
14516 	mutex_exit(&dtrace_provider_lock);
14517 }
14518 
14519 void
14520 dtrace_suspend(void)
14521 {
14522 	dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_suspend));
14523 }
14524 
14525 void
14526 dtrace_resume(void)
14527 {
14528 	dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_resume));
14529 }
14530 
14531 static int
14532 dtrace_cpu_setup(cpu_setup_t what, processorid_t cpu)
14533 {
14534 	ASSERT(MUTEX_HELD(&cpu_lock));
14535 	mutex_enter(&dtrace_lock);
14536 
14537 	switch (what) {
14538 	case CPU_CONFIG: {
14539 		dtrace_state_t *state;
14540 		dtrace_optval_t *opt, rs, c;
14541 
14542 		/*
14543 		 * For now, we only allocate a new buffer for anonymous state.
14544 		 */
14545 		if ((state = dtrace_anon.dta_state) == NULL)
14546 			break;
14547 
14548 		if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
14549 			break;
14550 
14551 		opt = state->dts_options;
14552 		c = opt[DTRACEOPT_CPU];
14553 
14554 		if (c != DTRACE_CPUALL && c != DTRACEOPT_UNSET && c != cpu)
14555 			break;
14556 
14557 		/*
14558 		 * Regardless of what the actual policy is, we're going to
14559 		 * temporarily set our resize policy to be manual.  We're
14560 		 * also going to temporarily set our CPU option to denote
14561 		 * the newly configured CPU.
14562 		 */
14563 		rs = opt[DTRACEOPT_BUFRESIZE];
14564 		opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_MANUAL;
14565 		opt[DTRACEOPT_CPU] = (dtrace_optval_t)cpu;
14566 
14567 		(void) dtrace_state_buffers(state);
14568 
14569 		opt[DTRACEOPT_BUFRESIZE] = rs;
14570 		opt[DTRACEOPT_CPU] = c;
14571 
14572 		break;
14573 	}
14574 
14575 	case CPU_UNCONFIG:
14576 		/*
14577 		 * We don't free the buffer in the CPU_UNCONFIG case.  (The
14578 		 * buffer will be freed when the consumer exits.)
14579 		 */
14580 		break;
14581 
14582 	default:
14583 		break;
14584 	}
14585 
14586 	mutex_exit(&dtrace_lock);
14587 	return (0);
14588 }
14589 
14590 static void
14591 dtrace_cpu_setup_initial(processorid_t cpu)
14592 {
14593 	(void) dtrace_cpu_setup(CPU_CONFIG, cpu);
14594 }
14595 
14596 static void
14597 dtrace_toxrange_add(uintptr_t base, uintptr_t limit)
14598 {
14599 	if (dtrace_toxranges >= dtrace_toxranges_max) {
14600 		int osize, nsize;
14601 		dtrace_toxrange_t *range;
14602 
14603 		osize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
14604 
14605 		if (osize == 0) {
14606 			ASSERT(dtrace_toxrange == NULL);
14607 			ASSERT(dtrace_toxranges_max == 0);
14608 			dtrace_toxranges_max = 1;
14609 		} else {
14610 			dtrace_toxranges_max <<= 1;
14611 		}
14612 
14613 		nsize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
14614 		range = kmem_zalloc(nsize, KM_SLEEP);
14615 
14616 		if (dtrace_toxrange != NULL) {
14617 			ASSERT(osize != 0);
14618 			bcopy(dtrace_toxrange, range, osize);
14619 			kmem_free(dtrace_toxrange, osize);
14620 		}
14621 
14622 		dtrace_toxrange = range;
14623 	}
14624 
14625 	ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_base == NULL);
14626 	ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_limit == NULL);
14627 
14628 	dtrace_toxrange[dtrace_toxranges].dtt_base = base;
14629 	dtrace_toxrange[dtrace_toxranges].dtt_limit = limit;
14630 	dtrace_toxranges++;
14631 }
14632 
14633 /*
14634  * DTrace Driver Cookbook Functions
14635  */
14636 /*ARGSUSED*/
14637 static int
14638 dtrace_attach(dev_info_t *devi, ddi_attach_cmd_t cmd)
14639 {
14640 	dtrace_provider_id_t id;
14641 	dtrace_state_t *state = NULL;
14642 	dtrace_enabling_t *enab;
14643 
14644 	mutex_enter(&cpu_lock);
14645 	mutex_enter(&dtrace_provider_lock);
14646 	mutex_enter(&dtrace_lock);
14647 
14648 	if (ddi_soft_state_init(&dtrace_softstate,
14649 	    sizeof (dtrace_state_t), 0) != 0) {
14650 		cmn_err(CE_NOTE, "/dev/dtrace failed to initialize soft state");
14651 		mutex_exit(&cpu_lock);
14652 		mutex_exit(&dtrace_provider_lock);
14653 		mutex_exit(&dtrace_lock);
14654 		return (DDI_FAILURE);
14655 	}
14656 
14657 	if (ddi_create_minor_node(devi, DTRACEMNR_DTRACE, S_IFCHR,
14658 	    DTRACEMNRN_DTRACE, DDI_PSEUDO, NULL) == DDI_FAILURE ||
14659 	    ddi_create_minor_node(devi, DTRACEMNR_HELPER, S_IFCHR,
14660 	    DTRACEMNRN_HELPER, DDI_PSEUDO, NULL) == DDI_FAILURE) {
14661 		cmn_err(CE_NOTE, "/dev/dtrace couldn't create minor nodes");
14662 		ddi_remove_minor_node(devi, NULL);
14663 		ddi_soft_state_fini(&dtrace_softstate);
14664 		mutex_exit(&cpu_lock);
14665 		mutex_exit(&dtrace_provider_lock);
14666 		mutex_exit(&dtrace_lock);
14667 		return (DDI_FAILURE);
14668 	}
14669 
14670 	ddi_report_dev(devi);
14671 	dtrace_devi = devi;
14672 
14673 	dtrace_modload = dtrace_module_loaded;
14674 	dtrace_modunload = dtrace_module_unloaded;
14675 	dtrace_cpu_init = dtrace_cpu_setup_initial;
14676 	dtrace_helpers_cleanup = dtrace_helpers_destroy;
14677 	dtrace_helpers_fork = dtrace_helpers_duplicate;
14678 	dtrace_cpustart_init = dtrace_suspend;
14679 	dtrace_cpustart_fini = dtrace_resume;
14680 	dtrace_debugger_init = dtrace_suspend;
14681 	dtrace_debugger_fini = dtrace_resume;
14682 
14683 	register_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
14684 
14685 	ASSERT(MUTEX_HELD(&cpu_lock));
14686 
14687 	dtrace_arena = vmem_create("dtrace", (void *)1, UINT32_MAX, 1,
14688 	    NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
14689 	dtrace_minor = vmem_create("dtrace_minor", (void *)DTRACEMNRN_CLONE,
14690 	    UINT32_MAX - DTRACEMNRN_CLONE, 1, NULL, NULL, NULL, 0,
14691 	    VM_SLEEP | VMC_IDENTIFIER);
14692 	dtrace_taskq = taskq_create("dtrace_taskq", 1, maxclsyspri,
14693 	    1, INT_MAX, 0);
14694 
14695 	dtrace_state_cache = kmem_cache_create("dtrace_state_cache",
14696 	    sizeof (dtrace_dstate_percpu_t) * NCPU, DTRACE_STATE_ALIGN,
14697 	    NULL, NULL, NULL, NULL, NULL, 0);
14698 
14699 	ASSERT(MUTEX_HELD(&cpu_lock));
14700 	dtrace_bymod = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_mod),
14701 	    offsetof(dtrace_probe_t, dtpr_nextmod),
14702 	    offsetof(dtrace_probe_t, dtpr_prevmod));
14703 
14704 	dtrace_byfunc = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_func),
14705 	    offsetof(dtrace_probe_t, dtpr_nextfunc),
14706 	    offsetof(dtrace_probe_t, dtpr_prevfunc));
14707 
14708 	dtrace_byname = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_name),
14709 	    offsetof(dtrace_probe_t, dtpr_nextname),
14710 	    offsetof(dtrace_probe_t, dtpr_prevname));
14711 
14712 	if (dtrace_retain_max < 1) {
14713 		cmn_err(CE_WARN, "illegal value (%lu) for dtrace_retain_max; "
14714 		    "setting to 1", dtrace_retain_max);
14715 		dtrace_retain_max = 1;
14716 	}
14717 
14718 	/*
14719 	 * Now discover our toxic ranges.
14720 	 */
14721 	dtrace_toxic_ranges(dtrace_toxrange_add);
14722 
14723 	/*
14724 	 * Before we register ourselves as a provider to our own framework,
14725 	 * we would like to assert that dtrace_provider is NULL -- but that's
14726 	 * not true if we were loaded as a dependency of a DTrace provider.
14727 	 * Once we've registered, we can assert that dtrace_provider is our
14728 	 * pseudo provider.
14729 	 */
14730 	(void) dtrace_register("dtrace", &dtrace_provider_attr,
14731 	    DTRACE_PRIV_NONE, 0, &dtrace_provider_ops, NULL, &id);
14732 
14733 	ASSERT(dtrace_provider != NULL);
14734 	ASSERT((dtrace_provider_id_t)dtrace_provider == id);
14735 
14736 	dtrace_probeid_begin = dtrace_probe_create((dtrace_provider_id_t)
14737 	    dtrace_provider, NULL, NULL, "BEGIN", 0, NULL);
14738 	dtrace_probeid_end = dtrace_probe_create((dtrace_provider_id_t)
14739 	    dtrace_provider, NULL, NULL, "END", 0, NULL);
14740 	dtrace_probeid_error = dtrace_probe_create((dtrace_provider_id_t)
14741 	    dtrace_provider, NULL, NULL, "ERROR", 1, NULL);
14742 
14743 	dtrace_anon_property();
14744 	mutex_exit(&cpu_lock);
14745 
14746 	/*
14747 	 * If DTrace helper tracing is enabled, we need to allocate the
14748 	 * trace buffer and initialize the values.
14749 	 */
14750 	if (dtrace_helptrace_enabled) {
14751 		ASSERT(dtrace_helptrace_buffer == NULL);
14752 		dtrace_helptrace_buffer =
14753 		    kmem_zalloc(dtrace_helptrace_bufsize, KM_SLEEP);
14754 		dtrace_helptrace_next = 0;
14755 	}
14756 
14757 	/*
14758 	 * If there are already providers, we must ask them to provide their
14759 	 * probes, and then match any anonymous enabling against them.  Note
14760 	 * that there should be no other retained enablings at this time:
14761 	 * the only retained enablings at this time should be the anonymous
14762 	 * enabling.
14763 	 */
14764 	if (dtrace_anon.dta_enabling != NULL) {
14765 		ASSERT(dtrace_retained == dtrace_anon.dta_enabling);
14766 
14767 		dtrace_enabling_provide(NULL);
14768 		state = dtrace_anon.dta_state;
14769 
14770 		/*
14771 		 * We couldn't hold cpu_lock across the above call to
14772 		 * dtrace_enabling_provide(), but we must hold it to actually
14773 		 * enable the probes.  We have to drop all of our locks, pick
14774 		 * up cpu_lock, and regain our locks before matching the
14775 		 * retained anonymous enabling.
14776 		 */
14777 		mutex_exit(&dtrace_lock);
14778 		mutex_exit(&dtrace_provider_lock);
14779 
14780 		mutex_enter(&cpu_lock);
14781 		mutex_enter(&dtrace_provider_lock);
14782 		mutex_enter(&dtrace_lock);
14783 
14784 		if ((enab = dtrace_anon.dta_enabling) != NULL)
14785 			(void) dtrace_enabling_match(enab, NULL);
14786 
14787 		mutex_exit(&cpu_lock);
14788 	}
14789 
14790 	mutex_exit(&dtrace_lock);
14791 	mutex_exit(&dtrace_provider_lock);
14792 
14793 	if (state != NULL) {
14794 		/*
14795 		 * If we created any anonymous state, set it going now.
14796 		 */
14797 		(void) dtrace_state_go(state, &dtrace_anon.dta_beganon);
14798 	}
14799 
14800 	return (DDI_SUCCESS);
14801 }
14802 
14803 /*ARGSUSED*/
14804 static int
14805 dtrace_open(dev_t *devp, int flag, int otyp, cred_t *cred_p)
14806 {
14807 	dtrace_state_t *state;
14808 	uint32_t priv;
14809 	uid_t uid;
14810 	zoneid_t zoneid;
14811 
14812 	if (getminor(*devp) == DTRACEMNRN_HELPER)
14813 		return (0);
14814 
14815 	/*
14816 	 * If this wasn't an open with the "helper" minor, then it must be
14817 	 * the "dtrace" minor.
14818 	 */
14819 	if (getminor(*devp) != DTRACEMNRN_DTRACE)
14820 		return (ENXIO);
14821 
14822 	/*
14823 	 * If no DTRACE_PRIV_* bits are set in the credential, then the
14824 	 * caller lacks sufficient permission to do anything with DTrace.
14825 	 */
14826 	dtrace_cred2priv(cred_p, &priv, &uid, &zoneid);
14827 	if (priv == DTRACE_PRIV_NONE)
14828 		return (EACCES);
14829 
14830 	/*
14831 	 * Ask all providers to provide all their probes.
14832 	 */
14833 	mutex_enter(&dtrace_provider_lock);
14834 	dtrace_probe_provide(NULL, NULL);
14835 	mutex_exit(&dtrace_provider_lock);
14836 
14837 	mutex_enter(&cpu_lock);
14838 	mutex_enter(&dtrace_lock);
14839 	dtrace_opens++;
14840 	dtrace_membar_producer();
14841 
14842 	/*
14843 	 * If the kernel debugger is active (that is, if the kernel debugger
14844 	 * modified text in some way), we won't allow the open.
14845 	 */
14846 	if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
14847 		dtrace_opens--;
14848 		mutex_exit(&cpu_lock);
14849 		mutex_exit(&dtrace_lock);
14850 		return (EBUSY);
14851 	}
14852 
14853 	state = dtrace_state_create(devp, cred_p);
14854 	mutex_exit(&cpu_lock);
14855 
14856 	if (state == NULL) {
14857 		if (--dtrace_opens == 0 && dtrace_anon.dta_enabling == NULL)
14858 			(void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
14859 		mutex_exit(&dtrace_lock);
14860 		return (EAGAIN);
14861 	}
14862 
14863 	mutex_exit(&dtrace_lock);
14864 
14865 	return (0);
14866 }
14867 
14868 /*ARGSUSED*/
14869 static int
14870 dtrace_close(dev_t dev, int flag, int otyp, cred_t *cred_p)
14871 {
14872 	minor_t minor = getminor(dev);
14873 	dtrace_state_t *state;
14874 
14875 	if (minor == DTRACEMNRN_HELPER)
14876 		return (0);
14877 
14878 	state = ddi_get_soft_state(dtrace_softstate, minor);
14879 
14880 	mutex_enter(&cpu_lock);
14881 	mutex_enter(&dtrace_lock);
14882 
14883 	if (state->dts_anon) {
14884 		/*
14885 		 * There is anonymous state. Destroy that first.
14886 		 */
14887 		ASSERT(dtrace_anon.dta_state == NULL);
14888 		dtrace_state_destroy(state->dts_anon);
14889 	}
14890 
14891 	dtrace_state_destroy(state);
14892 	ASSERT(dtrace_opens > 0);
14893 
14894 	/*
14895 	 * Only relinquish control of the kernel debugger interface when there
14896 	 * are no consumers and no anonymous enablings.
14897 	 */
14898 	if (--dtrace_opens == 0 && dtrace_anon.dta_enabling == NULL)
14899 		(void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
14900 
14901 	mutex_exit(&dtrace_lock);
14902 	mutex_exit(&cpu_lock);
14903 
14904 	return (0);
14905 }
14906 
14907 /*ARGSUSED*/
14908 static int
14909 dtrace_ioctl_helper(int cmd, intptr_t arg, int *rv)
14910 {
14911 	int rval;
14912 	dof_helper_t help, *dhp = NULL;
14913 
14914 	switch (cmd) {
14915 	case DTRACEHIOC_ADDDOF:
14916 		if (copyin((void *)arg, &help, sizeof (help)) != 0) {
14917 			dtrace_dof_error(NULL, "failed to copyin DOF helper");
14918 			return (EFAULT);
14919 		}
14920 
14921 		dhp = &help;
14922 		arg = (intptr_t)help.dofhp_dof;
14923 		/*FALLTHROUGH*/
14924 
14925 	case DTRACEHIOC_ADD: {
14926 		dof_hdr_t *dof = dtrace_dof_copyin(arg, &rval);
14927 
14928 		if (dof == NULL)
14929 			return (rval);
14930 
14931 		mutex_enter(&dtrace_lock);
14932 
14933 		/*
14934 		 * dtrace_helper_slurp() takes responsibility for the dof --
14935 		 * it may free it now or it may save it and free it later.
14936 		 */
14937 		if ((rval = dtrace_helper_slurp(dof, dhp)) != -1) {
14938 			*rv = rval;
14939 			rval = 0;
14940 		} else {
14941 			rval = EINVAL;
14942 		}
14943 
14944 		mutex_exit(&dtrace_lock);
14945 		return (rval);
14946 	}
14947 
14948 	case DTRACEHIOC_REMOVE: {
14949 		mutex_enter(&dtrace_lock);
14950 		rval = dtrace_helper_destroygen(arg);
14951 		mutex_exit(&dtrace_lock);
14952 
14953 		return (rval);
14954 	}
14955 
14956 	default:
14957 		break;
14958 	}
14959 
14960 	return (ENOTTY);
14961 }
14962 
14963 /*ARGSUSED*/
14964 static int
14965 dtrace_ioctl(dev_t dev, int cmd, intptr_t arg, int md, cred_t *cr, int *rv)
14966 {
14967 	minor_t minor = getminor(dev);
14968 	dtrace_state_t *state;
14969 	int rval;
14970 
14971 	if (minor == DTRACEMNRN_HELPER)
14972 		return (dtrace_ioctl_helper(cmd, arg, rv));
14973 
14974 	state = ddi_get_soft_state(dtrace_softstate, minor);
14975 
14976 	if (state->dts_anon) {
14977 		ASSERT(dtrace_anon.dta_state == NULL);
14978 		state = state->dts_anon;
14979 	}
14980 
14981 	switch (cmd) {
14982 	case DTRACEIOC_PROVIDER: {
14983 		dtrace_providerdesc_t pvd;
14984 		dtrace_provider_t *pvp;
14985 
14986 		if (copyin((void *)arg, &pvd, sizeof (pvd)) != 0)
14987 			return (EFAULT);
14988 
14989 		pvd.dtvd_name[DTRACE_PROVNAMELEN - 1] = '\0';
14990 		mutex_enter(&dtrace_provider_lock);
14991 
14992 		for (pvp = dtrace_provider; pvp != NULL; pvp = pvp->dtpv_next) {
14993 			if (strcmp(pvp->dtpv_name, pvd.dtvd_name) == 0)
14994 				break;
14995 		}
14996 
14997 		mutex_exit(&dtrace_provider_lock);
14998 
14999 		if (pvp == NULL)
15000 			return (ESRCH);
15001 
15002 		bcopy(&pvp->dtpv_priv, &pvd.dtvd_priv, sizeof (dtrace_ppriv_t));
15003 		bcopy(&pvp->dtpv_attr, &pvd.dtvd_attr, sizeof (dtrace_pattr_t));
15004 		if (copyout(&pvd, (void *)arg, sizeof (pvd)) != 0)
15005 			return (EFAULT);
15006 
15007 		return (0);
15008 	}
15009 
15010 	case DTRACEIOC_EPROBE: {
15011 		dtrace_eprobedesc_t epdesc;
15012 		dtrace_ecb_t *ecb;
15013 		dtrace_action_t *act;
15014 		void *buf;
15015 		size_t size;
15016 		uintptr_t dest;
15017 		int nrecs;
15018 
15019 		if (copyin((void *)arg, &epdesc, sizeof (epdesc)) != 0)
15020 			return (EFAULT);
15021 
15022 		mutex_enter(&dtrace_lock);
15023 
15024 		if ((ecb = dtrace_epid2ecb(state, epdesc.dtepd_epid)) == NULL) {
15025 			mutex_exit(&dtrace_lock);
15026 			return (EINVAL);
15027 		}
15028 
15029 		if (ecb->dte_probe == NULL) {
15030 			mutex_exit(&dtrace_lock);
15031 			return (EINVAL);
15032 		}
15033 
15034 		epdesc.dtepd_probeid = ecb->dte_probe->dtpr_id;
15035 		epdesc.dtepd_uarg = ecb->dte_uarg;
15036 		epdesc.dtepd_size = ecb->dte_size;
15037 
15038 		nrecs = epdesc.dtepd_nrecs;
15039 		epdesc.dtepd_nrecs = 0;
15040 		for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
15041 			if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
15042 				continue;
15043 
15044 			epdesc.dtepd_nrecs++;
15045 		}
15046 
15047 		/*
15048 		 * Now that we have the size, we need to allocate a temporary
15049 		 * buffer in which to store the complete description.  We need
15050 		 * the temporary buffer to be able to drop dtrace_lock()
15051 		 * across the copyout(), below.
15052 		 */
15053 		size = sizeof (dtrace_eprobedesc_t) +
15054 		    (epdesc.dtepd_nrecs * sizeof (dtrace_recdesc_t));
15055 
15056 		buf = kmem_alloc(size, KM_SLEEP);
15057 		dest = (uintptr_t)buf;
15058 
15059 		bcopy(&epdesc, (void *)dest, sizeof (epdesc));
15060 		dest += offsetof(dtrace_eprobedesc_t, dtepd_rec[0]);
15061 
15062 		for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
15063 			if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
15064 				continue;
15065 
15066 			if (nrecs-- == 0)
15067 				break;
15068 
15069 			bcopy(&act->dta_rec, (void *)dest,
15070 			    sizeof (dtrace_recdesc_t));
15071 			dest += sizeof (dtrace_recdesc_t);
15072 		}
15073 
15074 		mutex_exit(&dtrace_lock);
15075 
15076 		if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
15077 			kmem_free(buf, size);
15078 			return (EFAULT);
15079 		}
15080 
15081 		kmem_free(buf, size);
15082 		return (0);
15083 	}
15084 
15085 	case DTRACEIOC_AGGDESC: {
15086 		dtrace_aggdesc_t aggdesc;
15087 		dtrace_action_t *act;
15088 		dtrace_aggregation_t *agg;
15089 		int nrecs;
15090 		uint32_t offs;
15091 		dtrace_recdesc_t *lrec;
15092 		void *buf;
15093 		size_t size;
15094 		uintptr_t dest;
15095 
15096 		if (copyin((void *)arg, &aggdesc, sizeof (aggdesc)) != 0)
15097 			return (EFAULT);
15098 
15099 		mutex_enter(&dtrace_lock);
15100 
15101 		if ((agg = dtrace_aggid2agg(state, aggdesc.dtagd_id)) == NULL) {
15102 			mutex_exit(&dtrace_lock);
15103 			return (EINVAL);
15104 		}
15105 
15106 		aggdesc.dtagd_epid = agg->dtag_ecb->dte_epid;
15107 
15108 		nrecs = aggdesc.dtagd_nrecs;
15109 		aggdesc.dtagd_nrecs = 0;
15110 
15111 		offs = agg->dtag_base;
15112 		lrec = &agg->dtag_action.dta_rec;
15113 		aggdesc.dtagd_size = lrec->dtrd_offset + lrec->dtrd_size - offs;
15114 
15115 		for (act = agg->dtag_first; ; act = act->dta_next) {
15116 			ASSERT(act->dta_intuple ||
15117 			    DTRACEACT_ISAGG(act->dta_kind));
15118 
15119 			/*
15120 			 * If this action has a record size of zero, it
15121 			 * denotes an argument to the aggregating action.
15122 			 * Because the presence of this record doesn't (or
15123 			 * shouldn't) affect the way the data is interpreted,
15124 			 * we don't copy it out to save user-level the
15125 			 * confusion of dealing with a zero-length record.
15126 			 */
15127 			if (act->dta_rec.dtrd_size == 0) {
15128 				ASSERT(agg->dtag_hasarg);
15129 				continue;
15130 			}
15131 
15132 			aggdesc.dtagd_nrecs++;
15133 
15134 			if (act == &agg->dtag_action)
15135 				break;
15136 		}
15137 
15138 		/*
15139 		 * Now that we have the size, we need to allocate a temporary
15140 		 * buffer in which to store the complete description.  We need
15141 		 * the temporary buffer to be able to drop dtrace_lock()
15142 		 * across the copyout(), below.
15143 		 */
15144 		size = sizeof (dtrace_aggdesc_t) +
15145 		    (aggdesc.dtagd_nrecs * sizeof (dtrace_recdesc_t));
15146 
15147 		buf = kmem_alloc(size, KM_SLEEP);
15148 		dest = (uintptr_t)buf;
15149 
15150 		bcopy(&aggdesc, (void *)dest, sizeof (aggdesc));
15151 		dest += offsetof(dtrace_aggdesc_t, dtagd_rec[0]);
15152 
15153 		for (act = agg->dtag_first; ; act = act->dta_next) {
15154 			dtrace_recdesc_t rec = act->dta_rec;
15155 
15156 			/*
15157 			 * See the comment in the above loop for why we pass
15158 			 * over zero-length records.
15159 			 */
15160 			if (rec.dtrd_size == 0) {
15161 				ASSERT(agg->dtag_hasarg);
15162 				continue;
15163 			}
15164 
15165 			if (nrecs-- == 0)
15166 				break;
15167 
15168 			rec.dtrd_offset -= offs;
15169 			bcopy(&rec, (void *)dest, sizeof (rec));
15170 			dest += sizeof (dtrace_recdesc_t);
15171 
15172 			if (act == &agg->dtag_action)
15173 				break;
15174 		}
15175 
15176 		mutex_exit(&dtrace_lock);
15177 
15178 		if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
15179 			kmem_free(buf, size);
15180 			return (EFAULT);
15181 		}
15182 
15183 		kmem_free(buf, size);
15184 		return (0);
15185 	}
15186 
15187 	case DTRACEIOC_ENABLE: {
15188 		dof_hdr_t *dof;
15189 		dtrace_enabling_t *enab = NULL;
15190 		dtrace_vstate_t *vstate;
15191 		int err = 0;
15192 
15193 		*rv = 0;
15194 
15195 		/*
15196 		 * If a NULL argument has been passed, we take this as our
15197 		 * cue to reevaluate our enablings.
15198 		 */
15199 		if (arg == NULL) {
15200 			dtrace_enabling_matchall();
15201 
15202 			return (0);
15203 		}
15204 
15205 		if ((dof = dtrace_dof_copyin(arg, &rval)) == NULL)
15206 			return (rval);
15207 
15208 		mutex_enter(&cpu_lock);
15209 		mutex_enter(&dtrace_lock);
15210 		vstate = &state->dts_vstate;
15211 
15212 		if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
15213 			mutex_exit(&dtrace_lock);
15214 			mutex_exit(&cpu_lock);
15215 			dtrace_dof_destroy(dof);
15216 			return (EBUSY);
15217 		}
15218 
15219 		if (dtrace_dof_slurp(dof, vstate, cr, &enab, 0, B_TRUE) != 0) {
15220 			mutex_exit(&dtrace_lock);
15221 			mutex_exit(&cpu_lock);
15222 			dtrace_dof_destroy(dof);
15223 			return (EINVAL);
15224 		}
15225 
15226 		if ((rval = dtrace_dof_options(dof, state)) != 0) {
15227 			dtrace_enabling_destroy(enab);
15228 			mutex_exit(&dtrace_lock);
15229 			mutex_exit(&cpu_lock);
15230 			dtrace_dof_destroy(dof);
15231 			return (rval);
15232 		}
15233 
15234 		if ((err = dtrace_enabling_match(enab, rv)) == 0) {
15235 			err = dtrace_enabling_retain(enab);
15236 		} else {
15237 			dtrace_enabling_destroy(enab);
15238 		}
15239 
15240 		mutex_exit(&cpu_lock);
15241 		mutex_exit(&dtrace_lock);
15242 		dtrace_dof_destroy(dof);
15243 
15244 		return (err);
15245 	}
15246 
15247 	case DTRACEIOC_REPLICATE: {
15248 		dtrace_repldesc_t desc;
15249 		dtrace_probedesc_t *match = &desc.dtrpd_match;
15250 		dtrace_probedesc_t *create = &desc.dtrpd_create;
15251 		int err;
15252 
15253 		if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
15254 			return (EFAULT);
15255 
15256 		match->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
15257 		match->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
15258 		match->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
15259 		match->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
15260 
15261 		create->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
15262 		create->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
15263 		create->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
15264 		create->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
15265 
15266 		mutex_enter(&dtrace_lock);
15267 		err = dtrace_enabling_replicate(state, match, create);
15268 		mutex_exit(&dtrace_lock);
15269 
15270 		return (err);
15271 	}
15272 
15273 	case DTRACEIOC_PROBEMATCH:
15274 	case DTRACEIOC_PROBES: {
15275 		dtrace_probe_t *probe = NULL;
15276 		dtrace_probedesc_t desc;
15277 		dtrace_probekey_t pkey;
15278 		dtrace_id_t i;
15279 		int m = 0;
15280 		uint32_t priv;
15281 		uid_t uid;
15282 		zoneid_t zoneid;
15283 
15284 		if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
15285 			return (EFAULT);
15286 
15287 		desc.dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
15288 		desc.dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
15289 		desc.dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
15290 		desc.dtpd_name[DTRACE_NAMELEN - 1] = '\0';
15291 
15292 		/*
15293 		 * Before we attempt to match this probe, we want to give
15294 		 * all providers the opportunity to provide it.
15295 		 */
15296 		if (desc.dtpd_id == DTRACE_IDNONE) {
15297 			mutex_enter(&dtrace_provider_lock);
15298 			dtrace_probe_provide(&desc, NULL);
15299 			mutex_exit(&dtrace_provider_lock);
15300 			desc.dtpd_id++;
15301 		}
15302 
15303 		if (cmd == DTRACEIOC_PROBEMATCH)  {
15304 			dtrace_probekey(&desc, &pkey);
15305 			pkey.dtpk_id = DTRACE_IDNONE;
15306 		}
15307 
15308 		dtrace_cred2priv(cr, &priv, &uid, &zoneid);
15309 
15310 		mutex_enter(&dtrace_lock);
15311 
15312 		if (cmd == DTRACEIOC_PROBEMATCH) {
15313 			for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
15314 				if ((probe = dtrace_probes[i - 1]) != NULL &&
15315 				    (m = dtrace_match_probe(probe, &pkey,
15316 				    priv, uid, zoneid)) != 0)
15317 					break;
15318 			}
15319 
15320 			if (m < 0) {
15321 				mutex_exit(&dtrace_lock);
15322 				return (EINVAL);
15323 			}
15324 
15325 		} else {
15326 			for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
15327 				if ((probe = dtrace_probes[i - 1]) != NULL &&
15328 				    dtrace_match_priv(probe, priv, uid, zoneid))
15329 					break;
15330 			}
15331 		}
15332 
15333 		if (probe == NULL) {
15334 			mutex_exit(&dtrace_lock);
15335 			return (ESRCH);
15336 		}
15337 
15338 		dtrace_probe_description(probe, &desc);
15339 		mutex_exit(&dtrace_lock);
15340 
15341 		if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
15342 			return (EFAULT);
15343 
15344 		return (0);
15345 	}
15346 
15347 	case DTRACEIOC_PROBEARG: {
15348 		dtrace_argdesc_t desc;
15349 		dtrace_probe_t *probe;
15350 		dtrace_provider_t *prov;
15351 
15352 		if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
15353 			return (EFAULT);
15354 
15355 		if (desc.dtargd_id == DTRACE_IDNONE)
15356 			return (EINVAL);
15357 
15358 		if (desc.dtargd_ndx == DTRACE_ARGNONE)
15359 			return (EINVAL);
15360 
15361 		mutex_enter(&dtrace_provider_lock);
15362 		mutex_enter(&mod_lock);
15363 		mutex_enter(&dtrace_lock);
15364 
15365 		if (desc.dtargd_id > dtrace_nprobes) {
15366 			mutex_exit(&dtrace_lock);
15367 			mutex_exit(&mod_lock);
15368 			mutex_exit(&dtrace_provider_lock);
15369 			return (EINVAL);
15370 		}
15371 
15372 		if ((probe = dtrace_probes[desc.dtargd_id - 1]) == NULL) {
15373 			mutex_exit(&dtrace_lock);
15374 			mutex_exit(&mod_lock);
15375 			mutex_exit(&dtrace_provider_lock);
15376 			return (EINVAL);
15377 		}
15378 
15379 		mutex_exit(&dtrace_lock);
15380 
15381 		prov = probe->dtpr_provider;
15382 
15383 		if (prov->dtpv_pops.dtps_getargdesc == NULL) {
15384 			/*
15385 			 * There isn't any typed information for this probe.
15386 			 * Set the argument number to DTRACE_ARGNONE.
15387 			 */
15388 			desc.dtargd_ndx = DTRACE_ARGNONE;
15389 		} else {
15390 			desc.dtargd_native[0] = '\0';
15391 			desc.dtargd_xlate[0] = '\0';
15392 			desc.dtargd_mapping = desc.dtargd_ndx;
15393 
15394 			prov->dtpv_pops.dtps_getargdesc(prov->dtpv_arg,
15395 			    probe->dtpr_id, probe->dtpr_arg, &desc);
15396 		}
15397 
15398 		mutex_exit(&mod_lock);
15399 		mutex_exit(&dtrace_provider_lock);
15400 
15401 		if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
15402 			return (EFAULT);
15403 
15404 		return (0);
15405 	}
15406 
15407 	case DTRACEIOC_GO: {
15408 		processorid_t cpuid;
15409 		rval = dtrace_state_go(state, &cpuid);
15410 
15411 		if (rval != 0)
15412 			return (rval);
15413 
15414 		if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
15415 			return (EFAULT);
15416 
15417 		return (0);
15418 	}
15419 
15420 	case DTRACEIOC_STOP: {
15421 		processorid_t cpuid;
15422 
15423 		mutex_enter(&dtrace_lock);
15424 		rval = dtrace_state_stop(state, &cpuid);
15425 		mutex_exit(&dtrace_lock);
15426 
15427 		if (rval != 0)
15428 			return (rval);
15429 
15430 		if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
15431 			return (EFAULT);
15432 
15433 		return (0);
15434 	}
15435 
15436 	case DTRACEIOC_DOFGET: {
15437 		dof_hdr_t hdr, *dof;
15438 		uint64_t len;
15439 
15440 		if (copyin((void *)arg, &hdr, sizeof (hdr)) != 0)
15441 			return (EFAULT);
15442 
15443 		mutex_enter(&dtrace_lock);
15444 		dof = dtrace_dof_create(state);
15445 		mutex_exit(&dtrace_lock);
15446 
15447 		len = MIN(hdr.dofh_loadsz, dof->dofh_loadsz);
15448 		rval = copyout(dof, (void *)arg, len);
15449 		dtrace_dof_destroy(dof);
15450 
15451 		return (rval == 0 ? 0 : EFAULT);
15452 	}
15453 
15454 	case DTRACEIOC_AGGSNAP:
15455 	case DTRACEIOC_BUFSNAP: {
15456 		dtrace_bufdesc_t desc;
15457 		caddr_t cached;
15458 		dtrace_buffer_t *buf;
15459 
15460 		if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
15461 			return (EFAULT);
15462 
15463 		if (desc.dtbd_cpu < 0 || desc.dtbd_cpu >= NCPU)
15464 			return (EINVAL);
15465 
15466 		mutex_enter(&dtrace_lock);
15467 
15468 		if (cmd == DTRACEIOC_BUFSNAP) {
15469 			buf = &state->dts_buffer[desc.dtbd_cpu];
15470 		} else {
15471 			buf = &state->dts_aggbuffer[desc.dtbd_cpu];
15472 		}
15473 
15474 		if (buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL)) {
15475 			size_t sz = buf->dtb_offset;
15476 
15477 			if (state->dts_activity != DTRACE_ACTIVITY_STOPPED) {
15478 				mutex_exit(&dtrace_lock);
15479 				return (EBUSY);
15480 			}
15481 
15482 			/*
15483 			 * If this buffer has already been consumed, we're
15484 			 * going to indicate that there's nothing left here
15485 			 * to consume.
15486 			 */
15487 			if (buf->dtb_flags & DTRACEBUF_CONSUMED) {
15488 				mutex_exit(&dtrace_lock);
15489 
15490 				desc.dtbd_size = 0;
15491 				desc.dtbd_drops = 0;
15492 				desc.dtbd_errors = 0;
15493 				desc.dtbd_oldest = 0;
15494 				sz = sizeof (desc);
15495 
15496 				if (copyout(&desc, (void *)arg, sz) != 0)
15497 					return (EFAULT);
15498 
15499 				return (0);
15500 			}
15501 
15502 			/*
15503 			 * If this is a ring buffer that has wrapped, we want
15504 			 * to copy the whole thing out.
15505 			 */
15506 			if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
15507 				dtrace_buffer_polish(buf);
15508 				sz = buf->dtb_size;
15509 			}
15510 
15511 			if (copyout(buf->dtb_tomax, desc.dtbd_data, sz) != 0) {
15512 				mutex_exit(&dtrace_lock);
15513 				return (EFAULT);
15514 			}
15515 
15516 			desc.dtbd_size = sz;
15517 			desc.dtbd_drops = buf->dtb_drops;
15518 			desc.dtbd_errors = buf->dtb_errors;
15519 			desc.dtbd_oldest = buf->dtb_xamot_offset;
15520 
15521 			mutex_exit(&dtrace_lock);
15522 
15523 			if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
15524 				return (EFAULT);
15525 
15526 			buf->dtb_flags |= DTRACEBUF_CONSUMED;
15527 
15528 			return (0);
15529 		}
15530 
15531 		if (buf->dtb_tomax == NULL) {
15532 			ASSERT(buf->dtb_xamot == NULL);
15533 			mutex_exit(&dtrace_lock);
15534 			return (ENOENT);
15535 		}
15536 
15537 		cached = buf->dtb_tomax;
15538 		ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
15539 
15540 		dtrace_xcall(desc.dtbd_cpu,
15541 		    (dtrace_xcall_t)dtrace_buffer_switch, buf);
15542 
15543 		state->dts_errors += buf->dtb_xamot_errors;
15544 
15545 		/*
15546 		 * If the buffers did not actually switch, then the cross call
15547 		 * did not take place -- presumably because the given CPU is
15548 		 * not in the ready set.  If this is the case, we'll return
15549 		 * ENOENT.
15550 		 */
15551 		if (buf->dtb_tomax == cached) {
15552 			ASSERT(buf->dtb_xamot != cached);
15553 			mutex_exit(&dtrace_lock);
15554 			return (ENOENT);
15555 		}
15556 
15557 		ASSERT(cached == buf->dtb_xamot);
15558 
15559 		/*
15560 		 * We have our snapshot; now copy it out.
15561 		 */
15562 		if (copyout(buf->dtb_xamot, desc.dtbd_data,
15563 		    buf->dtb_xamot_offset) != 0) {
15564 			mutex_exit(&dtrace_lock);
15565 			return (EFAULT);
15566 		}
15567 
15568 		desc.dtbd_size = buf->dtb_xamot_offset;
15569 		desc.dtbd_drops = buf->dtb_xamot_drops;
15570 		desc.dtbd_errors = buf->dtb_xamot_errors;
15571 		desc.dtbd_oldest = 0;
15572 
15573 		mutex_exit(&dtrace_lock);
15574 
15575 		/*
15576 		 * Finally, copy out the buffer description.
15577 		 */
15578 		if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
15579 			return (EFAULT);
15580 
15581 		return (0);
15582 	}
15583 
15584 	case DTRACEIOC_CONF: {
15585 		dtrace_conf_t conf;
15586 
15587 		bzero(&conf, sizeof (conf));
15588 		conf.dtc_difversion = DIF_VERSION;
15589 		conf.dtc_difintregs = DIF_DIR_NREGS;
15590 		conf.dtc_diftupregs = DIF_DTR_NREGS;
15591 		conf.dtc_ctfmodel = CTF_MODEL_NATIVE;
15592 
15593 		if (copyout(&conf, (void *)arg, sizeof (conf)) != 0)
15594 			return (EFAULT);
15595 
15596 		return (0);
15597 	}
15598 
15599 	case DTRACEIOC_STATUS: {
15600 		dtrace_status_t stat;
15601 		dtrace_dstate_t *dstate;
15602 		int i, j;
15603 		uint64_t nerrs;
15604 
15605 		/*
15606 		 * See the comment in dtrace_state_deadman() for the reason
15607 		 * for setting dts_laststatus to INT64_MAX before setting
15608 		 * it to the correct value.
15609 		 */
15610 		state->dts_laststatus = INT64_MAX;
15611 		dtrace_membar_producer();
15612 		state->dts_laststatus = dtrace_gethrtime();
15613 
15614 		bzero(&stat, sizeof (stat));
15615 
15616 		mutex_enter(&dtrace_lock);
15617 
15618 		if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE) {
15619 			mutex_exit(&dtrace_lock);
15620 			return (ENOENT);
15621 		}
15622 
15623 		if (state->dts_activity == DTRACE_ACTIVITY_DRAINING)
15624 			stat.dtst_exiting = 1;
15625 
15626 		nerrs = state->dts_errors;
15627 		dstate = &state->dts_vstate.dtvs_dynvars;
15628 
15629 		for (i = 0; i < NCPU; i++) {
15630 			dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[i];
15631 
15632 			stat.dtst_dyndrops += dcpu->dtdsc_drops;
15633 			stat.dtst_dyndrops_dirty += dcpu->dtdsc_dirty_drops;
15634 			stat.dtst_dyndrops_rinsing += dcpu->dtdsc_rinsing_drops;
15635 
15636 			if (state->dts_buffer[i].dtb_flags & DTRACEBUF_FULL)
15637 				stat.dtst_filled++;
15638 
15639 			nerrs += state->dts_buffer[i].dtb_errors;
15640 
15641 			for (j = 0; j < state->dts_nspeculations; j++) {
15642 				dtrace_speculation_t *spec;
15643 				dtrace_buffer_t *buf;
15644 
15645 				spec = &state->dts_speculations[j];
15646 				buf = &spec->dtsp_buffer[i];
15647 				stat.dtst_specdrops += buf->dtb_xamot_drops;
15648 			}
15649 		}
15650 
15651 		stat.dtst_specdrops_busy = state->dts_speculations_busy;
15652 		stat.dtst_specdrops_unavail = state->dts_speculations_unavail;
15653 		stat.dtst_stkstroverflows = state->dts_stkstroverflows;
15654 		stat.dtst_dblerrors = state->dts_dblerrors;
15655 		stat.dtst_killed =
15656 		    (state->dts_activity == DTRACE_ACTIVITY_KILLED);
15657 		stat.dtst_errors = nerrs;
15658 
15659 		mutex_exit(&dtrace_lock);
15660 
15661 		if (copyout(&stat, (void *)arg, sizeof (stat)) != 0)
15662 			return (EFAULT);
15663 
15664 		return (0);
15665 	}
15666 
15667 	case DTRACEIOC_FORMAT: {
15668 		dtrace_fmtdesc_t fmt;
15669 		char *str;
15670 		int len;
15671 
15672 		if (copyin((void *)arg, &fmt, sizeof (fmt)) != 0)
15673 			return (EFAULT);
15674 
15675 		mutex_enter(&dtrace_lock);
15676 
15677 		if (fmt.dtfd_format == 0 ||
15678 		    fmt.dtfd_format > state->dts_nformats) {
15679 			mutex_exit(&dtrace_lock);
15680 			return (EINVAL);
15681 		}
15682 
15683 		/*
15684 		 * Format strings are allocated contiguously and they are
15685 		 * never freed; if a format index is less than the number
15686 		 * of formats, we can assert that the format map is non-NULL
15687 		 * and that the format for the specified index is non-NULL.
15688 		 */
15689 		ASSERT(state->dts_formats != NULL);
15690 		str = state->dts_formats[fmt.dtfd_format - 1];
15691 		ASSERT(str != NULL);
15692 
15693 		len = strlen(str) + 1;
15694 
15695 		if (len > fmt.dtfd_length) {
15696 			fmt.dtfd_length = len;
15697 
15698 			if (copyout(&fmt, (void *)arg, sizeof (fmt)) != 0) {
15699 				mutex_exit(&dtrace_lock);
15700 				return (EINVAL);
15701 			}
15702 		} else {
15703 			if (copyout(str, fmt.dtfd_string, len) != 0) {
15704 				mutex_exit(&dtrace_lock);
15705 				return (EINVAL);
15706 			}
15707 		}
15708 
15709 		mutex_exit(&dtrace_lock);
15710 		return (0);
15711 	}
15712 
15713 	default:
15714 		break;
15715 	}
15716 
15717 	return (ENOTTY);
15718 }
15719 
15720 /*ARGSUSED*/
15721 static int
15722 dtrace_detach(dev_info_t *dip, ddi_detach_cmd_t cmd)
15723 {
15724 	dtrace_state_t *state;
15725 
15726 	switch (cmd) {
15727 	case DDI_DETACH:
15728 		break;
15729 
15730 	case DDI_SUSPEND:
15731 		return (DDI_SUCCESS);
15732 
15733 	default:
15734 		return (DDI_FAILURE);
15735 	}
15736 
15737 	mutex_enter(&cpu_lock);
15738 	mutex_enter(&dtrace_provider_lock);
15739 	mutex_enter(&dtrace_lock);
15740 
15741 	ASSERT(dtrace_opens == 0);
15742 
15743 	if (dtrace_helpers > 0) {
15744 		mutex_exit(&dtrace_provider_lock);
15745 		mutex_exit(&dtrace_lock);
15746 		mutex_exit(&cpu_lock);
15747 		return (DDI_FAILURE);
15748 	}
15749 
15750 	if (dtrace_unregister((dtrace_provider_id_t)dtrace_provider) != 0) {
15751 		mutex_exit(&dtrace_provider_lock);
15752 		mutex_exit(&dtrace_lock);
15753 		mutex_exit(&cpu_lock);
15754 		return (DDI_FAILURE);
15755 	}
15756 
15757 	dtrace_provider = NULL;
15758 
15759 	if ((state = dtrace_anon_grab()) != NULL) {
15760 		/*
15761 		 * If there were ECBs on this state, the provider should
15762 		 * have not been allowed to detach; assert that there is
15763 		 * none.
15764 		 */
15765 		ASSERT(state->dts_necbs == 0);
15766 		dtrace_state_destroy(state);
15767 
15768 		/*
15769 		 * If we're being detached with anonymous state, we need to
15770 		 * indicate to the kernel debugger that DTrace is now inactive.
15771 		 */
15772 		(void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
15773 	}
15774 
15775 	bzero(&dtrace_anon, sizeof (dtrace_anon_t));
15776 	unregister_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
15777 	dtrace_cpu_init = NULL;
15778 	dtrace_helpers_cleanup = NULL;
15779 	dtrace_helpers_fork = NULL;
15780 	dtrace_cpustart_init = NULL;
15781 	dtrace_cpustart_fini = NULL;
15782 	dtrace_debugger_init = NULL;
15783 	dtrace_debugger_fini = NULL;
15784 	dtrace_modload = NULL;
15785 	dtrace_modunload = NULL;
15786 
15787 	mutex_exit(&cpu_lock);
15788 
15789 	if (dtrace_helptrace_enabled) {
15790 		kmem_free(dtrace_helptrace_buffer, dtrace_helptrace_bufsize);
15791 		dtrace_helptrace_buffer = NULL;
15792 	}
15793 
15794 	kmem_free(dtrace_probes, dtrace_nprobes * sizeof (dtrace_probe_t *));
15795 	dtrace_probes = NULL;
15796 	dtrace_nprobes = 0;
15797 
15798 	dtrace_hash_destroy(dtrace_bymod);
15799 	dtrace_hash_destroy(dtrace_byfunc);
15800 	dtrace_hash_destroy(dtrace_byname);
15801 	dtrace_bymod = NULL;
15802 	dtrace_byfunc = NULL;
15803 	dtrace_byname = NULL;
15804 
15805 	kmem_cache_destroy(dtrace_state_cache);
15806 	vmem_destroy(dtrace_minor);
15807 	vmem_destroy(dtrace_arena);
15808 
15809 	if (dtrace_toxrange != NULL) {
15810 		kmem_free(dtrace_toxrange,
15811 		    dtrace_toxranges_max * sizeof (dtrace_toxrange_t));
15812 		dtrace_toxrange = NULL;
15813 		dtrace_toxranges = 0;
15814 		dtrace_toxranges_max = 0;
15815 	}
15816 
15817 	ddi_remove_minor_node(dtrace_devi, NULL);
15818 	dtrace_devi = NULL;
15819 
15820 	ddi_soft_state_fini(&dtrace_softstate);
15821 
15822 	ASSERT(dtrace_vtime_references == 0);
15823 	ASSERT(dtrace_opens == 0);
15824 	ASSERT(dtrace_retained == NULL);
15825 
15826 	mutex_exit(&dtrace_lock);
15827 	mutex_exit(&dtrace_provider_lock);
15828 
15829 	/*
15830 	 * We don't destroy the task queue until after we have dropped our
15831 	 * locks (taskq_destroy() may block on running tasks).  To prevent
15832 	 * attempting to do work after we have effectively detached but before
15833 	 * the task queue has been destroyed, all tasks dispatched via the
15834 	 * task queue must check that DTrace is still attached before
15835 	 * performing any operation.
15836 	 */
15837 	taskq_destroy(dtrace_taskq);
15838 	dtrace_taskq = NULL;
15839 
15840 	return (DDI_SUCCESS);
15841 }
15842 
15843 /*ARGSUSED*/
15844 static int
15845 dtrace_info(dev_info_t *dip, ddi_info_cmd_t infocmd, void *arg, void **result)
15846 {
15847 	int error;
15848 
15849 	switch (infocmd) {
15850 	case DDI_INFO_DEVT2DEVINFO:
15851 		*result = (void *)dtrace_devi;
15852 		error = DDI_SUCCESS;
15853 		break;
15854 	case DDI_INFO_DEVT2INSTANCE:
15855 		*result = (void *)0;
15856 		error = DDI_SUCCESS;
15857 		break;
15858 	default:
15859 		error = DDI_FAILURE;
15860 	}
15861 	return (error);
15862 }
15863 
15864 static struct cb_ops dtrace_cb_ops = {
15865 	dtrace_open,		/* open */
15866 	dtrace_close,		/* close */
15867 	nulldev,		/* strategy */
15868 	nulldev,		/* print */
15869 	nodev,			/* dump */
15870 	nodev,			/* read */
15871 	nodev,			/* write */
15872 	dtrace_ioctl,		/* ioctl */
15873 	nodev,			/* devmap */
15874 	nodev,			/* mmap */
15875 	nodev,			/* segmap */
15876 	nochpoll,		/* poll */
15877 	ddi_prop_op,		/* cb_prop_op */
15878 	0,			/* streamtab  */
15879 	D_NEW | D_MP		/* Driver compatibility flag */
15880 };
15881 
15882 static struct dev_ops dtrace_ops = {
15883 	DEVO_REV,		/* devo_rev */
15884 	0,			/* refcnt */
15885 	dtrace_info,		/* get_dev_info */
15886 	nulldev,		/* identify */
15887 	nulldev,		/* probe */
15888 	dtrace_attach,		/* attach */
15889 	dtrace_detach,		/* detach */
15890 	nodev,			/* reset */
15891 	&dtrace_cb_ops,		/* driver operations */
15892 	NULL,			/* bus operations */
15893 	nodev,			/* dev power */
15894 	ddi_quiesce_not_needed,		/* quiesce */
15895 };
15896 
15897 static struct modldrv modldrv = {
15898 	&mod_driverops,		/* module type (this is a pseudo driver) */
15899 	"Dynamic Tracing",	/* name of module */
15900 	&dtrace_ops,		/* driver ops */
15901 };
15902 
15903 static struct modlinkage modlinkage = {
15904 	MODREV_1,
15905 	(void *)&modldrv,
15906 	NULL
15907 };
15908 
15909 int
15910 _init(void)
15911 {
15912 	return (mod_install(&modlinkage));
15913 }
15914 
15915 int
15916 _info(struct modinfo *modinfop)
15917 {
15918 	return (mod_info(&modlinkage, modinfop));
15919 }
15920 
15921 int
15922 _fini(void)
15923 {
15924 	return (mod_remove(&modlinkage));
15925 }
15926